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 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/AliasSetTracker.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Support/CFG.h"
28 #include "llvm/Support/StableBasicBlockNumbering.h"
29 #include "llvm/Support/Compiler.h"
33 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
34 /// This is true if there are only loads and stores to the alloca.
36 bool llvm::isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
37 // FIXME: If the memory unit is of pointer or integer type, we can permit
38 // assignments to subsections of the memory unit.
40 // Only allow direct loads and stores...
41 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
42 UI != UE; ++UI) // Loop over all of the uses of the alloca
43 if (isa<LoadInst>(*UI)) {
45 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
46 if (SI->getOperand(0) == AI)
47 return false; // Don't allow a store OF the AI, only INTO the AI.
49 return false; // Not a load or store.
56 struct VISIBILITY_HIDDEN PromoteMem2Reg {
57 /// Allocas - The alloca instructions being promoted.
59 std::vector<AllocaInst*> Allocas;
60 std::vector<AllocaInst*> &RetryList;
62 DominanceFrontier &DF;
65 /// AST - An AliasSetTracker object to update. If null, don't update it.
69 /// AllocaLookup - Reverse mapping of Allocas.
71 std::map<AllocaInst*, unsigned> AllocaLookup;
73 /// NewPhiNodes - The PhiNodes we're adding.
75 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
77 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
78 /// each alloca that is of pointer type, we keep track of what to copyValue
79 /// to the inserted PHI nodes here.
81 std::vector<Value*> PointerAllocaValues;
83 /// Visited - The set of basic blocks the renamer has already visited.
85 std::set<BasicBlock*> Visited;
87 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
88 /// non-determinstic behavior.
89 StableBasicBlockNumbering BBNumbers;
92 PromoteMem2Reg(const std::vector<AllocaInst*> &A,
93 std::vector<AllocaInst*> &Retry, DominatorTree &dt,
94 DominanceFrontier &df, const TargetData &td,
96 : Allocas(A), RetryList(Retry), DT(dt), DF(df), TD(td), AST(ast) {}
100 /// properlyDominates - Return true if I1 properly dominates I2.
102 bool properlyDominates(Instruction *I1, Instruction *I2) const {
103 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
104 I1 = II->getNormalDest()->begin();
105 return DT[I1->getParent()]->properlyDominates(DT[I2->getParent()]);
108 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
110 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
111 return DT[BB1]->dominates(DT[BB2]);
115 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
116 std::set<PHINode*> &DeadPHINodes);
117 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
118 void PromoteLocallyUsedAllocas(BasicBlock *BB,
119 const std::vector<AllocaInst*> &AIs);
121 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
122 std::vector<Value*> &IncVals);
123 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
124 std::set<PHINode*> &InsertedPHINodes);
126 } // end of anonymous namespace
128 void PromoteMem2Reg::run() {
129 Function &F = *DF.getRoot()->getParent();
131 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
132 // only used in a single basic block. These instructions can be efficiently
133 // promoted by performing a single linear scan over that one block. Since
134 // individual basic blocks are sometimes large, we group together all allocas
135 // that are live in a single basic block by the basic block they are live in.
136 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
138 if (AST) PointerAllocaValues.resize(Allocas.size());
140 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
141 AllocaInst *AI = Allocas[AllocaNum];
143 assert(isAllocaPromotable(AI, TD) &&
144 "Cannot promote non-promotable alloca!");
145 assert(AI->getParent()->getParent() == &F &&
146 "All allocas should be in the same function, which is same as DF!");
148 if (AI->use_empty()) {
149 // If there are no uses of the alloca, just delete it now.
150 if (AST) AST->deleteValue(AI);
151 AI->eraseFromParent();
153 // Remove the alloca from the Allocas list, since it has been processed
154 Allocas[AllocaNum] = Allocas.back();
160 // Calculate the set of read and write-locations for each alloca. This is
161 // analogous to finding the 'uses' and 'definitions' of each variable.
162 std::vector<BasicBlock*> DefiningBlocks;
163 std::vector<BasicBlock*> UsingBlocks;
165 StoreInst *OnlyStore = 0;
166 BasicBlock *OnlyBlock = 0;
167 bool OnlyUsedInOneBlock = true;
169 // As we scan the uses of the alloca instruction, keep track of stores, and
170 // decide whether all of the loads and stores to the alloca are within the
172 Value *AllocaPointerVal = 0;
173 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
174 Instruction *User = cast<Instruction>(*U);
175 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
176 // Remember the basic blocks which define new values for the alloca
177 DefiningBlocks.push_back(SI->getParent());
178 AllocaPointerVal = SI->getOperand(0);
181 LoadInst *LI = cast<LoadInst>(User);
182 // Otherwise it must be a load instruction, keep track of variable reads
183 UsingBlocks.push_back(LI->getParent());
184 AllocaPointerVal = LI;
187 if (OnlyUsedInOneBlock) {
189 OnlyBlock = User->getParent();
190 else if (OnlyBlock != User->getParent())
191 OnlyUsedInOneBlock = false;
195 // If the alloca is only read and written in one basic block, just perform a
196 // linear sweep over the block to eliminate it.
