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 const char *getPassName() const { return "Promote Memory to Register"; }
52 // runOnFunction - To run this pass, first we calculate the alloca
53 // instructions that are safe for promotion, then we promote each one.
55 virtual bool runOnFunction(Function &F);
57 // getAnalysisUsage - We need dominance frontiers
59 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
60 AU.addRequired(DominanceFrontier::ID);
65 void Traverse(BasicBlock *BB, BasicBlock *Pred, vector<Value*> &IncVals,
66 set<BasicBlock*> &Visited);
67 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx);
68 void FindSafeAllocas(Function &F);
71 } // end of anonymous namespace
74 // isSafeAlloca - This predicate controls what types of alloca instructions are
75 // allowed to be promoted...
77 static inline bool isSafeAlloca(const AllocaInst *AI) {
78 if (AI->isArrayAllocation()) return false;
80 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
81 UI != UE; ++UI) { // Loop over all of the uses of the alloca
83 // Only allow nonindexed memory access instructions...
84 if (MemAccessInst *MAI = dyn_cast<MemAccessInst>(*UI)) {
85 if (MAI->getPointerOperand() != (Value*)AI)
86 return false; // Reject stores of alloca pointer into some other loc.
88 if (MAI->hasIndices()) { // indexed?
89 // Allow the access if there is only one index and the index is
91 if (*MAI->idx_begin() != Constant::getNullValue(Type::UIntTy) ||
92 MAI->idx_begin()+1 != MAI->idx_end())
96 return false; // Not a load or store?
103 // FindSafeAllocas - Find allocas that are safe to promote
105 void PromotePass::FindSafeAllocas(Function &F) {
106 BasicBlock &BB = F.getEntryNode(); // Get the entry node for the function
108 // Look at all instructions in the entry node
109 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
110 if (AllocaInst *AI = dyn_cast<AllocaInst>(&*I)) // Is it an alloca?
111 if (isSafeAlloca(AI)) { // If safe alloca, add alloca to safe list
112 AllocaLookup[AI] = Allocas.size(); // Keep reverse mapping
113 Allocas.push_back(AI);
119 bool PromotePass::runOnFunction(Function &F) {
120 // Calculate the set of safe allocas
123 // If there is nothing to do, bail out...
124 if (Allocas.empty()) return false;
126 // Add each alloca to the KillList. Note: KillList is destroyed MOST recently
127 // added to least recently.
128 KillList.assign(Allocas.begin(), Allocas.end());
130 // Calculate the set of write-locations for each alloca. This is analogous to
131 // counting the number of 'redefinitions' of each variable.
132 vector<vector<BasicBlock*> > WriteSets; // index corresponds to Allocas
133 WriteSets.resize(Allocas.size());
134 for (unsigned i = 0; i != Allocas.size(); ++i) {
135 AllocaInst *AI = Allocas[i];
136 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E; ++U)
137 if (StoreInst *SI = dyn_cast<StoreInst>(*U))
138 // jot down the basic-block it came from
139 WriteSets[i].push_back(SI->getParent());
142 // Get dominance frontier information...
143 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
145 // Compute the locations where PhiNodes need to be inserted. Look at the
146 // dominance frontier of EACH basic-block we have a write in
148 PhiNodes.resize(Allocas.size());
149 for (unsigned i = 0; i != Allocas.size(); ++i) {
150 for (unsigned j = 0; j != WriteSets[i].size(); j++) {
151 // Look up the DF for this write, add it to PhiNodes
152 DominanceFrontier::const_iterator it = DF.find(WriteSets[i][j]);
153 DominanceFrontier::DomSetType S = it->second;
154 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end();
159 // Perform iterative step
160 for (unsigned k = 0; k != PhiNodes[i].size(); k++) {
161 DominanceFrontier::const_iterator it = DF.find(PhiNodes[i][k]);
162 DominanceFrontier::DomSetType S = it->second;
163 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end();
169 // Set the incoming values for the basic block to be null values for all of
170 // the alloca's. We do this in case there is a load of a value that has not
171 // been stored yet. In this case, it will get this null value.
