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) {
61 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
62 /// This is true if there are only loads and stores to the alloca.
64 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
65 // FIXME: If the memory unit is of pointer or integer type, we can permit
66 // assignments to subsections of the memory unit.
68 // Only allow direct and non-volatile loads and stores...
69 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
70 UI != UE; ++UI) // Loop over all of the uses of the alloca
71 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
74 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
75 if (SI->getOperand(0) == AI)
76 return false; // Don't allow a store OF the AI, only INTO the AI.
79 } else if (const BitCastInst *BC = dyn_cast<BitCastInst>(*UI)) {
80 // A bitcast that does not feed into debug info inhibits promotion.
81 if (!BC->hasOneUse() || !isa<DbgInfoIntrinsic>(*BC->use_begin()))
83 // If the only use is by debug info, this alloca will not exist in
84 // non-debug code, so don't try to promote; this ensures the same
85 // codegen with debug info. Otherwise, debug info should not
86 // inhibit promotion (but we must examine other uses).
99 // Data package used by RenamePass()
100 class RenamePassData {
102 typedef std::vector<Value *> ValVector;
104 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
105 RenamePassData(BasicBlock *B, BasicBlock *P,
106 const ValVector &V) : BB(B), Pred(P), Values(V) {}
111 void swap(RenamePassData &RHS) {
112 std::swap(BB, RHS.BB);
113 std::swap(Pred, RHS.Pred);
114 Values.swap(RHS.Values);
118 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
119 /// load/store instructions in the block that directly load or store an alloca.
121 /// This functionality is important because it avoids scanning large basic
122 /// blocks multiple times when promoting many allocas in the same block.
123 class LargeBlockInfo {
124 /// InstNumbers - For each instruction that we track, keep the index of the
125 /// instruction. The index starts out as the number of the instruction from
126 /// the start of the block.
127 DenseMap<const Instruction *, unsigned> InstNumbers;
130 /// isInterestingInstruction - This code only looks at accesses to allocas.
131 static bool isInterestingInstruction(const Instruction *I) {
132 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
133 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
136 /// getInstructionIndex - Get or calculate the index of the specified
138 unsigned getInstructionIndex(const Instruction *I) {
139 assert(isInterestingInstruction(I) &&
140 "Not a load/store to/from an alloca?");
142 // If we already have this instruction number, return it.
143 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
144 if (It != InstNumbers.end()) return It->second;
146 // Scan the whole block to get the instruction. This accumulates
147 // information for every interesting instruction in the block, in order to
148 // avoid gratuitus rescans.
149 const BasicBlock *BB = I->getParent();
151 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
153 if (isInterestingInstruction(BBI))
154 InstNumbers[BBI] = InstNo++;
155 It = InstNumbers.find(I);
157 assert(It != InstNumbers.end() && "Didn't insert instruction?");
161 void deleteValue(const Instruction *I) {
162 InstNumbers.erase(I);
170 struct PromoteMem2Reg {
171 /// Allocas - The alloca instructions being promoted.
173 std::vector<AllocaInst*> Allocas;
175 DominanceFrontier &DF;
177 /// AST - An AliasSetTracker object to update. If null, don't update it.
179 AliasSetTracker *AST;
181 /// AllocaLookup - Reverse mapping of Allocas.
183 std::map<AllocaInst*, unsigned> AllocaLookup;
185 /// NewPhiNodes - The PhiNodes we're adding.
187 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
189 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
190 /// it corresponds to.
191 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
193 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
194 /// each alloca that is of pointer type, we keep track of what to copyValue
195 /// to the inserted PHI nodes here.
197 std::vector<Value*> PointerAllocaValues;
199 /// Visited - The set of basic blocks the renamer has already visited.
201 SmallPtrSet<BasicBlock*, 16> Visited;
203 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
204 /// non-determinstic behavior.
205 DenseMap<BasicBlock*, unsigned> BBNumbers;
207 /// BBNumPreds - Lazily compute the number of predecessors a block has.
208 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
210 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
211 DominanceFrontier &df, AliasSetTracker *ast)
212 : Allocas(A), DT(dt), DF(df), AST(ast) {}
216 /// properlyDominates - Return true if I1 properly dominates I2.
