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 #define DEBUG_TYPE "mem2reg"
20 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/AliasSetTracker.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/Support/CFG.h"
33 #include "llvm/Support/Compiler.h"
37 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
38 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
39 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
41 // Provide DenseMapKeyInfo for all pointers.
44 struct DenseMapKeyInfo<std::pair<BasicBlock*, unsigned> > {
45 static inline std::pair<BasicBlock*, unsigned> getEmptyKey() {
46 return std::make_pair((BasicBlock*)-1, ~0U);
48 static inline std::pair<BasicBlock*, unsigned> getTombstoneKey() {
49 return std::make_pair((BasicBlock*)-2, 0U);
51 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
52 return DenseMapKeyInfo<void*>::getHashValue(Val.first) + Val.second*2;
54 static bool isPod() { return true; }
58 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
59 /// This is true if there are only loads and stores to the alloca.
61 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
62 // FIXME: If the memory unit is of pointer or integer type, we can permit
63 // assignments to subsections of the memory unit.
65 // Only allow direct loads and stores...
66 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
67 UI != UE; ++UI) // Loop over all of the uses of the alloca
68 if (isa<LoadInst>(*UI)) {
70 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
71 if (SI->getOperand(0) == AI)
72 return false; // Don't allow a store OF the AI, only INTO the AI.
74 return false; // Not a load or store.
82 // Data package used by RenamePass()
83 class VISIBILITY_HIDDEN RenamePassData {
85 typedef std::vector<Value *> ValVector;
88 RenamePassData(BasicBlock *B, BasicBlock *P,
89 const ValVector &V) : BB(B), Pred(P), Values(V) {}
94 void swap(RenamePassData &RHS) {
95 std::swap(BB, RHS.BB);
96 std::swap(Pred, RHS.Pred);
97 Values.swap(RHS.Values);
101 struct VISIBILITY_HIDDEN PromoteMem2Reg {
102 /// Allocas - The alloca instructions being promoted.
104 std::vector<AllocaInst*> Allocas;
105 SmallVector<AllocaInst*, 16> &RetryList;
107 DominanceFrontier &DF;
109 /// AST - An AliasSetTracker object to update. If null, don't update it.
111 AliasSetTracker *AST;
113 /// AllocaLookup - Reverse mapping of Allocas.
115 std::map<AllocaInst*, unsigned> AllocaLookup;
117 /// NewPhiNodes - The PhiNodes we're adding.
119 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
121 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
122 /// it corresponds to.
123 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
125 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
126 /// each alloca that is of pointer type, we keep track of what to copyValue
127 /// to the inserted PHI nodes here.
129 std::vector<Value*> PointerAllocaValues;
131 /// Visited - The set of basic blocks the renamer has already visited.
133 SmallPtrSet<BasicBlock*, 16> Visited;
135 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
136 /// non-determinstic behavior.
137 DenseMap<BasicBlock*, unsigned> BBNumbers;
140 PromoteMem2Reg(const std::vector<AllocaInst*> &A,
141 SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
142 DominanceFrontier &df, AliasSetTracker *ast)
143 : Allocas(A), RetryList(Retry), DT(dt), DF(df), AST(ast) {}
147 /// properlyDominates - Return true if I1 properly dominates I2.
149 bool properlyDominates(Instruction *I1, Instruction *I2) const {
150 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
151 I1 = II->getNormalDest()->begin();
152 return DT.properlyDominates(I1->getParent(), I2->getParent());
155 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
157 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
158 return DT.dominates(BB1, BB2);
162 void RemoveFromAllocasList(unsigned &AllocaIdx) {
163 Allocas[AllocaIdx] = Allocas.back();
168 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
169 SmallPtrSet<PHINode*, 16> &DeadPHINodes);
170 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
171 void PromoteLocallyUsedAllocas(BasicBlock *BB,
172 const std::vector<AllocaInst*> &AIs);
174 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
175 RenamePassData::ValVector &IncVals,
176 std::vector<RenamePassData> &Worklist);
177 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
178 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
182 std::vector<BasicBlock*> DefiningBlocks;
183 std::vector<BasicBlock*> UsingBlocks;
185 StoreInst *OnlyStore;
186 BasicBlock *OnlyBlock;
187 bool OnlyUsedInOneBlock;
189 Value *AllocaPointerVal;
192 DefiningBlocks.clear();
196 OnlyUsedInOneBlock = true;
197 AllocaPointerVal = 0;
200 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
202 void AnalyzeAlloca(AllocaInst *AI) {
205 // As we scan the uses of the alloca instruction, keep track of stores,
206 // and decide whether all of the loads and stores to the alloca are within
207 // the same basic block.
