| 1 | //===- GVN.cpp - Eliminate redundant values and loads ---------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This pass performs global value numbering to eliminate fully redundant |
| 10 | // instructions. It also performs simple dead load elimination. |
| 11 | // |
| 12 | // Note that this pass does the value numbering itself; it does not use the |
| 13 | // ValueNumbering analysis passes. |
| 14 | // |
| 15 | //===----------------------------------------------------------------------===// |
| 16 | |
| 17 | #include "llvm/Transforms/Scalar/GVN.h" |
| 18 | #include "llvm/ADT/DenseMap.h" |
| 19 | #include "llvm/ADT/DepthFirstIterator.h" |
| 20 | #include "llvm/ADT/Hashing.h" |
| 21 | #include "llvm/ADT/MapVector.h" |
| 22 | #include "llvm/ADT/PostOrderIterator.h" |
| 23 | #include "llvm/ADT/STLExtras.h" |
| 24 | #include "llvm/ADT/SetVector.h" |
| 25 | #include "llvm/ADT/SmallPtrSet.h" |
| 26 | #include "llvm/ADT/SmallVector.h" |
| 27 | #include "llvm/ADT/Statistic.h" |
| 28 | #include "llvm/Analysis/AliasAnalysis.h" |
| 29 | #include "llvm/Analysis/AssumeBundleQueries.h" |
| 30 | #include "llvm/Analysis/AssumptionCache.h" |
| 31 | #include "llvm/Analysis/CFG.h" |
| 32 | #include "llvm/Analysis/DomTreeUpdater.h" |
| 33 | #include "llvm/Analysis/GlobalsModRef.h" |
| 34 | #include "llvm/Analysis/InstructionPrecedenceTracking.h" |
| 35 | #include "llvm/Analysis/InstructionSimplify.h" |
| 36 | #include "llvm/Analysis/Loads.h" |
| 37 | #include "llvm/Analysis/LoopInfo.h" |
| 38 | #include "llvm/Analysis/MemoryBuiltins.h" |
| 39 | #include "llvm/Analysis/MemoryDependenceAnalysis.h" |
| 40 | #include "llvm/Analysis/MemorySSA.h" |
| 41 | #include "llvm/Analysis/MemorySSAUpdater.h" |
| 42 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 43 | #include "llvm/Analysis/PHITransAddr.h" |
| 44 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 45 | #include "llvm/Analysis/ValueTracking.h" |
| 46 | #include "llvm/IR/Attributes.h" |
| 47 | #include "llvm/IR/BasicBlock.h" |
| 48 | #include "llvm/IR/Constant.h" |
| 49 | #include "llvm/IR/Constants.h" |
| 50 | #include "llvm/IR/DebugLoc.h" |
| 51 | #include "llvm/IR/Dominators.h" |
| 52 | #include "llvm/IR/Function.h" |
| 53 | #include "llvm/IR/InstrTypes.h" |
| 54 | #include "llvm/IR/Instruction.h" |
| 55 | #include "llvm/IR/Instructions.h" |
| 56 | #include "llvm/IR/IntrinsicInst.h" |
| 57 | #include "llvm/IR/LLVMContext.h" |
| 58 | #include "llvm/IR/Metadata.h" |
| 59 | #include "llvm/IR/Module.h" |
| 60 | #include "llvm/IR/PassManager.h" |
| 61 | #include "llvm/IR/PatternMatch.h" |
| 62 | #include "llvm/IR/Type.h" |
| 63 | #include "llvm/IR/Use.h" |
| 64 | #include "llvm/IR/Value.h" |
| 65 | #include "llvm/InitializePasses.h" |
| 66 | #include "llvm/Pass.h" |
| 67 | #include "llvm/Support/Casting.h" |
| 68 | #include "llvm/Support/CommandLine.h" |
| 69 | #include "llvm/Support/Compiler.h" |
| 70 | #include "llvm/Support/Debug.h" |
| 71 | #include "llvm/Support/raw_ostream.h" |
| 72 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
| 73 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 74 | #include "llvm/Transforms/Utils/Local.h" |
| 75 | #include "llvm/Transforms/Utils/SSAUpdater.h" |
| 76 | #include "llvm/Transforms/Utils/VNCoercion.h" |
| 77 | #include <algorithm> |
| 78 | #include <cassert> |
| 79 | #include <cstdint> |
| 80 | #include <optional> |
| 81 | #include <utility> |
| 82 | |
| 83 | using namespace llvm; |
| 84 | using namespace llvm::VNCoercion; |
| 85 | using namespace PatternMatch; |
| 86 | |
| 87 | using AvailableValue = GVNPass::AvailableValue; |
| 88 | using AvailableValueInBlock = GVNPass::AvailableValueInBlock; |
| 89 | |
| 90 | #define DEBUG_TYPE "gvn" |
| 91 | |
| 92 | STATISTIC(NumGVNInstr, "Number of instructions deleted" ); |
| 93 | STATISTIC(NumGVNLoad, "Number of loads deleted" ); |
| 94 | STATISTIC(NumGVNPRE, "Number of instructions PRE'd" ); |
| 95 | STATISTIC(NumGVNBlocks, "Number of blocks merged" ); |
| 96 | STATISTIC(NumGVNSimpl, "Number of instructions simplified" ); |
| 97 | STATISTIC(NumGVNEqProp, "Number of equalities propagated" ); |
| 98 | STATISTIC(NumPRELoad, "Number of loads PRE'd" ); |
| 99 | STATISTIC(NumPRELoopLoad, "Number of loop loads PRE'd" ); |
| 100 | STATISTIC(NumPRELoadMoved2CEPred, |
| 101 | "Number of loads moved to predecessor of a critical edge in PRE" ); |
| 102 | |
| 103 | STATISTIC(IsValueFullyAvailableInBlockNumSpeculationsMax, |
| 104 | "Number of blocks speculated as available in " |
| 105 | "IsValueFullyAvailableInBlock(), max" ); |
| 106 | STATISTIC(MaxBBSpeculationCutoffReachedTimes, |
| 107 | "Number of times we we reached gvn-max-block-speculations cut-off " |
| 108 | "preventing further exploration" ); |
| 109 | |
| 110 | static cl::opt<bool> GVNEnableScalarPRE("enable-scalar-pre" , cl::init(Val: true), |
| 111 | cl::Hidden); |
| 112 | static cl::opt<bool> GVNEnableLoadPRE("enable-load-pre" , cl::init(Val: true)); |
| 113 | static cl::opt<bool> GVNEnableLoadInLoopPRE("enable-load-in-loop-pre" , |
| 114 | cl::init(Val: true)); |
| 115 | static cl::opt<bool> |
| 116 | GVNEnableSplitBackedgeInLoadPRE("enable-split-backedge-in-load-pre" , |
| 117 | cl::init(Val: false)); |
| 118 | static cl::opt<bool> GVNEnableMemDep("enable-gvn-memdep" , cl::init(Val: true)); |
| 119 | static cl::opt<bool> GVNEnableMemorySSA("enable-gvn-memoryssa" , |
| 120 | cl::init(Val: false)); |
| 121 | |
| 122 | static cl::opt<unsigned> ScanUsersLimit( |
| 123 | "gvn-scan-users-limit" , cl::Hidden, cl::init(Val: 100), |
| 124 | cl::desc("The number of memory accesses to scan in a block in reaching " |
| 125 | "memory values analysis (default = 100)" )); |
| 126 | |
| 127 | static cl::opt<uint32_t> MaxNumDeps( |
| 128 | "gvn-max-num-deps" , cl::Hidden, cl::init(Val: 100), |
| 129 | cl::desc("Max number of dependences to attempt Load PRE (default = 100)" )); |
| 130 | |
| 131 | // This is based on IsValueFullyAvailableInBlockNumSpeculationsMax stat. |
| 132 | static cl::opt<uint32_t> MaxBBSpeculations( |
| 133 | "gvn-max-block-speculations" , cl::Hidden, cl::init(Val: 600), |
| 134 | cl::desc("Max number of blocks we're willing to speculate on (and recurse " |
| 135 | "into) when deducing if a value is fully available or not in GVN " |
| 136 | "(default = 600)" )); |
| 137 | |
| 138 | static cl::opt<uint32_t> MaxNumVisitedInsts( |
| 139 | "gvn-max-num-visited-insts" , cl::Hidden, cl::init(Val: 100), |
| 140 | cl::desc("Max number of visited instructions when trying to find " |
| 141 | "dominating value of select dependency (default = 100)" )); |
| 142 | |
| 143 | static cl::opt<uint32_t> MaxNumInsnsPerBlock( |
| 144 | "gvn-max-num-insns" , cl::Hidden, cl::init(Val: 100), |
| 145 | cl::desc("Max number of instructions to scan in each basic block in GVN " |
| 146 | "(default = 100)" )); |
| 147 | |
| 148 | struct llvm::GVNPass::Expression { |
| 149 | uint32_t Opcode; |
| 150 | bool Commutative = false; |
| 151 | // The type is not necessarily the result type of the expression, it may be |
| 152 | // any additional type needed to disambiguate the expression. |
| 153 | Type *Ty = nullptr; |
| 154 | SmallVector<uint32_t, 4> VarArgs; |
| 155 | |
| 156 | AttributeList Attrs; |
| 157 | |
| 158 | Expression(uint32_t Op = ~2U) : Opcode(Op) {} |
| 159 | |
| 160 | bool operator==(const Expression &Other) const { |
| 161 | if (Opcode != Other.Opcode) |
| 162 | return false; |
| 163 | if (Opcode == ~0U || Opcode == ~1U) |
| 164 | return true; |
| 165 | if (Ty != Other.Ty) |
| 166 | return false; |
| 167 | if (VarArgs != Other.VarArgs) |
| 168 | return false; |
| 169 | if ((!Attrs.isEmpty() || !Other.Attrs.isEmpty()) && |
| 170 | !Attrs.intersectWith(C&: Ty->getContext(), Other: Other.Attrs).has_value()) |
| 171 | return false; |
| 172 | return true; |
| 173 | } |
| 174 | |
| 175 | friend hash_code hash_value(const Expression &Value) { |
| 176 | return hash_combine(args: Value.Opcode, args: Value.Ty, |
| 177 | args: hash_combine_range(R: Value.VarArgs)); |
| 178 | } |
| 179 | }; |
| 180 | |
| 181 | template <> struct llvm::DenseMapInfo<GVNPass::Expression> { |
| 182 | static unsigned getHashValue(const GVNPass::Expression &E) { |
| 183 | using llvm::hash_value; |
| 184 | |
| 185 | return static_cast<unsigned>(hash_value(Value: E)); |
| 186 | } |
| 187 | |
| 188 | static bool isEqual(const GVNPass::Expression &LHS, |
| 189 | const GVNPass::Expression &RHS) { |
| 190 | return LHS == RHS; |
| 191 | } |
| 192 | }; |
| 193 | |
| 194 | /// Represents a particular available value that we know how to materialize. |
| 195 | /// Materialization of an AvailableValue never fails. An AvailableValue is |
| 196 | /// implicitly associated with a rematerialization point which is the |
| 197 | /// location of the instruction from which it was formed. |
| 198 | struct llvm::GVNPass::AvailableValue { |
| 199 | enum class ValType { |
| 200 | SimpleVal, // A simple offsetted value that is accessed. |
| 201 | LoadVal, // A value produced by a load. |
| 202 | MemIntrin, // A memory intrinsic which is loaded from. |
| 203 | UndefVal, // A UndefValue representing a value from dead block (which |
| 204 | // is not yet physically removed from the CFG). |
| 205 | SelectVal, // A pointer select which is loaded from and for which the load |
| 206 | // can be replace by a value select. |
| 207 | }; |
| 208 | |
| 209 | /// Val - The value that is live out of the block. |
| 210 | Value *Val; |
| 211 | /// Kind of the live-out value. |
| 212 | ValType Kind; |
| 213 | |
| 214 | /// Offset - The byte offset in Val that is interesting for the load query. |
| 215 | unsigned Offset = 0; |
| 216 | /// V1, V2 - The dominating non-clobbered values of SelectVal. |
| 217 | Value *V1 = nullptr, *V2 = nullptr; |
| 218 | |
| 219 | static AvailableValue get(Value *V, unsigned Offset = 0) { |
| 220 | AvailableValue Res; |
| 221 | Res.Val = V; |
| 222 | Res.Kind = ValType::SimpleVal; |
| 223 | Res.Offset = Offset; |
| 224 | return Res; |
| 225 | } |
| 226 | |
| 227 | static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) { |
| 228 | AvailableValue Res; |
| 229 | Res.Val = MI; |
| 230 | Res.Kind = ValType::MemIntrin; |
| 231 | Res.Offset = Offset; |
| 232 | return Res; |
| 233 | } |
| 234 | |
| 235 | static AvailableValue getLoad(LoadInst *Load, unsigned Offset = 0) { |
| 236 | AvailableValue Res; |
| 237 | Res.Val = Load; |
| 238 | Res.Kind = ValType::LoadVal; |
| 239 | Res.Offset = Offset; |
| 240 | return Res; |
| 241 | } |
| 242 | |
| 243 | static AvailableValue getUndef() { |
| 244 | AvailableValue Res; |
| 245 | Res.Val = nullptr; |
| 246 | Res.Kind = ValType::UndefVal; |
| 247 | Res.Offset = 0; |
| 248 | return Res; |
| 249 | } |
| 250 | |
| 251 | static AvailableValue getSelect(Value *Cond, Value *V1, Value *V2) { |
| 252 | AvailableValue Res; |
| 253 | Res.Val = Cond; |
| 254 | Res.Kind = ValType::SelectVal; |
| 255 | Res.Offset = 0; |
| 256 | Res.V1 = V1; |
| 257 | Res.V2 = V2; |
| 258 | return Res; |
| 259 | } |
| 260 | |
| 261 | bool isSimpleValue() const { return Kind == ValType::SimpleVal; } |
| 262 | bool isCoercedLoadValue() const { return Kind == ValType::LoadVal; } |
| 263 | bool isMemIntrinValue() const { return Kind == ValType::MemIntrin; } |
| 264 | bool isUndefValue() const { return Kind == ValType::UndefVal; } |
| 265 | bool isSelectValue() const { return Kind == ValType::SelectVal; } |
| 266 | |
| 267 | Value *getSimpleValue() const { |
| 268 | assert(isSimpleValue() && "Wrong accessor" ); |
| 269 | return Val; |
| 270 | } |
| 271 | |
| 272 | LoadInst *getCoercedLoadValue() const { |
| 273 | assert(isCoercedLoadValue() && "Wrong accessor" ); |
| 274 | return cast<LoadInst>(Val); |
| 275 | } |
| 276 | |
| 277 | MemIntrinsic *getMemIntrinValue() const { |
| 278 | assert(isMemIntrinValue() && "Wrong accessor" ); |
| 279 | return cast<MemIntrinsic>(Val); |
| 280 | } |
| 281 | |
| 282 | Value *getSelectCondition() const { |
| 283 | assert(isSelectValue() && "Wrong accessor" ); |
| 284 | return Val; |
| 285 | } |
| 286 | |
| 287 | /// Emit code at the specified insertion point to adjust the value defined |
| 288 | /// here to the specified type. This handles various coercion cases. |
| 289 | Value *MaterializeAdjustedValue(LoadInst *Load, Instruction *InsertPt) const; |
| 290 | }; |
| 291 | |
| 292 | /// Represents an AvailableValue which can be rematerialized at the end of |
| 293 | /// the associated BasicBlock. |
| 294 | struct llvm::GVNPass::AvailableValueInBlock { |
| 295 | /// BB - The basic block in question. |
| 296 | BasicBlock *BB = nullptr; |
| 297 | |
| 298 | /// AV - The actual available value. |
| 299 | AvailableValue AV; |
| 300 | |
| 301 | static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) { |
| 302 | AvailableValueInBlock Res; |
| 303 | Res.BB = BB; |
| 304 | Res.AV = std::move(AV); |
| 305 | return Res; |
| 306 | } |
| 307 | |
| 308 | static AvailableValueInBlock get(BasicBlock *BB, Value *V, |
| 309 | unsigned Offset = 0) { |
| 310 | return get(BB, AV: AvailableValue::get(V, Offset)); |
| 311 | } |
| 312 | |
| 313 | static AvailableValueInBlock getUndef(BasicBlock *BB) { |
| 314 | return get(BB, AV: AvailableValue::getUndef()); |
| 315 | } |
| 316 | |
| 317 | /// Emit code at the end of this block to adjust the value defined here to |
| 318 | /// the specified type. This handles various coercion cases. |
| 319 | Value *MaterializeAdjustedValue(LoadInst *Load) const { |
| 320 | return AV.MaterializeAdjustedValue(Load, InsertPt: BB->getTerminator()); |
| 321 | } |
| 322 | }; |
| 323 | |
| 324 | //===----------------------------------------------------------------------===// |
| 325 | // ValueTable Internal Functions |
| 326 | //===----------------------------------------------------------------------===// |
| 327 | |
| 328 | GVNPass::Expression GVNPass::ValueTable::createExpr(Instruction *I) { |
| 329 | Expression E; |
| 330 | E.Ty = I->getType(); |
| 331 | E.Opcode = I->getOpcode(); |
| 332 | if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(Val: I)) { |
| 333 | // gc.relocate is 'special' call: its second and third operands are |
| 334 | // not real values, but indices into statepoint's argument list. |
| 335 | // Use the refered to values for purposes of identity. |
| 336 | E.VarArgs.push_back(Elt: lookupOrAdd(V: GCR->getOperand(i_nocapture: 0))); |
| 337 | E.VarArgs.push_back(Elt: lookupOrAdd(V: GCR->getBasePtr())); |
| 338 | E.VarArgs.push_back(Elt: lookupOrAdd(V: GCR->getDerivedPtr())); |
| 339 | } else { |
| 340 | for (Use &Op : I->operands()) |
| 341 | E.VarArgs.push_back(Elt: lookupOrAdd(V: Op)); |
| 342 | } |
| 343 | if (I->isCommutative()) { |
| 344 | // Ensure that commutative instructions that only differ by a permutation |
| 345 | // of their operands get the same value number by sorting the operand value |
| 346 | // numbers. Since commutative operands are the 1st two operands it is more |
| 347 | // efficient to sort by hand rather than using, say, std::sort. |
| 348 | assert(I->getNumOperands() >= 2 && "Unsupported commutative instruction!" ); |
| 349 | if (E.VarArgs[0] > E.VarArgs[1]) |
| 350 | std::swap(a&: E.VarArgs[0], b&: E.VarArgs[1]); |
| 351 | E.Commutative = true; |
| 352 | } |
| 353 | |
| 354 | if (auto *C = dyn_cast<CmpInst>(Val: I)) { |
| 355 | // Sort the operand value numbers so x<y and y>x get the same value number. |
| 356 | CmpInst::Predicate Predicate = C->getPredicate(); |
| 357 | if (E.VarArgs[0] > E.VarArgs[1]) { |
| 358 | std::swap(a&: E.VarArgs[0], b&: E.VarArgs[1]); |
| 359 | Predicate = CmpInst::getSwappedPredicate(pred: Predicate); |
| 360 | } |
| 361 | E.Opcode = (C->getOpcode() << 8) | Predicate; |
| 362 | E.Commutative = true; |
| 363 | } else if (auto *IVI = dyn_cast<InsertValueInst>(Val: I)) { |
| 364 | E.VarArgs.append(in_start: IVI->idx_begin(), in_end: IVI->idx_end()); |
| 365 | } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(Val: I)) { |
| 366 | ArrayRef<int> ShuffleMask = SVI->getShuffleMask(); |
| 367 | E.VarArgs.append(in_start: ShuffleMask.begin(), in_end: ShuffleMask.end()); |
| 368 | } else if (auto *CB = dyn_cast<CallBase>(Val: I)) { |
| 369 | E.Attrs = CB->getAttributes(); |
| 370 | } |
| 371 | |
| 372 | return E; |
| 373 | } |
| 374 | |
| 375 | GVNPass::Expression GVNPass::ValueTable::createCmpExpr( |
| 376 | unsigned Opcode, CmpInst::Predicate Predicate, Value *LHS, Value *RHS) { |
| 377 | assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && |
| 378 | "Not a comparison!" ); |
| 379 | Expression E; |
| 380 | E.Ty = CmpInst::makeCmpResultType(opnd_type: LHS->getType()); |
| 381 | E.VarArgs.push_back(Elt: lookupOrAdd(V: LHS)); |
| 382 | E.VarArgs.push_back(Elt: lookupOrAdd(V: RHS)); |
| 383 | |
| 384 | // Sort the operand value numbers so x<y and y>x get the same value number. |
| 385 | if (E.VarArgs[0] > E.VarArgs[1]) { |
| 386 | std::swap(a&: E.VarArgs[0], b&: E.VarArgs[1]); |
| 387 | Predicate = CmpInst::getSwappedPredicate(pred: Predicate); |
| 388 | } |
| 389 | E.Opcode = (Opcode << 8) | Predicate; |
| 390 | E.Commutative = true; |
| 391 | return E; |
| 392 | } |
| 393 | |
| 394 | GVNPass::Expression |
| 395 | GVNPass::ValueTable::(ExtractValueInst *EI) { |
| 396 | assert(EI && "Not an ExtractValueInst?" ); |
| 397 | Expression E; |
| 398 | E.Ty = EI->getType(); |
| 399 | E.Opcode = 0; |
| 400 | |
| 401 | WithOverflowInst *WO = dyn_cast<WithOverflowInst>(Val: EI->getAggregateOperand()); |
| 402 | if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) { |
| 403 | // EI is an extract from one of our with.overflow intrinsics. Synthesize |
| 404 | // a semantically equivalent expression instead of an extract value |
| 405 | // expression. |
| 406 | E.Opcode = WO->getBinaryOp(); |
| 407 | E.VarArgs.push_back(Elt: lookupOrAdd(V: WO->getLHS())); |
| 408 | E.VarArgs.push_back(Elt: lookupOrAdd(V: WO->getRHS())); |
| 409 | return E; |
| 410 | } |
| 411 | |
| 412 | // Not a recognised intrinsic. Fall back to producing an extract value |
| 413 | // expression. |
| 414 | E.Opcode = EI->getOpcode(); |
| 415 | for (Use &Op : EI->operands()) |
| 416 | E.VarArgs.push_back(Elt: lookupOrAdd(V: Op)); |
| 417 | |
| 418 | append_range(C&: E.VarArgs, R: EI->indices()); |
| 419 | |
| 420 | return E; |
| 421 | } |
| 422 | |
| 423 | GVNPass::Expression GVNPass::ValueTable::createGEPExpr(GetElementPtrInst *GEP) { |
| 424 | Expression E; |
| 425 | Type *PtrTy = GEP->getType()->getScalarType(); |
| 426 | const DataLayout &DL = GEP->getDataLayout(); |
| 427 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: PtrTy); |
| 428 | SmallMapVector<Value *, APInt, 4> VariableOffsets; |
| 429 | APInt ConstantOffset(BitWidth, 0); |
| 430 | if (GEP->collectOffset(DL, BitWidth, VariableOffsets, ConstantOffset)) { |
| 431 | // Convert into offset representation, to recognize equivalent address |
| 432 | // calculations that use different type encoding. |
| 433 | LLVMContext &Context = GEP->getContext(); |
| 434 | E.Opcode = GEP->getOpcode(); |
| 435 | E.Ty = nullptr; |
| 436 | E.VarArgs.push_back(Elt: lookupOrAdd(V: GEP->getPointerOperand())); |
| 437 | for (const auto &[V, Scale] : VariableOffsets) { |
| 438 | E.VarArgs.push_back(Elt: lookupOrAdd(V)); |
| 439 | E.VarArgs.push_back(Elt: lookupOrAdd(V: ConstantInt::get(Context, V: Scale))); |
| 440 | } |
| 441 | if (!ConstantOffset.isZero()) |
| 442 | E.VarArgs.push_back( |
| 443 | Elt: lookupOrAdd(V: ConstantInt::get(Context, V: ConstantOffset))); |
| 444 | } else { |
| 445 | // If converting to offset representation fails (for scalable vectors), |
| 446 | // fall back to type-based implementation. |
| 447 | E.Opcode = GEP->getOpcode(); |
| 448 | E.Ty = GEP->getSourceElementType(); |
| 449 | for (Use &Op : GEP->operands()) |
| 450 | E.VarArgs.push_back(Elt: lookupOrAdd(V: Op)); |
| 451 | } |
| 452 | return E; |
| 453 | } |
| 454 | |
| 455 | //===----------------------------------------------------------------------===// |
| 456 | // ValueTable External Functions |
| 457 | //===----------------------------------------------------------------------===// |
| 458 | |
| 459 | GVNPass::ValueTable::ValueTable() = default; |
| 460 | GVNPass::ValueTable::ValueTable(const ValueTable &) = default; |
| 461 | GVNPass::ValueTable::ValueTable(ValueTable &&) = default; |
| 462 | GVNPass::ValueTable::~ValueTable() = default; |
| 463 | GVNPass::ValueTable & |
| 464 | GVNPass::ValueTable::operator=(const GVNPass::ValueTable &Arg) = default; |
| 465 | |
| 466 | /// add - Insert a value into the table with a specified value number. |
| 467 | void GVNPass::ValueTable::add(Value *V, uint32_t Num) { |
| 468 | ValueNumbering.insert(KV: std::make_pair(x&: V, y&: Num)); |
| 469 | if (PHINode *PN = dyn_cast<PHINode>(Val: V)) |
| 470 | NumberingPhi[Num] = PN; |
| 471 | } |
| 472 | |
| 473 | /// Include the incoming memory state into the hash of the expression for the |
| 474 | /// given instruction. If the incoming memory state is: |
| 475 | /// * LiveOnEntry, add the value number of the entry block, |
| 476 | /// * a MemoryPhi, add the value number of the basic block corresponding to that |
| 477 | /// MemoryPhi, |
| 478 | /// * a MemoryDef, add the value number of the memory setting instruction. |
| 479 | void GVNPass::ValueTable::addMemoryStateToExp(Instruction *I, Expression &Exp) { |
| 480 | assert(MSSA && "addMemoryStateToExp should not be called without MemorySSA" ); |
| 481 | assert(MSSA->getMemoryAccess(I) && "Instruction does not access memory" ); |
| 482 | MemoryAccess *MA = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(I); |
| 483 | Exp.VarArgs.push_back(Elt: lookupOrAdd(MA)); |
| 484 | } |
| 485 | |
| 486 | uint32_t GVNPass::ValueTable::lookupOrAddCall(CallInst *C) { |
| 487 | // FIXME: Currently the calls which may access the thread id may |
| 488 | // be considered as not accessing the memory. But this is |
| 489 | // problematic for coroutines, since coroutines may resume in a |
| 490 | // different thread. So we disable the optimization here for the |
| 491 | // correctness. However, it may block many other correct |
| 492 | // optimizations. Revert this one when we detect the memory |
| 493 | // accessing kind more precisely. |
| 494 | if (C->getFunction()->isPresplitCoroutine()) { |
| 495 | ValueNumbering[C] = NextValueNumber; |
| 496 | return NextValueNumber++; |
| 497 | } |
| 498 | |
| 499 | // Do not combine convergent calls since they implicitly depend on the set of |
| 500 | // threads that is currently executing, and they might be in different basic |
| 501 | // blocks. |
| 502 | if (C->isConvergent()) { |
| 503 | ValueNumbering[C] = NextValueNumber; |
| 504 | return NextValueNumber++; |
| 505 | } |
| 506 | |
| 507 | if (AA->doesNotAccessMemory(Call: C)) { |
| 508 | Expression Exp = createExpr(I: C); |
| 509 | uint32_t E = assignExpNewValueNum(Exp).first; |
| 510 | ValueNumbering[C] = E; |
| 511 | return E; |
| 512 | } |
| 513 | |
| 514 | if (MD && AA->onlyReadsMemory(Call: C)) { |
| 515 | Expression Exp = createExpr(I: C); |
| 516 | auto [E, IsValNumNew] = assignExpNewValueNum(Exp); |
| 517 | if (IsValNumNew) { |
| 518 | ValueNumbering[C] = E; |
| 519 | return E; |
