| 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/MemoryBuiltins.h" |
| 23 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 24 | #include "llvm/Analysis/ScalarEvolution.h" |
| 25 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 26 | #include "llvm/Analysis/ScalarEvolutionPatternMatch.h" |
| 27 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 28 | #include "llvm/Analysis/ValueTracking.h" |
| 29 | #include "llvm/IR/DataLayout.h" |
| 30 | #include "llvm/IR/DebugInfo.h" |
| 31 | #include "llvm/IR/Dominators.h" |
| 32 | #include "llvm/IR/Function.h" |
| 33 | #include "llvm/IR/IRBuilder.h" |
| 34 | #include "llvm/IR/InstrTypes.h" |
| 35 | #include "llvm/IR/Instructions.h" |
| 36 | #include "llvm/IR/Module.h" |
| 37 | #include "llvm/IR/PatternMatch.h" |
| 38 | #include "llvm/IR/Verifier.h" |
| 39 | #include "llvm/Pass.h" |
| 40 | #include "llvm/Support/CommandLine.h" |
| 41 | #include "llvm/Support/Debug.h" |
| 42 | #include "llvm/Support/DebugCounter.h" |
| 43 | #include "llvm/Support/MathExtras.h" |
| 44 | #include "llvm/Transforms/Utils/Cloning.h" |
| 45 | #include "llvm/Transforms/Utils/ValueMapper.h" |
| 46 | |
| 47 | #include <optional> |
| 48 | #include <string> |
| 49 | |
| 50 | using namespace llvm; |
| 51 | using namespace PatternMatch; |
| 52 | using namespace SCEVPatternMatch; |
| 53 | |
| 54 | #define DEBUG_TYPE "constraint-elimination" |
| 55 | |
| 56 | STATISTIC(NumCondsRemoved, "Number of instructions removed" ); |
| 57 | DEBUG_COUNTER(EliminatedCounter, "conds-eliminated" , |
| 58 | "Controls which conditions are eliminated" ); |
| 59 | |
| 60 | static cl::opt<unsigned> |
| 61 | MaxRows("constraint-elimination-max-rows" , cl::init(Val: 500), cl::Hidden, |
| 62 | cl::desc("Maximum number of rows to keep in constraint system" )); |
| 63 | |
| 64 | static cl::opt<bool> DumpReproducers( |
| 65 | "constraint-elimination-dump-reproducers" , cl::init(Val: false), cl::Hidden, |
| 66 | cl::desc("Dump IR to reproduce successful transformations." )); |
| 67 | |
| 68 | static int64_t MaxConstraintValue = std::numeric_limits<int64_t>::max(); |
| 69 | static int64_t MinSignedConstraintValue = std::numeric_limits<int64_t>::min(); |
| 70 | |
| 71 | static Instruction *getContextInstForUse(Use &U) { |
| 72 | Instruction *UserI = cast<Instruction>(Val: U.getUser()); |
| 73 | if (auto *Phi = dyn_cast<PHINode>(Val: UserI)) |
| 74 | UserI = Phi->getIncomingBlock(U)->getTerminator(); |
| 75 | return UserI; |
| 76 | } |
| 77 | |
| 78 | /// Returns the closest program point dominating all uses of \p I. |
| 79 | static Instruction *findCommonDominatorOfUses(Instruction &I, |
| 80 | DominatorTree &DT) { |
| 81 | Instruction *CommonDom = nullptr; |
| 82 | unsigned NumUses = 0; |
| 83 | for (Use &U : I.uses()) { |
| 84 | // Conservatively use original instruction, if there are too many uses. |
| 85 | if (++NumUses == 16) |
| 86 | return &I; |
| 87 | Instruction *UserI = getContextInstForUse(U); |
| 88 | CommonDom = |
| 89 | CommonDom ? DT.findNearestCommonDominator(I1: CommonDom, I2: UserI) : UserI; |
| 90 | } |
| 91 | if (!CommonDom) |
| 92 | return &I; |
| 93 | // Uses in unreachable blocks are not in the dominator tree. |
| 94 | return DT.getNode(BB: CommonDom->getParent()) ? CommonDom : &I; |
| 95 | } |
| 96 | |
| 97 | namespace { |
| 98 | using Entry = ConstraintSystem::Entry; |
| 99 | using RowTy = ConstraintSystem::RowTy; |
| 100 | |
| 101 | /// Struct to express a condition of the form %Op0 Pred %Op1. |
| 102 | struct ConditionTy { |
| 103 | CmpPredicate Pred; |
| 104 | Value *Op0 = nullptr; |
| 105 | Value *Op1 = nullptr; |
| 106 | |
| 107 | ConditionTy() = default; |
| 108 | ConditionTy(CmpPredicate Pred, Value *Op0, Value *Op1) |
| 109 | : Pred(Pred), Op0(Op0), Op1(Op1) {} |
| 110 | }; |
| 111 | |
| 112 | /// Represents either |
| 113 | /// * a condition that holds on entry to a block (=condition fact) |
| 114 | /// * an assume (=assume fact) |
| 115 | /// * a use of a compare instruction to simplify. |
| 116 | /// It also tracks the Dominator DFS in and out numbers for each entry. |
| 117 | struct FactOrCheck { |
| 118 | enum class EntryTy { |
| 119 | ConditionFact, /// A condition that holds on entry to a block. |
| 120 | InstFact, /// A fact that holds after Inst executed (e.g. an assume or |
| 121 | /// min/mix intrinsic. |
| 122 | InstCheck, /// An instruction to simplify (e.g. an overflow math |
| 123 | /// intrinsics) or whose flags may be strengthened. |
| 124 | UseCheck /// An use of a compare instruction to simplify. |
| 125 | }; |
| 126 | |
| 127 | union { |
| 128 | Instruction *Inst; |
| 129 | Use *U; |
| 130 | ConditionTy Cond; |
| 131 | }; |
| 132 | |
| 133 | union { |
| 134 | /// A pre-condition that must hold for the current fact to be added to the |
| 135 | /// system. Only used by condition facts. |
| 136 | ConditionTy DoesHold; |
| 137 | |
| 138 | /// Context instruction for the point where conditions are checked for |
| 139 | /// InstCheck simplifications. |
| 140 | Instruction *ContextInst; |
| 141 | }; |
| 142 | |
| 143 | unsigned NumIn; |
| 144 | unsigned NumOut; |
| 145 | EntryTy Ty; |
| 146 | |
| 147 | FactOrCheck(EntryTy Ty, DomTreeNode *DTN, Instruction *Inst, |
| 148 | Instruction *ContextInst = nullptr) |
| 149 | : Inst(Inst), ContextInst(ContextInst ? ContextInst : Inst), |
| 150 | NumIn(DTN->getDFSNumIn()), NumOut(DTN->getDFSNumOut()), Ty(Ty) {} |
| 151 | |
| 152 | FactOrCheck(DomTreeNode *DTN, Use *U) |
| 153 | : U(U), ContextInst(nullptr), NumIn(DTN->getDFSNumIn()), |
| 154 | NumOut(DTN->getDFSNumOut()), Ty(EntryTy::UseCheck) {} |
| 155 | |
| 156 | FactOrCheck(DomTreeNode *DTN, CmpPredicate Pred, Value *Op0, Value *Op1, |
| 157 | ConditionTy Precond = {}) |
| 158 | : Cond(Pred, Op0, Op1), DoesHold(Precond), NumIn(DTN->getDFSNumIn()), |
| 159 | NumOut(DTN->getDFSNumOut()), Ty(EntryTy::ConditionFact) {} |
| 160 | |
| 161 | static FactOrCheck getConditionFact(DomTreeNode *DTN, CmpPredicate Pred, |
| 162 | Value *Op0, Value *Op1, |
| 163 | ConditionTy Precond = {}) { |
| 164 | return FactOrCheck(DTN, Pred, Op0, Op1, Precond); |
| 165 | } |
| 166 | |
| 167 | static FactOrCheck getInstFact(DomTreeNode *DTN, Instruction *Inst) { |
| 168 | return FactOrCheck(EntryTy::InstFact, DTN, Inst); |
| 169 | } |
| 170 | |
| 171 | static FactOrCheck getCheck(DomTreeNode *DTN, Use *U) { |
| 172 | return FactOrCheck(DTN, U); |
| 173 | } |
| 174 | |
| 175 | static FactOrCheck getCheck(DomTreeNode *DTN, Instruction *I, |
| 176 | Instruction *ContextInst = nullptr) { |
| 177 | assert((ContextInst ? ContextInst : I)->getParent() == DTN->getBlock() && |
| 178 | "anchoring instruction must be in DTN's block" ); |
| 179 | return FactOrCheck(EntryTy::InstCheck, DTN, I, ContextInst); |
| 180 | } |
| 181 | |
| 182 | bool isCheck() const { |
| 183 | return Ty == EntryTy::InstCheck || Ty == EntryTy::UseCheck; |
| 184 | } |
| 185 | |
| 186 | Instruction *getContextInst() const { |
| 187 | assert(!isConditionFact()); |
| 188 | if (Ty == EntryTy::UseCheck) |
| 189 | return getContextInstForUse(U&: *U); |
| 190 | return ContextInst; |
| 191 | } |
| 192 | |
| 193 | Instruction *getInstructionToSimplify() const { |
| 194 | assert(isCheck()); |
| 195 | if (Ty == EntryTy::InstCheck) |
| 196 | return Inst; |
| 197 | // The use may have been simplified to a constant already. |
| 198 | return dyn_cast<Instruction>(Val&: *U); |
| 199 | } |
| 200 | |
| 201 | bool isConditionFact() const { return Ty == EntryTy::ConditionFact; } |
| 202 | }; |
| 203 | |
| 204 | /// The senses in which an induction phi is monotonic, together with the |
| 205 | /// direction it moves in. |
| 206 | struct MonotonicInfo { |
| 207 | /// True if the phi steps by a negative constant. |
| 208 | bool Decreasing = false; |
| 209 | /// True if the phi is monotonic in the unsigned sense. |
| 210 | bool Unsigned = false; |
| 211 | /// True if the phi is monotonic in the signed sense. |
| 212 | bool Signed = false; |
| 213 | }; |
| 214 | |
| 215 | /// Keep state required to build worklist. |
| 216 | struct State { |
| 217 | DominatorTree &DT; |
| 218 | LoopInfo &LI; |
| 219 | ScalarEvolution &SE; |
| 220 | TargetLibraryInfo &TLI; |
| 221 | SmallVector<FactOrCheck, 64> WorkList; |
| 222 | |
| 223 | State(DominatorTree &DT, LoopInfo &LI, ScalarEvolution &SE, |
| 224 | TargetLibraryInfo &TLI) |
| 225 | : DT(DT), LI(LI), SE(SE), TLI(TLI) {} |
| 226 | |
| 227 | /// Process block \p BB and add known facts to work-list. |
| 228 | void addInfoFor(BasicBlock &BB); |
| 229 | |
| 230 | /// If \p BB is a loop header, bound each induction phi in it by its start |
| 231 | /// value. |
| 232 | void addBoundsForHeaderInductions(BasicBlock &BB); |
| 233 | |
| 234 | /// Try to add facts for loop inductions (AddRecs) in EQ/NE compares |
| 235 | /// controlling the loop header. |
| 236 | void addInfoForInductions(BasicBlock &BB); |
| 237 | |
| 238 | /// Returns the direction the induction phi \p PN with backedge value \p Step |
| 239 | /// moves in, and the senses in which it is monotonic in that direction. |
| 240 | MonotonicInfo getMonotonicityInfo(PHINode &PN, Value *Step); |
| 241 | |
| 242 | /// Returns true if we can add a known condition from BB to its successor |
| 243 | /// block Succ. |
| 244 | bool canAddSuccessor(BasicBlock &BB, BasicBlock *Succ) const { |
| 245 | return DT.dominates(BBE: BasicBlockEdge(&BB, Succ), BB: Succ); |
| 246 | } |
| 247 | }; |
| 248 | |
| 249 | class ConstraintInfo; |
| 250 | |
| 251 | struct StackEntry { |
| 252 | unsigned NumIn; |
| 253 | unsigned NumOut; |
| 254 | bool IsSigned = false; |
| 255 | /// Variables that can be removed from the system once the stack entry gets |
| 256 | /// removed. |
| 257 | SmallVector<Value *, 2> ValuesToRelease; |
| 258 | |
| 259 | StackEntry(unsigned NumIn, unsigned NumOut, bool IsSigned, |
| 260 | SmallVector<Value *, 2> ValuesToRelease) |
| 261 | : NumIn(NumIn), NumOut(NumOut), IsSigned(IsSigned), |
| 262 | ValuesToRelease(std::move(ValuesToRelease)) {} |
| 263 | }; |
| 264 | |
| 265 | struct ConstraintTy { |
| 266 | RowTy Coefficients; |
| 267 | |
| 268 | /// Number of variables the constraint is defined over. |
| 269 | unsigned NumVars = 0; |
| 270 | |
| 271 | bool IsSigned = false; |
| 272 | |
| 273 | ConstraintTy() = default; |
| 274 | |
| 275 | ConstraintTy(RowTy Coefficients, unsigned NumVars, bool IsSigned, bool IsEq, |
| 276 | bool IsNe) |
| 277 | : Coefficients(std::move(Coefficients)), NumVars(NumVars), |
| 278 | IsSigned(IsSigned), IsEq(IsEq), IsNe(IsNe) {} |
| 279 | |
| 280 | bool empty() const { return Coefficients.empty(); } |
| 281 | |
| 282 | bool isEq() const { return IsEq; } |
| 283 | |
| 284 | bool isNe() const { return IsNe; } |
| 285 | |
| 286 | /// Check if the current constraint is implied by the given ConstraintSystem. |
| 287 | /// |
| 288 | /// \return true or false if the constraint is proven to be respectively true, |
| 289 | /// or false. When the constraint cannot be proven to be either true or false, |
| 290 | /// std::nullopt is returned. |
| 291 | std::optional<bool> isImpliedBy(const ConstraintSystem &CS) const; |
| 292 | |
| 293 | private: |
| 294 | bool IsEq = false; |
| 295 | bool IsNe = false; |
| 296 | }; |
| 297 | |
| 298 | /// Wrapper encapsulating separate constraint systems and corresponding value |
| 299 | /// mappings for both unsigned and signed information. Facts are added to and |
| 300 | /// conditions are checked against the corresponding system depending on the |
| 301 | /// signed-ness of their predicates. While the information is kept separate |
| 302 | /// based on signed-ness, certain conditions can be transferred between the two |
| 303 | /// systems. |
| 304 | class ConstraintInfo { |
| 305 | |
| 306 | ConstraintSystem UnsignedCS; |
| 307 | ConstraintSystem SignedCS; |
| 308 | |
| 309 | const DataLayout &DL; |
| 310 | |
| 311 | public: |
| 312 | ConstraintInfo(const DataLayout &DL, ArrayRef<Value *> FunctionArgs) |
| 313 | : UnsignedCS(FunctionArgs), SignedCS(FunctionArgs), DL(DL) { |
| 314 | auto &Value2Index = getValue2Index(Signed: false); |
| 315 | // Add Arg > -1 constraints to unsigned system for all function arguments. |
| 316 | for (Value *Arg : FunctionArgs) |
| 317 | UnsignedCS.addRow(R: {Entry(0, 0), Entry(-1, Value2Index.at(Val: Arg))}, |
| 318 | NumVars: Value2Index.size()); |
| 319 | } |
| 320 | |
| 321 | DenseMap<Value *, unsigned> &getValue2Index(bool Signed) { |
| 322 | return Signed ? SignedCS.getValue2Index() : UnsignedCS.getValue2Index(); |
| 323 | } |
| 324 | const DenseMap<Value *, unsigned> &getValue2Index(bool Signed) const { |
| 325 | return Signed ? SignedCS.getValue2Index() : UnsignedCS.getValue2Index(); |
| 326 | } |
| 327 | |
| 328 | ConstraintSystem &getCS(bool Signed) { |
| 329 | return Signed ? SignedCS : UnsignedCS; |
| 330 | } |
| 331 | const ConstraintSystem &getCS(bool Signed) const { |
| 332 | return Signed ? SignedCS : UnsignedCS; |
| 333 | } |
| 334 | |
| 335 | void popLastConstraint(bool Signed) { getCS(Signed).popLastConstraint(); } |
| 336 | void popLastNVariables(bool Signed, unsigned N) { |
| 337 | getCS(Signed).popLastNVariables(N); |
| 338 | } |
| 339 | |
| 340 | bool doesHold(CmpInst::Predicate Pred, Value *A, Value *B) const; |
| 341 | |
| 342 | /// Returns true if \p V is known to be non-negative, either because the |
| 343 | /// signed system implies it or because ValueTracking can prove it. |
| 344 | bool isKnownNonNegative(Value *V) const; |
| 345 | |
| 346 | void addFact(CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn, |
| 347 | unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack); |
| 348 | |
| 349 | /// Turn a comparison of the form \p Op0 \p Pred \p Op1 into a vector of |
| 350 | /// constraints, using indices from the corresponding constraint system. |
| 351 | /// New variables that need to be added to the system are collected in |
| 352 | /// \p NewVariables. |
| 353 | ConstraintTy getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1, |
| 354 | SmallVectorImpl<Value *> &NewVariables, |
| 355 | bool ForceSignedSystem = false) const; |
| 356 | |
| 357 | /// Turns a comparison of the form \p Op0 \p Pred \p Op1 into a vector of |
| 358 | /// constraints using getConstraint. Returns an empty constraint if the result |
| 359 | /// cannot be used to query the existing constraint system, e.g. because it |
| 360 | /// would require adding new variables. Also tries to convert signed |
| 361 | /// predicates to unsigned ones if possible to allow using the unsigned system |
| 362 | /// which increases the effectiveness of the signed <-> unsigned transfer |
| 363 | /// logic. |
| 364 | ConstraintTy getConstraintForSolving(CmpInst::Predicate Pred, Value *Op0, |
| 365 | Value *Op1) const; |
