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/IntrinsicInst.h"
26 #include "llvm/Analysis/DebugInfo.h"
27 #include "llvm/Analysis/Dominators.h"
28 #include "llvm/Analysis/AliasSetTracker.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/Support/CFG.h"
38 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
39 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
40 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
41 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
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) {
62 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
63 /// This is true if there are only loads and stores to the alloca.
65 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
66 // FIXME: If the memory unit is of pointer or integer type, we can permit
67 // assignments to subsections of the memory unit.
69 // Only allow direct and non-volatile loads and stores...
70 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
71 UI != UE; ++UI) // Loop over all of the uses of the alloca
72 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
75 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
76 if (SI->getOperand(0) == AI)
77 return false; // Don't allow a store OF the AI, only INTO the AI.
90 // Data package used by RenamePass()
91 class RenamePassData {
93 typedef std::vector<Value *> ValVector;
95 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
96 RenamePassData(BasicBlock *B, BasicBlock *P,
97 const ValVector &V) : BB(B), Pred(P), Values(V) {}
102 void swap(RenamePassData &RHS) {
103 std::swap(BB, RHS.BB);
104 std::swap(Pred, RHS.Pred);
105 Values.swap(RHS.Values);
109 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
110 /// load/store instructions in the block that directly load or store an alloca.
112 /// This functionality is important because it avoids scanning large basic
113 /// blocks multiple times when promoting many allocas in the same block.
114 class LargeBlockInfo {
115 /// InstNumbers - For each instruction that we track, keep the index of the
116 /// instruction. The index starts out as the number of the instruction from
117 /// the start of the block.
118 DenseMap<const Instruction *, unsigned> InstNumbers;
121 /// isInterestingInstruction - This code only looks at accesses to allocas.
122 static bool isInterestingInstruction(const Instruction *I) {
123 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
124 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
127 /// getInstructionIndex - Get or calculate the index of the specified
129 unsigned getInstructionIndex(const Instruction *I) {
130 assert(isInterestingInstruction(I) &&
131 "Not a load/store to/from an alloca?");
133 // If we already have this instruction number, return it.
134 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
135 if (It != InstNumbers.end()) return It->second;
137 // Scan the whole block to get the instruction. This accumulates
138 // information for every interesting instruction in the block, in order to
139 // avoid gratuitus rescans.
140 const BasicBlock *BB = I->getParent();
142 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
144 if (isInterestingInstruction(BBI))
145 InstNumbers[BBI] = InstNo++;
146 It = InstNumbers.find(I);
148 assert(It != InstNumbers.end() && "Didn't insert instruction?");
152 void deleteValue(const Instruction *I) {
153 InstNumbers.erase(I);
161 struct PromoteMem2Reg {
162 /// Allocas - The alloca instructions being promoted.
164 std::vector<AllocaInst*> Allocas;
166 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), DIF(0), 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);
246 void ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, StoreInst* SI,
249 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
250 RenamePassData::ValVector &IncVals,
251 std::vector<RenamePassData> &Worklist);
252 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
253 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
257 std::vector<BasicBlock*> DefiningBlocks;
258 std::vector<BasicBlock*> UsingBlocks;
260 StoreInst *OnlyStore;
261 BasicBlock *OnlyBlock;
262 bool OnlyUsedInOneBlock;
264 Value *AllocaPointerVal;
267 DefiningBlocks.clear();
271 OnlyUsedInOneBlock = true;
272 AllocaPointerVal = 0;
275 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
277 void AnalyzeAlloca(AllocaInst *AI) {
280 // As we scan the uses of the alloca instruction, keep track of stores,
281 // and decide whether all of the loads and stores to the alloca are within
282 // the same basic block.
283 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
285 Instruction *User = cast<Instruction>(*UI++);
287 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
288 // Remember the basic blocks which define new values for the alloca
289 DefiningBlocks.push_back(SI->getParent());
290 AllocaPointerVal = SI->getOperand(0);
293 LoadInst *LI = cast<LoadInst>(User);
294 // Otherwise it must be a load instruction, keep track of variable
296 UsingBlocks.push_back(LI->getParent());
297 AllocaPointerVal = LI;
300 if (OnlyUsedInOneBlock) {
302 OnlyBlock = User->getParent();
303 else if (OnlyBlock != User->getParent())
304 OnlyUsedInOneBlock = false;
309 } // end of anonymous namespace
312 /// Finds the llvm.dbg.declare intrinsic corresponding to an alloca if any.
