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/EHPersonalities.h" 30 #include "llvm/Analysis/GuardUtils.h" 31 #include "llvm/Analysis/InstructionSimplify.h" 32 #include "llvm/Analysis/Loads.h" 33 #include "llvm/Analysis/LoopInfo.h" 34 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 35 #include "llvm/Analysis/TargetLibraryInfo.h" 36 #include "llvm/IR/Argument.h" 37 #include "llvm/IR/Attributes.h" 38 #include "llvm/IR/BasicBlock.h" 39 #include "llvm/IR/Constant.h" 40 #include "llvm/IR/ConstantRange.h" 41 #include "llvm/IR/Constants.h" 42 #include "llvm/IR/DerivedTypes.h" 43 #include "llvm/IR/DiagnosticInfo.h" 44 #include "llvm/IR/Dominators.h" 45 #include "llvm/IR/Function.h" 46 #include "llvm/IR/GetElementPtrTypeIterator.h" 47 #include "llvm/IR/GlobalAlias.h" 48 #include "llvm/IR/GlobalValue.h" 49 #include "llvm/IR/GlobalVariable.h" 50 #include "llvm/IR/InstrTypes.h" 51 #include "llvm/IR/Instruction.h" 52 #include "llvm/IR/Instructions.h" 53 #include "llvm/IR/IntrinsicInst.h" 54 #include "llvm/IR/Intrinsics.h" 55 #include "llvm/IR/IntrinsicsAArch64.h" 56 #include "llvm/IR/IntrinsicsRISCV.h" 57 #include "llvm/IR/IntrinsicsX86.h" 58 #include "llvm/IR/LLVMContext.h" 59 #include "llvm/IR/Metadata.h" 60 #include "llvm/IR/Module.h" 61 #include "llvm/IR/Operator.h" 62 #include "llvm/IR/PatternMatch.h" 63 #include "llvm/IR/Type.h" 64 #include "llvm/IR/User.h" 65 #include "llvm/IR/Value.h" 66 #include "llvm/Support/Casting.h" 67 #include "llvm/Support/CommandLine.h" 68 #include "llvm/Support/Compiler.h" 69 #include "llvm/Support/ErrorHandling.h" 70 #include "llvm/Support/KnownBits.h" 71 #include "llvm/Support/MathExtras.h" 72 #include <algorithm> 73 #include <array> 74 #include <cassert> 75 #include <cstdint> 76 #include <iterator> 77 #include <utility> 78 79 using namespace llvm; 80 using namespace llvm::PatternMatch; 81 82 // Controls the number of uses of the value searched for possible 83 // dominating comparisons. 84 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses", 85 cl::Hidden, cl::init(20)); 86 87 /// Returns the bitwidth of the given scalar or pointer type. For vector types, 88 /// returns the element type's bitwidth. 89 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { 90 if (unsigned BitWidth = Ty->getScalarSizeInBits()) 91 return BitWidth; 92 93 return DL.getPointerTypeSizeInBits(Ty); 94 } 95 96 namespace { 97 98 // Simplifying using an assume can only be done in a particular control-flow 99 // context (the context instruction provides that context). If an assume and 100 // the context instruction are not in the same block then the DT helps in 101 // figuring out if we can use it. 102 struct Query { 103 const DataLayout &DL; 104 AssumptionCache *AC; 105 const Instruction *CxtI; 106 const DominatorTree *DT; 107 108 // Unlike the other analyses, this may be a nullptr because not all clients 109 // provide it currently. 110 OptimizationRemarkEmitter *ORE; 111 112 /// If true, it is safe to use metadata during simplification. 113 InstrInfoQuery IIQ; 114 115 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI, 116 const DominatorTree *DT, bool UseInstrInfo, 117 OptimizationRemarkEmitter *ORE = nullptr) 118 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {} 119 }; 120 121 } // end anonymous namespace 122 123 // Given the provided Value and, potentially, a context instruction, return 124 // the preferred context instruction (if any). 125 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { 126 // If we've been provided with a context instruction, then use that (provided 127 // it has been inserted). 128 if (CxtI && CxtI->getParent()) 129 return CxtI; 130 131 // If the value is really an already-inserted instruction, then use that. 132 CxtI = dyn_cast<Instruction>(V); 133 if (CxtI && CxtI->getParent()) 134 return CxtI; 135 136 return nullptr; 137 } 138 139 static const Instruction *safeCxtI(const Value *V1, const Value *V2, const Instruction *CxtI) { 140 // If we've been provided with a context instruction, then use that (provided 141 // it has been inserted). 142 if (CxtI && CxtI->getParent()) 143 return CxtI; 144 145 // If the value is really an already-inserted instruction, then use that. 146 CxtI = dyn_cast<Instruction>(V1); 147 if (CxtI && CxtI->getParent()) 148 return CxtI; 149 150 CxtI = dyn_cast<Instruction>(V2); 151 if (CxtI && CxtI->getParent()) 152 return CxtI; 153 154 return nullptr; 155 } 156 157 static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf, 158 const APInt &DemandedElts, 159 APInt &DemandedLHS, APInt &DemandedRHS) { 160 // The length of scalable vectors is unknown at compile time, thus we 161 // cannot check their values 162 if (isa<ScalableVectorType>(Shuf->getType())) 163 return false; 164 165 int NumElts = 166 cast<FixedVectorType>(Shuf->getOperand(0)->getType())->getNumElements(); 167 int NumMaskElts = cast<FixedVectorType>(Shuf->getType())->getNumElements(); 168 DemandedLHS = DemandedRHS = APInt::getNullValue(NumElts); 169 if (DemandedElts.isNullValue()) 170 return true; 171 // Simple case of a shuffle with zeroinitializer. 172 if (all_of(Shuf->getShuffleMask(), [](int Elt) { return Elt == 0; })) { 173 DemandedLHS.setBit(0); 174 return true; 175 } 176 for (int i = 0; i != NumMaskElts; ++i) { 177 if (!DemandedElts[i]) 178 continue; 179 int M = Shuf->getMaskValue(i); 180 assert(M < (NumElts * 2) && "Invalid shuffle mask constant"); 181 182 // For undef elements, we don't know anything about the common state of 183 // the shuffle result. 184 if (M == -1) 185 return false; 186 if (M < NumElts) 187 DemandedLHS.setBit(M % NumElts); 188 else 189 DemandedRHS.setBit(M % NumElts); 190 } 191 192 return true; 193 } 194 195 static void computeKnownBits(const Value *V, const APInt &DemandedElts, 196 KnownBits &Known, unsigned Depth, const Query &Q); 197 198 static void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, 199 const Query &Q) { 200 // FIXME: We currently have no way to represent the DemandedElts of a scalable 201 // vector 202 if (isa<ScalableVectorType>(V->getType())) { 203 Known.resetAll(); 204 return; 205 } 206 207 auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); 208 APInt DemandedElts = 209 FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); 210 computeKnownBits(V, DemandedElts, Known, Depth, Q); 211 } 212 213 void llvm::computeKnownBits(const Value *V, KnownBits &Known, 214 const DataLayout &DL, unsigned Depth, 215 AssumptionCache *AC, const Instruction *CxtI, 216 const DominatorTree *DT, 217 OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { 218 ::computeKnownBits(V, Known, Depth, 219 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 220 } 221 222 void llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, 223 KnownBits &Known, const DataLayout &DL, 224 unsigned Depth, AssumptionCache *AC, 225 const Instruction *CxtI, const DominatorTree *DT, 226 OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { 227 ::computeKnownBits(V, DemandedElts, Known, Depth, 228 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 229 } 230 231 static KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, 232 unsigned Depth, const Query &Q); 233 234 static KnownBits computeKnownBits(const Value *V, unsigned Depth, 235 const Query &Q); 236 237 KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL, 238 unsigned Depth, AssumptionCache *AC, 239 const Instruction *CxtI, 240 const DominatorTree *DT, 241 OptimizationRemarkEmitter *ORE, 242 bool UseInstrInfo) { 243 return ::computeKnownBits( 244 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 245 } 246 247 KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, 248 const DataLayout &DL, unsigned Depth, 249 AssumptionCache *AC, const Instruction *CxtI, 250 const DominatorTree *DT, 251 OptimizationRemarkEmitter *ORE, 252 bool UseInstrInfo) { 253 return ::computeKnownBits( 254 V, DemandedElts, Depth, 255 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 256 } 257 258 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS, 259 const DataLayout &DL, AssumptionCache *AC, 260 const Instruction *CxtI, const DominatorTree *DT, 261 bool UseInstrInfo) { 262 assert(LHS->getType() == RHS->getType() && 263 "LHS and RHS should have the same type"); 264 assert(LHS->getType()->isIntOrIntVectorTy() && 265 "LHS and RHS should be integers"); 266 // Look for an inverted mask: (X & ~M) op (Y & M). 267 Value *M; 268 if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) && 269 match(RHS, m_c_And(m_Specific(M), m_Value()))) 270 return true; 271 if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) && 272 match(LHS, m_c_And(m_Specific(M), m_Value()))) 273 return true; 274 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType()); 275 KnownBits LHSKnown(IT->getBitWidth()); 276 KnownBits RHSKnown(IT->getBitWidth()); 277 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); 278 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); 279 return KnownBits::haveNoCommonBitsSet(LHSKnown, RHSKnown); 280 } 281 282 bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI) { 283 for (const User *U : CxtI->users()) { 284 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U)) 285 if (IC->isEquality()) 286 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1))) 287 if (C->isNullValue()) 288 continue; 289 return false; 290 } 291 return true; 292 } 293 294 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, 295 const Query &Q); 296 297 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, 298 bool OrZero, unsigned Depth, 299 AssumptionCache *AC, const Instruction *CxtI, 300 const DominatorTree *DT, bool UseInstrInfo) { 301 return ::isKnownToBeAPowerOfTwo( 302 V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 303 } 304 305 static bool isKnownNonZero(const Value *V, const APInt &DemandedElts, 306 unsigned Depth, const Query &Q); 307 308 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q); 309 310 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth, 311 AssumptionCache *AC, const Instruction *CxtI, 312 const DominatorTree *DT, bool UseInstrInfo) { 313 return ::isKnownNonZero(V, Depth, 314 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 315 } 316 317 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL, 318 unsigned Depth, AssumptionCache *AC, 319 const Instruction *CxtI, const DominatorTree *DT, 320 bool UseInstrInfo) { 321 KnownBits Known = 322 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); 323 return Known.isNonNegative(); 324 } 325 326 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth, 327 AssumptionCache *AC, const Instruction *CxtI, 328 const DominatorTree *DT, bool UseInstrInfo) { 329 if (auto *CI = dyn_cast<ConstantInt>(V)) 330 return CI->getValue().isStrictlyPositive(); 331 332 // TODO: We'd doing two recursive queries here. We should factor this such 333 // that only a single query is needed. 334 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) && 335 isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo); 336 } 337 338 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth, 339 AssumptionCache *AC, const Instruction *CxtI, 340 const DominatorTree *DT, bool UseInstrInfo) { 341 KnownBits Known = 342 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); 343 return Known.isNegative(); 344 } 345 346 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, 347 const Query &Q); 348 349 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2, 350 const DataLayout &DL, AssumptionCache *AC, 351 const Instruction *CxtI, const DominatorTree *DT, 352 bool UseInstrInfo) { 353 return ::isKnownNonEqual(V1, V2, 0, 354 Query(DL, AC, safeCxtI(V2, V1, CxtI), DT, 355 UseInstrInfo, /*ORE=*/nullptr)); 356 } 357 358 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, 359 const Query &Q); 360 361 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask, 362 const DataLayout &DL, unsigned Depth, 363 AssumptionCache *AC, const Instruction *CxtI, 364 const DominatorTree *DT, bool UseInstrInfo) { 365 return ::MaskedValueIsZero( 366 V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 367 } 368 369 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, 370 unsigned Depth, const Query &Q); 371 372 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth, 373 const Query &Q) { 374 // FIXME: We currently have no way to represent the DemandedElts of a scalable 375 // vector 376 if (isa<ScalableVectorType>(V->getType())) 377 return 1; 378 379 auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); 380 APInt DemandedElts = 381 FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); 382 return ComputeNumSignBits(V, DemandedElts, Depth, Q); 383 } 384 385 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL, 386 unsigned Depth, AssumptionCache *AC, 387 const Instruction *CxtI, 388 const DominatorTree *DT, bool UseInstrInfo) { 389 return ::ComputeNumSignBits( 390 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 391 } 392 393 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1, 394 bool NSW, const APInt &DemandedElts, 395 KnownBits &KnownOut, KnownBits &Known2, 396 unsigned Depth, const Query &Q) { 397 computeKnownBits(Op1, DemandedElts, KnownOut, Depth + 1, Q); 398 399 // If one operand is unknown and we have no nowrap information, 400 // the result will be unknown independently of the second operand. 401 if (KnownOut.isUnknown() && !NSW) 402 return; 403 404 computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); 405 KnownOut = KnownBits::computeForAddSub(Add, NSW, Known2, KnownOut); 406 } 407 408 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW, 409 const APInt &DemandedElts, KnownBits &Known, 410 KnownBits &Known2, unsigned Depth, 411 const Query &Q) { 412 computeKnownBits(Op1, DemandedElts, Known, Depth + 1, Q); 413 computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); 414 415 bool isKnownNegative = false; 416 bool isKnownNonNegative = false; 417 // If the multiplication is known not to overflow, compute the sign bit. 418 if (NSW) { 419 if (Op0 == Op1) { 420 // The product of a number with itself is non-negative. 421 isKnownNonNegative = true; 422 } else { 423 bool isKnownNonNegativeOp1 = Known.isNonNegative(); 424 bool isKnownNonNegativeOp0 = Known2.isNonNegative(); 425 bool isKnownNegativeOp1 = Known.isNegative(); 426 bool isKnownNegativeOp0 = Known2.isNegative(); 427 // The product of two numbers with the same sign is non-negative. 428 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || 429 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); 430 // The product of a negative number and a non-negative number is either 431 // negative or zero. 432 if (!isKnownNonNegative) 433 isKnownNegative = 434 (isKnownNegativeOp1 && isKnownNonNegativeOp0 && 435 Known2.isNonZero()) || 436 (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero()); 437 } 438 } 439 440 Known = KnownBits::mul(Known, Known2); 441 442 // Only make use of no-wrap flags if we failed to compute the sign bit 443 // directly. This matters if the multiplication always overflows, in 444 // which case we prefer to follow the result of the direct computation, 445 // though as the program is invoking undefined behaviour we can choose 446 // whatever we like here. 447 if (isKnownNonNegative && !Known.isNegative()) 448 Known.makeNonNegative(); 449 else if (isKnownNegative && !Known.isNonNegative()) 450 Known.makeNegative(); 451 } 452 453 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, 454 KnownBits &Known) { 455 unsigned BitWidth = Known.getBitWidth(); 456 unsigned NumRanges = Ranges.getNumOperands() / 2; 457 assert(NumRanges >= 1); 458 459 Known.Zero.setAllBits(); 460 Known.One.setAllBits(); 461 462 for (unsigned i = 0; i < NumRanges; ++i) { 463 ConstantInt *Lower = 464 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); 465 ConstantInt *Upper = 466 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); 467 ConstantRange Range(Lower->getValue(), Upper->getValue()); 468 469 // The first CommonPrefixBits of all values in Range are equal. 470 unsigned CommonPrefixBits = 471 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros(); 472 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits); 473 APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(BitWidth); 474 Known.One &= UnsignedMax & Mask; 475 Known.Zero &= ~UnsignedMax & Mask; 476 } 477 } 478 479 static bool isEphemeralValueOf(const Instruction *I, const Value *E) { 480 SmallVector<const Value *, 16> WorkSet(1, I); 481 SmallPtrSet<const Value *, 32> Visited; 482 SmallPtrSet<const Value *, 16> EphValues; 483 484 // The instruction defining an assumption's condition itself is always 485 // considered ephemeral to that assumption (even if it has other 486 // non-ephemeral users). See r246696's test case for an example. 487 if (is_contained(I->operands(), E)) 488 return true; 489 490 while (!WorkSet.empty()) { 491 const Value *V = WorkSet.pop_back_val(); 492 if (!Visited.insert(V).second) 493 continue; 494 495 // If all uses of this value are ephemeral, then so is this value. 496 if (llvm::all_of(V->users(), [&](const User *U) { 497 return EphValues.count(U); 498 })) { 499 if (V == E) 500 return true; 501 502 if (V == I || isSafeToSpeculativelyExecute(V)) { 503 EphValues.insert(V); 504 if (const User *U = dyn_cast<User>(V)) 505 append_range(WorkSet, U->operands()); 506 } 507 } 508 } 509 510 return false; 511 } 512 513 // Is this an intrinsic that cannot be speculated but also cannot trap? 514 bool llvm::isAssumeLikeIntrinsic(const Instruction *I) { 515 if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(I)) 516 return CI->isAssumeLikeIntrinsic(); 517 518 return false; 519 } 520 521 bool llvm::isValidAssumeForContext(const Instruction *Inv, 522 const Instruction *CxtI, 523 const DominatorTree *DT) { 524 // There are two restrictions on the use of an assume: 525 // 1. The assume must dominate the context (or the control flow must 526 // reach the assume whenever it reaches the context). 527 // 2. The context must not be in the assume's set of ephemeral values 528 // (otherwise we will use the assume to prove that the condition 529 // feeding the assume is trivially true, thus causing the removal of 530 // the assume). 531 532 if (Inv->getParent() == CxtI->getParent()) { 533 // If Inv and CtxI are in the same block, check if the assume (Inv) is first 534 // in the BB. 535 if (Inv->comesBefore(CxtI)) 536 return true; 537 538 // Don't let an assume affect itself - this would cause the problems 539 // `isEphemeralValueOf` is trying to prevent, and it would also make 540 // the loop below go out of bounds. 541 if (Inv == CxtI) 542 return false; 543 544 // The context comes first, but they're both in the same block. 545 // Make sure there is nothing in between that might interrupt 546 // the control flow, not even CxtI itself. 547 // We limit the scan distance between the assume and its context instruction 548 // to avoid a compile-time explosion. This limit is chosen arbitrarily, so 549 // it can be adjusted if needed (could be turned into a cl::opt). 550 unsigned ScanLimit = 15; 551 for (BasicBlock::const_iterator I(CxtI), IE(Inv); I != IE; ++I) 552 if (!isGuaranteedToTransferExecutionToSuccessor(&*I) || --ScanLimit == 0) 553 return false; 554 555 return !isEphemeralValueOf(Inv, CxtI); 556 } 557 558 // Inv and CxtI are in different blocks. 559 if (DT) { 560 if (DT->dominates(Inv, CxtI)) 561 return true; 562 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) { 563 // We don't have a DT, but this trivially dominates. 564 return true; 565 } 566 567 return false; 568 } 569 570 static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) { 571 // v u> y implies v != 0. 572 if (Pred == ICmpInst::ICMP_UGT) 573 return true; 574 575 // Special-case v != 0 to also handle v != null. 576 if (Pred == ICmpInst::ICMP_NE) 577 return match(RHS, m_Zero()); 578 579 // All other predicates - rely on generic ConstantRange handling. 580 const APInt *C; 581 if (!match(RHS, m_APInt(C))) 582 return false; 583 584 ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C); 585 return !TrueValues.contains(APInt::getNullValue(C->getBitWidth())); 586 } 587 588 static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) { 589 // Use of assumptions is context-sensitive. If we don't have a context, we 590 // cannot use them! 591 if (!Q.AC || !Q.CxtI) 592 return false; 593 594 if (Q.CxtI && V->getType()->isPointerTy()) { 595 SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NonNull}; 596 if (!NullPointerIsDefined(Q.CxtI->getFunction(), 597 V->getType()->getPointerAddressSpace())) 598 AttrKinds.push_back(Attribute::Dereferenceable); 599 600 if (getKnowledgeValidInContext(V, AttrKinds, Q.CxtI, Q.DT, Q.AC)) 601 return true; 602 } 603 604 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { 605 if (!AssumeVH) 606 continue; 607 CallInst *I = cast<CallInst>(AssumeVH); 608 assert(I->getFunction() == Q.CxtI->getFunction() && 609 "Got assumption for the wrong function!"); 610 611 // Warning: This loop can end up being somewhat performance sensitive. 612 // We're running this loop for once for each value queried resulting in a 613 // runtime of ~O(#assumes * #values). 614 615 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 616 "must be an assume intrinsic"); 617 618 Value *RHS; 619 CmpInst::Predicate Pred; 620 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); 621 if (!match(I->getArgOperand(0), m_c_ICmp(Pred, m_V, m_Value(RHS)))) 622 return false; 623 624 if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q.CxtI, Q.DT)) 625 return true; 626 } 627 628 return false; 629 } 630 631 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known, 632 unsigned Depth, const Query &Q) { 633 // Use of assumptions is context-sensitive. If we don't have a context, we 634 // cannot use them! 635 if (!Q.AC || !Q.CxtI) 636 return; 637 638 unsigned BitWidth = Known.getBitWidth(); 639 640 // Refine Known set if the pointer alignment is set by assume bundles. 641 if (V->getType()->isPointerTy()) { 642 if (RetainedKnowledge RK = getKnowledgeValidInContext( 643 V, {Attribute::Alignment}, Q.CxtI, Q.DT, Q.AC)) { 644 Known.Zero.setLowBits(Log2_32(RK.ArgValue)); 645 } 646 } 647 648 // Note that the patterns below need to be kept in sync with the code 649 // in AssumptionCache::updateAffectedValues. 650 651 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { 652 if (!AssumeVH) 653 continue; 654 CallInst *I = cast<CallInst>(AssumeVH); 655 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && 656 "Got assumption for the wrong function!"); 657 658 // Warning: This loop can end up being somewhat performance sensitive. 659 // We're running this loop for once for each value queried resulting in a 660 // runtime of ~O(#assumes * #values). 661 662 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 663 "must be an assume intrinsic"); 664 665 Value *Arg = I->getArgOperand(0); 666 667 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 668 assert(BitWidth == 1 && "assume operand is not i1?"); 669 Known.setAllOnes(); 670 return; 671 } 672 if (match(Arg, m_Not(m_Specific(V))) && 673 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 674 assert(BitWidth == 1 && "assume operand is not i1?"); 675 Known.setAllZero(); 676 return; 677 } 678 679 // The remaining tests are all recursive, so bail out if we hit the limit. 680 if (Depth == MaxAnalysisRecursionDepth) 681 continue; 682 683 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); 684 if (!Cmp) 685 continue; 686 687 // We are attempting to compute known bits for the operands of an assume. 688 // Do not try to use other assumptions for those recursive calls because 689 // that can lead to mutual recursion and a compile-time explosion. 690 // An example of the mutual recursion: computeKnownBits can call 691 // isKnownNonZero which calls computeKnownBitsFromAssume (this function) 692 // and so on. 693 Query QueryNoAC = Q; 694 QueryNoAC.AC = nullptr; 695 696 // Note that ptrtoint may change the bitwidth. 697 Value *A, *B; 698 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); 699 700 CmpInst::Predicate Pred; 701 uint64_t C; 702 switch (Cmp->getPredicate()) { 703 default: 704 break; 705 case ICmpInst::ICMP_EQ: 706 // assume(v = a) 707 if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A))) && 708 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 709 KnownBits RHSKnown = 710 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 711 Known.Zero |= RHSKnown.Zero; 712 Known.One |= RHSKnown.One; 713 // assume(v & b = a) 714 } else if (match(Cmp, 715 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) && 716 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 717 KnownBits RHSKnown = 718 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 719 KnownBits MaskKnown = 720 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 721 722 // For those bits in the mask that are known to be one, we can propagate 723 // known bits from the RHS to V. 724 Known.Zero |= RHSKnown.Zero & MaskKnown.One; 725 Known.One |= RHSKnown.One & MaskKnown.One; 726 // assume(~(v & b) = a) 727 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))), 728 m_Value(A))) && 729 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 730 KnownBits RHSKnown = 731 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 732 KnownBits MaskKnown = 733 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 734 735 // For those bits in the mask that are known to be one, we can propagate 736 // inverted known bits from the RHS to V. 737 Known.Zero |= RHSKnown.One & MaskKnown.One; 738 Known.One |= RHSKnown.Zero & MaskKnown.One; 739 // assume(v | b = a) 740 } else if (match(Cmp, 741 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) && 742 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 743 KnownBits RHSKnown = 744 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 745 KnownBits BKnown = 746 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 747 748 // For those bits in B that are known to be zero, we can propagate known 749 // bits from the RHS to V. 750 Known.Zero |= RHSKnown.Zero & BKnown.Zero; 751 Known.One |= RHSKnown.One & BKnown.Zero; 752 // assume(~(v | b) = a) 753 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))), 754 m_Value(A))) && 755 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 756 KnownBits RHSKnown = 757 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 758 KnownBits BKnown = 759 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 760 761 // For those bits in B that are known to be zero, we can propagate 762 // inverted known bits from the RHS to V. 763 Known.Zero |= RHSKnown.One & BKnown.Zero; 764 Known.One |= RHSKnown.Zero & BKnown.Zero; 765 // assume(v ^ b = a) 766 } else if (match(Cmp, 767 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) && 768 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 769 KnownBits RHSKnown = 770 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 771 KnownBits BKnown = 772 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 773 774 // For those bits in B that are known to be zero, we can propagate known 775 // bits from the RHS to V. For those bits in B that are known to be one, 776 // we can propagate inverted known bits from the RHS to V. 777 Known.Zero |= RHSKnown.Zero & BKnown.Zero; 778 Known.One |= RHSKnown.One & BKnown.Zero; 779 Known.Zero |= RHSKnown.One & BKnown.One; 780 Known.One |= RHSKnown.Zero & BKnown.One; 781 // assume(~(v ^ b) = a) 782 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))), 783 m_Value(A))) && 784 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 785 KnownBits RHSKnown = 786 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 787 KnownBits BKnown = 788 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 789 790 // For those bits in B that are known to be zero, we can propagate 791 // inverted known bits from the RHS to V. For those bits in B that are 792 // known to be one, we can propagate known bits from the RHS to V. 793 Known.Zero |= RHSKnown.One & BKnown.Zero; 794 Known.One |= RHSKnown.Zero & BKnown.Zero; 795 Known.Zero |= RHSKnown.Zero & BKnown.One; 796 Known.One |= RHSKnown.One & BKnown.One; 797 // assume(v << c = a) 798 } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)), 799 m_Value(A))) && 800 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 801 KnownBits RHSKnown = 802 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 803 804 // For those bits in RHS that are known, we can propagate them to known 805 // bits in V shifted to the right by C. 806 RHSKnown.Zero.lshrInPlace(C); 807 Known.Zero |= RHSKnown.Zero; 808 RHSKnown.One.lshrInPlace(C); 809 Known.One |= RHSKnown.One; 810 // assume(~(v << c) = a) 811 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))), 812 m_Value(A))) && 813 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 814 KnownBits RHSKnown = 815 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 816 // For those bits in RHS that are known, we can propagate them inverted 817 // to known bits in V shifted to the right by C. 818 RHSKnown.One.lshrInPlace(C); 819 Known.Zero |= RHSKnown.One; 820 RHSKnown.Zero.lshrInPlace(C); 821 Known.One |= RHSKnown.Zero; 822 // assume(v >> c = a) 823 } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)), 824 m_Value(A))) && 825 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 826 KnownBits RHSKnown = 827 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 828 // For those bits in RHS that are known, we can propagate them to known 829 // bits in V shifted to the right by C. 830 Known.Zero |= RHSKnown.Zero << C; 831 Known.One |= RHSKnown.One << C; 832 // assume(~(v >> c) = a) 833 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))), 834 m_Value(A))) && 835 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 836 KnownBits RHSKnown = 837 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 838 // For those bits in RHS that are known, we can propagate them inverted 839 // to known bits in V shifted to the right by C. 840 Known.Zero |= RHSKnown.One << C; 841 Known.One |= RHSKnown.Zero << C; 842 } 843 break; 844 case ICmpInst::ICMP_SGE: 845 // assume(v >=_s c) where c is non-negative 846 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 847 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 848 KnownBits RHSKnown = 849 computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); 850 851 if (RHSKnown.isNonNegative()) { 852 // We know that the sign bit is zero. 853 Known.makeNonNegative(); 854 } 855 } 856 break; 857 case ICmpInst::ICMP_SGT: 858 // assume(v >_s c) where c is at least -1. 859 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 860 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 861 KnownBits RHSKnown = 862 computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); 863 864 if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) { 865 // We know that the sign bit is zero. 866 Known.makeNonNegative(); 867 } 868 } 869 break; 870 case ICmpInst::ICMP_SLE: 871 // assume(v <=_s c) where c is negative 872 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 873 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 874 KnownBits RHSKnown = 875 computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); 876 877 if (RHSKnown.isNegative()) { 878 // We know that the sign bit is one. 879 Known.makeNegative(); 880 } 881 } 882 break; 883 case ICmpInst::ICMP_SLT: 884 // assume(v <_s c) where c is non-positive 885 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 886 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 887 KnownBits RHSKnown = 888 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 889 890 if (RHSKnown.isZero() || RHSKnown.isNegative()) { 891 // We know that the sign bit is one. 892 Known.makeNegative(); 893 } 894 } 895 break; 896 case ICmpInst::ICMP_ULE: 897 // assume(v <=_u c) 898 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 899 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 900 KnownBits RHSKnown = 901 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 902 903 // Whatever high bits in c are zero are known to be zero. 904 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); 905 } 906 break; 907 case ICmpInst::ICMP_ULT: 908 // assume(v <_u c) 909 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 910 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 911 KnownBits RHSKnown = 912 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); 913 914 // If the RHS is known zero, then this assumption must be wrong (nothing 915 // is unsigned less than zero). Signal a conflict and get out of here. 916 if (RHSKnown.isZero()) { 917 Known.Zero.setAllBits(); 918 Known.One.setAllBits(); 919 break; 920 } 921 922 // Whatever high bits in c are zero are known to be zero (if c is a power 923 // of 2, then one more). 