PromoteMemoryToRegister.cpp revision c6f3ae5c66c8e0dab6a2bd9601d0e253ef9ba794
1//===- PromoteMemoryToRegister.cpp - Convert memory refs to regs ----------===// 2// 3// This pass is used to promote memory references to be register references. A 4// simple example of the transformation performed by this pass is: 5// 6// FROM CODE TO CODE 7// %X = alloca int, uint 1 ret int 42 8// store int 42, int *%X 9// %Y = load int* %X 10// ret int %Y 11// 12// To do this transformation, a simple analysis is done to ensure it is safe. 13// Currently this just loops over all alloca instructions, looking for 14// instructions that are only used in simple load and stores. 15// 16// After this, the code is transformed by...something magical :) 17// 18//===----------------------------------------------------------------------===// 19 20#include "llvm/Transforms/Scalar/PromoteMemoryToRegister.h" 21#include "llvm/Analysis/Dominators.h" 22#include "llvm/iMemory.h" 23#include "llvm/iPHINode.h" 24#include "llvm/iTerminators.h" 25#include "llvm/Function.h" 26#include "llvm/BasicBlock.h" 27#include "llvm/Constant.h" 28 29using std::vector; 30using std::map; 31using std::set; 32 33namespace { 34 struct PromotePass : public FunctionPass { 35 vector<AllocaInst*> Allocas; // the alloca instruction.. 36 map<Instruction*, unsigned> AllocaLookup; // reverse mapping of above 37 38 vector<vector<BasicBlock*> > PhiNodes; // index corresponds to Allocas 39 40 // List of instructions to remove at end of pass 41 vector<Instruction *> KillList; 42 43 map<BasicBlock*,vector<PHINode*> > NewPhiNodes; // the PhiNodes we're adding 44 45 public: 46 const char *getPassName() const { return "Promote Memory to Register"; } 47 48 // runOnFunction - To run this pass, first we calculate the alloca 49 // instructions that are safe for promotion, then we promote each one. 50 // 51 virtual bool runOnFunction(Function *F); 52 53 // getAnalysisUsage - We need dominance frontiers 54 // 55 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 56 AU.addRequired(DominanceFrontier::ID); 57 AU.preservesCFG(); 58 } 59 60 private: 61 void Traverse(BasicBlock *BB, BasicBlock *Pred, vector<Value*> &IncVals, 62 set<BasicBlock*> &Visited); 63 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx); 64 void FindSafeAllocas(Function *F); 65 }; 66 67} // end of anonymous namespace 68 69 70// isSafeAlloca - This predicate controls what types of alloca instructions are 71// allowed to be promoted... 72// 73static inline bool isSafeAlloca(const AllocaInst *AI) { 74 if (AI->isArrayAllocation()) return false; 75 76 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end(); 77 UI != UE; ++UI) { // Loop over all of the uses of the alloca 78 79 // Only allow nonindexed memory access instructions... 80 if (MemAccessInst *MAI = dyn_cast<MemAccessInst>(*UI)) { 81 if (MAI->hasIndices()) { // indexed? 82 // Allow the access if there is only one index and the index is 83 // zero. 84 if (*MAI->idx_begin() != Constant::getNullValue(Type::UIntTy) || 85 MAI->idx_begin()+1 != MAI->idx_end()) 86 return false; 87 } 88 } else { 89 return false; // Not a load or store? 90 } 91 } 92 93 return true; 94} 95 96// FindSafeAllocas - Find allocas that are safe to promote 97// 98void PromotePass::FindSafeAllocas(Function *F) { 99 BasicBlock *BB = F->getEntryNode(); // Get the entry node for the function 100 101 // Look at all instructions in the entry node 102 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 103 if (AllocaInst *AI = dyn_cast<AllocaInst>(*I)) // Is it an alloca? 104 if (isSafeAlloca(AI)) { // If safe alloca, add alloca to safe list 105 AllocaLookup[AI] = Allocas.size(); // Keep reverse mapping 106 Allocas.push_back(AI); 107 } 108} 109 110 111 112bool PromotePass::runOnFunction(Function *F) { 113 // Calculate the set of safe allocas 114 FindSafeAllocas(F); 115 116 // If there is nothing to do, bail out... 117 if (Allocas.empty()) return false; 118 119 // Add each alloca to the KillList. Note: KillList is destroyed MOST recently 120 // added to least recently. 121 KillList.assign(Allocas.begin(), Allocas.end()); 122 123 // Calculate the set of write-locations for each alloca. This is analogous to 124 // counting the number of 'redefinitions' of each variable. 125 vector<vector<BasicBlock*> > WriteSets; // index corresponds to Allocas 126 WriteSets.resize(Allocas.size()); 127 for (unsigned i = 0; i != Allocas.size(); ++i) { 128 AllocaInst *AI = Allocas[i]; 129 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E; ++U) 130 if (StoreInst *SI = dyn_cast<StoreInst>(*U)) 131 // jot down the basic-block it came from 132 WriteSets[i].push_back(SI->getParent()); 133 } 134 135 // Get dominance frontier information... 