1 //===- PromoteMemoryToRegister.cpp - Convert memory refs to regs ----------===//
3 // This file is used to promote memory references to be register references. A
4 // simple example of the transformation performed by this function is:
7 // %X = alloca int, uint 1 ret int 42
8 // store int 42, int *%X
12 // The code is transformed by looping over all of the alloca instruction,
13 // calculating dominator frontiers, then inserting phi-nodes following the usual
14 // SSA construction algorithm. This code does not modify the CFG of the
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
20 #include "llvm/Analysis/Dominators.h"
21 #include "llvm/iMemory.h"
22 #include "llvm/iPHINode.h"
23 #include "llvm/iTerminators.h"
24 #include "llvm/Function.h"
25 #include "llvm/Constant.h"
26 #include "llvm/Type.h"
27 #include "llvm/Support/CFG.h"
28 #include "Support/StringExtras.h"
30 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
31 /// This is true if there are only loads and stores to the alloca...
33 bool isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
34 // FIXME: If the memory unit is of pointer or integer type, we can permit
35 // assignments to subsections of the memory unit.
37 // Only allow direct loads and stores...
38 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
39 UI != UE; ++UI) // Loop over all of the uses of the alloca
40 if (!isa<LoadInst>(*UI))
41 if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
42 if (SI->getOperand(0) == AI)
43 return false; // Don't allow a store of the AI, only INTO the AI.
45 return false; // Not a load or store?
53 struct PromoteMem2Reg {
54 const std::vector<AllocaInst*> &Allocas; // the alloca instructions..
55 std::vector<unsigned> VersionNumbers; // Current version counters
56 DominanceFrontier &DF;
59 std::map<Instruction*, unsigned> AllocaLookup; // reverse mapping of above
61 std::vector<std::vector<BasicBlock*> > PhiNodes;// Idx corresponds 2 Allocas
63 // NewPhiNodes - The PhiNodes we're adding.
64 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
66 // Visited - The set of basic blocks the renamer has already visited.
67 std::set<BasicBlock*> Visited;
70 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominanceFrontier &df,
71 const TargetData &td) : Allocas(A), DF(df), TD(td) {}
76 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
77 std::vector<Value*> &IncVals);
78 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx);
80 } // end of anonymous namespace
83 void PromoteMem2Reg::run() {
84 Function &F = *DF.getRoot()->getParent();
86 VersionNumbers.resize(Allocas.size());
88 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
89 assert(isAllocaPromotable(Allocas[i], TD) &&
90 "Cannot promote non-promotable alloca!");
91 assert(Allocas[i]->getParent()->getParent() == &F &&
92 "All allocas should be in the same function, which is same as DF!");
93 AllocaLookup[Allocas[i]] = i;
96 // Calculate the set of write-locations for each alloca. This is analogous to
97 // counting the number of 'redefinitions' of each variable.
98 std::vector<std::vector<BasicBlock*> > WriteSets;// Idx corresponds to Allocas
99 WriteSets.resize(Allocas.size());
100 for (unsigned i = 0; i != Allocas.size(); ++i) {
101 AllocaInst *AI = Allocas[i];
102 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E; ++U)
103 if (StoreInst *SI = dyn_cast<StoreInst>(*U))
104 // jot down the basic-block it came from
105 WriteSets[i].push_back(SI->getParent());
108 // Compute the locations where PhiNodes need to be inserted. Look at the
109 // dominance frontier of EACH basic-block we have a write in
111 PhiNodes.resize(Allocas.size());
112 for (unsigned i = 0; i != Allocas.size(); ++i) {
113 for (unsigned j = 0; j != WriteSets[i].size(); j++) {
114 // Look up the DF for this write, add it to PhiNodes
115 DominanceFrontier::const_iterator it = DF.find(WriteSets[i][j]);
116 if (it != DF.end()) {
117 const DominanceFrontier::DomSetType &S = it->second;
118 for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
124 // Perform iterative step
125 for (unsigned k = 0; k != PhiNodes[i].size(); k++) {
126 DominanceFrontier::const_iterator it = DF.find(PhiNodes[i][k]);
127 if (it != DF.end()) {
128 const DominanceFrontier::DomSetType &S = it->second;
129 for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
136 // Set the incoming values for the basic block to be null values for all of
137 // the alloca's. We do this in case there is a load of a value that has not
138 // been stored yet. In this case, it will get this null value.
