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/DenseMap.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/StringExtras.h"
30 #include "llvm/Support/CFG.h"
31 #include "llvm/Support/Compiler.h"
35 // Provide DenseMapKeyInfo for all pointers.
38 struct DenseMapKeyInfo<std::pair<BasicBlock*, unsigned> > {
39 static inline std::pair<BasicBlock*, unsigned> getEmptyKey() {
40 return std::make_pair((BasicBlock*)-1, ~0U);
42 static inline std::pair<BasicBlock*, unsigned> getTombstoneKey() {
43 return std::make_pair((BasicBlock*)-2, 0U);
45 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
46 return DenseMapKeyInfo<void*>::getHashValue(Val.first) + Val.second*2;
48 static bool isPod() { return true; }
52 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
53 /// This is true if there are only loads and stores to the alloca.
55 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
56 // FIXME: If the memory unit is of pointer or integer type, we can permit
57 // assignments to subsections of the memory unit.
59 // Only allow direct loads and stores...
60 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
61 UI != UE; ++UI) // Loop over all of the uses of the alloca
62 if (isa<LoadInst>(*UI)) {
64 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
65 if (SI->getOperand(0) == AI)
66 return false; // Don't allow a store OF the AI, only INTO the AI.
68 return false; // Not a load or store.
76 // Data package used by RenamePass()
77 class VISIBILITY_HIDDEN RenamePassData {
79 RenamePassData(BasicBlock *B, BasicBlock *P,
80 const std::vector<Value *> &V) : BB(B), Pred(P), Values(V) {}
83 std::vector<Value *> Values;
86 struct VISIBILITY_HIDDEN PromoteMem2Reg {
87 /// Allocas - The alloca instructions being promoted.
89 std::vector<AllocaInst*> Allocas;
90 SmallVector<AllocaInst*, 16> &RetryList;
92 DominanceFrontier &DF;
94 /// AST - An AliasSetTracker object to update. If null, don't update it.
98 /// AllocaLookup - Reverse mapping of Allocas.
100 std::map<AllocaInst*, unsigned> AllocaLookup;
102 /// NewPhiNodes - The PhiNodes we're adding.
104 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
106 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
107 /// it corresponds to.
108 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
110 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
111 /// each alloca that is of pointer type, we keep track of what to copyValue
112 /// to the inserted PHI nodes here.
114 std::vector<Value*> PointerAllocaValues;
116 /// Visited - The set of basic blocks the renamer has already visited.
118 SmallPtrSet<BasicBlock*, 16> Visited;
120 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
121 /// non-determinstic behavior.
122 DenseMap<BasicBlock*, unsigned> BBNumbers;
124 /// RenamePassWorkList - Worklist used by RenamePass()
125 std::vector<RenamePassData> RenamePassWorkList;
128 PromoteMem2Reg(const std::vector<AllocaInst*> &A,
129 SmallVector<AllocaInst*, 16> &Retry, ETForest &et,
130 DominanceFrontier &df, AliasSetTracker *ast)
131 : Allocas(A), RetryList(Retry), ET(et), DF(df), AST(ast) {}
135 /// properlyDominates - Return true if I1 properly dominates I2.
137 bool properlyDominates(Instruction *I1, Instruction *I2) const {
138 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
139 I1 = II->getNormalDest()->begin();
140 return ET.properlyDominates(I1->getParent(), I2->getParent());
143 /// dominates - Return true if BB1 dominates BB2 using the ETForest.
145 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
146 return ET.dominates(BB1, BB2);
150 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
151 SmallPtrSet<PHINode*, 16> &DeadPHINodes);
152 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
153 void PromoteLocallyUsedAllocas(BasicBlock *BB,
154 const std::vector<AllocaInst*> &AIs);
156 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
157 std::vector<Value*> &IncVals);
158 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
159 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
162 } // end of anonymous namespace
164 void PromoteMem2Reg::run() {
165 Function &F = *DF.getRoot()->getParent();
167 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
168 // only used in a single basic block. These instructions can be efficiently
169 // promoted by performing a single linear scan over that one block. Since
170 // individual basic blocks are sometimes large, we group together all allocas
171 // that are live in a single basic block by the basic block they are live in.
