1 //===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file promotes 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");
40 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
42 // Provide DenseMapInfo for all pointers.
45 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
46 typedef std::pair<BasicBlock*, unsigned> EltTy;
47 static inline EltTy getEmptyKey() {
48 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
50 static inline EltTy getTombstoneKey() {
51 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
53 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
54 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
56 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
59 static bool isPod() { return true; }
63 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
64 /// This is true if there are only loads and stores to the alloca.
66 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
67 // FIXME: If the memory unit is of pointer or integer type, we can permit
68 // assignments to subsections of the memory unit.
70 // Only allow direct and non-volatile loads and stores...
71 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
72 UI != UE; ++UI) // Loop over all of the uses of the alloca
73 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
76 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
77 if (SI->getOperand(0) == AI)
78 return false; // Don't allow a store OF the AI, only INTO the AI.
82 return false; // Not a load or store.
91 // Data package used by RenamePass()
92 class VISIBILITY_HIDDEN RenamePassData {
94 typedef std::vector<Value *> ValVector;
97 RenamePassData(BasicBlock *B, BasicBlock *P,
98 const ValVector &V) : BB(B), Pred(P), Values(V) {}
103 void swap(RenamePassData &RHS) {
104 std::swap(BB, RHS.BB);
105 std::swap(Pred, RHS.Pred);
106 Values.swap(RHS.Values);
110 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
111 /// load/store instructions in the block that directly load or store an alloca.
113 /// This functionality is important because it avoids scanning large basic
114 /// blocks multiple times when promoting many allocas in the same block.
115 class VISIBILITY_HIDDEN LargeBlockInfo {
116 /// InstNumbers - For each instruction that we track, keep the index of the
117 /// instruction. The index starts out as the number of the instruction from
118 /// the start of the block.
119 DenseMap<const Instruction *, unsigned> InstNumbers;
122 /// isInterestingInstruction - This code only looks at accesses to allocas.
123 static bool isInterestingInstruction(const Instruction *I) {
124 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
125 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
128 /// getInstructionIndex - Get or calculate the index of the specified
130 unsigned getInstructionIndex(const Instruction *I) {
131 assert(isInterestingInstruction(I) &&
132 "Not a load/store to/from an alloca?");
134 // If we already have this instruction number, return it.
135 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
136 if (It != InstNumbers.end()) return It->second;
138 // Scan the whole block to get the instruction. This accumulates
139 // information for every interesting instruction in the block, in order to
140 // avoid gratuitus rescans.
141 const BasicBlock *BB = I->getParent();
143 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
145 if (isInterestingInstruction(BBI))
146 InstNumbers[BBI] = InstNo++;
147 It = InstNumbers.find(I);
149 assert(It != InstNumbers.end() && "Didn't insert instruction?");
153 void deleteValue(const Instruction *I) {
154 InstNumbers.erase(I);
162 struct VISIBILITY_HIDDEN PromoteMem2Reg {
163 /// Allocas - The alloca instructions being promoted.
165 std::vector<AllocaInst*> Allocas;
167 DominanceFrontier &DF;
169 /// AST - An AliasSetTracker object to update. If null, don't update it.
171 AliasSetTracker *AST;
173 /// AllocaLookup - Reverse mapping of Allocas.
175 std::map<AllocaInst*, unsigned> AllocaLookup;
177 /// NewPhiNodes - The PhiNodes we're adding.
179 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
181 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
182 /// it corresponds to.
183 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
185 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
186 /// each alloca that is of pointer type, we keep track of what to copyValue
187 /// to the inserted PHI nodes here.
189 std::vector<Value*> PointerAllocaValues;
191 /// Visited - The set of basic blocks the renamer has already visited.
193 SmallPtrSet<BasicBlock*, 16> Visited;
195 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
196 /// non-determinstic behavior.
197 DenseMap<BasicBlock*, unsigned> BBNumbers;
199 /// BBNumPreds - Lazily compute the number of predecessors a block has.
200 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
202 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
203 DominanceFrontier &df, AliasSetTracker *ast)
204 : Allocas(A), DT(dt), DF(df), AST(ast) {}
208 /// properlyDominates - Return true if I1 properly dominates I2.