218 bool properlyDominates(Instruction *I1, Instruction *I2) const {
219 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
220 I1 = II->getNormalDest()->begin();
221 return DT.properlyDominates(I1->getParent(), I2->getParent());
224 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
226 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
227 return DT.dominates(BB1, BB2);
231 void RemoveFromAllocasList(unsigned &AllocaIdx) {
232 Allocas[AllocaIdx] = Allocas.back();
237 unsigned getNumPreds(const BasicBlock *BB) {
238 unsigned &NP = BBNumPreds[BB];
240 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
244 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
246 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
247 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
248 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
250 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
251 LargeBlockInfo &LBI);
252 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
253 LargeBlockInfo &LBI);
256 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
257 RenamePassData::ValVector &IncVals,
258 std::vector<RenamePassData> &Worklist);
259 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
260 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
264 std::vector<BasicBlock*> DefiningBlocks;
265 std::vector<BasicBlock*> UsingBlocks;
267 StoreInst *OnlyStore;
268 BasicBlock *OnlyBlock;
269 bool OnlyUsedInOneBlock;
271 Value *AllocaPointerVal;
274 DefiningBlocks.clear();
278 OnlyUsedInOneBlock = true;
279 AllocaPointerVal = 0;
282 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
284 void AnalyzeAlloca(AllocaInst *AI) {
287 // As we scan the uses of the alloca instruction, keep track of stores,
288 // and decide whether all of the loads and stores to the alloca are within
289 // the same basic block.
290 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
292 Instruction *User = cast<Instruction>(*UI++);
293 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
294 // Remove any uses of this alloca in DbgInfoInstrinsics.
295 assert(BC->hasOneUse() && "Unexpected alloca uses!");
296 DbgInfoIntrinsic *DI = cast<DbgInfoIntrinsic>(*BC->use_begin());
297 DI->eraseFromParent();
298 BC->eraseFromParent();
302 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
303 // Remember the basic blocks which define new values for the alloca
304 DefiningBlocks.push_back(SI->getParent());
305 AllocaPointerVal = SI->getOperand(0);
308 LoadInst *LI = cast<LoadInst>(User);
309 // Otherwise it must be a load instruction, keep track of variable
311 UsingBlocks.push_back(LI->getParent());
312 AllocaPointerVal = LI;
315 if (OnlyUsedInOneBlock) {
317 OnlyBlock = User->getParent();
318 else if (OnlyBlock != User->getParent())
319 OnlyUsedInOneBlock = false;
324 } // end of anonymous namespace
327 void PromoteMem2Reg::run() {
328 Function &F = *DF.getRoot()->getParent();
330 if (AST) PointerAllocaValues.resize(Allocas.size());
335 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
336 AllocaInst *AI = Allocas[AllocaNum];
338 assert(isAllocaPromotable(AI) &&
339 "Cannot promote non-promotable alloca!");
340 assert(AI->getParent()->getParent() == &F &&
341 "All allocas should be in the same function, which is same as DF!");
343 if (AI->use_empty()) {
344 // If there are no uses of the alloca, just delete it now.
345 if (AST) AST->deleteValue(AI);
346 AI->eraseFromParent();
348 // Remove the alloca from the Allocas list, since it has been processed
349 RemoveFromAllocasList(AllocaNum);
354 // Calculate the set of read and write-locations for each alloca. This is
355 // analogous to finding the 'uses' and 'definitions' of each variable.
356 Info.AnalyzeAlloca(AI);
358 // If there is only a single store to this value, replace any loads of
359 // it that are directly dominated by the definition with the value stored.
360 if (Info.DefiningBlocks.size() == 1) {
361 RewriteSingleStoreAlloca(AI, Info, LBI);
363 // Finally, after the scan, check to see if the store is all that is left.
364 if (Info.UsingBlocks.empty()) {
365 // Remove the (now dead) store and alloca.
366 Info.OnlyStore->eraseFromParent();
367 LBI.deleteValue(Info.OnlyStore);
369 if (AST) AST->deleteValue(AI);
370 AI->eraseFromParent();
373 // The alloca has been processed, move on.
374 RemoveFromAllocasList(AllocaNum);
381 // If the alloca is only read and written in one basic block, just perform a
382 // linear sweep over the block to eliminate it.
383 if (Info.OnlyUsedInOneBlock) {
384 PromoteSingleBlockAlloca(AI, Info, LBI);
386 // Finally, after the scan, check to see if the stores are all that is
388 if (Info.UsingBlocks.empty()) {
390 // Remove the (now dead) stores and alloca.
