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/Dominators.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/ADT/StringExtras.h"
33 #include "llvm/Support/CFG.h"
34 #include "llvm/Support/Compiler.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");
43 // Provide DenseMapInfo for all pointers.
46 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
47 typedef std::pair<BasicBlock*, unsigned> EltTy;
48 static inline EltTy getEmptyKey() {
49 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
51 static inline EltTy getTombstoneKey() {
52 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
54 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
55 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
57 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
60 static bool isPod() { return true; }
64 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
65 /// This is true if there are only loads and stores to the alloca.
67 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
68 // FIXME: If the memory unit is of pointer or integer type, we can permit
69 // assignments to subsections of the memory unit.
71 // Only allow direct and non-volatile loads and stores...
72 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
73 UI != UE; ++UI) // Loop over all of the uses of the alloca
74 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
77 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
78 if (SI->getOperand(0) == AI)
79 return false; // Don't allow a store OF the AI, only INTO the AI.
83 const BitCastInst *BC = dyn_cast<BitCastInst>(*UI);
85 return false; // Not a load or store or dbg intrinsic.
86 Value::use_const_iterator BCUI = BC->use_begin(), BCUE = BC->use_end();
88 return false; // Not a dbg intrinsic.
89 const DbgInfoIntrinsic *DI = dyn_cast<DbgInfoIntrinsic>(*BCUI);
91 return false; // Not a dbg intrinsic.
94 return false; // Not a dbg intrinsic use.
103 // Data package used by RenamePass()
104 class VISIBILITY_HIDDEN RenamePassData {
106 typedef std::vector<Value *> ValVector;
109 RenamePassData(BasicBlock *B, BasicBlock *P,
110 const ValVector &V) : BB(B), Pred(P), Values(V) {}
115 void swap(RenamePassData &RHS) {
116 std::swap(BB, RHS.BB);
117 std::swap(Pred, RHS.Pred);
118 Values.swap(RHS.Values);
122 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
123 /// load/store instructions in the block that directly load or store an alloca.
125 /// This functionality is important because it avoids scanning large basic
126 /// blocks multiple times when promoting many allocas in the same block.
127 class VISIBILITY_HIDDEN LargeBlockInfo {
128 /// InstNumbers - For each instruction that we track, keep the index of the
129 /// instruction. The index starts out as the number of the instruction from
130 /// the start of the block.
131 DenseMap<const Instruction *, unsigned> InstNumbers;
134 /// isInterestingInstruction - This code only looks at accesses to allocas.
135 static bool isInterestingInstruction(const Instruction *I) {
136 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
137 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
140 /// getInstructionIndex - Get or calculate the index of the specified
142 unsigned getInstructionIndex(const Instruction *I) {
143 assert(isInterestingInstruction(I) &&
144 "Not a load/store to/from an alloca?");
146 // If we already have this instruction number, return it.
147 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
148 if (It != InstNumbers.end()) return It->second;
150 // Scan the whole block to get the instruction. This accumulates
151 // information for every interesting instruction in the block, in order to
152 // avoid gratuitus rescans.
153 const BasicBlock *BB = I->getParent();
155 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
157 if (isInterestingInstruction(BBI))
158 InstNumbers[BBI] = InstNo++;
159 It = InstNumbers.find(I);
161 assert(It != InstNumbers.end() && "Didn't insert instruction?");
165 void deleteValue(const Instruction *I) {
166 InstNumbers.erase(I);
174 struct VISIBILITY_HIDDEN PromoteMem2Reg {
175 /// Allocas - The alloca instructions being promoted.
177 std::vector<AllocaInst*> Allocas;
179 DominanceFrontier &DF;
181 /// AST - An AliasSetTracker object to update. If null, don't update it.
183 AliasSetTracker *AST;
185 /// AllocaLookup - Reverse mapping of Allocas.
187 std::map<AllocaInst*, unsigned> AllocaLookup;
189 /// NewPhiNodes - The PhiNodes we're adding.
191 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
193 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
194 /// it corresponds to.
195 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
197 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
198 /// each alloca that is of pointer type, we keep track of what to copyValue
199 /// to the inserted PHI nodes here.
201 std::vector<Value*> PointerAllocaValues;
203 /// Visited - The set of basic blocks the renamer has already visited.
205 SmallPtrSet<BasicBlock*, 16> Visited;
207 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
208 /// non-determinstic behavior.
209 DenseMap<BasicBlock*, unsigned> BBNumbers;
211 /// BBNumPreds - Lazily compute the number of predecessors a block has.
