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, const TargetData &TD) {
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 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;
95 /// AST - An AliasSetTracker object to update. If null, don't update it.
99 /// AllocaLookup - Reverse mapping of Allocas.
101 std::map<AllocaInst*, unsigned> AllocaLookup;
103 /// NewPhiNodes - The PhiNodes we're adding.
105 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
107 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
108 /// it corresponds to.
109 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
111 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
112 /// each alloca that is of pointer type, we keep track of what to copyValue
113 /// to the inserted PHI nodes here.
115 std::vector<Value*> PointerAllocaValues;
117 /// Visited - The set of basic blocks the renamer has already visited.
119 SmallPtrSet<BasicBlock*, 16> Visited;
121 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
122 /// non-determinstic behavior.
123 DenseMap<BasicBlock*, unsigned> BBNumbers;
125 /// RenamePassWorkList - Worklist used by RenamePass()
126 std::vector<RenamePassData *> RenamePassWorkList;
129 PromoteMem2Reg(const std::vector<AllocaInst*> &A,
130 SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
131 DominanceFrontier &df, const TargetData &td,
132 AliasSetTracker *ast)
133 : Allocas(A), RetryList(Retry), DT(dt), DF(df), TD(td), AST(ast) {}
137 /// properlyDominates - Return true if I1 properly dominates I2.
139 bool properlyDominates(Instruction *I1, Instruction *I2) const {
140 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
141 I1 = II->getNormalDest()->begin();
142 return DT[I1->getParent()]->properlyDominates(DT[I2->getParent()]);
145 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
147 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
148 return DT[BB1]->dominates(DT[BB2]);
152 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
153 SmallPtrSet<PHINode*, 16> &DeadPHINodes);
154 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
155 void PromoteLocallyUsedAllocas(BasicBlock *BB,
156 const std::vector<AllocaInst*> &AIs);
158 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
159 std::vector<Value*> &IncVals);
160 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
161 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
164 } // end of anonymous namespace
166 void PromoteMem2Reg::run() {
167 Function &F = *DF.getRoot()->getParent();
169 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
170 // only used in a single basic block. These instructions can be efficiently
171 // promoted by performing a single linear scan over that one block. Since
172 // individual basic blocks are sometimes large, we group together all allocas
173 // that are live in a single basic block by the basic block they are live in.
174 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
176 if (AST) PointerAllocaValues.resize(Allocas.size());
178 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
179 AllocaInst *AI = Allocas[AllocaNum];
181 assert(isAllocaPromotable(AI, TD) &&
182 "Cannot promote non-promotable alloca!");
183 assert(AI->getParent()->getParent() == &F &&
184 "All allocas should be in the same function, which is same as DF!");
186 if (AI->use_empty()) {
187 // If there are no uses of the alloca, just delete it now.
188 if (AST) AST->deleteValue(AI);
189 AI->eraseFromParent();
191 // Remove the alloca from the Allocas list, since it has been processed
192 Allocas[AllocaNum] = Allocas.back();
198 // Calculate the set of read and write-locations for each alloca. This is
199 // analogous to finding the 'uses' and 'definitions' of each variable.
200 std::vector<BasicBlock*> DefiningBlocks;
201 std::vector<BasicBlock*> UsingBlocks;
203 StoreInst *OnlyStore = 0;
204 BasicBlock *OnlyBlock = 0;
205 bool OnlyUsedInOneBlock = true;
207 // As we scan the uses of the alloca instruction, keep track of stores, and
208 // decide whether all of the loads and stores to the alloca are within the
210 Value *AllocaPointerVal = 0;
211 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
212 Instruction *User = cast<Instruction>(*U);
213 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
214 // Remember the basic blocks which define new values for the alloca
215 DefiningBlocks.push_back(SI->getParent());
216 AllocaPointerVal = SI->getOperand(0);
219 LoadInst *LI = cast<LoadInst>(User);
220 // Otherwise it must be a load instruction, keep track of variable reads
221 UsingBlocks.push_back(LI->getParent());
222 AllocaPointerVal = LI;
225 if (OnlyUsedInOneBlock) {
227 OnlyBlock = User->getParent();
228 else if (OnlyBlock != User->getParent())
229 OnlyUsedInOneBlock = false;
233 // If the alloca is only read and written in one basic block, just perform a
234 // linear sweep over the block to eliminate it.
235 if (OnlyUsedInOneBlock) {
236 LocallyUsedAllocas[OnlyBlock].push_back(AI);
238 // Remove the alloca from the Allocas list, since it will be processed.
