1 //===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file promote 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/Analysis/Dominators.h"
26 #include "llvm/Analysis/AliasSetTracker.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/Support/CFG.h"
33 #include "llvm/Support/Compiler.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");
41 // Provide DenseMapKeyInfo for all pointers.
44 struct DenseMapKeyInfo<std::pair<BasicBlock*, unsigned> > {
45 static inline std::pair<BasicBlock*, unsigned> getEmptyKey() {
46 return std::make_pair((BasicBlock*)-1, ~0U);
48 static inline std::pair<BasicBlock*, unsigned> getTombstoneKey() {
49 return std::make_pair((BasicBlock*)-2, 0U);
51 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
52 return DenseMapKeyInfo<void*>::getHashValue(Val.first) + Val.second*2;
54 static bool isPod() { return true; }
58 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
59 /// This is true if there are only loads and stores to the alloca.
61 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
62 // FIXME: If the memory unit is of pointer or integer type, we can permit
63 // assignments to subsections of the memory unit.
65 // Only allow direct loads and stores...
66 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
67 UI != UE; ++UI) // Loop over all of the uses of the alloca
68 if (isa<LoadInst>(*UI)) {
70 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
71 if (SI->getOperand(0) == AI)
72 return false; // Don't allow a store OF the AI, only INTO the AI.
74 return false; // Not a load or store.
83 // Data package used by RenamePass()
84 class VISIBILITY_HIDDEN RenamePassData {
86 typedef std::vector<Value *> ValVector;
89 RenamePassData(BasicBlock *B, BasicBlock *P,
90 const ValVector &V) : BB(B), Pred(P), Values(V) {}
95 void swap(RenamePassData &RHS) {
96 std::swap(BB, RHS.BB);
97 std::swap(Pred, RHS.Pred);
98 Values.swap(RHS.Values);
102 struct VISIBILITY_HIDDEN PromoteMem2Reg {
103 /// Allocas - The alloca instructions being promoted.
105 std::vector<AllocaInst*> Allocas;
106 SmallVector<AllocaInst*, 16> &RetryList;
108 DominanceFrontier &DF;
110 /// AST - An AliasSetTracker object to update. If null, don't update it.
112 AliasSetTracker *AST;
114 /// AllocaLookup - Reverse mapping of Allocas.
116 std::map<AllocaInst*, unsigned> AllocaLookup;
118 /// NewPhiNodes - The PhiNodes we're adding.
120 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
122 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
123 /// it corresponds to.
124 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
126 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
127 /// each alloca that is of pointer type, we keep track of what to copyValue
128 /// to the inserted PHI nodes here.
130 std::vector<Value*> PointerAllocaValues;
132 /// Visited - The set of basic blocks the renamer has already visited.
134 SmallPtrSet<BasicBlock*, 16> Visited;
136 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
137 /// non-determinstic behavior.
138 DenseMap<BasicBlock*, unsigned> BBNumbers;
140 /// BBNumPreds - Lazily compute the number of predecessors a block has.
141 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
143 PromoteMem2Reg(const std::vector<AllocaInst*> &A,
144 SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
145 DominanceFrontier &df, AliasSetTracker *ast)
146 : Allocas(A), RetryList(Retry), DT(dt), DF(df), AST(ast) {}
150 /// properlyDominates - Return true if I1 properly dominates I2.
152 bool properlyDominates(Instruction *I1, Instruction *I2) const {
153 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
154 I1 = II->getNormalDest()->begin();
155 return DT.properlyDominates(I1->getParent(), I2->getParent());
158 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
160 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
161 return DT.dominates(BB1, BB2);
165 void RemoveFromAllocasList(unsigned &AllocaIdx) {
166 Allocas[AllocaIdx] = Allocas.back();
171 unsigned getNumPreds(const BasicBlock *BB) {
172 unsigned &NP = BBNumPreds[BB];
174 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
179 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info);
181 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
182 SmallPtrSet<PHINode*, 16> &DeadPHINodes);
183 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
184 void PromoteLocallyUsedAllocas(BasicBlock *BB,
185 const std::vector<AllocaInst*> &AIs);
187 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
188 RenamePassData::ValVector &IncVals,
189 std::vector<RenamePassData> &Worklist);
190 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
191 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
195 std::vector<BasicBlock*> DefiningBlocks;
196 std::vector<BasicBlock*> UsingBlocks;
198 StoreInst *OnlyStore;
199 BasicBlock *OnlyBlock;
200 bool OnlyUsedInOneBlock;
202 Value *AllocaPointerVal;
205 DefiningBlocks.clear();
209 OnlyUsedInOneBlock = true;
210 AllocaPointerVal = 0;
213 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
215 void AnalyzeAlloca(AllocaInst *AI) {
218 // As we scan the uses of the alloca instruction, keep track of stores,
219 // and decide whether all of the loads and stores to the alloca are within
220 // the same basic block.
