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/STLExtras.h"
33 #include "llvm/Support/CFG.h"
37 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
38 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
39 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
40 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
44 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
45 typedef std::pair<BasicBlock*, unsigned> EltTy;
46 static inline EltTy getEmptyKey() {
47 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
49 static inline EltTy getTombstoneKey() {
50 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
52 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
53 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
55 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
58 static bool isPod() { return true; }
62 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
63 /// This is true if there are only loads and stores to the alloca.
65 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
66 // FIXME: If the memory unit is of pointer or integer type, we can permit
67 // assignments to subsections of the memory unit.
69 // Only allow direct and non-volatile loads and stores...
70 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
71 UI != UE; ++UI) // Loop over all of the uses of the alloca
72 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
75 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
76 if (SI->getOperand(0) == AI)
77 return false; // Don't allow a store OF the AI, only INTO the AI.
80 } else if (const BitCastInst *BC = dyn_cast<BitCastInst>(*UI)) {
81 // A bitcast that does not feed into debug info inhibits promotion.
82 if (!BC->hasOneUse() || !isa<DbgInfoIntrinsic>(*BC->use_begin()))
84 // If the only use is by debug info, this alloca will not exist in
85 // non-debug code, so don't try to promote; this ensures the same
86 // codegen with debug info. Otherwise, debug info should not
87 // inhibit promotion (but we must examine other uses).
100 // Data package used by RenamePass()
101 class RenamePassData {
103 typedef std::vector<Value *> ValVector;
106 RenamePassData(BasicBlock *B, BasicBlock *P,
107 const ValVector &V) : BB(B), Pred(P), Values(V) {}
112 void swap(RenamePassData &RHS) {
113 std::swap(BB, RHS.BB);
114 std::swap(Pred, RHS.Pred);
115 Values.swap(RHS.Values);
119 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
120 /// load/store instructions in the block that directly load or store an alloca.
122 /// This functionality is important because it avoids scanning large basic
123 /// blocks multiple times when promoting many allocas in the same block.
124 class LargeBlockInfo {
125 /// InstNumbers - For each instruction that we track, keep the index of the
126 /// instruction. The index starts out as the number of the instruction from
127 /// the start of the block.
128 DenseMap<const Instruction *, unsigned> InstNumbers;
131 /// isInterestingInstruction - This code only looks at accesses to allocas.
132 static bool isInterestingInstruction(const Instruction *I) {
133 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
134 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
137 /// getInstructionIndex - Get or calculate the index of the specified
139 unsigned getInstructionIndex(const Instruction *I) {
140 assert(isInterestingInstruction(I) &&
141 "Not a load/store to/from an alloca?");
143 // If we already have this instruction number, return it.
144 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
145 if (It != InstNumbers.end()) return It->second;
147 // Scan the whole block to get the instruction. This accumulates
148 // information for every interesting instruction in the block, in order to
149 // avoid gratuitus rescans.
150 const BasicBlock *BB = I->getParent();
152 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
154 if (isInterestingInstruction(BBI))
155 InstNumbers[BBI] = InstNo++;
156 It = InstNumbers.find(I);
158 assert(It != InstNumbers.end() && "Didn't insert instruction?");
162 void deleteValue(const Instruction *I) {
163 InstNumbers.erase(I);
171 struct PromoteMem2Reg {
172 /// Allocas - The alloca instructions being promoted.
174 std::vector<AllocaInst*> Allocas;
176 DominanceFrontier &DF;
178 /// AST - An AliasSetTracker object to update. If null, don't update it.
180 AliasSetTracker *AST;
182 /// AllocaLookup - Reverse mapping of Allocas.
184 std::map<AllocaInst*, unsigned> AllocaLookup;
186 /// NewPhiNodes - The PhiNodes we're adding.
188 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
190 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
191 /// it corresponds to.
192 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
194 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
195 /// each alloca that is of pointer type, we keep track of what to copyValue
196 /// to the inserted PHI nodes here.
198 std::vector<Value*> PointerAllocaValues;
200 /// Visited - The set of basic blocks the renamer has already visited.
202 SmallPtrSet<BasicBlock*, 16> Visited;
204 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
205 /// non-determinstic behavior.
206 DenseMap<BasicBlock*, unsigned> BBNumbers;
208 /// BBNumPreds - Lazily compute the number of predecessors a block has.
209 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
211 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
212 DominanceFrontier &df, AliasSetTracker *ast)
213 : Allocas(A), DT(dt), DF(df), AST(ast) {}
217 /// properlyDominates - Return true if I1 properly dominates I2.
