1 //===-- ConstraintElimination.cpp - Eliminate conds using constraints. ----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // Eliminate conditions based on constraints collected from dominating 10 // conditions. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Scalar/ConstraintElimination.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/ScopeExit.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/ADT/Statistic.h" 19 #include "llvm/Analysis/ConstraintSystem.h" 20 #include "llvm/Analysis/GlobalsModRef.h" 21 #include "llvm/Analysis/LoopInfo.h" 22 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 23 #include "llvm/Analysis/ScalarEvolution.h" 24 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 25 #include "llvm/Analysis/ValueTracking.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/Dominators.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/IR/IRBuilder.h" 30 #include "llvm/IR/InstrTypes.h" 31 #include "llvm/IR/Instructions.h" 32 #include "llvm/IR/PatternMatch.h" 33 #include "llvm/IR/Verifier.h" 34 #include "llvm/Pass.h" 35 #include "llvm/Support/CommandLine.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/DebugCounter.h" 38 #include "llvm/Support/MathExtras.h" 39 #include "llvm/Transforms/Utils/Cloning.h" 40 #include "llvm/Transforms/Utils/ValueMapper.h" 41 42 #include <cmath> 43 #include <optional> 44 #include <string> 45 46 using namespace llvm; 47 using namespace PatternMatch; 48 49 #define DEBUG_TYPE "constraint-elimination" 50 51 STATISTIC(NumCondsRemoved, "Number of instructions removed"); 52 DEBUG_COUNTER(EliminatedCounter, "conds-eliminated", 53 "Controls which conditions are eliminated"); 54 55 static cl::opt<unsigned> 56 MaxRows("constraint-elimination-max-rows", cl::init(500), cl::Hidden, 57 cl::desc("Maximum number of rows to keep in constraint system")); 58 59 static cl::opt<bool> DumpReproducers( 60 "constraint-elimination-dump-reproducers", cl::init(false), cl::Hidden, 61 cl::desc("Dump IR to reproduce successful transformations.")); 62 63 static int64_t MaxConstraintValue = std::numeric_limits<int64_t>::max(); 64 static int64_t MinSignedConstraintValue = std::numeric_limits<int64_t>::min(); 65 66 // A helper to multiply 2 signed integers where overflowing is allowed. 67 static int64_t multiplyWithOverflow(int64_t A, int64_t B) { 68 int64_t Result; 69 MulOverflow(A, B, Result); 70 return Result; 71 } 72 73 // A helper to add 2 signed integers where overflowing is allowed. 74 static int64_t addWithOverflow(int64_t A, int64_t B) { 75 int64_t Result; 76 AddOverflow(A, B, Result); 77 return Result; 78 } 79 80 static Instruction *getContextInstForUse(Use &U) { 81 Instruction *UserI = cast<Instruction>(U.getUser()); 82 if (auto *Phi = dyn_cast<PHINode>(UserI)) 83 UserI = Phi->getIncomingBlock(U)->getTerminator(); 84 return UserI; 85 } 86 87 namespace { 88 /// Struct to express a condition of the form %Op0 Pred %Op1. 89 struct ConditionTy { 90 CmpInst::Predicate Pred; 91 Value *Op0; 92 Value *Op1; 93 94 ConditionTy() 95 : Pred(CmpInst::BAD_ICMP_PREDICATE), Op0(nullptr), Op1(nullptr) {} 96 ConditionTy(CmpInst::Predicate Pred, Value *Op0, Value *Op1) 97 : Pred(Pred), Op0(Op0), Op1(Op1) {} 98 }; 99 100 /// Represents either 101 /// * a condition that holds on entry to a block (=condition fact) 102 /// * an assume (=assume fact) 103 /// * a use of a compare instruction to simplify. 104 /// It also tracks the Dominator DFS in and out numbers for each entry. 105 struct FactOrCheck { 106 enum class EntryTy { 107 ConditionFact, /// A condition that holds on entry to a block. 108 InstFact, /// A fact that holds after Inst executed (e.g. an assume or 109 /// min/mix intrinsic. 110 InstCheck, /// An instruction to simplify (e.g. an overflow math 111 /// intrinsics). 112 UseCheck /// An use of a compare instruction to simplify. 113 }; 114 115 union { 116 Instruction *Inst; 117 Use *U; 118 ConditionTy Cond; 119 }; 120 121 /// A pre-condition that must hold for the current fact to be added to the 122 /// system. 123 ConditionTy DoesHold; 124 125 unsigned NumIn; 126 unsigned NumOut; 127 EntryTy Ty; 128 129 FactOrCheck(EntryTy Ty, DomTreeNode *DTN, Instruction *Inst) 130 : Inst(Inst), NumIn(DTN->getDFSNumIn()), NumOut(DTN->getDFSNumOut()), 131 Ty(Ty) {} 132 133 FactOrCheck(DomTreeNode *DTN, Use *U) 134 : U(U), DoesHold(CmpInst::BAD_ICMP_PREDICATE, nullptr, nullptr), 135 NumIn(DTN->getDFSNumIn()), NumOut(DTN->getDFSNumOut()), 136 Ty(EntryTy::UseCheck) {} 137 138 FactOrCheck(DomTreeNode *DTN, CmpInst::Predicate Pred, Value *Op0, Value *Op1, 139 ConditionTy Precond = ConditionTy()) 140 : Cond(Pred, Op0, Op1), DoesHold(Precond), NumIn(DTN->getDFSNumIn()), 141 NumOut(DTN->getDFSNumOut()), Ty(EntryTy::ConditionFact) {} 142 143 static FactOrCheck getConditionFact(DomTreeNode *DTN, CmpInst::Predicate Pred, 144 Value *Op0, Value *Op1, 145 ConditionTy Precond = ConditionTy()) { 146 return FactOrCheck(DTN, Pred, Op0, Op1, Precond); 147 } 148 149 static FactOrCheck getInstFact(DomTreeNode *DTN, Instruction *Inst) { 150 return FactOrCheck(EntryTy::InstFact, DTN, Inst); 151 } 152 153 static FactOrCheck getCheck(DomTreeNode *DTN, Use *U) { 154 return FactOrCheck(DTN, U); 155 } 156 157 static FactOrCheck getCheck(DomTreeNode *DTN, CallInst *CI) { 158 return FactOrCheck(EntryTy::InstCheck, DTN, CI); 159 } 160 161 bool isCheck() const { 162 return Ty == EntryTy::InstCheck || Ty == EntryTy::UseCheck; 163 } 164 165 Instruction *getContextInst() const { 166 if (Ty == EntryTy::UseCheck) 167 return getContextInstForUse(*U); 168 return Inst; 169 } 170 171 Instruction *getInstructionToSimplify() const { 172 assert(isCheck()); 173 if (Ty == EntryTy::InstCheck) 174 return Inst; 175 // The use may have been simplified to a constant already. 176 return dyn_cast<Instruction>(*U); 177 } 178 179 bool isConditionFact() const { return Ty == EntryTy::ConditionFact; } 180 }; 181 182 /// Keep state required to build worklist. 183 struct State { 184 DominatorTree &DT; 185 LoopInfo &LI; 186 ScalarEvolution &SE; 187 SmallVector<FactOrCheck, 64> WorkList; 188 189 State(DominatorTree &DT, LoopInfo &LI, ScalarEvolution &SE) 190 : DT(DT), LI(LI), SE(SE) {} 191 192 /// Process block \p BB and add known facts to work-list. 193 void addInfoFor(BasicBlock &BB); 194 195 /// Try to add facts for loop inductions (AddRecs) in EQ/NE compares 196 /// controlling the loop header. 197 void addInfoForInductions(BasicBlock &BB); 198 199 /// Returns true if we can add a known condition from BB to its successor 200 /// block Succ. 201 bool canAddSuccessor(BasicBlock &BB, BasicBlock *Succ) const { 202 return DT.dominates(BasicBlockEdge(&BB, Succ), Succ); 203 } 204 }; 205 206 class ConstraintInfo; 207 208 struct StackEntry { 209 unsigned NumIn; 210 unsigned NumOut; 211 bool IsSigned = false; 212 /// Variables that can be removed from the system once the stack entry gets 213 /// removed. 214 SmallVector<Value *, 2> ValuesToRelease; 215 216 StackEntry(unsigned NumIn, unsigned NumOut, bool IsSigned, 217 SmallVector<Value *, 2> ValuesToRelease) 218 : NumIn(NumIn), NumOut(NumOut), IsSigned(IsSigned), 219 ValuesToRelease(ValuesToRelease) {} 220 }; 221 222 struct ConstraintTy { 223 SmallVector<int64_t, 8> Coefficients; 224 SmallVector<ConditionTy, 2> Preconditions; 225 226 SmallVector<SmallVector<int64_t, 8>> ExtraInfo; 227 228 bool IsSigned = false; 229 230 ConstraintTy() = default; 231 232 ConstraintTy(SmallVector<int64_t, 8> Coefficients, bool IsSigned, bool IsEq, 233 bool IsNe) 234 : Coefficients(Coefficients), IsSigned(IsSigned), IsEq(IsEq), IsNe(IsNe) { 235 } 236 237 unsigned size() const { return Coefficients.size(); } 238 239 unsigned empty() const { return Coefficients.empty(); } 240 241 /// Returns true if all preconditions for this list of constraints are 242 /// satisfied given \p CS and the corresponding \p Value2Index mapping. 243 bool isValid(const ConstraintInfo &Info) const; 244 245 bool isEq() const { return IsEq; } 246 247 bool isNe() const { return IsNe; } 248 249 /// Check if the current constraint is implied by the given ConstraintSystem. 250 /// 251 /// \return true or false if the constraint is proven to be respectively true, 252 /// or false. When the constraint cannot be proven to be either true or false, 253 /// std::nullopt is returned. 