| 366 | |
| 367 | /// Try to add information from \p A \p Pred \p B to the unsigned/signed |
| 368 | /// system if \p Pred is signed/unsigned. |
| 369 | void transferToOtherSystem(CmpInst::Predicate Pred, Value *A, Value *B, |
| 370 | unsigned NumIn, unsigned NumOut, |
| 371 | SmallVectorImpl<StackEntry> &DFSInStack); |
| 372 | |
| 373 | private: |
| 374 | /// Adds facts into constraint system. \p ForceSignedSystem can be set when |
| 375 | /// the \p Pred is eq/ne, and signed constraint system is used when it's |
| 376 | /// specified. |
| 377 | void addFactImpl(CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn, |
| 378 | unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack, |
| 379 | bool ForceSignedSystem); |
| 380 | |
| 381 | /// Try to use the inequality \p A != \p B to tighten a non-strict bound the |
| 382 | /// system already implies to the corresponding strict bound. |
| 383 | void tightenBoundUsingNe(Value *A, Value *B, unsigned NumIn, unsigned NumOut, |
| 384 | SmallVectorImpl<StackEntry> &DFSInStack); |
| 385 | }; |
| 386 | |
| 387 | /// Represents a (Coefficient * Variable) entry after IR decomposition. |
| 388 | struct DecompEntry { |
| 389 | int64_t Coefficient; |
| 390 | Value *Variable; |
| 391 | |
| 392 | DecompEntry(int64_t Coefficient, Value *Variable) |
| 393 | : Coefficient(Coefficient), Variable(Variable) {} |
| 394 | }; |
| 395 | |
| 396 | /// Represents an Offset + Coefficient1 * Variable1 + ... decomposition. |
| 397 | struct Decomposition { |
| 398 | int64_t Offset = 0; |
| 399 | SmallVector<DecompEntry, 3> Vars; |
| 400 | |
| 401 | Decomposition(int64_t Offset) : Offset(Offset) {} |
| 402 | Decomposition(Value *V) { Vars.emplace_back(Args: 1, Args&: V); } |
| 403 | Decomposition(int64_t Offset, ArrayRef<DecompEntry> Vars) |
| 404 | : Offset(Offset), Vars(Vars) {} |
| 405 | |
| 406 | /// Add \p OtherOffset and return true if the operation overflows, i.e. the |
| 407 | /// new decomposition is invalid. |
| 408 | [[nodiscard]] bool add(int64_t OtherOffset) { |
| 409 | return AddOverflow(X: Offset, Y: OtherOffset, Result&: Offset); |
| 410 | } |
| 411 | |
| 412 | /// Add \p Other and return true if the operation overflows, i.e. the new |
| 413 | /// decomposition is invalid. |
| 414 | [[nodiscard]] bool add(const Decomposition &Other) { |
| 415 | if (add(OtherOffset: Other.Offset)) |
| 416 | return true; |
| 417 | append_range(C&: Vars, R: Other.Vars); |
| 418 | return false; |
| 419 | } |
| 420 | |
| 421 | /// Subtract \p Other and return true if the operation overflows, i.e. the new |
| 422 | /// decomposition is invalid. |
| 423 | [[nodiscard]] bool sub(const Decomposition &Other) { |
| 424 | Decomposition Tmp = Other; |
| 425 | if (Tmp.mul(Factor: -1)) |
| 426 | return true; |
| 427 | if (add(OtherOffset: Tmp.Offset)) |
| 428 | return true; |
| 429 | append_range(C&: Vars, R&: Tmp.Vars); |
| 430 | return false; |
| 431 | } |
| 432 | |
| 433 | /// Multiply all coefficients by \p Factor and return true if the operation |
| 434 | /// overflows, i.e. the new decomposition is invalid. |
| 435 | [[nodiscard]] bool mul(int64_t Factor) { |
| 436 | if (MulOverflow(X: Offset, Y: Factor, Result&: Offset)) |
| 437 | return true; |
| 438 | for (auto &Var : Vars) |
| 439 | if (MulOverflow(X: Var.Coefficient, Y: Factor, Result&: Var.Coefficient)) |
| 440 | return true; |
| 441 | return false; |
| 442 | } |
| 443 | }; |
| 444 | |
| 445 | // Variable and constant offsets for a chain of GEPs, with base pointer BasePtr. |
| 446 | struct OffsetResult { |
| 447 | Value *BasePtr; |
| 448 | APInt ConstantOffset; |
| 449 | SmallMapVector<Value *, APInt, 4> VariableOffsets; |
| 450 | GEPNoWrapFlags NW; |
| 451 | |
| 452 | OffsetResult() : BasePtr(nullptr), ConstantOffset(0, uint64_t(0)) {} |
| 453 | |
| 454 | OffsetResult(GEPOperator &GEP, const DataLayout &DL) |
| 455 | : BasePtr(GEP.getPointerOperand()), NW(GEP.getNoWrapFlags()) { |
| 456 | ConstantOffset = APInt(DL.getIndexTypeSizeInBits(Ty: BasePtr->getType()), 0); |
| 457 | } |
| 458 | }; |
| 459 | } // namespace |
| 460 | |
| 461 | // Try to collect variable and constant offsets for \p GEP, partly traversing |
| 462 | // nested GEPs. Returns an OffsetResult with nullptr as BasePtr of collecting |
| 463 | // the offset fails. |
| 464 | static OffsetResult collectOffsets(GEPOperator &GEP, const DataLayout &DL) { |
| 465 | OffsetResult Result(GEP, DL); |
| 466 | unsigned BitWidth = Result.ConstantOffset.getBitWidth(); |
| 467 | if (!GEP.collectOffset(DL, BitWidth, VariableOffsets&: Result.VariableOffsets, |
| 468 | ConstantOffset&: Result.ConstantOffset)) |
| 469 | return {}; |
| 470 | |
| 471 | // If we have a nested GEP, check if we can combine the constant offset of the |
| 472 | // inner GEP with the outer GEP. |
| 473 | if (auto *InnerGEP = dyn_cast<GetElementPtrInst>(Val: Result.BasePtr)) { |
| 474 | SmallMapVector<Value *, APInt, 4> VariableOffsets2; |
| 475 | APInt ConstantOffset2(BitWidth, 0); |
| 476 | bool CanCollectInner = InnerGEP->collectOffset( |
| 477 | DL, BitWidth, VariableOffsets&: VariableOffsets2, ConstantOffset&: ConstantOffset2); |
| 478 | // TODO: Support cases with more than 1 variable offset. |
| 479 | if (!CanCollectInner || Result.VariableOffsets.size() > 1 || |
| 480 | VariableOffsets2.size() > 1 || |
| 481 | (Result.VariableOffsets.size() >= 1 && VariableOffsets2.size() >= 1)) { |
| 482 | // More than 1 variable index, use outer result. |
| 483 | return Result; |
| 484 | } |
| 485 | Result.BasePtr = InnerGEP->getPointerOperand(); |
| 486 | Result.ConstantOffset += ConstantOffset2; |
| 487 | if (Result.VariableOffsets.size() == 0 && VariableOffsets2.size() == 1) |
| 488 | Result.VariableOffsets = std::move(VariableOffsets2); |
| 489 | Result.NW &= InnerGEP->getNoWrapFlags(); |
| 490 | } |
| 491 | return Result; |
| 492 | } |
| 493 | |
| 494 | static Decomposition decompose(Value *V, const ConstraintInfo &Info, |
| 495 | bool IsSigned, const DataLayout &DL); |
| 496 | |
| 497 | static bool canUseSExt(ConstantInt *CI) { |
| 498 | const APInt &Val = CI->getValue(); |
| 499 | return Val.sgt(RHS: MinSignedConstraintValue) && Val.slt(RHS: MaxConstraintValue); |
| 500 | } |
| 501 | |
| 502 | /// Returns true if \p Info implies that \p Op is in \p R, interpreting \p R as |
| 503 | /// a signed range if \p Signed is set and as an unsigned range otherwise. |
| 504 | static bool doesHoldInRange(const ConstraintInfo &Info, Value *Op, |
| 505 | const ConstantRange &R, bool Signed) { |
| 506 | if (R.isEmptySet() || (Signed ? R.isSignWrappedSet() : R.isWrappedSet())) |
| 507 | return false; |
| 508 | |
| 509 | if (R.isFullSet()) |
| 510 | return true; |
| 511 | |
| 512 | unsigned BitWidth = R.getBitWidth(); |
| 513 | APInt Min = Signed ? R.getSignedMin() : R.getUnsignedMin(); |
| 514 | APInt Max = Signed ? R.getSignedMax() : R.getUnsignedMax(); |
| 515 | APInt MinVal = Signed ? APInt::getSignedMinValue(numBits: BitWidth) |
| 516 | : APInt::getMinValue(numBits: BitWidth); |
| 517 | APInt MaxVal = Signed ? APInt::getSignedMaxValue(numBits: BitWidth) |
| 518 | : APInt::getMaxValue(numBits: BitWidth); |
| 519 | // Replace bound too large to be decomposed by the largest usable one. |
| 520 | if (!Signed && Max.uge(RHS: MaxConstraintValue)) |
| 521 | Max = APInt(BitWidth, MaxConstraintValue - 1); |
| 522 | |
| 523 | Type *Ty = Op->getType(); |
| 524 | if (Min != MinVal && |
| 525 | !Info.doesHold(Pred: Signed ? CmpInst::ICMP_SGE : CmpInst::ICMP_UGE, A: Op, |
| 526 | B: ConstantInt::get(Ty, V: Min))) |
| 527 | return false; |
| 528 | if (Max != MaxVal && |
| 529 | !Info.doesHold(Pred: Signed ? CmpInst::ICMP_SLE : CmpInst::ICMP_ULE, A: Op, |
| 530 | B: ConstantInt::get(Ty, V: Max))) |
| 531 | return false; |
| 532 | return true; |
| 533 | } |
| 534 | |
| 535 | /// Returns true if \p Opcode applied to \p Op0 and \p Op1 with \p NoWrapFlags |
| 536 | /// is known to not wrap in signed or unsigned, depending on \p Signed. |
| 537 | static bool isKnownNoWrap(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1, |
| 538 | unsigned NoWrapFlags, const ConstraintInfo &Info, |
| 539 | bool Signed) { |
| 540 | using OBO = OverflowingBinaryOperator; |
| 541 | |
| 542 | if (NoWrapFlags & (Signed ? OBO::NoSignedWrap : OBO::NoUnsignedWrap)) |
| 543 | return true; |
| 544 | |
| 545 | if (Opcode == Instruction::Sub) { |
| 546 | // Op0 - Op1 does not wrap unsigned if Op0 >=u Op1. |
| 547 | if (!Signed) |
| 548 | return Info.doesHold(Pred: CmpInst::ICMP_UGE, A: Op0, B: Op1); |
| 549 | |
| 550 | // Op0 - Op1 does not wrap signed if 0 <=s Op1 <=s Op0. |
| 551 | if (Info.isKnownNonNegative(V: Op1) && |
| 552 | Info.doesHold(Pred: CmpInst::ICMP_SGE, A: Op0, B: Op1)) |
| 553 | return true; |
| 554 | } |
| 555 | |
| 556 | if (!Signed && (NoWrapFlags & OBO::NoSignedWrap) && |
| 557 | (Opcode == Instruction::Shl || Info.isKnownNonNegative(V: Op1)) && |
| 558 | Info.isKnownNonNegative(V: Op0)) |
| 559 | return true; |
| 560 | |
| 561 | // For a constant Op1, the ranges of Op0 for which the operation does not |
| 562 | // wrap are known exactly; check if the systems imply one of them. |
| 563 | auto *C = dyn_cast<ConstantInt>(Val: Op1); |
| 564 | if (!C) |
| 565 | return false; |
| 566 | |
| 567 | return doesHoldInRange(Info, Op: Op0, |
| 568 | R: ConstantRange::makeExactNoWrapRegion( |
| 569 | BinOp: Opcode, Other: C->getValue(), |
| 570 | NoWrapKind: Signed ? OBO::NoSignedWrap : OBO::NoUnsignedWrap), |
| 571 | Signed); |
| 572 | } |
| 573 | |
| 574 | /// Returns true if \p V is known to not wrap in signed or unsigned, depending |
| 575 | /// on \p Signed. |
| 576 | static bool isKnownNoWrap(Value *V, const ConstraintInfo &Info, bool Signed) { |
| 577 | if (match(V, P: m_DisjointOr(L: m_Value(), R: m_Value()))) |
| 578 | return true; |
| 579 | |
| 580 | if (auto *Trunc = dyn_cast<TruncInst>(Val: V)) { |
| 581 | if (Signed) |
| 582 | return Trunc->hasNoSignedWrap(); |
| 583 | |
| 584 | // A trunc nsw only truncates without unsigned wrap if its operand is |
| 585 | // non-negative. |
| 586 | return Trunc->hasNoUnsignedWrap() || |
| 587 | (Trunc->hasNoSignedWrap() && |
| 588 | Info.isKnownNonNegative(V: Trunc->getOperand(i_nocapture: 0))); |
| 589 | } |
| 590 | |
| 591 | auto *BO = dyn_cast<OverflowingBinaryOperator>(Val: V); |
| 592 | return BO && |
| 593 | isKnownNoWrap(Opcode: static_cast<Instruction::BinaryOps>(BO->getOpcode()), |
| 594 | Op0: BO->getOperand(i_nocapture: 0), Op1: BO->getOperand(i_nocapture: 1), |
| 595 | NoWrapFlags: BO->getNoWrapKind(), Info, Signed); |
| 596 | } |
| 597 | |
| 598 | static Decomposition decomposeGEP(GEPOperator &GEP, const ConstraintInfo &Info, |
| 599 | bool IsSigned, const DataLayout &DL) { |
| 600 | // Do not reason about pointers where the index size is larger than 64 bits, |
| 601 | // as the coefficients used to encode constraints are 64 bit integers. |
| 602 | if (DL.getIndexTypeSizeInBits(Ty: GEP.getPointerOperand()->getType()) > 64) |
| 603 | return &GEP; |
| 604 | |
| 605 | assert(!IsSigned && "The logic below only supports decomposition for " |
| 606 | "unsigned predicates at the moment." ); |
| 607 | const auto &[BasePtr, ConstantOffset, VariableOffsets, NW] = |
| 608 | collectOffsets(GEP, DL); |
| 609 | // We support either plain gep nuw, or gep nusw with non-negative offset, |
| 610 | // which implies gep nuw. |
| 611 | if (!BasePtr || NW == GEPNoWrapFlags::none()) |
| 612 | return &GEP; |
| 613 | |
| 614 | // For a nuw-only GEP (nuw without nusw/inbounds), the offset must be |
| 615 | // interpreted as unsigned. |
| 616 | if (!NW.hasNoUnsignedSignedWrap() && ConstantOffset.isNegative()) |
| 617 | return &GEP; |
| 618 | |
| 619 | Decomposition Result(ConstantOffset.getSExtValue(), DecompEntry(1, BasePtr)); |
| 620 | for (auto [Index, Scale] : VariableOffsets) { |
| 621 | if (!NW.hasNoUnsignedWrap()) { |
| 622 | // Try to prove nuw from nusw and nneg. If the index cannot be proven |
| 623 | // non-negative, keep the GEP as-is instead of decomposing it. |
| 624 | assert(NW.hasNoUnsignedSignedWrap() && "Must have nusw flag" ); |
| 625 | if (!Info.isKnownNonNegative(V: Index)) |
| 626 | return &GEP; |
| 627 | } |
| 628 | |
| 629 | auto IdxResult = decompose(V: Index, Info, IsSigned, DL); |
| 630 | if (IdxResult.mul(Factor: Scale.getSExtValue())) |
| 631 | return &GEP; |
| 632 | if (Result.add(Other: IdxResult)) |
| 633 | return &GEP; |
| 634 | } |
| 635 | return Result; |
| 636 | } |
| 637 | |
| 638 | // Decomposes \p V into a constant offset + list of pairs { Coefficient, |
| 639 | // Variable } where Coefficient * Variable. The sum of the constant offset and |
| 640 | // pairs equals \p V. |
| 641 | // |
| 642 | // Looking through certain expressions is only valid if a pre-condition holds. |
| 643 | // Pre-conditions are checked against \p Info as needed. |
| 644 | static Decomposition decompose(Value *V, const ConstraintInfo &Info, |
| 645 | bool IsSigned, const DataLayout &DL) { |
| 646 | auto MergeResults = [&Info, IsSigned, |
| 647 | &DL](Value *A, Value *B, |
| 648 | bool IsSignedB) -> std::optional<Decomposition> { |
| 649 | auto ResA = decompose(V: A, Info, IsSigned, DL); |
| 650 | auto ResB = decompose(V: B, Info, IsSigned: IsSignedB, DL); |
| 651 | if (ResA.add(Other: ResB)) |
| 652 | return std::nullopt; |
| 653 | return ResA; |
| 654 | }; |
| 655 | |
| 656 | Type *Ty = V->getType()->getScalarType(); |
| 657 | if (Ty->isPointerTy() && !IsSigned) { |
| 658 | if (auto *GEP = dyn_cast<GEPOperator>(Val: V)) |
| 659 | return decomposeGEP(GEP&: *GEP, Info, IsSigned, DL); |
| 660 | if (isa<ConstantPointerNull>(Val: V)) |
| 661 | return int64_t(0); |
| 662 | |
| 663 | return V; |
| 664 | } |
| 665 | |
| 666 | // Don't handle integers > 64 bit. Our coefficients are 64-bit large, so |
| 667 | // coefficient add/mul may wrap, while the operation in the full bit width |
| 668 | // would not. |
| 669 | if (!Ty->isIntegerTy() || Ty->getIntegerBitWidth() > 64) |
| 670 | return V; |
| 671 | |
| 672 | if (auto *CI = dyn_cast<ConstantInt>(Val: V)) { |
| 673 | if (IsSigned) { |
| 674 | if (canUseSExt(CI)) |
| 675 | return CI->getSExtValue(); |
| 676 | } else if (!CI->uge(Num: MaxConstraintValue)) { |
| 677 | return int64_t(CI->getZExtValue()); |
| 678 | } |
| 679 | return V; |
| 680 | } |
| 681 | |
| 682 | Value *Op0; |
| 683 | Value *Op1; |
| 684 | ConstantInt *CI; |
| 685 | |
| 686 | if (match(V, P: m_ZExt(Op: m_Value(V&: Op0)))) { |
| 687 | // In the signed system, the ZExt must be non-negative. |
| 688 | if (IsSigned && !cast<ZExtInst>(Val: V)->hasNonNeg()) |
| 689 | return V; |
| 690 | V = Op0; |
| 691 | } else if (match(V, P: m_SExt(Op: m_Value(V&: Op0)))) { |
| 692 | // In the unsigned system, the SExt operand must be non-negative. |