172 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
174 if (AST) PointerAllocaValues.resize(Allocas.size());
176 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
177 AllocaInst *AI = Allocas[AllocaNum];
179 assert(isAllocaPromotable(AI) &&
180 "Cannot promote non-promotable alloca!");
181 assert(AI->getParent()->getParent() == &F &&
182 "All allocas should be in the same function, which is same as DF!");
184 if (AI->use_empty()) {
185 // If there are no uses of the alloca, just delete it now.
186 if (AST) AST->deleteValue(AI);
187 AI->eraseFromParent();
189 // Remove the alloca from the Allocas list, since it has been processed
190 Allocas[AllocaNum] = Allocas.back();
196 // Calculate the set of read and write-locations for each alloca. This is
197 // analogous to finding the 'uses' and 'definitions' of each variable.
198 std::vector<BasicBlock*> DefiningBlocks;
199 std::vector<BasicBlock*> UsingBlocks;
201 StoreInst *OnlyStore = 0;
202 BasicBlock *OnlyBlock = 0;
203 bool OnlyUsedInOneBlock = true;
205 // As we scan the uses of the alloca instruction, keep track of stores, and
206 // decide whether all of the loads and stores to the alloca are within the
208 Value *AllocaPointerVal = 0;
209 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
210 Instruction *User = cast<Instruction>(*U);
211 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
212 // Remember the basic blocks which define new values for the alloca
213 DefiningBlocks.push_back(SI->getParent());
214 AllocaPointerVal = SI->getOperand(0);
217 LoadInst *LI = cast<LoadInst>(User);
218 // Otherwise it must be a load instruction, keep track of variable reads
219 UsingBlocks.push_back(LI->getParent());
220 AllocaPointerVal = LI;
223 if (OnlyUsedInOneBlock) {
225 OnlyBlock = User->getParent();
226 else if (OnlyBlock != User->getParent())
227 OnlyUsedInOneBlock = false;
231 // If the alloca is only read and written in one basic block, just perform a
232 // linear sweep over the block to eliminate it.
233 if (OnlyUsedInOneBlock) {
234 LocallyUsedAllocas[OnlyBlock].push_back(AI);
236 // Remove the alloca from the Allocas list, since it will be processed.
237 Allocas[AllocaNum] = Allocas.back();
243 // If there is only a single store to this value, replace any loads of
244 // it that are directly dominated by the definition with the value stored.
245 if (DefiningBlocks.size() == 1) {
246 // Be aware of loads before the store.
247 std::set<BasicBlock*> ProcessedBlocks;
248 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
249 // If the store dominates the block and if we haven't processed it yet,
251 if (dominates(OnlyStore->getParent(), UsingBlocks[i]))
252 if (ProcessedBlocks.insert(UsingBlocks[i]).second) {
253 BasicBlock *UseBlock = UsingBlocks[i];
255 // If the use and store are in the same block, do a quick scan to
256 // verify that there are no uses before the store.
257 if (UseBlock == OnlyStore->getParent()) {
258 BasicBlock::iterator I = UseBlock->begin();
259 for (; &*I != OnlyStore; ++I) { // scan block for store.
260 if (isa<LoadInst>(I) && I->getOperand(0) == AI)
263 if (&*I != OnlyStore) break; // Do not handle this case.
266 // Otherwise, if this is a different block or if all uses happen
267 // after the store, do a simple linear scan to replace loads with
269 for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end();
271 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
272 if (LI->getOperand(0) == AI) {
273 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
274 if (AST && isa<PointerType>(LI->getType()))
275 AST->deleteValue(LI);
276 LI->eraseFromParent();
281 // Finally, remove this block from the UsingBlock set.
282 UsingBlocks[i] = UsingBlocks.back();
286 // Finally, after the scan, check to see if the store is all that is left.
287 if (UsingBlocks.empty()) {
288 // The alloca has been processed, move on.