391 while (!AI->use_empty()) {
392 StoreInst *SI = cast<StoreInst>(AI->use_back());
393 SI->eraseFromParent();
397 if (AST) AST->deleteValue(AI);
398 AI->eraseFromParent();
401 // The alloca has been processed, move on.
402 RemoveFromAllocasList(AllocaNum);
409 // If we haven't computed a numbering for the BB's in the function, do so
411 if (BBNumbers.empty()) {
413 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
417 // If we have an AST to keep updated, remember some pointer value that is
418 // stored into the alloca.
420 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
422 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
423 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
425 // At this point, we're committed to promoting the alloca using IDF's, and
426 // the standard SSA construction algorithm. Determine which blocks need PHI
427 // nodes and see if we can optimize out some work by avoiding insertion of
429 DetermineInsertionPoint(AI, AllocaNum, Info);
433 return; // All of the allocas must have been trivial!
438 // Set the incoming values for the basic block to be null values for all of
439 // the alloca's. We do this in case there is a load of a value that has not
440 // been stored yet. In this case, it will get this null value.
442 RenamePassData::ValVector Values(Allocas.size());
443 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
444 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
446 // Walks all basic blocks in the function performing the SSA rename algorithm
447 // and inserting the phi nodes we marked as necessary
449 std::vector<RenamePassData> RenamePassWorkList;
450 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
451 while (!RenamePassWorkList.empty()) {
453 RPD.swap(RenamePassWorkList.back());
454 RenamePassWorkList.pop_back();
455 // RenamePass may add new worklist entries.
456 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
459 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
462 // Remove the allocas themselves from the function.
463 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
464 Instruction *A = Allocas[i];
466 // If there are any uses of the alloca instructions left, they must be in
467 // sections of dead code that were not processed on the dominance frontier.
468 // Just delete the users now.
471 A->replaceAllUsesWith(UndefValue::get(A->getType()));
472 if (AST) AST->deleteValue(A);
473 A->eraseFromParent();
477 // Loop over all of the PHI nodes and see if there are any that we can get
478 // rid of because they merge all of the same incoming values. This can
479 // happen due to undef values coming into the PHI nodes. This process is
480 // iterative, because eliminating one PHI node can cause others to be removed.
481 bool EliminatedAPHI = true;
482 while (EliminatedAPHI) {
483 EliminatedAPHI = false;
485 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
486 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
487 PHINode *PN = I->second;
489 // If this PHI node merges one value and/or undefs, get the value.
490 if (Value *V = PN->hasConstantValue(&DT)) {
491 if (AST && isa<PointerType>(PN->getType()))
492 AST->deleteValue(PN);
493 PN->replaceAllUsesWith(V);
494 PN->eraseFromParent();
495 NewPhiNodes.erase(I++);
496 EliminatedAPHI = true;
503 // At this point, the renamer has added entries to PHI nodes for all reachable
504 // code. Unfortunately, there may be unreachable blocks which the renamer
505 // hasn't traversed. If this is the case, the PHI nodes may not
506 // have incoming values for all predecessors. Loop over all PHI nodes we have
507 // created, inserting undef values if they are missing any incoming values.
509 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
510 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
511 // We want to do this once per basic block. As such, only process a block
512 // when we find the PHI that is the first entry in the block.
513 PHINode *SomePHI = I->second;
514 BasicBlock *BB = SomePHI->getParent();
515 if (&BB->front() != SomePHI)
518 // Only do work here if there the PHI nodes are missing incoming values. We
519 // know that all PHI nodes that were inserted in a block will have the same
520 // number of incoming values, so we can just check any of them.
521 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
524 // Get the preds for BB.
525 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
527 // Ok, now we know that all of the PHI nodes are missing entries for some
528 // basic blocks. Start by sorting the incoming predecessors for efficient
530 std::sort(Preds.begin(), Preds.end());
532 // Now we loop through all BB's which have entries in SomePHI and remove
533 // them from the Preds list.
534 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
535 // Do a log(n) search of the Preds list for the entry we want.
