1//===- ValueTracking.cpp - Walk computations to compute properties --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file contains routines that help analyze properties that chains of 11// computations have. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Analysis/ValueTracking.h" 16#include "llvm/ADT/SmallPtrSet.h" 17#include "llvm/Analysis/AssumptionCache.h" 18#include "llvm/Analysis/InstructionSimplify.h" 19#include "llvm/Analysis/MemoryBuiltins.h" 20#include "llvm/IR/CallSite.h" 21#include "llvm/IR/ConstantRange.h" 22#include "llvm/IR/Constants.h" 23#include "llvm/IR/DataLayout.h" 24#include "llvm/IR/Dominators.h" 25#include "llvm/IR/GetElementPtrTypeIterator.h" 26#include "llvm/IR/GlobalAlias.h" 27#include "llvm/IR/GlobalVariable.h" 28#include "llvm/IR/Instructions.h" 29#include "llvm/IR/IntrinsicInst.h" 30#include "llvm/IR/LLVMContext.h" 31#include "llvm/IR/Metadata.h" 32#include "llvm/IR/Operator.h" 33#include "llvm/IR/PatternMatch.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/MathExtras.h" 36#include <cstring> 37using namespace llvm; 38using namespace llvm::PatternMatch; 39 40const unsigned MaxDepth = 6; 41 42/// Enable an experimental feature to leverage information about dominating 43/// conditions to compute known bits. The individual options below control how 44/// hard we search. The defaults are choosen to be fairly aggressive. If you 45/// run into compile time problems when testing, scale them back and report 46/// your findings. 47static cl::opt<bool> EnableDomConditions("value-tracking-dom-conditions", 48 cl::Hidden, cl::init(false)); 49 50// This is expensive, so we only do it for the top level query value. 51// (TODO: evaluate cost vs profit, consider higher thresholds) 52static cl::opt<unsigned> DomConditionsMaxDepth("dom-conditions-max-depth", 53 cl::Hidden, cl::init(1)); 54 55/// How many dominating blocks should be scanned looking for dominating 56/// conditions? 57static cl::opt<unsigned> DomConditionsMaxDomBlocks("dom-conditions-dom-blocks", 58 cl::Hidden, 59 cl::init(20000)); 60 61// Controls the number of uses of the value searched for possible 62// dominating comparisons. 63static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses", 64 cl::Hidden, cl::init(2000)); 65 66// If true, don't consider only compares whose only use is a branch. 67static cl::opt<bool> DomConditionsSingleCmpUse("dom-conditions-single-cmp-use", 68 cl::Hidden, cl::init(false)); 69 70/// Returns the bitwidth of the given scalar or pointer type (if unknown returns 71/// 0). For vector types, returns the element type's bitwidth. 72static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { 73 if (unsigned BitWidth = Ty->getScalarSizeInBits()) 74 return BitWidth; 75 76 return DL.getPointerTypeSizeInBits(Ty); 77} 78 79// Many of these functions have internal versions that take an assumption 80// exclusion set. This is because of the potential for mutual recursion to 81// cause computeKnownBits to repeatedly visit the same assume intrinsic. The 82// classic case of this is assume(x = y), which will attempt to determine 83// bits in x from bits in y, which will attempt to determine bits in y from 84// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call 85// isKnownNonZero, which calls computeKnownBits and ComputeSignBit and 86// isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on. 87typedef SmallPtrSet<const Value *, 8> ExclInvsSet; 88 89namespace { 90// Simplifying using an assume can only be done in a particular control-flow 91// context (the context instruction provides that context). If an assume and 92// the context instruction are not in the same block then the DT helps in 93// figuring out if we can use it. 94struct Query { 95 ExclInvsSet ExclInvs; 96 AssumptionCache *AC; 97 const Instruction *CxtI; 98 const DominatorTree *DT; 99 100 Query(AssumptionCache *AC = nullptr, const Instruction *CxtI = nullptr, 101 const DominatorTree *DT = nullptr) 102 : AC(AC), CxtI(CxtI), DT(DT) {} 103 104 Query(const Query &Q, const Value *NewExcl) 105 : ExclInvs(Q.ExclInvs), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT) { 106 ExclInvs.insert(NewExcl); 107 } 108}; 109} // end anonymous namespace 110 111// Given the provided Value and, potentially, a context instruction, return 112// the preferred context instruction (if any). 113static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { 114 // If we've been provided with a context instruction, then use that (provided 115 // it has been inserted). 116 if (CxtI && CxtI->getParent()) 117 return CxtI; 118 119 // If the value is really an already-inserted instruction, then use that. 120 CxtI = dyn_cast<Instruction>(V); 121 if (CxtI && CxtI->getParent()) 122 return CxtI; 123 124 return nullptr; 125} 126 127static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne, 128 const DataLayout &DL, unsigned Depth, 129 const Query &Q); 130 131void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne, 132 const DataLayout &DL, unsigned Depth, 133 AssumptionCache *AC, const Instruction *CxtI, 134 const DominatorTree *DT) { 135 ::computeKnownBits(V, KnownZero, KnownOne, DL, Depth, 136 Query(AC, safeCxtI(V, CxtI), DT)); 137} 138 139static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne, 140 const DataLayout &DL, unsigned Depth, 141 const Query &Q); 142 143void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne, 144 const DataLayout &DL, unsigned Depth, 145 AssumptionCache *AC, const Instruction *CxtI, 146 const DominatorTree *DT) { 147 ::ComputeSignBit(V, KnownZero, KnownOne, DL, Depth, 148 Query(AC, safeCxtI(V, CxtI), DT)); 149} 150 151static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth, 152 const Query &Q, const DataLayout &DL); 153 154bool llvm::isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, bool OrZero, 155 unsigned Depth, AssumptionCache *AC, 156 const Instruction *CxtI, 157 const DominatorTree *DT) { 158 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth, 159 Query(AC, safeCxtI(V, CxtI), DT), DL); 160} 161 162static bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth, 163 const Query &Q); 164 165bool llvm::isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth, 166 AssumptionCache *AC, const Instruction *CxtI, 167 const DominatorTree *DT) { 168 return ::isKnownNonZero(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT)); 169} 170 171static bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL, 172 unsigned Depth, const Query &Q); 173 174bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL, 175 unsigned Depth, AssumptionCache *AC, 176 const Instruction *CxtI, const DominatorTree *DT) { 177 return ::MaskedValueIsZero(V, Mask, DL, Depth, 178 Query(AC, safeCxtI(V, CxtI), DT)); 179} 180 181static unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, 182 unsigned Depth, const Query &Q); 183 184unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL, 185 unsigned Depth, AssumptionCache *AC, 186 const Instruction *CxtI, 187 const DominatorTree *DT) { 188 return ::ComputeNumSignBits(V, DL, Depth, Query(AC, safeCxtI(V, CxtI), DT)); 189} 190 191static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW, 192 APInt &KnownZero, APInt &KnownOne, 193 APInt &KnownZero2, APInt &KnownOne2, 194 const DataLayout &DL, unsigned Depth, 195 const Query &Q) { 196 if (!Add) { 197 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) { 198 // We know that the top bits of C-X are clear if X contains less bits 199 // than C (i.e. no wrap-around can happen). For example, 20-X is 200 // positive if we can prove that X is >= 0 and < 16. 201 if (!CLHS->getValue().isNegative()) { 202 unsigned BitWidth = KnownZero.getBitWidth(); 203 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros(); 204 // NLZ can't be BitWidth with no sign bit 205 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1); 206 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q); 207 208 // If all of the MaskV bits are known to be zero, then we know the 209 // output top bits are zero, because we now know that the output is 210 // from [0-C]. 211 if ((KnownZero2 & MaskV) == MaskV) { 212 unsigned NLZ2 = CLHS->getValue().countLeadingZeros(); 213 // Top bits known zero. 214 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2); 215 } 216 } 217 } 218 } 219 220 unsigned BitWidth = KnownZero.getBitWidth(); 221 222 // If an initial sequence of bits in the result is not needed, the 223 // corresponding bits in the operands are not needed. 224 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0); 225 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, DL, Depth + 1, Q); 226 computeKnownBits(Op1, KnownZero2, KnownOne2, DL, Depth + 1, Q); 227 228 // Carry in a 1 for a subtract, rather than a 0. 229 APInt CarryIn(BitWidth, 0); 230 if (!Add) { 231 // Sum = LHS + ~RHS + 1 232 std::swap(KnownZero2, KnownOne2); 233 CarryIn.setBit(0); 234 } 235 236 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn; 237 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn; 238 239 // Compute known bits of the carry. 240 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2); 241 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2; 242 243 // Compute set of known bits (where all three relevant bits are known). 244 APInt LHSKnown = LHSKnownZero | LHSKnownOne; 245 APInt RHSKnown = KnownZero2 | KnownOne2; 246 APInt CarryKnown = CarryKnownZero | CarryKnownOne; 247 APInt Known = LHSKnown & RHSKnown & CarryKnown; 248 249 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) && 250 "known bits of sum differ"); 251 252 // Compute known bits of the result. 253 KnownZero = ~PossibleSumOne & Known; 254 KnownOne = PossibleSumOne & Known; 255 256 // Are we still trying to solve for the sign bit? 257 if (!Known.isNegative()) { 258 if (NSW) { 259 // Adding two non-negative numbers, or subtracting a negative number from 260 // a non-negative one, can't wrap into negative. 261 if (LHSKnownZero.isNegative() && KnownZero2.isNegative()) 262 KnownZero |= APInt::getSignBit(BitWidth); 263 // Adding two negative numbers, or subtracting a non-negative number from 264 // a negative one, can't wrap into non-negative. 265 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative()) 266 KnownOne |= APInt::getSignBit(BitWidth); 267 } 268 } 269} 270 271static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW, 272 APInt &KnownZero, APInt &KnownOne, 273 APInt &KnownZero2, APInt &KnownOne2, 274 const DataLayout &DL, unsigned Depth, 275 const Query &Q) { 276 unsigned BitWidth = KnownZero.getBitWidth(); 277 computeKnownBits(Op1, KnownZero, KnownOne, DL, Depth + 1, Q); 278 computeKnownBits(Op0, KnownZero2, KnownOne2, DL, Depth + 1, Q); 279 280 bool isKnownNegative = false; 281 bool isKnownNonNegative = false; 282 // If the multiplication is known not to overflow, compute the sign bit. 283 if (NSW) { 284 if (Op0 == Op1) { 285 // The product of a number with itself is non-negative. 286 isKnownNonNegative = true; 287 } else { 288 bool isKnownNonNegativeOp1 = KnownZero.isNegative(); 289 bool isKnownNonNegativeOp0 = KnownZero2.isNegative(); 290 bool isKnownNegativeOp1 = KnownOne.isNegative(); 291 bool isKnownNegativeOp0 = KnownOne2.isNegative(); 292 // The product of two numbers with the same sign is non-negative. 293 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || 294 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); 295 // The product of a negative number and a non-negative number is either 296 // negative or zero. 297 if (!isKnownNonNegative) 298 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 && 299 isKnownNonZero(Op0, DL, Depth, Q)) || 300 (isKnownNegativeOp0 && isKnownNonNegativeOp1 && 301 isKnownNonZero(Op1, DL, Depth, Q)); 302 } 303 } 304 305 // If low bits are zero in either operand, output low known-0 bits. 306 // Also compute a conserative estimate for high known-0 bits. 307 // More trickiness is possible, but this is sufficient for the 308 // interesting case of alignment computation. 309 KnownOne.clearAllBits(); 310 unsigned TrailZ = KnownZero.countTrailingOnes() + 311 KnownZero2.countTrailingOnes(); 312 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() + 313 KnownZero2.countLeadingOnes(), 314 BitWidth) - BitWidth; 315 316 TrailZ = std::min(TrailZ, BitWidth); 317 LeadZ = std::min(LeadZ, BitWidth); 318 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) | 319 APInt::getHighBitsSet(BitWidth, LeadZ); 320 321 // Only make use of no-wrap flags if we failed to compute the sign bit 322 // directly. This matters if the multiplication always overflows, in 323 // which case we prefer to follow the result of the direct computation, 324 // though as the program is invoking undefined behaviour we can choose 325 // whatever we like here. 326 if (isKnownNonNegative && !KnownOne.isNegative()) 327 KnownZero.setBit(BitWidth - 1); 328 else if (isKnownNegative && !KnownZero.isNegative()) 329 KnownOne.setBit(BitWidth - 1); 330} 331 332void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, 333 APInt &KnownZero) { 334 unsigned BitWidth = KnownZero.getBitWidth(); 335 unsigned NumRanges = Ranges.getNumOperands() / 2; 336 assert(NumRanges >= 1); 337 338 // Use the high end of the ranges to find leading zeros. 339 unsigned MinLeadingZeros = BitWidth; 340 for (unsigned i = 0; i < NumRanges; ++i) { 341 ConstantInt *Lower = 342 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); 343 ConstantInt *Upper = 344 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); 345 ConstantRange Range(Lower->getValue(), Upper->getValue()); 346 if (Range.isWrappedSet()) 347 MinLeadingZeros = 0; // -1 has no zeros 348 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros(); 349 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros); 350 } 351 352 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros); 353} 354 355static bool isEphemeralValueOf(Instruction *I, const Value *E) { 356 SmallVector<const Value *, 16> WorkSet(1, I); 357 SmallPtrSet<const Value *, 32> Visited; 358 SmallPtrSet<const Value *, 16> EphValues; 359 360 while (!WorkSet.empty()) { 361 const Value *V = WorkSet.pop_back_val(); 362 if (!Visited.insert(V).second) 363 continue; 364 365 // If all uses of this value are ephemeral, then so is this value. 