313 static DbgDeclareInst *findDbgDeclare(AllocaInst *AI) {
314 Function *F = AI->getParent()->getParent();
315 for (Function::iterator FI = F->begin(), FE = F->end(); FI != FE; ++FI)
316 for (BasicBlock::iterator BI = (*FI).begin(), BE = (*FI).end();
318 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI))
319 if (DDI->getAddress() == AI)
325 void PromoteMem2Reg::run() {
326 Function &F = *DF.getRoot()->getParent();
328 if (AST) PointerAllocaValues.resize(Allocas.size());
333 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
334 AllocaInst *AI = Allocas[AllocaNum];
336 assert(isAllocaPromotable(AI) &&
337 "Cannot promote non-promotable alloca!");
338 assert(AI->getParent()->getParent() == &F &&
339 "All allocas should be in the same function, which is same as DF!");
341 if (AI->use_empty()) {
342 // If there are no uses of the alloca, just delete it now.
343 if (AST) AST->deleteValue(AI);
344 AI->eraseFromParent();
346 // Remove the alloca from the Allocas list, since it has been processed
347 RemoveFromAllocasList(AllocaNum);
352 // Calculate the set of read and write-locations for each alloca. This is
353 // analogous to finding the 'uses' and 'definitions' of each variable.
354 Info.AnalyzeAlloca(AI);
356 // If there is only a single store to this value, replace any loads of
357 // it that are directly dominated by the definition with the value stored.
358 if (Info.DefiningBlocks.size() == 1) {
359 RewriteSingleStoreAlloca(AI, Info, LBI);
361 // Finally, after the scan, check to see if the store is all that is left.
362 if (Info.UsingBlocks.empty()) {
363 // Record debuginfo for the store before removing it.
364 ConvertDebugDeclareToDebugValue(findDbgDeclare(AI), Info.OnlyStore, 0);
365 // Remove the (now dead) store and alloca.
366 Info.OnlyStore->eraseFromParent();
367 LBI.deleteValue(Info.OnlyStore);
369 if (AST) AST->deleteValue(AI);
370 AI->eraseFromParent();
373 // The alloca has been processed, move on.
374 RemoveFromAllocasList(AllocaNum);
381 // If the alloca is only read and written in one basic block, just perform a
382 // linear sweep over the block to eliminate it.
383 if (Info.OnlyUsedInOneBlock) {
384 PromoteSingleBlockAlloca(AI, Info, LBI);
386 // Finally, after the scan, check to see if the stores are all that is
388 if (Info.UsingBlocks.empty()) {
390 // Remove the (now dead) stores and alloca.
391 DbgDeclareInst *DDI = findDbgDeclare(AI);
392 while (!AI->use_empty()) {
393 StoreInst *SI = cast<StoreInst>(AI->use_back());
394 // Record debuginfo for the store before removing it.
395 ConvertDebugDeclareToDebugValue(DDI, SI, 0);
396 SI->eraseFromParent();
400 if (AST) AST->deleteValue(AI);
401 AI->eraseFromParent();
404 // The alloca has been processed, move on.
405 RemoveFromAllocasList(AllocaNum);
412 // If we haven't computed a numbering for the BB's in the function, do so
414 if (BBNumbers.empty()) {
416 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
420 // If we have an AST to keep updated, remember some pointer value that is
421 // stored into the alloca.
423 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
425 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
426 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
428 // At this point, we're committed to promoting the alloca using IDF's, and
429 // the standard SSA construction algorithm. Determine which blocks need PHI
430 // nodes and see if we can optimize out some work by avoiding insertion of
432 DetermineInsertionPoint(AI, AllocaNum, Info);
436 return; // All of the allocas must have been trivial!
441 // Set the incoming values for the basic block to be null values for all of
442 // the alloca's. We do this in case there is a load of a value that has not
443 // been stored yet. In this case, it will get this null value.