197 if (OnlyUsedInOneBlock) {
198 LocallyUsedAllocas[OnlyBlock].push_back(AI);
200 // Remove the alloca from the Allocas list, since it will be processed.
201 Allocas[AllocaNum] = Allocas.back();
207 // If there is only a single store to this value, replace any loads of
208 // it that are directly dominated by the definition with the value stored.
209 if (DefiningBlocks.size() == 1) {
210 // Be aware of loads before the store.
211 std::set<BasicBlock*> ProcessedBlocks;
212 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
213 // If the store dominates the block and if we haven't processed it yet,
215 if (dominates(OnlyStore->getParent(), UsingBlocks[i]))
216 if (ProcessedBlocks.insert(UsingBlocks[i]).second) {
217 BasicBlock *UseBlock = UsingBlocks[i];
219 // If the use and store are in the same block, do a quick scan to
220 // verify that there are no uses before the store.
221 if (UseBlock == OnlyStore->getParent()) {
222 BasicBlock::iterator I = UseBlock->begin();
223 for (; &*I != OnlyStore; ++I) { // scan block for store.
224 if (isa<LoadInst>(I) && I->getOperand(0) == AI)
227 if (&*I != OnlyStore) break; // Do not handle this case.
230 // Otherwise, if this is a different block or if all uses happen
231 // after the store, do a simple linear scan to replace loads with
233 for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end();
235 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
236 if (LI->getOperand(0) == AI) {
237 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
238 if (AST && isa<PointerType>(LI->getType()))
239 AST->deleteValue(LI);
240 LI->eraseFromParent();
245 // Finally, remove this block from the UsingBlock set.
246 UsingBlocks[i] = UsingBlocks.back();
250 // Finally, after the scan, check to see if the store is all that is left.
251 if (UsingBlocks.empty()) {
252 // The alloca has been processed, move on.
253 Allocas[AllocaNum] = Allocas.back();
262 PointerAllocaValues[AllocaNum] = AllocaPointerVal;
264 // If we haven't computed a numbering for the BB's in the function, do so
266 BBNumbers.compute(F);
268 // Compute the locations where PhiNodes need to be inserted. Look at the
269 // dominance frontier of EACH basic-block we have a write in.
271 unsigned CurrentVersion = 0;
272 std::set<PHINode*> InsertedPHINodes;
273 std::vector<unsigned> DFBlocks;
274 while (!DefiningBlocks.empty()) {
275 BasicBlock *BB = DefiningBlocks.back();
276 DefiningBlocks.pop_back();
278 // Look up the DF for this write, add it to PhiNodes
279 DominanceFrontier::const_iterator it = DF.find(BB);
280 if (it != DF.end()) {
281 const DominanceFrontier::DomSetType &S = it->second;
283 // In theory we don't need the indirection through the DFBlocks vector.
284 // In practice, the order of calling QueuePhiNode would depend on the
285 // (unspecified) ordering of basic blocks in the dominance frontier,
286 // which would give PHI nodes non-determinstic subscripts. Fix this by
287 // processing blocks in order of the occurance in the function.
288 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
289 PE = S.end(); P != PE; ++P)
290 DFBlocks.push_back(BBNumbers.getNumber(*P));
292 // Sort by which the block ordering in the function.
293 std::sort(DFBlocks.begin(), DFBlocks.end());
295 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
296 BasicBlock *BB = BBNumbers.getBlock(DFBlocks[i]);
297 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
298 DefiningBlocks.push_back(BB);
304 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
305 // of the writes to the variable, scan through the reads of the variable,
306 // marking PHI nodes which are actually necessary as alive (by removing them
307 // from the InsertedPHINodes set). This is not perfect: there may PHI
308 // marked alive because of loads which are dominated by stores, but there
309 // will be no unmarked PHI nodes which are actually used.
311 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
312 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
315 // If there are any PHI nodes which are now known to be dead, remove them!
