| File: | build/source/llvm/lib/Transforms/Scalar/GVN.cpp |
| Warning: | line 3095, column 29 Although the value stored to 'P' is used in the enclosing expression, the value is never actually read from 'P' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 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/LoopInfo.h" |
| 37 | #include "llvm/Analysis/MemoryBuiltins.h" |
| 38 | #include "llvm/Analysis/MemoryDependenceAnalysis.h" |
| 39 | #include "llvm/Analysis/MemorySSA.h" |
| 40 | #include "llvm/Analysis/MemorySSAUpdater.h" |
| 41 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 42 | #include "llvm/Analysis/PHITransAddr.h" |
| 43 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 44 | #include "llvm/Analysis/ValueTracking.h" |
| 45 | #include "llvm/IR/Attributes.h" |
| 46 | #include "llvm/IR/BasicBlock.h" |
| 47 | #include "llvm/IR/Constant.h" |
| 48 | #include "llvm/IR/Constants.h" |
| 49 | #include "llvm/IR/DebugLoc.h" |
| 50 | #include "llvm/IR/Dominators.h" |
| 51 | #include "llvm/IR/Function.h" |
| 52 | #include "llvm/IR/InstrTypes.h" |
| 53 | #include "llvm/IR/Instruction.h" |
| 54 | #include "llvm/IR/Instructions.h" |
| 55 | #include "llvm/IR/IntrinsicInst.h" |
| 56 | #include "llvm/IR/LLVMContext.h" |
| 57 | #include "llvm/IR/Metadata.h" |
| 58 | #include "llvm/IR/Module.h" |
| 59 | #include "llvm/IR/PassManager.h" |
| 60 | #include "llvm/IR/PatternMatch.h" |
| 61 | #include "llvm/IR/Type.h" |
| 62 | #include "llvm/IR/Use.h" |
| 63 | #include "llvm/IR/Value.h" |
| 64 | #include "llvm/InitializePasses.h" |
| 65 | #include "llvm/Pass.h" |
| 66 | #include "llvm/Support/Casting.h" |
| 67 | #include "llvm/Support/CommandLine.h" |
| 68 | #include "llvm/Support/Compiler.h" |
| 69 | #include "llvm/Support/Debug.h" |
| 70 | #include "llvm/Support/raw_ostream.h" |
| 71 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
| 72 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 73 | #include "llvm/Transforms/Utils/Local.h" |
| 74 | #include "llvm/Transforms/Utils/SSAUpdater.h" |
| 75 | #include "llvm/Transforms/Utils/VNCoercion.h" |
| 76 | #include <algorithm> |
| 77 | #include <cassert> |
| 78 | #include <cstdint> |
| 79 | #include <optional> |
| 80 | #include <utility> |
| 81 | |
| 82 | using namespace llvm; |
| 83 | using namespace llvm::gvn; |
| 84 | using namespace llvm::VNCoercion; |
| 85 | using namespace PatternMatch; |
| 86 | |
| 87 | #define DEBUG_TYPE"gvn" "gvn" |
| 88 | |
| 89 | STATISTIC(NumGVNInstr, "Number of instructions deleted")static llvm::Statistic NumGVNInstr = {"gvn", "NumGVNInstr", "Number of instructions deleted" }; |
| 90 | STATISTIC(NumGVNLoad, "Number of loads deleted")static llvm::Statistic NumGVNLoad = {"gvn", "NumGVNLoad", "Number of loads deleted" }; |
| 91 | STATISTIC(NumGVNPRE, "Number of instructions PRE'd")static llvm::Statistic NumGVNPRE = {"gvn", "NumGVNPRE", "Number of instructions PRE'd" }; |
| 92 | STATISTIC(NumGVNBlocks, "Number of blocks merged")static llvm::Statistic NumGVNBlocks = {"gvn", "NumGVNBlocks", "Number of blocks merged"}; |
| 93 | STATISTIC(NumGVNSimpl, "Number of instructions simplified")static llvm::Statistic NumGVNSimpl = {"gvn", "NumGVNSimpl", "Number of instructions simplified" }; |
| 94 | STATISTIC(NumGVNEqProp, "Number of equalities propagated")static llvm::Statistic NumGVNEqProp = {"gvn", "NumGVNEqProp", "Number of equalities propagated"}; |
| 95 | STATISTIC(NumPRELoad, "Number of loads PRE'd")static llvm::Statistic NumPRELoad = {"gvn", "NumPRELoad", "Number of loads PRE'd" }; |
| 96 | STATISTIC(NumPRELoopLoad, "Number of loop loads PRE'd")static llvm::Statistic NumPRELoopLoad = {"gvn", "NumPRELoopLoad" , "Number of loop loads PRE'd"}; |
| 97 | |
| 98 | STATISTIC(IsValueFullyAvailableInBlockNumSpeculationsMax,static llvm::Statistic IsValueFullyAvailableInBlockNumSpeculationsMax = {"gvn", "IsValueFullyAvailableInBlockNumSpeculationsMax", "Number of blocks speculated as available in " "IsValueFullyAvailableInBlock(), max"} |
| 99 | "Number of blocks speculated as available in "static llvm::Statistic IsValueFullyAvailableInBlockNumSpeculationsMax = {"gvn", "IsValueFullyAvailableInBlockNumSpeculationsMax", "Number of blocks speculated as available in " "IsValueFullyAvailableInBlock(), max"} |
| 100 | "IsValueFullyAvailableInBlock(), max")static llvm::Statistic IsValueFullyAvailableInBlockNumSpeculationsMax = {"gvn", "IsValueFullyAvailableInBlockNumSpeculationsMax", "Number of blocks speculated as available in " "IsValueFullyAvailableInBlock(), max"}; |
| 101 | STATISTIC(MaxBBSpeculationCutoffReachedTimes,static llvm::Statistic MaxBBSpeculationCutoffReachedTimes = { "gvn", "MaxBBSpeculationCutoffReachedTimes", "Number of times we we reached gvn-max-block-speculations cut-off " "preventing further exploration"} |
| 102 | "Number of times we we reached gvn-max-block-speculations cut-off "static llvm::Statistic MaxBBSpeculationCutoffReachedTimes = { "gvn", "MaxBBSpeculationCutoffReachedTimes", "Number of times we we reached gvn-max-block-speculations cut-off " "preventing further exploration"} |
| 103 | "preventing further exploration")static llvm::Statistic MaxBBSpeculationCutoffReachedTimes = { "gvn", "MaxBBSpeculationCutoffReachedTimes", "Number of times we we reached gvn-max-block-speculations cut-off " "preventing further exploration"}; |
| 104 | |
| 105 | static cl::opt<bool> GVNEnablePRE("enable-pre", cl::init(true), cl::Hidden); |
| 106 | static cl::opt<bool> GVNEnableLoadPRE("enable-load-pre", cl::init(true)); |
| 107 | static cl::opt<bool> GVNEnableLoadInLoopPRE("enable-load-in-loop-pre", |
| 108 | cl::init(true)); |
| 109 | static cl::opt<bool> |
| 110 | GVNEnableSplitBackedgeInLoadPRE("enable-split-backedge-in-load-pre", |
| 111 | cl::init(false)); |
| 112 | static cl::opt<bool> GVNEnableMemDep("enable-gvn-memdep", cl::init(true)); |
| 113 | |
| 114 | static cl::opt<uint32_t> MaxNumDeps( |
| 115 | "gvn-max-num-deps", cl::Hidden, cl::init(100), |
| 116 | cl::desc("Max number of dependences to attempt Load PRE (default = 100)")); |
| 117 | |
| 118 | // This is based on IsValueFullyAvailableInBlockNumSpeculationsMax stat. |
| 119 | static cl::opt<uint32_t> MaxBBSpeculations( |
| 120 | "gvn-max-block-speculations", cl::Hidden, cl::init(600), |
| 121 | cl::desc("Max number of blocks we're willing to speculate on (and recurse " |
| 122 | "into) when deducing if a value is fully available or not in GVN " |
| 123 | "(default = 600)")); |
| 124 | |
| 125 | static cl::opt<uint32_t> MaxNumVisitedInsts( |
| 126 | "gvn-max-num-visited-insts", cl::Hidden, cl::init(100), |
| 127 | cl::desc("Max number of visited instructions when trying to find " |
| 128 | "dominating value of select dependency (default = 100)")); |
| 129 | |
| 130 | struct llvm::GVNPass::Expression { |
| 131 | uint32_t opcode; |
| 132 | bool commutative = false; |
| 133 | // The type is not necessarily the result type of the expression, it may be |
| 134 | // any additional type needed to disambiguate the expression. |
| 135 | Type *type = nullptr; |
| 136 | SmallVector<uint32_t, 4> varargs; |
| 137 | |
| 138 | Expression(uint32_t o = ~2U) : opcode(o) {} |
| 139 | |
| 140 | bool operator==(const Expression &other) const { |
| 141 | if (opcode != other.opcode) |
| 142 | return false; |
| 143 | if (opcode == ~0U || opcode == ~1U) |
| 144 | return true; |
| 145 | if (type != other.type) |
| 146 | return false; |
| 147 | if (varargs != other.varargs) |
| 148 | return false; |
| 149 | return true; |
| 150 | } |
| 151 | |
| 152 | friend hash_code hash_value(const Expression &Value) { |
| 153 | return hash_combine( |
| 154 | Value.opcode, Value.type, |
| 155 | hash_combine_range(Value.varargs.begin(), Value.varargs.end())); |
| 156 | } |
| 157 | }; |
| 158 | |
| 159 | namespace llvm { |
| 160 | |
| 161 | template <> struct DenseMapInfo<GVNPass::Expression> { |
| 162 | static inline GVNPass::Expression getEmptyKey() { return ~0U; } |
| 163 | static inline GVNPass::Expression getTombstoneKey() { return ~1U; } |
| 164 | |
| 165 | static unsigned getHashValue(const GVNPass::Expression &e) { |
| 166 | using llvm::hash_value; |
| 167 | |
| 168 | return static_cast<unsigned>(hash_value(e)); |
| 169 | } |
| 170 | |
| 171 | static bool isEqual(const GVNPass::Expression &LHS, |
| 172 | const GVNPass::Expression &RHS) { |
| 173 | return LHS == RHS; |
| 174 | } |
| 175 | }; |
| 176 | |
| 177 | } // end namespace llvm |
| 178 | |
| 179 | /// Represents a particular available value that we know how to materialize. |
| 180 | /// Materialization of an AvailableValue never fails. An AvailableValue is |
| 181 | /// implicitly associated with a rematerialization point which is the |
| 182 | /// location of the instruction from which it was formed. |
| 183 | struct llvm::gvn::AvailableValue { |
| 184 | enum class ValType { |
| 185 | SimpleVal, // A simple offsetted value that is accessed. |
| 186 | LoadVal, // A value produced by a load. |
| 187 | MemIntrin, // A memory intrinsic which is loaded from. |
| 188 | UndefVal, // A UndefValue representing a value from dead block (which |
| 189 | // is not yet physically removed from the CFG). |
| 190 | SelectVal, // A pointer select which is loaded from and for which the load |
| 191 | // can be replace by a value select. |
| 192 | }; |
| 193 | |
| 194 | /// Val - The value that is live out of the block. |
| 195 | Value *Val; |
| 196 | /// Kind of the live-out value. |
| 197 | ValType Kind; |
| 198 | |
| 199 | /// Offset - The byte offset in Val that is interesting for the load query. |
| 200 | unsigned Offset = 0; |
| 201 | /// V1, V2 - The dominating non-clobbered values of SelectVal. |
| 202 | Value *V1 = nullptr, *V2 = nullptr; |
| 203 | |
| 204 | static AvailableValue get(Value *V, unsigned Offset = 0) { |
| 205 | AvailableValue Res; |
| 206 | Res.Val = V; |
| 207 | Res.Kind = ValType::SimpleVal; |
| 208 | Res.Offset = Offset; |
| 209 | return Res; |
| 210 | } |
| 211 | |
| 212 | static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) { |
| 213 | AvailableValue Res; |
| 214 | Res.Val = MI; |
| 215 | Res.Kind = ValType::MemIntrin; |
| 216 | Res.Offset = Offset; |
| 217 | return Res; |
| 218 | } |
| 219 | |
| 220 | static AvailableValue getLoad(LoadInst *Load, unsigned Offset = 0) { |
| 221 | AvailableValue Res; |
| 222 | Res.Val = Load; |
| 223 | Res.Kind = ValType::LoadVal; |
| 224 | Res.Offset = Offset; |
| 225 | return Res; |
| 226 | } |
| 227 | |
| 228 | static AvailableValue getUndef() { |
| 229 | AvailableValue Res; |
| 230 | Res.Val = nullptr; |
| 231 | Res.Kind = ValType::UndefVal; |
| 232 | Res.Offset = 0; |
| 233 | return Res; |
| 234 | } |
| 235 | |
| 236 | static AvailableValue getSelect(SelectInst *Sel, Value *V1, Value *V2) { |
| 237 | AvailableValue Res; |
| 238 | Res.Val = Sel; |
| 239 | Res.Kind = ValType::SelectVal; |
| 240 | Res.Offset = 0; |
| 241 | Res.V1 = V1; |
| 242 | Res.V2 = V2; |
| 243 | return Res; |
| 244 | } |
| 245 | |
| 246 | bool isSimpleValue() const { return Kind == ValType::SimpleVal; } |
| 247 | bool isCoercedLoadValue() const { return Kind == ValType::LoadVal; } |
| 248 | bool isMemIntrinValue() const { return Kind == ValType::MemIntrin; } |
| 249 | bool isUndefValue() const { return Kind == ValType::UndefVal; } |
| 250 | bool isSelectValue() const { return Kind == ValType::SelectVal; } |
| 251 | |
| 252 | Value *getSimpleValue() const { |
| 253 | assert(isSimpleValue() && "Wrong accessor")(static_cast <bool> (isSimpleValue() && "Wrong accessor" ) ? void (0) : __assert_fail ("isSimpleValue() && \"Wrong accessor\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 253, __extension__ __PRETTY_FUNCTION__ )); |
| 254 | return Val; |
| 255 | } |
| 256 | |
| 257 | LoadInst *getCoercedLoadValue() const { |
| 258 | assert(isCoercedLoadValue() && "Wrong accessor")(static_cast <bool> (isCoercedLoadValue() && "Wrong accessor" ) ? void (0) : __assert_fail ("isCoercedLoadValue() && \"Wrong accessor\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 258, __extension__ __PRETTY_FUNCTION__ )); |
| 259 | return cast<LoadInst>(Val); |
| 260 | } |
| 261 | |
| 262 | MemIntrinsic *getMemIntrinValue() const { |
| 263 | assert(isMemIntrinValue() && "Wrong accessor")(static_cast <bool> (isMemIntrinValue() && "Wrong accessor" ) ? void (0) : __assert_fail ("isMemIntrinValue() && \"Wrong accessor\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 263, __extension__ __PRETTY_FUNCTION__ )); |
| 264 | return cast<MemIntrinsic>(Val); |
| 265 | } |
| 266 | |
| 267 | SelectInst *getSelectValue() const { |
| 268 | assert(isSelectValue() && "Wrong accessor")(static_cast <bool> (isSelectValue() && "Wrong accessor" ) ? void (0) : __assert_fail ("isSelectValue() && \"Wrong accessor\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 268, __extension__ __PRETTY_FUNCTION__ )); |
| 269 | return cast<SelectInst>(Val); |
| 270 | } |
| 271 | |
| 272 | /// Emit code at the specified insertion point to adjust the value defined |
| 273 | /// here to the specified type. This handles various coercion cases. |
| 274 | Value *MaterializeAdjustedValue(LoadInst *Load, Instruction *InsertPt, |
| 275 | GVNPass &gvn) const; |
| 276 | }; |
| 277 | |
| 278 | /// Represents an AvailableValue which can be rematerialized at the end of |
| 279 | /// the associated BasicBlock. |
| 280 | struct llvm::gvn::AvailableValueInBlock { |
| 281 | /// BB - The basic block in question. |
| 282 | BasicBlock *BB = nullptr; |
| 283 | |
| 284 | /// AV - The actual available value |
| 285 | AvailableValue AV; |
| 286 | |
| 287 | static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) { |
| 288 | AvailableValueInBlock Res; |
| 289 | Res.BB = BB; |
| 290 | Res.AV = std::move(AV); |
| 291 | return Res; |
| 292 | } |
| 293 | |
| 294 | static AvailableValueInBlock get(BasicBlock *BB, Value *V, |
| 295 | unsigned Offset = 0) { |
| 296 | return get(BB, AvailableValue::get(V, Offset)); |
| 297 | } |
| 298 | |
| 299 | static AvailableValueInBlock getUndef(BasicBlock *BB) { |
| 300 | return get(BB, AvailableValue::getUndef()); |
| 301 | } |
| 302 | |
| 303 | static AvailableValueInBlock getSelect(BasicBlock *BB, SelectInst *Sel, |
| 304 | Value *V1, Value *V2) { |
| 305 | return get(BB, AvailableValue::getSelect(Sel, V1, V2)); |
| 306 | } |
| 307 | |
| 308 | /// Emit code at the end of this block to adjust the value defined here to |
| 309 | /// the specified type. This handles various coercion cases. |
| 310 | Value *MaterializeAdjustedValue(LoadInst *Load, GVNPass &gvn) const { |
| 311 | return AV.MaterializeAdjustedValue(Load, BB->getTerminator(), gvn); |
| 312 | } |
| 313 | }; |
| 314 | |
| 315 | //===----------------------------------------------------------------------===// |
| 316 | // ValueTable Internal Functions |
| 317 | //===----------------------------------------------------------------------===// |
| 318 | |
| 319 | GVNPass::Expression GVNPass::ValueTable::createExpr(Instruction *I) { |
| 320 | Expression e; |
| 321 | e.type = I->getType(); |
| 322 | e.opcode = I->getOpcode(); |
| 323 | if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(I)) { |
| 324 | // gc.relocate is 'special' call: its second and third operands are |
| 325 | // not real values, but indices into statepoint's argument list. |
| 326 | // Use the refered to values for purposes of identity. |
| 327 | e.varargs.push_back(lookupOrAdd(GCR->getOperand(0))); |
| 328 | e.varargs.push_back(lookupOrAdd(GCR->getBasePtr())); |
| 329 | e.varargs.push_back(lookupOrAdd(GCR->getDerivedPtr())); |
| 330 | } else { |
| 331 | for (Use &Op : I->operands()) |
| 332 | e.varargs.push_back(lookupOrAdd(Op)); |
| 333 | } |
| 334 | if (I->isCommutative()) { |
| 335 | // Ensure that commutative instructions that only differ by a permutation |
| 336 | // of their operands get the same value number by sorting the operand value |
| 337 | // numbers. Since commutative operands are the 1st two operands it is more |
| 338 | // efficient to sort by hand rather than using, say, std::sort. |
| 339 | assert(I->getNumOperands() >= 2 && "Unsupported commutative instruction!")(static_cast <bool> (I->getNumOperands() >= 2 && "Unsupported commutative instruction!") ? void (0) : __assert_fail ("I->getNumOperands() >= 2 && \"Unsupported commutative instruction!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 339, __extension__ __PRETTY_FUNCTION__ )); |
| 340 | if (e.varargs[0] > e.varargs[1]) |
| 341 | std::swap(e.varargs[0], e.varargs[1]); |
| 342 | e.commutative = true; |
| 343 | } |
| 344 | |
| 345 | if (auto *C = dyn_cast<CmpInst>(I)) { |
| 346 | // Sort the operand value numbers so x<y and y>x get the same value number. |
| 347 | CmpInst::Predicate Predicate = C->getPredicate(); |
| 348 | if (e.varargs[0] > e.varargs[1]) { |
| 349 | std::swap(e.varargs[0], e.varargs[1]); |
| 350 | Predicate = CmpInst::getSwappedPredicate(Predicate); |
| 351 | } |
| 352 | e.opcode = (C->getOpcode() << 8) | Predicate; |
| 353 | e.commutative = true; |
| 354 | } else if (auto *E = dyn_cast<InsertValueInst>(I)) { |
| 355 | e.varargs.append(E->idx_begin(), E->idx_end()); |
| 356 | } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(I)) { |
| 357 | ArrayRef<int> ShuffleMask = SVI->getShuffleMask(); |
| 358 | e.varargs.append(ShuffleMask.begin(), ShuffleMask.end()); |
| 359 | } |
| 360 | |
| 361 | return e; |
| 362 | } |
| 363 | |
| 364 | GVNPass::Expression GVNPass::ValueTable::createCmpExpr( |
| 365 | unsigned Opcode, CmpInst::Predicate Predicate, Value *LHS, Value *RHS) { |
| 366 | assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&(static_cast <bool> ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && "Not a comparison!") ? void (0) : __assert_fail ("(Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && \"Not a comparison!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 367, __extension__ __PRETTY_FUNCTION__ )) |
| 367 | "Not a comparison!")(static_cast <bool> ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && "Not a comparison!") ? void (0) : __assert_fail ("(Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && \"Not a comparison!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 367, __extension__ __PRETTY_FUNCTION__ )); |
| 368 | Expression e; |
| 369 | e.type = CmpInst::makeCmpResultType(LHS->getType()); |
| 370 | e.varargs.push_back(lookupOrAdd(LHS)); |
| 371 | e.varargs.push_back(lookupOrAdd(RHS)); |
| 372 | |
| 373 | // Sort the operand value numbers so x<y and y>x get the same value number. |
| 374 | if (e.varargs[0] > e.varargs[1]) { |
| 375 | std::swap(e.varargs[0], e.varargs[1]); |
| 376 | Predicate = CmpInst::getSwappedPredicate(Predicate); |
| 377 | } |
| 378 | e.opcode = (Opcode << 8) | Predicate; |
| 379 | e.commutative = true; |
| 380 | return e; |
| 381 | } |
| 382 | |
| 383 | GVNPass::Expression |
| 384 | GVNPass::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) { |
| 385 | assert(EI && "Not an ExtractValueInst?")(static_cast <bool> (EI && "Not an ExtractValueInst?" ) ? void (0) : __assert_fail ("EI && \"Not an ExtractValueInst?\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 385, __extension__ __PRETTY_FUNCTION__ )); |
| 386 | Expression e; |
| 387 | e.type = EI->getType(); |
| 388 | e.opcode = 0; |
| 389 | |
| 390 | WithOverflowInst *WO = dyn_cast<WithOverflowInst>(EI->getAggregateOperand()); |
| 391 | if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) { |
| 392 | // EI is an extract from one of our with.overflow intrinsics. Synthesize |
| 393 | // a semantically equivalent expression instead of an extract value |
| 394 | // expression. |
| 395 | e.opcode = WO->getBinaryOp(); |
| 396 | e.varargs.push_back(lookupOrAdd(WO->getLHS())); |
| 397 | e.varargs.push_back(lookupOrAdd(WO->getRHS())); |
| 398 | return e; |
| 399 | } |