208 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end();
210 Instruction *User = cast<Instruction>(*U);
211 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
212 // Remember the basic blocks which define new values for the alloca
213 DefiningBlocks.push_back(SI->getParent());
214 AllocaPointerVal = SI->getOperand(0);
217 LoadInst *LI = cast<LoadInst>(User);
218 // Otherwise it must be a load instruction, keep track of variable reads
219 UsingBlocks.push_back(LI->getParent());
220 AllocaPointerVal = LI;
223 if (OnlyUsedInOneBlock) {
225 OnlyBlock = User->getParent();
226 else if (OnlyBlock != User->getParent())
227 OnlyUsedInOneBlock = false;
233 } // end of anonymous namespace
235 void PromoteMem2Reg::run() {
236 Function &F = *DF.getRoot()->getParent();
238 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
239 // only used in a single basic block. These instructions can be efficiently
240 // promoted by performing a single linear scan over that one block. Since
241 // individual basic blocks are sometimes large, we group together all allocas
242 // that are live in a single basic block by the basic block they are live in.
243 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
245 if (AST) PointerAllocaValues.resize(Allocas.size());
249 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
250 AllocaInst *AI = Allocas[AllocaNum];
252 assert(isAllocaPromotable(AI) &&
253 "Cannot promote non-promotable alloca!");
254 assert(AI->getParent()->getParent() == &F &&
255 "All allocas should be in the same function, which is same as DF!");
257 if (AI->use_empty()) {
258 // If there are no uses of the alloca, just delete it now.
259 if (AST) AST->deleteValue(AI);
260 AI->eraseFromParent();
262 // Remove the alloca from the Allocas list, since it has been processed
263 RemoveFromAllocasList(AllocaNum);
268 // Calculate the set of read and write-locations for each alloca. This is
269 // analogous to finding the 'uses' and 'definitions' of each variable.
270 Info.AnalyzeAlloca(AI);
272 // If the alloca is only read and written in one basic block, just perform a
273 // linear sweep over the block to eliminate it.
274 if (Info.OnlyUsedInOneBlock) {
275 LocallyUsedAllocas[Info.OnlyBlock].push_back(AI);
277 // Remove the alloca from the Allocas list, since it will be processed.
278 RemoveFromAllocasList(AllocaNum);
282 // If there is only a single store to this value, replace any loads of
283 // it that are directly dominated by the definition with the value stored.
284 if (Info.DefiningBlocks.size() == 1) {
285 // Be aware of loads before the store.
286 std::set<BasicBlock*> ProcessedBlocks;
287 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i)
288 // If the store dominates the block and if we haven't processed it yet,
290 if (dominates(Info.OnlyStore->getParent(), Info.UsingBlocks[i]))
291 if (ProcessedBlocks.insert(Info.UsingBlocks[i]).second) {
292 BasicBlock *UseBlock = Info.UsingBlocks[i];
294 // If the use and store are in the same block, do a quick scan to
295 // verify that there are no uses before the store.
296 if (UseBlock == Info.OnlyStore->getParent()) {
297 BasicBlock::iterator I = UseBlock->begin();
298 for (; &*I != Info.OnlyStore; ++I) { // scan block for store.
299 if (isa<LoadInst>(I) && I->getOperand(0) == AI)
302 if (&*I != Info.OnlyStore) break; // Do not handle this case.
305 // Otherwise, if this is a different block or if all uses happen
306 // after the store, do a simple linear scan to replace loads with
308 for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end();
310 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
311 if (LI->getOperand(0) == AI) {
312 LI->replaceAllUsesWith(Info.OnlyStore->getOperand(0));
313 if (AST && isa<PointerType>(LI->getType()))
314 AST->deleteValue(LI);
315 LI->eraseFromParent();
320 // Finally, remove this block from the UsingBlock set.
