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/ADT/STLExtras.h"
34 #include "llvm/Support/CFG.h"
35 #include "llvm/Support/Compiler.h"
39 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
40 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
41 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
42 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
44 // Provide DenseMapInfo for all pointers.
47 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
48 typedef std::pair<BasicBlock*, unsigned> EltTy;
49 static inline EltTy getEmptyKey() {
50 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
52 static inline EltTy getTombstoneKey() {
53 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
55 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
56 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
58 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
61 static bool isPod() { return true; }
65 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
66 /// This is true if there are only loads and stores to the alloca.
68 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
69 // FIXME: If the memory unit is of pointer or integer type, we can permit
70 // assignments to subsections of the memory unit.
72 // Only allow direct and non-volatile loads and stores...
73 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
74 UI != UE; ++UI) // Loop over all of the uses of the alloca
75 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
78 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
79 if (SI->getOperand(0) == AI)
80 return false; // Don't allow a store OF the AI, only INTO the AI.
83 } else if (const BitCastInst *BC = dyn_cast<BitCastInst>(*UI)) {
84 // A bitcast that does not feed into debug info inhibits promotion.
85 if (!BC->hasOneUse() || !isa<DbgInfoIntrinsic>(*BC->use_begin()))
87 // If the only use is by debug info, this alloca will not exist in
88 // non-debug code, so don't try to promote; this ensures the same
89 // codegen with debug info. Otherwise, debug info should not
90 // inhibit promotion (but we must examine other uses).
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);
298 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
299 // Remove any uses of this alloca in DbgInfoInstrinsics.
300 assert(BC->hasOneUse() && "Unexpected alloca uses!");
301 DbgInfoIntrinsic *DI = cast<DbgInfoIntrinsic>(*BC->use_begin());
302 DI->eraseFromParent();
303 BC->eraseFromParent();
306 else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
307 // Remember the basic blocks which define new values for the alloca
308 DefiningBlocks.push_back(SI->getParent());
309 AllocaPointerVal = SI->getOperand(0);
312 LoadInst *LI = cast<LoadInst>(User);
313 // Otherwise it must be a load instruction, keep track of variable
315 UsingBlocks.push_back(LI->getParent());
316 AllocaPointerVal = LI;
319 if (OnlyUsedInOneBlock) {
321 OnlyBlock = User->getParent();
322 else if (OnlyBlock != User->getParent())
323 OnlyUsedInOneBlock = false;
328 } // end of anonymous namespace
331 void PromoteMem2Reg::run() {
332 Function &F = *DF.getRoot()->getParent();
334 if (AST) PointerAllocaValues.resize(Allocas.size());
339 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
340 AllocaInst *AI = Allocas[AllocaNum];
342 assert(isAllocaPromotable(AI) &&
343 "Cannot promote non-promotable alloca!");
344 assert(AI->getParent()->getParent() == &F &&
345 "All allocas should be in the same function, which is same as DF!");
347 if (AI->use_empty()) {
348 // If there are no uses of the alloca, just delete it now.
349 if (AST) AST->deleteValue(AI);
350 AI->eraseFromParent();
352 // Remove the alloca from the Allocas list, since it has been processed
353 RemoveFromAllocasList(AllocaNum);
358 // Calculate the set of read and write-locations for each alloca. This is
359 // analogous to finding the 'uses' and 'definitions' of each variable.
360 Info.AnalyzeAlloca(AI);
362 // If there is only a single store to this value, replace any loads of
363 // it that are directly dominated by the definition with the value stored.
364 if (Info.DefiningBlocks.size() == 1) {
365 RewriteSingleStoreAlloca(AI, Info, LBI);
367 // Finally, after the scan, check to see if the store is all that is left.
368 if (Info.UsingBlocks.empty()) {
369 // Remove the (now dead) store and alloca.
370 Info.OnlyStore->eraseFromParent();
371 LBI.deleteValue(Info.OnlyStore);
373 if (AST) AST->deleteValue(AI);
374 AI->eraseFromParent();
377 // The alloca has been processed, move on.
378 RemoveFromAllocasList(AllocaNum);
385 // If the alloca is only read and written in one basic block, just perform a
386 // linear sweep over the block to eliminate it.
387 if (Info.OnlyUsedInOneBlock) {
388 PromoteSingleBlockAlloca(AI, Info, LBI);
390 // Finally, after the scan, check to see if the stores are all that is
392 if (Info.UsingBlocks.empty()) {
394 // Remove the (now dead) stores and alloca.