210 bool properlyDominates(Instruction *I1, Instruction *I2) const {
211 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
212 I1 = II->getNormalDest()->begin();
213 return DT.properlyDominates(I1->getParent(), I2->getParent());
216 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
218 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
219 return DT.dominates(BB1, BB2);
223 void RemoveFromAllocasList(unsigned &AllocaIdx) {
224 Allocas[AllocaIdx] = Allocas.back();
229 unsigned getNumPreds(const BasicBlock *BB) {
230 unsigned &NP = BBNumPreds[BB];
232 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
236 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
238 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
239 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
240 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
242 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
243 LargeBlockInfo &LBI);
244 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
245 LargeBlockInfo &LBI);
248 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
249 RenamePassData::ValVector &IncVals,
250 std::vector<RenamePassData> &Worklist);
251 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
252 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
256 std::vector<BasicBlock*> DefiningBlocks;
257 std::vector<BasicBlock*> UsingBlocks;
259 StoreInst *OnlyStore;
260 BasicBlock *OnlyBlock;
261 bool OnlyUsedInOneBlock;
263 Value *AllocaPointerVal;
266 DefiningBlocks.clear();
270 OnlyUsedInOneBlock = true;
271 AllocaPointerVal = 0;
274 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
276 void AnalyzeAlloca(AllocaInst *AI) {
279 // As we scan the uses of the alloca instruction, keep track of stores,
280 // and decide whether all of the loads and stores to the alloca are within
281 // the same basic block.
282 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end();
284 Instruction *User = cast<Instruction>(*U);
285 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
286 // Remember the basic blocks which define new values for the alloca
287 DefiningBlocks.push_back(SI->getParent());
288 AllocaPointerVal = SI->getOperand(0);
291 LoadInst *LI = cast<LoadInst>(User);
292 // Otherwise it must be a load instruction, keep track of variable
294 UsingBlocks.push_back(LI->getParent());
295 AllocaPointerVal = LI;
298 if (OnlyUsedInOneBlock) {
300 OnlyBlock = User->getParent();
301 else if (OnlyBlock != User->getParent())
302 OnlyUsedInOneBlock = false;
307 } // end of anonymous namespace
310 void PromoteMem2Reg::run() {
311 Function &F = *DF.getRoot()->getParent();
313 if (AST) PointerAllocaValues.resize(Allocas.size());
318 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
319 AllocaInst *AI = Allocas[AllocaNum];
321 assert(isAllocaPromotable(AI) &&
322 "Cannot promote non-promotable alloca!");
323 assert(AI->getParent()->getParent() == &F &&
324 "All allocas should be in the same function, which is same as DF!");
326 if (AI->use_empty()) {
327 // If there are no uses of the alloca, just delete it now.
328 if (AST) AST->deleteValue(AI);
329 AI->eraseFromParent();
331 // Remove the alloca from the Allocas list, since it has been processed
332 RemoveFromAllocasList(AllocaNum);
337 // Calculate the set of read and write-locations for each alloca. This is
338 // analogous to finding the 'uses' and 'definitions' of each variable.
339 Info.AnalyzeAlloca(AI);
341 // If there is only a single store to this value, replace any loads of
342 // it that are directly dominated by the definition with the value stored.
343 if (Info.DefiningBlocks.size() == 1) {
344 RewriteSingleStoreAlloca(AI, Info, LBI);
346 // Finally, after the scan, check to see if the store is all that is left.
347 if (Info.UsingBlocks.empty()) {
348 // Remove the (now dead) store and alloca.
349 Info.OnlyStore->eraseFromParent();
350 LBI.deleteValue(Info.OnlyStore);
352 if (AST) AST->deleteValue(AI);
353 AI->eraseFromParent();
356 // The alloca has been processed, move on.
357 RemoveFromAllocasList(AllocaNum);
364 // If the alloca is only read and written in one basic block, just perform a
365 // linear sweep over the block to eliminate it.
366 if (Info.OnlyUsedInOneBlock) {
367 PromoteSingleBlockAlloca(AI, Info, LBI);
369 // Finally, after the scan, check to see if the stores are all that is
371 if (Info.UsingBlocks.empty()) {
373 // Remove the (now dead) stores and alloca.
374 while (!AI->use_empty()) {
375 StoreInst *SI = cast<StoreInst>(AI->use_back());
376 SI->eraseFromParent();
380 if (AST) AST->deleteValue(AI);
381 AI->eraseFromParent();
384 // The alloca has been processed, move on.
