1 //===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file promote memory references to be register references. It promotes
11 // alloca instructions which only have loads and stores as uses. An alloca is
12 // transformed by using dominator frontiers to place PHI nodes, then traversing
13 // the function in depth-first order to rewrite loads and stores as appropriate.
14 // This is just the standard SSA construction algorithm to construct "pruned"
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/Function.h"
24 #include "llvm/Constant.h"
25 #include "llvm/Support/CFG.h"
26 #include "Support/StringExtras.h"
28 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
29 /// This is true if there are only loads and stores to the alloca...
31 bool isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
32 // FIXME: If the memory unit is of pointer or integer type, we can permit
33 // assignments to subsections of the memory unit.
35 // Only allow direct loads and stores...
36 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
37 UI != UE; ++UI) // Loop over all of the uses of the alloca
38 if (!isa<LoadInst>(*UI))
39 if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
40 if (SI->getOperand(0) == AI)
41 return false; // Don't allow a store of the AI, only INTO the AI.
43 return false; // Not a load or store?
50 struct PromoteMem2Reg {
51 // Allocas - The alloca instructions being promoted
52 std::vector<AllocaInst*> Allocas;
54 DominanceFrontier &DF;
57 // AllocaLookup - Reverse mapping of Allocas
58 std::map<AllocaInst*, unsigned> AllocaLookup;
60 // NewPhiNodes - The PhiNodes we're adding.
61 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
63 // Visited - The set of basic blocks the renamer has already visited.
64 std::set<BasicBlock*> Visited;
67 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
68 DominanceFrontier &df, const TargetData &td)
69 : Allocas(A), DT(dt), DF(df), TD(td) {}
74 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
75 std::set<PHINode*> &DeadPHINodes);
76 void PromoteLocallyUsedAlloca(AllocaInst *AI);
78 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
79 std::vector<Value*> &IncVals);
80 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
81 std::set<PHINode*> &InsertedPHINodes);
83 } // end of anonymous namespace
85 void PromoteMem2Reg::run() {
86 Function &F = *DF.getRoot()->getParent();
88 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
89 AllocaInst *AI = Allocas[AllocaNum];
91 assert(isAllocaPromotable(AI, TD) &&
92 "Cannot promote non-promotable alloca!");
93 assert(AI->getParent()->getParent() == &F &&
94 "All allocas should be in the same function, which is same as DF!");
96 if (AI->use_empty()) {
97 // If there are no uses of the alloca, just delete it now.
98 AI->getParent()->getInstList().erase(AI);
100 // Remove the alloca from the Allocas list, since it has been processed
101 Allocas[AllocaNum] = Allocas.back();
107 // Calculate the set of read and write-locations for each alloca. This is
108 // analogous to counting the number of 'uses' and 'definitions' of each
110 std::vector<BasicBlock*> DefiningBlocks;
111 std::vector<BasicBlock*> UsingBlocks;
113 BasicBlock *OnlyBlock = 0;
114 bool OnlyUsedInOneBlock = true;
116 // As we scan the uses of the alloca instruction, keep track of stores, and
117 // decide whether all of the loads and stores to the alloca are within the
119 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
120 Instruction *User = cast<Instruction>(*U);
121 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
122 // Remember the basic blocks which define new values for the alloca
123 DefiningBlocks.push_back(SI->getParent());
125 // Otherwise it must be a load instruction, keep track of variable reads
126 UsingBlocks.push_back(cast<LoadInst>(User)->getParent());
129 if (OnlyUsedInOneBlock) {
131 OnlyBlock = User->getParent();
132 else if (OnlyBlock != User->getParent())
133 OnlyUsedInOneBlock = false;
137 // If the alloca is only read and written in one basic block, just perform a
138 // linear sweep over the block to eliminate it.
139 if (OnlyUsedInOneBlock) {
140 PromoteLocallyUsedAlloca(AI);
142 // Remove the alloca from the Allocas list, since it has been processed
143 Allocas[AllocaNum] = Allocas.back();
149 // Compute the locations where PhiNodes need to be inserted. Look at the
150 // dominance frontier of EACH basic-block we have a write in.
152 unsigned CurrentVersion = 0;
153 std::set<PHINode*> InsertedPHINodes;
154 while (!DefiningBlocks.empty()) {
155 BasicBlock *BB = DefiningBlocks.back();
156 DefiningBlocks.pop_back();
158 // Look up the DF for this write, add it to PhiNodes
159 DominanceFrontier::const_iterator it = DF.find(BB);
160 if (it != DF.end()) {
161 const DominanceFrontier::DomSetType &S = it->second;
162 for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
164 if (QueuePhiNode(*P, AllocaNum, CurrentVersion, InsertedPHINodes))
165 DefiningBlocks.push_back(*P);
169 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
170 // of the writes to the variable, scan through the reads of the variable,
171 // marking PHI nodes which are actually necessary as alive (by removing them
172 // from the InsertedPHINodes set). This is not perfect: there may PHI
173 // marked alive because of loads which are dominated by stores, but there
174 // will be no unmarked PHI nodes which are actually used.