289 Allocas[AllocaNum] = Allocas.back();
298 PointerAllocaValues[AllocaNum] = AllocaPointerVal;
300 // If we haven't computed a numbering for the BB's in the function, do so
302 if (BBNumbers.empty()) {
304 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
308 // Compute the locations where PhiNodes need to be inserted. Look at the
309 // dominance frontier of EACH basic-block we have a write in.
311 unsigned CurrentVersion = 0;
312 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
313 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
314 while (!DefiningBlocks.empty()) {
315 BasicBlock *BB = DefiningBlocks.back();
316 DefiningBlocks.pop_back();
318 // Look up the DF for this write, add it to PhiNodes
319 DominanceFrontier::const_iterator it = DF.find(BB);
320 if (it != DF.end()) {
321 const DominanceFrontier::DomSetType &S = it->second;
323 // In theory we don't need the indirection through the DFBlocks vector.
324 // In practice, the order of calling QueuePhiNode would depend on the
325 // (unspecified) ordering of basic blocks in the dominance frontier,
326 // which would give PHI nodes non-determinstic subscripts. Fix this by
327 // processing blocks in order of the occurance in the function.
328 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
329 PE = S.end(); P != PE; ++P)
330 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
332 // Sort by which the block ordering in the function.
333 std::sort(DFBlocks.begin(), DFBlocks.end());
335 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
336 BasicBlock *BB = DFBlocks[i].second;
337 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
338 DefiningBlocks.push_back(BB);
344 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
345 // of the writes to the variable, scan through the reads of the variable,
346 // marking PHI nodes which are actually necessary as alive (by removing them
347 // from the InsertedPHINodes set). This is not perfect: there may PHI
348 // marked alive because of loads which are dominated by stores, but there
349 // will be no unmarked PHI nodes which are actually used.
351 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
352 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
355 // If there are any PHI nodes which are now known to be dead, remove them!
356 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(),
357 E = InsertedPHINodes.end(); I != E; ++I) {
359 bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum));
361 assert(Erased && "PHI already removed?");
363 if (AST && isa<PointerType>(PN->getType()))
364 AST->deleteValue(PN);
365 PN->eraseFromParent();
366 PhiToAllocaMap.erase(PN);
369 // Keep the reverse mapping of the 'Allocas' array.
370 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
373 // Process all allocas which are only used in a single basic block.
374 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
375 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
376 const std::vector<AllocaInst*> &LocAllocas = I->second;
377 assert(!LocAllocas.empty() && "empty alloca list??");
379 // It's common for there to only be one alloca in the list. Handle it
381 if (LocAllocas.size() == 1) {
382 // If we can do the quick promotion pass, do so now.
383 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
384 RetryList.push_back(LocAllocas[0]); // Failed, retry later.
386 // Locally promote anything possible. Note that if this is unable to
387 // promote a particular alloca, it puts the alloca onto the Allocas vector
388 // for global processing.
389 PromoteLocallyUsedAllocas(I->first, LocAllocas);
394 return; // All of the allocas must have been trivial!
396 // Set the incoming values for the basic block to be null values for all of
397 // the alloca's. We do this in case there is a load of a value that has not
398 // been stored yet. In this case, it will get this null value.
400 std::vector<Value *> Values(Allocas.size());
401 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
402 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
404 // Walks all basic blocks in the function performing the SSA rename algorithm
405 // and inserting the phi nodes we marked as necessary
407 RenamePassWorkList.clear();
408 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
409 while(!RenamePassWorkList.empty()) {
410 RenamePassData RPD = RenamePassWorkList.back();
411 RenamePassWorkList.pop_back();
412 // RenamePass may add new worklist entries.
413 RenamePass(RPD.BB, RPD.Pred, RPD.Values);
416 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
419 // Remove the allocas themselves from the function.
420 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
421 Instruction *A = Allocas[i];
423 // If there are any uses of the alloca instructions left, they must be in
424 // sections of dead code that were not processed on the dominance frontier.
425 // Just delete the users now.
