| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp |
| Warning: | line 579, column 28 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- DFAJumpThreading.cpp - Threads a switch statement inside a loop ----===// | |||
| 2 | // | |||
| 3 | // The LLVM Compiler Infrastructure | |||
| 4 | // | |||
| 5 | // This file is distributed under the University of Illinois Open Source | |||
| 6 | // License. See LICENSE.TXT for details. | |||
| 7 | // | |||
| 8 | //===----------------------------------------------------------------------===// | |||
| 9 | // | |||
| 10 | // Transform each threading path to effectively jump thread the DFA. For | |||
| 11 | // example, the CFG below could be transformed as follows, where the cloned | |||
| 12 | // blocks unconditionally branch to the next correct case based on what is | |||
| 13 | // identified in the analysis. | |||
| 14 | // | |||
| 15 | // sw.bb sw.bb | |||
| 16 | // / | \ / | \ | |||
| 17 | // case1 case2 case3 case1 case2 case3 | |||
| 18 | // \ | / | | | | |||
| 19 | // determinator det.2 det.3 det.1 | |||
| 20 | // br sw.bb / | \ | |||
| 21 | // sw.bb.2 sw.bb.3 sw.bb.1 | |||
| 22 | // br case2 br case3 br case1ยง | |||
| 23 | // | |||
| 24 | // Definitions and Terminology: | |||
| 25 | // | |||
| 26 | // * Threading path: | |||
| 27 | // a list of basic blocks, the exit state, and the block that determines | |||
| 28 | // the next state, for which the following notation will be used: | |||
| 29 | // < path of BBs that form a cycle > [ state, determinator ] | |||
| 30 | // | |||
| 31 | // * Predictable switch: | |||
| 32 | // The switch variable is always a known constant so that all conditional | |||
| 33 | // jumps based on switch variable can be converted to unconditional jump. | |||
| 34 | // | |||
| 35 | // * Determinator: | |||
| 36 | // The basic block that determines the next state of the DFA. | |||
| 37 | // | |||
| 38 | // Representing the optimization in C-like pseudocode: the code pattern on the | |||
| 39 | // left could functionally be transformed to the right pattern if the switch | |||
| 40 | // condition is predictable. | |||
| 41 | // | |||
| 42 | // X = A goto A | |||
| 43 | // for (...) A: | |||
| 44 | // switch (X) ... | |||
| 45 | // case A goto B | |||
| 46 | // X = B B: | |||
| 47 | // case B ... | |||
| 48 | // X = C goto C | |||
| 49 | // | |||
| 50 | // The pass first checks that switch variable X is decided by the control flow | |||
| 51 | // path taken in the loop; for example, in case B, the next value of X is | |||
| 52 | // decided to be C. It then enumerates through all paths in the loop and labels | |||
| 53 | // the basic blocks where the next state is decided. | |||
| 54 | // | |||
| 55 | // Using this information it creates new paths that unconditionally branch to | |||
| 56 | // the next case. This involves cloning code, so it only gets triggered if the | |||
| 57 | // amount of code duplicated is below a threshold. | |||
| 58 | // | |||
| 59 | //===----------------------------------------------------------------------===// | |||
| 60 | ||||
| 61 | #include "llvm/Transforms/Scalar/DFAJumpThreading.h" | |||
| 62 | #include "llvm/ADT/APInt.h" | |||
| 63 | #include "llvm/ADT/DenseMap.h" | |||
| 64 | #include "llvm/ADT/DepthFirstIterator.h" | |||
| 65 | #include "llvm/ADT/SmallSet.h" | |||
| 66 | #include "llvm/ADT/Statistic.h" | |||
| 67 | #include "llvm/Analysis/AssumptionCache.h" | |||
| 68 | #include "llvm/Analysis/CodeMetrics.h" | |||
| 69 | #include "llvm/Analysis/LoopIterator.h" | |||
| 70 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
| 71 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
| 72 | #include "llvm/IR/CFG.h" | |||
| 73 | #include "llvm/IR/Constants.h" | |||
| 74 | #include "llvm/IR/IntrinsicInst.h" | |||
| 75 | #include "llvm/IR/Verifier.h" | |||
| 76 | #include "llvm/InitializePasses.h" | |||
| 77 | #include "llvm/Pass.h" | |||
| 78 | #include "llvm/Support/CommandLine.h" | |||
| 79 | #include "llvm/Support/Debug.h" | |||
| 80 | #include "llvm/Transforms/Scalar.h" | |||
| 81 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
| 82 | #include "llvm/Transforms/Utils/Cloning.h" | |||
| 83 | #include "llvm/Transforms/Utils/SSAUpdaterBulk.h" | |||
| 84 | #include "llvm/Transforms/Utils/ValueMapper.h" | |||
| 85 | #include <algorithm> | |||
| 86 | #include <deque> | |||
| 87 | #include <unordered_map> | |||
| 88 | #include <unordered_set> | |||
| 89 | ||||
| 90 | using namespace llvm; | |||
| 91 | ||||
| 92 | #define DEBUG_TYPE"dfa-jump-threading" "dfa-jump-threading" | |||
| 93 | ||||
| 94 | STATISTIC(NumTransforms, "Number of transformations done")static llvm::Statistic NumTransforms = {"dfa-jump-threading", "NumTransforms", "Number of transformations done"}; | |||
| 95 | STATISTIC(NumCloned, "Number of blocks cloned")static llvm::Statistic NumCloned = {"dfa-jump-threading", "NumCloned" , "Number of blocks cloned"}; | |||
| 96 | STATISTIC(NumPaths, "Number of individual paths threaded")static llvm::Statistic NumPaths = {"dfa-jump-threading", "NumPaths" , "Number of individual paths threaded"}; | |||
| 97 | ||||
| 98 | static cl::opt<bool> | |||
| 99 | ClViewCfgBefore("dfa-jump-view-cfg-before", | |||
| 100 | cl::desc("View the CFG before DFA Jump Threading"), | |||
| 101 | cl::Hidden, cl::init(false)); | |||
| 102 | ||||
| 103 | static cl::opt<unsigned> MaxPathLength( | |||
| 104 | "dfa-max-path-length", | |||
| 105 | cl::desc("Max number of blocks searched to find a threading path"), | |||
| 106 | cl::Hidden, cl::init(20)); | |||
| 107 | ||||
| 108 | static cl::opt<unsigned> | |||
| 109 | CostThreshold("dfa-cost-threshold", | |||
| 110 | cl::desc("Maximum cost accepted for the transformation"), | |||
| 111 | cl::Hidden, cl::init(50)); | |||
| 112 | ||||
| 113 | namespace { | |||
| 114 | ||||
| 115 | class SelectInstToUnfold { | |||
| 116 | SelectInst *SI; | |||
| 117 | PHINode *SIUse; | |||
| 118 | ||||
| 119 | public: | |||
| 120 | SelectInstToUnfold(SelectInst *SI, PHINode *SIUse) : SI(SI), SIUse(SIUse) {} | |||
| 121 | ||||
| 122 | SelectInst *getInst() { return SI; } | |||
| 123 | PHINode *getUse() { return SIUse; } | |||
| 124 | ||||
| 125 | explicit operator bool() const { return SI && SIUse; } | |||
| 126 | }; | |||
| 127 | ||||
| 128 | void unfold(DomTreeUpdater *DTU, SelectInstToUnfold SIToUnfold, | |||
| 129 | std::vector<SelectInstToUnfold> *NewSIsToUnfold, | |||
| 130 | std::vector<BasicBlock *> *NewBBs); | |||
| 131 | ||||
| 132 | class DFAJumpThreading { | |||
| 133 | public: | |||
| 134 | DFAJumpThreading(AssumptionCache *AC, DominatorTree *DT, | |||
| 135 | TargetTransformInfo *TTI, OptimizationRemarkEmitter *ORE) | |||
| 136 | : AC(AC), DT(DT), TTI(TTI), ORE(ORE) {} | |||
| 137 | ||||
| 138 | bool run(Function &F); | |||
| 139 | ||||
| 140 | private: | |||
| 141 | void | |||
| 142 | unfoldSelectInstrs(DominatorTree *DT, | |||
| 143 | const SmallVector<SelectInstToUnfold, 4> &SelectInsts) { | |||
| 144 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); | |||
| 145 | SmallVector<SelectInstToUnfold, 4> Stack; | |||
| 146 | for (SelectInstToUnfold SIToUnfold : SelectInsts) | |||
| 147 | Stack.push_back(SIToUnfold); | |||
| 148 | ||||
| 149 | while (!Stack.empty()) { | |||
| 150 | SelectInstToUnfold SIToUnfold = Stack.back(); | |||
| 151 | Stack.pop_back(); | |||
| 152 | ||||
| 153 | std::vector<SelectInstToUnfold> NewSIsToUnfold; | |||
| 154 | std::vector<BasicBlock *> NewBBs; | |||
| 155 | unfold(&DTU, SIToUnfold, &NewSIsToUnfold, &NewBBs); | |||
| 156 | ||||
| 157 | // Put newly discovered select instructions into the work list. | |||