| 400 | |
| 401 | // Not a recognised intrinsic. Fall back to producing an extract value |
| 402 | // expression. |
| 403 | e.opcode = EI->getOpcode(); |
| 404 | for (Use &Op : EI->operands()) |
| 405 | e.varargs.push_back(lookupOrAdd(Op)); |
| 406 | |
| 407 | append_range(e.varargs, EI->indices()); |
| 408 | |
| 409 | return e; |
| 410 | } |
| 411 | |
| 412 | GVNPass::Expression GVNPass::ValueTable::createGEPExpr(GetElementPtrInst *GEP) { |
| 413 | Expression E; |
| 414 | Type *PtrTy = GEP->getType()->getScalarType(); |
| 415 | const DataLayout &DL = GEP->getModule()->getDataLayout(); |
| 416 | unsigned BitWidth = DL.getIndexTypeSizeInBits(PtrTy); |
| 417 | MapVector<Value *, APInt> VariableOffsets; |
| 418 | APInt ConstantOffset(BitWidth, 0); |
| 419 | if (PtrTy->isOpaquePointerTy() && |
| 420 | GEP->collectOffset(DL, BitWidth, VariableOffsets, ConstantOffset)) { |
| 421 | // For opaque pointers, convert into offset representation, to recognize |
| 422 | // equivalent address calculations that use different type encoding. |
| 423 | LLVMContext &Context = GEP->getContext(); |
| 424 | E.opcode = GEP->getOpcode(); |
| 425 | E.type = nullptr; |
| 426 | E.varargs.push_back(lookupOrAdd(GEP->getPointerOperand())); |
| 427 | for (const auto &Pair : VariableOffsets) { |
| 428 | E.varargs.push_back(lookupOrAdd(Pair.first)); |
| 429 | E.varargs.push_back(lookupOrAdd(ConstantInt::get(Context, Pair.second))); |
| 430 | } |
| 431 | if (!ConstantOffset.isZero()) |
| 432 | E.varargs.push_back( |
| 433 | lookupOrAdd(ConstantInt::get(Context, ConstantOffset))); |
| 434 | } else { |
| 435 | // If converting to offset representation fails (for typed pointers and |
| 436 | // scalable vectors), fall back to type-based implementation: |
| 437 | E.opcode = GEP->getOpcode(); |
| 438 | E.type = GEP->getSourceElementType(); |
| 439 | for (Use &Op : GEP->operands()) |
| 440 | E.varargs.push_back(lookupOrAdd(Op)); |
| 441 | } |
| 442 | return E; |
| 443 | } |
| 444 | |
| 445 | //===----------------------------------------------------------------------===// |
| 446 | // ValueTable External Functions |
| 447 | //===----------------------------------------------------------------------===// |
| 448 | |
| 449 | GVNPass::ValueTable::ValueTable() = default; |
| 450 | GVNPass::ValueTable::ValueTable(const ValueTable &) = default; |
| 451 | GVNPass::ValueTable::ValueTable(ValueTable &&) = default; |
| 452 | GVNPass::ValueTable::~ValueTable() = default; |
| 453 | GVNPass::ValueTable & |
| 454 | GVNPass::ValueTable::operator=(const GVNPass::ValueTable &Arg) = default; |
| 455 | |
| 456 | /// add - Insert a value into the table with a specified value number. |
| 457 | void GVNPass::ValueTable::add(Value *V, uint32_t num) { |
| 458 | valueNumbering.insert(std::make_pair(V, num)); |
| 459 | if (PHINode *PN = dyn_cast<PHINode>(V)) |
| 460 | NumberingPhi[num] = PN; |
| 461 | } |
| 462 | |
| 463 | uint32_t GVNPass::ValueTable::lookupOrAddCall(CallInst *C) { |
| 464 | if (AA->doesNotAccessMemory(C) && |
| 465 | // FIXME: Currently the calls which may access the thread id may |
| 466 | // be considered as not accessing the memory. But this is |
| 467 | // problematic for coroutines, since coroutines may resume in a |
| 468 | // different thread. So we disable the optimization here for the |
| 469 | // correctness. However, it may block many other correct |
| 470 | // optimizations. Revert this one when we detect the memory |
| 471 | // accessing kind more precisely. |
| 472 | !C->getFunction()->isPresplitCoroutine()) { |
| 473 | Expression exp = createExpr(C); |
| 474 | uint32_t e = assignExpNewValueNum(exp).first; |
| 475 | valueNumbering[C] = e; |
| 476 | return e; |
| 477 | } else if (MD && AA->onlyReadsMemory(C) && |
| 478 | // FIXME: Currently the calls which may access the thread id may |
| 479 | // be considered as not accessing the memory. But this is |
| 480 | // problematic for coroutines, since coroutines may resume in a |
| 481 | // different thread. So we disable the optimization here for the |
| 482 | // correctness. However, it may block many other correct |
| 483 | // optimizations. Revert this one when we detect the memory |
| 484 | // accessing kind more precisely. |
| 485 | !C->getFunction()->isPresplitCoroutine()) { |
| 486 | Expression exp = createExpr(C); |
| 487 | auto ValNum = assignExpNewValueNum(exp); |
| 488 | if (ValNum.second) { |
| 489 | valueNumbering[C] = ValNum.first; |
| 490 | return ValNum.first; |
| 491 | } |
| 492 | |
| 493 | MemDepResult local_dep = MD->getDependency(C); |
| 494 | |
| 495 | if (!local_dep.isDef() && !local_dep.isNonLocal()) { |
| 496 | valueNumbering[C] = nextValueNumber; |
| 497 | return nextValueNumber++; |
| 498 | } |
| 499 | |
| 500 | if (local_dep.isDef()) { |
| 501 | // For masked load/store intrinsics, the local_dep may actually be |
| 502 | // a normal load or store instruction. |
| 503 | CallInst *local_cdep = dyn_cast<CallInst>(local_dep.getInst()); |
| 504 | |
| 505 | if (!local_cdep || local_cdep->arg_size() != C->arg_size()) { |
| 506 | valueNumbering[C] = nextValueNumber; |
| 507 | return nextValueNumber++; |
| 508 | } |
| 509 | |
| 510 | for (unsigned i = 0, e = C->arg_size(); i < e; ++i) { |
| 511 | uint32_t c_vn = lookupOrAdd(C->getArgOperand(i)); |
| 512 | uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i)); |
| 513 | if (c_vn != cd_vn) { |
| 514 | valueNumbering[C] = nextValueNumber; |
| 515 | return nextValueNumber++; |
| 516 | } |
| 517 | } |
| 518 | |
| 519 | uint32_t v = lookupOrAdd(local_cdep); |
| 520 | valueNumbering[C] = v; |
| 521 | return v; |
| 522 | } |
| 523 | |
| 524 | // Non-local case. |
| 525 | const MemoryDependenceResults::NonLocalDepInfo &deps = |
| 526 | MD->getNonLocalCallDependency(C); |
| 527 | // FIXME: Move the checking logic to MemDep! |
| 528 | CallInst* cdep = nullptr; |
| 529 | |
| 530 | // Check to see if we have a single dominating call instruction that is |
| 531 | // identical to C. |
| 532 | for (const NonLocalDepEntry &I : deps) { |
| 533 | if (I.getResult().isNonLocal()) |
| 534 | continue; |
| 535 | |
| 536 | // We don't handle non-definitions. If we already have a call, reject |
| 537 | // instruction dependencies. |
| 538 | if (!I.getResult().isDef() || cdep != nullptr) { |
| 539 | cdep = nullptr; |
| 540 | break; |
| 541 | } |
| 542 | |
| 543 | CallInst *NonLocalDepCall = dyn_cast<CallInst>(I.getResult().getInst()); |
| 544 | // FIXME: All duplicated with non-local case. |
| 545 | if (NonLocalDepCall && DT->properlyDominates(I.getBB(), C->getParent())) { |
| 546 | cdep = NonLocalDepCall; |
| 547 | continue; |
| 548 | } |
| 549 | |
| 550 | cdep = nullptr; |
| 551 | break; |
| 552 | } |
| 553 | |
| 554 | if (!cdep) { |
| 555 | valueNumbering[C] = nextValueNumber; |
| 556 | return nextValueNumber++; |
| 557 | } |
| 558 | |
| 559 | if (cdep->arg_size() != C->arg_size()) { |
| 560 | valueNumbering[C] = nextValueNumber; |
| 561 | return nextValueNumber++; |
| 562 | } |
| 563 | for (unsigned i = 0, e = C->arg_size(); i < e; ++i) { |
| 564 | uint32_t c_vn = lookupOrAdd(C->getArgOperand(i)); |
| 565 | uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i)); |
| 566 | if (c_vn != cd_vn) { |
| 567 | valueNumbering[C] = nextValueNumber; |
| 568 | return nextValueNumber++; |
| 569 | } |
| 570 | } |
| 571 | |
| 572 | uint32_t v = lookupOrAdd(cdep); |
| 573 | valueNumbering[C] = v; |
| 574 | return v; |
| 575 | } else { |
| 576 | valueNumbering[C] = nextValueNumber; |
| 577 | return nextValueNumber++; |
| 578 | } |
| 579 | } |
| 580 | |
| 581 | /// Returns true if a value number exists for the specified value. |
| 582 | bool GVNPass::ValueTable::exists(Value *V) const { |
| 583 | return valueNumbering.count(V) != 0; |
| 584 | } |
| 585 | |
| 586 | /// lookup_or_add - Returns the value number for the specified value, assigning |
| 587 | /// it a new number if it did not have one before. |
| 588 | uint32_t GVNPass::ValueTable::lookupOrAdd(Value *V) { |
| 589 | DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); |
| 590 | if (VI != valueNumbering.end()) |
| 591 | return VI->second; |
| 592 | |
| 593 | auto *I = dyn_cast<Instruction>(V); |
| 594 | if (!I) { |
| 595 | valueNumbering[V] = nextValueNumber; |
| 596 | return nextValueNumber++; |
| 597 | } |
| 598 | |
| 599 | Expression exp; |
| 600 | switch (I->getOpcode()) { |
| 601 | case Instruction::Call: |
| 602 | return lookupOrAddCall(cast<CallInst>(I)); |
| 603 | case Instruction::FNeg: |
| 604 | case Instruction::Add: |
| 605 | case Instruction::FAdd: |
| 606 | case Instruction::Sub: |
| 607 | case Instruction::FSub: |
| 608 | case Instruction::Mul: |
| 609 | case Instruction::FMul: |
| 610 | case Instruction::UDiv: |
| 611 | case Instruction::SDiv: |
| 612 | case Instruction::FDiv: |
| 613 | case Instruction::URem: |
| 614 | case Instruction::SRem: |
| 615 | case Instruction::FRem: |
| 616 | case Instruction::Shl: |
| 617 | case Instruction::LShr: |
| 618 | case Instruction::AShr: |
| 619 | case Instruction::And: |
| 620 | case Instruction::Or: |
| 621 | case Instruction::Xor: |
| 622 | case Instruction::ICmp: |
| 623 | case Instruction::FCmp: |
| 624 | case Instruction::Trunc: |
| 625 | case Instruction::ZExt: |
| 626 | case Instruction::SExt: |
| 627 | case Instruction::FPToUI: |
| 628 | case Instruction::FPToSI: |
| 629 | case Instruction::UIToFP: |
| 630 | case Instruction::SIToFP: |
| 631 | case Instruction::FPTrunc: |
| 632 | case Instruction::FPExt: |
| 633 | case Instruction::PtrToInt: |
| 634 | case Instruction::IntToPtr: |
| 635 | case Instruction::AddrSpaceCast: |
| 636 | case Instruction::BitCast: |
| 637 | case Instruction::Select: |
| 638 | case Instruction::Freeze: |
| 639 | case Instruction::ExtractElement: |
| 640 | case Instruction::InsertElement: |
| 641 | case Instruction::ShuffleVector: |
| 642 | case Instruction::InsertValue: |
| 643 | exp = createExpr(I); |
| 644 | break; |
| 645 | case Instruction::GetElementPtr: |
| 646 | exp = createGEPExpr(cast<GetElementPtrInst>(I)); |
| 647 | break; |
| 648 | case Instruction::ExtractValue: |
| 649 | exp = createExtractvalueExpr(cast<ExtractValueInst>(I)); |
| 650 | break; |
| 651 | case Instruction::PHI: |
| 652 | valueNumbering[V] = nextValueNumber; |
| 653 | NumberingPhi[nextValueNumber] = cast<PHINode>(V); |
| 654 | return nextValueNumber++; |
| 655 | default: |
| 656 | valueNumbering[V] = nextValueNumber; |
| 657 | return nextValueNumber++; |
| 658 | } |
| 659 | |
| 660 | uint32_t e = assignExpNewValueNum(exp).first; |
| 661 | valueNumbering[V] = e; |
| 662 | return e; |
| 663 | } |
| 664 | |
| 665 | /// Returns the value number of the specified value. Fails if |
| 666 | /// the value has not yet been numbered. |
| 667 | uint32_t GVNPass::ValueTable::lookup(Value *V, bool Verify) const { |
| 668 | DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); |
| 669 | if (Verify) { |
| 670 | assert(VI != valueNumbering.end() && "Value not numbered?")(static_cast <bool> (VI != valueNumbering.end() && "Value not numbered?") ? void (0) : __assert_fail ("VI != valueNumbering.end() && \"Value not numbered?\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 670, __extension__ __PRETTY_FUNCTION__ )); |
| 671 | return VI->second; |
| 672 | } |
| 673 | return (VI != valueNumbering.end()) ? VI->second : 0; |
| 674 | } |
| 675 | |
| 676 | /// Returns the value number of the given comparison, |
| 677 | /// assigning it a new number if it did not have one before. Useful when |
| 678 | /// we deduced the result of a comparison, but don't immediately have an |
| 679 | /// instruction realizing that comparison to hand. |
| 680 | uint32_t GVNPass::ValueTable::lookupOrAddCmp(unsigned Opcode, |
| 681 | CmpInst::Predicate Predicate, |
| 682 | Value *LHS, Value *RHS) { |
| 683 | Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS); |
| 684 | return assignExpNewValueNum(exp).first; |
| 685 | } |
| 686 | |
| 687 | /// Remove all entries from the ValueTable. |
| 688 | void GVNPass::ValueTable::clear() { |
| 689 | valueNumbering.clear(); |
| 690 | expressionNumbering.clear(); |
| 691 | NumberingPhi.clear(); |
| 692 | PhiTranslateTable.clear(); |
| 693 | nextValueNumber = 1; |
| 694 | Expressions.clear(); |
| 695 | ExprIdx.clear(); |
| 696 | nextExprNumber = 0; |
| 697 | } |
| 698 | |
| 699 | /// Remove a value from the value numbering. |
| 700 | void GVNPass::ValueTable::erase(Value *V) { |
| 701 | uint32_t Num = valueNumbering.lookup(V); |
| 702 | valueNumbering.erase(V); |
| 703 | // If V is PHINode, V <--> value number is an one-to-one mapping. |
| 704 | if (isa<PHINode>(V)) |
| 705 | NumberingPhi.erase(Num); |
| 706 | } |
| 707 | |
| 708 | /// verifyRemoved - Verify that the value is removed from all internal data |
| 709 | /// structures. |
| 710 | void GVNPass::ValueTable::verifyRemoved(const Value *V) const { |
| 711 | for (DenseMap<Value*, uint32_t>::const_iterator |
| 712 | I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { |
| 713 | assert(I->first != V && "Inst still occurs in value numbering map!")(static_cast <bool> (I->first != V && "Inst still occurs in value numbering map!" ) ? void (0) : __assert_fail ("I->first != V && \"Inst still occurs in value numbering map!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 713, __extension__ __PRETTY_FUNCTION__ )); |
| 714 | } |
| 715 | } |
| 716 | |
| 717 | //===----------------------------------------------------------------------===// |
| 718 | // GVN Pass |
| 719 | //===----------------------------------------------------------------------===// |
| 720 | |
| 721 | bool GVNPass::isPREEnabled() const { |
| 722 | return Options.AllowPRE.value_or(GVNEnablePRE); |
| 723 | } |
| 724 | |
| 725 | bool GVNPass::isLoadPREEnabled() const { |
| 726 | return Options.AllowLoadPRE.value_or(GVNEnableLoadPRE); |
| 727 | } |
| 728 | |
| 729 | bool GVNPass::isLoadInLoopPREEnabled() const { |
| 730 | return Options.AllowLoadInLoopPRE.value_or(GVNEnableLoadInLoopPRE); |
| 731 | } |
| 732 | |
| 733 | bool GVNPass::isLoadPRESplitBackedgeEnabled() const { |
| 734 | return Options.AllowLoadPRESplitBackedge.value_or( |
| 735 | GVNEnableSplitBackedgeInLoadPRE); |
| 736 | } |
| 737 | |
| 738 | bool GVNPass::isMemDepEnabled() const { |
| 739 | return Options.AllowMemDep.value_or(GVNEnableMemDep); |
| 740 | } |
| 741 | |
| 742 | PreservedAnalyses GVNPass::run(Function &F, FunctionAnalysisManager &AM) { |
| 743 | // FIXME: The order of evaluation of these 'getResult' calls is very |
| 744 | // significant! Re-ordering these variables will cause GVN when run alone to |
| 745 | // be less effective! We should fix memdep and basic-aa to not exhibit this |
| 746 | // behavior, but until then don't change the order here. |
| 747 | auto &AC = AM.getResult<AssumptionAnalysis>(F); |
| 748 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); |
| 749 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); |
| 750 | auto &AA = AM.getResult<AAManager>(F); |
| 751 | auto *MemDep = |
| 752 | isMemDepEnabled() ? &AM.getResult<MemoryDependenceAnalysis>(F) : nullptr; |
| 753 | auto *LI = AM.getCachedResult<LoopAnalysis>(F); |
| 754 | auto *MSSA = AM.getCachedResult<MemorySSAAnalysis>(F); |
| 755 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); |
| 756 | bool Changed = runImpl(F, AC, DT, TLI, AA, MemDep, LI, &ORE, |
| 757 | MSSA ? &MSSA->getMSSA() : nullptr); |
| 758 | if (!Changed) |
| 759 | return PreservedAnalyses::all(); |
| 760 | PreservedAnalyses PA; |
| 761 | PA.preserve<DominatorTreeAnalysis>(); |
| 762 | PA.preserve<TargetLibraryAnalysis>(); |
| 763 | if (MSSA) |
| 764 | PA.preserve<MemorySSAAnalysis>(); |
| 765 | if (LI) |
| 766 | PA.preserve<LoopAnalysis>(); |
| 767 | return PA; |
| 768 | } |
| 769 | |
| 770 | void GVNPass::printPipeline( |
| 771 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
| 772 | static_cast<PassInfoMixin<GVNPass> *>(this)->printPipeline( |
| 773 | OS, MapClassName2PassName); |
| 774 | |
| 775 | OS << '<'; |
| 776 | if (Options.AllowPRE != std::nullopt) |
| 777 | OS << (*Options.AllowPRE ? "" : "no-") << "pre;"; |
| 778 | if (Options.AllowLoadPRE != std::nullopt) |
| 779 | OS << (*Options.AllowLoadPRE ? "" : "no-") << "load-pre;"; |
| 780 | if (Options.AllowLoadPRESplitBackedge != std::nullopt) |
| 781 | OS << (*Options.AllowLoadPRESplitBackedge ? "" : "no-") |
| 782 | << "split-backedge-load-pre;"; |
| 783 | if (Options.AllowMemDep != std::nullopt) |
| 784 | OS << (*Options.AllowMemDep ? "" : "no-") << "memdep"; |
| 785 | OS << '>'; |
| 786 | } |
| 787 | |
| 788 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
| 789 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void GVNPass::dump(DenseMap<uint32_t, Value *> &d) const { |
| 790 | errs() << "{\n"; |
| 791 | for (auto &I : d) { |
| 792 | errs() << I.first << "\n"; |
| 793 | I.second->dump(); |
| 794 | } |
| 795 | errs() << "}\n"; |
| 796 | } |
| 797 | #endif |
| 798 | |
| 799 | enum class AvailabilityState : char { |
| 800 | /// We know the block *is not* fully available. This is a fixpoint. |
| 801 | Unavailable = 0, |
| 802 | /// We know the block *is* fully available. This is a fixpoint. |
| 803 | Available = 1, |
| 804 | /// We do not know whether the block is fully available or not, |
| 805 | /// but we are currently speculating that it will be. |
| 806 | /// If it would have turned out that the block was, in fact, not fully |
| 807 | /// available, this would have been cleaned up into an Unavailable. |
| 808 | SpeculativelyAvailable = 2, |
| 809 | }; |
| 810 | |
| 811 | /// Return true if we can prove that the value |
| 812 | /// we're analyzing is fully available in the specified block. As we go, keep |
| 813 | /// track of which blocks we know are fully alive in FullyAvailableBlocks. This |
| 814 | /// map is actually a tri-state map with the following values: |
| 815 | /// 0) we know the block *is not* fully available. |
| 816 | /// 1) we know the block *is* fully available. |
| 817 | /// 2) we do not know whether the block is fully available or not, but we are |
| 818 | /// currently speculating that it will be. |
| 819 | static bool IsValueFullyAvailableInBlock( |
| 820 | BasicBlock *BB, |
| 821 | DenseMap<BasicBlock *, AvailabilityState> &FullyAvailableBlocks) { |
| 822 | SmallVector<BasicBlock *, 32> Worklist; |
| 823 | std::optional<BasicBlock *> UnavailableBB; |
| 824 | |
| 825 | // The number of times we didn't find an entry for a block in a map and |
| 826 | // optimistically inserted an entry marking block as speculatively available. |
| 827 | unsigned NumNewNewSpeculativelyAvailableBBs = 0; |
| 828 | |
| 829 | #ifndef NDEBUG |
| 830 | SmallSet<BasicBlock *, 32> NewSpeculativelyAvailableBBs; |
| 831 | SmallVector<BasicBlock *, 32> AvailableBBs; |
| 832 | #endif |
| 833 | |
| 834 | Worklist.emplace_back(BB); |
| 835 | while (!Worklist.empty()) { |
| 836 | BasicBlock *CurrBB = Worklist.pop_back_val(); // LoadFO - depth-first! |
| 837 | // Optimistically assume that the block is Speculatively Available and check |
| 838 | // to see if we already know about this block in one lookup. |
| 839 | std::pair<DenseMap<BasicBlock *, AvailabilityState>::iterator, bool> IV = |
| 840 | FullyAvailableBlocks.try_emplace( |
| 841 | CurrBB, AvailabilityState::SpeculativelyAvailable); |
| 842 | AvailabilityState &State = IV.first->second; |
| 843 | |
| 844 | // Did the entry already exist for this block? |
| 845 | if (!IV.second) { |
| 846 | if (State == AvailabilityState::Unavailable) { |
| 847 | UnavailableBB = CurrBB; |
| 848 | break; // Backpropagate unavailability info. |
| 849 | } |
| 850 | |
| 851 | #ifndef NDEBUG |
| 852 | AvailableBBs.emplace_back(CurrBB); |
| 853 | #endif |
| 854 | continue; // Don't recurse further, but continue processing worklist. |
