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 // Uses by dbg info shouldn't inhibit promotion.
85 if (!BC->hasOneUse() || !isa<DbgInfoIntrinsic>(*BC->use_begin()))
97 // Data package used by RenamePass()
98 class VISIBILITY_HIDDEN RenamePassData {
100 typedef std::vector<Value *> ValVector;
103 RenamePassData(BasicBlock *B, BasicBlock *P,
104 const ValVector &V) : BB(B), Pred(P), Values(V) {}
109 void swap(RenamePassData &RHS) {
110 std::swap(BB, RHS.BB);
111 std::swap(Pred, RHS.Pred);
112 Values.swap(RHS.Values);
116 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
117 /// load/store instructions in the block that directly load or store an alloca.
119 /// This functionality is important because it avoids scanning large basic
120 /// blocks multiple times when promoting many allocas in the same block.
121 class VISIBILITY_HIDDEN LargeBlockInfo {
122 /// InstNumbers - For each instruction that we track, keep the index of the
123 /// instruction. The index starts out as the number of the instruction from
124 /// the start of the block.
125 DenseMap<const Instruction *, unsigned> InstNumbers;
128 /// isInterestingInstruction - This code only looks at accesses to allocas.
129 static bool isInterestingInstruction(const Instruction *I) {
130 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
131 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
134 /// getInstructionIndex - Get or calculate the index of the specified
136 unsigned getInstructionIndex(const Instruction *I) {
137 assert(isInterestingInstruction(I) &&
138 "Not a load/store to/from an alloca?");
140 // If we already have this instruction number, return it.
141 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
142 if (It != InstNumbers.end()) return It->second;
144 // Scan the whole block to get the instruction. This accumulates
145 // information for every interesting instruction in the block, in order to
146 // avoid gratuitus rescans.
147 const BasicBlock *BB = I->getParent();
149 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
151 if (isInterestingInstruction(BBI))
152 InstNumbers[BBI] = InstNo++;
153 It = InstNumbers.find(I);
155 assert(It != InstNumbers.end() && "Didn't insert instruction?");
159 void deleteValue(const Instruction *I) {
160 InstNumbers.erase(I);
168 struct VISIBILITY_HIDDEN PromoteMem2Reg {
169 /// Allocas - The alloca instructions being promoted.
171 std::vector<AllocaInst*> Allocas;
173 DominanceFrontier &DF;
175 /// AST - An AliasSetTracker object to update. If null, don't update it.
177 AliasSetTracker *AST;
179 /// AllocaLookup - Reverse mapping of Allocas.
181 std::map<AllocaInst*, unsigned> AllocaLookup;
183 /// NewPhiNodes - The PhiNodes we're adding.
185 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
187 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
188 /// it corresponds to.
189 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
191 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
192 /// each alloca that is of pointer type, we keep track of what to copyValue
193 /// to the inserted PHI nodes here.
195 std::vector<Value*> PointerAllocaValues;
197 /// Visited - The set of basic blocks the renamer has already visited.
199 SmallPtrSet<BasicBlock*, 16> Visited;
201 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
202 /// non-determinstic behavior.
203 DenseMap<BasicBlock*, unsigned> BBNumbers;
205 /// BBNumPreds - Lazily compute the number of predecessors a block has.
206 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
208 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
209 DominanceFrontier &df, AliasSetTracker *ast)
210 : Allocas(A), DT(dt), DF(df), AST(ast) {}
214 /// properlyDominates - Return true if I1 properly dominates I2.