385 RemoveFromAllocasList(AllocaNum);
392 // If we haven't computed a numbering for the BB's in the function, do so
394 if (BBNumbers.empty()) {
396 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
400 // If we have an AST to keep updated, remember some pointer value that is
401 // stored into the alloca.
403 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
405 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
406 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
408 // At this point, we're committed to promoting the alloca using IDF's, and
409 // the standard SSA construction algorithm. Determine which blocks need PHI
410 // nodes and see if we can optimize out some work by avoiding insertion of
412 DetermineInsertionPoint(AI, AllocaNum, Info);
416 return; // All of the allocas must have been trivial!
421 // Set the incoming values for the basic block to be null values for all of
422 // the alloca's. We do this in case there is a load of a value that has not
423 // been stored yet. In this case, it will get this null value.
425 RenamePassData::ValVector Values(Allocas.size());
426 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
427 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
429 // Walks all basic blocks in the function performing the SSA rename algorithm
430 // and inserting the phi nodes we marked as necessary
432 std::vector<RenamePassData> RenamePassWorkList;
433 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
434 while (!RenamePassWorkList.empty()) {
436 RPD.swap(RenamePassWorkList.back());
437 RenamePassWorkList.pop_back();
438 // RenamePass may add new worklist entries.
439 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
442 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
445 // Remove the allocas themselves from the function.
446 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
447 Instruction *A = Allocas[i];
449 // If there are any uses of the alloca instructions left, they must be in
450 // sections of dead code that were not processed on the dominance frontier.
451 // Just delete the users now.
454 A->replaceAllUsesWith(UndefValue::get(A->getType()));
455 if (AST) AST->deleteValue(A);
456 A->eraseFromParent();
460 // Loop over all of the PHI nodes and see if there are any that we can get
461 // rid of because they merge all of the same incoming values. This can
462 // happen due to undef values coming into the PHI nodes. This process is
463 // iterative, because eliminating one PHI node can cause others to be removed.
464 bool EliminatedAPHI = true;
465 while (EliminatedAPHI) {
466 EliminatedAPHI = false;
468 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
469 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
470 PHINode *PN = I->second;
472 // If this PHI node merges one value and/or undefs, get the value.
473 if (Value *V = PN->hasConstantValue(true)) {
474 if (!isa<Instruction>(V) ||
475 properlyDominates(cast<Instruction>(V), PN)) {
476 if (AST && isa<PointerType>(PN->getType()))
477 AST->deleteValue(PN);
478 PN->replaceAllUsesWith(V);
479 PN->eraseFromParent();
480 NewPhiNodes.erase(I++);
481 EliminatedAPHI = true;
489 // At this point, the renamer has added entries to PHI nodes for all reachable
490 // code. Unfortunately, there may be unreachable blocks which the renamer
491 // hasn't traversed. If this is the case, the PHI nodes may not
492 // have incoming values for all predecessors. Loop over all PHI nodes we have
493 // created, inserting undef values if they are missing any incoming values.
495 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
496 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
497 // We want to do this once per basic block. As such, only process a block
498 // when we find the PHI that is the first entry in the block.
499 PHINode *SomePHI = I->second;
500 BasicBlock *BB = SomePHI->getParent();
501 if (&BB->front() != SomePHI)
504 // Only do work here if there the PHI nodes are missing incoming values. We
505 // know that all PHI nodes that were inserted in a block will have the same
506 // number of incoming values, so we can just check any of them.
507 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
510 // Get the preds for BB.
511 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
513 // Ok, now we know that all of the PHI nodes are missing entries for some
514 // basic blocks. Start by sorting the incoming predecessors for efficient
516 std::sort(Preds.begin(), Preds.end());
518 // Now we loop through all BB's which have entries in SomePHI and remove
519 // them from the Preds list.
520 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
521 // Do a log(n) search of the Preds list for the entry we want.
522 SmallVector<BasicBlock*, 16>::iterator EntIt =
523 std::lower_bound(Preds.begin(), Preds.end(),
524 SomePHI->getIncomingBlock(i));
525 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
526 "PHI node has entry for a block which is not a predecessor!");
532 // At this point, the blocks left in the preds list must have dummy
533 // entries inserted into every PHI nodes for the block. Update all the phi
534 // nodes in this block that we are inserting (there could be phis before
536 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
537 BasicBlock::iterator BBI = BB->begin();
538 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
539 SomePHI->getNumIncomingValues() == NumBadPreds) {
540 Value *UndefVal = UndefValue::get(SomePHI->getType());
541 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
542 SomePHI->addIncoming(UndefVal, Preds[pred]);
550 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
551 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
552 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
554 void PromoteMem2Reg::
555 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
556 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
557 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
559 // To determine liveness, we must iterate through the predecessors of blocks
560 // where the def is live. Blocks are added to the worklist if we need to
561 // check their predecessors. Start with all the using blocks.