| 855 | } |
| 856 | |
| 857 | // No entry found for block. |
| 858 | ++NumNewNewSpeculativelyAvailableBBs; |
| 859 | bool OutOfBudget = NumNewNewSpeculativelyAvailableBBs > MaxBBSpeculations; |
| 860 | |
| 861 | // If we have exhausted our budget, mark this block as unavailable. |
| 862 | // Also, if this block has no predecessors, the value isn't live-in here. |
| 863 | if (OutOfBudget || pred_empty(CurrBB)) { |
| 864 | MaxBBSpeculationCutoffReachedTimes += (int)OutOfBudget; |
| 865 | State = AvailabilityState::Unavailable; |
| 866 | UnavailableBB = CurrBB; |
| 867 | break; // Backpropagate unavailability info. |
| 868 | } |
| 869 | |
| 870 | // Tentatively consider this block as speculatively available. |
| 871 | #ifndef NDEBUG |
| 872 | NewSpeculativelyAvailableBBs.insert(CurrBB); |
| 873 | #endif |
| 874 | // And further recurse into block's predecessors, in depth-first order! |
| 875 | Worklist.append(pred_begin(CurrBB), pred_end(CurrBB)); |
| 876 | } |
| 877 | |
| 878 | #if LLVM_ENABLE_STATS1 |
| 879 | IsValueFullyAvailableInBlockNumSpeculationsMax.updateMax( |
| 880 | NumNewNewSpeculativelyAvailableBBs); |
| 881 | #endif |
| 882 | |
| 883 | // If the block isn't marked as fixpoint yet |
| 884 | // (the Unavailable and Available states are fixpoints) |
| 885 | auto MarkAsFixpointAndEnqueueSuccessors = |
| 886 | [&](BasicBlock *BB, AvailabilityState FixpointState) { |
| 887 | auto It = FullyAvailableBlocks.find(BB); |
| 888 | if (It == FullyAvailableBlocks.end()) |
| 889 | return; // Never queried this block, leave as-is. |
| 890 | switch (AvailabilityState &State = It->second) { |
| 891 | case AvailabilityState::Unavailable: |
| 892 | case AvailabilityState::Available: |
| 893 | return; // Don't backpropagate further, continue processing worklist. |
| 894 | case AvailabilityState::SpeculativelyAvailable: // Fix it! |
| 895 | State = FixpointState; |
| 896 | #ifndef NDEBUG |
| 897 | assert(NewSpeculativelyAvailableBBs.erase(BB) &&(static_cast <bool> (NewSpeculativelyAvailableBBs.erase (BB) && "Found a speculatively available successor leftover?" ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.erase(BB) && \"Found a speculatively available successor leftover?\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 898, __extension__ __PRETTY_FUNCTION__ )) |
| 898 | "Found a speculatively available successor leftover?")(static_cast <bool> (NewSpeculativelyAvailableBBs.erase (BB) && "Found a speculatively available successor leftover?" ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.erase(BB) && \"Found a speculatively available successor leftover?\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 898, __extension__ __PRETTY_FUNCTION__ )); |
| 899 | #endif |
| 900 | // Queue successors for further processing. |
| 901 | Worklist.append(succ_begin(BB), succ_end(BB)); |
| 902 | return; |
| 903 | } |
| 904 | }; |
| 905 | |
| 906 | if (UnavailableBB) { |
| 907 | // Okay, we have encountered an unavailable block. |
| 908 | // Mark speculatively available blocks reachable from UnavailableBB as |
| 909 | // unavailable as well. Paths are terminated when they reach blocks not in |
| 910 | // FullyAvailableBlocks or they are not marked as speculatively available. |
| 911 | Worklist.clear(); |
| 912 | Worklist.append(succ_begin(*UnavailableBB), succ_end(*UnavailableBB)); |
| 913 | while (!Worklist.empty()) |
| 914 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
| 915 | AvailabilityState::Unavailable); |
| 916 | } |
| 917 | |
| 918 | #ifndef NDEBUG |
| 919 | Worklist.clear(); |
| 920 | for (BasicBlock *AvailableBB : AvailableBBs) |
| 921 | Worklist.append(succ_begin(AvailableBB), succ_end(AvailableBB)); |
| 922 | while (!Worklist.empty()) |
| 923 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
| 924 | AvailabilityState::Available); |
| 925 | |
| 926 | assert(NewSpeculativelyAvailableBBs.empty() &&(static_cast <bool> (NewSpeculativelyAvailableBBs.empty () && "Must have fixed all the new speculatively available blocks." ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.empty() && \"Must have fixed all the new speculatively available blocks.\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 927, __extension__ __PRETTY_FUNCTION__ )) |
| 927 | "Must have fixed all the new speculatively available blocks.")(static_cast <bool> (NewSpeculativelyAvailableBBs.empty () && "Must have fixed all the new speculatively available blocks." ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.empty() && \"Must have fixed all the new speculatively available blocks.\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 927, __extension__ __PRETTY_FUNCTION__ )); |
| 928 | #endif |
| 929 | |
| 930 | return !UnavailableBB; |
| 931 | } |
| 932 | |
| 933 | /// Given a set of loads specified by ValuesPerBlock, |
| 934 | /// construct SSA form, allowing us to eliminate Load. This returns the value |
| 935 | /// that should be used at Load's definition site. |
| 936 | static Value * |
| 937 | ConstructSSAForLoadSet(LoadInst *Load, |
| 938 | SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, |
| 939 | GVNPass &gvn) { |
| 940 | // Check for the fully redundant, dominating load case. In this case, we can |
| 941 | // just use the dominating value directly. |
| 942 | if (ValuesPerBlock.size() == 1 && |
| 943 | gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, |
| 944 | Load->getParent())) { |
| 945 | assert(!ValuesPerBlock[0].AV.isUndefValue() &&(static_cast <bool> (!ValuesPerBlock[0].AV.isUndefValue () && "Dead BB dominate this block") ? void (0) : __assert_fail ("!ValuesPerBlock[0].AV.isUndefValue() && \"Dead BB dominate this block\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 946, __extension__ __PRETTY_FUNCTION__ )) |
| 946 | "Dead BB dominate this block")(static_cast <bool> (!ValuesPerBlock[0].AV.isUndefValue () && "Dead BB dominate this block") ? void (0) : __assert_fail ("!ValuesPerBlock[0].AV.isUndefValue() && \"Dead BB dominate this block\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 946, __extension__ __PRETTY_FUNCTION__ )); |
| 947 | return ValuesPerBlock[0].MaterializeAdjustedValue(Load, gvn); |
| 948 | } |
| 949 | |
| 950 | // Otherwise, we have to construct SSA form. |
| 951 | SmallVector<PHINode*, 8> NewPHIs; |
| 952 | SSAUpdater SSAUpdate(&NewPHIs); |
| 953 | SSAUpdate.Initialize(Load->getType(), Load->getName()); |
| 954 | |
| 955 | for (const AvailableValueInBlock &AV : ValuesPerBlock) { |
| 956 | BasicBlock *BB = AV.BB; |
| 957 | |
| 958 | if (AV.AV.isUndefValue()) |
| 959 | continue; |
| 960 | |
| 961 | if (SSAUpdate.HasValueForBlock(BB)) |
| 962 | continue; |
| 963 | |
| 964 | // If the value is the load that we will be eliminating, and the block it's |
| 965 | // available in is the block that the load is in, then don't add it as |
| 966 | // SSAUpdater will resolve the value to the relevant phi which may let it |
| 967 | // avoid phi construction entirely if there's actually only one value. |
| 968 | if (BB == Load->getParent() && |
| 969 | ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == Load) || |
| 970 | (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == Load))) |
| 971 | continue; |
| 972 | |
| 973 | SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(Load, gvn)); |
| 974 | } |
| 975 | |
| 976 | // Perform PHI construction. |
| 977 | return SSAUpdate.GetValueInMiddleOfBlock(Load->getParent()); |
| 978 | } |
| 979 | |
| 980 | Value *AvailableValue::MaterializeAdjustedValue(LoadInst *Load, |
| 981 | Instruction *InsertPt, |
| 982 | GVNPass &gvn) const { |
| 983 | Value *Res; |
| 984 | Type *LoadTy = Load->getType(); |
| 985 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
| 986 | if (isSimpleValue()) { |
| 987 | Res = getSimpleValue(); |
| 988 | if (Res->getType() != LoadTy) { |
| 989 | Res = getValueForLoad(Res, Offset, LoadTy, InsertPt, DL); |
| 990 | |
| 991 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
| 992 | << " " << *getSimpleValue() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
| 993 | << *Res << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
| 994 | << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false); |
| 995 | } |
| 996 | } else if (isCoercedLoadValue()) { |
| 997 | LoadInst *CoercedLoad = getCoercedLoadValue(); |
| 998 | if (CoercedLoad->getType() == LoadTy && Offset == 0) { |
| 999 | Res = CoercedLoad; |
| 1000 | combineMetadataForCSE(CoercedLoad, Load, false); |
| 1001 | } else { |
| 1002 | Res = getValueForLoad(CoercedLoad, Offset, LoadTy, InsertPt, DL); |
| 1003 | // We are adding a new user for this load, for which the original |
| 1004 | // metadata may not hold. Additionally, the new load may have a different |
| 1005 | // size and type, so their metadata cannot be combined in any |
| 1006 | // straightforward way. |
| 1007 | // Drop all metadata that is not known to cause immediate UB on violation, |
| 1008 | // unless the load has !noundef, in which case all metadata violations |
| 1009 | // will be promoted to UB. |
| 1010 | // TODO: We can combine noalias/alias.scope metadata here, because it is |
| 1011 | // independent of the load type. |
| 1012 | if (!CoercedLoad->hasMetadata(LLVMContext::MD_noundef)) |
| 1013 | CoercedLoad->dropUnknownNonDebugMetadata( |
| 1014 | {LLVMContext::MD_dereferenceable, |
| 1015 | LLVMContext::MD_dereferenceable_or_null, |
| 1016 | LLVMContext::MD_invariant_load, LLVMContext::MD_invariant_group}); |
| 1017 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false) |
| 1018 | << " " << *getCoercedLoadValue() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false) |
| 1019 | << *Res << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false) |
| 1020 | << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false); |
| 1021 | } |
| 1022 | } else if (isMemIntrinValue()) { |
| 1023 | Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy, |
| 1024 | InsertPt, DL); |
| 1025 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
| 1026 | << " " << *getMemIntrinValue() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
| 1027 | << *Res << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
| 1028 | << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false); |
| 1029 | } else if (isSelectValue()) { |
| 1030 | // Introduce a new value select for a load from an eligible pointer select. |
| 1031 | SelectInst *Sel = getSelectValue(); |
| 1032 | assert(V1 && V2 && "both value operands of the select must be present")(static_cast <bool> (V1 && V2 && "both value operands of the select must be present" ) ? void (0) : __assert_fail ("V1 && V2 && \"both value operands of the select must be present\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1032, __extension__ __PRETTY_FUNCTION__ )); |
| 1033 | Res = SelectInst::Create(Sel->getCondition(), V1, V2, "", Sel); |
| 1034 | } else { |
| 1035 | llvm_unreachable("Should not materialize value from dead block")::llvm::llvm_unreachable_internal("Should not materialize value from dead block" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1035); |
| 1036 | } |
| 1037 | assert(Res && "failed to materialize?")(static_cast <bool> (Res && "failed to materialize?" ) ? void (0) : __assert_fail ("Res && \"failed to materialize?\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1037, __extension__ __PRETTY_FUNCTION__ )); |
| 1038 | return Res; |
| 1039 | } |
| 1040 | |
| 1041 | static bool isLifetimeStart(const Instruction *Inst) { |
| 1042 | if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) |
| 1043 | return II->getIntrinsicID() == Intrinsic::lifetime_start; |
| 1044 | return false; |
| 1045 | } |
| 1046 | |
| 1047 | /// Assuming To can be reached from both From and Between, does Between lie on |
| 1048 | /// every path from From to To? |
| 1049 | static bool liesBetween(const Instruction *From, Instruction *Between, |
| 1050 | const Instruction *To, DominatorTree *DT) { |
| 1051 | if (From->getParent() == Between->getParent()) |
| 1052 | return DT->dominates(From, Between); |
| 1053 | SmallSet<BasicBlock *, 1> Exclusion; |
| 1054 | Exclusion.insert(Between->getParent()); |
| 1055 | return !isPotentiallyReachable(From, To, &Exclusion, DT); |
| 1056 | } |
| 1057 | |
| 1058 | /// Try to locate the three instruction involved in a missed |
| 1059 | /// load-elimination case that is due to an intervening store. |
| 1060 | static void reportMayClobberedLoad(LoadInst *Load, MemDepResult DepInfo, |
| 1061 | DominatorTree *DT, |
| 1062 | OptimizationRemarkEmitter *ORE) { |
| 1063 | using namespace ore; |
| 1064 | |
| 1065 | Instruction *OtherAccess = nullptr; |
| 1066 | |
| 1067 | OptimizationRemarkMissed R(DEBUG_TYPE"gvn", "LoadClobbered", Load); |
| 1068 | R << "load of type " << NV("Type", Load->getType()) << " not eliminated" |
| 1069 | << setExtraArgs(); |
| 1070 | |
| 1071 | for (auto *U : Load->getPointerOperand()->users()) { |
| 1072 | if (U != Load && (isa<LoadInst>(U) || isa<StoreInst>(U))) { |
| 1073 | auto *I = cast<Instruction>(U); |
| 1074 | if (I->getFunction() == Load->getFunction() && DT->dominates(I, Load)) { |
| 1075 | // Use the most immediately dominating value |
| 1076 | if (OtherAccess) { |
| 1077 | if (DT->dominates(OtherAccess, I)) |
| 1078 | OtherAccess = I; |
| 1079 | else |
| 1080 | assert(U == OtherAccess || DT->dominates(I, OtherAccess))(static_cast <bool> (U == OtherAccess || DT->dominates (I, OtherAccess)) ? void (0) : __assert_fail ("U == OtherAccess || DT->dominates(I, OtherAccess)" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1080, __extension__ __PRETTY_FUNCTION__ )); |
| 1081 | } else |
| 1082 | OtherAccess = I; |
| 1083 | } |
| 1084 | } |
| 1085 | } |
| 1086 | |
| 1087 | if (!OtherAccess) { |
| 1088 | // There is no dominating use, check if we can find a closest non-dominating |
| 1089 | // use that lies between any other potentially available use and Load. |
| 1090 | for (auto *U : Load->getPointerOperand()->users()) { |
| 1091 | if (U != Load && (isa<LoadInst>(U) || isa<StoreInst>(U))) { |
| 1092 | auto *I = cast<Instruction>(U); |
| 1093 | if (I->getFunction() == Load->getFunction() && |
| 1094 | isPotentiallyReachable(I, Load, nullptr, DT)) { |
| 1095 | if (OtherAccess) { |
| 1096 | if (liesBetween(OtherAccess, I, Load, DT)) { |
| 1097 | OtherAccess = I; |
| 1098 | } else if (!liesBetween(I, OtherAccess, Load, DT)) { |
| 1099 | // These uses are both partially available at Load were it not for |
| 1100 | // the clobber, but neither lies strictly after the other. |
| 1101 | OtherAccess = nullptr; |
| 1102 | break; |
| 1103 | } // else: keep current OtherAccess since it lies between U and Load |
| 1104 | } else { |
| 1105 | OtherAccess = I; |
| 1106 | } |
| 1107 | } |
| 1108 | } |
| 1109 | } |
| 1110 | } |
| 1111 | |
| 1112 | if (OtherAccess) |
| 1113 | R << " in favor of " << NV("OtherAccess", OtherAccess); |
| 1114 | |
| 1115 | R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst()); |
| 1116 | |
| 1117 | ORE->emit(R); |
| 1118 | } |
| 1119 | |
| 1120 | // Find non-clobbered value for Loc memory location in extended basic block |
| 1121 | // (chain of basic blocks with single predecessors) starting From instruction. |
| 1122 | static Value *findDominatingValue(const MemoryLocation &Loc, Type *LoadTy, |
| 1123 | Instruction *From, AAResults *AA) { |
| 1124 | uint32_t NumVisitedInsts = 0; |
| 1125 | BasicBlock *FromBB = From->getParent(); |
| 1126 | BatchAAResults BatchAA(*AA); |
| 1127 | for (BasicBlock *BB = FromBB; BB; BB = BB->getSinglePredecessor()) |
| 1128 | for (auto I = BB == FromBB ? From->getReverseIterator() : BB->rbegin(), |
| 1129 | E = BB->rend(); |
| 1130 | I != E; ++I) { |
| 1131 | // Stop the search if limit is reached. |
| 1132 | if (++NumVisitedInsts > MaxNumVisitedInsts) |
| 1133 | return nullptr; |
| 1134 | Instruction *Inst = &*I; |
| 1135 | if (isModSet(BatchAA.getModRefInfo(Inst, Loc))) |
| 1136 | return nullptr; |
| 1137 | if (auto *LI = dyn_cast<LoadInst>(Inst)) |
| 1138 | if (LI->getPointerOperand() == Loc.Ptr && LI->getType() == LoadTy) |
| 1139 | return LI; |
| 1140 | } |
| 1141 | return nullptr; |
| 1142 | } |
| 1143 | |
| 1144 | std::optional<AvailableValue> |
| 1145 | GVNPass::AnalyzeLoadAvailability(LoadInst *Load, MemDepResult DepInfo, |
| 1146 | Value *Address) { |
| 1147 | assert(Load->isUnordered() && "rules below are incorrect for ordered access")(static_cast <bool> (Load->isUnordered() && "rules below are incorrect for ordered access" ) ? void (0) : __assert_fail ("Load->isUnordered() && \"rules below are incorrect for ordered access\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1147, __extension__ __PRETTY_FUNCTION__ )); |
| 1148 | assert(DepInfo.isLocal() && "expected a local dependence")(static_cast <bool> (DepInfo.isLocal() && "expected a local dependence" ) ? void (0) : __assert_fail ("DepInfo.isLocal() && \"expected a local dependence\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1148, __extension__ __PRETTY_FUNCTION__ )); |
| 1149 | |
| 1150 | Instruction *DepInst = DepInfo.getInst(); |
| 1151 | |
| 1152 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
| 1153 | if (DepInfo.isClobber()) { |
| 1154 | // If the dependence is to a store that writes to a superset of the bits |
| 1155 | // read by the load, we can extract the bits we need for the load from the |
| 1156 | // stored value. |
| 1157 | if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { |
| 1158 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1159 | if (Address && Load->isAtomic() <= DepSI->isAtomic()) { |
| 1160 | int Offset = |
| 1161 | analyzeLoadFromClobberingStore(Load->getType(), Address, DepSI, DL); |
| 1162 | if (Offset != -1) |
| 1163 | return AvailableValue::get(DepSI->getValueOperand(), Offset); |
| 1164 | } |
| 1165 | } |
| 1166 | |
| 1167 | // Check to see if we have something like this: |
| 1168 | // load i32* P |
| 1169 | // load i8* (P+1) |
| 1170 | // if we have this, replace the later with an extraction from the former. |
| 1171 | if (LoadInst *DepLoad = dyn_cast<LoadInst>(DepInst)) { |
| 1172 | // If this is a clobber and L is the first instruction in its block, then |
| 1173 | // we have the first instruction in the entry block. |
| 1174 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1175 | if (DepLoad != Load && Address && |
| 1176 | Load->isAtomic() <= DepLoad->isAtomic()) { |
| 1177 | Type *LoadType = Load->getType(); |
| 1178 | int Offset = -1; |
| 1179 | |
| 1180 | // If MD reported clobber, check it was nested. |
| 1181 | if (DepInfo.isClobber() && |
| 1182 | canCoerceMustAliasedValueToLoad(DepLoad, LoadType, DL)) { |
| 1183 | const auto ClobberOff = MD->getClobberOffset(DepLoad); |
| 1184 | // GVN has no deal with a negative offset. |
| 1185 | Offset = (ClobberOff == std::nullopt || *ClobberOff < 0) |
| 1186 | ? -1 |
| 1187 | : *ClobberOff; |
| 1188 | } |
| 1189 | if (Offset == -1) |
| 1190 | Offset = |
| 1191 | analyzeLoadFromClobberingLoad(LoadType, Address, DepLoad, DL); |
| 1192 | if (Offset != -1) |
| 1193 | return AvailableValue::getLoad(DepLoad, Offset); |
| 1194 | } |
| 1195 | } |
| 1196 | |
| 1197 | // If the clobbering value is a memset/memcpy/memmove, see if we can |
| 1198 | // forward a value on from it. |
| 1199 | if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) { |
| 1200 | if (Address && !Load->isAtomic()) { |
| 1201 | int Offset = analyzeLoadFromClobberingMemInst(Load->getType(), Address, |
| 1202 | DepMI, DL); |
| 1203 | if (Offset != -1) |
| 1204 | return AvailableValue::getMI(DepMI, Offset); |
| 1205 | } |
| 1206 | } |
| 1207 | |
| 1208 | // Nothing known about this clobber, have to be conservative |
| 1209 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false) |