| 693 | if (!IsSigned && !Info.isKnownNonNegative(V: Op0)) |
| 694 | return V; |
| 695 | V = Op0; |
| 696 | } else if (auto *Trunc = dyn_cast<TruncInst>(Val: V)) { |
| 697 | if (Trunc->getSrcTy()->getScalarSizeInBits() <= 64 && |
| 698 | isKnownNoWrap(V: Trunc, Info, Signed: IsSigned)) |
| 699 | V = Trunc->getOperand(i_nocapture: 0); |
| 700 | } |
| 701 | |
| 702 | if (match(V, P: m_AddLike(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
| 703 | if (isKnownNoWrap(V, Info, Signed: IsSigned)) { |
| 704 | if (auto Decomp = MergeResults(Op0, Op1, IsSigned)) |
| 705 | return *Decomp; |
| 706 | return V; |
| 707 | } |
| 708 | // In the unsigned system, adding a negative constant only wraps if Op0 is |
| 709 | // smaller than it. |
| 710 | if (!IsSigned && match(V: Op1, P: m_ConstantInt(CI)) && CI->isNegative() && |
| 711 | canUseSExt(CI) && |
| 712 | Info.doesHold(Pred: CmpInst::ICMP_UGE, A: Op0, |
| 713 | B: ConstantInt::get(Ty: Op0->getType(), V: -CI->getSExtValue()))) |
| 714 | if (auto Decomp = MergeResults(Op0, CI, /*IsSignedB=*/true)) |
| 715 | return *Decomp; |
| 716 | return V; |
| 717 | } |
| 718 | |
| 719 | // `xor %x, -1` is equivalent to `sub nsw -1, %x`. |
| 720 | if (IsSigned && match(V, P: m_Not(V: m_Value(V&: Op0)))) { |
| 721 | Decomposition Result(-1); |
| 722 | if (!Result.sub(Other: decompose(V: Op0, Info, IsSigned, DL))) |
| 723 | return Result; |
| 724 | return V; |
| 725 | } |
| 726 | |
| 727 | if (match(V, P: m_Sub(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
| 728 | if (isKnownNoWrap(V, Info, Signed: IsSigned)) { |
| 729 | auto ResA = decompose(V: Op0, Info, IsSigned, DL); |
| 730 | auto ResB = decompose(V: Op1, Info, IsSigned, DL); |
| 731 | if (!ResA.sub(Other: ResB)) |
| 732 | return ResA; |
| 733 | } |
| 734 | return V; |
| 735 | } |
| 736 | |
| 737 | if (match(V, P: m_Mul(L: m_Value(V&: Op0), R: m_ConstantInt(CI)))) { |
| 738 | // A negative constant is only a valid coefficient in the signed system; in |
| 739 | // the unsigned system the multiplier is the constant's unsigned value. |
| 740 | if (canUseSExt(CI) && (IsSigned || !CI->isNegative()) && |
| 741 | isKnownNoWrap(V, Info, Signed: IsSigned)) { |
| 742 | auto Result = decompose(V: Op0, Info, IsSigned, DL); |
| 743 | if (!Result.mul(Factor: CI->getSExtValue())) |
| 744 | return Result; |
| 745 | } |
| 746 | return V; |
| 747 | } |
| 748 | |
| 749 | if (match(V, P: m_Shl(L: m_Value(V&: Op0), R: m_ConstantInt(CI)))) { |
| 750 | // (shl x, shift) is (mul x, 1 << shift). The scale must fit in the signed |
| 751 | // coefficient, so reject shifts >= 63. Also reject a shift of bw-1, for |
| 752 | // which the product is not representable. |
| 753 | int64_t MaxShift = IsSigned ? Ty->getIntegerBitWidth() - 1 : 63; |
| 754 | if (!CI->isNegative() && CI->getSExtValue() < MaxShift && |
| 755 | isKnownNoWrap(V, Info, Signed: IsSigned)) { |
| 756 | auto Result = decompose(V: Op0, Info, IsSigned, DL); |
| 757 | if (!Result.mul(Factor: int64_t{1} << CI->getSExtValue())) |
| 758 | return Result; |
| 759 | } |
| 760 | return V; |
| 761 | } |
| 762 | |
| 763 | return V; |
| 764 | } |
| 765 | |
| 766 | /// Build the row for 'ADec <= BDec', using the indices from \p Value2Index. |
| 767 | /// Variables not in \p Value2Index are appended to \p NewVariables and get the |
| 768 | /// indices following the ones in \p Value2Index. Returns an empty row if the |
| 769 | /// coefficients overflow. |
| 770 | static RowTy getRowForLessEqual(const Decomposition &ADec, |
| 771 | const Decomposition &BDec, |
| 772 | const DenseMap<Value *, unsigned> &Value2Index, |
| 773 | SmallVectorImpl<Value *> &NewVariables) { |
| 774 | // Build the row, by first adding all coefficients from A and then subtracting |
| 775 | // all coefficients from B. |
| 776 | int64_t OffsetSum; |
| 777 | if (SubOverflow(X: BDec.Offset, Y: ADec.Offset, Result&: OffsetSum)) |
| 778 | return {}; |
| 779 | RowTy R(1, Entry(OffsetSum, 0)); |
| 780 | auto GetCoefficient = [&R](unsigned Idx) -> int64_t & { |
| 781 | // The entry for Idx, or the place to insert it at, is the first entry with |
| 782 | // an index >= Idx. |
| 783 | Entry *I = |
| 784 | find_if(Range: drop_begin(RangeOrContainer&: R), P: [Idx](const Entry &E) { return E.Id >= Idx; }); |
| 785 | if (I == R.end() || I->Id != Idx) |
| 786 | I = R.insert(I, Elt: Entry(0, Idx)); |
| 787 | return I->Coefficient; |
| 788 | }; |
| 789 | // First try to look up \p V in Value2Index and NewVariables. Otherwise add a |
| 790 | // new entry to NewVariables. |
| 791 | auto GetOrAddIndex = [&Value2Index, &NewVariables](Value *V) -> unsigned { |
| 792 | auto V2I = Value2Index.find(Val: V); |
| 793 | if (V2I != Value2Index.end()) |
| 794 | return V2I->second; |
| 795 | unsigned Idx = find(Range&: NewVariables, Val: V) - NewVariables.begin(); |
| 796 | if (Idx == NewVariables.size()) |
| 797 | NewVariables.push_back(Elt: V); |
| 798 | return Value2Index.size() + Idx + 1; |
| 799 | }; |
| 800 | for (const DecompEntry &KV : ADec.Vars) |
| 801 | GetCoefficient(GetOrAddIndex(KV.Variable)) += KV.Coefficient; |
| 802 | |
| 803 | for (const DecompEntry &KV : BDec.Vars) { |
| 804 | auto &Coeff = GetCoefficient(GetOrAddIndex(KV.Variable)); |
| 805 | if (SubOverflow(X: Coeff, Y: KV.Coefficient, Result&: Coeff)) |
| 806 | return {}; |
| 807 | } |
| 808 | |
| 809 | // Drop coefficients that cancelled out. |
| 810 | erase_if(C&: R, P: [](const Entry &E) { return E.Id != 0 && E.Coefficient == 0; }); |
| 811 | return R; |
| 812 | } |
| 813 | |
| 814 | ConstraintTy |
| 815 | ConstraintInfo::getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1, |
| 816 | SmallVectorImpl<Value *> &NewVariables, |
| 817 | bool ForceSignedSystem) const { |
| 818 | assert(NewVariables.empty() && "NewVariables must be empty when passed in" ); |
| 819 | assert((!ForceSignedSystem || CmpInst::isEquality(Pred)) && |
| 820 | "signed system can only be forced on eq/ne" ); |
| 821 | |
| 822 | bool IsEq = false; |
| 823 | bool IsNe = false; |
| 824 | |
| 825 | // Try to convert Pred to one of ULE/ULT/SLE/SLT. |
| 826 | switch (Pred) { |
| 827 | case CmpInst::ICMP_UGT: |
| 828 | case CmpInst::ICMP_UGE: |
| 829 | case CmpInst::ICMP_SGT: |
| 830 | case CmpInst::ICMP_SGE: { |
| 831 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
| 832 | std::swap(a&: Op0, b&: Op1); |
| 833 | break; |
| 834 | } |
| 835 | case CmpInst::ICMP_EQ: |
| 836 | if (!ForceSignedSystem && match(V: Op1, P: m_Zero())) { |
| 837 | Pred = CmpInst::ICMP_ULE; |
| 838 | } else { |
| 839 | IsEq = true; |
| 840 | Pred = CmpInst::ICMP_ULE; |
| 841 | } |
| 842 | break; |
| 843 | case CmpInst::ICMP_NE: |
| 844 | if (!ForceSignedSystem && match(V: Op1, P: m_Zero())) { |
| 845 | Pred = CmpInst::getSwappedPredicate(pred: CmpInst::ICMP_UGT); |
| 846 | std::swap(a&: Op0, b&: Op1); |
| 847 | } else { |
| 848 | IsNe = true; |
| 849 | Pred = CmpInst::ICMP_ULE; |
| 850 | } |
| 851 | break; |
| 852 | default: |
| 853 | break; |
| 854 | } |
| 855 | |
| 856 | if (Pred != CmpInst::ICMP_ULE && Pred != CmpInst::ICMP_ULT && |
| 857 | Pred != CmpInst::ICMP_SLE && Pred != CmpInst::ICMP_SLT) |
| 858 | return {}; |
| 859 | |
| 860 | bool IsSigned = ForceSignedSystem || CmpInst::isSigned(Pred); |
| 861 | auto &Value2Index = getValue2Index(Signed: IsSigned); |
| 862 | auto ADec = decompose(V: Op0->stripPointerCastsSameRepresentation(), Info: *this, |
| 863 | IsSigned, DL); |
| 864 | auto BDec = decompose(V: Op1->stripPointerCastsSameRepresentation(), Info: *this, |
| 865 | IsSigned, DL); |
| 866 | RowTy R = getRowForLessEqual(ADec, BDec, Value2Index, NewVariables); |
| 867 | if (R.empty()) |
| 868 | return {}; |
| 869 | |
| 870 | if (Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) |
| 871 | if (AddOverflow(X: R[0].Coefficient, Y: int64_t(-1), Result&: R[0].Coefficient)) |
| 872 | return {}; |
| 873 | |
| 874 | // Remove any new variable without a coefficient in the row. |
| 875 | unsigned NumV2I = Value2Index.size(); |
| 876 | NewVariables.truncate(N: R.back().Id > NumV2I ? R.back().Id - NumV2I : 0); |
| 877 | |
| 878 | return ConstraintTy(std::move(R), Value2Index.size() + NewVariables.size(), |
| 879 | IsSigned, IsEq, IsNe); |
| 880 | } |
| 881 | |
| 882 | ConstraintTy ConstraintInfo::getConstraintForSolving(CmpInst::Predicate Pred, |
| 883 | Value *Op0, |
| 884 | Value *Op1) const { |
| 885 | Constant *NullC = Constant::getNullValue(Ty: Op0->getType()); |
| 886 | // Handle trivially true compares directly to avoid adding V UGE 0 constraints |
| 887 | // for all variables in the unsigned system. |
| 888 | if ((Pred == CmpInst::ICMP_ULE && Op0 == NullC) || |
| 889 | (Pred == CmpInst::ICMP_UGE && Op1 == NullC)) { |
| 890 | // Return constraint that's trivially true. |
| 891 | return ConstraintTy(RowTy(1, Entry(0, 0)), /*NumVars=*/0, |
| 892 | /*IsSigned=*/false, /*IsEq=*/false, /*IsNe=*/false); |
| 893 | } |
| 894 | |
| 895 | // If both operands are known to be non-negative, change signed predicates to |
| 896 | // unsigned ones. This increases the reasoning effectiveness in combination |
| 897 | // with the signed <-> unsigned transfer logic. |
| 898 | if (CmpInst::isSigned(Pred) && |
| 899 | ::isKnownNonNegative(V: Op0, SQ: DL, /*Depth=*/MaxAnalysisRecursionDepth - 1) && |
| 900 | ::isKnownNonNegative(V: Op1, SQ: DL, /*Depth=*/MaxAnalysisRecursionDepth - 1)) |
| 901 | Pred = ICmpInst::getUnsignedPredicate(Pred); |
| 902 | |
| 903 | SmallVector<Value *> NewVariables; |
| 904 | ConstraintTy R = getConstraint(Pred, Op0, Op1, NewVariables); |
| 905 | if (!NewVariables.empty()) |
| 906 | return {}; |
| 907 | return R; |
| 908 | } |
| 909 | |
| 910 | std::optional<bool> |
| 911 | ConstraintTy::isImpliedBy(const ConstraintSystem &CS) const { |
| 912 | const auto &[SubCS, NewCoefficients] = CS.getSubSystem(R: Coefficients); |
| 913 | bool IsConditionImplied = SubCS.isConditionImplied(R: NewCoefficients); |
| 914 | |
| 915 | if (IsEq || IsNe) { |
| 916 | auto NegatedOrEqual = ConstraintSystem::negateOrEqual(R: NewCoefficients); |
| 917 | bool IsNegatedOrEqualImplied = |
| 918 | !NegatedOrEqual.empty() && SubCS.isConditionImplied(R: NegatedOrEqual); |
| 919 | |
| 920 | // In order to check that `%a == %b` is true (equality), both conditions `%a |
| 921 | // >= %b` and `%a <= %b` must hold true. When checking for equality (`IsEq` |
| 922 | // is true), we return true if they both hold, false in the other cases. |
| 923 | if (IsConditionImplied && IsNegatedOrEqualImplied) |
| 924 | return IsEq; |
| 925 | |
| 926 | auto Negated = ConstraintSystem::negate(R: NewCoefficients); |
| 927 | bool IsNegatedImplied = |
| 928 | !Negated.empty() && SubCS.isConditionImplied(R: Negated); |
| 929 | |
| 930 | auto StrictLessThan = ConstraintSystem::toStrictLessThan(R: NewCoefficients); |
| 931 | bool IsStrictLessThanImplied = |
| 932 | !StrictLessThan.empty() && SubCS.isConditionImplied(R: StrictLessThan); |
| 933 | |
| 934 | // In order to check that `%a != %b` is true (non-equality), either |
| 935 | // condition `%a > %b` or `%a < %b` must hold true. When checking for |
| 936 | // non-equality (`IsNe` is true), we return true if one of the two holds, |
| 937 | // false in the other cases. |
| 938 | if (IsNegatedImplied || IsStrictLessThanImplied) |
| 939 | return IsNe; |
| 940 | |
| 941 | return std::nullopt; |
| 942 | } |
| 943 | |
| 944 | if (IsConditionImplied) |
| 945 | return true; |
| 946 | |
| 947 | auto Negated = ConstraintSystem::negate(R: NewCoefficients); |
| 948 | auto IsNegatedImplied = !Negated.empty() && SubCS.isConditionImplied(R: Negated); |
| 949 | if (IsNegatedImplied) |
| 950 | return false; |
| 951 | |
| 952 | // Neither the condition nor its negated holds, did not prove anything. |
| 953 | return std::nullopt; |
| 954 | } |
| 955 | |
| 956 | bool ConstraintInfo::doesHold(CmpInst::Predicate Pred, Value *A, |
| 957 | Value *B) const { |
| 958 | auto R = getConstraintForSolving(Pred, Op0: A, Op1: B); |
| 959 | return !R.empty() && |
| 960 | getCS(Signed: R.IsSigned).isConditionImpliedInSubSystem(R: R.Coefficients); |
| 961 | } |
| 962 | |
| 963 | bool ConstraintInfo::isKnownNonNegative(Value *V) const { |
| 964 | if (auto *CI = dyn_cast<ConstantInt>(Val: V)) |
| 965 | return !CI->isNegative(); |
| 966 | return ::isKnownNonNegative(V, SQ: DL) || |
| 967 | doesHold(Pred: CmpInst::ICMP_SGE, A: V, B: ConstantInt::get(Ty: V->getType(), V: 0)); |
| 968 | } |
| 969 | |
| 970 | void ConstraintInfo::transferToOtherSystem( |
| 971 | CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn, |
| 972 | unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack) { |
| 973 | // Check if we can combine facts from the signed and unsigned systems to |
| 974 | // derive additional facts. |
| 975 | if (!A->getType()->isIntegerTy()) |
| 976 | return; |
| 977 | // FIXME: This currently depends on the order we add facts. Ideally we |
| 978 | // would first add all known facts and only then try to add additional |
| 979 | // facts. |
| 980 | switch (Pred) { |
| 981 | default: |
| 982 | break; |
| 983 | case CmpInst::ICMP_ULT: |
| 984 | case CmpInst::ICMP_ULE: |
| 985 | // If B is a signed positive constant, then A >=s 0 and A <s (or <=s) B. |
| 986 | if (isKnownNonNegative(V: B)) { |
| 987 | addFact(Pred: CmpInst::ICMP_SGE, A, B: ConstantInt::get(Ty: B->getType(), V: 0), NumIn, |
| 988 | NumOut, DFSInStack); |
| 989 | addFact(Pred: ICmpInst::getSignedPredicate(Pred), A, B, NumIn, NumOut, |
| 990 | DFSInStack); |
| 991 | } |
| 992 | break; |
| 993 | case CmpInst::ICMP_UGE: |
| 994 | case CmpInst::ICMP_UGT: |
| 995 | // If A is a signed positive constant, then B >=s 0 and A >s (or >=s) B. |
| 996 | if (isKnownNonNegative(V: A)) { |
| 997 | addFact(Pred: CmpInst::ICMP_SGE, A: B, B: ConstantInt::get(Ty: B->getType(), V: 0), NumIn, |
| 998 | NumOut, DFSInStack); |
| 999 | addFact(Pred: ICmpInst::getSignedPredicate(Pred), A, B, NumIn, NumOut, |
| 1000 | DFSInStack); |
| 1001 | } |
| 1002 | break; |
| 1003 | case CmpInst::ICMP_SLT: |
| 1004 | case CmpInst::ICMP_SLE: |
| 1005 | if (isKnownNonNegative(V: A)) |
| 1006 | addFact(Pred: ICmpInst::getUnsignedPredicate(Pred), A, B, NumIn, NumOut, |
| 1007 | DFSInStack); |
| 1008 | break; |
| 1009 | case CmpInst::ICMP_SGT: { |
| 1010 | if (doesHold(Pred: CmpInst::ICMP_SGE, A: B, B: Constant::getAllOnesValue(Ty: B->getType()))) |
| 1011 | addFact(Pred: CmpInst::ICMP_UGE, A, B: ConstantInt::get(Ty: B->getType(), V: 0), NumIn, |
| 1012 | NumOut, DFSInStack); |
| 1013 | if (isKnownNonNegative(V: B)) |
| 1014 | addFact(Pred: CmpInst::ICMP_UGT, A, B, NumIn, NumOut, DFSInStack); |
| 1015 | |
| 1016 | break; |
| 1017 | } |
| 1018 | case CmpInst::ICMP_SGE: |
| 1019 | if (isKnownNonNegative(V: B)) |