239 Allocas[AllocaNum] = Allocas.back();
245 // If there is only a single store to this value, replace any loads of
246 // it that are directly dominated by the definition with the value stored.
247 if (DefiningBlocks.size() == 1) {
248 // Be aware of loads before the store.
249 std::set<BasicBlock*> ProcessedBlocks;
250 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
251 // If the store dominates the block and if we haven't processed it yet,
253 if (dominates(OnlyStore->getParent(), UsingBlocks[i]))
254 if (ProcessedBlocks.insert(UsingBlocks[i]).second) {
255 BasicBlock *UseBlock = UsingBlocks[i];
257 // If the use and store are in the same block, do a quick scan to
258 // verify that there are no uses before the store.
259 if (UseBlock == OnlyStore->getParent()) {
260 BasicBlock::iterator I = UseBlock->begin();
261 for (; &*I != OnlyStore; ++I) { // scan block for store.
262 if (isa<LoadInst>(I) && I->getOperand(0) == AI)
265 if (&*I != OnlyStore) break; // Do not handle this case.
268 // Otherwise, if this is a different block or if all uses happen
269 // after the store, do a simple linear scan to replace loads with
271 for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end();
273 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
274 if (LI->getOperand(0) == AI) {
275 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
276 if (AST && isa<PointerType>(LI->getType()))
277 AST->deleteValue(LI);
278 LI->eraseFromParent();
283 // Finally, remove this block from the UsingBlock set.
284 UsingBlocks[i] = UsingBlocks.back();
288 // Finally, after the scan, check to see if the store is all that is left.
289 if (UsingBlocks.empty()) {
290 // The alloca has been processed, move on.
291 Allocas[AllocaNum] = Allocas.back();
300 PointerAllocaValues[AllocaNum] = AllocaPointerVal;
302 // If we haven't computed a numbering for the BB's in the function, do so
304 if (BBNumbers.empty()) {
306 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
310 // Compute the locations where PhiNodes need to be inserted. Look at the
311 // dominance frontier of EACH basic-block we have a write in.
313 unsigned CurrentVersion = 0;
314 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
315 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
316 while (!DefiningBlocks.empty()) {
317 BasicBlock *BB = DefiningBlocks.back();
318 DefiningBlocks.pop_back();
320 // Look up the DF for this write, add it to PhiNodes
321 DominanceFrontier::const_iterator it = DF.find(BB);
322 if (it != DF.end()) {
323 const DominanceFrontier::DomSetType &S = it->second;
325 // In theory we don't need the indirection through the DFBlocks vector.
326 // In practice, the order of calling QueuePhiNode would depend on the
327 // (unspecified) ordering of basic blocks in the dominance frontier,
328 // which would give PHI nodes non-determinstic subscripts. Fix this by
329 // processing blocks in order of the occurance in the function.
330 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
331 PE = S.end(); P != PE; ++P)
332 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
334 // Sort by which the block ordering in the function.
335 std::sort(DFBlocks.begin(), DFBlocks.end());
337 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
338 BasicBlock *BB = DFBlocks[i].second;
339 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
340 DefiningBlocks.push_back(BB);
346 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
347 // of the writes to the variable, scan through the reads of the variable,
348 // marking PHI nodes which are actually necessary as alive (by removing them
349 // from the InsertedPHINodes set). This is not perfect: there may PHI
350 // marked alive because of loads which are dominated by stores, but there
351 // will be no unmarked PHI nodes which are actually used.
353 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
354 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
357 // If there are any PHI nodes which are now known to be dead, remove them!
358 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(),
359 E = InsertedPHINodes.end(); I != E; ++I) {
361 bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum));
363 assert(Erased && "PHI already removed?");
365 if (AST && isa<PointerType>(PN->getType()))
366 AST->deleteValue(PN);
367 PN->eraseFromParent();
368 PhiToAllocaMap.erase(PN);
371 // Keep the reverse mapping of the 'Allocas' array.
372 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
375 // Process all allocas which are only used in a single basic block.
376 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
377 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
378 const std::vector<AllocaInst*> &LocAllocas = I->second;
379 assert(!LocAllocas.empty() && "empty alloca list??");
381 // It's common for there to only be one alloca in the list. Handle it
383 if (LocAllocas.size() == 1) {
384 // If we can do the quick promotion pass, do so now.
385 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
386 RetryList.push_back(LocAllocas[0]); // Failed, retry later.
388 // Locally promote anything possible. Note that if this is unable to
389 // promote a particular alloca, it puts the alloca onto the Allocas vector
390 // for global processing.