| 1210 | // fast print dep, using operator<< on instruction is too slow.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false) |
| 1211 | dbgs() << "GVN: load "; Load->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false) |
| 1212 | dbgs() << " is clobbered by " << *DepInst << '\n';)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false); |
| 1213 | if (ORE->allowExtraAnalysis(DEBUG_TYPE"gvn")) |
| 1214 | reportMayClobberedLoad(Load, DepInfo, DT, ORE); |
| 1215 | |
| 1216 | return std::nullopt; |
| 1217 | } |
| 1218 | assert(DepInfo.isDef() && "follows from above")(static_cast <bool> (DepInfo.isDef() && "follows from above" ) ? void (0) : __assert_fail ("DepInfo.isDef() && \"follows from above\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1218, __extension__ __PRETTY_FUNCTION__ )); |
| 1219 | |
| 1220 | // Loading the alloca -> undef. |
| 1221 | // Loading immediately after lifetime begin -> undef. |
| 1222 | if (isa<AllocaInst>(DepInst) || isLifetimeStart(DepInst)) |
| 1223 | return AvailableValue::get(UndefValue::get(Load->getType())); |
| 1224 | |
| 1225 | if (Constant *InitVal = |
| 1226 | getInitialValueOfAllocation(DepInst, TLI, Load->getType())) |
| 1227 | return AvailableValue::get(InitVal); |
| 1228 | |
| 1229 | if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { |
| 1230 | // Reject loads and stores that are to the same address but are of |
| 1231 | // different types if we have to. If the stored value is convertable to |
| 1232 | // the loaded value, we can reuse it. |
| 1233 | if (!canCoerceMustAliasedValueToLoad(S->getValueOperand(), Load->getType(), |
| 1234 | DL)) |
| 1235 | return std::nullopt; |
| 1236 | |
| 1237 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1238 | if (S->isAtomic() < Load->isAtomic()) |
| 1239 | return std::nullopt; |
| 1240 | |
| 1241 | return AvailableValue::get(S->getValueOperand()); |
| 1242 | } |
| 1243 | |
| 1244 | if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { |
| 1245 | // If the types mismatch and we can't handle it, reject reuse of the load. |
| 1246 | // If the stored value is larger or equal to the loaded value, we can reuse |
| 1247 | // it. |
| 1248 | if (!canCoerceMustAliasedValueToLoad(LD, Load->getType(), DL)) |
| 1249 | return std::nullopt; |
| 1250 | |
| 1251 | // Can't forward from non-atomic to atomic without violating memory model. |
| 1252 | if (LD->isAtomic() < Load->isAtomic()) |
| 1253 | return std::nullopt; |
| 1254 | |
| 1255 | return AvailableValue::getLoad(LD); |
| 1256 | } |
| 1257 | |
| 1258 | // Check if load with Addr dependent from select can be converted to select |
| 1259 | // between load values. There must be no instructions between the found |
| 1260 | // loads and DepInst that may clobber the loads. |
| 1261 | if (auto *Sel = dyn_cast<SelectInst>(DepInst)) { |
| 1262 | assert(Sel->getType() == Load->getPointerOperandType())(static_cast <bool> (Sel->getType() == Load->getPointerOperandType ()) ? void (0) : __assert_fail ("Sel->getType() == Load->getPointerOperandType()" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1262, __extension__ __PRETTY_FUNCTION__ )); |
| 1263 | auto Loc = MemoryLocation::get(Load); |
| 1264 | Value *V1 = |
| 1265 | findDominatingValue(Loc.getWithNewPtr(Sel->getTrueValue()), |
| 1266 | Load->getType(), DepInst, getAliasAnalysis()); |
| 1267 | if (!V1) |
| 1268 | return std::nullopt; |
| 1269 | Value *V2 = |
| 1270 | findDominatingValue(Loc.getWithNewPtr(Sel->getFalseValue()), |
| 1271 | Load->getType(), DepInst, getAliasAnalysis()); |
| 1272 | if (!V2) |
| 1273 | return std::nullopt; |
| 1274 | return AvailableValue::getSelect(Sel, V1, V2); |
| 1275 | } |
| 1276 | |
| 1277 | // Unknown def - must be conservative |
| 1278 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false) |
| 1279 | // fast print dep, using operator<< on instruction is too slow.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false) |
| 1280 | dbgs() << "GVN: load "; Load->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false) |
| 1281 | dbgs() << " has unknown def " << *DepInst << '\n';)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false); |
| 1282 | return std::nullopt; |
| 1283 | } |
| 1284 | |
| 1285 | void GVNPass::AnalyzeLoadAvailability(LoadInst *Load, LoadDepVect &Deps, |
| 1286 | AvailValInBlkVect &ValuesPerBlock, |
| 1287 | UnavailBlkVect &UnavailableBlocks) { |
| 1288 | // Filter out useless results (non-locals, etc). Keep track of the blocks |
| 1289 | // where we have a value available in repl, also keep track of whether we see |
| 1290 | // dependencies that produce an unknown value for the load (such as a call |
| 1291 | // that could potentially clobber the load). |
| 1292 | for (const auto &Dep : Deps) { |
| 1293 | BasicBlock *DepBB = Dep.getBB(); |
| 1294 | MemDepResult DepInfo = Dep.getResult(); |
| 1295 | |
| 1296 | if (DeadBlocks.count(DepBB)) { |
| 1297 | // Dead dependent mem-op disguise as a load evaluating the same value |
| 1298 | // as the load in question. |
| 1299 | ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB)); |
| 1300 | continue; |
| 1301 | } |
| 1302 | |
| 1303 | if (!DepInfo.isLocal()) { |
| 1304 | UnavailableBlocks.push_back(DepBB); |
| 1305 | continue; |
| 1306 | } |
| 1307 | |
| 1308 | // The address being loaded in this non-local block may not be the same as |
| 1309 | // the pointer operand of the load if PHI translation occurs. Make sure |
| 1310 | // to consider the right address. |
| 1311 | if (auto AV = AnalyzeLoadAvailability(Load, DepInfo, Dep.getAddress())) { |
| 1312 | // subtlety: because we know this was a non-local dependency, we know |
| 1313 | // it's safe to materialize anywhere between the instruction within |
| 1314 | // DepInfo and the end of it's block. |
| 1315 | ValuesPerBlock.push_back( |
| 1316 | AvailableValueInBlock::get(DepBB, std::move(*AV))); |
| 1317 | } else { |
| 1318 | UnavailableBlocks.push_back(DepBB); |
| 1319 | } |
| 1320 | } |
| 1321 | |
| 1322 | assert(Deps.size() == ValuesPerBlock.size() + UnavailableBlocks.size() &&(static_cast <bool> (Deps.size() == ValuesPerBlock.size () + UnavailableBlocks.size() && "post condition violation" ) ? void (0) : __assert_fail ("Deps.size() == ValuesPerBlock.size() + UnavailableBlocks.size() && \"post condition violation\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1323, __extension__ __PRETTY_FUNCTION__ )) |
| 1323 | "post condition violation")(static_cast <bool> (Deps.size() == ValuesPerBlock.size () + UnavailableBlocks.size() && "post condition violation" ) ? void (0) : __assert_fail ("Deps.size() == ValuesPerBlock.size() + UnavailableBlocks.size() && \"post condition violation\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1323, __extension__ __PRETTY_FUNCTION__ )); |
| 1324 | } |
| 1325 | |
| 1326 | void GVNPass::eliminatePartiallyRedundantLoad( |
| 1327 | LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
| 1328 | MapVector<BasicBlock *, Value *> &AvailableLoads) { |
| 1329 | for (const auto &AvailableLoad : AvailableLoads) { |
| 1330 | BasicBlock *UnavailableBlock = AvailableLoad.first; |
| 1331 | Value *LoadPtr = AvailableLoad.second; |
| 1332 | |
| 1333 | auto *NewLoad = |
| 1334 | new LoadInst(Load->getType(), LoadPtr, Load->getName() + ".pre", |
| 1335 | Load->isVolatile(), Load->getAlign(), Load->getOrdering(), |
| 1336 | Load->getSyncScopeID(), UnavailableBlock->getTerminator()); |
| 1337 | NewLoad->setDebugLoc(Load->getDebugLoc()); |
| 1338 | if (MSSAU) { |
| 1339 | auto *MSSA = MSSAU->getMemorySSA(); |
| 1340 | // Get the defining access of the original load or use the load if it is a |
| 1341 | // MemoryDef (e.g. because it is volatile). The inserted loads are |
| 1342 | // guaranteed to load from the same definition. |
| 1343 | auto *LoadAcc = MSSA->getMemoryAccess(Load); |
| 1344 | auto *DefiningAcc = |
| 1345 | isa<MemoryDef>(LoadAcc) ? LoadAcc : LoadAcc->getDefiningAccess(); |
| 1346 | auto *NewAccess = MSSAU->createMemoryAccessInBB( |
| 1347 | NewLoad, DefiningAcc, NewLoad->getParent(), |
| 1348 | MemorySSA::BeforeTerminator); |
| 1349 | if (auto *NewDef = dyn_cast<MemoryDef>(NewAccess)) |
| 1350 | MSSAU->insertDef(NewDef, /*RenameUses=*/true); |
| 1351 | else |
| 1352 | MSSAU->insertUse(cast<MemoryUse>(NewAccess), /*RenameUses=*/true); |
| 1353 | } |
| 1354 | |
| 1355 | // Transfer the old load's AA tags to the new load. |
| 1356 | AAMDNodes Tags = Load->getAAMetadata(); |
| 1357 | if (Tags) |
| 1358 | NewLoad->setAAMetadata(Tags); |
| 1359 | |
| 1360 | if (auto *MD = Load->getMetadata(LLVMContext::MD_invariant_load)) |
| 1361 | NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD); |
| 1362 | if (auto *InvGroupMD = Load->getMetadata(LLVMContext::MD_invariant_group)) |
| 1363 | NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD); |
| 1364 | if (auto *RangeMD = Load->getMetadata(LLVMContext::MD_range)) |
| 1365 | NewLoad->setMetadata(LLVMContext::MD_range, RangeMD); |
| 1366 | if (auto *AccessMD = Load->getMetadata(LLVMContext::MD_access_group)) |
| 1367 | if (LI && |
| 1368 | LI->getLoopFor(Load->getParent()) == LI->getLoopFor(UnavailableBlock)) |
| 1369 | NewLoad->setMetadata(LLVMContext::MD_access_group, AccessMD); |
| 1370 | |
| 1371 | // We do not propagate the old load's debug location, because the new |
| 1372 | // load now lives in a different BB, and we want to avoid a jumpy line |
| 1373 | // table. |
| 1374 | // FIXME: How do we retain source locations without causing poor debugging |
| 1375 | // behavior? |
| 1376 | |
| 1377 | // Add the newly created load. |
| 1378 | ValuesPerBlock.push_back( |
| 1379 | AvailableValueInBlock::get(UnavailableBlock, NewLoad)); |
| 1380 | MD->invalidateCachedPointerInfo(LoadPtr); |
| 1381 | LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN INSERTED " << *NewLoad << '\n'; } } while (false); |
| 1382 | } |
| 1383 | |
| 1384 | // Perform PHI construction. |
| 1385 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, *this); |
| 1386 | // ConstructSSAForLoadSet is responsible for combining metadata. |
| 1387 | Load->replaceAllUsesWith(V); |
| 1388 | if (isa<PHINode>(V)) |
| 1389 | V->takeName(Load); |
| 1390 | if (Instruction *I = dyn_cast<Instruction>(V)) |
| 1391 | I->setDebugLoc(Load->getDebugLoc()); |
| 1392 | if (V->getType()->isPtrOrPtrVectorTy()) |
| 1393 | MD->invalidateCachedPointerInfo(V); |
| 1394 | markInstructionForDeletion(Load); |
| 1395 | ORE->emit([&]() { |
| 1396 | return OptimizationRemark(DEBUG_TYPE"gvn", "LoadPRE", Load) |
| 1397 | << "load eliminated by PRE"; |
| 1398 | }); |
| 1399 | } |
| 1400 | |
| 1401 | bool GVNPass::PerformLoadPRE(LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
| 1402 | UnavailBlkVect &UnavailableBlocks) { |
| 1403 | // Okay, we have *some* definitions of the value. This means that the value |
| 1404 | // is available in some of our (transitive) predecessors. Lets think about |
| 1405 | // doing PRE of this load. This will involve inserting a new load into the |
| 1406 | // predecessor when it's not available. We could do this in general, but |
| 1407 | // prefer to not increase code size. As such, we only do this when we know |
| 1408 | // that we only have to insert *one* load (which means we're basically moving |
| 1409 | // the load, not inserting a new one). |
| 1410 | |
| 1411 | SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(), |
| 1412 | UnavailableBlocks.end()); |
| 1413 | |
| 1414 | // Let's find the first basic block with more than one predecessor. Walk |
| 1415 | // backwards through predecessors if needed. |
| 1416 | BasicBlock *LoadBB = Load->getParent(); |
| 1417 | BasicBlock *TmpBB = LoadBB; |
| 1418 | |
| 1419 | // Check that there is no implicit control flow instructions above our load in |
| 1420 | // its block. If there is an instruction that doesn't always pass the |
| 1421 | // execution to the following instruction, then moving through it may become |
| 1422 | // invalid. For example: |
| 1423 | // |
| 1424 | // int arr[LEN]; |
| 1425 | // int index = ???; |
| 1426 | // ... |
| 1427 | // guard(0 <= index && index < LEN); |
| 1428 | // use(arr[index]); |
| 1429 | // |
| 1430 | // It is illegal to move the array access to any point above the guard, |
| 1431 | // because if the index is out of bounds we should deoptimize rather than |
| 1432 | // access the array. |
| 1433 | // Check that there is no guard in this block above our instruction. |
| 1434 | bool MustEnsureSafetyOfSpeculativeExecution = |
| 1435 | ICF->isDominatedByICFIFromSameBlock(Load); |
| 1436 | |
| 1437 | while (TmpBB->getSinglePredecessor()) { |
| 1438 | TmpBB = TmpBB->getSinglePredecessor(); |
| 1439 | if (TmpBB == LoadBB) // Infinite (unreachable) loop. |
| 1440 | return false; |
| 1441 | if (Blockers.count(TmpBB)) |
| 1442 | return false; |
| 1443 | |
| 1444 | // If any of these blocks has more than one successor (i.e. if the edge we |
| 1445 | // just traversed was critical), then there are other paths through this |
| 1446 | // block along which the load may not be anticipated. Hoisting the load |
| 1447 | // above this block would be adding the load to execution paths along |
| 1448 | // which it was not previously executed. |
| 1449 | if (TmpBB->getTerminator()->getNumSuccessors() != 1) |
| 1450 | return false; |
| 1451 | |
| 1452 | // Check that there is no implicit control flow in a block above. |
| 1453 | MustEnsureSafetyOfSpeculativeExecution = |
| 1454 | MustEnsureSafetyOfSpeculativeExecution || ICF->hasICF(TmpBB); |
| 1455 | } |
| 1456 | |
| 1457 | assert(TmpBB)(static_cast <bool> (TmpBB) ? void (0) : __assert_fail ( "TmpBB", "llvm/lib/Transforms/Scalar/GVN.cpp", 1457, __extension__ __PRETTY_FUNCTION__)); |
| 1458 | LoadBB = TmpBB; |
| 1459 | |
| 1460 | // Check to see how many predecessors have the loaded value fully |
| 1461 | // available. |
| 1462 | MapVector<BasicBlock *, Value *> PredLoads; |
| 1463 | DenseMap<BasicBlock *, AvailabilityState> FullyAvailableBlocks; |
| 1464 | for (const AvailableValueInBlock &AV : ValuesPerBlock) |
| 1465 | FullyAvailableBlocks[AV.BB] = AvailabilityState::Available; |
| 1466 | for (BasicBlock *UnavailableBB : UnavailableBlocks) |
| 1467 | FullyAvailableBlocks[UnavailableBB] = AvailabilityState::Unavailable; |
| 1468 | |
| 1469 | SmallVector<BasicBlock *, 4> CriticalEdgePred; |
| 1470 | for (BasicBlock *Pred : predecessors(LoadBB)) { |
| 1471 | // If any predecessor block is an EH pad that does not allow non-PHI |
| 1472 | // instructions before the terminator, we can't PRE the load. |
| 1473 | if (Pred->getTerminator()->isEHPad()) { |
| 1474 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1475 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1476 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
| 1477 | return false; |
| 1478 | } |
| 1479 | |
| 1480 | if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { |
| 1481 | continue; |
| 1482 | } |
| 1483 | |
| 1484 | if (Pred->getTerminator()->getNumSuccessors() != 1) { |
| 1485 | if (isa<IndirectBrInst>(Pred->getTerminator())) { |
| 1486 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1487 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1488 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
| 1489 | return false; |
| 1490 | } |
| 1491 | |
| 1492 | if (LoadBB->isEHPad()) { |
| 1493 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1494 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1495 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
| 1496 | return false; |
| 1497 | } |
| 1498 | |
| 1499 | // Do not split backedge as it will break the canonical loop form. |
| 1500 | if (!isLoadPRESplitBackedgeEnabled()) |
| 1501 | if (DT->dominates(LoadBB, Pred)) { |
| 1502 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1503 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1504 | << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
| 1505 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
| 1506 | return false; |
| 1507 | } |
| 1508 | |
| 1509 | CriticalEdgePred.push_back(Pred); |
| 1510 | } else { |
| 1511 | // Only add the predecessors that will not be split for now. |
| 1512 | PredLoads[Pred] = nullptr; |
| 1513 | } |
| 1514 | } |
| 1515 | |
| 1516 | // Decide whether PRE is profitable for this load. |
| 1517 | unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size(); |
| 1518 | assert(NumUnavailablePreds != 0 &&(static_cast <bool> (NumUnavailablePreds != 0 && "Fully available value should already be eliminated!") ? void (0) : __assert_fail ("NumUnavailablePreds != 0 && \"Fully available value should already be eliminated!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1519, __extension__ __PRETTY_FUNCTION__ )) |
| 1519 | "Fully available value should already be eliminated!")(static_cast <bool> (NumUnavailablePreds != 0 && "Fully available value should already be eliminated!") ? void (0) : __assert_fail ("NumUnavailablePreds != 0 && \"Fully available value should already be eliminated!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1519, __extension__ __PRETTY_FUNCTION__ )); |
| 1520 | (void)NumUnavailablePreds; |
| 1521 | |
| 1522 | // If this load is unavailable in multiple predecessors, reject it. |
| 1523 | // FIXME: If we could restructure the CFG, we could make a common pred with |
| 1524 | // all the preds that don't have an available Load and insert a new load into |
| 1525 | // that one block. |
| 1526 | if (NumUnavailablePreds != 1) |
| 1527 | return false; |
| 1528 | |
| 1529 | // Now we know where we will insert load. We must ensure that it is safe |
| 1530 | // to speculatively execute the load at that points. |
| 1531 | if (MustEnsureSafetyOfSpeculativeExecution) { |
| 1532 | if (CriticalEdgePred.size()) |
| 1533 | if (!isSafeToSpeculativelyExecute(Load, LoadBB->getFirstNonPHI(), AC, DT)) |
| 1534 | return false; |
| 1535 | for (auto &PL : PredLoads) |
| 1536 | if (!isSafeToSpeculativelyExecute(Load, PL.first->getTerminator(), AC, |
| 1537 | DT)) |
| 1538 | return false; |
| 1539 | } |
| 1540 | |
| 1541 | // Split critical edges, and update the unavailable predecessors accordingly. |
| 1542 | for (BasicBlock *OrigPred : CriticalEdgePred) { |
| 1543 | BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB); |
| 1544 | assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!")(static_cast <bool> (!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!") ? void (0) : __assert_fail ("!PredLoads.count(OrigPred) && \"Split edges shouldn't be in map!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 1544, __extension__ __PRETTY_FUNCTION__ )); |
| 1545 | PredLoads[NewPred] = nullptr; |
| 1546 | LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Split critical edge " << OrigPred ->getName() << "->" << LoadBB->getName() << '\n'; } } while (false) |
| 1547 | << LoadBB->getName() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Split critical edge " << OrigPred ->getName() << "->" << LoadBB->getName() << '\n'; } } while (false); |
| 1548 | } |
| 1549 | |
| 1550 | // Check if the load can safely be moved to all the unavailable predecessors. |
| 1551 | bool CanDoPRE = true; |
| 1552 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
| 1553 | SmallVector<Instruction*, 8> NewInsts; |
| 1554 | for (auto &PredLoad : PredLoads) { |
| 1555 | BasicBlock *UnavailablePred = PredLoad.first; |
| 1556 | |
| 1557 | // Do PHI translation to get its value in the predecessor if necessary. The |
| 1558 | // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. |
| 1559 | // We do the translation for each edge we skipped by going from Load's block |
| 1560 | // to LoadBB, otherwise we might miss pieces needing translation. |
| 1561 | |
| 1562 | // If all preds have a single successor, then we know it is safe to insert |