| 1020 | addFact(Pred: CmpInst::ICMP_UGE, A, B, NumIn, NumOut, DFSInStack); |
| 1021 | break; |
| 1022 | } |
| 1023 | } |
| 1024 | |
| 1025 | #ifndef NDEBUG |
| 1026 | |
| 1027 | static void dumpConstraint(ArrayRef<Entry> C, |
| 1028 | const DenseMap<Value *, unsigned> &Value2Index) { |
| 1029 | ConstraintSystem CS(Value2Index); |
| 1030 | CS.addRow(C, Value2Index.size()); |
| 1031 | CS.dump(); |
| 1032 | } |
| 1033 | #endif |
| 1034 | |
| 1035 | /// Splits the induction phi \p PN into the start value, coming from the loop |
| 1036 | /// predecessor \p LoopPred, and the backedge value, coming from inside the |
| 1037 | /// loop. Returns {nullptr, nullptr} if \p PN has other incoming values. |
| 1038 | static std::pair<Value *, Value *> |
| 1039 | getStartAndBackedgeValue(const PHINode &PN, const BasicBlock *LoopPred) { |
| 1040 | assert(PN.getBasicBlockIndex(LoopPred) >= 0 && |
| 1041 | "LoopPred must be a predecessor of the phi's block" ); |
| 1042 | if (PN.getNumIncomingValues() != 2) |
| 1043 | return {nullptr, nullptr}; |
| 1044 | unsigned StartIdx = PN.getIncomingBlock(i: 0) == LoopPred ? 0 : 1; |
| 1045 | return {PN.getIncomingValue(i: StartIdx), PN.getIncomingValue(i: 1 - StartIdx)}; |
| 1046 | } |
| 1047 | |
| 1048 | /// Matches an increment of \p PhiM by a constant offset, captured in \p Off. |
| 1049 | /// The increment must be a plain IR add or [u|s]add.with.overflow. |
| 1050 | template <typename PhiMatchTy> |
| 1051 | static auto m_IncrementOf(const PhiMatchTy &PhiM, const APInt *&Off) { |
| 1052 | return m_CombineOr( |
| 1053 | m_c_Add(PhiM, m_APInt(Res&: Off)), |
| 1054 | m_ExtractValue<0>(m_CombineOr( |
| 1055 | m_c_Intrinsic<Intrinsic::uadd_with_overflow>(PhiM, m_APInt(Res&: Off)), |
| 1056 | m_c_Intrinsic<Intrinsic::sadd_with_overflow>(PhiM, m_APInt(Res&: Off))))); |
| 1057 | } |
| 1058 | |
| 1059 | MonotonicInfo State::getMonotonicityInfo(PHINode &PN, Value *Step) { |
| 1060 | MonotonicInfo Info; |
| 1061 | const APInt *StepOffset = nullptr; |
| 1062 | if (match(V: Step, P: m_IncrementOf(PhiM: m_Specific(V: &PN), Off&: StepOffset))) { |
| 1063 | Info.Decreasing = StepOffset->isNegative(); |
| 1064 | if (const auto *Add = dyn_cast<OverflowingBinaryOperator>(Val: Step)) { |
| 1065 | Info.Unsigned = !Info.Decreasing && Add->hasNoUnsignedWrap(); |
| 1066 | Info.Signed = Add->hasNoSignedWrap(); |
| 1067 | } |
| 1068 | } else if (const auto *GEP = dyn_cast<GEPOperator>(Val: Step)) { |
| 1069 | // TODO: Handle the non-increasing direction, which needs a nusw GEP with a |
| 1070 | // negative constant offset. |
| 1071 | const DataLayout &DL = PN.getDataLayout(); |
| 1072 | APInt GEPOffset(DL.getIndexTypeSizeInBits(Ty: GEP->getType()), 0); |
| 1073 | Info.Unsigned = GEP->getPointerOperand() == &PN && |
| 1074 | (GEP->hasNoUnsignedWrap() || |
| 1075 | ((GEP->hasNoUnsignedSignedWrap() && |
| 1076 | GEP->accumulateConstantOffset(DL, Offset&: GEPOffset) && |
| 1077 | !GEPOffset.isNegative()))); |
| 1078 | } |
| 1079 | |
| 1080 | // Forming the SCEV of a phi is expensive, so only consult it for a PN + C |
| 1081 | // step whose no-wrap flags prove nothing. |
| 1082 | if (Info.Unsigned || Info.Signed || !StepOffset) |
| 1083 | return Info; |
| 1084 | |
| 1085 | const auto *AR = dyn_cast<SCEVAddRecExpr>(Val: SE.getSCEV(V: &PN)); |
| 1086 | if (!AR) |
| 1087 | return Info; |
| 1088 | ScalarEvolution::MonotonicPredicateType Expected = |
| 1089 | Info.Decreasing ? ScalarEvolution::MonotonicallyDecreasing |
| 1090 | : ScalarEvolution::MonotonicallyIncreasing; |
| 1091 | auto IsMonotonic = [&](CmpInst::Predicate Pred) { |
| 1092 | return SE.getMonotonicPredicateType(LHS: AR, Pred) == Expected; |
| 1093 | }; |
| 1094 | Info.Signed = IsMonotonic(CmpInst::ICMP_SGT); |
| 1095 | Info.Unsigned = !Info.Decreasing && IsMonotonic(CmpInst::ICMP_UGT); |
| 1096 | return Info; |
| 1097 | } |
| 1098 | |
| 1099 | void State::(BasicBlock &BB) { |
| 1100 | Loop *L = LI.getLoopFor(BB: &BB); |
| 1101 | if (!L || L->getHeader() != &BB) |
| 1102 | return; |
| 1103 | BasicBlock *LoopPred = L->getLoopPredecessor(); |
| 1104 | if (!LoopPred) |
| 1105 | return; |
| 1106 | |
| 1107 | DomTreeNode *DTN = DT.getNode(BB: &BB); |
| 1108 | for (PHINode &PN : BB.phis()) { |
| 1109 | if (!PN.getType()->isIntegerTy() && !PN.getType()->isPointerTy()) |
| 1110 | continue; |
| 1111 | |
| 1112 | auto [Start, Step] = getStartAndBackedgeValue(PN, LoopPred); |
| 1113 | if (!Start) |
| 1114 | continue; |
| 1115 | |
| 1116 | MonotonicInfo Info = getMonotonicityInfo(PN, Step); |
| 1117 | // Every variable in the unsigned system already has a `V >= 0` row, so a |
| 1118 | // zero start value would just duplicate it. |
| 1119 | if (match(V: Start, P: m_Zero())) |
| 1120 | Info.Unsigned = false; |
| 1121 | if (!Info.Unsigned && !Info.Signed) |
| 1122 | continue; |
| 1123 | |
| 1124 | // A non-decreasing induction cannot step below its start value, and a |
| 1125 | // non-increasing one cannot step above it. |
| 1126 | Value *LHS = &PN, *RHS = Start; |
| 1127 | if (Info.Decreasing) |
| 1128 | std::swap(a&: LHS, b&: RHS); |
| 1129 | CmpPredicate Pred(Info.Unsigned ? CmpInst::ICMP_UGE : CmpInst::ICMP_SGE, |
| 1130 | /*HasSameSign=*/Info.Unsigned && Info.Signed); |
| 1131 | WorkList.push_back(Elt: FactOrCheck::getConditionFact(DTN, Pred, Op0: LHS, Op1: RHS)); |
| 1132 | } |
| 1133 | } |
| 1134 | |
| 1135 | void State::addInfoForInductions(BasicBlock &BB) { |
| 1136 | auto *L = LI.getLoopFor(BB: &BB); |
| 1137 | if (!L) |
| 1138 | return; |
| 1139 | |
| 1140 | BasicBlock * = L->getHeader(); |
| 1141 | BasicBlock *Latch = L->getLoopLatch(); |
| 1142 | if (Header != &BB && Latch != &BB) |
| 1143 | return; |
| 1144 | |
| 1145 | // A is either a phi or a post-increment PN + C with constant step. For the |
| 1146 | // latter, extract the constant IncStep. |
| 1147 | Value *A; |
| 1148 | Value *B; |
| 1149 | PHINode *PN = nullptr; |
| 1150 | const APInt *IncStep = nullptr; |
| 1151 | CmpPredicate Pred; |
| 1152 | auto IndValue = |
| 1153 | m_Value(V&: A, P: m_CombineOr(Ps: m_Phi(PN), Ps: m_IncrementOf(PhiM: m_Phi(PN), Off&: IncStep))); |
| 1154 | |
| 1155 | auto *Br = dyn_cast<CondBrInst>(Val: BB.getTerminator()); |
| 1156 | if (!Br) |
| 1157 | return; |
| 1158 | |
| 1159 | auto CountingCmp = m_c_ICmp(Pred, L: IndValue, R: m_Value(V&: B)); |
| 1160 | std::optional<bool> PeeledOnEdge; |
| 1161 | if (!match(V: Br->getCondition(), P: CountingCmp)) { |
| 1162 | // Look through AND/OR, and remember which edge requires all operands to be |
| 1163 | // true. |
| 1164 | if (match(V: Br->getCondition(), P: m_c_LogicalAnd(L: CountingCmp, R: m_Value()))) |
| 1165 | PeeledOnEdge = true; |
| 1166 | else if (match(V: Br->getCondition(), P: m_c_LogicalOr(L: CountingCmp, R: m_Value()))) |
| 1167 | PeeledOnEdge = false; |
| 1168 | else |
| 1169 | return; |
| 1170 | } |
| 1171 | |
| 1172 | if (PN->getParent() != Header || PN->getNumIncomingValues() != 2 || |
| 1173 | !SE.isSCEVable(Ty: PN->getType())) |
| 1174 | return; |
| 1175 | |
| 1176 | // For latch conditions, we need to inject the condition that holds for the |
| 1177 | // next iteration into the header. We limit to post-inc conditions, for which |
| 1178 | // an original PN + Step != B condition results in a PN < B constraint in the |
| 1179 | // header, which also holds for the next loop iteration. This would no longer |
| 1180 | // be correct if the post-inc handling would inject a more precise PN + Step < |
| 1181 | // B constraint instead. |
| 1182 | if (&BB == Latch && !IncStep) |
| 1183 | return; |
| 1184 | |
| 1185 | bool ContinueOnTrue = |
| 1186 | Pred == CmpInst::ICMP_NE || ICmpInst::isLT(P: Pred) || ICmpInst::isLE(P: Pred); |
| 1187 | CmpInst::Predicate ContinuePred = |
| 1188 | ContinueOnTrue ? Pred.dropSameSign() : CmpInst::getInversePredicate(pred: Pred); |
| 1189 | BasicBlock *InLoopSucc = Br->getSuccessor(i: ContinueOnTrue ? 0 : 1); |
| 1190 | |
| 1191 | // The peeled condition only implies the compare on the edge where the |
| 1192 | // combined condition forces its operands, which must be the in-loop edge. |
| 1193 | if (PeeledOnEdge && *PeeledOnEdge != ContinueOnTrue) |
| 1194 | return; |
| 1195 | |
| 1196 | if (!L->contains(BB: InLoopSucc) || !L->isLoopExiting(BB: &BB)) |
| 1197 | return; |
| 1198 | |
| 1199 | BasicBlock *LoopPred = L->getLoopPredecessor(); |
| 1200 | if (!LoopPred || !L->isLoopInvariant(V: B)) |
| 1201 | return; |
| 1202 | |
| 1203 | auto [StartValue, Backedge] = getStartAndBackedgeValue(PN: *PN, LoopPred); |
| 1204 | DomTreeNode *DTN = DT.getNode(BB: InLoopSucc); |
| 1205 | |
| 1206 | if (ICmpInst::isRelational(P: ContinuePred)) { |
| 1207 | if (A != Backedge) |
| 1208 | return; |
| 1209 | |
| 1210 | // The latch condition ensures ContinuePred holds in the header on each |
| 1211 | // iteration other than the first. Together with a precondition on the start |
| 1212 | // value (StartValue ContinuePred B), we can add B as bound of PN. |
| 1213 | WorkList.push_back(Elt: FactOrCheck::getConditionFact( |
| 1214 | DTN, Pred: ContinuePred, Op0: PN, Op1: B, Precond: ConditionTy(ContinuePred, StartValue, B))); |
| 1215 | |
| 1216 | // A signed bound can be translated to the unsigned system if PN is signed |
| 1217 | // non-decreasing (StartValue s<= PN s< B) and StartValue u< B holds. |
| 1218 | // Then StartValue, PN and B must all have the same sign. |
| 1219 | if (ICmpInst::isSigned(Pred: ContinuePred)) { |
| 1220 | assert((ContinuePred == CmpInst::ICMP_SLT || |
| 1221 | ContinuePred == CmpInst::ICMP_SLE) && |
| 1222 | "Expected a signed less-than continuation predicate" ); |
| 1223 | MonotonicInfo Info = getMonotonicityInfo(PN&: *PN, Step: Backedge); |
| 1224 | if (Info.Signed && !Info.Decreasing) { |
| 1225 | CmpInst::Predicate UPred = ICmpInst::getUnsignedPredicate(Pred: ContinuePred); |
| 1226 | WorkList.push_back(Elt: FactOrCheck::getConditionFact( |
| 1227 | DTN, Pred: UPred, Op0: PN, Op1: B, Precond: ConditionTy(UPred, StartValue, B))); |
| 1228 | } |
| 1229 | } |
| 1230 | |
| 1231 | // A relational latch steps past B rather than landing on it, so none of the |
| 1232 | // reasoning below applies. |
| 1233 | return; |
| 1234 | } |
| 1235 | |
| 1236 | const APInt *StepOffset = nullptr; |
| 1237 | const SCEV *StartSCEV = nullptr; |
| 1238 | if (match(V: Backedge, P: m_c_Add(L: m_Specific(V: PN), R: m_APInt(Res&: StepOffset)))) { |
| 1239 | if (StepOffset->isZero()) |
| 1240 | return; |
| 1241 | } else { |
| 1242 | const SCEV *Expr = SE.getSCEV(V: PN); |
| 1243 | if (!match(S: Expr, |
| 1244 | P: m_scev_AffineAddRec(Op0: m_SCEV(V&: StartSCEV), Op1: m_scev_APInt(C&: StepOffset), |
| 1245 | L: m_SpecificLoop(L)))) |
| 1246 | return; |
| 1247 | } |
| 1248 | |
| 1249 | // If we looked through `PN + C`, only derive facts when that add is |
| 1250 | // really the induction's post-increment or post-decrement. |
| 1251 | if (IncStep && *IncStep != *StepOffset) |
| 1252 | return; |
| 1253 | |
| 1254 | MonotonicInfo Info = getMonotonicityInfo(PN&: *PN, Step: Backedge); |
| 1255 | |
| 1256 | // Handle negative steps. |
| 1257 | if (StepOffset->isNegative()) { |
| 1258 | // TODO: Extend to allow steps > -1. |
| 1259 | if (!(-*StepOffset).isOne()) |
| 1260 | return; |
| 1261 | |
| 1262 | // AR may wrap. |
| 1263 | // The loop exits once the compared value reaches B, that is at PN == B when |
| 1264 | // comparing the phi, and at PN == B + 1 for a post-decrement. Use |
| 1265 | // non-strict predicate for the former, and a strict one for the latter to |
| 1266 | // ensure the loop exits before wrapping. |
| 1267 | CmpInst::Predicate UPrecond = |
| 1268 | IncStep ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE; |
| 1269 | ConditionTy BBeforeStartUnsigned = {UPrecond, B, StartValue}; |
| 1270 | ConditionTy BBeforeStartSigned = {ICmpInst::getSignedPredicate(Pred: UPrecond), B, |
| 1271 | StartValue}; |
| 1272 | |
| 1273 | // AR may wrap, so both facts are conditional on B being below StartValue. |
| 1274 | // Add StartValue >= PN, which holds as the loop exits before wrapping. |
| 1275 | WorkList.push_back(Elt: FactOrCheck::getConditionFact( |
| 1276 | DTN, Pred: CmpInst::ICMP_UGE, Op0: StartValue, Op1: PN, Precond: BBeforeStartUnsigned)); |
| 1277 | if (!(Info.Decreasing && Info.Signed)) |
| 1278 | WorkList.push_back(Elt: FactOrCheck::getConditionFact( |
| 1279 | DTN, Pred: CmpInst::ICMP_SGE, Op0: StartValue, Op1: PN, Precond: BBeforeStartSigned)); |
| 1280 | // Add PN > B, which holds as the loop exits when reaching B. |
| 1281 | WorkList.push_back(Elt: FactOrCheck::getConditionFact(DTN, Pred: CmpInst::ICMP_UGT, Op0: PN, |
| 1282 | Op1: B, Precond: BBeforeStartUnsigned)); |
| 1283 | WorkList.push_back(Elt: FactOrCheck::getConditionFact(DTN, Pred: CmpInst::ICMP_SGT, Op0: PN, |
| 1284 | Op1: B, Precond: BBeforeStartSigned)); |
| 1285 | return; |
| 1286 | } |
| 1287 | |
| 1288 | // Make sure AR either steps by 1 or that the value we compare against is a |
| 1289 | // GEP based on the same start value and all offsets are a multiple of the |
| 1290 | // step size, to guarantee that the induction will reach the value. |
| 1291 | if (StepOffset->isZero() || StepOffset->isNegative()) |
| 1292 | return; |
| 1293 | |
| 1294 | if (!StepOffset->isOne()) { |
| 1295 | // Check whether B-Start is known to be a multiple of StepOffset. |
| 1296 | if (!StartSCEV) |
| 1297 | StartSCEV = SE.getSCEV(V: StartValue); |
| 1298 | const SCEV *BMinusStart = SE.getMinusSCEV(LHS: SE.getSCEV(V: B), RHS: StartSCEV); |
| 1299 | if (isa<SCEVCouldNotCompute>(Val: BMinusStart) || |
| 1300 | !SE.getConstantMultiple(S: BMinusStart).urem(RHS: *StepOffset).isZero()) |
| 1301 | return; |
| 1302 | } |
| 1303 | |
| 1304 | // We already established that B - Start is a multiple of Step above. The loop |
| 1305 | // exits once the compared value reaches B, that is at PN == B when comparing |
| 1306 | // the phi, and at PN + Step == B for a post-increment. Together with the |
| 1307 | // added precondition StartValue <= B for the former and the strict |
| 1308 | // StartValue < B for the latter (which implies StartValue + Step <= B), |
| 1309 | // neither PN nor the increment can wrap. |
| 1310 | CmpInst::Predicate UPrecond = IncStep ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE; |
| 1311 | ConditionTy StartBeforeBoundUnsigned = {UPrecond, StartValue, B}; |
| 1312 | ConditionTy StartBeforeBoundSigned = {ICmpInst::getSignedPredicate(Pred: UPrecond), |
| 1313 | StartValue, B}; |
| 1314 | |
| 1315 | // Add PN >= StartValue, as the loop exits before wrapping. |
| 1316 | if (!Info.Unsigned) |
| 1317 | WorkList.push_back(Elt: FactOrCheck::getConditionFact( |
| 1318 | DTN, Pred: CmpInst::ICMP_UGE, Op0: PN, Op1: StartValue, Precond: StartBeforeBoundUnsigned)); |
| 1319 | if (!Info.Signed) |