428 A->replaceAllUsesWith(UndefValue::get(A->getType()));
429 if (AST) AST->deleteValue(A);
430 A->eraseFromParent();
434 // Loop over all of the PHI nodes and see if there are any that we can get
435 // rid of because they merge all of the same incoming values. This can
436 // happen due to undef values coming into the PHI nodes. This process is
437 // iterative, because eliminating one PHI node can cause others to be removed.
438 bool EliminatedAPHI = true;
439 while (EliminatedAPHI) {
440 EliminatedAPHI = false;
442 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
443 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
444 PHINode *PN = I->second;
446 // If this PHI node merges one value and/or undefs, get the value.
447 if (Value *V = PN->hasConstantValue(true)) {
448 if (!isa<Instruction>(V) ||
449 properlyDominates(cast<Instruction>(V), PN)) {
450 if (AST && isa<PointerType>(PN->getType()))
451 AST->deleteValue(PN);
452 PN->replaceAllUsesWith(V);
453 PN->eraseFromParent();
454 NewPhiNodes.erase(I++);
455 EliminatedAPHI = true;
463 // At this point, the renamer has added entries to PHI nodes for all reachable
464 // code. Unfortunately, there may be unreachable blocks which the renamer
465 // hasn't traversed. If this is the case, the PHI nodes may not
466 // have incoming values for all predecessors. Loop over all PHI nodes we have
467 // created, inserting undef values if they are missing any incoming values.
469 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
470 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
471 // We want to do this once per basic block. As such, only process a block
472 // when we find the PHI that is the first entry in the block.
473 PHINode *SomePHI = I->second;
474 BasicBlock *BB = SomePHI->getParent();
475 if (&BB->front() != SomePHI)
478 // Count the number of preds for BB.
479 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
481 // Only do work here if there the PHI nodes are missing incoming values. We
482 // know that all PHI nodes that were inserted in a block will have the same
483 // number of incoming values, so we can just check any of them.
484 if (SomePHI->getNumIncomingValues() == Preds.size())
487 // Ok, now we know that all of the PHI nodes are missing entries for some
488 // basic blocks. Start by sorting the incoming predecessors for efficient
490 std::sort(Preds.begin(), Preds.end());
492 // Now we loop through all BB's which have entries in SomePHI and remove
493 // them from the Preds list.
494 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
495 // Do a log(n) search of the Preds list for the entry we want.
496 SmallVector<BasicBlock*, 16>::iterator EntIt =
497 std::lower_bound(Preds.begin(), Preds.end(),
498 SomePHI->getIncomingBlock(i));
499 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
500 "PHI node has entry for a block which is not a predecessor!");
506 // At this point, the blocks left in the preds list must have dummy
507 // entries inserted into every PHI nodes for the block. Update all the phi
508 // nodes in this block that we are inserting (there could be phis before
510 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
511 BasicBlock::iterator BBI = BB->begin();
512 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
513 SomePHI->getNumIncomingValues() == NumBadPreds) {
514 Value *UndefVal = UndefValue::get(SomePHI->getType());
515 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
516 SomePHI->addIncoming(UndefVal, Preds[pred]);
523 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
524 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
525 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
526 // each read of the variable. For each block that reads the variable, this
527 // function is called, which removes used PHI nodes from the DeadPHINodes set.
528 // After all of the reads have been processed, any PHI nodes left in the
529 // DeadPHINodes set are removed.
531 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
532 SmallPtrSet<PHINode*, 16> &DeadPHINodes) {
533 // Scan the immediate dominators of this block looking for a block which has a
534 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
535 for (BasicBlock* DomBB = BB; DomBB; DomBB = ET.getIDom(DomBB)) {
536 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator
537 I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum));
538 if (I != NewPhiNodes.end()) {
539 // Ok, we found an inserted PHI node which dominates this value.
540 PHINode *DominatingPHI = I->second;
542 // Find out if we previously thought it was dead. If so, mark it as being
543 // live by removing it from the DeadPHINodes set.
544 if (DeadPHINodes.erase(DominatingPHI)) {
545 // Now that we have marked the PHI node alive, also mark any PHI nodes
546 // which it might use as being alive as well.
547 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
549 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
555 /// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
556 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
557 /// potentially useless PHI nodes by just performing a single linear pass over
558 /// the basic block using the Alloca.