562 SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
563 LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
564 Info.UsingBlocks.begin(), Info.UsingBlocks.end());
566 // If any of the using blocks is also a definition block, check to see if the
567 // definition occurs before or after the use. If it happens before the use,
568 // the value isn't really live-in.
569 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
570 BasicBlock *BB = LiveInBlockWorklist[i];
571 if (!DefBlocks.count(BB)) continue;
573 // Okay, this is a block that both uses and defines the value. If the first
574 // reference to the alloca is a def (store), then we know it isn't live-in.
575 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
576 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
577 if (SI->getOperand(1) != AI) continue;
579 // We found a store to the alloca before a load. The alloca is not
580 // actually live-in here.
581 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
582 LiveInBlockWorklist.pop_back();
585 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
586 if (LI->getOperand(0) != AI) continue;
588 // Okay, we found a load before a store to the alloca. It is actually
589 // live into this block.
595 // Now that we have a set of blocks where the phi is live-in, recursively add
596 // their predecessors until we find the full region the value is live.
597 while (!LiveInBlockWorklist.empty()) {
598 BasicBlock *BB = LiveInBlockWorklist.back();
599 LiveInBlockWorklist.pop_back();
601 // The block really is live in here, insert it into the set. If already in
602 // the set, then it has already been processed.
603 if (!LiveInBlocks.insert(BB))
606 // Since the value is live into BB, it is either defined in a predecessor or
607 // live into it to. Add the preds to the worklist unless they are a
609 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
612 // The value is not live into a predecessor if it defines the value.
613 if (DefBlocks.count(P))
616 // Otherwise it is, add to the worklist.
617 LiveInBlockWorklist.push_back(P);
622 /// DetermineInsertionPoint - At this point, we're committed to promoting the
623 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
624 /// which blocks need phi nodes and see if we can optimize out some work by
625 /// avoiding insertion of dead phi nodes.
626 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
629 // Unique the set of defining blocks for efficient lookup.
630 SmallPtrSet<BasicBlock*, 32> DefBlocks;
631 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
633 // Determine which blocks the value is live in. These are blocks which lead
635 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
636 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
638 // Compute the locations where PhiNodes need to be inserted. Look at the
639 // dominance frontier of EACH basic-block we have a write in.
640 unsigned CurrentVersion = 0;
641 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
642 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
643 while (!Info.DefiningBlocks.empty()) {
644 BasicBlock *BB = Info.DefiningBlocks.back();
645 Info.DefiningBlocks.pop_back();
647 // Look up the DF for this write, add it to defining blocks.
648 DominanceFrontier::const_iterator it = DF.find(BB);
649 if (it == DF.end()) continue;
651 const DominanceFrontier::DomSetType &S = it->second;
653 // In theory we don't need the indirection through the DFBlocks vector.
654 // In practice, the order of calling QueuePhiNode would depend on the
655 // (unspecified) ordering of basic blocks in the dominance frontier,
656 // which would give PHI nodes non-determinstic subscripts. Fix this by
657 // processing blocks in order of the occurance in the function.
658 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
659 PE = S.end(); P != PE; ++P) {
660 // If the frontier block is not in the live-in set for the alloca, don't
661 // bother processing it.
662 if (!LiveInBlocks.count(*P))
665 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
668 // Sort by which the block ordering in the function.
669 if (DFBlocks.size() > 1)
670 std::sort(DFBlocks.begin(), DFBlocks.end());
672 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
673 BasicBlock *BB = DFBlocks[i].second;
674 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
675 Info.DefiningBlocks.push_back(BB);
681 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
682 /// replace any loads of it that are directly dominated by the definition with
683 /// the value stored.
684 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
686 LargeBlockInfo &LBI) {
687 StoreInst *OnlyStore = Info.OnlyStore;
688 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
689 BasicBlock *StoreBB = OnlyStore->getParent();
692 // Clear out UsingBlocks. We will reconstruct it here if needed.