| 158 | for (const SelectInstToUnfold &NewSIToUnfold : NewSIsToUnfold) | |||
| 159 | Stack.push_back(NewSIToUnfold); | |||
| 160 | } | |||
| 161 | } | |||
| 162 | ||||
| 163 | AssumptionCache *AC; | |||
| 164 | DominatorTree *DT; | |||
| 165 | TargetTransformInfo *TTI; | |||
| 166 | OptimizationRemarkEmitter *ORE; | |||
| 167 | }; | |||
| 168 | ||||
| 169 | class DFAJumpThreadingLegacyPass : public FunctionPass { | |||
| 170 | public: | |||
| 171 | static char ID; // Pass identification | |||
| 172 | DFAJumpThreadingLegacyPass() : FunctionPass(ID) {} | |||
| 173 | ||||
| 174 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 175 | AU.addRequired<AssumptionCacheTracker>(); | |||
| 176 | AU.addRequired<DominatorTreeWrapperPass>(); | |||
| 177 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
| 178 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | |||
| 179 | } | |||
| 180 | ||||
| 181 | bool runOnFunction(Function &F) override { | |||
| 182 | if (skipFunction(F)) | |||
| 183 | return false; | |||
| 184 | ||||
| 185 | AssumptionCache *AC = | |||
| 186 | &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | |||
| 187 | DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | |||
| 188 | TargetTransformInfo *TTI = | |||
| 189 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
| 190 | OptimizationRemarkEmitter *ORE = | |||
| 191 | &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | |||
| 192 | ||||
| 193 | return DFAJumpThreading(AC, DT, TTI, ORE).run(F); | |||
| 194 | } | |||
| 195 | }; | |||
| 196 | } // end anonymous namespace | |||
| 197 | ||||
| 198 | char DFAJumpThreadingLegacyPass::ID = 0; | |||
| 199 | INITIALIZE_PASS_BEGIN(DFAJumpThreadingLegacyPass, "dfa-jump-threading",static void *initializeDFAJumpThreadingLegacyPassPassOnce(PassRegistry &Registry) { | |||
| 200 | "DFA Jump Threading", false, false)static void *initializeDFAJumpThreadingLegacyPassPassOnce(PassRegistry &Registry) { | |||
| 201 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | |||
| 202 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | |||
| 203 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
| 204 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | |||
| 205 | INITIALIZE_PASS_END(DFAJumpThreadingLegacyPass, "dfa-jump-threading",PassInfo *PI = new PassInfo( "DFA Jump Threading", "dfa-jump-threading" , &DFAJumpThreadingLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<DFAJumpThreadingLegacyPass>), false, false ); Registry.registerPass(*PI, true); return PI; } static llvm ::once_flag InitializeDFAJumpThreadingLegacyPassPassFlag; void llvm::initializeDFAJumpThreadingLegacyPassPass(PassRegistry & Registry) { llvm::call_once(InitializeDFAJumpThreadingLegacyPassPassFlag , initializeDFAJumpThreadingLegacyPassPassOnce, std::ref(Registry )); } | |||
| 206 | "DFA Jump Threading", false, false)PassInfo *PI = new PassInfo( "DFA Jump Threading", "dfa-jump-threading" , &DFAJumpThreadingLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<DFAJumpThreadingLegacyPass>), false, false ); Registry.registerPass(*PI, true); return PI; } static llvm ::once_flag InitializeDFAJumpThreadingLegacyPassPassFlag; void llvm::initializeDFAJumpThreadingLegacyPassPass(PassRegistry & Registry) { llvm::call_once(InitializeDFAJumpThreadingLegacyPassPassFlag , initializeDFAJumpThreadingLegacyPassPassOnce, std::ref(Registry )); } | |||
| 207 | ||||
| 208 | // Public interface to the DFA Jump Threading pass | |||
| 209 | FunctionPass *llvm::createDFAJumpThreadingPass() { | |||
| 210 | return new DFAJumpThreadingLegacyPass(); | |||
| 211 | } | |||
| 212 | ||||
| 213 | namespace { | |||
| 214 | ||||
| 215 | /// Create a new basic block and sink \p SIToSink into it. | |||
| 216 | void createBasicBlockAndSinkSelectInst( | |||
| 217 | DomTreeUpdater *DTU, SelectInst *SI, PHINode *SIUse, SelectInst *SIToSink, | |||
| 218 | BasicBlock *EndBlock, StringRef NewBBName, BasicBlock **NewBlock, | |||
| 219 | BranchInst **NewBranch, std::vector<SelectInstToUnfold> *NewSIsToUnfold, | |||
| 220 | std::vector<BasicBlock *> *NewBBs) { | |||
| 221 | assert(SIToSink->hasOneUse())((void)0); | |||
| 222 | assert(NewBlock)((void)0); | |||
| 223 | assert(NewBranch)((void)0); | |||
| 224 | *NewBlock = BasicBlock::Create(SI->getContext(), NewBBName, | |||
| 225 | EndBlock->getParent(), EndBlock); | |||
| 226 | NewBBs->push_back(*NewBlock); | |||
| 227 | *NewBranch = BranchInst::Create(EndBlock, *NewBlock); | |||
| 228 | SIToSink->moveBefore(*NewBranch); | |||
| 229 | NewSIsToUnfold->push_back(SelectInstToUnfold(SIToSink, SIUse)); | |||
| 230 | DTU->applyUpdates({{DominatorTree::Insert, *NewBlock, EndBlock}}); | |||
| 231 | } | |||
| 232 | ||||
| 233 | /// Unfold the select instruction held in \p SIToUnfold by replacing it with | |||
| 234 | /// control flow. | |||
| 235 | /// | |||
| 236 | /// Put newly discovered select instructions into \p NewSIsToUnfold. Put newly | |||
| 237 | /// created basic blocks into \p NewBBs. | |||
| 238 | /// | |||
| 239 | /// TODO: merge it with CodeGenPrepare::optimizeSelectInst() if possible. | |||
| 240 | void unfold(DomTreeUpdater *DTU, SelectInstToUnfold SIToUnfold, | |||
| 241 | std::vector<SelectInstToUnfold> *NewSIsToUnfold, | |||
| 242 | std::vector<BasicBlock *> *NewBBs) { | |||
| 243 | SelectInst *SI = SIToUnfold.getInst(); | |||
| 244 | PHINode *SIUse = SIToUnfold.getUse(); | |||
| 245 | BasicBlock *StartBlock = SI->getParent(); | |||
| 246 | BasicBlock *EndBlock = SIUse->getParent(); | |||
| 247 | BranchInst *StartBlockTerm = | |||
| 248 | dyn_cast<BranchInst>(StartBlock->getTerminator()); | |||
| 249 | ||||
| 250 | assert(StartBlockTerm && StartBlockTerm->isUnconditional())((void)0); | |||
| 251 | assert(SI->hasOneUse())((void)0); | |||
| 252 | ||||
| 253 | // These are the new basic blocks for the conditional branch. | |||
| 254 | // At least one will become an actual new basic block. | |||
| 255 | BasicBlock *TrueBlock = nullptr; | |||
| 256 | BasicBlock *FalseBlock = nullptr; | |||
| 257 | BranchInst *TrueBranch = nullptr; | |||
| 258 | BranchInst *FalseBranch = nullptr; | |||
| 259 | ||||
| 260 | // Sink select instructions to be able to unfold them later. | |||
| 261 | if (SelectInst *SIOp = dyn_cast<SelectInst>(SI->getTrueValue())) { | |||
| 262 | createBasicBlockAndSinkSelectInst(DTU, SI, SIUse, SIOp, EndBlock, | |||
| 263 | "si.unfold.true", &TrueBlock, &TrueBranch, | |||
| 264 | NewSIsToUnfold, NewBBs); | |||
| 265 | } | |||
| 266 | if (SelectInst *SIOp = dyn_cast<SelectInst>(SI->getFalseValue())) { | |||
| 267 | createBasicBlockAndSinkSelectInst(DTU, SI, SIUse, SIOp, EndBlock, | |||
| 268 | "si.unfold.false", &FalseBlock, | |||
| 269 | &FalseBranch, NewSIsToUnfold, NewBBs); | |||
| 270 | } | |||
| 271 | ||||
| 272 | // If there was nothing to sink, then arbitrarily choose the 'false' side | |||
| 273 | // for a new input value to the PHI. | |||
| 274 | if (!TrueBlock && !FalseBlock) { | |||
| 275 | FalseBlock = BasicBlock::Create(SI->getContext(), "si.unfold.false", | |||
| 276 | EndBlock->getParent(), EndBlock); | |||
| 277 | NewBBs->push_back(FalseBlock); | |||
| 278 | BranchInst::Create(EndBlock, FalseBlock); | |||
| 279 | DTU->applyUpdates({{DominatorTree::Insert, FalseBlock, EndBlock}}); | |||
| 280 | } | |||
| 281 | ||||
| 282 | // Insert the real conditional branch based on the original condition. | |||
| 283 | // If we did not create a new block for one of the 'true' or 'false' paths | |||
| 284 | // of the condition, it means that side of the branch goes to the end block | |||