391 PromoteLocallyUsedAllocas(I->first, LocAllocas);
396 return; // All of the allocas must have been trivial!
398 // Set the incoming values for the basic block to be null values for all of
399 // the alloca's. We do this in case there is a load of a value that has not
400 // been stored yet. In this case, it will get this null value.
402 std::vector<Value *> Values(Allocas.size());
403 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
404 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
406 // Walks all basic blocks in the function performing the SSA rename algorithm
407 // and inserting the phi nodes we marked as necessary
409 //RenamePass(F.begin(), 0, Values);
410 RenamePassWorkList.clear();
411 RenamePassData *RPD = new RenamePassData(F.begin(), 0, Values);
412 RenamePassWorkList.push_back(RPD);
413 while(!RenamePassWorkList.empty()) {
414 RenamePassData *RPD = RenamePassWorkList.back(); RenamePassWorkList.pop_back();
415 // RenamePass may add new worklist entries.
416 RenamePass(RPD->BB, RPD->Pred, RPD->Values);
420 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
423 // Remove the allocas themselves from the function.
424 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
425 Instruction *A = Allocas[i];
427 // If there are any uses of the alloca instructions left, they must be in
428 // sections of dead code that were not processed on the dominance frontier.
429 // Just delete the users now.
432 A->replaceAllUsesWith(UndefValue::get(A->getType()));
433 if (AST) AST->deleteValue(A);
434 A->eraseFromParent();
438 // Loop over all of the PHI nodes and see if there are any that we can get
439 // rid of because they merge all of the same incoming values. This can
440 // happen due to undef values coming into the PHI nodes. This process is
441 // iterative, because eliminating one PHI node can cause others to be removed.
442 bool EliminatedAPHI = true;
443 while (EliminatedAPHI) {
444 EliminatedAPHI = false;
446 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
447 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
448 PHINode *PN = I->second;
450 // If this PHI node merges one value and/or undefs, get the value.
451 if (Value *V = PN->hasConstantValue(true)) {
452 if (!isa<Instruction>(V) ||
453 properlyDominates(cast<Instruction>(V), PN)) {
454 if (AST && isa<PointerType>(PN->getType()))
455 AST->deleteValue(PN);
456 PN->replaceAllUsesWith(V);
457 PN->eraseFromParent();
458 NewPhiNodes.erase(I++);
459 EliminatedAPHI = true;
467 // At this point, the renamer has added entries to PHI nodes for all reachable
468 // code. Unfortunately, there may be unreachable blocks which the renamer
469 // hasn't traversed. If this is the case, the PHI nodes may not
470 // have incoming values for all predecessors. Loop over all PHI nodes we have
471 // created, inserting undef values if they are missing any incoming values.
473 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
474 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
475 // We want to do this once per basic block. As such, only process a block
476 // when we find the PHI that is the first entry in the block.
477 PHINode *SomePHI = I->second;
478 BasicBlock *BB = SomePHI->getParent();
479 if (&BB->front() != SomePHI)
482 // Count the number of preds for BB.
483 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
485 // Only do work here if there the PHI nodes are missing incoming values. We
486 // know that all PHI nodes that were inserted in a block will have the same
487 // number of incoming values, so we can just check any of them.
488 if (SomePHI->getNumIncomingValues() == Preds.size())
491 // Ok, now we know that all of the PHI nodes are missing entries for some
492 // basic blocks. Start by sorting the incoming predecessors for efficient
494 std::sort(Preds.begin(), Preds.end());
496 // Now we loop through all BB's which have entries in SomePHI and remove
497 // them from the Preds list.
498 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
499 // Do a log(n) search of the Preds list for the entry we want.
500 SmallVector<BasicBlock*, 16>::iterator EntIt =
501 std::lower_bound(Preds.begin(), Preds.end(),
502 SomePHI->getIncomingBlock(i));
503 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
504 "PHI node has entry for a block which is not a predecessor!");
510 // At this point, the blocks left in the preds list must have dummy
511 // entries inserted into every PHI nodes for the block. Update all the phi
512 // nodes in this block that we are inserting (there could be phis before
514 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
515 BasicBlock::iterator BBI = BB->begin();
516 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
517 SomePHI->getNumIncomingValues() == NumBadPreds) {
518 Value *UndefVal = UndefValue::get(SomePHI->getType());
519 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
520 SomePHI->addIncoming(UndefVal, Preds[pred]);
527 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
528 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
529 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
530 // each read of the variable. For each block that reads the variable, this
531 // function is called, which removes used PHI nodes from the DeadPHINodes set.
532 // After all of the reads have been processed, any PHI nodes left in the
533 // DeadPHINodes set are removed.