212 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
214 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
215 DominanceFrontier &df, AliasSetTracker *ast)
216 : Allocas(A), DT(dt), DF(df), AST(ast) {}
220 /// properlyDominates - Return true if I1 properly dominates I2.
222 bool properlyDominates(Instruction *I1, Instruction *I2) const {
223 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
224 I1 = II->getNormalDest()->begin();
225 return DT.properlyDominates(I1->getParent(), I2->getParent());
228 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
230 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
231 return DT.dominates(BB1, BB2);
235 void RemoveFromAllocasList(unsigned &AllocaIdx) {
236 Allocas[AllocaIdx] = Allocas.back();
241 unsigned getNumPreds(const BasicBlock *BB) {
242 unsigned &NP = BBNumPreds[BB];
244 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
248 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
250 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
251 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
252 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
254 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
255 LargeBlockInfo &LBI);
256 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
257 LargeBlockInfo &LBI);
260 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
261 RenamePassData::ValVector &IncVals,
262 std::vector<RenamePassData> &Worklist);
263 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
264 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
268 std::vector<BasicBlock*> DefiningBlocks;
269 std::vector<BasicBlock*> UsingBlocks;
271 StoreInst *OnlyStore;
272 BasicBlock *OnlyBlock;
273 bool OnlyUsedInOneBlock;
275 Value *AllocaPointerVal;
278 DefiningBlocks.clear();
282 OnlyUsedInOneBlock = true;
283 AllocaPointerVal = 0;
286 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
288 void AnalyzeAlloca(AllocaInst *AI) {
291 // As we scan the uses of the alloca instruction, keep track of stores,
292 // and decide whether all of the loads and stores to the alloca are within
293 // the same basic block.
294 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end();
296 Instruction *User = cast<Instruction>(*U);
297 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
298 // Remove dbg intrinsic uses now.
299 Value::use_iterator BCUI = BC->use_begin();
300 DbgInfoIntrinsic *DI = cast<DbgInfoIntrinsic>(*BCUI);
301 assert (BCUI + 1 == BC->use_end() && "Unexpected alloca uses!");
302 DI->eraseFromParent();
303 BC->eraseFromParent();
304 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
305 // Remember the basic blocks which define new values for the alloca
306 DefiningBlocks.push_back(SI->getParent());
307 AllocaPointerVal = SI->getOperand(0);
310 LoadInst *LI = cast<LoadInst>(User);
311 // Otherwise it must be a load instruction, keep track of variable
313 UsingBlocks.push_back(LI->getParent());
314 AllocaPointerVal = LI;
317 if (OnlyUsedInOneBlock) {
319 OnlyBlock = User->getParent();
320 else if (OnlyBlock != User->getParent())
321 OnlyUsedInOneBlock = false;
326 } // end of anonymous namespace
329 void PromoteMem2Reg::run() {
330 Function &F = *DF.getRoot()->getParent();
332 if (AST) PointerAllocaValues.resize(Allocas.size());
337 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
338 AllocaInst *AI = Allocas[AllocaNum];
340 assert(isAllocaPromotable(AI) &&
341 "Cannot promote non-promotable alloca!");
342 assert(AI->getParent()->getParent() == &F &&
343 "All allocas should be in the same function, which is same as DF!");
345 if (AI->use_empty()) {
346 // If there are no uses of the alloca, just delete it now.
347 if (AST) AST->deleteValue(AI);
348 AI->eraseFromParent();
350 // Remove the alloca from the Allocas list, since it has been processed
351 RemoveFromAllocasList(AllocaNum);
356 // Calculate the set of read and write-locations for each alloca. This is
357 // analogous to finding the 'uses' and 'definitions' of each variable.
358 Info.AnalyzeAlloca(AI);
360 // If there is only a single store to this value, replace any loads of
361 // it that are directly dominated by the definition with the value stored.
362 if (Info.DefiningBlocks.size() == 1) {
363 RewriteSingleStoreAlloca(AI, Info, LBI);
365 // Finally, after the scan, check to see if the store is all that is left.
366 if (Info.UsingBlocks.empty()) {
367 // Remove the (now dead) store and alloca.
368 Info.OnlyStore->eraseFromParent();
369 LBI.deleteValue(Info.OnlyStore);
371 if (AST) AST->deleteValue(AI);
372 AI->eraseFromParent();
375 // The alloca has been processed, move on.