366 bool FoundNEUse = false; 367 for (const User *I : V->users()) 368 if (!EphValues.count(I)) { 369 FoundNEUse = true; 370 break; 371 } 372 373 if (!FoundNEUse) { 374 if (V == E) 375 return true; 376 377 EphValues.insert(V); 378 if (const User *U = dyn_cast<User>(V)) 379 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end(); 380 J != JE; ++J) { 381 if (isSafeToSpeculativelyExecute(*J)) 382 WorkSet.push_back(*J); 383 } 384 } 385 } 386 387 return false; 388} 389 390// Is this an intrinsic that cannot be speculated but also cannot trap? 391static bool isAssumeLikeIntrinsic(const Instruction *I) { 392 if (const CallInst *CI = dyn_cast<CallInst>(I)) 393 if (Function *F = CI->getCalledFunction()) 394 switch (F->getIntrinsicID()) { 395 default: break; 396 // FIXME: This list is repeated from NoTTI::getIntrinsicCost. 397 case Intrinsic::assume: 398 case Intrinsic::dbg_declare: 399 case Intrinsic::dbg_value: 400 case Intrinsic::invariant_start: 401 case Intrinsic::invariant_end: 402 case Intrinsic::lifetime_start: 403 case Intrinsic::lifetime_end: 404 case Intrinsic::objectsize: 405 case Intrinsic::ptr_annotation: 406 case Intrinsic::var_annotation: 407 return true; 408 } 409 410 return false; 411} 412 413static bool isValidAssumeForContext(Value *V, const Query &Q) { 414 Instruction *Inv = cast<Instruction>(V); 415 416 // There are two restrictions on the use of an assume: 417 // 1. The assume must dominate the context (or the control flow must 418 // reach the assume whenever it reaches the context). 419 // 2. The context must not be in the assume's set of ephemeral values 420 // (otherwise we will use the assume to prove that the condition 421 // feeding the assume is trivially true, thus causing the removal of 422 // the assume). 423 424 if (Q.DT) { 425 if (Q.DT->dominates(Inv, Q.CxtI)) { 426 return true; 427 } else if (Inv->getParent() == Q.CxtI->getParent()) { 428 // The context comes first, but they're both in the same block. Make sure 429 // there is nothing in between that might interrupt the control flow. 430 for (BasicBlock::const_iterator I = 431 std::next(BasicBlock::const_iterator(Q.CxtI)), 432 IE(Inv); I != IE; ++I) 433 if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I)) 434 return false; 435 436 return !isEphemeralValueOf(Inv, Q.CxtI); 437 } 438 439 return false; 440 } 441 442 // When we don't have a DT, we do a limited search... 443 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) { 444 return true; 445 } else if (Inv->getParent() == Q.CxtI->getParent()) { 446 // Search forward from the assume until we reach the context (or the end 447 // of the block); the common case is that the assume will come first. 448 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)), 449 IE = Inv->getParent()->end(); I != IE; ++I) 450 if (I == Q.CxtI) 451 return true; 452 453 // The context must come first... 454 for (BasicBlock::const_iterator I = 455 std::next(BasicBlock::const_iterator(Q.CxtI)), 456 IE(Inv); I != IE; ++I) 457 if (!isSafeToSpeculativelyExecute(I) && !isAssumeLikeIntrinsic(I)) 458 return false; 459 460 return !isEphemeralValueOf(Inv, Q.CxtI); 461 } 462 463 return false; 464} 465 466bool llvm::isValidAssumeForContext(const Instruction *I, 467 const Instruction *CxtI, 468 const DominatorTree *DT) { 469 return ::isValidAssumeForContext(const_cast<Instruction *>(I), 470 Query(nullptr, CxtI, DT)); 471} 472 473template<typename LHS, typename RHS> 474inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>, 475 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>> 476m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { 477 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L)); 478} 479 480template<typename LHS, typename RHS> 481inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>, 482 BinaryOp_match<RHS, LHS, Instruction::And>> 483m_c_And(const LHS &L, const RHS &R) { 484 return m_CombineOr(m_And(L, R), m_And(R, L)); 485} 486 487template<typename LHS, typename RHS> 488inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>, 489 BinaryOp_match<RHS, LHS, Instruction::Or>> 490m_c_Or(const LHS &L, const RHS &R) { 491 return m_CombineOr(m_Or(L, R), m_Or(R, L)); 492} 493 494template<typename LHS, typename RHS> 495inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>, 496 BinaryOp_match<RHS, LHS, Instruction::Xor>> 497m_c_Xor(const LHS &L, const RHS &R) { 498 return m_CombineOr(m_Xor(L, R), m_Xor(R, L)); 499} 500 501/// Compute known bits in 'V' under the assumption that the condition 'Cmp' is 502/// true (at the context instruction.) This is mostly a utility function for 503/// the prototype dominating conditions reasoning below. 504static void computeKnownBitsFromTrueCondition(Value *V, ICmpInst *Cmp, 505 APInt &KnownZero, 506 APInt &KnownOne, 507 const DataLayout &DL, 508 unsigned Depth, const Query &Q) { 509 Value *LHS = Cmp->getOperand(0); 510 Value *RHS = Cmp->getOperand(1); 511 // TODO: We could potentially be more aggressive here. This would be worth 512 // evaluating. If we can, explore commoning this code with the assume 513 // handling logic. 514 if (LHS != V && RHS != V) 515 return; 516 517 const unsigned BitWidth = KnownZero.getBitWidth(); 518 519 switch (Cmp->getPredicate()) { 520 default: 521 // We know nothing from this condition 522 break; 523 // TODO: implement unsigned bound from below (known one bits) 524 // TODO: common condition check implementations with assumes 525 // TODO: implement other patterns from assume (e.g. V & B == A) 526 case ICmpInst::ICMP_SGT: 527 if (LHS == V) { 528 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0); 529 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q); 530 if (KnownOneTemp.isAllOnesValue() || KnownZeroTemp.isNegative()) { 531 // We know that the sign bit is zero. 532 KnownZero |= APInt::getSignBit(BitWidth); 533 } 534 } 535 break; 536 case ICmpInst::ICMP_EQ: 537 if (LHS == V) 538 computeKnownBits(RHS, KnownZero, KnownOne, DL, Depth + 1, Q); 539 else if (RHS == V) 540 computeKnownBits(LHS, KnownZero, KnownOne, DL, Depth + 1, Q); 541 else 542 llvm_unreachable("missing use?"); 543 break; 544 case ICmpInst::ICMP_ULE: 545 if (LHS == V) { 546 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0); 547 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q); 548 // The known zero bits carry over 549 unsigned SignBits = KnownZeroTemp.countLeadingOnes(); 550 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits); 551 } 552 break; 553 case ICmpInst::ICMP_ULT: 554 if (LHS == V) { 555 APInt KnownZeroTemp(BitWidth, 0), KnownOneTemp(BitWidth, 0); 556 computeKnownBits(RHS, KnownZeroTemp, KnownOneTemp, DL, Depth + 1, Q); 557 // Whatever high bits in rhs are zero are known to be zero (if rhs is a 558 // power of 2, then one more). 559 unsigned SignBits = KnownZeroTemp.countLeadingOnes(); 560 if (isKnownToBeAPowerOfTwo(RHS, false, Depth + 1, Query(Q, Cmp), DL)) 561 SignBits++; 562 KnownZero |= APInt::getHighBitsSet(BitWidth, SignBits); 563 } 564 break; 565 }; 566} 567 568/// Compute known bits in 'V' from conditions which are known to be true along 569/// all paths leading to the context instruction. In particular, look for 570/// cases where one branch of an interesting condition dominates the context 571/// instruction. This does not do general dataflow. 572/// NOTE: This code is EXPERIMENTAL and currently off by default. 573static void computeKnownBitsFromDominatingCondition(Value *V, APInt &KnownZero, 574 APInt &KnownOne, 575 const DataLayout &DL, 576 unsigned Depth, 577 const Query &Q) { 578 // Need both the dominator tree and the query location to do anything useful 579 if (!Q.DT || !Q.CxtI) 580 return; 581 Instruction *Cxt = const_cast<Instruction *>(Q.CxtI); 582 583 // Avoid useless work 584 if (auto VI = dyn_cast<Instruction>(V)) 585 if (VI->getParent() == Cxt->getParent()) 586 return; 587 588 // Note: We currently implement two options. It's not clear which of these 589 // will survive long term, we need data for that. 590 // Option 1 - Try walking the dominator tree looking for conditions which 591 // might apply. This works well for local conditions (loop guards, etc..), 592 // but not as well for things far from the context instruction (presuming a 593 // low max blocks explored). If we can set an high enough limit, this would 594 // be all we need. 595 // Option 2 - We restrict out search to those conditions which are uses of 596 // the value we're interested in. This is independent of dom structure, 597 // but is slightly less powerful without looking through lots of use chains. 598 // It does handle conditions far from the context instruction (e.g. early 599 // function exits on entry) really well though. 600 601 // Option 1 - Search the dom tree 602 unsigned NumBlocksExplored = 0; 603 BasicBlock *Current = Cxt->getParent(); 604 while (true) { 605 // Stop searching if we've gone too far up the chain 606 if (NumBlocksExplored >= DomConditionsMaxDomBlocks) 607 break; 608 NumBlocksExplored++; 609 610 if (!Q.DT->getNode(Current)->getIDom()) 611 break; 612 Current = Q.DT->getNode(Current)->getIDom()->getBlock(); 613 if (!Current) 614 // found function entry 615 break; 616 617 BranchInst *BI = dyn_cast<BranchInst>(Current->getTerminator()); 618 if (!BI || BI->isUnconditional()) 619 continue; 620 ICmpInst *Cmp = dyn_cast<ICmpInst>(BI->getCondition()); 621 if (!Cmp) 622 continue; 623 624 // We're looking for conditions that are guaranteed to hold at the context 625 // instruction. Finding a condition where one path dominates the context 626 // isn't enough because both the true and false cases could merge before 627 // the context instruction we're actually interested in. Instead, we need 628 // to ensure that the taken *edge* dominates the context instruction. 629 BasicBlock *BB0 = BI->getSuccessor(0); 630 BasicBlockEdge Edge(BI->getParent(), BB0); 631 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent())) 632 continue; 633 634 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth, 635 Q); 636 } 637 638 // Option 2 - Search the other uses of V 639 unsigned NumUsesExplored = 0; 640 for (auto U : V->users()) { 641 // Avoid massive lists 642 if (NumUsesExplored >= DomConditionsMaxUses) 643 break; 644 NumUsesExplored++; 645 // Consider only compare instructions uniquely controlling a branch 646 ICmpInst *Cmp = dyn_cast<ICmpInst>(U); 647 if (!Cmp) 648 continue; 649 650 if (DomConditionsSingleCmpUse && !Cmp->hasOneUse()) 651 continue; 652 653 for (auto *CmpU : Cmp->users()) { 654 BranchInst *BI = dyn_cast<BranchInst>(CmpU); 655 if (!BI || BI->isUnconditional()) 656 continue; 657 // We're looking for conditions that are guaranteed to hold at the 658 // context instruction. Finding a condition where one path dominates 659 // the context isn't enough because both the true and false cases could 660 // merge before the context instruction we're actually interested in. 661 // Instead, we need to ensure that the taken *edge* dominates the context 662 // instruction. 663 BasicBlock *BB0 = BI->getSuccessor(0); 664 BasicBlockEdge Edge(BI->getParent(), BB0); 665 if (!Edge.isSingleEdge() || !Q.DT->dominates(Edge, Q.CxtI->getParent())) 666 continue; 667 668 computeKnownBitsFromTrueCondition(V, Cmp, KnownZero, KnownOne, DL, Depth, 669 Q); 670 } 671 } 672} 673 674static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero, 675 APInt &KnownOne, const DataLayout &DL, 676 unsigned Depth, const Query &Q) { 677 // Use of assumptions is context-sensitive. If we don't have a context, we 678 // cannot use them! 679 if (!Q.AC || !Q.CxtI) 680 return; 681 682 unsigned BitWidth = KnownZero.getBitWidth(); 683 684 for (auto &AssumeVH : Q.AC->assumptions()) { 685 if (!AssumeVH) 686 continue; 687 CallInst *I = cast<CallInst>(AssumeVH); 688 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && 689 "Got assumption for the wrong function!"); 690 if (Q.ExclInvs.count(I)) 691 continue; 692 693 // Warning: This loop can end up being somewhat performance sensetive. 694 // We're running this loop for once for each value queried resulting in a 695 // runtime of ~O(#assumes * #values). 696 697 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 698 "must be an assume intrinsic"); 699 700 Value *Arg = I->getArgOperand(0); 701 702 if (Arg == V && isValidAssumeForContext(I, Q)) { 703 assert(BitWidth == 1 && "assume operand is not i1?"); 704 KnownZero.clearAllBits(); 705 KnownOne.setAllBits(); 706 return; 707 } 708 709 // The remaining tests are all recursive, so bail out if we hit the limit. 710 if (Depth == MaxDepth) 711 continue; 712 713 Value *A, *B; 714 auto m_V = m_CombineOr(m_Specific(V), 715 m_CombineOr(m_PtrToInt(m_Specific(V)), 716 m_BitCast(m_Specific(V)))); 717 718 CmpInst::Predicate Pred; 719 ConstantInt *C; 720 // assume(v = a) 721 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) && 722 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 723 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 724 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 725 KnownZero |= RHSKnownZero; 726 KnownOne |= RHSKnownOne; 727 // assume(v & b = a) 728 } else if (match(Arg, 729 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) && 730 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 731 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 732 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 733 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0); 734 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I)); 735 736 // For those bits in the mask that are known to be one, we can propagate 737 // known bits from the RHS to V. 738 KnownZero |= RHSKnownZero & MaskKnownOne; 739 KnownOne |= RHSKnownOne & MaskKnownOne; 740 // assume(~(v & b) = a) 741 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))), 742 m_Value(A))) && 743 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 744 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 745 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 746 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0); 747 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I)); 748 749 // For those bits in the mask that are known to be one, we can propagate 750 // inverted known bits from the RHS to V. 751 KnownZero |= RHSKnownOne & MaskKnownOne; 752 KnownOne |= RHSKnownZero & MaskKnownOne; 753 // assume(v | b = a) 754 } else if (match(Arg, 755 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) && 756 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 757 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 758 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 759 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 760 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I)); 