321 Info.UsingBlocks[i] = Info.UsingBlocks.back();
325 // Finally, after the scan, check to see if the store is all that is left.
326 if (Info.UsingBlocks.empty()) {
328 // The alloca has been processed, move on.
329 RemoveFromAllocasList(AllocaNum);
336 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
338 // If we haven't computed a numbering for the BB's in the function, do so
340 if (BBNumbers.empty()) {
342 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
346 // Compute the locations where PhiNodes need to be inserted. Look at the
347 // dominance frontier of EACH basic-block we have a write in.
349 unsigned CurrentVersion = 0;
350 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
351 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
352 while (!Info.DefiningBlocks.empty()) {
353 BasicBlock *BB = Info.DefiningBlocks.back();
354 Info.DefiningBlocks.pop_back();
356 // Look up the DF for this write, add it to PhiNodes
357 DominanceFrontier::const_iterator it = DF.find(BB);
358 if (it != DF.end()) {
359 const DominanceFrontier::DomSetType &S = it->second;
361 // In theory we don't need the indirection through the DFBlocks vector.
362 // In practice, the order of calling QueuePhiNode would depend on the
363 // (unspecified) ordering of basic blocks in the dominance frontier,
364 // which would give PHI nodes non-determinstic subscripts. Fix this by
365 // processing blocks in order of the occurance in the function.
366 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
367 PE = S.end(); P != PE; ++P)
368 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
370 // Sort by which the block ordering in the function.
371 std::sort(DFBlocks.begin(), DFBlocks.end());
373 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
374 BasicBlock *BB = DFBlocks[i].second;
375 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
376 Info.DefiningBlocks.push_back(BB);
382 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
383 // of the writes to the variable, scan through the reads of the variable,
384 // marking PHI nodes which are actually necessary as alive (by removing them
385 // from the InsertedPHINodes set). This is not perfect: there may PHI
386 // marked alive because of loads which are dominated by stores, but there
387 // will be no unmarked PHI nodes which are actually used.
389 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i)
390 MarkDominatingPHILive(Info.UsingBlocks[i], AllocaNum, InsertedPHINodes);
391 Info.UsingBlocks.clear();
393 // If there are any PHI nodes which are now known to be dead, remove them!
394 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(),
395 E = InsertedPHINodes.end(); I != E; ++I) {
397 bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum));
399 assert(Erased && "PHI already removed?");
401 if (AST && isa<PointerType>(PN->getType()))
402 AST->deleteValue(PN);
403 PN->eraseFromParent();
404 PhiToAllocaMap.erase(PN);
407 // Keep the reverse mapping of the 'Allocas' array.
408 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
411 // Process all allocas which are only used in a single basic block.
412 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
413 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
414 const std::vector<AllocaInst*> &LocAllocas = I->second;
415 assert(!LocAllocas.empty() && "empty alloca list??");
417 // It's common for there to only be one alloca in the list. Handle it
419 if (LocAllocas.size() == 1) {
420 // If we can do the quick promotion pass, do so now.
421 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
422 RetryList.push_back(LocAllocas[0]); // Failed, retry later.
424 // Locally promote anything possible. Note that if this is unable to
425 // promote a particular alloca, it puts the alloca onto the Allocas vector
426 // for global processing.
427 PromoteLocallyUsedAllocas(I->first, LocAllocas);
432 return; // All of the allocas must have been trivial!
434 // Set the incoming values for the basic block to be null values for all of
435 // the alloca's. We do this in case there is a load of a value that has not
436 // been stored yet. In this case, it will get this null value.
438 RenamePassData::ValVector Values(Allocas.size());
439 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
440 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
442 // Walks all basic blocks in the function performing the SSA rename algorithm
443 // and inserting the phi nodes we marked as necessary
445 std::vector<RenamePassData> RenamePassWorkList;
446 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
447 while (!RenamePassWorkList.empty()) {
449 RPD.swap(RenamePassWorkList.back());
450 RenamePassWorkList.pop_back();
451 // RenamePass may add new worklist entries.
452 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
455 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
458 // Remove the allocas themselves from the function.
459 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
460 Instruction *A = Allocas[i];
462 // If there are any uses of the alloca instructions left, they must be in
463 // sections of dead code that were not processed on the dominance frontier.