254 std::optional<bool> isImpliedBy(const ConstraintSystem &CS) const; 255 256 private: 257 bool IsEq = false; 258 bool IsNe = false; 259 }; 260 261 /// Wrapper encapsulating separate constraint systems and corresponding value 262 /// mappings for both unsigned and signed information. Facts are added to and 263 /// conditions are checked against the corresponding system depending on the 264 /// signed-ness of their predicates. While the information is kept separate 265 /// based on signed-ness, certain conditions can be transferred between the two 266 /// systems. 267 class ConstraintInfo { 268 269 ConstraintSystem UnsignedCS; 270 ConstraintSystem SignedCS; 271 272 const DataLayout &DL; 273 274 public: 275 ConstraintInfo(const DataLayout &DL, ArrayRef<Value *> FunctionArgs) 276 : UnsignedCS(FunctionArgs), SignedCS(FunctionArgs), DL(DL) { 277 auto &Value2Index = getValue2Index(false); 278 // Add Arg > -1 constraints to unsigned system for all function arguments. 279 for (Value *Arg : FunctionArgs) { 280 ConstraintTy VarPos(SmallVector<int64_t, 8>(Value2Index.size() + 1, 0), 281 false, false, false); 282 VarPos.Coefficients[Value2Index[Arg]] = -1; 283 UnsignedCS.addVariableRow(VarPos.Coefficients); 284 } 285 } 286 287 DenseMap<Value *, unsigned> &getValue2Index(bool Signed) { 288 return Signed ? SignedCS.getValue2Index() : UnsignedCS.getValue2Index(); 289 } 290 const DenseMap<Value *, unsigned> &getValue2Index(bool Signed) const { 291 return Signed ? SignedCS.getValue2Index() : UnsignedCS.getValue2Index(); 292 } 293 294 ConstraintSystem &getCS(bool Signed) { 295 return Signed ? SignedCS : UnsignedCS; 296 } 297 const ConstraintSystem &getCS(bool Signed) const { 298 return Signed ? SignedCS : UnsignedCS; 299 } 300 301 void popLastConstraint(bool Signed) { getCS(Signed).popLastConstraint(); } 302 void popLastNVariables(bool Signed, unsigned N) { 303 getCS(Signed).popLastNVariables(N); 304 } 305 306 bool doesHold(CmpInst::Predicate Pred, Value *A, Value *B) const; 307 308 void addFact(CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn, 309 unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack); 310 311 /// Turn a comparison of the form \p Op0 \p Pred \p Op1 into a vector of 312 /// constraints, using indices from the corresponding constraint system. 313 /// New variables that need to be added to the system are collected in 314 /// \p NewVariables. 315 ConstraintTy getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1, 316 SmallVectorImpl<Value *> &NewVariables) const; 317 318 /// Turns a comparison of the form \p Op0 \p Pred \p Op1 into a vector of 319 /// constraints using getConstraint. Returns an empty constraint if the result 320 /// cannot be used to query the existing constraint system, e.g. because it 321 /// would require adding new variables. Also tries to convert signed 322 /// predicates to unsigned ones if possible to allow using the unsigned system 323 /// which increases the effectiveness of the signed <-> unsigned transfer 324 /// logic. 325 ConstraintTy getConstraintForSolving(CmpInst::Predicate Pred, Value *Op0, 326 Value *Op1) const; 327 328 /// Try to add information from \p A \p Pred \p B to the unsigned/signed 329 /// system if \p Pred is signed/unsigned. 330 void transferToOtherSystem(CmpInst::Predicate Pred, Value *A, Value *B, 331 unsigned NumIn, unsigned NumOut, 332 SmallVectorImpl<StackEntry> &DFSInStack); 333 }; 334 335 /// Represents a (Coefficient * Variable) entry after IR decomposition. 336 struct DecompEntry { 337 int64_t Coefficient; 338 Value *Variable; 339 /// True if the variable is known positive in the current constraint. 340 bool IsKnownNonNegative; 341 342 DecompEntry(int64_t Coefficient, Value *Variable, 343 bool IsKnownNonNegative = false) 344 : Coefficient(Coefficient), Variable(Variable), 345 IsKnownNonNegative(IsKnownNonNegative) {} 346 }; 347 348 /// Represents an Offset + Coefficient1 * Variable1 + ... decomposition. 349 struct Decomposition { 350 int64_t Offset = 0; 351 SmallVector<DecompEntry, 3> Vars; 352 353 Decomposition(int64_t Offset) : Offset(Offset) {} 354 Decomposition(Value *V, bool IsKnownNonNegative = false) { 355 Vars.emplace_back(1, V, IsKnownNonNegative); 356 } 357 Decomposition(int64_t Offset, ArrayRef<DecompEntry> Vars) 358 : Offset(Offset), Vars(Vars) {} 359 360 void add(int64_t OtherOffset) { 361 Offset = addWithOverflow(Offset, OtherOffset); 362 } 363 364 void add(const Decomposition &Other) { 365 add(Other.Offset); 366 append_range(Vars, Other.Vars); 367 } 368 369 void sub(const Decomposition &Other) { 370 Decomposition Tmp = Other; 371 Tmp.mul(-1); 372 add(Tmp.Offset); 373 append_range(Vars, Tmp.Vars); 374 } 375 376 void mul(int64_t Factor) { 377 Offset = multiplyWithOverflow(Offset, Factor); 378 for (auto &Var : Vars) 379 Var.Coefficient = multiplyWithOverflow(Var.Coefficient, Factor); 380 } 381 }; 382 383 // Variable and constant offsets for a chain of GEPs, with base pointer BasePtr. 384 struct OffsetResult { 385 Value *BasePtr; 386 APInt ConstantOffset; 387 MapVector<Value *, APInt> VariableOffsets; 388 bool AllInbounds; 389 390 OffsetResult() : BasePtr(nullptr), ConstantOffset(0, uint64_t(0)) {} 391 392 OffsetResult(GEPOperator &GEP, const DataLayout &DL) 393 : BasePtr(GEP.getPointerOperand()), AllInbounds(GEP.isInBounds()) { 394 ConstantOffset = APInt(DL.getIndexTypeSizeInBits(BasePtr->getType()), 0); 395 } 396 }; 397 } // namespace 398 399 // Try to collect variable and constant offsets for \p GEP, partly traversing 400 // nested GEPs. Returns an OffsetResult with nullptr as BasePtr of collecting 401 // the offset fails. 402 static OffsetResult collectOffsets(GEPOperator &GEP, const DataLayout &DL) { 403 OffsetResult Result(GEP, DL); 404 unsigned BitWidth = Result.ConstantOffset.getBitWidth(); 405 if (!GEP.collectOffset(DL, BitWidth, Result.VariableOffsets, 406 Result.ConstantOffset)) 407 return {}; 408 409 // If we have a nested GEP, check if we can combine the constant offset of the 410 // inner GEP with the outer GEP. 411 if (auto *InnerGEP = dyn_cast<GetElementPtrInst>(Result.BasePtr)) { 412 MapVector<Value *, APInt> VariableOffsets2; 413 APInt ConstantOffset2(BitWidth, 0); 414 bool CanCollectInner = InnerGEP->collectOffset( 415 DL, BitWidth, VariableOffsets2, ConstantOffset2); 416 // TODO: Support cases with more than 1 variable offset. 417 if (!CanCollectInner || Result.VariableOffsets.size() > 1 || 418 VariableOffsets2.size() > 1 || 419 (Result.VariableOffsets.size() >= 1 && VariableOffsets2.size() >= 1)) { 420 // More than 1 variable index, use outer result. 421 return Result; 422 } 423 Result.BasePtr = InnerGEP->getPointerOperand(); 424 Result.ConstantOffset += ConstantOffset2; 425 if (Result.VariableOffsets.size() == 0 && VariableOffsets2.size() == 1) 426 Result.VariableOffsets = VariableOffsets2; 427 Result.AllInbounds &= InnerGEP->isInBounds(); 428 } 429 return Result; 430 } 431 432 static Decomposition decompose(Value *V, 433 SmallVectorImpl<ConditionTy> &Preconditions, 434 bool IsSigned, const DataLayout &DL); 435 436 static bool canUseSExt(ConstantInt *CI) { 437 const APInt &Val = CI->getValue(); 438 return Val.sgt(MinSignedConstraintValue) && Val.slt(MaxConstraintValue); 439 } 440 441 static Decomposition decomposeGEP(GEPOperator &GEP, 442 SmallVectorImpl<ConditionTy> &Preconditions, 443 bool IsSigned, const DataLayout &DL) { 444 // Do not reason about pointers where the index size is larger than 64 bits, 445 // as the coefficients used to encode constraints are 64 bit integers. 446 if (DL.getIndexTypeSizeInBits(GEP.getPointerOperand()->getType()) > 64) 447 return &GEP; 448 449 assert(!IsSigned && "The logic below only supports decomposition for " 450 "unsigned predicates at the moment."); 451 const auto &[BasePtr, ConstantOffset, VariableOffsets, AllInbounds] = 452 collectOffsets(GEP, DL); 453 if (!BasePtr || !AllInbounds) 454 return &GEP; 455 456 Decomposition Result(ConstantOffset.getSExtValue(), DecompEntry(1, BasePtr)); 457 for (auto [Index, Scale] : VariableOffsets) { 458 auto IdxResult = decompose(Index, Preconditions, IsSigned, DL); 459 IdxResult.mul(Scale.getSExtValue()); 460 Result.add(IdxResult); 461 462 // If Op0 is signed non-negative, the GEP is increasing monotonically and 463 // can be de-composed. 