| 1563 | // the load on the pred (?!?), so we can insert code to materialize the |
| 1564 | // pointer if it is not available. |
| 1565 | Value *LoadPtr = Load->getPointerOperand(); |
| 1566 | BasicBlock *Cur = Load->getParent(); |
| 1567 | while (Cur != LoadBB) { |
| 1568 | PHITransAddr Address(LoadPtr, DL, AC); |
| 1569 | LoadPtr = Address.translateWithInsertion(Cur, Cur->getSinglePredecessor(), |
| 1570 | *DT, NewInsts); |
| 1571 | if (!LoadPtr) { |
| 1572 | CanDoPRE = false; |
| 1573 | break; |
| 1574 | } |
| 1575 | Cur = Cur->getSinglePredecessor(); |
| 1576 | } |
| 1577 | |
| 1578 | if (LoadPtr) { |
| 1579 | PHITransAddr Address(LoadPtr, DL, AC); |
| 1580 | LoadPtr = Address.translateWithInsertion(LoadBB, UnavailablePred, *DT, |
| 1581 | NewInsts); |
| 1582 | } |
| 1583 | // If we couldn't find or insert a computation of this phi translated value, |
| 1584 | // we fail PRE. |
| 1585 | if (!LoadPtr) { |
| 1586 | LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " << *Load->getPointerOperand() << "\n"; } } while (false) |
| 1587 | << *Load->getPointerOperand() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " << *Load->getPointerOperand() << "\n"; } } while (false); |
| 1588 | CanDoPRE = false; |
| 1589 | break; |
| 1590 | } |
| 1591 | |
| 1592 | PredLoad.second = LoadPtr; |
| 1593 | } |
| 1594 | |
| 1595 | if (!CanDoPRE) { |
| 1596 | while (!NewInsts.empty()) { |
| 1597 | // Erase instructions generated by the failed PHI translation before |
| 1598 | // trying to number them. PHI translation might insert instructions |
| 1599 | // in basic blocks other than the current one, and we delete them |
| 1600 | // directly, as markInstructionForDeletion only allows removing from the |
| 1601 | // current basic block. |
| 1602 | NewInsts.pop_back_val()->eraseFromParent(); |
| 1603 | } |
| 1604 | // HINT: Don't revert the edge-splitting as following transformation may |
| 1605 | // also need to split these critical edges. |
| 1606 | return !CriticalEdgePred.empty(); |
| 1607 | } |
| 1608 | |
| 1609 | // Okay, we can eliminate this load by inserting a reload in the predecessor |
| 1610 | // and using PHI construction to get the value in the other predecessors, do |
| 1611 | // it. |
| 1612 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN REMOVING PRE LOAD: " << *Load << '\n'; } } while (false); |
| 1613 | LLVM_DEBUG(if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts .back() << '\n'; } } while (false) |
| 1614 | << " INSTS: " << *NewInsts.back()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts .back() << '\n'; } } while (false) |
| 1615 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts .back() << '\n'; } } while (false); |
| 1616 | |
| 1617 | // Assign value numbers to the new instructions. |
| 1618 | for (Instruction *I : NewInsts) { |
| 1619 | // Instructions that have been inserted in predecessor(s) to materialize |
| 1620 | // the load address do not retain their original debug locations. Doing |
| 1621 | // so could lead to confusing (but correct) source attributions. |
| 1622 | I->updateLocationAfterHoist(); |
| 1623 | |
| 1624 | // FIXME: We really _ought_ to insert these value numbers into their |
| 1625 | // parent's availability map. However, in doing so, we risk getting into |
| 1626 | // ordering issues. If a block hasn't been processed yet, we would be |
| 1627 | // marking a value as AVAIL-IN, which isn't what we intend. |
| 1628 | VN.lookupOrAdd(I); |
| 1629 | } |
| 1630 | |
| 1631 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, PredLoads); |
| 1632 | ++NumPRELoad; |
| 1633 | return true; |
| 1634 | } |
| 1635 | |
| 1636 | bool GVNPass::performLoopLoadPRE(LoadInst *Load, |
| 1637 | AvailValInBlkVect &ValuesPerBlock, |
| 1638 | UnavailBlkVect &UnavailableBlocks) { |
| 1639 | if (!LI) |
| 1640 | return false; |
| 1641 | |
| 1642 | const Loop *L = LI->getLoopFor(Load->getParent()); |
| 1643 | // TODO: Generalize to other loop blocks that dominate the latch. |
| 1644 | if (!L || L->getHeader() != Load->getParent()) |
| 1645 | return false; |
| 1646 | |
| 1647 | BasicBlock *Preheader = L->getLoopPreheader(); |
| 1648 | BasicBlock *Latch = L->getLoopLatch(); |
| 1649 | if (!Preheader || !Latch) |
| 1650 | return false; |
| 1651 | |
| 1652 | Value *LoadPtr = Load->getPointerOperand(); |
| 1653 | // Must be available in preheader. |
| 1654 | if (!L->isLoopInvariant(LoadPtr)) |
| 1655 | return false; |
| 1656 | |
| 1657 | // We plan to hoist the load to preheader without introducing a new fault. |
| 1658 | // In order to do it, we need to prove that we cannot side-exit the loop |
| 1659 | // once loop header is first entered before execution of the load. |
| 1660 | if (ICF->isDominatedByICFIFromSameBlock(Load)) |
| 1661 | return false; |
| 1662 | |
| 1663 | BasicBlock *LoopBlock = nullptr; |
| 1664 | for (auto *Blocker : UnavailableBlocks) { |
| 1665 | // Blockers from outside the loop are handled in preheader. |
| 1666 | if (!L->contains(Blocker)) |
| 1667 | continue; |
| 1668 | |
| 1669 | // Only allow one loop block. Loop header is not less frequently executed |
| 1670 | // than each loop block, and likely it is much more frequently executed. But |
| 1671 | // in case of multiple loop blocks, we need extra information (such as block |
| 1672 | // frequency info) to understand whether it is profitable to PRE into |
| 1673 | // multiple loop blocks. |
| 1674 | if (LoopBlock) |
| 1675 | return false; |
| 1676 | |
| 1677 | // Do not sink into inner loops. This may be non-profitable. |
| 1678 | if (L != LI->getLoopFor(Blocker)) |
| 1679 | return false; |
| 1680 | |
| 1681 | // Blocks that dominate the latch execute on every single iteration, maybe |
| 1682 | // except the last one. So PREing into these blocks doesn't make much sense |
| 1683 | // in most cases. But the blocks that do not necessarily execute on each |
| 1684 | // iteration are sometimes much colder than the header, and this is when |
| 1685 | // PRE is potentially profitable. |
| 1686 | if (DT->dominates(Blocker, Latch)) |
| 1687 | return false; |
| 1688 | |
| 1689 | // Make sure that the terminator itself doesn't clobber. |
| 1690 | if (Blocker->getTerminator()->mayWriteToMemory()) |
| 1691 | return false; |
| 1692 | |
| 1693 | LoopBlock = Blocker; |
| 1694 | } |
| 1695 | |
| 1696 | if (!LoopBlock) |
| 1697 | return false; |
| 1698 | |
| 1699 | // Make sure the memory at this pointer cannot be freed, therefore we can |
| 1700 | // safely reload from it after clobber. |
| 1701 | if (LoadPtr->canBeFreed()) |
| 1702 | return false; |
| 1703 | |
| 1704 | // TODO: Support critical edge splitting if blocker has more than 1 successor. |
| 1705 | MapVector<BasicBlock *, Value *> AvailableLoads; |
| 1706 | AvailableLoads[LoopBlock] = LoadPtr; |
| 1707 | AvailableLoads[Preheader] = LoadPtr; |
| 1708 | |
| 1709 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOOP LOAD: " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN REMOVING PRE LOOP LOAD: " << *Load << '\n'; } } while (false); |
| 1710 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, AvailableLoads); |
| 1711 | ++NumPRELoopLoad; |
| 1712 | return true; |
| 1713 | } |
| 1714 | |
| 1715 | static void reportLoadElim(LoadInst *Load, Value *AvailableValue, |
| 1716 | OptimizationRemarkEmitter *ORE) { |
| 1717 | using namespace ore; |
| 1718 | |
| 1719 | ORE->emit([&]() { |
| 1720 | return OptimizationRemark(DEBUG_TYPE"gvn", "LoadElim", Load) |
| 1721 | << "load of type " << NV("Type", Load->getType()) << " eliminated" |
| 1722 | << setExtraArgs() << " in favor of " |
| 1723 | << NV("InfavorOfValue", AvailableValue); |
| 1724 | }); |
| 1725 | } |
| 1726 | |
| 1727 | /// Attempt to eliminate a load whose dependencies are |
| 1728 | /// non-local by performing PHI construction. |
| 1729 | bool GVNPass::processNonLocalLoad(LoadInst *Load) { |
| 1730 | // non-local speculations are not allowed under asan. |
| 1731 | if (Load->getParent()->getParent()->hasFnAttribute( |
| 1732 | Attribute::SanitizeAddress) || |
| 1733 | Load->getParent()->getParent()->hasFnAttribute( |
| 1734 | Attribute::SanitizeHWAddress)) |
| 1735 | return false; |
| 1736 | |
| 1737 | // Step 1: Find the non-local dependencies of the load. |
| 1738 | LoadDepVect Deps; |
| 1739 | MD->getNonLocalPointerDependency(Load, Deps); |
| 1740 | |
| 1741 | // If we had to process more than one hundred blocks to find the |
| 1742 | // dependencies, this load isn't worth worrying about. Optimizing |
| 1743 | // it will be too expensive. |
| 1744 | unsigned NumDeps = Deps.size(); |
| 1745 | if (NumDeps > MaxNumDeps) |
| 1746 | return false; |
| 1747 | |
| 1748 | // If we had a phi translation failure, we'll have a single entry which is a |
| 1749 | // clobber in the current block. Reject this early. |
| 1750 | if (NumDeps == 1 && |
| 1751 | !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { |
| 1752 | LLVM_DEBUG(dbgs() << "GVN: non-local load "; Load->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: non-local load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown dependencies\n";; } } while (false) |
| 1753 | dbgs() << " has unknown dependencies\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: non-local load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown dependencies\n";; } } while (false); |
| 1754 | return false; |
| 1755 | } |
| 1756 | |
| 1757 | bool Changed = false; |
| 1758 | // If this load follows a GEP, see if we can PRE the indices before analyzing. |
| 1759 | if (GetElementPtrInst *GEP = |
| 1760 | dyn_cast<GetElementPtrInst>(Load->getOperand(0))) { |
| 1761 | for (Use &U : GEP->indices()) |
| 1762 | if (Instruction *I = dyn_cast<Instruction>(U.get())) |
| 1763 | Changed |= performScalarPRE(I); |
| 1764 | } |
| 1765 | |
| 1766 | // Step 2: Analyze the availability of the load |
| 1767 | AvailValInBlkVect ValuesPerBlock; |
| 1768 | UnavailBlkVect UnavailableBlocks; |
| 1769 | AnalyzeLoadAvailability(Load, Deps, ValuesPerBlock, UnavailableBlocks); |
| 1770 | |
| 1771 | // If we have no predecessors that produce a known value for this load, exit |
| 1772 | // early. |
| 1773 | if (ValuesPerBlock.empty()) |
| 1774 | return Changed; |
| 1775 | |
| 1776 | // Step 3: Eliminate fully redundancy. |
| 1777 | // |
| 1778 | // If all of the instructions we depend on produce a known value for this |
| 1779 | // load, then it is fully redundant and we can use PHI insertion to compute |
| 1780 | // its value. Insert PHIs and remove the fully redundant value now. |
| 1781 | if (UnavailableBlocks.empty()) { |
| 1782 | LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *Load << '\n'; } } while (false); |
| 1783 | |
| 1784 | // Perform PHI construction. |
| 1785 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, *this); |
| 1786 | // ConstructSSAForLoadSet is responsible for combining metadata. |
| 1787 | Load->replaceAllUsesWith(V); |
| 1788 | |
| 1789 | if (isa<PHINode>(V)) |
| 1790 | V->takeName(Load); |
| 1791 | if (Instruction *I = dyn_cast<Instruction>(V)) |
| 1792 | // If instruction I has debug info, then we should not update it. |
| 1793 | // Also, if I has a null DebugLoc, then it is still potentially incorrect |
| 1794 | // to propagate Load's DebugLoc because Load may not post-dominate I. |
| 1795 | if (Load->getDebugLoc() && Load->getParent() == I->getParent()) |
| 1796 | I->setDebugLoc(Load->getDebugLoc()); |
| 1797 | if (V->getType()->isPtrOrPtrVectorTy()) |
| 1798 | MD->invalidateCachedPointerInfo(V); |
| 1799 | markInstructionForDeletion(Load); |
| 1800 | ++NumGVNLoad; |
| 1801 | reportLoadElim(Load, V, ORE); |
| 1802 | return true; |
| 1803 | } |
| 1804 | |
| 1805 | // Step 4: Eliminate partial redundancy. |
| 1806 | if (!isPREEnabled() || !isLoadPREEnabled()) |
| 1807 | return Changed; |
| 1808 | if (!isLoadInLoopPREEnabled() && LI && LI->getLoopFor(Load->getParent())) |
| 1809 | return Changed; |
| 1810 | |
| 1811 | if (performLoopLoadPRE(Load, ValuesPerBlock, UnavailableBlocks) || |
| 1812 | PerformLoadPRE(Load, ValuesPerBlock, UnavailableBlocks)) |
| 1813 | return true; |
| 1814 | |
| 1815 | return Changed; |
| 1816 | } |
| 1817 | |
| 1818 | static bool impliesEquivalanceIfTrue(CmpInst* Cmp) { |
| 1819 | if (Cmp->getPredicate() == CmpInst::Predicate::ICMP_EQ) |
| 1820 | return true; |
| 1821 | |
| 1822 | // Floating point comparisons can be equal, but not equivalent. Cases: |
| 1823 | // NaNs for unordered operators |
| 1824 | // +0.0 vs 0.0 for all operators |
| 1825 | if (Cmp->getPredicate() == CmpInst::Predicate::FCMP_OEQ || |
| 1826 | (Cmp->getPredicate() == CmpInst::Predicate::FCMP_UEQ && |
| 1827 | Cmp->getFastMathFlags().noNaNs())) { |
| 1828 | Value *LHS = Cmp->getOperand(0); |
| 1829 | Value *RHS = Cmp->getOperand(1); |
| 1830 | // If we can prove either side non-zero, then equality must imply |
| 1831 | // equivalence. |
| 1832 | // FIXME: We should do this optimization if 'no signed zeros' is |
| 1833 | // applicable via an instruction-level fast-math-flag or some other |
| 1834 | // indicator that relaxed FP semantics are being used. |
| 1835 | if (isa<ConstantFP>(LHS) && !cast<ConstantFP>(LHS)->isZero()) |
| 1836 | return true; |
| 1837 | if (isa<ConstantFP>(RHS) && !cast<ConstantFP>(RHS)->isZero()) |
| 1838 | return true;; |
| 1839 | // TODO: Handle vector floating point constants |
| 1840 | } |
| 1841 | return false; |
| 1842 | } |
| 1843 | |
| 1844 | static bool impliesEquivalanceIfFalse(CmpInst* Cmp) { |
| 1845 | if (Cmp->getPredicate() == CmpInst::Predicate::ICMP_NE) |
| 1846 | return true; |
| 1847 | |
| 1848 | // Floating point comparisons can be equal, but not equivelent. Cases: |
| 1849 | // NaNs for unordered operators |
| 1850 | // +0.0 vs 0.0 for all operators |
| 1851 | if ((Cmp->getPredicate() == CmpInst::Predicate::FCMP_ONE && |
| 1852 | Cmp->getFastMathFlags().noNaNs()) || |
| 1853 | Cmp->getPredicate() == CmpInst::Predicate::FCMP_UNE) { |
| 1854 | Value *LHS = Cmp->getOperand(0); |
| 1855 | Value *RHS = Cmp->getOperand(1); |
| 1856 | // If we can prove either side non-zero, then equality must imply |
| 1857 | // equivalence. |
| 1858 | // FIXME: We should do this optimization if 'no signed zeros' is |
| 1859 | // applicable via an instruction-level fast-math-flag or some other |
| 1860 | // indicator that relaxed FP semantics are being used. |
| 1861 | if (isa<ConstantFP>(LHS) && !cast<ConstantFP>(LHS)->isZero()) |
| 1862 | return true; |
| 1863 | if (isa<ConstantFP>(RHS) && !cast<ConstantFP>(RHS)->isZero()) |
| 1864 | return true;; |
| 1865 | // TODO: Handle vector floating point constants |
| 1866 | } |
| 1867 | return false; |
| 1868 | } |
| 1869 | |
| 1870 | |
| 1871 | static bool hasUsersIn(Value *V, BasicBlock *BB) { |
| 1872 | return llvm::any_of(V->users(), [BB](User *U) { |
| 1873 | auto *I = dyn_cast<Instruction>(U); |
| 1874 | return I && I->getParent() == BB; |
| 1875 | }); |
| 1876 | } |
| 1877 | |
| 1878 | bool GVNPass::processAssumeIntrinsic(AssumeInst *IntrinsicI) { |
| 1879 | Value *V = IntrinsicI->getArgOperand(0); |
| 1880 | |
| 1881 | if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) { |
| 1882 | if (Cond->isZero()) { |
| 1883 | Type *Int8Ty = Type::getInt8Ty(V->getContext()); |
| 1884 | // Insert a new store to null instruction before the load to indicate that |
| 1885 | // this code is not reachable. FIXME: We could insert unreachable |
| 1886 | // instruction directly because we can modify the CFG. |
| 1887 | auto *NewS = new StoreInst(PoisonValue::get(Int8Ty), |
| 1888 | Constant::getNullValue(Int8Ty->getPointerTo()), |
| 1889 | IntrinsicI); |
| 1890 | if (MSSAU) { |
| 1891 | const MemoryUseOrDef *FirstNonDom = nullptr; |
| 1892 | const auto *AL = |
| 1893 | MSSAU->getMemorySSA()->getBlockAccesses(IntrinsicI->getParent()); |
| 1894 | |
| 1895 | // If there are accesses in the current basic block, find the first one |
| 1896 | // that does not come before NewS. The new memory access is inserted |
| 1897 | // after the found access or before the terminator if no such access is |
| 1898 | // found. |
| 1899 | if (AL) { |
| 1900 | for (const auto &Acc : *AL) { |
| 1901 | if (auto *Current = dyn_cast<MemoryUseOrDef>(&Acc)) |
| 1902 | if (!Current->getMemoryInst()->comesBefore(NewS)) { |
| 1903 | FirstNonDom = Current; |
| 1904 | break; |
| 1905 | } |
| 1906 | } |
| 1907 | } |
| 1908 | |
| 1909 | // This added store is to null, so it will never executed and we can |
| 1910 | // just use the LiveOnEntry def as defining access. |
| 1911 | auto *NewDef = |
| 1912 | FirstNonDom ? MSSAU->createMemoryAccessBefore( |
| 1913 | NewS, MSSAU->getMemorySSA()->getLiveOnEntryDef(), |
| 1914 | const_cast<MemoryUseOrDef *>(FirstNonDom)) |
| 1915 | : MSSAU->createMemoryAccessInBB( |
| 1916 | NewS, MSSAU->getMemorySSA()->getLiveOnEntryDef(), |
| 1917 | NewS->getParent(), MemorySSA::BeforeTerminator); |
| 1918 | |
| 1919 | MSSAU->insertDef(cast<MemoryDef>(NewDef), /*RenameUses=*/false); |
| 1920 | } |
| 1921 | } |
| 1922 | if (isAssumeWithEmptyBundle(*IntrinsicI)) { |
| 1923 | markInstructionForDeletion(IntrinsicI); |
| 1924 | return true; |
| 1925 | } |
| 1926 | return false; |
| 1927 | } else if (isa<Constant>(V)) { |
| 1928 | // If it's not false, and constant, it must evaluate to true. This means our |
| 1929 | // assume is assume(true), and thus, pointless, and we don't want to do |
| 1930 | // anything more here. |
| 1931 | return false; |
| 1932 | } |
| 1933 | |
| 1934 | Constant *True = ConstantInt::getTrue(V->getContext()); |
| 1935 | bool Changed = false; |
| 1936 | |
| 1937 | for (BasicBlock *Successor : successors(IntrinsicI->getParent())) { |
| 1938 | BasicBlockEdge Edge(IntrinsicI->getParent(), Successor); |
| 1939 | |
| 1940 | // This property is only true in dominated successors, propagateEquality |
| 1941 | // will check dominance for us. |
| 1942 | Changed |= propagateEquality(V, True, Edge, false); |
| 1943 | } |
| 1944 | |
| 1945 | // We can replace assume value with true, which covers cases like this: |
| 1946 | // call void @llvm.assume(i1 %cmp) |
| 1947 | // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true |
| 1948 | ReplaceOperandsWithMap[V] = True; |
| 1949 | |
| 1950 | // Similarly, after assume(!NotV) we know that NotV == false. |
| 1951 | Value *NotV; |
| 1952 | if (match(V, m_Not(m_Value(NotV)))) |
| 1953 | ReplaceOperandsWithMap[NotV] = ConstantInt::getFalse(V->getContext()); |
| 1954 | |
| 1955 | // If we find an equality fact, canonicalize all dominated uses in this block |
| 1956 | // to one of the two values. We heuristically choice the "oldest" of the |
| 1957 | // two where age is determined by value number. (Note that propagateEquality |
| 1958 | // above handles the cross block case.) |
| 1959 | // |
| 1960 | // Key case to cover are: |
| 1961 | // 1) |
| 1962 | // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen |
| 1963 | // call void @llvm.assume(i1 %cmp) |
| 1964 | // ret float %0 ; will change it to ret float 3.000000e+00 |
| 1965 | // 2) |
| 1966 | // %load = load float, float* %addr |
| 1967 | // %cmp = fcmp oeq float %load, %0 |
| 1968 | // call void @llvm.assume(i1 %cmp) |
| 1969 | // ret float %load ; will change it to ret float %0 |
| 1970 | if (auto *CmpI = dyn_cast<CmpInst>(V)) { |
| 1971 | if (impliesEquivalanceIfTrue(CmpI)) { |
| 1972 | Value *CmpLHS = CmpI->getOperand(0); |
| 1973 | Value *CmpRHS = CmpI->getOperand(1); |
| 1974 | // Heuristically pick the better replacement -- the choice of heuristic |
| 1975 | // isn't terribly important here, but the fact we canonicalize on some |
| 1976 | // replacement is for exposing other simplifications. |
| 1977 | // TODO: pull this out as a helper function and reuse w/existing |
| 1978 | // (slightly different) logic. |
| 1979 | if (isa<Constant>(CmpLHS) && !isa<Constant>(CmpRHS)) |