924 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, QueryNoAC)) 925 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1); 926 else 927 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); 928 } 929 break; 930 } 931 } 932 933 // If assumptions conflict with each other or previous known bits, then we 934 // have a logical fallacy. It's possible that the assumption is not reachable, 935 // so this isn't a real bug. On the other hand, the program may have undefined 936 // behavior, or we might have a bug in the compiler. We can't assert/crash, so 937 // clear out the known bits, try to warn the user, and hope for the best. 938 if (Known.Zero.intersects(Known.One)) { 939 Known.resetAll(); 940 941 if (Q.ORE) 942 Q.ORE->emit([&]() { 943 auto *CxtI = const_cast<Instruction *>(Q.CxtI); 944 return OptimizationRemarkAnalysis("value-tracking", "BadAssumption", 945 CxtI) 946 << "Detected conflicting code assumptions. Program may " 947 "have undefined behavior, or compiler may have " 948 "internal error."; 949 }); 950 } 951 } 952 953 /// Compute known bits from a shift operator, including those with a 954 /// non-constant shift amount. Known is the output of this function. Known2 is a 955 /// pre-allocated temporary with the same bit width as Known and on return 956 /// contains the known bit of the shift value source. KF is an 957 /// operator-specific function that, given the known-bits and a shift amount, 958 /// compute the implied known-bits of the shift operator's result respectively 959 /// for that shift amount. The results from calling KF are conservatively 960 /// combined for all permitted shift amounts. 961 static void computeKnownBitsFromShiftOperator( 962 const Operator *I, const APInt &DemandedElts, KnownBits &Known, 963 KnownBits &Known2, unsigned Depth, const Query &Q, 964 function_ref<KnownBits(const KnownBits &, const KnownBits &)> KF) { 965 unsigned BitWidth = Known.getBitWidth(); 966 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 967 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 968 969 // Note: We cannot use Known.Zero.getLimitedValue() here, because if 970 // BitWidth > 64 and any upper bits are known, we'll end up returning the 971 // limit value (which implies all bits are known). 972 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue(); 973 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue(); 974 bool ShiftAmtIsConstant = Known.isConstant(); 975 bool MaxShiftAmtIsOutOfRange = Known.getMaxValue().uge(BitWidth); 976 977 if (ShiftAmtIsConstant) { 978 Known = KF(Known2, Known); 979 980 // If the known bits conflict, this must be an overflowing left shift, so 981 // the shift result is poison. We can return anything we want. Choose 0 for 982 // the best folding opportunity. 983 if (Known.hasConflict()) 984 Known.setAllZero(); 985 986 return; 987 } 988 989 // If the shift amount could be greater than or equal to the bit-width of the 990 // LHS, the value could be poison, but bail out because the check below is 991 // expensive. 992 // TODO: Should we just carry on? 993 if (MaxShiftAmtIsOutOfRange) { 994 Known.resetAll(); 995 return; 996 } 997 998 // It would be more-clearly correct to use the two temporaries for this 999 // calculation. Reusing the APInts here to prevent unnecessary allocations. 1000 Known.resetAll(); 1001 1002 // If we know the shifter operand is nonzero, we can sometimes infer more 1003 // known bits. However this is expensive to compute, so be lazy about it and 1004 // only compute it when absolutely necessary. 1005 Optional<bool> ShifterOperandIsNonZero; 1006 1007 // Early exit if we can't constrain any well-defined shift amount. 1008 if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) && 1009 !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) { 1010 ShifterOperandIsNonZero = 1011 isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); 1012 if (!*ShifterOperandIsNonZero) 1013 return; 1014 } 1015 1016 Known.Zero.setAllBits(); 1017 Known.One.setAllBits(); 1018 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) { 1019 // Combine the shifted known input bits only for those shift amounts 1020 // compatible with its known constraints. 1021 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt) 1022 continue; 1023 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt) 1024 continue; 1025 // If we know the shifter is nonzero, we may be able to infer more known 1026 // bits. This check is sunk down as far as possible to avoid the expensive 1027 // call to isKnownNonZero if the cheaper checks above fail. 1028 if (ShiftAmt == 0) { 1029 if (!ShifterOperandIsNonZero.hasValue()) 1030 ShifterOperandIsNonZero = 1031 isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); 1032 if (*ShifterOperandIsNonZero) 1033 continue; 1034 } 1035 1036 Known = KnownBits::commonBits( 1037 Known, KF(Known2, KnownBits::makeConstant(APInt(32, ShiftAmt)))); 1038 } 1039 1040 // If the known bits conflict, the result is poison. Return a 0 and hope the 1041 // caller can further optimize that. 1042 if (Known.hasConflict()) 1043 Known.setAllZero(); 1044 } 1045 1046 static void computeKnownBitsFromOperator(const Operator *I, 1047 const APInt &DemandedElts, 1048 KnownBits &Known, unsigned Depth, 1049 const Query &Q) { 1050 unsigned BitWidth = Known.getBitWidth(); 1051 1052 KnownBits Known2(BitWidth); 1053 switch (I->getOpcode()) { 1054 default: break; 1055 case Instruction::Load: 1056 if (MDNode *MD = 1057 Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range)) 1058 computeKnownBitsFromRangeMetadata(*MD, Known); 1059 break; 1060 case Instruction::And: { 1061 // If either the LHS or the RHS are Zero, the result is zero. 1062 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 1063 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1064 1065 Known &= Known2; 1066 1067 // and(x, add (x, -1)) is a common idiom that always clears the low bit; 1068 // here we handle the more general case of adding any odd number by 1069 // matching the form add(x, add(x, y)) where y is odd. 1070 // TODO: This could be generalized to clearing any bit set in y where the 1071 // following bit is known to be unset in y. 1072 Value *X = nullptr, *Y = nullptr; 1073 if (!Known.Zero[0] && !Known.One[0] && 1074 match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) { 1075 Known2.resetAll(); 1076 computeKnownBits(Y, DemandedElts, Known2, Depth + 1, Q); 1077 if (Known2.countMinTrailingOnes() > 0) 1078 Known.Zero.setBit(0); 1079 } 1080 break; 1081 } 1082 case Instruction::Or: 1083 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 1084 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1085 1086 Known |= Known2; 1087 break; 1088 case Instruction::Xor: 1089 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 1090 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1091 1092 Known ^= Known2; 1093 break; 1094 case Instruction::Mul: { 1095 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1096 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, DemandedElts, 1097 Known, Known2, Depth, Q); 1098 break; 1099 } 1100 case Instruction::UDiv: { 1101 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1102 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1103 Known = KnownBits::udiv(Known, Known2); 1104 break; 1105 } 1106 case Instruction::Select: { 1107 const Value *LHS = nullptr, *RHS = nullptr; 1108 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor; 1109 if (SelectPatternResult::isMinOrMax(SPF)) { 1110 computeKnownBits(RHS, Known, Depth + 1, Q); 1111 computeKnownBits(LHS, Known2, Depth + 1, Q); 1112 switch (SPF) { 1113 default: 1114 llvm_unreachable("Unhandled select pattern flavor!"); 1115 case SPF_SMAX: 1116 Known = KnownBits::smax(Known, Known2); 1117 break; 1118 case SPF_SMIN: 1119 Known = KnownBits::smin(Known, Known2); 1120 break; 1121 case SPF_UMAX: 1122 Known = KnownBits::umax(Known, Known2); 1123 break; 1124 case SPF_UMIN: 1125 Known = KnownBits::umin(Known, Known2); 1126 break; 1127 } 1128 break; 1129 } 1130 1131 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q); 1132 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1133 1134 // Only known if known in both the LHS and RHS. 1135 Known = KnownBits::commonBits(Known, Known2); 1136 1137 if (SPF == SPF_ABS) { 1138 // RHS from matchSelectPattern returns the negation part of abs pattern. 1139 // If the negate has an NSW flag we can assume the sign bit of the result 1140 // will be 0 because that makes abs(INT_MIN) undefined. 1141 if (match(RHS, m_Neg(m_Specific(LHS))) && 1142 Q.IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) 1143 Known.Zero.setSignBit(); 1144 } 1145 1146 break; 1147 } 1148 case Instruction::FPTrunc: 1149 case Instruction::FPExt: 1150 case Instruction::FPToUI: 1151 case Instruction::FPToSI: 1152 case Instruction::SIToFP: 1153 case Instruction::UIToFP: 1154 break; // Can't work with floating point. 1155 case Instruction::PtrToInt: 1156 case Instruction::IntToPtr: 1157 // Fall through and handle them the same as zext/trunc. 1158 LLVM_FALLTHROUGH; 1159 case Instruction::ZExt: 1160 case Instruction::Trunc: { 1161 Type *SrcTy = I->getOperand(0)->getType(); 1162 1163 unsigned SrcBitWidth; 1164 // Note that we handle pointer operands here because of inttoptr/ptrtoint 1165 // which fall through here. 1166 Type *ScalarTy = SrcTy->getScalarType(); 1167 SrcBitWidth = ScalarTy->isPointerTy() ? 1168 Q.DL.getPointerTypeSizeInBits(ScalarTy) : 1169 Q.DL.getTypeSizeInBits(ScalarTy); 1170 1171 assert(SrcBitWidth && "SrcBitWidth can't be zero"); 1172 Known = Known.anyextOrTrunc(SrcBitWidth); 1173 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1174 Known = Known.zextOrTrunc(BitWidth); 1175 break; 1176 } 1177 case Instruction::BitCast: { 1178 Type *SrcTy = I->getOperand(0)->getType(); 1179 if (SrcTy->isIntOrPtrTy() && 1180 // TODO: For now, not handling conversions like: 1181 // (bitcast i64 %x to <2 x i32>) 1182 !I->getType()->isVectorTy()) { 1183 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1184 break; 1185 } 1186 1187 // Handle cast from vector integer type to scalar or vector integer. 1188 auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy); 1189 if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() || 1190 !I->getType()->isIntOrIntVectorTy()) 1191 break; 1192 1193 // Look through a cast from narrow vector elements to wider type. 1194 // Examples: v4i32 -> v2i64, v3i8 -> v24 1195 unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits(); 1196 if (BitWidth % SubBitWidth == 0) { 1197 // Known bits are automatically intersected across demanded elements of a 1198 // vector. So for example, if a bit is computed as known zero, it must be 1199 // zero across all demanded elements of the vector. 1200 // 1201 // For this bitcast, each demanded element of the output is sub-divided 1202 // across a set of smaller vector elements in the source vector. To get 1203 // the known bits for an entire element of the output, compute the known 1204 // bits for each sub-element sequentially. This is done by shifting the 1205 // one-set-bit demanded elements parameter across the sub-elements for 1206 // consecutive calls to computeKnownBits. We are using the demanded 1207 // elements parameter as a mask operator. 1208 // 1209 // The known bits of each sub-element are then inserted into place 1210 // (dependent on endian) to form the full result of known bits. 1211 unsigned NumElts = DemandedElts.getBitWidth(); 1212 unsigned SubScale = BitWidth / SubBitWidth; 1213 APInt SubDemandedElts = APInt::getNullValue(NumElts * SubScale); 1214 for (unsigned i = 0; i != NumElts; ++i) { 1215 if (DemandedElts[i]) 1216 SubDemandedElts.setBit(i * SubScale); 1217 } 1218 1219 KnownBits KnownSrc(SubBitWidth); 1220 for (unsigned i = 0; i != SubScale; ++i) { 1221 computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc, 1222 Depth + 1, Q); 1223 unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i; 1224 Known.insertBits(KnownSrc, ShiftElt * SubBitWidth); 1225 } 1226 } 1227 break; 1228 } 1229 case Instruction::SExt: { 1230 // Compute the bits in the result that are not present in the input. 1231 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); 1232 1233 Known = Known.trunc(SrcBitWidth); 1234 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1235 // If the sign bit of the input is known set or clear, then we know the 1236 // top bits of the result. 1237 Known = Known.sext(BitWidth); 1238 break; 1239 } 1240 case Instruction::Shl: { 1241 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1242 auto KF = [NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt) { 1243 KnownBits Result = KnownBits::shl(KnownVal, KnownAmt); 1244 // If this shift has "nsw" keyword, then the result is either a poison 1245 // value or has the same sign bit as the first operand. 1246 if (NSW) { 1247 if (KnownVal.Zero.isSignBitSet()) 1248 Result.Zero.setSignBit(); 1249 if (KnownVal.One.isSignBitSet()) 1250 Result.One.setSignBit(); 1251 } 1252 return Result; 1253 }; 1254 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, 1255 KF); 1256 // Trailing zeros of a right-shifted constant never decrease. 1257 const APInt *C; 1258 if (match(I->getOperand(0), m_APInt(C))) 1259 Known.Zero.setLowBits(C->countTrailingZeros()); 1260 break; 1261 } 1262 case Instruction::LShr: { 1263 auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { 1264 return KnownBits::lshr(KnownVal, KnownAmt); 1265 }; 1266 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, 1267 KF); 1268 // Leading zeros of a left-shifted constant never decrease. 1269 const APInt *C; 1270 if (match(I->getOperand(0), m_APInt(C))) 1271 Known.Zero.setHighBits(C->countLeadingZeros()); 1272 break; 1273 } 1274 case Instruction::AShr: { 1275 auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { 1276 return KnownBits::ashr(KnownVal, KnownAmt); 1277 }; 1278 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, 1279 KF); 1280 break; 1281 } 1282 case Instruction::Sub: { 1283 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1284 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW, 1285 DemandedElts, Known, Known2, Depth, Q); 1286 break; 1287 } 1288 case Instruction::Add: { 1289 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1290 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW, 1291 DemandedElts, Known, Known2, Depth, Q); 1292 break; 1293 } 1294 case Instruction::SRem: 1295 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1296 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1297 Known = KnownBits::srem(Known, Known2); 1298 break; 1299 1300 case Instruction::URem: 1301 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1302 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1303 Known = KnownBits::urem(Known, Known2); 1304 break; 1305 case Instruction::Alloca: 1306 Known.Zero.setLowBits(Log2(cast<AllocaInst>(I)->getAlign())); 1307 break; 1308 case Instruction::GetElementPtr: { 1309 // Analyze all of the subscripts of this getelementptr instruction 1310 // to determine if we can prove known low zero bits. 1311 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1312 // Accumulate the constant indices in a separate variable 1313 // to minimize the number of calls to computeForAddSub. 1314 APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true); 1315 1316 gep_type_iterator GTI = gep_type_begin(I); 1317 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { 1318 // TrailZ can only become smaller, short-circuit if we hit zero. 1319 if (Known.isUnknown()) 1320 break; 1321 1322 Value *Index = I->getOperand(i); 1323 1324 // Handle case when index is zero. 1325 Constant *CIndex = dyn_cast<Constant>(Index); 1326 if (CIndex && CIndex->isZeroValue()) 1327 continue; 1328 1329 if (StructType *STy = GTI.getStructTypeOrNull()) { 1330 // Handle struct member offset arithmetic. 1331 1332 assert(CIndex && 1333 "Access to structure field must be known at compile time"); 1334 1335 if (CIndex->getType()->isVectorTy()) 1336 Index = CIndex->getSplatValue(); 1337 1338 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue(); 1339 const StructLayout *SL = Q.DL.getStructLayout(STy); 1340 uint64_t Offset = SL->getElementOffset(Idx); 1341 AccConstIndices += Offset; 1342 continue; 1343 } 1344 1345 // Handle array index arithmetic. 1346 Type *IndexedTy = GTI.getIndexedType(); 1347 if (!IndexedTy->isSized()) { 1348 Known.resetAll(); 1349 break; 1350 } 1351 1352 unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits(); 1353 KnownBits IndexBits(IndexBitWidth); 1354 computeKnownBits(Index, IndexBits, Depth + 1, Q); 1355 TypeSize IndexTypeSize = Q.DL.getTypeAllocSize(IndexedTy); 1356 uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinSize(); 1357 KnownBits ScalingFactor(IndexBitWidth); 1358 // Multiply by current sizeof type. 1359 // &A[i] == A + i * sizeof(*A[i]). 1360 if (IndexTypeSize.isScalable()) { 1361 // For scalable types the only thing we know about sizeof is 1362 // that this is a multiple of the minimum size. 1363 ScalingFactor.Zero.setLowBits(countTrailingZeros(TypeSizeInBytes)); 1364 } else if (IndexBits.isConstant()) { 1365 APInt IndexConst = IndexBits.getConstant(); 1366 APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes); 1367 IndexConst *= ScalingFactor; 1368 AccConstIndices += IndexConst.sextOrTrunc(BitWidth); 1369 continue; 1370 } else { 1371 ScalingFactor = 1372 KnownBits::makeConstant(APInt(IndexBitWidth, TypeSizeInBytes)); 1373 } 1374 IndexBits = KnownBits::mul(IndexBits, ScalingFactor); 1375 1376 // If the offsets have a different width from the pointer, according 1377 // to the language reference we need to sign-extend or truncate them 1378 // to the width of the pointer. 1379 IndexBits = IndexBits.sextOrTrunc(BitWidth); 1380 1381 // Note that inbounds does *not* guarantee nsw for the addition, as only 1382 // the offset is signed, while the base address is unsigned. 1383 Known = KnownBits::computeForAddSub( 1384 /*Add=*/true, /*NSW=*/false, Known, IndexBits); 1385 } 1386 if (!Known.isUnknown() && !AccConstIndices.isNullValue()) { 1387 KnownBits Index = KnownBits::makeConstant(AccConstIndices); 1388 Known = KnownBits::computeForAddSub( 1389 /*Add=*/true, /*NSW=*/false, Known, Index); 1390 } 1391 break; 1392 } 1393 case Instruction::PHI: { 1394 const PHINode *P = cast<PHINode>(I); 1395 BinaryOperator *BO = nullptr; 1396 Value *R = nullptr, *L = nullptr; 1397 if (matchSimpleRecurrence(P, BO, R, L)) { 1398 // Handle the case of a simple two-predecessor recurrence PHI. 1399 // There's a lot more that could theoretically be done here, but 1400 // this is sufficient to catch some interesting cases. 1401 unsigned Opcode = BO->getOpcode(); 1402 1403 // If this is a shift recurrence, we know the bits being shifted in. 1404 // We can combine that with information about the start value of the 1405 // recurrence to conclude facts about the result. 1406 if ((Opcode == Instruction::LShr || Opcode == Instruction::AShr || 1407 Opcode == Instruction::Shl) && 1408 BO->getOperand(0) == I) { 1409 1410 // We have matched a recurrence of the form: 1411 // %iv = [R, %entry], [%iv.next, %backedge] 1412 // %iv.next = shift_op %iv, L 1413 1414 // Recurse with the phi context to avoid concern about whether facts 1415 // inferred hold at original context instruction. TODO: It may be 1416 // correct to use the original context. IF warranted, explore and 1417 // add sufficient tests to cover. 1418 Query RecQ = Q; 1419 RecQ.CxtI = P; 1420 computeKnownBits(R, DemandedElts, Known2, Depth + 1, RecQ); 1421 switch (Opcode) { 1422 case Instruction::Shl: 1423 // A shl recurrence will only increase the tailing zeros 1424 Known.Zero.setLowBits(Known2.countMinTrailingZeros()); 1425 break; 1426 case Instruction::LShr: 1427 // A lshr recurrence will preserve the leading zeros of the 1428 // start value 1429 Known.Zero.setHighBits(Known2.countMinLeadingZeros()); 1430 break; 1431 case Instruction::AShr: 1432 // An ashr recurrence will extend the initial sign bit 1433 Known.Zero.setHighBits(Known2.countMinLeadingZeros()); 1434 Known.One.setHighBits(Known2.countMinLeadingOnes()); 1435 break; 1436 }; 1437 } 1438 1439 // Check for operations that have the property that if 1440 // both their operands have low zero bits, the result 1441 // will have low zero bits. 1442 if (Opcode == Instruction::Add || 1443 Opcode == Instruction::Sub || 1444 Opcode == Instruction::And || 1445 Opcode == Instruction::Or || 1446 Opcode == Instruction::Mul) { 1447 // Change the context instruction to the "edge" that flows into the 1448 // phi. This is important because that is where the value is actually 1449 // "evaluated" even though it is used later somewhere else. (see also 1450 // D69571). 1451 Query RecQ = Q; 1452 1453 unsigned OpNum = P->getOperand(0) == R ? 0 : 1; 1454 Instruction *RInst = P->getIncomingBlock(OpNum)->getTerminator(); 1455 Instruction *LInst = P->getIncomingBlock(1-OpNum)->getTerminator(); 1456 1457 // Ok, we have a PHI of the form L op= R. Check for low 1458 // zero bits. 1459 RecQ.CxtI = RInst; 1460 computeKnownBits(R, Known2, Depth + 1, RecQ); 1461 1462 // We need to take the minimum number of known bits 1463 KnownBits Known3(BitWidth); 1464 RecQ.CxtI = LInst; 1465 computeKnownBits(L, Known3, Depth + 1, RecQ); 1466 1467 Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(), 1468 Known3.countMinTrailingZeros())); 1469 1470 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(BO); 1471 if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) { 1472 // If initial value of recurrence is nonnegative, and we are adding 1473 // a nonnegative number with nsw, the result can only be nonnegative 1474 // or poison value regardless of the number of times we execute the 1475 // add in phi recurrence. If initial value is negative and we are 1476 // adding a negative number with nsw, the result can only be 1477 // negative or poison value. Similar arguments apply to sub and mul. 1478 // 1479 // (add non-negative, non-negative) --> non-negative 1480 // (add negative, negative) --> negative 1481 if (Opcode == Instruction::Add) { 1482 if (Known2.isNonNegative() && Known3.isNonNegative()) 1483 Known.makeNonNegative(); 1484 else if (Known2.isNegative() && Known3.isNegative()) 1485 Known.makeNegative(); 1486 } 1487 1488 // (sub nsw non-negative, negative) --> non-negative 1489 // (sub nsw negative, non-negative) --> negative 1490 else if (Opcode == Instruction::Sub && BO->getOperand(0) == I) { 1491 if (Known2.isNonNegative() && Known3.isNegative()) 1492 Known.makeNonNegative(); 1493 else if (Known2.isNegative() && Known3.isNonNegative()) 1494 Known.makeNegative(); 1495 } 1496 1497 // (mul nsw non-negative, non-negative) --> non-negative 1498 else if (Opcode == Instruction::Mul && Known2.isNonNegative() && 1499 Known3.isNonNegative()) 1500 Known.makeNonNegative(); 1501 } 1502 1503 break; 1504 } 1505 } 1506 1507 // Unreachable blocks may have zero-operand PHI nodes. 1508 if (P->getNumIncomingValues() == 0) 1509 break; 1510 1511 // Otherwise take the unions of the known bit sets of the operands, 1512 // taking conservative care to avoid excessive recursion. 1513 if (Depth < MaxAnalysisRecursionDepth - 1 && !Known.Zero && !Known.One) { 1514 // Skip if every incoming value references to ourself. 1515 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue())) 1516 break; 1517 1518 Known.Zero.setAllBits(); 1519 Known.One.setAllBits(); 1520 for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) { 1521 Value *IncValue = P->getIncomingValue(u); 1522 // Skip direct self references. 1523 if (IncValue == P) continue; 1524 1525 // Change the context instruction to the "edge" that flows into the 1526 // phi. This is important because that is where the value is actually 1527 // "evaluated" even though it is used later somewhere else. (see also 1528 // D69571). 1529 Query RecQ = Q; 1530 RecQ.CxtI = P->getIncomingBlock(u)->getTerminator(); 1531 1532 Known2 = KnownBits(BitWidth); 1533 // Recurse, but cap the recursion to one level, because we don't 1534 // want to waste time spinning around in loops. 1535 computeKnownBits(IncValue, Known2, MaxAnalysisRecursionDepth - 1, RecQ); 1536 Known = KnownBits::commonBits(Known, Known2); 1537 // If all bits have been ruled out, there's no need to check 1538 // more operands. 1539 if (Known.isUnknown()) 1540 break; 1541 } 1542 } 1543 break; 1544 } 1545 case Instruction::Call: 1546 case Instruction::Invoke: 1547 // If range metadata is attached to this call, set known bits from that, 1548 // and then intersect with known bits based on other properties of the 1549 // function. 1550 if (MDNode *MD = 1551 Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range)) 1552 computeKnownBitsFromRangeMetadata(*MD, Known); 1553 if (const Value *RV = cast<CallBase>(I)->getReturnedArgOperand()) { 1554 computeKnownBits(RV, Known2, Depth + 1, Q); 1555 Known.Zero |= Known2.Zero; 1556 Known.One |= Known2.One; 1557 } 1558 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1559 switch (II->getIntrinsicID()) { 1560 default: break; 1561 case Intrinsic::abs: { 1562 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1563 bool IntMinIsPoison = match(II->getArgOperand(1), m_One()); 1564 Known = Known2.abs(IntMinIsPoison); 1565 break; 1566 } 1567 case Intrinsic::bitreverse: 1568 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1569 Known.Zero |= Known2.Zero.reverseBits(); 1570 Known.One |= Known2.One.reverseBits(); 1571 break; 1572 case Intrinsic::bswap: 1573 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1574 Known.Zero |= Known2.Zero.byteSwap(); 1575 Known.One |= Known2.One.byteSwap(); 1576 break; 1577 case Intrinsic::ctlz: { 1578 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1579 // If we have a known 1, its position is our upper bound. 1580 unsigned PossibleLZ = Known2.countMaxLeadingZeros(); 1581 // If this call is undefined for 0, the result will be less than 2^n. 1582 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) 1583 PossibleLZ = std::min(PossibleLZ, BitWidth - 1); 1584 unsigned LowBits = Log2_32(PossibleLZ)+1; 1585 Known.Zero.setBitsFrom(LowBits); 1586 break; 1587 } 1588 case Intrinsic::cttz: { 1589 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1590 // If we have a known 1, its position is our upper bound. 1591 unsigned PossibleTZ = Known2.countMaxTrailingZeros(); 1592 // If this call is undefined for 0, the result will be less than 2^n. 1593 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) 1594 PossibleTZ = std::min(PossibleTZ, BitWidth - 1); 1595 unsigned LowBits = Log2_32(PossibleTZ)+1; 1596 Known.Zero.setBitsFrom(LowBits); 1597 break; 1598 } 1599 case Intrinsic::ctpop: { 1600 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1601 // We can bound the space the count needs. Also, bits known to be zero 1602 // can't contribute to the population. 1603 unsigned BitsPossiblySet = Known2.countMaxPopulation(); 1604 unsigned LowBits = Log2_32(BitsPossiblySet)+1; 1605 Known.Zero.setBitsFrom(LowBits); 1606 // TODO: we could bound KnownOne using the lower bound on the number 1607 // of bits which might be set provided by popcnt KnownOne2. 1608 break; 1609 } 1610 case Intrinsic::fshr: 1611 case Intrinsic::fshl: { 1612 const APInt *SA; 1613 if (!match(I->getOperand(2), m_APInt(SA))) 1614 break; 1615 1616 // Normalize to funnel shift left. 1617 uint64_t ShiftAmt = SA->urem(BitWidth); 1618 if (II->getIntrinsicID() == Intrinsic::fshr) 1619 ShiftAmt = BitWidth - ShiftAmt; 1620 1621 KnownBits Known3(BitWidth); 1622 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1623 computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q); 1624 1625 Known.Zero = 1626 Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt); 1627 Known.One = 1628 Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt); 1629 break; 1630 } 1631 case Intrinsic::uadd_sat: 1632 case Intrinsic::usub_sat: { 1633 bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat; 1634 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1635 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1636 1637 // Add: Leading ones of either operand are preserved. 1638 // Sub: Leading zeros of LHS and leading ones of RHS are preserved 1639 // as leading zeros in the result. 1640 unsigned LeadingKnown; 1641 if (IsAdd) 1642 LeadingKnown = std::max(Known.countMinLeadingOnes(), 1643 Known2.countMinLeadingOnes()); 1644 else 1645 LeadingKnown = std::max(Known.countMinLeadingZeros(), 1646 Known2.countMinLeadingOnes()); 1647 1648 Known = KnownBits::computeForAddSub( 1649 IsAdd, /* NSW */ false, Known, Known2); 1650 1651 // We select between the operation result and all-ones/zero 1652 // respectively, so we can preserve known ones/zeros. 1653 if (IsAdd) { 1654 Known.One.setHighBits(LeadingKnown); 1655 Known.Zero.clearAllBits(); 1656 } else { 1657 Known.Zero.setHighBits(LeadingKnown); 1658 Known.One.clearAllBits(); 1659 } 1660 break; 1661 } 1662 case Intrinsic::umin: 1663 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1664 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1665 Known = KnownBits::umin(Known, Known2); 1666 break; 1667 case Intrinsic::umax: 1668 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1669 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1670 Known = KnownBits::umax(Known, Known2); 1671 break; 1672 case Intrinsic::smin: 1673 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1674 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1675 Known = KnownBits::smin(Known, Known2); 1676 break; 1677 case Intrinsic::smax: 1678 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1679 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1680 Known = KnownBits::smax(Known, Known2); 1681 break; 1682 case Intrinsic::x86_sse42_crc32_64_64: 1683 Known.Zero.setBitsFrom(32); 1684 break; 1685 case Intrinsic::riscv_vsetvli: 1686 case Intrinsic::riscv_vsetvlimax: 1687 // Assume that VL output is positive and would fit in an int32_t. 1688 // TODO: VLEN might be capped at 16 bits in a future V spec update. 1689 if (BitWidth >= 32) 1690 Known.Zero.setBitsFrom(31); 1691 break; 1692 } 1693 } 1694 break; 1695 case Instruction::ShuffleVector: { 1696 auto *Shuf = dyn_cast<ShuffleVectorInst>(I); 1697 // FIXME: Do we need to handle ConstantExpr involving shufflevectors? 1698 if (!Shuf) { 1699 Known.resetAll(); 1700 return; 1701 } 1702 // For undef elements, we don't know anything about the common state of 1703 // the shuffle result. 1704 APInt DemandedLHS, DemandedRHS; 1705 if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) { 1706 Known.resetAll(); 1707 return; 1708 } 1709 Known.One.setAllBits(); 1710 Known.Zero.setAllBits(); 1711 if (!!DemandedLHS) { 1712 const Value *LHS = Shuf->getOperand(0); 1713 computeKnownBits(LHS, DemandedLHS, Known, Depth + 1, Q); 1714 // If we don't know any bits, early out. 1715 if (Known.isUnknown()) 1716 break; 1717 } 1718 if (!!DemandedRHS) { 1719 const Value *RHS = Shuf->getOperand(1); 1720 computeKnownBits(RHS, DemandedRHS, Known2, Depth + 1, Q); 1721 Known = KnownBits::commonBits(Known, Known2); 1722 } 1723 break; 1724 } 1725 case Instruction::InsertElement: { 1726 const Value *Vec = I->getOperand(0); 1727 const Value *Elt = I->getOperand(1); 1728 auto *CIdx = dyn_cast<ConstantInt>(I->getOperand(2)); 1729 // Early out if the index is non-constant or out-of-range. 1730 unsigned NumElts = DemandedElts.getBitWidth(); 1731 if (!CIdx || CIdx->getValue().uge(NumElts)) { 1732 Known.resetAll(); 1733 return; 1734 } 1735 Known.One.setAllBits(); 1736 Known.Zero.setAllBits(); 1737 unsigned EltIdx = CIdx->getZExtValue(); 1738 // Do we demand the inserted element? 1739 if (DemandedElts[EltIdx]) { 1740 computeKnownBits(Elt, Known, Depth + 1, Q); 1741 // If we don't know any bits, early out. 1742 if (Known.isUnknown()) 1743 break; 1744 } 1745 // We don't need the base vector element that has been inserted. 1746 APInt DemandedVecElts = DemandedElts; 1747 DemandedVecElts.clearBit(EltIdx); 1748 if (!!DemandedVecElts) { 1749 computeKnownBits(Vec, DemandedVecElts, Known2, Depth + 1, Q); 1750 Known = KnownBits::commonBits(Known, Known2); 1751 } 1752 break; 1753 } 1754 case Instruction::ExtractElement: { 1755 // Look through extract element. If the index is non-constant or 1756 // out-of-range demand all elements, otherwise just the extracted element. 