761 762 // For those bits in B that are known to be zero, we can propagate known 763 // bits from the RHS to V. 764 KnownZero |= RHSKnownZero & BKnownZero; 765 KnownOne |= RHSKnownOne & BKnownZero; 766 // assume(~(v | b) = a) 767 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))), 768 m_Value(A))) && 769 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 770 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 771 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 772 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 773 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I)); 774 775 // For those bits in B that are known to be zero, we can propagate 776 // inverted known bits from the RHS to V. 777 KnownZero |= RHSKnownOne & BKnownZero; 778 KnownOne |= RHSKnownZero & BKnownZero; 779 // assume(v ^ b = a) 780 } else if (match(Arg, 781 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) && 782 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 783 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 784 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 785 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 786 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I)); 787 788 // For those bits in B that are known to be zero, we can propagate known 789 // bits from the RHS to V. For those bits in B that are known to be one, 790 // we can propagate inverted known bits from the RHS to V. 791 KnownZero |= RHSKnownZero & BKnownZero; 792 KnownOne |= RHSKnownOne & BKnownZero; 793 KnownZero |= RHSKnownOne & BKnownOne; 794 KnownOne |= RHSKnownZero & BKnownOne; 795 // assume(~(v ^ b) = a) 796 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))), 797 m_Value(A))) && 798 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 799 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 800 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 801 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 802 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I)); 803 804 // For those bits in B that are known to be zero, we can propagate 805 // inverted known bits from the RHS to V. For those bits in B that are 806 // known to be one, we can propagate known bits from the RHS to V. 807 KnownZero |= RHSKnownOne & BKnownZero; 808 KnownOne |= RHSKnownZero & BKnownZero; 809 KnownZero |= RHSKnownZero & BKnownOne; 810 KnownOne |= RHSKnownOne & BKnownOne; 811 // assume(v << c = a) 812 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)), 813 m_Value(A))) && 814 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 815 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 816 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 817 // For those bits in RHS that are known, we can propagate them to known 818 // bits in V shifted to the right by C. 819 KnownZero |= RHSKnownZero.lshr(C->getZExtValue()); 820 KnownOne |= RHSKnownOne.lshr(C->getZExtValue()); 821 // assume(~(v << c) = a) 822 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))), 823 m_Value(A))) && 824 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 825 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 826 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 827 // For those bits in RHS that are known, we can propagate them inverted 828 // to known bits in V shifted to the right by C. 829 KnownZero |= RHSKnownOne.lshr(C->getZExtValue()); 830 KnownOne |= RHSKnownZero.lshr(C->getZExtValue()); 831 // assume(v >> c = a) 832 } else if (match(Arg, 833 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)), 834 m_AShr(m_V, m_ConstantInt(C))), 835 m_Value(A))) && 836 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 837 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 838 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 839 // For those bits in RHS that are known, we can propagate them to known 840 // bits in V shifted to the right by C. 841 KnownZero |= RHSKnownZero << C->getZExtValue(); 842 KnownOne |= RHSKnownOne << C->getZExtValue(); 843 // assume(~(v >> c) = a) 844 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr( 845 m_LShr(m_V, m_ConstantInt(C)), 846 m_AShr(m_V, m_ConstantInt(C)))), 847 m_Value(A))) && 848 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q)) { 849 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 850 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 851 // For those bits in RHS that are known, we can propagate them inverted 852 // to known bits in V shifted to the right by C. 853 KnownZero |= RHSKnownOne << C->getZExtValue(); 854 KnownOne |= RHSKnownZero << C->getZExtValue(); 855 // assume(v >=_s c) where c is non-negative 856 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 857 Pred == ICmpInst::ICMP_SGE && isValidAssumeForContext(I, Q)) { 858 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 859 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 860 861 if (RHSKnownZero.isNegative()) { 862 // We know that the sign bit is zero. 863 KnownZero |= APInt::getSignBit(BitWidth); 864 } 865 // assume(v >_s c) where c is at least -1. 866 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 867 Pred == ICmpInst::ICMP_SGT && isValidAssumeForContext(I, Q)) { 868 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 869 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 870 871 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) { 872 // We know that the sign bit is zero. 873 KnownZero |= APInt::getSignBit(BitWidth); 874 } 875 // assume(v <=_s c) where c is negative 876 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 877 Pred == ICmpInst::ICMP_SLE && isValidAssumeForContext(I, Q)) { 878 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 879 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 880 881 if (RHSKnownOne.isNegative()) { 882 // We know that the sign bit is one. 883 KnownOne |= APInt::getSignBit(BitWidth); 884 } 885 // assume(v <_s c) where c is non-positive 886 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 887 Pred == ICmpInst::ICMP_SLT && isValidAssumeForContext(I, Q)) { 888 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 889 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 890 891 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) { 892 // We know that the sign bit is one. 893 KnownOne |= APInt::getSignBit(BitWidth); 894 } 895 // assume(v <=_u c) 896 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 897 Pred == ICmpInst::ICMP_ULE && isValidAssumeForContext(I, Q)) { 898 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 899 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 900 901 // Whatever high bits in c are zero are known to be zero. 902 KnownZero |= 903 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()); 904 // assume(v <_u c) 905 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 906 Pred == ICmpInst::ICMP_ULT && isValidAssumeForContext(I, Q)) { 907 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 908 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I)); 909 910 // Whatever high bits in c are zero are known to be zero (if c is a power 911 // of 2, then one more). 912 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I), DL)) 913 KnownZero |= 914 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1); 915 else 916 KnownZero |= 917 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()); 918 } 919 } 920} 921 922/// Determine which bits of V are known to be either zero or one and return 923/// them in the KnownZero/KnownOne bit sets. 924/// 925/// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that 926/// we cannot optimize based on the assumption that it is zero without changing 927/// it to be an explicit zero. If we don't change it to zero, other code could 928/// optimized based on the contradictory assumption that it is non-zero. 929/// Because instcombine aggressively folds operations with undef args anyway, 930/// this won't lose us code quality. 931/// 932/// This function is defined on values with integer type, values with pointer 933/// type, and vectors of integers. In the case 934/// where V is a vector, known zero, and known one values are the 935/// same width as the vector element, and the bit is set only if it is true 936/// for all of the elements in the vector. 937void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne, 938 const DataLayout &DL, unsigned Depth, const Query &Q) { 939 assert(V && "No Value?"); 940 assert(Depth <= MaxDepth && "Limit Search Depth"); 941 unsigned BitWidth = KnownZero.getBitWidth(); 942 943 assert((V->getType()->isIntOrIntVectorTy() || 944 V->getType()->getScalarType()->isPointerTy()) && 945 "Not integer or pointer type!"); 946 assert((DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) && 947 (!V->getType()->isIntOrIntVectorTy() || 948 V->getType()->getScalarSizeInBits() == BitWidth) && 949 KnownZero.getBitWidth() == BitWidth && 950 KnownOne.getBitWidth() == BitWidth && 951 "V, KnownOne and KnownZero should have same BitWidth"); 952 953 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 954 // We know all of the bits for a constant! 955 KnownOne = CI->getValue(); 956 KnownZero = ~KnownOne; 957 return; 958 } 959 // Null and aggregate-zero are all-zeros. 960 if (isa<ConstantPointerNull>(V) || 961 isa<ConstantAggregateZero>(V)) { 962 KnownOne.clearAllBits(); 963 KnownZero = APInt::getAllOnesValue(BitWidth); 964 return; 965 } 966 // Handle a constant vector by taking the intersection of the known bits of 967 // each element. There is no real need to handle ConstantVector here, because 968 // we don't handle undef in any particularly useful way. 969 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) { 970 // We know that CDS must be a vector of integers. Take the intersection of 971 // each element. 972 KnownZero.setAllBits(); KnownOne.setAllBits(); 973 APInt Elt(KnownZero.getBitWidth(), 0); 974 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) { 975 Elt = CDS->getElementAsInteger(i); 976 KnownZero &= ~Elt; 977 KnownOne &= Elt; 978 } 979 return; 980 } 981 982 // The address of an aligned GlobalValue has trailing zeros. 983 if (auto *GO = dyn_cast<GlobalObject>(V)) { 984 unsigned Align = GO->getAlignment(); 985 if (Align == 0) { 986 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) { 987 Type *ObjectType = GVar->getType()->getElementType(); 988 if (ObjectType->isSized()) { 989 // If the object is defined in the current Module, we'll be giving 990 // it the preferred alignment. Otherwise, we have to assume that it 991 // may only have the minimum ABI alignment. 992 if (!GVar->isDeclaration() && !GVar->isWeakForLinker()) 993 Align = DL.getPreferredAlignment(GVar); 994 else 995 Align = DL.getABITypeAlignment(ObjectType); 996 } 997 } 998 } 999 if (Align > 0) 1000 KnownZero = APInt::getLowBitsSet(BitWidth, 1001 countTrailingZeros(Align)); 1002 else 1003 KnownZero.clearAllBits(); 1004 KnownOne.clearAllBits(); 1005 return; 1006 } 1007 1008 if (Argument *A = dyn_cast<Argument>(V)) { 1009 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0; 1010 1011 if (!Align && A->hasStructRetAttr()) { 1012 // An sret parameter has at least the ABI alignment of the return type. 1013 Type *EltTy = cast<PointerType>(A->getType())->getElementType(); 1014 if (EltTy->isSized()) 1015 Align = DL.getABITypeAlignment(EltTy); 1016 } 1017 1018 if (Align) 1019 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align)); 1020 else 1021 KnownZero.clearAllBits(); 1022 KnownOne.clearAllBits(); 1023 1024 // Don't give up yet... there might be an assumption that provides more 1025 // information... 1026 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q); 1027 1028 // Or a dominating condition for that matter 1029 if (EnableDomConditions && Depth <= DomConditionsMaxDepth) 1030 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, 1031 Depth, Q); 1032 return; 1033 } 1034 1035 // Start out not knowing anything. 1036 KnownZero.clearAllBits(); KnownOne.clearAllBits(); 1037 1038 // Limit search depth. 1039 // All recursive calls that increase depth must come after this. 1040 if (Depth == MaxDepth) 1041 return; 1042 1043 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has 1044 // the bits of its aliasee. 1045 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 1046 if (!GA->mayBeOverridden()) 1047 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, DL, Depth + 1, Q); 1048 return; 1049 } 1050 1051 // Check whether a nearby assume intrinsic can determine some known bits. 1052 computeKnownBitsFromAssume(V, KnownZero, KnownOne, DL, Depth, Q); 1053 1054 // Check whether there's a dominating condition which implies something about 1055 // this value at the given context. 1056 if (EnableDomConditions && Depth <= DomConditionsMaxDepth) 1057 computeKnownBitsFromDominatingCondition(V, KnownZero, KnownOne, DL, Depth, 1058 Q); 1059 1060 Operator *I = dyn_cast<Operator>(V); 1061 if (!I) return; 1062 1063 APInt KnownZero2(KnownZero), KnownOne2(KnownOne); 1064 switch (I->getOpcode()) { 1065 default: break; 1066 case Instruction::Load: 1067 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range)) 1068 computeKnownBitsFromRangeMetadata(*MD, KnownZero); 1069 break; 1070 case Instruction::And: { 1071 // If either the LHS or the RHS are Zero, the result is zero. 1072 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q); 1073 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q); 1074 1075 // Output known-1 bits are only known if set in both the LHS & RHS. 1076 KnownOne &= KnownOne2; 1077 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1078 KnownZero |= KnownZero2; 1079 break; 1080 } 1081 case Instruction::Or: { 1082 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q); 1083 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q); 1084 1085 // Output known-0 bits are only known if clear in both the LHS & RHS. 1086 KnownZero &= KnownZero2; 1087 // Output known-1 are known to be set if set in either the LHS | RHS. 1088 KnownOne |= KnownOne2; 1089 break; 1090 } 1091 case Instruction::Xor: { 1092 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q); 1093 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q); 1094 1095 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1096 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1097 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1098 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1099 KnownZero = KnownZeroOut; 1100 break; 1101 } 1102 case Instruction::Mul: { 1103 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 1104 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero, 1105 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q); 1106 break; 1107 } 1108 case Instruction::UDiv: { 1109 // For the purposes of computing leading zeros we can conservatively 1110 // treat a udiv as a logical right shift by the power of 2 known to 1111 // be less than the denominator. 