1 //===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file promotes memory references to be register references. It promotes
11 // alloca instructions which only have loads and stores as uses. An alloca is
12 // transformed by using dominator frontiers to place PHI nodes, then traversing
13 // the function in depth-first order to rewrite loads and stores as appropriate.
14 // This is just the standard SSA construction algorithm to construct "pruned"
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "mem2reg"
20 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/IntrinsicInst.h"
26 #include "llvm/Metadata.h"
27 #include "llvm/Analysis/DebugInfo.h"
28 #include "llvm/Analysis/Dominators.h"
29 #include "llvm/Analysis/AliasSetTracker.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/STLExtras.h"
35 #include "llvm/Support/CFG.h"
39 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
40 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
41 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
42 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
46 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
47 typedef std::pair<BasicBlock*, unsigned> EltTy;
48 static inline EltTy getEmptyKey() {
49 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
51 static inline EltTy getTombstoneKey() {
52 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
54 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
55 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
57 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
63 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
64 /// This is true if there are only loads and stores to the alloca.
66 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
67 // FIXME: If the memory unit is of pointer or integer type, we can permit
68 // assignments to subsections of the memory unit.
70 // Only allow direct and non-volatile loads and stores...
71 for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end();
72 UI != UE; ++UI) { // Loop over all of the uses of the alloca
74 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
77 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
78 if (SI->getOperand(0) == AI)
79 return false; // Don't allow a store OF the AI, only INTO the AI.
90 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
91 /// alloca 'V', if any.
92 static DbgDeclareInst *FindAllocaDbgDeclare(Value *V) {
93 if (MDNode *DebugNode = MDNode::getIfExists(V->getContext(), &V, 1))
94 for (Value::use_iterator UI = DebugNode->use_begin(),
95 E = DebugNode->use_end(); UI != E; ++UI)
96 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
105 // Data package used by RenamePass()
106 class RenamePassData {
108 typedef std::vector<Value *> ValVector;
110 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
111 RenamePassData(BasicBlock *B, BasicBlock *P,
112 const ValVector &V) : BB(B), Pred(P), Values(V) {}
117 void swap(RenamePassData &RHS) {
118 std::swap(BB, RHS.BB);
119 std::swap(Pred, RHS.Pred);
120 Values.swap(RHS.Values);
124 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
125 /// load/store instructions in the block that directly load or store an alloca.
127 /// This functionality is important because it avoids scanning large basic
128 /// blocks multiple times when promoting many allocas in the same block.
129 class LargeBlockInfo {
130 /// InstNumbers - For each instruction that we track, keep the index of the
131 /// instruction. The index starts out as the number of the instruction from
132 /// the start of the block.
133 DenseMap<const Instruction *, unsigned> InstNumbers;
136 /// isInterestingInstruction - This code only looks at accesses to allocas.
137 static bool isInterestingInstruction(const Instruction *I) {
138 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
139 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
142 /// getInstructionIndex - Get or calculate the index of the specified
144 unsigned getInstructionIndex(const Instruction *I) {
145 assert(isInterestingInstruction(I) &&
146 "Not a load/store to/from an alloca?");
148 // If we already have this instruction number, return it.
149 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
150 if (It != InstNumbers.end()) return It->second;
152 // Scan the whole block to get the instruction. This accumulates
153 // information for every interesting instruction in the block, in order to
154 // avoid gratuitus rescans.
155 const BasicBlock *BB = I->getParent();
157 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
159 if (isInterestingInstruction(BBI))
160 InstNumbers[BBI] = InstNo++;
161 It = InstNumbers.find(I);
163 assert(It != InstNumbers.end() && "Didn't insert instruction?");
167 void deleteValue(const Instruction *I) {
168 InstNumbers.erase(I);
176 struct PromoteMem2Reg {
177 /// Allocas - The alloca instructions being promoted.
179 std::vector<AllocaInst*> Allocas;
181 DominanceFrontier &DF;
184 /// AST - An AliasSetTracker object to update. If null, don't update it.