316 for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(),
317 E = InsertedPHINodes.end(); I != E; ++I) {
319 std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()];
320 BBPNs[AllocaNum] = 0;
322 // Check to see if we just removed the last inserted PHI node from this
323 // basic block. If so, remove the entry for the basic block.
324 bool HasOtherPHIs = false;
325 for (unsigned i = 0, e = BBPNs.size(); i != e; ++i)
331 NewPhiNodes.erase(PN->getParent());
333 if (AST && isa<PointerType>(PN->getType()))
334 AST->deleteValue(PN);
335 PN->eraseFromParent();
338 // Keep the reverse mapping of the 'Allocas' array.
339 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
342 // Process all allocas which are only used in a single basic block.
343 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
344 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
345 const std::vector<AllocaInst*> &LocAllocas = I->second;
346 assert(!LocAllocas.empty() && "empty alloca list??");
348 // It's common for there to only be one alloca in the list. Handle it
350 if (LocAllocas.size() == 1) {
351 // If we can do the quick promotion pass, do so now.
352 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
353 RetryList.push_back(LocAllocas[0]); // Failed, retry later.
355 // Locally promote anything possible. Note that if this is unable to
356 // promote a particular alloca, it puts the alloca onto the Allocas vector
357 // for global processing.
358 PromoteLocallyUsedAllocas(I->first, LocAllocas);
363 return; // All of the allocas must have been trivial!
365 // Set the incoming values for the basic block to be null values for all of
366 // the alloca's. We do this in case there is a load of a value that has not
367 // been stored yet. In this case, it will get this null value.
369 std::vector<Value *> Values(Allocas.size());
370 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
371 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
373 // Walks all basic blocks in the function performing the SSA rename algorithm
374 // and inserting the phi nodes we marked as necessary
376 RenamePass(F.begin(), 0, Values);
378 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
381 // Remove the allocas themselves from the function.
382 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
383 Instruction *A = Allocas[i];
385 // If there are any uses of the alloca instructions left, they must be in
386 // sections of dead code that were not processed on the dominance frontier.
387 // Just delete the users now.
390 A->replaceAllUsesWith(UndefValue::get(A->getType()));
391 if (AST) AST->deleteValue(A);
392 A->eraseFromParent();
396 // Loop over all of the PHI nodes and see if there are any that we can get
397 // rid of because they merge all of the same incoming values. This can
398 // happen due to undef values coming into the PHI nodes. This process is
399 // iterative, because eliminating one PHI node can cause others to be removed.
400 bool EliminatedAPHI = true;
401 while (EliminatedAPHI) {
402 EliminatedAPHI = false;
404 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
405 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
406 std::vector<PHINode*> &PNs = I->second;
407 for (unsigned i = 0, e = PNs.size(); i != e; ++i) {
408 if (!PNs[i]) continue;
410 // If this PHI node merges one value and/or undefs, get the value.
411 if (Value *V = PNs[i]->hasConstantValue(true)) {
412 if (!isa<Instruction>(V) ||
413 properlyDominates(cast<Instruction>(V), PNs[i])) {
414 if (AST && isa<PointerType>(PNs[i]->getType()))
415 AST->deleteValue(PNs[i]);
416 PNs[i]->replaceAllUsesWith(V);
417 PNs[i]->eraseFromParent();
419 EliminatedAPHI = true;
427 // At this point, the renamer has added entries to PHI nodes for all reachable
428 // code. Unfortunately, there may be blocks which are not reachable, which
429 // the renamer hasn't traversed. If this is the case, the PHI nodes may not
430 // have incoming values for all predecessors. Loop over all PHI nodes we have
431 // created, inserting undef values if they are missing any incoming values.
433 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
434 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
436 std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first));
437 std::vector<PHINode*> &PNs = I->second;
438 assert(!PNs.empty() && "Empty PHI node list??");
439 PHINode *SomePHI = 0;
440 for (unsigned i = 0, e = PNs.size(); i != e; ++i)
446 // Only do work here if there the PHI nodes are missing incoming values. We
447 // know that all PHI nodes that were inserted in a block will have the same
448 // number of incoming values, so we can just check any PHI node.
449 if (SomePHI && Preds.size() != SomePHI->getNumIncomingValues()) {
450 // Ok, now we know that all of the PHI nodes are missing entries for some
451 // basic blocks. Start by sorting the incoming predecessors for efficient
453 std::sort(Preds.begin(), Preds.end());
455 // Now we loop through all BB's which have entries in SomePHI and remove
456 // them from the Preds list.
457 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
458 // Do a log(n) search of the Preds list for the entry we want.