693 Info.UsingBlocks.clear();
695 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
696 Instruction *UserInst = cast<Instruction>(*UI++);
697 if (!isa<LoadInst>(UserInst)) {
698 assert(UserInst == OnlyStore && "Should only have load/stores");
701 LoadInst *LI = cast<LoadInst>(UserInst);
703 // Okay, if we have a load from the alloca, we want to replace it with the
704 // only value stored to the alloca. We can do this if the value is
705 // dominated by the store. If not, we use the rest of the mem2reg machinery
706 // to insert the phi nodes as needed.
707 if (!StoringGlobalVal) { // Non-instructions are always dominated.
708 if (LI->getParent() == StoreBB) {
709 // If we have a use that is in the same block as the store, compare the
710 // indices of the two instructions to see which one came first. If the
711 // load came before the store, we can't handle it.
712 if (StoreIndex == -1)
713 StoreIndex = LBI.getInstructionIndex(OnlyStore);
715 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
716 // Can't handle this load, bail out.
717 Info.UsingBlocks.push_back(StoreBB);
721 } else if (LI->getParent() != StoreBB &&
722 !dominates(StoreBB, LI->getParent())) {
723 // If the load and store are in different blocks, use BB dominance to
724 // check their relationships. If the store doesn't dom the use, bail
726 Info.UsingBlocks.push_back(LI->getParent());
731 // Otherwise, we *can* safely rewrite this load.
732 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
733 if (AST && isa<PointerType>(LI->getType()))
734 AST->deleteValue(LI);
735 LI->eraseFromParent();
741 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
742 /// first element of a pair.
743 struct StoreIndexSearchPredicate {
744 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
745 const std::pair<unsigned, StoreInst*> &RHS) {
746 return LHS.first < RHS.first;
750 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
751 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
752 /// potentially useless PHI nodes by just performing a single linear pass over
753 /// the basic block using the Alloca.
755 /// If we cannot promote this alloca (because it is read before it is written),
756 /// return true. This is necessary in cases where, due to control flow, the
757 /// alloca is potentially undefined on some control flow paths. e.g. code like
758 /// this is potentially correct:
760 /// for (...) { if (c) { A = undef; undef = B; } }
762 /// ... so long as A is not used before undef is set.
764 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
765 LargeBlockInfo &LBI) {
766 // The trickiest case to handle is when we have large blocks. Because of this,
767 // this code is optimized assuming that large blocks happen. This does not
768 // significantly pessimize the small block case. This uses LargeBlockInfo to
769 // make it efficient to get the index of various operations in the block.
771 // Clear out UsingBlocks. We will reconstruct it here if needed.
772 Info.UsingBlocks.clear();
774 // Walk the use-def list of the alloca, getting the locations of all stores.
775 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
776 StoresByIndexTy StoresByIndex;
778 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
780 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
781 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
783 // If there are no stores to the alloca, just replace any loads with undef.
784 if (StoresByIndex.empty()) {
785 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
786 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
787 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
788 if (AST && isa<PointerType>(LI->getType()))
789 AST->deleteValue(LI);
791 LI->eraseFromParent();
796 // Sort the stores by their index, making it efficient to do a lookup with a
798 std::sort(StoresByIndex.begin(), StoresByIndex.end());
800 // Walk all of the loads from this alloca, replacing them with the nearest
801 // store above them, if any.
802 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
803 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
806 unsigned LoadIdx = LBI.getInstructionIndex(LI);
808 // Find the nearest store that has a lower than this load.
809 StoresByIndexTy::iterator I =
810 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
811 std::pair<unsigned, StoreInst*>(LoadIdx, 0),
812 StoreIndexSearchPredicate());
814 // If there is no store before this load, then we can't promote this load.
815 if (I == StoresByIndex.begin()) {
816 // Can't handle this load, bail out.
817 Info.UsingBlocks.push_back(LI->getParent());
821 // Otherwise, there was a store before this load, the load takes its value.
823 LI->replaceAllUsesWith(I->second->getOperand(0));
824 if (AST && isa<PointerType>(LI->getType()))
825 AST->deleteValue(LI);
826 LI->eraseFromParent();
832 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
833 // Alloca returns true if there wasn't already a phi-node for that variable
835 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
837 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
838 // Look up the basic-block in question.