445 RenamePassData::ValVector Values(Allocas.size());
446 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
447 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
449 // Walks all basic blocks in the function performing the SSA rename algorithm
450 // and inserting the phi nodes we marked as necessary
452 std::vector<RenamePassData> RenamePassWorkList;
453 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
456 RPD.swap(RenamePassWorkList.back());
457 RenamePassWorkList.pop_back();
458 // RenamePass may add new worklist entries.
459 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
460 } while (!RenamePassWorkList.empty());
462 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
465 // Remove the allocas themselves from the function.
466 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
467 Instruction *A = Allocas[i];
469 // If there are any uses of the alloca instructions left, they must be in
470 // sections of dead code that were not processed on the dominance frontier.
471 // Just delete the users now.
474 A->replaceAllUsesWith(UndefValue::get(A->getType()));
475 if (AST) AST->deleteValue(A);
476 A->eraseFromParent();
480 // Loop over all of the PHI nodes and see if there are any that we can get
481 // rid of because they merge all of the same incoming values. This can
482 // happen due to undef values coming into the PHI nodes. This process is
483 // iterative, because eliminating one PHI node can cause others to be removed.
484 bool EliminatedAPHI = true;
485 while (EliminatedAPHI) {
486 EliminatedAPHI = false;
488 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
489 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
490 PHINode *PN = I->second;
492 // If this PHI node merges one value and/or undefs, get the value.
493 if (Value *V = PN->hasConstantValue(&DT)) {
494 if (AST && isa<PointerType>(PN->getType()))
495 AST->deleteValue(PN);
496 PN->replaceAllUsesWith(V);
497 PN->eraseFromParent();
498 NewPhiNodes.erase(I++);
499 EliminatedAPHI = true;
506 // At this point, the renamer has added entries to PHI nodes for all reachable
507 // code. Unfortunately, there may be unreachable blocks which the renamer
508 // hasn't traversed. If this is the case, the PHI nodes may not
509 // have incoming values for all predecessors. Loop over all PHI nodes we have
510 // created, inserting undef values if they are missing any incoming values.
512 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
513 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
514 // We want to do this once per basic block. As such, only process a block
515 // when we find the PHI that is the first entry in the block.
516 PHINode *SomePHI = I->second;
517 BasicBlock *BB = SomePHI->getParent();
518 if (&BB->front() != SomePHI)
521 // Only do work here if there the PHI nodes are missing incoming values. We
522 // know that all PHI nodes that were inserted in a block will have the same
523 // number of incoming values, so we can just check any of them.
524 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
527 // Get the preds for BB.
528 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
530 // Ok, now we know that all of the PHI nodes are missing entries for some
531 // basic blocks. Start by sorting the incoming predecessors for efficient
533 std::sort(Preds.begin(), Preds.end());
535 // Now we loop through all BB's which have entries in SomePHI and remove
536 // them from the Preds list.
537 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
538 // Do a log(n) search of the Preds list for the entry we want.
539 SmallVector<BasicBlock*, 16>::iterator EntIt =
540 std::lower_bound(Preds.begin(), Preds.end(),
541 SomePHI->getIncomingBlock(i));
542 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
543 "PHI node has entry for a block which is not a predecessor!");
549 // At this point, the blocks left in the preds list must have dummy
550 // entries inserted into every PHI nodes for the block. Update all the phi
551 // nodes in this block that we are inserting (there could be phis before
553 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
554 BasicBlock::iterator BBI = BB->begin();
555 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
556 SomePHI->getNumIncomingValues() == NumBadPreds) {
557 Value *UndefVal = UndefValue::get(SomePHI->getType());
558 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
559 SomePHI->addIncoming(UndefVal, Preds[pred]);
567 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
568 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
569 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
571 void PromoteMem2Reg::
572 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
573 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
574 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
576 // To determine liveness, we must iterate through the predecessors of blocks
577 // where the def is live. Blocks are added to the worklist if we need to
578 // check their predecessors. Start with all the using blocks.
579 SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
580 LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
581 Info.UsingBlocks.begin(), Info.UsingBlocks.end());
583 // If any of the using blocks is also a definition block, check to see if the
584 // definition occurs before or after the use. If it happens before the use,
585 // the value isn't really live-in.