186 AliasSetTracker *AST;
188 /// AllocaLookup - Reverse mapping of Allocas.
190 std::map<AllocaInst*, unsigned> AllocaLookup;
192 /// NewPhiNodes - The PhiNodes we're adding.
194 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
196 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
197 /// it corresponds to.
198 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
200 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
201 /// each alloca that is of pointer type, we keep track of what to copyValue
202 /// to the inserted PHI nodes here.
204 std::vector<Value*> PointerAllocaValues;
206 /// AllocaDbgDeclares - For each alloca, we keep track of the dbg.declare
207 /// intrinsic that describes it, if any, so that we can convert it to a
208 /// dbg.value intrinsic if the alloca gets promoted.
209 SmallVector<DbgDeclareInst*, 8> AllocaDbgDeclares;
211 /// Visited - The set of basic blocks the renamer has already visited.
213 SmallPtrSet<BasicBlock*, 16> Visited;
215 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
216 /// non-determinstic behavior.
217 DenseMap<BasicBlock*, unsigned> BBNumbers;
219 /// BBNumPreds - Lazily compute the number of predecessors a block has.
220 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
222 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
223 DominanceFrontier &df, AliasSetTracker *ast)
224 : Allocas(A), DT(dt), DF(df), DIF(0), AST(ast) {}
231 /// properlyDominates - Return true if I1 properly dominates I2.
233 bool properlyDominates(Instruction *I1, Instruction *I2) const {
234 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
235 I1 = II->getNormalDest()->begin();
236 return DT.properlyDominates(I1->getParent(), I2->getParent());
239 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
241 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
242 return DT.dominates(BB1, BB2);
246 void RemoveFromAllocasList(unsigned &AllocaIdx) {
247 Allocas[AllocaIdx] = Allocas.back();
252 unsigned getNumPreds(const BasicBlock *BB) {
253 unsigned &NP = BBNumPreds[BB];
255 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
259 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
261 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
262 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
263 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
265 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
266 LargeBlockInfo &LBI);
267 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
268 LargeBlockInfo &LBI);
269 void ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, StoreInst *SI);
272 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
273 RenamePassData::ValVector &IncVals,
274 std::vector<RenamePassData> &Worklist);
275 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
276 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
280 std::vector<BasicBlock*> DefiningBlocks;
281 std::vector<BasicBlock*> UsingBlocks;
283 StoreInst *OnlyStore;
284 BasicBlock *OnlyBlock;
285 bool OnlyUsedInOneBlock;
287 Value *AllocaPointerVal;
288 DbgDeclareInst *DbgDeclare;
291 DefiningBlocks.clear();
295 OnlyUsedInOneBlock = true;
296 AllocaPointerVal = 0;
300 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
302 void AnalyzeAlloca(AllocaInst *AI) {
305 // As we scan the uses of the alloca instruction, keep track of stores,
306 // and decide whether all of the loads and stores to the alloca are within
307 // the same basic block.
308 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
310 Instruction *User = cast<Instruction>(*UI++);
312 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
313 // Remember the basic blocks which define new values for the alloca
314 DefiningBlocks.push_back(SI->getParent());
315 AllocaPointerVal = SI->getOperand(0);
318 LoadInst *LI = cast<LoadInst>(User);
319 // Otherwise it must be a load instruction, keep track of variable
321 UsingBlocks.push_back(LI->getParent());
322 AllocaPointerVal = LI;
325 if (OnlyUsedInOneBlock) {
327 OnlyBlock = User->getParent();
328 else if (OnlyBlock != User->getParent())
329 OnlyUsedInOneBlock = false;
333 DbgDeclare = FindAllocaDbgDeclare(AI);
336 } // end of anonymous namespace
339 void PromoteMem2Reg::run() {
340 Function &F = *DF.getRoot()->getParent();
342 if (AST) PointerAllocaValues.resize(Allocas.size());
343 AllocaDbgDeclares.resize(Allocas.size());
348 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
349 AllocaInst *AI = Allocas[AllocaNum];
351 assert(isAllocaPromotable(AI) &&
352 "Cannot promote non-promotable alloca!");
353 assert(AI->getParent()->getParent() == &F &&
354 "All allocas should be in the same function, which is same as DF!");
356 if (AI->use_empty()) {
357 // If there are no uses of the alloca, just delete it now.