136 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 137 138 // Compute the locations where PhiNodes need to be inserted. Look at the 139 // dominance frontier of EACH basic-block we have a write in 140 // 141 PhiNodes.resize(Allocas.size()); 142 for (unsigned i = 0; i != Allocas.size(); ++i) { 143 for (unsigned j = 0; j != WriteSets[i].size(); j++) { 144 // Look up the DF for this write, add it to PhiNodes 145 DominanceFrontier::const_iterator it = DF.find(WriteSets[i][j]); 146 DominanceFrontier::DomSetType S = it->second; 147 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end(); 148 P != PE; ++P) 149 QueuePhiNode(*P, i); 150 } 151 152 // Perform iterative step 153 for (unsigned k = 0; k != PhiNodes[i].size(); k++) { 154 DominanceFrontier::const_iterator it = DF.find(PhiNodes[i][k]); 155 DominanceFrontier::DomSetType S = it->second; 156 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end(); 157 P != PE; ++P) 158 QueuePhiNode(*P, i); 159 } 160 } 161 162 // Set the incoming values for the basic block to be null values for all of 163 // the alloca's. We do this in case there is a load of a value that has not 164 // been stored yet. In this case, it will get this null value. 165 // 166 vector<Value *> Values(Allocas.size()); 167 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) 168 Values[i] = Constant::getNullValue(Allocas[i]->getType()->getElementType()); 169 170 // Walks all basic blocks in the function performing the SSA rename algorithm 171 // and inserting the phi nodes we marked as necessary 172 // 173 set<BasicBlock*> Visited; // The basic blocks we've already visited 174 Traverse(F->front(), 0, Values, Visited); 175 176 // Remove all instructions marked by being placed in the KillList... 177 // 178 while (!KillList.empty()) { 179 Instruction *I = KillList.back(); 180 KillList.pop_back(); 181 182 I->getParent()->getInstList().remove(I); 183 delete I; 184 } 185 186 // Purge data structurse so they are available the next iteration... 187 Allocas.clear(); 188 AllocaLookup.clear(); 189 PhiNodes.clear(); 190 NewPhiNodes.clear(); 191 return true; 192} 193 194 195// QueuePhiNode - queues a phi-node to be added to a basic-block for a specific 196// Alloca returns true if there wasn't already a phi-node for that variable 197// 198bool PromotePass::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo) { 199 // Look up the basic-block in question 200 vector<PHINode*> &BBPNs = NewPhiNodes[BB]; 201 if (BBPNs.empty()) BBPNs.resize(Allocas.size()); 202 203 // If the BB already has a phi node added for the i'th alloca then we're done! 204 if (BBPNs[AllocaNo]) return false; 205 206 // Create a PhiNode using the dereferenced type... 207 PHINode *PN = new PHINode(Allocas[AllocaNo]->getType()->getElementType(), 208 Allocas[AllocaNo]->getName()+".mem2reg"); 209 BBPNs[AllocaNo] = PN; 210 211 // Add the phi-node to the basic-block 212 BB->getInstList().push_front(PN); 213 214 PhiNodes[AllocaNo].push_back(BB); 215 return true; 216} 217 218void PromotePass::Traverse(BasicBlock *BB, BasicBlock *Pred, 219 vector<Value*> &IncomingVals, 220 set<BasicBlock*> &Visited) { 221 // If this is a BB needing a phi node, lookup/create the phinode for each 222 // variable we need phinodes for. 223 vector<PHINode *> &BBPNs = NewPhiNodes[BB]; 224 for (unsigned k = 0; k != BBPNs.size(); ++k) 225 if (PHINode *PN = BBPNs[k]) { 226 // at this point we can assume that the array has phi nodes.. let's add 227 // the incoming data 228 PN->addIncoming(IncomingVals[k], Pred); 229 230 // also note that the active variable IS designated by the phi node 231 IncomingVals[k] = PN; 232 } 233 234 // don't revisit nodes 235 if (Visited.count(BB)) return; 236 237 // mark as visited 238 Visited.insert(BB); 239 240 // keep track of the value of each variable we're watching.. how? 241 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II) { 242 Instruction *I = *II; //get the instruction 243 244 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 245 Value *Ptr = LI->getPointerOperand(); 246 247 if (AllocaInst *Src = dyn_cast<AllocaInst>(Ptr)) { 248 map<Instruction*, unsigned>::iterator AI = AllocaLookup.find(Src); 249 if (AI != AllocaLookup.end()) { 250 Value *V = IncomingVals[AI->second]; 251 252 // walk the use list of this load and replace all uses with r 253 LI->replaceAllUsesWith(V); 254 KillList.push_back(LI); // Mark the load to be deleted 255 } 256 } 257 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 258 // delete this instruction and mark the name as the current holder of the 259 // value 260 Value *Ptr = SI->getPointerOperand(); 261 if (AllocaInst *Dest = dyn_cast<AllocaInst>(Ptr)) { 262 map<Instruction *, unsigned>::iterator ai = AllocaLookup.find(Dest); 263 if (ai != AllocaLookup.end()) { 264 // what value were we writing? 265 IncomingVals[ai->second] = SI->getOperand(0); 266 KillList.push_back(SI); // Mark the store to be deleted 267 } 268 } 269 270 } else if (TerminatorInst *TI = dyn_cast<TerminatorInst>(I)) { 271 // Recurse across our successors 272 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) { 273 vector<Value*> OutgoingVals(IncomingVals); 274 Traverse(TI->getSuccessor(i), BB, OutgoingVals, Visited); 275 } 276 } 277 } 278} 279 280 281// createPromoteMemoryToRegister - Provide an entry point to create this pass. 282// 283Pass *createPromoteMemoryToRegister() { 284 return new PromotePass(); 285} 286