536 SmallVector<BasicBlock*, 16>::iterator EntIt =
537 std::lower_bound(Preds.begin(), Preds.end(),
538 SomePHI->getIncomingBlock(i));
539 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
540 "PHI node has entry for a block which is not a predecessor!");
546 // At this point, the blocks left in the preds list must have dummy
547 // entries inserted into every PHI nodes for the block. Update all the phi
548 // nodes in this block that we are inserting (there could be phis before
550 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
551 BasicBlock::iterator BBI = BB->begin();
552 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
553 SomePHI->getNumIncomingValues() == NumBadPreds) {
554 Value *UndefVal = UndefValue::get(SomePHI->getType());
555 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
556 SomePHI->addIncoming(UndefVal, Preds[pred]);
564 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
565 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
566 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
568 void PromoteMem2Reg::
569 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
570 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
571 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
573 // To determine liveness, we must iterate through the predecessors of blocks
574 // where the def is live. Blocks are added to the worklist if we need to
575 // check their predecessors. Start with all the using blocks.
576 SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
577 LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
578 Info.UsingBlocks.begin(), Info.UsingBlocks.end());
580 // If any of the using blocks is also a definition block, check to see if the
581 // definition occurs before or after the use. If it happens before the use,
582 // the value isn't really live-in.
583 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
584 BasicBlock *BB = LiveInBlockWorklist[i];
585 if (!DefBlocks.count(BB)) continue;
587 // Okay, this is a block that both uses and defines the value. If the first
588 // reference to the alloca is a def (store), then we know it isn't live-in.
589 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
590 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
591 if (SI->getOperand(1) != AI) continue;
593 // We found a store to the alloca before a load. The alloca is not
594 // actually live-in here.
595 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
596 LiveInBlockWorklist.pop_back();
601 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
602 if (LI->getOperand(0) != AI) continue;
604 // Okay, we found a load before a store to the alloca. It is actually
605 // live into this block.
611 // Now that we have a set of blocks where the phi is live-in, recursively add
612 // their predecessors until we find the full region the value is live.
613 while (!LiveInBlockWorklist.empty()) {
614 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
616 // The block really is live in here, insert it into the set. If already in
617 // the set, then it has already been processed.
618 if (!LiveInBlocks.insert(BB))
621 // Since the value is live into BB, it is either defined in a predecessor or
622 // live into it to. Add the preds to the worklist unless they are a
624 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
627 // The value is not live into a predecessor if it defines the value.
628 if (DefBlocks.count(P))
631 // Otherwise it is, add to the worklist.
632 LiveInBlockWorklist.push_back(P);
637 /// DetermineInsertionPoint - At this point, we're committed to promoting the
638 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
639 /// which blocks need phi nodes and see if we can optimize out some work by
640 /// avoiding insertion of dead phi nodes.
641 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
644 // Unique the set of defining blocks for efficient lookup.
645 SmallPtrSet<BasicBlock*, 32> DefBlocks;
646 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
648 // Determine which blocks the value is live in. These are blocks which lead
650 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
651 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
653 // Compute the locations where PhiNodes need to be inserted. Look at the
654 // dominance frontier of EACH basic-block we have a write in.
655 unsigned CurrentVersion = 0;
656 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
657 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
658 while (!Info.DefiningBlocks.empty()) {
659 BasicBlock *BB = Info.DefiningBlocks.back();
660 Info.DefiningBlocks.pop_back();
662 // Look up the DF for this write, add it to defining blocks.
663 DominanceFrontier::const_iterator it = DF.find(BB);
664 if (it == DF.end()) continue;
666 const DominanceFrontier::DomSetType &S = it->second;
668 // In theory we don't need the indirection through the DFBlocks vector.
669 // In practice, the order of calling QueuePhiNode would depend on the
670 // (unspecified) ordering of basic blocks in the dominance frontier,
671 // which would give PHI nodes non-determinstic subscripts. Fix this by
672 // processing blocks in order of the occurance in the function.
673 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
674 PE = S.end(); P != PE; ++P) {
675 // If the frontier block is not in the live-in set for the alloca, don't
676 // bother processing it.
677 if (!LiveInBlocks.count(*P))
680 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
683 // Sort by which the block ordering in the function.
684 if (DFBlocks.size() > 1)
685 std::sort(DFBlocks.begin(), DFBlocks.end());
687 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
688 BasicBlock *BB = DFBlocks[i].second;
689 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
690 Info.DefiningBlocks.push_back(BB);
696 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
697 /// replace any loads of it that are directly dominated by the definition with
698 /// the value stored.