560 /// If we cannot promote this alloca (because it is read before it is written),
561 /// return true. This is necessary in cases where, due to control flow, the
562 /// alloca is potentially undefined on some control flow paths. e.g. code like
563 /// this is potentially correct:
565 /// for (...) { if (c) { A = undef; undef = B; } }
567 /// ... so long as A is not used before undef is set.
569 bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
570 assert(!AI->use_empty() && "There are no uses of the alloca!");
572 // Handle degenerate cases quickly.
573 if (AI->hasOneUse()) {
574 Instruction *U = cast<Instruction>(AI->use_back());
575 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
576 // Must be a load of uninitialized value.
577 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
578 if (AST && isa<PointerType>(LI->getType()))
579 AST->deleteValue(LI);
581 // Otherwise it must be a store which is never read.
582 assert(isa<StoreInst>(U));
584 BB->getInstList().erase(U);
586 // Uses of the uninitialized memory location shall get undef.
589 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
590 Instruction *Inst = I++;
591 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
592 if (LI->getOperand(0) == AI) {
593 if (!CurVal) return true; // Could not locally promote!
595 // Loads just returns the "current value"...
596 LI->replaceAllUsesWith(CurVal);
597 if (AST && isa<PointerType>(LI->getType()))
598 AST->deleteValue(LI);
599 BB->getInstList().erase(LI);
601 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
602 if (SI->getOperand(1) == AI) {
603 // Store updates the "current value"...
604 CurVal = SI->getOperand(0);
605 BB->getInstList().erase(SI);
611 // After traversing the basic block, there should be no more uses of the
612 // alloca, remove it now.
613 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
614 if (AST) AST->deleteValue(AI);
615 AI->getParent()->getInstList().erase(AI);
619 /// PromoteLocallyUsedAllocas - This method is just like
620 /// PromoteLocallyUsedAlloca, except that it processes multiple alloca
621 /// instructions in parallel. This is important in cases where we have large
622 /// basic blocks, as we don't want to rescan the entire basic block for each
623 /// alloca which is locally used in it (which might be a lot).
624 void PromoteMem2Reg::
625 PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
626 std::map<AllocaInst*, Value*> CurValues;
627 for (unsigned i = 0, e = AIs.size(); i != e; ++i)
628 CurValues[AIs[i]] = 0; // Insert with null value
630 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
631 Instruction *Inst = I++;
632 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
633 // Is this a load of an alloca we are tracking?
634 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
635 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
636 if (AIt != CurValues.end()) {
637 // If loading an uninitialized value, allow the inter-block case to
638 // handle it. Due to control flow, this might actually be ok.
639 if (AIt->second == 0) { // Use of locally uninitialized value??
640 RetryList.push_back(AI); // Retry elsewhere.
641 CurValues.erase(AIt); // Stop tracking this here.
642 if (CurValues.empty()) return;
644 // Loads just returns the "current value"...
645 LI->replaceAllUsesWith(AIt->second);
646 if (AST && isa<PointerType>(LI->getType()))
647 AST->deleteValue(LI);
648 BB->getInstList().erase(LI);
652 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
653 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
654 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
655 if (AIt != CurValues.end()) {
656 // Store updates the "current value"...
657 AIt->second = SI->getOperand(0);
658 BB->getInstList().erase(SI);
667 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
668 // Alloca returns true if there wasn't already a phi-node for that variable
670 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
672 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
673 // Look up the basic-block in question.
674 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
676 // If the BB already has a phi node added for the i'th alloca then we're done!
677 if (PN) return false;
679 // Create a PhiNode using the dereferenced type... and add the phi-node to the
681 PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
682 Allocas[AllocaNo]->getName() + "." +
683 utostr(Version++), BB->begin());
684 PhiToAllocaMap[PN] = AllocaNo;
686 InsertedPHINodes.insert(PN);
688 if (AST && isa<PointerType>(PN->getType()))
689 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
695 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
696 // stores to the allocas which we are promoting. IncomingVals indicates what
697 // value each Alloca contains on exit from the predecessor block Pred.