586 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
587 BasicBlock *BB = LiveInBlockWorklist[i];
588 if (!DefBlocks.count(BB)) continue;
590 // Okay, this is a block that both uses and defines the value. If the first
591 // reference to the alloca is a def (store), then we know it isn't live-in.
592 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
593 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
594 if (SI->getOperand(1) != AI) continue;
596 // We found a store to the alloca before a load. The alloca is not
597 // actually live-in here.
598 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
599 LiveInBlockWorklist.pop_back();
604 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
605 if (LI->getOperand(0) != AI) continue;
607 // Okay, we found a load before a store to the alloca. It is actually
608 // live into this block.
614 // Now that we have a set of blocks where the phi is live-in, recursively add
615 // their predecessors until we find the full region the value is live.
616 while (!LiveInBlockWorklist.empty()) {
617 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
619 // The block really is live in here, insert it into the set. If already in
620 // the set, then it has already been processed.
621 if (!LiveInBlocks.insert(BB))
624 // Since the value is live into BB, it is either defined in a predecessor or
625 // live into it to. Add the preds to the worklist unless they are a
627 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
630 // The value is not live into a predecessor if it defines the value.
631 if (DefBlocks.count(P))
634 // Otherwise it is, add to the worklist.
635 LiveInBlockWorklist.push_back(P);
640 /// DetermineInsertionPoint - At this point, we're committed to promoting the
641 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
642 /// which blocks need phi nodes and see if we can optimize out some work by
643 /// avoiding insertion of dead phi nodes.
644 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
647 // Unique the set of defining blocks for efficient lookup.
648 SmallPtrSet<BasicBlock*, 32> DefBlocks;
649 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
651 // Determine which blocks the value is live in. These are blocks which lead
653 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
654 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
656 // Compute the locations where PhiNodes need to be inserted. Look at the
657 // dominance frontier of EACH basic-block we have a write in.
658 unsigned CurrentVersion = 0;
659 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
660 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
661 while (!Info.DefiningBlocks.empty()) {
662 BasicBlock *BB = Info.DefiningBlocks.back();
663 Info.DefiningBlocks.pop_back();
665 // Look up the DF for this write, add it to defining blocks.
666 DominanceFrontier::const_iterator it = DF.find(BB);
667 if (it == DF.end()) continue;
669 const DominanceFrontier::DomSetType &S = it->second;
671 // In theory we don't need the indirection through the DFBlocks vector.
672 // In practice, the order of calling QueuePhiNode would depend on the
673 // (unspecified) ordering of basic blocks in the dominance frontier,
674 // which would give PHI nodes non-determinstic subscripts. Fix this by
675 // processing blocks in order of the occurance in the function.
676 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
677 PE = S.end(); P != PE; ++P) {
678 // If the frontier block is not in the live-in set for the alloca, don't
679 // bother processing it.
680 if (!LiveInBlocks.count(*P))
683 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
686 // Sort by which the block ordering in the function.
687 if (DFBlocks.size() > 1)
688 std::sort(DFBlocks.begin(), DFBlocks.end());
690 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
691 BasicBlock *BB = DFBlocks[i].second;
692 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
693 Info.DefiningBlocks.push_back(BB);
699 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
700 /// replace any loads of it that are directly dominated by the definition with
701 /// the value stored.
702 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
704 LargeBlockInfo &LBI) {
705 StoreInst *OnlyStore = Info.OnlyStore;
706 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
707 BasicBlock *StoreBB = OnlyStore->getParent();
710 // Clear out UsingBlocks. We will reconstruct it here if needed.
711 Info.UsingBlocks.clear();
713 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
714 Instruction *UserInst = cast<Instruction>(*UI++);
715 if (!isa<LoadInst>(UserInst)) {
716 assert(UserInst == OnlyStore && "Should only have load/stores");
719 LoadInst *LI = cast<LoadInst>(UserInst);
721 // Okay, if we have a load from the alloca, we want to replace it with the
722 // only value stored to the alloca. We can do this if the value is
723 // dominated by the store. If not, we use the rest of the mem2reg machinery
724 // to insert the phi nodes as needed.
725 if (!StoringGlobalVal) { // Non-instructions are always dominated.