358 if (AST) AST->deleteValue(AI);
359 AI->eraseFromParent();
361 // Remove the alloca from the Allocas list, since it has been processed
362 RemoveFromAllocasList(AllocaNum);
367 // Calculate the set of read and write-locations for each alloca. This is
368 // analogous to finding the 'uses' and 'definitions' of each variable.
369 Info.AnalyzeAlloca(AI);
371 // If there is only a single store to this value, replace any loads of
372 // it that are directly dominated by the definition with the value stored.
373 if (Info.DefiningBlocks.size() == 1) {
374 RewriteSingleStoreAlloca(AI, Info, LBI);
376 // Finally, after the scan, check to see if the store is all that is left.
377 if (Info.UsingBlocks.empty()) {
378 // Record debuginfo for the store and remove the declaration's debuginfo.
379 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
380 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore);
381 DDI->eraseFromParent();
383 // Remove the (now dead) store and alloca.
384 Info.OnlyStore->eraseFromParent();
385 LBI.deleteValue(Info.OnlyStore);
387 if (AST) AST->deleteValue(AI);
388 AI->eraseFromParent();
391 // The alloca has been processed, move on.
392 RemoveFromAllocasList(AllocaNum);
399 // If the alloca is only read and written in one basic block, just perform a
400 // linear sweep over the block to eliminate it.
401 if (Info.OnlyUsedInOneBlock) {
402 PromoteSingleBlockAlloca(AI, Info, LBI);
404 // Finally, after the scan, check to see if the stores are all that is
406 if (Info.UsingBlocks.empty()) {
408 // Remove the (now dead) stores and alloca.
409 while (!AI->use_empty()) {
410 StoreInst *SI = cast<StoreInst>(AI->use_back());
411 // Record debuginfo for the store before removing it.
412 if (DbgDeclareInst *DDI = Info.DbgDeclare)
413 ConvertDebugDeclareToDebugValue(DDI, SI);
414 SI->eraseFromParent();
418 if (AST) AST->deleteValue(AI);
419 AI->eraseFromParent();
422 // The alloca has been processed, move on.
423 RemoveFromAllocasList(AllocaNum);
425 // The alloca's debuginfo can be removed as well.
426 if (DbgDeclareInst *DDI = Info.DbgDeclare)
427 DDI->eraseFromParent();
434 // If we haven't computed a numbering for the BB's in the function, do so
436 if (BBNumbers.empty()) {
438 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
442 // If we have an AST to keep updated, remember some pointer value that is
443 // stored into the alloca.
445 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
447 // Remember the dbg.declare intrinsic describing this alloca, if any.
448 if (Info.DbgDeclare) AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
450 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
451 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
453 // At this point, we're committed to promoting the alloca using IDF's, and
454 // the standard SSA construction algorithm. Determine which blocks need PHI
455 // nodes and see if we can optimize out some work by avoiding insertion of
457 DetermineInsertionPoint(AI, AllocaNum, Info);
461 return; // All of the allocas must have been trivial!
466 // Set the incoming values for the basic block to be null values for all of
467 // the alloca's. We do this in case there is a load of a value that has not
468 // been stored yet. In this case, it will get this null value.
470 RenamePassData::ValVector Values(Allocas.size());
471 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
472 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
474 // Walks all basic blocks in the function performing the SSA rename algorithm
475 // and inserting the phi nodes we marked as necessary
477 std::vector<RenamePassData> RenamePassWorkList;
478 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
481 RPD.swap(RenamePassWorkList.back());
482 RenamePassWorkList.pop_back();
483 // RenamePass may add new worklist entries.
484 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
485 } while (!RenamePassWorkList.empty());
487 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
490 // Remove the allocas themselves from the function.