1757 const Value *Vec = I->getOperand(0); 1758 const Value *Idx = I->getOperand(1); 1759 auto *CIdx = dyn_cast<ConstantInt>(Idx); 1760 if (isa<ScalableVectorType>(Vec->getType())) { 1761 // FIXME: there's probably *something* we can do with scalable vectors 1762 Known.resetAll(); 1763 break; 1764 } 1765 unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); 1766 APInt DemandedVecElts = APInt::getAllOnesValue(NumElts); 1767 if (CIdx && CIdx->getValue().ult(NumElts)) 1768 DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); 1769 computeKnownBits(Vec, DemandedVecElts, Known, Depth + 1, Q); 1770 break; 1771 } 1772 case Instruction::ExtractValue: 1773 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) { 1774 const ExtractValueInst *EVI = cast<ExtractValueInst>(I); 1775 if (EVI->getNumIndices() != 1) break; 1776 if (EVI->getIndices()[0] == 0) { 1777 switch (II->getIntrinsicID()) { 1778 default: break; 1779 case Intrinsic::uadd_with_overflow: 1780 case Intrinsic::sadd_with_overflow: 1781 computeKnownBitsAddSub(true, II->getArgOperand(0), 1782 II->getArgOperand(1), false, DemandedElts, 1783 Known, Known2, Depth, Q); 1784 break; 1785 case Intrinsic::usub_with_overflow: 1786 case Intrinsic::ssub_with_overflow: 1787 computeKnownBitsAddSub(false, II->getArgOperand(0), 1788 II->getArgOperand(1), false, DemandedElts, 1789 Known, Known2, Depth, Q); 1790 break; 1791 case Intrinsic::umul_with_overflow: 1792 case Intrinsic::smul_with_overflow: 1793 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false, 1794 DemandedElts, Known, Known2, Depth, Q); 1795 break; 1796 } 1797 } 1798 } 1799 break; 1800 case Instruction::Freeze: 1801 if (isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT, 1802 Depth + 1)) 1803 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1804 break; 1805 } 1806 } 1807 1808 /// Determine which bits of V are known to be either zero or one and return 1809 /// them. 1810 KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, 1811 unsigned Depth, const Query &Q) { 1812 KnownBits Known(getBitWidth(V->getType(), Q.DL)); 1813 computeKnownBits(V, DemandedElts, Known, Depth, Q); 1814 return Known; 1815 } 1816 1817 /// Determine which bits of V are known to be either zero or one and return 1818 /// them. 1819 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) { 1820 KnownBits Known(getBitWidth(V->getType(), Q.DL)); 1821 computeKnownBits(V, Known, Depth, Q); 1822 return Known; 1823 } 1824 1825 /// Determine which bits of V are known to be either zero or one and return 1826 /// them in the Known bit set. 1827 /// 1828 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that 1829 /// we cannot optimize based on the assumption that it is zero without changing 1830 /// it to be an explicit zero. If we don't change it to zero, other code could 1831 /// optimized based on the contradictory assumption that it is non-zero. 1832 /// Because instcombine aggressively folds operations with undef args anyway, 1833 /// this won't lose us code quality. 1834 /// 1835 /// This function is defined on values with integer type, values with pointer 1836 /// type, and vectors of integers. In the case 1837 /// where V is a vector, known zero, and known one values are the 1838 /// same width as the vector element, and the bit is set only if it is true 1839 /// for all of the demanded elements in the vector specified by DemandedElts. 1840 void computeKnownBits(const Value *V, const APInt &DemandedElts, 1841 KnownBits &Known, unsigned Depth, const Query &Q) { 1842 if (!DemandedElts || isa<ScalableVectorType>(V->getType())) { 1843 // No demanded elts or V is a scalable vector, better to assume we don't 1844 // know anything. 1845 Known.resetAll(); 1846 return; 1847 } 1848 1849 assert(V && "No Value?"); 1850 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 1851 1852 #ifndef NDEBUG 1853 Type *Ty = V->getType(); 1854 unsigned BitWidth = Known.getBitWidth(); 1855 1856 assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && 1857 "Not integer or pointer type!"); 1858 1859 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { 1860 assert( 1861 FVTy->getNumElements() == DemandedElts.getBitWidth() && 1862 "DemandedElt width should equal the fixed vector number of elements"); 1863 } else { 1864 assert(DemandedElts == APInt(1, 1) && 1865 "DemandedElt width should be 1 for scalars"); 1866 } 1867 1868 Type *ScalarTy = Ty->getScalarType(); 1869 if (ScalarTy->isPointerTy()) { 1870 assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && 1871 "V and Known should have same BitWidth"); 1872 } else { 1873 assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && 1874 "V and Known should have same BitWidth"); 1875 } 1876 #endif 1877 1878 const APInt *C; 1879 if (match(V, m_APInt(C))) { 1880 // We know all of the bits for a scalar constant or a splat vector constant! 1881 Known = KnownBits::makeConstant(*C); 1882 return; 1883 } 1884 // Null and aggregate-zero are all-zeros. 1885 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) { 1886 Known.setAllZero(); 1887 return; 1888 } 1889 // Handle a constant vector by taking the intersection of the known bits of 1890 // each element. 1891 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) { 1892 // We know that CDV must be a vector of integers. Take the intersection of 1893 // each element. 1894 Known.Zero.setAllBits(); Known.One.setAllBits(); 1895 for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) { 1896 if (!DemandedElts[i]) 1897 continue; 1898 APInt Elt = CDV->getElementAsAPInt(i); 1899 Known.Zero &= ~Elt; 1900 Known.One &= Elt; 1901 } 1902 return; 1903 } 1904 1905 if (const auto *CV = dyn_cast<ConstantVector>(V)) { 1906 // We know that CV must be a vector of integers. Take the intersection of 1907 // each element. 1908 Known.Zero.setAllBits(); Known.One.setAllBits(); 1909 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { 1910 if (!DemandedElts[i]) 1911 continue; 1912 Constant *Element = CV->getAggregateElement(i); 1913 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element); 1914 if (!ElementCI) { 1915 Known.resetAll(); 1916 return; 1917 } 1918 const APInt &Elt = ElementCI->getValue(); 1919 Known.Zero &= ~Elt; 1920 Known.One &= Elt; 1921 } 1922 return; 1923 } 1924 1925 // Start out not knowing anything. 1926 Known.resetAll(); 1927 1928 // We can't imply anything about undefs. 1929 if (isa<UndefValue>(V)) 1930 return; 1931 1932 // There's no point in looking through other users of ConstantData for 1933 // assumptions. Confirm that we've handled them all. 1934 assert(!isa<ConstantData>(V) && "Unhandled constant data!"); 1935 1936 // All recursive calls that increase depth must come after this. 1937 if (Depth == MaxAnalysisRecursionDepth) 1938 return; 1939 1940 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has 1941 // the bits of its aliasee. 1942 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 1943 if (!GA->isInterposable()) 1944 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q); 1945 return; 1946 } 1947 1948 if (const Operator *I = dyn_cast<Operator>(V)) 1949 computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q); 1950 1951 // Aligned pointers have trailing zeros - refine Known.Zero set 1952 if (isa<PointerType>(V->getType())) { 1953 Align Alignment = V->getPointerAlignment(Q.DL); 1954 Known.Zero.setLowBits(Log2(Alignment)); 1955 } 1956 1957 // computeKnownBitsFromAssume strictly refines Known. 1958 // Therefore, we run them after computeKnownBitsFromOperator. 1959 1960 // Check whether a nearby assume intrinsic can determine some known bits. 1961 computeKnownBitsFromAssume(V, Known, Depth, Q); 1962 1963 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?"); 1964 } 1965 1966 /// Return true if the given value is known to have exactly one 1967 /// bit set when defined. For vectors return true if every element is known to 1968 /// be a power of two when defined. Supports values with integer or pointer 1969 /// types and vectors of integers. 1970 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, 1971 const Query &Q) { 1972 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 1973 1974 // Attempt to match against constants. 1975 if (OrZero && match(V, m_Power2OrZero())) 1976 return true; 1977 if (match(V, m_Power2())) 1978 return true; 1979 1980 // 1 << X is clearly a power of two if the one is not shifted off the end. If 1981 // it is shifted off the end then the result is undefined. 1982 if (match(V, m_Shl(m_One(), m_Value()))) 1983 return true; 1984 1985 // (signmask) >>l X is clearly a power of two if the one is not shifted off 1986 // the bottom. If it is shifted off the bottom then the result is undefined. 1987 if (match(V, m_LShr(m_SignMask(), m_Value()))) 1988 return true; 1989 1990 // The remaining tests are all recursive, so bail out if we hit the limit. 1991 if (Depth++ == MaxAnalysisRecursionDepth) 1992 return false; 1993 1994 Value *X = nullptr, *Y = nullptr; 1995 // A shift left or a logical shift right of a power of two is a power of two 1996 // or zero. 1997 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) || 1998 match(V, m_LShr(m_Value(X), m_Value())))) 1999 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q); 2000 2001 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V)) 2002 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q); 2003 2004 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) 2005 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) && 2006 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q); 2007 2008 // Peek through min/max. 2009 if (match(V, m_MaxOrMin(m_Value(X), m_Value(Y)))) { 2010 return isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q) && 2011 isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q); 2012 } 2013 2014 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) { 2015 // A power of two and'd with anything is a power of two or zero. 2016 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) || 2017 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q)) 2018 return true; 2019 // X & (-X) is always a power of two or zero. 2020 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X)))) 2021 return true; 2022 return false; 2023 } 2024 2025 // Adding a power-of-two or zero to the same power-of-two or zero yields 2026 // either the original power-of-two, a larger power-of-two or zero. 2027 if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 2028 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V); 2029 if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) || 2030 Q.IIQ.hasNoSignedWrap(VOBO)) { 2031 if (match(X, m_And(m_Specific(Y), m_Value())) || 2032 match(X, m_And(m_Value(), m_Specific(Y)))) 2033 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q)) 2034 return true; 2035 if (match(Y, m_And(m_Specific(X), m_Value())) || 2036 match(Y, m_And(m_Value(), m_Specific(X)))) 2037 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q)) 2038 return true; 2039 2040 unsigned BitWidth = V->getType()->getScalarSizeInBits(); 2041 KnownBits LHSBits(BitWidth); 2042 computeKnownBits(X, LHSBits, Depth, Q); 2043 2044 KnownBits RHSBits(BitWidth); 2045 computeKnownBits(Y, RHSBits, Depth, Q); 2046 // If i8 V is a power of two or zero: 2047 // ZeroBits: 1 1 1 0 1 1 1 1 2048 // ~ZeroBits: 0 0 0 1 0 0 0 0 2049 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2()) 2050 // If OrZero isn't set, we cannot give back a zero result. 2051 // Make sure either the LHS or RHS has a bit set. 2052 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue()) 2053 return true; 2054 } 2055 } 2056 2057 // An exact divide or right shift can only shift off zero bits, so the result 2058 // is a power of two only if the first operand is a power of two and not 2059 // copying a sign bit (sdiv int_min, 2). 2060 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) || 2061 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) { 2062 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, 2063 Depth, Q); 2064 } 2065 2066 return false; 2067 } 2068 2069 /// Test whether a GEP's result is known to be non-null. 2070 /// 2071 /// Uses properties inherent in a GEP to try to determine whether it is known 2072 /// to be non-null. 2073 /// 2074 /// Currently this routine does not support vector GEPs. 2075 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth, 2076 const Query &Q) { 2077 const Function *F = nullptr; 2078 if (const Instruction *I = dyn_cast<Instruction>(GEP)) 2079 F = I->getFunction(); 2080 2081 if (!GEP->isInBounds() || 2082 NullPointerIsDefined(F, GEP->getPointerAddressSpace())) 2083 return false; 2084 2085 // FIXME: Support vector-GEPs. 2086 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP"); 2087 2088 // If the base pointer is non-null, we cannot walk to a null address with an 2089 // inbounds GEP in address space zero. 2090 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q)) 2091 return true; 2092 2093 // Walk the GEP operands and see if any operand introduces a non-zero offset. 2094 // If so, then the GEP cannot produce a null pointer, as doing so would 2095 // inherently violate the inbounds contract within address space zero. 2096 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); 2097 GTI != GTE; ++GTI) { 2098 // Struct types are easy -- they must always be indexed by a constant. 2099 if (StructType *STy = GTI.getStructTypeOrNull()) { 2100 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand()); 2101 unsigned ElementIdx = OpC->getZExtValue(); 2102 const StructLayout *SL = Q.DL.getStructLayout(STy); 2103 uint64_t ElementOffset = SL->getElementOffset(ElementIdx); 2104 if (ElementOffset > 0) 2105 return true; 2106 continue; 2107 } 2108 2109 // If we have a zero-sized type, the index doesn't matter. Keep looping. 2110 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()).getKnownMinSize() == 0) 2111 continue; 2112 2113 // Fast path the constant operand case both for efficiency and so we don't 2114 // increment Depth when just zipping down an all-constant GEP. 2115 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) { 2116 if (!OpC->isZero()) 2117 return true; 2118 continue; 2119 } 2120 2121 // We post-increment Depth here because while isKnownNonZero increments it 2122 // as well, when we pop back up that increment won't persist. We don't want 2123 // to recurse 10k times just because we have 10k GEP operands. We don't 2124 // bail completely out because we want to handle constant GEPs regardless 2125 // of depth. 2126 if (Depth++ >= MaxAnalysisRecursionDepth) 2127 continue; 2128 2129 if (isKnownNonZero(GTI.getOperand(), Depth, Q)) 2130 return true; 2131 } 2132 2133 return false; 2134 } 2135 2136 static bool isKnownNonNullFromDominatingCondition(const Value *V, 2137 const Instruction *CtxI, 2138 const DominatorTree *DT) { 2139 if (isa<Constant>(V)) 2140 return false; 2141 2142 if (!CtxI || !DT) 2143 return false; 2144 2145 unsigned NumUsesExplored = 0; 2146 for (auto *U : V->users()) { 2147 // Avoid massive lists 2148 if (NumUsesExplored >= DomConditionsMaxUses) 2149 break; 2150 NumUsesExplored++; 2151 2152 // If the value is used as an argument to a call or invoke, then argument 2153 // attributes may provide an answer about null-ness. 2154 if (const auto *CB = dyn_cast<CallBase>(U)) 2155 if (auto *CalledFunc = CB->getCalledFunction()) 2156 for (const Argument &Arg : CalledFunc->args()) 2157 if (CB->getArgOperand(Arg.getArgNo()) == V && 2158 Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) && 2159 DT->dominates(CB, CtxI)) 2160 return true; 2161 2162 // If the value is used as a load/store, then the pointer must be non null. 2163 if (V == getLoadStorePointerOperand(U)) { 2164 const Instruction *I = cast<Instruction>(U); 2165 if (!NullPointerIsDefined(I->getFunction(), 2166 V->getType()->getPointerAddressSpace()) && 2167 DT->dominates(I, CtxI)) 2168 return true; 2169 } 2170 2171 // Consider only compare instructions uniquely controlling a branch 2172 Value *RHS; 2173 CmpInst::Predicate Pred; 2174 if (!match(U, m_c_ICmp(Pred, m_Specific(V), m_Value(RHS)))) 2175 continue; 2176 2177 bool NonNullIfTrue; 2178 if (cmpExcludesZero(Pred, RHS)) 2179 NonNullIfTrue = true; 2180 else if (cmpExcludesZero(CmpInst::getInversePredicate(Pred), RHS)) 2181 NonNullIfTrue = false; 2182 else 2183 continue; 2184 2185 SmallVector<const User *, 4> WorkList; 2186 SmallPtrSet<const User *, 4> Visited; 2187 for (auto *CmpU : U->users()) { 2188 assert(WorkList.empty() && "Should be!"); 2189 if (Visited.insert(CmpU).second) 2190 WorkList.push_back(CmpU); 2191 2192 while (!WorkList.empty()) { 2193 auto *Curr = WorkList.pop_back_val(); 2194 2195 // If a user is an AND, add all its users to the work list. We only 2196 // propagate "pred != null" condition through AND because it is only 2197 // correct to assume that all conditions of AND are met in true branch. 2198 // TODO: Support similar logic of OR and EQ predicate? 2199 if (NonNullIfTrue) 2200 if (match(Curr, m_LogicalAnd(m_Value(), m_Value()))) { 2201 for (auto *CurrU : Curr->users()) 2202 if (Visited.insert(CurrU).second) 2203 WorkList.push_back(CurrU); 2204 continue; 2205 } 2206 2207 if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) { 2208 assert(BI->isConditional() && "uses a comparison!"); 2209 2210 BasicBlock *NonNullSuccessor = 2211 BI->getSuccessor(NonNullIfTrue ? 0 : 1); 2212 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); 2213 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent())) 2214 return true; 2215 } else if (NonNullIfTrue && isGuard(Curr) && 2216 DT->dominates(cast<Instruction>(Curr), CtxI)) { 2217 return true; 2218 } 2219 } 2220 } 2221 } 2222 2223 return false; 2224 } 2225 2226 /// Does the 'Range' metadata (which must be a valid MD_range operand list) 2227 /// ensure that the value it's attached to is never Value? 'RangeType' is 2228 /// is the type of the value described by the range. 2229 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) { 2230 const unsigned NumRanges = Ranges->getNumOperands() / 2; 2231 assert(NumRanges >= 1); 2232 for (unsigned i = 0; i < NumRanges; ++i) { 2233 ConstantInt *Lower = 2234 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); 2235 ConstantInt *Upper = 2236 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); 2237 ConstantRange Range(Lower->getValue(), Upper->getValue()); 2238 if (Range.contains(Value)) 2239 return false; 2240 } 2241 return true; 2242 } 2243 2244 /// Try to detect a recurrence that monotonically increases/decreases from a 2245 /// non-zero starting value. These are common as induction variables. 2246 static bool isNonZeroRecurrence(const PHINode *PN) { 2247 BinaryOperator *BO = nullptr; 2248 Value *Start = nullptr, *Step = nullptr; 2249 const APInt *StartC, *StepC; 2250 if (!matchSimpleRecurrence(PN, BO, Start, Step) || 2251 !match(Start, m_APInt(StartC)) || StartC->isNullValue()) 2252 return false; 2253 2254 switch (BO->getOpcode()) { 2255 case Instruction::Add: 2256 // Starting from non-zero and stepping away from zero can never wrap back 2257 // to zero. 2258 return BO->hasNoUnsignedWrap() || 2259 (BO->hasNoSignedWrap() && match(Step, m_APInt(StepC)) && 2260 StartC->isNegative() == StepC->isNegative()); 2261 case Instruction::Mul: 2262 return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) && 2263 match(Step, m_APInt(StepC)) && !StepC->isNullValue(); 2264 case Instruction::Shl: 2265 return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap(); 2266 case Instruction::AShr: 2267 case Instruction::LShr: 2268 return BO->isExact(); 2269 default: 2270 return false; 2271 } 2272 } 2273 2274 /// Return true if the given value is known to be non-zero when defined. For 2275 /// vectors, return true if every demanded element is known to be non-zero when 2276 /// defined. For pointers, if the context instruction and dominator tree are 2277 /// specified, perform context-sensitive analysis and return true if the 2278 /// pointer couldn't possibly be null at the specified instruction. 2279 /// Supports values with integer or pointer type and vectors of integers. 2280 bool isKnownNonZero(const Value *V, const APInt &DemandedElts, unsigned Depth, 2281 const Query &Q) { 2282 // FIXME: We currently have no way to represent the DemandedElts of a scalable 2283 // vector 2284 if (isa<ScalableVectorType>(V->getType())) 2285 return false; 2286 2287 if (auto *C = dyn_cast<Constant>(V)) { 2288 if (C->isNullValue()) 2289 return false; 2290 if (isa<ConstantInt>(C)) 2291 // Must be non-zero due to null test above. 2292 return true; 2293 2294 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 2295 // See the comment for IntToPtr/PtrToInt instructions below. 2296 if (CE->getOpcode() == Instruction::IntToPtr || 2297 CE->getOpcode() == Instruction::PtrToInt) 2298 if (Q.DL.getTypeSizeInBits(CE->getOperand(0)->getType()) 2299 .getFixedSize() <= 2300 Q.DL.getTypeSizeInBits(CE->getType()).getFixedSize()) 2301 return isKnownNonZero(CE->getOperand(0), Depth, Q); 2302 } 2303 2304 // For constant vectors, check that all elements are undefined or known 2305 // non-zero to determine that the whole vector is known non-zero. 2306 if (auto *VecTy = dyn_cast<FixedVectorType>(C->getType())) { 2307 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { 2308 if (!DemandedElts[i]) 2309 continue; 2310 Constant *Elt = C->getAggregateElement(i); 2311 if (!Elt || Elt->isNullValue()) 2312 return false; 2313 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt)) 2314 return false; 2315 } 2316 return true; 2317 } 2318 2319 // A global variable in address space 0 is non null unless extern weak 2320 // or an absolute symbol reference. Other address spaces may have null as a 2321 // valid address for a global, so we can't assume anything. 2322 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { 2323 if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() && 2324 GV->getType()->getAddressSpace() == 0) 2325 return true; 2326 } else 2327 return false; 2328 } 2329 2330 if (auto *I = dyn_cast<Instruction>(V)) { 2331 if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) { 2332 // If the possible ranges don't contain zero, then the value is 2333 // definitely non-zero. 2334 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) { 2335 const APInt ZeroValue(Ty->getBitWidth(), 0); 2336 if (rangeMetadataExcludesValue(Ranges, ZeroValue)) 2337 return true; 2338 } 2339 } 2340 } 2341 2342 if (isKnownNonZeroFromAssume(V, Q)) 2343 return true; 2344 2345 // Some of the tests below are recursive, so bail out if we hit the limit. 2346 if (Depth++ >= MaxAnalysisRecursionDepth) 2347 return false; 2348 2349 // Check for pointer simplifications. 2350 2351 if (PointerType *PtrTy = dyn_cast<PointerType>(V->getType())) { 2352 // Alloca never returns null, malloc might. 2353 if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0) 2354 return true; 2355 2356 // A byval, inalloca may not be null in a non-default addres space. A 2357 // nonnull argument is assumed never 0. 2358 if (const Argument *A = dyn_cast<Argument>(V)) { 2359 if (((A->hasPassPointeeByValueCopyAttr() && 2360 !NullPointerIsDefined(A->getParent(), PtrTy->getAddressSpace())) || 2361 A->hasNonNullAttr())) 2362 return true; 2363 } 2364 2365 // A Load tagged with nonnull metadata is never null. 2366 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) 2367 if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull)) 2368 return true; 2369 2370 if (const auto *Call = dyn_cast<CallBase>(V)) { 2371 if (Call->isReturnNonNull()) 2372 return true; 2373 if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true)) 2374 return isKnownNonZero(RP, Depth, Q); 2375 } 2376 } 2377 2378 if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT)) 2379 return true; 2380 2381 // Check for recursive pointer simplifications. 2382 if (V->getType()->isPointerTy()) { 2383 // Look through bitcast operations, GEPs, and int2ptr instructions as they 2384 // do not alter the value, or at least not the nullness property of the 2385 // value, e.g., int2ptr is allowed to zero/sign extend the value. 2386 // 2387 // Note that we have to take special care to avoid looking through 2388 // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well 2389 // as casts that can alter the value, e.g., AddrSpaceCasts. 2390 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 2391 return isGEPKnownNonNull(GEP, Depth, Q); 2392 2393 if (auto *BCO = dyn_cast<BitCastOperator>(V)) 2394 return isKnownNonZero(BCO->getOperand(0), Depth, Q); 2395 2396 if (auto *I2P = dyn_cast<IntToPtrInst>(V)) 2397 if (Q.DL.getTypeSizeInBits(I2P->getSrcTy()).getFixedSize() <= 2398 Q.DL.getTypeSizeInBits(I2P->getDestTy()).getFixedSize()) 2399 return isKnownNonZero(I2P->getOperand(0), Depth, Q); 2400 } 2401 2402 // Similar to int2ptr above, we can look through ptr2int here if the cast 2403 // is a no-op or an extend and not a truncate. 2404 if (auto *P2I = dyn_cast<PtrToIntInst>(V)) 2405 if (Q.DL.getTypeSizeInBits(P2I->getSrcTy()).getFixedSize() <= 2406 Q.DL.getTypeSizeInBits(P2I->getDestTy()).getFixedSize()) 2407 return isKnownNonZero(P2I->getOperand(0), Depth, Q); 2408 2409 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL); 2410 2411 // X | Y != 0 if X != 0 or Y != 0. 2412 Value *X = nullptr, *Y = nullptr; 2413 if (match(V, m_Or(m_Value(X), m_Value(Y)))) 2414 return isKnownNonZero(X, DemandedElts, Depth, Q) || 2415 isKnownNonZero(Y, DemandedElts, Depth, Q); 2416 2417 // ext X != 0 if X != 0. 2418 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) 2419 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q); 2420 2421 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined 2422 // if the lowest bit is shifted off the end. 2423 if (match(V, m_Shl(m_Value(X), m_Value(Y)))) { 2424 // shl nuw can't remove any non-zero bits. 2425 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 2426 if (Q.IIQ.hasNoUnsignedWrap(BO)) 2427 return isKnownNonZero(X, Depth, Q); 2428 2429 KnownBits Known(BitWidth); 2430 computeKnownBits(X, DemandedElts, Known, Depth, Q); 2431 if (Known.One[0]) 2432 return true; 2433 } 2434 // shr X, Y != 0 if X is negative. Note that the value of the shift is not 2435 // defined if the sign bit is shifted off the end. 2436 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) { 2437 // shr exact can only shift out zero bits. 2438 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V); 2439 if (BO->isExact()) 2440 return isKnownNonZero(X, Depth, Q); 2441 2442 KnownBits Known = computeKnownBits(X, DemandedElts, Depth, Q); 2443 if (Known.isNegative()) 2444 return true; 2445 2446 // If the shifter operand is a constant, and all of the bits shifted 2447 // out are known to be zero, and X is known non-zero then at least one 2448 // non-zero bit must remain. 2449 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) { 2450 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1); 2451 // Is there a known one in the portion not shifted out? 2452 if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal) 2453 return true; 2454 // Are all the bits to be shifted out known zero? 2455 if (Known.countMinTrailingZeros() >= ShiftVal) 2456 return isKnownNonZero(X, DemandedElts, Depth, Q); 2457 } 2458 } 2459 // div exact can only produce a zero if the dividend is zero. 2460 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) { 2461 return isKnownNonZero(X, DemandedElts, Depth, Q); 2462 } 2463 // X + Y. 2464 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 2465 KnownBits XKnown = computeKnownBits(X, DemandedElts, Depth, Q); 2466 KnownBits YKnown = computeKnownBits(Y, DemandedElts, Depth, Q); 2467 2468 // If X and Y are both non-negative (as signed values) then their sum is not 2469 // zero unless both X and Y are zero. 2470 if (XKnown.isNonNegative() && YKnown.isNonNegative()) 2471 if (isKnownNonZero(X, DemandedElts, Depth, Q) || 2472 isKnownNonZero(Y, DemandedElts, Depth, Q)) 2473 return true; 2474 2475 // If X and Y are both negative (as signed values) then their sum is not 2476 // zero unless both X and Y equal INT_MIN. 2477 if (XKnown.isNegative() && YKnown.isNegative()) { 2478 APInt Mask = APInt::getSignedMaxValue(BitWidth); 2479 // The sign bit of X is set. If some other bit is set then X is not equal 2480 // to INT_MIN. 2481 if (XKnown.One.intersects(Mask)) 2482 return true; 2483 // The sign bit of Y is set. If some other bit is set then Y is not equal 2484 // to INT_MIN. 2485 if (YKnown.One.intersects(Mask)) 2486 return true; 2487 } 2488 2489 // The sum of a non-negative number and a power of two is not zero. 2490 if (XKnown.isNonNegative() && 2491 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q)) 2492 return true; 2493 if (YKnown.isNonNegative() && 2494 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q)) 2495 return true; 2496 } 2497 // X * Y. 2498 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) { 2499 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 2500 // If X and Y are non-zero then so is X * Y as long as the multiplication 2501 // does not overflow. 2502 if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) && 2503 isKnownNonZero(X, DemandedElts, Depth, Q) && 2504 isKnownNonZero(Y, DemandedElts, Depth, Q)) 2505 return true; 2506 } 2507 // (C ? X : Y) != 0 if X != 0 and Y != 0. 2508 else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { 2509 if (isKnownNonZero(SI->getTrueValue(), DemandedElts, Depth, Q) && 2510 isKnownNonZero(SI->getFalseValue(), DemandedElts, Depth, Q)) 2511 return true; 2512 } 2513 // PHI 2514 else if (const PHINode *PN = dyn_cast<PHINode>(V)) { 2515 if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN)) 2516 return true; 2517 2518 // Check if all incoming values are non-zero using recursion. 2519 Query RecQ = Q; 2520 unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1); 2521 return llvm::all_of(PN->operands(), [&](const Use &U) { 2522 if (U.get() == PN) 2523 return true; 2524 RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); 2525 return isKnownNonZero(U.get(), DemandedElts, NewDepth, RecQ); 2526 }); 2527 } 2528 // ExtractElement 2529 else if (const auto *EEI = dyn_cast<ExtractElementInst>(V)) { 2530 const Value *Vec = EEI->getVectorOperand(); 2531 const Value *Idx = EEI->getIndexOperand(); 2532 auto *CIdx = dyn_cast<ConstantInt>(Idx); 2533 if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) { 2534 unsigned NumElts = VecTy->getNumElements(); 2535 APInt DemandedVecElts = APInt::getAllOnesValue(NumElts); 2536 if (CIdx && CIdx->getValue().ult(NumElts)) 2537 DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); 2538 return isKnownNonZero(Vec, DemandedVecElts, Depth, Q); 2539 } 2540 } 2541 // Freeze 2542 else if (const FreezeInst *FI = dyn_cast<FreezeInst>(V)) { 2543 auto *Op = FI->getOperand(0); 2544 if (isKnownNonZero(Op, Depth, Q) && 2545 isGuaranteedNotToBePoison(Op, Q.AC, Q.CxtI, Q.DT, Depth)) 2546 return true; 2547 } 2548 2549 KnownBits Known(BitWidth); 2550 computeKnownBits(V, DemandedElts, Known, Depth, Q); 2551 return Known.One != 0; 2552 } 2553 2554 bool isKnownNonZero(const Value* V, unsigned Depth, const Query& Q) { 2555 // FIXME: We currently have no way to represent the DemandedElts of a scalable 2556 // vector 2557 if (isa<ScalableVectorType>(V->getType())) 2558 return false; 2559 2560 auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); 2561 APInt DemandedElts = 2562 FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); 2563 return isKnownNonZero(V, DemandedElts, Depth, Q); 2564 } 2565 2566 /// If the pair of operators are the same invertible function, return the 2567 /// the operands of the function corresponding to each input. Otherwise, 2568 /// return None. An invertible function is one that is 1-to-1 and maps 2569 /// every input value to exactly one output value. This is equivalent to 2570 /// saying that Op1 and Op2 are equal exactly when the specified pair of 2571 /// operands are equal, (except that Op1 and Op2 may be poison more often.) 