839 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
841 // If the BB already has a phi node added for the i'th alloca then we're done!
842 if (PN) return false;
844 // Create a PhiNode using the dereferenced type... and add the phi-node to the
846 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
847 Allocas[AllocaNo]->getName() + "." +
848 utostr(Version++), BB->begin());
850 PhiToAllocaMap[PN] = AllocaNo;
851 PN->reserveOperandSpace(getNumPreds(BB));
853 InsertedPHINodes.insert(PN);
855 if (AST && isa<PointerType>(PN->getType()))
856 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
861 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
862 // stores to the allocas which we are promoting. IncomingVals indicates what
863 // value each Alloca contains on exit from the predecessor block Pred.
865 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
866 RenamePassData::ValVector &IncomingVals,
867 std::vector<RenamePassData> &Worklist) {
869 // If we are inserting any phi nodes into this BB, they will already be in the
871 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
872 // Pred may have multiple edges to BB. If so, we want to add N incoming
873 // values to each PHI we are inserting on the first time we see the edge.
874 // Check to see if APN already has incoming values from Pred. This also
875 // prevents us from modifying PHI nodes that are not currently being
877 bool HasPredEntries = false;
878 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
879 if (APN->getIncomingBlock(i) == Pred) {
880 HasPredEntries = true;
885 // If we have PHI nodes to update, compute the number of edges from Pred to
887 if (!HasPredEntries) {
888 // We want to be able to distinguish between PHI nodes being inserted by
889 // this invocation of mem2reg from those phi nodes that already existed in
890 // the IR before mem2reg was run. We determine that APN is being inserted
891 // because it is missing incoming edges. All other PHI nodes being
892 // inserted by this pass of mem2reg will have the same number of incoming
893 // operands so far. Remember this count.
894 unsigned NewPHINumOperands = APN->getNumOperands();
896 unsigned NumEdges = 0;
897 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
900 assert(NumEdges && "Must be at least one edge from Pred to BB!");
902 // Add entries for all the phis.
903 BasicBlock::iterator PNI = BB->begin();
905 unsigned AllocaNo = PhiToAllocaMap[APN];
907 // Add N incoming values to the PHI node.
908 for (unsigned i = 0; i != NumEdges; ++i)
909 APN->addIncoming(IncomingVals[AllocaNo], Pred);
911 // The currently active variable for this block is now the PHI.
912 IncomingVals[AllocaNo] = APN;
914 // Get the next phi node.
916 APN = dyn_cast<PHINode>(PNI);
919 // Verify that it is missing entries. If not, it is not being inserted
920 // by this mem2reg invocation so we want to ignore it.
921 } while (APN->getNumOperands() == NewPHINumOperands);
925 // Don't revisit blocks.
926 if (!Visited.insert(BB)) return;
928 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
929 Instruction *I = II++; // get the instruction, increment iterator
931 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
932 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
935 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
936 if (AI == AllocaLookup.end()) continue;
938 Value *V = IncomingVals[AI->second];
940 // Anything using the load now uses the current value.
941 LI->replaceAllUsesWith(V);
942 if (AST && isa<PointerType>(LI->getType()))
943 AST->deleteValue(LI);
944 BB->getInstList().erase(LI);
945 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
946 // Delete this instruction and mark the name as the current holder of the
948 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
951 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
952 if (ai == AllocaLookup.end())
955 // what value were we writing?
956 IncomingVals[ai->second] = SI->getOperand(0);
957 BB->getInstList().erase(SI);
961 // 'Recurse' to our successors.
962 succ_iterator I = succ_begin(BB), E = succ_end(BB);
965 // Handle the last successor without using the worklist. This allows us to
966 // handle unconditional branches directly, for example.
969 Worklist.push_back(RenamePassData(*I, BB, IncomingVals));
976 /// PromoteMemToReg - Promote the specified list of alloca instructions into
977 /// scalar registers, inserting PHI nodes as appropriate. This function makes
978 /// use of DominanceFrontier information. This function does not modify the CFG
979 /// of the function at all. All allocas must be from the same function.
981 /// If AST is specified, the specified tracker is updated to reflect changes
984 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
985 DominatorTree &DT, DominanceFrontier &DF,
986 AliasSetTracker *AST) {
987 // If there is nothing to do, bail out...
988 if (Allocas.empty()) return;
990 PromoteMem2Reg(Allocas, DT, DF, AST).run();