221 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end();
223 Instruction *User = cast<Instruction>(*U);
224 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
225 // Remember the basic blocks which define new values for the alloca
226 DefiningBlocks.push_back(SI->getParent());
227 AllocaPointerVal = SI->getOperand(0);
230 LoadInst *LI = cast<LoadInst>(User);
231 // Otherwise it must be a load instruction, keep track of variable
233 UsingBlocks.push_back(LI->getParent());
234 AllocaPointerVal = LI;
237 if (OnlyUsedInOneBlock) {
239 OnlyBlock = User->getParent();
240 else if (OnlyBlock != User->getParent())
241 OnlyUsedInOneBlock = false;
247 } // end of anonymous namespace
250 void PromoteMem2Reg::run() {
251 Function &F = *DF.getRoot()->getParent();
253 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
254 // only used in a single basic block. These instructions can be efficiently
255 // promoted by performing a single linear scan over that one block. Since
256 // individual basic blocks are sometimes large, we group together all allocas
257 // that are live in a single basic block by the basic block they are live in.
258 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
260 if (AST) PointerAllocaValues.resize(Allocas.size());
264 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
265 AllocaInst *AI = Allocas[AllocaNum];
267 assert(isAllocaPromotable(AI) &&
268 "Cannot promote non-promotable alloca!");
269 assert(AI->getParent()->getParent() == &F &&
270 "All allocas should be in the same function, which is same as DF!");
272 if (AI->use_empty()) {
273 // If there are no uses of the alloca, just delete it now.
274 if (AST) AST->deleteValue(AI);
275 AI->eraseFromParent();
277 // Remove the alloca from the Allocas list, since it has been processed
278 RemoveFromAllocasList(AllocaNum);
283 // Calculate the set of read and write-locations for each alloca. This is
284 // analogous to finding the 'uses' and 'definitions' of each variable.
285 Info.AnalyzeAlloca(AI);
287 // If there is only a single store to this value, replace any loads of
288 // it that are directly dominated by the definition with the value stored.
289 if (Info.DefiningBlocks.size() == 1) {
290 RewriteSingleStoreAlloca(AI, Info);
292 // Finally, after the scan, check to see if the store is all that is left.
293 if (Info.UsingBlocks.empty()) {
294 // Remove the (now dead) store and alloca.
295 Info.OnlyStore->eraseFromParent();
296 if (AST) AST->deleteValue(AI);
297 AI->eraseFromParent();
299 // The alloca has been processed, move on.
300 RemoveFromAllocasList(AllocaNum);
307 // If the alloca is only read and written in one basic block, just perform a
308 // linear sweep over the block to eliminate it.
309 if (Info.OnlyUsedInOneBlock) {
310 LocallyUsedAllocas[Info.OnlyBlock].push_back(AI);
312 // Remove the alloca from the Allocas list, since it will be processed.
313 RemoveFromAllocasList(AllocaNum);
318 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
320 // If we haven't computed a numbering for the BB's in the function, do so
322 if (BBNumbers.empty()) {
324 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
328 // Compute the locations where PhiNodes need to be inserted. Look at the
329 // dominance frontier of EACH basic-block we have a write in.
331 unsigned CurrentVersion = 0;
332 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
333 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
334 while (!Info.DefiningBlocks.empty()) {
335 BasicBlock *BB = Info.DefiningBlocks.back();
336 Info.DefiningBlocks.pop_back();
338 // Look up the DF for this write, add it to PhiNodes
339 DominanceFrontier::const_iterator it = DF.find(BB);
340 if (it != DF.end()) {
341 const DominanceFrontier::DomSetType &S = it->second;
343 // In theory we don't need the indirection through the DFBlocks vector.
344 // In practice, the order of calling QueuePhiNode would depend on the
345 // (unspecified) ordering of basic blocks in the dominance frontier,
346 // which would give PHI nodes non-determinstic subscripts. Fix this by
347 // processing blocks in order of the occurance in the function.
348 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
349 PE = S.end(); P != PE; ++P)
350 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
352 // Sort by which the block ordering in the function.
