1 //===- PromoteMemoryToRegister.cpp - Convert memory refs to regs ----------===//
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:
7 // %X = alloca int, uint 1 ret int 42
8 // store int 42, int *%X
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.
16 // After this, the code is transformed by...something magical :)
18 //===----------------------------------------------------------------------===//
20 #include "llvm/Transforms/Scalar.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 #include "llvm/Type.h"
29 #include "Support/StatisticReporter.h"
31 static Statistic<> NumPromoted("mem2reg\t\t- Number of alloca's promoted");
38 struct PromotePass : public FunctionPass {
39 vector<AllocaInst*> Allocas; // the alloca instruction..
40 map<Instruction*, unsigned> AllocaLookup; // reverse mapping of above
42 vector<vector<BasicBlock*> > PhiNodes; // index corresponds to Allocas
44 // List of instructions to remove at end of pass
45 vector<Instruction *> KillList;
47 map<BasicBlock*,vector<PHINode*> > NewPhiNodes; // the PhiNodes we're adding
50 // runOnFunction - To run this pass, first we calculate the alloca
51 // instructions that are safe for promotion, then we promote each one.
53 virtual bool runOnFunction(Function &F);
55 // getAnalysisUsage - We need dominance frontiers
57 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
58 AU.addRequired(DominanceFrontier::ID);
63 void Traverse(BasicBlock *BB, BasicBlock *Pred, vector<Value*> &IncVals,
64 set<BasicBlock*> &Visited);
65 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx);
66 void FindSafeAllocas(Function &F);
69 RegisterPass<PromotePass> X("mem2reg", "Promote Memory to Register");
70 } // end of anonymous namespace
73 // isSafeAlloca - This predicate controls what types of alloca instructions are
74 // allowed to be promoted...
76 static inline bool isSafeAlloca(const AllocaInst *AI) {
77 if (AI->isArrayAllocation()) return false;
79 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
80 UI != UE; ++UI) { // Loop over all of the uses of the alloca
82 // Only allow nonindexed memory access instructions...
83 if (MemAccessInst *MAI = dyn_cast<MemAccessInst>(*UI)) {
84 if (MAI->getPointerOperand() != (Value*)AI)
85 return false; // Reject stores of alloca pointer into some other loc.
87 if (MAI->hasIndices()) { // indexed?
88 // Allow the access if there is only one index and the index is
90 if (*MAI->idx_begin() != Constant::getNullValue(Type::UIntTy) ||
91 MAI->idx_begin()+1 != MAI->idx_end())
95 return false; // Not a load or store?
102 // FindSafeAllocas - Find allocas that are safe to promote
104 void PromotePass::FindSafeAllocas(Function &F) {
105 BasicBlock &BB = F.getEntryNode(); // Get the entry node for the function
107 // Look at all instructions in the entry node
108 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
109 if (AllocaInst *AI = dyn_cast<AllocaInst>(&*I)) // Is it an alloca?
110 if (isSafeAlloca(AI)) { // If safe alloca, add alloca to safe list
111 AllocaLookup[AI] = Allocas.size(); // Keep reverse mapping
112 Allocas.push_back(AI);
118 bool PromotePass::runOnFunction(Function &F) {
119 // Calculate the set of safe allocas
122 // If there is nothing to do, bail out...
123 if (Allocas.empty()) return false;
125 // Add each alloca to the KillList. Note: KillList is destroyed MOST recently
126 // added to least recently.
127 KillList.assign(Allocas.begin(), Allocas.end());
129 // Calculate the set of write-locations for each alloca. This is analogous to
130 // counting the number of 'redefinitions' of each variable.
131 vector<vector<BasicBlock*> > WriteSets; // index corresponds to Allocas
132 WriteSets.resize(Allocas.size());
133 for (unsigned i = 0; i != Allocas.size(); ++i) {
134 AllocaInst *AI = Allocas[i];
135 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E; ++U)
136 if (StoreInst *SI = dyn_cast<StoreInst>(*U))
137 // jot down the basic-block it came from
138 WriteSets[i].push_back(SI->getParent());
141 // Get dominance frontier information...