216 bool properlyDominates(Instruction *I1, Instruction *I2) const {
217 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
218 I1 = II->getNormalDest()->begin();
219 return DT.properlyDominates(I1->getParent(), I2->getParent());
222 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
224 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
225 return DT.dominates(BB1, BB2);
229 void RemoveFromAllocasList(unsigned &AllocaIdx) {
230 Allocas[AllocaIdx] = Allocas.back();
235 unsigned getNumPreds(const BasicBlock *BB) {
236 unsigned &NP = BBNumPreds[BB];
238 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
242 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
244 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
245 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
246 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
248 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
249 LargeBlockInfo &LBI);
250 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
251 LargeBlockInfo &LBI);
254 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
255 RenamePassData::ValVector &IncVals,
256 std::vector<RenamePassData> &Worklist);
257 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
258 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
262 std::vector<BasicBlock*> DefiningBlocks;
263 std::vector<BasicBlock*> UsingBlocks;
265 StoreInst *OnlyStore;
266 BasicBlock *OnlyBlock;
267 bool OnlyUsedInOneBlock;
269 Value *AllocaPointerVal;
272 DefiningBlocks.clear();
276 OnlyUsedInOneBlock = true;
277 AllocaPointerVal = 0;
280 /// RemoveDebugUses - Remove uses of the alloca in DbgInfoInstrinsics.
281 void RemoveDebugUses(AllocaInst *AI) {
282 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end();
284 Instruction *User = cast<Instruction>(*U);
286 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
287 assert(BC->hasOneUse() && "Unexpected alloca uses!");
288 DbgInfoIntrinsic *DI = cast<DbgInfoIntrinsic>(*BC->use_begin());
289 DI->eraseFromParent();
290 BC->eraseFromParent();
295 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
297 void AnalyzeAlloca(AllocaInst *AI) {
300 // As we scan the uses of the alloca instruction, keep track of stores,
301 // and decide whether all of the loads and stores to the alloca are within
302 // the same basic block.
303 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end();
305 Instruction *User = cast<Instruction>(*U);
306 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 // Remove any uses of this alloca in DbgInfoInstrinsics.
348 Info.RemoveDebugUses(AI);
350 if (AI->use_empty()) {
351 // If there are no uses of the alloca, just delete it now.
352 if (AST) AST->deleteValue(AI);
353 AI->eraseFromParent();
355 // Remove the alloca from the Allocas list, since it has been processed
356 RemoveFromAllocasList(AllocaNum);
361 // Calculate the set of read and write-locations for each alloca. This is
362 // analogous to finding the 'uses' and 'definitions' of each variable.
363 Info.AnalyzeAlloca(AI);
365 // If there is only a single store to this value, replace any loads of
366 // it that are directly dominated by the definition with the value stored.
367 if (Info.DefiningBlocks.size() == 1) {
368 RewriteSingleStoreAlloca(AI, Info, LBI);
370 // Finally, after the scan, check to see if the store is all that is left.
371 if (Info.UsingBlocks.empty()) {
372 // Remove the (now dead) store and alloca.
373 Info.OnlyStore->eraseFromParent();
374 LBI.deleteValue(Info.OnlyStore);
376 if (AST) AST->deleteValue(AI);
377 AI->eraseFromParent();
380 // The alloca has been processed, move on.
381 RemoveFromAllocasList(AllocaNum);
388 // If the alloca is only read and written in one basic block, just perform a
389 // linear sweep over the block to eliminate it.
390 if (Info.OnlyUsedInOneBlock) {
391 PromoteSingleBlockAlloca(AI, Info, LBI);
393 // Finally, after the scan, check to see if the stores are all that is
395 if (Info.UsingBlocks.empty()) {
397 // Remove the (now dead) stores and alloca.
398 while (!AI->use_empty()) {
399 StoreInst *SI = cast<StoreInst>(AI->use_back());
400 SI->eraseFromParent();
404 if (AST) AST->deleteValue(AI);
405 AI->eraseFromParent();
408 // The alloca has been processed, move on.
409 RemoveFromAllocasList(AllocaNum);
416 // If we haven't computed a numbering for the BB's in the function, do so
418 if (BBNumbers.empty()) {
420 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
424 // If we have an AST to keep updated, remember some pointer value that is
425 // stored into the alloca.
427 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
429 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
430 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
432 // At this point, we're committed to promoting the alloca using IDF's, and
433 // the standard SSA construction algorithm. Determine which blocks need PHI
434 // nodes and see if we can optimize out some work by avoiding insertion of
436 DetermineInsertionPoint(AI, AllocaNum, Info);
440 return; // All of the allocas must have been trivial!
445 // Set the incoming values for the basic block to be null values for all of
446 // the alloca's. We do this in case there is a load of a value that has not
447 // been stored yet. In this case, it will get this null value.