| 520 | } |
| 521 | |
| 522 | MemDepResult LocalDep = MD->getDependency(QueryInst: C); |
| 523 | |
| 524 | if (!LocalDep.isDef() && !LocalDep.isNonLocal()) { |
| 525 | ValueNumbering[C] = NextValueNumber; |
| 526 | return NextValueNumber++; |
| 527 | } |
| 528 | |
| 529 | if (LocalDep.isDef()) { |
| 530 | // For masked load/store intrinsics, the local_dep may actually be |
| 531 | // a normal load or store instruction. |
| 532 | CallInst *LocalDepCall = dyn_cast<CallInst>(Val: LocalDep.getInst()); |
| 533 | |
| 534 | if (!LocalDepCall || LocalDepCall->arg_size() != C->arg_size()) { |
| 535 | ValueNumbering[C] = NextValueNumber; |
| 536 | return NextValueNumber++; |
| 537 | } |
| 538 | |
| 539 | for (unsigned I = 0, E = C->arg_size(); I < E; ++I) { |
| 540 | uint32_t CVN = lookupOrAdd(V: C->getArgOperand(i: I)); |
| 541 | uint32_t LocalDepCallVN = lookupOrAdd(V: LocalDepCall->getArgOperand(i: I)); |
| 542 | if (CVN != LocalDepCallVN) { |
| 543 | ValueNumbering[C] = NextValueNumber; |
| 544 | return NextValueNumber++; |
| 545 | } |
| 546 | } |
| 547 | |
| 548 | uint32_t V = lookupOrAdd(V: LocalDepCall); |
| 549 | ValueNumbering[C] = V; |
| 550 | return V; |
| 551 | } |
| 552 | |
| 553 | // Non-local case. |
| 554 | const MemoryDependenceResults::NonLocalDepInfo &Deps = |
| 555 | MD->getNonLocalCallDependency(QueryCall: C); |
| 556 | // FIXME: Move the checking logic to MemDep! |
| 557 | CallInst *CDep = nullptr; |
| 558 | |
| 559 | // Check to see if we have a single dominating call instruction that is |
| 560 | // identical to C. |
| 561 | for (const NonLocalDepEntry &I : Deps) { |
| 562 | if (I.getResult().isNonLocal()) |
| 563 | continue; |
| 564 | |
| 565 | // We don't handle non-definitions. If we already have a call, reject |
| 566 | // instruction dependencies. |
| 567 | if (!I.getResult().isDef() || CDep != nullptr) { |
| 568 | CDep = nullptr; |
| 569 | break; |
| 570 | } |
| 571 | |
| 572 | CallInst *NonLocalDepCall = dyn_cast<CallInst>(Val: I.getResult().getInst()); |
| 573 | // FIXME: All duplicated with non-local case. |
| 574 | if (NonLocalDepCall && DT->properlyDominates(A: I.getBB(), B: C->getParent())) { |
| 575 | CDep = NonLocalDepCall; |
| 576 | continue; |
| 577 | } |
| 578 | |
| 579 | CDep = nullptr; |
| 580 | break; |
| 581 | } |
| 582 | |
| 583 | if (!CDep) { |
| 584 | ValueNumbering[C] = NextValueNumber; |
| 585 | return NextValueNumber++; |
| 586 | } |
| 587 | |
| 588 | if (CDep->arg_size() != C->arg_size()) { |
| 589 | ValueNumbering[C] = NextValueNumber; |
| 590 | return NextValueNumber++; |
| 591 | } |
| 592 | for (unsigned I = 0, E = C->arg_size(); I < E; ++I) { |
| 593 | uint32_t CVN = lookupOrAdd(V: C->getArgOperand(i: I)); |
| 594 | uint32_t CDepVN = lookupOrAdd(V: CDep->getArgOperand(i: I)); |
| 595 | if (CVN != CDepVN) { |
| 596 | ValueNumbering[C] = NextValueNumber; |
| 597 | return NextValueNumber++; |
| 598 | } |
| 599 | } |
| 600 | |
| 601 | uint32_t V = lookupOrAdd(V: CDep); |
| 602 | ValueNumbering[C] = V; |
| 603 | return V; |
| 604 | } |
| 605 | |
| 606 | if (MSSA && IsMSSAEnabled && AA->onlyReadsMemory(Call: C)) { |
| 607 | Expression Exp = createExpr(I: C); |
| 608 | addMemoryStateToExp(I: C, Exp); |
| 609 | auto [V, _] = assignExpNewValueNum(Exp); |
| 610 | ValueNumbering[C] = V; |
| 611 | return V; |
| 612 | } |
| 613 | |
| 614 | ValueNumbering[C] = NextValueNumber; |
| 615 | return NextValueNumber++; |
| 616 | } |
| 617 | |
| 618 | /// Returns the value number for the specified load or store instruction. |
| 619 | uint32_t GVNPass::ValueTable::computeLoadStoreVN(Instruction *I) { |
| 620 | if (!MSSA || !IsMSSAEnabled) { |
| 621 | ValueNumbering[I] = NextValueNumber; |
| 622 | return NextValueNumber++; |
| 623 | } |
| 624 | |
| 625 | Expression Exp; |
| 626 | Exp.Ty = I->getType(); |
| 627 | Exp.Opcode = I->getOpcode(); |
| 628 | for (Use &Op : I->operands()) |
| 629 | Exp.VarArgs.push_back(Elt: lookupOrAdd(V: Op)); |
| 630 | addMemoryStateToExp(I, Exp); |
| 631 | |
| 632 | auto [V, _] = assignExpNewValueNum(Exp); |
| 633 | ValueNumbering[I] = V; |
| 634 | return V; |
| 635 | } |
| 636 | |
| 637 | /// Returns true if a value number exists for the specified value. |
| 638 | bool GVNPass::ValueTable::exists(Value *V) const { |
| 639 | return ValueNumbering.contains(Val: V); |
| 640 | } |
| 641 | |
| 642 | uint32_t GVNPass::ValueTable::lookupOrAdd(MemoryAccess *MA) { |
| 643 | return MSSA->isLiveOnEntryDef(MA) || isa<MemoryPhi>(Val: MA) |
| 644 | ? lookupOrAdd(V: MA->getBlock()) |
| 645 | : lookupOrAdd(V: cast<MemoryUseOrDef>(Val: MA)->getMemoryInst()); |
| 646 | } |
| 647 | |
| 648 | /// lookupOrAdd - Returns the value number for the specified value, assigning |
| 649 | /// it a new number if it did not have one before. |
| 650 | uint32_t GVNPass::ValueTable::lookupOrAdd(Value *V) { |
| 651 | auto VI = ValueNumbering.find(Val: V); |
| 652 | if (VI != ValueNumbering.end()) |
| 653 | return VI->second; |
| 654 | |
| 655 | auto *I = dyn_cast<Instruction>(Val: V); |
| 656 | if (!I) { |
| 657 | ValueNumbering[V] = NextValueNumber; |
| 658 | if (isa<BasicBlock>(Val: V)) |
| 659 | NumberingBB[NextValueNumber] = cast<BasicBlock>(Val: V); |
| 660 | return NextValueNumber++; |
| 661 | } |
| 662 | |
| 663 | Expression Exp; |
| 664 | switch (I->getOpcode()) { |
| 665 | case Instruction::Call: |
| 666 | return lookupOrAddCall(C: cast<CallInst>(Val: I)); |
| 667 | case Instruction::FNeg: |
| 668 | case Instruction::Add: |
| 669 | case Instruction::FAdd: |
| 670 | case Instruction::Sub: |
| 671 | case Instruction::FSub: |
| 672 | case Instruction::Mul: |
| 673 | case Instruction::FMul: |
| 674 | case Instruction::UDiv: |
| 675 | case Instruction::SDiv: |
| 676 | case Instruction::FDiv: |
| 677 | case Instruction::URem: |
| 678 | case Instruction::SRem: |
| 679 | case Instruction::FRem: |
| 680 | case Instruction::Shl: |
| 681 | case Instruction::LShr: |
| 682 | case Instruction::AShr: |
| 683 | case Instruction::And: |
| 684 | case Instruction::Or: |
| 685 | case Instruction::Xor: |
| 686 | case Instruction::ICmp: |
| 687 | case Instruction::FCmp: |
| 688 | case Instruction::Trunc: |
| 689 | case Instruction::ZExt: |
| 690 | case Instruction::SExt: |
| 691 | case Instruction::FPToUI: |
| 692 | case Instruction::FPToSI: |
| 693 | case Instruction::UIToFP: |
| 694 | case Instruction::SIToFP: |
| 695 | case Instruction::FPTrunc: |
| 696 | case Instruction::FPExt: |
| 697 | case Instruction::PtrToInt: |
| 698 | case Instruction::PtrToAddr: |
| 699 | case Instruction::IntToPtr: |
| 700 | case Instruction::AddrSpaceCast: |
| 701 | case Instruction::BitCast: |
| 702 | case Instruction::Select: |
| 703 | case Instruction::Freeze: |
| 704 | case Instruction::ExtractElement: |
| 705 | case Instruction::InsertElement: |
| 706 | case Instruction::ShuffleVector: |
| 707 | case Instruction::InsertValue: |
| 708 | Exp = createExpr(I); |
| 709 | break; |
| 710 | case Instruction::GetElementPtr: |
| 711 | Exp = createGEPExpr(GEP: cast<GetElementPtrInst>(Val: I)); |
| 712 | break; |
| 713 | case Instruction::ExtractValue: |
| 714 | Exp = createExtractvalueExpr(EI: cast<ExtractValueInst>(Val: I)); |
| 715 | break; |
| 716 | case Instruction::PHI: |
| 717 | ValueNumbering[V] = NextValueNumber; |
| 718 | NumberingPhi[NextValueNumber] = cast<PHINode>(Val: V); |
| 719 | return NextValueNumber++; |
| 720 | case Instruction::Load: |
| 721 | case Instruction::Store: |
| 722 | return computeLoadStoreVN(I); |
| 723 | default: |
| 724 | ValueNumbering[V] = NextValueNumber; |
| 725 | return NextValueNumber++; |
| 726 | } |
| 727 | |
| 728 | uint32_t E = assignExpNewValueNum(Exp).first; |
| 729 | ValueNumbering[V] = E; |
| 730 | return E; |
| 731 | } |
| 732 | |
| 733 | /// Returns the value number of the specified value. Fails if |
| 734 | /// the value has not yet been numbered. |
| 735 | uint32_t GVNPass::ValueTable::lookup(Value *V, bool Verify) const { |
| 736 | auto VI = ValueNumbering.find(Val: V); |
| 737 | if (Verify) { |
| 738 | assert(VI != ValueNumbering.end() && "Value not numbered?" ); |
| 739 | return VI->second; |
| 740 | } |
| 741 | return (VI != ValueNumbering.end()) ? VI->second : 0; |
| 742 | } |
| 743 | |
| 744 | /// Returns the value number of the given comparison, |
| 745 | /// assigning it a new number if it did not have one before. Useful when |
| 746 | /// we deduced the result of a comparison, but don't immediately have an |
| 747 | /// instruction realizing that comparison to hand. |
| 748 | uint32_t GVNPass::ValueTable::lookupOrAddCmp(unsigned Opcode, |
| 749 | CmpInst::Predicate Predicate, |
| 750 | Value *LHS, Value *RHS) { |
| 751 | Expression Exp = createCmpExpr(Opcode, Predicate, LHS, RHS); |
| 752 | return assignExpNewValueNum(Exp).first; |
| 753 | } |
| 754 | |
| 755 | /// Returns the value number of ptrtoint \p Ptr to \Ty. |
| 756 | uint32_t GVNPass::ValueTable::lookupPtrToInt(Value *Ptr, Type *Ty) { |
| 757 | Expression Exp(Instruction::PtrToInt); |
| 758 | Exp.Ty = Ty; |
| 759 | Exp.VarArgs.push_back(Elt: lookupOrAdd(V: Ptr)); |
| 760 | return ExpressionNumbering.lookup(Val: Exp); |
| 761 | } |
| 762 | |
| 763 | /// Remove all entries from the ValueTable. |
| 764 | void GVNPass::ValueTable::clear() { |
| 765 | ValueNumbering.clear(); |
| 766 | ExpressionNumbering.clear(); |
| 767 | NumberingPhi.clear(); |
| 768 | NumberingBB.clear(); |
| 769 | PhiTranslateTable.clear(); |
| 770 | NextValueNumber = 1; |
| 771 | Expressions.clear(); |
| 772 | ExprIdx.clear(); |
| 773 | NextExprNumber = 0; |
| 774 | } |
| 775 | |
| 776 | /// Remove a value from the value numbering. |
| 777 | void GVNPass::ValueTable::erase(Value *V) { |
| 778 | uint32_t Num = ValueNumbering.lookup(Val: V); |
| 779 | ValueNumbering.erase(Val: V); |
| 780 | // If V is PHINode, V <--> value number is an one-to-one mapping. |
| 781 | if (isa<PHINode>(Val: V)) |
| 782 | NumberingPhi.erase(Val: Num); |
| 783 | else if (isa<BasicBlock>(Val: V)) |
| 784 | NumberingBB.erase(Val: Num); |
| 785 | } |
| 786 | |
| 787 | /// verifyRemoved - Verify that the value is removed from all internal data |
| 788 | /// structures. |
| 789 | void GVNPass::ValueTable::verifyRemoved(const Value *V) const { |
| 790 | assert(!ValueNumbering.contains(V) && |
| 791 | "Inst still occurs in value numbering map!" ); |
| 792 | } |
| 793 | |
| 794 | //===----------------------------------------------------------------------===// |
| 795 | // LeaderMap External Functions |
| 796 | //===----------------------------------------------------------------------===// |
| 797 | |
| 798 | /// Push a new Value to the LeaderTable onto the list for its value number. |
| 799 | void GVNPass::LeaderMap::insert(uint32_t N, Value *V, const BasicBlock *BB) { |
| 800 | const auto &[It, Inserted] = NumToLeaders.try_emplace(Key: N, Args&: V, Args&: BB, Args: nullptr); |
| 801 | if (!Inserted) { |
| 802 | // Key already exists: insert new node after the head. |
| 803 | auto *NewSlot = TableAllocator.Allocate<LeaderListNode>(); |
| 804 | new (NewSlot) LeaderListNode(V, BB, It->second.Next); |
| 805 | It->second.Next = NewSlot; |
| 806 | } |
| 807 | } |
| 808 | |
| 809 | /// Scan the list of values corresponding to a given |
| 810 | /// value number, and remove the given instruction if encountered. |
| 811 | void GVNPass::LeaderMap::erase(uint32_t N, Instruction *I, |
| 812 | const BasicBlock *BB) { |
| 813 | auto It = NumToLeaders.find(Val: N); |
| 814 | if (It == NumToLeaders.end()) |
| 815 | return; |
| 816 | |
| 817 | LeaderListNode *Prev = nullptr; |
| 818 | LeaderListNode *Curr = &It->second; |
| 819 | |
| 820 | while (Curr && (Curr->Entry.Val != I || Curr->Entry.BB != BB)) { |
| 821 | Prev = Curr; |
| 822 | Curr = Curr->Next; |
| 823 | } |
| 824 | |
| 825 | if (!Curr) |
| 826 | return; |
| 827 | |
| 828 | if (Prev) { |
| 829 | // Non-head node: unlink and destroy. |
| 830 | Prev->Next = Curr->Next; |
| 831 | Curr->~LeaderListNode(); |
| 832 | TableAllocator.Deallocate<LeaderListNode>(Ptr: Curr); |
| 833 | } else { |
| 834 | // Head node (stored by value in DenseMap). |
| 835 | if (!Curr->Next) { |
| 836 | // Only node; erase from map (DenseMap calls the destructor). |
| 837 | NumToLeaders.erase(I: It); |
| 838 | } else { |
| 839 | // Move second node's data into head, then destroy second node. |
| 840 | LeaderListNode *Next = Curr->Next; |
| 841 | Curr->Entry.Val = std::move(Next->Entry.Val); |
| 842 | Curr->Entry.BB = Next->Entry.BB; |
| 843 | Curr->Next = Next->Next; |
| 844 | Next->~LeaderListNode(); |
| 845 | TableAllocator.Deallocate<LeaderListNode>(Ptr: Next); |
| 846 | } |
| 847 | } |
| 848 | } |
| 849 | |
| 850 | //===----------------------------------------------------------------------===// |
| 851 | // GVN Pass |
| 852 | //===----------------------------------------------------------------------===// |
| 853 | |
| 854 | bool GVNPass::isScalarPREEnabled() const { |
| 855 | return Options.AllowScalarPRE.value_or(u&: GVNEnableScalarPRE); |
| 856 | } |
| 857 | |
| 858 | bool GVNPass::isLoadPREEnabled() const { |
| 859 | return Options.AllowLoadPRE.value_or(u&: GVNEnableLoadPRE); |
| 860 | } |
| 861 | |
| 862 | bool GVNPass::isLoadInLoopPREEnabled() const { |
| 863 | return Options.AllowLoadInLoopPRE.value_or(u&: GVNEnableLoadInLoopPRE); |
| 864 | } |
| 865 | |
| 866 | bool GVNPass::isLoadPRESplitBackedgeEnabled() const { |
| 867 | return Options.AllowLoadPRESplitBackedge.value_or( |
| 868 | u&: GVNEnableSplitBackedgeInLoadPRE); |
| 869 | } |
| 870 | |
| 871 | bool GVNPass::isMemDepEnabled() const { |
| 872 | return Options.AllowMemDep.value_or(u&: GVNEnableMemDep); |
| 873 | } |
| 874 | |
| 875 | bool GVNPass::isMemorySSAEnabled() const { |
| 876 | return Options.AllowMemorySSA.value_or(u&: GVNEnableMemorySSA); |
| 877 | } |
| 878 | |
| 879 | PreservedAnalyses GVNPass::run(Function &F, FunctionAnalysisManager &AM) { |
| 880 | // FIXME: The order of evaluation of these 'getResult' calls is very |
| 881 | // significant! Re-ordering these variables will cause GVN when run alone to |
| 882 | // be less effective! We should fix memdep and basic-aa to not exhibit this |
| 883 | // behavior, but until then don't change the order here. |
| 884 | auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F); |
| 885 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 886 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 887 | auto &AA = AM.getResult<AAManager>(IR&: F); |
| 888 | auto *MemDep = |
| 889 | isMemDepEnabled() ? &AM.getResult<MemoryDependenceAnalysis>(IR&: F) : nullptr; |
| 890 | auto &LI = AM.getResult<LoopAnalysis>(IR&: F); |
| 891 | auto *MSSA = AM.getCachedResult<MemorySSAAnalysis>(IR&: F); |
| 892 | if (isMemorySSAEnabled() && !MSSA) { |
| 893 | assert(!MemDep && |
| 894 | "On-demand computation of MemSSA implies that MemDep is disabled!" ); |
| 895 | MSSA = &AM.getResult<MemorySSAAnalysis>(IR&: F); |
| 896 | } |
| 897 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
| 898 | bool Changed = runImpl(F, RunAC&: AC, RunDT&: DT, RunTLI: TLI, RunAA&: AA, RunMD: MemDep, LI, ORE: &ORE, |
| 899 | MSSA: MSSA ? &MSSA->getMSSA() : nullptr); |
| 900 | if (!Changed) |
| 901 | return PreservedAnalyses::all(); |
| 902 | PreservedAnalyses PA; |
| 903 | PA.preserve<DominatorTreeAnalysis>(); |
| 904 | PA.preserve<TargetLibraryAnalysis>(); |
| 905 | if (MSSA) |
| 906 | PA.preserve<MemorySSAAnalysis>(); |
| 907 | PA.preserve<LoopAnalysis>(); |
| 908 | return PA; |
| 909 | } |
| 910 | |
| 911 | void GVNPass::printPipeline( |
| 912 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
| 913 | static_cast<PassInfoMixin<GVNPass> *>(this)->printPipeline( |
| 914 | OS, MapClassName2PassName); |
| 915 | |
| 916 | OS << '<'; |
| 917 | if (Options.AllowScalarPRE != std::nullopt) |
| 918 | OS << (*Options.AllowScalarPRE ? "" : "no-" ) << "scalar-pre;" ; |
| 919 | if (Options.AllowLoadPRE != std::nullopt) |
| 920 | OS << (*Options.AllowLoadPRE ? "" : "no-" ) << "load-pre;" ; |
| 921 | if (Options.AllowLoadPRESplitBackedge != std::nullopt) |
| 922 | OS << (*Options.AllowLoadPRESplitBackedge ? "" : "no-" ) |
| 923 | << "split-backedge-load-pre;" ; |
| 924 | if (Options.AllowMemDep != std::nullopt) |
| 925 | OS << (*Options.AllowMemDep ? "" : "no-" ) << "memdep;" ; |
| 926 | if (Options.AllowMemorySSA != std::nullopt) |
| 927 | OS << (*Options.AllowMemorySSA ? "" : "no-" ) << "memoryssa" ; |
| 928 | OS << '>'; |
| 929 | } |
| 930 | |
| 931 | void GVNPass::salvageAndRemoveInstruction(Instruction *I) { |
| 932 | salvageKnowledge(I, AC); |
| 933 | salvageDebugInfo(I&: *I); |
| 934 | removeInstruction(I); |
| 935 | } |
| 936 | |
| 937 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
| 938 | LLVM_DUMP_METHOD void GVNPass::dump(DenseMap<uint32_t, Value *> &Map) const { |
| 939 | errs() << "{\n" ; |
| 940 | for (const auto &[Num, Exp] : Map) { |
| 941 | errs() << Num << "\n" ; |
| 942 | Exp->dump(); |
| 943 | } |
| 944 | errs() << "}\n" ; |
| 945 | } |
| 946 | #endif |
| 947 | |
| 948 | enum class AvailabilityState : char { |
| 949 | /// We know the block *is not* fully available. This is a fixpoint. |
| 950 | Unavailable = 0, |
| 951 | /// We know the block *is* fully available. This is a fixpoint. |
| 952 | Available = 1, |
| 953 | /// We do not know whether the block is fully available or not, |
| 954 | /// but we are currently speculating that it will be. |
| 955 | /// If it would have turned out that the block was, in fact, not fully |
| 956 | /// available, this would have been cleaned up into an Unavailable. |
| 957 | SpeculativelyAvailable = 2, |
| 958 | }; |
| 959 | |
| 960 | /// Return true if we can prove that the value |
| 961 | /// we're analyzing is fully available in the specified block. As we go, keep |
| 962 | /// track of which blocks we know are fully alive in FullyAvailableBlocks. This |
| 963 | /// map is actually a tri-state map with the following values: |
| 964 | /// 0) we know the block *is not* fully available. |
| 965 | /// 1) we know the block *is* fully available. |
| 966 | /// 2) we do not know whether the block is fully available or not, but we are |
| 967 | /// currently speculating that it will be. |
| 968 | static bool IsValueFullyAvailableInBlock( |
| 969 | BasicBlock *BB, |
| 970 | DenseMap<BasicBlock *, AvailabilityState> &FullyAvailableBlocks) { |
| 971 | SmallVector<BasicBlock *, 32> Worklist; |
| 972 | std::optional<BasicBlock *> UnavailableBB; |
| 973 | |
| 974 | // The number of times we didn't find an entry for a block in a map and |
| 975 | // optimistically inserted an entry marking block as speculatively available. |
| 976 | unsigned NumNewNewSpeculativelyAvailableBBs = 0; |
| 977 | |
| 978 | #ifndef NDEBUG |
| 979 | SmallPtrSet<BasicBlock *, 32> NewSpeculativelyAvailableBBs; |
| 980 | SmallVector<BasicBlock *, 32> AvailableBBs; |
| 981 | #endif |
| 982 | |
| 983 | Worklist.emplace_back(Args&: BB); |
| 984 | while (!Worklist.empty()) { |
| 985 | BasicBlock *CurrBB = Worklist.pop_back_val(); // LoadFO - depth-first! |
| 986 | // Optimistically assume that the block is Speculatively Available and check |
| 987 | // to see if we already know about this block in one lookup. |
| 988 | std::pair<DenseMap<BasicBlock *, AvailabilityState>::iterator, bool> IV = |
| 989 | FullyAvailableBlocks.try_emplace( |
| 990 | Key: CurrBB, Args: AvailabilityState::SpeculativelyAvailable); |
| 991 | AvailabilityState &State = IV.first->second; |
| 992 | |
| 993 | // Did the entry already exist for this block? |
| 994 | if (!IV.second) { |
| 995 | if (State == AvailabilityState::Unavailable) { |
| 996 | UnavailableBB = CurrBB; |
| 997 | break; // Backpropagate unavailability info. |
| 998 | } |
| 999 | |
| 1000 | #ifndef NDEBUG |
| 1001 | AvailableBBs.emplace_back(CurrBB); |
| 1002 | #endif |
| 1003 | continue; // Don't recurse further, but continue processing worklist. |
| 1004 | } |
| 1005 | |
| 1006 | // No entry found for block. |
| 1007 | ++NumNewNewSpeculativelyAvailableBBs; |
| 1008 | bool OutOfBudget = NumNewNewSpeculativelyAvailableBBs > MaxBBSpeculations; |
| 1009 | |
| 1010 | // If we have exhausted our budget, mark this block as unavailable. |
| 1011 | // Also, if this block has no predecessors, the value isn't live-in here. |
| 1012 | if (OutOfBudget || pred_empty(BB: CurrBB)) { |
| 1013 | MaxBBSpeculationCutoffReachedTimes += (int)OutOfBudget; |
| 1014 | State = AvailabilityState::Unavailable; |
| 1015 | UnavailableBB = CurrBB; |
| 1016 | break; // Backpropagate unavailability info. |
| 1017 | } |
| 1018 | |
| 1019 | // Tentatively consider this block as speculatively available. |
| 1020 | #ifndef NDEBUG |
| 1021 | NewSpeculativelyAvailableBBs.insert(CurrBB); |
| 1022 | #endif |
| 1023 | // And further recurse into block's predecessors, in depth-first order! |
| 1024 | Worklist.append(in_start: pred_begin(BB: CurrBB), in_end: pred_end(BB: CurrBB)); |
| 1025 | } |
| 1026 | |
| 1027 | #if LLVM_ENABLE_STATS |
| 1028 | IsValueFullyAvailableInBlockNumSpeculationsMax.updateMax( |
| 1029 | NumNewNewSpeculativelyAvailableBBs); |
| 1030 | #endif |
| 1031 | |
| 1032 | // If the block isn't marked as fixpoint yet |
| 1033 | // (the Unavailable and Available states are fixpoints). |
| 1034 | auto MarkAsFixpointAndEnqueueSuccessors = |
| 1035 | [&](BasicBlock *BB, AvailabilityState FixpointState) { |
| 1036 | auto It = FullyAvailableBlocks.find(Val: BB); |
| 1037 | if (It == FullyAvailableBlocks.end()) |
| 1038 | return; // Never queried this block, leave as-is. |
| 1039 | switch (AvailabilityState &State = It->second) { |
| 1040 | case AvailabilityState::Unavailable: |
| 1041 | case AvailabilityState::Available: |
| 1042 | return; // Don't backpropagate further, continue processing worklist. |
| 1043 | case AvailabilityState::SpeculativelyAvailable: // Fix it! |
| 1044 | State = FixpointState; |
| 1045 | #ifndef NDEBUG |
| 1046 | assert(NewSpeculativelyAvailableBBs.erase(BB) && |
| 1047 | "Found a speculatively available successor leftover?" ); |
| 1048 | #endif |
| 1049 | // Queue successors for further processing. |
| 1050 | Worklist.append(in_start: succ_begin(BB), in_end: succ_end(BB)); |
| 1051 | return; |
| 1052 | } |
| 1053 | }; |
| 1054 | |
| 1055 | if (UnavailableBB) { |
| 1056 | // Okay, we have encountered an unavailable block. |
| 1057 | // Mark speculatively available blocks reachable from UnavailableBB as |
| 1058 | // unavailable as well. Paths are terminated when they reach blocks not in |
| 1059 | // FullyAvailableBlocks or they are not marked as speculatively available. |
| 1060 | Worklist.clear(); |
| 1061 | Worklist.append(in_start: succ_begin(BB: *UnavailableBB), in_end: succ_end(BB: *UnavailableBB)); |
| 1062 | while (!Worklist.empty()) |
| 1063 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
| 1064 | AvailabilityState::Unavailable); |
| 1065 | } |
| 1066 | |
| 1067 | #ifndef NDEBUG |
| 1068 | Worklist.clear(); |
| 1069 | for (BasicBlock *AvailableBB : AvailableBBs) |
| 1070 | Worklist.append(succ_begin(AvailableBB), succ_end(AvailableBB)); |
| 1071 | while (!Worklist.empty()) |
| 1072 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
| 1073 | AvailabilityState::Available); |