2572 static Optional<std::pair<Value*, Value*>> 2573 getInvertibleOperands(const Operator *Op1, 2574 const Operator *Op2) { 2575 if (Op1->getOpcode() != Op2->getOpcode()) 2576 return None; 2577 2578 auto getOperands = [&](unsigned OpNum) -> auto { 2579 return std::make_pair(Op1->getOperand(OpNum), Op2->getOperand(OpNum)); 2580 }; 2581 2582 switch (Op1->getOpcode()) { 2583 default: 2584 break; 2585 case Instruction::Add: 2586 case Instruction::Sub: 2587 if (Op1->getOperand(0) == Op2->getOperand(0)) 2588 return getOperands(1); 2589 if (Op1->getOperand(1) == Op2->getOperand(1)) 2590 return getOperands(0); 2591 break; 2592 case Instruction::Mul: { 2593 // invertible if A * B == (A * B) mod 2^N where A, and B are integers 2594 // and N is the bitwdith. The nsw case is non-obvious, but proven by 2595 // alive2: https://alive2.llvm.org/ce/z/Z6D5qK 2596 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); 2597 auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); 2598 if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && 2599 (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) 2600 break; 2601 2602 // Assume operand order has been canonicalized 2603 if (Op1->getOperand(1) == Op2->getOperand(1) && 2604 isa<ConstantInt>(Op1->getOperand(1)) && 2605 !cast<ConstantInt>(Op1->getOperand(1))->isZero()) 2606 return getOperands(0); 2607 break; 2608 } 2609 case Instruction::Shl: { 2610 // Same as multiplies, with the difference that we don't need to check 2611 // for a non-zero multiply. Shifts always multiply by non-zero. 2612 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); 2613 auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); 2614 if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && 2615 (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) 2616 break; 2617 2618 if (Op1->getOperand(1) == Op2->getOperand(1)) 2619 return getOperands(0); 2620 break; 2621 } 2622 case Instruction::AShr: 2623 case Instruction::LShr: { 2624 auto *PEO1 = cast<PossiblyExactOperator>(Op1); 2625 auto *PEO2 = cast<PossiblyExactOperator>(Op2); 2626 if (!PEO1->isExact() || !PEO2->isExact()) 2627 break; 2628 2629 if (Op1->getOperand(1) == Op2->getOperand(1)) 2630 return getOperands(0); 2631 break; 2632 } 2633 case Instruction::SExt: 2634 case Instruction::ZExt: 2635 if (Op1->getOperand(0)->getType() == Op2->getOperand(0)->getType()) 2636 return getOperands(0); 2637 break; 2638 case Instruction::PHI: { 2639 const PHINode *PN1 = cast<PHINode>(Op1); 2640 const PHINode *PN2 = cast<PHINode>(Op2); 2641 2642 // If PN1 and PN2 are both recurrences, can we prove the entire recurrences 2643 // are a single invertible function of the start values? Note that repeated 2644 // application of an invertible function is also invertible 2645 BinaryOperator *BO1 = nullptr; 2646 Value *Start1 = nullptr, *Step1 = nullptr; 2647 BinaryOperator *BO2 = nullptr; 2648 Value *Start2 = nullptr, *Step2 = nullptr; 2649 if (PN1->getParent() != PN2->getParent() || 2650 !matchSimpleRecurrence(PN1, BO1, Start1, Step1) || 2651 !matchSimpleRecurrence(PN2, BO2, Start2, Step2)) 2652 break; 2653 2654 auto Values = getInvertibleOperands(cast<Operator>(BO1), 2655 cast<Operator>(BO2)); 2656 if (!Values) 2657 break; 2658 2659 // We have to be careful of mutually defined recurrences here. Ex: 2660 // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V 2661 // * X_i = Y_i = X_(i-1) OP Y_(i-1) 2662 // The invertibility of these is complicated, and not worth reasoning 2663 // about (yet?). 2664 if (Values->first != PN1 || Values->second != PN2) 2665 break; 2666 2667 return std::make_pair(Start1, Start2); 2668 } 2669 } 2670 return None; 2671 } 2672 2673 /// Return true if V2 == V1 + X, where X is known non-zero. 2674 static bool isAddOfNonZero(const Value *V1, const Value *V2, unsigned Depth, 2675 const Query &Q) { 2676 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1); 2677 if (!BO || BO->getOpcode() != Instruction::Add) 2678 return false; 2679 Value *Op = nullptr; 2680 if (V2 == BO->getOperand(0)) 2681 Op = BO->getOperand(1); 2682 else if (V2 == BO->getOperand(1)) 2683 Op = BO->getOperand(0); 2684 else 2685 return false; 2686 return isKnownNonZero(Op, Depth + 1, Q); 2687 } 2688 2689 /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and 2690 /// the multiplication is nuw or nsw. 2691 static bool isNonEqualMul(const Value *V1, const Value *V2, unsigned Depth, 2692 const Query &Q) { 2693 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { 2694 const APInt *C; 2695 return match(OBO, m_Mul(m_Specific(V1), m_APInt(C))) && 2696 (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && 2697 !C->isNullValue() && !C->isOneValue() && 2698 isKnownNonZero(V1, Depth + 1, Q); 2699 } 2700 return false; 2701 } 2702 2703 /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and 2704 /// the shift is nuw or nsw. 2705 static bool isNonEqualShl(const Value *V1, const Value *V2, unsigned Depth, 2706 const Query &Q) { 2707 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { 2708 const APInt *C; 2709 return match(OBO, m_Shl(m_Specific(V1), m_APInt(C))) && 2710 (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && 2711 !C->isNullValue() && isKnownNonZero(V1, Depth + 1, Q); 2712 } 2713 return false; 2714 } 2715 2716 static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2, 2717 unsigned Depth, const Query &Q) { 2718 // Check two PHIs are in same block. 2719 if (PN1->getParent() != PN2->getParent()) 2720 return false; 2721 2722 SmallPtrSet<const BasicBlock *, 8> VisitedBBs; 2723 bool UsedFullRecursion = false; 2724 for (const BasicBlock *IncomBB : PN1->blocks()) { 2725 if (!VisitedBBs.insert(IncomBB).second) 2726 continue; // Don't reprocess blocks that we have dealt with already. 2727 const Value *IV1 = PN1->getIncomingValueForBlock(IncomBB); 2728 const Value *IV2 = PN2->getIncomingValueForBlock(IncomBB); 2729 const APInt *C1, *C2; 2730 if (match(IV1, m_APInt(C1)) && match(IV2, m_APInt(C2)) && *C1 != *C2) 2731 continue; 2732 2733 // Only one pair of phi operands is allowed for full recursion. 2734 if (UsedFullRecursion) 2735 return false; 2736 2737 Query RecQ = Q; 2738 RecQ.CxtI = IncomBB->getTerminator(); 2739 if (!isKnownNonEqual(IV1, IV2, Depth + 1, RecQ)) 2740 return false; 2741 UsedFullRecursion = true; 2742 } 2743 return true; 2744 } 2745 2746 /// Return true if it is known that V1 != V2. 2747 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, 2748 const Query &Q) { 2749 if (V1 == V2) 2750 return false; 2751 if (V1->getType() != V2->getType()) 2752 // We can't look through casts yet. 2753 return false; 2754 2755 if (Depth >= MaxAnalysisRecursionDepth) 2756 return false; 2757 2758 // See if we can recurse through (exactly one of) our operands. This 2759 // requires our operation be 1-to-1 and map every input value to exactly 2760 // one output value. Such an operation is invertible. 2761 auto *O1 = dyn_cast<Operator>(V1); 2762 auto *O2 = dyn_cast<Operator>(V2); 2763 if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) { 2764 if (auto Values = getInvertibleOperands(O1, O2)) 2765 return isKnownNonEqual(Values->first, Values->second, Depth + 1, Q); 2766 2767 if (const PHINode *PN1 = dyn_cast<PHINode>(V1)) { 2768 const PHINode *PN2 = cast<PHINode>(V2); 2769 // FIXME: This is missing a generalization to handle the case where one is 2770 // a PHI and another one isn't. 2771 if (isNonEqualPHIs(PN1, PN2, Depth, Q)) 2772 return true; 2773 }; 2774 } 2775 2776 if (isAddOfNonZero(V1, V2, Depth, Q) || isAddOfNonZero(V2, V1, Depth, Q)) 2777 return true; 2778 2779 if (isNonEqualMul(V1, V2, Depth, Q) || isNonEqualMul(V2, V1, Depth, Q)) 2780 return true; 2781 2782 if (isNonEqualShl(V1, V2, Depth, Q) || isNonEqualShl(V2, V1, Depth, Q)) 2783 return true; 2784 2785 if (V1->getType()->isIntOrIntVectorTy()) { 2786 // Are any known bits in V1 contradictory to known bits in V2? If V1 2787 // has a known zero where V2 has a known one, they must not be equal. 2788 KnownBits Known1 = computeKnownBits(V1, Depth, Q); 2789 KnownBits Known2 = computeKnownBits(V2, Depth, Q); 2790 2791 if (Known1.Zero.intersects(Known2.One) || 2792 Known2.Zero.intersects(Known1.One)) 2793 return true; 2794 } 2795 return false; 2796 } 2797 2798 /// Return true if 'V & Mask' is known to be zero. We use this predicate to 2799 /// simplify operations downstream. Mask is known to be zero for bits that V 2800 /// cannot have. 2801 /// 2802 /// This function is defined on values with integer type, values with pointer 2803 /// type, and vectors of integers. In the case 2804 /// where V is a vector, the mask, known zero, and known one values are the 2805 /// same width as the vector element, and the bit is set only if it is true 2806 /// for all of the elements in the vector. 2807 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, 2808 const Query &Q) { 2809 KnownBits Known(Mask.getBitWidth()); 2810 computeKnownBits(V, Known, Depth, Q); 2811 return Mask.isSubsetOf(Known.Zero); 2812 } 2813 2814 // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow). 2815 // Returns the input and lower/upper bounds. 2816 static bool isSignedMinMaxClamp(const Value *Select, const Value *&In, 2817 const APInt *&CLow, const APInt *&CHigh) { 2818 assert(isa<Operator>(Select) && 2819 cast<Operator>(Select)->getOpcode() == Instruction::Select && 2820 "Input should be a Select!"); 2821 2822 const Value *LHS = nullptr, *RHS = nullptr; 2823 SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor; 2824 if (SPF != SPF_SMAX && SPF != SPF_SMIN) 2825 return false; 2826 2827 if (!match(RHS, m_APInt(CLow))) 2828 return false; 2829 2830 const Value *LHS2 = nullptr, *RHS2 = nullptr; 2831 SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor; 2832 if (getInverseMinMaxFlavor(SPF) != SPF2) 2833 return false; 2834 2835 if (!match(RHS2, m_APInt(CHigh))) 2836 return false; 2837 2838 if (SPF == SPF_SMIN) 2839 std::swap(CLow, CHigh); 2840 2841 In = LHS2; 2842 return CLow->sle(*CHigh); 2843 } 2844 2845 /// For vector constants, loop over the elements and find the constant with the 2846 /// minimum number of sign bits. Return 0 if the value is not a vector constant 2847 /// or if any element was not analyzed; otherwise, return the count for the 2848 /// element with the minimum number of sign bits. 2849 static unsigned computeNumSignBitsVectorConstant(const Value *V, 2850 const APInt &DemandedElts, 2851 unsigned TyBits) { 2852 const auto *CV = dyn_cast<Constant>(V); 2853 if (!CV || !isa<FixedVectorType>(CV->getType())) 2854 return 0; 2855 2856 unsigned MinSignBits = TyBits; 2857 unsigned NumElts = cast<FixedVectorType>(CV->getType())->getNumElements(); 2858 for (unsigned i = 0; i != NumElts; ++i) { 2859 if (!DemandedElts[i]) 2860 continue; 2861 // If we find a non-ConstantInt, bail out. 2862 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i)); 2863 if (!Elt) 2864 return 0; 2865 2866 MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits()); 2867 } 2868 2869 return MinSignBits; 2870 } 2871 2872 static unsigned ComputeNumSignBitsImpl(const Value *V, 2873 const APInt &DemandedElts, 2874 unsigned Depth, const Query &Q); 2875 2876 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, 2877 unsigned Depth, const Query &Q) { 2878 unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q); 2879 assert(Result > 0 && "At least one sign bit needs to be present!"); 2880 return Result; 2881 } 2882 2883 /// Return the number of times the sign bit of the register is replicated into 2884 /// the other bits. We know that at least 1 bit is always equal to the sign bit 2885 /// (itself), but other cases can give us information. For example, immediately 2886 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each 2887 /// other, so we return 3. For vectors, return the number of sign bits for the 2888 /// vector element with the minimum number of known sign bits of the demanded 2889 /// elements in the vector specified by DemandedElts. 2890 static unsigned ComputeNumSignBitsImpl(const Value *V, 2891 const APInt &DemandedElts, 2892 unsigned Depth, const Query &Q) { 2893 Type *Ty = V->getType(); 2894 2895 // FIXME: We currently have no way to represent the DemandedElts of a scalable 2896 // vector 2897 if (isa<ScalableVectorType>(Ty)) 2898 return 1; 2899 2900 #ifndef NDEBUG 2901 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 2902 2903 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { 2904 assert( 2905 FVTy->getNumElements() == DemandedElts.getBitWidth() && 2906 "DemandedElt width should equal the fixed vector number of elements"); 2907 } else { 2908 assert(DemandedElts == APInt(1, 1) && 2909 "DemandedElt width should be 1 for scalars"); 2910 } 2911 #endif 2912 2913 // We return the minimum number of sign bits that are guaranteed to be present 2914 // in V, so for undef we have to conservatively return 1. We don't have the 2915 // same behavior for poison though -- that's a FIXME today. 2916 2917 Type *ScalarTy = Ty->getScalarType(); 2918 unsigned TyBits = ScalarTy->isPointerTy() ? 2919 Q.DL.getPointerTypeSizeInBits(ScalarTy) : 2920 Q.DL.getTypeSizeInBits(ScalarTy); 2921 2922 unsigned Tmp, Tmp2; 2923 unsigned FirstAnswer = 1; 2924 2925 // Note that ConstantInt is handled by the general computeKnownBits case 2926 // below. 2927 2928 if (Depth == MaxAnalysisRecursionDepth) 2929 return 1; 2930 2931 if (auto *U = dyn_cast<Operator>(V)) { 2932 switch (Operator::getOpcode(V)) { 2933 default: break; 2934 case Instruction::SExt: 2935 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits(); 2936 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp; 2937 2938 case Instruction::SDiv: { 2939 const APInt *Denominator; 2940 // sdiv X, C -> adds log(C) sign bits. 2941 if (match(U->getOperand(1), m_APInt(Denominator))) { 2942 2943 // Ignore non-positive denominator. 2944 if (!Denominator->isStrictlyPositive()) 2945 break; 2946 2947 // Calculate the incoming numerator bits. 2948 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2949 2950 // Add floor(log(C)) bits to the numerator bits. 2951 return std::min(TyBits, NumBits + Denominator->logBase2()); 2952 } 2953 break; 2954 } 2955 2956 case Instruction::SRem: { 2957 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2958 2959 const APInt *Denominator; 2960 // srem X, C -> we know that the result is within [-C+1,C) when C is a 2961 // positive constant. This let us put a lower bound on the number of sign 2962 // bits. 2963 if (match(U->getOperand(1), m_APInt(Denominator))) { 2964 2965 // Ignore non-positive denominator. 2966 if (Denominator->isStrictlyPositive()) { 2967 // Calculate the leading sign bit constraints by examining the 2968 // denominator. Given that the denominator is positive, there are two 2969 // cases: 2970 // 2971 // 1. The numerator is positive. The result range is [0,C) and 2972 // [0,C) u< (1 << ceilLogBase2(C)). 2973 // 2974 // 2. The numerator is negative. Then the result range is (-C,0] and 2975 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). 2976 // 2977 // Thus a lower bound on the number of sign bits is `TyBits - 2978 // ceilLogBase2(C)`. 2979 2980 unsigned ResBits = TyBits - Denominator->ceilLogBase2(); 2981 Tmp = std::max(Tmp, ResBits); 2982 } 2983 } 2984 return Tmp; 2985 } 2986 2987 case Instruction::AShr: { 2988 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2989 // ashr X, C -> adds C sign bits. Vectors too. 2990 const APInt *ShAmt; 2991 if (match(U->getOperand(1), m_APInt(ShAmt))) { 2992 if (ShAmt->uge(TyBits)) 2993 break; // Bad shift. 2994 unsigned ShAmtLimited = ShAmt->getZExtValue(); 2995 Tmp += ShAmtLimited; 2996 if (Tmp > TyBits) Tmp = TyBits; 2997 } 2998 return Tmp; 2999 } 3000 case Instruction::Shl: { 3001 const APInt *ShAmt; 3002 if (match(U->getOperand(1), m_APInt(ShAmt))) { 3003 // shl destroys sign bits. 3004 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 3005 if (ShAmt->uge(TyBits) || // Bad shift. 3006 ShAmt->uge(Tmp)) break; // Shifted all sign bits out. 3007 Tmp2 = ShAmt->getZExtValue(); 3008 return Tmp - Tmp2; 3009 } 3010 break; 3011 } 3012 case Instruction::And: 3013 case Instruction::Or: 3014 case Instruction::Xor: // NOT is handled here. 3015 // Logical binary ops preserve the number of sign bits at the worst. 3016 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 3017 if (Tmp != 1) { 3018 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 3019 FirstAnswer = std::min(Tmp, Tmp2); 3020 // We computed what we know about the sign bits as our first 3021 // answer. Now proceed to the generic code that uses 3022 // computeKnownBits, and pick whichever answer is better. 3023 } 3024 break; 3025 3026 case Instruction::Select: { 3027 // If we have a clamp pattern, we know that the number of sign bits will 3028 // be the minimum of the clamp min/max range. 3029 const Value *X; 3030 const APInt *CLow, *CHigh; 3031 if (isSignedMinMaxClamp(U, X, CLow, CHigh)) 3032 return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits()); 3033 3034 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 3035 if (Tmp == 1) break; 3036 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q); 3037 return std::min(Tmp, Tmp2); 3038 } 3039 3040 case Instruction::Add: 3041 // Add can have at most one carry bit. Thus we know that the output 3042 // is, at worst, one more bit than the inputs. 3043 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 3044 if (Tmp == 1) break; 3045 3046 // Special case decrementing a value (ADD X, -1): 3047 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1))) 3048 if (CRHS->isAllOnesValue()) { 3049 KnownBits Known(TyBits); 3050 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q); 3051 3052 // If the input is known to be 0 or 1, the output is 0/-1, which is 3053 // all sign bits set. 3054 if ((Known.Zero | 1).isAllOnesValue()) 3055 return TyBits; 3056 3057 // If we are subtracting one from a positive number, there is no carry 3058 // out of the result. 3059 if (Known.isNonNegative()) 3060 return Tmp; 3061 } 3062 3063 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 3064 if (Tmp2 == 1) break; 3065 return std::min(Tmp, Tmp2) - 1; 3066 3067 case Instruction::Sub: 3068 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 3069 if (Tmp2 == 1) break; 3070 3071 // Handle NEG. 3072 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0))) 3073 if (CLHS->isNullValue()) { 3074 KnownBits Known(TyBits); 3075 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q); 3076 // If the input is known to be 0 or 1, the output is 0/-1, which is 3077 // all sign bits set. 3078 if ((Known.Zero | 1).isAllOnesValue()) 3079 return TyBits; 3080 3081 // If the input is known to be positive (the sign bit is known clear), 3082 // the output of the NEG has the same number of sign bits as the 3083 // input. 3084 if (Known.isNonNegative()) 3085 return Tmp2; 3086 3087 // Otherwise, we treat this like a SUB. 3088 } 3089 3090 // Sub can have at most one carry bit. Thus we know that the output 3091 // is, at worst, one more bit than the inputs. 3092 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 3093 if (Tmp == 1) break; 3094 return std::min(Tmp, Tmp2) - 1; 3095 3096 case Instruction::Mul: { 3097 // The output of the Mul can be at most twice the valid bits in the 3098 // inputs. 3099 unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 3100 if (SignBitsOp0 == 1) break; 3101 unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 3102 if (SignBitsOp1 == 1) break; 3103 unsigned OutValidBits = 3104 (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1); 3105 return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1; 3106 } 3107 3108 case Instruction::PHI: { 3109 const PHINode *PN = cast<PHINode>(U); 3110 unsigned NumIncomingValues = PN->getNumIncomingValues(); 3111 // Don't analyze large in-degree PHIs. 3112 if (NumIncomingValues > 4) break; 3113 // Unreachable blocks may have zero-operand PHI nodes. 3114 if (NumIncomingValues == 0) break; 3115 3116 // Take the minimum of all incoming values. This can't infinitely loop 3117 // because of our depth threshold. 3118 Query RecQ = Q; 3119 Tmp = TyBits; 3120 for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) { 3121 if (Tmp == 1) return Tmp; 3122 RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator(); 3123 Tmp = std::min( 3124 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, RecQ)); 3125 } 3126 return Tmp; 3127 } 3128 3129 case Instruction::Trunc: 3130 // FIXME: it's tricky to do anything useful for this, but it is an 3131 // important case for targets like X86. 3132 break; 3133 3134 case Instruction::ExtractElement: 3135 // Look through extract element. At the moment we keep this simple and 3136 // skip tracking the specific element. But at least we might find 3137 // information valid for all elements of the vector (for example if vector 3138 // is sign extended, shifted, etc). 3139 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 3140 3141 case Instruction::ShuffleVector: { 3142 // Collect the minimum number of sign bits that are shared by every vector 3143 // element referenced by the shuffle. 3144 auto *Shuf = dyn_cast<ShuffleVectorInst>(U); 3145 if (!Shuf) { 3146 // FIXME: Add support for shufflevector constant expressions. 3147 return 1; 3148 } 3149 APInt DemandedLHS, DemandedRHS; 3150 // For undef elements, we don't know anything about the common state of 3151 // the shuffle result. 3152 if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) 3153 return 1; 3154 Tmp = std::numeric_limits<unsigned>::max(); 3155 if (!!DemandedLHS) { 3156 const Value *LHS = Shuf->getOperand(0); 3157 Tmp = ComputeNumSignBits(LHS, DemandedLHS, Depth + 1, Q); 3158 } 3159 // If we don't know anything, early out and try computeKnownBits 3160 // fall-back. 3161 if (Tmp == 1) 3162 break; 3163 if (!!DemandedRHS) { 3164 const Value *RHS = Shuf->getOperand(1); 3165 Tmp2 = ComputeNumSignBits(RHS, DemandedRHS, Depth + 1, Q); 3166 Tmp = std::min(Tmp, Tmp2); 3167 } 3168 // If we don't know anything, early out and try computeKnownBits 3169 // fall-back. 3170 if (Tmp == 1) 3171 break; 3172 assert(Tmp <= TyBits && "Failed to determine minimum sign bits"); 3173 return Tmp; 3174 } 3175 case Instruction::Call: { 3176 if (const auto *II = dyn_cast<IntrinsicInst>(U)) { 3177 switch (II->getIntrinsicID()) { 3178 default: break; 3179 case Intrinsic::abs: 3180 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 3181 if (Tmp == 1) break; 3182 3183 // Absolute value reduces number of sign bits by at most 1. 3184 return Tmp - 1; 3185 } 3186 } 3187 } 3188 } 3189 } 3190 3191 // Finally, if we can prove that the top bits of the result are 0's or 1's, 3192 // use this information. 3193 3194 // If we can examine all elements of a vector constant successfully, we're 3195 // done (we can't do any better than that). If not, keep trying. 3196 if (unsigned VecSignBits = 3197 computeNumSignBitsVectorConstant(V, DemandedElts, TyBits)) 3198 return VecSignBits; 3199 3200 KnownBits Known(TyBits); 3201 computeKnownBits(V, DemandedElts, Known, Depth, Q); 3202 3203 // If we know that the sign bit is either zero or one, determine the number of 3204 // identical bits in the top of the input value. 3205 return std::max(FirstAnswer, Known.countMinSignBits()); 3206 } 3207 3208 /// This function computes the integer multiple of Base that equals V. 3209 /// If successful, it returns true and returns the multiple in 3210 /// Multiple. If unsuccessful, it returns false. It looks 3211 /// through SExt instructions only if LookThroughSExt is true. 3212 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple, 3213 bool LookThroughSExt, unsigned Depth) { 3214 assert(V && "No Value?"); 3215 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 3216 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!"); 3217 3218 Type *T = V->getType(); 3219 3220 ConstantInt *CI = dyn_cast<ConstantInt>(V); 3221 3222 if (Base == 0) 3223 return false; 3224 3225 if (Base == 1) { 3226 Multiple = V; 3227 return true; 3228 } 3229 3230 ConstantExpr *CO = dyn_cast<ConstantExpr>(V); 3231 Constant *BaseVal = ConstantInt::get(T, Base); 3232 if (CO && CO == BaseVal) { 3233 // Multiple is 1. 3234 Multiple = ConstantInt::get(T, 1); 3235 return true; 3236 } 3237 3238 if (CI && CI->getZExtValue() % Base == 0) { 3239 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base); 3240 return true; 3241 } 3242 3243 if (Depth == MaxAnalysisRecursionDepth) return false; 3244 3245 Operator *I = dyn_cast<Operator>(V); 3246 if (!I) return false; 3247 3248 switch (I->getOpcode()) { 3249 default: break; 3250 case Instruction::SExt: 3251 if (!LookThroughSExt) return false; 3252 // otherwise fall through to ZExt 3253 LLVM_FALLTHROUGH; 3254 case Instruction::ZExt: 3255 return ComputeMultiple(I->getOperand(0), Base, Multiple, 3256 LookThroughSExt, Depth+1); 3257 case Instruction::Shl: 3258 case Instruction::Mul: { 3259 Value *Op0 = I->getOperand(0); 3260 Value *Op1 = I->getOperand(1); 3261 3262 if (I->getOpcode() == Instruction::Shl) { 3263 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1); 3264 if (!Op1CI) return false; 3265 // Turn Op0 << Op1 into Op0 * 2^Op1 3266 APInt Op1Int = Op1CI->getValue(); 3267 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1); 3268 APInt API(Op1Int.getBitWidth(), 0); 3269 API.setBit(BitToSet); 3270 Op1 = ConstantInt::get(V->getContext(), API); 3271 } 3272 3273 Value *Mul0 = nullptr; 3274 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) { 3275 if (Constant *Op1C = dyn_cast<Constant>(Op1)) 3276 if (Constant *MulC = dyn_cast<Constant>(Mul0)) { 3277 if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() < 3278 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3279 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType()); 3280 if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() > 3281 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3282 MulC = ConstantExpr::getZExt(MulC, Op1C->getType()); 3283 3284 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1) 3285 Multiple = ConstantExpr::getMul(MulC, Op1C); 3286 return true; 3287 } 3288 3289 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0)) 3290 if (Mul0CI->getValue() == 1) { 3291 // V == Base * Op1, so return Op1 3292 Multiple = Op1; 3293 return true; 3294 } 3295 } 3296 3297 Value *Mul1 = nullptr; 3298 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) { 3299 if (Constant *Op0C = dyn_cast<Constant>(Op0)) 3300 if (Constant *MulC = dyn_cast<Constant>(Mul1)) { 3301 if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() < 3302 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3303 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType()); 3304 if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() > 3305 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3306 MulC = ConstantExpr::getZExt(MulC, Op0C->getType()); 3307 3308 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0) 3309 Multiple = ConstantExpr::getMul(MulC, Op0C); 3310 return true; 3311 } 3312 3313 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1)) 3314 if (Mul1CI->getValue() == 1) { 3315 // V == Base * Op0, so return Op0 3316 Multiple = Op0; 3317 return true; 3318 } 3319 } 3320 } 3321 } 3322 3323 // We could not determine if V is a multiple of Base. 3324 return false; 3325 } 3326 3327 Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB, 3328 const TargetLibraryInfo *TLI) { 3329 const Function *F = CB.getCalledFunction(); 3330 if (!F) 3331 return Intrinsic::not_intrinsic; 3332 3333 if (F->isIntrinsic()) 3334 return F->getIntrinsicID(); 3335 3336 // We are going to infer semantics of a library function based on mapping it 3337 // to an LLVM intrinsic. Check that the library function is available from 3338 // this callbase and in this environment. 3339 LibFunc Func; 3340 if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) || 3341 !CB.onlyReadsMemory()) 3342 return Intrinsic::not_intrinsic; 3343 3344 switch (Func) { 3345 default: 3346 break; 3347 case LibFunc_sin: 3348 case LibFunc_sinf: 3349 case LibFunc_sinl: 3350 return Intrinsic::sin; 3351 case LibFunc_cos: 3352 case LibFunc_cosf: 3353 case LibFunc_cosl: 3354 return Intrinsic::cos; 3355 case LibFunc_exp: 3356 case LibFunc_expf: 3357 case LibFunc_expl: 3358 return Intrinsic::exp; 3359 case LibFunc_exp2: 3360 case LibFunc_exp2f: 3361 case LibFunc_exp2l: 3362 return Intrinsic::exp2; 3363 case LibFunc_log: 3364 case LibFunc_logf: 3365 case LibFunc_logl: 3366 return Intrinsic::log; 3367 case LibFunc_log10: 3368 case LibFunc_log10f: 3369 case LibFunc_log10l: 3370 return Intrinsic::log10; 3371 case LibFunc_log2: 3372 case LibFunc_log2f: 3373 case LibFunc_log2l: 3374 return Intrinsic::log2; 3375 case LibFunc_fabs: 3376 case LibFunc_fabsf: 3377 case LibFunc_fabsl: 3378 return Intrinsic::fabs; 3379 case LibFunc_fmin: 3380 case LibFunc_fminf: 3381 case LibFunc_fminl: 3382 return Intrinsic::minnum; 3383 case LibFunc_fmax: 3384 case LibFunc_fmaxf: 3385 case LibFunc_fmaxl: 3386 return Intrinsic::maxnum; 3387 case LibFunc_copysign: 3388 case LibFunc_copysignf: 3389 case LibFunc_copysignl: 3390 return Intrinsic::copysign; 3391 case LibFunc_floor: 3392 case LibFunc_floorf: 3393 case LibFunc_floorl: 3394 return Intrinsic::floor; 3395 case LibFunc_ceil: 3396 case LibFunc_ceilf: 3397 case LibFunc_ceill: 3398 return Intrinsic::ceil; 3399 case LibFunc_trunc: 3400 case LibFunc_truncf: 3401 case LibFunc_truncl: 3402 return Intrinsic::trunc; 3403 case LibFunc_rint: 3404 case LibFunc_rintf: 3405 case LibFunc_rintl: 3406 return Intrinsic::rint; 3407 case LibFunc_nearbyint: 3408 case LibFunc_nearbyintf: 3409 case LibFunc_nearbyintl: 3410 return Intrinsic::nearbyint; 3411 case LibFunc_round: 3412 case LibFunc_roundf: 3413 case LibFunc_roundl: 3414 return Intrinsic::round; 3415 case LibFunc_roundeven: 3416 case LibFunc_roundevenf: 3417 case LibFunc_roundevenl: 3418 return Intrinsic::roundeven; 3419 case LibFunc_pow: 3420 case LibFunc_powf: 3421 case LibFunc_powl: 3422 return Intrinsic::pow; 3423 case LibFunc_sqrt: 3424 case LibFunc_sqrtf: 3425 case LibFunc_sqrtl: 3426 return Intrinsic::sqrt; 3427 } 3428 3429 return Intrinsic::not_intrinsic; 3430 } 3431 3432 /// Return true if we can prove that the specified FP value is never equal to 3433 /// -0.0. 3434 /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee 3435 /// that a value is not -0.0. It only guarantees that -0.0 may be treated 3436 /// the same as +0.0 in floating-point ops. 3437 /// 3438 /// NOTE: this function will need to be revisited when we support non-default 3439 /// rounding modes! 3440 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, 3441 unsigned Depth) { 3442 if (auto *CFP = dyn_cast<ConstantFP>(V)) 3443 return !CFP->getValueAPF().isNegZero(); 3444 3445 if (Depth == MaxAnalysisRecursionDepth) 3446 return false; 3447 3448 auto *Op = dyn_cast<Operator>(V); 3449 if (!Op) 3450 return false; 3451 3452 // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0. 3453 if (match(Op, m_FAdd(m_Value(), m_PosZeroFP()))) 3454 return true; 3455 3456 // sitofp and uitofp turn into +0.0 for zero. 3457 if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) 3458 return true; 3459 3460 if (auto *Call = dyn_cast<CallInst>(Op)) { 3461 Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI); 3462 switch (IID) { 3463 default: 3464 break; 3465 // sqrt(-0.0) = -0.0, no other negative results are possible. 3466 case Intrinsic::sqrt: 3467 case Intrinsic::canonicalize: 3468 return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1); 3469 // fabs(x) != -0.0 3470 case Intrinsic::fabs: 3471 return true; 3472 } 3473 } 3474 3475 return false; 3476 } 3477 3478 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a 3479 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign 3480 /// bit despite comparing equal. 3481 static bool cannotBeOrderedLessThanZeroImpl(const Value *V, 3482 const TargetLibraryInfo *TLI, 3483 bool SignBitOnly, 3484 unsigned Depth) { 3485 // TODO: This function does not do the right thing when SignBitOnly is true 3486 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform 3487 // which flips the sign bits of NaNs. See 3488 // https://llvm.org/bugs/show_bug.cgi?id=31702. 3489 3490 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 3491 return !CFP->getValueAPF().isNegative() || 3492 (!SignBitOnly && CFP->getValueAPF().isZero()); 3493 } 3494 3495 // Handle vector of constants. 3496 if (auto *CV = dyn_cast<Constant>(V)) { 3497 if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) { 3498 unsigned NumElts = CVFVTy->getNumElements(); 3499 for (unsigned i = 0; i != NumElts; ++i) { 3500 auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i)); 3501 if (!CFP) 3502 return false; 3503 if (CFP->getValueAPF().isNegative() && 3504 (SignBitOnly || !CFP->getValueAPF().isZero())) 3505 return false; 3506 } 3507 3508 // All non-negative ConstantFPs. 3509 return true; 3510 } 3511 } 3512 3513 if (Depth == MaxAnalysisRecursionDepth) 3514 return false; 3515 3516 const Operator *I = dyn_cast<Operator>(V); 3517 if (!I) 3518 return false; 3519 3520 switch (I->getOpcode()) { 3521 default: 3522 break; 3523 // Unsigned integers are always nonnegative. 