376 RemoveFromAllocasList(AllocaNum);
383 // If the alloca is only read and written in one basic block, just perform a
384 // linear sweep over the block to eliminate it.
385 if (Info.OnlyUsedInOneBlock) {
386 PromoteSingleBlockAlloca(AI, Info, LBI);
388 // Finally, after the scan, check to see if the stores are all that is
390 if (Info.UsingBlocks.empty()) {
392 // Remove the (now dead) stores and alloca.
393 while (!AI->use_empty()) {
394 StoreInst *SI = cast<StoreInst>(AI->use_back());
395 SI->eraseFromParent();
399 if (AST) AST->deleteValue(AI);
400 AI->eraseFromParent();
403 // The alloca has been processed, move on.
404 RemoveFromAllocasList(AllocaNum);
411 // If we haven't computed a numbering for the BB's in the function, do so
413 if (BBNumbers.empty()) {
415 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
419 // If we have an AST to keep updated, remember some pointer value that is
420 // stored into the alloca.
422 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
424 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
425 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
427 // At this point, we're committed to promoting the alloca using IDF's, and
428 // the standard SSA construction algorithm. Determine which blocks need PHI
429 // nodes and see if we can optimize out some work by avoiding insertion of
431 DetermineInsertionPoint(AI, AllocaNum, Info);
435 return; // All of the allocas must have been trivial!
440 // Set the incoming values for the basic block to be null values for all of
441 // the alloca's. We do this in case there is a load of a value that has not
442 // been stored yet. In this case, it will get this null value.
444 RenamePassData::ValVector Values(Allocas.size());
445 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
446 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
448 // Walks all basic blocks in the function performing the SSA rename algorithm
449 // and inserting the phi nodes we marked as necessary
451 std::vector<RenamePassData> RenamePassWorkList;
452 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
453 while (!RenamePassWorkList.empty()) {
455 RPD.swap(RenamePassWorkList.back());
456 RenamePassWorkList.pop_back();
457 // RenamePass may add new worklist entries.
458 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
461 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
464 // Remove the allocas themselves from the function.
465 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
466 Instruction *A = Allocas[i];
468 // If there are any uses of the alloca instructions left, they must be in
469 // sections of dead code that were not processed on the dominance frontier.
470 // Just delete the users now.
473 A->replaceAllUsesWith(UndefValue::get(A->getType()));
474 if (AST) AST->deleteValue(A);
475 A->eraseFromParent();
479 // Loop over all of the PHI nodes and see if there are any that we can get
480 // rid of because they merge all of the same incoming values. This can
481 // happen due to undef values coming into the PHI nodes. This process is
482 // iterative, because eliminating one PHI node can cause others to be removed.
483 bool EliminatedAPHI = true;
484 while (EliminatedAPHI) {
485 EliminatedAPHI = false;
487 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
488 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
489 PHINode *PN = I->second;
491 // If this PHI node merges one value and/or undefs, get the value.
492 if (Value *V = PN->hasConstantValue(true)) {
493 if (!isa<Instruction>(V) ||
494 properlyDominates(cast<Instruction>(V), PN)) {
495 if (AST && isa<PointerType>(PN->getType()))
496 AST->deleteValue(PN);
497 PN->replaceAllUsesWith(V);
498 PN->eraseFromParent();
499 NewPhiNodes.erase(I++);
500 EliminatedAPHI = true;
508 // At this point, the renamer has added entries to PHI nodes for all reachable
509 // code. Unfortunately, there may be unreachable blocks which the renamer
510 // hasn't traversed. If this is the case, the PHI nodes may not
511 // have incoming values for all predecessors. Loop over all PHI nodes we have
512 // created, inserting undef values if they are missing any incoming values.
514 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
515 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
516 // We want to do this once per basic block. As such, only process a block
517 // when we find the PHI that is the first entry in the block.
518 PHINode *SomePHI = I->second;
519 BasicBlock *BB = SomePHI->getParent();
520 if (&BB->front() != SomePHI)
523 // Only do work here if there the PHI nodes are missing incoming values. We
524 // know that all PHI nodes that were inserted in a block will have the same
525 // number of incoming values, so we can just check any of them.
526 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
529 // Get the preds for BB.
530 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
532 // Ok, now we know that all of the PHI nodes are missing entries for some
533 // basic blocks. Start by sorting the incoming predecessors for efficient
535 std::sort(Preds.begin(), Preds.end());
537 // Now we loop through all BB's which have entries in SomePHI and remove
538 // them from the Preds list.