219 bool properlyDominates(Instruction *I1, Instruction *I2) const {
220 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
221 I1 = II->getNormalDest()->begin();
222 return DT.properlyDominates(I1->getParent(), I2->getParent());
225 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
227 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
228 return DT.dominates(BB1, BB2);
232 void RemoveFromAllocasList(unsigned &AllocaIdx) {
233 Allocas[AllocaIdx] = Allocas.back();
238 unsigned getNumPreds(const BasicBlock *BB) {
239 unsigned &NP = BBNumPreds[BB];
241 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
245 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
247 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
248 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
249 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
251 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
252 LargeBlockInfo &LBI);
253 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
254 LargeBlockInfo &LBI);
257 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
258 RenamePassData::ValVector &IncVals,
259 std::vector<RenamePassData> &Worklist);
260 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
261 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
265 std::vector<BasicBlock*> DefiningBlocks;
266 std::vector<BasicBlock*> UsingBlocks;
268 StoreInst *OnlyStore;
269 BasicBlock *OnlyBlock;
270 bool OnlyUsedInOneBlock;
272 Value *AllocaPointerVal;
275 DefiningBlocks.clear();
279 OnlyUsedInOneBlock = true;
280 AllocaPointerVal = 0;
283 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
285 void AnalyzeAlloca(AllocaInst *AI) {
288 // As we scan the uses of the alloca instruction, keep track of stores,
289 // and decide whether all of the loads and stores to the alloca are within
290 // the same basic block.
291 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
293 Instruction *User = cast<Instruction>(*UI++);
294 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
295 // Remove any uses of this alloca in DbgInfoInstrinsics.
296 assert(BC->hasOneUse() && "Unexpected alloca uses!");
297 DbgInfoIntrinsic *DI = cast<DbgInfoIntrinsic>(*BC->use_begin());
298 DI->eraseFromParent();
299 BC->eraseFromParent();
303 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
304 // Remember the basic blocks which define new values for the alloca
305 DefiningBlocks.push_back(SI->getParent());
306 AllocaPointerVal = SI->getOperand(0);
309 LoadInst *LI = cast<LoadInst>(User);
310 // Otherwise it must be a load instruction, keep track of variable
312 UsingBlocks.push_back(LI->getParent());
313 AllocaPointerVal = LI;
316 if (OnlyUsedInOneBlock) {
318 OnlyBlock = User->getParent();
319 else if (OnlyBlock != User->getParent())
320 OnlyUsedInOneBlock = false;
325 } // end of anonymous namespace
328 void PromoteMem2Reg::run() {
329 Function &F = *DF.getRoot()->getParent();
331 if (AST) PointerAllocaValues.resize(Allocas.size());
336 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
337 AllocaInst *AI = Allocas[AllocaNum];
339 assert(isAllocaPromotable(AI) &&
340 "Cannot promote non-promotable alloca!");
341 assert(AI->getParent()->getParent() == &F &&
342 "All allocas should be in the same function, which is same as DF!");
344 if (AI->use_empty()) {
345 // If there are no uses of the alloca, just delete it now.
346 if (AST) AST->deleteValue(AI);
347 AI->eraseFromParent();
349 // Remove the alloca from the Allocas list, since it has been processed
350 RemoveFromAllocasList(AllocaNum);
355 // Calculate the set of read and write-locations for each alloca. This is
356 // analogous to finding the 'uses' and 'definitions' of each variable.
357 Info.AnalyzeAlloca(AI);
359 // If there is only a single store to this value, replace any loads of
360 // it that are directly dominated by the definition with the value stored.
361 if (Info.DefiningBlocks.size() == 1) {
362 RewriteSingleStoreAlloca(AI, Info, LBI);
364 // Finally, after the scan, check to see if the store is all that is left.
365 if (Info.UsingBlocks.empty()) {
366 // Remove the (now dead) store and alloca.
367 Info.OnlyStore->eraseFromParent();
368 LBI.deleteValue(Info.OnlyStore);
370 if (AST) AST->deleteValue(AI);
371 AI->eraseFromParent();
374 // The alloca has been processed, move on.
375 RemoveFromAllocasList(AllocaNum);
382 // If the alloca is only read and written in one basic block, just perform a
383 // linear sweep over the block to eliminate it.
384 if (Info.OnlyUsedInOneBlock) {
385 PromoteSingleBlockAlloca(AI, Info, LBI);
387 // Finally, after the scan, check to see if the stores are all that is
389 if (Info.UsingBlocks.empty()) {
391 // Remove the (now dead) stores and alloca.
392 while (!AI->use_empty()) {
393 StoreInst *SI = cast<StoreInst>(AI->use_back());
394 SI->eraseFromParent();
398 if (AST) AST->deleteValue(AI);
399 AI->eraseFromParent();
402 // The alloca has been processed, move on.