| 285 | // directly and the path originates from the start block from the point of | |||
| 286 | // view of the new PHI. | |||
| 287 | BasicBlock *TT = EndBlock; | |||
| 288 | BasicBlock *FT = EndBlock; | |||
| 289 | if (TrueBlock && FalseBlock) { | |||
| 290 | // A diamond. | |||
| 291 | TT = TrueBlock; | |||
| 292 | FT = FalseBlock; | |||
| 293 | ||||
| 294 | // Update the phi node of SI. | |||
| 295 | SIUse->removeIncomingValue(StartBlock, /* DeletePHIIfEmpty = */ false); | |||
| 296 | SIUse->addIncoming(SI->getTrueValue(), TrueBlock); | |||
| 297 | SIUse->addIncoming(SI->getFalseValue(), FalseBlock); | |||
| 298 | ||||
| 299 | // Update any other PHI nodes in EndBlock. | |||
| 300 | for (PHINode &Phi : EndBlock->phis()) { | |||
| 301 | if (&Phi != SIUse) { | |||
| 302 | Phi.addIncoming(Phi.getIncomingValueForBlock(StartBlock), TrueBlock); | |||
| 303 | Phi.addIncoming(Phi.getIncomingValueForBlock(StartBlock), FalseBlock); | |||
| 304 | } | |||
| 305 | } | |||
| 306 | } else { | |||
| 307 | BasicBlock *NewBlock = nullptr; | |||
| 308 | Value *SIOp1 = SI->getTrueValue(); | |||
| 309 | Value *SIOp2 = SI->getFalseValue(); | |||
| 310 | ||||
| 311 | // A triangle pointing right. | |||
| 312 | if (!TrueBlock) { | |||
| 313 | NewBlock = FalseBlock; | |||
| 314 | FT = FalseBlock; | |||
| 315 | } | |||
| 316 | // A triangle pointing left. | |||
| 317 | else { | |||
| 318 | NewBlock = TrueBlock; | |||
| 319 | TT = TrueBlock; | |||
| 320 | std::swap(SIOp1, SIOp2); | |||
| 321 | } | |||
| 322 | ||||
| 323 | // Update the phi node of SI. | |||
| 324 | for (unsigned Idx = 0; Idx < SIUse->getNumIncomingValues(); ++Idx) { | |||
| 325 | if (SIUse->getIncomingBlock(Idx) == StartBlock) | |||
| 326 | SIUse->setIncomingValue(Idx, SIOp1); | |||
| 327 | } | |||
| 328 | SIUse->addIncoming(SIOp2, NewBlock); | |||
| 329 | ||||
| 330 | // Update any other PHI nodes in EndBlock. | |||
| 331 | for (auto II = EndBlock->begin(); PHINode *Phi = dyn_cast<PHINode>(II); | |||
| 332 | ++II) { | |||
| 333 | if (Phi != SIUse) | |||
| 334 | Phi->addIncoming(Phi->getIncomingValueForBlock(StartBlock), NewBlock); | |||
| 335 | } | |||
| 336 | } | |||
| 337 | StartBlockTerm->eraseFromParent(); | |||
| 338 | BranchInst::Create(TT, FT, SI->getCondition(), StartBlock); | |||
| 339 | DTU->applyUpdates({{DominatorTree::Insert, StartBlock, TT}, | |||
| 340 | {DominatorTree::Insert, StartBlock, FT}}); | |||
| 341 | ||||
| 342 | // The select is now dead. | |||
| 343 | SI->eraseFromParent(); | |||
| 344 | } | |||
| 345 | ||||
| 346 | struct ClonedBlock { | |||
| 347 | BasicBlock *BB; | |||
| 348 | uint64_t State; ///< \p State corresponds to the next value of a switch stmnt. | |||
| 349 | }; | |||
| 350 | ||||
| 351 | typedef std::deque<BasicBlock *> PathType; | |||
| 352 | typedef std::vector<PathType> PathsType; | |||
| 353 | typedef std::set<const BasicBlock *> VisitedBlocks; | |||
| 354 | typedef std::vector<ClonedBlock> CloneList; | |||
| 355 | ||||
| 356 | // This data structure keeps track of all blocks that have been cloned. If two | |||
| 357 | // different ThreadingPaths clone the same block for a certain state it should | |||
| 358 | // be reused, and it can be looked up in this map. | |||
| 359 | typedef DenseMap<BasicBlock *, CloneList> DuplicateBlockMap; | |||
| 360 | ||||
| 361 | // This map keeps track of all the new definitions for an instruction. This | |||
| 362 | // information is needed when restoring SSA form after cloning blocks. | |||
| 363 | typedef DenseMap<Instruction *, std::vector<Instruction *>> DefMap; | |||
| 364 | ||||
| 365 | inline raw_ostream &operator<<(raw_ostream &OS, const PathType &Path) { | |||
| 366 | OS << "< "; | |||
| 367 | for (const BasicBlock *BB : Path) { | |||
| 368 | std::string BBName; | |||
| 369 | if (BB->hasName()) | |||
| 370 | raw_string_ostream(BBName) << BB->getName(); | |||
| 371 | else | |||
| 372 | raw_string_ostream(BBName) << BB; | |||
| 373 | OS << BBName << " "; | |||
| 374 | } | |||
| 375 | OS << ">"; | |||
| 376 | return OS; | |||
| 377 | } | |||
| 378 | ||||
| 379 | /// ThreadingPath is a path in the control flow of a loop that can be threaded | |||
| 380 | /// by cloning necessary basic blocks and replacing conditional branches with | |||
| 381 | /// unconditional ones. A threading path includes a list of basic blocks, the | |||
| 382 | /// exit state, and the block that determines the next state. | |||
| 383 | struct ThreadingPath { | |||
| 384 | /// Exit value is DFA's exit state for the given path. | |||
| 385 | uint64_t getExitValue() const { return ExitVal; } | |||
| 386 | void setExitValue(const ConstantInt *V) { | |||
| 387 | ExitVal = V->getZExtValue(); | |||
| 388 | IsExitValSet = true; | |||
| 389 | } | |||
| 390 | bool isExitValueSet() const { return IsExitValSet; } | |||
| 391 | ||||
| 392 | /// Determinator is the basic block that determines the next state of the DFA. | |||
| 393 | const BasicBlock *getDeterminatorBB() const { return DBB; } | |||
| 394 | void setDeterminator(const BasicBlock *BB) { DBB = BB; } | |||
| 395 | ||||
| 396 | /// Path is a list of basic blocks. | |||
| 397 | const PathType &getPath() const { return Path; } | |||
| 398 | void setPath(const PathType &NewPath) { Path = NewPath; } | |||
| 399 | ||||
| 400 | void print(raw_ostream &OS) const { | |||
| 401 | OS << Path << " [ " << ExitVal << ", " << DBB->getName() << " ]"; | |||
| 402 | } | |||
| 403 | ||||
| 404 | private: | |||
| 405 | PathType Path; | |||
| 406 | uint64_t ExitVal; | |||
| 407 | const BasicBlock *DBB = nullptr; | |||
| 408 | bool IsExitValSet = false; | |||
| 409 | }; | |||
| 410 | ||||
| 411 | #ifndef NDEBUG1 | |||
| 412 | inline raw_ostream &operator<<(raw_ostream &OS, const ThreadingPath &TPath) { | |||
| 413 | TPath.print(OS); | |||
| 414 | return OS; | |||
| 415 | } | |||
| 416 | #endif | |||
| 417 | ||||
| 418 | struct MainSwitch { | |||
| 419 | MainSwitch(SwitchInst *SI, OptimizationRemarkEmitter *ORE) { | |||
| 420 | if (isPredictable(SI)) { | |||
| 421 | Instr = SI; | |||
| 422 | } else { | |||
| 423 | ORE->emit([&]() { | |||
| 424 | return OptimizationRemarkMissed(DEBUG_TYPE"dfa-jump-threading", "SwitchNotPredictable", SI) | |||
| 425 | << "Switch instruction is not predictable."; | |||
| 426 | }); | |||
| 427 | } | |||
| 428 | } | |||
| 429 | ||||
| 430 | virtual ~MainSwitch() = default; | |||
| 431 | ||||
| 432 | SwitchInst *getInstr() const { return Instr; } | |||
| 433 | const SmallVector<SelectInstToUnfold, 4> getSelectInsts() { | |||
| 434 | return SelectInsts; | |||
| 435 | } | |||
| 436 | ||||
| 437 | private: | |||
| 438 | /// Do a use-def chain traversal. Make sure the value of the switch variable | |||
| 439 | /// is always a known constant. This means that all conditional jumps based on | |||
| 440 | /// switch variable can be converted to unconditional jumps. | |||
| 441 | bool isPredictable(const SwitchInst *SI) { | |||
| 442 | std::deque<Instruction *> Q; | |||
| 443 | SmallSet<Value *, 16> SeenValues; | |||
| 444 | SelectInsts.clear(); | |||
| 445 | ||||
| 446 | Value *FirstDef = SI->getOperand(0); | |||
| 447 | auto *Inst = dyn_cast<Instruction>(FirstDef); | |||
| 448 | ||||
| 449 | // If this is a function argument or another non-instruction, then give up. | |||
| 450 | // We are interested in loop local variables. | |||
| 451 | if (!Inst) | |||
| 452 | return false; | |||
| 453 | ||||
| 454 | // Require the first definition to be a PHINode | |||
| 455 | if (!isa<PHINode>(Inst)) | |||
| 456 | return false; | |||
| 457 | ||||
| 458 | LLVM_DEBUG(dbgs() << "\tisPredictable() FirstDef: " << *Inst << "\n")do { } while (false); | |||