395 while (!AI->use_empty()) {
396 StoreInst *SI = cast<StoreInst>(AI->use_back());
397 SI->eraseFromParent();
401 if (AST) AST->deleteValue(AI);
402 AI->eraseFromParent();
405 // The alloca has been processed, move on.
406 RemoveFromAllocasList(AllocaNum);
413 // If we haven't computed a numbering for the BB's in the function, do so
415 if (BBNumbers.empty()) {
417 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
421 // If we have an AST to keep updated, remember some pointer value that is
422 // stored into the alloca.
424 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
426 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
427 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
429 // At this point, we're committed to promoting the alloca using IDF's, and
430 // the standard SSA construction algorithm. Determine which blocks need PHI
431 // nodes and see if we can optimize out some work by avoiding insertion of
433 DetermineInsertionPoint(AI, AllocaNum, Info);
437 return; // All of the allocas must have been trivial!
442 // Set the incoming values for the basic block to be null values for all of
443 // the alloca's. We do this in case there is a load of a value that has not
444 // been stored yet. In this case, it will get this null value.
446 RenamePassData::ValVector Values(Allocas.size());
447 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
448 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
450 // Walks all basic blocks in the function performing the SSA rename algorithm
451 // and inserting the phi nodes we marked as necessary
453 std::vector<RenamePassData> RenamePassWorkList;
454 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
455 while (!RenamePassWorkList.empty()) {
457 RPD.swap(RenamePassWorkList.back());
458 RenamePassWorkList.pop_back();
459 // RenamePass may add new worklist entries.
460 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
463 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
466 // Remove the allocas themselves from the function.
467 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
468 Instruction *A = Allocas[i];
470 // If there are any uses of the alloca instructions left, they must be in
471 // sections of dead code that were not processed on the dominance frontier.
472 // Just delete the users now.
475 A->replaceAllUsesWith(UndefValue::get(A->getType()));
476 if (AST) AST->deleteValue(A);
477 A->eraseFromParent();
481 // Loop over all of the PHI nodes and see if there are any that we can get
482 // rid of because they merge all of the same incoming values. This can
483 // happen due to undef values coming into the PHI nodes. This process is
484 // iterative, because eliminating one PHI node can cause others to be removed.
485 bool EliminatedAPHI = true;
486 while (EliminatedAPHI) {
487 EliminatedAPHI = false;
489 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
490 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
491 PHINode *PN = I->second;
493 // If this PHI node merges one value and/or undefs, get the value.
494 if (Value *V = PN->hasConstantValue(true)) {
495 if (!isa<Instruction>(V) ||
496 properlyDominates(cast<Instruction>(V), PN)) {
497 if (AST && isa<PointerType>(PN->getType()))
498 AST->deleteValue(PN);
499 PN->replaceAllUsesWith(V);
500 PN->eraseFromParent();
501 NewPhiNodes.erase(I++);
502 EliminatedAPHI = true;
510 // At this point, the renamer has added entries to PHI nodes for all reachable
511 // code. Unfortunately, there may be unreachable blocks which the renamer
512 // hasn't traversed. If this is the case, the PHI nodes may not
513 // have incoming values for all predecessors. Loop over all PHI nodes we have
514 // created, inserting undef values if they are missing any incoming values.
516 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
517 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
518 // We want to do this once per basic block. As such, only process a block
519 // when we find the PHI that is the first entry in the block.
520 PHINode *SomePHI = I->second;
521 BasicBlock *BB = SomePHI->getParent();
522 if (&BB->front() != SomePHI)
525 // Only do work here if there the PHI nodes are missing incoming values. We
526 // know that all PHI nodes that were inserted in a block will have the same
527 // number of incoming values, so we can just check any of them.
528 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
531 // Get the preds for BB.
532 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
534 // Ok, now we know that all of the PHI nodes are missing entries for some
535 // basic blocks. Start by sorting the incoming predecessors for efficient
537 std::sort(Preds.begin(), Preds.end());
539 // Now we loop through all BB's which have entries in SomePHI and remove
540 // them from the Preds list.
541 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
542 // Do a log(n) search of the Preds list for the entry we want.