| 1074 | |
| 1075 | assert(NewSpeculativelyAvailableBBs.empty() && |
| 1076 | "Must have fixed all the new speculatively available blocks." ); |
| 1077 | #endif |
| 1078 | |
| 1079 | return !UnavailableBB; |
| 1080 | } |
| 1081 | |
| 1082 | /// If the specified OldValue exists in ValuesPerBlock, replace its value with |
| 1083 | /// NewValue. |
| 1084 | static void replaceValuesPerBlockEntry( |
| 1085 | SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, Value *OldValue, |
| 1086 | Value *NewValue) { |
| 1087 | for (AvailableValueInBlock &V : ValuesPerBlock) { |
| 1088 | if (V.AV.Val == OldValue) |
| 1089 | V.AV.Val = NewValue; |
| 1090 | if (V.AV.isSelectValue()) { |
| 1091 | if (V.AV.V1 == OldValue) |
| 1092 | V.AV.V1 = NewValue; |
| 1093 | if (V.AV.V2 == OldValue) |
| 1094 | V.AV.V2 = NewValue; |
| 1095 | } |
| 1096 | } |
| 1097 | } |
| 1098 | |
| 1099 | /// Given a set of loads specified by ValuesPerBlock, |
| 1100 | /// construct SSA form, allowing us to eliminate Load. This returns the value |
| 1101 | /// that should be used at Load's definition site. |
| 1102 | static Value * |
| 1103 | ConstructSSAForLoadSet(LoadInst *Load, |
| 1104 | SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, |
| 1105 | GVNPass &GVN) { |
| 1106 | // Check for the fully redundant, dominating load case. In this case, we can |
| 1107 | // just use the dominating value directly. |
| 1108 | if (ValuesPerBlock.size() == 1 && |
| 1109 | GVN.getDominatorTree().properlyDominates(A: ValuesPerBlock[0].BB, |
| 1110 | B: Load->getParent())) { |
| 1111 | assert(!ValuesPerBlock[0].AV.isUndefValue() && |
| 1112 | "Dead BB dominate this block" ); |
| 1113 | return ValuesPerBlock[0].MaterializeAdjustedValue(Load); |
| 1114 | } |
| 1115 | |
| 1116 | // Otherwise, we have to construct SSA form. |
| 1117 | SmallVector<PHINode*, 8> NewPHIs; |
| 1118 | SSAUpdater SSAUpdate(&NewPHIs); |
| 1119 | SSAUpdate.Initialize(Ty: Load->getType(), Name: Load->getName()); |
| 1120 | |
| 1121 | for (const AvailableValueInBlock &AV : ValuesPerBlock) { |
| 1122 | BasicBlock *BB = AV.BB; |
| 1123 | |
| 1124 | if (AV.AV.isUndefValue()) |
| 1125 | continue; |
| 1126 | |
| 1127 | if (SSAUpdate.HasValueForBlock(BB)) |
| 1128 | continue; |
| 1129 | |
| 1130 | // If the value is the load that we will be eliminating, and the block it's |
| 1131 | // available in is the block that the load is in, then don't add it as |
| 1132 | // SSAUpdater will resolve the value to the relevant phi which may let it |
| 1133 | // avoid phi construction entirely if there's actually only one value. |
| 1134 | if (BB == Load->getParent() && |
| 1135 | ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == Load) || |
| 1136 | (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == Load))) |
| 1137 | continue; |
| 1138 | |
| 1139 | SSAUpdate.AddAvailableValue(BB, V: AV.MaterializeAdjustedValue(Load)); |
| 1140 | } |
| 1141 | |
| 1142 | // Perform PHI construction. |
| 1143 | return SSAUpdate.GetValueInMiddleOfBlock(BB: Load->getParent()); |
| 1144 | } |
| 1145 | |
| 1146 | Value *AvailableValue::MaterializeAdjustedValue(LoadInst *Load, |
| 1147 | Instruction *InsertPt) const { |
| 1148 | Value *Res; |
| 1149 | Type *LoadTy = Load->getType(); |
| 1150 | const DataLayout &DL = Load->getDataLayout(); |
| 1151 | if (isSimpleValue()) { |
| 1152 | Res = getSimpleValue(); |
| 1153 | if (Res->getType() != LoadTy) { |
| 1154 | Res = getValueForLoad(SrcVal: Res, Offset, LoadTy, InsertPt, F: Load->getFunction()); |
| 1155 | |
| 1156 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset |
| 1157 | << " " << *getSimpleValue() << '\n' |
| 1158 | << *Res << '\n' |
| 1159 | << "\n\n\n" ); |
| 1160 | } |
| 1161 | } else if (isCoercedLoadValue()) { |
| 1162 | LoadInst *CoercedLoad = getCoercedLoadValue(); |
| 1163 | if (CoercedLoad->getType() == LoadTy && Offset == 0) { |
| 1164 | Res = CoercedLoad; |
| 1165 | combineMetadataForCSE(K: CoercedLoad, J: Load, DoesKMove: false); |
| 1166 | } else { |
| 1167 | Res = getValueForLoad(SrcVal: CoercedLoad, Offset, LoadTy, InsertPt, |
| 1168 | F: Load->getFunction()); |
| 1169 | // We are adding a new user for this load, for which the original |
| 1170 | // metadata may not hold. Additionally, the new load may have a different |
| 1171 | // size and type, so their metadata cannot be combined in any |
| 1172 | // straightforward way. |
| 1173 | // Drop all metadata that is not known to cause immediate UB on violation, |
| 1174 | // unless the load has !noundef, in which case all metadata violations |
| 1175 | // will be promoted to UB. |
| 1176 | // !noalias and !alias.scope are kept: the load is not moved and still |
| 1177 | // accesses the same memory, and these are independent of the load type |
| 1178 | // and offset, so they remain valid for the coerced result. |
| 1179 | if (!CoercedLoad->hasMetadata(KindID: LLVMContext::MD_noundef)) |
| 1180 | CoercedLoad->dropUnknownNonDebugMetadata( |
| 1181 | KnownIDs: {LLVMContext::MD_dereferenceable, |
| 1182 | LLVMContext::MD_dereferenceable_or_null, |
| 1183 | LLVMContext::MD_invariant_load, LLVMContext::MD_invariant_group, |
| 1184 | LLVMContext::MD_alias_scope, LLVMContext::MD_noalias}); |
| 1185 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset |
| 1186 | << " " << *getCoercedLoadValue() << '\n' |
| 1187 | << *Res << '\n' |
| 1188 | << "\n\n\n" ); |
| 1189 | } |
| 1190 | } else if (isMemIntrinValue()) { |
| 1191 | Res = getMemInstValueForLoad(SrcInst: getMemIntrinValue(), Offset, LoadTy, |
| 1192 | InsertPt, DL); |
| 1193 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset |
| 1194 | << " " << *getMemIntrinValue() << '\n' |
| 1195 | << *Res << '\n' |
| 1196 | << "\n\n\n" ); |
| 1197 | } else if (isSelectValue()) { |
| 1198 | // Introduce a new value select for a load from an eligible pointer select. |
| 1199 | Value *Cond = getSelectCondition(); |
| 1200 | assert(V1 && V2 && "both value operands of the select must be present" ); |
| 1201 | Res = SelectInst::Create(C: Cond, S1: V1, S2: V2, NameStr: "" , InsertBefore: InsertPt->getIterator()); |
| 1202 | // We use the DebugLoc from the original load here, as this instruction |
| 1203 | // materializes the value that would previously have been loaded. |
| 1204 | cast<SelectInst>(Val: Res)->setDebugLoc(Load->getDebugLoc()); |
| 1205 | } else { |
| 1206 | llvm_unreachable("Should not materialize value from dead block" ); |
| 1207 | } |
| 1208 | assert(Res && "failed to materialize?" ); |
| 1209 | return Res; |
| 1210 | } |
| 1211 | |
| 1212 | static bool isLifetimeStart(const Instruction *Inst) { |
| 1213 | if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Val: Inst)) |
| 1214 | return II->getIntrinsicID() == Intrinsic::lifetime_start; |
| 1215 | return false; |
| 1216 | } |
| 1217 | |
| 1218 | /// Assuming To can be reached from both From and Between, does Between lie on |
| 1219 | /// every path from From to To? |
| 1220 | static bool liesBetween(const Instruction *From, Instruction *Between, |
| 1221 | const Instruction *To, const DominatorTree *DT) { |
| 1222 | if (From->getParent() == Between->getParent()) |
| 1223 | return DT->dominates(Def: From, User: Between); |
| 1224 | SmallPtrSet<BasicBlock *, 1> Exclusion; |
| 1225 | Exclusion.insert(Ptr: Between->getParent()); |
| 1226 | return !isPotentiallyReachable(From, To, ExclusionSet: &Exclusion, DT); |
| 1227 | } |
| 1228 | |
| 1229 | static const Instruction *findMayClobberedPtrAccess(LoadInst *Load, |
| 1230 | const DominatorTree *DT) { |
| 1231 | Value *PtrOp = Load->getPointerOperand(); |
| 1232 | if (!PtrOp->hasUseList()) |
| 1233 | return nullptr; |
| 1234 | |
| 1235 | Instruction *OtherAccess = nullptr; |
| 1236 | |
| 1237 | for (auto *U : PtrOp->users()) { |
| 1238 | if (U != Load && (isa<LoadInst>(Val: U) || isa<StoreInst>(Val: U))) { |
| 1239 | auto *I = cast<Instruction>(Val: U); |
| 1240 | if (I->getFunction() == Load->getFunction() && DT->dominates(Def: I, User: Load)) { |
| 1241 | // Use the most immediately dominating value. |
| 1242 | if (OtherAccess) { |
| 1243 | if (DT->dominates(Def: OtherAccess, User: I)) |
| 1244 | OtherAccess = I; |
| 1245 | else |
| 1246 | assert(U == OtherAccess || DT->dominates(I, OtherAccess)); |
| 1247 | } else |
| 1248 | OtherAccess = I; |
| 1249 | } |
| 1250 | } |
| 1251 | } |
| 1252 | |
| 1253 | if (OtherAccess) |
| 1254 | return OtherAccess; |
| 1255 | |
| 1256 | // There is no dominating use, check if we can find a closest non-dominating |
| 1257 | // use that lies between any other potentially available use and Load. |
| 1258 | for (auto *U : PtrOp->users()) { |
| 1259 | if (U != Load && (isa<LoadInst>(Val: U) || isa<StoreInst>(Val: U))) { |
| 1260 | auto *I = cast<Instruction>(Val: U); |
| 1261 | if (I->getFunction() == Load->getFunction() && |
| 1262 | isPotentiallyReachable(From: I, To: Load, ExclusionSet: nullptr, DT)) { |
| 1263 | if (OtherAccess) { |
| 1264 | if (liesBetween(From: OtherAccess, Between: I, To: Load, DT)) { |
| 1265 | OtherAccess = I; |
| 1266 | } else if (!liesBetween(From: I, Between: OtherAccess, To: Load, DT)) { |
| 1267 | // These uses are both partially available at Load were it not for |
| 1268 | // the clobber, but neither lies strictly after the other. |
| 1269 | OtherAccess = nullptr; |
| 1270 | break; |
| 1271 | } // else: keep current OtherAccess since it lies between U and |
| 1272 | // Load. |
| 1273 | } else { |
| 1274 | OtherAccess = I; |
| 1275 | } |
| 1276 | } |
| 1277 | } |
| 1278 | } |
| 1279 | |
| 1280 | return OtherAccess; |
| 1281 | } |
| 1282 | |
| 1283 | /// Try to locate the three instruction involved in a missed |
| 1284 | /// load-elimination case that is due to an intervening store. |
| 1285 | static void (LoadInst *Load, Instruction *DepInst, |
| 1286 | const DominatorTree *DT, |
| 1287 | OptimizationRemarkEmitter *ORE) { |
| 1288 | using namespace ore; |
| 1289 | |
| 1290 | OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered" , Load); |
| 1291 | R << "load of type " << NV("Type" , Load->getType()) << " not eliminated" |
| 1292 | << setExtraArgs(); |
| 1293 | |
| 1294 | const Instruction *OtherAccess = findMayClobberedPtrAccess(Load, DT); |
| 1295 | if (OtherAccess) |
| 1296 | R << " in favor of " << NV("OtherAccess" , OtherAccess); |
| 1297 | |
| 1298 | R << " because it is clobbered by " << NV("ClobberedBy" , DepInst); |
| 1299 | |
| 1300 | ORE->emit(OptDiag&: R); |
| 1301 | } |
| 1302 | |
| 1303 | // Find a dominating value for Loc memory location in the extended basic block |
| 1304 | // (chain of basic blocks with single predecessors) starting From instruction. |
| 1305 | // Returns the value from a matching load or a simple store to the same pointer. |
| 1306 | static Value *findDominatingValue(const MemoryLocation &Loc, Type *LoadTy, |
| 1307 | Instruction *From, AAResults *AA) { |
| 1308 | uint32_t NumVisitedInsts = 0; |
| 1309 | BasicBlock *FromBB = From->getParent(); |
| 1310 | BatchAAResults BatchAA(*AA); |
| 1311 | for (BasicBlock *BB = FromBB; BB; BB = BB->getSinglePredecessor()) |
| 1312 | for (auto *Inst = BB == FromBB ? From : BB->getTerminator(); |
| 1313 | Inst != nullptr; Inst = Inst->getPrevNode()) { |
| 1314 | // Stop the search if limit is reached. |
| 1315 | if (++NumVisitedInsts > MaxNumVisitedInsts) |
| 1316 | return nullptr; |
| 1317 | if (isModSet(MRI: BatchAA.getModRefInfo(I: Inst, OptLoc: Loc))) { |
| 1318 | // A simple store to the exact location can forward its value. |
| 1319 | if (auto *SI = dyn_cast<StoreInst>(Val: Inst)) |
| 1320 | if (SI->isSimple() && SI->getPointerOperand() == Loc.Ptr && |
| 1321 | SI->getValueOperand()->getType() == LoadTy) |
| 1322 | return SI->getValueOperand(); |
| 1323 | return nullptr; |
| 1324 | } |
| 1325 | if (auto *LI = dyn_cast<LoadInst>(Val: Inst)) |
| 1326 | if (LI->getPointerOperand() == Loc.Ptr && LI->getType() == LoadTy) |
| 1327 | return LI; |
| 1328 | } |
| 1329 | return nullptr; |
| 1330 | } |
| 1331 | |
| 1332 | std::optional<AvailableValue> |
| 1333 | GVNPass::AnalyzeSelectAvailability(LoadInst *Load, Value *Cond, Value *TrueAddr, |
| 1334 | Value *FalseAddr, Instruction *From) { |
| 1335 | assert(TrueAddr->getType() == Load->getPointerOperandType() && |
| 1336 | "Invalid address type of true side of select dependency" ); |
| 1337 | assert(FalseAddr->getType() == Load->getPointerOperandType() && |
| 1338 | "Invalid address type of false side of select dependency" ); |
| 1339 | // We can convert a load through a select address into a select of the two |
| 1340 | // loaded values only if both sides have a dominating, non-clobbered value of |
| 1341 | // the right type in the extended basic block ending at From. |
| 1342 | auto Loc = MemoryLocation::get(LI: Load); |
| 1343 | Value *V1 = findDominatingValue(Loc: Loc.getWithNewPtr(NewPtr: TrueAddr), LoadTy: Load->getType(), |
| 1344 | From, AA: getAliasAnalysis()); |
| 1345 | if (!V1) |
| 1346 | return std::nullopt; |
| 1347 | Value *V2 = findDominatingValue(Loc: Loc.getWithNewPtr(NewPtr: FalseAddr), LoadTy: Load->getType(), |
| 1348 | From, AA: getAliasAnalysis()); |
| 1349 | if (!V2) |
| 1350 | return std::nullopt; |
| 1351 | return AvailableValue::getSelect(Cond, V1, V2); |
| 1352 | } |
| 1353 | |
| 1354 | std::optional<AvailableValue> |
| 1355 | GVNPass::AnalyzeLoadAvailability(LoadInst *Load, const ReachingMemVal &Dep, |
| 1356 | Value *Address) { |
| 1357 | assert(Load->isUnordered() && "rules below are incorrect for ordered access" ); |
| 1358 | assert((Dep.Kind == DepKind::Def || Dep.Kind == DepKind::Clobber) && |
| 1359 | "expected a local dependence" ); |
| 1360 | |
| 1361 | Instruction *DepInst = Dep.Inst; |
| 1362 | |
| 1363 | const DataLayout &DL = Load->getDataLayout(); |
| 1364 | if (Dep.Kind == DepKind::Clobber) { |
| 1365 | // If the dependence is to a store that writes to a superset of the bits |
| 1366 | // read by the load, we can extract the bits we need for the load from the |
| 1367 | // stored value. |
| 1368 | if (StoreInst *DepSI = dyn_cast<StoreInst>(Val: DepInst)) { |
| 1369 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1370 | if (Address && Load->isAtomic() <= DepSI->isAtomic()) { |
| 1371 | int Offset = |
| 1372 | analyzeLoadFromClobberingStore(LoadTy: Load->getType(), LoadPtr: Address, DepSI, DL); |
| 1373 | if (Offset != -1) |
| 1374 | return AvailableValue::get(V: DepSI->getValueOperand(), Offset); |
| 1375 | } |
| 1376 | } |
| 1377 | |
| 1378 | // Check to see if we have something like this: |
| 1379 | // load i32* P |
| 1380 | // load i8* (P+1) |
| 1381 | // if we have this, replace the later with an extraction from the former. |
| 1382 | if (LoadInst *DepLoad = dyn_cast<LoadInst>(Val: DepInst)) { |
| 1383 | // If this is a clobber and L is the first instruction in its block, then |
| 1384 | // we have the first instruction in the entry block. |
| 1385 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1386 | if (DepLoad != Load && Address && |
| 1387 | Load->isAtomic() <= DepLoad->isAtomic()) { |
| 1388 | Type *LoadType = Load->getType(); |
| 1389 | int Offset = Dep.Offset; |
| 1390 | |
| 1391 | if (!isMemorySSAEnabled()) { |
| 1392 | // If MD reported clobber, check it was nested. |
| 1393 | if (canCoerceMustAliasedValueToLoad(StoredVal: DepLoad, LoadTy: LoadType, |
| 1394 | F: DepLoad->getFunction())) { |
| 1395 | const auto ClobberOff = MD->getClobberOffset(DepInst: DepLoad); |
| 1396 | // GVN has no deal with a negative offset. |
| 1397 | Offset = (ClobberOff == std::nullopt || *ClobberOff < 0) |
| 1398 | ? -1 |
| 1399 | : *ClobberOff; |
| 1400 | } |
| 1401 | } else { |
| 1402 | if (!canCoerceMustAliasedValueToLoad(StoredVal: DepLoad, LoadTy: LoadType, |
| 1403 | F: DepLoad->getFunction()) || |
| 1404 | Offset < 0) |
| 1405 | Offset = -1; |
| 1406 | } |
| 1407 | if (Offset == -1) |
| 1408 | Offset = |
| 1409 | analyzeLoadFromClobberingLoad(LoadTy: LoadType, LoadPtr: Address, DepLI: DepLoad, DL); |
| 1410 | if (Offset != -1) |
| 1411 | return AvailableValue::getLoad(Load: DepLoad, Offset); |
| 1412 | } |
| 1413 | } |
| 1414 | |
| 1415 | // If the clobbering value is a memset/memcpy/memmove, see if we can |
| 1416 | // forward a value on from it. |
| 1417 | if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Val: DepInst)) { |
| 1418 | if (Address && !Load->isAtomic()) { |
| 1419 | int Offset = analyzeLoadFromClobberingMemInst(LoadTy: Load->getType(), LoadPtr: Address, |
| 1420 | DepMI, DL); |
| 1421 | if (Offset != -1) |
| 1422 | return AvailableValue::getMI(MI: DepMI, Offset); |
| 1423 | } |
| 1424 | } |
| 1425 | |
| 1426 | // Nothing known about this clobber, have to be conservative. |
| 1427 | LLVM_DEBUG( |
| 1428 | // fast print dep, using operator<< on instruction is too slow. |
| 1429 | dbgs() << "GVN: load " ; Load->printAsOperand(dbgs()); |
| 1430 | dbgs() << " is clobbered by " << *DepInst << '\n';); |
| 1431 | if (ORE->allowExtraAnalysis(DEBUG_TYPE)) |
| 1432 | reportMayClobberedLoad(Load, DepInst, DT, ORE); |
| 1433 | |
| 1434 | return std::nullopt; |
| 1435 | } |
| 1436 | assert(Dep.Kind == DepKind::Def && "follows from above" ); |
| 1437 | |
| 1438 | // Loading the alloca -> undef. |
| 1439 | // Loading immediately after lifetime begin -> undef. |
| 1440 | if (isa<AllocaInst>(Val: DepInst) || isLifetimeStart(Inst: DepInst)) |
| 1441 | return AvailableValue::get(V: UndefValue::get(T: Load->getType())); |
| 1442 | |
| 1443 | if (Constant *InitVal = |
| 1444 | getInitialValueOfAllocation(V: DepInst, TLI, Ty: Load->getType())) |
| 1445 | return AvailableValue::get(V: InitVal); |
| 1446 | |
| 1447 | if (StoreInst *S = dyn_cast<StoreInst>(Val: DepInst)) { |
| 1448 | // Reject loads and stores that are to the same address but are of |
| 1449 | // different types if we have to. If the stored value is convertable to |
| 1450 | // the loaded value, we can reuse it. |
| 1451 | if (!canCoerceMustAliasedValueToLoad(StoredVal: S->getValueOperand(), LoadTy: Load->getType(), |
| 1452 | F: S->getFunction())) |
| 1453 | return std::nullopt; |
| 1454 | |
| 1455 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1456 | if (S->isAtomic() < Load->isAtomic()) |
| 1457 | return std::nullopt; |
| 1458 | |
| 1459 | return AvailableValue::get(V: S->getValueOperand()); |
| 1460 | } |
| 1461 | |
| 1462 | if (LoadInst *LD = dyn_cast<LoadInst>(Val: DepInst)) { |
| 1463 | // If the types mismatch and we can't handle it, reject reuse of the load. |
| 1464 | // If the stored value is larger or equal to the loaded value, we can reuse |
| 1465 | // it. |
| 1466 | if (!canCoerceMustAliasedValueToLoad(StoredVal: LD, LoadTy: Load->getType(), |
| 1467 | F: LD->getFunction())) |
| 1468 | return std::nullopt; |
| 1469 | |
| 1470 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1471 | if (LD->isAtomic() < Load->isAtomic()) |
| 1472 | return std::nullopt; |
| 1473 | |
| 1474 | return AvailableValue::getLoad(Load: LD); |
| 1475 | } |
| 1476 | |
| 1477 | // Check if load with Addr dependent from select can be converted to select |
| 1478 | // between load values. There must be no instructions between the found |
| 1479 | // loads and DepInst that may clobber the loads. |
| 1480 | if (auto *Sel = dyn_cast<SelectInst>(Val: DepInst)) { |
| 1481 | assert(Sel->getType() == Load->getPointerOperandType()); |
| 1482 | if (auto AV = AnalyzeSelectAvailability(Load, Cond: Sel->getCondition(), |
| 1483 | TrueAddr: Sel->getTrueValue(), |
| 1484 | FalseAddr: Sel->getFalseValue(), From: DepInst)) |
| 1485 | return AV; |
| 1486 | return std::nullopt; |
| 1487 | } |
| 1488 | |
| 1489 | // Unknown def - must be conservative. |
| 1490 | LLVM_DEBUG( |
| 1491 | // fast print dep, using operator<< on instruction is too slow. |
| 1492 | dbgs() << "GVN: load " ; Load->printAsOperand(dbgs()); |
| 1493 | dbgs() << " has unknown def " << *DepInst << '\n';); |
| 1494 | return std::nullopt; |
| 1495 | } |
| 1496 | |
| 1497 | void GVNPass::AnalyzeLoadAvailability(LoadInst *Load, |
| 1498 | SmallVectorImpl<ReachingMemVal> &Deps, |
| 1499 | AvailValInBlkVect &ValuesPerBlock, |
| 1500 | UnavailBlkVect &UnavailableBlocks) { |
| 1501 | // Filter out useless results (non-locals, etc). Keep track of the blocks |
| 1502 | // where we have a value available in repl, also keep track of whether we see |
| 1503 | // dependencies that produce an unknown value for the load (such as a call |
| 1504 | // that could potentially clobber the load). |
| 1505 | for (const auto &Dep : Deps) { |
| 1506 | BasicBlock *DepBB = Dep.Block; |
| 1507 | |
| 1508 | if (DeadBlocks.count(key: DepBB)) { |
| 1509 | // Dead dependent mem-op disguise as a load evaluating the same value |
| 1510 | // as the load in question. |
| 1511 | ValuesPerBlock.push_back(Elt: AvailableValueInBlock::getUndef(BB: DepBB)); |
| 1512 | continue; |
| 1513 | } |
| 1514 | |
| 1515 | if (Dep.Kind == DepKind::Other) { |
| 1516 | UnavailableBlocks.push_back(Elt: DepBB); |
| 1517 | continue; |
| 1518 | } |
| 1519 | |
| 1520 | // The load address is a select in this block: try to rematerialize the |
| 1521 | // load as a select of the two reaching values (one per side). The values |
| 1522 | // are searched for at the end of DepBB. |
| 1523 | if (Dep.Kind == DepKind::Select) { |
| 1524 | if (auto AV = AnalyzeSelectAvailability( |
| 1525 | Load, Cond: const_cast<Value *>(Dep.SelCond), |
| 1526 | TrueAddr: const_cast<Value *>(Dep.SelTrueAddr), |
| 1527 | FalseAddr: const_cast<Value *>(Dep.SelFalseAddr), From: DepBB->getTerminator())) { |
| 1528 | ValuesPerBlock.push_back( |
| 1529 | Elt: AvailableValueInBlock::get(BB: DepBB, AV: std::move(*AV))); |
| 1530 | } else { |
| 1531 | UnavailableBlocks.push_back(Elt: DepBB); |
| 1532 | } |
| 1533 | continue; |
| 1534 | } |
| 1535 | |
| 1536 | // The address being loaded in this non-local block may not be the same as |
| 1537 | // the pointer operand of the load if PHI translation occurs. Make sure |
| 1538 | // to consider the right address. |
| 1539 | if (auto AV = |
| 1540 | AnalyzeLoadAvailability(Load, Dep, Address: const_cast<Value *>(Dep.Addr))) { |
| 1541 | // subtlety: because we know this was a non-local dependency, we know |
| 1542 | // it's safe to materialize anywhere between the instruction within |
| 1543 | // DepInfo and the end of it's block. |
| 1544 | ValuesPerBlock.push_back( |
| 1545 | Elt: AvailableValueInBlock::get(BB: DepBB, AV: std::move(*AV))); |
| 1546 | } else { |
| 1547 | UnavailableBlocks.push_back(Elt: DepBB); |
| 1548 | } |
| 1549 | } |
| 1550 | |
| 1551 | assert(Deps.size() == ValuesPerBlock.size() + UnavailableBlocks.size() && |
| 1552 | "post condition violation" ); |
| 1553 | } |
| 1554 | |
| 1555 | /// Given the following code, v1 is partially available on some edges, but not |
| 1556 | /// available on the edge from PredBB. This function tries to find if there is |
| 1557 | /// another identical load in the other successor of PredBB. |
| 1558 | /// |
| 1559 | /// v0 = load %addr |
| 1560 | /// br %LoadBB |
| 1561 | /// |
| 1562 | /// LoadBB: |
| 1563 | /// v1 = load %addr |
| 1564 | /// ... |
| 1565 | /// |
| 1566 | /// PredBB: |
| 1567 | /// ... |
| 1568 | /// br %cond, label %LoadBB, label %SuccBB |