535 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
536 SmallPtrSet<PHINode*, 16> &DeadPHINodes) {
537 // Scan the immediate dominators of this block looking for a block which has a
538 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
539 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) {
540 BasicBlock *DomBB = N->getBlock();
541 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator
542 I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum));
543 if (I != NewPhiNodes.end()) {
544 // Ok, we found an inserted PHI node which dominates this value.
545 PHINode *DominatingPHI = I->second;
547 // Find out if we previously thought it was dead. If so, mark it as being
548 // live by removing it from the DeadPHINodes set.
549 if (DeadPHINodes.erase(DominatingPHI)) {
550 // Now that we have marked the PHI node alive, also mark any PHI nodes
551 // which it might use as being alive as well.
552 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
554 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
560 /// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
561 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
562 /// potentially useless PHI nodes by just performing a single linear pass over
563 /// the basic block using the Alloca.
565 /// If we cannot promote this alloca (because it is read before it is written),
566 /// return true. This is necessary in cases where, due to control flow, the
567 /// alloca is potentially undefined on some control flow paths. e.g. code like
568 /// this is potentially correct:
570 /// for (...) { if (c) { A = undef; undef = B; } }
572 /// ... so long as A is not used before undef is set.
574 bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
575 assert(!AI->use_empty() && "There are no uses of the alloca!");
577 // Handle degenerate cases quickly.
578 if (AI->hasOneUse()) {
579 Instruction *U = cast<Instruction>(AI->use_back());
580 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
581 // Must be a load of uninitialized value.
582 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
583 if (AST && isa<PointerType>(LI->getType()))
584 AST->deleteValue(LI);
586 // Otherwise it must be a store which is never read.
587 assert(isa<StoreInst>(U));
589 BB->getInstList().erase(U);
591 // Uses of the uninitialized memory location shall get undef.
594 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
595 Instruction *Inst = I++;
596 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
597 if (LI->getOperand(0) == AI) {
598 if (!CurVal) return true; // Could not locally promote!
600 // Loads just returns the "current value"...
601 LI->replaceAllUsesWith(CurVal);
602 if (AST && isa<PointerType>(LI->getType()))
603 AST->deleteValue(LI);
604 BB->getInstList().erase(LI);
606 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
607 if (SI->getOperand(1) == AI) {
608 // Store updates the "current value"...
609 CurVal = SI->getOperand(0);
610 BB->getInstList().erase(SI);
616 // After traversing the basic block, there should be no more uses of the
617 // alloca, remove it now.
618 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
619 if (AST) AST->deleteValue(AI);
620 AI->getParent()->getInstList().erase(AI);
624 /// PromoteLocallyUsedAllocas - This method is just like
625 /// PromoteLocallyUsedAlloca, except that it processes multiple alloca
626 /// instructions in parallel. This is important in cases where we have large
627 /// basic blocks, as we don't want to rescan the entire basic block for each
628 /// alloca which is locally used in it (which might be a lot).
629 void PromoteMem2Reg::
630 PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
631 std::map<AllocaInst*, Value*> CurValues;
632 for (unsigned i = 0, e = AIs.size(); i != e; ++i)
633 CurValues[AIs[i]] = 0; // Insert with null value
635 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
636 Instruction *Inst = I++;
637 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
638 // Is this a load of an alloca we are tracking?
639 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
640 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
641 if (AIt != CurValues.end()) {
642 // If loading an uninitialized value, allow the inter-block case to
643 // handle it. Due to control flow, this might actually be ok.
644 if (AIt->second == 0) { // Use of locally uninitialized value??
645 RetryList.push_back(AI); // Retry elsewhere.
646 CurValues.erase(AIt); // Stop tracking this here.
647 if (CurValues.empty()) return;
649 // Loads just returns the "current value"...
650 LI->replaceAllUsesWith(AIt->second);
651 if (AST && isa<PointerType>(LI->getType()))
652 AST->deleteValue(LI);
653 BB->getInstList().erase(LI);
657 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
658 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
659 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
660 if (AIt != CurValues.end()) {
661 // Store updates the "current value"...
662 AIt->second = SI->getOperand(0);
663 BB->getInstList().erase(SI);
672 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
673 // Alloca returns true if there wasn't already a phi-node for that variable
675 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
677 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
678 // Look up the basic-block in question.
679 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
681 // If the BB already has a phi node added for the i'th alloca then we're done!