1112 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q); 1113 unsigned LeadZ = KnownZero2.countLeadingOnes(); 1114 1115 KnownOne2.clearAllBits(); 1116 KnownZero2.clearAllBits(); 1117 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q); 1118 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros(); 1119 if (RHSUnknownLeadingOnes != BitWidth) 1120 LeadZ = std::min(BitWidth, 1121 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1); 1122 1123 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ); 1124 break; 1125 } 1126 case Instruction::Select: 1127 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, DL, Depth + 1, Q); 1128 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q); 1129 1130 // Only known if known in both the LHS and RHS. 1131 KnownOne &= KnownOne2; 1132 KnownZero &= KnownZero2; 1133 break; 1134 case Instruction::FPTrunc: 1135 case Instruction::FPExt: 1136 case Instruction::FPToUI: 1137 case Instruction::FPToSI: 1138 case Instruction::SIToFP: 1139 case Instruction::UIToFP: 1140 break; // Can't work with floating point. 1141 case Instruction::PtrToInt: 1142 case Instruction::IntToPtr: 1143 case Instruction::AddrSpaceCast: // Pointers could be different sizes. 1144 // FALL THROUGH and handle them the same as zext/trunc. 1145 case Instruction::ZExt: 1146 case Instruction::Trunc: { 1147 Type *SrcTy = I->getOperand(0)->getType(); 1148 1149 unsigned SrcBitWidth; 1150 // Note that we handle pointer operands here because of inttoptr/ptrtoint 1151 // which fall through here. 1152 SrcBitWidth = DL.getTypeSizeInBits(SrcTy->getScalarType()); 1153 1154 assert(SrcBitWidth && "SrcBitWidth can't be zero"); 1155 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth); 1156 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth); 1157 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q); 1158 KnownZero = KnownZero.zextOrTrunc(BitWidth); 1159 KnownOne = KnownOne.zextOrTrunc(BitWidth); 1160 // Any top bits are known to be zero. 1161 if (BitWidth > SrcBitWidth) 1162 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); 1163 break; 1164 } 1165 case Instruction::BitCast: { 1166 Type *SrcTy = I->getOperand(0)->getType(); 1167 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 1168 // TODO: For now, not handling conversions like: 1169 // (bitcast i64 %x to <2 x i32>) 1170 !I->getType()->isVectorTy()) { 1171 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q); 1172 break; 1173 } 1174 break; 1175 } 1176 case Instruction::SExt: { 1177 // Compute the bits in the result that are not present in the input. 1178 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); 1179 1180 KnownZero = KnownZero.trunc(SrcBitWidth); 1181 KnownOne = KnownOne.trunc(SrcBitWidth); 1182 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q); 1183 KnownZero = KnownZero.zext(BitWidth); 1184 KnownOne = KnownOne.zext(BitWidth); 1185 1186 // If the sign bit of the input is known set or clear, then we know the 1187 // top bits of the result. 1188 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero 1189 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); 1190 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set 1191 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); 1192 break; 1193 } 1194 case Instruction::Shl: 1195 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0 1196 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) { 1197 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth); 1198 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q); 1199 KnownZero <<= ShiftAmt; 1200 KnownOne <<= ShiftAmt; 1201 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0 1202 } 1203 break; 1204 case Instruction::LShr: 1205 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0 1206 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) { 1207 // Compute the new bits that are at the top now. 1208 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth); 1209 1210 // Unsigned shift right. 1211 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q); 1212 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt); 1213 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt); 1214 // high bits known zero. 1215 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt); 1216 } 1217 break; 1218 case Instruction::AShr: 1219 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0 1220 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) { 1221 // Compute the new bits that are at the top now. 1222 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1); 1223 1224 // Signed shift right. 1225 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q); 1226 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt); 1227 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt); 1228 1229 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt)); 1230 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero. 1231 KnownZero |= HighBits; 1232 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one. 1233 KnownOne |= HighBits; 1234 } 1235 break; 1236 case Instruction::Sub: { 1237 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 1238 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW, 1239 KnownZero, KnownOne, KnownZero2, KnownOne2, DL, 1240 Depth, Q); 1241 break; 1242 } 1243 case Instruction::Add: { 1244 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 1245 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW, 1246 KnownZero, KnownOne, KnownZero2, KnownOne2, DL, 1247 Depth, Q); 1248 break; 1249 } 1250 case Instruction::SRem: 1251 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) { 1252 APInt RA = Rem->getValue().abs(); 1253 if (RA.isPowerOf2()) { 1254 APInt LowBits = RA - 1; 1255 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, 1256 Q); 1257 1258 // The low bits of the first operand are unchanged by the srem. 1259 KnownZero = KnownZero2 & LowBits; 1260 KnownOne = KnownOne2 & LowBits; 1261 1262 // If the first operand is non-negative or has all low bits zero, then 1263 // the upper bits are all zero. 1264 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits)) 1265 KnownZero |= ~LowBits; 1266 1267 // If the first operand is negative and not all low bits are zero, then 1268 // the upper bits are all one. 1269 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0)) 1270 KnownOne |= ~LowBits; 1271 1272 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1273 } 1274 } 1275 1276 // The sign bit is the LHS's sign bit, except when the result of the 1277 // remainder is zero. 1278 if (KnownZero.isNonNegative()) { 1279 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0); 1280 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, DL, 1281 Depth + 1, Q); 1282 // If it's known zero, our sign bit is also zero. 1283 if (LHSKnownZero.isNegative()) 1284 KnownZero.setBit(BitWidth - 1); 1285 } 1286 1287 break; 1288 case Instruction::URem: { 1289 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) { 1290 APInt RA = Rem->getValue(); 1291 if (RA.isPowerOf2()) { 1292 APInt LowBits = (RA - 1); 1293 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, 1294 Q); 1295 KnownZero |= ~LowBits; 1296 KnownOne &= LowBits; 1297 break; 1298 } 1299 } 1300 1301 // Since the result is less than or equal to either operand, any leading 1302 // zero bits in either operand must also exist in the result. 1303 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, Q); 1304 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, DL, Depth + 1, Q); 1305 1306 unsigned Leaders = std::max(KnownZero.countLeadingOnes(), 1307 KnownZero2.countLeadingOnes()); 1308 KnownOne.clearAllBits(); 1309 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders); 1310 break; 1311 } 1312 1313 case Instruction::Alloca: { 1314 AllocaInst *AI = cast<AllocaInst>(V); 1315 unsigned Align = AI->getAlignment(); 1316 if (Align == 0) 1317 Align = DL.getABITypeAlignment(AI->getType()->getElementType()); 1318 1319 if (Align > 0) 1320 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align)); 1321 break; 1322 } 1323 case Instruction::GetElementPtr: { 1324 // Analyze all of the subscripts of this getelementptr instruction 1325 // to determine if we can prove known low zero bits. 1326 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0); 1327 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, DL, 1328 Depth + 1, Q); 1329 unsigned TrailZ = LocalKnownZero.countTrailingOnes(); 1330 1331 gep_type_iterator GTI = gep_type_begin(I); 1332 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { 1333 Value *Index = I->getOperand(i); 1334 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 1335 // Handle struct member offset arithmetic. 1336 1337 // Handle case when index is vector zeroinitializer 1338 Constant *CIndex = cast<Constant>(Index); 1339 if (CIndex->isZeroValue()) 1340 continue; 1341 1342 if (CIndex->getType()->isVectorTy()) 1343 Index = CIndex->getSplatValue(); 1344 1345 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue(); 1346 const StructLayout *SL = DL.getStructLayout(STy); 1347 uint64_t Offset = SL->getElementOffset(Idx); 1348 TrailZ = std::min<unsigned>(TrailZ, 1349 countTrailingZeros(Offset)); 1350 } else { 1351 // Handle array index arithmetic. 1352 Type *IndexedTy = GTI.getIndexedType(); 1353 if (!IndexedTy->isSized()) { 1354 TrailZ = 0; 1355 break; 1356 } 1357 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits(); 1358 uint64_t TypeSize = DL.getTypeAllocSize(IndexedTy); 1359 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0); 1360 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, DL, Depth + 1, 1361 Q); 1362 TrailZ = std::min(TrailZ, 1363 unsigned(countTrailingZeros(TypeSize) + 1364 LocalKnownZero.countTrailingOnes())); 1365 } 1366 } 1367 1368 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ); 1369 break; 1370 } 1371 case Instruction::PHI: { 1372 PHINode *P = cast<PHINode>(I); 1373 // Handle the case of a simple two-predecessor recurrence PHI. 1374 // There's a lot more that could theoretically be done here, but 1375 // this is sufficient to catch some interesting cases. 1376 if (P->getNumIncomingValues() == 2) { 1377 for (unsigned i = 0; i != 2; ++i) { 1378 Value *L = P->getIncomingValue(i); 1379 Value *R = P->getIncomingValue(!i); 1380 Operator *LU = dyn_cast<Operator>(L); 1381 if (!LU) 1382 continue; 1383 unsigned Opcode = LU->getOpcode(); 1384 // Check for operations that have the property that if 1385 // both their operands have low zero bits, the result 1386 // will have low zero bits. 1387 if (Opcode == Instruction::Add || 1388 Opcode == Instruction::Sub || 1389 Opcode == Instruction::And || 1390 Opcode == Instruction::Or || 1391 Opcode == Instruction::Mul) { 1392 Value *LL = LU->getOperand(0); 1393 Value *LR = LU->getOperand(1); 1394 // Find a recurrence. 1395 if (LL == I) 1396 L = LR; 1397 else if (LR == I) 1398 L = LL; 1399 else 1400 break; 1401 // Ok, we have a PHI of the form L op= R. Check for low 1402 // zero bits. 1403 computeKnownBits(R, KnownZero2, KnownOne2, DL, Depth + 1, Q); 1404 1405 // We need to take the minimum number of known bits 1406 APInt KnownZero3(KnownZero), KnownOne3(KnownOne); 1407 computeKnownBits(L, KnownZero3, KnownOne3, DL, Depth + 1, Q); 1408 1409 KnownZero = APInt::getLowBitsSet(BitWidth, 1410 std::min(KnownZero2.countTrailingOnes(), 1411 KnownZero3.countTrailingOnes())); 1412 break; 1413 } 1414 } 1415 } 1416 1417 // Unreachable blocks may have zero-operand PHI nodes. 1418 if (P->getNumIncomingValues() == 0) 1419 break; 1420 1421 // Otherwise take the unions of the known bit sets of the operands, 1422 // taking conservative care to avoid excessive recursion. 1423 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) { 1424 // Skip if every incoming value references to ourself. 1425 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue())) 1426 break; 1427 1428 KnownZero = APInt::getAllOnesValue(BitWidth); 1429 KnownOne = APInt::getAllOnesValue(BitWidth); 1430 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) { 1431 // Skip direct self references. 1432 if (P->getIncomingValue(i) == P) continue; 1433 1434 KnownZero2 = APInt(BitWidth, 0); 1435 KnownOne2 = APInt(BitWidth, 0); 1436 // Recurse, but cap the recursion to one level, because we don't 1437 // want to waste time spinning around in loops. 1438 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, DL, 1439 MaxDepth - 1, Q); 1440 KnownZero &= KnownZero2; 1441 KnownOne &= KnownOne2; 1442 // If all bits have been ruled out, there's no need to check 1443 // more operands. 1444 if (!KnownZero && !KnownOne) 1445 break; 1446 } 1447 } 1448 break; 1449 } 1450 case Instruction::Call: 1451 case Instruction::Invoke: 1452 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range)) 1453 computeKnownBitsFromRangeMetadata(*MD, KnownZero); 1454 // If a range metadata is attached to this IntrinsicInst, intersect the 1455 // explicit range specified by the metadata and the implicit range of 1456 // the intrinsic. 1457 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1458 switch (II->getIntrinsicID()) { 1459 default: break; 1460 case Intrinsic::ctlz: 1461 case Intrinsic::cttz: { 1462 unsigned LowBits = Log2_32(BitWidth)+1; 1463 // If this call is undefined for 0, the result will be less than 2^n. 1464 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) 1465 LowBits -= 1; 1466 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits); 1467 break; 1468 } 1469 case Intrinsic::ctpop: { 1470 unsigned LowBits = Log2_32(BitWidth)+1; 1471 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits); 1472 break; 1473 } 1474 case Intrinsic::x86_sse42_crc32_64_64: 1475 KnownZero |= APInt::getHighBitsSet(64, 32); 1476 break; 1477 } 1478 } 1479 break; 1480 case Instruction::ExtractValue: 1481 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) { 1482 ExtractValueInst *EVI = cast<ExtractValueInst>(I); 1483 if (EVI->getNumIndices() != 1) break; 1484 if (EVI->getIndices()[0] == 0) { 1485 switch (II->getIntrinsicID()) { 1486 default: break; 1487 case Intrinsic::uadd_with_overflow: 1488 case Intrinsic::sadd_with_overflow: 1489 computeKnownBitsAddSub(true, II->getArgOperand(0), 1490 II->getArgOperand(1), false, KnownZero, 1491 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q); 1492 break; 1493 case Intrinsic::usub_with_overflow: 1494 case Intrinsic::ssub_with_overflow: 1495 computeKnownBitsAddSub(false, II->getArgOperand(0), 1496 II->getArgOperand(1), false, KnownZero, 1497 KnownOne, KnownZero2, KnownOne2, DL, Depth, Q); 1498 break; 1499 case Intrinsic::umul_with_overflow: 1500 case Intrinsic::smul_with_overflow: 1501 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false, 1502 KnownZero, KnownOne, KnownZero2, KnownOne2, DL, 1503 Depth, Q); 1504 break; 1505 } 1506 } 1507 } 1508 } 1509 1510 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1511} 1512 1513/// Determine whether the sign bit is known to be zero or one. 1514/// Convenience wrapper around computeKnownBits. 