| 1320 | WorkList.push_back(Elt: FactOrCheck::getConditionFact( |
| 1321 | DTN, Pred: CmpInst::ICMP_SGE, Op0: PN, Op1: StartValue, Precond: StartBeforeBoundSigned)); |
| 1322 | // Add PN < B, as the loop exits once the compared value reaches B. |
| 1323 | WorkList.push_back(Elt: FactOrCheck::getConditionFact(DTN, Pred: CmpInst::ICMP_SLT, Op0: PN, |
| 1324 | Op1: B, Precond: StartBeforeBoundSigned)); |
| 1325 | WorkList.push_back(Elt: FactOrCheck::getConditionFact( |
| 1326 | DTN, Pred: CmpInst::ICMP_ULT, Op0: PN, Op1: B, Precond: StartBeforeBoundUnsigned)); |
| 1327 | |
| 1328 | // Try to add condition from the header or latch to the dedicated exit |
| 1329 | // blocks. When exiting either with EQ or NE, we know that the induction value |
| 1330 | // must be u<= B, as other exits may only exit earlier. |
| 1331 | assert(!StepOffset->isNegative() && "induction must be increasing" ); |
| 1332 | assert(ContinuePred == CmpInst::ICMP_NE && "unsupported predicate" ); |
| 1333 | SmallVector<BasicBlock *> ExitBBs; |
| 1334 | L->getExitBlocks(ExitBlocks&: ExitBBs); |
| 1335 | for (BasicBlock *EB : ExitBBs) { |
| 1336 | // Bail out on non-dedicated exits. |
| 1337 | if (DT.dominates(A: &BB, B: EB)) { |
| 1338 | WorkList.emplace_back(Args: FactOrCheck::getConditionFact( |
| 1339 | DTN: DT.getNode(BB: EB), Pred: CmpInst::ICMP_ULE, Op0: A, Op1: B, Precond: StartBeforeBoundUnsigned)); |
| 1340 | } |
| 1341 | } |
| 1342 | } |
| 1343 | |
| 1344 | static bool getConstraintFromMemoryAccess(GetElementPtrInst &GEP, |
| 1345 | uint64_t AccessSize, |
| 1346 | CmpPredicate &Pred, Value *&A, |
| 1347 | Value *&B, const DataLayout &DL, |
| 1348 | const TargetLibraryInfo &TLI) { |
| 1349 | auto Offset = collectOffsets(GEP&: cast<GEPOperator>(Val&: GEP), DL); |
| 1350 | if (!Offset.NW.hasNoUnsignedWrap()) |
| 1351 | return false; |
| 1352 | |
| 1353 | if (Offset.VariableOffsets.size() != 1) |
| 1354 | return false; |
| 1355 | |
| 1356 | uint64_t BitWidth = Offset.ConstantOffset.getBitWidth(); |
| 1357 | auto &[Index, Scale] = Offset.VariableOffsets.front(); |
| 1358 | // Bail out on non-canonical GEPs. |
| 1359 | if (Index->getType()->getScalarSizeInBits() != BitWidth) |
| 1360 | return false; |
| 1361 | |
| 1362 | ObjectSizeOpts Opts; |
| 1363 | // Workaround for gep inbounds, ptr null, idx. |
| 1364 | Opts.NullIsUnknownSize = true; |
| 1365 | // Be conservative since we are not clear on whether an out of bounds access |
| 1366 | // to the padding is UB or not. |
| 1367 | Opts.RoundToAlign = true; |
| 1368 | std::optional<TypeSize> Size = |
| 1369 | getBaseObjectSize(Ptr: Offset.BasePtr, DL, TLI: &TLI, Opts); |
| 1370 | if (!Size || Size->isScalable()) |
| 1371 | return false; |
| 1372 | |
| 1373 | // Index * Scale + ConstOffset + AccessSize <= AllocSize |
| 1374 | // With nuw flag, we know that the index addition doesn't have unsigned wrap. |
| 1375 | // If (AllocSize - (ConstOffset + AccessSize)) wraps around, there is no valid |
| 1376 | // value for Index. |
| 1377 | APInt MaxIndex = (APInt(BitWidth, Size->getFixedValue() - AccessSize, |
| 1378 | /*isSigned=*/false, /*implicitTrunc=*/true) - |
| 1379 | Offset.ConstantOffset) |
| 1380 | .udiv(RHS: Scale); |
| 1381 | Pred = ICmpInst::ICMP_ULE; |
| 1382 | A = Index; |
| 1383 | B = ConstantInt::get(Ty: Index->getType(), V: MaxIndex); |
| 1384 | return true; |
| 1385 | } |
| 1386 | |
| 1387 | /// Returns true if \p I is a candidate whose poison-generating flags may be |
| 1388 | /// strengthened using the constraint systems. |
| 1389 | static bool canStrengthenFlags(Instruction *I) { |
| 1390 | if (auto *Trunc = dyn_cast<TruncInst>(Val: I)) |
| 1391 | return Trunc->getType()->isIntegerTy() && Trunc->hasNoSignedWrap() && |
| 1392 | !Trunc->hasNoUnsignedWrap(); |
| 1393 | |
| 1394 | auto *BO = dyn_cast<BinaryOperator>(Val: I); |
| 1395 | if (!BO || !BO->getType()->isIntegerTy()) |
| 1396 | return false; |
| 1397 | |
| 1398 | switch (BO->getOpcode()) { |
| 1399 | case Instruction::Sub: |
| 1400 | if (BO->hasNoUnsignedWrap() && BO->hasNoSignedWrap()) |
| 1401 | return false; |
| 1402 | // A - B does not wrap unsigned if A >=u B, and does not wrap signed if |
| 1403 | // 0 <=s B <=s A. With a constant B, bounds on A can refine both flags. |
| 1404 | return true; |
| 1405 | case Instruction::Add: |
| 1406 | case Instruction::Mul: |
| 1407 | case Instruction::Shl: |
| 1408 | if (BO->hasNoUnsignedWrap() && BO->hasNoSignedWrap()) |
| 1409 | return false; |
| 1410 | // With a constant second operand, we can use bounds on the first operand to |
| 1411 | // refine no-wrap flags. Independently, nuw can be added for nsw if the |
| 1412 | // operands are non-negative. |
| 1413 | return isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 1)) || BO->hasNoSignedWrap(); |
| 1414 | default: |
| 1415 | return false; |
| 1416 | } |
| 1417 | } |
| 1418 | |
| 1419 | /// Try to strengthen \p I's poison generating flags using \p Info. Returns |
| 1420 | /// true if \p I was modified. |
| 1421 | static bool tryToStrengthenFlags(Instruction *I, ConstraintInfo &Info) { |
| 1422 | assert(canStrengthenFlags(I) && "not a candidate for flag strengthening" ); |
| 1423 | |
| 1424 | bool Changed = false; |
| 1425 | if (!I->hasNoSignedWrap() && isKnownNoWrap(V: I, Info, /*Signed=*/true)) { |
| 1426 | LLVM_DEBUG(dbgs() << "Adding nsw to " << *I << "\n" ); |
| 1427 | I->setHasNoSignedWrap(); |
| 1428 | Changed = true; |
| 1429 | } |
| 1430 | if (!I->hasNoUnsignedWrap() && isKnownNoWrap(V: I, Info, /*Signed=*/false)) { |
| 1431 | LLVM_DEBUG(dbgs() << "Adding nuw to " << *I << "\n" ); |
| 1432 | I->setHasNoUnsignedWrap(); |
| 1433 | Changed = true; |
| 1434 | } |
| 1435 | return Changed; |
| 1436 | } |
| 1437 | |
| 1438 | void State::addInfoFor(BasicBlock &BB) { |
| 1439 | addBoundsForHeaderInductions(BB); |
| 1440 | addInfoForInductions(BB); |
| 1441 | auto &DL = BB.getDataLayout(); |
| 1442 | |
| 1443 | Value *A, *B; |
| 1444 | CmpPredicate Pred; |
| 1445 | // True as long as the current instruction is guaranteed to execute. |
| 1446 | bool GuaranteedToExecute = true; |
| 1447 | // Queue conditions and assumes. |
| 1448 | for (Instruction &I : BB) { |
| 1449 | if (match(V: &I, P: m_ICmpLike(Pred, L: m_Value(), R: m_Value()))) { |
| 1450 | for (Use &U : I.uses()) { |
| 1451 | auto *UserI = getContextInstForUse(U); |
| 1452 | auto *DTN = DT.getNode(BB: UserI->getParent()); |
| 1453 | if (!DTN) |
| 1454 | continue; |
| 1455 | WorkList.push_back(Elt: FactOrCheck::getCheck(DTN, U: &U)); |
| 1456 | } |
| 1457 | continue; |
| 1458 | } |
| 1459 | |
| 1460 | auto AddFactFromMemoryAccess = [&](Value *Ptr, Type *AccessType) { |
| 1461 | auto *GEP = dyn_cast<GetElementPtrInst>(Val: Ptr); |
| 1462 | if (!GEP) |
| 1463 | return; |
| 1464 | TypeSize AccessSize = DL.getTypeStoreSize(Ty: AccessType); |
| 1465 | if (!AccessSize.isFixed()) |
| 1466 | return; |
| 1467 | if (GuaranteedToExecute) { |
| 1468 | if (getConstraintFromMemoryAccess(GEP&: *GEP, AccessSize: AccessSize.getFixedValue(), |
| 1469 | Pred, A, B, DL, TLI)) { |
| 1470 | // The memory access is guaranteed to execute when BB is entered, |
| 1471 | // hence the constraint holds on entry to BB. |
| 1472 | WorkList.emplace_back(Args: FactOrCheck::getConditionFact( |
| 1473 | DTN: DT.getNode(BB: I.getParent()), Pred, Op0: A, Op1: B)); |
| 1474 | } |
| 1475 | } else { |
| 1476 | WorkList.emplace_back( |
| 1477 | Args: FactOrCheck::getInstFact(DTN: DT.getNode(BB: I.getParent()), Inst: &I)); |
| 1478 | } |
| 1479 | }; |
| 1480 | |
| 1481 | if (auto *LI = dyn_cast<LoadInst>(Val: &I)) { |
| 1482 | if (!LI->isVolatile()) |
| 1483 | AddFactFromMemoryAccess(LI->getPointerOperand(), LI->getAccessType()); |
| 1484 | } |
| 1485 | if (auto *SI = dyn_cast<StoreInst>(Val: &I)) { |
| 1486 | if (!SI->isVolatile()) |
| 1487 | AddFactFromMemoryAccess(SI->getPointerOperand(), SI->getAccessType()); |
| 1488 | } |
| 1489 | |
| 1490 | auto *II = dyn_cast<IntrinsicInst>(Val: &I); |
| 1491 | Intrinsic::ID ID = II ? II->getIntrinsicID() : Intrinsic::not_intrinsic; |
| 1492 | switch (ID) { |
| 1493 | case Intrinsic::assume: { |
| 1494 | if (!match(V: I.getOperand(i: 0), P: m_ICmpLike(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) |
| 1495 | break; |
| 1496 | if (GuaranteedToExecute) { |
| 1497 | // The assume is guaranteed to execute when BB is entered, hence Cond |
| 1498 | // holds on entry to BB. |
| 1499 | WorkList.emplace_back(Args: FactOrCheck::getConditionFact( |
| 1500 | DTN: DT.getNode(BB: I.getParent()), Pred, Op0: A, Op1: B)); |
| 1501 | } else { |
| 1502 | WorkList.emplace_back( |
| 1503 | Args: FactOrCheck::getInstFact(DTN: DT.getNode(BB: I.getParent()), Inst: &I)); |
| 1504 | } |
| 1505 | break; |
| 1506 | } |
| 1507 | // Enqueue intrinsics for simplification. |
| 1508 | case Intrinsic::sadd_with_overflow: |
| 1509 | case Intrinsic::ssub_with_overflow: |
| 1510 | case Intrinsic::ucmp: |
| 1511 | case Intrinsic::scmp: |
| 1512 | WorkList.push_back( |
| 1513 | Elt: FactOrCheck::getCheck(DTN: DT.getNode(BB: &BB), I: cast<CallInst>(Val: &I))); |
| 1514 | break; |
| 1515 | // Enqueue the intrinsics to add extra info. |
| 1516 | case Intrinsic::umin: |
| 1517 | case Intrinsic::umax: |
| 1518 | case Intrinsic::smin: |
| 1519 | case Intrinsic::smax: |
| 1520 | case Intrinsic::usub_sat: |
| 1521 | // TODO: handle llvm.abs as well |
| 1522 | WorkList.push_back( |
| 1523 | Elt: FactOrCheck::getCheck(DTN: DT.getNode(BB: &BB), I: cast<CallInst>(Val: &I))); |
| 1524 | [[fallthrough]]; |
| 1525 | case Intrinsic::uadd_sat: |
| 1526 | // TODO: Check if it is possible to instead only added the min/max facts |
| 1527 | // when simplifying uses of the min/max intrinsics. |
| 1528 | if (!isGuaranteedNotToBePoison(V: &I)) |
| 1529 | break; |
| 1530 | [[fallthrough]]; |
| 1531 | case Intrinsic::abs: |
| 1532 | WorkList.push_back(Elt: FactOrCheck::getInstFact(DTN: DT.getNode(BB: &BB), Inst: &I)); |
| 1533 | break; |
| 1534 | } |
| 1535 | |
| 1536 | // Add facts from unsigned division, remainder and logical shift right, and |
| 1537 | // from signed remainder. |
| 1538 | // urem x, n: result < n and result <= x |
| 1539 | // udiv x, n: result <= x |
| 1540 | // lshr x, n: result <= x |
| 1541 | // srem x, n: result >= 0 and result <= x, if x >= 0 |
| 1542 | // result < n, if n > 0 |
| 1543 | if (auto *BO = dyn_cast<BinaryOperator>(Val: &I)) { |
| 1544 | if ((BO->getOpcode() == Instruction::URem || |
| 1545 | BO->getOpcode() == Instruction::UDiv || |
| 1546 | BO->getOpcode() == Instruction::LShr || |
| 1547 | BO->getOpcode() == Instruction::SRem) && |
| 1548 | isGuaranteedNotToBePoison(V: BO)) |
| 1549 | WorkList.push_back(Elt: FactOrCheck::getInstFact(DTN: DT.getNode(BB: &BB), Inst: BO)); |
| 1550 | } |
| 1551 | |
| 1552 | // Queue instructions whose flags may be strengthened, checked at the |
| 1553 | // closest point dominating all uses. |
| 1554 | if (canStrengthenFlags(I: &I)) { |
| 1555 | Instruction *CommonDom = findCommonDominatorOfUses(I, DT); |
| 1556 | WorkList.push_back(Elt: FactOrCheck::getCheck( |
| 1557 | DTN: DT.getNode(BB: CommonDom->getParent()), I: &I, ContextInst: CommonDom)); |
| 1558 | } |
| 1559 | |
| 1560 | GuaranteedToExecute &= isGuaranteedToTransferExecutionToSuccessor(I: &I); |
| 1561 | } |
| 1562 | |
| 1563 | if (auto *Switch = dyn_cast<SwitchInst>(Val: BB.getTerminator())) { |
| 1564 | for (auto &Case : Switch->cases()) { |
| 1565 | BasicBlock *Succ = Case.getCaseSuccessor(); |
| 1566 | Value *V = Case.getCaseValue(); |
| 1567 | if (!canAddSuccessor(BB, Succ)) |
| 1568 | continue; |
| 1569 | WorkList.emplace_back(Args: FactOrCheck::getConditionFact( |
| 1570 | DTN: DT.getNode(BB: Succ), Pred: CmpInst::ICMP_EQ, Op0: Switch->getCondition(), Op1: V)); |
| 1571 | } |
| 1572 | return; |
| 1573 | } |
| 1574 | |
| 1575 | auto *Br = dyn_cast<CondBrInst>(Val: BB.getTerminator()); |
| 1576 | if (!Br) |
| 1577 | return; |
| 1578 | |
| 1579 | Value *Cond = Br->getCondition(); |
| 1580 | |
| 1581 | // If the condition is a chain of ORs/AND and the successor only has the |
| 1582 | // current block as predecessor, queue conditions for the successor. |
| 1583 | Value *Op0, *Op1; |
| 1584 | if (match(V: Cond, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1))) || |
| 1585 | match(V: Cond, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
| 1586 | bool IsOr = match(V: Cond, P: m_LogicalOr()); |
| 1587 | bool IsAnd = match(V: Cond, P: m_LogicalAnd()); |
| 1588 | // If there's a select that matches both AND and OR, we need to commit to |
| 1589 | // one of the options. Arbitrarily pick OR. |
| 1590 | if (IsOr && IsAnd) |
| 1591 | IsAnd = false; |
| 1592 | |
| 1593 | BasicBlock *Successor = Br->getSuccessor(i: IsOr ? 1 : 0); |
| 1594 | if (canAddSuccessor(BB, Succ: Successor)) { |
| 1595 | SmallVector<Value *> CondWorkList; |
| 1596 | SmallPtrSet<Value *, 8> SeenCond; |
| 1597 | auto QueueValue = [&CondWorkList, &SeenCond](Value *V) { |
| 1598 | if (SeenCond.insert(Ptr: V).second) |
| 1599 | CondWorkList.push_back(Elt: V); |
| 1600 | }; |
| 1601 | QueueValue(Op1); |
| 1602 | QueueValue(Op0); |
| 1603 | while (!CondWorkList.empty()) { |
| 1604 | Value *Cur = CondWorkList.pop_back_val(); |
| 1605 | if (match(V: Cur, P: m_ICmpLike(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) { |
| 1606 | WorkList.emplace_back(Args: FactOrCheck::getConditionFact( |
| 1607 | DTN: DT.getNode(BB: Successor), |
| 1608 | Pred: IsOr ? CmpPredicate::getInverse(P: Pred) : Pred, Op0: A, Op1: B)); |
| 1609 | continue; |
| 1610 | } |
| 1611 | if (IsOr && match(V: Cur, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
| 1612 | QueueValue(Op1); |
| 1613 | QueueValue(Op0); |
| 1614 | continue; |
| 1615 | } |
| 1616 | if (IsAnd && match(V: Cur, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
| 1617 | QueueValue(Op1); |
| 1618 | QueueValue(Op0); |
| 1619 | continue; |
| 1620 | } |
| 1621 | } |
| 1622 | } |
| 1623 | return; |
| 1624 | } |
| 1625 | |
| 1626 | if (!match(V: Br->getCondition(), P: m_ICmpLike(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) |
| 1627 | return; |
| 1628 | if (canAddSuccessor(BB, Succ: Br->getSuccessor(i: 0))) |
| 1629 | WorkList.emplace_back(Args: FactOrCheck::getConditionFact( |
| 1630 | DTN: DT.getNode(BB: Br->getSuccessor(i: 0)), Pred, Op0: A, Op1: B)); |
| 1631 | if (canAddSuccessor(BB, Succ: Br->getSuccessor(i: 1))) |
| 1632 | WorkList.emplace_back(Args: FactOrCheck::getConditionFact( |
| 1633 | DTN: DT.getNode(BB: Br->getSuccessor(i: 1)), Pred: CmpPredicate::getInverse(P: Pred), Op0: A, Op1: B)); |
| 1634 | } |
| 1635 | |
| 1636 | #ifndef NDEBUG |
| 1637 | static void dumpUnpackedICmp(raw_ostream &OS, ICmpInst::Predicate Pred, |
| 1638 | Value *LHS, Value *RHS) { |
| 1639 | OS << "icmp " << Pred << ' '; |