449 RenamePassData::ValVector Values(Allocas.size());
450 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
451 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
453 // Walks all basic blocks in the function performing the SSA rename algorithm
454 // and inserting the phi nodes we marked as necessary
456 std::vector<RenamePassData> RenamePassWorkList;
457 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
458 while (!RenamePassWorkList.empty()) {
460 RPD.swap(RenamePassWorkList.back());
461 RenamePassWorkList.pop_back();
462 // RenamePass may add new worklist entries.
463 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
466 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
469 // Remove the allocas themselves from the function.
470 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
471 Instruction *A = Allocas[i];
473 // If there are any uses of the alloca instructions left, they must be in
474 // sections of dead code that were not processed on the dominance frontier.
475 // Just delete the users now.
478 A->replaceAllUsesWith(UndefValue::get(A->getType()));
479 if (AST) AST->deleteValue(A);
480 A->eraseFromParent();
484 // Loop over all of the PHI nodes and see if there are any that we can get
485 // rid of because they merge all of the same incoming values. This can
486 // happen due to undef values coming into the PHI nodes. This process is
487 // iterative, because eliminating one PHI node can cause others to be removed.
488 bool EliminatedAPHI = true;
489 while (EliminatedAPHI) {
490 EliminatedAPHI = false;
492 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
493 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
494 PHINode *PN = I->second;
496 // If this PHI node merges one value and/or undefs, get the value.
497 if (Value *V = PN->hasConstantValue(true)) {
498 if (!isa<Instruction>(V) ||
499 properlyDominates(cast<Instruction>(V), PN)) {
500 if (AST && isa<PointerType>(PN->getType()))
501 AST->deleteValue(PN);
502 PN->replaceAllUsesWith(V);
503 PN->eraseFromParent();
504 NewPhiNodes.erase(I++);
505 EliminatedAPHI = true;
513 // At this point, the renamer has added entries to PHI nodes for all reachable
514 // code. Unfortunately, there may be unreachable blocks which the renamer
515 // hasn't traversed. If this is the case, the PHI nodes may not
516 // have incoming values for all predecessors. Loop over all PHI nodes we have
517 // created, inserting undef values if they are missing any incoming values.
519 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
520 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
521 // We want to do this once per basic block. As such, only process a block
522 // when we find the PHI that is the first entry in the block.
523 PHINode *SomePHI = I->second;
524 BasicBlock *BB = SomePHI->getParent();
525 if (&BB->front() != SomePHI)
528 // Only do work here if there the PHI nodes are missing incoming values. We
529 // know that all PHI nodes that were inserted in a block will have the same
530 // number of incoming values, so we can just check any of them.
531 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
534 // Get the preds for BB.
535 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
537 // Ok, now we know that all of the PHI nodes are missing entries for some
538 // basic blocks. Start by sorting the incoming predecessors for efficient
540 std::sort(Preds.begin(), Preds.end());
542 // Now we loop through all BB's which have entries in SomePHI and remove
543 // them from the Preds list.
544 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
545 // Do a log(n) search of the Preds list for the entry we want.
546 SmallVector<BasicBlock*, 16>::iterator EntIt =
547 std::lower_bound(Preds.begin(), Preds.end(),
548 SomePHI->getIncomingBlock(i));
549 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
550 "PHI node has entry for a block which is not a predecessor!");
556 // At this point, the blocks left in the preds list must have dummy
557 // entries inserted into every PHI nodes for the block. Update all the phi
558 // nodes in this block that we are inserting (there could be phis before
560 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
561 BasicBlock::iterator BBI = BB->begin();
562 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
563 SomePHI->getNumIncomingValues() == NumBadPreds) {
564 Value *UndefVal = UndefValue::get(SomePHI->getType());
565 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
566 SomePHI->addIncoming(UndefVal, Preds[pred]);
574 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
575 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
576 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
578 void PromoteMem2Reg::
579 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
580 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
581 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
583 // To determine liveness, we must iterate through the predecessors of blocks
584 // where the def is live. Blocks are added to the worklist if we need to
585 // check their predecessors. Start with all the using blocks.