1515void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne, 1516 const DataLayout &DL, unsigned Depth, const Query &Q) { 1517 unsigned BitWidth = getBitWidth(V->getType(), DL); 1518 if (!BitWidth) { 1519 KnownZero = false; 1520 KnownOne = false; 1521 return; 1522 } 1523 APInt ZeroBits(BitWidth, 0); 1524 APInt OneBits(BitWidth, 0); 1525 computeKnownBits(V, ZeroBits, OneBits, DL, Depth, Q); 1526 KnownOne = OneBits[BitWidth - 1]; 1527 KnownZero = ZeroBits[BitWidth - 1]; 1528} 1529 1530/// Return true if the given value is known to have exactly one 1531/// bit set when defined. For vectors return true if every element is known to 1532/// be a power of two when defined. Supports values with integer or pointer 1533/// types and vectors of integers. 1534bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth, 1535 const Query &Q, const DataLayout &DL) { 1536 if (Constant *C = dyn_cast<Constant>(V)) { 1537 if (C->isNullValue()) 1538 return OrZero; 1539 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1540 return CI->getValue().isPowerOf2(); 1541 // TODO: Handle vector constants. 1542 } 1543 1544 // 1 << X is clearly a power of two if the one is not shifted off the end. If 1545 // it is shifted off the end then the result is undefined. 1546 if (match(V, m_Shl(m_One(), m_Value()))) 1547 return true; 1548 1549 // (signbit) >>l X is clearly a power of two if the one is not shifted off the 1550 // bottom. If it is shifted off the bottom then the result is undefined. 1551 if (match(V, m_LShr(m_SignBit(), m_Value()))) 1552 return true; 1553 1554 // The remaining tests are all recursive, so bail out if we hit the limit. 1555 if (Depth++ == MaxDepth) 1556 return false; 1557 1558 Value *X = nullptr, *Y = nullptr; 1559 // A shift of a power of two is a power of two or zero. 1560 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) || 1561 match(V, m_Shr(m_Value(X), m_Value())))) 1562 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL); 1563 1564 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V)) 1565 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q, DL); 1566 1567 if (SelectInst *SI = dyn_cast<SelectInst>(V)) 1568 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q, DL) && 1569 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q, DL); 1570 1571 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) { 1572 // A power of two and'd with anything is a power of two or zero. 1573 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q, DL) || 1574 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q, DL)) 1575 return true; 1576 // X & (-X) is always a power of two or zero. 1577 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X)))) 1578 return true; 1579 return false; 1580 } 1581 1582 // Adding a power-of-two or zero to the same power-of-two or zero yields 1583 // either the original power-of-two, a larger power-of-two or zero. 1584 if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 1585 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V); 1586 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) { 1587 if (match(X, m_And(m_Specific(Y), m_Value())) || 1588 match(X, m_And(m_Value(), m_Specific(Y)))) 1589 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q, DL)) 1590 return true; 1591 if (match(Y, m_And(m_Specific(X), m_Value())) || 1592 match(Y, m_And(m_Value(), m_Specific(X)))) 1593 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q, DL)) 1594 return true; 1595 1596 unsigned BitWidth = V->getType()->getScalarSizeInBits(); 1597 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0); 1598 computeKnownBits(X, LHSZeroBits, LHSOneBits, DL, Depth, Q); 1599 1600 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0); 1601 computeKnownBits(Y, RHSZeroBits, RHSOneBits, DL, Depth, Q); 1602 // If i8 V is a power of two or zero: 1603 // ZeroBits: 1 1 1 0 1 1 1 1 1604 // ~ZeroBits: 0 0 0 1 0 0 0 0 1605 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2()) 1606 // If OrZero isn't set, we cannot give back a zero result. 1607 // Make sure either the LHS or RHS has a bit set. 1608 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue()) 1609 return true; 1610 } 1611 } 1612 1613 // An exact divide or right shift can only shift off zero bits, so the result 1614 // is a power of two only if the first operand is a power of two and not 1615 // copying a sign bit (sdiv int_min, 2). 1616 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) || 1617 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) { 1618 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, 1619 Depth, Q, DL); 1620 } 1621 1622 return false; 1623} 1624 1625/// \brief Test whether a GEP's result is known to be non-null. 1626/// 1627/// Uses properties inherent in a GEP to try to determine whether it is known 1628/// to be non-null. 1629/// 1630/// Currently this routine does not support vector GEPs. 1631static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout &DL, 1632 unsigned Depth, const Query &Q) { 1633 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0) 1634 return false; 1635 1636 // FIXME: Support vector-GEPs. 1637 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP"); 1638 1639 // If the base pointer is non-null, we cannot walk to a null address with an 1640 // inbounds GEP in address space zero. 1641 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q)) 1642 return true; 1643 1644 // Walk the GEP operands and see if any operand introduces a non-zero offset. 1645 // If so, then the GEP cannot produce a null pointer, as doing so would 1646 // inherently violate the inbounds contract within address space zero. 1647 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); 1648 GTI != GTE; ++GTI) { 1649 // Struct types are easy -- they must always be indexed by a constant. 1650 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 1651 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand()); 1652 unsigned ElementIdx = OpC->getZExtValue(); 1653 const StructLayout *SL = DL.getStructLayout(STy); 1654 uint64_t ElementOffset = SL->getElementOffset(ElementIdx); 1655 if (ElementOffset > 0) 1656 return true; 1657 continue; 1658 } 1659 1660 // If we have a zero-sized type, the index doesn't matter. Keep looping. 1661 if (DL.getTypeAllocSize(GTI.getIndexedType()) == 0) 1662 continue; 1663 1664 // Fast path the constant operand case both for efficiency and so we don't 1665 // increment Depth when just zipping down an all-constant GEP. 1666 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) { 1667 if (!OpC->isZero()) 1668 return true; 1669 continue; 1670 } 1671 1672 // We post-increment Depth here because while isKnownNonZero increments it 1673 // as well, when we pop back up that increment won't persist. We don't want 1674 // to recurse 10k times just because we have 10k GEP operands. We don't 1675 // bail completely out because we want to handle constant GEPs regardless 1676 // of depth. 1677 if (Depth++ >= MaxDepth) 1678 continue; 1679 1680 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q)) 1681 return true; 1682 } 1683 1684 return false; 1685} 1686 1687/// Does the 'Range' metadata (which must be a valid MD_range operand list) 1688/// ensure that the value it's attached to is never Value? 'RangeType' is 1689/// is the type of the value described by the range. 1690static bool rangeMetadataExcludesValue(MDNode* Ranges, 1691 const APInt& Value) { 1692 const unsigned NumRanges = Ranges->getNumOperands() / 2; 1693 assert(NumRanges >= 1); 1694 for (unsigned i = 0; i < NumRanges; ++i) { 1695 ConstantInt *Lower = 1696 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); 1697 ConstantInt *Upper = 1698 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); 1699 ConstantRange Range(Lower->getValue(), Upper->getValue()); 1700 if (Range.contains(Value)) 1701 return false; 1702 } 1703 return true; 1704} 1705 1706/// Return true if the given value is known to be non-zero when defined. 1707/// For vectors return true if every element is known to be non-zero when 1708/// defined. Supports values with integer or pointer type and vectors of 1709/// integers. 1710bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth, 1711 const Query &Q) { 1712 if (Constant *C = dyn_cast<Constant>(V)) { 1713 if (C->isNullValue()) 1714 return false; 1715 if (isa<ConstantInt>(C)) 1716 // Must be non-zero due to null test above. 1717 return true; 1718 // TODO: Handle vectors 1719 return false; 1720 } 1721 1722 if (Instruction* I = dyn_cast<Instruction>(V)) { 1723 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) { 1724 // If the possible ranges don't contain zero, then the value is 1725 // definitely non-zero. 1726 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) { 1727 const APInt ZeroValue(Ty->getBitWidth(), 0); 1728 if (rangeMetadataExcludesValue(Ranges, ZeroValue)) 1729 return true; 1730 } 1731 } 1732 } 1733 1734 // The remaining tests are all recursive, so bail out if we hit the limit. 1735 if (Depth++ >= MaxDepth) 1736 return false; 1737 1738 // Check for pointer simplifications. 1739 if (V->getType()->isPointerTy()) { 1740 if (isKnownNonNull(V)) 1741 return true; 1742 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 1743 if (isGEPKnownNonNull(GEP, DL, Depth, Q)) 1744 return true; 1745 } 1746 1747 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), DL); 1748 1749 // X | Y != 0 if X != 0 or Y != 0. 1750 Value *X = nullptr, *Y = nullptr; 1751 if (match(V, m_Or(m_Value(X), m_Value(Y)))) 1752 return isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q); 1753 1754 // ext X != 0 if X != 0. 1755 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) 1756 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), DL, Depth, Q); 1757 1758 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined 1759 // if the lowest bit is shifted off the end. 1760 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) { 1761 // shl nuw can't remove any non-zero bits. 1762 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 1763 if (BO->hasNoUnsignedWrap()) 1764 return isKnownNonZero(X, DL, Depth, Q); 1765 1766 APInt KnownZero(BitWidth, 0); 1767 APInt KnownOne(BitWidth, 0); 1768 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q); 1769 if (KnownOne[0]) 1770 return true; 1771 } 1772 // shr X, Y != 0 if X is negative. Note that the value of the shift is not 1773 // defined if the sign bit is shifted off the end. 1774 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) { 1775 // shr exact can only shift out zero bits. 1776 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V); 1777 if (BO->isExact()) 1778 return isKnownNonZero(X, DL, Depth, Q); 1779 1780 bool XKnownNonNegative, XKnownNegative; 1781 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q); 1782 if (XKnownNegative) 1783 return true; 1784 } 1785 // div exact can only produce a zero if the dividend is zero. 1786 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) { 1787 return isKnownNonZero(X, DL, Depth, Q); 1788 } 1789 // X + Y. 1790 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 1791 bool XKnownNonNegative, XKnownNegative; 1792 bool YKnownNonNegative, YKnownNegative; 1793 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, DL, Depth, Q); 1794 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, DL, Depth, Q); 1795 1796 // If X and Y are both non-negative (as signed values) then their sum is not 1797 // zero unless both X and Y are zero. 1798 if (XKnownNonNegative && YKnownNonNegative) 1799 if (isKnownNonZero(X, DL, Depth, Q) || isKnownNonZero(Y, DL, Depth, Q)) 1800 return true; 1801 1802 // If X and Y are both negative (as signed values) then their sum is not 1803 // zero unless both X and Y equal INT_MIN. 1804 if (BitWidth && XKnownNegative && YKnownNegative) { 1805 APInt KnownZero(BitWidth, 0); 1806 APInt KnownOne(BitWidth, 0); 1807 APInt Mask = APInt::getSignedMaxValue(BitWidth); 1808 // The sign bit of X is set. If some other bit is set then X is not equal 1809 // to INT_MIN. 1810 computeKnownBits(X, KnownZero, KnownOne, DL, Depth, Q); 1811 if ((KnownOne & Mask) != 0) 1812 return true; 1813 // The sign bit of Y is set. If some other bit is set then Y is not equal 1814 // to INT_MIN. 1815 computeKnownBits(Y, KnownZero, KnownOne, DL, Depth, Q); 1816 if ((KnownOne & Mask) != 0) 1817 return true; 1818 } 1819 1820 // The sum of a non-negative number and a power of two is not zero. 1821 if (XKnownNonNegative && 1822 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q, DL)) 1823 return true; 1824 if (YKnownNonNegative && 1825 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q, DL)) 1826 return true; 1827 } 1828 // X * Y. 1829 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) { 1830 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 1831 // If X and Y are non-zero then so is X * Y as long as the multiplication 1832 // does not overflow. 1833 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) && 1834 isKnownNonZero(X, DL, Depth, Q) && isKnownNonZero(Y, DL, Depth, Q)) 1835 return true; 1836 } 1837 // (C ? X : Y) != 0 if X != 0 and Y != 0. 