| 459 | ||||
| 460 | Q.push_back(Inst); | |||
| 461 | SeenValues.insert(FirstDef); | |||
| 462 | ||||
| 463 | while (!Q.empty()) { | |||
| 464 | Instruction *Current = Q.front(); | |||
| 465 | Q.pop_front(); | |||
| 466 | ||||
| 467 | if (auto *Phi = dyn_cast<PHINode>(Current)) { | |||
| 468 | for (Value *Incoming : Phi->incoming_values()) { | |||
| 469 | if (!isPredictableValue(Incoming, SeenValues)) | |||
| 470 | return false; | |||
| 471 | addInstToQueue(Incoming, Q, SeenValues); | |||
| 472 | } | |||
| 473 | LLVM_DEBUG(dbgs() << "\tisPredictable() phi: " << *Phi << "\n")do { } while (false); | |||
| 474 | } else if (SelectInst *SelI = dyn_cast<SelectInst>(Current)) { | |||
| 475 | if (!isValidSelectInst(SelI)) | |||
| 476 | return false; | |||
| 477 | if (!isPredictableValue(SelI->getTrueValue(), SeenValues) || | |||
| 478 | !isPredictableValue(SelI->getFalseValue(), SeenValues)) { | |||
| 479 | return false; | |||
| 480 | } | |||
| 481 | addInstToQueue(SelI->getTrueValue(), Q, SeenValues); | |||
| 482 | addInstToQueue(SelI->getFalseValue(), Q, SeenValues); | |||
| 483 | LLVM_DEBUG(dbgs() << "\tisPredictable() select: " << *SelI << "\n")do { } while (false); | |||
| 484 | if (auto *SelIUse = dyn_cast<PHINode>(SelI->user_back())) | |||
| 485 | SelectInsts.push_back(SelectInstToUnfold(SelI, SelIUse)); | |||
| 486 | } else { | |||
| 487 | // If it is neither a phi nor a select, then we give up. | |||
| 488 | return false; | |||
| 489 | } | |||
| 490 | } | |||
| 491 | ||||
| 492 | return true; | |||
| 493 | } | |||
| 494 | ||||
| 495 | bool isPredictableValue(Value *InpVal, SmallSet<Value *, 16> &SeenValues) { | |||
| 496 | if (SeenValues.find(InpVal) != SeenValues.end()) | |||
| 497 | return true; | |||
| 498 | ||||
| 499 | if (isa<ConstantInt>(InpVal)) | |||
| 500 | return true; | |||
| 501 | ||||
| 502 | // If this is a function argument or another non-instruction, then give up. | |||
| 503 | if (!isa<Instruction>(InpVal)) | |||
| 504 | return false; | |||
| 505 | ||||
| 506 | return true; | |||
| 507 | } | |||
| 508 | ||||
| 509 | void addInstToQueue(Value *Val, std::deque<Instruction *> &Q, | |||
| 510 | SmallSet<Value *, 16> &SeenValues) { | |||
| 511 | if (SeenValues.find(Val) != SeenValues.end()) | |||
| 512 | return; | |||
| 513 | if (Instruction *I = dyn_cast<Instruction>(Val)) | |||
| 514 | Q.push_back(I); | |||
| 515 | SeenValues.insert(Val); | |||
| 516 | } | |||
| 517 | ||||
| 518 | bool isValidSelectInst(SelectInst *SI) { | |||
| 519 | if (!SI->hasOneUse()) | |||
| 520 | return false; | |||
| 521 | ||||
| 522 | Instruction *SIUse = dyn_cast<Instruction>(SI->user_back()); | |||
| 523 | // The use of the select inst should be either a phi or another select. | |||
| 524 | if (!SIUse && !(isa<PHINode>(SIUse) || isa<SelectInst>(SIUse))) | |||
| 525 | return false; | |||
| 526 | ||||
| 527 | BasicBlock *SIBB = SI->getParent(); | |||
| 528 | ||||
| 529 | // Currently, we can only expand select instructions in basic blocks with | |||
| 530 | // one successor. | |||
| 531 | BranchInst *SITerm = dyn_cast<BranchInst>(SIBB->getTerminator()); | |||
| 532 | if (!SITerm || !SITerm->isUnconditional()) | |||
| 533 | return false; | |||
| 534 | ||||
| 535 | if (isa<PHINode>(SIUse) && | |||
| 536 | SIBB->getSingleSuccessor() != dyn_cast<Instruction>(SIUse)->getParent()) | |||
| 537 | return false; | |||
| 538 | ||||
| 539 | // If select will not be sunk during unfolding, and it is in the same basic | |||
| 540 | // block as another state defining select, then cannot unfold both. | |||
| 541 | for (SelectInstToUnfold SIToUnfold : SelectInsts) { | |||
| 542 | SelectInst *PrevSI = SIToUnfold.getInst(); | |||
| 543 | if (PrevSI->getTrueValue() != SI && PrevSI->getFalseValue() != SI && | |||
| 544 | PrevSI->getParent() == SI->getParent()) | |||
| 545 | return false; | |||
| 546 | } | |||
| 547 | ||||
| 548 | return true; | |||
| 549 | } | |||
| 550 | ||||
| 551 | SwitchInst *Instr = nullptr; | |||
| 552 | SmallVector<SelectInstToUnfold, 4> SelectInsts; | |||
| 553 | }; | |||
| 554 | ||||
| 555 | struct AllSwitchPaths { | |||
| 556 | AllSwitchPaths(const MainSwitch *MSwitch, OptimizationRemarkEmitter *ORE) | |||
| 557 | : Switch(MSwitch->getInstr()), SwitchBlock(Switch->getParent()), | |||
| 558 | ORE(ORE) {} | |||
| 559 | ||||
| 560 | std::vector<ThreadingPath> &getThreadingPaths() { return TPaths; } | |||
| 561 | unsigned getNumThreadingPaths() { return TPaths.size(); } | |||
| 562 | SwitchInst *getSwitchInst() { return Switch; } | |||
| 563 | BasicBlock *getSwitchBlock() { return SwitchBlock; } | |||
| 564 | ||||
| 565 | void run() { | |||
| 566 | VisitedBlocks Visited; | |||
| 567 | PathsType LoopPaths = paths(SwitchBlock, Visited, /* PathDepth = */ 1); | |||
| 568 | StateDefMap StateDef = getStateDefMap(); | |||
| 569 | ||||
| 570 | for (PathType Path : LoopPaths) { | |||
| 571 | ThreadingPath TPath; | |||
| 572 | ||||
| 573 | const BasicBlock *PrevBB = Path.back(); | |||
| 574 | for (const BasicBlock *BB : Path) { | |||
| 575 | if (StateDef.count(BB) != 0) { | |||
| 576 | const PHINode *Phi = dyn_cast<PHINode>(StateDef[BB]); | |||
| 577 | assert(Phi && "Expected a state-defining instr to be a phi node.")((void)0); | |||
| 578 | ||||
| 579 | const Value *V = Phi->getIncomingValueForBlock(PrevBB); | |||
| ||||
| 580 | if (const ConstantInt *C = dyn_cast<const ConstantInt>(V)) { | |||
| 581 | TPath.setExitValue(C); | |||
| 582 | TPath.setDeterminator(BB); | |||
| 583 | TPath.setPath(Path); | |||
| 584 | } | |||
| 585 | } | |||
| 586 | ||||
| 587 | // Switch block is the determinator, this is the final exit value. | |||
| 588 | if (TPath.isExitValueSet() && BB == Path.front()) | |||
| 589 | break; | |||
| 590 | ||||
| 591 | PrevBB = BB; | |||
| 592 | } | |||
| 593 | ||||
| 594 | if (TPath.isExitValueSet()) | |||
| 595 | TPaths.push_back(TPath); | |||
| 596 | } | |||
| 597 | } | |||
| 598 | ||||
| 599 | private: | |||
| 600 | // Value: an instruction that defines a switch state; | |||
| 601 | // Key: the parent basic block of that instruction. | |||
| 602 | typedef DenseMap<const BasicBlock *, const PHINode *> StateDefMap; | |||
| 603 | ||||
| 604 | PathsType paths(BasicBlock *BB, VisitedBlocks &Visited, | |||
| 605 | unsigned PathDepth) const { | |||
| 606 | PathsType Res; | |||
| 607 | ||||
| 608 | // Stop exploring paths after visiting MaxPathLength blocks | |||
| 609 | if (PathDepth > MaxPathLength) { | |||
| 610 | ORE->emit([&]() { | |||
| 611 | return OptimizationRemarkAnalysis(DEBUG_TYPE"dfa-jump-threading", "MaxPathLengthReached", | |||
| 612 | Switch) | |||
| 613 | << "Exploration stopped after visiting MaxPathLength=" | |||
| 614 | << ore::NV("MaxPathLength", MaxPathLength) << " blocks."; | |||
| 615 | }); | |||
| 616 | return Res; | |||
| 617 | } | |||
| 618 | ||||
| 619 | Visited.insert(BB); | |||
| 620 | ||||
| 621 | // Some blocks have multiple edges to the same successor, and this set | |||
| 622 | // is used to prevent a duplicate path from being generated | |||
| 623 | SmallSet<BasicBlock *, 4> Successors; | |||
| 624 | ||||
| 625 | for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { | |||
| 626 | BasicBlock *Succ = *SI; | |||
| 627 | ||||
| 628 | if (Successors.find(Succ) != Successors.end()) | |||
| 629 | continue; | |||
| 630 | Successors.insert(Succ); | |||
| 631 | ||||
| 632 | // Found a cycle through the SwitchBlock | |||
| 633 | if (Succ == SwitchBlock) { | |||
| 634 | Res.push_back({BB}); | |||
| 635 | continue; | |||
| 636 | } | |||
| 637 | ||||
| 638 | // We have encountered a cycle, do not get caught in it | |||