586 SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
587 LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
588 Info.UsingBlocks.begin(), Info.UsingBlocks.end());
590 // If any of the using blocks is also a definition block, check to see if the
591 // definition occurs before or after the use. If it happens before the use,
592 // the value isn't really live-in.
593 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
594 BasicBlock *BB = LiveInBlockWorklist[i];
595 if (!DefBlocks.count(BB)) continue;
597 // Okay, this is a block that both uses and defines the value. If the first
598 // reference to the alloca is a def (store), then we know it isn't live-in.
599 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
600 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
601 if (SI->getOperand(1) != AI) continue;
603 // We found a store to the alloca before a load. The alloca is not
604 // actually live-in here.
605 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
606 LiveInBlockWorklist.pop_back();
609 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
610 if (LI->getOperand(0) != AI) continue;
612 // Okay, we found a load before a store to the alloca. It is actually
613 // live into this block.
619 // Now that we have a set of blocks where the phi is live-in, recursively add
620 // their predecessors until we find the full region the value is live.
621 while (!LiveInBlockWorklist.empty()) {
622 BasicBlock *BB = LiveInBlockWorklist.back();
623 LiveInBlockWorklist.pop_back();
625 // The block really is live in here, insert it into the set. If already in
626 // the set, then it has already been processed.
627 if (!LiveInBlocks.insert(BB))
630 // Since the value is live into BB, it is either defined in a predecessor or
631 // live into it to. Add the preds to the worklist unless they are a
633 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
636 // The value is not live into a predecessor if it defines the value.
637 if (DefBlocks.count(P))
640 // Otherwise it is, add to the worklist.
641 LiveInBlockWorklist.push_back(P);
646 /// DetermineInsertionPoint - At this point, we're committed to promoting the
647 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
648 /// which blocks need phi nodes and see if we can optimize out some work by
649 /// avoiding insertion of dead phi nodes.
650 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
653 // Unique the set of defining blocks for efficient lookup.
654 SmallPtrSet<BasicBlock*, 32> DefBlocks;
655 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
657 // Determine which blocks the value is live in. These are blocks which lead
659 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
660 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
662 // Compute the locations where PhiNodes need to be inserted. Look at the
663 // dominance frontier of EACH basic-block we have a write in.
664 unsigned CurrentVersion = 0;
665 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
666 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
667 while (!Info.DefiningBlocks.empty()) {
668 BasicBlock *BB = Info.DefiningBlocks.back();
669 Info.DefiningBlocks.pop_back();
671 // Look up the DF for this write, add it to defining blocks.
672 DominanceFrontier::const_iterator it = DF.find(BB);
673 if (it == DF.end()) continue;
675 const DominanceFrontier::DomSetType &S = it->second;
677 // In theory we don't need the indirection through the DFBlocks vector.
678 // In practice, the order of calling QueuePhiNode would depend on the
679 // (unspecified) ordering of basic blocks in the dominance frontier,
680 // which would give PHI nodes non-determinstic subscripts. Fix this by
681 // processing blocks in order of the occurance in the function.
682 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
683 PE = S.end(); P != PE; ++P) {
684 // If the frontier block is not in the live-in set for the alloca, don't
685 // bother processing it.
686 if (!LiveInBlocks.count(*P))
689 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
692 // Sort by which the block ordering in the function.
693 if (DFBlocks.size() > 1)
694 std::sort(DFBlocks.begin(), DFBlocks.end());
696 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
697 BasicBlock *BB = DFBlocks[i].second;
698 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
699 Info.DefiningBlocks.push_back(BB);
705 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
706 /// replace any loads of it that are directly dominated by the definition with
707 /// the value stored.
708 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
710 LargeBlockInfo &LBI) {
711 StoreInst *OnlyStore = Info.OnlyStore;
712 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
713 BasicBlock *StoreBB = OnlyStore->getParent();
716 // Clear out UsingBlocks. We will reconstruct it here if needed.