| 1569 | /// |
| 1570 | /// SuccBB: |
| 1571 | /// v2 = load %addr |
| 1572 | /// ... |
| 1573 | /// |
| 1574 | LoadInst *GVNPass::findLoadToHoistIntoPred(BasicBlock *Pred, BasicBlock *LoadBB, |
| 1575 | LoadInst *Load) { |
| 1576 | // For simplicity we handle a Pred has 2 successors only. |
| 1577 | auto *Term = Pred->getTerminator(); |
| 1578 | if (Term->getNumSuccessors() != 2 || Term->isSpecialTerminator()) |
| 1579 | return nullptr; |
| 1580 | auto *SuccBB = Term->getSuccessor(Idx: 0); |
| 1581 | if (SuccBB == LoadBB) |
| 1582 | SuccBB = Term->getSuccessor(Idx: 1); |
| 1583 | if (!SuccBB->getSinglePredecessor()) |
| 1584 | return nullptr; |
| 1585 | |
| 1586 | unsigned int NumInsts = MaxNumInsnsPerBlock; |
| 1587 | for (Instruction &Inst : *SuccBB) { |
| 1588 | if (Inst.isDebugOrPseudoInst()) |
| 1589 | continue; |
| 1590 | if (--NumInsts == 0) |
| 1591 | return nullptr; |
| 1592 | |
| 1593 | if (!Inst.isIdenticalTo(I: Load)) |
| 1594 | continue; |
| 1595 | |
| 1596 | bool HasLocalDep = true; |
| 1597 | if (!isMemorySSAEnabled()) { |
| 1598 | MemDepResult Dep = MD->getDependency(QueryInst: &Inst); |
| 1599 | HasLocalDep = !Dep.isNonLocal(); |
| 1600 | } else { |
| 1601 | auto *MSSA = MSSAU->getMemorySSA(); |
| 1602 | // Do not hoist if the identical load has ordering constraint. |
| 1603 | if (auto *MA = MSSA->getMemoryAccess(I: &Inst); MA && isa<MemoryUse>(Val: MA)) { |
| 1604 | auto *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(MA); |
| 1605 | HasLocalDep = Clobber->getBlock() == SuccBB; |
| 1606 | } |
| 1607 | } |
| 1608 | |
| 1609 | // If an identical load doesn't depends on any local instructions, it can |
| 1610 | // be safely moved to PredBB. |
| 1611 | // Also check for the implicit control flow instructions. See the comments |
| 1612 | // in PerformLoadPRE for details. |
| 1613 | if (!HasLocalDep && !ICF->isDominatedByICFIFromSameBlock(Insn: &Inst)) |
| 1614 | return cast<LoadInst>(Val: &Inst); |
| 1615 | |
| 1616 | // Otherwise there is something in the same BB clobbers the memory, we can't |
| 1617 | // move this and later load to PredBB. |
| 1618 | return nullptr; |
| 1619 | } |
| 1620 | |
| 1621 | return nullptr; |
| 1622 | } |
| 1623 | |
| 1624 | void GVNPass::eliminatePartiallyRedundantLoad( |
| 1625 | LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
| 1626 | MapVector<BasicBlock *, Value *> &AvailableLoads, |
| 1627 | MapVector<BasicBlock *, LoadInst *> *CriticalEdgePredAndLoad) { |
| 1628 | for (const auto &AvailableLoad : AvailableLoads) { |
| 1629 | BasicBlock *UnavailableBlock = AvailableLoad.first; |
| 1630 | Value *LoadPtr = AvailableLoad.second; |
| 1631 | |
| 1632 | auto *NewLoad = |
| 1633 | new LoadInst(Load->getType(), LoadPtr, Load->getName() + ".pre" , |
| 1634 | Load->getProperties(), |
| 1635 | UnavailableBlock->getTerminator()->getIterator()); |
| 1636 | NewLoad->setDebugLoc(Load->getDebugLoc()); |
| 1637 | if (MSSAU) { |
| 1638 | auto *NewAccess = MSSAU->createMemoryAccessInBB( |
| 1639 | I: NewLoad, Definition: nullptr, BB: NewLoad->getParent(), Point: MemorySSA::BeforeTerminator); |
| 1640 | if (auto *NewDef = dyn_cast<MemoryDef>(Val: NewAccess)) |
| 1641 | MSSAU->insertDef(Def: NewDef, /*RenameUses=*/true); |
| 1642 | else |
| 1643 | MSSAU->insertUse(Use: cast<MemoryUse>(Val: NewAccess), /*RenameUses=*/true); |
| 1644 | } |
| 1645 | |
| 1646 | // Transfer the old load's AA tags to the new load. |
| 1647 | AAMDNodes Tags = Load->getAAMetadata(); |
| 1648 | if (Tags) |
| 1649 | NewLoad->setAAMetadata(Tags); |
| 1650 | |
| 1651 | if (auto *MD = Load->getMetadata(KindID: LLVMContext::MD_invariant_load)) |
| 1652 | NewLoad->setMetadata(KindID: LLVMContext::MD_invariant_load, Node: MD); |
| 1653 | if (auto *InvGroupMD = Load->getMetadata(KindID: LLVMContext::MD_invariant_group)) |
| 1654 | NewLoad->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: InvGroupMD); |
| 1655 | if (auto *RangeMD = Load->getMetadata(KindID: LLVMContext::MD_range)) |
| 1656 | NewLoad->setMetadata(KindID: LLVMContext::MD_range, Node: RangeMD); |
| 1657 | if (auto *NoFPClassMD = Load->getMetadata(KindID: LLVMContext::MD_nofpclass)) |
| 1658 | NewLoad->setMetadata(KindID: LLVMContext::MD_nofpclass, Node: NoFPClassMD); |
| 1659 | |
| 1660 | if (auto *AccessMD = Load->getMetadata(KindID: LLVMContext::MD_access_group)) |
| 1661 | if (LI->getLoopFor(BB: Load->getParent()) == LI->getLoopFor(BB: UnavailableBlock)) |
| 1662 | NewLoad->setMetadata(KindID: LLVMContext::MD_access_group, Node: AccessMD); |
| 1663 | |
| 1664 | // We do not propagate the old load's debug location, because the new |
| 1665 | // load now lives in a different BB, and we want to avoid a jumpy line |
| 1666 | // table. |
| 1667 | // FIXME: How do we retain source locations without causing poor debugging |
| 1668 | // behavior? |
| 1669 | |
| 1670 | // Add the newly created load. |
| 1671 | ValuesPerBlock.push_back( |
| 1672 | Elt: AvailableValueInBlock::get(BB: UnavailableBlock, V: NewLoad)); |
| 1673 | if (MD) |
| 1674 | MD->invalidateCachedPointerInfo(Ptr: LoadPtr); |
| 1675 | LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); |
| 1676 | |
| 1677 | // For PredBB in CriticalEdgePredAndLoad we need to replace the uses of old |
| 1678 | // load instruction with the new created load instruction. |
| 1679 | if (CriticalEdgePredAndLoad) { |
| 1680 | auto It = CriticalEdgePredAndLoad->find(Key: UnavailableBlock); |
| 1681 | if (It != CriticalEdgePredAndLoad->end()) { |
| 1682 | ++NumPRELoadMoved2CEPred; |
| 1683 | ICF->insertInstructionTo(Inst: NewLoad, BB: UnavailableBlock); |
| 1684 | LoadInst *OldLoad = It->second; |
| 1685 | combineMetadataForCSE(K: NewLoad, J: OldLoad, /*DoesKMove=*/true); |
| 1686 | OldLoad->replaceAllUsesWith(V: NewLoad); |
| 1687 | replaceValuesPerBlockEntry(ValuesPerBlock, OldValue: OldLoad, NewValue: NewLoad); |
| 1688 | if (uint32_t ValNo = VN.lookup(V: OldLoad, Verify: false)) |
| 1689 | LeaderTable.erase(N: ValNo, I: OldLoad, BB: OldLoad->getParent()); |
| 1690 | removeInstruction(I: OldLoad); |
| 1691 | } |
| 1692 | } |
| 1693 | } |
| 1694 | |
| 1695 | // Perform PHI construction. |
| 1696 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, GVN&: *this); |
| 1697 | // ConstructSSAForLoadSet is responsible for combining metadata. |
| 1698 | ICF->removeUsersOf(Inst: Load); |
| 1699 | Load->replaceAllUsesWith(V); |
| 1700 | if (isa<PHINode>(Val: V)) |
| 1701 | V->takeName(V: Load); |
| 1702 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) |
| 1703 | I->setDebugLoc(Load->getDebugLoc()); |
| 1704 | if (MD && V->getType()->isPtrOrPtrVectorTy()) |
| 1705 | MD->invalidateCachedPointerInfo(Ptr: V); |
| 1706 | ORE->emit(RemarkBuilder: [&]() { |
| 1707 | return OptimizationRemark(DEBUG_TYPE, "LoadPRE" , Load) |
| 1708 | << "load eliminated by PRE" ; |
| 1709 | }); |
| 1710 | salvageAndRemoveInstruction(I: Load); |
| 1711 | } |
| 1712 | |
| 1713 | bool GVNPass::PerformLoadPRE(LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
| 1714 | UnavailBlkVect &UnavailableBlocks) { |
| 1715 | // Okay, we have *some* definitions of the value. This means that the value |
| 1716 | // is available in some of our (transitive) predecessors. Lets think about |
| 1717 | // doing PRE of this load. This will involve inserting a new load into the |
| 1718 | // predecessor when it's not available. We could do this in general, but |
| 1719 | // prefer to not increase code size. As such, we only do this when we know |
| 1720 | // that we only have to insert *one* load (which means we're basically moving |
| 1721 | // the load, not inserting a new one). |
| 1722 | |
| 1723 | SmallPtrSet<BasicBlock *, 4> Blockers(llvm::from_range, UnavailableBlocks); |
| 1724 | |
| 1725 | // Let's find the first basic block with more than one predecessor. Walk |
| 1726 | // backwards through predecessors if needed. |
| 1727 | BasicBlock *LoadBB = Load->getParent(); |
| 1728 | BasicBlock *TmpBB = LoadBB; |
| 1729 | |
| 1730 | // Check that there is no implicit control flow instructions above our load in |
| 1731 | // its block. If there is an instruction that doesn't always pass the |
| 1732 | // execution to the following instruction, then moving through it may become |
| 1733 | // invalid. For example: |
| 1734 | // |
| 1735 | // int arr[LEN]; |
| 1736 | // int index = ???; |
| 1737 | // ... |
| 1738 | // guard(0 <= index && index < LEN); |
| 1739 | // use(arr[index]); |
| 1740 | // |
| 1741 | // It is illegal to move the array access to any point above the guard, |
| 1742 | // because if the index is out of bounds we should deoptimize rather than |
| 1743 | // access the array. |
| 1744 | // Check that there is no guard in this block above our instruction. |
| 1745 | bool MustEnsureSafetyOfSpeculativeExecution = |
| 1746 | ICF->isDominatedByICFIFromSameBlock(Insn: Load); |
| 1747 | |
| 1748 | while (TmpBB->getSinglePredecessor()) { |
| 1749 | TmpBB = TmpBB->getSinglePredecessor(); |
| 1750 | if (TmpBB == LoadBB) // Infinite (unreachable) loop. |
| 1751 | return false; |
| 1752 | if (Blockers.count(Ptr: TmpBB)) |
| 1753 | return false; |
| 1754 | |
| 1755 | // If any of these blocks has more than one successor (i.e. if the edge we |
| 1756 | // just traversed was critical), then there are other paths through this |
| 1757 | // block along which the load may not be anticipated. Hoisting the load |
| 1758 | // above this block would be adding the load to execution paths along |
| 1759 | // which it was not previously executed. |
| 1760 | if (TmpBB->getTerminator()->getNumSuccessors() != 1) |
| 1761 | return false; |
| 1762 | |
| 1763 | // Check that there is no implicit control flow in a block above. |
| 1764 | MustEnsureSafetyOfSpeculativeExecution = |
| 1765 | MustEnsureSafetyOfSpeculativeExecution || ICF->hasICF(BB: TmpBB); |
| 1766 | } |
| 1767 | |
| 1768 | assert(TmpBB); |
| 1769 | LoadBB = TmpBB; |
| 1770 | |
| 1771 | // Check to see how many predecessors have the loaded value fully |
| 1772 | // available. |
| 1773 | MapVector<BasicBlock *, Value *> PredLoads; |
| 1774 | DenseMap<BasicBlock *, AvailabilityState> FullyAvailableBlocks; |
| 1775 | for (const AvailableValueInBlock &AV : ValuesPerBlock) |
| 1776 | FullyAvailableBlocks[AV.BB] = AvailabilityState::Available; |
| 1777 | for (BasicBlock *UnavailableBB : UnavailableBlocks) |
| 1778 | FullyAvailableBlocks[UnavailableBB] = AvailabilityState::Unavailable; |
| 1779 | |
| 1780 | // The edge from Pred to LoadBB is a critical edge will be splitted. |
| 1781 | SmallVector<BasicBlock *, 4> CriticalEdgePredSplit; |
| 1782 | // The edge from Pred to LoadBB is a critical edge, another successor of Pred |
| 1783 | // contains a load can be moved to Pred. This data structure maps the Pred to |
| 1784 | // the movable load. |
| 1785 | MapVector<BasicBlock *, LoadInst *> CriticalEdgePredAndLoad; |
| 1786 | for (BasicBlock *Pred : predecessors(BB: LoadBB)) { |
| 1787 | // If any predecessor block is an EH pad that does not allow non-PHI |
| 1788 | // instructions before the terminator, we can't PRE the load. |
| 1789 | if (Pred->getTerminator()->isEHPad()) { |
| 1790 | LLVM_DEBUG( |
| 1791 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" |
| 1792 | << Pred->getName() << "': " << *Load << '\n'); |
| 1793 | return false; |
| 1794 | } |
| 1795 | |
| 1796 | if (IsValueFullyAvailableInBlock(BB: Pred, FullyAvailableBlocks)) { |
| 1797 | continue; |
| 1798 | } |
| 1799 | |
| 1800 | if (Pred->getTerminator()->getNumSuccessors() != 1) { |
| 1801 | if (isa<IndirectBrInst>(Val: Pred->getTerminator())) { |
| 1802 | LLVM_DEBUG( |
| 1803 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" |
| 1804 | << Pred->getName() << "': " << *Load << '\n'); |
| 1805 | return false; |
| 1806 | } |
| 1807 | |
| 1808 | if (LoadBB->isEHPad()) { |
| 1809 | LLVM_DEBUG( |
| 1810 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" |
| 1811 | << Pred->getName() << "': " << *Load << '\n'); |
| 1812 | return false; |
| 1813 | } |
| 1814 | |
| 1815 | // Do not split backedge as it will break the canonical loop form. |
| 1816 | if (!isLoadPRESplitBackedgeEnabled()) |
| 1817 | if (DT->dominates(A: LoadBB, B: Pred)) { |
| 1818 | LLVM_DEBUG( |
| 1819 | dbgs() |
| 1820 | << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" |
| 1821 | << Pred->getName() << "': " << *Load << '\n'); |
| 1822 | return false; |
| 1823 | } |
| 1824 | |
| 1825 | if (LoadInst *LI = findLoadToHoistIntoPred(Pred, LoadBB, Load)) |
| 1826 | CriticalEdgePredAndLoad[Pred] = LI; |
| 1827 | else |
| 1828 | CriticalEdgePredSplit.push_back(Elt: Pred); |
| 1829 | } else { |
| 1830 | // Only add the predecessors that will not be split for now. |
| 1831 | PredLoads[Pred] = nullptr; |
| 1832 | } |
| 1833 | } |
| 1834 | |
| 1835 | // Decide whether PRE is profitable for this load. |
| 1836 | unsigned NumInsertPreds = PredLoads.size() + CriticalEdgePredSplit.size(); |
| 1837 | unsigned NumUnavailablePreds = NumInsertPreds + |
| 1838 | CriticalEdgePredAndLoad.size(); |
| 1839 | assert(NumUnavailablePreds != 0 && |
| 1840 | "Fully available value should already be eliminated!" ); |
| 1841 | (void)NumUnavailablePreds; |
| 1842 | |
| 1843 | // If we need to insert new load in multiple predecessors, reject it. |
| 1844 | // FIXME: If we could restructure the CFG, we could make a common pred with |
| 1845 | // all the preds that don't have an available Load and insert a new load into |
| 1846 | // that one block. |
| 1847 | if (NumInsertPreds > 1) |
| 1848 | return false; |
| 1849 | |
| 1850 | // Now we know where we will insert load. We must ensure that it is safe |
| 1851 | // to speculatively execute the load at that points. |
| 1852 | if (MustEnsureSafetyOfSpeculativeExecution) { |
| 1853 | if (CriticalEdgePredSplit.size()) |
| 1854 | if (!isSafeToSpeculativelyExecute(I: Load, CtxI: &*LoadBB->getFirstNonPHIIt(), AC, |
| 1855 | DT)) |
| 1856 | return false; |
| 1857 | for (auto &PL : PredLoads) |
| 1858 | if (!isSafeToSpeculativelyExecute(I: Load, CtxI: PL.first->getTerminator(), AC, |
| 1859 | DT)) |
| 1860 | return false; |
| 1861 | for (auto &CEP : CriticalEdgePredAndLoad) |
| 1862 | if (!isSafeToSpeculativelyExecute(I: Load, CtxI: CEP.first->getTerminator(), AC, |
| 1863 | DT)) |
| 1864 | return false; |
| 1865 | } |
| 1866 | |
| 1867 | // Split critical edges, and update the unavailable predecessors accordingly. |
| 1868 | for (BasicBlock *OrigPred : CriticalEdgePredSplit) { |
| 1869 | BasicBlock *NewPred = splitCriticalEdges(Pred: OrigPred, Succ: LoadBB); |
| 1870 | assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!" ); |
| 1871 | PredLoads[NewPred] = nullptr; |
| 1872 | LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->" |
| 1873 | << LoadBB->getName() << '\n'); |
| 1874 | } |
| 1875 | |
| 1876 | for (auto &CEP : CriticalEdgePredAndLoad) |
| 1877 | PredLoads[CEP.first] = nullptr; |
| 1878 | |
| 1879 | // Check if the load can safely be moved to all the unavailable predecessors. |
| 1880 | bool CanDoPRE = true; |
| 1881 | const DataLayout &DL = Load->getDataLayout(); |
| 1882 | SmallVector<Instruction*, 8> NewInsts; |
| 1883 | for (auto &PredLoad : PredLoads) { |
| 1884 | BasicBlock *UnavailablePred = PredLoad.first; |
| 1885 | |
| 1886 | // Do PHI translation to get its value in the predecessor if necessary. The |
| 1887 | // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. |
| 1888 | // We do the translation for each edge we skipped by going from Load's block |
| 1889 | // to LoadBB, otherwise we might miss pieces needing translation. |
| 1890 | |
| 1891 | // If all preds have a single successor, then we know it is safe to insert |
| 1892 | // the load on the pred (?!?), so we can insert code to materialize the |
| 1893 | // pointer if it is not available. |
| 1894 | Value *LoadPtr = Load->getPointerOperand(); |
| 1895 | BasicBlock *Cur = Load->getParent(); |
| 1896 | while (Cur != LoadBB) { |
| 1897 | PHITransAddr Address(LoadPtr, DL, AC); |
| 1898 | LoadPtr = Address.translateWithInsertion(CurBB: Cur, PredBB: Cur->getSinglePredecessor(), |
| 1899 | DT: *DT, NewInsts); |
| 1900 | if (!LoadPtr) { |
| 1901 | CanDoPRE = false; |
| 1902 | break; |
| 1903 | } |
| 1904 | Cur = Cur->getSinglePredecessor(); |
| 1905 | } |
| 1906 | |
| 1907 | if (LoadPtr) { |
| 1908 | PHITransAddr Address(LoadPtr, DL, AC); |
| 1909 | LoadPtr = Address.translateWithInsertion(CurBB: LoadBB, PredBB: UnavailablePred, DT: *DT, |
| 1910 | NewInsts); |
| 1911 | } |
| 1912 | // If we couldn't find or insert a computation of this phi translated value, |
| 1913 | // we fail PRE. |
| 1914 | if (!LoadPtr) { |
| 1915 | LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " |
| 1916 | << *Load->getPointerOperand() << "\n" ); |
| 1917 | CanDoPRE = false; |
| 1918 | break; |
| 1919 | } |
| 1920 | |
| 1921 | PredLoad.second = LoadPtr; |
| 1922 | } |
| 1923 | |
| 1924 | if (!CanDoPRE) { |
| 1925 | while (!NewInsts.empty()) { |
| 1926 | // Erase instructions generated by the failed PHI translation before |
| 1927 | // trying to number them. PHI translation might insert instructions |
| 1928 | // in basic blocks other than the current one, and we delete them |
| 1929 | // directly, as salvageAndRemoveInstruction only allows removing from the |
| 1930 | // current basic block. |
| 1931 | NewInsts.pop_back_val()->eraseFromParent(); |
| 1932 | } |
| 1933 | // HINT: Don't revert the edge-splitting as following transformation may |
| 1934 | // also need to split these critical edges. |
| 1935 | return !CriticalEdgePredSplit.empty(); |
| 1936 | } |
| 1937 | |
| 1938 | // Okay, we can eliminate this load by inserting a reload in the predecessor |
| 1939 | // and using PHI construction to get the value in the other predecessors, do |
| 1940 | // it. |
| 1941 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *Load << '\n'); |
| 1942 | LLVM_DEBUG(if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() |
| 1943 | << " INSTS: " << *NewInsts.back() |
| 1944 | << '\n'); |
| 1945 | |
| 1946 | // Assign value numbers to the new instructions. |
| 1947 | for (Instruction *I : NewInsts) { |
| 1948 | // Instructions that have been inserted in predecessor(s) to materialize |
| 1949 | // the load address do not retain their original debug locations. Doing |
| 1950 | // so could lead to confusing (but correct) source attributions. |
| 1951 | I->updateLocationAfterHoist(); |
| 1952 | |
| 1953 | // FIXME: We really _ought_ to insert these value numbers into their |
| 1954 | // parent's availability map. However, in doing so, we risk getting into |
| 1955 | // ordering issues. If a block hasn't been processed yet, we would be |
| 1956 | // marking a value as AVAIL-IN, which isn't what we intend. |
| 1957 | VN.lookupOrAdd(V: I); |
| 1958 | } |
| 1959 | |
| 1960 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, AvailableLoads&: PredLoads, |
| 1961 | CriticalEdgePredAndLoad: &CriticalEdgePredAndLoad); |
| 1962 | ++NumPRELoad; |
| 1963 | return true; |
| 1964 | } |
| 1965 | |
| 1966 | bool GVNPass::performLoopLoadPRE(LoadInst *Load, |
| 1967 | AvailValInBlkVect &ValuesPerBlock, |
| 1968 | UnavailBlkVect &UnavailableBlocks) { |
| 1969 | const Loop *L = LI->getLoopFor(BB: Load->getParent()); |
| 1970 | // TODO: Generalize to other loop blocks that dominate the latch. |
| 1971 | if (!L || L->getHeader() != Load->getParent()) |
| 1972 | return false; |
| 1973 | |
| 1974 | BasicBlock * = L->getLoopPreheader(); |
| 1975 | BasicBlock *Latch = L->getLoopLatch(); |
| 1976 | if (!Preheader || !Latch) |
| 1977 | return false; |
| 1978 | |
| 1979 | Value *LoadPtr = Load->getPointerOperand(); |
| 1980 | // Must be available in preheader. |
| 1981 | if (!L->isLoopInvariant(V: LoadPtr)) |
| 1982 | return false; |
| 1983 | |
| 1984 | // We plan to hoist the load to preheader without introducing a new fault. |
| 1985 | // In order to do it, we need to prove that we cannot side-exit the loop |
| 1986 | // once loop header is first entered before execution of the load. |
| 1987 | if (ICF->isDominatedByICFIFromSameBlock(Insn: Load)) |
| 1988 | return false; |
| 1989 | |
| 1990 | BasicBlock *LoopBlock = nullptr; |
| 1991 | for (auto *Blocker : UnavailableBlocks) { |
| 1992 | // Blockers from outside the loop are handled in preheader. |
| 1993 | if (!L->contains(BB: Blocker)) |
| 1994 | continue; |
| 1995 | |
| 1996 | // Only allow one loop block. Loop header is not less frequently executed |
| 1997 | // than each loop block, and likely it is much more frequently executed. But |
| 1998 | // in case of multiple loop blocks, we need extra information (such as block |
| 1999 | // frequency info) to understand whether it is profitable to PRE into |
| 2000 | // multiple loop blocks. |
| 2001 | if (LoopBlock) |
| 2002 | return false; |
| 2003 | |
| 2004 | // Do not sink into inner loops. This may be non-profitable. |
| 2005 | if (L != LI->getLoopFor(BB: Blocker)) |
| 2006 | return false; |
| 2007 | |
| 2008 | // Blocks that dominate the latch execute on every single iteration, maybe |
| 2009 | // except the last one. So PREing into these blocks doesn't make much sense |
| 2010 | // in most cases. But the blocks that do not necessarily execute on each |
| 2011 | // iteration are sometimes much colder than the header, and this is when |
| 2012 | // PRE is potentially profitable. |
| 2013 | if (DT->dominates(A: Blocker, B: Latch)) |
| 2014 | return false; |
| 2015 | |
| 2016 | // Make sure that the terminator itself doesn't clobber. |
| 2017 | if (Blocker->getTerminator()->mayWriteToMemory()) |
| 2018 | return false; |
| 2019 | |
| 2020 | LoopBlock = Blocker; |
| 2021 | } |
| 2022 | |
| 2023 | if (!LoopBlock) |
| 2024 | return false; |
| 2025 | |
| 2026 | // Make sure the memory at this pointer cannot be freed, therefore we can |
| 2027 | // safely reload from it after clobber. |
| 2028 | if (LoadPtr->canBeFreed()) |
| 2029 | return false; |
| 2030 | |
| 2031 | // TODO: Support critical edge splitting if blocker has more than 1 successor. |
| 2032 | MapVector<BasicBlock *, Value *> AvailableLoads; |
| 2033 | AvailableLoads[LoopBlock] = LoadPtr; |
| 2034 | AvailableLoads[Preheader] = LoadPtr; |
| 2035 | |
| 2036 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOOP LOAD: " << *Load << '\n'); |
| 2037 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, AvailableLoads, |
| 2038 | /*CriticalEdgePredAndLoad*/ nullptr); |
| 2039 | ++NumPRELoopLoad; |
| 2040 | return true; |
| 2041 | } |
| 2042 | |
| 2043 | static void (LoadInst *Load, Value *AvailableValue, |
| 2044 | OptimizationRemarkEmitter *ORE) { |
| 2045 | using namespace ore; |
| 2046 | |
| 2047 | ORE->emit(RemarkBuilder: [&]() { |
| 2048 | return OptimizationRemark(DEBUG_TYPE, "LoadElim" , Load) |
| 2049 | << "load of type " << NV("Type" , Load->getType()) << " eliminated" |
| 2050 | << setExtraArgs() << " in favor of " |
| 2051 | << NV("InfavorOfValue" , AvailableValue); |
| 2052 | }); |
| 2053 | } |
| 2054 | |
| 2055 | /// Attempt to eliminate a load whose dependencies are |
| 2056 | /// non-local by performing PHI construction. |
| 2057 | bool GVNPass::processNonLocalLoad(LoadInst *Load) { |