| 639 | if (Visited.find(Succ) != Visited.end()) | |||
| 640 | continue; | |||
| 641 | ||||
| 642 | PathsType SuccPaths = paths(Succ, Visited, PathDepth + 1); | |||
| 643 | for (PathType Path : SuccPaths) { | |||
| 644 | PathType NewPath(Path); | |||
| 645 | NewPath.push_front(BB); | |||
| 646 | Res.push_back(NewPath); | |||
| 647 | } | |||
| 648 | } | |||
| 649 | // This block could now be visited again from a different predecessor. Note | |||
| 650 | // that this will result in exponential runtime. Subpaths could possibly be | |||
| 651 | // cached but it takes a lot of memory to store them. | |||
| 652 | Visited.erase(BB); | |||
| 653 | return Res; | |||
| 654 | } | |||
| 655 | ||||
| 656 | /// Walk the use-def chain and collect all the state-defining instructions. | |||
| 657 | StateDefMap getStateDefMap() const { | |||
| 658 | StateDefMap Res; | |||
| 659 | ||||
| 660 | Value *FirstDef = Switch->getOperand(0); | |||
| 661 | ||||
| 662 | assert(isa<PHINode>(FirstDef) && "After select unfolding, all state "((void)0) | |||
| 663 | "definitions are expected to be phi "((void)0) | |||
| 664 | "nodes.")((void)0); | |||
| 665 | ||||
| 666 | SmallVector<PHINode *, 8> Stack; | |||
| 667 | Stack.push_back(dyn_cast<PHINode>(FirstDef)); | |||
| 668 | SmallSet<Value *, 16> SeenValues; | |||
| 669 | ||||
| 670 | while (!Stack.empty()) { | |||
| 671 | PHINode *CurPhi = Stack.back(); | |||
| 672 | Stack.pop_back(); | |||
| 673 | ||||
| 674 | Res[CurPhi->getParent()] = CurPhi; | |||
| 675 | SeenValues.insert(CurPhi); | |||
| 676 | ||||
| 677 | for (Value *Incoming : CurPhi->incoming_values()) { | |||
| 678 | if (Incoming == FirstDef || isa<ConstantInt>(Incoming) || | |||
| 679 | SeenValues.find(Incoming) != SeenValues.end()) { | |||
| 680 | continue; | |||
| 681 | } | |||
| 682 | ||||
| 683 | assert(isa<PHINode>(Incoming) && "After select unfolding, all state "((void)0) | |||
| 684 | "definitions are expected to be phi "((void)0) | |||
| 685 | "nodes.")((void)0); | |||
| 686 | ||||
| 687 | Stack.push_back(cast<PHINode>(Incoming)); | |||
| 688 | } | |||
| 689 | } | |||
| 690 | ||||
| 691 | return Res; | |||
| 692 | } | |||
| 693 | ||||
| 694 | SwitchInst *Switch; | |||
| 695 | BasicBlock *SwitchBlock; | |||
| 696 | OptimizationRemarkEmitter *ORE; | |||
| 697 | std::vector<ThreadingPath> TPaths; | |||
| 698 | }; | |||
| 699 | ||||
| 700 | struct TransformDFA { | |||
| 701 | TransformDFA(AllSwitchPaths *SwitchPaths, DominatorTree *DT, | |||
| 702 | AssumptionCache *AC, TargetTransformInfo *TTI, | |||
| 703 | OptimizationRemarkEmitter *ORE, | |||
| 704 | SmallPtrSet<const Value *, 32> EphValues) | |||
| 705 | : SwitchPaths(SwitchPaths), DT(DT), AC(AC), TTI(TTI), ORE(ORE), | |||
| 706 | EphValues(EphValues) {} | |||
| 707 | ||||
| 708 | void run() { | |||
| 709 | if (isLegalAndProfitableToTransform()) { | |||
| 710 | createAllExitPaths(); | |||
| 711 | NumTransforms++; | |||
| 712 | } | |||
| 713 | } | |||
| 714 | ||||
| 715 | private: | |||
| 716 | /// This function performs both a legality check and profitability check at | |||
| 717 | /// the same time since it is convenient to do so. It iterates through all | |||
| 718 | /// blocks that will be cloned, and keeps track of the duplication cost. It | |||
| 719 | /// also returns false if it is illegal to clone some required block. | |||
| 720 | bool isLegalAndProfitableToTransform() { | |||
| 721 | CodeMetrics Metrics; | |||
| 722 | SwitchInst *Switch = SwitchPaths->getSwitchInst(); | |||
| 723 | ||||
| 724 | // Note that DuplicateBlockMap is not being used as intended here. It is | |||
| 725 | // just being used to ensure (BB, State) pairs are only counted once. | |||
| 726 | DuplicateBlockMap DuplicateMap; | |||
| 727 | ||||
| 728 | for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) { | |||
| 729 | PathType PathBBs = TPath.getPath(); | |||
| 730 | uint64_t NextState = TPath.getExitValue(); | |||
| 731 | const BasicBlock *Determinator = TPath.getDeterminatorBB(); | |||
| 732 | ||||
| 733 | // Update Metrics for the Switch block, this is always cloned | |||
| 734 | BasicBlock *BB = SwitchPaths->getSwitchBlock(); | |||
| 735 | BasicBlock *VisitedBB = getClonedBB(BB, NextState, DuplicateMap); | |||
| 736 | if (!VisitedBB) { | |||
| 737 | Metrics.analyzeBasicBlock(BB, *TTI, EphValues); | |||
| 738 | DuplicateMap[BB].push_back({BB, NextState}); | |||
| 739 | } | |||
| 740 | ||||
| 741 | // If the Switch block is the Determinator, then we can continue since | |||
| 742 | // this is the only block that is cloned and we already counted for it. | |||
| 743 | if (PathBBs.front() == Determinator) | |||
| 744 | continue; | |||
| 745 | ||||
| 746 | // Otherwise update Metrics for all blocks that will be cloned. If any | |||
| 747 | // block is already cloned and would be reused, don't double count it. | |||
| 748 | auto DetIt = std::find(PathBBs.begin(), PathBBs.end(), Determinator); | |||
| 749 | for (auto BBIt = DetIt; BBIt != PathBBs.end(); BBIt++) { | |||
| 750 | BB = *BBIt; | |||
| 751 | VisitedBB = getClonedBB(BB, NextState, DuplicateMap); | |||
| 752 | if (VisitedBB) | |||
| 753 | continue; | |||
| 754 | Metrics.analyzeBasicBlock(BB, *TTI, EphValues); | |||
| 755 | DuplicateMap[BB].push_back({BB, NextState}); | |||
| 756 | } | |||
| 757 | ||||
| 758 | if (Metrics.notDuplicatable) { | |||
| 759 | LLVM_DEBUG(dbgs() << "DFA Jump Threading: Not jump threading, contains "do { } while (false) | |||
| 760 | << "non-duplicatable instructions.\n")do { } while (false); | |||
| 761 | ORE->emit([&]() { | |||
| 762 | return OptimizationRemarkMissed(DEBUG_TYPE"dfa-jump-threading", "NonDuplicatableInst", | |||
| 763 | Switch) | |||
| 764 | << "Contains non-duplicatable instructions."; | |||
| 765 | }); | |||
| 766 | return false; | |||
| 767 | } | |||
| 768 | ||||
| 769 | if (Metrics.convergent) { | |||
| 770 | LLVM_DEBUG(dbgs() << "DFA Jump Threading: Not jump threading, contains "do { } while (false) | |||
| 771 | << "convergent instructions.\n")do { } while (false); | |||
| 772 | ORE->emit([&]() { | |||
| 773 | return OptimizationRemarkMissed(DEBUG_TYPE"dfa-jump-threading", "ConvergentInst", Switch) | |||
| 774 | << "Contains convergent instructions."; | |||
| 775 | }); | |||
| 776 | return false; | |||
| 777 | } | |||
| 778 | } | |||
| 779 | ||||
| 780 | unsigned DuplicationCost = 0; | |||
| 781 | ||||
| 782 | unsigned JumpTableSize = 0; | |||
| 783 | TTI->getEstimatedNumberOfCaseClusters(*Switch, JumpTableSize, nullptr, | |||
| 784 | nullptr); | |||
| 785 | if (JumpTableSize == 0) { | |||
| 786 | // Factor in the number of conditional branches reduced from jump | |||
| 787 | // threading. Assume that lowering the switch block is implemented by | |||
| 788 | // using binary search, hence the LogBase2(). | |||
| 789 | unsigned CondBranches = | |||
| 790 | APInt(32, Switch->getNumSuccessors()).ceilLogBase2(); | |||
| 791 | DuplicationCost = Metrics.NumInsts / CondBranches; | |||
| 792 | } else { | |||
| 793 | // Compared with jump tables, the DFA optimizer removes an indirect branch | |||
| 794 | // on each loop iteration, thus making branch prediction more precise. The | |||
| 795 | // more branch targets there are, the more likely it is for the branch | |||
| 796 | // predictor to make a mistake, and the more benefit there is in the DFA | |||
| 797 | // optimizer. Thus, the more branch targets there are, the lower is the | |||
| 798 | // cost of the DFA opt. | |||
| 799 | DuplicationCost = Metrics.NumInsts / JumpTableSize; | |||
| 800 | } | |||
| 801 | ||||