464 // Just delete the users now.
467 A->replaceAllUsesWith(UndefValue::get(A->getType()));
468 if (AST) AST->deleteValue(A);
469 A->eraseFromParent();
473 // Loop over all of the PHI nodes and see if there are any that we can get
474 // rid of because they merge all of the same incoming values. This can
475 // happen due to undef values coming into the PHI nodes. This process is
476 // iterative, because eliminating one PHI node can cause others to be removed.
477 bool EliminatedAPHI = true;
478 while (EliminatedAPHI) {
479 EliminatedAPHI = false;
481 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
482 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
483 PHINode *PN = I->second;
485 // If this PHI node merges one value and/or undefs, get the value.
486 if (Value *V = PN->hasConstantValue(true)) {
487 if (!isa<Instruction>(V) ||
488 properlyDominates(cast<Instruction>(V), PN)) {
489 if (AST && isa<PointerType>(PN->getType()))
490 AST->deleteValue(PN);
491 PN->replaceAllUsesWith(V);
492 PN->eraseFromParent();
493 NewPhiNodes.erase(I++);
494 EliminatedAPHI = true;
502 // At this point, the renamer has added entries to PHI nodes for all reachable
503 // code. Unfortunately, there may be unreachable blocks which the renamer
504 // hasn't traversed. If this is the case, the PHI nodes may not
505 // have incoming values for all predecessors. Loop over all PHI nodes we have
506 // created, inserting undef values if they are missing any incoming values.
508 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
509 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
510 // We want to do this once per basic block. As such, only process a block
511 // when we find the PHI that is the first entry in the block.
512 PHINode *SomePHI = I->second;
513 BasicBlock *BB = SomePHI->getParent();
514 if (&BB->front() != SomePHI)
517 // Count the number of preds for BB.
518 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
520 // Only do work here if there the PHI nodes are missing incoming values. We
521 // know that all PHI nodes that were inserted in a block will have the same
522 // number of incoming values, so we can just check any of them.
523 if (SomePHI->getNumIncomingValues() == Preds.size())
526 // Ok, now we know that all of the PHI nodes are missing entries for some
527 // basic blocks. Start by sorting the incoming predecessors for efficient
529 std::sort(Preds.begin(), Preds.end());
531 // Now we loop through all BB's which have entries in SomePHI and remove
532 // them from the Preds list.
533 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
534 // Do a log(n) search of the Preds list for the entry we want.
535 SmallVector<BasicBlock*, 16>::iterator EntIt =
536 std::lower_bound(Preds.begin(), Preds.end(),
537 SomePHI->getIncomingBlock(i));
538 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
539 "PHI node has entry for a block which is not a predecessor!");
545 // At this point, the blocks left in the preds list must have dummy
546 // entries inserted into every PHI nodes for the block. Update all the phi
547 // nodes in this block that we are inserting (there could be phis before
549 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
550 BasicBlock::iterator BBI = BB->begin();
551 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
552 SomePHI->getNumIncomingValues() == NumBadPreds) {
553 Value *UndefVal = UndefValue::get(SomePHI->getType());
554 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
555 SomePHI->addIncoming(UndefVal, Preds[pred]);
562 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
563 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
564 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
565 // each read of the variable. For each block that reads the variable, this
566 // function is called, which removes used PHI nodes from the DeadPHINodes set.
567 // After all of the reads have been processed, any PHI nodes left in the
568 // DeadPHINodes set are removed.
570 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
571 SmallPtrSet<PHINode*, 16> &DeadPHINodes) {
572 // Scan the immediate dominators of this block looking for a block which has a
573 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
574 DomTreeNode *IDomNode = DT.getNode(BB);
575 for (DomTreeNode *IDom = IDomNode; IDom; IDom = IDom->getIDom()) {
576 BasicBlock *DomBB = IDom->getBlock();
577 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator
578 I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum));
579 if (I != NewPhiNodes.end()) {
580 // Ok, we found an inserted PHI node which dominates this value.
581 PHINode *DominatingPHI = I->second;
583 // Find out if we previously thought it was dead. If so, mark it as being
584 // live by removing it from the DeadPHINodes set.