| 2058 | // Non-local speculations are not allowed under asan. |
| 2059 | if (Load->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeAddress) || |
| 2060 | Load->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeHWAddress)) |
| 2061 | return false; |
| 2062 | |
| 2063 | // Find the non-local dependencies of the load. |
| 2064 | LoadDepVect Deps; |
| 2065 | MD->getNonLocalPointerDependency(QueryInst: Load, Result&: Deps); |
| 2066 | |
| 2067 | // If we had to process more than one hundred blocks to find the |
| 2068 | // dependencies, this load isn't worth worrying about. Optimizing |
| 2069 | // it will be too expensive. |
| 2070 | unsigned NumDeps = Deps.size(); |
| 2071 | if (NumDeps > MaxNumDeps) |
| 2072 | return false; |
| 2073 | |
| 2074 | SmallVector<ReachingMemVal, 64> MemVals; |
| 2075 | MemVals.reserve(N: Deps.size()); |
| 2076 | |
| 2077 | for (const NonLocalDepResult &Dep : Deps) { |
| 2078 | const auto &R = Dep.getResult(); |
| 2079 | SelectAddr SelAddr = Dep.getAddress(); |
| 2080 | BasicBlock *BB = Dep.getBB(); |
| 2081 | Instruction *Inst = R.getInst(); |
| 2082 | if (R.isSelect()) { |
| 2083 | auto [Cond, Addrs] = SelAddr.getSelectCondAndAddrs(); |
| 2084 | MemVals.emplace_back( |
| 2085 | Args: ReachingMemVal::getSelect(BB, Cond, TrueAddr: Addrs.first, FalseAddr: Addrs.second)); |
| 2086 | continue; |
| 2087 | } |
| 2088 | Value *Address = SelAddr.getAddr(); |
| 2089 | if (R.isClobber()) |
| 2090 | MemVals.emplace_back(Args: ReachingMemVal::getClobber(Addr: Address, Inst)); |
| 2091 | else if (R.isDef()) |
| 2092 | MemVals.emplace_back(Args: ReachingMemVal::getDef(Addr: Address, Inst)); |
| 2093 | else |
| 2094 | MemVals.emplace_back(Args: ReachingMemVal::getUnknown(BB, Addr: Address, Inst)); |
| 2095 | } |
| 2096 | |
| 2097 | return processNonLocalLoad(L: Load, Deps&: MemVals); |
| 2098 | } |
| 2099 | |
| 2100 | bool GVNPass::processNonLocalLoad(LoadInst *Load, |
| 2101 | SmallVectorImpl<ReachingMemVal> &Deps) { |
| 2102 | // If we had a phi translation failure, we'll have a single entry which is a |
| 2103 | // clobber in the current block. Reject this early. |
| 2104 | if (Deps.size() == 1 && Deps[0].Kind == DepKind::Other) { |
| 2105 | LLVM_DEBUG(dbgs() << "GVN: non-local load " ; Load->printAsOperand(dbgs()); |
| 2106 | dbgs() << " has unknown dependencies\n" ;); |
| 2107 | return false; |
| 2108 | } |
| 2109 | |
| 2110 | bool Changed = false; |
| 2111 | // This is a limited form of scalar PRE for load indices. If this load follows |
| 2112 | // a GEP, see if we can PRE the indices before analyzing. |
| 2113 | if (isScalarPREEnabled()) { |
| 2114 | if (GetElementPtrInst *GEP = |
| 2115 | dyn_cast<GetElementPtrInst>(Val: Load->getOperand(i_nocapture: 0))) { |
| 2116 | for (Use &U : GEP->indices()) |
| 2117 | if (Instruction *I = dyn_cast<Instruction>(Val: U.get())) |
| 2118 | Changed |= performScalarPRE(I); |
| 2119 | } |
| 2120 | } |
| 2121 | |
| 2122 | // Step 1: Analyze the availability of the load. |
| 2123 | AvailValInBlkVect ValuesPerBlock; |
| 2124 | UnavailBlkVect UnavailableBlocks; |
| 2125 | AnalyzeLoadAvailability(Load, Deps, ValuesPerBlock, UnavailableBlocks); |
| 2126 | |
| 2127 | // If we have no predecessors that produce a known value for this load, exit |
| 2128 | // early. |
| 2129 | if (ValuesPerBlock.empty()) |
| 2130 | return Changed; |
| 2131 | |
| 2132 | // Step 2: Eliminate fully redundancy. |
| 2133 | // |
| 2134 | // If all of the instructions we depend on produce a known value for this |
| 2135 | // load, then it is fully redundant and we can use PHI insertion to compute |
| 2136 | // its value. Insert PHIs and remove the fully redundant value now. |
| 2137 | if (UnavailableBlocks.empty()) { |
| 2138 | LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *Load << '\n'); |
| 2139 | |
| 2140 | // Perform PHI construction. |
| 2141 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, GVN&: *this); |
| 2142 | // ConstructSSAForLoadSet is responsible for combining metadata. |
| 2143 | ICF->removeUsersOf(Inst: Load); |
| 2144 | Load->replaceAllUsesWith(V); |
| 2145 | |
| 2146 | if (isa<PHINode>(Val: V)) |
| 2147 | V->takeName(V: Load); |
| 2148 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) |
| 2149 | // If instruction I has debug info, then we should not update it. |
| 2150 | // Also, if I has a null DebugLoc, then it is still potentially incorrect |
| 2151 | // to propagate Load's DebugLoc because Load may not post-dominate I. |
| 2152 | if (Load->getDebugLoc() && Load->getParent() == I->getParent()) |
| 2153 | I->setDebugLoc(Load->getDebugLoc()); |
| 2154 | if (MD && V->getType()->isPtrOrPtrVectorTy()) |
| 2155 | MD->invalidateCachedPointerInfo(Ptr: V); |
| 2156 | ++NumGVNLoad; |
| 2157 | reportLoadElim(Load, AvailableValue: V, ORE); |
| 2158 | salvageAndRemoveInstruction(I: Load); |
| 2159 | return true; |
| 2160 | } |
| 2161 | |
| 2162 | // Step 3: Eliminate partial redundancy. |
| 2163 | if (!isLoadPREEnabled()) |
| 2164 | return Changed; |
| 2165 | if (!isLoadInLoopPREEnabled() && LI->getLoopFor(BB: Load->getParent())) |
| 2166 | return Changed; |
| 2167 | |
| 2168 | if (performLoopLoadPRE(Load, ValuesPerBlock, UnavailableBlocks) || |
| 2169 | PerformLoadPRE(Load, ValuesPerBlock, UnavailableBlocks)) |
| 2170 | return true; |
| 2171 | |
| 2172 | return Changed; |
| 2173 | } |
| 2174 | |
| 2175 | bool GVNPass::processAssumeIntrinsic(AssumeInst *IntrinsicI) { |
| 2176 | Value *V = IntrinsicI->getArgOperand(i: 0); |
| 2177 | |
| 2178 | if (ConstantInt *Cond = dyn_cast<ConstantInt>(Val: V)) { |
| 2179 | if (Cond->isZero()) { |
| 2180 | Type *Int8Ty = Type::getInt8Ty(C&: V->getContext()); |
| 2181 | Type *PtrTy = PointerType::get(C&: V->getContext(), AddressSpace: 0); |
| 2182 | // Insert a new store to null instruction before the load to indicate that |
| 2183 | // this code is not reachable. FIXME: We could insert unreachable |
| 2184 | // instruction directly because we can modify the CFG. |
| 2185 | auto *NewS = |
| 2186 | new StoreInst(PoisonValue::get(T: Int8Ty), Constant::getNullValue(Ty: PtrTy), |
| 2187 | IntrinsicI->getIterator()); |
| 2188 | if (MSSAU) { |
| 2189 | const MemoryUseOrDef *FirstNonDom = nullptr; |
| 2190 | const auto *AL = |
| 2191 | MSSAU->getMemorySSA()->getBlockAccesses(BB: IntrinsicI->getParent()); |
| 2192 | |
| 2193 | // If there are accesses in the current basic block, find the first one |
| 2194 | // that does not come before NewS. The new memory access is inserted |
| 2195 | // after the found access or before the terminator if no such access is |
| 2196 | // found. |
| 2197 | if (AL) { |
| 2198 | for (const auto &Acc : *AL) { |
| 2199 | if (auto *Current = dyn_cast<MemoryUseOrDef>(Val: &Acc)) |
| 2200 | if (!Current->getMemoryInst()->comesBefore(Other: NewS)) { |
| 2201 | FirstNonDom = Current; |
| 2202 | break; |
| 2203 | } |
| 2204 | } |
| 2205 | } |
| 2206 | |
| 2207 | auto *NewDef = |
| 2208 | FirstNonDom ? MSSAU->createMemoryAccessBefore( |
| 2209 | I: NewS, Definition: nullptr, |
| 2210 | InsertPt: const_cast<MemoryUseOrDef *>(FirstNonDom)) |
| 2211 | : MSSAU->createMemoryAccessInBB( |
| 2212 | I: NewS, Definition: nullptr, |
| 2213 | BB: NewS->getParent(), Point: MemorySSA::BeforeTerminator); |
| 2214 | |
| 2215 | MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewDef), /*RenameUses=*/false); |
| 2216 | } |
| 2217 | } |
| 2218 | if (isAssumeWithEmptyBundle(Assume: *IntrinsicI)) { |
| 2219 | salvageAndRemoveInstruction(I: IntrinsicI); |
| 2220 | return true; |
| 2221 | } |
| 2222 | return false; |
| 2223 | } |
| 2224 | |
| 2225 | if (isa<Constant>(Val: V)) { |
| 2226 | // If it's not false, and constant, it must evaluate to true. This means our |
| 2227 | // assume is assume(true), and thus, pointless, and we don't want to do |
| 2228 | // anything more here. |
| 2229 | return false; |
| 2230 | } |
| 2231 | |
| 2232 | Constant *True = ConstantInt::getTrue(Context&: V->getContext()); |
| 2233 | return propagateEquality(LHS: V, RHS: True, Root: IntrinsicI); |
| 2234 | } |
| 2235 | |
| 2236 | static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { |
| 2237 | patchReplacementInstruction(I, Repl); |
| 2238 | I->replaceAllUsesWith(V: Repl); |
| 2239 | } |
| 2240 | |
| 2241 | /// If a load has !invariant.group, try to find the most-dominating instruction |
| 2242 | /// with the same metadata and equivalent pointer (modulo bitcasts and zero |
| 2243 | /// GEPs). If one is found that dominates the load, its value can be reused. |
| 2244 | static Instruction *findInvariantGroupValue(LoadInst *L, DominatorTree &DT) { |
| 2245 | Value *PointerOperand = L->getPointerOperand()->stripPointerCasts(); |
| 2246 | |
| 2247 | // It's not safe to walk the use list of a global value because function |
| 2248 | // passes aren't allowed to look outside their functions. |
| 2249 | // FIXME: this could be fixed by filtering instructions from outside of |
| 2250 | // current function. |
| 2251 | if (isa<Constant>(Val: PointerOperand)) |
| 2252 | return nullptr; |
| 2253 | |
| 2254 | // Queue to process all pointers that are equivalent to load operand. |
| 2255 | SmallVector<Value *, 8> PointerUsesQueue; |
| 2256 | PointerUsesQueue.push_back(Elt: PointerOperand); |
| 2257 | |
| 2258 | Instruction *MostDominatingInstruction = L; |
| 2259 | |
| 2260 | // FIXME: This loop is potentially O(n^2) due to repeated dominates checks. |
| 2261 | while (!PointerUsesQueue.empty()) { |
| 2262 | Value *Ptr = PointerUsesQueue.pop_back_val(); |
| 2263 | assert(Ptr && !isa<GlobalValue>(Ptr) && |
| 2264 | "Null or GlobalValue should not be inserted" ); |
| 2265 | |
| 2266 | for (User *U : Ptr->users()) { |
| 2267 | auto *I = dyn_cast<Instruction>(Val: U); |
| 2268 | if (!I || I == L || !DT.dominates(Def: I, User: MostDominatingInstruction)) |
| 2269 | continue; |
| 2270 | |
| 2271 | // Add bitcasts and zero GEPs to queue. |
| 2272 | // TODO: Should drop bitcast? |
| 2273 | if (isa<BitCastInst>(Val: I) || |
| 2274 | (isa<GetElementPtrInst>(Val: I) && |
| 2275 | cast<GetElementPtrInst>(Val: I)->hasAllZeroIndices())) { |
| 2276 | PointerUsesQueue.push_back(Elt: I); |
| 2277 | continue; |
| 2278 | } |
| 2279 | |
| 2280 | // If we hit a load/store with an invariant.group metadata and the same |
| 2281 | // pointer operand, we can assume that value pointed to by the pointer |
| 2282 | // operand didn't change. |
| 2283 | if (I->hasMetadata(KindID: LLVMContext::MD_invariant_group) && |
| 2284 | Ptr == getLoadStorePointerOperand(V: I) && !I->isVolatile()) |
| 2285 | MostDominatingInstruction = I; |
| 2286 | } |
| 2287 | } |
| 2288 | |
| 2289 | return MostDominatingInstruction != L ? MostDominatingInstruction : nullptr; |
| 2290 | } |
| 2291 | |
| 2292 | /// Return the memory location accessed by the (masked) load/store instruction |
| 2293 | /// `I`, if the instruction could potentially provide a useful value for |
| 2294 | /// eliminating the load. |
| 2295 | static std::optional<MemoryLocation> |
| 2296 | maybeLoadStoreLocation(Instruction *I, bool AllowStores, |
| 2297 | const TargetLibraryInfo *TLI) { |
| 2298 | if (auto *LI = dyn_cast<LoadInst>(Val: I)) |
| 2299 | return MemoryLocation::get(LI); |
| 2300 | |
| 2301 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) { |
| 2302 | switch (II->getIntrinsicID()) { |
| 2303 | case Intrinsic::masked_load: |
| 2304 | return MemoryLocation::getForArgument(Call: II, ArgIdx: 0, TLI); |
| 2305 | case Intrinsic::masked_store: |
| 2306 | if (AllowStores) |
| 2307 | return MemoryLocation::getForArgument(Call: II, ArgIdx: 1, TLI); |
| 2308 | return std::nullopt; |
| 2309 | default: |
| 2310 | break; |
| 2311 | } |
| 2312 | } |
| 2313 | |
| 2314 | if (!AllowStores) |
| 2315 | return std::nullopt; |
| 2316 | |
| 2317 | if (auto *SI = dyn_cast<StoreInst>(Val: I)) |
| 2318 | return MemoryLocation::get(SI); |
| 2319 | return std::nullopt; |
| 2320 | } |
| 2321 | |
| 2322 | /// Scan the users of each MemoryAccess in `ClobbersList` that belong to `BB`, |
| 2323 | /// looking for memory reads whose location aliases `Loc` and dominates our |
| 2324 | /// load. |
| 2325 | std::optional<GVNPass::ReachingMemVal> GVNPass::scanMemoryAccessesUsers( |
| 2326 | const MemoryLocation &Loc, bool IsInvariantLoad, BasicBlock *BB, |
| 2327 | const SmallVectorImpl<MemoryAccess *> &ClobbersList, MemorySSA &MSSA, |
| 2328 | BatchAAResults &AA, LoadInst *L) { |
| 2329 | |
| 2330 | // Prefer a candidate that is closer to the load within the same block. |
| 2331 | auto UpdateChoice = [&](std::optional<ReachingMemVal> &Choice, |
| 2332 | AliasResult &AR, Instruction *Candidate) { |
| 2333 | if (!Choice) { |
| 2334 | if (AR == AliasResult::PartialAlias) |
| 2335 | Choice = ReachingMemVal::getClobber(Addr: Loc.Ptr, Inst: Candidate, Offset: AR.getOffset()); |
| 2336 | else |
| 2337 | Choice = ReachingMemVal::getDef(Addr: Loc.Ptr, Inst: Candidate); |
| 2338 | return; |
| 2339 | } |
| 2340 | if (!MSSA.locallyDominates(A: MSSA.getMemoryAccess(I: Choice->Inst), |
| 2341 | B: MSSA.getMemoryAccess(I: Candidate))) |
| 2342 | return; |
| 2343 | |
| 2344 | if (AR == AliasResult::PartialAlias) { |
| 2345 | Choice->Kind = DepKind::Clobber; |
| 2346 | Choice->Offset = AR.getOffset(); |
| 2347 | } else { |
| 2348 | Choice->Kind = DepKind::Def; |
| 2349 | Choice->Offset = -1; |
| 2350 | } |
| 2351 | |
| 2352 | Choice->Inst = Candidate; |
| 2353 | Choice->Block = Candidate->getParent(); |
| 2354 | }; |
| 2355 | |
| 2356 | std::optional<ReachingMemVal> ReachingVal; |
| 2357 | for (MemoryAccess *MA : ClobbersList) { |
| 2358 | unsigned Scanned = 0; |
| 2359 | for (User *U : MA->users()) { |
| 2360 | if (++Scanned >= ScanUsersLimit) |
| 2361 | return ReachingMemVal::getUnknown(BB, Addr: Loc.Ptr); |
| 2362 | |
| 2363 | auto *UseOrDef = dyn_cast<MemoryUseOrDef>(Val: U); |
| 2364 | if (!UseOrDef || UseOrDef->getBlock() != BB) |
| 2365 | continue; |
| 2366 | |
| 2367 | Instruction *MemI = UseOrDef->getMemoryInst(); |
| 2368 | if (MemI == L || |
| 2369 | (L && !MSSA.locallyDominates(A: UseOrDef, B: MSSA.getMemoryAccess(I: L)))) |
| 2370 | continue; |
| 2371 | |
| 2372 | if (auto MaybeLoc = maybeLoadStoreLocation(I: MemI, AllowStores: IsInvariantLoad, TLI)) { |
| 2373 | AliasResult AR = AA.alias(LocA: *MaybeLoc, LocB: Loc); |
| 2374 | // If the locations do not certainly alias, we cannot possibly infer the |
| 2375 | // following load loads the same value. |
| 2376 | if (AR == AliasResult::NoAlias || AR == AliasResult::MayAlias) |
| 2377 | continue; |
| 2378 | |
| 2379 | // Locations partially overlap, but neither is a subset of the other, or |
| 2380 | // the second location is before the first. |
| 2381 | if (AR == AliasResult::PartialAlias && |
| 2382 | (!AR.hasOffset() || AR.getOffset() < 0)) |
| 2383 | continue; |
| 2384 | |
| 2385 | // Found candidate, the new load memory location and the given location |
| 2386 | // must alias: precise overlap, or subset with non-negative offset. |
| 2387 | UpdateChoice(ReachingVal, AR, MemI); |
| 2388 | } |
| 2389 | } |
| 2390 | if (ReachingVal) |
| 2391 | break; |
| 2392 | } |
| 2393 | |
| 2394 | return ReachingVal; |
| 2395 | } |
| 2396 | |
| 2397 | /// Check if a given MemoryAccess (usually a MemoryDef) actually modifies a |
| 2398 | /// given location. Returns a ReachingMemVal describing the dependency. |
| 2399 | std::optional<GVNPass::ReachingMemVal> GVNPass::accessMayModifyLocation( |
| 2400 | MemoryAccess *ClobberMA, const MemoryLocation &Loc, bool IsInvariantLoad, |
| 2401 | BasicBlock *BB, MemorySSA &MSSA, BatchAAResults &AA) { |
| 2402 | assert(ClobberMA->getBlock() == BB); |
| 2403 | |
| 2404 | // If the clobbering access is the entry memory state, we cannot say anything |
| 2405 | // about the content of the memory, except when we are accessing a local |
| 2406 | // object, which can be turned later into producing `undef`. |
| 2407 | if (MSSA.isLiveOnEntryDef(MA: ClobberMA)) { |
| 2408 | if (auto *Alloc = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Loc.Ptr))) |
| 2409 | if (Alloc->getParent() == BB) |
| 2410 | return ReachingMemVal::getDef(Addr: Loc.Ptr, Inst: const_cast<AllocaInst *>(Alloc)); |
| 2411 | return ReachingMemVal::getUnknown(BB, Addr: Loc.Ptr); |
| 2412 | } |
| 2413 | |
| 2414 | // Loads from "constant" memory can't be clobbered. |
| 2415 | if (IsInvariantLoad || AA.pointsToConstantMemory(Loc)) |
| 2416 | return std::nullopt; |
| 2417 | |
| 2418 | auto GetOrdering = [](const Instruction *I) { |
| 2419 | if (auto *L = dyn_cast<LoadInst>(Val: I)) |
| 2420 | return L->getOrdering(); |
| 2421 | return cast<StoreInst>(Val: I)->getOrdering(); |
| 2422 | }; |
| 2423 | Instruction *ClobberI = cast<MemoryDef>(Val: ClobberMA)->getMemoryInst(); |
| 2424 | |
| 2425 | // Check if the clobbering access is a load or a store that we can reuse. |
| 2426 | if (auto MaybeLoc = maybeLoadStoreLocation(I: ClobberI, AllowStores: true, TLI)) { |
| 2427 | AliasResult AR = AA.alias(LocA: *MaybeLoc, LocB: Loc); |
| 2428 | if (AR == AliasResult::MustAlias) |
| 2429 | return ReachingMemVal::getDef(Addr: Loc.Ptr, Inst: ClobberI); |
| 2430 | |
| 2431 | if (AR == AliasResult::NoAlias) { |
| 2432 | // If the locations do not alias we may still be able to skip over the |
| 2433 | // clobbering instruction, even if it is atomic. |
| 2434 | // The original load is either non-atomic or unordered. We can reorder |
| 2435 | // these across non-atomic, unordered or monotonic loads or across any |
| 2436 | // store. |
| 2437 | if (!ClobberI->isAtomic() || |
| 2438 | !isStrongerThan(AO: GetOrdering(ClobberI), Other: AtomicOrdering::Monotonic) || |
| 2439 | isa<StoreInst>(Val: ClobberI)) |
| 2440 | return std::nullopt; |
| 2441 | return ReachingMemVal::getClobber(Addr: Loc.Ptr, Inst: ClobberI); |
| 2442 | } |
| 2443 | |
| 2444 | // Skip over volatile loads (the original load is non-volatile, non-atomic). |
| 2445 | if (!ClobberI->isAtomic() && isa<LoadInst>(Val: ClobberI)) |
| 2446 | return std::nullopt; |
| 2447 | |
| 2448 | if (AR == AliasResult::MayAlias || |
| 2449 | (AR == AliasResult::PartialAlias && |
| 2450 | (!AR.hasOffset() || AR.getOffset() < 0))) |
| 2451 | return ReachingMemVal::getClobber(Addr: Loc.Ptr, Inst: ClobberI); |
| 2452 | |
| 2453 | // The only option left is a store of the superset of the required bits. |
| 2454 | assert(AR == AliasResult::PartialAlias && AR.hasOffset() && |
| 2455 | AR.getOffset() > 0 && |
| 2456 | "Must be the superset/partial overlap case with positive offset" ); |
| 2457 | return ReachingMemVal::getClobber(Addr: Loc.Ptr, Inst: ClobberI, Offset: AR.getOffset()); |
| 2458 | } |
| 2459 | |
| 2460 | if (auto *II = dyn_cast<IntrinsicInst>(Val: ClobberI)) { |
| 2461 | if (isa<DbgInfoIntrinsic>(Val: II)) |
| 2462 | return std::nullopt; |
| 2463 | if (II->getIntrinsicID() == Intrinsic::lifetime_start) { |
| 2464 | MemoryLocation IIObjLoc = MemoryLocation::getForArgument(Call: II, ArgIdx: 0, TLI); |
| 2465 | if (AA.isMustAlias(LocA: IIObjLoc, LocB: Loc)) |
| 2466 | return ReachingMemVal::getDef(Addr: Loc.Ptr, Inst: ClobberI); |
| 2467 | return std::nullopt; |
| 2468 | } |
| 2469 | } |
| 2470 | |
| 2471 | // If we are at a malloc-like function call, we can turn the load into `undef` |
| 2472 | // or zero. |
| 2473 | if (isNoAliasCall(V: ClobberI)) { |
| 2474 | const Value *Obj = getUnderlyingObject(V: Loc.Ptr); |
| 2475 | if (Obj == ClobberI || AA.isMustAlias(V1: ClobberI, V2: Loc.Ptr)) |
| 2476 | return ReachingMemVal::getDef(Addr: Loc.Ptr, Inst: ClobberI); |
| 2477 | } |
| 2478 | |
| 2479 | // Can reorder loads across a release fence. |
| 2480 | if (auto *FI = dyn_cast<FenceInst>(Val: ClobberI)) |
| 2481 | if (FI->getOrdering() == AtomicOrdering::Release) |
| 2482 | return std::nullopt; |
| 2483 | |
| 2484 | // See if the clobber instruction (e.g., a generic call) may modify the |
| 2485 | // location. |
| 2486 | ModRefInfo MR = AA.getModRefInfo(I: ClobberI, OptLoc: Loc); |
| 2487 | // If may modify the location, analyze deeper, to exclude accesses to |
| 2488 | // non-escaping local allocations. |
| 2489 | if (MR == ModRefInfo::NoModRef || MR == ModRefInfo::Ref) |
| 2490 | return std::nullopt; |
| 2491 | |
| 2492 | // Conservatively assume the clobbering memory access may overwrite the |
| 2493 | // location. |
| 2494 | return ReachingMemVal::getClobber(Addr: Loc.Ptr, Inst: ClobberI); |
| 2495 | } |
| 2496 | |
| 2497 | /// Collect the predecessors of block, while doing phi-translation of the memory |
| 2498 | /// address and the memory clobber. Return false if the block should be marked |
| 2499 | /// as clobbering the memory location in an unknown way. |
| 2500 | bool GVNPass::collectPredecessors(BasicBlock *BB, const PHITransAddr &Addr, |
| 2501 | MemoryAccess *ClobberMA, |
| 2502 | DependencyBlockSet &Blocks, |
| 2503 | SmallVectorImpl<BasicBlock *> &Worklist) { |
| 2504 | if (Addr.needsPHITranslationFromBlock(BB) && |
| 2505 | !Addr.isPotentiallyPHITranslatable()) |
| 2506 | return false; |
| 2507 | |
| 2508 | auto *MPhi = |
| 2509 | ClobberMA->getBlock() == BB ? dyn_cast<MemoryPhi>(Val: ClobberMA) : nullptr; |
| 2510 | SmallVector<std::pair<BasicBlock *, DependencyBlockInfo>, 8> Preds; |
| 2511 | for (BasicBlock *Pred : predecessors(BB)) { |
| 2512 | // Skip unreachable predecessors. |
| 2513 | if (!DT->isReachableFromEntry(A: Pred)) |
| 2514 | continue; |
| 2515 | |
| 2516 | // Skip already visited predecessors. |
| 2517 | if (llvm::any_of(Range&: Preds, P: [Pred](const auto &P) { return P.first == Pred; })) |
| 2518 | continue; |
| 2519 | |
| 2520 | PHITransAddr TransAddr = Addr; |
| 2521 | if (TransAddr.needsPHITranslationFromBlock(BB)) |
| 2522 | TransAddr.translateValue(CurBB: BB, PredBB: Pred, DT, MustDominate: false); |
| 2523 | |
| 2524 | auto It = Blocks.find(Val: Pred); |
| 2525 | if (It != Blocks.end()) { |
| 2526 | // If we reach a visited block with a different address, set the |
| 2527 | // current block as clobbering the memory location in an unknown way |
| 2528 | // (by returning false). |
| 2529 | if (It->second.Addr.getAddr() != TransAddr.getAddr()) |
| 2530 | return false; |
| 2531 | // Otherwise, just stop the traversal. |
| 2532 | continue; |
| 2533 | } |
| 2534 | |
| 2535 | Preds.emplace_back( |
| 2536 | Args&: Pred, Args: DependencyBlockInfo(TransAddr, |
| 2537 | MPhi ? MPhi->getIncomingValueForBlock(BB: Pred) |
| 2538 | : ClobberMA)); |
| 2539 | } |
| 2540 | |
| 2541 | // We collected the predecessors and stored them in Preds. Now, populate the |
| 2542 | // worklist with the predecessors found, and cache the eventual translated |
| 2543 | // address for each block. |
| 2544 | for (auto &P : Preds) { |
| 2545 | [[maybe_unused]] auto It = |
| 2546 | Blocks.try_emplace(Key: P.first, Args: std::move(P.second)).first; |
| 2547 | Worklist.push_back(Elt: P.first); |
| 2548 | } |
| 2549 | |
| 2550 | return true; |
| 2551 | } |
| 2552 | |
| 2553 | /// Build a list of MemoryAccesses whose users could potentially alias the |
| 2554 | /// memory location being queried. Starts from StartInfo's initial clobber, |
| 2555 | /// walk the use-def chain to the final clobber. If the chain extends beyond |