464 if (!isKnownNonNegative(Index, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1)) 465 Preconditions.emplace_back(CmpInst::ICMP_SGE, Index, 466 ConstantInt::get(Index->getType(), 0)); 467 } 468 return Result; 469 } 470 471 // Decomposes \p V into a constant offset + list of pairs { Coefficient, 472 // Variable } where Coefficient * Variable. The sum of the constant offset and 473 // pairs equals \p V. 474 static Decomposition decompose(Value *V, 475 SmallVectorImpl<ConditionTy> &Preconditions, 476 bool IsSigned, const DataLayout &DL) { 477 478 auto MergeResults = [&Preconditions, IsSigned, &DL](Value *A, Value *B, 479 bool IsSignedB) { 480 auto ResA = decompose(A, Preconditions, IsSigned, DL); 481 auto ResB = decompose(B, Preconditions, IsSignedB, DL); 482 ResA.add(ResB); 483 return ResA; 484 }; 485 486 Type *Ty = V->getType()->getScalarType(); 487 if (Ty->isPointerTy() && !IsSigned) { 488 if (auto *GEP = dyn_cast<GEPOperator>(V)) 489 return decomposeGEP(*GEP, Preconditions, IsSigned, DL); 490 if (isa<ConstantPointerNull>(V)) 491 return int64_t(0); 492 493 return V; 494 } 495 496 // Don't handle integers > 64 bit. Our coefficients are 64-bit large, so 497 // coefficient add/mul may wrap, while the operation in the full bit width 498 // would not. 499 if (!Ty->isIntegerTy() || Ty->getIntegerBitWidth() > 64) 500 return V; 501 502 // Decompose \p V used with a signed predicate. 503 if (IsSigned) { 504 if (auto *CI = dyn_cast<ConstantInt>(V)) { 505 if (canUseSExt(CI)) 506 return CI->getSExtValue(); 507 } 508 Value *Op0; 509 Value *Op1; 510 if (match(V, m_NSWAdd(m_Value(Op0), m_Value(Op1)))) 511 return MergeResults(Op0, Op1, IsSigned); 512 513 ConstantInt *CI; 514 if (match(V, m_NSWMul(m_Value(Op0), m_ConstantInt(CI))) && canUseSExt(CI)) { 515 auto Result = decompose(Op0, Preconditions, IsSigned, DL); 516 Result.mul(CI->getSExtValue()); 517 return Result; 518 } 519 520 // (shl nsw x, shift) is (mul nsw x, (1<<shift)), with the exception of 521 // shift == bw-1. 522 if (match(V, m_NSWShl(m_Value(Op0), m_ConstantInt(CI)))) { 523 uint64_t Shift = CI->getValue().getLimitedValue(); 524 if (Shift < Ty->getIntegerBitWidth() - 1) { 525 assert(Shift < 64 && "Would overflow"); 526 auto Result = decompose(Op0, Preconditions, IsSigned, DL); 527 Result.mul(int64_t(1) << Shift); 528 return Result; 529 } 530 } 531 532 return V; 533 } 534 535 if (auto *CI = dyn_cast<ConstantInt>(V)) { 536 if (CI->uge(MaxConstraintValue)) 537 return V; 538 return int64_t(CI->getZExtValue()); 539 } 540 541 Value *Op0; 542 bool IsKnownNonNegative = false; 543 if (match(V, m_ZExt(m_Value(Op0)))) { 544 IsKnownNonNegative = true; 545 V = Op0; 546 } 547 548 Value *Op1; 549 ConstantInt *CI; 550 if (match(V, m_NUWAdd(m_Value(Op0), m_Value(Op1)))) { 551 return MergeResults(Op0, Op1, IsSigned); 552 } 553 if (match(V, m_NSWAdd(m_Value(Op0), m_Value(Op1)))) { 554 if (!isKnownNonNegative(Op0, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1)) 555 Preconditions.emplace_back(CmpInst::ICMP_SGE, Op0, 556 ConstantInt::get(Op0->getType(), 0)); 557 if (!isKnownNonNegative(Op1, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1)) 558 Preconditions.emplace_back(CmpInst::ICMP_SGE, Op1, 559 ConstantInt::get(Op1->getType(), 0)); 560 561 return MergeResults(Op0, Op1, IsSigned); 562 } 563 564 if (match(V, m_Add(m_Value(Op0), m_ConstantInt(CI))) && CI->isNegative() && 565 canUseSExt(CI)) { 566 Preconditions.emplace_back( 567 CmpInst::ICMP_UGE, Op0, 568 ConstantInt::get(Op0->getType(), CI->getSExtValue() * -1)); 569 return MergeResults(Op0, CI, true); 570 } 571 572 // Decompose or as an add if there are no common bits between the operands. 573 if (match(V, m_DisjointOr(m_Value(Op0), m_ConstantInt(CI)))) 574 return MergeResults(Op0, CI, IsSigned); 575 576 if (match(V, m_NUWShl(m_Value(Op1), m_ConstantInt(CI))) && canUseSExt(CI)) { 577 if (CI->getSExtValue() < 0 || CI->getSExtValue() >= 64) 578 return {V, IsKnownNonNegative}; 579 auto Result = decompose(Op1, Preconditions, IsSigned, DL); 580 Result.mul(int64_t{1} << CI->getSExtValue()); 581 return Result; 582 } 583 584 if (match(V, m_NUWMul(m_Value(Op1), m_ConstantInt(CI))) && canUseSExt(CI) && 585 (!CI->isNegative())) { 586 auto Result = decompose(Op1, Preconditions, IsSigned, DL); 587 Result.mul(CI->getSExtValue()); 588 return Result; 589 } 590 591 if (match(V, m_NUWSub(m_Value(Op0), m_Value(Op1)))) { 592 auto ResA = decompose(Op0, Preconditions, IsSigned, DL); 593 auto ResB = decompose(Op1, Preconditions, IsSigned, DL); 594 ResA.sub(ResB); 595 return ResA; 596 } 597 598 return {V, IsKnownNonNegative}; 599 } 600 601 ConstraintTy 602 ConstraintInfo::getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1, 603 SmallVectorImpl<Value *> &NewVariables) const { 604 assert(NewVariables.empty() && "NewVariables must be empty when passed in"); 605 bool IsEq = false; 606 bool IsNe = false; 607 608 // Try to convert Pred to one of ULE/SLT/SLE/SLT. 609 switch (Pred) { 610 case CmpInst::ICMP_UGT: 611 case CmpInst::ICMP_UGE: 612 case CmpInst::ICMP_SGT: 613 case CmpInst::ICMP_SGE: { 614 Pred = CmpInst::getSwappedPredicate(Pred); 615 std::swap(Op0, Op1); 616 break; 617 } 618 case CmpInst::ICMP_EQ: 619 if (match(Op1, m_Zero())) { 620 Pred = CmpInst::ICMP_ULE; 621 } else { 622 IsEq = true; 623 Pred = CmpInst::ICMP_ULE; 624 } 625 break; 626 case CmpInst::ICMP_NE: 627 if (match(Op1, m_Zero())) { 628 Pred = CmpInst::getSwappedPredicate(CmpInst::ICMP_UGT); 629 std::swap(Op0, Op1); 630 } else { 631 IsNe = true; 632 Pred = CmpInst::ICMP_ULE; 633 } 634 break; 635 default: 636 break; 637 } 638 639 if (Pred != CmpInst::ICMP_ULE && Pred != CmpInst::ICMP_ULT && 640 Pred != CmpInst::ICMP_SLE && Pred != CmpInst::ICMP_SLT) 641 return {}; 642 643 SmallVector<ConditionTy, 4> Preconditions; 644 bool IsSigned = CmpInst::isSigned(Pred); 645 auto &Value2Index = getValue2Index(IsSigned); 646 auto ADec = decompose(Op0->stripPointerCastsSameRepresentation(), 647 Preconditions, IsSigned, DL); 648 auto BDec = decompose(Op1->stripPointerCastsSameRepresentation(), 649 Preconditions, IsSigned, DL); 650 int64_t Offset1 = ADec.Offset; 651 int64_t Offset2 = BDec.Offset; 652 Offset1 *= -1; 653 654 auto &VariablesA = ADec.Vars; 655 auto &VariablesB = BDec.Vars; 656 657 // First try to look up \p V in Value2Index and NewVariables. Otherwise add a 658 // new entry to NewVariables. 659 SmallDenseMap<Value *, unsigned> NewIndexMap; 660 auto GetOrAddIndex = [&Value2Index, &NewVariables, 661 &NewIndexMap](Value *V) -> unsigned { 662 auto V2I = Value2Index.find(V); 663 if (V2I != Value2Index.end()) 664 return V2I->second; 665 auto Insert = 666 NewIndexMap.insert({V, Value2Index.size() + NewVariables.size() + 1}); 667 if (Insert.second) 668 NewVariables.push_back(V); 669 return Insert.first->second; 670 }; 671 672 // Make sure all variables have entries in Value2Index or NewVariables. 673 for (const auto &KV : concat<DecompEntry>(VariablesA, VariablesB)) 674 GetOrAddIndex(KV.Variable); 675 676 // Build result constraint, by first adding all coefficients from A and then 677 // subtracting all coefficients from B. 678 ConstraintTy Res( 679 SmallVector<int64_t, 8>(Value2Index.size() + NewVariables.size() + 1, 0), 680 IsSigned, IsEq, IsNe); 681 // Collect variables that are known to be positive in all uses in the 682 // constraint. 683 SmallDenseMap<Value *, bool> KnownNonNegativeVariables; 684 auto &R = Res.Coefficients; 685 for (const auto &KV : VariablesA) { 686 R[GetOrAddIndex(KV.Variable)] += KV.Coefficient; 687 auto I = 688 KnownNonNegativeVariables.insert({KV.Variable, KV.IsKnownNonNegative}); 689 I.first->second &= KV.IsKnownNonNegative; 690 } 691 692 for (const auto &KV : VariablesB) { 693 if (SubOverflow(R[GetOrAddIndex(KV.Variable)], KV.Coefficient, 694 R[GetOrAddIndex(KV.Variable)])) 695 return {}; 696 auto I = 697 KnownNonNegativeVariables.insert({KV.Variable, KV.IsKnownNonNegative}); 698 I.first->second &= KV.IsKnownNonNegative; 699 } 700 701 int64_t OffsetSum; 702 if (AddOverflow(Offset1, Offset2, OffsetSum)) 703 return {}; 704 if (Pred == (IsSigned ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT)) 705 if (AddOverflow(OffsetSum, int64_t(-1), OffsetSum)) 706 return {}; 707 R[0] = OffsetSum; 708 Res.Preconditions = std::move(Preconditions); 709 710 // Remove any (Coefficient, Variable) entry where the Coefficient is 0 for new 711 // variables. 