699 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
701 LargeBlockInfo &LBI) {
702 StoreInst *OnlyStore = Info.OnlyStore;
703 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
704 BasicBlock *StoreBB = OnlyStore->getParent();
707 // Clear out UsingBlocks. We will reconstruct it here if needed.
708 Info.UsingBlocks.clear();
710 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
711 Instruction *UserInst = cast<Instruction>(*UI++);
712 if (!isa<LoadInst>(UserInst)) {
713 assert(UserInst == OnlyStore && "Should only have load/stores");
716 LoadInst *LI = cast<LoadInst>(UserInst);
718 // Okay, if we have a load from the alloca, we want to replace it with the
719 // only value stored to the alloca. We can do this if the value is
720 // dominated by the store. If not, we use the rest of the mem2reg machinery
721 // to insert the phi nodes as needed.
722 if (!StoringGlobalVal) { // Non-instructions are always dominated.
723 if (LI->getParent() == StoreBB) {
724 // If we have a use that is in the same block as the store, compare the
725 // indices of the two instructions to see which one came first. If the
726 // load came before the store, we can't handle it.
727 if (StoreIndex == -1)
728 StoreIndex = LBI.getInstructionIndex(OnlyStore);
730 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
731 // Can't handle this load, bail out.
732 Info.UsingBlocks.push_back(StoreBB);
736 } else if (LI->getParent() != StoreBB &&
737 !dominates(StoreBB, LI->getParent())) {
738 // If the load and store are in different blocks, use BB dominance to
739 // check their relationships. If the store doesn't dom the use, bail
741 Info.UsingBlocks.push_back(LI->getParent());
746 // Otherwise, we *can* safely rewrite this load.
747 Value *ReplVal = OnlyStore->getOperand(0);
748 // If the replacement value is the load, this must occur in unreachable
751 ReplVal = UndefValue::get(LI->getType());
752 LI->replaceAllUsesWith(ReplVal);
753 if (AST && isa<PointerType>(LI->getType()))
754 AST->deleteValue(LI);
755 LI->eraseFromParent();
762 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
763 /// first element of a pair.
764 struct StoreIndexSearchPredicate {
765 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
766 const std::pair<unsigned, StoreInst*> &RHS) {
767 return LHS.first < RHS.first;
773 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
774 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
775 /// potentially useless PHI nodes by just performing a single linear pass over
776 /// the basic block using the Alloca.
778 /// If we cannot promote this alloca (because it is read before it is written),
779 /// return true. This is necessary in cases where, due to control flow, the
780 /// alloca is potentially undefined on some control flow paths. e.g. code like
781 /// this is potentially correct:
783 /// for (...) { if (c) { A = undef; undef = B; } }
785 /// ... so long as A is not used before undef is set.
787 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
788 LargeBlockInfo &LBI) {
789 // The trickiest case to handle is when we have large blocks. Because of this,
790 // this code is optimized assuming that large blocks happen. This does not
791 // significantly pessimize the small block case. This uses LargeBlockInfo to
792 // make it efficient to get the index of various operations in the block.
794 // Clear out UsingBlocks. We will reconstruct it here if needed.
795 Info.UsingBlocks.clear();
797 // Walk the use-def list of the alloca, getting the locations of all stores.
798 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
799 StoresByIndexTy StoresByIndex;
801 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
803 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
804 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
806 // If there are no stores to the alloca, just replace any loads with undef.
807 if (StoresByIndex.empty()) {
808 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
809 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
810 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
811 if (AST && isa<PointerType>(LI->getType()))
812 AST->deleteValue(LI);
814 LI->eraseFromParent();
819 // Sort the stores by their index, making it efficient to do a lookup with a
821 std::sort(StoresByIndex.begin(), StoresByIndex.end());
823 // Walk all of the loads from this alloca, replacing them with the nearest
824 // store above them, if any.
825 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
826 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
829 unsigned LoadIdx = LBI.getInstructionIndex(LI);
831 // Find the nearest store that has a lower than this load.
832 StoresByIndexTy::iterator I =
833 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
834 std::pair<unsigned, StoreInst*>(LoadIdx, 0),
835 StoreIndexSearchPredicate());
837 // If there is no store before this load, then we can't promote this load.
838 if (I == StoresByIndex.begin()) {
839 // Can't handle this load, bail out.
840 Info.UsingBlocks.push_back(LI->getParent());
844 // Otherwise, there was a store before this load, the load takes its value.