173 vector<Value *> Values(Allocas.size());
174 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
175 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
177 // Walks all basic blocks in the function performing the SSA rename algorithm
178 // and inserting the phi nodes we marked as necessary
180 set<BasicBlock*> Visited; // The basic blocks we've already visited
181 Traverse(F.begin(), 0, Values, Visited);
183 // Remove all instructions marked by being placed in the KillList...
185 while (!KillList.empty()) {
186 Instruction *I = KillList.back();
189 I->getParent()->getInstList().erase(I);
192 NumPromoted += Allocas.size();
194 // Purge data structurse so they are available the next iteration...
196 AllocaLookup.clear();
203 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
204 // Alloca returns true if there wasn't already a phi-node for that variable
206 bool PromotePass::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo) {
207 // Look up the basic-block in question
208 vector<PHINode*> &BBPNs = NewPhiNodes[BB];
209 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
211 // If the BB already has a phi node added for the i'th alloca then we're done!
212 if (BBPNs[AllocaNo]) return false;
214 // Create a PhiNode using the dereferenced type...
215 PHINode *PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
216 Allocas[AllocaNo]->getName()+".mem2reg");
217 BBPNs[AllocaNo] = PN;
219 // Add the phi-node to the basic-block
220 BB->getInstList().push_front(PN);
222 PhiNodes[AllocaNo].push_back(BB);
226 void PromotePass::Traverse(BasicBlock *BB, BasicBlock *Pred,
227 vector<Value*> &IncomingVals,
228 set<BasicBlock*> &Visited) {
229 // If this is a BB needing a phi node, lookup/create the phinode for each
230 // variable we need phinodes for.
231 vector<PHINode *> &BBPNs = NewPhiNodes[BB];
232 for (unsigned k = 0; k != BBPNs.size(); ++k)
233 if (PHINode *PN = BBPNs[k]) {
234 // at this point we can assume that the array has phi nodes.. let's add
236 PN->addIncoming(IncomingVals[k], Pred);
238 // also note that the active variable IS designated by the phi node
239 IncomingVals[k] = PN;
242 // don't revisit nodes
243 if (Visited.count(BB)) return;
248 // keep track of the value of each variable we're watching.. how?
249 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II) {
250 Instruction *I = II; // get the instruction
252 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
253 Value *Ptr = LI->getPointerOperand();
255 if (AllocaInst *Src = dyn_cast<AllocaInst>(Ptr)) {
256 map<Instruction*, unsigned>::iterator AI = AllocaLookup.find(Src);
257 if (AI != AllocaLookup.end()) {
258 Value *V = IncomingVals[AI->second];
260 // walk the use list of this load and replace all uses with r
261 LI->replaceAllUsesWith(V);
262 KillList.push_back(LI); // Mark the load to be deleted
265 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
266 // delete this instruction and mark the name as the current holder of the
268 Value *Ptr = SI->getPointerOperand();
269 if (AllocaInst *Dest = dyn_cast<AllocaInst>(Ptr)) {
270 map<Instruction *, unsigned>::iterator ai = AllocaLookup.find(Dest);
271 if (ai != AllocaLookup.end()) {
272 // what value were we writing?
273 IncomingVals[ai->second] = SI->getOperand(0);
274 KillList.push_back(SI); // Mark the store to be deleted
278 } else if (TerminatorInst *TI = dyn_cast<TerminatorInst>(I)) {
279 // Recurse across our successors
280 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
281 vector<Value*> OutgoingVals(IncomingVals);
282 Traverse(TI->getSuccessor(i), BB, OutgoingVals, Visited);
289 // createPromoteMemoryToRegister - Provide an entry point to create this pass.
291 Pass *createPromoteMemoryToRegister() {
292 return new PromotePass();