403 RemoveFromAllocasList(AllocaNum);
410 // If we haven't computed a numbering for the BB's in the function, do so
412 if (BBNumbers.empty()) {
414 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
418 // If we have an AST to keep updated, remember some pointer value that is
419 // stored into the alloca.
421 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
423 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
424 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
426 // At this point, we're committed to promoting the alloca using IDF's, and
427 // the standard SSA construction algorithm. Determine which blocks need PHI
428 // nodes and see if we can optimize out some work by avoiding insertion of
430 DetermineInsertionPoint(AI, AllocaNum, Info);
434 return; // All of the allocas must have been trivial!
439 // Set the incoming values for the basic block to be null values for all of
440 // the alloca's. We do this in case there is a load of a value that has not
441 // been stored yet. In this case, it will get this null value.
443 RenamePassData::ValVector Values(Allocas.size());
444 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
445 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
447 // Walks all basic blocks in the function performing the SSA rename algorithm
448 // and inserting the phi nodes we marked as necessary
450 std::vector<RenamePassData> RenamePassWorkList;
451 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
452 while (!RenamePassWorkList.empty()) {
454 RPD.swap(RenamePassWorkList.back());
455 RenamePassWorkList.pop_back();
456 // RenamePass may add new worklist entries.
457 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
460 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
463 // Remove the allocas themselves from the function.
464 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
465 Instruction *A = Allocas[i];
467 // If there are any uses of the alloca instructions left, they must be in
468 // sections of dead code that were not processed on the dominance frontier.
469 // Just delete the users now.
472 A->replaceAllUsesWith(UndefValue::get(A->getType()));
473 if (AST) AST->deleteValue(A);
474 A->eraseFromParent();
478 // Loop over all of the PHI nodes and see if there are any that we can get
479 // rid of because they merge all of the same incoming values. This can
480 // happen due to undef values coming into the PHI nodes. This process is
481 // iterative, because eliminating one PHI node can cause others to be removed.
482 bool EliminatedAPHI = true;
483 while (EliminatedAPHI) {
484 EliminatedAPHI = false;
486 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
487 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
488 PHINode *PN = I->second;
490 // If this PHI node merges one value and/or undefs, get the value.
491 if (Value *V = PN->hasConstantValue(&DT)) {
492 if (AST && isa<PointerType>(PN->getType()))
493 AST->deleteValue(PN);
494 PN->replaceAllUsesWith(V);
495 PN->eraseFromParent();
496 NewPhiNodes.erase(I++);
497 EliminatedAPHI = true;
504 // At this point, the renamer has added entries to PHI nodes for all reachable
505 // code. Unfortunately, there may be unreachable blocks which the renamer
506 // hasn't traversed. If this is the case, the PHI nodes may not
507 // have incoming values for all predecessors. Loop over all PHI nodes we have
508 // created, inserting undef values if they are missing any incoming values.
510 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
511 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
512 // We want to do this once per basic block. As such, only process a block
513 // when we find the PHI that is the first entry in the block.
514 PHINode *SomePHI = I->second;
515 BasicBlock *BB = SomePHI->getParent();
516 if (&BB->front() != SomePHI)
519 // Only do work here if there the PHI nodes are missing incoming values. We
520 // know that all PHI nodes that were inserted in a block will have the same
521 // number of incoming values, so we can just check any of them.
522 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
525 // Get the preds for BB.
526 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
528 // Ok, now we know that all of the PHI nodes are missing entries for some
529 // basic blocks. Start by sorting the incoming predecessors for efficient
531 std::sort(Preds.begin(), Preds.end());
533 // Now we loop through all BB's which have entries in SomePHI and remove
534 // them from the Preds list.
535 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
536 // Do a log(n) search of the Preds list for the entry we want.
537 SmallVector<BasicBlock*, 16>::iterator EntIt =
538 std::lower_bound(Preds.begin(), Preds.end(),
539 SomePHI->getIncomingBlock(i));
540 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
541 "PHI node has entry for a block which is not a predecessor!");
547 // At this point, the blocks left in the preds list must have dummy
548 // entries inserted into every PHI nodes for the block. Update all the phi
549 // nodes in this block that we are inserting (there could be phis before
551 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
552 BasicBlock::iterator BBI = BB->begin();
553 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
554 SomePHI->getNumIncomingValues() == NumBadPreds) {
555 Value *UndefVal = UndefValue::get(SomePHI->getType());
556 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
557 SomePHI->addIncoming(UndefVal, Preds[pred]);
565 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
566 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
567 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
569 void PromoteMem2Reg::
570 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
571 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
572 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
574 // To determine liveness, we must iterate through the predecessors of blocks
575 // where the def is live. Blocks are added to the worklist if we need to
576 // check their predecessors. Start with all the using blocks.