| 1640 | LHS->printAsOperand(OS, /*PrintType=*/true); |
| 1641 | OS << ", " ; |
| 1642 | RHS->printAsOperand(OS, /*PrintType=*/false); |
| 1643 | } |
| 1644 | #endif |
| 1645 | |
| 1646 | namespace { |
| 1647 | /// Helper to keep track of a condition and if it should be treated as negated |
| 1648 | /// for reproducer construction. |
| 1649 | /// Pred == Predicate::BAD_ICMP_PREDICATE indicates that this entry is a |
| 1650 | /// placeholder to keep the ReproducerCondStack in sync with DFSInStack. |
| 1651 | struct ReproducerEntry { |
| 1652 | ICmpInst::Predicate Pred; |
| 1653 | Value *LHS; |
| 1654 | Value *RHS; |
| 1655 | |
| 1656 | ReproducerEntry(ICmpInst::Predicate Pred, Value *LHS, Value *RHS) |
| 1657 | : Pred(Pred), LHS(LHS), RHS(RHS) {} |
| 1658 | }; |
| 1659 | } // namespace |
| 1660 | |
| 1661 | /// Helper function to generate a reproducer function for simplifying \p Cond. |
| 1662 | /// The reproducer function contains a series of @llvm.assume calls, one for |
| 1663 | /// each condition in \p Stack. For each condition, the operand instruction are |
| 1664 | /// cloned until we reach operands that have an entry in \p Value2Index. Those |
| 1665 | /// will then be added as function arguments. \p DT is used to order cloned |
| 1666 | /// instructions. The reproducer function will get added to \p M, if it is |
| 1667 | /// non-null. Otherwise no reproducer function is generated. |
| 1668 | static void generateReproducer(Instruction *Cond, bool IsSigned, Module *M, |
| 1669 | ArrayRef<ReproducerEntry> Stack, |
| 1670 | ConstraintInfo &Info, DominatorTree &DT) { |
| 1671 | if (!M) |
| 1672 | return; |
| 1673 | |
| 1674 | LLVMContext &Ctx = Cond->getContext(); |
| 1675 | |
| 1676 | LLVM_DEBUG(dbgs() << "Creating reproducer for " << *Cond << "\n" ); |
| 1677 | |
| 1678 | ValueToValueMapTy Old2New; |
| 1679 | SmallVector<Value *> Args; |
| 1680 | SmallPtrSet<Value *, 8> Seen; |
| 1681 | // Traverse Cond and its operands recursively until we reach a value that's in |
| 1682 | // Value2Index or not an instruction, or not a operation that |
| 1683 | // ConstraintElimination can decompose. Such values will be considered as |
| 1684 | // external inputs to the reproducer, they are collected and added as function |
| 1685 | // arguments later. |
| 1686 | auto CollectArguments = [&](ArrayRef<Value *> Ops, bool IsSigned) { |
| 1687 | auto &Value2Index = Info.getValue2Index(Signed: IsSigned); |
| 1688 | SmallVector<Value *, 4> WorkList(Ops); |
| 1689 | while (!WorkList.empty()) { |
| 1690 | Value *V = WorkList.pop_back_val(); |
| 1691 | if (!Seen.insert(Ptr: V).second) |
| 1692 | continue; |
| 1693 | if (Old2New.find(Val: V) != Old2New.end()) |
| 1694 | continue; |
| 1695 | if (isa<Constant>(Val: V)) |
| 1696 | continue; |
| 1697 | |
| 1698 | auto *I = dyn_cast<Instruction>(Val: V); |
| 1699 | if (Value2Index.contains(Val: V) || !I || |
| 1700 | !isa<CmpInst, BinaryOperator, GEPOperator, CastInst>(Val: V)) { |
| 1701 | Old2New[V] = V; |
| 1702 | Args.push_back(Elt: V); |
| 1703 | LLVM_DEBUG(dbgs() << " found external input " << *V << "\n" ); |
| 1704 | } else { |
| 1705 | append_range(C&: WorkList, R: I->operands()); |
| 1706 | } |
| 1707 | } |
| 1708 | }; |
| 1709 | |
| 1710 | for (auto &Entry : Stack) |
| 1711 | if (Entry.Pred != ICmpInst::BAD_ICMP_PREDICATE) |
| 1712 | CollectArguments({Entry.LHS, Entry.RHS}, ICmpInst::isSigned(Pred: Entry.Pred)); |
| 1713 | CollectArguments(Cond, IsSigned); |
| 1714 | |
| 1715 | SmallVector<Type *> ParamTys; |
| 1716 | for (auto *P : Args) |
| 1717 | ParamTys.push_back(Elt: P->getType()); |
| 1718 | |
| 1719 | FunctionType *FTy = FunctionType::get(Result: Cond->getType(), Params: ParamTys, |
| 1720 | /*isVarArg=*/false); |
| 1721 | Function *F = Function::Create(Ty: FTy, Linkage: Function::ExternalLinkage, |
| 1722 | N: Cond->getModule()->getName() + |
| 1723 | Cond->getFunction()->getName() + "repro" , |
| 1724 | M); |
| 1725 | // Add arguments to the reproducer function for each external value collected. |
| 1726 | for (unsigned I = 0; I < Args.size(); ++I) { |
| 1727 | F->getArg(i: I)->setName(Args[I]->getName()); |
| 1728 | Old2New[Args[I]] = F->getArg(i: I); |
| 1729 | } |
| 1730 | |
| 1731 | BasicBlock *Entry = BasicBlock::Create(Context&: Ctx, Name: "entry" , Parent: F); |
| 1732 | IRBuilder<> Builder(Entry); |
| 1733 | Builder.CreateRet(V: Builder.getTrue()); |
| 1734 | Builder.SetInsertPoint(Entry->getTerminator()); |
| 1735 | |
| 1736 | // Clone instructions in \p Ops and their operands recursively until reaching |
| 1737 | // an value in Value2Index (external input to the reproducer). Update Old2New |
| 1738 | // mapping for the original and cloned instructions. Sort instructions to |
| 1739 | // clone by dominance, then insert the cloned instructions in the function. |
| 1740 | auto CloneInstructions = [&](ArrayRef<Value *> Ops, bool IsSigned) { |
| 1741 | SmallVector<Value *, 4> WorkList(Ops); |
| 1742 | SmallVector<Instruction *> ToClone; |
| 1743 | auto &Value2Index = Info.getValue2Index(Signed: IsSigned); |
| 1744 | while (!WorkList.empty()) { |
| 1745 | Value *V = WorkList.pop_back_val(); |
| 1746 | if (Old2New.find(Val: V) != Old2New.end()) |
| 1747 | continue; |
| 1748 | |
| 1749 | auto *I = dyn_cast<Instruction>(Val: V); |
| 1750 | if (!Value2Index.contains(Val: V) && I) { |
| 1751 | Old2New[V] = nullptr; |
| 1752 | ToClone.push_back(Elt: I); |
| 1753 | append_range(C&: WorkList, R: I->operands()); |
| 1754 | } |
| 1755 | } |
| 1756 | |
| 1757 | sort(C&: ToClone, |
| 1758 | Comp: [&DT](Instruction *A, Instruction *B) { return DT.dominates(Def: A, User: B); }); |
| 1759 | for (Instruction *I : ToClone) { |
| 1760 | Instruction *Cloned = I->clone(); |
| 1761 | Old2New[I] = Cloned; |
| 1762 | Old2New[I]->setName(I->getName()); |
| 1763 | Cloned->insertBefore(InsertPos: Builder.GetInsertPoint()); |
| 1764 | Cloned->dropUnknownNonDebugMetadata(); |
| 1765 | Cloned->setDebugLoc({}); |
| 1766 | } |
| 1767 | }; |
| 1768 | |
| 1769 | // Materialize the assumptions for the reproducer using the entries in Stack. |
| 1770 | // That is, first clone the operands of the condition recursively until we |
| 1771 | // reach an external input to the reproducer and add them to the reproducer |
| 1772 | // function. Then add an ICmp for the condition (with the inverse predicate if |
| 1773 | // the entry is negated) and an assert using the ICmp. |
| 1774 | for (auto &Entry : Stack) { |
| 1775 | if (Entry.Pred == ICmpInst::BAD_ICMP_PREDICATE) |
| 1776 | continue; |
| 1777 | |
| 1778 | LLVM_DEBUG(dbgs() << " Materializing assumption " ; |
| 1779 | dumpUnpackedICmp(dbgs(), Entry.Pred, Entry.LHS, Entry.RHS); |
| 1780 | dbgs() << "\n" ); |
| 1781 | CloneInstructions({Entry.LHS, Entry.RHS}, CmpInst::isSigned(Pred: Entry.Pred)); |
| 1782 | |
| 1783 | auto *Cmp = Builder.CreateICmp(P: Entry.Pred, LHS: Entry.LHS, RHS: Entry.RHS); |
| 1784 | Builder.CreateAssumption(Cond: Cmp); |
| 1785 | } |
| 1786 | |
| 1787 | // Finally, clone the condition to reproduce and remap instruction operands in |
| 1788 | // the reproducer using Old2New. |
| 1789 | CloneInstructions(Cond, IsSigned); |
| 1790 | Entry->getTerminator()->setOperand(i: 0, Val: Cond); |
| 1791 | remapInstructionsInBlocks(Blocks: {Entry}, VMap&: Old2New); |
| 1792 | |
| 1793 | assert(!verifyFunction(*F, &dbgs())); |
| 1794 | } |
| 1795 | |
| 1796 | static std::optional<bool> checkCondition(CmpInst::Predicate Pred, Value *A, |
| 1797 | Value *B, Instruction *CheckInst, |
| 1798 | ConstraintInfo &Info) { |
| 1799 | LLVM_DEBUG(dbgs() << "Checking " << *CheckInst << "\n" ); |
| 1800 | |
| 1801 | auto TryWithConstraint = [&](const ConstraintTy &R) -> std::optional<bool> { |
| 1802 | if (R.empty()) { |
| 1803 | LLVM_DEBUG(dbgs() << " failed to decompose condition\n" ); |
| 1804 | return std::nullopt; |
| 1805 | } |
| 1806 | |
| 1807 | auto &CSToUse = Info.getCS(Signed: R.IsSigned); |
| 1808 | if (auto ImpliedCondition = R.isImpliedBy(CS: CSToUse)) { |
| 1809 | if (!DebugCounter::shouldExecute(Counter&: EliminatedCounter)) |
| 1810 | return std::nullopt; |
| 1811 | LLVM_DEBUG({ |
| 1812 | dbgs() << "Condition " ; |
| 1813 | dumpUnpackedICmp(dbgs(), |
| 1814 | *ImpliedCondition ? Pred |
| 1815 | : CmpInst::getInversePredicate(Pred), |
| 1816 | A, B); |
| 1817 | dbgs() << " implied by dominating constraints\n" ; |
| 1818 | CSToUse.dump(); |
| 1819 | }); |
| 1820 | return ImpliedCondition; |
| 1821 | } |
| 1822 | return std::nullopt; |
| 1823 | }; |
| 1824 | |
| 1825 | auto R = Info.getConstraintForSolving(Pred, Op0: A, Op1: B); |
| 1826 | if (auto ImpliedCondition = TryWithConstraint(R)) |
| 1827 | return ImpliedCondition; |
| 1828 | |
| 1829 | // For non-negative operands unsigned queries can also be checked against the |
| 1830 | // signed system. |
| 1831 | if (CmpInst::isUnsigned(Pred) && A->getType()->isIntegerTy()) { |
| 1832 | SmallVector<Value *> NewVariables; |
| 1833 | auto SR = Info.getConstraint(Pred: ICmpInst::getSignedPredicate(Pred), Op0: A, Op1: B, |
| 1834 | NewVariables); |
| 1835 | if (NewVariables.empty() && !SR.empty() && Info.isKnownNonNegative(V: A) && |
| 1836 | Info.isKnownNonNegative(V: B)) |
| 1837 | if (auto ImpliedCondition = TryWithConstraint(SR)) |
| 1838 | return ImpliedCondition; |
| 1839 | } |
| 1840 | |
| 1841 | // Additionally, query the signed system for eq/ne predicates if we know about |
| 1842 | // A or B. |
| 1843 | if (CmpInst::isEquality(pred: Pred)) { |
| 1844 | const auto &Value2Index = Info.getValue2Index(/*Signed=*/true); |
| 1845 | if (!Value2Index.contains(Val: A) && !Value2Index.contains(Val: B)) |
| 1846 | return std::nullopt; |
| 1847 | |
| 1848 | SmallVector<Value *> NewVariables; |
| 1849 | auto SR = Info.getConstraint(Pred, Op0: A, Op1: B, NewVariables, |
| 1850 | /*ForceSignedSystem=*/true); |
| 1851 | if (NewVariables.empty()) |
| 1852 | if (auto ImpliedCondition = TryWithConstraint(SR)) |
| 1853 | return ImpliedCondition; |
| 1854 | } |
| 1855 | return std::nullopt; |
| 1856 | } |
| 1857 | |
| 1858 | static bool checkAndReplaceCondition( |
| 1859 | CmpPredicate Pred, Value *A, Value *B, Instruction *CheckInst, |
| 1860 | ConstraintInfo &Info, unsigned NumIn, unsigned NumOut, |
| 1861 | Instruction *ContextInst, Module *ReproducerModule, |
| 1862 | ArrayRef<ReproducerEntry> ReproducerCondStack, DominatorTree &DT, |
| 1863 | SmallVectorImpl<Instruction *> &ToRemove) { |
| 1864 | auto ReplaceCmpWithConstant = [&](Instruction *CheckInst, bool IsTrue) { |
| 1865 | generateReproducer(Cond: CheckInst, IsSigned: ICmpInst::isSigned(Pred), M: ReproducerModule, |
| 1866 | Stack: ReproducerCondStack, Info, DT); |
| 1867 | Constant *ConstantC = ConstantInt::getBool( |
| 1868 | Ty: CmpInst::makeCmpResultType(opnd_type: CheckInst->getType()), V: IsTrue); |
| 1869 | bool Changed = CheckInst->replaceUsesWithIf(New: ConstantC, ShouldReplace: [&](Use &U) { |
| 1870 | auto *UserI = getContextInstForUse(U); |
| 1871 | auto *DTN = DT.getNode(BB: UserI->getParent()); |
| 1872 | if (!DTN || DTN->getDFSNumIn() < NumIn || DTN->getDFSNumOut() > NumOut) |
| 1873 | return false; |
| 1874 | if (UserI->getParent() == ContextInst->getParent() && |
| 1875 | UserI->comesBefore(Other: ContextInst)) |
| 1876 | return false; |
| 1877 | |
| 1878 | // Conditions in an assume trivially simplify to true. Skip uses |
| 1879 | // in assume calls to not destroy the available information. |
| 1880 | auto *II = dyn_cast<IntrinsicInst>(Val: U.getUser()); |
| 1881 | return !II || II->getIntrinsicID() != Intrinsic::assume; |
| 1882 | }); |
| 1883 | NumCondsRemoved++; |
| 1884 | |
| 1885 | // Update the debug value records that satisfy the same condition used |
| 1886 | // in replaceUsesWithIf. |
| 1887 | SmallVector<DbgVariableRecord *> DVRUsers; |
| 1888 | findDbgUsers(V: CheckInst, DbgVariableRecords&: DVRUsers); |
| 1889 | |
| 1890 | for (auto *DVR : DVRUsers) { |
| 1891 | auto *DTN = DT.getNode(BB: DVR->getParent()); |
| 1892 | if (!DTN || DTN->getDFSNumIn() < NumIn || DTN->getDFSNumOut() > NumOut) |
| 1893 | continue; |
| 1894 | |
| 1895 | auto *MarkedI = DVR->getInstruction(); |
| 1896 | if (MarkedI->getParent() == ContextInst->getParent() && |
| 1897 | MarkedI->comesBefore(Other: ContextInst)) |
| 1898 | continue; |
| 1899 | |
| 1900 | DVR->replaceVariableLocationOp(OldValue: CheckInst, NewValue: ConstantC); |
| 1901 | } |
| 1902 | |
| 1903 | if (CheckInst->use_empty()) |
| 1904 | ToRemove.push_back(Elt: CheckInst); |
| 1905 | |
| 1906 | return Changed; |
| 1907 | }; |
| 1908 | |
| 1909 | if (auto ImpliedCondition = checkCondition(Pred, A, B, CheckInst, Info)) |
| 1910 | return ReplaceCmpWithConstant(CheckInst, *ImpliedCondition); |
| 1911 | |
| 1912 | // When the predicate is samesign and unsigned, we can also make use of the |
| 1913 | // signed predicate information. |
| 1914 | if (Pred.hasSameSign() && ICmpInst::isUnsigned(Pred)) |
| 1915 | if (auto ImpliedCondition = checkCondition( |
| 1916 | Pred: ICmpInst::getSignedPredicate(Pred), A, B, CheckInst, Info)) |
| 1917 | return ReplaceCmpWithConstant(CheckInst, *ImpliedCondition); |
| 1918 | |
| 1919 | return false; |
| 1920 | } |
| 1921 | |
| 1922 | static bool checkAndReplaceMinMax(MinMaxIntrinsic *MinMax, ConstraintInfo &Info, |
| 1923 | SmallVectorImpl<Instruction *> &ToRemove) { |
| 1924 | auto ReplaceMinMaxWithOperand = [&](MinMaxIntrinsic *MinMax, bool UseLHS) { |
| 1925 | // TODO: generate reproducer for min/max. |
| 1926 | MinMax->replaceAllUsesWith(V: MinMax->getOperand(i_nocapture: UseLHS ? 0 : 1)); |
| 1927 | ToRemove.push_back(Elt: MinMax); |
| 1928 | return true; |
| 1929 | }; |
| 1930 | |
| 1931 | ICmpInst::Predicate Pred = |
| 1932 | ICmpInst::getNonStrictPredicate(pred: MinMax->getPredicate()); |
| 1933 | if (auto ImpliedCondition = checkCondition( |
| 1934 | Pred, A: MinMax->getOperand(i_nocapture: 0), B: MinMax->getOperand(i_nocapture: 1), CheckInst: MinMax, Info)) |
| 1935 | return ReplaceMinMaxWithOperand(MinMax, *ImpliedCondition); |
| 1936 | if (auto ImpliedCondition = checkCondition( |
| 1937 | Pred, A: MinMax->getOperand(i_nocapture: 1), B: MinMax->getOperand(i_nocapture: 0), CheckInst: MinMax, Info)) |
| 1938 | return ReplaceMinMaxWithOperand(MinMax, !*ImpliedCondition); |
| 1939 | return false; |
| 1940 | } |
| 1941 | |
| 1942 | static bool checkAndReplaceCmp(CmpIntrinsic *I, ConstraintInfo &Info, |
| 1943 | SmallVectorImpl<Instruction *> &ToRemove) { |
| 1944 | Value *LHS = I->getOperand(i_nocapture: 0); |
| 1945 | Value *RHS = I->getOperand(i_nocapture: 1); |
| 1946 | if (checkCondition(Pred: I->getGTPredicate(), A: LHS, B: RHS, CheckInst: I, Info).value_or(u: false)) { |
| 1947 | I->replaceAllUsesWith(V: ConstantInt::get(Ty: I->getType(), V: 1)); |