712 while (!NewVariables.empty()) { 713 int64_t Last = R.back(); 714 if (Last != 0) 715 break; 716 R.pop_back(); 717 Value *RemovedV = NewVariables.pop_back_val(); 718 NewIndexMap.erase(RemovedV); 719 } 720 721 // Add extra constraints for variables that are known positive. 722 for (auto &KV : KnownNonNegativeVariables) { 723 if (!KV.second || 724 (!Value2Index.contains(KV.first) && !NewIndexMap.contains(KV.first))) 725 continue; 726 SmallVector<int64_t, 8> C(Value2Index.size() + NewVariables.size() + 1, 0); 727 C[GetOrAddIndex(KV.first)] = -1; 728 Res.ExtraInfo.push_back(C); 729 } 730 return Res; 731 } 732 733 ConstraintTy ConstraintInfo::getConstraintForSolving(CmpInst::Predicate Pred, 734 Value *Op0, 735 Value *Op1) const { 736 Constant *NullC = Constant::getNullValue(Op0->getType()); 737 // Handle trivially true compares directly to avoid adding V UGE 0 constraints 738 // for all variables in the unsigned system. 739 if ((Pred == CmpInst::ICMP_ULE && Op0 == NullC) || 740 (Pred == CmpInst::ICMP_UGE && Op1 == NullC)) { 741 auto &Value2Index = getValue2Index(false); 742 // Return constraint that's trivially true. 743 return ConstraintTy(SmallVector<int64_t, 8>(Value2Index.size(), 0), false, 744 false, false); 745 } 746 747 // If both operands are known to be non-negative, change signed predicates to 748 // unsigned ones. This increases the reasoning effectiveness in combination 749 // with the signed <-> unsigned transfer logic. 750 if (CmpInst::isSigned(Pred) && 751 isKnownNonNegative(Op0, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1) && 752 isKnownNonNegative(Op1, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1)) 753 Pred = CmpInst::getUnsignedPredicate(Pred); 754 755 SmallVector<Value *> NewVariables; 756 ConstraintTy R = getConstraint(Pred, Op0, Op1, NewVariables); 757 if (!NewVariables.empty()) 758 return {}; 759 return R; 760 } 761 762 bool ConstraintTy::isValid(const ConstraintInfo &Info) const { 763 return Coefficients.size() > 0 && 764 all_of(Preconditions, [&Info](const ConditionTy &C) { 765 return Info.doesHold(C.Pred, C.Op0, C.Op1); 766 }); 767 } 768 769 std::optional<bool> 770 ConstraintTy::isImpliedBy(const ConstraintSystem &CS) const { 771 bool IsConditionImplied = CS.isConditionImplied(Coefficients); 772 773 if (IsEq || IsNe) { 774 auto NegatedOrEqual = ConstraintSystem::negateOrEqual(Coefficients); 775 bool IsNegatedOrEqualImplied = 776 !NegatedOrEqual.empty() && CS.isConditionImplied(NegatedOrEqual); 777 778 // In order to check that `%a == %b` is true (equality), both conditions `%a 779 // >= %b` and `%a <= %b` must hold true. When checking for equality (`IsEq` 780 // is true), we return true if they both hold, false in the other cases. 781 if (IsConditionImplied && IsNegatedOrEqualImplied) 782 return IsEq; 783 784 auto Negated = ConstraintSystem::negate(Coefficients); 785 bool IsNegatedImplied = !Negated.empty() && CS.isConditionImplied(Negated); 786 787 auto StrictLessThan = ConstraintSystem::toStrictLessThan(Coefficients); 788 bool IsStrictLessThanImplied = 789 !StrictLessThan.empty() && CS.isConditionImplied(StrictLessThan); 790 791 // In order to check that `%a != %b` is true (non-equality), either 792 // condition `%a > %b` or `%a < %b` must hold true. When checking for 793 // non-equality (`IsNe` is true), we return true if one of the two holds, 794 // false in the other cases. 795 if (IsNegatedImplied || IsStrictLessThanImplied) 796 return IsNe; 797 798 return std::nullopt; 799 } 800 801 if (IsConditionImplied) 802 return true; 803 804 auto Negated = ConstraintSystem::negate(Coefficients); 805 auto IsNegatedImplied = !Negated.empty() && CS.isConditionImplied(Negated); 806 if (IsNegatedImplied) 807 return false; 808 809 // Neither the condition nor its negated holds, did not prove anything. 810 return std::nullopt; 811 } 812 813 bool ConstraintInfo::doesHold(CmpInst::Predicate Pred, Value *A, 814 Value *B) const { 815 auto R = getConstraintForSolving(Pred, A, B); 816 return R.isValid(*this) && 817 getCS(R.IsSigned).isConditionImplied(R.Coefficients); 818 } 819 820 void ConstraintInfo::transferToOtherSystem( 821 CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn, 822 unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack) { 823 auto IsKnownNonNegative = [this](Value *V) { 824 return doesHold(CmpInst::ICMP_SGE, V, ConstantInt::get(V->getType(), 0)) || 825 isKnownNonNegative(V, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1); 826 }; 827 // Check if we can combine facts from the signed and unsigned systems to 828 // derive additional facts. 829 if (!A->getType()->isIntegerTy()) 830 return; 831 // FIXME: This currently depends on the order we add facts. Ideally we 832 // would first add all known facts and only then try to add additional 833 // facts. 834 switch (Pred) { 835 default: 836 break; 837 case CmpInst::ICMP_ULT: 838 case CmpInst::ICMP_ULE: 839 // If B is a signed positive constant, then A >=s 0 and A <s (or <=s) B. 840 if (IsKnownNonNegative(B)) { 841 addFact(CmpInst::ICMP_SGE, A, ConstantInt::get(B->getType(), 0), NumIn, 842 NumOut, DFSInStack); 843 addFact(CmpInst::getSignedPredicate(Pred), A, B, NumIn, NumOut, 844 DFSInStack); 845 } 846 break; 847 case CmpInst::ICMP_UGE: 848 case CmpInst::ICMP_UGT: 849 // If A is a signed positive constant, then B >=s 0 and A >s (or >=s) B. 850 if (IsKnownNonNegative(A)) { 851 addFact(CmpInst::ICMP_SGE, B, ConstantInt::get(B->getType(), 0), NumIn, 852 NumOut, DFSInStack); 853 addFact(CmpInst::getSignedPredicate(Pred), A, B, NumIn, NumOut, 854 DFSInStack); 855 } 856 break; 857 case CmpInst::ICMP_SLT: 858 if (IsKnownNonNegative(A)) 859 addFact(CmpInst::ICMP_ULT, A, B, NumIn, NumOut, DFSInStack); 860 break; 861 case CmpInst::ICMP_SGT: { 862 if (doesHold(CmpInst::ICMP_SGE, B, ConstantInt::get(B->getType(), -1))) 863 addFact(CmpInst::ICMP_UGE, A, ConstantInt::get(B->getType(), 0), NumIn, 864 NumOut, DFSInStack); 865 if (IsKnownNonNegative(B)) 866 addFact(CmpInst::ICMP_UGT, A, B, NumIn, NumOut, DFSInStack); 867 868 break; 869 } 870 case CmpInst::ICMP_SGE: 871 if (IsKnownNonNegative(B)) 872 addFact(CmpInst::ICMP_UGE, A, B, NumIn, NumOut, DFSInStack); 873 break; 874 } 875 } 876 877 #ifndef NDEBUG 878 879 static void dumpConstraint(ArrayRef<int64_t> C, 880 const DenseMap<Value *, unsigned> &Value2Index) { 881 ConstraintSystem CS(Value2Index); 882 CS.addVariableRowFill(C); 883 CS.dump(); 884 } 885 #endif 886 887 void State::addInfoForInductions(BasicBlock &BB) { 888 auto *L = LI.getLoopFor(&BB); 889 if (!L || L->getHeader() != &BB) 890 return; 891 892 Value *A; 893 Value *B; 894 CmpInst::Predicate Pred; 895 896 if (!match(BB.getTerminator(), 897 m_Br(m_ICmp(Pred, m_Value(A), m_Value(B)), m_Value(), m_Value()))) 898 return; 899 PHINode *PN = dyn_cast<PHINode>(A); 900 if (!PN) { 901 Pred = CmpInst::getSwappedPredicate(Pred); 902 std::swap(A, B); 903 PN = dyn_cast<PHINode>(A); 904 } 905 906 if (!PN || PN->getParent() != &BB || PN->getNumIncomingValues() != 2 || 907 !SE.isSCEVable(PN->getType())) 908 return; 909 910 BasicBlock *InLoopSucc = nullptr; 911 if (Pred == CmpInst::ICMP_NE) 912 InLoopSucc = cast<BranchInst>(BB.getTerminator())->getSuccessor(0); 913 else if (Pred == CmpInst::ICMP_EQ) 914 InLoopSucc = cast<BranchInst>(BB.getTerminator())->getSuccessor(1); 915 else 916 return; 917 918 if (!L->contains(InLoopSucc) || !L->isLoopExiting(&BB) || InLoopSucc == &BB) 919 return; 920 921 auto *AR = dyn_cast_or_null<SCEVAddRecExpr>(SE.getSCEV(PN)); 922 BasicBlock *LoopPred = L->getLoopPredecessor(); 923 if (!AR || AR->getLoop() != L || !LoopPred) 924 return; 925 926 const SCEV *StartSCEV = AR->getStart(); 927 Value *StartValue = nullptr; 928 if (auto *C = dyn_cast<SCEVConstant>(StartSCEV)) { 929 StartValue = C->getValue(); 930 } else { 931 StartValue = PN->getIncomingValueForBlock(LoopPred); 932 assert(SE.getSCEV(StartValue) == StartSCEV && "inconsistent start value"); 933 } 934 935 DomTreeNode *DTN = DT.getNode(InLoopSucc); 936 auto IncUnsigned = SE.getMonotonicPredicateType(AR, CmpInst::ICMP_UGT); 937 auto IncSigned = SE.getMonotonicPredicateType(AR, CmpInst::ICMP_SGT); 938 bool MonotonicallyIncreasingUnsigned = 939 IncUnsigned && *IncUnsigned == ScalarEvolution::MonotonicallyIncreasing; 940 bool MonotonicallyIncreasingSigned = 941 IncSigned && *IncSigned == ScalarEvolution::MonotonicallyIncreasing; 942 // If SCEV guarantees that AR does not wrap, PN >= StartValue can be added 943 // unconditionally. 