| 1980 | std::swap(CmpLHS, CmpRHS); |
| 1981 | if (!isa<Instruction>(CmpLHS) && isa<Instruction>(CmpRHS)) |
| 1982 | std::swap(CmpLHS, CmpRHS); |
| 1983 | if ((isa<Argument>(CmpLHS) && isa<Argument>(CmpRHS)) || |
| 1984 | (isa<Instruction>(CmpLHS) && isa<Instruction>(CmpRHS))) { |
| 1985 | // Move the 'oldest' value to the right-hand side, using the value |
| 1986 | // number as a proxy for age. |
| 1987 | uint32_t LVN = VN.lookupOrAdd(CmpLHS); |
| 1988 | uint32_t RVN = VN.lookupOrAdd(CmpRHS); |
| 1989 | if (LVN < RVN) |
| 1990 | std::swap(CmpLHS, CmpRHS); |
| 1991 | } |
| 1992 | |
| 1993 | // Handle degenerate case where we either haven't pruned a dead path or a |
| 1994 | // removed a trivial assume yet. |
| 1995 | if (isa<Constant>(CmpLHS) && isa<Constant>(CmpRHS)) |
| 1996 | return Changed; |
| 1997 | |
| 1998 | LLVM_DEBUG(dbgs() << "Replacing dominated uses of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false) |
| 1999 | << *CmpLHS << " with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false) |
| 2000 | << *CmpRHS << " in block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false) |
| 2001 | << IntrinsicI->getParent()->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false); |
| 2002 | |
| 2003 | |
| 2004 | // Setup the replacement map - this handles uses within the same block |
| 2005 | if (hasUsersIn(CmpLHS, IntrinsicI->getParent())) |
| 2006 | ReplaceOperandsWithMap[CmpLHS] = CmpRHS; |
| 2007 | |
| 2008 | // NOTE: The non-block local cases are handled by the call to |
| 2009 | // propagateEquality above; this block is just about handling the block |
| 2010 | // local cases. TODO: There's a bunch of logic in propagateEqualiy which |
| 2011 | // isn't duplicated for the block local case, can we share it somehow? |
| 2012 | } |
| 2013 | } |
| 2014 | return Changed; |
| 2015 | } |
| 2016 | |
| 2017 | static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { |
| 2018 | patchReplacementInstruction(I, Repl); |
| 2019 | I->replaceAllUsesWith(Repl); |
| 2020 | } |
| 2021 | |
| 2022 | /// Attempt to eliminate a load, first by eliminating it |
| 2023 | /// locally, and then attempting non-local elimination if that fails. |
| 2024 | bool GVNPass::processLoad(LoadInst *L) { |
| 2025 | if (!MD) |
| 2026 | return false; |
| 2027 | |
| 2028 | // This code hasn't been audited for ordered or volatile memory access |
| 2029 | if (!L->isUnordered()) |
| 2030 | return false; |
| 2031 | |
| 2032 | if (L->use_empty()) { |
| 2033 | markInstructionForDeletion(L); |
| 2034 | return true; |
| 2035 | } |
| 2036 | |
| 2037 | // ... to a pointer that has been loaded from before... |
| 2038 | MemDepResult Dep = MD->getDependency(L); |
| 2039 | |
| 2040 | // If it is defined in another block, try harder. |
| 2041 | if (Dep.isNonLocal()) |
| 2042 | return processNonLocalLoad(L); |
| 2043 | |
| 2044 | // Only handle the local case below |
| 2045 | if (!Dep.isLocal()) { |
| 2046 | // This might be a NonFuncLocal or an Unknown |
| 2047 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false) |
| 2048 | // fast print dep, using operator<< on instruction is too slow.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false) |
| 2049 | dbgs() << "GVN: load "; L->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false) |
| 2050 | dbgs() << " has unknown dependence\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false); |
| 2051 | return false; |
| 2052 | } |
| 2053 | |
| 2054 | auto AV = AnalyzeLoadAvailability(L, Dep, L->getPointerOperand()); |
| 2055 | if (!AV) |
| 2056 | return false; |
| 2057 | |
| 2058 | Value *AvailableValue = AV->MaterializeAdjustedValue(L, L, *this); |
| 2059 | |
| 2060 | // MaterializeAdjustedValue is responsible for combining metadata. |
| 2061 | L->replaceAllUsesWith(AvailableValue); |
| 2062 | markInstructionForDeletion(L); |
| 2063 | if (MSSAU) |
| 2064 | MSSAU->removeMemoryAccess(L); |
| 2065 | ++NumGVNLoad; |
| 2066 | reportLoadElim(L, AvailableValue, ORE); |
| 2067 | // Tell MDA to reexamine the reused pointer since we might have more |
| 2068 | // information after forwarding it. |
| 2069 | if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy()) |
| 2070 | MD->invalidateCachedPointerInfo(AvailableValue); |
| 2071 | return true; |
| 2072 | } |
| 2073 | |
| 2074 | /// Return a pair the first field showing the value number of \p Exp and the |
| 2075 | /// second field showing whether it is a value number newly created. |
| 2076 | std::pair<uint32_t, bool> |
| 2077 | GVNPass::ValueTable::assignExpNewValueNum(Expression &Exp) { |
| 2078 | uint32_t &e = expressionNumbering[Exp]; |
| 2079 | bool CreateNewValNum = !e; |
| 2080 | if (CreateNewValNum) { |
| 2081 | Expressions.push_back(Exp); |
| 2082 | if (ExprIdx.size() < nextValueNumber + 1) |
| 2083 | ExprIdx.resize(nextValueNumber * 2); |
| 2084 | e = nextValueNumber; |
| 2085 | ExprIdx[nextValueNumber++] = nextExprNumber++; |
| 2086 | } |
| 2087 | return {e, CreateNewValNum}; |
| 2088 | } |
| 2089 | |
| 2090 | /// Return whether all the values related with the same \p num are |
| 2091 | /// defined in \p BB. |
| 2092 | bool GVNPass::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB, |
| 2093 | GVNPass &Gvn) { |
| 2094 | LeaderTableEntry *Vals = &Gvn.LeaderTable[Num]; |
| 2095 | while (Vals && Vals->BB == BB) |
| 2096 | Vals = Vals->Next; |
| 2097 | return !Vals; |
| 2098 | } |
| 2099 | |
| 2100 | /// Wrap phiTranslateImpl to provide caching functionality. |
| 2101 | uint32_t GVNPass::ValueTable::phiTranslate(const BasicBlock *Pred, |
| 2102 | const BasicBlock *PhiBlock, |
| 2103 | uint32_t Num, GVNPass &Gvn) { |
| 2104 | auto FindRes = PhiTranslateTable.find({Num, Pred}); |
| 2105 | if (FindRes != PhiTranslateTable.end()) |
| 2106 | return FindRes->second; |
| 2107 | uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn); |
| 2108 | PhiTranslateTable.insert({{Num, Pred}, NewNum}); |
| 2109 | return NewNum; |
| 2110 | } |
| 2111 | |
| 2112 | // Return true if the value number \p Num and NewNum have equal value. |
| 2113 | // Return false if the result is unknown. |
| 2114 | bool GVNPass::ValueTable::areCallValsEqual(uint32_t Num, uint32_t NewNum, |
| 2115 | const BasicBlock *Pred, |
| 2116 | const BasicBlock *PhiBlock, |
| 2117 | GVNPass &Gvn) { |
| 2118 | CallInst *Call = nullptr; |
| 2119 | LeaderTableEntry *Vals = &Gvn.LeaderTable[Num]; |
| 2120 | while (Vals) { |
| 2121 | Call = dyn_cast<CallInst>(Vals->Val); |
| 2122 | if (Call && Call->getParent() == PhiBlock) |
| 2123 | break; |
| 2124 | Vals = Vals->Next; |
| 2125 | } |
| 2126 | |
| 2127 | if (AA->doesNotAccessMemory(Call)) |
| 2128 | return true; |
| 2129 | |
| 2130 | if (!MD || !AA->onlyReadsMemory(Call)) |
| 2131 | return false; |
| 2132 | |
| 2133 | MemDepResult local_dep = MD->getDependency(Call); |
| 2134 | if (!local_dep.isNonLocal()) |
| 2135 | return false; |
| 2136 | |
| 2137 | const MemoryDependenceResults::NonLocalDepInfo &deps = |
| 2138 | MD->getNonLocalCallDependency(Call); |
| 2139 | |
| 2140 | // Check to see if the Call has no function local clobber. |
| 2141 | for (const NonLocalDepEntry &D : deps) { |
| 2142 | if (D.getResult().isNonFuncLocal()) |
| 2143 | return true; |
| 2144 | } |
| 2145 | return false; |
| 2146 | } |
| 2147 | |
| 2148 | /// Translate value number \p Num using phis, so that it has the values of |
| 2149 | /// the phis in BB. |
| 2150 | uint32_t GVNPass::ValueTable::phiTranslateImpl(const BasicBlock *Pred, |
| 2151 | const BasicBlock *PhiBlock, |
| 2152 | uint32_t Num, GVNPass &Gvn) { |
| 2153 | if (PHINode *PN = NumberingPhi[Num]) { |
| 2154 | for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { |
| 2155 | if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred) |
| 2156 | if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false)) |
| 2157 | return TransVal; |
| 2158 | } |
| 2159 | return Num; |
| 2160 | } |
| 2161 | |
| 2162 | // If there is any value related with Num is defined in a BB other than |
| 2163 | // PhiBlock, it cannot depend on a phi in PhiBlock without going through |
| 2164 | // a backedge. We can do an early exit in that case to save compile time. |
| 2165 | if (!areAllValsInBB(Num, PhiBlock, Gvn)) |
| 2166 | return Num; |
| 2167 | |
| 2168 | if (Num >= ExprIdx.size() || ExprIdx[Num] == 0) |
| 2169 | return Num; |
| 2170 | Expression Exp = Expressions[ExprIdx[Num]]; |
| 2171 | |
| 2172 | for (unsigned i = 0; i < Exp.varargs.size(); i++) { |
| 2173 | // For InsertValue and ExtractValue, some varargs are index numbers |
| 2174 | // instead of value numbers. Those index numbers should not be |
| 2175 | // translated. |
| 2176 | if ((i > 1 && Exp.opcode == Instruction::InsertValue) || |
| 2177 | (i > 0 && Exp.opcode == Instruction::ExtractValue) || |
| 2178 | (i > 1 && Exp.opcode == Instruction::ShuffleVector)) |
| 2179 | continue; |
| 2180 | Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn); |
| 2181 | } |
| 2182 | |
| 2183 | if (Exp.commutative) { |
| 2184 | assert(Exp.varargs.size() >= 2 && "Unsupported commutative instruction!")(static_cast <bool> (Exp.varargs.size() >= 2 && "Unsupported commutative instruction!") ? void (0) : __assert_fail ("Exp.varargs.size() >= 2 && \"Unsupported commutative instruction!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2184, __extension__ __PRETTY_FUNCTION__ )); |
| 2185 | if (Exp.varargs[0] > Exp.varargs[1]) { |
| 2186 | std::swap(Exp.varargs[0], Exp.varargs[1]); |
| 2187 | uint32_t Opcode = Exp.opcode >> 8; |
| 2188 | if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) |
| 2189 | Exp.opcode = (Opcode << 8) | |
| 2190 | CmpInst::getSwappedPredicate( |
| 2191 | static_cast<CmpInst::Predicate>(Exp.opcode & 255)); |
| 2192 | } |
| 2193 | } |
| 2194 | |
| 2195 | if (uint32_t NewNum = expressionNumbering[Exp]) { |
| 2196 | if (Exp.opcode == Instruction::Call && NewNum != Num) |
| 2197 | return areCallValsEqual(Num, NewNum, Pred, PhiBlock, Gvn) ? NewNum : Num; |
| 2198 | return NewNum; |
| 2199 | } |
| 2200 | return Num; |
| 2201 | } |
| 2202 | |
| 2203 | /// Erase stale entry from phiTranslate cache so phiTranslate can be computed |
| 2204 | /// again. |
| 2205 | void GVNPass::ValueTable::eraseTranslateCacheEntry( |
| 2206 | uint32_t Num, const BasicBlock &CurrBlock) { |
| 2207 | for (const BasicBlock *Pred : predecessors(&CurrBlock)) |
| 2208 | PhiTranslateTable.erase({Num, Pred}); |
| 2209 | } |
| 2210 | |
| 2211 | // In order to find a leader for a given value number at a |
| 2212 | // specific basic block, we first obtain the list of all Values for that number, |
| 2213 | // and then scan the list to find one whose block dominates the block in |
| 2214 | // question. This is fast because dominator tree queries consist of only |
| 2215 | // a few comparisons of DFS numbers. |
| 2216 | Value *GVNPass::findLeader(const BasicBlock *BB, uint32_t num) { |
| 2217 | LeaderTableEntry Vals = LeaderTable[num]; |
| 2218 | if (!Vals.Val) return nullptr; |
| 2219 | |
| 2220 | Value *Val = nullptr; |
| 2221 | if (DT->dominates(Vals.BB, BB)) { |
| 2222 | Val = Vals.Val; |
| 2223 | if (isa<Constant>(Val)) return Val; |
| 2224 | } |
| 2225 | |
| 2226 | LeaderTableEntry* Next = Vals.Next; |
| 2227 | while (Next) { |
| 2228 | if (DT->dominates(Next->BB, BB)) { |
| 2229 | if (isa<Constant>(Next->Val)) return Next->Val; |
| 2230 | if (!Val) Val = Next->Val; |
| 2231 | } |
| 2232 | |
| 2233 | Next = Next->Next; |
| 2234 | } |
| 2235 | |
| 2236 | return Val; |
| 2237 | } |
| 2238 | |
| 2239 | /// There is an edge from 'Src' to 'Dst'. Return |
| 2240 | /// true if every path from the entry block to 'Dst' passes via this edge. In |
| 2241 | /// particular 'Dst' must not be reachable via another edge from 'Src'. |
| 2242 | static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, |
| 2243 | DominatorTree *DT) { |
| 2244 | // While in theory it is interesting to consider the case in which Dst has |
| 2245 | // more than one predecessor, because Dst might be part of a loop which is |
| 2246 | // only reachable from Src, in practice it is pointless since at the time |
| 2247 | // GVN runs all such loops have preheaders, which means that Dst will have |
| 2248 | // been changed to have only one predecessor, namely Src. |
| 2249 | const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); |
| 2250 | assert((!Pred || Pred == E.getStart()) &&(static_cast <bool> ((!Pred || Pred == E.getStart()) && "No edge between these basic blocks!") ? void (0) : __assert_fail ("(!Pred || Pred == E.getStart()) && \"No edge between these basic blocks!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2251, __extension__ __PRETTY_FUNCTION__ )) |
| 2251 | "No edge between these basic blocks!")(static_cast <bool> ((!Pred || Pred == E.getStart()) && "No edge between these basic blocks!") ? void (0) : __assert_fail ("(!Pred || Pred == E.getStart()) && \"No edge between these basic blocks!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2251, __extension__ __PRETTY_FUNCTION__ )); |
| 2252 | return Pred != nullptr; |
| 2253 | } |
| 2254 | |
| 2255 | void GVNPass::assignBlockRPONumber(Function &F) { |
| 2256 | BlockRPONumber.clear(); |
| 2257 | uint32_t NextBlockNumber = 1; |
| 2258 | ReversePostOrderTraversal<Function *> RPOT(&F); |
| 2259 | for (BasicBlock *BB : RPOT) |
| 2260 | BlockRPONumber[BB] = NextBlockNumber++; |
| 2261 | InvalidBlockRPONumbers = false; |
| 2262 | } |
| 2263 | |
| 2264 | bool GVNPass::replaceOperandsForInBlockEquality(Instruction *Instr) const { |
| 2265 | bool Changed = false; |
| 2266 | for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) { |
| 2267 | Value *Operand = Instr->getOperand(OpNum); |
| 2268 | auto it = ReplaceOperandsWithMap.find(Operand); |
| 2269 | if (it != ReplaceOperandsWithMap.end()) { |
| 2270 | LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN replacing: " << *Operand << " with " << *it->second << " in instruction " << *Instr << '\n'; } } while (false) |
| 2271 | << *it->second << " in instruction " << *Instr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN replacing: " << *Operand << " with " << *it->second << " in instruction " << *Instr << '\n'; } } while (false); |
| 2272 | Instr->setOperand(OpNum, it->second); |
| 2273 | Changed = true; |
| 2274 | } |
| 2275 | } |
| 2276 | return Changed; |
| 2277 | } |
| 2278 | |
| 2279 | /// The given values are known to be equal in every block |
| 2280 | /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with |
| 2281 | /// 'RHS' everywhere in the scope. Returns whether a change was made. |
| 2282 | /// If DominatesByEdge is false, then it means that we will propagate the RHS |
| 2283 | /// value starting from the end of Root.Start. |
| 2284 | bool GVNPass::propagateEquality(Value *LHS, Value *RHS, |
| 2285 | const BasicBlockEdge &Root, |
| 2286 | bool DominatesByEdge) { |
| 2287 | SmallVector<std::pair<Value*, Value*>, 4> Worklist; |
| 2288 | Worklist.push_back(std::make_pair(LHS, RHS)); |
| 2289 | bool Changed = false; |
| 2290 | // For speed, compute a conservative fast approximation to |
| 2291 | // DT->dominates(Root, Root.getEnd()); |
| 2292 | const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); |
| 2293 | |
| 2294 | while (!Worklist.empty()) { |
| 2295 | std::pair<Value*, Value*> Item = Worklist.pop_back_val(); |
| 2296 | LHS = Item.first; RHS = Item.second; |
| 2297 | |
| 2298 | if (LHS == RHS) |
| 2299 | continue; |
| 2300 | assert(LHS->getType() == RHS->getType() && "Equality but unequal types!")(static_cast <bool> (LHS->getType() == RHS->getType () && "Equality but unequal types!") ? void (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"Equality but unequal types!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2300, __extension__ __PRETTY_FUNCTION__ )); |
| 2301 | |
| 2302 | // Don't try to propagate equalities between constants. |
| 2303 | if (isa<Constant>(LHS) && isa<Constant>(RHS)) |
| 2304 | continue; |
| 2305 | |
| 2306 | // Prefer a constant on the right-hand side, or an Argument if no constants. |
| 2307 | if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) |
| 2308 | std::swap(LHS, RHS); |
| 2309 | assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!")(static_cast <bool> ((isa<Argument>(LHS) || isa< Instruction>(LHS)) && "Unexpected value!") ? void ( 0) : __assert_fail ("(isa<Argument>(LHS) || isa<Instruction>(LHS)) && \"Unexpected value!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2309, __extension__ __PRETTY_FUNCTION__ )); |
| 2310 | |
| 2311 | // If there is no obvious reason to prefer the left-hand side over the |
| 2312 | // right-hand side, ensure the longest lived term is on the right-hand side, |
| 2313 | // so the shortest lived term will be replaced by the longest lived. |
| 2314 | // This tends to expose more simplifications. |
| 2315 | uint32_t LVN = VN.lookupOrAdd(LHS); |
| 2316 | if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || |
| 2317 | (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { |
| 2318 | // Move the 'oldest' value to the right-hand side, using the value number |
| 2319 | // as a proxy for age. |
| 2320 | uint32_t RVN = VN.lookupOrAdd(RHS); |
| 2321 | if (LVN < RVN) { |
| 2322 | std::swap(LHS, RHS); |
| 2323 | LVN = RVN; |
| 2324 | } |
| 2325 | } |
| 2326 | |
| 2327 | // If value numbering later sees that an instruction in the scope is equal |
| 2328 | // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve |
| 2329 | // the invariant that instructions only occur in the leader table for their |
| 2330 | // own value number (this is used by removeFromLeaderTable), do not do this |
| 2331 | // if RHS is an instruction (if an instruction in the scope is morphed into |
| 2332 | // LHS then it will be turned into RHS by the next GVN iteration anyway, so |
| 2333 | // using the leader table is about compiling faster, not optimizing better). |
| 2334 | // The leader table only tracks basic blocks, not edges. Only add to if we |
| 2335 | // have the simple case where the edge dominates the end. |
| 2336 | if (RootDominatesEnd && !isa<Instruction>(RHS)) |
| 2337 | addToLeaderTable(LVN, RHS, Root.getEnd()); |
| 2338 | |
| 2339 | // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As |
| 2340 | // LHS always has at least one use that is not dominated by Root, this will |
| 2341 | // never do anything if LHS has only one use. |
| 2342 | if (!LHS->hasOneUse()) { |
| 2343 | unsigned NumReplacements = |
| 2344 | DominatesByEdge |
| 2345 | ? replaceDominatedUsesWith(LHS, RHS, *DT, Root) |
| 2346 | : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart()); |
| 2347 | |
| 2348 | Changed |= NumReplacements > 0; |
| 2349 | NumGVNEqProp += NumReplacements; |
| 2350 | // Cached information for anything that uses LHS will be invalid. |
| 2351 | if (MD) |
| 2352 | MD->invalidateCachedPointerInfo(LHS); |
| 2353 | } |
| 2354 | |
| 2355 | // Now try to deduce additional equalities from this one. For example, if |
| 2356 | // the known equality was "(A != B)" == "false" then it follows that A and B |
| 2357 | // are equal in the scope. Only boolean equalities with an explicit true or |
| 2358 | // false RHS are currently supported. |
| 2359 | if (!RHS->getType()->isIntegerTy(1)) |
| 2360 | // Not a boolean equality - bail out. |
| 2361 | continue; |
| 2362 | ConstantInt *CI = dyn_cast<ConstantInt>(RHS); |
| 2363 | if (!CI) |
| 2364 | // RHS neither 'true' nor 'false' - bail out. |
| 2365 | continue; |
| 2366 | // Whether RHS equals 'true'. Otherwise it equals 'false'. |
| 2367 | bool isKnownTrue = CI->isMinusOne(); |
| 2368 | bool isKnownFalse = !isKnownTrue; |
| 2369 | |
| 2370 | // If "A && B" is known true then both A and B are known true. If "A || B" |
| 2371 | // is known false then both A and B are known false. |