543 SmallVector<BasicBlock*, 16>::iterator EntIt =
544 std::lower_bound(Preds.begin(), Preds.end(),
545 SomePHI->getIncomingBlock(i));
546 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
547 "PHI node has entry for a block which is not a predecessor!");
553 // At this point, the blocks left in the preds list must have dummy
554 // entries inserted into every PHI nodes for the block. Update all the phi
555 // nodes in this block that we are inserting (there could be phis before
557 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
558 BasicBlock::iterator BBI = BB->begin();
559 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
560 SomePHI->getNumIncomingValues() == NumBadPreds) {
561 Value *UndefVal = UndefValue::get(SomePHI->getType());
562 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
563 SomePHI->addIncoming(UndefVal, Preds[pred]);
571 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
572 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
573 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
575 void PromoteMem2Reg::
576 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
577 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
578 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
580 // To determine liveness, we must iterate through the predecessors of blocks
581 // where the def is live. Blocks are added to the worklist if we need to
582 // check their predecessors. Start with all the using blocks.
583 SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
584 LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
585 Info.UsingBlocks.begin(), Info.UsingBlocks.end());
587 // If any of the using blocks is also a definition block, check to see if the
588 // definition occurs before or after the use. If it happens before the use,
589 // the value isn't really live-in.
590 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
591 BasicBlock *BB = LiveInBlockWorklist[i];
592 if (!DefBlocks.count(BB)) continue;
594 // Okay, this is a block that both uses and defines the value. If the first
595 // reference to the alloca is a def (store), then we know it isn't live-in.
596 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
597 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
598 if (SI->getOperand(1) != AI) continue;
600 // We found a store to the alloca before a load. The alloca is not
601 // actually live-in here.
602 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
603 LiveInBlockWorklist.pop_back();
606 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
607 if (LI->getOperand(0) != AI) continue;
609 // Okay, we found a load before a store to the alloca. It is actually
610 // live into this block.
616 // Now that we have a set of blocks where the phi is live-in, recursively add
617 // their predecessors until we find the full region the value is live.
618 while (!LiveInBlockWorklist.empty()) {
619 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
621 // The block really is live in here, insert it into the set. If already in
622 // the set, then it has already been processed.
623 if (!LiveInBlocks.insert(BB))
626 // Since the value is live into BB, it is either defined in a predecessor or
627 // live into it to. Add the preds to the worklist unless they are a
629 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
632 // The value is not live into a predecessor if it defines the value.
633 if (DefBlocks.count(P))
636 // Otherwise it is, add to the worklist.
637 LiveInBlockWorklist.push_back(P);
642 /// DetermineInsertionPoint - At this point, we're committed to promoting the
643 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
644 /// which blocks need phi nodes and see if we can optimize out some work by
645 /// avoiding insertion of dead phi nodes.
646 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
649 // Unique the set of defining blocks for efficient lookup.
650 SmallPtrSet<BasicBlock*, 32> DefBlocks;
651 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
653 // Determine which blocks the value is live in. These are blocks which lead
655 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
656 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
658 // Compute the locations where PhiNodes need to be inserted. Look at the
659 // dominance frontier of EACH basic-block we have a write in.
660 unsigned CurrentVersion = 0;
661 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
662 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
663 while (!Info.DefiningBlocks.empty()) {
664 BasicBlock *BB = Info.DefiningBlocks.back();
665 Info.DefiningBlocks.pop_back();
667 // Look up the DF for this write, add it to defining blocks.
668 DominanceFrontier::const_iterator it = DF.find(BB);
669 if (it == DF.end()) continue;
671 const DominanceFrontier::DomSetType &S = it->second;
673 // In theory we don't need the indirection through the DFBlocks vector.
674 // In practice, the order of calling QueuePhiNode would depend on the
675 // (unspecified) ordering of basic blocks in the dominance frontier,
676 // which would give PHI nodes non-determinstic subscripts. Fix this by
677 // processing blocks in order of the occurance in the function.
678 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
679 PE = S.end(); P != PE; ++P) {
680 // If the frontier block is not in the live-in set for the alloca, don't
681 // bother processing it.
682 if (!LiveInBlocks.count(*P))
685 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
688 // Sort by which the block ordering in the function.
689 if (DFBlocks.size() > 1)
690 std::sort(DFBlocks.begin(), DFBlocks.end());
692 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
693 BasicBlock *BB = DFBlocks[i].second;
694 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
695 Info.DefiningBlocks.push_back(BB);
701 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
702 /// replace any loads of it that are directly dominated by the definition with
703 /// the value stored.