585 if (DeadPHINodes.erase(DominatingPHI)) {
586 // Now that we have marked the PHI node alive, also mark any PHI nodes
587 // which it might use as being alive as well.
588 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
590 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
596 /// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
597 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
598 /// potentially useless PHI nodes by just performing a single linear pass over
599 /// the basic block using the Alloca.
601 /// If we cannot promote this alloca (because it is read before it is written),
602 /// return true. This is necessary in cases where, due to control flow, the
603 /// alloca is potentially undefined on some control flow paths. e.g. code like
604 /// this is potentially correct:
606 /// for (...) { if (c) { A = undef; undef = B; } }
608 /// ... so long as A is not used before undef is set.
610 bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
611 assert(!AI->use_empty() && "There are no uses of the alloca!");
613 // Handle degenerate cases quickly.
614 if (AI->hasOneUse()) {
615 Instruction *U = cast<Instruction>(AI->use_back());
616 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
617 // Must be a load of uninitialized value.
618 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
619 if (AST && isa<PointerType>(LI->getType()))
620 AST->deleteValue(LI);
622 // Otherwise it must be a store which is never read.
623 assert(isa<StoreInst>(U));
625 BB->getInstList().erase(U);
627 // Uses of the uninitialized memory location shall get undef.
630 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
631 Instruction *Inst = I++;
632 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
633 if (LI->getOperand(0) == AI) {
634 if (!CurVal) return true; // Could not locally promote!
636 // Loads just returns the "current value"...
637 LI->replaceAllUsesWith(CurVal);
638 if (AST && isa<PointerType>(LI->getType()))
639 AST->deleteValue(LI);
640 BB->getInstList().erase(LI);
642 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
643 if (SI->getOperand(1) == AI) {
644 // Store updates the "current value"...
645 CurVal = SI->getOperand(0);
646 BB->getInstList().erase(SI);
652 // After traversing the basic block, there should be no more uses of the
653 // alloca, remove it now.
654 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
655 if (AST) AST->deleteValue(AI);
656 AI->getParent()->getInstList().erase(AI);
662 /// PromoteLocallyUsedAllocas - This method is just like
663 /// PromoteLocallyUsedAlloca, except that it processes multiple alloca
664 /// instructions in parallel. This is important in cases where we have large
665 /// basic blocks, as we don't want to rescan the entire basic block for each
666 /// alloca which is locally used in it (which might be a lot).
667 void PromoteMem2Reg::
668 PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
669 std::map<AllocaInst*, Value*> CurValues;
670 for (unsigned i = 0, e = AIs.size(); i != e; ++i)
671 CurValues[AIs[i]] = 0; // Insert with null value
673 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
674 Instruction *Inst = I++;
675 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
676 // Is this a load of an alloca we are tracking?
677 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
678 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
679 if (AIt != CurValues.end()) {
680 // If loading an uninitialized value, allow the inter-block case to
681 // handle it. Due to control flow, this might actually be ok.
682 if (AIt->second == 0) { // Use of locally uninitialized value??
683 RetryList.push_back(AI); // Retry elsewhere.
684 CurValues.erase(AIt); // Stop tracking this here.
685 if (CurValues.empty()) return;
687 // Loads just returns the "current value"...
688 LI->replaceAllUsesWith(AIt->second);
689 if (AST && isa<PointerType>(LI->getType()))
690 AST->deleteValue(LI);
691 BB->getInstList().erase(LI);
695 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
696 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
697 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
698 if (AIt != CurValues.end()) {
699 // Store updates the "current value"...
700 AIt->second = SI->getOperand(0);
701 BB->getInstList().erase(SI);
710 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
711 // Alloca returns true if there wasn't already a phi-node for that variable
713 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
715 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
716 // Look up the basic-block in question.
717 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
719 // If the BB already has a phi node added for the i'th alloca then we're done!
720 if (PN) return false;
722 // Create a PhiNode using the dereferenced type... and add the phi-node to the
724 PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
725 Allocas[AllocaNo]->getName() + "." +
726 utostr(Version++), BB->begin());
727 PhiToAllocaMap[PN] = AllocaNo;
729 InsertedPHINodes.insert(PN);
731 if (AST && isa<PointerType>(PN->getType()))
732 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
738 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
739 // stores to the allocas which we are promoting. IncomingVals indicates what
740 // value each Alloca contains on exit from the predecessor block Pred.