353 std::sort(DFBlocks.begin(), DFBlocks.end());
355 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
356 BasicBlock *BB = DFBlocks[i].second;
357 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
358 Info.DefiningBlocks.push_back(BB);
364 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
365 // of the writes to the variable, scan through the reads of the variable,
366 // marking PHI nodes which are actually necessary as alive (by removing them
367 // from the InsertedPHINodes set). This is not perfect: there may PHI
368 // marked alive because of loads which are dominated by stores, but there
369 // will be no unmarked PHI nodes which are actually used.
371 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i)
372 MarkDominatingPHILive(Info.UsingBlocks[i], AllocaNum, InsertedPHINodes);
373 Info.UsingBlocks.clear();
375 // If there are any PHI nodes which are now known to be dead, remove them!
376 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(),
377 E = InsertedPHINodes.end(); I != E; ++I) {
379 bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum));
381 assert(Erased && "PHI already removed?");
383 if (AST && isa<PointerType>(PN->getType()))
384 AST->deleteValue(PN);
385 PN->eraseFromParent();
386 PhiToAllocaMap.erase(PN);
389 // Keep the reverse mapping of the 'Allocas' array.
390 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
393 // Process all allocas which are only used in a single basic block.
394 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
395 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
396 const std::vector<AllocaInst*> &LocAllocas = I->second;
397 assert(!LocAllocas.empty() && "empty alloca list??");
399 // It's common for there to only be one alloca in the list. Handle it
401 if (LocAllocas.size() == 1) {
402 // If we can do the quick promotion pass, do so now.
403 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
404 RetryList.push_back(LocAllocas[0]); // Failed, retry later.
406 // Locally promote anything possible. Note that if this is unable to
407 // promote a particular alloca, it puts the alloca onto the Allocas vector
408 // for global processing.
409 PromoteLocallyUsedAllocas(I->first, LocAllocas);
414 return; // All of the allocas must have been trivial!
416 // Set the incoming values for the basic block to be null values for all of
417 // the alloca's. We do this in case there is a load of a value that has not
418 // been stored yet. In this case, it will get this null value.
420 RenamePassData::ValVector Values(Allocas.size());
421 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
422 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
424 // Walks all basic blocks in the function performing the SSA rename algorithm
425 // and inserting the phi nodes we marked as necessary
427 std::vector<RenamePassData> RenamePassWorkList;
428 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
429 while (!RenamePassWorkList.empty()) {
431 RPD.swap(RenamePassWorkList.back());
432 RenamePassWorkList.pop_back();
433 // RenamePass may add new worklist entries.
434 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
437 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
440 // Remove the allocas themselves from the function.
441 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
442 Instruction *A = Allocas[i];
444 // If there are any uses of the alloca instructions left, they must be in
445 // sections of dead code that were not processed on the dominance frontier.
446 // Just delete the users now.
449 A->replaceAllUsesWith(UndefValue::get(A->getType()));
450 if (AST) AST->deleteValue(A);
451 A->eraseFromParent();
455 // Loop over all of the PHI nodes and see if there are any that we can get
456 // rid of because they merge all of the same incoming values. This can
457 // happen due to undef values coming into the PHI nodes. This process is
458 // iterative, because eliminating one PHI node can cause others to be removed.
459 bool EliminatedAPHI = true;
460 while (EliminatedAPHI) {
461 EliminatedAPHI = false;
463 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
464 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
465 PHINode *PN = I->second;
467 // If this PHI node merges one value and/or undefs, get the value.
468 if (Value *V = PN->hasConstantValue(true)) {
469 if (!isa<Instruction>(V) ||
470 properlyDominates(cast<Instruction>(V), PN)) {
471 if (AST && isa<PointerType>(PN->getType()))
472 AST->deleteValue(PN);
473 PN->replaceAllUsesWith(V);
474 PN->eraseFromParent();
475 NewPhiNodes.erase(I++);
476 EliminatedAPHI = true;
484 // At this point, the renamer has added entries to PHI nodes for all reachable
485 // code. Unfortunately, there may be unreachable blocks which the renamer
486 // hasn't traversed. If this is the case, the PHI nodes may not
487 // have incoming values for all predecessors. Loop over all PHI nodes we have
488 // created, inserting undef values if they are missing any incoming values.
490 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
491 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
492 // We want to do this once per basic block. As such, only process a block
493 // when we find the PHI that is the first entry in the block.