142 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
144 // Compute the locations where PhiNodes need to be inserted. Look at the
145 // dominance frontier of EACH basic-block we have a write in
147 PhiNodes.resize(Allocas.size());
148 for (unsigned i = 0; i != Allocas.size(); ++i) {
149 for (unsigned j = 0; j != WriteSets[i].size(); j++) {
150 // Look up the DF for this write, add it to PhiNodes
151 DominanceFrontier::const_iterator it = DF.find(WriteSets[i][j]);
152 DominanceFrontier::DomSetType S = it->second;
153 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end();
158 // Perform iterative step
159 for (unsigned k = 0; k != PhiNodes[i].size(); k++) {
160 DominanceFrontier::const_iterator it = DF.find(PhiNodes[i][k]);
161 DominanceFrontier::DomSetType S = it->second;
162 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end();
168 // Set the incoming values for the basic block to be null values for all of
169 // the alloca's. We do this in case there is a load of a value that has not
170 // been stored yet. In this case, it will get this null value.
172 vector<Value *> Values(Allocas.size());
173 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
174 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
176 // Walks all basic blocks in the function performing the SSA rename algorithm
177 // and inserting the phi nodes we marked as necessary
179 set<BasicBlock*> Visited; // The basic blocks we've already visited
180 Traverse(F.begin(), 0, Values, Visited);
182 // Remove all instructions marked by being placed in the KillList...
184 while (!KillList.empty()) {
185 Instruction *I = KillList.back();
188 I->getParent()->getInstList().erase(I);
191 NumPromoted += Allocas.size();
193 // Purge data structurse so they are available the next iteration...
195 AllocaLookup.clear();
202 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
203 // Alloca returns true if there wasn't already a phi-node for that variable
205 bool PromotePass::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo) {
206 // Look up the basic-block in question
207 vector<PHINode*> &BBPNs = NewPhiNodes[BB];
208 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
210 // If the BB already has a phi node added for the i'th alloca then we're done!
211 if (BBPNs[AllocaNo]) return false;
213 // Create a PhiNode using the dereferenced type...
214 PHINode *PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
215 Allocas[AllocaNo]->getName()+".mem2reg");
216 BBPNs[AllocaNo] = PN;
218 // Add the phi-node to the basic-block
219 BB->getInstList().push_front(PN);
221 PhiNodes[AllocaNo].push_back(BB);
225 void PromotePass::Traverse(BasicBlock *BB, BasicBlock *Pred,
226 vector<Value*> &IncomingVals,
227 set<BasicBlock*> &Visited) {
228 // If this is a BB needing a phi node, lookup/create the phinode for each
229 // variable we need phinodes for.
230 vector<PHINode *> &BBPNs = NewPhiNodes[BB];
231 for (unsigned k = 0; k != BBPNs.size(); ++k)
232 if (PHINode *PN = BBPNs[k]) {
233 // at this point we can assume that the array has phi nodes.. let's add
235 PN->addIncoming(IncomingVals[k], Pred);
237 // also note that the active variable IS designated by the phi node
238 IncomingVals[k] = PN;
241 // don't revisit nodes
242 if (Visited.count(BB)) return;
247 // keep track of the value of each variable we're watching.. how?
248 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II) {
249 Instruction *I = II; // get the instruction
251 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
252 Value *Ptr = LI->getPointerOperand();
254 if (AllocaInst *Src = dyn_cast<AllocaInst>(Ptr)) {
255 map<Instruction*, unsigned>::iterator AI = AllocaLookup.find(Src);
256 if (AI != AllocaLookup.end()) {
257 Value *V = IncomingVals[AI->second];
259 // walk the use list of this load and replace all uses with r
260 LI->replaceAllUsesWith(V);
261 KillList.push_back(LI); // Mark the load to be deleted
264 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
265 // delete this instruction and mark the name as the current holder of the
267 Value *Ptr = SI->getPointerOperand();
268 if (AllocaInst *Dest = dyn_cast<AllocaInst>(Ptr)) {
269 map<Instruction *, unsigned>::iterator ai = AllocaLookup.find(Dest);
270 if (ai != AllocaLookup.end()) {
271 // what value were we writing?
272 IncomingVals[ai->second] = SI->getOperand(0);
273 KillList.push_back(SI); // Mark the store to be deleted
277 } else if (TerminatorInst *TI = dyn_cast<TerminatorInst>(I)) {
278 // Recurse across our successors
279 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
280 vector<Value*> OutgoingVals(IncomingVals);
281 Traverse(TI->getSuccessor(i), BB, OutgoingVals, Visited);
288 // createPromoteMemoryToRegister - Provide an entry point to create this pass.
290 Pass *createPromoteMemoryToRegister() {
291 return new PromotePass();