140 std::vector<Value *> Values(Allocas.size());
141 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
142 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
144 // Walks all basic blocks in the function performing the SSA rename algorithm
145 // and inserting the phi nodes we marked as necessary
147 RenamePass(F.begin(), 0, Values);
150 // Remove the allocas themselves from the function...
151 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
152 Instruction *A = Allocas[i];
154 // If there are any uses of the alloca instructions left, they must be in
155 // sections of dead code that were not processed on the dominance frontier.
156 // Just delete the users now.
159 A->replaceAllUsesWith(Constant::getNullValue(A->getType()));
160 A->getParent()->getInstList().erase(A);
165 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
166 // Alloca returns true if there wasn't already a phi-node for that variable
168 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo) {
169 // Look up the basic-block in question
170 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB];
171 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
173 // If the BB already has a phi node added for the i'th alloca then we're done!
174 if (BBPNs[AllocaNo]) return false;
176 // Create a PhiNode using the dereferenced type... and add the phi-node to the
178 PHINode *PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
179 Allocas[AllocaNo]->getName() + "." +
180 utostr(VersionNumbers[AllocaNo]++),
183 // Add null incoming values for all predecessors. This ensures that if one of
184 // the predecessors is not found in the depth-first traversal of the CFG (ie,
185 // because it is an unreachable predecessor), that all PHI nodes will have the
186 // correct number of entries for their predecessors.
187 Value *NullVal = Constant::getNullValue(PN->getType());
189 // This is necessary because adding incoming values to the PHI node adds uses
190 // to the basic blocks being used, which can invalidate the predecessor
192 std::vector<BasicBlock*> Preds(pred_begin(BB), pred_end(BB));
193 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
194 PN->addIncoming(NullVal, Preds[i]);
196 BBPNs[AllocaNo] = PN;
197 PhiNodes[AllocaNo].push_back(BB);
201 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
202 std::vector<Value*> &IncomingVals) {
203 // If this is a BB needing a phi node, lookup/create the phinode for each
204 // variable we need phinodes for.
205 std::vector<PHINode *> &BBPNs = NewPhiNodes[BB];
206 for (unsigned k = 0; k != BBPNs.size(); ++k)
207 if (PHINode *PN = BBPNs[k]) {
208 // The PHI node may have multiple entries for this predecessor. We must
209 // make sure we update all of them.
210 for (unsigned i = 0, e = PN->getNumOperands(); i != e; i += 2) {
211 if (PN->getOperand(i+1) == Pred)
212 // At this point we can assume that the array has phi nodes.. let's
213 // update the incoming data.
214 PN->setOperand(i, IncomingVals[k]);
216 // also note that the active variable IS designated by the phi node
217 IncomingVals[k] = PN;
220 // don't revisit nodes
221 if (Visited.count(BB)) return;
226 BasicBlock::iterator II = BB->begin();
228 Instruction *I = II++; // get the instruction, increment iterator
230 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
231 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
232 std::map<Instruction*, unsigned>::iterator AI = AllocaLookup.find(Src);
233 if (AI != AllocaLookup.end()) {
234 Value *V = IncomingVals[AI->second];
236 // walk the use list of this load and replace all uses with r
237 LI->replaceAllUsesWith(V);
238 BB->getInstList().erase(LI);
241 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
242 // Delete this instruction and mark the name as the current holder of the
244 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
245 std::map<Instruction *, unsigned>::iterator ai =AllocaLookup.find(Dest);
246 if (ai != AllocaLookup.end()) {
247 // what value were we writing?
248 IncomingVals[ai->second] = SI->getOperand(0);
249 BB->getInstList().erase(SI);
253 } else if (TerminatorInst *TI = dyn_cast<TerminatorInst>(I)) {
254 // Recurse across our successors
255 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
256 std::vector<Value*> OutgoingVals(IncomingVals);
257 RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
264 /// PromoteMemToReg - Promote the specified list of alloca instructions into
265 /// scalar registers, inserting PHI nodes as appropriate. This function makes
266 /// use of DominanceFrontier information. This function does not modify the CFG
267 /// of the function at all. All allocas must be from the same function.
269 void PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
270 DominanceFrontier &DF, const TargetData &TD) {
271 // If there is nothing to do, bail out...
272 if (Allocas.empty()) return;
273 PromoteMem2Reg(Allocas, DF, TD).run();