494 PHINode *SomePHI = I->second;
495 BasicBlock *BB = SomePHI->getParent();
496 if (&BB->front() != SomePHI)
499 // Count the number of preds for BB.
500 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
502 // Only do work here if there the PHI nodes are missing incoming values. We
503 // know that all PHI nodes that were inserted in a block will have the same
504 // number of incoming values, so we can just check any of them.
505 if (SomePHI->getNumIncomingValues() == Preds.size())
508 // Ok, now we know that all of the PHI nodes are missing entries for some
509 // basic blocks. Start by sorting the incoming predecessors for efficient
511 std::sort(Preds.begin(), Preds.end());
513 // Now we loop through all BB's which have entries in SomePHI and remove
514 // them from the Preds list.
515 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
516 // Do a log(n) search of the Preds list for the entry we want.
517 SmallVector<BasicBlock*, 16>::iterator EntIt =
518 std::lower_bound(Preds.begin(), Preds.end(),
519 SomePHI->getIncomingBlock(i));
520 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
521 "PHI node has entry for a block which is not a predecessor!");
527 // At this point, the blocks left in the preds list must have dummy
528 // entries inserted into every PHI nodes for the block. Update all the phi
529 // nodes in this block that we are inserting (there could be phis before
531 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
532 BasicBlock::iterator BBI = BB->begin();
533 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
534 SomePHI->getNumIncomingValues() == NumBadPreds) {
535 Value *UndefVal = UndefValue::get(SomePHI->getType());
536 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
537 SomePHI->addIncoming(UndefVal, Preds[pred]);
545 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
546 /// replace any loads of it that are directly dominated by the definition with
547 /// the value stored.
548 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
550 StoreInst *OnlyStore = Info.OnlyStore;
551 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
553 // Be aware of loads before the store.
554 SmallPtrSet<BasicBlock*, 32> ProcessedBlocks;
555 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) {
556 BasicBlock *UseBlock = Info.UsingBlocks[i];
558 // If we already processed this block, don't reprocess it.
559 if (!ProcessedBlocks.insert(UseBlock)) {
560 Info.UsingBlocks[i] = Info.UsingBlocks.back();
561 Info.UsingBlocks.pop_back();
566 // If the store dominates the block and if we haven't processed it yet,
567 // do so now. We can't handle the case where the store doesn't dominate a
568 // block because there may be a path between the store and the use, but we
569 // may need to insert phi nodes to handle dominance properly.
570 if (!StoringGlobalVal && !dominates(OnlyStore->getParent(), UseBlock))
573 // If the use and store are in the same block, do a quick scan to
574 // verify that there are no uses before the store.
575 if (UseBlock == OnlyStore->getParent()) {
576 BasicBlock::iterator I = UseBlock->begin();
577 for (; &*I != OnlyStore; ++I) { // scan block for store.
578 if (isa<LoadInst>(I) && I->getOperand(0) == AI)
581 if (&*I != OnlyStore)
582 continue; // Do not promote the uses of this in this block.
585 // Otherwise, if this is a different block or if all uses happen
586 // after the store, do a simple linear scan to replace loads with
588 for (BasicBlock::iterator I = UseBlock->begin(), E = UseBlock->end();
590 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
591 if (LI->getOperand(0) == AI) {
592 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
593 if (AST && isa<PointerType>(LI->getType()))
594 AST->deleteValue(LI);
595 LI->eraseFromParent();
600 // Finally, remove this block from the UsingBlock set.
601 Info.UsingBlocks[i] = Info.UsingBlocks.back();
602 Info.UsingBlocks.pop_back();
608 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
609 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
610 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
611 // each read of the variable. For each block that reads the variable, this
612 // function is called, which removes used PHI nodes from the DeadPHINodes set.
613 // After all of the reads have been processed, any PHI nodes left in the
614 // DeadPHINodes set are removed.
616 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
617 SmallPtrSet<PHINode*, 16> &DeadPHINodes) {
618 // Scan the immediate dominators of this block looking for a block which has a
619 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
620 DomTreeNode *IDomNode = DT.getNode(BB);
621 for (DomTreeNode *IDom = IDomNode; IDom; IDom = IDom->getIDom()) {
622 BasicBlock *DomBB = IDom->getBlock();
623 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator
624 I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum));
625 if (I == NewPhiNodes.end()) continue;
627 // Ok, we found an inserted PHI node which dominates this value.
628 PHINode *DominatingPHI = I->second;
630 // Find out if we previously thought it was dead. If so, mark it as being
631 // live by removing it from the DeadPHINodes set.