| 1948 | ToRemove.push_back(Elt: I); |
| 1949 | return true; |
| 1950 | } |
| 1951 | if (checkCondition(Pred: I->getLTPredicate(), A: LHS, B: RHS, CheckInst: I, Info).value_or(u: false)) { |
| 1952 | I->replaceAllUsesWith(V: ConstantInt::getSigned(Ty: I->getType(), V: -1)); |
| 1953 | ToRemove.push_back(Elt: I); |
| 1954 | return true; |
| 1955 | } |
| 1956 | if (checkCondition(Pred: ICmpInst::ICMP_EQ, A: LHS, B: RHS, CheckInst: I, Info).value_or(u: false)) { |
| 1957 | I->replaceAllUsesWith(V: ConstantInt::get(Ty: I->getType(), V: 0)); |
| 1958 | ToRemove.push_back(Elt: I); |
| 1959 | return true; |
| 1960 | } |
| 1961 | return false; |
| 1962 | } |
| 1963 | |
| 1964 | /// Try to replace \p USub by a plain subtract, if \p Info proves it cannot |
| 1965 | /// saturate. Returns true if \p USub was replaced. |
| 1966 | static bool checkAndReplaceUSubSat(SaturatingInst *USub, ConstraintInfo &Info, |
| 1967 | SmallVectorImpl<Instruction *> &ToRemove) { |
| 1968 | // usub.sat(A, B) is A - B exactly when A >=u B. |
| 1969 | Value *A = USub->getLHS(); |
| 1970 | Value *B = USub->getRHS(); |
| 1971 | if (!checkCondition(Pred: CmpInst::ICMP_UGE, A, B, CheckInst: USub, Info).value_or(u: false)) |
| 1972 | return false; |
| 1973 | |
| 1974 | IRBuilder<> Builder(USub); |
| 1975 | Value *Sub = Builder.CreateSub(LHS: A, RHS: B, Name: "" , /*HasNUW=*/true, |
| 1976 | /*HasNSW=*/Info.isKnownNonNegative(V: A)); |
| 1977 | USub->replaceAllUsesWith(V: Sub); |
| 1978 | Sub->takeName(V: USub); |
| 1979 | ToRemove.push_back(Elt: USub); |
| 1980 | return true; |
| 1981 | } |
| 1982 | |
| 1983 | static void |
| 1984 | removeEntryFromStack(const StackEntry &E, ConstraintInfo &Info, |
| 1985 | Module *ReproducerModule, |
| 1986 | SmallVectorImpl<ReproducerEntry> &ReproducerCondStack, |
| 1987 | SmallVectorImpl<StackEntry> &DFSInStack) { |
| 1988 | Info.popLastConstraint(Signed: E.IsSigned); |
| 1989 | // Remove variables in the system that went out of scope. |
| 1990 | auto &Mapping = Info.getValue2Index(Signed: E.IsSigned); |
| 1991 | for (Value *V : E.ValuesToRelease) |
| 1992 | Mapping.erase(Val: V); |
| 1993 | Info.popLastNVariables(Signed: E.IsSigned, N: E.ValuesToRelease.size()); |
| 1994 | DFSInStack.pop_back(); |
| 1995 | if (ReproducerModule) |
| 1996 | ReproducerCondStack.pop_back(); |
| 1997 | } |
| 1998 | |
| 1999 | /// Check if either the first condition of an AND or OR is implied by the |
| 2000 | /// (negated in case of OR) second condition or vice versa. |
| 2001 | static bool checkOrAndOpImpliedByOther( |
| 2002 | FactOrCheck &CB, ConstraintInfo &Info, Module *ReproducerModule, |
| 2003 | SmallVectorImpl<ReproducerEntry> &ReproducerCondStack, |
| 2004 | SmallVectorImpl<StackEntry> &DFSInStack, |
| 2005 | SmallVectorImpl<Instruction *> &ToRemove) { |
| 2006 | Instruction *JoinOp = CB.getContextInst(); |
| 2007 | if (JoinOp->use_empty()) |
| 2008 | return false; |
| 2009 | |
| 2010 | Instruction *CmpToCheck = cast<Instruction>(Val: CB.getInstructionToSimplify()); |
| 2011 | unsigned OtherOpIdx = JoinOp->getOperand(i: 0) == CmpToCheck ? 1 : 0; |
| 2012 | |
| 2013 | // Don't try to simplify the first condition of a select by the second, as |
| 2014 | // this may make the select more poisonous than the original one. |
| 2015 | // TODO: check if the first operand may be poison. |
| 2016 | if (OtherOpIdx != 0 && isa<SelectInst>(Val: JoinOp)) |
| 2017 | return false; |
| 2018 | |
| 2019 | unsigned OldSize = DFSInStack.size(); |
| 2020 | llvm::scope_exit InfoRestorer([&]() { |
| 2021 | // Remove entries again. |
| 2022 | while (OldSize < DFSInStack.size()) { |
| 2023 | StackEntry E = DFSInStack.back(); |
| 2024 | removeEntryFromStack(E, Info, ReproducerModule, ReproducerCondStack, |
| 2025 | DFSInStack); |
| 2026 | } |
| 2027 | }); |
| 2028 | bool IsOr = match(V: JoinOp, P: m_LogicalOr()); |
| 2029 | SmallVector<Value *, 4> Worklist({JoinOp->getOperand(i: OtherOpIdx)}); |
| 2030 | // Do a traversal of the AND/OR tree to add facts from leaf compares. |
| 2031 | while (!Worklist.empty()) { |
| 2032 | Value *Val = Worklist.pop_back_val(); |
| 2033 | Value *LHS, *RHS; |
| 2034 | CmpPredicate Pred; |
| 2035 | if (match(V: Val, P: m_ICmpLike(Pred, L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
| 2036 | // For OR, check if the negated condition implies CmpToCheck. |
| 2037 | if (IsOr) |
| 2038 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 2039 | // Optimistically add fact from the other compares in the AND/OR. |
| 2040 | Info.addFact(Pred, A: LHS, B: RHS, NumIn: CB.NumIn, NumOut: CB.NumOut, DFSInStack); |
| 2041 | continue; |
| 2042 | } |
| 2043 | if (IsOr ? match(V: Val, P: m_LogicalOr(L: m_Value(V&: LHS), R: m_Value(V&: RHS))) |
| 2044 | : match(V: Val, P: m_LogicalAnd(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
| 2045 | Worklist.push_back(Elt: LHS); |
| 2046 | Worklist.push_back(Elt: RHS); |
| 2047 | } |
| 2048 | } |
| 2049 | if (OldSize == DFSInStack.size()) |
| 2050 | return false; |
| 2051 | |
| 2052 | Value *A, *B; |
| 2053 | CmpPredicate Pred; |
| 2054 | [[maybe_unused]] bool Matched = |
| 2055 | match(V: CmpToCheck, P: m_ICmpLike(Pred, L: m_Value(V&: A), R: m_Value(V&: B))); |
| 2056 | assert(Matched && "expected icmp-like match" ); |
| 2057 | // Check if the second condition can be simplified now. |
| 2058 | if (auto ImpliedCondition = checkCondition(Pred, A, B, CheckInst: CmpToCheck, Info)) { |
| 2059 | if (IsOr == *ImpliedCondition) |
| 2060 | JoinOp->replaceAllUsesWith( |
| 2061 | V: ConstantInt::getBool(Ty: JoinOp->getType(), V: *ImpliedCondition)); |
| 2062 | else |
| 2063 | JoinOp->replaceAllUsesWith(V: JoinOp->getOperand(i: OtherOpIdx)); |
| 2064 | ToRemove.push_back(Elt: JoinOp); |
| 2065 | return true; |
| 2066 | } |
| 2067 | |
| 2068 | return false; |
| 2069 | } |
| 2070 | |
| 2071 | void ConstraintInfo::addFact(CmpInst::Predicate Pred, Value *A, Value *B, |
| 2072 | unsigned NumIn, unsigned NumOut, |
| 2073 | SmallVectorImpl<StackEntry> &DFSInStack) { |
| 2074 | addFactImpl(Pred, A, B, NumIn, NumOut, DFSInStack, ForceSignedSystem: false); |
| 2075 | // If the Pred is eq/ne, also add the fact to signed system. |
| 2076 | if (CmpInst::isEquality(pred: Pred)) |
| 2077 | addFactImpl(Pred, A, B, NumIn, NumOut, DFSInStack, ForceSignedSystem: true); |
| 2078 | if (Pred == CmpInst::ICMP_NE) |
| 2079 | tightenBoundUsingNe(A, B, NumIn, NumOut, DFSInStack); |
| 2080 | } |
| 2081 | |
| 2082 | void ConstraintInfo::tightenBoundUsingNe( |
| 2083 | Value *A, Value *B, unsigned NumIn, unsigned NumOut, |
| 2084 | SmallVectorImpl<StackEntry> &DFSInStack) { |
| 2085 | if (!A->getType()->isIntegerTy()) |
| 2086 | return; |
| 2087 | |
| 2088 | for (bool IsSigned : {false, true}) { |
| 2089 | // In the unsigned system `A u>= 0` holds for every A, so getConstraint |
| 2090 | // already turned `A != 0` into `A u> 0`. |
| 2091 | if (!IsSigned && match(V: B, P: m_Zero())) |
| 2092 | continue; |
| 2093 | |
| 2094 | // Skip if there are any unknown variables. |
| 2095 | const auto &Value2Index = getValue2Index(Signed: IsSigned); |
| 2096 | if (any_of(Range: decompose(V: A, Info: *this, IsSigned, DL).Vars, |
| 2097 | P: [&Value2Index](const DecompEntry &E) { |
| 2098 | return !Value2Index.contains(Val: E.Variable); |
| 2099 | })) |
| 2100 | continue; |
| 2101 | |
| 2102 | // If the system implies `A >= B` then together with `A != B` we get the |
| 2103 | // strict `A > B`; symmetrically `A <= B` becomes `A < B`. |
| 2104 | CmpInst::Predicate GEPred = |
| 2105 | IsSigned ? CmpInst::ICMP_SGE : CmpInst::ICMP_UGE; |
| 2106 | CmpInst::Predicate LEPred = |
| 2107 | IsSigned ? CmpInst::ICMP_SLE : CmpInst::ICMP_ULE; |
| 2108 | for (CmpInst::Predicate NonStrict : {GEPred, LEPred}) { |
| 2109 | if (!doesHold(Pred: NonStrict, A, B)) |
| 2110 | continue; |
| 2111 | CmpInst::Predicate Strict = CmpInst::getStrictPredicate(pred: NonStrict); |
| 2112 | LLVM_DEBUG(dbgs() << "Tightening '" ; |
| 2113 | dumpUnpackedICmp(dbgs(), NonStrict, A, B); dbgs() << "' to '" ; |
| 2114 | dumpUnpackedICmp(dbgs(), Strict, A, B); |
| 2115 | dbgs() << "' using inequality\n" ); |
| 2116 | addFactImpl(Pred: Strict, A, B, NumIn, NumOut, DFSInStack, |
| 2117 | /*ForceSignedSystem=*/false); |
| 2118 | break; |
| 2119 | } |
| 2120 | } |
| 2121 | } |
| 2122 | |
| 2123 | void ConstraintInfo::addFactImpl(CmpInst::Predicate Pred, Value *A, Value *B, |
| 2124 | unsigned NumIn, unsigned NumOut, |
| 2125 | SmallVectorImpl<StackEntry> &DFSInStack, |
| 2126 | bool ForceSignedSystem) { |
| 2127 | SmallVector<Value *> NewVariables; |
| 2128 | auto R = getConstraint(Pred, Op0: A, Op1: B, NewVariables, ForceSignedSystem); |
| 2129 | |
| 2130 | // TODO: Support non-equality for facts as well. |
| 2131 | if (R.empty() || R.isNe()) |
| 2132 | return; |
| 2133 | |
| 2134 | LLVM_DEBUG(dbgs() << "Adding '" ; dumpUnpackedICmp(dbgs(), Pred, A, B); |
| 2135 | dbgs() << "'\n" ); |
| 2136 | auto &CSToUse = getCS(Signed: R.IsSigned); |
| 2137 | bool Added = CSToUse.addRow(R: R.Coefficients, NumVars: R.NumVars); |
| 2138 | if (!Added) |
| 2139 | return; |
| 2140 | |
| 2141 | // If R has been added to the system, add the new variables and queue it for |
| 2142 | // removal once it goes out-of-scope. |
| 2143 | SmallVector<Value *, 2> ValuesToRelease; |
| 2144 | auto &Value2Index = getValue2Index(Signed: R.IsSigned); |
| 2145 | for (Value *V : NewVariables) { |
| 2146 | Value2Index.try_emplace(Key: V, Args: Value2Index.size() + 1); |
| 2147 | ValuesToRelease.push_back(Elt: V); |
| 2148 | } |
| 2149 | |
| 2150 | LLVM_DEBUG({ |
| 2151 | dbgs() << " constraint: " ; |
| 2152 | dumpConstraint(R.Coefficients, getValue2Index(R.IsSigned)); |
| 2153 | dbgs() << "\n" ; |
| 2154 | }); |
| 2155 | |
| 2156 | DFSInStack.emplace_back(Args&: NumIn, Args&: NumOut, Args&: R.IsSigned, |
| 2157 | Args: std::move(ValuesToRelease)); |
| 2158 | |
| 2159 | if (!R.IsSigned) { |
| 2160 | for (Value *V : NewVariables) { |
| 2161 | // Add V > -1 constraints for all new variables. |
| 2162 | CSToUse.addRow(R: {Entry(0, 0), Entry(-1, Value2Index.at(Val: V))}, |
| 2163 | NumVars: Value2Index.size()); |
| 2164 | DFSInStack.emplace_back(Args&: NumIn, Args&: NumOut, Args&: R.IsSigned, |
| 2165 | Args: SmallVector<Value *, 2>()); |
| 2166 | } |
| 2167 | } |
| 2168 | |
| 2169 | if (R.isEq()) { |
| 2170 | // Also add the inverted constraint for equality constraints. |
| 2171 | for (Entry &E : R.Coefficients) |
| 2172 | if (MulOverflow(X: E.Coefficient, Y: int64_t(-1), Result&: E.Coefficient)) |
| 2173 | return; |
| 2174 | CSToUse.addRow(R: R.Coefficients, NumVars: R.NumVars); |
| 2175 | |
| 2176 | DFSInStack.emplace_back(Args&: NumIn, Args&: NumOut, Args&: R.IsSigned, |
| 2177 | Args: SmallVector<Value *, 2>()); |
| 2178 | } |
| 2179 | } |
| 2180 | |
| 2181 | /// Replace the uses of the overflow intrinsic \p II, which has been proven not |
| 2182 | /// to signed-overflow, by (Opcode A, B). |
| 2183 | static bool replaceOverflowUses(IntrinsicInst *II, |
| 2184 | Instruction::BinaryOps Opcode, Value *A, |
| 2185 | Value *B, |
| 2186 | SmallVectorImpl<Instruction *> &ToRemove) { |
| 2187 | bool Changed = false; |
| 2188 | IRBuilder<> Builder(II->getParent(), II->getIterator()); |
| 2189 | Value *Res = nullptr; |
| 2190 | for (User *U : make_early_inc_range(Range: II->users())) { |
| 2191 | if (match(V: U, P: m_ExtractValue<0>(V: m_Value()))) { |
| 2192 | if (!Res) |
| 2193 | Res = Builder.CreateNoWrapBinOp(Opc: Opcode, LHS: A, RHS: B, /*IsNUW=*/false, |
| 2194 | /*IsNSW=*/true); |
| 2195 | U->replaceAllUsesWith(V: Res); |
| 2196 | Changed = true; |
| 2197 | } else if (match(V: U, P: m_ExtractValue<1>(V: m_Value()))) { |
| 2198 | U->replaceAllUsesWith(V: Builder.getFalse()); |
| 2199 | Changed = true; |
| 2200 | } else |
| 2201 | continue; |
| 2202 | |
| 2203 | if (U->use_empty()) { |
| 2204 | auto *I = cast<Instruction>(Val: U); |
| 2205 | ToRemove.push_back(Elt: I); |
| 2206 | I->setOperand(i: 0, Val: PoisonValue::get(T: II->getType())); |
| 2207 | Changed = true; |
| 2208 | } |
| 2209 | } |
| 2210 | |
| 2211 | if (II->use_empty()) { |
| 2212 | // Do not erase II here: the worklist may still hold Uses of II's operands. |
| 2213 | for (Use &Arg : II->args()) |
| 2214 | Arg.set(PoisonValue::get(T: Arg->getType())); |
| 2215 | ToRemove.push_back(Elt: II); |
| 2216 | Changed = true; |
| 2217 | } |
| 2218 | return Changed; |
| 2219 | } |
| 2220 | |
| 2221 | static bool |
| 2222 | tryToSimplifyOverflowMath(IntrinsicInst *II, ConstraintInfo &Info, |
| 2223 | SmallVectorImpl<Instruction *> &ToRemove) { |
| 2224 | switch (II->getIntrinsicID()) { |
| 2225 | case Intrinsic::ssub_with_overflow: { |
| 2226 | // If A s>= B && B s>= 0, ssub.with.overflow(a, b) should not overflow and |
| 2227 | // can be simplified to a regular sub. |
| 2228 | Value *A = II->getArgOperand(i: 0); |
| 2229 | Value *B = II->getArgOperand(i: 1); |
| 2230 | if (!Info.doesHold(Pred: CmpInst::ICMP_SGE, A, B) || |
| 2231 | !Info.doesHold(Pred: CmpInst::ICMP_SGE, A: B, B: ConstantInt::get(Ty: A->getType(), V: 0))) |
| 2232 | return false; |
| 2233 | return replaceOverflowUses(II, Opcode: Instruction::Sub, A, B, ToRemove); |
| 2234 | } |
| 2235 | case Intrinsic::sadd_with_overflow: { |
| 2236 | Value *A = II->getArgOperand(i: 0); |
| 2237 | Value *B = II->getArgOperand(i: 1); |
| 2238 | auto *C = dyn_cast<ConstantInt>(Val: B); |
| 2239 | if (!C || |
| 2240 | !doesHoldInRange(Info, Op: A, |
| 2241 | R: ConstantRange::makeGuaranteedNoWrapRegion( |
| 2242 | BinOp: Instruction::Add, Other: ConstantRange(C->getValue()), |
| 2243 | NoWrapKind: OverflowingBinaryOperator::NoSignedWrap), |
| 2244 | /*Signed=*/true)) |
| 2245 | return false; |
| 2246 | return replaceOverflowUses(II, Opcode: Instruction::Add, A, B, ToRemove); |
| 2247 | } |
| 2248 | default: |
| 2249 | return false; |
| 2250 | } |
| 2251 | } |
| 2252 | |
| 2253 | static bool (Function &F, DominatorTree &DT, LoopInfo &LI, |
| 2254 | ScalarEvolution &SE, |
| 2255 | OptimizationRemarkEmitter &ORE, |
| 2256 | TargetLibraryInfo &TLI) { |
| 2257 | bool Changed = false; |
| 2258 | DT.updateDFSNumbers(); |
| 2259 | SmallVector<Value *> FunctionArgs(llvm::make_pointer_range(Range: F.args())); |
| 2260 | ConstraintInfo Info(F.getDataLayout(), FunctionArgs); |
| 2261 | State S(DT, LI, SE, TLI); |
| 2262 | std::unique_ptr<Module> ReproducerModule( |
| 2263 | DumpReproducers ? new Module(F.getName(), F.getContext()) : nullptr); |
| 2264 | |
| 2265 | // First, collect conditions implied by branches and blocks with their |
| 2266 | // Dominator DFS in and out numbers. |
| 2267 | for (BasicBlock &BB : F) { |
| 2268 | if (!DT.getNode(BB: &BB)) |
| 2269 | continue; |
| 2270 | S.addInfoFor(BB); |
| 2271 | } |
| 2272 | |