| 802 | LLVM_DEBUG(dbgs() << "\nDFA Jump Threading: Cost to jump thread block "do { } while (false) | |||
| 803 | << SwitchPaths->getSwitchBlock()->getName()do { } while (false) | |||
| 804 | << " is: " << DuplicationCost << "\n\n")do { } while (false); | |||
| 805 | ||||
| 806 | if (DuplicationCost > CostThreshold) { | |||
| 807 | LLVM_DEBUG(dbgs() << "Not jump threading, duplication cost exceeds the "do { } while (false) | |||
| 808 | << "cost threshold.\n")do { } while (false); | |||
| 809 | ORE->emit([&]() { | |||
| 810 | return OptimizationRemarkMissed(DEBUG_TYPE"dfa-jump-threading", "NotProfitable", Switch) | |||
| 811 | << "Duplication cost exceeds the cost threshold (cost=" | |||
| 812 | << ore::NV("Cost", DuplicationCost) | |||
| 813 | << ", threshold=" << ore::NV("Threshold", CostThreshold) << ")."; | |||
| 814 | }); | |||
| 815 | return false; | |||
| 816 | } | |||
| 817 | ||||
| 818 | ORE->emit([&]() { | |||
| 819 | return OptimizationRemark(DEBUG_TYPE"dfa-jump-threading", "JumpThreaded", Switch) | |||
| 820 | << "Switch statement jump-threaded."; | |||
| 821 | }); | |||
| 822 | ||||
| 823 | return true; | |||
| 824 | } | |||
| 825 | ||||
| 826 | /// Transform each threading path to effectively jump thread the DFA. | |||
| 827 | void createAllExitPaths() { | |||
| 828 | DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Eager); | |||
| 829 | ||||
| 830 | // Move the switch block to the end of the path, since it will be duplicated | |||
| 831 | BasicBlock *SwitchBlock = SwitchPaths->getSwitchBlock(); | |||
| 832 | for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) { | |||
| 833 | LLVM_DEBUG(dbgs() << TPath << "\n")do { } while (false); | |||
| 834 | PathType NewPath(TPath.getPath()); | |||
| 835 | NewPath.push_back(SwitchBlock); | |||
| 836 | TPath.setPath(NewPath); | |||
| 837 | } | |||
| 838 | ||||
| 839 | // Transform the ThreadingPaths and keep track of the cloned values | |||
| 840 | DuplicateBlockMap DuplicateMap; | |||
| 841 | DefMap NewDefs; | |||
| 842 | ||||
| 843 | SmallSet<BasicBlock *, 16> BlocksToClean; | |||
| 844 | for (BasicBlock *BB : successors(SwitchBlock)) | |||
| 845 | BlocksToClean.insert(BB); | |||
| 846 | ||||
| 847 | for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) { | |||
| 848 | createExitPath(NewDefs, TPath, DuplicateMap, BlocksToClean, &DTU); | |||
| 849 | NumPaths++; | |||
| 850 | } | |||
| 851 | ||||
| 852 | // After all paths are cloned, now update the last successor of the cloned | |||
| 853 | // path so it skips over the switch statement | |||
| 854 | for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) | |||
| 855 | updateLastSuccessor(TPath, DuplicateMap, &DTU); | |||
| 856 | ||||
| 857 | // For each instruction that was cloned and used outside, update its uses | |||
| 858 | updateSSA(NewDefs); | |||
| 859 | ||||
| 860 | // Clean PHI Nodes for the newly created blocks | |||
| 861 | for (BasicBlock *BB : BlocksToClean) | |||
| 862 | cleanPhiNodes(BB); | |||
| 863 | } | |||
| 864 | ||||
| 865 | /// For a specific ThreadingPath \p Path, create an exit path starting from | |||
| 866 | /// the determinator block. | |||
| 867 | /// | |||
| 868 | /// To remember the correct destination, we have to duplicate blocks | |||
| 869 | /// corresponding to each state. Also update the terminating instruction of | |||
| 870 | /// the predecessors, and phis in the successor blocks. | |||
| 871 | void createExitPath(DefMap &NewDefs, ThreadingPath &Path, | |||
| 872 | DuplicateBlockMap &DuplicateMap, | |||
| 873 | SmallSet<BasicBlock *, 16> &BlocksToClean, | |||
| 874 | DomTreeUpdater *DTU) { | |||
| 875 | uint64_t NextState = Path.getExitValue(); | |||
| 876 | const BasicBlock *Determinator = Path.getDeterminatorBB(); | |||
| 877 | PathType PathBBs = Path.getPath(); | |||
| 878 | ||||
| 879 | // Don't select the placeholder block in front | |||
| 880 | if (PathBBs.front() == Determinator) | |||
| 881 | PathBBs.pop_front(); | |||
| 882 | ||||
| 883 | auto DetIt = std::find(PathBBs.begin(), PathBBs.end(), Determinator); | |||
| 884 | auto Prev = std::prev(DetIt); | |||
| 885 | BasicBlock *PrevBB = *Prev; | |||
| 886 | for (auto BBIt = DetIt; BBIt != PathBBs.end(); BBIt++) { | |||
| 887 | BasicBlock *BB = *BBIt; | |||
| 888 | BlocksToClean.insert(BB); | |||
| 889 | ||||
| 890 | // We already cloned BB for this NextState, now just update the branch | |||
| 891 | // and continue. | |||
| 892 | BasicBlock *NextBB = getClonedBB(BB, NextState, DuplicateMap); | |||
| 893 | if (NextBB) { | |||
| 894 | updatePredecessor(PrevBB, BB, NextBB, DTU); | |||
| 895 | PrevBB = NextBB; | |||
| 896 | continue; | |||
| 897 | } | |||
| 898 | ||||
| 899 | // Clone the BB and update the successor of Prev to jump to the new block | |||
| 900 | BasicBlock *NewBB = cloneBlockAndUpdatePredecessor( | |||
| 901 | BB, PrevBB, NextState, DuplicateMap, NewDefs, DTU); | |||
| 902 | DuplicateMap[BB].push_back({NewBB, NextState}); | |||
| 903 | BlocksToClean.insert(NewBB); | |||
| 904 | PrevBB = NewBB; | |||
| 905 | } | |||
| 906 | } | |||
| 907 | ||||
| 908 | /// Restore SSA form after cloning blocks. | |||
| 909 | /// | |||
| 910 | /// Each cloned block creates new defs for a variable, and the uses need to be | |||
| 911 | /// updated to reflect this. The uses may be replaced with a cloned value, or | |||
| 912 | /// some derived phi instruction. Note that all uses of a value defined in the | |||
| 913 | /// same block were already remapped when cloning the block. | |||
| 914 | void updateSSA(DefMap &NewDefs) { | |||
| 915 | SSAUpdaterBulk SSAUpdate; | |||
| 916 | SmallVector<Use *, 16> UsesToRename; | |||
| 917 | ||||
| 918 | for (auto KV : NewDefs) { | |||
| 919 | Instruction *I = KV.first; | |||
| 920 | BasicBlock *BB = I->getParent(); | |||
| 921 | std::vector<Instruction *> Cloned = KV.second; | |||
| 922 | ||||
| 923 | // Scan all uses of this instruction to see if it is used outside of its | |||
| 924 | // block, and if so, record them in UsesToRename. | |||
| 925 | for (Use &U : I->uses()) { | |||
| 926 | Instruction *User = cast<Instruction>(U.getUser()); | |||
| 927 | if (PHINode *UserPN = dyn_cast<PHINode>(User)) { | |||
| 928 | if (UserPN->getIncomingBlock(U) == BB) | |||
| 929 | continue; | |||
| 930 | } else if (User->getParent() == BB) { | |||
| 931 | continue; | |||
| 932 | } | |||
| 933 | ||||
| 934 | UsesToRename.push_back(&U); | |||
| 935 | } | |||
| 936 | ||||
| 937 | // If there are no uses outside the block, we're done with this | |||
| 938 | // instruction. | |||
| 939 | if (UsesToRename.empty()) | |||
| 940 | continue; | |||
| 941 | LLVM_DEBUG(dbgs() << "DFA-JT: Renaming non-local uses of: " << *Ido { } while (false) | |||
| 942 | << "\n")do { } while (false); | |||
| 943 | ||||
| 944 | // We found a use of I outside of BB. Rename all uses of I that are | |||
| 945 | // outside its block to be uses of the appropriate PHI node etc. See | |||
| 946 | // ValuesInBlocks with the values we know. | |||
| 947 | unsigned VarNum = SSAUpdate.AddVariable(I->getName(), I->getType()); | |||
| 948 | SSAUpdate.AddAvailableValue(VarNum, BB, I); | |||
| 949 | for (Instruction *New : Cloned) | |||
| 950 | SSAUpdate.AddAvailableValue(VarNum, New->getParent(), New); | |||
| 951 | ||||
| 952 | while (!UsesToRename.empty()) | |||
| 953 | SSAUpdate.AddUse(VarNum, UsesToRename.pop_back_val()); | |||
| 954 | ||||
| 955 | LLVM_DEBUG(dbgs() << "\n")do { } while (false); | |||
| 956 | } | |||
| 957 | // SSAUpdater handles phi placement and renaming uses with the appropriate | |||
| 958 | // value. | |||
| 959 | SSAUpdate.RewriteAllUses(DT); | |||