3524 case Instruction::UIToFP: 3525 return true; 3526 case Instruction::FMul: 3527 case Instruction::FDiv: 3528 // X * X is always non-negative or a NaN. 3529 // X / X is always exactly 1.0 or a NaN. 3530 if (I->getOperand(0) == I->getOperand(1) && 3531 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs())) 3532 return true; 3533 3534 LLVM_FALLTHROUGH; 3535 case Instruction::FAdd: 3536 case Instruction::FRem: 3537 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3538 Depth + 1) && 3539 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3540 Depth + 1); 3541 case Instruction::Select: 3542 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3543 Depth + 1) && 3544 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, 3545 Depth + 1); 3546 case Instruction::FPExt: 3547 case Instruction::FPTrunc: 3548 // Widening/narrowing never change sign. 3549 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3550 Depth + 1); 3551 case Instruction::ExtractElement: 3552 // Look through extract element. At the moment we keep this simple and skip 3553 // tracking the specific element. But at least we might find information 3554 // valid for all elements of the vector. 3555 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3556 Depth + 1); 3557 case Instruction::Call: 3558 const auto *CI = cast<CallInst>(I); 3559 Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI); 3560 switch (IID) { 3561 default: 3562 break; 3563 case Intrinsic::maxnum: { 3564 Value *V0 = I->getOperand(0), *V1 = I->getOperand(1); 3565 auto isPositiveNum = [&](Value *V) { 3566 if (SignBitOnly) { 3567 // With SignBitOnly, this is tricky because the result of 3568 // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is 3569 // a constant strictly greater than 0.0. 3570 const APFloat *C; 3571 return match(V, m_APFloat(C)) && 3572 *C > APFloat::getZero(C->getSemantics()); 3573 } 3574 3575 // -0.0 compares equal to 0.0, so if this operand is at least -0.0, 3576 // maxnum can't be ordered-less-than-zero. 3577 return isKnownNeverNaN(V, TLI) && 3578 cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1); 3579 }; 3580 3581 // TODO: This could be improved. We could also check that neither operand 3582 // has its sign bit set (and at least 1 is not-NAN?). 3583 return isPositiveNum(V0) || isPositiveNum(V1); 3584 } 3585 3586 case Intrinsic::maximum: 3587 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3588 Depth + 1) || 3589 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3590 Depth + 1); 3591 case Intrinsic::minnum: 3592 case Intrinsic::minimum: 3593 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3594 Depth + 1) && 3595 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3596 Depth + 1); 3597 case Intrinsic::exp: 3598 case Intrinsic::exp2: 3599 case Intrinsic::fabs: 3600 return true; 3601 3602 case Intrinsic::sqrt: 3603 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0. 3604 if (!SignBitOnly) 3605 return true; 3606 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() || 3607 CannotBeNegativeZero(CI->getOperand(0), TLI)); 3608 3609 case Intrinsic::powi: 3610 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) { 3611 // powi(x,n) is non-negative if n is even. 3612 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0) 3613 return true; 3614 } 3615 // TODO: This is not correct. Given that exp is an integer, here are the 3616 // ways that pow can return a negative value: 3617 // 3618 // pow(x, exp) --> negative if exp is odd and x is negative. 3619 // pow(-0, exp) --> -inf if exp is negative odd. 3620 // pow(-0, exp) --> -0 if exp is positive odd. 3621 // pow(-inf, exp) --> -0 if exp is negative odd. 3622 // pow(-inf, exp) --> -inf if exp is positive odd. 3623 // 3624 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN, 3625 // but we must return false if x == -0. Unfortunately we do not currently 3626 // have a way of expressing this constraint. See details in 3627 // https://llvm.org/bugs/show_bug.cgi?id=31702. 3628 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3629 Depth + 1); 3630 3631 case Intrinsic::fma: 3632 case Intrinsic::fmuladd: 3633 // x*x+y is non-negative if y is non-negative. 3634 return I->getOperand(0) == I->getOperand(1) && 3635 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) && 3636 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, 3637 Depth + 1); 3638 } 3639 break; 3640 } 3641 return false; 3642 } 3643 3644 bool llvm::CannotBeOrderedLessThanZero(const Value *V, 3645 const TargetLibraryInfo *TLI) { 3646 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0); 3647 } 3648 3649 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) { 3650 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0); 3651 } 3652 3653 bool llvm::isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI, 3654 unsigned Depth) { 3655 assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type"); 3656 3657 // If we're told that infinities won't happen, assume they won't. 3658 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) 3659 if (FPMathOp->hasNoInfs()) 3660 return true; 3661 3662 // Handle scalar constants. 3663 if (auto *CFP = dyn_cast<ConstantFP>(V)) 3664 return !CFP->isInfinity(); 3665 3666 if (Depth == MaxAnalysisRecursionDepth) 3667 return false; 3668 3669 if (auto *Inst = dyn_cast<Instruction>(V)) { 3670 switch (Inst->getOpcode()) { 3671 case Instruction::Select: { 3672 return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) && 3673 isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1); 3674 } 3675 case Instruction::SIToFP: 3676 case Instruction::UIToFP: { 3677 // Get width of largest magnitude integer (remove a bit if signed). 3678 // This still works for a signed minimum value because the largest FP 3679 // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). 3680 int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits(); 3681 if (Inst->getOpcode() == Instruction::SIToFP) 3682 --IntSize; 3683 3684 // If the exponent of the largest finite FP value can hold the largest 3685 // integer, the result of the cast must be finite. 3686 Type *FPTy = Inst->getType()->getScalarType(); 3687 return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize; 3688 } 3689 default: 3690 break; 3691 } 3692 } 3693 3694 // try to handle fixed width vector constants 3695 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); 3696 if (VFVTy && isa<Constant>(V)) { 3697 // For vectors, verify that each element is not infinity. 3698 unsigned NumElts = VFVTy->getNumElements(); 3699 for (unsigned i = 0; i != NumElts; ++i) { 3700 Constant *Elt = cast<Constant>(V)->getAggregateElement(i); 3701 if (!Elt) 3702 return false; 3703 if (isa<UndefValue>(Elt)) 3704 continue; 3705 auto *CElt = dyn_cast<ConstantFP>(Elt); 3706 if (!CElt || CElt->isInfinity()) 3707 return false; 3708 } 3709 // All elements were confirmed non-infinity or undefined. 3710 return true; 3711 } 3712 3713 // was not able to prove that V never contains infinity 3714 return false; 3715 } 3716 3717 bool llvm::isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI, 3718 unsigned Depth) { 3719 assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type"); 3720 3721 // If we're told that NaNs won't happen, assume they won't. 3722 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) 3723 if (FPMathOp->hasNoNaNs()) 3724 return true; 3725 3726 // Handle scalar constants. 3727 if (auto *CFP = dyn_cast<ConstantFP>(V)) 3728 return !CFP->isNaN(); 3729 3730 if (Depth == MaxAnalysisRecursionDepth) 3731 return false; 3732 3733 if (auto *Inst = dyn_cast<Instruction>(V)) { 3734 switch (Inst->getOpcode()) { 3735 case Instruction::FAdd: 3736 case Instruction::FSub: 3737 // Adding positive and negative infinity produces NaN. 3738 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && 3739 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && 3740 (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) || 3741 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1)); 3742 3743 case Instruction::FMul: 3744 // Zero multiplied with infinity produces NaN. 3745 // FIXME: If neither side can be zero fmul never produces NaN. 3746 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && 3747 isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) && 3748 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && 3749 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1); 3750 3751 case Instruction::FDiv: 3752 case Instruction::FRem: 3753 // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN. 3754 return false; 3755 3756 case Instruction::Select: { 3757 return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && 3758 isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1); 3759 } 3760 case Instruction::SIToFP: 3761 case Instruction::UIToFP: 3762 return true; 3763 case Instruction::FPTrunc: 3764 case Instruction::FPExt: 3765 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1); 3766 default: 3767 break; 3768 } 3769 } 3770 3771 if (const auto *II = dyn_cast<IntrinsicInst>(V)) { 3772 switch (II->getIntrinsicID()) { 3773 case Intrinsic::canonicalize: 3774 case Intrinsic::fabs: 3775 case Intrinsic::copysign: 3776 case Intrinsic::exp: 3777 case Intrinsic::exp2: 3778 case Intrinsic::floor: 3779 case Intrinsic::ceil: 3780 case Intrinsic::trunc: 3781 case Intrinsic::rint: 3782 case Intrinsic::nearbyint: 3783 case Intrinsic::round: 3784 case Intrinsic::roundeven: 3785 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1); 3786 case Intrinsic::sqrt: 3787 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) && 3788 CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI); 3789 case Intrinsic::minnum: 3790 case Intrinsic::maxnum: 3791 // If either operand is not NaN, the result is not NaN. 3792 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) || 3793 isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1); 3794 default: 3795 return false; 3796 } 3797 } 3798 3799 // Try to handle fixed width vector constants 3800 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); 3801 if (VFVTy && isa<Constant>(V)) { 3802 // For vectors, verify that each element is not NaN. 3803 unsigned NumElts = VFVTy->getNumElements(); 3804 for (unsigned i = 0; i != NumElts; ++i) { 3805 Constant *Elt = cast<Constant>(V)->getAggregateElement(i); 3806 if (!Elt) 3807 return false; 3808 if (isa<UndefValue>(Elt)) 3809 continue; 3810 auto *CElt = dyn_cast<ConstantFP>(Elt); 3811 if (!CElt || CElt->isNaN()) 3812 return false; 3813 } 3814 // All elements were confirmed not-NaN or undefined. 3815 return true; 3816 } 3817 3818 // Was not able to prove that V never contains NaN 3819 return false; 3820 } 3821 3822 Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) { 3823 3824 // All byte-wide stores are splatable, even of arbitrary variables. 3825 if (V->getType()->isIntegerTy(8)) 3826 return V; 3827 3828 LLVMContext &Ctx = V->getContext(); 3829 3830 // Undef don't care. 3831 auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx)); 3832 if (isa<UndefValue>(V)) 3833 return UndefInt8; 3834 3835 // Return Undef for zero-sized type. 3836 if (!DL.getTypeStoreSize(V->getType()).isNonZero()) 3837 return UndefInt8; 3838 3839 Constant *C = dyn_cast<Constant>(V); 3840 if (!C) { 3841 // Conceptually, we could handle things like: 3842 // %a = zext i8 %X to i16 3843 // %b = shl i16 %a, 8 3844 // %c = or i16 %a, %b 3845 // but until there is an example that actually needs this, it doesn't seem 3846 // worth worrying about. 3847 return nullptr; 3848 } 3849 3850 // Handle 'null' ConstantArrayZero etc. 3851 if (C->isNullValue()) 3852 return Constant::getNullValue(Type::getInt8Ty(Ctx)); 3853 3854 // Constant floating-point values can be handled as integer values if the 3855 // corresponding integer value is "byteable". An important case is 0.0. 3856 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 3857 Type *Ty = nullptr; 3858 if (CFP->getType()->isHalfTy()) 3859 Ty = Type::getInt16Ty(Ctx); 3860 else if (CFP->getType()->isFloatTy()) 3861 Ty = Type::getInt32Ty(Ctx); 3862 else if (CFP->getType()->isDoubleTy()) 3863 Ty = Type::getInt64Ty(Ctx); 3864 // Don't handle long double formats, which have strange constraints. 3865 return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL) 3866 : nullptr; 3867 } 3868 3869 // We can handle constant integers that are multiple of 8 bits. 3870 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 3871 if (CI->getBitWidth() % 8 == 0) { 3872 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!"); 3873 if (!CI->getValue().isSplat(8)) 3874 return nullptr; 3875 return ConstantInt::get(Ctx, CI->getValue().trunc(8)); 3876 } 3877 } 3878 3879 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 3880 if (CE->getOpcode() == Instruction::IntToPtr) { 3881 if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) { 3882 unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace()); 3883 return isBytewiseValue( 3884 ConstantExpr::getIntegerCast(CE->getOperand(0), 3885 Type::getIntNTy(Ctx, BitWidth), false), 3886 DL); 3887 } 3888 } 3889 } 3890 3891 auto Merge = [&](Value *LHS, Value *RHS) -> Value * { 3892 if (LHS == RHS) 3893 return LHS; 3894 if (!LHS || !RHS) 3895 return nullptr; 3896 if (LHS == UndefInt8) 3897 return RHS; 3898 if (RHS == UndefInt8) 3899 return LHS; 3900 return nullptr; 3901 }; 3902 3903 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) { 3904 Value *Val = UndefInt8; 3905 for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I) 3906 if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL)))) 3907 return nullptr; 3908 return Val; 3909 } 3910 3911 if (isa<ConstantAggregate>(C)) { 3912 Value *Val = UndefInt8; 3913 for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I) 3914 if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL)))) 3915 return nullptr; 3916 return Val; 3917 } 3918 3919 // Don't try to handle the handful of other constants. 3920 return nullptr; 3921 } 3922 3923 // This is the recursive version of BuildSubAggregate. It takes a few different 3924 // arguments. Idxs is the index within the nested struct From that we are 3925 // looking at now (which is of type IndexedType). IdxSkip is the number of 3926 // indices from Idxs that should be left out when inserting into the resulting 3927 // struct. To is the result struct built so far, new insertvalue instructions 3928 // build on that. 3929 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType, 3930 SmallVectorImpl<unsigned> &Idxs, 3931 unsigned IdxSkip, 3932 Instruction *InsertBefore) { 3933 StructType *STy = dyn_cast<StructType>(IndexedType); 3934 if (STy) { 3935 // Save the original To argument so we can modify it 3936 Value *OrigTo = To; 3937 // General case, the type indexed by Idxs is a struct 3938 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3939 // Process each struct element recursively 3940 Idxs.push_back(i); 3941 Value *PrevTo = To; 3942 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip, 3943 InsertBefore); 3944 Idxs.pop_back(); 3945 if (!To) { 3946 // Couldn't find any inserted value for this index? Cleanup 3947 while (PrevTo != OrigTo) { 3948 InsertValueInst* Del = cast<InsertValueInst>(PrevTo); 3949 PrevTo = Del->getAggregateOperand(); 3950 Del->eraseFromParent(); 3951 } 3952 // Stop processing elements 3953 break; 3954 } 3955 } 3956 // If we successfully found a value for each of our subaggregates 3957 if (To) 3958 return To; 3959 } 3960 // Base case, the type indexed by SourceIdxs is not a struct, or not all of 3961 // the struct's elements had a value that was inserted directly. In the latter 3962 // case, perhaps we can't determine each of the subelements individually, but 3963 // we might be able to find the complete struct somewhere. 3964 3965 // Find the value that is at that particular spot 3966 Value *V = FindInsertedValue(From, Idxs); 3967 3968 if (!V) 3969 return nullptr; 3970 3971 // Insert the value in the new (sub) aggregate 3972 return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip), 3973 "tmp", InsertBefore); 3974 } 3975 3976 // This helper takes a nested struct and extracts a part of it (which is again a 3977 // struct) into a new value. For example, given the struct: 3978 // { a, { b, { c, d }, e } } 3979 // and the indices "1, 1" this returns 3980 // { c, d }. 3981 // 3982 // It does this by inserting an insertvalue for each element in the resulting 3983 // struct, as opposed to just inserting a single struct. This will only work if 3984 // each of the elements of the substruct are known (ie, inserted into From by an 3985 // insertvalue instruction somewhere). 3986 // 3987 // All inserted insertvalue instructions are inserted before InsertBefore 3988 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, 3989 Instruction *InsertBefore) { 3990 assert(InsertBefore && "Must have someplace to insert!"); 3991 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(), 3992 idx_range); 3993 Value *To = UndefValue::get(IndexedType); 3994 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); 3995 unsigned IdxSkip = Idxs.size(); 3996 3997 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); 3998 } 3999 4000 /// Given an aggregate and a sequence of indices, see if the scalar value 4001 /// indexed is already around as a register, for example if it was inserted 4002 /// directly into the aggregate. 4003 /// 4004 /// If InsertBefore is not null, this function will duplicate (modified) 4005 /// insertvalues when a part of a nested struct is extracted. 4006 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, 4007 Instruction *InsertBefore) { 4008 // Nothing to index? Just return V then (this is useful at the end of our 4009 // recursion). 4010 if (idx_range.empty()) 4011 return V; 4012 // We have indices, so V should have an indexable type. 4013 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) && 4014 "Not looking at a struct or array?"); 4015 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) && 4016 "Invalid indices for type?"); 4017 4018 if (Constant *C = dyn_cast<Constant>(V)) { 4019 C = C->getAggregateElement(idx_range[0]); 4020 if (!C) return nullptr; 4021 return FindInsertedValue(C, idx_range.slice(1), InsertBefore); 4022 } 4023 4024 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) { 4025 // Loop the indices for the insertvalue instruction in parallel with the 4026 // requested indices 4027 const unsigned *req_idx = idx_range.begin(); 4028 for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); 4029 i != e; ++i, ++req_idx) { 4030 if (req_idx == idx_range.end()) { 4031 // We can't handle this without inserting insertvalues 4032 if (!InsertBefore) 4033 return nullptr; 4034 4035 // The requested index identifies a part of a nested aggregate. Handle 4036 // this specially. For example, 4037 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 4038 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 4039 // %C = extractvalue {i32, { i32, i32 } } %B, 1 4040 // This can be changed into 4041 // %A = insertvalue {i32, i32 } undef, i32 10, 0 4042 // %C = insertvalue {i32, i32 } %A, i32 11, 1 4043 // which allows the unused 0,0 element from the nested struct to be 4044 // removed. 4045 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx), 4046 InsertBefore); 4047 } 4048 4049 // This insert value inserts something else than what we are looking for. 4050 // See if the (aggregate) value inserted into has the value we are 4051 // looking for, then. 4052 if (*req_idx != *i) 4053 return FindInsertedValue(I->getAggregateOperand(), idx_range, 4054 InsertBefore); 4055 } 4056 // If we end up here, the indices of the insertvalue match with those 4057 // requested (though possibly only partially). Now we recursively look at 4058 // the inserted value, passing any remaining indices. 4059 return FindInsertedValue(I->getInsertedValueOperand(), 4060 makeArrayRef(req_idx, idx_range.end()), 4061 InsertBefore); 4062 } 4063 4064 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) { 4065 // If we're extracting a value from an aggregate that was extracted from 4066 // something else, we can extract from that something else directly instead. 4067 // However, we will need to chain I's indices with the requested indices. 4068 4069 // Calculate the number of indices required 4070 unsigned size = I->getNumIndices() + idx_range.size(); 4071 // Allocate some space to put the new indices in 4072 SmallVector<unsigned, 5> Idxs; 4073 Idxs.reserve(size); 4074 // Add indices from the extract value instruction 4075 Idxs.append(I->idx_begin(), I->idx_end()); 4076 4077 // Add requested indices 4078 Idxs.append(idx_range.begin(), idx_range.end()); 4079 4080 assert(Idxs.size() == size 4081 && "Number of indices added not correct?"); 4082 4083 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore); 4084 } 4085 // Otherwise, we don't know (such as, extracting from a function return value 4086 // or load instruction) 4087 return nullptr; 4088 } 4089 4090 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP, 4091 unsigned CharSize) { 4092 // Make sure the GEP has exactly three arguments. 4093 if (GEP->getNumOperands() != 3) 4094 return false; 4095 4096 // Make sure the index-ee is a pointer to array of \p CharSize integers. 4097 // CharSize. 4098 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType()); 4099 if (!AT || !AT->getElementType()->isIntegerTy(CharSize)) 4100 return false; 4101 4102 // Check to make sure that the first operand of the GEP is an integer and 4103 // has value 0 so that we are sure we're indexing into the initializer. 4104 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1)); 4105 if (!FirstIdx || !FirstIdx->isZero()) 4106 return false; 4107 4108 return true; 4109 } 4110 4111 bool llvm::getConstantDataArrayInfo(const Value *V, 4112 ConstantDataArraySlice &Slice, 4113 unsigned ElementSize, uint64_t Offset) { 4114 assert(V); 4115 4116 // Look through bitcast instructions and geps. 4117 V = V->stripPointerCasts(); 4118 4119 // If the value is a GEP instruction or constant expression, treat it as an 4120 // offset. 4121 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 4122 // The GEP operator should be based on a pointer to string constant, and is 4123 // indexing into the string constant. 4124 if (!isGEPBasedOnPointerToString(GEP, ElementSize)) 4125 return false; 4126 4127 // If the second index isn't a ConstantInt, then this is a variable index 4128 // into the array. If this occurs, we can't say anything meaningful about 4129 // the string. 4130 uint64_t StartIdx = 0; 4131 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2))) 4132 StartIdx = CI->getZExtValue(); 4133 else 4134 return false; 4135 return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize, 4136 StartIdx + Offset); 4137 } 4138 4139 // The GEP instruction, constant or instruction, must reference a global 4140 // variable that is a constant and is initialized. The referenced constant 4141 // initializer is the array that we'll use for optimization. 4142 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V); 4143 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) 4144 return false; 4145 4146 const ConstantDataArray *Array; 4147 ArrayType *ArrayTy; 4148 if (GV->getInitializer()->isNullValue()) { 4149 Type *GVTy = GV->getValueType(); 4150 if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) { 4151 // A zeroinitializer for the array; there is no ConstantDataArray. 4152 Array = nullptr; 4153 } else { 4154 const DataLayout &DL = GV->getParent()->getDataLayout(); 4155 uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedSize(); 4156 uint64_t Length = SizeInBytes / (ElementSize / 8); 4157 if (Length <= Offset) 4158 return false; 4159 4160 Slice.Array = nullptr; 4161 Slice.Offset = 0; 4162 Slice.Length = Length - Offset; 4163 return true; 4164 } 4165 } else { 4166 // This must be a ConstantDataArray. 4167 Array = dyn_cast<ConstantDataArray>(GV->getInitializer()); 4168 if (!Array) 4169 return false; 4170 ArrayTy = Array->getType(); 4171 } 4172 if (!ArrayTy->getElementType()->isIntegerTy(ElementSize)) 4173 return false; 4174 4175 uint64_t NumElts = ArrayTy->getArrayNumElements(); 4176 if (Offset > NumElts) 4177 return false; 4178 4179 Slice.Array = Array; 4180 Slice.Offset = Offset; 4181 Slice.Length = NumElts - Offset; 4182 return true; 4183 } 4184 4185 /// This function computes the length of a null-terminated C string pointed to 4186 /// by V. If successful, it returns true and returns the string in Str. 4187 /// If unsuccessful, it returns false. 4188 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, 4189 uint64_t Offset, bool TrimAtNul) { 4190 ConstantDataArraySlice Slice; 4191 if (!getConstantDataArrayInfo(V, Slice, 8, Offset)) 4192 return false; 4193 4194 if (Slice.Array == nullptr) { 4195 if (TrimAtNul) { 4196 Str = StringRef(); 4197 return true; 4198 } 4199 if (Slice.Length == 1) { 4200 Str = StringRef("", 1); 4201 return true; 4202 } 4203 // We cannot instantiate a StringRef as we do not have an appropriate string 4204 // of 0s at hand. 4205 return false; 4206 } 4207 4208 // Start out with the entire array in the StringRef. 4209 Str = Slice.Array->getAsString(); 4210 // Skip over 'offset' bytes. 4211 Str = Str.substr(Slice.Offset); 4212 4213 if (TrimAtNul) { 4214 // Trim off the \0 and anything after it. If the array is not nul 4215 // terminated, we just return the whole end of string. The client may know 4216 // some other way that the string is length-bound. 4217 Str = Str.substr(0, Str.find('\0')); 4218 } 4219 return true; 4220 } 4221 4222 // These next two are very similar to the above, but also look through PHI 4223 // nodes. 4224 // TODO: See if we can integrate these two together. 4225 4226 /// If we can compute the length of the string pointed to by 4227 /// the specified pointer, return 'len+1'. If we can't, return 0. 4228 static uint64_t GetStringLengthH(const Value *V, 4229 SmallPtrSetImpl<const PHINode*> &PHIs, 4230 unsigned CharSize) { 4231 // Look through noop bitcast instructions. 4232 V = V->stripPointerCasts(); 4233 4234 // If this is a PHI node, there are two cases: either we have already seen it 4235 // or we haven't. 4236 if (const PHINode *PN = dyn_cast<PHINode>(V)) { 4237 if (!PHIs.insert(PN).second) 4238 return ~0ULL; // already in the set. 4239 4240 // If it was new, see if all the input strings are the same length. 4241 uint64_t LenSoFar = ~0ULL; 4242 for (Value *IncValue : PN->incoming_values()) { 4243 uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize); 4244 if (Len == 0) return 0; // Unknown length -> unknown. 4245 4246 if (Len == ~0ULL) continue; 4247 4248 if (Len != LenSoFar && LenSoFar != ~0ULL) 4249 return 0; // Disagree -> unknown. 4250 LenSoFar = Len; 4251 } 4252 4253 // Success, all agree. 4254 return LenSoFar; 4255 } 4256 4257 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) 4258 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { 4259 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize); 4260 if (Len1 == 0) return 0; 4261 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize); 4262 if (Len2 == 0) return 0; 4263 if (Len1 == ~0ULL) return Len2; 4264 if (Len2 == ~0ULL) return Len1; 4265 if (Len1 != Len2) return 0; 4266 return Len1; 4267 } 4268 4269 // Otherwise, see if we can read the string. 4270 ConstantDataArraySlice Slice; 4271 if (!getConstantDataArrayInfo(V, Slice, CharSize)) 4272 return 0; 4273 4274 if (Slice.Array == nullptr) 4275 return 1; 4276 4277 // Search for nul characters 4278 unsigned NullIndex = 0; 4279 for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) { 4280 if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0) 4281 break; 4282 } 4283 4284 return NullIndex + 1; 4285 } 4286 4287 /// If we can compute the length of the string pointed to by 4288 /// the specified pointer, return 'len+1'. If we can't, return 0. 4289 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) { 4290 if (!V->getType()->isPointerTy()) 4291 return 0; 4292 4293 SmallPtrSet<const PHINode*, 32> PHIs; 4294 uint64_t Len = GetStringLengthH(V, PHIs, CharSize); 4295 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return 4296 // an empty string as a length. 4297 return Len == ~0ULL ? 1 : Len; 4298 } 4299 4300 const Value * 4301 llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call, 4302 bool MustPreserveNullness) { 4303 assert(Call && 4304 "getArgumentAliasingToReturnedPointer only works on nonnull calls"); 4305 if (const Value *RV = Call->getReturnedArgOperand()) 4306 return RV; 4307 // This can be used only as a aliasing property. 4308 if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( 4309 Call, MustPreserveNullness)) 4310 return Call->getArgOperand(0); 4311 return nullptr; 4312 } 4313 4314 bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( 4315 const CallBase *Call, bool MustPreserveNullness) { 4316 switch (Call->getIntrinsicID()) { 4317 case Intrinsic::launder_invariant_group: 4318 case Intrinsic::strip_invariant_group: 4319 case Intrinsic::aarch64_irg: 4320 case Intrinsic::aarch64_tagp: 4321 return true; 4322 case Intrinsic::ptrmask: 4323 return !MustPreserveNullness; 4324 default: 4325 return false; 4326 } 4327 } 4328 4329 /// \p PN defines a loop-variant pointer to an object. Check if the 4330 /// previous iteration of the loop was referring to the same object as \p PN. 4331 static bool isSameUnderlyingObjectInLoop(const PHINode *PN, 4332 const LoopInfo *LI) { 4333 // Find the loop-defined value. 4334 Loop *L = LI->getLoopFor(PN->getParent()); 4335 if (PN->getNumIncomingValues() != 2) 4336 return true; 4337 4338 // Find the value from previous iteration. 4339 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0)); 4340 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) 4341 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1)); 4342 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) 4343 return true; 4344 4345 // If a new pointer is loaded in the loop, the pointer references a different 4346 // object in every iteration. E.g.: 4347 // for (i) 4348 // int *p = a[i]; 4349 // ... 4350 if (auto *Load = dyn_cast<LoadInst>(PrevValue)) 4351 if (!L->isLoopInvariant(Load->getPointerOperand())) 4352 return false; 4353 return true; 4354 } 4355 4356 const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) { 4357 if (!V->getType()->isPointerTy()) 4358 return V; 4359 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { 4360 if (auto *GEP = dyn_cast<GEPOperator>(V)) { 4361 V = GEP->getPointerOperand(); 4362 } else if (Operator::getOpcode(V) == Instruction::BitCast || 4363 Operator::getOpcode(V) == Instruction::AddrSpaceCast) { 4364 V = cast<Operator>(V)->getOperand(0); 4365 if (!V->getType()->isPointerTy()) 4366 return V; 4367 } else if (auto *GA = dyn_cast<GlobalAlias>(V)) { 4368 if (GA->isInterposable()) 4369 return V; 4370 V = GA->getAliasee(); 4371 } else { 4372 if (auto *PHI = dyn_cast<PHINode>(V)) { 4373 // Look through single-arg phi nodes created by LCSSA. 4374 if (PHI->getNumIncomingValues() == 1) { 4375 V = PHI->getIncomingValue(0); 4376 continue; 4377 } 4378 } else if (auto *Call = dyn_cast<CallBase>(V)) { 4379 // CaptureTracking can know about special capturing properties of some 4380 // intrinsics like launder.invariant.group, that can't be expressed with 4381 // the attributes, but have properties like returning aliasing pointer. 4382 // Because some analysis may assume that nocaptured pointer is not 4383 // returned from some special intrinsic (because function would have to 4384 // be marked with returns attribute), it is crucial to use this function 4385 // because it should be in sync with CaptureTracking. Not using it may 4386 // cause weird miscompilations where 2 aliasing pointers are assumed to 4387 // noalias. 4388 if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) { 4389 V = RP; 4390 continue; 4391 } 4392 } 4393 4394 return V; 4395 } 4396 assert(V->getType()->isPointerTy() && "Unexpected operand type!"); 4397 } 4398 return V; 4399 } 4400 4401 void llvm::getUnderlyingObjects(const Value *V, 4402 SmallVectorImpl<const Value *> &Objects, 4403 LoopInfo *LI, unsigned MaxLookup) { 4404 SmallPtrSet<const Value *, 4> Visited; 4405 SmallVector<const Value *, 4> Worklist; 4406 Worklist.push_back(V); 4407 do { 4408 const Value *P = Worklist.pop_back_val(); 4409 P = getUnderlyingObject(P, MaxLookup); 4410 4411 if (!Visited.insert(P).second) 4412 continue; 4413 4414 if (auto *SI = dyn_cast<SelectInst>(P)) { 4415 Worklist.push_back(SI->getTrueValue()); 4416 Worklist.push_back(SI->getFalseValue()); 4417 continue; 4418 } 4419 4420 if (auto *PN = dyn_cast<PHINode>(P)) { 4421 // If this PHI changes the underlying object in every iteration of the 4422 // loop, don't look through it. Consider: 4423 // int **A; 4424 // for (i) { 4425 // Prev = Curr; // Prev = PHI (Prev_0, Curr) 4426 // Curr = A[i]; 4427 // *Prev, *Curr; 4428 // 4429 // Prev is tracking Curr one iteration behind so they refer to different 4430 // underlying objects. 