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/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 #include "llvm/Type.h"
35 struct PromotePass : public FunctionPass {
36 vector<AllocaInst*> Allocas; // the alloca instruction..
37 map<Instruction*, unsigned> AllocaLookup; // reverse mapping of above
39 vector<vector<BasicBlock*> > PhiNodes; // index corresponds to Allocas
41 // List of instructions to remove at end of pass
42 vector<Instruction *> KillList;
44 map<BasicBlock*,vector<PHINode*> > NewPhiNodes; // the PhiNodes we're adding
47 const char *getPassName() const { return "Promote Memory to Register"; }
49 // runOnFunction - To run this pass, first we calculate the alloca
50 // instructions that are safe for promotion, then we promote each one.
52 virtual bool runOnFunction(Function *F);
54 // getAnalysisUsage - We need dominance frontiers
56 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
57 AU.addRequired(DominanceFrontier::ID);
62 void Traverse(BasicBlock *BB, BasicBlock *Pred, vector<Value*> &IncVals,
63 set<BasicBlock*> &Visited);
64 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx);
65 void FindSafeAllocas(Function *F);
68 } // end of anonymous namespace
71 // isSafeAlloca - This predicate controls what types of alloca instructions are
72 // allowed to be promoted...
74 static inline bool isSafeAlloca(const AllocaInst *AI) {
75 if (AI->isArrayAllocation()) return false;
77 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
78 UI != UE; ++UI) { // Loop over all of the uses of the alloca
80 // Only allow nonindexed memory access instructions...
81 if (MemAccessInst *MAI = dyn_cast<MemAccessInst>(*UI)) {
82 if (MAI->hasIndices()) { // indexed?
83 // Allow the access if there is only one index and the index is
85 if (*MAI->idx_begin() != Constant::getNullValue(Type::UIntTy) ||
86 MAI->idx_begin()+1 != MAI->idx_end())
90 return false; // Not a load or store?
97 // FindSafeAllocas - Find allocas that are safe to promote
99 void PromotePass::FindSafeAllocas(Function *F) {
100 BasicBlock *BB = F->getEntryNode(); // Get the entry node for the function
102 // Look at all instructions in the entry node
103 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
104 if (AllocaInst *AI = dyn_cast<AllocaInst>(*I)) // Is it an alloca?
105 if (isSafeAlloca(AI)) { // If safe alloca, add alloca to safe list
106 AllocaLookup[AI] = Allocas.size(); // Keep reverse mapping
107 Allocas.push_back(AI);
113 bool PromotePass::runOnFunction(Function *F) {
114 // Calculate the set of safe allocas
117 // If there is nothing to do, bail out...
118 if (Allocas.empty()) return false;
120 // Add each alloca to the KillList. Note: KillList is destroyed MOST recently
121 // added to least recently.
122 KillList.assign(Allocas.begin(), Allocas.end());
124 // Calculate the set of write-locations for each alloca. This is analogous to
125 // counting the number of 'redefinitions' of each variable.
126 vector<vector<BasicBlock*> > WriteSets; // index corresponds to Allocas
127 WriteSets.resize(Allocas.size());
128 for (unsigned i = 0; i != Allocas.size(); ++i) {
129 AllocaInst *AI = Allocas[i];
130 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E; ++U)
131 if (StoreInst *SI = dyn_cast<StoreInst>(*U))
132 // jot down the basic-block it came from
133 WriteSets[i].push_back(SI->getParent());
136 // Get dominance frontier information...
137 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
139 // Compute the locations where PhiNodes need to be inserted. Look at the
140 // dominance frontier of EACH basic-block we have a write in
142 PhiNodes.resize(Allocas.size());
143 for (unsigned i = 0; i != Allocas.size(); ++i) {
144 for (unsigned j = 0; j != WriteSets[i].size(); j++) {
145 // Look up the DF for this write, add it to PhiNodes
146 DominanceFrontier::const_iterator it = DF.find(WriteSets[i][j]);
147 DominanceFrontier::DomSetType S = it->second;
148 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end();
153 // Perform iterative step
154 for (unsigned k = 0; k != PhiNodes[i].size(); k++) {
155 DominanceFrontier::const_iterator it = DF.find(PhiNodes[i][k]);
156 DominanceFrontier::DomSetType S = it->second;
157 for (DominanceFrontier::DomSetType::iterator P = S.begin(), PE = S.end();
163 // Set the incoming values for the basic block to be null values for all of
164 // the alloca's. We do this in case there is a load of a value that has not
165 // been stored yet. In this case, it will get this null value.