704 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
706 LargeBlockInfo &LBI) {
707 StoreInst *OnlyStore = Info.OnlyStore;
708 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
709 BasicBlock *StoreBB = OnlyStore->getParent();
712 // Clear out UsingBlocks. We will reconstruct it here if needed.
713 Info.UsingBlocks.clear();
715 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
716 Instruction *UserInst = cast<Instruction>(*UI++);
717 if (!isa<LoadInst>(UserInst)) {
718 assert(UserInst == OnlyStore && "Should only have load/stores");
721 LoadInst *LI = cast<LoadInst>(UserInst);
723 // Okay, if we have a load from the alloca, we want to replace it with the
724 // only value stored to the alloca. We can do this if the value is
725 // dominated by the store. If not, we use the rest of the mem2reg machinery
726 // to insert the phi nodes as needed.
727 if (!StoringGlobalVal) { // Non-instructions are always dominated.
728 if (LI->getParent() == StoreBB) {
729 // If we have a use that is in the same block as the store, compare the
730 // indices of the two instructions to see which one came first. If the
731 // load came before the store, we can't handle it.
732 if (StoreIndex == -1)
733 StoreIndex = LBI.getInstructionIndex(OnlyStore);
735 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
736 // Can't handle this load, bail out.
737 Info.UsingBlocks.push_back(StoreBB);
741 } else if (LI->getParent() != StoreBB &&
742 !dominates(StoreBB, LI->getParent())) {
743 // If the load and store are in different blocks, use BB dominance to
744 // check their relationships. If the store doesn't dom the use, bail
746 Info.UsingBlocks.push_back(LI->getParent());
751 // Otherwise, we *can* safely rewrite this load.
752 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
753 if (AST && isa<PointerType>(LI->getType()))
754 AST->deleteValue(LI);
755 LI->eraseFromParent();
761 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
762 /// first element of a pair.
763 struct StoreIndexSearchPredicate {
764 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
765 const std::pair<unsigned, StoreInst*> &RHS) {
766 return LHS.first < RHS.first;
770 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
771 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
772 /// potentially useless PHI nodes by just performing a single linear pass over
773 /// the basic block using the Alloca.
775 /// If we cannot promote this alloca (because it is read before it is written),
776 /// return true. This is necessary in cases where, due to control flow, the
777 /// alloca is potentially undefined on some control flow paths. e.g. code like
778 /// this is potentially correct:
780 /// for (...) { if (c) { A = undef; undef = B; } }
782 /// ... so long as A is not used before undef is set.
784 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
785 LargeBlockInfo &LBI) {
786 // The trickiest case to handle is when we have large blocks. Because of this,
787 // this code is optimized assuming that large blocks happen. This does not
788 // significantly pessimize the small block case. This uses LargeBlockInfo to
789 // make it efficient to get the index of various operations in the block.
791 // Clear out UsingBlocks. We will reconstruct it here if needed.
792 Info.UsingBlocks.clear();
794 // Walk the use-def list of the alloca, getting the locations of all stores.
795 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
796 StoresByIndexTy StoresByIndex;
798 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
800 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
801 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
803 // If there are no stores to the alloca, just replace any loads with undef.
804 if (StoresByIndex.empty()) {
805 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
806 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
807 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
808 if (AST && isa<PointerType>(LI->getType()))
809 AST->deleteValue(LI);
811 LI->eraseFromParent();
816 // Sort the stores by their index, making it efficient to do a lookup with a
818 std::sort(StoresByIndex.begin(), StoresByIndex.end());
820 // Walk all of the loads from this alloca, replacing them with the nearest
821 // store above them, if any.
822 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
823 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
826 unsigned LoadIdx = LBI.getInstructionIndex(LI);
828 // Find the nearest store that has a lower than this load.
829 StoresByIndexTy::iterator I =
830 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
831 std::pair<unsigned, StoreInst*>(LoadIdx, 0),
832 StoreIndexSearchPredicate());
834 // If there is no store before this load, then we can't promote this load.
835 if (I == StoresByIndex.begin()) {
836 // Can't handle this load, bail out.
837 Info.UsingBlocks.push_back(LI->getParent());
841 // Otherwise, there was a store before this load, the load takes its value.