459 std::vector<BasicBlock*>::iterator EntIt =
460 std::lower_bound(Preds.begin(), Preds.end(),
461 SomePHI->getIncomingBlock(i));
462 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
463 "PHI node has entry for a block which is not a predecessor!");
469 // At this point, the blocks left in the preds list must have dummy
470 // entries inserted into every PHI nodes for the block.
471 for (unsigned i = 0, e = PNs.size(); i != e; ++i)
472 if (PHINode *PN = PNs[i]) {
473 Value *UndefVal = UndefValue::get(PN->getType());
474 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
475 PN->addIncoming(UndefVal, Preds[pred]);
481 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
482 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
483 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
484 // each read of the variable. For each block that reads the variable, this
485 // function is called, which removes used PHI nodes from the DeadPHINodes set.
486 // After all of the reads have been processed, any PHI nodes left in the
487 // DeadPHINodes set are removed.
489 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
490 std::set<PHINode*> &DeadPHINodes) {
491 // Scan the immediate dominators of this block looking for a block which has a
492 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
493 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) {
494 BasicBlock *DomBB = N->getBlock();
495 std::map<BasicBlock*, std::vector<PHINode*> >::iterator
496 I = NewPhiNodes.find(DomBB);
497 if (I != NewPhiNodes.end() && I->second[AllocaNum]) {
498 // Ok, we found an inserted PHI node which dominates this value.
499 PHINode *DominatingPHI = I->second[AllocaNum];
501 // Find out if we previously thought it was dead.
502 std::set<PHINode*>::iterator DPNI = DeadPHINodes.find(DominatingPHI);
503 if (DPNI != DeadPHINodes.end()) {
504 // Ok, until now, we thought this PHI node was dead. Mark it as being
506 DeadPHINodes.erase(DPNI);
508 // Now that we have marked the PHI node alive, also mark any PHI nodes
509 // which it might use as being alive as well.
510 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
512 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
518 /// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
519 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
520 /// potentially useless PHI nodes by just performing a single linear pass over
521 /// the basic block using the Alloca.
523 /// If we cannot promote this alloca (because it is read before it is written),
524 /// return true. This is necessary in cases where, due to control flow, the
525 /// alloca is potentially undefined on some control flow paths. e.g. code like
526 /// this is potentially correct:
528 /// for (...) { if (c) { A = undef; undef = B; } }
530 /// ... so long as A is not used before undef is set.
532 bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
533 assert(!AI->use_empty() && "There are no uses of the alloca!");
535 // Handle degenerate cases quickly.
536 if (AI->hasOneUse()) {
537 Instruction *U = cast<Instruction>(AI->use_back());
538 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
539 // Must be a load of uninitialized value.
540 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
541 if (AST && isa<PointerType>(LI->getType()))
542 AST->deleteValue(LI);
544 // Otherwise it must be a store which is never read.
545 assert(isa<StoreInst>(U));
547 BB->getInstList().erase(U);
549 // Uses of the uninitialized memory location shall get undef.
552 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
553 Instruction *Inst = I++;
554 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
555 if (LI->getOperand(0) == AI) {
556 if (!CurVal) return true; // Could not locally promote!
558 // Loads just returns the "current value"...
559 LI->replaceAllUsesWith(CurVal);
560 if (AST && isa<PointerType>(LI->getType()))
561 AST->deleteValue(LI);
562 BB->getInstList().erase(LI);
564 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
565 if (SI->getOperand(1) == AI) {
566 // Store updates the "current value"...
567 CurVal = SI->getOperand(0);
568 BB->getInstList().erase(SI);
574 // After traversing the basic block, there should be no more uses of the
575 // alloca, remove it now.
576 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
577 if (AST) AST->deleteValue(AI);
578 AI->getParent()->getInstList().erase(AI);
582 /// PromoteLocallyUsedAllocas - This method is just like
583 /// PromoteLocallyUsedAlloca, except that it processes multiple alloca
584 /// instructions in parallel. This is important in cases where we have large
585 /// basic blocks, as we don't want to rescan the entire basic block for each
586 /// alloca which is locally used in it (which might be a lot).
587 void PromoteMem2Reg::
588 PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
589 std::map<AllocaInst*, Value*> CurValues;
590 for (unsigned i = 0, e = AIs.size(); i != e; ++i)
591 CurValues[AIs[i]] = 0; // Insert with null value
593 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
594 Instruction *Inst = I++;
595 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
596 // Is this a load of an alloca we are tracking?
597 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
598 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
599 if (AIt != CurValues.end()) {
600 // If loading an uninitialized value, allow the inter-block case to
601 // handle it. Due to control flow, this might actually be ok.