176 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
177 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
180 // If there are any PHI nodes which are now known to be dead, remove them!
181 for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(),
182 E = InsertedPHINodes.end(); I != E; ++I) {
184 std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()];
185 BBPNs[AllocaNum] = 0;
187 // Check to see if we just removed the last inserted PHI node from this
188 // basic block. If so, remove the entry for the basic block.
189 bool HasOtherPHIs = false;
190 for (unsigned i = 0, e = BBPNs.size(); i != e; ++i)
196 NewPhiNodes.erase(PN->getParent());
198 PN->getParent()->getInstList().erase(PN);
201 // Keep the reverse mapping of the 'Allocas' array.
202 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
206 return; // All of the allocas must have been trivial!
208 // Set the incoming values for the basic block to be null values for all of
209 // the alloca's. We do this in case there is a load of a value that has not
210 // been stored yet. In this case, it will get this null value.
212 std::vector<Value *> Values(Allocas.size());
213 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
214 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
216 // Walks all basic blocks in the function performing the SSA rename algorithm
217 // and inserting the phi nodes we marked as necessary
219 RenamePass(F.begin(), 0, Values);
221 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
224 // Remove the allocas themselves from the function...
225 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
226 Instruction *A = Allocas[i];
228 // If there are any uses of the alloca instructions left, they must be in
229 // sections of dead code that were not processed on the dominance frontier.
230 // Just delete the users now.
233 A->replaceAllUsesWith(Constant::getNullValue(A->getType()));
234 A->getParent()->getInstList().erase(A);
237 // At this point, the renamer has added entries to PHI nodes for all reachable
238 // code. Unfortunately, there may be blocks which are not reachable, which
239 // the renamer hasn't traversed. If this is the case, the PHI nodes may not
240 // have incoming values for all predecessors. Loop over all PHI nodes we have
241 // created, inserting null constants if they are missing any incoming values.
243 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
244 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
246 std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first));
247 std::vector<PHINode*> &PNs = I->second;
248 assert(!PNs.empty() && "Empty PHI node list??");
250 // Only do work here if there the PHI nodes are missing incoming values. We
251 // know that all PHI nodes that were inserted in a block will have the same
252 // number of incoming values, so we can just check any PHI node.
254 for (unsigned i = 0; (FirstPHI = PNs[i]) == 0; ++i)
257 if (Preds.size() != FirstPHI->getNumIncomingValues()) {
258 // Ok, now we know that all of the PHI nodes are missing entries for some
259 // basic blocks. Start by sorting the incoming predecessors for efficient
261 std::sort(Preds.begin(), Preds.end());
263 // Now we loop through all BB's which have entries in FirstPHI and remove
264 // them from the Preds list.
265 for (unsigned i = 0, e = FirstPHI->getNumIncomingValues(); i != e; ++i) {
266 // Do a log(n) search of the Preds list for the entry we want.
267 std::vector<BasicBlock*>::iterator EntIt =
268 std::lower_bound(Preds.begin(), Preds.end(),
269 FirstPHI->getIncomingBlock(i));
270 assert(EntIt != Preds.end() && *EntIt == FirstPHI->getIncomingBlock(i)&&
271 "PHI node has entry for a block which is not a predecessor!");
277 // At this point, the blocks left in the preds list must have dummy
278 // entries inserted into every PHI nodes for the block.
279 for (unsigned i = 0, e = PNs.size(); i != e; ++i)
280 if (PHINode *PN = PNs[i]) {
281 Value *NullVal = Constant::getNullValue(PN->getType());
282 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
283 PN->addIncoming(NullVal, Preds[pred]);
289 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
290 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
291 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
292 // each read of the variable. For each block that reads the variable, this
293 // function is called, which removes used PHI nodes from the DeadPHINodes set.
294 // After all of the reads have been processed, any PHI nodes left in the
295 // DeadPHINodes set are removed.
297 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
298 std::set<PHINode*> &DeadPHINodes) {
299 // Scan the immediate dominators of this block looking for a block which has a
300 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
301 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) {
302 BasicBlock *DomBB = N->getBlock();
303 std::map<BasicBlock*, std::vector<PHINode*> >::iterator
304 I = NewPhiNodes.find(DomBB);
305 if (I != NewPhiNodes.end() && I->second[AllocaNum]) {
306 // Ok, we found an inserted PHI node which dominates this value.