944 if (MonotonicallyIncreasingUnsigned) 945 WorkList.push_back( 946 FactOrCheck::getConditionFact(DTN, CmpInst::ICMP_UGE, PN, StartValue)); 947 if (MonotonicallyIncreasingSigned) 948 WorkList.push_back( 949 FactOrCheck::getConditionFact(DTN, CmpInst::ICMP_SGE, PN, StartValue)); 950 951 APInt StepOffset; 952 if (auto *C = dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE))) 953 StepOffset = C->getAPInt(); 954 else 955 return; 956 957 // Make sure the bound B is loop-invariant. 958 if (!L->isLoopInvariant(B)) 959 return; 960 961 // Handle negative steps. 962 if (StepOffset.isNegative()) { 963 // TODO: Extend to allow steps > -1. 964 if (!(-StepOffset).isOne()) 965 return; 966 967 // AR may wrap. 968 // Add StartValue >= PN conditional on B <= StartValue which guarantees that 969 // the loop exits before wrapping with a step of -1. 970 WorkList.push_back(FactOrCheck::getConditionFact( 971 DTN, CmpInst::ICMP_UGE, StartValue, PN, 972 ConditionTy(CmpInst::ICMP_ULE, B, StartValue))); 973 WorkList.push_back(FactOrCheck::getConditionFact( 974 DTN, CmpInst::ICMP_SGE, StartValue, PN, 975 ConditionTy(CmpInst::ICMP_SLE, B, StartValue))); 976 // Add PN > B conditional on B <= StartValue which guarantees that the loop 977 // exits when reaching B with a step of -1. 978 WorkList.push_back(FactOrCheck::getConditionFact( 979 DTN, CmpInst::ICMP_UGT, PN, B, 980 ConditionTy(CmpInst::ICMP_ULE, B, StartValue))); 981 WorkList.push_back(FactOrCheck::getConditionFact( 982 DTN, CmpInst::ICMP_SGT, PN, B, 983 ConditionTy(CmpInst::ICMP_SLE, B, StartValue))); 984 return; 985 } 986 987 // Make sure AR either steps by 1 or that the value we compare against is a 988 // GEP based on the same start value and all offsets are a multiple of the 989 // step size, to guarantee that the induction will reach the value. 990 if (StepOffset.isZero() || StepOffset.isNegative()) 991 return; 992 993 if (!StepOffset.isOne()) { 994 // Check whether B-Start is known to be a multiple of StepOffset. 995 const SCEV *BMinusStart = SE.getMinusSCEV(SE.getSCEV(B), StartSCEV); 996 if (isa<SCEVCouldNotCompute>(BMinusStart) || 997 !SE.getConstantMultiple(BMinusStart).urem(StepOffset).isZero()) 998 return; 999 } 1000 1001 // AR may wrap. Add PN >= StartValue conditional on StartValue <= B which 1002 // guarantees that the loop exits before wrapping in combination with the 1003 // restrictions on B and the step above. 1004 if (!MonotonicallyIncreasingUnsigned) 1005 WorkList.push_back(FactOrCheck::getConditionFact( 1006 DTN, CmpInst::ICMP_UGE, PN, StartValue, 1007 ConditionTy(CmpInst::ICMP_ULE, StartValue, B))); 1008 if (!MonotonicallyIncreasingSigned) 1009 WorkList.push_back(FactOrCheck::getConditionFact( 1010 DTN, CmpInst::ICMP_SGE, PN, StartValue, 1011 ConditionTy(CmpInst::ICMP_SLE, StartValue, B))); 1012 1013 WorkList.push_back(FactOrCheck::getConditionFact( 1014 DTN, CmpInst::ICMP_ULT, PN, B, 1015 ConditionTy(CmpInst::ICMP_ULE, StartValue, B))); 1016 WorkList.push_back(FactOrCheck::getConditionFact( 1017 DTN, CmpInst::ICMP_SLT, PN, B, 1018 ConditionTy(CmpInst::ICMP_SLE, StartValue, B))); 1019 } 1020 1021 void State::addInfoFor(BasicBlock &BB) { 1022 addInfoForInductions(BB); 1023 1024 // True as long as long as the current instruction is guaranteed to execute. 1025 bool GuaranteedToExecute = true; 1026 // Queue conditions and assumes. 1027 for (Instruction &I : BB) { 1028 if (auto Cmp = dyn_cast<ICmpInst>(&I)) { 1029 for (Use &U : Cmp->uses()) { 1030 auto *UserI = getContextInstForUse(U); 1031 auto *DTN = DT.getNode(UserI->getParent()); 1032 if (!DTN) 1033 continue; 1034 WorkList.push_back(FactOrCheck::getCheck(DTN, &U)); 1035 } 1036 continue; 1037 } 1038 1039 auto *II = dyn_cast<IntrinsicInst>(&I); 1040 Intrinsic::ID ID = II ? II->getIntrinsicID() : Intrinsic::not_intrinsic; 1041 switch (ID) { 1042 case Intrinsic::assume: { 1043 Value *A, *B; 1044 CmpInst::Predicate Pred; 1045 if (!match(I.getOperand(0), m_ICmp(Pred, m_Value(A), m_Value(B)))) 1046 break; 1047 if (GuaranteedToExecute) { 1048 // The assume is guaranteed to execute when BB is entered, hence Cond 1049 // holds on entry to BB. 1050 WorkList.emplace_back(FactOrCheck::getConditionFact( 1051 DT.getNode(I.getParent()), Pred, A, B)); 1052 } else { 1053 WorkList.emplace_back( 1054 FactOrCheck::getInstFact(DT.getNode(I.getParent()), &I)); 1055 } 1056 break; 1057 } 1058 // Enqueue ssub_with_overflow for simplification. 1059 case Intrinsic::ssub_with_overflow: 1060 WorkList.push_back( 1061 FactOrCheck::getCheck(DT.getNode(&BB), cast<CallInst>(&I))); 1062 break; 1063 // Enqueue the intrinsics to add extra info. 1064 case Intrinsic::umin: 1065 case Intrinsic::umax: 1066 case Intrinsic::smin: 1067 case Intrinsic::smax: 1068 // TODO: Check if it is possible to instead only added the min/max facts 1069 // when simplifying uses of the min/max intrinsics. 1070 if (!isGuaranteedNotToBePoison(&I)) 1071 break; 1072 [[fallthrough]]; 1073 case Intrinsic::abs: 1074 WorkList.push_back(FactOrCheck::getInstFact(DT.getNode(&BB), &I)); 1075 break; 1076 } 1077 1078 GuaranteedToExecute &= isGuaranteedToTransferExecutionToSuccessor(&I); 1079 } 1080 1081 if (auto *Switch = dyn_cast<SwitchInst>(BB.getTerminator())) { 1082 for (auto &Case : Switch->cases()) { 1083 BasicBlock *Succ = Case.getCaseSuccessor(); 1084 Value *V = Case.getCaseValue(); 1085 if (!canAddSuccessor(BB, Succ)) 1086 continue; 1087 WorkList.emplace_back(FactOrCheck::getConditionFact( 1088 DT.getNode(Succ), CmpInst::ICMP_EQ, Switch->getCondition(), V)); 1089 } 1090 return; 1091 } 1092 1093 auto *Br = dyn_cast<BranchInst>(BB.getTerminator()); 1094 if (!Br || !Br->isConditional()) 1095 return; 1096 1097 Value *Cond = Br->getCondition(); 1098 1099 // If the condition is a chain of ORs/AND and the successor only has the 1100 // current block as predecessor, queue conditions for the successor. 1101 Value *Op0, *Op1; 1102 if (match(Cond, m_LogicalOr(m_Value(Op0), m_Value(Op1))) || 1103 match(Cond, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) { 1104 bool IsOr = match(Cond, m_LogicalOr()); 1105 bool IsAnd = match(Cond, m_LogicalAnd()); 1106 // If there's a select that matches both AND and OR, we need to commit to 1107 // one of the options. Arbitrarily pick OR. 1108 if (IsOr && IsAnd) 1109 IsAnd = false; 1110 1111 BasicBlock *Successor = Br->getSuccessor(IsOr ? 1 : 0); 1112 if (canAddSuccessor(BB, Successor)) { 1113 SmallVector<Value *> CondWorkList; 1114 SmallPtrSet<Value *, 8> SeenCond; 1115 auto QueueValue = [&CondWorkList, &SeenCond](Value *V) { 1116 if (SeenCond.insert(V).second) 1117 CondWorkList.push_back(V); 1118 }; 1119 QueueValue(Op1); 1120 QueueValue(Op0); 1121 while (!CondWorkList.empty()) { 1122 Value *Cur = CondWorkList.pop_back_val(); 1123 if (auto *Cmp = dyn_cast<ICmpInst>(Cur)) { 1124 WorkList.emplace_back(FactOrCheck::getConditionFact( 1125 DT.getNode(Successor), 1126 IsOr ? CmpInst::getInversePredicate(Cmp->getPredicate()) 1127 : Cmp->getPredicate(), 1128 Cmp->getOperand(0), Cmp->getOperand(1))); 1129 continue; 1130 } 1131 if (IsOr && match(Cur, m_LogicalOr(m_Value(Op0), m_Value(Op1)))) { 1132 QueueValue(Op1); 1133 QueueValue(Op0); 1134 continue; 1135 } 1136 if (IsAnd && match(Cur, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) { 1137 QueueValue(Op1); 1138 QueueValue(Op0); 1139 continue; 1140 } 1141 } 1142 } 1143 return; 1144 } 1145 1146 auto *CmpI = dyn_cast<ICmpInst>(Br->getCondition()); 1147 if (!CmpI) 1148 return; 1149 if (canAddSuccessor(BB, Br->getSuccessor(0))) 1150 WorkList.emplace_back(FactOrCheck::getConditionFact( 1151 DT.getNode(Br->getSuccessor(0)), CmpI->getPredicate(), 1152 CmpI->getOperand(0), CmpI->getOperand(1))); 1153 if (canAddSuccessor(BB, Br->getSuccessor(1))) 1154 WorkList.emplace_back(FactOrCheck::getConditionFact( 1155 DT.getNode(Br->getSuccessor(1)), 1156 CmpInst::getInversePredicate(CmpI->getPredicate()), CmpI->getOperand(0), 1157 CmpI->getOperand(1))); 1158 } 1159 1160 #ifndef NDEBUG 1161 static void dumpUnpackedICmp(raw_ostream &OS, ICmpInst::Predicate Pred, 1162 Value *LHS, Value *RHS) { 1163 OS << "icmp " << Pred << ' '; 1164 LHS->printAsOperand(OS, /*PrintType=*/true); 1165 OS << ", "; 1166 RHS->printAsOperand(OS, /*PrintType=*/false); 1167 } 1168 #endif 1169 1170 namespace { 1171 /// Helper to keep track of a condition and if it should be treated as negated 1172 /// for reproducer construction. 