742 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
743 RenamePassData::ValVector &IncomingVals,
744 std::vector<RenamePassData> &Worklist) {
745 // If we are inserting any phi nodes into this BB, they will already be in the
747 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
748 // Pred may have multiple edges to BB. If so, we want to add N incoming
749 // values to each PHI we are inserting on the first time we see the edge.
750 // Check to see if APN already has incoming values from Pred. This also
751 // prevents us from modifying PHI nodes that are not currently being
753 bool HasPredEntries = false;
754 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
755 if (APN->getIncomingBlock(i) == Pred) {
756 HasPredEntries = true;
761 // If we have PHI nodes to update, compute the number of edges from Pred to
763 if (!HasPredEntries) {
764 TerminatorInst *PredTerm = Pred->getTerminator();
765 unsigned NumEdges = 0;
766 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
767 if (PredTerm->getSuccessor(i) == BB)
770 assert(NumEdges && "Must be at least one edge from Pred to BB!");
772 // Add entries for all the phis.
773 BasicBlock::iterator PNI = BB->begin();
775 unsigned AllocaNo = PhiToAllocaMap[APN];
777 // Add N incoming values to the PHI node.
778 for (unsigned i = 0; i != NumEdges; ++i)
779 APN->addIncoming(IncomingVals[AllocaNo], Pred);
781 // The currently active variable for this block is now the PHI.
782 IncomingVals[AllocaNo] = APN;
784 // Get the next phi node.
786 APN = dyn_cast<PHINode>(PNI);
789 // Verify it doesn't already have entries for Pred. If it does, it is
790 // not being inserted by this mem2reg invocation.
791 HasPredEntries = false;
792 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
793 if (APN->getIncomingBlock(i) == Pred) {
794 HasPredEntries = true;
798 } while (!HasPredEntries);
802 // Don't revisit blocks.
803 if (!Visited.insert(BB)) return;
805 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
806 Instruction *I = II++; // get the instruction, increment iterator
808 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
809 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
810 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
811 if (AI != AllocaLookup.end()) {
812 Value *V = IncomingVals[AI->second];
814 // walk the use list of this load and replace all uses with r
815 LI->replaceAllUsesWith(V);
816 if (AST && isa<PointerType>(LI->getType()))
817 AST->deleteValue(LI);
818 BB->getInstList().erase(LI);
821 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
822 // Delete this instruction and mark the name as the current holder of the
824 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
825 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
826 if (ai != AllocaLookup.end()) {
827 // what value were we writing?
828 IncomingVals[ai->second] = SI->getOperand(0);
829 BB->getInstList().erase(SI);
835 // Recurse to our successors.
836 TerminatorInst *TI = BB->getTerminator();
837 for (unsigned i = 0; i != TI->getNumSuccessors(); i++)
838 Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
841 /// PromoteMemToReg - Promote the specified list of alloca instructions into
842 /// scalar registers, inserting PHI nodes as appropriate. This function makes
843 /// use of DominanceFrontier information. This function does not modify the CFG
844 /// of the function at all. All allocas must be from the same function.
846 /// If AST is specified, the specified tracker is updated to reflect changes
849 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
850 DominatorTree &DT, DominanceFrontier &DF,
851 AliasSetTracker *AST) {
852 // If there is nothing to do, bail out...
853 if (Allocas.empty()) return;
855 SmallVector<AllocaInst*, 16> RetryList;
856 PromoteMem2Reg(Allocas, RetryList, DT, DF, AST).run();
858 // PromoteMem2Reg may not have been able to promote all of the allocas in one
859 // pass, run it again if needed.
860 std::vector<AllocaInst*> NewAllocas;
861 while (!RetryList.empty()) {
862 // If we need to retry some allocas, this is due to there being no store
863 // before a read in a local block. To counteract this, insert a store of
864 // undef into the alloca right after the alloca itself.
865 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
866 BasicBlock::iterator BBI = RetryList[i];
868 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
869 RetryList[i], ++BBI);
872 NewAllocas.assign(RetryList.begin(), RetryList.end());
874 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, AST).run();