699 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
700 std::vector<Value*> &IncomingVals) {
701 // If we are inserting any phi nodes into this BB, they will already be in the
703 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
704 // Pred may have multiple edges to BB. If so, we want to add N incoming
705 // values to each PHI we are inserting on the first time we see the edge.
706 // Check to see if APN already has incoming values from Pred. This also
707 // prevents us from modifying PHI nodes that are not currently being
709 bool HasPredEntries = false;
710 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
711 if (APN->getIncomingBlock(i) == Pred) {
712 HasPredEntries = true;
717 // If we have PHI nodes to update, compute the number of edges from Pred to
719 if (!HasPredEntries) {
720 TerminatorInst *PredTerm = Pred->getTerminator();
721 unsigned NumEdges = 0;
722 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
723 if (PredTerm->getSuccessor(i) == BB)
726 assert(NumEdges && "Must be at least one edge from Pred to BB!");
728 // Add entries for all the phis.
729 BasicBlock::iterator PNI = BB->begin();
731 unsigned AllocaNo = PhiToAllocaMap[APN];
733 // Add N incoming values to the PHI node.
734 for (unsigned i = 0; i != NumEdges; ++i)
735 APN->addIncoming(IncomingVals[AllocaNo], Pred);
737 // The currently active variable for this block is now the PHI.
738 IncomingVals[AllocaNo] = APN;
740 // Get the next phi node.
742 APN = dyn_cast<PHINode>(PNI);
745 // Verify it doesn't already have entries for Pred. If it does, it is
746 // not being inserted by this mem2reg invocation.
747 HasPredEntries = false;
748 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
749 if (APN->getIncomingBlock(i) == Pred) {
750 HasPredEntries = true;
754 } while (!HasPredEntries);
758 // Don't revisit blocks.
759 if (!Visited.insert(BB)) return;
761 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
762 Instruction *I = II++; // get the instruction, increment iterator
764 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
765 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
766 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
767 if (AI != AllocaLookup.end()) {
768 Value *V = IncomingVals[AI->second];
770 // walk the use list of this load and replace all uses with r
771 LI->replaceAllUsesWith(V);
772 if (AST && isa<PointerType>(LI->getType()))
773 AST->deleteValue(LI);
774 BB->getInstList().erase(LI);
777 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
778 // Delete this instruction and mark the name as the current holder of the
780 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
781 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
782 if (ai != AllocaLookup.end()) {
783 // what value were we writing?
784 IncomingVals[ai->second] = SI->getOperand(0);
785 BB->getInstList().erase(SI);
791 // Recurse to our successors.
792 TerminatorInst *TI = BB->getTerminator();
793 for (unsigned i = 0; i != TI->getNumSuccessors(); i++)
794 RenamePassWorkList.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
797 /// PromoteMemToReg - Promote the specified list of alloca instructions into
798 /// scalar registers, inserting PHI nodes as appropriate. This function makes
799 /// use of DominanceFrontier information. This function does not modify the CFG
800 /// of the function at all. All allocas must be from the same function.
802 /// If AST is specified, the specified tracker is updated to reflect changes
805 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
806 ETForest &ET, DominanceFrontier &DF,
807 AliasSetTracker *AST) {
808 // If there is nothing to do, bail out...
809 if (Allocas.empty()) return;
811 SmallVector<AllocaInst*, 16> RetryList;
812 PromoteMem2Reg(Allocas, RetryList, ET, DF, AST).run();
814 // PromoteMem2Reg may not have been able to promote all of the allocas in one
815 // pass, run it again if needed.
816 std::vector<AllocaInst*> NewAllocas;
817 while (!RetryList.empty()) {
818 // If we need to retry some allocas, this is due to there being no store
819 // before a read in a local block. To counteract this, insert a store of
820 // undef into the alloca right after the alloca itself.
821 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
822 BasicBlock::iterator BBI = RetryList[i];
824 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
825 RetryList[i], ++BBI);
828 NewAllocas.assign(RetryList.begin(), RetryList.end());
830 PromoteMem2Reg(NewAllocas, RetryList, ET, DF, AST).run();