1173 /// Pred == Predicate::BAD_ICMP_PREDICATE indicates that this entry is a 1174 /// placeholder to keep the ReproducerCondStack in sync with DFSInStack. 1175 struct ReproducerEntry { 1176 ICmpInst::Predicate Pred; 1177 Value *LHS; 1178 Value *RHS; 1179 1180 ReproducerEntry(ICmpInst::Predicate Pred, Value *LHS, Value *RHS) 1181 : Pred(Pred), LHS(LHS), RHS(RHS) {} 1182 }; 1183 } // namespace 1184 1185 /// Helper function to generate a reproducer function for simplifying \p Cond. 1186 /// The reproducer function contains a series of @llvm.assume calls, one for 1187 /// each condition in \p Stack. For each condition, the operand instruction are 1188 /// cloned until we reach operands that have an entry in \p Value2Index. Those 1189 /// will then be added as function arguments. \p DT is used to order cloned 1190 /// instructions. The reproducer function will get added to \p M, if it is 1191 /// non-null. Otherwise no reproducer function is generated. 1192 static void generateReproducer(CmpInst *Cond, Module *M, 1193 ArrayRef<ReproducerEntry> Stack, 1194 ConstraintInfo &Info, DominatorTree &DT) { 1195 if (!M) 1196 return; 1197 1198 LLVMContext &Ctx = Cond->getContext(); 1199 1200 LLVM_DEBUG(dbgs() << "Creating reproducer for " << *Cond << "\n"); 1201 1202 ValueToValueMapTy Old2New; 1203 SmallVector<Value *> Args; 1204 SmallPtrSet<Value *, 8> Seen; 1205 // Traverse Cond and its operands recursively until we reach a value that's in 1206 // Value2Index or not an instruction, or not a operation that 1207 // ConstraintElimination can decompose. Such values will be considered as 1208 // external inputs to the reproducer, they are collected and added as function 1209 // arguments later. 1210 auto CollectArguments = [&](ArrayRef<Value *> Ops, bool IsSigned) { 1211 auto &Value2Index = Info.getValue2Index(IsSigned); 1212 SmallVector<Value *, 4> WorkList(Ops); 1213 while (!WorkList.empty()) { 1214 Value *V = WorkList.pop_back_val(); 1215 if (!Seen.insert(V).second) 1216 continue; 1217 if (Old2New.find(V) != Old2New.end()) 1218 continue; 1219 if (isa<Constant>(V)) 1220 continue; 1221 1222 auto *I = dyn_cast<Instruction>(V); 1223 if (Value2Index.contains(V) || !I || 1224 !isa<CmpInst, BinaryOperator, GEPOperator, CastInst>(V)) { 1225 Old2New[V] = V; 1226 Args.push_back(V); 1227 LLVM_DEBUG(dbgs() << " found external input " << *V << "\n"); 1228 } else { 1229 append_range(WorkList, I->operands()); 1230 } 1231 } 1232 }; 1233 1234 for (auto &Entry : Stack) 1235 if (Entry.Pred != ICmpInst::BAD_ICMP_PREDICATE) 1236 CollectArguments({Entry.LHS, Entry.RHS}, ICmpInst::isSigned(Entry.Pred)); 1237 CollectArguments(Cond, ICmpInst::isSigned(Cond->getPredicate())); 1238 1239 SmallVector<Type *> ParamTys; 1240 for (auto *P : Args) 1241 ParamTys.push_back(P->getType()); 1242 1243 FunctionType *FTy = FunctionType::get(Cond->getType(), ParamTys, 1244 /*isVarArg=*/false); 1245 Function *F = Function::Create(FTy, Function::ExternalLinkage, 1246 Cond->getModule()->getName() + 1247 Cond->getFunction()->getName() + "repro", 1248 M); 1249 // Add arguments to the reproducer function for each external value collected. 1250 for (unsigned I = 0; I < Args.size(); ++I) { 1251 F->getArg(I)->setName(Args[I]->getName()); 1252 Old2New[Args[I]] = F->getArg(I); 1253 } 1254 1255 BasicBlock *Entry = BasicBlock::Create(Ctx, "entry", F); 1256 IRBuilder<> Builder(Entry); 1257 Builder.CreateRet(Builder.getTrue()); 1258 Builder.SetInsertPoint(Entry->getTerminator()); 1259 1260 // Clone instructions in \p Ops and their operands recursively until reaching 1261 // an value in Value2Index (external input to the reproducer). Update Old2New 1262 // mapping for the original and cloned instructions. Sort instructions to 1263 // clone by dominance, then insert the cloned instructions in the function. 1264 auto CloneInstructions = [&](ArrayRef<Value *> Ops, bool IsSigned) { 1265 SmallVector<Value *, 4> WorkList(Ops); 1266 SmallVector<Instruction *> ToClone; 1267 auto &Value2Index = Info.getValue2Index(IsSigned); 1268 while (!WorkList.empty()) { 1269 Value *V = WorkList.pop_back_val(); 1270 if (Old2New.find(V) != Old2New.end()) 1271 continue; 1272 1273 auto *I = dyn_cast<Instruction>(V); 1274 if (!Value2Index.contains(V) && I) { 1275 Old2New[V] = nullptr; 1276 ToClone.push_back(I); 1277 append_range(WorkList, I->operands()); 1278 } 1279 } 1280 1281 sort(ToClone, 1282 [&DT](Instruction *A, Instruction *B) { return DT.dominates(A, B); }); 1283 for (Instruction *I : ToClone) { 1284 Instruction *Cloned = I->clone(); 1285 Old2New[I] = Cloned; 1286 Old2New[I]->setName(I->getName()); 1287 Cloned->insertBefore(&*Builder.GetInsertPoint()); 1288 Cloned->dropUnknownNonDebugMetadata(); 1289 Cloned->setDebugLoc({}); 1290 } 1291 }; 1292 1293 // Materialize the assumptions for the reproducer using the entries in Stack. 1294 // That is, first clone the operands of the condition recursively until we 1295 // reach an external input to the reproducer and add them to the reproducer 1296 // function. Then add an ICmp for the condition (with the inverse predicate if 1297 // the entry is negated) and an assert using the ICmp. 1298 for (auto &Entry : Stack) { 1299 if (Entry.Pred == ICmpInst::BAD_ICMP_PREDICATE) 1300 continue; 1301 1302 LLVM_DEBUG(dbgs() << " Materializing assumption "; 1303 dumpUnpackedICmp(dbgs(), Entry.Pred, Entry.LHS, Entry.RHS); 1304 dbgs() << "\n"); 1305 CloneInstructions({Entry.LHS, Entry.RHS}, CmpInst::isSigned(Entry.Pred)); 1306 1307 auto *Cmp = Builder.CreateICmp(Entry.Pred, Entry.LHS, Entry.RHS); 1308 Builder.CreateAssumption(Cmp); 1309 } 1310 1311 // Finally, clone the condition to reproduce and remap instruction operands in 1312 // the reproducer using Old2New. 1313 CloneInstructions(Cond, CmpInst::isSigned(Cond->getPredicate())); 1314 Entry->getTerminator()->setOperand(0, Cond); 1315 remapInstructionsInBlocks({Entry}, Old2New); 1316 1317 assert(!verifyFunction(*F, &dbgs())); 1318 } 1319 1320 static std::optional<bool> checkCondition(CmpInst::Predicate Pred, Value *A, 1321 Value *B, Instruction *CheckInst, 1322 ConstraintInfo &Info) { 1323 LLVM_DEBUG(dbgs() << "Checking " << *CheckInst << "\n"); 1324 1325 auto R = Info.getConstraintForSolving(Pred, A, B); 1326 if (R.empty() || !R.isValid(Info)){ 1327 LLVM_DEBUG(dbgs() << " failed to decompose condition\n"); 1328 return std::nullopt; 1329 } 1330 1331 auto &CSToUse = Info.getCS(R.IsSigned); 1332 1333 // If there was extra information collected during decomposition, apply 1334 // it now and remove it immediately once we are done with reasoning 1335 // about the constraint. 1336 for (auto &Row : R.ExtraInfo) 1337 CSToUse.addVariableRow(Row); 1338 auto InfoRestorer = make_scope_exit([&]() { 1339 for (unsigned I = 0; I < R.ExtraInfo.size(); ++I) 1340 CSToUse.popLastConstraint(); 1341 }); 1342 1343 if (auto ImpliedCondition = R.isImpliedBy(CSToUse)) { 1344 if (!DebugCounter::shouldExecute(EliminatedCounter)) 1345 return std::nullopt; 1346 1347 LLVM_DEBUG({ 1348 dbgs() << "Condition "; 1349 dumpUnpackedICmp( 1350 dbgs(), *ImpliedCondition ? Pred : CmpInst::getInversePredicate(Pred), 1351 A, B); 1352 dbgs() << " implied by dominating constraints\n"; 1353 CSToUse.dump(); 1354 }); 1355 return ImpliedCondition; 1356 } 1357 1358 return std::nullopt; 1359 } 1360 1361 static bool checkAndReplaceCondition( 1362 CmpInst *Cmp, ConstraintInfo &Info, unsigned NumIn, unsigned NumOut, 1363 Instruction *ContextInst, Module *ReproducerModule, 1364 ArrayRef<ReproducerEntry> ReproducerCondStack, DominatorTree &DT, 1365 SmallVectorImpl<Instruction *> &ToRemove) { 1366 auto ReplaceCmpWithConstant = [&](CmpInst *Cmp, bool IsTrue) { 1367 generateReproducer(Cmp, ReproducerModule, ReproducerCondStack, Info, DT); 1368 Constant *ConstantC = ConstantInt::getBool( 1369 CmpInst::makeCmpResultType(Cmp->getType()), IsTrue); 1370 Cmp->replaceUsesWithIf(ConstantC, [&DT, NumIn, NumOut, 1371 ContextInst](Use &U) { 1372 auto *UserI = getContextInstForUse(U); 1373 auto *DTN = DT.getNode(UserI->getParent()); 1374 if (!DTN || DTN->getDFSNumIn() < NumIn || DTN->getDFSNumOut() > NumOut) 1375 return false; 1376 if (UserI->getParent() == ContextInst->getParent() && 1377 UserI->comesBefore(ContextInst)) 1378 return false; 1379 1380 // Conditions in an assume trivially simplify to true. Skip uses 1381 // in assume calls to not destroy the available information. 