577 SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
578 LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
579 Info.UsingBlocks.begin(), Info.UsingBlocks.end());
581 // If any of the using blocks is also a definition block, check to see if the
582 // definition occurs before or after the use. If it happens before the use,
583 // the value isn't really live-in.
584 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
585 BasicBlock *BB = LiveInBlockWorklist[i];
586 if (!DefBlocks.count(BB)) continue;
588 // Okay, this is a block that both uses and defines the value. If the first
589 // reference to the alloca is a def (store), then we know it isn't live-in.
590 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
591 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
592 if (SI->getOperand(1) != AI) continue;
594 // We found a store to the alloca before a load. The alloca is not
595 // actually live-in here.
596 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
597 LiveInBlockWorklist.pop_back();
602 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
603 if (LI->getOperand(0) != AI) continue;
605 // Okay, we found a load before a store to the alloca. It is actually
606 // live into this block.
612 // Now that we have a set of blocks where the phi is live-in, recursively add
613 // their predecessors until we find the full region the value is live.
614 while (!LiveInBlockWorklist.empty()) {
615 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
617 // The block really is live in here, insert it into the set. If already in
618 // the set, then it has already been processed.
619 if (!LiveInBlocks.insert(BB))
622 // Since the value is live into BB, it is either defined in a predecessor or
623 // live into it to. Add the preds to the worklist unless they are a
625 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
628 // The value is not live into a predecessor if it defines the value.
629 if (DefBlocks.count(P))
632 // Otherwise it is, add to the worklist.
633 LiveInBlockWorklist.push_back(P);
638 /// DetermineInsertionPoint - At this point, we're committed to promoting the
639 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
640 /// which blocks need phi nodes and see if we can optimize out some work by
641 /// avoiding insertion of dead phi nodes.
642 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
645 // Unique the set of defining blocks for efficient lookup.
646 SmallPtrSet<BasicBlock*, 32> DefBlocks;
647 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
649 // Determine which blocks the value is live in. These are blocks which lead
651 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
652 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
654 // Compute the locations where PhiNodes need to be inserted. Look at the
655 // dominance frontier of EACH basic-block we have a write in.
656 unsigned CurrentVersion = 0;
657 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
658 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
659 while (!Info.DefiningBlocks.empty()) {
660 BasicBlock *BB = Info.DefiningBlocks.back();
661 Info.DefiningBlocks.pop_back();
663 // Look up the DF for this write, add it to defining blocks.
664 DominanceFrontier::const_iterator it = DF.find(BB);
665 if (it == DF.end()) continue;
667 const DominanceFrontier::DomSetType &S = it->second;
669 // In theory we don't need the indirection through the DFBlocks vector.
670 // In practice, the order of calling QueuePhiNode would depend on the
671 // (unspecified) ordering of basic blocks in the dominance frontier,
672 // which would give PHI nodes non-determinstic subscripts. Fix this by
673 // processing blocks in order of the occurance in the function.
674 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
675 PE = S.end(); P != PE; ++P) {
676 // If the frontier block is not in the live-in set for the alloca, don't
677 // bother processing it.
678 if (!LiveInBlocks.count(*P))
681 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
684 // Sort by which the block ordering in the function.
685 if (DFBlocks.size() > 1)
686 std::sort(DFBlocks.begin(), DFBlocks.end());
688 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
689 BasicBlock *BB = DFBlocks[i].second;
690 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
691 Info.DefiningBlocks.push_back(BB);
697 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
698 /// replace any loads of it that are directly dominated by the definition with
699 /// the value stored.
700 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
702 LargeBlockInfo &LBI) {
703 StoreInst *OnlyStore = Info.OnlyStore;
704 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
705 BasicBlock *StoreBB = OnlyStore->getParent();
708 // Clear out UsingBlocks. We will reconstruct it here if needed.
709 Info.UsingBlocks.clear();
711 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
712 Instruction *UserInst = cast<Instruction>(*UI++);
713 if (!isa<LoadInst>(UserInst)) {
714 assert(UserInst == OnlyStore && "Should only have load/stores");
717 LoadInst *LI = cast<LoadInst>(UserInst);
719 // Okay, if we have a load from the alloca, we want to replace it with the
720 // only value stored to the alloca. We can do this if the value is
721 // dominated by the store. If not, we use the rest of the mem2reg machinery
722 // to insert the phi nodes as needed.