4431 if (!LI || !LI->isLoopHeader(PN->getParent()) || 4432 isSameUnderlyingObjectInLoop(PN, LI)) 4433 append_range(Worklist, PN->incoming_values()); 4434 continue; 4435 } 4436 4437 Objects.push_back(P); 4438 } while (!Worklist.empty()); 4439 } 4440 4441 /// This is the function that does the work of looking through basic 4442 /// ptrtoint+arithmetic+inttoptr sequences. 4443 static const Value *getUnderlyingObjectFromInt(const Value *V) { 4444 do { 4445 if (const Operator *U = dyn_cast<Operator>(V)) { 4446 // If we find a ptrtoint, we can transfer control back to the 4447 // regular getUnderlyingObjectFromInt. 4448 if (U->getOpcode() == Instruction::PtrToInt) 4449 return U->getOperand(0); 4450 // If we find an add of a constant, a multiplied value, or a phi, it's 4451 // likely that the other operand will lead us to the base 4452 // object. We don't have to worry about the case where the 4453 // object address is somehow being computed by the multiply, 4454 // because our callers only care when the result is an 4455 // identifiable object. 4456 if (U->getOpcode() != Instruction::Add || 4457 (!isa<ConstantInt>(U->getOperand(1)) && 4458 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul && 4459 !isa<PHINode>(U->getOperand(1)))) 4460 return V; 4461 V = U->getOperand(0); 4462 } else { 4463 return V; 4464 } 4465 assert(V->getType()->isIntegerTy() && "Unexpected operand type!"); 4466 } while (true); 4467 } 4468 4469 /// This is a wrapper around getUnderlyingObjects and adds support for basic 4470 /// ptrtoint+arithmetic+inttoptr sequences. 4471 /// It returns false if unidentified object is found in getUnderlyingObjects. 4472 bool llvm::getUnderlyingObjectsForCodeGen(const Value *V, 4473 SmallVectorImpl<Value *> &Objects) { 4474 SmallPtrSet<const Value *, 16> Visited; 4475 SmallVector<const Value *, 4> Working(1, V); 4476 do { 4477 V = Working.pop_back_val(); 4478 4479 SmallVector<const Value *, 4> Objs; 4480 getUnderlyingObjects(V, Objs); 4481 4482 for (const Value *V : Objs) { 4483 if (!Visited.insert(V).second) 4484 continue; 4485 if (Operator::getOpcode(V) == Instruction::IntToPtr) { 4486 const Value *O = 4487 getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); 4488 if (O->getType()->isPointerTy()) { 4489 Working.push_back(O); 4490 continue; 4491 } 4492 } 4493 // If getUnderlyingObjects fails to find an identifiable object, 4494 // getUnderlyingObjectsForCodeGen also fails for safety. 4495 if (!isIdentifiedObject(V)) { 4496 Objects.clear(); 4497 return false; 4498 } 4499 Objects.push_back(const_cast<Value *>(V)); 4500 } 4501 } while (!Working.empty()); 4502 return true; 4503 } 4504 4505 AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) { 4506 AllocaInst *Result = nullptr; 4507 SmallPtrSet<Value *, 4> Visited; 4508 SmallVector<Value *, 4> Worklist; 4509 4510 auto AddWork = [&](Value *V) { 4511 if (Visited.insert(V).second) 4512 Worklist.push_back(V); 4513 }; 4514 4515 AddWork(V); 4516 do { 4517 V = Worklist.pop_back_val(); 4518 assert(Visited.count(V)); 4519 4520 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 4521 if (Result && Result != AI) 4522 return nullptr; 4523 Result = AI; 4524 } else if (CastInst *CI = dyn_cast<CastInst>(V)) { 4525 AddWork(CI->getOperand(0)); 4526 } else if (PHINode *PN = dyn_cast<PHINode>(V)) { 4527 for (Value *IncValue : PN->incoming_values()) 4528 AddWork(IncValue); 4529 } else if (auto *SI = dyn_cast<SelectInst>(V)) { 4530 AddWork(SI->getTrueValue()); 4531 AddWork(SI->getFalseValue()); 4532 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 4533 if (OffsetZero && !GEP->hasAllZeroIndices()) 4534 return nullptr; 4535 AddWork(GEP->getPointerOperand()); 4536 } else { 4537 return nullptr; 4538 } 4539 } while (!Worklist.empty()); 4540 4541 return Result; 4542 } 4543 4544 static bool onlyUsedByLifetimeMarkersOrDroppableInstsHelper( 4545 const Value *V, bool AllowLifetime, bool AllowDroppable) { 4546 for (const User *U : V->users()) { 4547 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); 4548 if (!II) 4549 return false; 4550 4551 if (AllowLifetime && II->isLifetimeStartOrEnd()) 4552 continue; 4553 4554 if (AllowDroppable && II->isDroppable()) 4555 continue; 4556 4557 return false; 4558 } 4559 return true; 4560 } 4561 4562 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { 4563 return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( 4564 V, /* AllowLifetime */ true, /* AllowDroppable */ false); 4565 } 4566 bool llvm::onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V) { 4567 return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( 4568 V, /* AllowLifetime */ true, /* AllowDroppable */ true); 4569 } 4570 4571 bool llvm::mustSuppressSpeculation(const LoadInst &LI) { 4572 if (!LI.isUnordered()) 4573 return true; 4574 const Function &F = *LI.getFunction(); 4575 // Speculative load may create a race that did not exist in the source. 4576 return F.hasFnAttribute(Attribute::SanitizeThread) || 4577 // Speculative load may load data from dirty regions. 4578 F.hasFnAttribute(Attribute::SanitizeAddress) || 4579 F.hasFnAttribute(Attribute::SanitizeHWAddress); 4580 } 4581 4582 4583 bool llvm::isSafeToSpeculativelyExecute(const Value *V, 4584 const Instruction *CtxI, 4585 const DominatorTree *DT, 4586 const TargetLibraryInfo *TLI) { 4587 const Operator *Inst = dyn_cast<Operator>(V); 4588 if (!Inst) 4589 return false; 4590 4591 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) 4592 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i))) 4593 if (C->canTrap()) 4594 return false; 4595 4596 switch (Inst->getOpcode()) { 4597 default: 4598 return true; 4599 case Instruction::UDiv: 4600 case Instruction::URem: { 4601 // x / y is undefined if y == 0. 4602 const APInt *V; 4603 if (match(Inst->getOperand(1), m_APInt(V))) 4604 return *V != 0; 4605 return false; 4606 } 4607 case Instruction::SDiv: 4608 case Instruction::SRem: { 4609 // x / y is undefined if y == 0 or x == INT_MIN and y == -1 4610 const APInt *Numerator, *Denominator; 4611 if (!match(Inst->getOperand(1), m_APInt(Denominator))) 4612 return false; 4613 // We cannot hoist this division if the denominator is 0. 4614 if (*Denominator == 0) 4615 return false; 4616 // It's safe to hoist if the denominator is not 0 or -1. 4617 if (!Denominator->isAllOnesValue()) 4618 return true; 4619 // At this point we know that the denominator is -1. It is safe to hoist as 4620 // long we know that the numerator is not INT_MIN. 4621 if (match(Inst->getOperand(0), m_APInt(Numerator))) 4622 return !Numerator->isMinSignedValue(); 4623 // The numerator *might* be MinSignedValue. 4624 return false; 4625 } 4626 case Instruction::Load: { 4627 const LoadInst *LI = cast<LoadInst>(Inst); 4628 if (mustSuppressSpeculation(*LI)) 4629 return false; 4630 const DataLayout &DL = LI->getModule()->getDataLayout(); 4631 return isDereferenceableAndAlignedPointer( 4632 LI->getPointerOperand(), LI->getType(), MaybeAlign(LI->getAlignment()), 4633 DL, CtxI, DT, TLI); 4634 } 4635 case Instruction::Call: { 4636 auto *CI = cast<const CallInst>(Inst); 4637 const Function *Callee = CI->getCalledFunction(); 4638 4639 // The called function could have undefined behavior or side-effects, even 4640 // if marked readnone nounwind. 4641 return Callee && Callee->isSpeculatable(); 4642 } 4643 case Instruction::VAArg: 4644 case Instruction::Alloca: 4645 case Instruction::Invoke: 4646 case Instruction::CallBr: 4647 case Instruction::PHI: 4648 case Instruction::Store: 4649 case Instruction::Ret: 4650 case Instruction::Br: 4651 case Instruction::IndirectBr: 4652 case Instruction::Switch: 4653 case Instruction::Unreachable: 4654 case Instruction::Fence: 4655 case Instruction::AtomicRMW: 4656 case Instruction::AtomicCmpXchg: 4657 case Instruction::LandingPad: 4658 case Instruction::Resume: 4659 case Instruction::CatchSwitch: 4660 case Instruction::CatchPad: 4661 case Instruction::CatchRet: 4662 case Instruction::CleanupPad: 4663 case Instruction::CleanupRet: 4664 return false; // Misc instructions which have effects 4665 } 4666 } 4667 4668 bool llvm::mayBeMemoryDependent(const Instruction &I) { 4669 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I); 4670 } 4671 4672 /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult. 4673 static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) { 4674 switch (OR) { 4675 case ConstantRange::OverflowResult::MayOverflow: 4676 return OverflowResult::MayOverflow; 4677 case ConstantRange::OverflowResult::AlwaysOverflowsLow: 4678 return OverflowResult::AlwaysOverflowsLow; 4679 case ConstantRange::OverflowResult::AlwaysOverflowsHigh: 4680 return OverflowResult::AlwaysOverflowsHigh; 4681 case ConstantRange::OverflowResult::NeverOverflows: 4682 return OverflowResult::NeverOverflows; 4683 } 4684 llvm_unreachable("Unknown OverflowResult"); 4685 } 4686 4687 /// Combine constant ranges from computeConstantRange() and computeKnownBits(). 4688 static ConstantRange computeConstantRangeIncludingKnownBits( 4689 const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth, 4690 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, 4691 OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) { 4692 KnownBits Known = computeKnownBits( 4693 V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo); 4694 ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned); 4695 ConstantRange CR2 = computeConstantRange(V, UseInstrInfo); 4696 ConstantRange::PreferredRangeType RangeType = 4697 ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned; 4698 return CR1.intersectWith(CR2, RangeType); 4699 } 4700 4701 OverflowResult llvm::computeOverflowForUnsignedMul( 4702 const Value *LHS, const Value *RHS, const DataLayout &DL, 4703 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, 4704 bool UseInstrInfo) { 4705 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, 4706 nullptr, UseInstrInfo); 4707 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, 4708 nullptr, UseInstrInfo); 4709 ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false); 4710 ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false); 4711 return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange)); 4712 } 4713 4714 OverflowResult 4715 llvm::computeOverflowForSignedMul(const Value *LHS, const Value *RHS, 4716 const DataLayout &DL, AssumptionCache *AC, 4717 const Instruction *CxtI, 4718 const DominatorTree *DT, bool UseInstrInfo) { 4719 // Multiplying n * m significant bits yields a result of n + m significant 4720 // bits. If the total number of significant bits does not exceed the 4721 // result bit width (minus 1), there is no overflow. 4722 // This means if we have enough leading sign bits in the operands 4723 // we can guarantee that the result does not overflow. 4724 // Ref: "Hacker's Delight" by Henry Warren 4725 unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); 4726 4727 // Note that underestimating the number of sign bits gives a more 4728 // conservative answer. 4729 unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) + 4730 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT); 4731 4732 // First handle the easy case: if we have enough sign bits there's 4733 // definitely no overflow. 4734 if (SignBits > BitWidth + 1) 4735 return OverflowResult::NeverOverflows; 4736 4737 // There are two ambiguous cases where there can be no overflow: 4738 // SignBits == BitWidth + 1 and 4739 // SignBits == BitWidth 4740 // The second case is difficult to check, therefore we only handle the 4741 // first case. 4742 if (SignBits == BitWidth + 1) { 4743 // It overflows only when both arguments are negative and the true 4744 // product is exactly the minimum negative number. 4745 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000 4746 // For simplicity we just check if at least one side is not negative. 4747 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, 4748 nullptr, UseInstrInfo); 4749 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, 4750 nullptr, UseInstrInfo); 4751 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) 4752 return OverflowResult::NeverOverflows; 4753 } 4754 return OverflowResult::MayOverflow; 4755 } 4756 4757 OverflowResult llvm::computeOverflowForUnsignedAdd( 4758 const Value *LHS, const Value *RHS, const DataLayout &DL, 4759 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, 4760 bool UseInstrInfo) { 4761 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4762 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, 4763 nullptr, UseInstrInfo); 4764 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4765 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, 4766 nullptr, UseInstrInfo); 4767 return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange)); 4768 } 4769 4770 static OverflowResult computeOverflowForSignedAdd(const Value *LHS, 4771 const Value *RHS, 4772 const AddOperator *Add, 4773 const DataLayout &DL, 4774 AssumptionCache *AC, 4775 const Instruction *CxtI, 4776 const DominatorTree *DT) { 4777 if (Add && Add->hasNoSignedWrap()) { 4778 return OverflowResult::NeverOverflows; 4779 } 4780 4781 // If LHS and RHS each have at least two sign bits, the addition will look 4782 // like 4783 // 4784 // XX..... + 4785 // YY..... 4786 // 4787 // If the carry into the most significant position is 0, X and Y can't both 4788 // be 1 and therefore the carry out of the addition is also 0. 4789 // 4790 // If the carry into the most significant position is 1, X and Y can't both 4791 // be 0 and therefore the carry out of the addition is also 1. 4792 // 4793 // Since the carry into the most significant position is always equal to 4794 // the carry out of the addition, there is no signed overflow. 4795 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && 4796 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) 4797 return OverflowResult::NeverOverflows; 4798 4799 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4800 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4801 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4802 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4803 OverflowResult OR = 4804 mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange)); 4805 if (OR != OverflowResult::MayOverflow) 4806 return OR; 4807 4808 // The remaining code needs Add to be available. Early returns if not so. 4809 if (!Add) 4810 return OverflowResult::MayOverflow; 4811 4812 // If the sign of Add is the same as at least one of the operands, this add 4813 // CANNOT overflow. If this can be determined from the known bits of the 4814 // operands the above signedAddMayOverflow() check will have already done so. 4815 // The only other way to improve on the known bits is from an assumption, so 4816 // call computeKnownBitsFromAssume() directly. 4817 bool LHSOrRHSKnownNonNegative = 4818 (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative()); 4819 bool LHSOrRHSKnownNegative = 4820 (LHSRange.isAllNegative() || RHSRange.isAllNegative()); 4821 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { 4822 KnownBits AddKnown(LHSRange.getBitWidth()); 4823 computeKnownBitsFromAssume( 4824 Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true)); 4825 if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) || 4826 (AddKnown.isNegative() && LHSOrRHSKnownNegative)) 4827 return OverflowResult::NeverOverflows; 4828 } 4829 4830 return OverflowResult::MayOverflow; 4831 } 4832 4833 OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS, 4834 const Value *RHS, 4835 const DataLayout &DL, 4836 AssumptionCache *AC, 4837 const Instruction *CxtI, 4838 const DominatorTree *DT) { 4839 // Checking for conditions implied by dominating conditions may be expensive. 4840 // Limit it to usub_with_overflow calls for now. 4841 if (match(CxtI, 4842 m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value()))) 4843 if (auto C = 4844 isImpliedByDomCondition(CmpInst::ICMP_UGE, LHS, RHS, CxtI, DL)) { 4845 if (*C) 4846 return OverflowResult::NeverOverflows; 4847 return OverflowResult::AlwaysOverflowsLow; 4848 } 4849 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4850 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); 4851 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4852 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); 4853 return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange)); 4854 } 4855 4856 OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS, 4857 const Value *RHS, 4858 const DataLayout &DL, 4859 AssumptionCache *AC, 4860 const Instruction *CxtI, 4861 const DominatorTree *DT) { 4862 // If LHS and RHS each have at least two sign bits, the subtraction 4863 // cannot overflow. 4864 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && 4865 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) 4866 return OverflowResult::NeverOverflows; 4867 4868 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4869 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4870 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4871 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4872 return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange)); 4873 } 4874 4875 bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO, 4876 const DominatorTree &DT) { 4877 SmallVector<const BranchInst *, 2> GuardingBranches; 4878 SmallVector<const ExtractValueInst *, 2> Results; 4879 4880 for (const User *U : WO->users()) { 4881 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) { 4882 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type"); 4883 4884 if (EVI->getIndices()[0] == 0) 4885 Results.push_back(EVI); 4886 else { 4887 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type"); 4888 4889 for (const auto *U : EVI->users()) 4890 if (const auto *B = dyn_cast<BranchInst>(U)) { 4891 assert(B->isConditional() && "How else is it using an i1?"); 4892 GuardingBranches.push_back(B); 4893 } 4894 } 4895 } else { 4896 // We are using the aggregate directly in a way we don't want to analyze 4897 // here (storing it to a global, say). 4898 return false; 4899 } 4900 } 4901 4902 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) { 4903 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1)); 4904 if (!NoWrapEdge.isSingleEdge()) 4905 return false; 4906 4907 // Check if all users of the add are provably no-wrap. 4908 for (const auto *Result : Results) { 4909 // If the extractvalue itself is not executed on overflow, the we don't 4910 // need to check each use separately, since domination is transitive. 4911 if (DT.dominates(NoWrapEdge, Result->getParent())) 4912 continue; 4913 4914 for (auto &RU : Result->uses()) 4915 if (!DT.dominates(NoWrapEdge, RU)) 4916 return false; 4917 } 4918 4919 return true; 4920 }; 4921 4922 return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch); 4923 } 4924 4925 static bool canCreateUndefOrPoison(const Operator *Op, bool PoisonOnly) { 4926 // See whether I has flags that may create poison 4927 if (const auto *OvOp = dyn_cast<OverflowingBinaryOperator>(Op)) { 4928 if (OvOp->hasNoSignedWrap() || OvOp->hasNoUnsignedWrap()) 4929 return true; 4930 } 4931 if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(Op)) 4932 if (ExactOp->isExact()) 4933 return true; 4934 if (const auto *FP = dyn_cast<FPMathOperator>(Op)) { 4935 auto FMF = FP->getFastMathFlags(); 4936 if (FMF.noNaNs() || FMF.noInfs()) 4937 return true; 4938 } 4939 4940 unsigned Opcode = Op->getOpcode(); 4941 4942 // Check whether opcode is a poison/undef-generating operation 4943 switch (Opcode) { 4944 case Instruction::Shl: 4945 case Instruction::AShr: 4946 case Instruction::LShr: { 4947 // Shifts return poison if shiftwidth is larger than the bitwidth. 4948 if (auto *C = dyn_cast<Constant>(Op->getOperand(1))) { 4949 SmallVector<Constant *, 4> ShiftAmounts; 4950 if (auto *FVTy = dyn_cast<FixedVectorType>(C->getType())) { 4951 unsigned NumElts = FVTy->getNumElements(); 4952 for (unsigned i = 0; i < NumElts; ++i) 4953 ShiftAmounts.push_back(C->getAggregateElement(i)); 4954 } else if (isa<ScalableVectorType>(C->getType())) 4955 return true; // Can't tell, just return true to be safe 4956 else 4957 ShiftAmounts.push_back(C); 4958 4959 bool Safe = llvm::all_of(ShiftAmounts, [](Constant *C) { 4960 auto *CI = dyn_cast_or_null<ConstantInt>(C); 4961 return CI && CI->getValue().ult(C->getType()->getIntegerBitWidth()); 4962 }); 4963 return !Safe; 4964 } 4965 return true; 4966 } 4967 case Instruction::FPToSI: 4968 case Instruction::FPToUI: 4969 // fptosi/ui yields poison if the resulting value does not fit in the 4970 // destination type. 4971 return true; 4972 case Instruction::Call: 4973 if (auto *II = dyn_cast<IntrinsicInst>(Op)) { 4974 switch (II->getIntrinsicID()) { 4975 // TODO: Add more intrinsics. 4976 case Intrinsic::ctpop: 4977 case Intrinsic::sadd_with_overflow: 4978 case Intrinsic::ssub_with_overflow: 4979 case Intrinsic::smul_with_overflow: 4980 case Intrinsic::uadd_with_overflow: 4981 case Intrinsic::usub_with_overflow: 4982 case Intrinsic::umul_with_overflow: 4983 return false; 4984 } 4985 } 4986 LLVM_FALLTHROUGH; 4987 case Instruction::CallBr: 4988 case Instruction::Invoke: { 4989 const auto *CB = cast<CallBase>(Op); 4990 return !CB->hasRetAttr(Attribute::NoUndef); 4991 } 4992 case Instruction::InsertElement: 4993 case Instruction::ExtractElement: { 4994 // If index exceeds the length of the vector, it returns poison 4995 auto *VTy = cast<VectorType>(Op->getOperand(0)->getType()); 4996 unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1; 4997 auto *Idx = dyn_cast<ConstantInt>(Op->getOperand(IdxOp)); 4998 if (!Idx || Idx->getValue().uge(VTy->getElementCount().getKnownMinValue())) 4999 return true; 5000 return false; 5001 } 5002 case Instruction::ShuffleVector: { 5003 // shufflevector may return undef. 5004 if (PoisonOnly) 5005 return false; 5006 ArrayRef<int> Mask = isa<ConstantExpr>(Op) 5007 ? cast<ConstantExpr>(Op)->getShuffleMask() 5008 : cast<ShuffleVectorInst>(Op)->getShuffleMask(); 5009 return is_contained(Mask, UndefMaskElem); 5010 } 5011 case Instruction::FNeg: 5012 case Instruction::PHI: 5013 case Instruction::Select: 5014 case Instruction::URem: 5015 case Instruction::SRem: 5016 case Instruction::ExtractValue: 5017 case Instruction::InsertValue: 5018 case Instruction::Freeze: 5019 case Instruction::ICmp: 5020 case Instruction::FCmp: 5021 return false; 5022 case Instruction::GetElementPtr: { 5023 const auto *GEP = cast<GEPOperator>(Op); 5024 return GEP->isInBounds(); 5025 } 5026 default: { 5027 const auto *CE = dyn_cast<ConstantExpr>(Op); 5028 if (isa<CastInst>(Op) || (CE && CE->isCast())) 5029 return false; 5030 else if (Instruction::isBinaryOp(Opcode)) 5031 return false; 5032 // Be conservative and return true. 5033 return true; 5034 } 5035 } 5036 } 5037 5038 bool llvm::canCreateUndefOrPoison(const Operator *Op) { 5039 return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/false); 5040 } 5041 5042 bool llvm::canCreatePoison(const Operator *Op) { 5043 return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/true); 5044 } 5045 5046 static bool directlyImpliesPoison(const Value *ValAssumedPoison, 5047 const Value *V, unsigned Depth) { 5048 if (ValAssumedPoison == V) 5049 return true; 5050 5051 const unsigned MaxDepth = 2; 5052 if (Depth >= MaxDepth) 5053 return false; 5054 5055 if (const auto *I = dyn_cast<Instruction>(V)) { 5056 if (propagatesPoison(cast<Operator>(I))) 5057 return any_of(I->operands(), [=](const Value *Op) { 5058 return directlyImpliesPoison(ValAssumedPoison, Op, Depth + 1); 5059 }); 5060 5061 // 'select ValAssumedPoison, _, _' is poison. 5062 if (const auto *SI = dyn_cast<SelectInst>(I)) 5063 return directlyImpliesPoison(ValAssumedPoison, SI->getCondition(), 5064 Depth + 1); 5065 // V = extractvalue V0, idx 5066 // V2 = extractvalue V0, idx2 5067 // V0's elements are all poison or not. (e.g., add_with_overflow) 5068 const WithOverflowInst *II; 5069 if (match(I, m_ExtractValue(m_WithOverflowInst(II))) && 5070 (match(ValAssumedPoison, m_ExtractValue(m_Specific(II))) || 5071 llvm::is_contained(II->arg_operands(), ValAssumedPoison))) 5072 return true; 5073 } 5074 return false; 5075 } 5076 5077 static bool impliesPoison(const Value *ValAssumedPoison, const Value *V, 5078 unsigned Depth) { 5079 if (isGuaranteedNotToBeUndefOrPoison(ValAssumedPoison)) 5080 return true; 5081 5082 if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0)) 5083 return true; 5084 5085 const unsigned MaxDepth = 2; 5086 if (Depth >= MaxDepth) 5087 return false; 5088 5089 const auto *I = dyn_cast<Instruction>(ValAssumedPoison); 5090 if (I && !canCreatePoison(cast<Operator>(I))) { 5091 return all_of(I->operands(), [=](const Value *Op) { 5092 return impliesPoison(Op, V, Depth + 1); 5093 }); 5094 } 5095 return false; 5096 } 5097 5098 bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) { 5099 return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0); 5100 } 5101 5102 static bool programUndefinedIfUndefOrPoison(const Value *V, 5103 bool PoisonOnly); 5104 5105 static bool isGuaranteedNotToBeUndefOrPoison(const Value *V, 5106 AssumptionCache *AC, 5107 const Instruction *CtxI, 5108 const DominatorTree *DT, 5109 unsigned Depth, bool PoisonOnly) { 5110 if (Depth >= MaxAnalysisRecursionDepth) 5111 return false; 5112 5113 if (isa<MetadataAsValue>(V)) 5114 return false; 5115 5116 if (const auto *A = dyn_cast<Argument>(V)) { 5117 if (A->hasAttribute(Attribute::NoUndef)) 5118 return true; 5119 } 5120 5121 if (auto *C = dyn_cast<Constant>(V)) { 5122 if (isa<UndefValue>(C)) 5123 return PoisonOnly && !isa<PoisonValue>(C); 5124 5125 if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(V) || 5126 isa<ConstantPointerNull>(C) || isa<Function>(C)) 5127 return true; 5128 5129 if (C->getType()->isVectorTy() && !isa<ConstantExpr>(C)) 5130 return (PoisonOnly ? !C->containsPoisonElement() 5131 : !C->containsUndefOrPoisonElement()) && 5132 !C->containsConstantExpression(); 5133 } 5134 5135 // Strip cast operations from a pointer value. 5136 // Note that stripPointerCastsSameRepresentation can strip off getelementptr 5137 // inbounds with zero offset. To guarantee that the result isn't poison, the 5138 // stripped pointer is checked as it has to be pointing into an allocated 5139 // object or be null `null` to ensure `inbounds` getelement pointers with a 5140 // zero offset could not produce poison. 5141 // It can strip off addrspacecast that do not change bit representation as 5142 // well. We believe that such addrspacecast is equivalent to no-op. 5143 auto *StrippedV = V->stripPointerCastsSameRepresentation(); 5144 if (isa<AllocaInst>(StrippedV) || isa<GlobalVariable>(StrippedV) || 5145 isa<Function>(StrippedV) || isa<ConstantPointerNull>(StrippedV)) 5146 return true; 5147 5148 auto OpCheck = [&](const Value *V) { 5149 return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth + 1, 5150 PoisonOnly); 5151 }; 5152 5153 if (auto *Opr = dyn_cast<Operator>(V)) { 5154 // If the value is a freeze instruction, then it can never 5155 // be undef or poison. 5156 if (isa<FreezeInst>(V)) 5157 return true; 5158 5159 if (const auto *CB = dyn_cast<CallBase>(V)) { 5160 if (CB->hasRetAttr(Attribute::NoUndef)) 5161 return true; 5162 } 5163 5164 if (const auto *PN = dyn_cast<PHINode>(V)) { 5165 unsigned Num = PN->getNumIncomingValues(); 5166 bool IsWellDefined = true; 5167 for (unsigned i = 0; i < Num; ++i) { 5168 auto *TI = PN->getIncomingBlock(i)->getTerminator(); 5169 if (!isGuaranteedNotToBeUndefOrPoison(PN->getIncomingValue(i), AC, TI, 5170 DT, Depth + 1, PoisonOnly)) { 5171 IsWellDefined = false; 5172 break; 5173 } 5174 } 5175 if (IsWellDefined) 5176 return true; 5177 } else if (!canCreateUndefOrPoison(Opr) && all_of(Opr->operands(), OpCheck)) 5178 return true; 5179 } 5180 5181 if (auto *I = dyn_cast<LoadInst>(V)) 5182 if (I->getMetadata(LLVMContext::MD_noundef)) 5183 return true; 5184 5185 if (programUndefinedIfUndefOrPoison(V, PoisonOnly)) 5186 return true; 5187 5188 // CxtI may be null or a cloned instruction. 5189 if (!CtxI || !CtxI->getParent() || !DT) 5190 return false; 5191 5192 auto *DNode = DT->getNode(CtxI->getParent()); 5193 if (!DNode) 5194 // Unreachable block 5195 return false; 5196 5197 // If V is used as a branch condition before reaching CtxI, V cannot be 5198 // undef or poison. 5199 // br V, BB1, BB2 5200 // BB1: 5201 // CtxI ; V cannot be undef or poison here 5202 auto *Dominator = DNode->getIDom(); 5203 while (Dominator) { 5204 auto *TI = Dominator->getBlock()->getTerminator(); 5205 5206 Value *Cond = nullptr; 5207 if (auto BI = dyn_cast<BranchInst>(TI)) { 5208 if (BI->isConditional()) 5209 Cond = BI->getCondition(); 5210 } else if (auto SI = dyn_cast<SwitchInst>(TI)) { 5211 Cond = SI->getCondition(); 5212 } 5213 5214 if (Cond) { 5215 if (Cond == V) 5216 return true; 5217 else if (PoisonOnly && isa<Operator>(Cond)) { 5218 // For poison, we can analyze further 5219 auto *Opr = cast<Operator>(Cond); 5220 if (propagatesPoison(Opr) && is_contained(Opr->operand_values(), V)) 5221 return true; 5222 } 5223 } 5224 5225 Dominator = Dominator->getIDom(); 5226 } 5227 5228 SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NoUndef}; 5229 if (getKnowledgeValidInContext(V, AttrKinds, CtxI, DT, AC)) 5230 return true; 5231 5232 return false; 5233 } 5234 5235 bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC, 5236 const Instruction *CtxI, 5237 const DominatorTree *DT, 5238 unsigned Depth) { 5239 return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, false); 5240 } 5241 5242 bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC, 5243 const Instruction *CtxI, 5244 const DominatorTree *DT, unsigned Depth) { 5245 return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, true); 5246 } 5247 5248 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add, 5249 const DataLayout &DL, 5250 AssumptionCache *AC, 5251 const Instruction *CxtI, 5252 const DominatorTree *DT) { 5253 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1), 5254 Add, DL, AC, CxtI, DT); 5255 } 5256 5257 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS, 5258 const Value *RHS, 5259 const DataLayout &DL, 5260 AssumptionCache *AC, 5261 const Instruction *CxtI, 5262 const DominatorTree *DT) { 5263 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT); 5264 } 5265 5266 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { 5267 // Note: An atomic operation isn't guaranteed to return in a reasonable amount 5268 // of time because it's possible for another thread to interfere with it for an 5269 // arbitrary length of time, but programs aren't allowed to rely on that. 5270 5271 // If there is no successor, then execution can't transfer to it. 5272 if (isa<ReturnInst>(I)) 5273 return false; 5274 if (isa<UnreachableInst>(I)) 5275 return false; 5276 5277 // Note: Do not add new checks here; instead, change Instruction::mayThrow or 5278 // Instruction::willReturn. 5279 // 5280 // FIXME: Move this check into Instruction::willReturn. 5281 if (isa<CatchPadInst>(I)) { 5282 switch (classifyEHPersonality(I->getFunction()->getPersonalityFn())) { 5283 default: 5284 // A catchpad may invoke exception object constructors and such, which 5285 // in some languages can be arbitrary code, so be conservative by default. 5286 return false; 5287 case EHPersonality::CoreCLR: 5288 // For CoreCLR, it just involves a type test. 5289 return true; 5290 } 5291 } 5292 5293 // An instruction that returns without throwing must transfer control flow 5294 // to a successor. 5295 return !I->mayThrow() && I->willReturn(); 5296 } 5297 5298 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) { 5299 // TODO: This is slightly conservative for invoke instruction since exiting 5300 // via an exception *is* normal control for them. 5301 for (const Instruction &I : *BB) 5302 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) 5303 return false; 5304 return true; 5305 } 5306 5307 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, 5308 const Loop *L) { 5309 // The loop header is guaranteed to be executed for every iteration. 5310 // 5311 // FIXME: Relax this constraint to cover all basic blocks that are 5312 // guaranteed to be executed at every iteration. 5313 if (I->getParent() != L->getHeader()) return false; 5314 5315 for (const Instruction &LI : *L->getHeader()) { 5316 if (&LI == I) return true; 5317 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false; 5318 } 5319 llvm_unreachable("Instruction not contained in its own parent basic block."); 5320 } 5321 5322 bool llvm::propagatesPoison(const Operator *I) { 5323 switch (I->getOpcode()) { 5324 case Instruction::Freeze: 5325 case Instruction::Select: 5326 case Instruction::PHI: 5327 case Instruction::Invoke: 5328 return false; 5329 case Instruction::Call: 5330 if (auto *II = dyn_cast<IntrinsicInst>(I)) { 5331 switch (II->getIntrinsicID()) { 5332 // TODO: Add more intrinsics. 