602 if (AIt->second == 0) { // Use of locally uninitialized value??
603 RetryList.push_back(AI); // Retry elsewhere.
604 CurValues.erase(AIt); // Stop tracking this here.
605 if (CurValues.empty()) return;
607 // Loads just returns the "current value"...
608 LI->replaceAllUsesWith(AIt->second);
609 if (AST && isa<PointerType>(LI->getType()))
610 AST->deleteValue(LI);
611 BB->getInstList().erase(LI);
615 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
616 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
617 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
618 if (AIt != CurValues.end()) {
619 // Store updates the "current value"...
620 AIt->second = SI->getOperand(0);
621 BB->getInstList().erase(SI);
630 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
631 // Alloca returns true if there wasn't already a phi-node for that variable
633 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
635 std::set<PHINode*> &InsertedPHINodes) {
636 // Look up the basic-block in question.
637 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB];
638 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
640 // If the BB already has a phi node added for the i'th alloca then we're done!
641 if (BBPNs[AllocaNo]) return false;
643 // Create a PhiNode using the dereferenced type... and add the phi-node to the
645 PHINode *PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
646 Allocas[AllocaNo]->getName() + "." +
647 utostr(Version++), BB->begin());
648 BBPNs[AllocaNo] = PN;
649 InsertedPHINodes.insert(PN);
651 if (AST && isa<PointerType>(PN->getType()))
652 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
658 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
659 // stores to the allocas which we are promoting. IncomingVals indicates what
660 // value each Alloca contains on exit from the predecessor block Pred.
662 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
663 std::vector<Value*> &IncomingVals) {
665 // If this BB needs a PHI node, update the PHI node for each variable we need
667 std::map<BasicBlock*, std::vector<PHINode *> >::iterator
668 BBPNI = NewPhiNodes.find(BB);
669 if (BBPNI != NewPhiNodes.end()) {
670 std::vector<PHINode *> &BBPNs = BBPNI->second;
671 for (unsigned k = 0; k != BBPNs.size(); ++k)
672 if (PHINode *PN = BBPNs[k]) {
673 // Add this incoming value to the PHI node.
674 PN->addIncoming(IncomingVals[k], Pred);
676 // The currently active variable for this block is now the PHI.
677 IncomingVals[k] = PN;
681 // don't revisit nodes
682 if (Visited.count(BB)) return;
687 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
688 Instruction *I = II++; // get the instruction, increment iterator
690 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
691 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
692 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
693 if (AI != AllocaLookup.end()) {
694 Value *V = IncomingVals[AI->second];
696 // walk the use list of this load and replace all uses with r
697 LI->replaceAllUsesWith(V);
698 if (AST && isa<PointerType>(LI->getType()))
699 AST->deleteValue(LI);
700 BB->getInstList().erase(LI);
703 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
704 // Delete this instruction and mark the name as the current holder of the
706 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
707 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
708 if (ai != AllocaLookup.end()) {
709 // what value were we writing?
710 IncomingVals[ai->second] = SI->getOperand(0);
711 BB->getInstList().erase(SI);
717 // Recurse to our successors.
718 TerminatorInst *TI = BB->getTerminator();
719 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
720 std::vector<Value*> OutgoingVals(IncomingVals);
721 RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
725 /// PromoteMemToReg - Promote the specified list of alloca instructions into
726 /// scalar registers, inserting PHI nodes as appropriate. This function makes
727 /// use of DominanceFrontier information. This function does not modify the CFG
728 /// of the function at all. All allocas must be from the same function.
730 /// If AST is specified, the specified tracker is updated to reflect changes
733 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
734 DominatorTree &DT, DominanceFrontier &DF,
735 const TargetData &TD, AliasSetTracker *AST) {
736 // If there is nothing to do, bail out...
737 if (Allocas.empty()) return;
739 std::vector<AllocaInst*> RetryList;
740 PromoteMem2Reg(Allocas, RetryList, DT, DF, TD, AST).run();
742 // PromoteMem2Reg may not have been able to promote all of the allocas in one
743 // pass, run it again if needed.
744 while (!RetryList.empty()) {
745 // If we need to retry some allocas, this is due to there being no store
746 // before a read in a local block. To counteract this, insert a store of
747 // undef into the alloca right after the alloca itself.
748 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
749 BasicBlock::iterator BBI = RetryList[i];
751 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
752 RetryList[i], ++BBI);
755 std::vector<AllocaInst*> NewAllocas;
756 std::swap(NewAllocas, RetryList);
757 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, TD, AST).run();