539 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
540 // Do a log(n) search of the Preds list for the entry we want.
541 SmallVector<BasicBlock*, 16>::iterator EntIt =
542 std::lower_bound(Preds.begin(), Preds.end(),
543 SomePHI->getIncomingBlock(i));
544 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
545 "PHI node has entry for a block which is not a predecessor!");
551 // At this point, the blocks left in the preds list must have dummy
552 // entries inserted into every PHI nodes for the block. Update all the phi
553 // nodes in this block that we are inserting (there could be phis before
555 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
556 BasicBlock::iterator BBI = BB->begin();
557 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
558 SomePHI->getNumIncomingValues() == NumBadPreds) {
559 Value *UndefVal = UndefValue::get(SomePHI->getType());
560 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
561 SomePHI->addIncoming(UndefVal, Preds[pred]);
569 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
570 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
571 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
573 void PromoteMem2Reg::
574 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
575 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
576 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
578 // To determine liveness, we must iterate through the predecessors of blocks
579 // where the def is live. Blocks are added to the worklist if we need to
580 // check their predecessors. Start with all the using blocks.
581 SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
582 LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
583 Info.UsingBlocks.begin(), Info.UsingBlocks.end());
585 // If any of the using blocks is also a definition block, check to see if the
586 // definition occurs before or after the use. If it happens before the use,
587 // the value isn't really live-in.
588 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
589 BasicBlock *BB = LiveInBlockWorklist[i];
590 if (!DefBlocks.count(BB)) continue;
592 // Okay, this is a block that both uses and defines the value. If the first
593 // reference to the alloca is a def (store), then we know it isn't live-in.
594 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
595 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
596 if (SI->getOperand(1) != AI) continue;
598 // We found a store to the alloca before a load. The alloca is not
599 // actually live-in here.
600 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
601 LiveInBlockWorklist.pop_back();
604 } else 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.back();
618 LiveInBlockWorklist.pop_back();
620 // The block really is live in here, insert it into the set. If already in
621 // the set, then it has already been processed.
622 if (!LiveInBlocks.insert(BB))
625 // Since the value is live into BB, it is either defined in a predecessor or
626 // live into it to. Add the preds to the worklist unless they are a
628 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
631 // The value is not live into a predecessor if it defines the value.
632 if (DefBlocks.count(P))
635 // Otherwise it is, add to the worklist.
636 LiveInBlockWorklist.push_back(P);
641 /// DetermineInsertionPoint - At this point, we're committed to promoting the
642 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
643 /// which blocks need phi nodes and see if we can optimize out some work by
644 /// avoiding insertion of dead phi nodes.
645 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
648 // Unique the set of defining blocks for efficient lookup.
649 SmallPtrSet<BasicBlock*, 32> DefBlocks;
650 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
652 // Determine which blocks the value is live in. These are blocks which lead
654 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
655 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
657 // Compute the locations where PhiNodes need to be inserted. Look at the
658 // dominance frontier of EACH basic-block we have a write in.
659 unsigned CurrentVersion = 0;
660 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
661 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
662 while (!Info.DefiningBlocks.empty()) {
663 BasicBlock *BB = Info.DefiningBlocks.back();
664 Info.DefiningBlocks.pop_back();
666 // Look up the DF for this write, add it to defining blocks.
667 DominanceFrontier::const_iterator it = DF.find(BB);
668 if (it == DF.end()) continue;
670 const DominanceFrontier::DomSetType &S = it->second;
672 // In theory we don't need the indirection through the DFBlocks vector.
673 // In practice, the order of calling QueuePhiNode would depend on the
674 // (unspecified) ordering of basic blocks in the dominance frontier,
675 // which would give PHI nodes non-determinstic subscripts. Fix this by
676 // processing blocks in order of the occurance in the function.
677 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
678 PE = S.end(); P != PE; ++P) {
679 // If the frontier block is not in the live-in set for the alloca, don't
680 // bother processing it.
681 if (!LiveInBlocks.count(*P))
684 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
687 // Sort by which the block ordering in the function.
688 if (DFBlocks.size() > 1)
689 std::sort(DFBlocks.begin(), DFBlocks.end());
691 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
692 BasicBlock *BB = DFBlocks[i].second;
693 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
694 Info.DefiningBlocks.push_back(BB);
700 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
701 /// replace any loads of it that are directly dominated by the definition with
702 /// the value stored.