846 LI->replaceAllUsesWith(I->second->getOperand(0));
847 if (AST && isa<PointerType>(LI->getType()))
848 AST->deleteValue(LI);
849 LI->eraseFromParent();
855 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
856 // Alloca returns true if there wasn't already a phi-node for that variable
858 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
860 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
861 // Look up the basic-block in question.
862 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
864 // If the BB already has a phi node added for the i'th alloca then we're done!
865 if (PN) return false;
867 // Create a PhiNode using the dereferenced type... and add the phi-node to the
869 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
870 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
873 PhiToAllocaMap[PN] = AllocaNo;
874 PN->reserveOperandSpace(getNumPreds(BB));
876 InsertedPHINodes.insert(PN);
878 if (AST && isa<PointerType>(PN->getType()))
879 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
884 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
885 // stores to the allocas which we are promoting. IncomingVals indicates what
886 // value each Alloca contains on exit from the predecessor block Pred.
888 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
889 RenamePassData::ValVector &IncomingVals,
890 std::vector<RenamePassData> &Worklist) {
892 // If we are inserting any phi nodes into this BB, they will already be in the
894 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
895 // If we have PHI nodes to update, compute the number of edges from Pred to
897 if (PhiToAllocaMap.count(APN)) {
898 // We want to be able to distinguish between PHI nodes being inserted by
899 // this invocation of mem2reg from those phi nodes that already existed in
900 // the IR before mem2reg was run. We determine that APN is being inserted
901 // because it is missing incoming edges. All other PHI nodes being
902 // inserted by this pass of mem2reg will have the same number of incoming
903 // operands so far. Remember this count.
904 unsigned NewPHINumOperands = APN->getNumOperands();
906 unsigned NumEdges = 0;
907 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
910 assert(NumEdges && "Must be at least one edge from Pred to BB!");
912 // Add entries for all the phis.
913 BasicBlock::iterator PNI = BB->begin();
915 unsigned AllocaNo = PhiToAllocaMap[APN];
917 // Add N incoming values to the PHI node.
918 for (unsigned i = 0; i != NumEdges; ++i)
919 APN->addIncoming(IncomingVals[AllocaNo], Pred);
921 // The currently active variable for this block is now the PHI.
922 IncomingVals[AllocaNo] = APN;
924 // Get the next phi node.
926 APN = dyn_cast<PHINode>(PNI);
929 // Verify that it is missing entries. If not, it is not being inserted
930 // by this mem2reg invocation so we want to ignore it.
931 } while (APN->getNumOperands() == NewPHINumOperands);
935 // Don't revisit blocks.
936 if (!Visited.insert(BB)) return;
938 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
939 Instruction *I = II++; // get the instruction, increment iterator
941 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
942 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
945 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
946 if (AI == AllocaLookup.end()) continue;
948 Value *V = IncomingVals[AI->second];
950 // Anything using the load now uses the current value.
951 LI->replaceAllUsesWith(V);
952 if (AST && isa<PointerType>(LI->getType()))
953 AST->deleteValue(LI);
954 BB->getInstList().erase(LI);
955 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
956 // Delete this instruction and mark the name as the current holder of the
958 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
961 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
962 if (ai == AllocaLookup.end())
965 // what value were we writing?
966 IncomingVals[ai->second] = SI->getOperand(0);
967 BB->getInstList().erase(SI);
971 // 'Recurse' to our successors.
972 succ_iterator I = succ_begin(BB), E = succ_end(BB);
975 // Keep track of the successors so we don't visit the same successor twice
976 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
978 // Handle the first successor without using the worklist.
979 VisitedSuccs.insert(*I);
985 if (VisitedSuccs.insert(*I))
986 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
991 /// PromoteMemToReg - Promote the specified list of alloca instructions into
992 /// scalar registers, inserting PHI nodes as appropriate. This function makes
993 /// use of DominanceFrontier information. This function does not modify the CFG
994 /// of the function at all. All allocas must be from the same function.
996 /// If AST is specified, the specified tracker is updated to reflect changes
999 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1000 DominatorTree &DT, DominanceFrontier &DF,
1001 AliasSetTracker *AST) {
1002 // If there is nothing to do, bail out...
1003 if (Allocas.empty()) return;
1005 PromoteMem2Reg(Allocas, DT, DF, AST).run();