843 LI->replaceAllUsesWith(I->second->getOperand(0));
844 if (AST && isa<PointerType>(LI->getType()))
845 AST->deleteValue(LI);
846 LI->eraseFromParent();
852 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
853 // Alloca returns true if there wasn't already a phi-node for that variable
855 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
857 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
858 // Look up the basic-block in question.
859 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
861 // If the BB already has a phi node added for the i'th alloca then we're done!
862 if (PN) return false;
864 // Create a PhiNode using the dereferenced type... and add the phi-node to the
866 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
867 Allocas[AllocaNo]->getName() + "." +
868 utostr(Version++), BB->begin());
870 PhiToAllocaMap[PN] = AllocaNo;
871 PN->reserveOperandSpace(getNumPreds(BB));
873 InsertedPHINodes.insert(PN);
875 if (AST && isa<PointerType>(PN->getType()))
876 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
881 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
882 // stores to the allocas which we are promoting. IncomingVals indicates what
883 // value each Alloca contains on exit from the predecessor block Pred.
885 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
886 RenamePassData::ValVector &IncomingVals,
887 std::vector<RenamePassData> &Worklist) {
889 // If we are inserting any phi nodes into this BB, they will already be in the
891 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
892 // If we have PHI nodes to update, compute the number of edges from Pred to
894 if (PhiToAllocaMap.count(APN)) {
895 // We want to be able to distinguish between PHI nodes being inserted by
896 // this invocation of mem2reg from those phi nodes that already existed in
897 // the IR before mem2reg was run. We determine that APN is being inserted
898 // because it is missing incoming edges. All other PHI nodes being
899 // inserted by this pass of mem2reg will have the same number of incoming
900 // operands so far. Remember this count.
901 unsigned NewPHINumOperands = APN->getNumOperands();
903 unsigned NumEdges = 0;
904 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
907 assert(NumEdges && "Must be at least one edge from Pred to BB!");
909 // Add entries for all the phis.
910 BasicBlock::iterator PNI = BB->begin();
912 unsigned AllocaNo = PhiToAllocaMap[APN];
914 // Add N incoming values to the PHI node.
915 for (unsigned i = 0; i != NumEdges; ++i)
916 APN->addIncoming(IncomingVals[AllocaNo], Pred);
918 // The currently active variable for this block is now the PHI.
919 IncomingVals[AllocaNo] = APN;
921 // Get the next phi node.
923 APN = dyn_cast<PHINode>(PNI);
926 // Verify that it is missing entries. If not, it is not being inserted
927 // by this mem2reg invocation so we want to ignore it.
928 } while (APN->getNumOperands() == NewPHINumOperands);
932 // Don't revisit blocks.
933 if (!Visited.insert(BB)) return;
935 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
936 Instruction *I = II++; // get the instruction, increment iterator
938 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
939 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
942 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
943 if (AI == AllocaLookup.end()) continue;
945 Value *V = IncomingVals[AI->second];
947 // Anything using the load now uses the current value.
948 LI->replaceAllUsesWith(V);
949 if (AST && isa<PointerType>(LI->getType()))
950 AST->deleteValue(LI);
951 BB->getInstList().erase(LI);
952 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
953 // Delete this instruction and mark the name as the current holder of the
955 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
958 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
959 if (ai == AllocaLookup.end())
962 // what value were we writing?
963 IncomingVals[ai->second] = SI->getOperand(0);
964 BB->getInstList().erase(SI);
968 // 'Recurse' to our successors.
969 succ_iterator I = succ_begin(BB), E = succ_end(BB);
972 // Keep track of the successors so we don't visit the same successor twice
973 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
975 // Handle the first successor without using the worklist.
976 VisitedSuccs.insert(*I);
982 if (VisitedSuccs.insert(*I))
983 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
988 /// PromoteMemToReg - Promote the specified list of alloca instructions into
989 /// scalar registers, inserting PHI nodes as appropriate. This function makes
990 /// use of DominanceFrontier information. This function does not modify the CFG
991 /// of the function at all. All allocas must be from the same function.
993 /// If AST is specified, the specified tracker is updated to reflect changes
996 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
997 DominatorTree &DT, DominanceFrontier &DF,
998 AliasSetTracker *AST) {
999 // If there is nothing to do, bail out...
1000 if (Allocas.empty()) return;
1002 PromoteMem2Reg(Allocas, DT, DF, AST).run();