703 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
705 LargeBlockInfo &LBI) {
706 StoreInst *OnlyStore = Info.OnlyStore;
707 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
708 BasicBlock *StoreBB = OnlyStore->getParent();
711 // Clear out UsingBlocks. We will reconstruct it here if needed.
712 Info.UsingBlocks.clear();
714 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
715 Instruction *UserInst = cast<Instruction>(*UI++);
716 if (!isa<LoadInst>(UserInst)) {
717 assert(UserInst == OnlyStore && "Should only have load/stores");
720 LoadInst *LI = cast<LoadInst>(UserInst);
722 // Okay, if we have a load from the alloca, we want to replace it with the
723 // only value stored to the alloca. We can do this if the value is
724 // dominated by the store. If not, we use the rest of the mem2reg machinery
725 // to insert the phi nodes as needed.
726 if (!StoringGlobalVal) { // Non-instructions are always dominated.
727 if (LI->getParent() == StoreBB) {
728 // If we have a use that is in the same block as the store, compare the
729 // indices of the two instructions to see which one came first. If the
730 // load came before the store, we can't handle it.
731 if (StoreIndex == -1)
732 StoreIndex = LBI.getInstructionIndex(OnlyStore);
734 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
735 // Can't handle this load, bail out.
736 Info.UsingBlocks.push_back(StoreBB);
740 } else if (LI->getParent() != StoreBB &&
741 !dominates(StoreBB, LI->getParent())) {
742 // If the load and store are in different blocks, use BB dominance to
743 // check their relationships. If the store doesn't dom the use, bail
745 Info.UsingBlocks.push_back(LI->getParent());
750 // Otherwise, we *can* safely rewrite this load.
751 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
752 if (AST && isa<PointerType>(LI->getType()))
753 AST->deleteValue(LI);
754 LI->eraseFromParent();
760 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
761 /// first element of a pair.
762 struct StoreIndexSearchPredicate {
763 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
764 const std::pair<unsigned, StoreInst*> &RHS) {
765 return LHS.first < RHS.first;
769 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
770 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
771 /// potentially useless PHI nodes by just performing a single linear pass over
772 /// the basic block using the Alloca.
774 /// If we cannot promote this alloca (because it is read before it is written),
775 /// return true. This is necessary in cases where, due to control flow, the
776 /// alloca is potentially undefined on some control flow paths. e.g. code like
777 /// this is potentially correct:
779 /// for (...) { if (c) { A = undef; undef = B; } }
781 /// ... so long as A is not used before undef is set.
783 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
784 LargeBlockInfo &LBI) {
785 // The trickiest case to handle is when we have large blocks. Because of this,
786 // this code is optimized assuming that large blocks happen. This does not
787 // significantly pessimize the small block case. This uses LargeBlockInfo to
788 // make it efficient to get the index of various operations in the block.
790 // Clear out UsingBlocks. We will reconstruct it here if needed.
791 Info.UsingBlocks.clear();
793 // Walk the use-def list of the alloca, getting the locations of all stores.
794 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
795 StoresByIndexTy StoresByIndex;
797 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
799 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
800 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
802 // If there are no stores to the alloca, just replace any loads with undef.
803 if (StoresByIndex.empty()) {
804 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
805 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
806 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
807 if (AST && isa<PointerType>(LI->getType()))
808 AST->deleteValue(LI);
810 LI->eraseFromParent();
815 // Sort the stores by their index, making it efficient to do a lookup with a
817 std::sort(StoresByIndex.begin(), StoresByIndex.end());
819 // Walk all of the loads from this alloca, replacing them with the nearest
820 // store above them, if any.
821 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
822 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
825 unsigned LoadIdx = LBI.getInstructionIndex(LI);
827 // Find the nearest store that has a lower than this load.
828 StoresByIndexTy::iterator I =
829 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
830 std::pair<unsigned, StoreInst*>(LoadIdx, 0),
831 StoreIndexSearchPredicate());
833 // If there is no store before this load, then we can't promote this load.
834 if (I == StoresByIndex.begin()) {
835 // Can't handle this load, bail out.
836 Info.UsingBlocks.push_back(LI->getParent());
840 // Otherwise, there was a store before this load, the load takes its value.
842 LI->replaceAllUsesWith(I->second->getOperand(0));
843 if (AST && isa<PointerType>(LI->getType()))
844 AST->deleteValue(LI);
845 LI->eraseFromParent();
851 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
852 // Alloca returns true if there wasn't already a phi-node for that variable
854 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
856 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
857 // Look up the basic-block in question.