1382 auto *II = dyn_cast<IntrinsicInst>(U.getUser()); 1383 return !II || II->getIntrinsicID() != Intrinsic::assume; 1384 }); 1385 NumCondsRemoved++; 1386 if (Cmp->use_empty()) 1387 ToRemove.push_back(Cmp); 1388 return true; 1389 }; 1390 1391 if (auto ImpliedCondition = 1392 checkCondition(Cmp->getPredicate(), Cmp->getOperand(0), 1393 Cmp->getOperand(1), Cmp, Info)) 1394 return ReplaceCmpWithConstant(Cmp, *ImpliedCondition); 1395 return false; 1396 } 1397 1398 static void 1399 removeEntryFromStack(const StackEntry &E, ConstraintInfo &Info, 1400 Module *ReproducerModule, 1401 SmallVectorImpl<ReproducerEntry> &ReproducerCondStack, 1402 SmallVectorImpl<StackEntry> &DFSInStack) { 1403 Info.popLastConstraint(E.IsSigned); 1404 // Remove variables in the system that went out of scope. 1405 auto &Mapping = Info.getValue2Index(E.IsSigned); 1406 for (Value *V : E.ValuesToRelease) 1407 Mapping.erase(V); 1408 Info.popLastNVariables(E.IsSigned, E.ValuesToRelease.size()); 1409 DFSInStack.pop_back(); 1410 if (ReproducerModule) 1411 ReproducerCondStack.pop_back(); 1412 } 1413 1414 /// Check if either the first condition of an AND or OR is implied by the 1415 /// (negated in case of OR) second condition or vice versa. 1416 static bool checkOrAndOpImpliedByOther( 1417 FactOrCheck &CB, ConstraintInfo &Info, Module *ReproducerModule, 1418 SmallVectorImpl<ReproducerEntry> &ReproducerCondStack, 1419 SmallVectorImpl<StackEntry> &DFSInStack) { 1420 1421 CmpInst::Predicate Pred; 1422 Value *A, *B; 1423 Instruction *JoinOp = CB.getContextInst(); 1424 CmpInst *CmpToCheck = cast<CmpInst>(CB.getInstructionToSimplify()); 1425 unsigned OtherOpIdx = JoinOp->getOperand(0) == CmpToCheck ? 1 : 0; 1426 1427 // Don't try to simplify the first condition of a select by the second, as 1428 // this may make the select more poisonous than the original one. 1429 // TODO: check if the first operand may be poison. 1430 if (OtherOpIdx != 0 && isa<SelectInst>(JoinOp)) 1431 return false; 1432 1433 if (!match(JoinOp->getOperand(OtherOpIdx), 1434 m_ICmp(Pred, m_Value(A), m_Value(B)))) 1435 return false; 1436 1437 // For OR, check if the negated condition implies CmpToCheck. 1438 bool IsOr = match(JoinOp, m_LogicalOr()); 1439 if (IsOr) 1440 Pred = CmpInst::getInversePredicate(Pred); 1441 1442 // Optimistically add fact from first condition. 1443 unsigned OldSize = DFSInStack.size(); 1444 Info.addFact(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack); 1445 if (OldSize == DFSInStack.size()) 1446 return false; 1447 1448 bool Changed = false; 1449 // Check if the second condition can be simplified now. 1450 if (auto ImpliedCondition = 1451 checkCondition(CmpToCheck->getPredicate(), CmpToCheck->getOperand(0), 1452 CmpToCheck->getOperand(1), CmpToCheck, Info)) { 1453 if (IsOr && isa<SelectInst>(JoinOp)) { 1454 JoinOp->setOperand( 1455 OtherOpIdx == 0 ? 2 : 0, 1456 ConstantInt::getBool(JoinOp->getType(), *ImpliedCondition)); 1457 } else 1458 JoinOp->setOperand( 1459 1 - OtherOpIdx, 1460 ConstantInt::getBool(JoinOp->getType(), *ImpliedCondition)); 1461 1462 Changed = true; 1463 } 1464 1465 // Remove entries again. 1466 while (OldSize < DFSInStack.size()) { 1467 StackEntry E = DFSInStack.back(); 1468 removeEntryFromStack(E, Info, ReproducerModule, ReproducerCondStack, 1469 DFSInStack); 1470 } 1471 return Changed; 1472 } 1473 1474 void ConstraintInfo::addFact(CmpInst::Predicate Pred, Value *A, Value *B, 1475 unsigned NumIn, unsigned NumOut, 1476 SmallVectorImpl<StackEntry> &DFSInStack) { 1477 // If the constraint has a pre-condition, skip the constraint if it does not 1478 // hold. 1479 SmallVector<Value *> NewVariables; 1480 auto R = getConstraint(Pred, A, B, NewVariables); 1481 1482 // TODO: Support non-equality for facts as well. 1483 if (!R.isValid(*this) || R.isNe()) 1484 return; 1485 1486 LLVM_DEBUG(dbgs() << "Adding '"; dumpUnpackedICmp(dbgs(), Pred, A, B); 1487 dbgs() << "'\n"); 1488 bool Added = false; 1489 auto &CSToUse = getCS(R.IsSigned); 1490 if (R.Coefficients.empty()) 1491 return; 1492 1493 Added |= CSToUse.addVariableRowFill(R.Coefficients); 1494 1495 // If R has been added to the system, add the new variables and queue it for 1496 // removal once it goes out-of-scope. 1497 if (Added) { 1498 SmallVector<Value *, 2> ValuesToRelease; 1499 auto &Value2Index = getValue2Index(R.IsSigned); 1500 for (Value *V : NewVariables) { 1501 Value2Index.insert({V, Value2Index.size() + 1}); 1502 ValuesToRelease.push_back(V); 1503 } 1504 1505 LLVM_DEBUG({ 1506 dbgs() << " constraint: "; 1507 dumpConstraint(R.Coefficients, getValue2Index(R.IsSigned)); 1508 dbgs() << "\n"; 1509 }); 1510 1511 DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned, 1512 std::move(ValuesToRelease)); 1513 1514 if (!R.IsSigned) { 1515 for (Value *V : NewVariables) { 1516 ConstraintTy VarPos(SmallVector<int64_t, 8>(Value2Index.size() + 1, 0), 1517 false, false, false); 1518 VarPos.Coefficients[Value2Index[V]] = -1; 1519 CSToUse.addVariableRow(VarPos.Coefficients); 1520 DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned, 1521 SmallVector<Value *, 2>()); 1522 } 1523 } 1524 1525 if (R.isEq()) { 1526 // Also add the inverted constraint for equality constraints. 1527 for (auto &Coeff : R.Coefficients) 1528 Coeff *= -1; 1529 CSToUse.addVariableRowFill(R.Coefficients); 1530 1531 DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned, 1532 SmallVector<Value *, 2>()); 1533 } 1534 } 1535 } 1536 1537 static bool replaceSubOverflowUses(IntrinsicInst *II, Value *A, Value *B, 1538 SmallVectorImpl<Instruction *> &ToRemove) { 1539 bool Changed = false; 1540 IRBuilder<> Builder(II->getParent(), II->getIterator()); 1541 Value *Sub = nullptr; 1542 for (User *U : make_early_inc_range(II->users())) { 1543 if (match(U, m_ExtractValue<0>(m_Value()))) { 1544 if (!Sub) 1545 Sub = Builder.CreateSub(A, B); 1546 U->replaceAllUsesWith(Sub); 1547 Changed = true; 1548 } else if (match(U, m_ExtractValue<1>(m_Value()))) { 1549 U->replaceAllUsesWith(Builder.getFalse()); 1550 Changed = true; 1551 } else 1552 continue; 1553 1554 if (U->use_empty()) { 1555 auto *I = cast<Instruction>(U); 1556 ToRemove.push_back(I); 1557 I->setOperand(0, PoisonValue::get(II->getType())); 1558 Changed = true; 1559 } 1560 } 1561 1562 if (II->use_empty()) { 1563 II->eraseFromParent(); 1564 Changed = true; 1565 } 1566 return Changed; 1567 } 1568 1569 static bool 1570 tryToSimplifyOverflowMath(IntrinsicInst *II, ConstraintInfo &Info, 1571 SmallVectorImpl<Instruction *> &ToRemove) { 1572 auto DoesConditionHold = [](CmpInst::Predicate Pred, Value *A, Value *B, 1573 ConstraintInfo &Info) { 1574 auto R = Info.getConstraintForSolving(Pred, A, B); 1575 if (R.size() < 2 || !R.isValid(Info)) 1576 return false; 1577 1578 auto &CSToUse = Info.getCS(R.IsSigned); 1579 return CSToUse.isConditionImplied(R.Coefficients); 1580 }; 1581 1582 bool Changed = false; 1583 if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow) { 1584 // If A s>= B && B s>= 0, ssub.with.overflow(a, b) should not overflow and 1585 // can be simplified to a regular sub. 1586 Value *A = II->getArgOperand(0); 1587 Value *B = II->getArgOperand(1); 1588 if (!DoesConditionHold(CmpInst::ICMP_SGE, A, B, Info) || 1589 !DoesConditionHold(CmpInst::ICMP_SGE, B, 1590 ConstantInt::get(A->getType(), 0), Info)) 1591 return false; 1592 Changed = replaceSubOverflowUses(II, A, B, ToRemove); 1593 } 1594 return Changed; 1595 } 1596 1597 static bool eliminateConstraints(Function &F, DominatorTree &DT, LoopInfo &LI, 1598 ScalarEvolution &SE, 1599 OptimizationRemarkEmitter &ORE) { 1600 bool Changed = false; 1601 DT.updateDFSNumbers(); 1602 SmallVector<Value *> FunctionArgs; 1603 for (Value &Arg : F.args()) 1604 FunctionArgs.push_back(&Arg); 1605 ConstraintInfo Info(F.getParent()->getDataLayout(), FunctionArgs); 1606 State S(DT, LI, SE); 1607 std::unique_ptr<Module> ReproducerModule( 1608 DumpReproducers ? new Module(F.getName(), F.getContext()) : nullptr); 1609 1610 // First, collect conditions implied by branches and blocks with their 1611 // Dominator DFS in and out numbers. 