723 if (!StoringGlobalVal) { // Non-instructions are always dominated.
724 if (LI->getParent() == StoreBB) {
725 // If we have a use that is in the same block as the store, compare the
726 // indices of the two instructions to see which one came first. If the
727 // load came before the store, we can't handle it.
728 if (StoreIndex == -1)
729 StoreIndex = LBI.getInstructionIndex(OnlyStore);
731 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
732 // Can't handle this load, bail out.
733 Info.UsingBlocks.push_back(StoreBB);
737 } else if (LI->getParent() != StoreBB &&
738 !dominates(StoreBB, LI->getParent())) {
739 // If the load and store are in different blocks, use BB dominance to
740 // check their relationships. If the store doesn't dom the use, bail
742 Info.UsingBlocks.push_back(LI->getParent());
747 // Otherwise, we *can* safely rewrite this load.
748 Value *ReplVal = OnlyStore->getOperand(0);
749 // If the replacement value is the load, this must occur in unreachable
752 ReplVal = UndefValue::get(LI->getType());
753 LI->replaceAllUsesWith(ReplVal);
754 if (AST && isa<PointerType>(LI->getType()))
755 AST->deleteValue(LI);
756 LI->eraseFromParent();
763 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
764 /// first element of a pair.
765 struct StoreIndexSearchPredicate {
766 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
767 const std::pair<unsigned, StoreInst*> &RHS) {
768 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() + "." + Twine(Version++),
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 // If we have PHI nodes to update, compute the number of edges from Pred to
898 if (PhiToAllocaMap.count(APN)) {
899 // We want to be able to distinguish between PHI nodes being inserted by
900 // this invocation of mem2reg from those phi nodes that already existed in
901 // the IR before mem2reg was run. We determine that APN is being inserted
902 // because it is missing incoming edges. All other PHI nodes being
903 // inserted by this pass of mem2reg will have the same number of incoming
904 // operands so far. Remember this count.
905 unsigned NewPHINumOperands = APN->getNumOperands();
907 unsigned NumEdges = 0;
908 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
911 assert(NumEdges && "Must be at least one edge from Pred to BB!");
913 // Add entries for all the phis.
914 BasicBlock::iterator PNI = BB->begin();
916 unsigned AllocaNo = PhiToAllocaMap[APN];
918 // Add N incoming values to the PHI node.
919 for (unsigned i = 0; i != NumEdges; ++i)
920 APN->addIncoming(IncomingVals[AllocaNo], Pred);
922 // The currently active variable for this block is now the PHI.
923 IncomingVals[AllocaNo] = APN;
925 // Get the next phi node.
927 APN = dyn_cast<PHINode>(PNI);
930 // Verify that it is missing entries. If not, it is not being inserted
931 // by this mem2reg invocation so we want to ignore it.
932 } while (APN->getNumOperands() == NewPHINumOperands);
936 // Don't revisit blocks.
937 if (!Visited.insert(BB)) return;
939 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
940 Instruction *I = II++; // get the instruction, increment iterator
942 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
943 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
946 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
947 if (AI == AllocaLookup.end()) continue;
949 Value *V = IncomingVals[AI->second];
951 // Anything using the load now uses the current value.
952 LI->replaceAllUsesWith(V);
953 if (AST && isa<PointerType>(LI->getType()))
954 AST->deleteValue(LI);
955 BB->getInstList().erase(LI);
956 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
957 // Delete this instruction and mark the name as the current holder of the
959 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
962 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
963 if (ai == AllocaLookup.end())
966 // what value were we writing?
967 IncomingVals[ai->second] = SI->getOperand(0);
968 BB->getInstList().erase(SI);
972 // 'Recurse' to our successors.
973 succ_iterator I = succ_begin(BB), E = succ_end(BB);
976 // Keep track of the successors so we don't visit the same successor twice
977 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
979 // Handle the first successor without using the worklist.
980 VisitedSuccs.insert(*I);
986 if (VisitedSuccs.insert(*I))
987 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
992 /// PromoteMemToReg - Promote the specified list of alloca instructions into
993 /// scalar registers, inserting PHI nodes as appropriate. This function makes
994 /// use of DominanceFrontier information. This function does not modify the CFG
995 /// of the function at all. All allocas must be from the same function.
997 /// If AST is specified, the specified tracker is updated to reflect changes
1000 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1001 DominatorTree &DT, DominanceFrontier &DF,
1002 AliasSetTracker *AST) {
1003 // If there is nothing to do, bail out...
1004 if (Allocas.empty()) return;
1006 PromoteMem2Reg(Allocas, DT, DF, AST).run();