| 2372 | Value *A, *B; |
| 2373 | if ((isKnownTrue && match(LHS, m_LogicalAnd(m_Value(A), m_Value(B)))) || |
| 2374 | (isKnownFalse && match(LHS, m_LogicalOr(m_Value(A), m_Value(B))))) { |
| 2375 | Worklist.push_back(std::make_pair(A, RHS)); |
| 2376 | Worklist.push_back(std::make_pair(B, RHS)); |
| 2377 | continue; |
| 2378 | } |
| 2379 | |
| 2380 | // If we are propagating an equality like "(A == B)" == "true" then also |
| 2381 | // propagate the equality A == B. When propagating a comparison such as |
| 2382 | // "(A >= B)" == "true", replace all instances of "A < B" with "false". |
| 2383 | if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) { |
| 2384 | Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); |
| 2385 | |
| 2386 | // If "A == B" is known true, or "A != B" is known false, then replace |
| 2387 | // A with B everywhere in the scope. For floating point operations, we |
| 2388 | // have to be careful since equality does not always imply equivalance. |
| 2389 | if ((isKnownTrue && impliesEquivalanceIfTrue(Cmp)) || |
| 2390 | (isKnownFalse && impliesEquivalanceIfFalse(Cmp))) |
| 2391 | Worklist.push_back(std::make_pair(Op0, Op1)); |
| 2392 | |
| 2393 | // If "A >= B" is known true, replace "A < B" with false everywhere. |
| 2394 | CmpInst::Predicate NotPred = Cmp->getInversePredicate(); |
| 2395 | Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); |
| 2396 | // Since we don't have the instruction "A < B" immediately to hand, work |
| 2397 | // out the value number that it would have and use that to find an |
| 2398 | // appropriate instruction (if any). |
| 2399 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
| 2400 | uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1); |
| 2401 | // If the number we were assigned was brand new then there is no point in |
| 2402 | // looking for an instruction realizing it: there cannot be one! |
| 2403 | if (Num < NextNum) { |
| 2404 | Value *NotCmp = findLeader(Root.getEnd(), Num); |
| 2405 | if (NotCmp && isa<Instruction>(NotCmp)) { |
| 2406 | unsigned NumReplacements = |
| 2407 | DominatesByEdge |
| 2408 | ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root) |
| 2409 | : replaceDominatedUsesWith(NotCmp, NotVal, *DT, |
| 2410 | Root.getStart()); |
| 2411 | Changed |= NumReplacements > 0; |
| 2412 | NumGVNEqProp += NumReplacements; |
| 2413 | // Cached information for anything that uses NotCmp will be invalid. |
| 2414 | if (MD) |
| 2415 | MD->invalidateCachedPointerInfo(NotCmp); |
| 2416 | } |
| 2417 | } |
| 2418 | // Ensure that any instruction in scope that gets the "A < B" value number |
| 2419 | // is replaced with false. |
| 2420 | // The leader table only tracks basic blocks, not edges. Only add to if we |
| 2421 | // have the simple case where the edge dominates the end. |
| 2422 | if (RootDominatesEnd) |
| 2423 | addToLeaderTable(Num, NotVal, Root.getEnd()); |
| 2424 | |
| 2425 | continue; |
| 2426 | } |
| 2427 | } |
| 2428 | |
| 2429 | return Changed; |
| 2430 | } |
| 2431 | |
| 2432 | /// When calculating availability, handle an instruction |
| 2433 | /// by inserting it into the appropriate sets |
| 2434 | bool GVNPass::processInstruction(Instruction *I) { |
| 2435 | // Ignore dbg info intrinsics. |
| 2436 | if (isa<DbgInfoIntrinsic>(I)) |
| 2437 | return false; |
| 2438 | |
| 2439 | // If the instruction can be easily simplified then do so now in preference |
| 2440 | // to value numbering it. Value numbering often exposes redundancies, for |
| 2441 | // example if it determines that %y is equal to %x then the instruction |
| 2442 | // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. |
| 2443 | const DataLayout &DL = I->getModule()->getDataLayout(); |
| 2444 | if (Value *V = simplifyInstruction(I, {DL, TLI, DT, AC})) { |
| 2445 | bool Changed = false; |
| 2446 | if (!I->use_empty()) { |
| 2447 | // Simplification can cause a special instruction to become not special. |
| 2448 | // For example, devirtualization to a willreturn function. |
| 2449 | ICF->removeUsersOf(I); |
| 2450 | I->replaceAllUsesWith(V); |
| 2451 | Changed = true; |
| 2452 | } |
| 2453 | if (isInstructionTriviallyDead(I, TLI)) { |
| 2454 | markInstructionForDeletion(I); |
| 2455 | Changed = true; |
| 2456 | } |
| 2457 | if (Changed) { |
| 2458 | if (MD && V->getType()->isPtrOrPtrVectorTy()) |
| 2459 | MD->invalidateCachedPointerInfo(V); |
| 2460 | ++NumGVNSimpl; |
| 2461 | return true; |
| 2462 | } |
| 2463 | } |
| 2464 | |
| 2465 | if (auto *Assume = dyn_cast<AssumeInst>(I)) |
| 2466 | return processAssumeIntrinsic(Assume); |
| 2467 | |
| 2468 | if (LoadInst *Load = dyn_cast<LoadInst>(I)) { |
| 2469 | if (processLoad(Load)) |
| 2470 | return true; |
| 2471 | |
| 2472 | unsigned Num = VN.lookupOrAdd(Load); |
| 2473 | addToLeaderTable(Num, Load, Load->getParent()); |
| 2474 | return false; |
| 2475 | } |
| 2476 | |
| 2477 | // For conditional branches, we can perform simple conditional propagation on |
| 2478 | // the condition value itself. |
| 2479 | if (BranchInst *BI = dyn_cast<BranchInst>(I)) { |
| 2480 | if (!BI->isConditional()) |
| 2481 | return false; |
| 2482 | |
| 2483 | if (isa<Constant>(BI->getCondition())) |
| 2484 | return processFoldableCondBr(BI); |
| 2485 | |
| 2486 | Value *BranchCond = BI->getCondition(); |
| 2487 | BasicBlock *TrueSucc = BI->getSuccessor(0); |
| 2488 | BasicBlock *FalseSucc = BI->getSuccessor(1); |
| 2489 | // Avoid multiple edges early. |
| 2490 | if (TrueSucc == FalseSucc) |
| 2491 | return false; |
| 2492 | |
| 2493 | BasicBlock *Parent = BI->getParent(); |
| 2494 | bool Changed = false; |
| 2495 | |
| 2496 | Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); |
| 2497 | BasicBlockEdge TrueE(Parent, TrueSucc); |
| 2498 | Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true); |
| 2499 | |
| 2500 | Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); |
| 2501 | BasicBlockEdge FalseE(Parent, FalseSucc); |
| 2502 | Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true); |
| 2503 | |
| 2504 | return Changed; |
| 2505 | } |
| 2506 | |
| 2507 | // For switches, propagate the case values into the case destinations. |
| 2508 | if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { |
| 2509 | Value *SwitchCond = SI->getCondition(); |
| 2510 | BasicBlock *Parent = SI->getParent(); |
| 2511 | bool Changed = false; |
| 2512 | |
| 2513 | // Remember how many outgoing edges there are to every successor. |
| 2514 | SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; |
| 2515 | for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) |
| 2516 | ++SwitchEdges[SI->getSuccessor(i)]; |
| 2517 | |
| 2518 | for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); |
| 2519 | i != e; ++i) { |
| 2520 | BasicBlock *Dst = i->getCaseSuccessor(); |
| 2521 | // If there is only a single edge, propagate the case value into it. |
| 2522 | if (SwitchEdges.lookup(Dst) == 1) { |
| 2523 | BasicBlockEdge E(Parent, Dst); |
| 2524 | Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true); |
| 2525 | } |
| 2526 | } |
| 2527 | return Changed; |
| 2528 | } |
| 2529 | |
| 2530 | // Instructions with void type don't return a value, so there's |
| 2531 | // no point in trying to find redundancies in them. |
| 2532 | if (I->getType()->isVoidTy()) |
| 2533 | return false; |
| 2534 | |
| 2535 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
| 2536 | unsigned Num = VN.lookupOrAdd(I); |
| 2537 | |
| 2538 | // Allocations are always uniquely numbered, so we can save time and memory |
| 2539 | // by fast failing them. |
| 2540 | if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) { |
| 2541 | addToLeaderTable(Num, I, I->getParent()); |
| 2542 | return false; |
| 2543 | } |
| 2544 | |
| 2545 | // If the number we were assigned was a brand new VN, then we don't |
| 2546 | // need to do a lookup to see if the number already exists |
| 2547 | // somewhere in the domtree: it can't! |
| 2548 | if (Num >= NextNum) { |
| 2549 | addToLeaderTable(Num, I, I->getParent()); |
| 2550 | return false; |
| 2551 | } |
| 2552 | |
| 2553 | // Perform fast-path value-number based elimination of values inherited from |
| 2554 | // dominators. |
| 2555 | Value *Repl = findLeader(I->getParent(), Num); |
| 2556 | if (!Repl) { |
| 2557 | // Failure, just remember this instance for future use. |
| 2558 | addToLeaderTable(Num, I, I->getParent()); |
| 2559 | return false; |
| 2560 | } else if (Repl == I) { |
| 2561 | // If I was the result of a shortcut PRE, it might already be in the table |
| 2562 | // and the best replacement for itself. Nothing to do. |
| 2563 | return false; |
| 2564 | } |
| 2565 | |
| 2566 | // Remove it! |
| 2567 | patchAndReplaceAllUsesWith(I, Repl); |
| 2568 | if (MD && Repl->getType()->isPtrOrPtrVectorTy()) |
| 2569 | MD->invalidateCachedPointerInfo(Repl); |
| 2570 | markInstructionForDeletion(I); |
| 2571 | return true; |
| 2572 | } |
| 2573 | |
| 2574 | /// runOnFunction - This is the main transformation entry point for a function. |
| 2575 | bool GVNPass::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT, |
| 2576 | const TargetLibraryInfo &RunTLI, AAResults &RunAA, |
| 2577 | MemoryDependenceResults *RunMD, LoopInfo *LI, |
| 2578 | OptimizationRemarkEmitter *RunORE, MemorySSA *MSSA) { |
| 2579 | AC = &RunAC; |
| 2580 | DT = &RunDT; |
| 2581 | VN.setDomTree(DT); |
| 2582 | TLI = &RunTLI; |
| 2583 | VN.setAliasAnalysis(&RunAA); |
| 2584 | MD = RunMD; |
| 2585 | ImplicitControlFlowTracking ImplicitCFT; |
| 2586 | ICF = &ImplicitCFT; |
| 2587 | this->LI = LI; |
| 2588 | VN.setMemDep(MD); |
| 2589 | ORE = RunORE; |
| 2590 | InvalidBlockRPONumbers = true; |
| 2591 | MemorySSAUpdater Updater(MSSA); |
| 2592 | MSSAU = MSSA ? &Updater : nullptr; |
| 2593 | |
| 2594 | bool Changed = false; |
| 2595 | bool ShouldContinue = true; |
| 2596 | |
| 2597 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); |
| 2598 | // Merge unconditional branches, allowing PRE to catch more |
| 2599 | // optimization opportunities. |
| 2600 | for (BasicBlock &BB : llvm::make_early_inc_range(F)) { |
| 2601 | bool removedBlock = MergeBlockIntoPredecessor(&BB, &DTU, LI, MSSAU, MD); |
| 2602 | if (removedBlock) |
| 2603 | ++NumGVNBlocks; |
| 2604 | |
| 2605 | Changed |= removedBlock; |
| 2606 | } |
| 2607 | |
| 2608 | unsigned Iteration = 0; |
| 2609 | while (ShouldContinue) { |
| 2610 | LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN iteration: " << Iteration << "\n"; } } while (false); |
| 2611 | (void) Iteration; |
| 2612 | ShouldContinue = iterateOnFunction(F); |
| 2613 | Changed |= ShouldContinue; |
| 2614 | ++Iteration; |
| 2615 | } |
| 2616 | |
| 2617 | if (isPREEnabled()) { |
| 2618 | // Fabricate val-num for dead-code in order to suppress assertion in |
| 2619 | // performPRE(). |
| 2620 | assignValNumForDeadCode(); |
| 2621 | bool PREChanged = true; |
| 2622 | while (PREChanged) { |
| 2623 | PREChanged = performPRE(F); |
| 2624 | Changed |= PREChanged; |
| 2625 | } |
| 2626 | } |
| 2627 | |
| 2628 | // FIXME: Should perform GVN again after PRE does something. PRE can move |
| 2629 | // computations into blocks where they become fully redundant. Note that |
| 2630 | // we can't do this until PRE's critical edge splitting updates memdep. |
| 2631 | // Actually, when this happens, we should just fully integrate PRE into GVN. |
| 2632 | |
| 2633 | cleanupGlobalSets(); |
| 2634 | // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each |
| 2635 | // iteration. |
| 2636 | DeadBlocks.clear(); |
| 2637 | |
| 2638 | if (MSSA && VerifyMemorySSA) |
| 2639 | MSSA->verifyMemorySSA(); |
| 2640 | |
| 2641 | return Changed; |
| 2642 | } |
| 2643 | |
| 2644 | bool GVNPass::processBlock(BasicBlock *BB) { |
| 2645 | // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function |
| 2646 | // (and incrementing BI before processing an instruction). |
| 2647 | assert(InstrsToErase.empty() &&(static_cast <bool> (InstrsToErase.empty() && "We expect InstrsToErase to be empty across iterations" ) ? void (0) : __assert_fail ("InstrsToErase.empty() && \"We expect InstrsToErase to be empty across iterations\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2648, __extension__ __PRETTY_FUNCTION__ )) |
| 2648 | "We expect InstrsToErase to be empty across iterations")(static_cast <bool> (InstrsToErase.empty() && "We expect InstrsToErase to be empty across iterations" ) ? void (0) : __assert_fail ("InstrsToErase.empty() && \"We expect InstrsToErase to be empty across iterations\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2648, __extension__ __PRETTY_FUNCTION__ )); |
| 2649 | if (DeadBlocks.count(BB)) |
| 2650 | return false; |
| 2651 | |
| 2652 | // Clearing map before every BB because it can be used only for single BB. |
| 2653 | ReplaceOperandsWithMap.clear(); |
| 2654 | bool ChangedFunction = false; |
| 2655 | |
| 2656 | // Since we may not have visited the input blocks of the phis, we can't |
| 2657 | // use our normal hash approach for phis. Instead, simply look for |
| 2658 | // obvious duplicates. The first pass of GVN will tend to create |
| 2659 | // identical phis, and the second or later passes can eliminate them. |
| 2660 | ChangedFunction |= EliminateDuplicatePHINodes(BB); |
| 2661 | |
| 2662 | for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); |
| 2663 | BI != BE;) { |
| 2664 | if (!ReplaceOperandsWithMap.empty()) |
| 2665 | ChangedFunction |= replaceOperandsForInBlockEquality(&*BI); |
| 2666 | ChangedFunction |= processInstruction(&*BI); |
| 2667 | |
| 2668 | if (InstrsToErase.empty()) { |
| 2669 | ++BI; |
| 2670 | continue; |
| 2671 | } |
| 2672 | |
| 2673 | // If we need some instructions deleted, do it now. |
| 2674 | NumGVNInstr += InstrsToErase.size(); |
| 2675 | |
| 2676 | // Avoid iterator invalidation. |
| 2677 | bool AtStart = BI == BB->begin(); |
| 2678 | if (!AtStart) |
| 2679 | --BI; |
| 2680 | |
| 2681 | for (auto *I : InstrsToErase) { |
| 2682 | assert(I->getParent() == BB && "Removing instruction from wrong block?")(static_cast <bool> (I->getParent() == BB && "Removing instruction from wrong block?") ? void (0) : __assert_fail ("I->getParent() == BB && \"Removing instruction from wrong block?\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2682, __extension__ __PRETTY_FUNCTION__ )); |
| 2683 | LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN removed: " << *I << '\n'; } } while (false); |
| 2684 | salvageKnowledge(I, AC); |
| 2685 | salvageDebugInfo(*I); |
| 2686 | if (MD) MD->removeInstruction(I); |
| 2687 | if (MSSAU) |
| 2688 | MSSAU->removeMemoryAccess(I); |
| 2689 | LLVM_DEBUG(verifyRemoved(I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { verifyRemoved(I); } } while (false); |
| 2690 | ICF->removeInstruction(I); |
| 2691 | I->eraseFromParent(); |
| 2692 | } |
| 2693 | InstrsToErase.clear(); |
| 2694 | |
| 2695 | if (AtStart) |
| 2696 | BI = BB->begin(); |
| 2697 | else |
| 2698 | ++BI; |
| 2699 | } |
| 2700 | |
| 2701 | return ChangedFunction; |
| 2702 | } |
| 2703 | |
| 2704 | // Instantiate an expression in a predecessor that lacked it. |
| 2705 | bool GVNPass::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred, |
| 2706 | BasicBlock *Curr, unsigned int ValNo) { |
| 2707 | // Because we are going top-down through the block, all value numbers |
| 2708 | // will be available in the predecessor by the time we need them. Any |
| 2709 | // that weren't originally present will have been instantiated earlier |
| 2710 | // in this loop. |
| 2711 | bool success = true; |
| 2712 | for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) { |
| 2713 | Value *Op = Instr->getOperand(i); |
| 2714 | if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) |
| 2715 | continue; |
| 2716 | // This could be a newly inserted instruction, in which case, we won't |
| 2717 | // find a value number, and should give up before we hurt ourselves. |
| 2718 | // FIXME: Rewrite the infrastructure to let it easier to value number |
| 2719 | // and process newly inserted instructions. |
| 2720 | if (!VN.exists(Op)) { |
| 2721 | success = false; |
| 2722 | break; |
| 2723 | } |
| 2724 | uint32_t TValNo = |
| 2725 | VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this); |
| 2726 | if (Value *V = findLeader(Pred, TValNo)) { |
| 2727 | Instr->setOperand(i, V); |
| 2728 | } else { |
| 2729 | success = false; |
| 2730 | break; |
| 2731 | } |
| 2732 | } |
| 2733 | |
| 2734 | // Fail out if we encounter an operand that is not available in |
| 2735 | // the PRE predecessor. This is typically because of loads which |
| 2736 | // are not value numbered precisely. |
| 2737 | if (!success) |
| 2738 | return false; |
| 2739 | |
| 2740 | Instr->insertBefore(Pred->getTerminator()); |
| 2741 | Instr->setName(Instr->getName() + ".pre"); |
| 2742 | Instr->setDebugLoc(Instr->getDebugLoc()); |
| 2743 | |
| 2744 | ICF->insertInstructionTo(Instr, Pred); |
| 2745 | |
| 2746 | unsigned Num = VN.lookupOrAdd(Instr); |
| 2747 | VN.add(Instr, Num); |
| 2748 | |
| 2749 | // Update the availability map to include the new instruction. |
| 2750 | addToLeaderTable(Num, Instr, Pred); |
| 2751 | return true; |
| 2752 | } |
| 2753 | |
| 2754 | bool GVNPass::performScalarPRE(Instruction *CurInst) { |
| 2755 | if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() || |
| 2756 | isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() || |
| 2757 | CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || |
| 2758 | isa<DbgInfoIntrinsic>(CurInst)) |
| 2759 | return false; |
| 2760 | |
| 2761 | // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from |
| 2762 | // sinking the compare again, and it would force the code generator to |
| 2763 | // move the i1 from processor flags or predicate registers into a general |
| 2764 | // purpose register. |
| 2765 | if (isa<CmpInst>(CurInst)) |
| 2766 | return false; |
| 2767 | |
| 2768 | // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from |
| 2769 | // sinking the addressing mode computation back to its uses. Extending the |
| 2770 | // GEP's live range increases the register pressure, and therefore it can |
| 2771 | // introduce unnecessary spills. |
| 2772 | // |
| 2773 | // This doesn't prevent Load PRE. PHI translation will make the GEP available |
| 2774 | // to the load by moving it to the predecessor block if necessary. |
| 2775 | if (isa<GetElementPtrInst>(CurInst)) |
| 2776 | return false; |
| 2777 | |
| 2778 | if (auto *CallB = dyn_cast<CallBase>(CurInst)) { |
| 2779 | // We don't currently value number ANY inline asm calls. |
| 2780 | if (CallB->isInlineAsm()) |
| 2781 | return false; |
| 2782 | // Don't do PRE on convergent calls. |
| 2783 | if (CallB->isConvergent()) |
| 2784 | return false; |
| 2785 | } |
| 2786 | |
| 2787 | uint32_t ValNo = VN.lookup(CurInst); |
| 2788 | |
| 2789 | // Look for the predecessors for PRE opportunities. We're |
| 2790 | // only trying to solve the basic diamond case, where |
| 2791 | // a value is computed in the successor and one predecessor, |
| 2792 | // but not the other. We also explicitly disallow cases |
| 2793 | // where the successor is its own predecessor, because they're |
| 2794 | // more complicated to get right. |
| 2795 | unsigned NumWith = 0; |
| 2796 | unsigned NumWithout = 0; |
| 2797 | BasicBlock *PREPred = nullptr; |
| 2798 | BasicBlock *CurrentBlock = CurInst->getParent(); |
| 2799 | |
| 2800 | // Update the RPO numbers for this function. |
| 2801 | if (InvalidBlockRPONumbers) |
| 2802 | assignBlockRPONumber(*CurrentBlock->getParent()); |
| 2803 | |
| 2804 | SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap; |
| 2805 | for (BasicBlock *P : predecessors(CurrentBlock)) { |
| 2806 | // We're not interested in PRE where blocks with predecessors that are |