| 2556 | /// `BB`, continue into that block but only if it is in the previously collected |
| 2557 | /// set. |
| 2558 | void GVNPass::collectClobberList(SmallVectorImpl<MemoryAccess *> &Clobbers, |
| 2559 | BasicBlock *BB, |
| 2560 | const DependencyBlockInfo &StartInfo, |
| 2561 | const DependencyBlockSet &Blocks, |
| 2562 | MemorySSA &MSSA) { |
| 2563 | MemoryAccess *MA = StartInfo.InitialClobberMA; |
| 2564 | MemoryAccess *LastMA = StartInfo.ClobberMA; |
| 2565 | |
| 2566 | for (;;) { |
| 2567 | while (MA != LastMA) { |
| 2568 | Clobbers.push_back(Elt: MA); |
| 2569 | MA = cast<MemoryUseOrDef>(Val: MA)->getDefiningAccess(); |
| 2570 | } |
| 2571 | Clobbers.push_back(Elt: MA); |
| 2572 | |
| 2573 | if (MSSA.isLiveOnEntryDef(MA) || |
| 2574 | (MA->getBlock() == BB && !isa<MemoryPhi>(Val: MA))) |
| 2575 | break; |
| 2576 | |
| 2577 | // If the final clobber in the current block is a MemoryPhi, go to the |
| 2578 | // immediate dominator; otherwise, just get to the block containing the |
| 2579 | // final clobber. |
| 2580 | if (MA->getBlock() == BB) |
| 2581 | BB = DT->getNode(BB)->getIDom()->getBlock(); |
| 2582 | else |
| 2583 | BB = MA->getBlock(); |
| 2584 | |
| 2585 | auto It = Blocks.find(Val: BB); |
| 2586 | if (It == Blocks.end()) |
| 2587 | break; |
| 2588 | |
| 2589 | MA = It->second.InitialClobberMA; |
| 2590 | LastMA = It->second.ClobberMA; |
| 2591 | if (MA == Clobbers.back()) |
| 2592 | Clobbers.pop_back(); |
| 2593 | } |
| 2594 | } |
| 2595 | |
| 2596 | /// Entrypoint for the MemorySSA-based redundant load elimination algorithm. |
| 2597 | /// Given as input a load instruction, the function computes the set of reaching |
| 2598 | /// memory values, one per predecessor path, that AnalyzeLoadAvailability can |
| 2599 | /// later use to establish whether the load may be eliminated. A reaching value |
| 2600 | /// may be of the following descriptor kind: |
| 2601 | /// * Def: a precise instruction that produces the exact bits the load would |
| 2602 | /// read (e.g., an equivalent load or a MustAlias store); |
| 2603 | /// * Clobber: a write that clobbers a superset of the bits the load would read |
| 2604 | /// (e.g., a memset over a larger region); |
| 2605 | /// * Other: we know which block defines the memory location in some way, but |
| 2606 | /// could not identify a precise instruction (e.g., memory already live at |
| 2607 | /// function entry). |
| 2608 | bool GVNPass::findReachingValuesForLoad(LoadInst *L, |
| 2609 | SmallVectorImpl<ReachingMemVal> &Values, |
| 2610 | MemorySSA &MSSA, AAResults &AAR) { |
| 2611 | EarliestEscapeAnalysis EA(*DT, LI); |
| 2612 | BatchAAResults AA(AAR, &EA); |
| 2613 | BasicBlock *StartBlock = L->getParent(); |
| 2614 | bool IsInvariantLoad = L->hasMetadata(KindID: LLVMContext::MD_invariant_load); |
| 2615 | // TODO: Simplify later work by just getClobberingMemoryAccess(). |
| 2616 | MemoryAccess *ClobberMA = MSSA.getMemoryAccess(I: L)->getDefiningAccess(); |
| 2617 | const MemoryLocation Loc = MemoryLocation::get(LI: L); |
| 2618 | |
| 2619 | // Fast path for load tagged with !invariant.group. |
| 2620 | if (L->hasMetadata(KindID: LLVMContext::MD_invariant_group)) { |
| 2621 | if (Instruction *G = findInvariantGroupValue(L, DT&: *DT)) { |
| 2622 | Values.emplace_back( |
| 2623 | Args: ReachingMemVal::getDef(Addr: getLoadStorePointerOperand(V: G), Inst: G)); |
| 2624 | return true; |
| 2625 | } |
| 2626 | } |
| 2627 | |
| 2628 | // Phase 1. First off, look for a local dependency to avoid having to |
| 2629 | // disambiguate between before the load and after the load of the starting |
| 2630 | // block (as the load may be visited from a backedge). |
| 2631 | do { |
| 2632 | // Scan users of the clobbering memory access. |
| 2633 | if (auto RMV = scanMemoryAccessesUsers( |
| 2634 | Loc, IsInvariantLoad, BB: StartBlock, |
| 2635 | ClobbersList: SmallVector<MemoryAccess *, 1>{ClobberMA}, MSSA, AA, L)) { |
| 2636 | Values.emplace_back(Args&: *RMV); |
| 2637 | return true; |
| 2638 | } |
| 2639 | |
| 2640 | // Exit from here, and proceed visiting predecessors if the clobbering |
| 2641 | // access is non-local or is a MemoryPhi. |
| 2642 | if (ClobberMA->getBlock() != StartBlock || isa<MemoryPhi>(Val: ClobberMA)) |
| 2643 | break; |
| 2644 | |
| 2645 | // Check if the clobber actually aliases the load location. |
| 2646 | if (auto RMV = accessMayModifyLocation(ClobberMA, Loc, IsInvariantLoad, |
| 2647 | BB: StartBlock, MSSA, AA)) { |
| 2648 | Values.emplace_back(Args&: *RMV); |
| 2649 | return true; |
| 2650 | } |
| 2651 | |
| 2652 | // It may happen that the clobbering memory access does not actually |
| 2653 | // clobber our load location, transition to its defining memory access. |
| 2654 | ClobberMA = cast<MemoryUseOrDef>(Val: ClobberMA)->getDefiningAccess(); |
| 2655 | } while (ClobberMA->getBlock() == StartBlock); |
| 2656 | |
| 2657 | // Non-local speculations are not allowed under ASan. |
| 2658 | if (L->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeAddress) || |
| 2659 | L->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeHWAddress)) |
| 2660 | return false; |
| 2661 | |
| 2662 | // Phase 2. Walk backwards through the CFG, collecting all the blocks that |
| 2663 | // contain an instruction that modifies the load memory location, or that lie |
| 2664 | // on a path between a clobbering block and our load. Start off by collecting |
| 2665 | // the predecessors of `StartBlock`. All the visited blocks are stored in a |
| 2666 | // the set `Blocks`. If possible, the memory address maintained for the block |
| 2667 | // visited does get phi-translated. |
| 2668 | DependencyBlockSet Blocks; |
| 2669 | SmallVector<BasicBlock *, 16> InitialWorklist; |
| 2670 | const DataLayout &DL = L->getModule()->getDataLayout(); |
| 2671 | if (!collectPredecessors(BB: StartBlock, |
| 2672 | Addr: PHITransAddr(L->getPointerOperand(), DL, AC), |
| 2673 | ClobberMA, Blocks, Worklist&: InitialWorklist)) |
| 2674 | return false; |
| 2675 | |
| 2676 | // Do a bottom-up DFS. |
| 2677 | auto Worklist = InitialWorklist; |
| 2678 | while (!Worklist.empty()) { |
| 2679 | auto *BB = Worklist.pop_back_val(); |
| 2680 | DependencyBlockInfo &Info = Blocks.find(Val: BB)->second; |
| 2681 | |
| 2682 | // Phi-translation may have failed. |
| 2683 | if (!Info.Addr.getAddr()) |
| 2684 | continue; |
| 2685 | |
| 2686 | // If the clobbering memory access is in the current block and it indeed |
| 2687 | // clobbers our load location, record the dependency and do not visit the |
| 2688 | // predecessors of this block further, continue with the blocks in the |
| 2689 | // worklist. |
| 2690 | if (Info.ClobberMA->getBlock() == BB && !isa<MemoryPhi>(Val: Info.ClobberMA)) { |
| 2691 | if (auto RMV = accessMayModifyLocation( |
| 2692 | ClobberMA: Info.ClobberMA, Loc: Loc.getWithNewPtr(NewPtr: Info.Addr.getAddr()), |
| 2693 | IsInvariantLoad, BB, MSSA, AA)) { |
| 2694 | Info.MemVal = RMV; |
| 2695 | continue; |
| 2696 | } |
| 2697 | assert(!MSSA.isLiveOnEntryDef(Info.ClobberMA) && |
| 2698 | "LiveOnEntry aliases everything" ); |
| 2699 | |
| 2700 | // If, however, the clobbering memory access does not actually clobber |
| 2701 | // our load location, transition to its defining memory access, but |
| 2702 | // keep examining the same basic block. |
| 2703 | Info.ClobberMA = |
| 2704 | cast<MemoryUseOrDef>(Val: Info.ClobberMA)->getDefiningAccess(); |
| 2705 | Worklist.emplace_back(Args&: BB); |
| 2706 | continue; |
| 2707 | } |
| 2708 | |
| 2709 | // At this point we know the current block is "transparent", i.e. the memory |
| 2710 | // location is not modified when execution goes through this block. |
| 2711 | // Continue to its predecessors, unless a predecessor has already been |
| 2712 | // visited with a different address. We currently cannot represent such a |
| 2713 | // dependency. |
| 2714 | if (BB == StartBlock && Info.Addr.getAddr() != L->getPointerOperand()) { |
| 2715 | Info.ForceUnknown = true; |
| 2716 | continue; |
| 2717 | } |
| 2718 | if (BB != StartBlock && |
| 2719 | !collectPredecessors(BB, Addr: Info.Addr, ClobberMA: Info.ClobberMA, Blocks, Worklist)) |
| 2720 | Info.ForceUnknown = true; |
| 2721 | } |
| 2722 | |
| 2723 | // Phase 3. We have collected all the blocks that either write a value to the |
| 2724 | // memory location of the load, or there exists a path to the load, along |
| 2725 | // which the memory location is not modified. Perform a second DFS to find |
| 2726 | // load-to-load dependencies; namely, look at the dominating memory reads, |
| 2727 | // that alias our load. These are the MemoryUses that are users of the |
| 2728 | // MemoryDefs we previously identified. If no memory read is encountered, |
| 2729 | // either confirm the clobbering write found before or set to unknown. |
| 2730 | Worklist = InitialWorklist; |
| 2731 | for (BasicBlock *BB : Worklist) { |
| 2732 | DependencyBlockInfo &Info = Blocks.find(Val: BB)->second; |
| 2733 | Info.Visited = true; |
| 2734 | } |
| 2735 | |
| 2736 | SmallVector<MemoryAccess *> Clobbers; |
| 2737 | while (!Worklist.empty()) { |
| 2738 | auto *BB = Worklist.pop_back_val(); |
| 2739 | DependencyBlockInfo &Info = Blocks.find(Val: BB)->second; |
| 2740 | |
| 2741 | // If phi-translation failed, assume the memory location is modified in |
| 2742 | // unknown way. |
| 2743 | if (!Info.Addr.getAddr()) { |
| 2744 | Values.push_back(Elt: ReachingMemVal::getUnknown(BB, Addr: nullptr)); |
| 2745 | continue; |
| 2746 | } |
| 2747 | |
| 2748 | Clobbers.clear(); |
| 2749 | collectClobberList(Clobbers, BB, StartInfo: Info, Blocks, MSSA); |
| 2750 | if (auto RMV = |
| 2751 | scanMemoryAccessesUsers(Loc: Loc.getWithNewPtr(NewPtr: Info.Addr.getAddr()), |
| 2752 | IsInvariantLoad, BB, ClobbersList: Clobbers, MSSA, AA)) { |
| 2753 | Values.push_back(Elt: *RMV); |
| 2754 | continue; |
| 2755 | } |
| 2756 | |
| 2757 | // If no reusable memory use was found, and the current block is not |
| 2758 | // transparent, use the already established memory def. |
| 2759 | if (Info.MemVal) { |
| 2760 | Values.push_back(Elt: *Info.MemVal); |
| 2761 | continue; |
| 2762 | } |
| 2763 | |
| 2764 | if (Info.ForceUnknown) { |
| 2765 | Values.push_back(Elt: ReachingMemVal::getUnknown(BB, Addr: Info.Addr.getAddr())); |
| 2766 | continue; |
| 2767 | } |
| 2768 | |
| 2769 | // If the current block is transparent, continue to its predecessors. |
| 2770 | for (BasicBlock *Pred : predecessors(BB)) { |
| 2771 | auto It = Blocks.find(Val: Pred); |
| 2772 | if (It == Blocks.end()) |
| 2773 | continue; |
| 2774 | DependencyBlockInfo &PredInfo = It->second; |
| 2775 | if (PredInfo.Visited) |
| 2776 | continue; |
| 2777 | PredInfo.Visited = true; |
| 2778 | Worklist.push_back(Elt: Pred); |
| 2779 | } |
| 2780 | } |
| 2781 | |
| 2782 | return true; |
| 2783 | } |
| 2784 | |
| 2785 | /// Attempt to eliminate a load, first by eliminating it |
| 2786 | /// locally, and then attempting non-local elimination if that fails. |
| 2787 | bool GVNPass::processLoad(LoadInst *L) { |
| 2788 | if (!MD && !isMemorySSAEnabled()) |
| 2789 | return false; |
| 2790 | |
| 2791 | // This code hasn't been audited for ordered or volatile memory access. |
| 2792 | if (!L->isUnordered()) |
| 2793 | return false; |
| 2794 | |
| 2795 | if (L->getType()->isTokenLikeTy()) |
| 2796 | return false; |
| 2797 | |
| 2798 | if (L->use_empty()) { |
| 2799 | salvageAndRemoveInstruction(I: L); |
| 2800 | return true; |
| 2801 | } |
| 2802 | |
| 2803 | ReachingMemVal MemVal = ReachingMemVal::getUnknown(BB: nullptr, Addr: nullptr); |
| 2804 | if (!isMemorySSAEnabled()) { |
| 2805 | // ... to a pointer that has been loaded from before... |
| 2806 | MemDepResult Dep = MD->getDependency(QueryInst: L); |
| 2807 | |
| 2808 | // If it is defined in another block, try harder. |
| 2809 | if (Dep.isNonLocal()) |
| 2810 | return processNonLocalLoad(Load: L); |
| 2811 | |
| 2812 | // Only handle the local case below. |
| 2813 | if (Dep.isDef()) |
| 2814 | MemVal = ReachingMemVal::getDef(Addr: L->getPointerOperand(), Inst: Dep.getInst()); |
| 2815 | else if (Dep.isClobber()) |
| 2816 | MemVal = |
| 2817 | ReachingMemVal::getClobber(Addr: L->getPointerOperand(), Inst: Dep.getInst()); |
| 2818 | } else { |
| 2819 | SmallVector<ReachingMemVal, 8> MemVals; |
| 2820 | if (!findReachingValuesForLoad(L, Values&: MemVals, MSSA&: *MSSAU->getMemorySSA(), AAR&: *AA)) |
| 2821 | return false; // Too many dependencies. |
| 2822 | assert(MemVals.size() && "Expected at least an unknown value" ); |
| 2823 | if (MemVals.size() > 1 || MemVals[0].Block != L->getParent()) |
| 2824 | return processNonLocalLoad(Load: L, Deps&: MemVals); |
| 2825 | |
| 2826 | MemVal = MemVals[0]; |
| 2827 | } |
| 2828 | |
| 2829 | if (MemVal.Kind == DepKind::Other) { |
| 2830 | // This might be a NonFuncLocal or an Unknown. |
| 2831 | LLVM_DEBUG( |
| 2832 | // fast print dep, using operator<< on instruction is too slow. |
| 2833 | dbgs() << "GVN: load " ; L->printAsOperand(dbgs()); |
| 2834 | dbgs() << " has unknown dependence\n" ;); |
| 2835 | return false; |
| 2836 | } |
| 2837 | |
| 2838 | auto AV = AnalyzeLoadAvailability(Load: L, Dep: MemVal, Address: L->getPointerOperand()); |
| 2839 | if (!AV) |
| 2840 | return false; |
| 2841 | |
| 2842 | Value *AvailableValue = AV->MaterializeAdjustedValue(Load: L, InsertPt: L); |
| 2843 | |
| 2844 | // MaterializeAdjustedValue is responsible for combining metadata. |
| 2845 | ICF->removeUsersOf(Inst: L); |
| 2846 | L->replaceAllUsesWith(V: AvailableValue); |
| 2847 | if (MSSAU) |
| 2848 | MSSAU->removeMemoryAccess(I: L); |
| 2849 | ++NumGVNLoad; |
| 2850 | reportLoadElim(Load: L, AvailableValue, ORE); |
| 2851 | salvageAndRemoveInstruction(I: L); |
| 2852 | // Tell MDA to reexamine the reused pointer since we might have more |
| 2853 | // information after forwarding it. |
| 2854 | if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy()) |
| 2855 | MD->invalidateCachedPointerInfo(Ptr: AvailableValue); |
| 2856 | return true; |
| 2857 | } |
| 2858 | |
| 2859 | // Attempt to process masked loads which have loaded from |
| 2860 | // masked stores with the same mask |
| 2861 | bool GVNPass::processMaskedLoad(IntrinsicInst *I) { |
| 2862 | if (!MD) |
| 2863 | return false; |
| 2864 | MemDepResult Dep = MD->getDependency(QueryInst: I); |
| 2865 | Instruction *DepInst = Dep.getInst(); |
| 2866 | if (!DepInst || !Dep.isLocal() || !Dep.isDef()) |
| 2867 | return false; |
| 2868 | |
| 2869 | Value *Mask = I->getOperand(i_nocapture: 1); |
| 2870 | Value *Passthrough = I->getOperand(i_nocapture: 2); |
| 2871 | Value *StoreVal; |
| 2872 | if (!match(V: DepInst, |
| 2873 | P: m_MaskedStore(Op0: m_Value(V&: StoreVal), Op1: m_Value(), Op2: m_Specific(V: Mask))) || |
| 2874 | StoreVal->getType() != I->getType()) |
| 2875 | return false; |
| 2876 | |
| 2877 | // Remove the load but generate a select for the passthrough |
| 2878 | Value *OpToForward = llvm::SelectInst::Create(C: Mask, S1: StoreVal, S2: Passthrough, NameStr: "" , |
| 2879 | InsertBefore: I->getIterator()); |
| 2880 | |
| 2881 | ICF->removeUsersOf(Inst: I); |
| 2882 | I->replaceAllUsesWith(V: OpToForward); |
| 2883 | salvageAndRemoveInstruction(I); |
| 2884 | ++NumGVNLoad; |
| 2885 | return true; |
| 2886 | } |
| 2887 | |
| 2888 | /// Return a pair the first field showing the value number of \p Exp and the |
| 2889 | /// second field showing whether it is a value number newly created. |
| 2890 | std::pair<uint32_t, bool> |
| 2891 | GVNPass::ValueTable::assignExpNewValueNum(Expression &Exp) { |
| 2892 | uint32_t &E = ExpressionNumbering[Exp]; |
| 2893 | bool CreateNewValNum = !E; |
| 2894 | if (CreateNewValNum) { |
| 2895 | Expressions.push_back(x: Exp); |
| 2896 | if (ExprIdx.size() < NextValueNumber + 1) |
| 2897 | ExprIdx.resize(new_size: NextValueNumber * 2); |
| 2898 | E = NextValueNumber; |
| 2899 | ExprIdx[NextValueNumber++] = NextExprNumber++; |
| 2900 | } |
| 2901 | return {E, CreateNewValNum}; |
| 2902 | } |
| 2903 | |
| 2904 | /// Return whether all the values related with the same \p num are |
| 2905 | /// defined in \p BB. |
| 2906 | bool GVNPass::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB, |
| 2907 | GVNPass &GVN) { |
| 2908 | return all_of( |
| 2909 | Range: GVN.LeaderTable.getLeaders(N: Num), |
| 2910 | P: [=](const LeaderMap::LeaderTableEntry &L) { return L.BB == BB; }); |
| 2911 | } |
| 2912 | |
| 2913 | /// Wrap phiTranslateImpl to provide caching functionality. |
| 2914 | uint32_t GVNPass::ValueTable::phiTranslate(const BasicBlock *Pred, |
| 2915 | const BasicBlock *PhiBlock, |
| 2916 | uint32_t Num, GVNPass &GVN) { |
| 2917 | auto FindRes = PhiTranslateTable.find(Val: {Num, Pred}); |
| 2918 | if (FindRes != PhiTranslateTable.end()) |
| 2919 | return FindRes->second; |
| 2920 | uint32_t NewNum = phiTranslateImpl(BB: Pred, PhiBlock, Num, GVN); |
| 2921 | PhiTranslateTable.insert(KV: {{Num, Pred}, NewNum}); |
| 2922 | return NewNum; |
| 2923 | } |
| 2924 | |
| 2925 | // Return true if the value number \p Num and NewNum have equal value. |
| 2926 | // Return false if the result is unknown. |
| 2927 | bool GVNPass::ValueTable::areCallValsEqual(uint32_t Num, uint32_t NewNum, |
| 2928 | const BasicBlock *Pred, |
| 2929 | const BasicBlock *PhiBlock, |
| 2930 | GVNPass &GVN) { |
| 2931 | CallInst *Call = nullptr; |
| 2932 | auto Leaders = GVN.LeaderTable.getLeaders(N: Num); |
| 2933 | for (const auto &Entry : Leaders) { |
| 2934 | Call = dyn_cast<CallInst>(Val: &*Entry.Val); |
| 2935 | if (Call && Call->getParent() == PhiBlock) |
| 2936 | break; |
| 2937 | } |
| 2938 | |
| 2939 | if (AA->doesNotAccessMemory(Call)) |
| 2940 | return true; |
| 2941 | |
| 2942 | if (!MD || !AA->onlyReadsMemory(Call)) |
| 2943 | return false; |
| 2944 | |
| 2945 | MemDepResult LocalDep = MD->getDependency(QueryInst: Call); |
| 2946 | if (!LocalDep.isNonLocal()) |
| 2947 | return false; |
| 2948 | |
| 2949 | const MemoryDependenceResults::NonLocalDepInfo &Deps = |
| 2950 | MD->getNonLocalCallDependency(QueryCall: Call); |
| 2951 | |
| 2952 | // Check to see if the Call has no function local clobber. |
| 2953 | for (const NonLocalDepEntry &D : Deps) { |
| 2954 | if (D.getResult().isNonFuncLocal()) |
| 2955 | return true; |
| 2956 | } |
| 2957 | return false; |
| 2958 | } |
| 2959 | |
| 2960 | /// Translate value number \p Num using phis, so that it has the values of |
| 2961 | /// the phis in BB. |
| 2962 | uint32_t GVNPass::ValueTable::phiTranslateImpl(const BasicBlock *Pred, |
| 2963 | const BasicBlock *PhiBlock, |
| 2964 | uint32_t Num, GVNPass &GVN) { |
| 2965 | // See if we can refine the value number by looking at the PN incoming value |
| 2966 | // for the given predecessor. |
| 2967 | if (PHINode *PN = NumberingPhi[Num]) { |
| 2968 | if (PN->getParent() != PhiBlock) |
| 2969 | return Num; |
| 2970 | for (unsigned I = 0; I != PN->getNumIncomingValues(); ++I) { |
| 2971 | if (PN->getIncomingBlock(i: I) != Pred) |
| 2972 | continue; |
| 2973 | if (uint32_t TransVal = lookup(V: PN->getIncomingValue(i: I), Verify: false)) |
| 2974 | return TransVal; |
| 2975 | } |
| 2976 | return Num; |
| 2977 | } |
| 2978 | |
| 2979 | if (BasicBlock *BB = NumberingBB[Num]) { |
| 2980 | assert(MSSA && "NumberingBB is non-empty only when using MemorySSA" ); |
| 2981 | // Value numbers of basic blocks are used to represent memory state in |
| 2982 | // load/store instructions and read-only function calls when said state is |
| 2983 | // set by a MemoryPhi. |
| 2984 | if (BB != PhiBlock) |
| 2985 | return Num; |
| 2986 | MemoryPhi *MPhi = MSSA->getMemoryAccess(BB); |
| 2987 | for (unsigned i = 0, N = MPhi->getNumIncomingValues(); i != N; ++i) { |
| 2988 | if (MPhi->getIncomingBlock(I: i) != Pred) |
| 2989 | continue; |
| 2990 | MemoryAccess *MA = MPhi->getIncomingValue(I: i); |
| 2991 | if (auto *PredPhi = dyn_cast<MemoryPhi>(Val: MA)) |
| 2992 | return lookupOrAdd(V: PredPhi->getBlock()); |
| 2993 | if (MSSA->isLiveOnEntryDef(MA)) |
| 2994 | return lookupOrAdd(V: &BB->getParent()->getEntryBlock()); |
| 2995 | return lookupOrAdd(V: cast<MemoryUseOrDef>(Val: MA)->getMemoryInst()); |
| 2996 | } |
| 2997 | llvm_unreachable( |
| 2998 | "CFG/MemorySSA mismatch: predecessor not found among incoming blocks" ); |
| 2999 | } |
| 3000 | |
| 3001 | // If there is any value related with Num is defined in a BB other than |
| 3002 | // PhiBlock, it cannot depend on a phi in PhiBlock without going through |
| 3003 | // a backedge. We can do an early exit in that case to save compile time. |
| 3004 | if (!areAllValsInBB(Num, BB: PhiBlock, GVN)) |
| 3005 | return Num; |
| 3006 | |
| 3007 | if (Num >= ExprIdx.size() || ExprIdx[Num] == 0) |
| 3008 | return Num; |
| 3009 | Expression Exp = Expressions[ExprIdx[Num]]; |
| 3010 | |
| 3011 | for (unsigned I = 0; I < Exp.VarArgs.size(); I++) { |
| 3012 | // For InsertValue and ExtractValue, some varargs are index numbers |
| 3013 | // instead of value numbers. Those index numbers should not be |
| 3014 | // translated. |
| 3015 | if ((I > 1 && Exp.Opcode == Instruction::InsertValue) || |
| 3016 | (I > 0 && Exp.Opcode == Instruction::ExtractValue) || |
| 3017 | (I > 1 && Exp.Opcode == Instruction::ShuffleVector)) |
| 3018 | continue; |
| 3019 | Exp.VarArgs[I] = phiTranslate(Pred, PhiBlock, Num: Exp.VarArgs[I], GVN); |
| 3020 | } |
| 3021 | |
| 3022 | if (Exp.Commutative) { |
| 3023 | assert(Exp.VarArgs.size() >= 2 && "Unsupported commutative instruction!" ); |
| 3024 | if (Exp.VarArgs[0] > Exp.VarArgs[1]) { |
| 3025 | std::swap(a&: Exp.VarArgs[0], b&: Exp.VarArgs[1]); |
| 3026 | uint32_t Opcode = Exp.Opcode >> 8; |
| 3027 | if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) |
| 3028 | Exp.Opcode = (Opcode << 8) | |
| 3029 | CmpInst::getSwappedPredicate( |
| 3030 | pred: static_cast<CmpInst::Predicate>(Exp.Opcode & 255)); |
| 3031 | } |
| 3032 | } |
| 3033 | |
| 3034 | if (uint32_t NewNum = ExpressionNumbering[Exp]) { |
| 3035 | if (Exp.Opcode == Instruction::Call && NewNum != Num) |
| 3036 | return areCallValsEqual(Num, NewNum, Pred, PhiBlock, GVN) ? NewNum : Num; |
| 3037 | return NewNum; |
| 3038 | } |
| 3039 | return Num; |
| 3040 | } |
| 3041 | |
| 3042 | /// Erase stale entry from phiTranslate cache so phiTranslate can be computed |
| 3043 | /// again. |
| 3044 | void GVNPass::ValueTable::eraseTranslateCacheEntry( |
| 3045 | uint32_t Num, const BasicBlock &CurrBlock) { |
| 3046 | for (const BasicBlock *Pred : predecessors(BB: &CurrBlock)) |
| 3047 | PhiTranslateTable.erase(Val: {Num, Pred}); |
| 3048 | } |
| 3049 | |
| 3050 | // In order to find a leader for a given value number at a |
| 3051 | // specific basic block, we first obtain the list of all Values for that number, |
| 3052 | // and then scan the list to find one whose block dominates the block in |
| 3053 | // question. This is fast because dominator tree queries consist of only |
| 3054 | // a few comparisons of DFS numbers. |
| 3055 | Value *GVNPass::findLeader(const BasicBlock *BB, uint32_t Num) { |
| 3056 | auto Leaders = LeaderTable.getLeaders(N: Num); |
| 3057 | if (Leaders.empty()) |
| 3058 | return nullptr; |
| 3059 | |