167 vector<Value *> Values(Allocas.size());
168 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
169 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
171 // Walks all basic blocks in the function performing the SSA rename algorithm
172 // and inserting the phi nodes we marked as necessary
174 set<BasicBlock*> Visited; // The basic blocks we've already visited
175 Traverse(F->front(), 0, Values, Visited);
177 // Remove all instructions marked by being placed in the KillList...
179 while (!KillList.empty()) {
180 Instruction *I = KillList.back();
183 I->getParent()->getInstList().remove(I);
187 // Purge data structurse so they are available the next iteration...
189 AllocaLookup.clear();
196 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
197 // Alloca returns true if there wasn't already a phi-node for that variable
199 bool PromotePass::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo) {
200 // Look up the basic-block in question
201 vector<PHINode*> &BBPNs = NewPhiNodes[BB];
202 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
204 // If the BB already has a phi node added for the i'th alloca then we're done!
205 if (BBPNs[AllocaNo]) return false;
207 // Create a PhiNode using the dereferenced type...
208 PHINode *PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
209 Allocas[AllocaNo]->getName()+".mem2reg");
210 BBPNs[AllocaNo] = PN;
212 // Add the phi-node to the basic-block
213 BB->getInstList().push_front(PN);
215 PhiNodes[AllocaNo].push_back(BB);
219 void PromotePass::Traverse(BasicBlock *BB, BasicBlock *Pred,
220 vector<Value*> &IncomingVals,
221 set<BasicBlock*> &Visited) {
222 // If this is a BB needing a phi node, lookup/create the phinode for each
223 // variable we need phinodes for.
224 vector<PHINode *> &BBPNs = NewPhiNodes[BB];
225 for (unsigned k = 0; k != BBPNs.size(); ++k)
226 if (PHINode *PN = BBPNs[k]) {
227 // at this point we can assume that the array has phi nodes.. let's add
229 PN->addIncoming(IncomingVals[k], Pred);
231 // also note that the active variable IS designated by the phi node
232 IncomingVals[k] = PN;
235 // don't revisit nodes
236 if (Visited.count(BB)) return;
241 // keep track of the value of each variable we're watching.. how?
242 for (BasicBlock::iterator II = BB->begin(); II != BB->end(); ++II) {
243 Instruction *I = *II; //get the instruction
245 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
246 Value *Ptr = LI->getPointerOperand();
248 if (AllocaInst *Src = dyn_cast<AllocaInst>(Ptr)) {
249 map<Instruction*, unsigned>::iterator AI = AllocaLookup.find(Src);
250 if (AI != AllocaLookup.end()) {
251 Value *V = IncomingVals[AI->second];
253 // walk the use list of this load and replace all uses with r
254 LI->replaceAllUsesWith(V);
255 KillList.push_back(LI); // Mark the load to be deleted
258 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
259 // delete this instruction and mark the name as the current holder of the
261 Value *Ptr = SI->getPointerOperand();
262 if (AllocaInst *Dest = dyn_cast<AllocaInst>(Ptr)) {
263 map<Instruction *, unsigned>::iterator ai = AllocaLookup.find(Dest);
264 if (ai != AllocaLookup.end()) {
265 // what value were we writing?
266 IncomingVals[ai->second] = SI->getOperand(0);
267 KillList.push_back(SI); // Mark the store to be deleted
271 } else if (TerminatorInst *TI = dyn_cast<TerminatorInst>(I)) {
272 // Recurse across our successors
273 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
274 vector<Value*> OutgoingVals(IncomingVals);
275 Traverse(TI->getSuccessor(i), BB, OutgoingVals, Visited);
282 // createPromoteMemoryToRegister - Provide an entry point to create this pass.
284 Pass *createPromoteMemoryToRegister() {
285 return new PromotePass();