307 PHINode *DominatingPHI = I->second[AllocaNum];
309 // Find out if we previously thought it was dead.
310 std::set<PHINode*>::iterator DPNI = DeadPHINodes.find(DominatingPHI);
311 if (DPNI != DeadPHINodes.end()) {
312 // Ok, until now, we thought this PHI node was dead. Mark it as being
314 DeadPHINodes.erase(DPNI);
316 // Now that we have marked the PHI node alive, also mark any PHI nodes
317 // which it might use as being alive as well.
318 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
320 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
326 // PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
327 // block. If this is the case, avoid traversing the CFG and inserting a lot of
328 // potentially useless PHI nodes by just performing a single linear pass over
329 // the basic block using the Alloca.
331 void PromoteMem2Reg::PromoteLocallyUsedAlloca(AllocaInst *AI) {
332 assert(!AI->use_empty() && "There are no uses of the alloca!");
334 // Uses of the uninitialized memory location shall get zero...
335 Value *CurVal = Constant::getNullValue(AI->getAllocatedType());
337 BasicBlock *BB = cast<Instruction>(AI->use_back())->getParent();
339 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
340 Instruction *Inst = I++;
341 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
342 if (LI->getOperand(0) == AI) {
343 // Loads just return the "current value"...
344 LI->replaceAllUsesWith(CurVal);
345 BB->getInstList().erase(LI);
347 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
348 if (SI->getOperand(1) == AI) {
349 // Loads just update the "current value"...
350 CurVal = SI->getOperand(0);
351 BB->getInstList().erase(SI);
356 // After traversing the basic block, there should be no more uses of the
357 // alloca, remove it now.
358 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
359 AI->getParent()->getInstList().erase(AI);
362 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
363 // Alloca returns true if there wasn't already a phi-node for that variable
365 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
367 std::set<PHINode*> &InsertedPHINodes) {
368 // Look up the basic-block in question
369 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB];
370 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
372 // If the BB already has a phi node added for the i'th alloca then we're done!
373 if (BBPNs[AllocaNo]) return false;
375 // Create a PhiNode using the dereferenced type... and add the phi-node to the
377 BBPNs[AllocaNo] = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
378 Allocas[AllocaNo]->getName() + "." +
379 utostr(Version++), BB->begin());
380 InsertedPHINodes.insert(BBPNs[AllocaNo]);
385 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
386 // stores to the allocas which we are promoting. IncomingVals indicates what
387 // value each Alloca contains on exit from the predecessor block Pred.
389 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
390 std::vector<Value*> &IncomingVals) {
392 // If this BB needs a PHI node, update the PHI node for each variable we need
394 std::map<BasicBlock*, std::vector<PHINode *> >::iterator
395 BBPNI = NewPhiNodes.find(BB);
396 if (BBPNI != NewPhiNodes.end()) {
397 std::vector<PHINode *> &BBPNs = BBPNI->second;
398 for (unsigned k = 0; k != BBPNs.size(); ++k)
399 if (PHINode *PN = BBPNs[k]) {
400 // Add this incoming value to the PHI node.
401 PN->addIncoming(IncomingVals[k], Pred);
403 // The currently active variable for this block is now the PHI.
404 IncomingVals[k] = PN;
408 // don't revisit nodes
409 if (Visited.count(BB)) return;
414 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
415 Instruction *I = II++; // get the instruction, increment iterator
417 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
418 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
419 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
420 if (AI != AllocaLookup.end()) {
421 Value *V = IncomingVals[AI->second];
423 // walk the use list of this load and replace all uses with r
424 LI->replaceAllUsesWith(V);
425 BB->getInstList().erase(LI);
428 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
429 // Delete this instruction and mark the name as the current holder of the
431 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
432 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
433 if (ai != AllocaLookup.end()) {
434 // what value were we writing?
435 IncomingVals[ai->second] = SI->getOperand(0);
436 BB->getInstList().erase(SI);
442 // Recurse to our successors.
443 TerminatorInst *TI = BB->getTerminator();
444 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
445 std::vector<Value*> OutgoingVals(IncomingVals);
446 RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
450 /// PromoteMemToReg - Promote the specified list of alloca instructions into
451 /// scalar registers, inserting PHI nodes as appropriate. This function makes
452 /// use of DominanceFrontier information. This function does not modify the CFG
453 /// of the function at all. All allocas must be from the same function.
455 void PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
456 DominatorTree &DT, DominanceFrontier &DF,
457 const TargetData &TD) {
458 // If there is nothing to do, bail out...
459 if (Allocas.empty()) return;
460 PromoteMem2Reg(Allocas, DT, DF, TD).run();