858 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
860 // If the BB already has a phi node added for the i'th alloca then we're done!
861 if (PN) return false;
863 // Create a PhiNode using the dereferenced type... and add the phi-node to the
865 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
866 Allocas[AllocaNo]->getName() + "." +
867 utostr(Version++), BB->begin());
869 PhiToAllocaMap[PN] = AllocaNo;
870 PN->reserveOperandSpace(getNumPreds(BB));
872 InsertedPHINodes.insert(PN);
874 if (AST && isa<PointerType>(PN->getType()))
875 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
880 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
881 // stores to the allocas which we are promoting. IncomingVals indicates what
882 // value each Alloca contains on exit from the predecessor block Pred.
884 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
885 RenamePassData::ValVector &IncomingVals,
886 std::vector<RenamePassData> &Worklist) {
888 // If we are inserting any phi nodes into this BB, they will already be in the
890 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
891 // Pred may have multiple edges to BB. If so, we want to add N incoming
892 // values to each PHI we are inserting on the first time we see the edge.
893 // Check to see if APN already has incoming values from Pred. This also
894 // prevents us from modifying PHI nodes that are not currently being
896 bool HasPredEntries = false;
897 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
898 if (APN->getIncomingBlock(i) == Pred) {
899 HasPredEntries = true;
904 // If we have PHI nodes to update, compute the number of edges from Pred to
906 if (!HasPredEntries) {
907 // We want to be able to distinguish between PHI nodes being inserted by
908 // this invocation of mem2reg from those phi nodes that already existed in
909 // the IR before mem2reg was run. We determine that APN is being inserted
910 // because it is missing incoming edges. All other PHI nodes being
911 // inserted by this pass of mem2reg will have the same number of incoming
912 // operands so far. Remember this count.
913 unsigned NewPHINumOperands = APN->getNumOperands();
915 unsigned NumEdges = 0;
916 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
919 assert(NumEdges && "Must be at least one edge from Pred to BB!");
921 // Add entries for all the phis.
922 BasicBlock::iterator PNI = BB->begin();
924 unsigned AllocaNo = PhiToAllocaMap[APN];
926 // Add N incoming values to the PHI node.
927 for (unsigned i = 0; i != NumEdges; ++i)
928 APN->addIncoming(IncomingVals[AllocaNo], Pred);
930 // The currently active variable for this block is now the PHI.
931 IncomingVals[AllocaNo] = APN;
933 // Get the next phi node.
935 APN = dyn_cast<PHINode>(PNI);
938 // Verify that it is missing entries. If not, it is not being inserted
939 // by this mem2reg invocation so we want to ignore it.
940 } while (APN->getNumOperands() == NewPHINumOperands);
944 // Don't revisit blocks.
945 if (!Visited.insert(BB)) return;
947 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
948 Instruction *I = II++; // get the instruction, increment iterator
950 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
951 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
954 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
955 if (AI == AllocaLookup.end()) continue;
957 Value *V = IncomingVals[AI->second];
959 // Anything using the load now uses the current value.
960 LI->replaceAllUsesWith(V);
961 if (AST && isa<PointerType>(LI->getType()))
962 AST->deleteValue(LI);
963 BB->getInstList().erase(LI);
964 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
965 // Delete this instruction and mark the name as the current holder of the
967 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
970 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
971 if (ai == AllocaLookup.end())
974 // what value were we writing?
975 IncomingVals[ai->second] = SI->getOperand(0);
976 BB->getInstList().erase(SI);
980 // 'Recurse' to our successors.
981 succ_iterator I = succ_begin(BB), E = succ_end(BB);
984 // Handle the last successor without using the worklist. This allows us to
985 // handle unconditional branches directly, for example.
988 Worklist.push_back(RenamePassData(*I, BB, IncomingVals));
995 /// PromoteMemToReg - Promote the specified list of alloca instructions into
996 /// scalar registers, inserting PHI nodes as appropriate. This function makes
997 /// use of DominanceFrontier information. This function does not modify the CFG
998 /// of the function at all. All allocas must be from the same function.
1000 /// If AST is specified, the specified tracker is updated to reflect changes
1003 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1004 DominatorTree &DT, DominanceFrontier &DF,
1005 AliasSetTracker *AST) {
1006 // If there is nothing to do, bail out...
1007 if (Allocas.empty()) return;
1009 PromoteMem2Reg(Allocas, DT, DF, AST).run();