| 3060 | Value *Val = nullptr; |
| 3061 | for (const auto &Entry : Leaders) { |
| 3062 | if (DT->dominates(A: Entry.BB, B: BB)) { |
| 3063 | Val = Entry.Val; |
| 3064 | if (isa<Constant>(Val)) |
| 3065 | return Val; |
| 3066 | } |
| 3067 | } |
| 3068 | |
| 3069 | return Val; |
| 3070 | } |
| 3071 | |
| 3072 | /// There is an edge from 'Src' to 'Dst'. Return |
| 3073 | /// true if every path from the entry block to 'Dst' passes via this edge. In |
| 3074 | /// particular 'Dst' must not be reachable via another edge from 'Src'. |
| 3075 | static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, |
| 3076 | DominatorTree *DT) { |
| 3077 | // While in theory it is interesting to consider the case in which Dst has |
| 3078 | // more than one predecessor, because Dst might be part of a loop which is |
| 3079 | // only reachable from Src, in practice it is pointless since at the time |
| 3080 | // GVN runs all such loops have preheaders, which means that Dst will have |
| 3081 | // been changed to have only one predecessor, namely Src. |
| 3082 | const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); |
| 3083 | assert((!Pred || Pred == E.getStart()) && |
| 3084 | "No edge between these basic blocks!" ); |
| 3085 | return Pred != nullptr; |
| 3086 | } |
| 3087 | |
| 3088 | void GVNPass::assignBlockRPONumber(Function &F) { |
| 3089 | BlockRPONumber.clear(); |
| 3090 | uint32_t NextBlockNumber = 1; |
| 3091 | ReversePostOrderTraversal<Function *> RPOT(&F); |
| 3092 | for (BasicBlock *BB : RPOT) |
| 3093 | BlockRPONumber[BB] = NextBlockNumber++; |
| 3094 | InvalidBlockRPONumbers = false; |
| 3095 | } |
| 3096 | |
| 3097 | /// The given values are known to be equal in every use |
| 3098 | /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with |
| 3099 | /// 'RHS' everywhere in the scope. Returns whether a change was made. |
| 3100 | /// The Root may either be a basic block edge (for conditions) or an |
| 3101 | /// instruction (for assumes). |
| 3102 | bool GVNPass::propagateEquality( |
| 3103 | Value *LHS, Value *RHS, |
| 3104 | const std::variant<BasicBlockEdge, Instruction *> &Root) { |
| 3105 | SmallVector<std::pair<Value*, Value*>, 4> Worklist; |
| 3106 | Worklist.push_back(Elt: std::make_pair(x&: LHS, y&: RHS)); |
| 3107 | bool Changed = false; |
| 3108 | SmallVector<const BasicBlock *> DominatedBlocks; |
| 3109 | if (const BasicBlockEdge *Edge = std::get_if<BasicBlockEdge>(ptr: &Root)) { |
| 3110 | // For speed, compute a conservative fast approximation to |
| 3111 | // DT->dominates(Root, Root.getEnd()); |
| 3112 | if (isOnlyReachableViaThisEdge(E: *Edge, DT)) |
| 3113 | DominatedBlocks.push_back(Elt: Edge->getEnd()); |
| 3114 | } else { |
| 3115 | Instruction *I = std::get<Instruction *>(v: Root); |
| 3116 | for (const auto *Node : DT->getNode(BB: I->getParent())->children()) |
| 3117 | DominatedBlocks.push_back(Elt: Node->getBlock()); |
| 3118 | } |
| 3119 | |
| 3120 | while (!Worklist.empty()) { |
| 3121 | std::pair<Value*, Value*> Item = Worklist.pop_back_val(); |
| 3122 | LHS = Item.first; RHS = Item.second; |
| 3123 | |
| 3124 | if (LHS == RHS) |
| 3125 | continue; |
| 3126 | assert(LHS->getType() == RHS->getType() && "Equality but unequal types!" ); |
| 3127 | |
| 3128 | // Don't try to propagate equalities between constants. |
| 3129 | if (isa<Constant>(Val: LHS) && isa<Constant>(Val: RHS)) |
| 3130 | continue; |
| 3131 | |
| 3132 | // Prefer a constant on the right-hand side, or an Argument if no constants. |
| 3133 | if (isa<Constant>(Val: LHS) || (isa<Argument>(Val: LHS) && !isa<Constant>(Val: RHS))) |
| 3134 | std::swap(a&: LHS, b&: RHS); |
| 3135 | assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!" ); |
| 3136 | const DataLayout &DL = |
| 3137 | isa<Argument>(Val: LHS) |
| 3138 | ? cast<Argument>(Val: LHS)->getParent()->getDataLayout() |
| 3139 | : cast<Instruction>(Val: LHS)->getDataLayout(); |
| 3140 | |
| 3141 | // If there is no obvious reason to prefer the left-hand side over the |
| 3142 | // right-hand side, ensure the longest lived term is on the right-hand side, |
| 3143 | // so the shortest lived term will be replaced by the longest lived. |
| 3144 | // This tends to expose more simplifications. |
| 3145 | uint32_t LVN = VN.lookupOrAdd(V: LHS); |
| 3146 | if ((isa<Argument>(Val: LHS) && isa<Argument>(Val: RHS)) || |
| 3147 | (isa<Instruction>(Val: LHS) && isa<Instruction>(Val: RHS))) { |
| 3148 | // Move the 'oldest' value to the right-hand side, using the value number |
| 3149 | // as a proxy for age. |
| 3150 | uint32_t RVN = VN.lookupOrAdd(V: RHS); |
| 3151 | if (LVN < RVN) { |
| 3152 | std::swap(a&: LHS, b&: RHS); |
| 3153 | LVN = RVN; |
| 3154 | } |
| 3155 | } |
| 3156 | |
| 3157 | // If value numbering later sees that an instruction in the scope is equal |
| 3158 | // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve |
| 3159 | // the invariant that instructions only occur in the leader table for their |
| 3160 | // own value number (this is used by removeFromLeaderTable), do not do this |
| 3161 | // if RHS is an instruction (if an instruction in the scope is morphed into |
| 3162 | // LHS then it will be turned into RHS by the next GVN iteration anyway, so |
| 3163 | // using the leader table is about compiling faster, not optimizing better). |
| 3164 | // The leader table only tracks basic blocks, not edges. Only add to if we |
| 3165 | // have the simple case where the edge dominates the end. |
| 3166 | if (!isa<Instruction>(Val: RHS) && canReplacePointersIfEqual(From: LHS, To: RHS, DL)) |
| 3167 | for (const BasicBlock *BB : DominatedBlocks) |
| 3168 | LeaderTable.insert(N: LVN, V: RHS, BB); |
| 3169 | |
| 3170 | // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As |
| 3171 | // LHS always has at least one use that is not dominated by Root, this will |
| 3172 | // never do anything if LHS has only one use. |
| 3173 | if (!LHS->hasOneUse()) { |
| 3174 | // Create a callback that captures the DL. |
| 3175 | auto CanReplacePointersCallBack = [&DL](const Use &U, const Value *To) { |
| 3176 | return canReplacePointersInUseIfEqual(U, To, DL); |
| 3177 | }; |
| 3178 | unsigned NumReplacements; |
| 3179 | if (const BasicBlockEdge *Edge = std::get_if<BasicBlockEdge>(ptr: &Root)) |
| 3180 | NumReplacements = replaceDominatedUsesWithIf( |
| 3181 | From: LHS, To: RHS, DT&: *DT, Edge: *Edge, ShouldReplace: CanReplacePointersCallBack); |
| 3182 | else |
| 3183 | NumReplacements = replaceDominatedUsesWithIf( |
| 3184 | From: LHS, To: RHS, DT&: *DT, I: std::get<Instruction *>(v: Root), |
| 3185 | ShouldReplace: CanReplacePointersCallBack); |
| 3186 | |
| 3187 | if (NumReplacements > 0) { |
| 3188 | Changed = true; |
| 3189 | NumGVNEqProp += NumReplacements; |
| 3190 | // Cached information for anything that uses LHS will be invalid. |
| 3191 | if (MD) |
| 3192 | MD->invalidateCachedPointerInfo(Ptr: LHS); |
| 3193 | } |
| 3194 | } |
| 3195 | |
| 3196 | // Now try to deduce additional equalities from this one. For example, if |
| 3197 | // the known equality was "(A != B)" == "false" then it follows that A and B |
| 3198 | // are equal in the scope. Only boolean equalities with an explicit true or |
| 3199 | // false RHS are currently supported. |
| 3200 | if (!RHS->getType()->isIntegerTy(BitWidth: 1)) |
| 3201 | // Not a boolean equality - bail out. |
| 3202 | continue; |
| 3203 | ConstantInt *CI = dyn_cast<ConstantInt>(Val: RHS); |
| 3204 | if (!CI) |
| 3205 | // RHS neither 'true' nor 'false' - bail out. |
| 3206 | continue; |
| 3207 | // Whether RHS equals 'true'. Otherwise it equals 'false'. |
| 3208 | bool IsKnownTrue = CI->isMinusOne(); |
| 3209 | bool IsKnownFalse = !IsKnownTrue; |
| 3210 | |
| 3211 | // If "A && B" is known true then both A and B are known true. If "A || B" |
| 3212 | // is known false then both A and B are known false. |
| 3213 | Value *A, *B; |
| 3214 | if ((IsKnownTrue && match(V: LHS, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B)))) || |
| 3215 | (IsKnownFalse && match(V: LHS, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))))) { |
| 3216 | Worklist.push_back(Elt: std::make_pair(x&: A, y&: RHS)); |
| 3217 | Worklist.push_back(Elt: std::make_pair(x&: B, y&: RHS)); |
| 3218 | continue; |
| 3219 | } |
| 3220 | |
| 3221 | // If we are propagating an equality like "(A == B)" == "true" then also |
| 3222 | // propagate the equality A == B. When propagating a comparison such as |
| 3223 | // "(A >= B)" == "true", replace all instances of "A < B" with "false". |
| 3224 | if (CmpInst *Cmp = dyn_cast<CmpInst>(Val: LHS)) { |
| 3225 | Value *Op0 = Cmp->getOperand(i_nocapture: 0), *Op1 = Cmp->getOperand(i_nocapture: 1); |
| 3226 | |
| 3227 | // If "A == B" is known true, or "A != B" is known false, then replace |
| 3228 | // A with B everywhere in the scope. For floating point operations, we |
| 3229 | // have to be careful since equality does not always imply equivalance. |
| 3230 | if (Cmp->isEquivalence(Invert: IsKnownFalse)) |
| 3231 | Worklist.push_back(Elt: std::make_pair(x&: Op0, y&: Op1)); |
| 3232 | |
| 3233 | // If "A >= B" is known true, replace "A < B" with false everywhere. |
| 3234 | CmpInst::Predicate NotPred = Cmp->getInversePredicate(); |
| 3235 | Constant *NotVal = ConstantInt::get(Ty: Cmp->getType(), V: IsKnownFalse); |
| 3236 | // Since we don't have the instruction "A < B" immediately to hand, work |
| 3237 | // out the value number that it would have and use that to find an |
| 3238 | // appropriate instruction (if any). |
| 3239 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
| 3240 | uint32_t Num = VN.lookupOrAddCmp(Opcode: Cmp->getOpcode(), Predicate: NotPred, LHS: Op0, RHS: Op1); |
| 3241 | // If the number we were assigned was brand new then there is no point in |
| 3242 | // looking for an instruction realizing it: there cannot be one! |
| 3243 | if (Num < NextNum) { |
| 3244 | for (const auto &Entry : LeaderTable.getLeaders(N: Num)) { |
| 3245 | // Only look at leaders that either dominate the start of the edge, |
| 3246 | // or are dominated by the end. This check is not necessary for |
| 3247 | // correctness, it only discards cases for which the following |
| 3248 | // use replacement will not work anyway. |
| 3249 | if (const BasicBlockEdge *Edge = std::get_if<BasicBlockEdge>(ptr: &Root)) { |
| 3250 | if (!DT->dominates(A: Entry.BB, B: Edge->getStart()) && |
| 3251 | !DT->dominates(A: Edge->getEnd(), B: Entry.BB)) |
| 3252 | continue; |
| 3253 | } else { |
| 3254 | auto *InstBB = std::get<Instruction *>(v: Root)->getParent(); |
| 3255 | if (!DT->dominates(A: Entry.BB, B: InstBB) && |
| 3256 | !DT->dominates(A: InstBB, B: Entry.BB)) |
| 3257 | continue; |
| 3258 | } |
| 3259 | |
| 3260 | Value *NotCmp = Entry.Val; |
| 3261 | if (NotCmp && isa<Instruction>(Val: NotCmp)) { |
| 3262 | unsigned NumReplacements; |
| 3263 | if (const BasicBlockEdge *Edge = std::get_if<BasicBlockEdge>(ptr: &Root)) |
| 3264 | NumReplacements = |
| 3265 | replaceDominatedUsesWith(From: NotCmp, To: NotVal, DT&: *DT, Edge: *Edge); |
| 3266 | else |
| 3267 | NumReplacements = replaceDominatedUsesWith( |
| 3268 | From: NotCmp, To: NotVal, DT&: *DT, I: std::get<Instruction *>(v: Root)); |
| 3269 | Changed |= NumReplacements > 0; |
| 3270 | NumGVNEqProp += NumReplacements; |
| 3271 | // Cached information for anything that uses NotCmp will be invalid. |
| 3272 | if (MD) |
| 3273 | MD->invalidateCachedPointerInfo(Ptr: NotCmp); |
| 3274 | } |
| 3275 | } |
| 3276 | } |
| 3277 | // Ensure that any instruction in scope that gets the "A < B" value number |
| 3278 | // is replaced with false. |
| 3279 | // The leader table only tracks basic blocks, not edges. Only add to if we |
| 3280 | // have the simple case where the edge dominates the end. |
| 3281 | for (const BasicBlock *BB : DominatedBlocks) |
| 3282 | LeaderTable.insert(N: Num, V: NotVal, BB); |
| 3283 | |
| 3284 | continue; |
| 3285 | } |
| 3286 | |
| 3287 | // Propagate equalities that results from truncation with no unsigned wrap |
| 3288 | // like (trunc nuw i64 %v to i1) == "true" or (trunc nuw i64 %v to i1) == |
| 3289 | // "false" |
| 3290 | if (match(V: LHS, P: m_NUWTrunc(Op: m_Value(V&: A)))) { |
| 3291 | Worklist.emplace_back(Args&: A, Args: ConstantInt::get(Ty: A->getType(), V: IsKnownTrue)); |
| 3292 | continue; |
| 3293 | } |
| 3294 | |
| 3295 | if (match(V: LHS, P: m_Not(V: m_Value(V&: A)))) { |
| 3296 | Worklist.emplace_back(Args&: A, Args: ConstantInt::get(Ty: A->getType(), V: !IsKnownTrue)); |
| 3297 | continue; |
| 3298 | } |
| 3299 | } |
| 3300 | |
| 3301 | return Changed; |
| 3302 | } |
| 3303 | |
| 3304 | /// When calculating availability, handle an instruction |
| 3305 | /// by inserting it into the appropriate sets. |
| 3306 | bool GVNPass::processInstruction(Instruction *I) { |
| 3307 | // If the instruction can be easily simplified then do so now in preference |
| 3308 | // to value numbering it. Value numbering often exposes redundancies, for |
| 3309 | // example if it determines that %y is equal to %x then the instruction |
| 3310 | // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. |
| 3311 | const DataLayout &DL = I->getDataLayout(); |
| 3312 | if (Value *V = simplifyInstruction(I, Q: {DL, TLI, DT, AC})) { |
| 3313 | bool Changed = false; |
| 3314 | if (!I->use_empty()) { |
| 3315 | // Simplification can cause a special instruction to become not special. |
| 3316 | // For example, devirtualization to a willreturn function. |
| 3317 | ICF->removeUsersOf(Inst: I); |
| 3318 | I->replaceAllUsesWith(V); |
| 3319 | Changed = true; |
| 3320 | } |
| 3321 | if (isInstructionTriviallyDead(I, TLI)) { |
| 3322 | salvageAndRemoveInstruction(I); |
| 3323 | Changed = true; |
| 3324 | } |
| 3325 | if (Changed) { |
| 3326 | if (MD && V->getType()->isPtrOrPtrVectorTy()) |
| 3327 | MD->invalidateCachedPointerInfo(Ptr: V); |
| 3328 | ++NumGVNSimpl; |
| 3329 | return true; |
| 3330 | } |
| 3331 | } |
| 3332 | |
| 3333 | if (auto *Assume = dyn_cast<AssumeInst>(Val: I)) |
| 3334 | return processAssumeIntrinsic(IntrinsicI: Assume); |
| 3335 | |
| 3336 | if (LoadInst *Load = dyn_cast<LoadInst>(Val: I)) { |
| 3337 | if (processLoad(L: Load)) |
| 3338 | return true; |
| 3339 | |
| 3340 | unsigned Num = VN.lookupOrAdd(V: Load); |
| 3341 | LeaderTable.insert(N: Num, V: Load, BB: Load->getParent()); |
| 3342 | return false; |
| 3343 | } |
| 3344 | |
| 3345 | if (match(V: I, P: m_Intrinsic<Intrinsic::masked_load>()) && |
| 3346 | processMaskedLoad(I: cast<IntrinsicInst>(Val: I))) |
| 3347 | return true; |
| 3348 | |
| 3349 | // For conditional branches, we can perform simple conditional propagation on |
| 3350 | // the condition value itself. |
| 3351 | if (CondBrInst *BI = dyn_cast<CondBrInst>(Val: I)) { |
| 3352 | if (isa<Constant>(Val: BI->getCondition())) |
| 3353 | return processFoldableCondBr(BI); |
| 3354 | |
| 3355 | Value *BranchCond = BI->getCondition(); |
| 3356 | BasicBlock *TrueSucc = BI->getSuccessor(i: 0); |
| 3357 | BasicBlock *FalseSucc = BI->getSuccessor(i: 1); |
| 3358 | // Avoid multiple edges early. |
| 3359 | if (TrueSucc == FalseSucc) |
| 3360 | return false; |
| 3361 | |
| 3362 | BasicBlock *Parent = BI->getParent(); |
| 3363 | bool Changed = false; |
| 3364 | |
| 3365 | Value *TrueVal = ConstantInt::getTrue(Context&: TrueSucc->getContext()); |
| 3366 | BasicBlockEdge TrueE(Parent, TrueSucc); |
| 3367 | Changed |= propagateEquality(LHS: BranchCond, RHS: TrueVal, Root: TrueE); |
| 3368 | |
| 3369 | Value *FalseVal = ConstantInt::getFalse(Context&: FalseSucc->getContext()); |
| 3370 | BasicBlockEdge FalseE(Parent, FalseSucc); |
| 3371 | Changed |= propagateEquality(LHS: BranchCond, RHS: FalseVal, Root: FalseE); |
| 3372 | |
| 3373 | return Changed; |
| 3374 | } |
| 3375 | |
| 3376 | // For switches, propagate the case values into the case destinations. |
| 3377 | if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: I)) { |
| 3378 | Value *SwitchCond = SI->getCondition(); |
| 3379 | BasicBlock *Parent = SI->getParent(); |
| 3380 | bool Changed = false; |
| 3381 | |
| 3382 | // Remember how many outgoing edges there are to every successor. |
| 3383 | SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; |
| 3384 | for (BasicBlock *Succ : successors(BB: Parent)) |
| 3385 | ++SwitchEdges[Succ]; |
| 3386 | |
| 3387 | for (const auto &Case : SI->cases()) { |
| 3388 | BasicBlock *Dst = Case.getCaseSuccessor(); |
| 3389 | // If there is only a single edge, propagate the case value into it. |
| 3390 | if (SwitchEdges.lookup(Val: Dst) == 1) { |
| 3391 | BasicBlockEdge E(Parent, Dst); |
| 3392 | Changed |= propagateEquality(LHS: SwitchCond, RHS: Case.getCaseValue(), Root: E); |
| 3393 | } |
| 3394 | } |
| 3395 | return Changed; |
| 3396 | } |
| 3397 | |
| 3398 | // Instructions with void type don't return a value, so there's |
| 3399 | // no point in trying to find redundancies in them. |
| 3400 | if (I->getType()->isVoidTy()) |
| 3401 | return false; |
| 3402 | |
| 3403 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
| 3404 | unsigned Num = VN.lookupOrAdd(V: I); |
| 3405 | |
| 3406 | // Allocations are always uniquely numbered, so we can save time and memory |
| 3407 | // by fast failing them. |
| 3408 | if (isa<AllocaInst>(Val: I) || I->isTerminator() || isa<PHINode>(Val: I)) { |
| 3409 | LeaderTable.insert(N: Num, V: I, BB: I->getParent()); |
| 3410 | return false; |
| 3411 | } |
| 3412 | |
| 3413 | // A ptrtoaddr and a ptrtoint of the same pointer compute the same value when |
| 3414 | // the address width equals the pointer representation width. |
| 3415 | if (auto *PTA = dyn_cast<PtrToAddrInst>(Val: I)) { |
| 3416 | const DataLayout &DL = I->getDataLayout(); |
| 3417 | unsigned AS = PTA->getPointerAddressSpace(); |
| 3418 | if (DL.getAddressSizeInBits(AS) == DL.getPointerSizeInBits(AS) && |
| 3419 | !DL.hasUnstableRepresentation(AddrSpace: AS)) { |
| 3420 | uint32_t PTINum = |
| 3421 | VN.lookupPtrToInt(Ptr: PTA->getPointerOperand(), Ty: PTA->getType()); |
| 3422 | if (Value *PTI = findLeader(BB: I->getParent(), Num: PTINum)) { |
| 3423 | patchAndReplaceAllUsesWith(I, Repl: PTI); |
| 3424 | salvageAndRemoveInstruction(I); |
| 3425 | return true; |
| 3426 | } |
| 3427 | } |
| 3428 | } |
| 3429 | |
| 3430 | // If the number we were assigned was a brand new VN, then we don't |
| 3431 | // need to do a lookup to see if the number already exists |
| 3432 | // somewhere in the domtree: it can't! |
| 3433 | if (Num >= NextNum) { |
| 3434 | LeaderTable.insert(N: Num, V: I, BB: I->getParent()); |
| 3435 | return false; |
| 3436 | } |
| 3437 | |
| 3438 | // Perform fast-path value-number based elimination of values inherited from |
| 3439 | // dominators. |
| 3440 | Value *Repl = findLeader(BB: I->getParent(), Num); |
| 3441 | if (!Repl) { |
| 3442 | // Failure, just remember this instance for future use. |
| 3443 | LeaderTable.insert(N: Num, V: I, BB: I->getParent()); |
| 3444 | return false; |
| 3445 | } |
| 3446 | |
| 3447 | if (Repl == I) { |
| 3448 | // If I was the result of a shortcut PRE, it might already be in the table |
| 3449 | // and the best replacement for itself. Nothing to do. |
| 3450 | return false; |
| 3451 | } |
| 3452 | |
| 3453 | // Remove it! |
| 3454 | patchAndReplaceAllUsesWith(I, Repl); |
| 3455 | if (MD && Repl->getType()->isPtrOrPtrVectorTy()) |
| 3456 | MD->invalidateCachedPointerInfo(Ptr: Repl); |
| 3457 | salvageAndRemoveInstruction(I); |
| 3458 | return true; |
| 3459 | } |
| 3460 | |
| 3461 | /// runOnFunction - This is the main transformation entry point for a function. |
| 3462 | bool GVNPass::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT, |
| 3463 | const TargetLibraryInfo &RunTLI, AAResults &RunAA, |
| 3464 | MemoryDependenceResults *RunMD, LoopInfo &LI, |
| 3465 | OptimizationRemarkEmitter *RunORE, MemorySSA *MSSA) { |
| 3466 | AC = &RunAC; |
| 3467 | DT = &RunDT; |
| 3468 | VN.setDomTree(DT); |
| 3469 | TLI = &RunTLI; |
| 3470 | AA = &RunAA; |
| 3471 | VN.setAliasAnalysis(&RunAA); |
| 3472 | MD = RunMD; |
| 3473 | ImplicitControlFlowTracking ImplicitCFT; |
| 3474 | ICF = &ImplicitCFT; |
| 3475 | this->LI = &LI; |
| 3476 | VN.setMemDep(M: MD); |
| 3477 | // Propagate the MSSA-enabled flag so the value-numbering paths in |
| 3478 | // lookupOrAddCall() and computeLoadStoreVN(), which depends on whether |
| 3479 | // IsMSSAEnabled is turned on. |
| 3480 | VN.setMemorySSA(M: MSSA, MSSAEnabled: isMemorySSAEnabled()); |
| 3481 | ORE = RunORE; |
| 3482 | InvalidBlockRPONumbers = true; |
| 3483 | MemorySSAUpdater Updater(MSSA); |
| 3484 | MSSAU = MSSA ? &Updater : nullptr; |
| 3485 | |
| 3486 | bool Changed = false; |
| 3487 | bool ShouldContinue = true; |
| 3488 | |
| 3489 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); |
| 3490 | // Merge unconditional branches, allowing PRE to catch more |
| 3491 | // optimization opportunities. |
| 3492 | for (BasicBlock &BB : make_early_inc_range(Range&: F)) { |
| 3493 | bool RemovedBlock = MergeBlockIntoPredecessor(BB: &BB, DTU: &DTU, LI: &LI, MSSAU, MemDep: MD); |
| 3494 | if (RemovedBlock) |
| 3495 | ++NumGVNBlocks; |
| 3496 | |
| 3497 | Changed |= RemovedBlock; |
| 3498 | } |
| 3499 | DTU.flush(); |
| 3500 | |
| 3501 | unsigned Iteration = 0; |
| 3502 | while (ShouldContinue) { |
| 3503 | LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n" ); |
| 3504 | (void) Iteration; |
| 3505 | ShouldContinue = iterateOnFunction(F); |
| 3506 | Changed |= ShouldContinue; |
| 3507 | ++Iteration; |
| 3508 | } |
| 3509 | |
| 3510 | if (isScalarPREEnabled()) { |
| 3511 | // Fabricate val-num for dead-code in order to suppress assertion in |
| 3512 | // performPRE(). |
| 3513 | assignValNumForDeadCode(); |
| 3514 | bool PREChanged = true; |
| 3515 | while (PREChanged) { |
| 3516 | PREChanged = performPRE(F); |
| 3517 | Changed |= PREChanged; |
| 3518 | } |
| 3519 | } |
| 3520 | |
| 3521 | // FIXME: Should perform GVN again after PRE does something. PRE can move |
| 3522 | // computations into blocks where they become fully redundant. Note that |
| 3523 | // we can't do this until PRE's critical edge splitting updates memdep. |
| 3524 | // Actually, when this happens, we should just fully integrate PRE into GVN. |
| 3525 | |
| 3526 | cleanupGlobalSets(); |
| 3527 | // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each |
| 3528 | // iteration. |
| 3529 | DeadBlocks.clear(); |
| 3530 | |
| 3531 | if (MSSA && VerifyMemorySSA) |
| 3532 | MSSA->verifyMemorySSA(); |
| 3533 | |
| 3534 | return Changed; |
| 3535 | } |
| 3536 | |
| 3537 | bool GVNPass::processBlock(BasicBlock *BB) { |
| 3538 | if (DeadBlocks.count(key: BB)) |
| 3539 | return false; |
| 3540 | |
| 3541 | bool ChangedFunction = false; |
| 3542 | |
| 3543 | // Since we may not have visited the input blocks of the phis, we can't |
| 3544 | // use our normal hash approach for phis. Instead, simply look for |
| 3545 | // obvious duplicates. The first pass of GVN will tend to create |
| 3546 | // identical phis, and the second or later passes can eliminate them. |
| 3547 | SmallPtrSet<PHINode *, 8> PHINodesToRemove; |
| 3548 | ChangedFunction |= EliminateDuplicatePHINodes(BB, ToRemove&: PHINodesToRemove); |