5333 case Intrinsic::sadd_with_overflow: 5334 case Intrinsic::ssub_with_overflow: 5335 case Intrinsic::smul_with_overflow: 5336 case Intrinsic::uadd_with_overflow: 5337 case Intrinsic::usub_with_overflow: 5338 case Intrinsic::umul_with_overflow: 5339 // If an input is a vector containing a poison element, the 5340 // two output vectors (calculated results, overflow bits)' 5341 // corresponding lanes are poison. 5342 return true; 5343 case Intrinsic::ctpop: 5344 return true; 5345 } 5346 } 5347 return false; 5348 case Instruction::ICmp: 5349 case Instruction::FCmp: 5350 case Instruction::GetElementPtr: 5351 return true; 5352 default: 5353 if (isa<BinaryOperator>(I) || isa<UnaryOperator>(I) || isa<CastInst>(I)) 5354 return true; 5355 5356 // Be conservative and return false. 5357 return false; 5358 } 5359 } 5360 5361 void llvm::getGuaranteedWellDefinedOps( 5362 const Instruction *I, SmallPtrSetImpl<const Value *> &Operands) { 5363 switch (I->getOpcode()) { 5364 case Instruction::Store: 5365 Operands.insert(cast<StoreInst>(I)->getPointerOperand()); 5366 break; 5367 5368 case Instruction::Load: 5369 Operands.insert(cast<LoadInst>(I)->getPointerOperand()); 5370 break; 5371 5372 // Since dereferenceable attribute imply noundef, atomic operations 5373 // also implicitly have noundef pointers too 5374 case Instruction::AtomicCmpXchg: 5375 Operands.insert(cast<AtomicCmpXchgInst>(I)->getPointerOperand()); 5376 break; 5377 5378 case Instruction::AtomicRMW: 5379 Operands.insert(cast<AtomicRMWInst>(I)->getPointerOperand()); 5380 break; 5381 5382 case Instruction::Call: 5383 case Instruction::Invoke: { 5384 const CallBase *CB = cast<CallBase>(I); 5385 if (CB->isIndirectCall()) 5386 Operands.insert(CB->getCalledOperand()); 5387 for (unsigned i = 0; i < CB->arg_size(); ++i) { 5388 if (CB->paramHasAttr(i, Attribute::NoUndef) || 5389 CB->paramHasAttr(i, Attribute::Dereferenceable)) 5390 Operands.insert(CB->getArgOperand(i)); 5391 } 5392 break; 5393 } 5394 5395 default: 5396 break; 5397 } 5398 } 5399 5400 void llvm::getGuaranteedNonPoisonOps(const Instruction *I, 5401 SmallPtrSetImpl<const Value *> &Operands) { 5402 getGuaranteedWellDefinedOps(I, Operands); 5403 switch (I->getOpcode()) { 5404 // Divisors of these operations are allowed to be partially undef. 5405 case Instruction::UDiv: 5406 case Instruction::SDiv: 5407 case Instruction::URem: 5408 case Instruction::SRem: 5409 Operands.insert(I->getOperand(1)); 5410 break; 5411 5412 default: 5413 break; 5414 } 5415 } 5416 5417 bool llvm::mustTriggerUB(const Instruction *I, 5418 const SmallSet<const Value *, 16>& KnownPoison) { 5419 SmallPtrSet<const Value *, 4> NonPoisonOps; 5420 getGuaranteedNonPoisonOps(I, NonPoisonOps); 5421 5422 for (const auto *V : NonPoisonOps) 5423 if (KnownPoison.count(V)) 5424 return true; 5425 5426 return false; 5427 } 5428 5429 static bool programUndefinedIfUndefOrPoison(const Value *V, 5430 bool PoisonOnly) { 5431 // We currently only look for uses of values within the same basic 5432 // block, as that makes it easier to guarantee that the uses will be 5433 // executed given that Inst is executed. 5434 // 5435 // FIXME: Expand this to consider uses beyond the same basic block. To do 5436 // this, look out for the distinction between post-dominance and strong 5437 // post-dominance. 5438 const BasicBlock *BB = nullptr; 5439 BasicBlock::const_iterator Begin; 5440 if (const auto *Inst = dyn_cast<Instruction>(V)) { 5441 BB = Inst->getParent(); 5442 Begin = Inst->getIterator(); 5443 Begin++; 5444 } else if (const auto *Arg = dyn_cast<Argument>(V)) { 5445 BB = &Arg->getParent()->getEntryBlock(); 5446 Begin = BB->begin(); 5447 } else { 5448 return false; 5449 } 5450 5451 // Limit number of instructions we look at, to avoid scanning through large 5452 // blocks. The current limit is chosen arbitrarily. 5453 unsigned ScanLimit = 32; 5454 BasicBlock::const_iterator End = BB->end(); 5455 5456 if (!PoisonOnly) { 5457 // Since undef does not propagate eagerly, be conservative & just check 5458 // whether a value is directly passed to an instruction that must take 5459 // well-defined operands. 5460 5461 for (auto &I : make_range(Begin, End)) { 5462 if (isa<DbgInfoIntrinsic>(I)) 5463 continue; 5464 if (--ScanLimit == 0) 5465 break; 5466 5467 SmallPtrSet<const Value *, 4> WellDefinedOps; 5468 getGuaranteedWellDefinedOps(&I, WellDefinedOps); 5469 if (WellDefinedOps.contains(V)) 5470 return true; 5471 5472 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) 5473 break; 5474 } 5475 return false; 5476 } 5477 5478 // Set of instructions that we have proved will yield poison if Inst 5479 // does. 5480 SmallSet<const Value *, 16> YieldsPoison; 5481 SmallSet<const BasicBlock *, 4> Visited; 5482 5483 YieldsPoison.insert(V); 5484 auto Propagate = [&](const User *User) { 5485 if (propagatesPoison(cast<Operator>(User))) 5486 YieldsPoison.insert(User); 5487 }; 5488 for_each(V->users(), Propagate); 5489 Visited.insert(BB); 5490 5491 while (true) { 5492 for (auto &I : make_range(Begin, End)) { 5493 if (isa<DbgInfoIntrinsic>(I)) 5494 continue; 5495 if (--ScanLimit == 0) 5496 return false; 5497 if (mustTriggerUB(&I, YieldsPoison)) 5498 return true; 5499 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) 5500 return false; 5501 5502 // Mark poison that propagates from I through uses of I. 5503 if (YieldsPoison.count(&I)) 5504 for_each(I.users(), Propagate); 5505 } 5506 5507 BB = BB->getSingleSuccessor(); 5508 if (!BB || !Visited.insert(BB).second) 5509 break; 5510 5511 Begin = BB->getFirstNonPHI()->getIterator(); 5512 End = BB->end(); 5513 } 5514 return false; 5515 } 5516 5517 bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) { 5518 return ::programUndefinedIfUndefOrPoison(Inst, false); 5519 } 5520 5521 bool llvm::programUndefinedIfPoison(const Instruction *Inst) { 5522 return ::programUndefinedIfUndefOrPoison(Inst, true); 5523 } 5524 5525 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) { 5526 if (FMF.noNaNs()) 5527 return true; 5528 5529 if (auto *C = dyn_cast<ConstantFP>(V)) 5530 return !C->isNaN(); 5531 5532 if (auto *C = dyn_cast<ConstantDataVector>(V)) { 5533 if (!C->getElementType()->isFloatingPointTy()) 5534 return false; 5535 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { 5536 if (C->getElementAsAPFloat(I).isNaN()) 5537 return false; 5538 } 5539 return true; 5540 } 5541 5542 if (isa<ConstantAggregateZero>(V)) 5543 return true; 5544 5545 return false; 5546 } 5547 5548 static bool isKnownNonZero(const Value *V) { 5549 if (auto *C = dyn_cast<ConstantFP>(V)) 5550 return !C->isZero(); 5551 5552 if (auto *C = dyn_cast<ConstantDataVector>(V)) { 5553 if (!C->getElementType()->isFloatingPointTy()) 5554 return false; 5555 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { 5556 if (C->getElementAsAPFloat(I).isZero()) 5557 return false; 5558 } 5559 return true; 5560 } 5561 5562 return false; 5563 } 5564 5565 /// Match clamp pattern for float types without care about NaNs or signed zeros. 5566 /// Given non-min/max outer cmp/select from the clamp pattern this 5567 /// function recognizes if it can be substitued by a "canonical" min/max 5568 /// pattern. 5569 static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred, 5570 Value *CmpLHS, Value *CmpRHS, 5571 Value *TrueVal, Value *FalseVal, 5572 Value *&LHS, Value *&RHS) { 5573 // Try to match 5574 // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2)) 5575 // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2)) 5576 // and return description of the outer Max/Min. 5577 5578 // First, check if select has inverse order: 5579 if (CmpRHS == FalseVal) { 5580 std::swap(TrueVal, FalseVal); 5581 Pred = CmpInst::getInversePredicate(Pred); 5582 } 5583 5584 // Assume success now. If there's no match, callers should not use these anyway. 5585 LHS = TrueVal; 5586 RHS = FalseVal; 5587 5588 const APFloat *FC1; 5589 if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite()) 5590 return {SPF_UNKNOWN, SPNB_NA, false}; 5591 5592 const APFloat *FC2; 5593 switch (Pred) { 5594 case CmpInst::FCMP_OLT: 5595 case CmpInst::FCMP_OLE: 5596 case CmpInst::FCMP_ULT: 5597 case CmpInst::FCMP_ULE: 5598 if (match(FalseVal, 5599 m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)), 5600 m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) && 5601 *FC1 < *FC2) 5602 return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false}; 5603 break; 5604 case CmpInst::FCMP_OGT: 5605 case CmpInst::FCMP_OGE: 5606 case CmpInst::FCMP_UGT: 5607 case CmpInst::FCMP_UGE: 5608 if (match(FalseVal, 5609 m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)), 5610 m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) && 5611 *FC1 > *FC2) 5612 return {SPF_FMINNUM, SPNB_RETURNS_ANY, false}; 5613 break; 5614 default: 5615 break; 5616 } 5617 5618 return {SPF_UNKNOWN, SPNB_NA, false}; 5619 } 5620 5621 /// Recognize variations of: 5622 /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v))) 5623 static SelectPatternResult matchClamp(CmpInst::Predicate Pred, 5624 Value *CmpLHS, Value *CmpRHS, 5625 Value *TrueVal, Value *FalseVal) { 5626 // Swap the select operands and predicate to match the patterns below. 5627 if (CmpRHS != TrueVal) { 5628 Pred = ICmpInst::getSwappedPredicate(Pred); 5629 std::swap(TrueVal, FalseVal); 5630 } 5631 const APInt *C1; 5632 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) { 5633 const APInt *C2; 5634 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1) 5635 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) && 5636 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT) 5637 return {SPF_SMAX, SPNB_NA, false}; 5638 5639 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1) 5640 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) && 5641 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT) 5642 return {SPF_SMIN, SPNB_NA, false}; 5643 5644 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1) 5645 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) && 5646 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT) 5647 return {SPF_UMAX, SPNB_NA, false}; 5648 5649 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1) 5650 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) && 5651 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT) 5652 return {SPF_UMIN, SPNB_NA, false}; 5653 } 5654 return {SPF_UNKNOWN, SPNB_NA, false}; 5655 } 5656 5657 /// Recognize variations of: 5658 /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c)) 5659 static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred, 5660 Value *CmpLHS, Value *CmpRHS, 5661 Value *TVal, Value *FVal, 5662 unsigned Depth) { 5663 // TODO: Allow FP min/max with nnan/nsz. 5664 assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison"); 5665 5666 Value *A = nullptr, *B = nullptr; 5667 SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1); 5668 if (!SelectPatternResult::isMinOrMax(L.Flavor)) 5669 return {SPF_UNKNOWN, SPNB_NA, false}; 5670 5671 Value *C = nullptr, *D = nullptr; 5672 SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1); 5673 if (L.Flavor != R.Flavor) 5674 return {SPF_UNKNOWN, SPNB_NA, false}; 5675 5676 // We have something like: x Pred y ? min(a, b) : min(c, d). 5677 // Try to match the compare to the min/max operations of the select operands. 5678 // First, make sure we have the right compare predicate. 5679 switch (L.Flavor) { 5680 case SPF_SMIN: 5681 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { 5682 Pred = ICmpInst::getSwappedPredicate(Pred); 5683 std::swap(CmpLHS, CmpRHS); 5684 } 5685 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 5686 break; 5687 return {SPF_UNKNOWN, SPNB_NA, false}; 5688 case SPF_SMAX: 5689 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { 5690 Pred = ICmpInst::getSwappedPredicate(Pred); 5691 std::swap(CmpLHS, CmpRHS); 5692 } 5693 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) 5694 break; 5695 return {SPF_UNKNOWN, SPNB_NA, false}; 5696 case SPF_UMIN: 5697 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { 5698 Pred = ICmpInst::getSwappedPredicate(Pred); 5699 std::swap(CmpLHS, CmpRHS); 5700 } 5701 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) 5702 break; 5703 return {SPF_UNKNOWN, SPNB_NA, false}; 5704 case SPF_UMAX: 5705 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 5706 Pred = ICmpInst::getSwappedPredicate(Pred); 5707 std::swap(CmpLHS, CmpRHS); 5708 } 5709 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 5710 break; 5711 return {SPF_UNKNOWN, SPNB_NA, false}; 5712 default: 5713 return {SPF_UNKNOWN, SPNB_NA, false}; 5714 } 5715 5716 // If there is a common operand in the already matched min/max and the other 5717 // min/max operands match the compare operands (either directly or inverted), 5718 // then this is min/max of the same flavor. 5719 5720 // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) 5721 // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) 5722 if (D == B) { 5723 if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && 5724 match(A, m_Not(m_Specific(CmpRHS))))) 5725 return {L.Flavor, SPNB_NA, false}; 5726 } 5727 // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) 5728 // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) 5729 if (C == B) { 5730 if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && 5731 match(A, m_Not(m_Specific(CmpRHS))))) 5732 return {L.Flavor, SPNB_NA, false}; 5733 } 5734 // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) 5735 // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) 5736 if (D == A) { 5737 if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && 5738 match(B, m_Not(m_Specific(CmpRHS))))) 5739 return {L.Flavor, SPNB_NA, false}; 5740 } 5741 // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) 5742 // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) 5743 if (C == A) { 5744 if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && 5745 match(B, m_Not(m_Specific(CmpRHS))))) 5746 return {L.Flavor, SPNB_NA, false}; 5747 } 5748 5749 return {SPF_UNKNOWN, SPNB_NA, false}; 5750 } 5751 5752 /// If the input value is the result of a 'not' op, constant integer, or vector 5753 /// splat of a constant integer, return the bitwise-not source value. 5754 /// TODO: This could be extended to handle non-splat vector integer constants. 5755 static Value *getNotValue(Value *V) { 5756 Value *NotV; 5757 if (match(V, m_Not(m_Value(NotV)))) 5758 return NotV; 5759 5760 const APInt *C; 5761 if (match(V, m_APInt(C))) 5762 return ConstantInt::get(V->getType(), ~(*C)); 5763 5764 return nullptr; 5765 } 5766 5767 /// Match non-obvious integer minimum and maximum sequences. 5768 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred, 5769 Value *CmpLHS, Value *CmpRHS, 5770 Value *TrueVal, Value *FalseVal, 5771 Value *&LHS, Value *&RHS, 5772 unsigned Depth) { 5773 // Assume success. If there's no match, callers should not use these anyway. 5774 LHS = TrueVal; 5775 RHS = FalseVal; 5776 5777 SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal); 5778 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) 5779 return SPR; 5780 5781 SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth); 5782 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) 5783 return SPR; 5784 5785 // Look through 'not' ops to find disguised min/max. 5786 // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y) 5787 // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y) 5788 if (CmpLHS == getNotValue(TrueVal) && CmpRHS == getNotValue(FalseVal)) { 5789 switch (Pred) { 5790 case CmpInst::ICMP_SGT: return {SPF_SMIN, SPNB_NA, false}; 5791 case CmpInst::ICMP_SLT: return {SPF_SMAX, SPNB_NA, false}; 5792 case CmpInst::ICMP_UGT: return {SPF_UMIN, SPNB_NA, false}; 5793 case CmpInst::ICMP_ULT: return {SPF_UMAX, SPNB_NA, false}; 5794 default: break; 5795 } 5796 } 5797 5798 // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X) 5799 // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X) 5800 if (CmpLHS == getNotValue(FalseVal) && CmpRHS == getNotValue(TrueVal)) { 5801 switch (Pred) { 5802 case CmpInst::ICMP_SGT: return {SPF_SMAX, SPNB_NA, false}; 5803 case CmpInst::ICMP_SLT: return {SPF_SMIN, SPNB_NA, false}; 5804 case CmpInst::ICMP_UGT: return {SPF_UMAX, SPNB_NA, false}; 5805 case CmpInst::ICMP_ULT: return {SPF_UMIN, SPNB_NA, false}; 5806 default: break; 5807 } 5808 } 5809 5810 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT) 5811 return {SPF_UNKNOWN, SPNB_NA, false}; 5812 5813 // Z = X -nsw Y 5814 // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0) 5815 // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0) 5816 if (match(TrueVal, m_Zero()) && 5817 match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS)))) 5818 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false}; 5819 5820 // Z = X -nsw Y 5821 // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0) 5822 // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0) 5823 if (match(FalseVal, m_Zero()) && 5824 match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS)))) 5825 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false}; 5826 5827 const APInt *C1; 5828 if (!match(CmpRHS, m_APInt(C1))) 5829 return {SPF_UNKNOWN, SPNB_NA, false}; 5830 5831 // An unsigned min/max can be written with a signed compare. 5832 const APInt *C2; 5833 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) || 5834 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) { 5835 // Is the sign bit set? 5836 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX 5837 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN 5838 if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() && 5839 C2->isMaxSignedValue()) 5840 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; 5841 5842 // Is the sign bit clear? 5843 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX 5844 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN 5845 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() && 5846 C2->isMinSignedValue()) 5847 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; 5848 } 5849 5850 return {SPF_UNKNOWN, SPNB_NA, false}; 5851 } 5852 5853 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) { 5854 assert(X && Y && "Invalid operand"); 5855 5856 // X = sub (0, Y) || X = sub nsw (0, Y) 5857 if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) || 5858 (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y))))) 5859 return true; 5860 5861 // Y = sub (0, X) || Y = sub nsw (0, X) 5862 if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) || 5863 (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X))))) 5864 return true; 5865 5866 // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A) 5867 Value *A, *B; 5868 return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) && 5869 match(Y, m_Sub(m_Specific(B), m_Specific(A))))) || 5870 (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) && 5871 match(Y, m_NSWSub(m_Specific(B), m_Specific(A))))); 5872 } 5873 5874 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, 5875 FastMathFlags FMF, 5876 Value *CmpLHS, Value *CmpRHS, 5877 Value *TrueVal, Value *FalseVal, 5878 Value *&LHS, Value *&RHS, 5879 unsigned Depth) { 5880 if (CmpInst::isFPPredicate(Pred)) { 5881 // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one 5882 // 0.0 operand, set the compare's 0.0 operands to that same value for the 5883 // purpose of identifying min/max. Disregard vector constants with undefined 5884 // elements because those can not be back-propagated for analysis. 5885 Value *OutputZeroVal = nullptr; 5886 if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) && 5887 !cast<Constant>(TrueVal)->containsUndefOrPoisonElement()) 5888 OutputZeroVal = TrueVal; 5889 else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) && 5890 !cast<Constant>(FalseVal)->containsUndefOrPoisonElement()) 5891 OutputZeroVal = FalseVal; 5892 5893 if (OutputZeroVal) { 5894 if (match(CmpLHS, m_AnyZeroFP())) 5895 CmpLHS = OutputZeroVal; 5896 if (match(CmpRHS, m_AnyZeroFP())) 5897 CmpRHS = OutputZeroVal; 5898 } 5899 } 5900 5901 LHS = CmpLHS; 5902 RHS = CmpRHS; 5903 5904 // Signed zero may return inconsistent results between implementations. 5905 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 5906 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) 5907 // Therefore, we behave conservatively and only proceed if at least one of the 5908 // operands is known to not be zero or if we don't care about signed zero. 5909 switch (Pred) { 5910 default: break; 5911 // FIXME: Include OGT/OLT/UGT/ULT. 5912 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE: 5913 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE: 5914 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && 5915 !isKnownNonZero(CmpRHS)) 5916 return {SPF_UNKNOWN, SPNB_NA, false}; 5917 } 5918 5919 SelectPatternNaNBehavior NaNBehavior = SPNB_NA; 5920 bool Ordered = false; 5921 5922 // When given one NaN and one non-NaN input: 5923 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. 5924 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the 5925 // ordered comparison fails), which could be NaN or non-NaN. 5926 // so here we discover exactly what NaN behavior is required/accepted. 5927 if (CmpInst::isFPPredicate(Pred)) { 5928 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF); 5929 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF); 5930 5931 if (LHSSafe && RHSSafe) { 5932 // Both operands are known non-NaN. 5933 NaNBehavior = SPNB_RETURNS_ANY; 5934 } else if (CmpInst::isOrdered(Pred)) { 5935 // An ordered comparison will return false when given a NaN, so it 5936 // returns the RHS. 5937 Ordered = true; 5938 if (LHSSafe) 5939 // LHS is non-NaN, so if RHS is NaN then NaN will be returned. 5940 NaNBehavior = SPNB_RETURNS_NAN; 5941 else if (RHSSafe) 5942 NaNBehavior = SPNB_RETURNS_OTHER; 5943 else 5944 // Completely unsafe. 5945 return {SPF_UNKNOWN, SPNB_NA, false}; 5946 } else { 5947 Ordered = false; 5948 // An unordered comparison will return true when given a NaN, so it 5949 // returns the LHS. 5950 if (LHSSafe) 5951 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. 5952 NaNBehavior = SPNB_RETURNS_OTHER; 5953 else if (RHSSafe) 5954 NaNBehavior = SPNB_RETURNS_NAN; 5955 else 5956 // Completely unsafe. 5957 return {SPF_UNKNOWN, SPNB_NA, false}; 5958 } 5959 } 5960 5961 if (TrueVal == CmpRHS && FalseVal == CmpLHS) { 5962 std::swap(CmpLHS, CmpRHS); 5963 Pred = CmpInst::getSwappedPredicate(Pred); 5964 if (NaNBehavior == SPNB_RETURNS_NAN) 5965 NaNBehavior = SPNB_RETURNS_OTHER; 5966 else if (NaNBehavior == SPNB_RETURNS_OTHER) 5967 NaNBehavior = SPNB_RETURNS_NAN; 5968 Ordered = !Ordered; 5969 } 5970 5971 // ([if]cmp X, Y) ? X : Y 5972 if (TrueVal == CmpLHS && FalseVal == CmpRHS) { 5973 switch (Pred) { 5974 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality. 5975 case ICmpInst::ICMP_UGT: 5976 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false}; 5977 case ICmpInst::ICMP_SGT: 5978 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false}; 5979 case ICmpInst::ICMP_ULT: 5980 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false}; 5981 case ICmpInst::ICMP_SLT: 5982 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false}; 5983 case FCmpInst::FCMP_UGT: 5984 case FCmpInst::FCMP_UGE: 5985 case FCmpInst::FCMP_OGT: 5986 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered}; 5987 case FCmpInst::FCMP_ULT: 5988 case FCmpInst::FCMP_ULE: 5989 case FCmpInst::FCMP_OLT: 5990 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered}; 5991 } 5992 } 5993 5994 if (isKnownNegation(TrueVal, FalseVal)) { 5995 // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can 5996 // match against either LHS or sext(LHS). 5997 auto MaybeSExtCmpLHS = 5998 m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS))); 5999 auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes()); 6000 auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One()); 6001 if (match(TrueVal, MaybeSExtCmpLHS)) { 6002 // Set the return values. If the compare uses the negated value (-X >s 0), 6003 // swap the return values because the negated value is always 'RHS'. 6004 LHS = TrueVal; 6005 RHS = FalseVal; 6006 if (match(CmpLHS, m_Neg(m_Specific(FalseVal)))) 6007 std::swap(LHS, RHS); 6008 6009 // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X) 6010 // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X) 6011 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) 6012 return {SPF_ABS, SPNB_NA, false}; 6013 6014 // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X) 6015 if (Pred == ICmpInst::ICMP_SGE && match(CmpRHS, ZeroOrOne)) 6016 return {SPF_ABS, SPNB_NA, false}; 6017 6018 // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X) 6019 // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X) 6020 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) 6021 return {SPF_NABS, SPNB_NA, false}; 6022 } 6023 else if (match(FalseVal, MaybeSExtCmpLHS)) { 6024 // Set the return values. If the compare uses the negated value (-X >s 0), 6025 // swap the return values because the negated value is always 'RHS'. 6026 LHS = FalseVal; 6027 RHS = TrueVal; 6028 if (match(CmpLHS, m_Neg(m_Specific(TrueVal)))) 6029 std::swap(LHS, RHS); 6030 6031 // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X) 6032 // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X) 6033 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) 6034 return {SPF_NABS, SPNB_NA, false}; 6035 6036 // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X) 6037 // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X) 6038 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) 6039 return {SPF_ABS, SPNB_NA, false}; 6040 } 6041 } 6042 6043 if (CmpInst::isIntPredicate(Pred)) 6044 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth); 6045 6046 // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar 6047 // may return either -0.0 or 0.0, so fcmp/select pair has stricter 6048 // semantics than minNum. Be conservative in such case. 6049 if (NaNBehavior != SPNB_RETURNS_ANY || 6050 (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && 6051 !isKnownNonZero(CmpRHS))) 6052 return {SPF_UNKNOWN, SPNB_NA, false}; 6053 6054 return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); 6055 } 6056 6057 /// Helps to match a select pattern in case of a type mismatch. 6058 /// 6059 /// The function processes the case when type of true and false values of a 6060 /// select instruction differs from type of the cmp instruction operands because 6061 /// of a cast instruction. The function checks if it is legal to move the cast 6062 /// operation after "select". If yes, it returns the new second value of 6063 /// "select" (with the assumption that cast is moved): 6064 /// 1. As operand of cast instruction when both values of "select" are same cast 6065 /// instructions. 6066 /// 2. As restored constant (by applying reverse cast operation) when the first 6067 /// value of the "select" is a cast operation and the second value is a 6068 /// constant. 6069 /// NOTE: We return only the new second value because the first value could be 6070 /// accessed as operand of cast instruction. 6071 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, 6072 Instruction::CastOps *CastOp) { 6073 auto *Cast1 = dyn_cast<CastInst>(V1); 6074 if (!Cast1) 6075 return nullptr; 6076 6077 *CastOp = Cast1->getOpcode(); 6078 Type *SrcTy = Cast1->getSrcTy(); 6079 if (auto *Cast2 = dyn_cast<CastInst>(V2)) { 6080 // If V1 and V2 are both the same cast from the same type, look through V1. 6081 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy()) 6082 return Cast2->getOperand(0); 6083 return nullptr; 6084 } 6085 6086 auto *C = dyn_cast<Constant>(V2); 6087 if (!C) 6088 return nullptr; 6089 6090 Constant *CastedTo = nullptr; 6091 switch (*CastOp) { 6092 case Instruction::ZExt: 6093 if (CmpI->isUnsigned()) 6094 CastedTo = ConstantExpr::getTrunc(C, SrcTy); 6095 break; 6096 case Instruction::SExt: 6097 if (CmpI->isSigned()) 6098 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true); 6099 break; 6100 case Instruction::Trunc: 6101 Constant *CmpConst; 6102 if (match(CmpI->getOperand(1), m_Constant(CmpConst)) && 6103 CmpConst->getType() == SrcTy) { 6104 // Here we have the following case: 6105 // 6106 // %cond = cmp iN %x, CmpConst 6107 // %tr = trunc iN %x to iK 6108 // %narrowsel = select i1 %cond, iK %t, iK C 6109 // 6110 // We can always move trunc after select operation: 6111 // 6112 // %cond = cmp iN %x, CmpConst 6113 // %widesel = select i1 %cond, iN %x, iN CmpConst 6114 // %tr = trunc iN %widesel to iK 6115 // 6116 // Note that C could be extended in any way because we don't care about 6117 // upper bits after truncation. It can't be abs pattern, because it would 6118 // look like: 6119 // 6120 // select i1 %cond, x, -x. 6121 // 6122 // So only min/max pattern could be matched. Such match requires widened C 6123 // == CmpConst. That is why set widened C = CmpConst, condition trunc 6124 // CmpConst == C is checked below. 6125 CastedTo = CmpConst; 6126 } else { 6127 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned()); 6128 } 6129 break; 6130 case Instruction::FPTrunc: 6131 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true); 6132 break; 6133 case Instruction::FPExt: 6134 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true); 6135 break; 6136 case Instruction::FPToUI: 6137 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true); 6138 break; 6139 case Instruction::FPToSI: 6140 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true); 6141 break; 6142 case Instruction::UIToFP: 6143 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true); 6144 break; 6145 case Instruction::SIToFP: 6146 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true); 6147 break; 6148 default: 6149 break; 6150 } 6151 6152 if (!CastedTo) 6153 return nullptr; 6154 6155 // Make sure the cast doesn't lose any information. 6156 Constant *CastedBack = 6157 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true); 6158 if (CastedBack != C) 6159 return nullptr; 6160 6161 return CastedTo; 6162 } 6163 6164 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, 6165 Instruction::CastOps *CastOp, 6166 unsigned Depth) { 6167 if (Depth >= MaxAnalysisRecursionDepth) 6168 return {SPF_UNKNOWN, SPNB_NA, false}; 6169 6170 SelectInst *SI = dyn_cast<SelectInst>(V); 6171 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false}; 6172 6173 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition()); 6174 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false}; 6175 6176 Value *TrueVal = SI->getTrueValue(); 6177 Value *FalseVal = SI->getFalseValue(); 6178 6179 return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS, 6180 CastOp, Depth); 6181 } 6182 6183 SelectPatternResult llvm::matchDecomposedSelectPattern( 6184 CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS, 6185 Instruction::CastOps *CastOp, unsigned Depth) { 6186 CmpInst::Predicate Pred = CmpI->getPredicate(); 6187 Value *CmpLHS = CmpI->getOperand(0); 6188 Value *CmpRHS = CmpI->getOperand(1); 6189 FastMathFlags FMF; 6190 if (isa<FPMathOperator>(CmpI)) 6191 FMF = CmpI->getFastMathFlags(); 6192 6193 // Bail out early. 6194 if (CmpI->isEquality()) 6195 return {SPF_UNKNOWN, SPNB_NA, false}; 6196 6197 // Deal with type mismatches. 6198 if (CastOp && CmpLHS->getType() != TrueVal->getType()) { 6199 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) { 6200 // If this is a potential fmin/fmax with a cast to integer, then ignore 6201 // -0.0 because there is no corresponding integer value. 6202 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) 6203 FMF.setNoSignedZeros(); 6204 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, 6205 cast<CastInst>(TrueVal)->getOperand(0), C, 6206 LHS, RHS, Depth); 6207 } 6208 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) { 6209 // If this is a potential fmin/fmax with a cast to integer, then ignore 6210 // -0.0 because there is no corresponding integer value. 