726 if (LI->getParent() == StoreBB) {
727 // If we have a use that is in the same block as the store, compare the
728 // indices of the two instructions to see which one came first. If the
729 // load came before the store, we can't handle it.
730 if (StoreIndex == -1)
731 StoreIndex = LBI.getInstructionIndex(OnlyStore);
733 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
734 // Can't handle this load, bail out.
735 Info.UsingBlocks.push_back(StoreBB);
739 } else if (LI->getParent() != StoreBB &&
740 !dominates(StoreBB, LI->getParent())) {
741 // If the load and store are in different blocks, use BB dominance to
742 // check their relationships. If the store doesn't dom the use, bail
744 Info.UsingBlocks.push_back(LI->getParent());
749 // Otherwise, we *can* safely rewrite this load.
750 Value *ReplVal = OnlyStore->getOperand(0);
751 // If the replacement value is the load, this must occur in unreachable
754 ReplVal = UndefValue::get(LI->getType());
755 LI->replaceAllUsesWith(ReplVal);
756 if (AST && isa<PointerType>(LI->getType()))
757 AST->deleteValue(LI);
758 LI->eraseFromParent();
765 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
766 /// first element of a pair.
767 struct StoreIndexSearchPredicate {
768 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
769 const std::pair<unsigned, StoreInst*> &RHS) {
770 return LHS.first < RHS.first;
776 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
777 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
778 /// potentially useless PHI nodes by just performing a single linear pass over
779 /// the basic block using the Alloca.
781 /// If we cannot promote this alloca (because it is read before it is written),
782 /// return true. This is necessary in cases where, due to control flow, the
783 /// alloca is potentially undefined on some control flow paths. e.g. code like
784 /// this is potentially correct:
786 /// for (...) { if (c) { A = undef; undef = B; } }
788 /// ... so long as A is not used before undef is set.
790 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
791 LargeBlockInfo &LBI) {
792 // The trickiest case to handle is when we have large blocks. Because of this,
793 // this code is optimized assuming that large blocks happen. This does not
794 // significantly pessimize the small block case. This uses LargeBlockInfo to
795 // make it efficient to get the index of various operations in the block.
797 // Clear out UsingBlocks. We will reconstruct it here if needed.
798 Info.UsingBlocks.clear();
800 // Walk the use-def list of the alloca, getting the locations of all stores.
801 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
802 StoresByIndexTy StoresByIndex;
804 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
806 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
807 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
809 // If there are no stores to the alloca, just replace any loads with undef.
810 if (StoresByIndex.empty()) {
811 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
812 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
813 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
814 if (AST && isa<PointerType>(LI->getType()))
815 AST->deleteValue(LI);
817 LI->eraseFromParent();
822 // Sort the stores by their index, making it efficient to do a lookup with a
824 std::sort(StoresByIndex.begin(), StoresByIndex.end());
826 // Walk all of the loads from this alloca, replacing them with the nearest
827 // store above them, if any.
828 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
829 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
832 unsigned LoadIdx = LBI.getInstructionIndex(LI);
834 // Find the nearest store that has a lower than this load.
835 StoresByIndexTy::iterator I =
836 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
837 std::pair<unsigned, StoreInst*>(LoadIdx, 0),
838 StoreIndexSearchPredicate());
840 // If there is no store before this load, then we can't promote this load.
841 if (I == StoresByIndex.begin()) {
842 // Can't handle this load, bail out.
843 Info.UsingBlocks.push_back(LI->getParent());
847 // Otherwise, there was a store before this load, the load takes its value.
849 LI->replaceAllUsesWith(I->second->getOperand(0));
850 if (AST && isa<PointerType>(LI->getType()))
851 AST->deleteValue(LI);
852 LI->eraseFromParent();
857 // Inserts a llvm.dbg.value instrinsic before the stores to an alloca'd value
858 // that has an associated llvm.dbg.decl intrinsic.
859 void PromoteMem2Reg::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
865 DIF = new DIFactory(*SI->getParent()->getParent()->getParent());
866 DIF->InsertDbgValueIntrinsic(SI->getOperand(0), Offset,
867 DIVariable(DDI->getVariable()), SI);
870 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
871 // Alloca returns true if there wasn't already a phi-node for that variable
873 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
875 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
876 // Look up the basic-block in question.