717 Info.UsingBlocks.clear();
719 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
720 Instruction *UserInst = cast<Instruction>(*UI++);
721 if (!isa<LoadInst>(UserInst)) {
722 assert(UserInst == OnlyStore && "Should only have load/stores");
725 LoadInst *LI = cast<LoadInst>(UserInst);
727 // Okay, if we have a load from the alloca, we want to replace it with the
728 // only value stored to the alloca. We can do this if the value is
729 // dominated by the store. If not, we use the rest of the mem2reg machinery
730 // to insert the phi nodes as needed.
731 if (!StoringGlobalVal) { // Non-instructions are always dominated.
732 if (LI->getParent() == StoreBB) {
733 // If we have a use that is in the same block as the store, compare the
734 // indices of the two instructions to see which one came first. If the
735 // load came before the store, we can't handle it.
736 if (StoreIndex == -1)
737 StoreIndex = LBI.getInstructionIndex(OnlyStore);
739 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
740 // Can't handle this load, bail out.
741 Info.UsingBlocks.push_back(StoreBB);
745 } else if (LI->getParent() != StoreBB &&
746 !dominates(StoreBB, LI->getParent())) {
747 // If the load and store are in different blocks, use BB dominance to
748 // check their relationships. If the store doesn't dom the use, bail
750 Info.UsingBlocks.push_back(LI->getParent());
755 // Otherwise, we *can* safely rewrite this load.
756 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
757 if (AST && isa<PointerType>(LI->getType()))
758 AST->deleteValue(LI);
759 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;
774 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
775 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
776 /// potentially useless PHI nodes by just performing a single linear pass over
777 /// the basic block using the Alloca.
779 /// If we cannot promote this alloca (because it is read before it is written),
780 /// return true. This is necessary in cases where, due to control flow, the
781 /// alloca is potentially undefined on some control flow paths. e.g. code like
782 /// this is potentially correct:
784 /// for (...) { if (c) { A = undef; undef = B; } }
786 /// ... so long as A is not used before undef is set.
788 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
789 LargeBlockInfo &LBI) {
790 // The trickiest case to handle is when we have large blocks. Because of this,
791 // this code is optimized assuming that large blocks happen. This does not
792 // significantly pessimize the small block case. This uses LargeBlockInfo to
793 // make it efficient to get the index of various operations in the block.
795 // Clear out UsingBlocks. We will reconstruct it here if needed.
796 Info.UsingBlocks.clear();
798 // Walk the use-def list of the alloca, getting the locations of all stores.
799 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
800 StoresByIndexTy StoresByIndex;
802 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
804 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
805 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
807 // If there are no stores to the alloca, just replace any loads with undef.
808 if (StoresByIndex.empty()) {
809 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
810 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
811 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
812 if (AST && isa<PointerType>(LI->getType()))
813 AST->deleteValue(LI);
815 LI->eraseFromParent();
820 // Sort the stores by their index, making it efficient to do a lookup with a
822 std::sort(StoresByIndex.begin(), StoresByIndex.end());
824 // Walk all of the loads from this alloca, replacing them with the nearest
825 // store above them, if any.
826 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
827 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
830 unsigned LoadIdx = LBI.getInstructionIndex(LI);
832 // Find the nearest store that has a lower than this load.
833 StoresByIndexTy::iterator I =
834 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
835 std::pair<unsigned, StoreInst*>(LoadIdx, 0),
836 StoreIndexSearchPredicate());
838 // If there is no store before this load, then we can't promote this load.
839 if (I == StoresByIndex.begin()) {
840 // Can't handle this load, bail out.
841 Info.UsingBlocks.push_back(LI->getParent());
845 // Otherwise, there was a store before this load, the load takes its value.
847 LI->replaceAllUsesWith(I->second->getOperand(0));
848 if (AST && isa<PointerType>(LI->getType()))
849 AST->deleteValue(LI);
850 LI->eraseFromParent();
856 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
857 // Alloca returns true if there wasn't already a phi-node for that variable
859 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
861 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
862 // Look up the basic-block in question.
863 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
865 // If the BB already has a phi node added for the i'th alloca then we're done!