1838 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) { 1839 if (isKnownNonZero(SI->getTrueValue(), DL, Depth, Q) && 1840 isKnownNonZero(SI->getFalseValue(), DL, Depth, Q)) 1841 return true; 1842 } 1843 1844 if (!BitWidth) return false; 1845 APInt KnownZero(BitWidth, 0); 1846 APInt KnownOne(BitWidth, 0); 1847 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q); 1848 return KnownOne != 0; 1849} 1850 1851/// Return true if 'V & Mask' is known to be zero. We use this predicate to 1852/// simplify operations downstream. Mask is known to be zero for bits that V 1853/// cannot have. 1854/// 1855/// This function is defined on values with integer type, values with pointer 1856/// type, and vectors of integers. In the case 1857/// where V is a vector, the mask, known zero, and known one values are the 1858/// same width as the vector element, and the bit is set only if it is true 1859/// for all of the elements in the vector. 1860bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL, 1861 unsigned Depth, const Query &Q) { 1862 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0); 1863 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q); 1864 return (KnownZero & Mask) == Mask; 1865} 1866 1867 1868 1869/// Return the number of times the sign bit of the register is replicated into 1870/// the other bits. We know that at least 1 bit is always equal to the sign bit 1871/// (itself), but other cases can give us information. For example, immediately 1872/// after an "ashr X, 2", we know that the top 3 bits are all equal to each 1873/// other, so we return 3. 1874/// 1875/// 'Op' must have a scalar integer type. 1876/// 1877unsigned ComputeNumSignBits(Value *V, const DataLayout &DL, unsigned Depth, 1878 const Query &Q) { 1879 unsigned TyBits = DL.getTypeSizeInBits(V->getType()->getScalarType()); 1880 unsigned Tmp, Tmp2; 1881 unsigned FirstAnswer = 1; 1882 1883 // Note that ConstantInt is handled by the general computeKnownBits case 1884 // below. 1885 1886 if (Depth == 6) 1887 return 1; // Limit search depth. 1888 1889 Operator *U = dyn_cast<Operator>(V); 1890 switch (Operator::getOpcode(V)) { 1891 default: break; 1892 case Instruction::SExt: 1893 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits(); 1894 return ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q) + Tmp; 1895 1896 case Instruction::SDiv: { 1897 const APInt *Denominator; 1898 // sdiv X, C -> adds log(C) sign bits. 1899 if (match(U->getOperand(1), m_APInt(Denominator))) { 1900 1901 // Ignore non-positive denominator. 1902 if (!Denominator->isStrictlyPositive()) 1903 break; 1904 1905 // Calculate the incoming numerator bits. 1906 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q); 1907 1908 // Add floor(log(C)) bits to the numerator bits. 1909 return std::min(TyBits, NumBits + Denominator->logBase2()); 1910 } 1911 break; 1912 } 1913 1914 case Instruction::SRem: { 1915 const APInt *Denominator; 1916 // srem X, C -> we know that the result is within [-C+1,C) when C is a 1917 // positive constant. This let us put a lower bound on the number of sign 1918 // bits. 1919 if (match(U->getOperand(1), m_APInt(Denominator))) { 1920 1921 // Ignore non-positive denominator. 1922 if (!Denominator->isStrictlyPositive()) 1923 break; 1924 1925 // Calculate the incoming numerator bits. SRem by a positive constant 1926 // can't lower the number of sign bits. 1927 unsigned NumrBits = 1928 ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q); 1929 1930 // Calculate the leading sign bit constraints by examining the 1931 // denominator. Given that the denominator is positive, there are two 1932 // cases: 1933 // 1934 // 1. the numerator is positive. The result range is [0,C) and [0,C) u< 1935 // (1 << ceilLogBase2(C)). 1936 // 1937 // 2. the numerator is negative. Then the result range is (-C,0] and 1938 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). 1939 // 1940 // Thus a lower bound on the number of sign bits is `TyBits - 1941 // ceilLogBase2(C)`. 1942 1943 unsigned ResBits = TyBits - Denominator->ceilLogBase2(); 1944 return std::max(NumrBits, ResBits); 1945 } 1946 break; 1947 } 1948 1949 case Instruction::AShr: { 1950 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q); 1951 // ashr X, C -> adds C sign bits. Vectors too. 1952 const APInt *ShAmt; 1953 if (match(U->getOperand(1), m_APInt(ShAmt))) { 1954 Tmp += ShAmt->getZExtValue(); 1955 if (Tmp > TyBits) Tmp = TyBits; 1956 } 1957 return Tmp; 1958 } 1959 case Instruction::Shl: { 1960 const APInt *ShAmt; 1961 if (match(U->getOperand(1), m_APInt(ShAmt))) { 1962 // shl destroys sign bits. 1963 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q); 1964 Tmp2 = ShAmt->getZExtValue(); 1965 if (Tmp2 >= TyBits || // Bad shift. 1966 Tmp2 >= Tmp) break; // Shifted all sign bits out. 1967 return Tmp - Tmp2; 1968 } 1969 break; 1970 } 1971 case Instruction::And: 1972 case Instruction::Or: 1973 case Instruction::Xor: // NOT is handled here. 1974 // Logical binary ops preserve the number of sign bits at the worst. 1975 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q); 1976 if (Tmp != 1) { 1977 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q); 1978 FirstAnswer = std::min(Tmp, Tmp2); 1979 // We computed what we know about the sign bits as our first 1980 // answer. Now proceed to the generic code that uses 1981 // computeKnownBits, and pick whichever answer is better. 1982 } 1983 break; 1984 1985 case Instruction::Select: 1986 Tmp = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q); 1987 if (Tmp == 1) return 1; // Early out. 1988 Tmp2 = ComputeNumSignBits(U->getOperand(2), DL, Depth + 1, Q); 1989 return std::min(Tmp, Tmp2); 1990 1991 case Instruction::Add: 1992 // Add can have at most one carry bit. Thus we know that the output 1993 // is, at worst, one more bit than the inputs. 1994 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q); 1995 if (Tmp == 1) return 1; // Early out. 1996 1997 // Special case decrementing a value (ADD X, -1): 1998 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1))) 1999 if (CRHS->isAllOnesValue()) { 2000 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0); 2001 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, Depth + 1, 2002 Q); 2003 2004 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2005 // sign bits set. 2006 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue()) 2007 return TyBits; 2008 2009 // If we are subtracting one from a positive number, there is no carry 2010 // out of the result. 2011 if (KnownZero.isNegative()) 2012 return Tmp; 2013 } 2014 2015 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q); 2016 if (Tmp2 == 1) return 1; 2017 return std::min(Tmp, Tmp2)-1; 2018 2019 case Instruction::Sub: 2020 Tmp2 = ComputeNumSignBits(U->getOperand(1), DL, Depth + 1, Q); 2021 if (Tmp2 == 1) return 1; 2022 2023 // Handle NEG. 2024 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0))) 2025 if (CLHS->isNullValue()) { 2026 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0); 2027 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, 2028 Q); 2029 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2030 // sign bits set. 2031 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue()) 2032 return TyBits; 2033 2034 // If the input is known to be positive (the sign bit is known clear), 2035 // the output of the NEG has the same number of sign bits as the input. 2036 if (KnownZero.isNegative()) 2037 return Tmp2; 2038 2039 // Otherwise, we treat this like a SUB. 2040 } 2041 2042 // Sub can have at most one carry bit. Thus we know that the output 2043 // is, at worst, one more bit than the inputs. 2044 Tmp = ComputeNumSignBits(U->getOperand(0), DL, Depth + 1, Q); 2045 if (Tmp == 1) return 1; // Early out. 2046 return std::min(Tmp, Tmp2)-1; 2047 2048 case Instruction::PHI: { 2049 PHINode *PN = cast<PHINode>(U); 2050 unsigned NumIncomingValues = PN->getNumIncomingValues(); 2051 // Don't analyze large in-degree PHIs. 2052 if (NumIncomingValues > 4) break; 2053 // Unreachable blocks may have zero-operand PHI nodes. 2054 if (NumIncomingValues == 0) break; 2055 2056 // Take the minimum of all incoming values. This can't infinitely loop 2057 // because of our depth threshold. 2058 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), DL, Depth + 1, Q); 2059 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) { 2060 if (Tmp == 1) return Tmp; 2061 Tmp = std::min( 2062 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), DL, Depth + 1, Q)); 2063 } 2064 return Tmp; 2065 } 2066 2067 case Instruction::Trunc: 2068 // FIXME: it's tricky to do anything useful for this, but it is an important 2069 // case for targets like X86. 2070 break; 2071 } 2072 2073 // Finally, if we can prove that the top bits of the result are 0's or 1's, 2074 // use this information. 2075 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0); 2076 APInt Mask; 2077 computeKnownBits(V, KnownZero, KnownOne, DL, Depth, Q); 2078 2079 if (KnownZero.isNegative()) { // sign bit is 0 2080 Mask = KnownZero; 2081 } else if (KnownOne.isNegative()) { // sign bit is 1; 2082 Mask = KnownOne; 2083 } else { 2084 // Nothing known. 2085 return FirstAnswer; 2086 } 2087 2088 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine 2089 // the number of identical bits in the top of the input value. 2090 Mask = ~Mask; 2091 Mask <<= Mask.getBitWidth()-TyBits; 2092 // Return # leading zeros. We use 'min' here in case Val was zero before 2093 // shifting. We don't want to return '64' as for an i32 "0". 2094 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros())); 2095} 2096 2097/// This function computes the integer multiple of Base that equals V. 2098/// If successful, it returns true and returns the multiple in 2099/// Multiple. If unsuccessful, it returns false. It looks 2100/// through SExt instructions only if LookThroughSExt is true. 2101bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple, 2102 bool LookThroughSExt, unsigned Depth) { 2103 const unsigned MaxDepth = 6; 2104 2105 assert(V && "No Value?"); 2106 assert(Depth <= MaxDepth && "Limit Search Depth"); 2107 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!"); 2108 2109 Type *T = V->getType(); 2110 2111 ConstantInt *CI = dyn_cast<ConstantInt>(V); 2112 2113 if (Base == 0) 2114 return false; 2115 2116 if (Base == 1) { 2117 Multiple = V; 2118 return true; 2119 } 2120 2121 ConstantExpr *CO = dyn_cast<ConstantExpr>(V); 2122 Constant *BaseVal = ConstantInt::get(T, Base); 2123 if (CO && CO == BaseVal) { 2124 // Multiple is 1. 2125 Multiple = ConstantInt::get(T, 1); 2126 return true; 2127 } 2128 2129 if (CI && CI->getZExtValue() % Base == 0) { 2130 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base); 2131 return true; 2132 } 2133 2134 if (Depth == MaxDepth) return false; // Limit search depth. 2135 2136 Operator *I = dyn_cast<Operator>(V); 2137 if (!I) return false; 2138 2139 switch (I->getOpcode()) { 2140 default: break; 2141 case Instruction::SExt: 2142 if (!LookThroughSExt) return false; 2143 // otherwise fall through to ZExt 2144 case Instruction::ZExt: 2145 return ComputeMultiple(I->getOperand(0), Base, Multiple, 2146 LookThroughSExt, Depth+1); 2147 case Instruction::Shl: 2148 case Instruction::Mul: { 2149 Value *Op0 = I->getOperand(0); 2150 Value *Op1 = I->getOperand(1); 2151 2152 if (I->getOpcode() == Instruction::Shl) { 2153 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1); 2154 if (!Op1CI) return false; 2155 // Turn Op0 << Op1 into Op0 * 2^Op1 2156 APInt Op1Int = Op1CI->getValue(); 2157 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1); 2158 APInt API(Op1Int.getBitWidth(), 0); 2159 API.setBit(BitToSet); 2160 Op1 = ConstantInt::get(V->getContext(), API); 2161 } 2162 2163 Value *Mul0 = nullptr; 2164 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) { 2165 if (Constant *Op1C = dyn_cast<Constant>(Op1)) 2166 if (Constant *MulC = dyn_cast<Constant>(Mul0)) { 2167 if (Op1C->getType()->getPrimitiveSizeInBits() < 2168 MulC->getType()->getPrimitiveSizeInBits()) 2169 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType()); 2170 if (Op1C->getType()->getPrimitiveSizeInBits() > 2171 MulC->getType()->getPrimitiveSizeInBits()) 2172 MulC = ConstantExpr::getZExt(MulC, Op1C->getType()); 2173 2174 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1) 2175 Multiple = ConstantExpr::getMul(MulC, Op1C); 2176 return true; 2177 } 2178 2179 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0)) 2180 if (Mul0CI->getValue() == 1) { 2181 // V == Base * Op1, so return Op1 2182 Multiple = Op1; 2183 return true; 2184 } 2185 } 2186 2187 Value *Mul1 = nullptr; 2188 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) { 2189 if (Constant *Op0C = dyn_cast<Constant>(Op0)) 2190 if (Constant *MulC = dyn_cast<Constant>(Mul1)) { 2191 if (Op0C->getType()->getPrimitiveSizeInBits() < 2192 MulC->getType()->getPrimitiveSizeInBits()) 2193 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType()); 2194 if (Op0C->getType()->getPrimitiveSizeInBits() > 2195 MulC->getType()->getPrimitiveSizeInBits()) 2196 MulC = ConstantExpr::getZExt(MulC, Op0C->getType()); 2197 2198 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0) 2199 Multiple = ConstantExpr::getMul(MulC, Op0C); 2200 return true; 2201 } 2202 2203 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1)) 2204 if (Mul1CI->getValue() == 1) { 2205 // V == Base * Op0, so return Op0 2206 Multiple = Op0; 2207 return true; 2208 } 2209 } 2210 } 2211 } 2212 2213 // We could not determine if V is a multiple of Base. 2214 return false; 2215} 2216 2217/// Return true if we can prove that the specified FP value is never equal to 2218/// -0.0. 2219/// 2220/// NOTE: this function will need to be revisited when we support non-default 2221/// rounding modes! 2222/// 2223bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) { 2224 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) 2225 return !CFP->getValueAPF().isNegZero(); 2226 2227 // FIXME: Magic number! At the least, this should be given a name because it's 2228 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to 2229 // expose it as a parameter, so it can be used for testing / experimenting. 2230 if (Depth == 6) 2231 return false; // Limit search depth. 2232 2233 const Operator *I = dyn_cast<Operator>(V); 2234 if (!I) return false; 2235 2236 // Check if the nsz fast-math flag is set 2237 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I)) 2238 if (FPO->hasNoSignedZeros()) 2239 return true; 2240 2241 // (add x, 0.0) is guaranteed to return +0.0, not -0.0. 2242 if (I->getOpcode() == Instruction::FAdd) 2243 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1))) 2244 if (CFP->isNullValue()) 2245 return true; 2246 2247 // sitofp and uitofp turn into +0.0 for zero. 2248 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I)) 2249 return true; 2250 2251 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 2252 // sqrt(-0.0) = -0.0, no other negative results are possible. 