1612 for (BasicBlock &BB : F) { 1613 if (!DT.getNode(&BB)) 1614 continue; 1615 S.addInfoFor(BB); 1616 } 1617 1618 // Next, sort worklist by dominance, so that dominating conditions to check 1619 // and facts come before conditions and facts dominated by them. If a 1620 // condition to check and a fact have the same numbers, conditional facts come 1621 // first. Assume facts and checks are ordered according to their relative 1622 // order in the containing basic block. Also make sure conditions with 1623 // constant operands come before conditions without constant operands. This 1624 // increases the effectiveness of the current signed <-> unsigned fact 1625 // transfer logic. 1626 stable_sort(S.WorkList, [](const FactOrCheck &A, const FactOrCheck &B) { 1627 auto HasNoConstOp = [](const FactOrCheck &B) { 1628 Value *V0 = B.isConditionFact() ? B.Cond.Op0 : B.Inst->getOperand(0); 1629 Value *V1 = B.isConditionFact() ? B.Cond.Op1 : B.Inst->getOperand(1); 1630 return !isa<ConstantInt>(V0) && !isa<ConstantInt>(V1); 1631 }; 1632 // If both entries have the same In numbers, conditional facts come first. 1633 // Otherwise use the relative order in the basic block. 1634 if (A.NumIn == B.NumIn) { 1635 if (A.isConditionFact() && B.isConditionFact()) { 1636 bool NoConstOpA = HasNoConstOp(A); 1637 bool NoConstOpB = HasNoConstOp(B); 1638 return NoConstOpA < NoConstOpB; 1639 } 1640 if (A.isConditionFact()) 1641 return true; 1642 if (B.isConditionFact()) 1643 return false; 1644 auto *InstA = A.getContextInst(); 1645 auto *InstB = B.getContextInst(); 1646 return InstA->comesBefore(InstB); 1647 } 1648 return A.NumIn < B.NumIn; 1649 }); 1650 1651 SmallVector<Instruction *> ToRemove; 1652 1653 // Finally, process ordered worklist and eliminate implied conditions. 1654 SmallVector<StackEntry, 16> DFSInStack; 1655 SmallVector<ReproducerEntry> ReproducerCondStack; 1656 for (FactOrCheck &CB : S.WorkList) { 1657 // First, pop entries from the stack that are out-of-scope for CB. Remove 1658 // the corresponding entry from the constraint system. 1659 while (!DFSInStack.empty()) { 1660 auto &E = DFSInStack.back(); 1661 LLVM_DEBUG(dbgs() << "Top of stack : " << E.NumIn << " " << E.NumOut 1662 << "\n"); 1663 LLVM_DEBUG(dbgs() << "CB: " << CB.NumIn << " " << CB.NumOut << "\n"); 1664 assert(E.NumIn <= CB.NumIn); 1665 if (CB.NumOut <= E.NumOut) 1666 break; 1667 LLVM_DEBUG({ 1668 dbgs() << "Removing "; 1669 dumpConstraint(Info.getCS(E.IsSigned).getLastConstraint(), 1670 Info.getValue2Index(E.IsSigned)); 1671 dbgs() << "\n"; 1672 }); 1673 removeEntryFromStack(E, Info, ReproducerModule.get(), ReproducerCondStack, 1674 DFSInStack); 1675 } 1676 1677 // For a block, check if any CmpInsts become known based on the current set 1678 // of constraints. 1679 if (CB.isCheck()) { 1680 Instruction *Inst = CB.getInstructionToSimplify(); 1681 if (!Inst) 1682 continue; 1683 LLVM_DEBUG(dbgs() << "Processing condition to simplify: " << *Inst 1684 << "\n"); 1685 if (auto *II = dyn_cast<WithOverflowInst>(Inst)) { 1686 Changed |= tryToSimplifyOverflowMath(II, Info, ToRemove); 1687 } else if (auto *Cmp = dyn_cast<ICmpInst>(Inst)) { 1688 bool Simplified = checkAndReplaceCondition( 1689 Cmp, Info, CB.NumIn, CB.NumOut, CB.getContextInst(), 1690 ReproducerModule.get(), ReproducerCondStack, S.DT, ToRemove); 1691 if (!Simplified && 1692 match(CB.getContextInst(), m_LogicalOp(m_Value(), m_Value()))) { 1693 Simplified = 1694 checkOrAndOpImpliedByOther(CB, Info, ReproducerModule.get(), 1695 ReproducerCondStack, DFSInStack); 1696 } 1697 Changed |= Simplified; 1698 } 1699 continue; 1700 } 1701 1702 auto AddFact = [&](CmpInst::Predicate Pred, Value *A, Value *B) { 1703 LLVM_DEBUG(dbgs() << "Processing fact to add to the system: "; 1704 dumpUnpackedICmp(dbgs(), Pred, A, B); dbgs() << "\n"); 1705 if (Info.getCS(CmpInst::isSigned(Pred)).size() > MaxRows) { 1706 LLVM_DEBUG( 1707 dbgs() 1708 << "Skip adding constraint because system has too many rows.\n"); 1709 return; 1710 } 1711 1712 Info.addFact(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack); 1713 if (ReproducerModule && DFSInStack.size() > ReproducerCondStack.size()) 1714 ReproducerCondStack.emplace_back(Pred, A, B); 1715 1716 Info.transferToOtherSystem(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack); 1717 if (ReproducerModule && DFSInStack.size() > ReproducerCondStack.size()) { 1718 // Add dummy entries to ReproducerCondStack to keep it in sync with 1719 // DFSInStack. 1720 for (unsigned I = 0, 1721 E = (DFSInStack.size() - ReproducerCondStack.size()); 1722 I < E; ++I) { 1723 ReproducerCondStack.emplace_back(ICmpInst::BAD_ICMP_PREDICATE, 1724 nullptr, nullptr); 1725 } 1726 } 1727 }; 1728 1729 ICmpInst::Predicate Pred; 1730 if (!CB.isConditionFact()) { 1731 Value *X; 1732 if (match(CB.Inst, m_Intrinsic<Intrinsic::abs>(m_Value(X)))) { 1733 // TODO: Add CB.Inst >= 0 fact. 1734 AddFact(CmpInst::ICMP_SGE, CB.Inst, X); 1735 continue; 1736 } 1737 1738 if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(CB.Inst)) { 1739 Pred = ICmpInst::getNonStrictPredicate(MinMax->getPredicate()); 1740 AddFact(Pred, MinMax, MinMax->getLHS()); 1741 AddFact(Pred, MinMax, MinMax->getRHS()); 1742 continue; 1743 } 1744 } 1745 1746 Value *A = nullptr, *B = nullptr; 1747 if (CB.isConditionFact()) { 1748 Pred = CB.Cond.Pred; 1749 A = CB.Cond.Op0; 1750 B = CB.Cond.Op1; 1751 if (CB.DoesHold.Pred != CmpInst::BAD_ICMP_PREDICATE && 1752 !Info.doesHold(CB.DoesHold.Pred, CB.DoesHold.Op0, CB.DoesHold.Op1)) { 1753 LLVM_DEBUG({ 1754 dbgs() << "Not adding fact "; 1755 dumpUnpackedICmp(dbgs(), Pred, A, B); 1756 dbgs() << " because precondition "; 1757 dumpUnpackedICmp(dbgs(), CB.DoesHold.Pred, CB.DoesHold.Op0, 1758 CB.DoesHold.Op1); 1759 dbgs() << " does not hold.\n"; 1760 }); 1761 continue; 1762 } 1763 } else { 1764 bool Matched = match(CB.Inst, m_Intrinsic<Intrinsic::assume>( 1765 m_ICmp(Pred, m_Value(A), m_Value(B)))); 1766 (void)Matched; 1767 assert(Matched && "Must have an assume intrinsic with a icmp operand"); 1768 } 1769 AddFact(Pred, A, B); 1770 } 1771 1772 if (ReproducerModule && !ReproducerModule->functions().empty()) { 1773 std::string S; 1774 raw_string_ostream StringS(S); 1775 ReproducerModule->print(StringS, nullptr); 1776 StringS.flush(); 1777 OptimizationRemark Rem(DEBUG_TYPE, "Reproducer", &F); 1778 Rem << ore::NV("module") << S; 1779 ORE.emit(Rem); 1780 } 1781 1782 #ifndef NDEBUG 1783 unsigned SignedEntries = 1784 count_if(DFSInStack, [](const StackEntry &E) { return E.IsSigned; }); 1785 assert(Info.getCS(false).size() - FunctionArgs.size() == 1786 DFSInStack.size() - SignedEntries && 1787 "updates to CS and DFSInStack are out of sync"); 1788 assert(Info.getCS(true).size() == SignedEntries && 1789 "updates to CS and DFSInStack are out of sync"); 1790 #endif 1791 1792 for (Instruction *I : ToRemove) 1793 I->eraseFromParent(); 1794 return Changed; 1795 } 1796 1797 PreservedAnalyses ConstraintEliminationPass::run(Function &F, 1798 FunctionAnalysisManager &AM) { 1799 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1800 auto &LI = AM.getResult<LoopAnalysis>(F); 1801 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 1802 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); 1803 if (!eliminateConstraints(F, DT, LI, SE, ORE)) 1804 return PreservedAnalyses::all(); 1805 1806 PreservedAnalyses PA; 1807 PA.preserve<DominatorTreeAnalysis>(); 1808 PA.preserve<LoopAnalysis>(); 1809 PA.preserve<ScalarEvolutionAnalysis>(); 1810 PA.preserveSet<CFGAnalyses>(); 1811 return PA; 1812 } 1813