866 if (PN) return false;
868 // Create a PhiNode using the dereferenced type... and add the phi-node to the
870 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
871 Allocas[AllocaNo]->getName() + "." +
872 utostr(Version++), BB->begin());
874 PhiToAllocaMap[PN] = AllocaNo;
875 PN->reserveOperandSpace(getNumPreds(BB));
877 InsertedPHINodes.insert(PN);
879 if (AST && isa<PointerType>(PN->getType()))
880 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
885 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
886 // stores to the allocas which we are promoting. IncomingVals indicates what
887 // value each Alloca contains on exit from the predecessor block Pred.
889 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
890 RenamePassData::ValVector &IncomingVals,
891 std::vector<RenamePassData> &Worklist) {
893 // If we are inserting any phi nodes into this BB, they will already be in the
895 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
896 // Pred may have multiple edges to BB. If so, we want to add N incoming
897 // values to each PHI we are inserting on the first time we see the edge.
898 // Check to see if APN already has incoming values from Pred. This also
899 // prevents us from modifying PHI nodes that are not currently being
901 bool HasPredEntries = false;
902 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
903 if (APN->getIncomingBlock(i) == Pred) {
904 HasPredEntries = true;
909 // If we have PHI nodes to update, compute the number of edges from Pred to
911 if (!HasPredEntries) {
912 // We want to be able to distinguish between PHI nodes being inserted by
913 // this invocation of mem2reg from those phi nodes that already existed in
914 // the IR before mem2reg was run. We determine that APN is being inserted
915 // because it is missing incoming edges. All other PHI nodes being
916 // inserted by this pass of mem2reg will have the same number of incoming
917 // operands so far. Remember this count.
918 unsigned NewPHINumOperands = APN->getNumOperands();
920 unsigned NumEdges = 0;
921 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
924 assert(NumEdges && "Must be at least one edge from Pred to BB!");
926 // Add entries for all the phis.
927 BasicBlock::iterator PNI = BB->begin();
929 unsigned AllocaNo = PhiToAllocaMap[APN];
931 // Add N incoming values to the PHI node.
932 for (unsigned i = 0; i != NumEdges; ++i)
933 APN->addIncoming(IncomingVals[AllocaNo], Pred);
935 // The currently active variable for this block is now the PHI.
936 IncomingVals[AllocaNo] = APN;
938 // Get the next phi node.
940 APN = dyn_cast<PHINode>(PNI);
943 // Verify that it is missing entries. If not, it is not being inserted
944 // by this mem2reg invocation so we want to ignore it.
945 } while (APN->getNumOperands() == NewPHINumOperands);
949 // Don't revisit blocks.
950 if (!Visited.insert(BB)) return;
952 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
953 Instruction *I = II++; // get the instruction, increment iterator
955 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
956 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
959 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
960 if (AI == AllocaLookup.end()) continue;
962 Value *V = IncomingVals[AI->second];
964 // Anything using the load now uses the current value.
965 LI->replaceAllUsesWith(V);
966 if (AST && isa<PointerType>(LI->getType()))
967 AST->deleteValue(LI);
968 BB->getInstList().erase(LI);
969 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
970 // Delete this instruction and mark the name as the current holder of the
972 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
975 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
976 if (ai == AllocaLookup.end())
979 // what value were we writing?
980 IncomingVals[ai->second] = SI->getOperand(0);
981 BB->getInstList().erase(SI);
985 // 'Recurse' to our successors.
986 succ_iterator I = succ_begin(BB), E = succ_end(BB);
989 // Handle the last successor without using the worklist. This allows us to
990 // handle unconditional branches directly, for example.
993 Worklist.push_back(RenamePassData(*I, BB, IncomingVals));
1000 /// PromoteMemToReg - Promote the specified list of alloca instructions into
1001 /// scalar registers, inserting PHI nodes as appropriate. This function makes
1002 /// use of DominanceFrontier information. This function does not modify the CFG
1003 /// of the function at all. All allocas must be from the same function.
1005 /// If AST is specified, the specified tracker is updated to reflect changes
1008 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1009 DominatorTree &DT, DominanceFrontier &DF,
1010 AliasSetTracker *AST) {
1011 // If there is nothing to do, bail out...
1012 if (Allocas.empty()) return;
1014 PromoteMem2Reg(Allocas, DT, DF, AST).run();