6211 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) 6212 FMF.setNoSignedZeros(); 6213 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, 6214 C, cast<CastInst>(FalseVal)->getOperand(0), 6215 LHS, RHS, Depth); 6216 } 6217 } 6218 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, 6219 LHS, RHS, Depth); 6220 } 6221 6222 CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) { 6223 if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT; 6224 if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT; 6225 if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT; 6226 if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT; 6227 if (SPF == SPF_FMINNUM) 6228 return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; 6229 if (SPF == SPF_FMAXNUM) 6230 return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; 6231 llvm_unreachable("unhandled!"); 6232 } 6233 6234 SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) { 6235 if (SPF == SPF_SMIN) return SPF_SMAX; 6236 if (SPF == SPF_UMIN) return SPF_UMAX; 6237 if (SPF == SPF_SMAX) return SPF_SMIN; 6238 if (SPF == SPF_UMAX) return SPF_UMIN; 6239 llvm_unreachable("unhandled!"); 6240 } 6241 6242 Intrinsic::ID llvm::getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID) { 6243 switch (MinMaxID) { 6244 case Intrinsic::smax: return Intrinsic::smin; 6245 case Intrinsic::smin: return Intrinsic::smax; 6246 case Intrinsic::umax: return Intrinsic::umin; 6247 case Intrinsic::umin: return Intrinsic::umax; 6248 default: llvm_unreachable("Unexpected intrinsic"); 6249 } 6250 } 6251 6252 CmpInst::Predicate llvm::getInverseMinMaxPred(SelectPatternFlavor SPF) { 6253 return getMinMaxPred(getInverseMinMaxFlavor(SPF)); 6254 } 6255 6256 APInt llvm::getMinMaxLimit(SelectPatternFlavor SPF, unsigned BitWidth) { 6257 switch (SPF) { 6258 case SPF_SMAX: return APInt::getSignedMaxValue(BitWidth); 6259 case SPF_SMIN: return APInt::getSignedMinValue(BitWidth); 6260 case SPF_UMAX: return APInt::getMaxValue(BitWidth); 6261 case SPF_UMIN: return APInt::getMinValue(BitWidth); 6262 default: llvm_unreachable("Unexpected flavor"); 6263 } 6264 } 6265 6266 std::pair<Intrinsic::ID, bool> 6267 llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) { 6268 // Check if VL contains select instructions that can be folded into a min/max 6269 // vector intrinsic and return the intrinsic if it is possible. 6270 // TODO: Support floating point min/max. 6271 bool AllCmpSingleUse = true; 6272 SelectPatternResult SelectPattern; 6273 SelectPattern.Flavor = SPF_UNKNOWN; 6274 if (all_of(VL, [&SelectPattern, &AllCmpSingleUse](Value *I) { 6275 Value *LHS, *RHS; 6276 auto CurrentPattern = matchSelectPattern(I, LHS, RHS); 6277 if (!SelectPatternResult::isMinOrMax(CurrentPattern.Flavor) || 6278 CurrentPattern.Flavor == SPF_FMINNUM || 6279 CurrentPattern.Flavor == SPF_FMAXNUM || 6280 !I->getType()->isIntOrIntVectorTy()) 6281 return false; 6282 if (SelectPattern.Flavor != SPF_UNKNOWN && 6283 SelectPattern.Flavor != CurrentPattern.Flavor) 6284 return false; 6285 SelectPattern = CurrentPattern; 6286 AllCmpSingleUse &= 6287 match(I, m_Select(m_OneUse(m_Value()), m_Value(), m_Value())); 6288 return true; 6289 })) { 6290 switch (SelectPattern.Flavor) { 6291 case SPF_SMIN: 6292 return {Intrinsic::smin, AllCmpSingleUse}; 6293 case SPF_UMIN: 6294 return {Intrinsic::umin, AllCmpSingleUse}; 6295 case SPF_SMAX: 6296 return {Intrinsic::smax, AllCmpSingleUse}; 6297 case SPF_UMAX: 6298 return {Intrinsic::umax, AllCmpSingleUse}; 6299 default: 6300 llvm_unreachable("unexpected select pattern flavor"); 6301 } 6302 } 6303 return {Intrinsic::not_intrinsic, false}; 6304 } 6305 6306 bool llvm::matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO, 6307 Value *&Start, Value *&Step) { 6308 // Handle the case of a simple two-predecessor recurrence PHI. 6309 // There's a lot more that could theoretically be done here, but 6310 // this is sufficient to catch some interesting cases. 6311 if (P->getNumIncomingValues() != 2) 6312 return false; 6313 6314 for (unsigned i = 0; i != 2; ++i) { 6315 Value *L = P->getIncomingValue(i); 6316 Value *R = P->getIncomingValue(!i); 6317 Operator *LU = dyn_cast<Operator>(L); 6318 if (!LU) 6319 continue; 6320 unsigned Opcode = LU->getOpcode(); 6321 6322 switch (Opcode) { 6323 default: 6324 continue; 6325 // TODO: Expand list -- xor, div, gep, uaddo, etc.. 6326 case Instruction::LShr: 6327 case Instruction::AShr: 6328 case Instruction::Shl: 6329 case Instruction::Add: 6330 case Instruction::Sub: 6331 case Instruction::And: 6332 case Instruction::Or: 6333 case Instruction::Mul: { 6334 Value *LL = LU->getOperand(0); 6335 Value *LR = LU->getOperand(1); 6336 // Find a recurrence. 6337 if (LL == P) 6338 L = LR; 6339 else if (LR == P) 6340 L = LL; 6341 else 6342 continue; // Check for recurrence with L and R flipped. 6343 6344 break; // Match! 6345 } 6346 }; 6347 6348 // We have matched a recurrence of the form: 6349 // %iv = [R, %entry], [%iv.next, %backedge] 6350 // %iv.next = binop %iv, L 6351 // OR 6352 // %iv = [R, %entry], [%iv.next, %backedge] 6353 // %iv.next = binop L, %iv 6354 BO = cast<BinaryOperator>(LU); 6355 Start = R; 6356 Step = L; 6357 return true; 6358 } 6359 return false; 6360 } 6361 6362 bool llvm::matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P, 6363 Value *&Start, Value *&Step) { 6364 BinaryOperator *BO = nullptr; 6365 P = dyn_cast<PHINode>(I->getOperand(0)); 6366 if (!P) 6367 P = dyn_cast<PHINode>(I->getOperand(1)); 6368 return P && matchSimpleRecurrence(P, BO, Start, Step) && BO == I; 6369 } 6370 6371 /// Return true if "icmp Pred LHS RHS" is always true. 6372 static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, 6373 const Value *RHS, const DataLayout &DL, 6374 unsigned Depth) { 6375 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!"); 6376 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS) 6377 return true; 6378 6379 switch (Pred) { 6380 default: 6381 return false; 6382 6383 case CmpInst::ICMP_SLE: { 6384 const APInt *C; 6385 6386 // LHS s<= LHS +_{nsw} C if C >= 0 6387 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C)))) 6388 return !C->isNegative(); 6389 return false; 6390 } 6391 6392 case CmpInst::ICMP_ULE: { 6393 const APInt *C; 6394 6395 // LHS u<= LHS +_{nuw} C for any C 6396 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C)))) 6397 return true; 6398 6399 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) 6400 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B, 6401 const Value *&X, 6402 const APInt *&CA, const APInt *&CB) { 6403 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) && 6404 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB)))) 6405 return true; 6406 6407 // If X & C == 0 then (X | C) == X +_{nuw} C 6408 if (match(A, m_Or(m_Value(X), m_APInt(CA))) && 6409 match(B, m_Or(m_Specific(X), m_APInt(CB)))) { 6410 KnownBits Known(CA->getBitWidth()); 6411 computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr, 6412 /*CxtI*/ nullptr, /*DT*/ nullptr); 6413 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero)) 6414 return true; 6415 } 6416 6417 return false; 6418 }; 6419 6420 const Value *X; 6421 const APInt *CLHS, *CRHS; 6422 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS)) 6423 return CLHS->ule(*CRHS); 6424 6425 return false; 6426 } 6427 } 6428 } 6429 6430 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred 6431 /// ALHS ARHS" is true. Otherwise, return None. 6432 static Optional<bool> 6433 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS, 6434 const Value *ARHS, const Value *BLHS, const Value *BRHS, 6435 const DataLayout &DL, unsigned Depth) { 6436 switch (Pred) { 6437 default: 6438 return None; 6439 6440 case CmpInst::ICMP_SLT: 6441 case CmpInst::ICMP_SLE: 6442 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) && 6443 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth)) 6444 return true; 6445 return None; 6446 6447 case CmpInst::ICMP_ULT: 6448 case CmpInst::ICMP_ULE: 6449 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) && 6450 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth)) 6451 return true; 6452 return None; 6453 } 6454 } 6455 6456 /// Return true if the operands of the two compares match. IsSwappedOps is true 6457 /// when the operands match, but are swapped. 6458 static bool isMatchingOps(const Value *ALHS, const Value *ARHS, 6459 const Value *BLHS, const Value *BRHS, 6460 bool &IsSwappedOps) { 6461 6462 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS); 6463 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS); 6464 return IsMatchingOps || IsSwappedOps; 6465 } 6466 6467 /// Return true if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is true. 6468 /// Return false if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is false. 6469 /// Otherwise, return None if we can't infer anything. 6470 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred, 6471 CmpInst::Predicate BPred, 6472 bool AreSwappedOps) { 6473 // Canonicalize the predicate as if the operands were not commuted. 6474 if (AreSwappedOps) 6475 BPred = ICmpInst::getSwappedPredicate(BPred); 6476 6477 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred)) 6478 return true; 6479 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred)) 6480 return false; 6481 6482 return None; 6483 } 6484 6485 /// Return true if "icmp APred X, C1" implies "icmp BPred X, C2" is true. 6486 /// Return false if "icmp APred X, C1" implies "icmp BPred X, C2" is false. 6487 /// Otherwise, return None if we can't infer anything. 6488 static Optional<bool> 6489 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, 6490 const ConstantInt *C1, 6491 CmpInst::Predicate BPred, 6492 const ConstantInt *C2) { 6493 ConstantRange DomCR = 6494 ConstantRange::makeExactICmpRegion(APred, C1->getValue()); 6495 ConstantRange CR = ConstantRange::makeExactICmpRegion(BPred, C2->getValue()); 6496 ConstantRange Intersection = DomCR.intersectWith(CR); 6497 ConstantRange Difference = DomCR.difference(CR); 6498 if (Intersection.isEmptySet()) 6499 return false; 6500 if (Difference.isEmptySet()) 6501 return true; 6502 return None; 6503 } 6504 6505 /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is 6506 /// false. Otherwise, return None if we can't infer anything. 6507 static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS, 6508 CmpInst::Predicate BPred, 6509 const Value *BLHS, const Value *BRHS, 6510 const DataLayout &DL, bool LHSIsTrue, 6511 unsigned Depth) { 6512 Value *ALHS = LHS->getOperand(0); 6513 Value *ARHS = LHS->getOperand(1); 6514 6515 // The rest of the logic assumes the LHS condition is true. If that's not the 6516 // case, invert the predicate to make it so. 6517 CmpInst::Predicate APred = 6518 LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate(); 6519 6520 // Can we infer anything when the two compares have matching operands? 6521 bool AreSwappedOps; 6522 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, AreSwappedOps)) { 6523 if (Optional<bool> Implication = isImpliedCondMatchingOperands( 6524 APred, BPred, AreSwappedOps)) 6525 return Implication; 6526 // No amount of additional analysis will infer the second condition, so 6527 // early exit. 6528 return None; 6529 } 6530 6531 // Can we infer anything when the LHS operands match and the RHS operands are 6532 // constants (not necessarily matching)? 6533 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) { 6534 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands( 6535 APred, cast<ConstantInt>(ARHS), BPred, cast<ConstantInt>(BRHS))) 6536 return Implication; 6537 // No amount of additional analysis will infer the second condition, so 6538 // early exit. 6539 return None; 6540 } 6541 6542 if (APred == BPred) 6543 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth); 6544 return None; 6545 } 6546 6547 /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is 6548 /// false. Otherwise, return None if we can't infer anything. We expect the 6549 /// RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select' instruction. 6550 static Optional<bool> 6551 isImpliedCondAndOr(const Instruction *LHS, CmpInst::Predicate RHSPred, 6552 const Value *RHSOp0, const Value *RHSOp1, 6553 const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { 6554 // The LHS must be an 'or', 'and', or a 'select' instruction. 6555 assert((LHS->getOpcode() == Instruction::And || 6556 LHS->getOpcode() == Instruction::Or || 6557 LHS->getOpcode() == Instruction::Select) && 6558 "Expected LHS to be 'and', 'or', or 'select'."); 6559 6560 assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit"); 6561 6562 // If the result of an 'or' is false, then we know both legs of the 'or' are 6563 // false. Similarly, if the result of an 'and' is true, then we know both 6564 // legs of the 'and' are true. 6565 const Value *ALHS, *ARHS; 6566 if ((!LHSIsTrue && match(LHS, m_LogicalOr(m_Value(ALHS), m_Value(ARHS)))) || 6567 (LHSIsTrue && match(LHS, m_LogicalAnd(m_Value(ALHS), m_Value(ARHS))))) { 6568 // FIXME: Make this non-recursion. 6569 if (Optional<bool> Implication = isImpliedCondition( 6570 ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) 6571 return Implication; 6572 if (Optional<bool> Implication = isImpliedCondition( 6573 ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) 6574 return Implication; 6575 return None; 6576 } 6577 return None; 6578 } 6579 6580 Optional<bool> 6581 llvm::isImpliedCondition(const Value *LHS, CmpInst::Predicate RHSPred, 6582 const Value *RHSOp0, const Value *RHSOp1, 6583 const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { 6584 // Bail out when we hit the limit. 6585 if (Depth == MaxAnalysisRecursionDepth) 6586 return None; 6587 6588 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for 6589 // example. 6590 if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy()) 6591 return None; 6592 6593 Type *OpTy = LHS->getType(); 6594 assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!"); 6595 6596 // FIXME: Extending the code below to handle vectors. 6597 if (OpTy->isVectorTy()) 6598 return None; 6599 6600 assert(OpTy->isIntegerTy(1) && "implied by above"); 6601 6602 // Both LHS and RHS are icmps. 6603 const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS); 6604 if (LHSCmp) 6605 return isImpliedCondICmps(LHSCmp, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, 6606 Depth); 6607 6608 /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect 6609 /// the RHS to be an icmp. 6610 /// FIXME: Add support for and/or/select on the RHS. 6611 if (const Instruction *LHSI = dyn_cast<Instruction>(LHS)) { 6612 if ((LHSI->getOpcode() == Instruction::And || 6613 LHSI->getOpcode() == Instruction::Or || 6614 LHSI->getOpcode() == Instruction::Select)) 6615 return isImpliedCondAndOr(LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, 6616 Depth); 6617 } 6618 return None; 6619 } 6620 6621 Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS, 6622 const DataLayout &DL, bool LHSIsTrue, 6623 unsigned Depth) { 6624 // LHS ==> RHS by definition 6625 if (LHS == RHS) 6626 return LHSIsTrue; 6627 6628 const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS); 6629 if (RHSCmp) 6630 return isImpliedCondition(LHS, RHSCmp->getPredicate(), 6631 RHSCmp->getOperand(0), RHSCmp->getOperand(1), DL, 6632 LHSIsTrue, Depth); 6633 return None; 6634 } 6635 6636 // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch 6637 // condition dominating ContextI or nullptr, if no condition is found. 6638 static std::pair<Value *, bool> 6639 getDomPredecessorCondition(const Instruction *ContextI) { 6640 if (!ContextI || !ContextI->getParent()) 6641 return {nullptr, false}; 6642 6643 // TODO: This is a poor/cheap way to determine dominance. Should we use a 6644 // dominator tree (eg, from a SimplifyQuery) instead? 6645 const BasicBlock *ContextBB = ContextI->getParent(); 6646 const BasicBlock *PredBB = ContextBB->getSinglePredecessor(); 6647 if (!PredBB) 6648 return {nullptr, false}; 6649 6650 // We need a conditional branch in the predecessor. 6651 Value *PredCond; 6652 BasicBlock *TrueBB, *FalseBB; 6653 if (!match(PredBB->getTerminator(), m_Br(m_Value(PredCond), TrueBB, FalseBB))) 6654 return {nullptr, false}; 6655 6656 // The branch should get simplified. Don't bother simplifying this condition. 6657 if (TrueBB == FalseBB) 6658 return {nullptr, false}; 6659 6660 assert((TrueBB == ContextBB || FalseBB == ContextBB) && 6661 "Predecessor block does not point to successor?"); 6662 6663 // Is this condition implied by the predecessor condition? 6664 return {PredCond, TrueBB == ContextBB}; 6665 } 6666 6667 Optional<bool> llvm::isImpliedByDomCondition(const Value *Cond, 6668 const Instruction *ContextI, 6669 const DataLayout &DL) { 6670 assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool"); 6671 auto PredCond = getDomPredecessorCondition(ContextI); 6672 if (PredCond.first) 6673 return isImpliedCondition(PredCond.first, Cond, DL, PredCond.second); 6674 return None; 6675 } 6676 6677 Optional<bool> llvm::isImpliedByDomCondition(CmpInst::Predicate Pred, 6678 const Value *LHS, const Value *RHS, 6679 const Instruction *ContextI, 6680 const DataLayout &DL) { 6681 auto PredCond = getDomPredecessorCondition(ContextI); 6682 if (PredCond.first) 6683 return isImpliedCondition(PredCond.first, Pred, LHS, RHS, DL, 6684 PredCond.second); 6685 return None; 6686 } 6687 6688 static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower, 6689 APInt &Upper, const InstrInfoQuery &IIQ) { 6690 unsigned Width = Lower.getBitWidth(); 6691 const APInt *C; 6692 switch (BO.getOpcode()) { 6693 case Instruction::Add: 6694 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) { 6695 // FIXME: If we have both nuw and nsw, we should reduce the range further. 6696 if (IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(&BO))) { 6697 // 'add nuw x, C' produces [C, UINT_MAX]. 6698 Lower = *C; 6699 } else if (IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(&BO))) { 6700 if (C->isNegative()) { 6701 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. 6702 Lower = APInt::getSignedMinValue(Width); 6703 Upper = APInt::getSignedMaxValue(Width) + *C + 1; 6704 } else { 6705 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. 6706 Lower = APInt::getSignedMinValue(Width) + *C; 6707 Upper = APInt::getSignedMaxValue(Width) + 1; 6708 } 6709 } 6710 } 6711 break; 6712 6713 case Instruction::And: 6714 if (match(BO.getOperand(1), m_APInt(C))) 6715 // 'and x, C' produces [0, C]. 6716 Upper = *C + 1; 6717 break; 6718 6719 case Instruction::Or: 6720 if (match(BO.getOperand(1), m_APInt(C))) 6721 // 'or x, C' produces [C, UINT_MAX]. 6722 Lower = *C; 6723 break; 6724 6725 case Instruction::AShr: 6726 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { 6727 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. 6728 Lower = APInt::getSignedMinValue(Width).ashr(*C); 6729 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1; 6730 } else if (match(BO.getOperand(0), m_APInt(C))) { 6731 unsigned ShiftAmount = Width - 1; 6732 if (!C->isNullValue() && IIQ.isExact(&BO)) 6733 ShiftAmount = C->countTrailingZeros(); 6734 if (C->isNegative()) { 6735 // 'ashr C, x' produces [C, C >> (Width-1)] 6736 Lower = *C; 6737 Upper = C->ashr(ShiftAmount) + 1; 6738 } else { 6739 // 'ashr C, x' produces [C >> (Width-1), C] 6740 Lower = C->ashr(ShiftAmount); 6741 Upper = *C + 1; 6742 } 6743 } 6744 break; 6745 6746 case Instruction::LShr: 6747 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { 6748 // 'lshr x, C' produces [0, UINT_MAX >> C]. 6749 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1; 6750 } else if (match(BO.getOperand(0), m_APInt(C))) { 6751 // 'lshr C, x' produces [C >> (Width-1), C]. 6752 unsigned ShiftAmount = Width - 1; 6753 if (!C->isNullValue() && IIQ.isExact(&BO)) 6754 ShiftAmount = C->countTrailingZeros(); 6755 Lower = C->lshr(ShiftAmount); 6756 Upper = *C + 1; 6757 } 6758 break; 6759 6760 case Instruction::Shl: 6761 if (match(BO.getOperand(0), m_APInt(C))) { 6762 if (IIQ.hasNoUnsignedWrap(&BO)) { 6763 // 'shl nuw C, x' produces [C, C << CLZ(C)] 6764 Lower = *C; 6765 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 6766 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? 6767 if (C->isNegative()) { 6768 // 'shl nsw C, x' produces [C << CLO(C)-1, C] 6769 unsigned ShiftAmount = C->countLeadingOnes() - 1; 6770 Lower = C->shl(ShiftAmount); 6771 Upper = *C + 1; 6772 } else { 6773 // 'shl nsw C, x' produces [C, C << CLZ(C)-1] 6774 unsigned ShiftAmount = C->countLeadingZeros() - 1; 6775 Lower = *C; 6776 Upper = C->shl(ShiftAmount) + 1; 6777 } 6778 } 6779 } 6780 break; 6781 6782 case Instruction::SDiv: 6783 if (match(BO.getOperand(1), m_APInt(C))) { 6784 APInt IntMin = APInt::getSignedMinValue(Width); 6785 APInt IntMax = APInt::getSignedMaxValue(Width); 6786 if (C->isAllOnesValue()) { 6787 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 6788 // where C != -1 and C != 0 and C != 1 6789 Lower = IntMin + 1; 6790 Upper = IntMax + 1; 6791 } else if (C->countLeadingZeros() < Width - 1) { 6792 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] 6793 // where C != -1 and C != 0 and C != 1 6794 Lower = IntMin.sdiv(*C); 6795 Upper = IntMax.sdiv(*C); 6796 if (Lower.sgt(Upper)) 6797 std::swap(Lower, Upper); 6798 Upper = Upper + 1; 6799 assert(Upper != Lower && "Upper part of range has wrapped!"); 6800 } 6801 } else if (match(BO.getOperand(0), m_APInt(C))) { 6802 if (C->isMinSignedValue()) { 6803 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 6804 Lower = *C; 6805 Upper = Lower.lshr(1) + 1; 6806 } else { 6807 // 'sdiv C, x' produces [-|C|, |C|]. 6808 Upper = C->abs() + 1; 6809 Lower = (-Upper) + 1; 6810 } 6811 } 6812 break; 6813 6814 case Instruction::UDiv: 6815 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) { 6816 // 'udiv x, C' produces [0, UINT_MAX / C]. 6817 Upper = APInt::getMaxValue(Width).udiv(*C) + 1; 6818 } else if (match(BO.getOperand(0), m_APInt(C))) { 6819 // 'udiv C, x' produces [0, C]. 6820 Upper = *C + 1; 6821 } 6822 break; 6823 6824 case Instruction::SRem: 6825 if (match(BO.getOperand(1), m_APInt(C))) { 6826 // 'srem x, C' produces (-|C|, |C|). 6827 Upper = C->abs(); 6828 Lower = (-Upper) + 1; 6829 } 6830 break; 6831 6832 case Instruction::URem: 6833 if (match(BO.getOperand(1), m_APInt(C))) 6834 // 'urem x, C' produces [0, C). 6835 Upper = *C; 6836 break; 6837 6838 default: 6839 break; 6840 } 6841 } 6842 6843 static void setLimitsForIntrinsic(const IntrinsicInst &II, APInt &Lower, 6844 APInt &Upper) { 6845 unsigned Width = Lower.getBitWidth(); 6846 const APInt *C; 6847 switch (II.getIntrinsicID()) { 6848 case Intrinsic::ctpop: 6849 case Intrinsic::ctlz: 6850 case Intrinsic::cttz: 6851 // Maximum of set/clear bits is the bit width. 6852 assert(Lower == 0 && "Expected lower bound to be zero"); 6853 Upper = Width + 1; 6854 break; 6855 case Intrinsic::uadd_sat: 6856 // uadd.sat(x, C) produces [C, UINT_MAX]. 6857 if (match(II.getOperand(0), m_APInt(C)) || 6858 match(II.getOperand(1), m_APInt(C))) 6859 Lower = *C; 6860 break; 6861 case Intrinsic::sadd_sat: 6862 if (match(II.getOperand(0), m_APInt(C)) || 6863 match(II.getOperand(1), m_APInt(C))) { 6864 if (C->isNegative()) { 6865 // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)]. 6866 Lower = APInt::getSignedMinValue(Width); 6867 Upper = APInt::getSignedMaxValue(Width) + *C + 1; 6868 } else { 6869 // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX]. 6870 Lower = APInt::getSignedMinValue(Width) + *C; 6871 Upper = APInt::getSignedMaxValue(Width) + 1; 6872 } 6873 } 6874 break; 6875 case Intrinsic::usub_sat: 6876 // usub.sat(C, x) produces [0, C]. 6877 if (match(II.getOperand(0), m_APInt(C))) 6878 Upper = *C + 1; 6879 // usub.sat(x, C) produces [0, UINT_MAX - C]. 6880 else if (match(II.getOperand(1), m_APInt(C))) 6881 Upper = APInt::getMaxValue(Width) - *C + 1; 6882 break; 6883 case Intrinsic::ssub_sat: 6884 if (match(II.getOperand(0), m_APInt(C))) { 6885 if (C->isNegative()) { 6886 // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)]. 6887 Lower = APInt::getSignedMinValue(Width); 6888 Upper = *C - APInt::getSignedMinValue(Width) + 1; 6889 } else { 6890 // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX]. 6891 Lower = *C - APInt::getSignedMaxValue(Width); 6892 Upper = APInt::getSignedMaxValue(Width) + 1; 6893 } 6894 } else if (match(II.getOperand(1), m_APInt(C))) { 6895 if (C->isNegative()) { 6896 // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]: 6897 Lower = APInt::getSignedMinValue(Width) - *C; 6898 Upper = APInt::getSignedMaxValue(Width) + 1; 6899 } else { 6900 // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C]. 6901 Lower = APInt::getSignedMinValue(Width); 6902 Upper = APInt::getSignedMaxValue(Width) - *C + 1; 6903 } 6904 } 6905 break; 6906 case Intrinsic::umin: 6907 case Intrinsic::umax: 6908 case Intrinsic::smin: 6909 case Intrinsic::smax: 6910 if (!match(II.getOperand(0), m_APInt(C)) && 6911 !match(II.getOperand(1), m_APInt(C))) 6912 break; 6913 6914 switch (II.getIntrinsicID()) { 6915 case Intrinsic::umin: 6916 Upper = *C + 1; 6917 break; 6918 case Intrinsic::umax: 6919 Lower = *C; 6920 break; 6921 case Intrinsic::smin: 6922 Lower = APInt::getSignedMinValue(Width); 6923 Upper = *C + 1; 6924 break; 6925 case Intrinsic::smax: 6926 Lower = *C; 6927 Upper = APInt::getSignedMaxValue(Width) + 1; 6928 break; 6929 default: 6930 llvm_unreachable("Must be min/max intrinsic"); 6931 } 6932 break; 6933 case Intrinsic::abs: 6934 // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX], 6935 // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. 6936 if (match(II.getOperand(1), m_One())) 6937 Upper = APInt::getSignedMaxValue(Width) + 1; 6938 else 6939 Upper = APInt::getSignedMinValue(Width) + 1; 6940 break; 6941 default: 6942 break; 6943 } 6944 } 6945 6946 static void setLimitsForSelectPattern(const SelectInst &SI, APInt &Lower, 6947 APInt &Upper, const InstrInfoQuery &IIQ) { 6948 const Value *LHS = nullptr, *RHS = nullptr; 6949 SelectPatternResult R = matchSelectPattern(&SI, LHS, RHS); 6950 if (R.Flavor == SPF_UNKNOWN) 6951 return; 6952 6953 unsigned BitWidth = SI.getType()->getScalarSizeInBits(); 6954 6955 if (R.Flavor == SelectPatternFlavor::SPF_ABS) { 6956 // If the negation part of the abs (in RHS) has the NSW flag, 6957 // then the result of abs(X) is [0..SIGNED_MAX], 6958 // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. 6959 Lower = APInt::getNullValue(BitWidth); 6960 if (match(RHS, m_Neg(m_Specific(LHS))) && 6961 IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) 6962 Upper = APInt::getSignedMaxValue(BitWidth) + 1; 6963 else 6964 Upper = APInt::getSignedMinValue(BitWidth) + 1; 6965 return; 6966 } 6967 6968 if (R.Flavor == SelectPatternFlavor::SPF_NABS) { 6969 // The result of -abs(X) is <= 0. 6970 Lower = APInt::getSignedMinValue(BitWidth); 6971 Upper = APInt(BitWidth, 1); 6972 return; 6973 } 6974 6975 const APInt *C; 6976 if (!match(LHS, m_APInt(C)) && !match(RHS, m_APInt(C))) 6977 return; 6978 6979 switch (R.Flavor) { 6980 case SPF_UMIN: 6981 Upper = *C + 1; 6982 break; 6983 case SPF_UMAX: 6984 Lower = *C; 6985 break; 6986 case SPF_SMIN: 6987 Lower = APInt::getSignedMinValue(BitWidth); 6988 Upper = *C + 1; 6989 break; 6990 case SPF_SMAX: 6991 Lower = *C; 6992 Upper = APInt::getSignedMaxValue(BitWidth) + 1; 6993 break; 6994 default: 6995 break; 6996 } 6997 } 6998 6999 ConstantRange llvm::computeConstantRange(const Value *V, bool UseInstrInfo, 7000 AssumptionCache *AC, 7001 const Instruction *CtxI, 7002 unsigned Depth) { 7003 assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction"); 7004 7005 if (Depth == MaxAnalysisRecursionDepth) 7006 return ConstantRange::getFull(V->getType()->getScalarSizeInBits()); 7007 7008 const APInt *C; 7009 if (match(V, m_APInt(C))) 7010 return ConstantRange(*C); 7011 7012 InstrInfoQuery IIQ(UseInstrInfo); 7013 unsigned BitWidth = V->getType()->getScalarSizeInBits(); 7014 APInt Lower = APInt(BitWidth, 0); 7015 APInt Upper = APInt(BitWidth, 0); 7016 if (auto *BO = dyn_cast<BinaryOperator>(V)) 7017 setLimitsForBinOp(*BO, Lower, Upper, IIQ); 7018 else if (auto *II = dyn_cast<IntrinsicInst>(V)) 7019 setLimitsForIntrinsic(*II, Lower, Upper); 7020 else if (auto *SI = dyn_cast<SelectInst>(V)) 7021 setLimitsForSelectPattern(*SI, Lower, Upper, IIQ); 7022 7023 ConstantRange CR = ConstantRange::getNonEmpty(Lower, Upper); 7024 7025 if (auto *I = dyn_cast<Instruction>(V)) 7026 if (auto *Range = IIQ.getMetadata(I, LLVMContext::MD_range)) 7027 CR = CR.intersectWith(getConstantRangeFromMetadata(*Range)); 7028 7029 if (CtxI && AC) { 7030 // Try to restrict the range based on information from assumptions. 7031 for (auto &AssumeVH : AC->assumptionsFor(V)) { 7032 if (!AssumeVH) 7033 continue; 7034 CallInst *I = cast<CallInst>(AssumeVH); 7035 assert(I->getParent()->getParent() == CtxI->getParent()->getParent() && 7036 "Got assumption for the wrong function!"); 7037 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 7038 "must be an assume intrinsic"); 7039 7040 if (!isValidAssumeForContext(I, CtxI, nullptr)) 7041 continue; 7042 Value *Arg = I->getArgOperand(0); 7043 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); 7044 // Currently we just use information from comparisons. 7045 if (!Cmp || Cmp->getOperand(0) != V) 7046 continue; 7047 ConstantRange RHS = computeConstantRange(Cmp->getOperand(1), UseInstrInfo, 7048 AC, I, Depth + 1); 7049 CR = CR.intersectWith( 7050 ConstantRange::makeSatisfyingICmpRegion(Cmp->getPredicate(), RHS)); 7051 } 7052 } 7053 7054 return CR; 7055 } 7056 7057 static Optional<int64_t> 7058 getOffsetFromIndex(const GEPOperator *GEP, unsigned Idx, const DataLayout &DL) { 7059 // Skip over the first indices. 7060 gep_type_iterator GTI = gep_type_begin(GEP); 7061 for (unsigned i = 1; i != Idx; ++i, ++GTI) 7062 /*skip along*/; 7063 7064 // Compute the offset implied by the rest of the indices. 7065 int64_t Offset = 0; 7066 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 7067 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 7068 if (!OpC) 7069 return None; 7070 if (OpC->isZero()) 7071 continue; // No offset. 7072 7073 // Handle struct indices, which add their field offset to the pointer. 7074 if (StructType *STy = GTI.getStructTypeOrNull()) { 7075 Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 7076 continue; 7077 } 7078 7079 // Otherwise, we have a sequential type like an array or fixed-length 7080 // vector. Multiply the index by the ElementSize. 7081 TypeSize Size = DL.getTypeAllocSize(GTI.getIndexedType()); 7082 if (Size.isScalable()) 7083 return None; 7084 Offset += Size.getFixedSize() * OpC->getSExtValue(); 7085 } 7086 7087 return Offset; 7088 } 7089 7090 Optional<int64_t> llvm::isPointerOffset(const Value *Ptr1, const Value *Ptr2, 7091 const DataLayout &DL) { 7092 Ptr1 = Ptr1->stripPointerCasts(); 7093 Ptr2 = Ptr2->stripPointerCasts(); 7094 7095 // Handle the trivial case first. 7096 if (Ptr1 == Ptr2) { 7097 return 0; 7098 } 7099 7100 const GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1); 7101 const GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2); 7102 7103 // If one pointer is a GEP see if the GEP is a constant offset from the base, 7104 // as in "P" and "gep P, 1". 7105 // Also do this iteratively to handle the the following case: 7106 // Ptr_t1 = GEP Ptr1, c1 7107 // Ptr_t2 = GEP Ptr_t1, c2 7108 // Ptr2 = GEP Ptr_t2, c3 7109 // where we will return c1+c2+c3. 7110 // TODO: Handle the case when both Ptr1 and Ptr2 are GEPs of some common base 7111 // -- replace getOffsetFromBase with getOffsetAndBase, check that the bases 7112 // are the same, and return the difference between offsets. 7113 auto getOffsetFromBase = [&DL](const GEPOperator *GEP, 7114 const Value *Ptr) -> Optional<int64_t> { 7115 const GEPOperator *GEP_T = GEP; 7116 int64_t OffsetVal = 0; 7117 bool HasSameBase = false; 7118 while (GEP_T) { 7119 auto Offset = getOffsetFromIndex(GEP_T, 1, DL); 7120 if (!Offset) 7121 return None; 7122 OffsetVal += *Offset; 7123 auto Op0 = GEP_T->getOperand(0)->stripPointerCasts(); 7124 if (Op0 == Ptr) { 7125 HasSameBase = true; 7126 break; 7127 } 7128 GEP_T = dyn_cast<GEPOperator>(Op0); 7129 } 7130 if (!HasSameBase) 7131 return None; 7132 return OffsetVal; 7133 }; 7134 7135 if (GEP1) { 7136 auto Offset = getOffsetFromBase(GEP1, Ptr2); 7137 if (Offset) 7138 return -*Offset; 7139 } 7140 if (GEP2) { 7141 auto Offset = getOffsetFromBase(GEP2, Ptr1); 7142 if (Offset) 7143 return Offset; 7144 } 7145 7146 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 7147 // base. After that base, they may have some number of common (and 7148 // potentially variable) indices. After that they handle some constant 7149 // offset, which determines their offset from each other. At this point, we 7150 // handle no other case. 7151 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 7152 return None; 7153 7154 // Skip any common indices and track the GEP types. 7155 unsigned Idx = 1; 7156 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 7157 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 7158 break; 7159 7160 auto Offset1 = getOffsetFromIndex(GEP1, Idx, DL); 7161 auto Offset2 = getOffsetFromIndex(GEP2, Idx, DL); 7162 if (!Offset1 || !Offset2) 7163 return None; 7164 return *Offset2 - *Offset1; 7165 } 7166