| 2807 | // not reachable. |
| 2808 | if (!DT->isReachableFromEntry(P)) { |
| 2809 | NumWithout = 2; |
| 2810 | break; |
| 2811 | } |
| 2812 | // It is not safe to do PRE when P->CurrentBlock is a loop backedge. |
| 2813 | assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&(static_cast <bool> (BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && "Invalid BlockRPONumber map." ) ? void (0) : __assert_fail ("BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && \"Invalid BlockRPONumber map.\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2814, __extension__ __PRETTY_FUNCTION__ )) |
| 2814 | "Invalid BlockRPONumber map.")(static_cast <bool> (BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && "Invalid BlockRPONumber map." ) ? void (0) : __assert_fail ("BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && \"Invalid BlockRPONumber map.\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2814, __extension__ __PRETTY_FUNCTION__ )); |
| 2815 | if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock]) { |
| 2816 | NumWithout = 2; |
| 2817 | break; |
| 2818 | } |
| 2819 | |
| 2820 | uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this); |
| 2821 | Value *predV = findLeader(P, TValNo); |
| 2822 | if (!predV) { |
| 2823 | predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P)); |
| 2824 | PREPred = P; |
| 2825 | ++NumWithout; |
| 2826 | } else if (predV == CurInst) { |
| 2827 | /* CurInst dominates this predecessor. */ |
| 2828 | NumWithout = 2; |
| 2829 | break; |
| 2830 | } else { |
| 2831 | predMap.push_back(std::make_pair(predV, P)); |
| 2832 | ++NumWith; |
| 2833 | } |
| 2834 | } |
| 2835 | |
| 2836 | // Don't do PRE when it might increase code size, i.e. when |
| 2837 | // we would need to insert instructions in more than one pred. |
| 2838 | if (NumWithout > 1 || NumWith == 0) |
| 2839 | return false; |
| 2840 | |
| 2841 | // We may have a case where all predecessors have the instruction, |
| 2842 | // and we just need to insert a phi node. Otherwise, perform |
| 2843 | // insertion. |
| 2844 | Instruction *PREInstr = nullptr; |
| 2845 | |
| 2846 | if (NumWithout != 0) { |
| 2847 | if (!isSafeToSpeculativelyExecute(CurInst)) { |
| 2848 | // It is only valid to insert a new instruction if the current instruction |
| 2849 | // is always executed. An instruction with implicit control flow could |
| 2850 | // prevent us from doing it. If we cannot speculate the execution, then |
| 2851 | // PRE should be prohibited. |
| 2852 | if (ICF->isDominatedByICFIFromSameBlock(CurInst)) |
| 2853 | return false; |
| 2854 | } |
| 2855 | |
| 2856 | // Don't do PRE across indirect branch. |
| 2857 | if (isa<IndirectBrInst>(PREPred->getTerminator())) |
| 2858 | return false; |
| 2859 | |
| 2860 | // We can't do PRE safely on a critical edge, so instead we schedule |
| 2861 | // the edge to be split and perform the PRE the next time we iterate |
| 2862 | // on the function. |
| 2863 | unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); |
| 2864 | if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { |
| 2865 | toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); |
| 2866 | return false; |
| 2867 | } |
| 2868 | // We need to insert somewhere, so let's give it a shot |
| 2869 | PREInstr = CurInst->clone(); |
| 2870 | if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) { |
| 2871 | // If we failed insertion, make sure we remove the instruction. |
| 2872 | LLVM_DEBUG(verifyRemoved(PREInstr))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { verifyRemoved(PREInstr); } } while (false); |
| 2873 | PREInstr->deleteValue(); |
| 2874 | return false; |
| 2875 | } |
| 2876 | } |
| 2877 | |
| 2878 | // Either we should have filled in the PRE instruction, or we should |
| 2879 | // not have needed insertions. |
| 2880 | assert(PREInstr != nullptr || NumWithout == 0)(static_cast <bool> (PREInstr != nullptr || NumWithout == 0) ? void (0) : __assert_fail ("PREInstr != nullptr || NumWithout == 0" , "llvm/lib/Transforms/Scalar/GVN.cpp", 2880, __extension__ __PRETTY_FUNCTION__ )); |
| 2881 | |
| 2882 | ++NumGVNPRE; |
| 2883 | |
| 2884 | // Create a PHI to make the value available in this block. |
| 2885 | PHINode *Phi = |
| 2886 | PHINode::Create(CurInst->getType(), predMap.size(), |
| 2887 | CurInst->getName() + ".pre-phi", &CurrentBlock->front()); |
| 2888 | for (unsigned i = 0, e = predMap.size(); i != e; ++i) { |
| 2889 | if (Value *V = predMap[i].first) { |
| 2890 | // If we use an existing value in this phi, we have to patch the original |
| 2891 | // value because the phi will be used to replace a later value. |
| 2892 | patchReplacementInstruction(CurInst, V); |
| 2893 | Phi->addIncoming(V, predMap[i].second); |
| 2894 | } else |
| 2895 | Phi->addIncoming(PREInstr, PREPred); |
| 2896 | } |
| 2897 | |
| 2898 | VN.add(Phi, ValNo); |
| 2899 | // After creating a new PHI for ValNo, the phi translate result for ValNo will |
| 2900 | // be changed, so erase the related stale entries in phi translate cache. |
| 2901 | VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock); |
| 2902 | addToLeaderTable(ValNo, Phi, CurrentBlock); |
| 2903 | Phi->setDebugLoc(CurInst->getDebugLoc()); |
| 2904 | CurInst->replaceAllUsesWith(Phi); |
| 2905 | if (MD && Phi->getType()->isPtrOrPtrVectorTy()) |
| 2906 | MD->invalidateCachedPointerInfo(Phi); |
| 2907 | VN.erase(CurInst); |
| 2908 | removeFromLeaderTable(ValNo, CurInst, CurrentBlock); |
| 2909 | |
| 2910 | LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN PRE removed: " << *CurInst << '\n'; } } while (false); |
| 2911 | if (MD) |
| 2912 | MD->removeInstruction(CurInst); |
| 2913 | if (MSSAU) |
| 2914 | MSSAU->removeMemoryAccess(CurInst); |
| 2915 | LLVM_DEBUG(verifyRemoved(CurInst))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { verifyRemoved(CurInst); } } while (false); |
| 2916 | // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes |
| 2917 | // some assertion failures. |
| 2918 | ICF->removeInstruction(CurInst); |
| 2919 | CurInst->eraseFromParent(); |
| 2920 | ++NumGVNInstr; |
| 2921 | |
| 2922 | return true; |
| 2923 | } |
| 2924 | |
| 2925 | /// Perform a purely local form of PRE that looks for diamond |
| 2926 | /// control flow patterns and attempts to perform simple PRE at the join point. |
| 2927 | bool GVNPass::performPRE(Function &F) { |
| 2928 | bool Changed = false; |
| 2929 | for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) { |
| 2930 | // Nothing to PRE in the entry block. |
| 2931 | if (CurrentBlock == &F.getEntryBlock()) |
| 2932 | continue; |
| 2933 | |
| 2934 | // Don't perform PRE on an EH pad. |
| 2935 | if (CurrentBlock->isEHPad()) |
| 2936 | continue; |
| 2937 | |
| 2938 | for (BasicBlock::iterator BI = CurrentBlock->begin(), |
| 2939 | BE = CurrentBlock->end(); |
| 2940 | BI != BE;) { |
| 2941 | Instruction *CurInst = &*BI++; |
| 2942 | Changed |= performScalarPRE(CurInst); |
| 2943 | } |
| 2944 | } |
| 2945 | |
| 2946 | if (splitCriticalEdges()) |
| 2947 | Changed = true; |
| 2948 | |
| 2949 | return Changed; |
| 2950 | } |
| 2951 | |
| 2952 | /// Split the critical edge connecting the given two blocks, and return |
| 2953 | /// the block inserted to the critical edge. |
| 2954 | BasicBlock *GVNPass::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { |
| 2955 | // GVN does not require loop-simplify, do not try to preserve it if it is not |
| 2956 | // possible. |
| 2957 | BasicBlock *BB = SplitCriticalEdge( |
| 2958 | Pred, Succ, |
| 2959 | CriticalEdgeSplittingOptions(DT, LI, MSSAU).unsetPreserveLoopSimplify()); |
| 2960 | if (BB) { |
| 2961 | if (MD) |
| 2962 | MD->invalidateCachedPredecessors(); |
| 2963 | InvalidBlockRPONumbers = true; |
| 2964 | } |
| 2965 | return BB; |
| 2966 | } |
| 2967 | |
| 2968 | /// Split critical edges found during the previous |
| 2969 | /// iteration that may enable further optimization. |
| 2970 | bool GVNPass::splitCriticalEdges() { |
| 2971 | if (toSplit.empty()) |
| 2972 | return false; |
| 2973 | |
| 2974 | bool Changed = false; |
| 2975 | do { |
| 2976 | std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val(); |
| 2977 | Changed |= SplitCriticalEdge(Edge.first, Edge.second, |
| 2978 | CriticalEdgeSplittingOptions(DT, LI, MSSAU)) != |
| 2979 | nullptr; |
| 2980 | } while (!toSplit.empty()); |
| 2981 | if (Changed) { |
| 2982 | if (MD) |
| 2983 | MD->invalidateCachedPredecessors(); |
| 2984 | InvalidBlockRPONumbers = true; |
| 2985 | } |
| 2986 | return Changed; |
| 2987 | } |
| 2988 | |
| 2989 | /// Executes one iteration of GVN |
| 2990 | bool GVNPass::iterateOnFunction(Function &F) { |
| 2991 | cleanupGlobalSets(); |
| 2992 | |
| 2993 | // Top-down walk of the dominator tree |
| 2994 | bool Changed = false; |
| 2995 | // Needed for value numbering with phi construction to work. |
| 2996 | // RPOT walks the graph in its constructor and will not be invalidated during |
| 2997 | // processBlock. |
| 2998 | ReversePostOrderTraversal<Function *> RPOT(&F); |
| 2999 | |
| 3000 | for (BasicBlock *BB : RPOT) |
| 3001 | Changed |= processBlock(BB); |
| 3002 | |
| 3003 | return Changed; |
| 3004 | } |
| 3005 | |
| 3006 | void GVNPass::cleanupGlobalSets() { |
| 3007 | VN.clear(); |
| 3008 | LeaderTable.clear(); |
| 3009 | BlockRPONumber.clear(); |
| 3010 | TableAllocator.Reset(); |
| 3011 | ICF->clear(); |
| 3012 | InvalidBlockRPONumbers = true; |
| 3013 | } |
| 3014 | |
| 3015 | /// Verify that the specified instruction does not occur in our |
| 3016 | /// internal data structures. |
| 3017 | void GVNPass::verifyRemoved(const Instruction *Inst) const { |
| 3018 | VN.verifyRemoved(Inst); |
| 3019 | |
| 3020 | // Walk through the value number scope to make sure the instruction isn't |
| 3021 | // ferreted away in it. |
| 3022 | for (const auto &I : LeaderTable) { |
| 3023 | const LeaderTableEntry *Node = &I.second; |
| 3024 | assert(Node->Val != Inst && "Inst still in value numbering scope!")(static_cast <bool> (Node->Val != Inst && "Inst still in value numbering scope!" ) ? void (0) : __assert_fail ("Node->Val != Inst && \"Inst still in value numbering scope!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 3024, __extension__ __PRETTY_FUNCTION__ )); |
| 3025 | |
| 3026 | while (Node->Next) { |
| 3027 | Node = Node->Next; |
| 3028 | assert(Node->Val != Inst && "Inst still in value numbering scope!")(static_cast <bool> (Node->Val != Inst && "Inst still in value numbering scope!" ) ? void (0) : __assert_fail ("Node->Val != Inst && \"Inst still in value numbering scope!\"" , "llvm/lib/Transforms/Scalar/GVN.cpp", 3028, __extension__ __PRETTY_FUNCTION__ )); |
| 3029 | } |
| 3030 | } |
| 3031 | } |
| 3032 | |
| 3033 | /// BB is declared dead, which implied other blocks become dead as well. This |
| 3034 | /// function is to add all these blocks to "DeadBlocks". For the dead blocks' |
| 3035 | /// live successors, update their phi nodes by replacing the operands |
| 3036 | /// corresponding to dead blocks with UndefVal. |
| 3037 | void GVNPass::addDeadBlock(BasicBlock *BB) { |
| 3038 | SmallVector<BasicBlock *, 4> NewDead; |
| 3039 | SmallSetVector<BasicBlock *, 4> DF; |
| 3040 | |
| 3041 | NewDead.push_back(BB); |
| 3042 | while (!NewDead.empty()) { |
| 3043 | BasicBlock *D = NewDead.pop_back_val(); |
| 3044 | if (DeadBlocks.count(D)) |
| 3045 | continue; |
| 3046 | |
| 3047 | // All blocks dominated by D are dead. |
| 3048 | SmallVector<BasicBlock *, 8> Dom; |
| 3049 | DT->getDescendants(D, Dom); |
| 3050 | DeadBlocks.insert(Dom.begin(), Dom.end()); |
| 3051 | |
| 3052 | // Figure out the dominance-frontier(D). |
| 3053 | for (BasicBlock *B : Dom) { |
| 3054 | for (BasicBlock *S : successors(B)) { |
| 3055 | if (DeadBlocks.count(S)) |
| 3056 | continue; |
| 3057 | |
| 3058 | bool AllPredDead = true; |
| 3059 | for (BasicBlock *P : predecessors(S)) |
| 3060 | if (!DeadBlocks.count(P)) { |
| 3061 | AllPredDead = false; |
| 3062 | break; |
| 3063 | } |
| 3064 | |
| 3065 | if (!AllPredDead) { |
| 3066 | // S could be proved dead later on. That is why we don't update phi |
| 3067 | // operands at this moment. |
| 3068 | DF.insert(S); |
| 3069 | } else { |
| 3070 | // While S is not dominated by D, it is dead by now. This could take |
| 3071 | // place if S already have a dead predecessor before D is declared |
| 3072 | // dead. |
| 3073 | NewDead.push_back(S); |
| 3074 | } |
| 3075 | } |
| 3076 | } |
| 3077 | } |
| 3078 | |
| 3079 | // For the dead blocks' live successors, update their phi nodes by replacing |
| 3080 | // the operands corresponding to dead blocks with UndefVal. |
| 3081 | for (BasicBlock *B : DF) { |
| 3082 | if (DeadBlocks.count(B)) |
| 3083 | continue; |
| 3084 | |
| 3085 | // First, split the critical edges. This might also create additional blocks |
| 3086 | // to preserve LoopSimplify form and adjust edges accordingly. |
| 3087 | SmallVector<BasicBlock *, 4> Preds(predecessors(B)); |
| 3088 | for (BasicBlock *P : Preds) { |
| 3089 | if (!DeadBlocks.count(P)) |
| 3090 | continue; |
| 3091 | |
| 3092 | if (llvm::is_contained(successors(P), B) && |
| 3093 | isCriticalEdge(P->getTerminator(), B)) { |
| 3094 | if (BasicBlock *S = splitCriticalEdges(P, B)) |
| 3095 | DeadBlocks.insert(P = S); |
Although the value stored to 'P' is used in the enclosing expression, the value is never actually read from 'P' | |
| 3096 | } |
| 3097 | } |
| 3098 | |
| 3099 | // Now poison the incoming values from the dead predecessors. |
| 3100 | for (BasicBlock *P : predecessors(B)) { |
| 3101 | if (!DeadBlocks.count(P)) |
| 3102 | continue; |
| 3103 | for (PHINode &Phi : B->phis()) { |
| 3104 | Phi.setIncomingValueForBlock(P, PoisonValue::get(Phi.getType())); |
| 3105 | if (MD) |
| 3106 | MD->invalidateCachedPointerInfo(&Phi); |
| 3107 | } |
| 3108 | } |
| 3109 | } |
| 3110 | } |
| 3111 | |
| 3112 | // If the given branch is recognized as a foldable branch (i.e. conditional |
| 3113 | // branch with constant condition), it will perform following analyses and |
| 3114 | // transformation. |
| 3115 | // 1) If the dead out-coming edge is a critical-edge, split it. Let |
| 3116 | // R be the target of the dead out-coming edge. |
| 3117 | // 1) Identify the set of dead blocks implied by the branch's dead outcoming |
| 3118 | // edge. The result of this step will be {X| X is dominated by R} |
| 3119 | // 2) Identify those blocks which haves at least one dead predecessor. The |
| 3120 | // result of this step will be dominance-frontier(R). |
| 3121 | // 3) Update the PHIs in DF(R) by replacing the operands corresponding to |
| 3122 | // dead blocks with "UndefVal" in an hope these PHIs will optimized away. |
| 3123 | // |
| 3124 | // Return true iff *NEW* dead code are found. |
| 3125 | bool GVNPass::processFoldableCondBr(BranchInst *BI) { |
| 3126 | if (!BI || BI->isUnconditional()) |
| 3127 | return false; |
| 3128 | |
| 3129 | // If a branch has two identical successors, we cannot declare either dead. |
| 3130 | if (BI->getSuccessor(0) == BI->getSuccessor(1)) |
| 3131 | return false; |
| 3132 | |
| 3133 | ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition()); |
| 3134 | if (!Cond) |
| 3135 | return false; |
| 3136 | |
| 3137 | BasicBlock *DeadRoot = |
| 3138 | Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0); |
| 3139 | if (DeadBlocks.count(DeadRoot)) |
| 3140 | return false; |
| 3141 | |
| 3142 | if (!DeadRoot->getSinglePredecessor()) |
| 3143 | DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot); |
| 3144 | |
| 3145 | addDeadBlock(DeadRoot); |
| 3146 | return true; |
| 3147 | } |
| 3148 | |
| 3149 | // performPRE() will trigger assert if it comes across an instruction without |
| 3150 | // associated val-num. As it normally has far more live instructions than dead |
| 3151 | // instructions, it makes more sense just to "fabricate" a val-number for the |
| 3152 | // dead code than checking if instruction involved is dead or not. |
| 3153 | void GVNPass::assignValNumForDeadCode() { |
| 3154 | for (BasicBlock *BB : DeadBlocks) { |
| 3155 | for (Instruction &Inst : *BB) { |
| 3156 | unsigned ValNum = VN.lookupOrAdd(&Inst); |
| 3157 | addToLeaderTable(ValNum, &Inst, BB); |
| 3158 | } |
| 3159 | } |
| 3160 | } |
| 3161 | |
| 3162 | class llvm::gvn::GVNLegacyPass : public FunctionPass { |
| 3163 | public: |
| 3164 | static char ID; // Pass identification, replacement for typeid |
| 3165 | |
| 3166 | explicit GVNLegacyPass(bool NoMemDepAnalysis = !GVNEnableMemDep) |
| 3167 | : FunctionPass(ID), Impl(GVNOptions().setMemDep(!NoMemDepAnalysis)) { |
| 3168 | initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry()); |
| 3169 | } |
| 3170 | |
| 3171 | bool runOnFunction(Function &F) override { |
| 3172 | if (skipFunction(F)) |
| 3173 | return false; |
| 3174 | |
| 3175 | auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>(); |
| 3176 | |
| 3177 | auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
| 3178 | return Impl.runImpl( |
| 3179 | F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F), |
| 3180 | getAnalysis<DominatorTreeWrapperPass>().getDomTree(), |
| 3181 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F), |
| 3182 | getAnalysis<AAResultsWrapperPass>().getAAResults(), |
| 3183 | Impl.isMemDepEnabled() |
| 3184 | ? &getAnalysis<MemoryDependenceWrapperPass>().getMemDep() |
| 3185 | : nullptr, |
| 3186 | LIWP ? &LIWP->getLoopInfo() : nullptr, |
| 3187 | &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(), |
| 3188 | MSSAWP ? &MSSAWP->getMSSA() : nullptr); |
| 3189 | } |
| 3190 | |
| 3191 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 3192 | AU.addRequired<AssumptionCacheTracker>(); |
| 3193 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 3194 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 3195 | AU.addRequired<LoopInfoWrapperPass>(); |
| 3196 | if (Impl.isMemDepEnabled()) |
| 3197 | AU.addRequired<MemoryDependenceWrapperPass>(); |
| 3198 | AU.addRequired<AAResultsWrapperPass>(); |
| 3199 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 3200 | AU.addPreserved<GlobalsAAWrapperPass>(); |
| 3201 | AU.addPreserved<TargetLibraryInfoWrapperPass>(); |
| 3202 | AU.addPreserved<LoopInfoWrapperPass>(); |
| 3203 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); |
| 3204 | AU.addPreserved<MemorySSAWrapperPass>(); |
| 3205 | } |
| 3206 | |
| 3207 | private: |
| 3208 | GVNPass Impl; |
| 3209 | }; |
| 3210 | |
| 3211 | char GVNLegacyPass::ID = 0; |
| 3212 | |
| 3213 | INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)static void *initializeGVNLegacyPassPassOnce(PassRegistry & Registry) { |
| 3214 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); |
| 3215 | INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)initializeMemoryDependenceWrapperPassPass(Registry); |
| 3216 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); |
| 3217 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
| 3218 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); |
| 3219 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); |
| 3220 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); |
| 3221 | INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)PassInfo *PI = new PassInfo( "Global Value Numbering", "gvn", &GVNLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <GVNLegacyPass>), false, false); Registry.registerPass( *PI, true); return PI; } static llvm::once_flag InitializeGVNLegacyPassPassFlag ; void llvm::initializeGVNLegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeGVNLegacyPassPassFlag, initializeGVNLegacyPassPassOnce , std::ref(Registry)); } |
| 3222 | |
| 3223 | // The public interface to this file... |
| 3224 | FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) { |
| 3225 | return new GVNLegacyPass(NoMemDepAnalysis); |
| 3226 | } |