682 if (PN) return false;
684 // Create a PhiNode using the dereferenced type... and add the phi-node to the
686 PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
687 Allocas[AllocaNo]->getName() + "." +
688 utostr(Version++), BB->begin());
689 PhiToAllocaMap[PN] = AllocaNo;
691 InsertedPHINodes.insert(PN);
693 if (AST && isa<PointerType>(PN->getType()))
694 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
700 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
701 // stores to the allocas which we are promoting. IncomingVals indicates what
702 // value each Alloca contains on exit from the predecessor block Pred.
704 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
705 std::vector<Value*> &IncomingVals) {
706 // If we are inserting any phi nodes into this BB, they will already be in the
708 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
709 // Pred may have multiple edges to BB. If so, we want to add N incoming
710 // values to each PHI we are inserting on the first time we see the edge.
711 // Check to see if APN already has incoming values from Pred. This also
712 // prevents us from modifying PHI nodes that are not currently being
714 bool HasPredEntries = false;
715 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
716 if (APN->getIncomingBlock(i) == Pred) {
717 HasPredEntries = true;
722 // If we have PHI nodes to update, compute the number of edges from Pred to
724 if (!HasPredEntries) {
725 TerminatorInst *PredTerm = Pred->getTerminator();
726 unsigned NumEdges = 0;
727 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
728 if (PredTerm->getSuccessor(i) == BB)
731 assert(NumEdges && "Must be at least one edge from Pred to BB!");
733 // Add entries for all the phis.
734 BasicBlock::iterator PNI = BB->begin();
736 unsigned AllocaNo = PhiToAllocaMap[APN];
738 // Add N incoming values to the PHI node.
739 for (unsigned i = 0; i != NumEdges; ++i)
740 APN->addIncoming(IncomingVals[AllocaNo], Pred);
742 // The currently active variable for this block is now the PHI.
743 IncomingVals[AllocaNo] = APN;
745 // Get the next phi node.
747 APN = dyn_cast<PHINode>(PNI);
750 // Verify it doesn't already have entries for Pred. If it does, it is
751 // not being inserted by this mem2reg invocation.
752 HasPredEntries = false;
753 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
754 if (APN->getIncomingBlock(i) == Pred) {
755 HasPredEntries = true;
759 } while (!HasPredEntries);
763 // Don't revisit blocks.
764 if (!Visited.insert(BB)) return;
766 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
767 Instruction *I = II++; // get the instruction, increment iterator
769 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
770 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
771 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
772 if (AI != AllocaLookup.end()) {
773 Value *V = IncomingVals[AI->second];
775 // walk the use list of this load and replace all uses with r
776 LI->replaceAllUsesWith(V);
777 if (AST && isa<PointerType>(LI->getType()))
778 AST->deleteValue(LI);
779 BB->getInstList().erase(LI);
782 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
783 // Delete this instruction and mark the name as the current holder of the
785 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
786 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
787 if (ai != AllocaLookup.end()) {
788 // what value were we writing?
789 IncomingVals[ai->second] = SI->getOperand(0);
790 BB->getInstList().erase(SI);
796 // Recurse to our successors.
797 TerminatorInst *TI = BB->getTerminator();
798 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
799 RenamePassData *RPD = new RenamePassData(TI->getSuccessor(i), BB,
801 RenamePassWorkList.push_back(RPD);
802 // std::vector<Value*> OutgoingVals(IncomingVals);
803 // RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
807 /// PromoteMemToReg - Promote the specified list of alloca instructions into
808 /// scalar registers, inserting PHI nodes as appropriate. This function makes
809 /// use of DominanceFrontier information. This function does not modify the CFG
810 /// of the function at all. All allocas must be from the same function.
812 /// If AST is specified, the specified tracker is updated to reflect changes
815 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
816 DominatorTree &DT, DominanceFrontier &DF,
817 const TargetData &TD, AliasSetTracker *AST) {
818 // If there is nothing to do, bail out...
819 if (Allocas.empty()) return;
821 SmallVector<AllocaInst*, 16> RetryList;
822 PromoteMem2Reg(Allocas, RetryList, DT, DF, TD, AST).run();
824 // PromoteMem2Reg may not have been able to promote all of the allocas in one
825 // pass, run it again if needed.
826 std::vector<AllocaInst*> NewAllocas;
827 while (!RetryList.empty()) {
828 // If we need to retry some allocas, this is due to there being no store
829 // before a read in a local block. To counteract this, insert a store of
830 // undef into the alloca right after the alloca itself.
831 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
832 BasicBlock::iterator BBI = RetryList[i];
834 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
835 RetryList[i], ++BBI);
838 NewAllocas.assign(RetryList.begin(), RetryList.end());
840 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, TD, AST).run();