877 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
879 // If the BB already has a phi node added for the i'th alloca then we're done!
880 if (PN) return false;
882 // Create a PhiNode using the dereferenced type... and add the phi-node to the
884 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
885 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
888 PhiToAllocaMap[PN] = AllocaNo;
889 PN->reserveOperandSpace(getNumPreds(BB));
891 InsertedPHINodes.insert(PN);
893 if (AST && isa<PointerType>(PN->getType()))
894 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
899 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
900 // stores to the allocas which we are promoting. IncomingVals indicates what
901 // value each Alloca contains on exit from the predecessor block Pred.
903 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
904 RenamePassData::ValVector &IncomingVals,
905 std::vector<RenamePassData> &Worklist) {
907 // If we are inserting any phi nodes into this BB, they will already be in the
909 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
910 // If we have PHI nodes to update, compute the number of edges from Pred to
912 if (PhiToAllocaMap.count(APN)) {
913 // We want to be able to distinguish between PHI nodes being inserted by
914 // this invocation of mem2reg from those phi nodes that already existed in
915 // the IR before mem2reg was run. We determine that APN is being inserted
916 // because it is missing incoming edges. All other PHI nodes being
917 // inserted by this pass of mem2reg will have the same number of incoming
918 // operands so far. Remember this count.
919 unsigned NewPHINumOperands = APN->getNumOperands();
921 unsigned NumEdges = 0;
922 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
925 assert(NumEdges && "Must be at least one edge from Pred to BB!");
927 // Add entries for all the phis.
928 BasicBlock::iterator PNI = BB->begin();
930 unsigned AllocaNo = PhiToAllocaMap[APN];
932 // Add N incoming values to the PHI node.
933 for (unsigned i = 0; i != NumEdges; ++i)
934 APN->addIncoming(IncomingVals[AllocaNo], Pred);
936 // The currently active variable for this block is now the PHI.
937 IncomingVals[AllocaNo] = APN;
939 // Get the next phi node.
941 APN = dyn_cast<PHINode>(PNI);
944 // Verify that it is missing entries. If not, it is not being inserted
945 // by this mem2reg invocation so we want to ignore it.
946 } while (APN->getNumOperands() == NewPHINumOperands);
950 // Don't revisit blocks.
951 if (!Visited.insert(BB)) return;
953 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
954 Instruction *I = II++; // get the instruction, increment iterator
956 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
957 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
960 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
961 if (AI == AllocaLookup.end()) continue;
963 Value *V = IncomingVals[AI->second];
965 // Anything using the load now uses the current value.
966 LI->replaceAllUsesWith(V);
967 if (AST && isa<PointerType>(LI->getType()))
968 AST->deleteValue(LI);
969 BB->getInstList().erase(LI);
970 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
971 // Delete this instruction and mark the name as the current holder of the
973 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
976 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
977 if (ai == AllocaLookup.end())
980 // what value were we writing?
981 IncomingVals[ai->second] = SI->getOperand(0);
982 // Record debuginfo for the store before removing it.
983 ConvertDebugDeclareToDebugValue(findDbgDeclare(Dest), SI, 0);
984 BB->getInstList().erase(SI);
988 // 'Recurse' to our successors.
989 succ_iterator I = succ_begin(BB), E = succ_end(BB);
992 // Keep track of the successors so we don't visit the same successor twice
993 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
995 // Handle the first successor without using the worklist.
996 VisitedSuccs.insert(*I);
1002 if (VisitedSuccs.insert(*I))
1003 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1008 /// PromoteMemToReg - Promote the specified list of alloca instructions into
1009 /// scalar registers, inserting PHI nodes as appropriate. This function makes
1010 /// use of DominanceFrontier information. This function does not modify the CFG
1011 /// of the function at all. All allocas must be from the same function.
1013 /// If AST is specified, the specified tracker is updated to reflect changes
1016 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1017 DominatorTree &DT, DominanceFrontier &DF,
1018 AliasSetTracker *AST) {
1019 // If there is nothing to do, bail out...
1020 if (Allocas.empty()) return;
1022 PromoteMem2Reg(Allocas, DT, DF, AST).run();