2253 if (II->getIntrinsicID() == Intrinsic::sqrt) 2254 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1); 2255 2256 if (const CallInst *CI = dyn_cast<CallInst>(I)) 2257 if (const Function *F = CI->getCalledFunction()) { 2258 if (F->isDeclaration()) { 2259 // abs(x) != -0.0 2260 if (F->getName() == "abs") return true; 2261 // fabs[lf](x) != -0.0 2262 if (F->getName() == "fabs") return true; 2263 if (F->getName() == "fabsf") return true; 2264 if (F->getName() == "fabsl") return true; 2265 if (F->getName() == "sqrt" || F->getName() == "sqrtf" || 2266 F->getName() == "sqrtl") 2267 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1); 2268 } 2269 } 2270 2271 return false; 2272} 2273 2274bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) { 2275 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) 2276 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero(); 2277 2278 // FIXME: Magic number! At the least, this should be given a name because it's 2279 // used similarly in CannotBeNegativeZero(). A better fix may be to 2280 // expose it as a parameter, so it can be used for testing / experimenting. 2281 if (Depth == 6) 2282 return false; // Limit search depth. 2283 2284 const Operator *I = dyn_cast<Operator>(V); 2285 if (!I) return false; 2286 2287 switch (I->getOpcode()) { 2288 default: break; 2289 case Instruction::FMul: 2290 // x*x is always non-negative or a NaN. 2291 if (I->getOperand(0) == I->getOperand(1)) 2292 return true; 2293 // Fall through 2294 case Instruction::FAdd: 2295 case Instruction::FDiv: 2296 case Instruction::FRem: 2297 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) && 2298 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1); 2299 case Instruction::FPExt: 2300 case Instruction::FPTrunc: 2301 // Widening/narrowing never change sign. 2302 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1); 2303 case Instruction::Call: 2304 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 2305 switch (II->getIntrinsicID()) { 2306 default: break; 2307 case Intrinsic::exp: 2308 case Intrinsic::exp2: 2309 case Intrinsic::fabs: 2310 case Intrinsic::sqrt: 2311 return true; 2312 case Intrinsic::powi: 2313 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 2314 // powi(x,n) is non-negative if n is even. 2315 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0) 2316 return true; 2317 } 2318 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1); 2319 case Intrinsic::fma: 2320 case Intrinsic::fmuladd: 2321 // x*x+y is non-negative if y is non-negative. 2322 return I->getOperand(0) == I->getOperand(1) && 2323 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1); 2324 } 2325 break; 2326 } 2327 return false; 2328} 2329 2330/// If the specified value can be set by repeating the same byte in memory, 2331/// return the i8 value that it is represented with. This is 2332/// true for all i8 values obviously, but is also true for i32 0, i32 -1, 2333/// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated 2334/// byte store (e.g. i16 0x1234), return null. 2335Value *llvm::isBytewiseValue(Value *V) { 2336 // All byte-wide stores are splatable, even of arbitrary variables. 2337 if (V->getType()->isIntegerTy(8)) return V; 2338 2339 // Handle 'null' ConstantArrayZero etc. 2340 if (Constant *C = dyn_cast<Constant>(V)) 2341 if (C->isNullValue()) 2342 return Constant::getNullValue(Type::getInt8Ty(V->getContext())); 2343 2344 // Constant float and double values can be handled as integer values if the 2345 // corresponding integer value is "byteable". An important case is 0.0. 2346 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 2347 if (CFP->getType()->isFloatTy()) 2348 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext())); 2349 if (CFP->getType()->isDoubleTy()) 2350 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext())); 2351 // Don't handle long double formats, which have strange constraints. 2352 } 2353 2354 // We can handle constant integers that are multiple of 8 bits. 2355 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 2356 if (CI->getBitWidth() % 8 == 0) { 2357 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!"); 2358 2359 if (!CI->getValue().isSplat(8)) 2360 return nullptr; 2361 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8)); 2362 } 2363 } 2364 2365 // A ConstantDataArray/Vector is splatable if all its members are equal and 2366 // also splatable. 2367 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) { 2368 Value *Elt = CA->getElementAsConstant(0); 2369 Value *Val = isBytewiseValue(Elt); 2370 if (!Val) 2371 return nullptr; 2372 2373 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I) 2374 if (CA->getElementAsConstant(I) != Elt) 2375 return nullptr; 2376 2377 return Val; 2378 } 2379 2380 // Conceptually, we could handle things like: 2381 // %a = zext i8 %X to i16 2382 // %b = shl i16 %a, 8 2383 // %c = or i16 %a, %b 2384 // but until there is an example that actually needs this, it doesn't seem 2385 // worth worrying about. 2386 return nullptr; 2387} 2388 2389 2390// This is the recursive version of BuildSubAggregate. It takes a few different 2391// arguments. Idxs is the index within the nested struct From that we are 2392// looking at now (which is of type IndexedType). IdxSkip is the number of 2393// indices from Idxs that should be left out when inserting into the resulting 2394// struct. To is the result struct built so far, new insertvalue instructions 2395// build on that. 2396static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType, 2397 SmallVectorImpl<unsigned> &Idxs, 2398 unsigned IdxSkip, 2399 Instruction *InsertBefore) { 2400 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType); 2401 if (STy) { 2402 // Save the original To argument so we can modify it 2403 Value *OrigTo = To; 2404 // General case, the type indexed by Idxs is a struct 2405 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2406 // Process each struct element recursively 2407 Idxs.push_back(i); 2408 Value *PrevTo = To; 2409 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip, 2410 InsertBefore); 2411 Idxs.pop_back(); 2412 if (!To) { 2413 // Couldn't find any inserted value for this index? Cleanup 2414 while (PrevTo != OrigTo) { 2415 InsertValueInst* Del = cast<InsertValueInst>(PrevTo); 2416 PrevTo = Del->getAggregateOperand(); 2417 Del->eraseFromParent(); 2418 } 2419 // Stop processing elements 2420 break; 2421 } 2422 } 2423 // If we successfully found a value for each of our subaggregates 2424 if (To) 2425 return To; 2426 } 2427 // Base case, the type indexed by SourceIdxs is not a struct, or not all of 2428 // the struct's elements had a value that was inserted directly. In the latter 2429 // case, perhaps we can't determine each of the subelements individually, but 2430 // we might be able to find the complete struct somewhere. 2431 2432 // Find the value that is at that particular spot 2433 Value *V = FindInsertedValue(From, Idxs); 2434 2435 if (!V) 2436 return nullptr; 2437 2438 // Insert the value in the new (sub) aggregrate 2439 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip), 2440 "tmp", InsertBefore); 2441} 2442 2443// This helper takes a nested struct and extracts a part of it (which is again a 2444// struct) into a new value. For example, given the struct: 2445// { a, { b, { c, d }, e } } 2446// and the indices "1, 1" this returns 2447// { c, d }. 2448// 2449// It does this by inserting an insertvalue for each element in the resulting 2450// struct, as opposed to just inserting a single struct. This will only work if 2451// each of the elements of the substruct are known (ie, inserted into From by an 2452// insertvalue instruction somewhere). 2453// 2454// All inserted insertvalue instructions are inserted before InsertBefore 2455static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, 2456 Instruction *InsertBefore) { 2457 assert(InsertBefore && "Must have someplace to insert!"); 2458 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(), 2459 idx_range); 2460 Value *To = UndefValue::get(IndexedType); 2461 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); 2462 unsigned IdxSkip = Idxs.size(); 2463 2464 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); 2465} 2466 2467/// Given an aggregrate and an sequence of indices, see if 2468/// the scalar value indexed is already around as a register, for example if it 2469/// were inserted directly into the aggregrate. 2470/// 2471/// If InsertBefore is not null, this function will duplicate (modified) 2472/// insertvalues when a part of a nested struct is extracted. 2473Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, 2474 Instruction *InsertBefore) { 2475 // Nothing to index? Just return V then (this is useful at the end of our 2476 // recursion). 2477 if (idx_range.empty()) 2478 return V; 2479 // We have indices, so V should have an indexable type. 2480 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) && 2481 "Not looking at a struct or array?"); 2482 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) && 2483 "Invalid indices for type?"); 2484 2485 if (Constant *C = dyn_cast<Constant>(V)) { 2486 C = C->getAggregateElement(idx_range[0]); 2487 if (!C) return nullptr; 2488 return FindInsertedValue(C, idx_range.slice(1), InsertBefore); 2489 } 2490 2491 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) { 2492 // Loop the indices for the insertvalue instruction in parallel with the 2493 // requested indices 2494 const unsigned *req_idx = idx_range.begin(); 2495 for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); 2496 i != e; ++i, ++req_idx) { 2497 if (req_idx == idx_range.end()) { 2498 // We can't handle this without inserting insertvalues 2499 if (!InsertBefore) 2500 return nullptr; 2501 2502 // The requested index identifies a part of a nested aggregate. Handle 2503 // this specially. For example, 2504 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 2505 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 2506 // %C = extractvalue {i32, { i32, i32 } } %B, 1 2507 // This can be changed into 2508 // %A = insertvalue {i32, i32 } undef, i32 10, 0 2509 // %C = insertvalue {i32, i32 } %A, i32 11, 1 2510 // which allows the unused 0,0 element from the nested struct to be 2511 // removed. 2512 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx), 2513 InsertBefore); 2514 } 2515 2516 // This insert value inserts something else than what we are looking for. 2517 // See if the (aggregrate) value inserted into has the value we are 2518 // looking for, then. 2519 if (*req_idx != *i) 2520 return FindInsertedValue(I->getAggregateOperand(), idx_range, 2521 InsertBefore); 2522 } 2523 // If we end up here, the indices of the insertvalue match with those 2524 // requested (though possibly only partially). Now we recursively look at 2525 // the inserted value, passing any remaining indices. 2526 return FindInsertedValue(I->getInsertedValueOperand(), 2527 makeArrayRef(req_idx, idx_range.end()), 2528 InsertBefore); 2529 } 2530 2531 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) { 2532 // If we're extracting a value from an aggregrate that was extracted from 2533 // something else, we can extract from that something else directly instead. 2534 // However, we will need to chain I's indices with the requested indices. 2535 2536 // Calculate the number of indices required 2537 unsigned size = I->getNumIndices() + idx_range.size(); 2538 // Allocate some space to put the new indices in 2539 SmallVector<unsigned, 5> Idxs; 2540 Idxs.reserve(size); 2541 // Add indices from the extract value instruction 2542 Idxs.append(I->idx_begin(), I->idx_end()); 2543 2544 // Add requested indices 2545 Idxs.append(idx_range.begin(), idx_range.end()); 2546 2547 assert(Idxs.size() == size 2548 && "Number of indices added not correct?"); 2549 2550 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore); 2551 } 2552 // Otherwise, we don't know (such as, extracting from a function return value 2553 // or load instruction) 2554 return nullptr; 2555} 2556 2557/// Analyze the specified pointer to see if it can be expressed as a base 2558/// pointer plus a constant offset. Return the base and offset to the caller. 2559Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 2560 const DataLayout &DL) { 2561 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType()); 2562 APInt ByteOffset(BitWidth, 0); 2563 while (1) { 2564 if (Ptr->getType()->isVectorTy()) 2565 break; 2566 2567 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 2568 APInt GEPOffset(BitWidth, 0); 2569 if (!GEP->accumulateConstantOffset(DL, GEPOffset)) 2570 break; 2571 2572 ByteOffset += GEPOffset; 2573 2574 Ptr = GEP->getPointerOperand(); 2575 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast || 2576 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) { 2577 Ptr = cast<Operator>(Ptr)->getOperand(0); 2578 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) { 2579 if (GA->mayBeOverridden()) 2580 break; 2581 Ptr = GA->getAliasee(); 2582 } else { 2583 break; 2584 } 2585 } 2586 Offset = ByteOffset.getSExtValue(); 2587 return Ptr; 2588} 2589 2590 2591/// This function computes the length of a null-terminated C string pointed to 2592/// by V. If successful, it returns true and returns the string in Str. 2593/// If unsuccessful, it returns false. 2594bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, 2595 uint64_t Offset, bool TrimAtNul) { 2596 assert(V); 2597 2598 // Look through bitcast instructions and geps. 2599 V = V->stripPointerCasts(); 2600 2601 // If the value is a GEP instruction or constant expression, treat it as an 2602 // offset. 2603 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 2604 // Make sure the GEP has exactly three arguments. 2605 if (GEP->getNumOperands() != 3) 2606 return false; 2607 2608 // Make sure the index-ee is a pointer to array of i8. 2609 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType()); 2610 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType()); 2611 if (!AT || !AT->getElementType()->isIntegerTy(8)) 2612 return false; 2613 2614 // Check to make sure that the first operand of the GEP is an integer and 2615 // has value 0 so that we are sure we're indexing into the initializer. 