632 if (!DeadPHINodes.erase(DominatingPHI))
635 // Now that we have marked the PHI node alive, also mark any PHI nodes
636 // which it might use as being alive as well.
637 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
639 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
643 /// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
644 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
645 /// potentially useless PHI nodes by just performing a single linear pass over
646 /// the basic block using the Alloca.
648 /// If we cannot promote this alloca (because it is read before it is written),
649 /// return true. This is necessary in cases where, due to control flow, the
650 /// alloca is potentially undefined on some control flow paths. e.g. code like
651 /// this is potentially correct:
653 /// for (...) { if (c) { A = undef; undef = B; } }
655 /// ... so long as A is not used before undef is set.
657 bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
658 assert(!AI->use_empty() && "There are no uses of the alloca!");
660 // Handle degenerate cases quickly.
661 if (AI->hasOneUse()) {
662 Instruction *U = cast<Instruction>(AI->use_back());
663 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
664 // Must be a load of uninitialized value.
665 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
666 if (AST && isa<PointerType>(LI->getType()))
667 AST->deleteValue(LI);
669 // Otherwise it must be a store which is never read.
670 assert(isa<StoreInst>(U));
672 BB->getInstList().erase(U);
674 // Uses of the uninitialized memory location shall get undef.
677 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
678 Instruction *Inst = I++;
679 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
680 if (LI->getOperand(0) == AI) {
681 if (!CurVal) return true; // Could not locally promote!
683 // Loads just returns the "current value"...
684 LI->replaceAllUsesWith(CurVal);
685 if (AST && isa<PointerType>(LI->getType()))
686 AST->deleteValue(LI);
687 BB->getInstList().erase(LI);
689 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
690 if (SI->getOperand(1) == AI) {
691 // Store updates the "current value"...
692 CurVal = SI->getOperand(0);
693 BB->getInstList().erase(SI);
699 // After traversing the basic block, there should be no more uses of the
700 // alloca: remove it now.
701 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
702 if (AST) AST->deleteValue(AI);
703 AI->eraseFromParent();
709 /// PromoteLocallyUsedAllocas - This method is just like
710 /// PromoteLocallyUsedAlloca, except that it processes multiple alloca
711 /// instructions in parallel. This is important in cases where we have large
712 /// basic blocks, as we don't want to rescan the entire basic block for each
713 /// alloca which is locally used in it (which might be a lot).
714 void PromoteMem2Reg::
715 PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
716 DenseMap<AllocaInst*, Value*> CurValues;
717 for (unsigned i = 0, e = AIs.size(); i != e; ++i)
718 CurValues[AIs[i]] = 0; // Insert with null value
720 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
721 Instruction *Inst = I++;
722 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
723 // Is this a load of an alloca we are tracking?
724 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
725 DenseMap<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
726 if (AIt != CurValues.end()) {
727 // If loading an uninitialized value, allow the inter-block case to
728 // handle it. Due to control flow, this might actually be ok.
729 if (AIt->second == 0) { // Use of locally uninitialized value??
730 RetryList.push_back(AI); // Retry elsewhere.
731 CurValues.erase(AIt); // Stop tracking this here.
732 if (CurValues.empty()) return;
734 // Loads just returns the "current value"...
735 LI->replaceAllUsesWith(AIt->second);
736 if (AST && isa<PointerType>(LI->getType()))
737 AST->deleteValue(LI);
738 BB->getInstList().erase(LI);
742 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
743 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
744 DenseMap<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
745 if (AIt != CurValues.end()) {
746 // Store updates the "current value"...
747 AIt->second = SI->getOperand(0);
748 SI->eraseFromParent();
754 // At the end of the block scan, all allocas in CurValues are dead.
755 for (DenseMap<AllocaInst*, Value*>::iterator I = CurValues.begin(),
756 E = CurValues.end(); I != E; ++I) {
757 AllocaInst *AI = I->first;
758 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
759 if (AST) AST->deleteValue(AI);
760 AI->eraseFromParent();
763 NumLocalPromoted += CurValues.size();
768 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
769 // Alloca returns true if there wasn't already a phi-node for that variable
771 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
773 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
774 // Look up the basic-block in question.
775 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
777 // If the BB already has a phi node added for the i'th alloca then we're done!