| 2273 | // Next, sort worklist by dominance, so that dominating conditions to check |
| 2274 | // and facts come before conditions and facts dominated by them. If a |
| 2275 | // condition to check and a fact have the same numbers, conditional facts come |
| 2276 | // first. Assume facts and checks are ordered according to their relative |
| 2277 | // order in the containing basic block. Also make sure conditions with |
| 2278 | // constant operands come before conditions without constant operands. This |
| 2279 | // increases the effectiveness of the current signed <-> unsigned fact |
| 2280 | // transfer logic. |
| 2281 | stable_sort(Range&: S.WorkList, C: [](const FactOrCheck &A, const FactOrCheck &B) { |
| 2282 | auto HasNoConstOp = [](const FactOrCheck &B) { |
| 2283 | Value *V0 = B.isConditionFact() ? B.Cond.Op0 : B.Inst->getOperand(i: 0); |
| 2284 | Value *V1 = B.isConditionFact() ? B.Cond.Op1 : B.Inst->getOperand(i: 1); |
| 2285 | return !isa<ConstantInt>(Val: V0) && !isa<ConstantInt>(Val: V1); |
| 2286 | }; |
| 2287 | // If both entries have the same In numbers, conditional facts come first. |
| 2288 | // Otherwise use the relative order in the basic block. |
| 2289 | if (A.NumIn == B.NumIn) { |
| 2290 | if (A.isConditionFact() && B.isConditionFact()) { |
| 2291 | bool NoConstOpA = HasNoConstOp(A); |
| 2292 | bool NoConstOpB = HasNoConstOp(B); |
| 2293 | return NoConstOpA < NoConstOpB; |
| 2294 | } |
| 2295 | if (A.isConditionFact()) |
| 2296 | return true; |
| 2297 | if (B.isConditionFact()) |
| 2298 | return false; |
| 2299 | auto *InstA = A.getContextInst(); |
| 2300 | auto *InstB = B.getContextInst(); |
| 2301 | return InstA->comesBefore(Other: InstB); |
| 2302 | } |
| 2303 | return A.NumIn < B.NumIn; |
| 2304 | }); |
| 2305 | |
| 2306 | SmallVector<Instruction *> ToRemove; |
| 2307 | |
| 2308 | // Finally, process ordered worklist and eliminate implied conditions. |
| 2309 | SmallVector<StackEntry, 16> DFSInStack; |
| 2310 | SmallVector<ReproducerEntry> ReproducerCondStack; |
| 2311 | for (FactOrCheck &CB : S.WorkList) { |
| 2312 | // First, pop entries from the stack that are out-of-scope for CB. Remove |
| 2313 | // the corresponding entry from the constraint system. |
| 2314 | while (!DFSInStack.empty()) { |
| 2315 | auto &E = DFSInStack.back(); |
| 2316 | LLVM_DEBUG(dbgs() << "Top of stack : " << E.NumIn << " " << E.NumOut |
| 2317 | << "\n" ); |
| 2318 | LLVM_DEBUG(dbgs() << "CB: " << CB.NumIn << " " << CB.NumOut << "\n" ); |
| 2319 | assert(E.NumIn <= CB.NumIn); |
| 2320 | if (CB.NumOut <= E.NumOut) |
| 2321 | break; |
| 2322 | LLVM_DEBUG({ |
| 2323 | dbgs() << "Removing " ; |
| 2324 | dumpConstraint(Info.getCS(E.IsSigned).getLastConstraint(), |
| 2325 | Info.getValue2Index(E.IsSigned)); |
| 2326 | dbgs() << "\n" ; |
| 2327 | }); |
| 2328 | removeEntryFromStack(E, Info, ReproducerModule: ReproducerModule.get(), ReproducerCondStack, |
| 2329 | DFSInStack); |
| 2330 | } |
| 2331 | |
| 2332 | CmpPredicate Pred; |
| 2333 | Value *A, *B; |
| 2334 | // For a block, check if any CmpInsts become known based on the current set |
| 2335 | // of constraints. |
| 2336 | if (CB.isCheck()) { |
| 2337 | Instruction *Inst = CB.getInstructionToSimplify(); |
| 2338 | if (!Inst) |
| 2339 | continue; |
| 2340 | if (canStrengthenFlags(I: Inst)) { |
| 2341 | Changed |= tryToStrengthenFlags(I: Inst, Info); |
| 2342 | continue; |
| 2343 | } |
| 2344 | LLVM_DEBUG(dbgs() << "Processing condition to simplify: " << *Inst |
| 2345 | << "\n" ); |
| 2346 | if (auto *II = dyn_cast<WithOverflowInst>(Val: Inst)) { |
| 2347 | Changed |= tryToSimplifyOverflowMath(II, Info, ToRemove); |
| 2348 | } else if (match(V: Inst, P: m_ICmpLike(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) { |
| 2349 | bool Simplified = checkAndReplaceCondition( |
| 2350 | Pred, A, B, CheckInst: Inst, Info, NumIn: CB.NumIn, NumOut: CB.NumOut, ContextInst: CB.getContextInst(), |
| 2351 | ReproducerModule: ReproducerModule.get(), ReproducerCondStack, DT&: S.DT, ToRemove); |
| 2352 | if (!Simplified && |
| 2353 | match(V: CB.getContextInst(), P: m_LogicalOp(L: m_Value(), R: m_Value()))) { |
| 2354 | Simplified = checkOrAndOpImpliedByOther( |
| 2355 | CB, Info, ReproducerModule: ReproducerModule.get(), ReproducerCondStack, DFSInStack, |
| 2356 | ToRemove); |
| 2357 | } |
| 2358 | Changed |= Simplified; |
| 2359 | } else if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Inst)) { |
| 2360 | Changed |= checkAndReplaceMinMax(MinMax, Info, ToRemove); |
| 2361 | } else if (auto *CmpIntr = dyn_cast<CmpIntrinsic>(Val: Inst)) { |
| 2362 | Changed |= checkAndReplaceCmp(I: CmpIntr, Info, ToRemove); |
| 2363 | } else if (match(V: Inst, P: m_Intrinsic<Intrinsic::usub_sat>())) { |
| 2364 | Changed |= |
| 2365 | checkAndReplaceUSubSat(USub: cast<SaturatingInst>(Val: Inst), Info, ToRemove); |
| 2366 | } |
| 2367 | continue; |
| 2368 | } |
| 2369 | |
| 2370 | auto AddFact = [&](CmpPredicate Pred, Value *A, Value *B) { |
| 2371 | LLVM_DEBUG(dbgs() << "Processing fact to add to the system: " ; |
| 2372 | dumpUnpackedICmp(dbgs(), Pred, A, B); dbgs() << "\n" ); |
| 2373 | if (Info.getCS(Signed: CmpInst::isSigned(Pred)).size() > MaxRows) { |
| 2374 | LLVM_DEBUG( |
| 2375 | dbgs() |
| 2376 | << "Skip adding constraint because system has too many rows.\n" ); |
| 2377 | return; |
| 2378 | } |
| 2379 | |
| 2380 | Info.addFact(Pred, A, B, NumIn: CB.NumIn, NumOut: CB.NumOut, DFSInStack); |
| 2381 | if (ReproducerModule && DFSInStack.size() > ReproducerCondStack.size()) |
| 2382 | ReproducerCondStack.emplace_back(Args&: Pred, Args&: A, Args&: B); |
| 2383 | |
| 2384 | if (ICmpInst::isRelational(P: Pred)) { |
| 2385 | // If samesign is present on the ICmp, simply flip the sign of the |
| 2386 | // predicate, transferring the information from the signed system to the |
| 2387 | // unsigned system, and viceversa. |
| 2388 | if (Pred.hasSameSign()) |
| 2389 | Info.addFact(Pred: ICmpInst::getFlippedSignednessPredicate(Pred), A, B, |
| 2390 | NumIn: CB.NumIn, NumOut: CB.NumOut, DFSInStack); |
| 2391 | else |
| 2392 | Info.transferToOtherSystem(Pred, A, B, NumIn: CB.NumIn, NumOut: CB.NumOut, |
| 2393 | DFSInStack); |
| 2394 | } |
| 2395 | |
| 2396 | // (X | Y) >s -1 implies X >s -1 and Y >s -1, because the sign bit of an |
| 2397 | // OR is the OR of the operand sign bits. Similarly, (X & Y) <s 0 implies |
| 2398 | // X <s 0 and Y <s 0. Look through these canonical forms produced by |
| 2399 | // InstCombine so the sign facts on the operands are available to the |
| 2400 | // solver. |
| 2401 | if ((Pred == CmpInst::ICMP_SGT && match(V: B, P: m_AllOnes())) || |
| 2402 | (Pred == CmpInst::ICMP_SLT && match(V: B, P: m_Zero()))) { |
| 2403 | unsigned Opc = |
| 2404 | Pred == CmpInst::ICMP_SGT ? Instruction::Or : Instruction::And; |
| 2405 | SmallVector<Value *> Worklist = {A}; |
| 2406 | SmallPtrSet<Value *, 4> Seen; |
| 2407 | while (!Worklist.empty()) { |
| 2408 | Value *Cur = Worklist.pop_back_val(); |
| 2409 | auto *BO = dyn_cast<BinaryOperator>(Val: Cur); |
| 2410 | if (!BO || BO->getOpcode() != Opc) |
| 2411 | continue; |
| 2412 | for (Value *Op : {BO->getOperand(i_nocapture: 0), BO->getOperand(i_nocapture: 1)}) { |
| 2413 | if (!Seen.insert(Ptr: Op).second) |
| 2414 | continue; |
| 2415 | Worklist.push_back(Elt: Op); |
| 2416 | Info.addFact(Pred, A: Op, B, NumIn: CB.NumIn, NumOut: CB.NumOut, DFSInStack); |
| 2417 | } |
| 2418 | } |
| 2419 | } |
| 2420 | |
| 2421 | if (ReproducerModule && DFSInStack.size() > ReproducerCondStack.size()) { |
| 2422 | // Add dummy entries to ReproducerCondStack to keep it in sync with |
| 2423 | // DFSInStack. |
| 2424 | for (unsigned I = 0, |
| 2425 | E = (DFSInStack.size() - ReproducerCondStack.size()); |
| 2426 | I < E; ++I) { |
| 2427 | ReproducerCondStack.emplace_back(Args: ICmpInst::BAD_ICMP_PREDICATE, |
| 2428 | Args: nullptr, Args: nullptr); |
| 2429 | } |
| 2430 | } |
| 2431 | }; |
| 2432 | |
| 2433 | if (!CB.isConditionFact()) { |
| 2434 | Value *X; |
| 2435 | if (match(V: CB.Inst, P: m_Intrinsic<Intrinsic::abs>(Ops: m_Value(V&: X)))) { |
| 2436 | // If is_int_min_poison is true then we may assume llvm.abs >= 0. |
| 2437 | if (cast<ConstantInt>(Val: CB.Inst->getOperand(i: 1))->isOne()) |
| 2438 | AddFact(CmpInst::ICMP_SGE, CB.Inst, |
| 2439 | ConstantInt::get(Ty: CB.Inst->getType(), V: 0)); |
| 2440 | AddFact(CmpInst::ICMP_SGE, CB.Inst, X); |
| 2441 | continue; |
| 2442 | } |
| 2443 | |
| 2444 | if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: CB.Inst)) { |
| 2445 | Pred = ICmpInst::getNonStrictPredicate(pred: MinMax->getPredicate()); |
| 2446 | AddFact(Pred, MinMax, MinMax->getLHS()); |
| 2447 | AddFact(Pred, MinMax, MinMax->getRHS()); |
| 2448 | continue; |
| 2449 | } |
| 2450 | if (auto *USatI = dyn_cast<SaturatingInst>(Val: CB.Inst)) { |
| 2451 | switch (USatI->getIntrinsicID()) { |
| 2452 | default: |
| 2453 | llvm_unreachable("Unexpected intrinsic." ); |
| 2454 | case Intrinsic::uadd_sat: |
| 2455 | AddFact(ICmpInst::ICMP_UGE, USatI, USatI->getLHS()); |
| 2456 | AddFact(ICmpInst::ICMP_UGE, USatI, USatI->getRHS()); |
| 2457 | break; |
| 2458 | case Intrinsic::usub_sat: |
| 2459 | AddFact(ICmpInst::ICMP_ULE, USatI, USatI->getLHS()); |
| 2460 | break; |
| 2461 | } |
| 2462 | continue; |
| 2463 | } |
| 2464 | |
| 2465 | if (auto *BO = dyn_cast<BinaryOperator>(Val: CB.Inst)) { |
| 2466 | if (BO->getOpcode() == Instruction::URem) { |
| 2467 | // urem x, n: result < n (remainder is always less than divisor) |
| 2468 | AddFact(CmpInst::ICMP_ULT, BO, BO->getOperand(i_nocapture: 1)); |
| 2469 | // urem x, n: result <= x (remainder is at most the dividend) |
| 2470 | AddFact(CmpInst::ICMP_ULE, BO, BO->getOperand(i_nocapture: 0)); |
| 2471 | continue; |
| 2472 | } |
| 2473 | if (BO->getOpcode() == Instruction::UDiv) { |
| 2474 | // udiv x, n: result <= x (quotient is at most the dividend) |
| 2475 | AddFact(CmpInst::ICMP_ULE, BO, BO->getOperand(i_nocapture: 0)); |
| 2476 | continue; |
| 2477 | } |
| 2478 | if (BO->getOpcode() == Instruction::LShr) { |
| 2479 | // lshr x, n: result <= x (right shift cannot increase the value) |
| 2480 | AddFact(CmpInst::ICMP_ULE, BO, BO->getOperand(i_nocapture: 0)); |
| 2481 | continue; |
| 2482 | } |
| 2483 | if (BO->getOpcode() == Instruction::SRem) { |
| 2484 | Value *X = BO->getOperand(i_nocapture: 0); |
| 2485 | Value *N = BO->getOperand(i_nocapture: 1); |
| 2486 | Constant *Zero = Constant::getNullValue(Ty: BO->getType()); |
| 2487 | if (Info.doesHold(Pred: CmpInst::ICMP_SGE, A: X, B: Zero) || |
| 2488 | isKnownNonNegative(V: X, SQ: F.getDataLayout())) { |
| 2489 | // srem x, n: result >= 0, if x >= 0 (result has the sign of x) |
| 2490 | AddFact(CmpInst::ICMP_SGE, BO, Zero); |
| 2491 | // srem x, n: result <= x, if x >= 0 (|result| <= |x| and both are |
| 2492 | // non-negative) |
| 2493 | AddFact(CmpInst::ICMP_SLE, BO, X); |
| 2494 | } |
| 2495 | if (Info.doesHold(Pred: CmpInst::ICMP_SGE, A: N, B: Zero) || |
| 2496 | isKnownPositive(V: N, SQ: F.getDataLayout())) { |
| 2497 | // srem x, n: result <= n, if n >= 0 (|result| < n, so result <= n - |
| 2498 | // 1 |
| 2499 | AddFact(CmpInst::ICMP_SLT, BO, N); |
| 2500 | } |
| 2501 | continue; |
| 2502 | } |
| 2503 | } |
| 2504 | |
| 2505 | auto &DL = F.getDataLayout(); |
| 2506 | auto AddFactsAboutIndices = [&](Value *Ptr, Type *AccessType) { |
| 2507 | CmpPredicate Pred; |
| 2508 | Value *A, *B; |
| 2509 | if (getConstraintFromMemoryAccess( |
| 2510 | GEP&: *cast<GetElementPtrInst>(Val: Ptr), |
| 2511 | AccessSize: DL.getTypeStoreSize(Ty: AccessType).getFixedValue(), Pred, A, B, DL, |
| 2512 | TLI)) |
| 2513 | AddFact(Pred, A, B); |
| 2514 | }; |
| 2515 | |
| 2516 | if (auto *LI = dyn_cast<LoadInst>(Val: CB.Inst)) { |
| 2517 | AddFactsAboutIndices(LI->getPointerOperand(), LI->getAccessType()); |
| 2518 | continue; |
| 2519 | } |
| 2520 | if (auto *SI = dyn_cast<StoreInst>(Val: CB.Inst)) { |
| 2521 | AddFactsAboutIndices(SI->getPointerOperand(), SI->getAccessType()); |
| 2522 | continue; |
| 2523 | } |
| 2524 | } |
| 2525 | |
| 2526 | if (CB.isConditionFact()) { |
| 2527 | Pred = CB.Cond.Pred; |
| 2528 | A = CB.Cond.Op0; |
| 2529 | B = CB.Cond.Op1; |
| 2530 | if (CB.DoesHold.Pred != CmpInst::BAD_ICMP_PREDICATE && |
| 2531 | !Info.doesHold(Pred: CB.DoesHold.Pred, A: CB.DoesHold.Op0, B: CB.DoesHold.Op1)) { |
| 2532 | LLVM_DEBUG({ |
| 2533 | dbgs() << "Not adding fact " ; |
| 2534 | dumpUnpackedICmp(dbgs(), Pred, A, B); |
| 2535 | dbgs() << " because precondition " ; |
| 2536 | dumpUnpackedICmp(dbgs(), CB.DoesHold.Pred, CB.DoesHold.Op0, |
| 2537 | CB.DoesHold.Op1); |
| 2538 | dbgs() << " does not hold.\n" ; |
| 2539 | }); |
| 2540 | continue; |
| 2541 | } |
| 2542 | } else { |
| 2543 | [[maybe_unused]] bool Matched = |
| 2544 | match(V: CB.Inst, P: m_Intrinsic<Intrinsic::assume>( |
| 2545 | Ops: m_ICmpLike(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))); |
| 2546 | assert(Matched && |
| 2547 | "Must have an assume intrinsic with a icmp like operand" ); |
| 2548 | } |
| 2549 | AddFact(Pred, A, B); |
| 2550 | } |
| 2551 | |
| 2552 | if (ReproducerModule && !ReproducerModule->functions().empty()) { |
| 2553 | std::string S; |
| 2554 | raw_string_ostream StringS(S); |
| 2555 | ReproducerModule->print(OS&: StringS, AAW: nullptr); |
| 2556 | OptimizationRemark Rem(DEBUG_TYPE, "Reproducer" , &F); |
| 2557 | Rem << ore::NV("module" ) << S; |
| 2558 | ORE.emit(OptDiag&: Rem); |
| 2559 | } |
| 2560 | |
| 2561 | #ifndef NDEBUG |
| 2562 | unsigned SignedEntries = |
| 2563 | count_if(DFSInStack, [](const StackEntry &E) { return E.IsSigned; }); |
| 2564 | assert(Info.getCS(false).size() - FunctionArgs.size() == |
| 2565 | DFSInStack.size() - SignedEntries && |
| 2566 | "updates to CS and DFSInStack are out of sync" ); |
| 2567 | assert(Info.getCS(true).size() == SignedEntries && |
| 2568 | "updates to CS and DFSInStack are out of sync" ); |
| 2569 | #endif |
| 2570 | |
| 2571 | for (Instruction *I : ToRemove) |
| 2572 | I->eraseFromParent(); |
| 2573 | return Changed; |
| 2574 | } |
| 2575 | |
| 2576 | PreservedAnalyses ConstraintEliminationPass::run(Function &F, |
| 2577 | FunctionAnalysisManager &AM) { |
| 2578 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 2579 | auto &LI = AM.getResult<LoopAnalysis>(IR&: F); |
| 2580 | auto &SE = AM.getResult<ScalarEvolutionAnalysis>(IR&: F); |
| 2581 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
| 2582 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 2583 | if (!eliminateConstraints(F, DT, LI, SE, ORE, TLI)) |
| 2584 | return PreservedAnalyses::all(); |
| 2585 | |
| 2586 | PreservedAnalyses PA; |
| 2587 | PA.preserve<ScalarEvolutionAnalysis>(); |
| 2588 | PA.preserveSet<CFGAnalyses>(); |
| 2589 | return PA; |
| 2590 | } |
| 2591 | |