2616 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1)); 2617 if (!FirstIdx || !FirstIdx->isZero()) 2618 return false; 2619 2620 // If the second index isn't a ConstantInt, then this is a variable index 2621 // into the array. If this occurs, we can't say anything meaningful about 2622 // the string. 2623 uint64_t StartIdx = 0; 2624 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2))) 2625 StartIdx = CI->getZExtValue(); 2626 else 2627 return false; 2628 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset, 2629 TrimAtNul); 2630 } 2631 2632 // The GEP instruction, constant or instruction, must reference a global 2633 // variable that is a constant and is initialized. The referenced constant 2634 // initializer is the array that we'll use for optimization. 2635 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V); 2636 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) 2637 return false; 2638 2639 // Handle the all-zeros case 2640 if (GV->getInitializer()->isNullValue()) { 2641 // This is a degenerate case. The initializer is constant zero so the 2642 // length of the string must be zero. 2643 Str = ""; 2644 return true; 2645 } 2646 2647 // Must be a Constant Array 2648 const ConstantDataArray *Array = 2649 dyn_cast<ConstantDataArray>(GV->getInitializer()); 2650 if (!Array || !Array->isString()) 2651 return false; 2652 2653 // Get the number of elements in the array 2654 uint64_t NumElts = Array->getType()->getArrayNumElements(); 2655 2656 // Start out with the entire array in the StringRef. 2657 Str = Array->getAsString(); 2658 2659 if (Offset > NumElts) 2660 return false; 2661 2662 // Skip over 'offset' bytes. 2663 Str = Str.substr(Offset); 2664 2665 if (TrimAtNul) { 2666 // Trim off the \0 and anything after it. If the array is not nul 2667 // terminated, we just return the whole end of string. The client may know 2668 // some other way that the string is length-bound. 2669 Str = Str.substr(0, Str.find('\0')); 2670 } 2671 return true; 2672} 2673 2674// These next two are very similar to the above, but also look through PHI 2675// nodes. 2676// TODO: See if we can integrate these two together. 2677 2678/// If we can compute the length of the string pointed to by 2679/// the specified pointer, return 'len+1'. If we can't, return 0. 2680static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) { 2681 // Look through noop bitcast instructions. 2682 V = V->stripPointerCasts(); 2683 2684 // If this is a PHI node, there are two cases: either we have already seen it 2685 // or we haven't. 2686 if (PHINode *PN = dyn_cast<PHINode>(V)) { 2687 if (!PHIs.insert(PN).second) 2688 return ~0ULL; // already in the set. 2689 2690 // If it was new, see if all the input strings are the same length. 2691 uint64_t LenSoFar = ~0ULL; 2692 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 2693 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs); 2694 if (Len == 0) return 0; // Unknown length -> unknown. 2695 2696 if (Len == ~0ULL) continue; 2697 2698 if (Len != LenSoFar && LenSoFar != ~0ULL) 2699 return 0; // Disagree -> unknown. 2700 LenSoFar = Len; 2701 } 2702 2703 // Success, all agree. 2704 return LenSoFar; 2705 } 2706 2707 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) 2708 if (SelectInst *SI = dyn_cast<SelectInst>(V)) { 2709 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs); 2710 if (Len1 == 0) return 0; 2711 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs); 2712 if (Len2 == 0) return 0; 2713 if (Len1 == ~0ULL) return Len2; 2714 if (Len2 == ~0ULL) return Len1; 2715 if (Len1 != Len2) return 0; 2716 return Len1; 2717 } 2718 2719 // Otherwise, see if we can read the string. 2720 StringRef StrData; 2721 if (!getConstantStringInfo(V, StrData)) 2722 return 0; 2723 2724 return StrData.size()+1; 2725} 2726 2727/// If we can compute the length of the string pointed to by 2728/// the specified pointer, return 'len+1'. If we can't, return 0. 2729uint64_t llvm::GetStringLength(Value *V) { 2730 if (!V->getType()->isPointerTy()) return 0; 2731 2732 SmallPtrSet<PHINode*, 32> PHIs; 2733 uint64_t Len = GetStringLengthH(V, PHIs); 2734 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return 2735 // an empty string as a length. 2736 return Len == ~0ULL ? 1 : Len; 2737} 2738 2739Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL, 2740 unsigned MaxLookup) { 2741 if (!V->getType()->isPointerTy()) 2742 return V; 2743 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { 2744 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 2745 V = GEP->getPointerOperand(); 2746 } else if (Operator::getOpcode(V) == Instruction::BitCast || 2747 Operator::getOpcode(V) == Instruction::AddrSpaceCast) { 2748 V = cast<Operator>(V)->getOperand(0); 2749 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 2750 if (GA->mayBeOverridden()) 2751 return V; 2752 V = GA->getAliasee(); 2753 } else { 2754 // See if InstructionSimplify knows any relevant tricks. 2755 if (Instruction *I = dyn_cast<Instruction>(V)) 2756 // TODO: Acquire a DominatorTree and AssumptionCache and use them. 2757 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) { 2758 V = Simplified; 2759 continue; 2760 } 2761 2762 return V; 2763 } 2764 assert(V->getType()->isPointerTy() && "Unexpected operand type!"); 2765 } 2766 return V; 2767} 2768 2769void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects, 2770 const DataLayout &DL, unsigned MaxLookup) { 2771 SmallPtrSet<Value *, 4> Visited; 2772 SmallVector<Value *, 4> Worklist; 2773 Worklist.push_back(V); 2774 do { 2775 Value *P = Worklist.pop_back_val(); 2776 P = GetUnderlyingObject(P, DL, MaxLookup); 2777 2778 if (!Visited.insert(P).second) 2779 continue; 2780 2781 if (SelectInst *SI = dyn_cast<SelectInst>(P)) { 2782 Worklist.push_back(SI->getTrueValue()); 2783 Worklist.push_back(SI->getFalseValue()); 2784 continue; 2785 } 2786 2787 if (PHINode *PN = dyn_cast<PHINode>(P)) { 2788 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 2789 Worklist.push_back(PN->getIncomingValue(i)); 2790 continue; 2791 } 2792 2793 Objects.push_back(P); 2794 } while (!Worklist.empty()); 2795} 2796 2797/// Return true if the only users of this pointer are lifetime markers. 2798bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { 2799 for (const User *U : V->users()) { 2800 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); 2801 if (!II) return false; 2802 2803 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 2804 II->getIntrinsicID() != Intrinsic::lifetime_end) 2805 return false; 2806 } 2807 return true; 2808} 2809 2810bool llvm::isSafeToSpeculativelyExecute(const Value *V) { 2811 const Operator *Inst = dyn_cast<Operator>(V); 2812 if (!Inst) 2813 return false; 2814 2815 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) 2816 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i))) 2817 if (C->canTrap()) 2818 return false; 2819 2820 switch (Inst->getOpcode()) { 2821 default: 2822 return true; 2823 case Instruction::UDiv: 2824 case Instruction::URem: { 2825 // x / y is undefined if y == 0. 2826 const APInt *V; 2827 if (match(Inst->getOperand(1), m_APInt(V))) 2828 return *V != 0; 2829 return false; 2830 } 2831 case Instruction::SDiv: 2832 case Instruction::SRem: { 2833 // x / y is undefined if y == 0 or x == INT_MIN and y == -1 2834 const APInt *Numerator, *Denominator; 2835 if (!match(Inst->getOperand(1), m_APInt(Denominator))) 2836 return false; 2837 // We cannot hoist this division if the denominator is 0. 2838 if (*Denominator == 0) 2839 return false; 2840 // It's safe to hoist if the denominator is not 0 or -1. 2841 if (*Denominator != -1) 2842 return true; 2843 // At this point we know that the denominator is -1. It is safe to hoist as 2844 // long we know that the numerator is not INT_MIN. 2845 if (match(Inst->getOperand(0), m_APInt(Numerator))) 2846 return !Numerator->isMinSignedValue(); 2847 // The numerator *might* be MinSignedValue. 2848 return false; 2849 } 2850 case Instruction::Load: { 2851 const LoadInst *LI = cast<LoadInst>(Inst); 2852 if (!LI->isUnordered() || 2853 // Speculative load may create a race that did not exist in the source. 2854 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread)) 2855 return false; 2856 const DataLayout &DL = LI->getModule()->getDataLayout(); 2857 return LI->getPointerOperand()->isDereferenceablePointer(DL); 2858 } 2859 case Instruction::Call: { 2860 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 2861 switch (II->getIntrinsicID()) { 2862 // These synthetic intrinsics have no side-effects and just mark 2863 // information about their operands. 2864 // FIXME: There are other no-op synthetic instructions that potentially 2865 // should be considered at least *safe* to speculate... 2866 case Intrinsic::dbg_declare: 2867 case Intrinsic::dbg_value: 2868 return true; 2869 2870 case Intrinsic::bswap: 2871 case Intrinsic::ctlz: 2872 case Intrinsic::ctpop: 2873 case Intrinsic::cttz: 2874 case Intrinsic::objectsize: 2875 case Intrinsic::sadd_with_overflow: 2876 case Intrinsic::smul_with_overflow: 2877 case Intrinsic::ssub_with_overflow: 2878 case Intrinsic::uadd_with_overflow: 2879 case Intrinsic::umul_with_overflow: 2880 case Intrinsic::usub_with_overflow: 2881 return true; 2882 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set 2883 // errno like libm sqrt would. 2884 case Intrinsic::sqrt: 2885 case Intrinsic::fma: 2886 case Intrinsic::fmuladd: 2887 case Intrinsic::fabs: 2888 case Intrinsic::minnum: 2889 case Intrinsic::maxnum: 2890 return true; 2891 // TODO: some fp intrinsics are marked as having the same error handling 2892 // as libm. They're safe to speculate when they won't error. 2893 // TODO: are convert_{from,to}_fp16 safe? 2894 // TODO: can we list target-specific intrinsics here? 2895 default: break; 2896 } 2897 } 2898 return false; // The called function could have undefined behavior or 2899 // side-effects, even if marked readnone nounwind. 2900 } 2901 case Instruction::VAArg: 2902 case Instruction::Alloca: 2903 case Instruction::Invoke: 2904 case Instruction::PHI: 2905 case Instruction::Store: 2906 case Instruction::Ret: 2907 case Instruction::Br: 2908 case Instruction::IndirectBr: 2909 case Instruction::Switch: 2910 case Instruction::Unreachable: 2911 case Instruction::Fence: 2912 case Instruction::LandingPad: 2913 case Instruction::AtomicRMW: 2914 case Instruction::AtomicCmpXchg: 2915 case Instruction::Resume: 2916 return false; // Misc instructions which have effects 2917 } 2918} 2919 2920/// Return true if we know that the specified value is never null. 2921bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) { 2922 // Alloca never returns null, malloc might. 2923 if (isa<AllocaInst>(V)) return true; 2924 2925 // A byval, inalloca, or nonnull argument is never null. 2926 if (const Argument *A = dyn_cast<Argument>(V)) 2927 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr(); 2928 2929 // Global values are not null unless extern weak. 2930 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 2931 return !GV->hasExternalWeakLinkage(); 2932 2933 // A Load tagged w/nonnull metadata is never null. 2934 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) 2935 return LI->getMetadata(LLVMContext::MD_nonnull); 2936 2937 if (auto CS = ImmutableCallSite(V)) 2938 if (CS.isReturnNonNull()) 2939 return true; 2940 2941 // operator new never returns null. 2942 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true)) 2943 return true; 2944 2945 return false; 2946} 2947 2948OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS, 2949 const DataLayout &DL, 2950 AssumptionCache *AC, 2951 const Instruction *CxtI, 2952 const DominatorTree *DT) { 2953 // Multiplying n * m significant bits yields a result of n + m significant 2954 // bits. If the total number of significant bits does not exceed the 2955 // result bit width (minus 1), there is no overflow. 2956 // This means if we have enough leading zero bits in the operands 2957 // we can guarantee that the result does not overflow. 2958 // Ref: "Hacker's Delight" by Henry Warren 2959 unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); 2960 APInt LHSKnownZero(BitWidth, 0); 2961 APInt LHSKnownOne(BitWidth, 0); 2962 APInt RHSKnownZero(BitWidth, 0); 2963 APInt RHSKnownOne(BitWidth, 0); 2964 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI, 2965 DT); 2966 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI, 2967 DT); 2968 // Note that underestimating the number of zero bits gives a more 2969 // conservative answer. 2970 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() + 2971 RHSKnownZero.countLeadingOnes(); 2972 // First handle the easy case: if we have enough zero bits there's 2973 // definitely no overflow. 2974 if (ZeroBits >= BitWidth) 2975 return OverflowResult::NeverOverflows; 2976 2977 // Get the largest possible values for each operand. 2978 APInt LHSMax = ~LHSKnownZero; 2979 APInt RHSMax = ~RHSKnownZero; 2980 2981 // We know the multiply operation doesn't overflow if the maximum values for 2982 // each operand will not overflow after we multiply them together. 2983 bool MaxOverflow; 2984 LHSMax.umul_ov(RHSMax, MaxOverflow); 2985 if (!MaxOverflow) 2986 return OverflowResult::NeverOverflows; 2987 2988 // We know it always overflows if multiplying the smallest possible values for 2989 // the operands also results in overflow. 2990 bool MinOverflow; 2991 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow); 2992 if (MinOverflow) 2993 return OverflowResult::AlwaysOverflows; 2994 2995 return OverflowResult::MayOverflow; 2996} 2997 2998OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS, 2999 const DataLayout &DL, 3000 AssumptionCache *AC, 3001 const Instruction *CxtI, 3002 const DominatorTree *DT) { 3003 bool LHSKnownNonNegative, LHSKnownNegative; 3004 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0, 3005 AC, CxtI, DT); 3006 if (LHSKnownNonNegative || LHSKnownNegative) { 3007 bool RHSKnownNonNegative, RHSKnownNegative; 3008 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0, 3009 AC, CxtI, DT); 3010 3011 if (LHSKnownNegative && RHSKnownNegative) { 3012 // The sign bit is set in both cases: this MUST overflow. 3013 // Create a simple add instruction, and insert it into the struct. 3014 return OverflowResult::AlwaysOverflows; 3015 } 3016 3017 if (LHSKnownNonNegative && RHSKnownNonNegative) { 3018 // The sign bit is clear in both cases: this CANNOT overflow. 3019 // Create a simple add instruction, and insert it into the struct. 3020 return OverflowResult::NeverOverflows; 3021 } 3022 } 3023 3024 return OverflowResult::MayOverflow; 3025} 3026