| 960 | } | |||
| 961 | ||||
| 962 | /// Clones a basic block, and adds it to the CFG. | |||
| 963 | /// | |||
| 964 | /// This function also includes updating phi nodes in the successors of the | |||
| 965 | /// BB, and remapping uses that were defined locally in the cloned BB. | |||
| 966 | BasicBlock *cloneBlockAndUpdatePredecessor(BasicBlock *BB, BasicBlock *PrevBB, | |||
| 967 | uint64_t NextState, | |||
| 968 | DuplicateBlockMap &DuplicateMap, | |||
| 969 | DefMap &NewDefs, | |||
| 970 | DomTreeUpdater *DTU) { | |||
| 971 | ValueToValueMapTy VMap; | |||
| 972 | BasicBlock *NewBB = CloneBasicBlock( | |||
| 973 | BB, VMap, ".jt" + std::to_string(NextState), BB->getParent()); | |||
| 974 | NewBB->moveAfter(BB); | |||
| 975 | NumCloned++; | |||
| 976 | ||||
| 977 | for (Instruction &I : *NewBB) { | |||
| 978 | // Do not remap operands of PHINode in case a definition in BB is an | |||
| 979 | // incoming value to a phi in the same block. This incoming value will | |||
| 980 | // be renamed later while restoring SSA. | |||
| 981 | if (isa<PHINode>(&I)) | |||
| 982 | continue; | |||
| 983 | RemapInstruction(&I, VMap, | |||
| 984 | RF_IgnoreMissingLocals | RF_NoModuleLevelChanges); | |||
| 985 | if (AssumeInst *II = dyn_cast<AssumeInst>(&I)) | |||
| 986 | AC->registerAssumption(II); | |||
| 987 | } | |||
| 988 | ||||
| 989 | updateSuccessorPhis(BB, NewBB, NextState, VMap, DuplicateMap); | |||
| 990 | updatePredecessor(PrevBB, BB, NewBB, DTU); | |||
| 991 | updateDefMap(NewDefs, VMap); | |||
| 992 | ||||
| 993 | // Add all successors to the DominatorTree | |||
| 994 | SmallPtrSet<BasicBlock *, 4> SuccSet; | |||
| 995 | for (auto *SuccBB : successors(NewBB)) { | |||
| 996 | if (SuccSet.insert(SuccBB).second) | |||
| 997 | DTU->applyUpdates({{DominatorTree::Insert, NewBB, SuccBB}}); | |||
| 998 | } | |||
| 999 | SuccSet.clear(); | |||
| 1000 | return NewBB; | |||
| 1001 | } | |||
| 1002 | ||||
| 1003 | /// Update the phi nodes in BB's successors. | |||
| 1004 | /// | |||
| 1005 | /// This means creating a new incoming value from NewBB with the new | |||
| 1006 | /// instruction wherever there is an incoming value from BB. | |||
| 1007 | void updateSuccessorPhis(BasicBlock *BB, BasicBlock *ClonedBB, | |||
| 1008 | uint64_t NextState, ValueToValueMapTy &VMap, | |||
| 1009 | DuplicateBlockMap &DuplicateMap) { | |||
| 1010 | std::vector<BasicBlock *> BlocksToUpdate; | |||
| 1011 | ||||
| 1012 | // If BB is the last block in the path, we can simply update the one case | |||
| 1013 | // successor that will be reached. | |||
| 1014 | if (BB == SwitchPaths->getSwitchBlock()) { | |||
| 1015 | SwitchInst *Switch = SwitchPaths->getSwitchInst(); | |||
| 1016 | BasicBlock *NextCase = getNextCaseSuccessor(Switch, NextState); | |||
| 1017 | BlocksToUpdate.push_back(NextCase); | |||
| 1018 | BasicBlock *ClonedSucc = getClonedBB(NextCase, NextState, DuplicateMap); | |||
| 1019 | if (ClonedSucc) | |||
| 1020 | BlocksToUpdate.push_back(ClonedSucc); | |||
| 1021 | } | |||
| 1022 | // Otherwise update phis in all successors. | |||
| 1023 | else { | |||
| 1024 | for (BasicBlock *Succ : successors(BB)) { | |||
| 1025 | BlocksToUpdate.push_back(Succ); | |||
| 1026 | ||||
| 1027 | // Check if a successor has already been cloned for the particular exit | |||
| 1028 | // value. In this case if a successor was already cloned, the phi nodes | |||
| 1029 | // in the cloned block should be updated directly. | |||
| 1030 | BasicBlock *ClonedSucc = getClonedBB(Succ, NextState, DuplicateMap); | |||
| 1031 | if (ClonedSucc) | |||
| 1032 | BlocksToUpdate.push_back(ClonedSucc); | |||
| 1033 | } | |||
| 1034 | } | |||
| 1035 | ||||
| 1036 | // If there is a phi with an incoming value from BB, create a new incoming | |||
| 1037 | // value for the new predecessor ClonedBB. The value will either be the same | |||
| 1038 | // value from BB or a cloned value. | |||
| 1039 | for (BasicBlock *Succ : BlocksToUpdate) { | |||
| 1040 | for (auto II = Succ->begin(); PHINode *Phi = dyn_cast<PHINode>(II); | |||
| 1041 | ++II) { | |||
| 1042 | Value *Incoming = Phi->getIncomingValueForBlock(BB); | |||
| 1043 | if (Incoming) { | |||
| 1044 | if (isa<Constant>(Incoming)) { | |||
| 1045 | Phi->addIncoming(Incoming, ClonedBB); | |||
| 1046 | continue; | |||
| 1047 | } | |||
| 1048 | Value *ClonedVal = VMap[Incoming]; | |||
| 1049 | if (ClonedVal) | |||
| 1050 | Phi->addIncoming(ClonedVal, ClonedBB); | |||
| 1051 | else | |||
| 1052 | Phi->addIncoming(Incoming, ClonedBB); | |||
| 1053 | } | |||
| 1054 | } | |||
| 1055 | } | |||
| 1056 | } | |||
| 1057 | ||||
| 1058 | /// Sets the successor of PrevBB to be NewBB instead of OldBB. Note that all | |||
| 1059 | /// other successors are kept as well. | |||
| 1060 | void updatePredecessor(BasicBlock *PrevBB, BasicBlock *OldBB, | |||
| 1061 | BasicBlock *NewBB, DomTreeUpdater *DTU) { | |||
| 1062 | // When a path is reused, there is a chance that predecessors were already | |||
| 1063 | // updated before. Check if the predecessor needs to be updated first. | |||
| 1064 | if (!isPredecessor(OldBB, PrevBB)) | |||
| 1065 | return; | |||
| 1066 | ||||
| 1067 | Instruction *PrevTerm = PrevBB->getTerminator(); | |||
| 1068 | for (unsigned Idx = 0; Idx < PrevTerm->getNumSuccessors(); Idx++) { | |||
| 1069 | if (PrevTerm->getSuccessor(Idx) == OldBB) { | |||
| 1070 | OldBB->removePredecessor(PrevBB, /* KeepOneInputPHIs = */ true); | |||
| 1071 | PrevTerm->setSuccessor(Idx, NewBB); | |||
| 1072 | } | |||
| 1073 | } | |||
| 1074 | DTU->applyUpdates({{DominatorTree::Delete, PrevBB, OldBB}, | |||
| 1075 | {DominatorTree::Insert, PrevBB, NewBB}}); | |||
| 1076 | } | |||
| 1077 | ||||
| 1078 | /// Add new value mappings to the DefMap to keep track of all new definitions | |||
| 1079 | /// for a particular instruction. These will be used while updating SSA form. | |||
| 1080 | void updateDefMap(DefMap &NewDefs, ValueToValueMapTy &VMap) { | |||
| 1081 | for (auto Entry : VMap) { | |||
| 1082 | Instruction *Inst = | |||
| 1083 | dyn_cast<Instruction>(const_cast<Value *>(Entry.first)); | |||
| 1084 | if (!Inst || !Entry.second || isa<BranchInst>(Inst) || | |||
| 1085 | isa<SwitchInst>(Inst)) { | |||
| 1086 | continue; | |||
| 1087 | } | |||
| 1088 | ||||
| 1089 | Instruction *Cloned = dyn_cast<Instruction>(Entry.second); | |||
| 1090 | if (!Cloned) | |||
| 1091 | continue; | |||
| 1092 | ||||
| 1093 | if (NewDefs.find(Inst) == NewDefs.end()) | |||
| 1094 | NewDefs[Inst] = {Cloned}; | |||
| 1095 | else | |||
| 1096 | NewDefs[Inst].push_back(Cloned); | |||
| 1097 | } | |||
| 1098 | } | |||
| 1099 | ||||
| 1100 | /// Update the last branch of a particular cloned path to point to the correct | |||
| 1101 | /// case successor. | |||
| 1102 | /// | |||
| 1103 | /// Note that this is an optional step and would have been done in later | |||
| 1104 | /// optimizations, but it makes the CFG significantly easier to work with. | |||
| 1105 | void updateLastSuccessor(ThreadingPath &TPath, | |||
| 1106 | DuplicateBlockMap &DuplicateMap, | |||
| 1107 | DomTreeUpdater *DTU) { | |||
| 1108 | uint64_t NextState = TPath.getExitValue(); | |||
| 1109 | BasicBlock *BB = TPath.getPath().back(); | |||
| 1110 | BasicBlock *LastBlock = getClonedBB(BB, NextState, DuplicateMap); | |||
| 1111 | ||||
| 1112 | // Note multiple paths can end at the same block so check that it is not | |||
| 1113 | // updated yet | |||
| 1114 | if (!isa<SwitchInst>(LastBlock->getTerminator())) | |||
| 1115 | return; | |||
| 1116 | SwitchInst *Switch = cast<SwitchInst>(LastBlock->getTerminator()); | |||