778 if (PN) return false;
780 // Create a PhiNode using the dereferenced type... and add the phi-node to the
782 PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
783 Allocas[AllocaNo]->getName() + "." +
784 utostr(Version++), BB->begin());
785 PhiToAllocaMap[PN] = AllocaNo;
786 PN->reserveOperandSpace(getNumPreds(BB));
788 InsertedPHINodes.insert(PN);
790 if (AST && isa<PointerType>(PN->getType()))
791 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
797 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
798 // stores to the allocas which we are promoting. IncomingVals indicates what
799 // value each Alloca contains on exit from the predecessor block Pred.
801 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
802 RenamePassData::ValVector &IncomingVals,
803 std::vector<RenamePassData> &Worklist) {
804 // If we are inserting any phi nodes into this BB, they will already be in the
806 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
807 // Pred may have multiple edges to BB. If so, we want to add N incoming
808 // values to each PHI we are inserting on the first time we see the edge.
809 // Check to see if APN already has incoming values from Pred. This also
810 // prevents us from modifying PHI nodes that are not currently being
812 bool HasPredEntries = false;
813 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
814 if (APN->getIncomingBlock(i) == Pred) {
815 HasPredEntries = true;
820 // If we have PHI nodes to update, compute the number of edges from Pred to
822 if (!HasPredEntries) {
823 TerminatorInst *PredTerm = Pred->getTerminator();
824 unsigned NumEdges = 0;
825 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
826 if (PredTerm->getSuccessor(i) == BB)
829 assert(NumEdges && "Must be at least one edge from Pred to BB!");
831 // Add entries for all the phis.
832 BasicBlock::iterator PNI = BB->begin();
834 unsigned AllocaNo = PhiToAllocaMap[APN];
836 // Add N incoming values to the PHI node.
837 for (unsigned i = 0; i != NumEdges; ++i)
838 APN->addIncoming(IncomingVals[AllocaNo], Pred);
840 // The currently active variable for this block is now the PHI.
841 IncomingVals[AllocaNo] = APN;
843 // Get the next phi node.
845 APN = dyn_cast<PHINode>(PNI);
848 // Verify it doesn't already have entries for Pred. If it does, it is
849 // not being inserted by this mem2reg invocation.
850 HasPredEntries = false;
851 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
852 if (APN->getIncomingBlock(i) == Pred) {
853 HasPredEntries = true;
857 } while (!HasPredEntries);
861 // Don't revisit blocks.
862 if (!Visited.insert(BB)) return;
864 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
865 Instruction *I = II++; // get the instruction, increment iterator
867 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
868 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
869 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
870 if (AI != AllocaLookup.end()) {
871 Value *V = IncomingVals[AI->second];
873 // walk the use list of this load and replace all uses with r
874 LI->replaceAllUsesWith(V);
875 if (AST && isa<PointerType>(LI->getType()))
876 AST->deleteValue(LI);
877 BB->getInstList().erase(LI);
880 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
881 // Delete this instruction and mark the name as the current holder of the
883 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
884 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
885 if (ai != AllocaLookup.end()) {
886 // what value were we writing?
887 IncomingVals[ai->second] = SI->getOperand(0);
888 BB->getInstList().erase(SI);
894 // Recurse to our successors.
895 TerminatorInst *TI = BB->getTerminator();
896 for (unsigned i = 0; i != TI->getNumSuccessors(); i++)
897 Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
900 /// PromoteMemToReg - Promote the specified list of alloca instructions into
901 /// scalar registers, inserting PHI nodes as appropriate. This function makes
902 /// use of DominanceFrontier information. This function does not modify the CFG
903 /// of the function at all. All allocas must be from the same function.
905 /// If AST is specified, the specified tracker is updated to reflect changes
908 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
909 DominatorTree &DT, DominanceFrontier &DF,
910 AliasSetTracker *AST) {
911 // If there is nothing to do, bail out...
912 if (Allocas.empty()) return;
914 SmallVector<AllocaInst*, 16> RetryList;
915 PromoteMem2Reg(Allocas, RetryList, DT, DF, AST).run();
917 // PromoteMem2Reg may not have been able to promote all of the allocas in one
918 // pass, run it again if needed.
919 std::vector<AllocaInst*> NewAllocas;
920 while (!RetryList.empty()) {
921 // If we need to retry some allocas, this is due to there being no store
922 // before a read in a local block. To counteract this, insert a store of
923 // undef into the alloca right after the alloca itself.
924 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
925 BasicBlock::iterator BBI = RetryList[i];
927 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
928 RetryList[i], ++BBI);
931 NewAllocas.assign(RetryList.begin(), RetryList.end());
933 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, AST).run();