| 3549 | for (PHINode *PN : PHINodesToRemove) { |
| 3550 | removeInstruction(I: PN); |
| 3551 | } |
| 3552 | for (Instruction &Inst : make_early_inc_range(Range&: *BB)) |
| 3553 | ChangedFunction |= processInstruction(I: &Inst); |
| 3554 | return ChangedFunction; |
| 3555 | } |
| 3556 | |
| 3557 | // Instantiate an expression in a predecessor that lacked it. |
| 3558 | bool GVNPass::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred, |
| 3559 | BasicBlock *Curr, unsigned int ValNo) { |
| 3560 | // Because we are going top-down through the block, all value numbers |
| 3561 | // will be available in the predecessor by the time we need them. Any |
| 3562 | // that weren't originally present will have been instantiated earlier |
| 3563 | // in this loop. |
| 3564 | bool Success = true; |
| 3565 | for (unsigned I = 0, E = Instr->getNumOperands(); I != E; ++I) { |
| 3566 | Value *Op = Instr->getOperand(i: I); |
| 3567 | if (isa<Argument>(Val: Op) || isa<Constant>(Val: Op) || isa<GlobalValue>(Val: Op)) |
| 3568 | continue; |
| 3569 | // This could be a newly inserted instruction, in which case, we won't |
| 3570 | // find a value number, and should give up before we hurt ourselves. |
| 3571 | // FIXME: Rewrite the infrastructure to let it easier to value number |
| 3572 | // and process newly inserted instructions. |
| 3573 | if (!VN.exists(V: Op)) { |
| 3574 | Success = false; |
| 3575 | break; |
| 3576 | } |
| 3577 | uint32_t TValNo = |
| 3578 | VN.phiTranslate(Pred, PhiBlock: Curr, Num: VN.lookup(V: Op), GVN&: *this); |
| 3579 | if (Value *V = findLeader(BB: Pred, Num: TValNo)) { |
| 3580 | Instr->setOperand(i: I, Val: V); |
| 3581 | } else { |
| 3582 | Success = false; |
| 3583 | break; |
| 3584 | } |
| 3585 | } |
| 3586 | |
| 3587 | // Fail out if we encounter an operand that is not available in |
| 3588 | // the PRE predecessor. This is typically because of loads which |
| 3589 | // are not value numbered precisely. |
| 3590 | if (!Success) |
| 3591 | return false; |
| 3592 | |
| 3593 | Instr->insertBefore(InsertPos: Pred->getTerminator()->getIterator()); |
| 3594 | Instr->setName(Instr->getName() + ".pre" ); |
| 3595 | Instr->setDebugLoc(Instr->getDebugLoc()); |
| 3596 | |
| 3597 | ICF->insertInstructionTo(Inst: Instr, BB: Pred); |
| 3598 | |
| 3599 | unsigned Num = VN.lookupOrAdd(V: Instr); |
| 3600 | VN.add(V: Instr, Num); |
| 3601 | |
| 3602 | // Update the availability map to include the new instruction. |
| 3603 | LeaderTable.insert(N: Num, V: Instr, BB: Pred); |
| 3604 | return true; |
| 3605 | } |
| 3606 | |
| 3607 | bool GVNPass::performScalarPRE(Instruction *CurInst) { |
| 3608 | if (isa<AllocaInst>(Val: CurInst) || CurInst->isTerminator() || |
| 3609 | isa<PHINode>(Val: CurInst) || CurInst->getType()->isVoidTy() || |
| 3610 | CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || |
| 3611 | CurInst->getType()->isTokenLikeTy()) |
| 3612 | return false; |
| 3613 | |
| 3614 | // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from |
| 3615 | // sinking the compare again, and it would force the code generator to |
| 3616 | // move the i1 from processor flags or predicate registers into a general |
| 3617 | // purpose register. |
| 3618 | if (isa<CmpInst>(Val: CurInst)) |
| 3619 | return false; |
| 3620 | |
| 3621 | // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from |
| 3622 | // sinking the addressing mode computation back to its uses. Extending the |
| 3623 | // GEP's live range increases the register pressure, and therefore it can |
| 3624 | // introduce unnecessary spills. |
| 3625 | // |
| 3626 | // This doesn't prevent Load PRE. PHI translation will make the GEP available |
| 3627 | // to the load by moving it to the predecessor block if necessary. |
| 3628 | if (isa<GetElementPtrInst>(Val: CurInst)) |
| 3629 | return false; |
| 3630 | |
| 3631 | if (auto *CallB = dyn_cast<CallBase>(Val: CurInst)) { |
| 3632 | // We don't currently value number ANY inline asm calls. |
| 3633 | if (CallB->isInlineAsm()) |
| 3634 | return false; |
| 3635 | } |
| 3636 | |
| 3637 | uint32_t ValNo = VN.lookup(V: CurInst); |
| 3638 | |
| 3639 | // Look for the predecessors for PRE opportunities. We're |
| 3640 | // only trying to solve the basic diamond case, where |
| 3641 | // a value is computed in the successor and one predecessor, |
| 3642 | // but not the other. We also explicitly disallow cases |
| 3643 | // where the successor is its own predecessor, because they're |
| 3644 | // more complicated to get right. |
| 3645 | unsigned NumWith = 0; |
| 3646 | unsigned NumWithout = 0; |
| 3647 | BasicBlock *PREPred = nullptr; |
| 3648 | BasicBlock *CurrentBlock = CurInst->getParent(); |
| 3649 | |
| 3650 | // Update the RPO numbers for this function. |
| 3651 | if (InvalidBlockRPONumbers) |
| 3652 | assignBlockRPONumber(F&: *CurrentBlock->getParent()); |
| 3653 | |
| 3654 | SmallVector<std::pair<Value *, BasicBlock *>, 8> PredMap; |
| 3655 | for (BasicBlock *P : predecessors(BB: CurrentBlock)) { |
| 3656 | // We're not interested in PRE where blocks with predecessors that are |
| 3657 | // not reachable. |
| 3658 | if (!DT->isReachableFromEntry(A: P)) { |
| 3659 | NumWithout = 2; |
| 3660 | break; |
| 3661 | } |
| 3662 | // It is not safe to do PRE when P->CurrentBlock is a loop backedge. |
| 3663 | assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && |
| 3664 | "Invalid BlockRPONumber map." ); |
| 3665 | if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock]) { |
| 3666 | NumWithout = 2; |
| 3667 | break; |
| 3668 | } |
| 3669 | |
| 3670 | uint32_t TValNo = VN.phiTranslate(Pred: P, PhiBlock: CurrentBlock, Num: ValNo, GVN&: *this); |
| 3671 | Value *PredV = findLeader(BB: P, Num: TValNo); |
| 3672 | if (!PredV) { |
| 3673 | PredMap.push_back(Elt: std::make_pair(x: static_cast<Value *>(nullptr), y&: P)); |
| 3674 | PREPred = P; |
| 3675 | ++NumWithout; |
| 3676 | } else if (PredV == CurInst) { |
| 3677 | // CurInst dominates this predecessor. |
| 3678 | NumWithout = 2; |
| 3679 | break; |
| 3680 | } else { |
| 3681 | PredMap.push_back(Elt: std::make_pair(x&: PredV, y&: P)); |
| 3682 | ++NumWith; |
| 3683 | } |
| 3684 | } |
| 3685 | |
| 3686 | // Don't do PRE when it might increase code size, i.e. when |
| 3687 | // we would need to insert instructions in more than one pred. |
| 3688 | if (NumWithout > 1 || NumWith == 0) |
| 3689 | return false; |
| 3690 | |
| 3691 | // We may have a case where all predecessors have the instruction, |
| 3692 | // and we just need to insert a phi node. Otherwise, perform |
| 3693 | // insertion. |
| 3694 | Instruction *PREInstr = nullptr; |
| 3695 | |
| 3696 | if (NumWithout != 0) { |
| 3697 | if (!isSafeToSpeculativelyExecute(I: CurInst)) { |
| 3698 | // It is only valid to insert a new instruction if the current instruction |
| 3699 | // is always executed. An instruction with implicit control flow could |
| 3700 | // prevent us from doing it. If we cannot speculate the execution, then |
| 3701 | // PRE should be prohibited. |
| 3702 | if (ICF->isDominatedByICFIFromSameBlock(Insn: CurInst)) |
| 3703 | return false; |
| 3704 | } |
| 3705 | |
| 3706 | // Don't do PRE across indirect branch. |
| 3707 | if (isa<IndirectBrInst>(Val: PREPred->getTerminator())) |
| 3708 | return false; |
| 3709 | |
| 3710 | // We can't do PRE safely on a critical edge, so instead we schedule |
| 3711 | // the edge to be split and perform the PRE the next time we iterate |
| 3712 | // on the function. |
| 3713 | unsigned SuccNum = GetSuccessorNumber(BB: PREPred, Succ: CurrentBlock); |
| 3714 | if (isCriticalEdge(TI: PREPred->getTerminator(), SuccNum)) { |
| 3715 | ToSplit.push_back(Elt: std::make_pair(x: PREPred->getTerminator(), y&: SuccNum)); |
| 3716 | return false; |
| 3717 | } |
| 3718 | // We need to insert somewhere, so let's give it a shot. |
| 3719 | PREInstr = CurInst->clone(); |
| 3720 | if (!performScalarPREInsertion(Instr: PREInstr, Pred: PREPred, Curr: CurrentBlock, ValNo)) { |
| 3721 | // If we failed insertion, make sure we remove the instruction. |
| 3722 | #ifndef NDEBUG |
| 3723 | verifyRemoved(PREInstr); |
| 3724 | #endif |
| 3725 | PREInstr->deleteValue(); |
| 3726 | return false; |
| 3727 | } |
| 3728 | } |
| 3729 | |
| 3730 | // Either we should have filled in the PRE instruction, or we should |
| 3731 | // not have needed insertions. |
| 3732 | assert(PREInstr != nullptr || NumWithout == 0); |
| 3733 | |
| 3734 | ++NumGVNPRE; |
| 3735 | |
| 3736 | // Create a PHI to make the value available in this block. |
| 3737 | PHINode *Phi = PHINode::Create(Ty: CurInst->getType(), NumReservedValues: PredMap.size(), |
| 3738 | NameStr: CurInst->getName() + ".pre-phi" ); |
| 3739 | Phi->insertBefore(InsertPos: CurrentBlock->begin()); |
| 3740 | for (auto &[V, BB] : PredMap) { |
| 3741 | if (V) { |
| 3742 | // If we use an existing value in this phi, we have to patch the original |
| 3743 | // value because the phi will be used to replace a later value. |
| 3744 | patchReplacementInstruction(I: CurInst, Repl: V); |
| 3745 | Phi->addIncoming(V, BB); |
| 3746 | } else |
| 3747 | Phi->addIncoming(V: PREInstr, BB: PREPred); |
| 3748 | } |
| 3749 | |
| 3750 | VN.add(V: Phi, Num: ValNo); |
| 3751 | // After creating a new PHI for ValNo, the phi translate result for ValNo will |
| 3752 | // be changed, so erase the related stale entries in phi translate cache. |
| 3753 | VN.eraseTranslateCacheEntry(Num: ValNo, CurrBlock: *CurrentBlock); |
| 3754 | LeaderTable.insert(N: ValNo, V: Phi, BB: CurrentBlock); |
| 3755 | Phi->setDebugLoc(CurInst->getDebugLoc()); |
| 3756 | CurInst->replaceAllUsesWith(V: Phi); |
| 3757 | if (MD && Phi->getType()->isPtrOrPtrVectorTy()) |
| 3758 | MD->invalidateCachedPointerInfo(Ptr: Phi); |
| 3759 | LeaderTable.erase(N: ValNo, I: CurInst, BB: CurrentBlock); |
| 3760 | |
| 3761 | LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); |
| 3762 | removeInstruction(I: CurInst); |
| 3763 | |
| 3764 | return true; |
| 3765 | } |
| 3766 | |
| 3767 | /// Perform a purely local form of PRE that looks for diamond |
| 3768 | /// control flow patterns and attempts to perform simple PRE at the join point. |
| 3769 | bool GVNPass::performPRE(Function &F) { |
| 3770 | bool Changed = false; |
| 3771 | for (BasicBlock *CurrentBlock : depth_first(G: &F.getEntryBlock())) { |
| 3772 | // Nothing to PRE in the entry block. |
| 3773 | if (CurrentBlock == &F.getEntryBlock()) |
| 3774 | continue; |
| 3775 | |
| 3776 | // Don't perform PRE on an EH pad. |
| 3777 | if (CurrentBlock->isEHPad()) |
| 3778 | continue; |
| 3779 | |
| 3780 | for (BasicBlock::iterator BI = CurrentBlock->begin(), |
| 3781 | BE = CurrentBlock->end(); |
| 3782 | BI != BE;) { |
| 3783 | Instruction *CurInst = &*BI++; |
| 3784 | Changed |= performScalarPRE(CurInst); |
| 3785 | } |
| 3786 | } |
| 3787 | |
| 3788 | if (splitCriticalEdges()) |
| 3789 | Changed = true; |
| 3790 | |
| 3791 | return Changed; |
| 3792 | } |
| 3793 | |
| 3794 | /// Split the critical edge connecting the given two blocks, and return |
| 3795 | /// the block inserted to the critical edge. |
| 3796 | BasicBlock *GVNPass::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { |
| 3797 | // GVN does not require loop-simplify, do not try to preserve it if it is not |
| 3798 | // possible. |
| 3799 | BasicBlock *BB = SplitCriticalEdge( |
| 3800 | Src: Pred, Dst: Succ, |
| 3801 | Options: CriticalEdgeSplittingOptions(DT, LI, MSSAU).unsetPreserveLoopSimplify()); |
| 3802 | if (BB) { |
| 3803 | if (MD) |
| 3804 | MD->invalidateCachedPredecessors(); |
| 3805 | InvalidBlockRPONumbers = true; |
| 3806 | } |
| 3807 | return BB; |
| 3808 | } |
| 3809 | |
| 3810 | /// Split critical edges found during the previous |
| 3811 | /// iteration that may enable further optimization. |
| 3812 | bool GVNPass::splitCriticalEdges() { |
| 3813 | if (ToSplit.empty()) |
| 3814 | return false; |
| 3815 | |
| 3816 | bool Changed = false; |
| 3817 | do { |
| 3818 | std::pair<Instruction *, unsigned> Edge = ToSplit.pop_back_val(); |
| 3819 | Changed |= SplitCriticalEdge(TI: Edge.first, SuccNum: Edge.second, |
| 3820 | Options: CriticalEdgeSplittingOptions(DT, LI, MSSAU)) != |
| 3821 | nullptr; |
| 3822 | } while (!ToSplit.empty()); |
| 3823 | if (Changed) { |
| 3824 | if (MD) |
| 3825 | MD->invalidateCachedPredecessors(); |
| 3826 | InvalidBlockRPONumbers = true; |
| 3827 | } |
| 3828 | return Changed; |
| 3829 | } |
| 3830 | |
| 3831 | /// Executes one iteration of GVN. |
| 3832 | bool GVNPass::iterateOnFunction(Function &F) { |
| 3833 | cleanupGlobalSets(); |
| 3834 | |
| 3835 | // Top-down walk of the dominator tree. |
| 3836 | bool Changed = false; |
| 3837 | // Needed for value numbering with phi construction to work. |
| 3838 | // RPOT walks the graph in its constructor and will not be invalidated during |
| 3839 | // processBlock. |
| 3840 | ReversePostOrderTraversal<Function *> RPOT(&F); |
| 3841 | |
| 3842 | for (BasicBlock *BB : RPOT) |
| 3843 | Changed |= processBlock(BB); |
| 3844 | |
| 3845 | return Changed; |
| 3846 | } |
| 3847 | |
| 3848 | void GVNPass::cleanupGlobalSets() { |
| 3849 | VN.clear(); |
| 3850 | LeaderTable.clear(); |
| 3851 | BlockRPONumber.clear(); |
| 3852 | ICF->clear(); |
| 3853 | InvalidBlockRPONumbers = true; |
| 3854 | } |
| 3855 | |
| 3856 | void GVNPass::removeInstruction(Instruction *I) { |
| 3857 | VN.erase(V: I); |
| 3858 | if (MD) MD->removeInstruction(InstToRemove: I); |
| 3859 | if (MSSAU) |
| 3860 | MSSAU->removeMemoryAccess(I); |
| 3861 | #ifndef NDEBUG |
| 3862 | verifyRemoved(I); |
| 3863 | #endif |
| 3864 | ICF->removeInstruction(Inst: I); |
| 3865 | I->eraseFromParent(); |
| 3866 | ++NumGVNInstr; |
| 3867 | } |
| 3868 | |
| 3869 | /// Verify that the specified instruction does not occur in our |
| 3870 | /// internal data structures. |
| 3871 | void GVNPass::verifyRemoved(const Instruction *Inst) const { |
| 3872 | VN.verifyRemoved(V: Inst); |
| 3873 | } |
| 3874 | |
| 3875 | /// BB is declared dead, which implied other blocks become dead as well. This |
| 3876 | /// function is to add all these blocks to "DeadBlocks". For the dead blocks' |
| 3877 | /// live successors, update their phi nodes by replacing the operands |
| 3878 | /// corresponding to dead blocks with UndefVal. |
| 3879 | void GVNPass::addDeadBlock(BasicBlock *BB) { |
| 3880 | SmallVector<BasicBlock *, 4> NewDead; |
| 3881 | SmallSetVector<BasicBlock *, 4> DF; |
| 3882 | |
| 3883 | NewDead.push_back(Elt: BB); |
| 3884 | while (!NewDead.empty()) { |
| 3885 | BasicBlock *D = NewDead.pop_back_val(); |
| 3886 | if (DeadBlocks.count(key: D)) |
| 3887 | continue; |
| 3888 | |
| 3889 | // All blocks dominated by D are dead. |
| 3890 | SmallVector<BasicBlock *, 8> Dom; |
| 3891 | DT->getDescendants(R: D, Result&: Dom); |
| 3892 | DeadBlocks.insert_range(R&: Dom); |
| 3893 | |
| 3894 | // Figure out the dominance-frontier(D). |
| 3895 | for (BasicBlock *B : Dom) { |
| 3896 | for (BasicBlock *S : successors(BB: B)) { |
| 3897 | if (DeadBlocks.count(key: S)) |
| 3898 | continue; |
| 3899 | |
| 3900 | bool AllPredDead = true; |
| 3901 | for (BasicBlock *P : predecessors(BB: S)) |
| 3902 | if (!DeadBlocks.count(key: P)) { |
| 3903 | AllPredDead = false; |
| 3904 | break; |
| 3905 | } |
| 3906 | |
| 3907 | if (!AllPredDead) { |
| 3908 | // S could be proved dead later on. That is why we don't update phi |
| 3909 | // operands at this moment. |
| 3910 | DF.insert(X: S); |
| 3911 | } else { |
| 3912 | // While S is not dominated by D, it is dead by now. This could take |
| 3913 | // place if S already have a dead predecessor before D is declared |
| 3914 | // dead. |
| 3915 | NewDead.push_back(Elt: S); |
| 3916 | } |
| 3917 | } |
| 3918 | } |
| 3919 | } |
| 3920 | |
| 3921 | // For the dead blocks' live successors, update their phi nodes by replacing |
| 3922 | // the operands corresponding to dead blocks with UndefVal. |
| 3923 | for (BasicBlock *B : DF) { |
| 3924 | if (DeadBlocks.count(key: B)) |
| 3925 | continue; |
| 3926 | |
| 3927 | // First, split the critical edges. This might also create additional blocks |
| 3928 | // to preserve LoopSimplify form and adjust edges accordingly. |
| 3929 | SmallVector<BasicBlock *, 4> Preds(predecessors(BB: B)); |
| 3930 | for (BasicBlock *P : Preds) { |
| 3931 | if (!DeadBlocks.count(key: P)) |
| 3932 | continue; |
| 3933 | |
| 3934 | if (is_contained(Range: successors(BB: P), Element: B) && |
| 3935 | isCriticalEdge(TI: P->getTerminator(), Succ: B)) { |
| 3936 | if (BasicBlock *S = splitCriticalEdges(Pred: P, Succ: B)) |
| 3937 | DeadBlocks.insert(X: P = S); |
| 3938 | } |
| 3939 | } |
| 3940 | |
| 3941 | // Now poison the incoming values from the dead predecessors. |
| 3942 | for (BasicBlock *P : predecessors(BB: B)) { |
| 3943 | if (!DeadBlocks.count(key: P)) |
| 3944 | continue; |
| 3945 | for (PHINode &Phi : B->phis()) { |
| 3946 | Phi.setIncomingValueForBlock(BB: P, V: PoisonValue::get(T: Phi.getType())); |
| 3947 | if (MD) |
| 3948 | MD->invalidateCachedPointerInfo(Ptr: &Phi); |
| 3949 | } |
| 3950 | } |
| 3951 | } |
| 3952 | } |
| 3953 | |
| 3954 | // If the given branch is recognized as a foldable branch (i.e. conditional |
| 3955 | // branch with constant condition), it will perform following analyses and |
| 3956 | // transformation. |
| 3957 | // 1) If the dead out-coming edge is a critical-edge, split it. Let |
| 3958 | // R be the target of the dead out-coming edge. |
| 3959 | // 1) Identify the set of dead blocks implied by the branch's dead outcoming |
| 3960 | // edge. The result of this step will be {X| X is dominated by R} |
| 3961 | // 2) Identify those blocks which haves at least one dead predecessor. The |
| 3962 | // result of this step will be dominance-frontier(R). |
| 3963 | // 3) Update the PHIs in DF(R) by replacing the operands corresponding to |
| 3964 | // dead blocks with "UndefVal" in an hope these PHIs will optimized away. |
| 3965 | // |
| 3966 | // Return true iff *NEW* dead code are found. |
| 3967 | bool GVNPass::processFoldableCondBr(CondBrInst *BI) { |
| 3968 | // If a branch has two identical successors, we cannot declare either dead. |
| 3969 | if (BI->getSuccessor(i: 0) == BI->getSuccessor(i: 1)) |
| 3970 | return false; |
| 3971 | |
| 3972 | ConstantInt *Cond = dyn_cast<ConstantInt>(Val: BI->getCondition()); |
| 3973 | if (!Cond) |
| 3974 | return false; |
| 3975 | |
| 3976 | BasicBlock *DeadRoot = |
| 3977 | Cond->getZExtValue() ? BI->getSuccessor(i: 1) : BI->getSuccessor(i: 0); |
| 3978 | if (DeadBlocks.count(key: DeadRoot)) |
| 3979 | return false; |
| 3980 | |
| 3981 | if (!DeadRoot->getSinglePredecessor()) |
| 3982 | DeadRoot = splitCriticalEdges(Pred: BI->getParent(), Succ: DeadRoot); |
| 3983 | |
| 3984 | addDeadBlock(BB: DeadRoot); |
| 3985 | return true; |
| 3986 | } |
| 3987 | |
| 3988 | // performPRE() will trigger assert if it comes across an instruction without |
| 3989 | // associated val-num. As it normally has far more live instructions than dead |
| 3990 | // instructions, it makes more sense just to "fabricate" a val-number for the |
| 3991 | // dead code than checking if instruction involved is dead or not. |
| 3992 | void GVNPass::assignValNumForDeadCode() { |
| 3993 | for (BasicBlock *BB : DeadBlocks) { |
| 3994 | for (Instruction &Inst : *BB) { |
| 3995 | unsigned ValNum = VN.lookupOrAdd(V: &Inst); |
| 3996 | LeaderTable.insert(N: ValNum, V: &Inst, BB); |
| 3997 | } |
| 3998 | } |
| 3999 | } |
| 4000 | |
| 4001 | class llvm::GVNLegacyPass : public FunctionPass { |
| 4002 | public: |
| 4003 | static char ID; // Pass identification, replacement for typeid. |
| 4004 | |
| 4005 | explicit GVNLegacyPass(bool MemDepAnalysis = GVNEnableMemDep, |
| 4006 | bool MemSSAAnalysis = GVNEnableMemorySSA, |
| 4007 | bool ScalarPRE = true) |
| 4008 | : FunctionPass(ID), Impl(GVNOptions() |
| 4009 | .setMemDep(MemDepAnalysis) |
| 4010 | .setMemorySSA(MemSSAAnalysis) |
| 4011 | .setScalarPRE(ScalarPRE)) { |
| 4012 | initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry()); |
| 4013 | } |
| 4014 | |
| 4015 | bool runOnFunction(Function &F) override { |
| 4016 | if (skipFunction(F)) |
| 4017 | return false; |
| 4018 | |
| 4019 | auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
| 4020 | if (Impl.isMemorySSAEnabled() && !MSSAWP) |
| 4021 | MSSAWP = &getAnalysis<MemorySSAWrapperPass>(); |
| 4022 | |
| 4023 | return Impl.runImpl( |
| 4024 | F, RunAC&: getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F), |
| 4025 | RunDT&: getAnalysis<DominatorTreeWrapperPass>().getDomTree(), |
| 4026 | RunTLI: getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F), |
| 4027 | RunAA&: getAnalysis<AAResultsWrapperPass>().getAAResults(), |
| 4028 | RunMD: Impl.isMemDepEnabled() |
| 4029 | ? &getAnalysis<MemoryDependenceWrapperPass>().getMemDep() |
| 4030 | : nullptr, |
| 4031 | LI&: getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), |
| 4032 | RunORE: &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(), |
| 4033 | MSSA: MSSAWP ? &MSSAWP->getMSSA() : nullptr); |
| 4034 | } |
| 4035 | |
| 4036 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 4037 | AU.addRequired<AssumptionCacheTracker>(); |
| 4038 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 4039 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 4040 | AU.addRequired<LoopInfoWrapperPass>(); |
| 4041 | if (Impl.isMemDepEnabled()) |
| 4042 | AU.addRequired<MemoryDependenceWrapperPass>(); |
| 4043 | AU.addRequired<AAResultsWrapperPass>(); |
| 4044 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 4045 | AU.addPreserved<GlobalsAAWrapperPass>(); |
| 4046 | AU.addPreserved<TargetLibraryInfoWrapperPass>(); |
| 4047 | AU.addPreserved<LoopInfoWrapperPass>(); |
| 4048 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); |
| 4049 | AU.addPreserved<MemorySSAWrapperPass>(); |
| 4050 | if (Impl.isMemorySSAEnabled()) |
| 4051 | AU.addRequired<MemorySSAWrapperPass>(); |
| 4052 | } |
| 4053 | |
| 4054 | private: |
| 4055 | GVNPass Impl; |
| 4056 | }; |
| 4057 | |
| 4058 | char GVNLegacyPass::ID = 0; |
| 4059 | |
| 4060 | INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn" , "Global Value Numbering" , false, false) |
| 4061 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) |
| 4062 | INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass) |
| 4063 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) |
| 4064 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
| 4065 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
| 4066 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
| 4067 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) |
| 4068 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) |
| 4069 | INITIALIZE_PASS_END(GVNLegacyPass, "gvn" , "Global Value Numbering" , false, false) |
| 4070 | |
| 4071 | // The public interface to this file... |
| 4072 | FunctionPass *llvm::createGVNPass() { return new GVNLegacyPass(); } |
| 4073 | FunctionPass *llvm::createGVNPass(bool ScalarPRE) { |
| 4074 | return new GVNLegacyPass(GVNEnableMemDep, GVNEnableMemorySSA, ScalarPRE); |
| 4075 | } |
| 4076 | |