| 1117 | BasicBlock *NextCase = getNextCaseSuccessor(Switch, NextState); | |||
| 1118 | ||||
| 1119 | std::vector<DominatorTree::UpdateType> DTUpdates; | |||
| 1120 | SmallPtrSet<BasicBlock *, 4> SuccSet; | |||
| 1121 | for (BasicBlock *Succ : successors(LastBlock)) { | |||
| 1122 | if (Succ != NextCase && SuccSet.insert(Succ).second) | |||
| 1123 | DTUpdates.push_back({DominatorTree::Delete, LastBlock, Succ}); | |||
| 1124 | } | |||
| 1125 | ||||
| 1126 | Switch->eraseFromParent(); | |||
| 1127 | BranchInst::Create(NextCase, LastBlock); | |||
| 1128 | ||||
| 1129 | DTU->applyUpdates(DTUpdates); | |||
| 1130 | } | |||
| 1131 | ||||
| 1132 | /// After cloning blocks, some of the phi nodes have extra incoming values | |||
| 1133 | /// that are no longer used. This function removes them. | |||
| 1134 | void cleanPhiNodes(BasicBlock *BB) { | |||
| 1135 | // If BB is no longer reachable, remove any remaining phi nodes | |||
| 1136 | if (pred_empty(BB)) { | |||
| 1137 | std::vector<PHINode *> PhiToRemove; | |||
| 1138 | for (auto II = BB->begin(); PHINode *Phi = dyn_cast<PHINode>(II); ++II) { | |||
| 1139 | PhiToRemove.push_back(Phi); | |||
| 1140 | } | |||
| 1141 | for (PHINode *PN : PhiToRemove) { | |||
| 1142 | PN->replaceAllUsesWith(UndefValue::get(PN->getType())); | |||
| 1143 | PN->eraseFromParent(); | |||
| 1144 | } | |||
| 1145 | return; | |||
| 1146 | } | |||
| 1147 | ||||
| 1148 | // Remove any incoming values that come from an invalid predecessor | |||
| 1149 | for (auto II = BB->begin(); PHINode *Phi = dyn_cast<PHINode>(II); ++II) { | |||
| 1150 | std::vector<BasicBlock *> BlocksToRemove; | |||
| 1151 | for (BasicBlock *IncomingBB : Phi->blocks()) { | |||
| 1152 | if (!isPredecessor(BB, IncomingBB)) | |||
| 1153 | BlocksToRemove.push_back(IncomingBB); | |||
| 1154 | } | |||
| 1155 | for (BasicBlock *BB : BlocksToRemove) | |||
| 1156 | Phi->removeIncomingValue(BB); | |||
| 1157 | } | |||
| 1158 | } | |||
| 1159 | ||||
| 1160 | /// Checks if BB was already cloned for a particular next state value. If it | |||
| 1161 | /// was then it returns this cloned block, and otherwise null. | |||
| 1162 | BasicBlock *getClonedBB(BasicBlock *BB, uint64_t NextState, | |||
| 1163 | DuplicateBlockMap &DuplicateMap) { | |||
| 1164 | CloneList ClonedBBs = DuplicateMap[BB]; | |||
| 1165 | ||||
| 1166 | // Find an entry in the CloneList with this NextState. If it exists then | |||
| 1167 | // return the corresponding BB | |||
| 1168 | auto It = llvm::find_if(ClonedBBs, [NextState](const ClonedBlock &C) { | |||
| 1169 | return C.State == NextState; | |||
| 1170 | }); | |||
| 1171 | return It != ClonedBBs.end() ? (*It).BB : nullptr; | |||
| 1172 | } | |||
| 1173 | ||||
| 1174 | /// Helper to get the successor corresponding to a particular case value for | |||
| 1175 | /// a switch statement. | |||
| 1176 | BasicBlock *getNextCaseSuccessor(SwitchInst *Switch, uint64_t NextState) { | |||
| 1177 | BasicBlock *NextCase = nullptr; | |||
| 1178 | for (auto Case : Switch->cases()) { | |||
| 1179 | if (Case.getCaseValue()->getZExtValue() == NextState) { | |||
| 1180 | NextCase = Case.getCaseSuccessor(); | |||
| 1181 | break; | |||
| 1182 | } | |||
| 1183 | } | |||
| 1184 | if (!NextCase) | |||
| 1185 | NextCase = Switch->getDefaultDest(); | |||
| 1186 | return NextCase; | |||
| 1187 | } | |||
| 1188 | ||||
| 1189 | /// Returns true if IncomingBB is a predecessor of BB. | |||
| 1190 | bool isPredecessor(BasicBlock *BB, BasicBlock *IncomingBB) { | |||
| 1191 | return llvm::find(predecessors(BB), IncomingBB) != pred_end(BB); | |||
| 1192 | } | |||
| 1193 | ||||
| 1194 | AllSwitchPaths *SwitchPaths; | |||
| 1195 | DominatorTree *DT; | |||
| 1196 | AssumptionCache *AC; | |||
| 1197 | TargetTransformInfo *TTI; | |||
| 1198 | OptimizationRemarkEmitter *ORE; | |||
| 1199 | SmallPtrSet<const Value *, 32> EphValues; | |||
| 1200 | std::vector<ThreadingPath> TPaths; | |||
| 1201 | }; | |||
| 1202 | ||||
| 1203 | bool DFAJumpThreading::run(Function &F) { | |||
| 1204 | LLVM_DEBUG(dbgs() << "\nDFA Jump threading: " << F.getName() << "\n")do { } while (false); | |||
| 1205 | ||||
| 1206 | if (F.hasOptSize()) { | |||
| 1207 | LLVM_DEBUG(dbgs() << "Skipping due to the 'minsize' attribute\n")do { } while (false); | |||
| 1208 | return false; | |||
| 1209 | } | |||
| 1210 | ||||
| 1211 | if (ClViewCfgBefore) | |||
| 1212 | F.viewCFG(); | |||
| 1213 | ||||
| 1214 | SmallVector<AllSwitchPaths, 2> ThreadableLoops; | |||
| 1215 | bool MadeChanges = false; | |||
| 1216 | ||||
| 1217 | for (BasicBlock &BB : F) { | |||
| 1218 | auto *SI = dyn_cast<SwitchInst>(BB.getTerminator()); | |||
| 1219 | if (!SI
| |||
| 1220 | continue; | |||
| 1221 | ||||
| 1222 | LLVM_DEBUG(dbgs() << "\nCheck if SwitchInst in BB " << BB.getName()do { } while (false) | |||
| 1223 | << " is predictable\n")do { } while (false); | |||
| 1224 | MainSwitch Switch(SI, ORE); | |||
| 1225 | ||||
| 1226 | if (!Switch.getInstr()) | |||
| 1227 | continue; | |||
| 1228 | ||||
| 1229 | LLVM_DEBUG(dbgs() << "\nSwitchInst in BB " << BB.getName() << " is a "do { } while (false) | |||
| 1230 | << "candidate for jump threading\n")do { } while (false); | |||
| 1231 | LLVM_DEBUG(SI->dump())do { } while (false); | |||
| 1232 | ||||
| 1233 | unfoldSelectInstrs(DT, Switch.getSelectInsts()); | |||
| 1234 | if (!Switch.getSelectInsts().empty()) | |||
| 1235 | MadeChanges = true; | |||
| 1236 | ||||
| 1237 | AllSwitchPaths SwitchPaths(&Switch, ORE); | |||
| 1238 | SwitchPaths.run(); | |||
| 1239 | ||||
| 1240 | if (SwitchPaths.getNumThreadingPaths() > 0) { | |||
| 1241 | ThreadableLoops.push_back(SwitchPaths); | |||
| 1242 | ||||
| 1243 | // For the time being limit this optimization to occurring once in a | |||
| 1244 | // function since it can change the CFG significantly. This is not a | |||
| 1245 | // strict requirement but it can cause buggy behavior if there is an | |||
| 1246 | // overlap of blocks in different opportunities. There is a lot of room to | |||
| 1247 | // experiment with catching more opportunities here. | |||
| 1248 | break; | |||
| 1249 | } | |||
| 1250 | } | |||
| 1251 | ||||
| 1252 | SmallPtrSet<const Value *, 32> EphValues; | |||
| 1253 | if (ThreadableLoops.size() > 0) | |||
| 1254 | CodeMetrics::collectEphemeralValues(&F, AC, EphValues); | |||
| 1255 | ||||
| 1256 | for (AllSwitchPaths SwitchPaths : ThreadableLoops) { | |||
| 1257 | TransformDFA Transform(&SwitchPaths, DT, AC, TTI, ORE, EphValues); | |||
| 1258 | Transform.run(); | |||
| 1259 | MadeChanges = true; | |||
| 1260 | } | |||
| 1261 | ||||
| 1262 | #ifdef EXPENSIVE_CHECKS | |||
| 1263 | assert(DT->verify(DominatorTree::VerificationLevel::Full))((void)0); | |||
| 1264 | verifyFunction(F, &dbgs()); | |||
| 1265 | #endif | |||
| 1266 | ||||
| 1267 | return MadeChanges; | |||
| 1268 | } | |||
| 1269 | ||||
| 1270 | } // end anonymous namespace | |||
| 1271 | ||||
| 1272 | /// Integrate with the new Pass Manager | |||
| 1273 | PreservedAnalyses DFAJumpThreadingPass::run(Function &F, | |||
| 1274 | FunctionAnalysisManager &AM) { | |||
| 1275 | AssumptionCache &AC = AM.getResult<AssumptionAnalysis>(F); | |||
| 1276 | DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F); | |||
| 1277 | TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F); | |||
| 1278 | OptimizationRemarkEmitter ORE(&F); | |||
| 1279 | ||||
| 1280 | if (!DFAJumpThreading(&AC, &DT, &TTI, &ORE).run(F)) | |||
| ||||
| 1281 | return PreservedAnalyses::all(); | |||
| 1282 | ||||
| 1283 | PreservedAnalyses PA; | |||
| 1284 | PA.preserve<DominatorTreeAnalysis>(); | |||
| 1285 | return PA; | |||
| 1286 | } |