491 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
492 Instruction *A = Allocas[i];
494 // If there are any uses of the alloca instructions left, they must be in
495 // sections of dead code that were not processed on the dominance frontier.
496 // Just delete the users now.
499 A->replaceAllUsesWith(UndefValue::get(A->getType()));
500 if (AST) AST->deleteValue(A);
501 A->eraseFromParent();
504 // Remove alloca's dbg.declare instrinsics from the function.
505 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
506 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
507 DDI->eraseFromParent();
509 // Loop over all of the PHI nodes and see if there are any that we can get
510 // rid of because they merge all of the same incoming values. This can
511 // happen due to undef values coming into the PHI nodes. This process is
512 // iterative, because eliminating one PHI node can cause others to be removed.
513 bool EliminatedAPHI = true;
514 while (EliminatedAPHI) {
515 EliminatedAPHI = false;
517 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
518 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
519 PHINode *PN = I->second;
521 // If this PHI node merges one value and/or undefs, get the value.
522 if (Value *V = PN->hasConstantValue(&DT)) {
523 if (AST && PN->getType()->isPointerTy())
524 AST->deleteValue(PN);
525 PN->replaceAllUsesWith(V);
526 PN->eraseFromParent();
527 NewPhiNodes.erase(I++);
528 EliminatedAPHI = true;
535 // At this point, the renamer has added entries to PHI nodes for all reachable
536 // code. Unfortunately, there may be unreachable blocks which the renamer
537 // hasn't traversed. If this is the case, the PHI nodes may not
538 // have incoming values for all predecessors. Loop over all PHI nodes we have
539 // created, inserting undef values if they are missing any incoming values.
541 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
542 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
543 // We want to do this once per basic block. As such, only process a block
544 // when we find the PHI that is the first entry in the block.
545 PHINode *SomePHI = I->second;
546 BasicBlock *BB = SomePHI->getParent();
547 if (&BB->front() != SomePHI)
550 // Only do work here if there the PHI nodes are missing incoming values. We
551 // know that all PHI nodes that were inserted in a block will have the same
552 // number of incoming values, so we can just check any of them.
553 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
556 // Get the preds for BB.
557 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
559 // Ok, now we know that all of the PHI nodes are missing entries for some
560 // basic blocks. Start by sorting the incoming predecessors for efficient
562 std::sort(Preds.begin(), Preds.end());
564 // Now we loop through all BB's which have entries in SomePHI and remove
565 // them from the Preds list.
566 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
567 // Do a log(n) search of the Preds list for the entry we want.
568 SmallVector<BasicBlock*, 16>::iterator EntIt =
569 std::lower_bound(Preds.begin(), Preds.end(),
570 SomePHI->getIncomingBlock(i));
571 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
572 "PHI node has entry for a block which is not a predecessor!");
578 // At this point, the blocks left in the preds list must have dummy
579 // entries inserted into every PHI nodes for the block. Update all the phi
580 // nodes in this block that we are inserting (there could be phis before
582 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
583 BasicBlock::iterator BBI = BB->begin();
584 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
585 SomePHI->getNumIncomingValues() == NumBadPreds) {
586 Value *UndefVal = UndefValue::get(SomePHI->getType());
587 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
588 SomePHI->addIncoming(UndefVal, Preds[pred]);
596 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
597 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
598 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
600 void PromoteMem2Reg::
601 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
602 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
603 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
605 // To determine liveness, we must iterate through the predecessors of blocks
606 // where the def is live. Blocks are added to the worklist if we need to
607 // check their predecessors. Start with all the using blocks.
608 SmallVector<BasicBlock*, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
609 Info.UsingBlocks.end());
611 // If any of the using blocks is also a definition block, check to see if the
612 // definition occurs before or after the use. If it happens before the use,
613 // the value isn't really live-in.
614 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
615 BasicBlock *BB = LiveInBlockWorklist[i];
616 if (!DefBlocks.count(BB)) continue;
618 // Okay, this is a block that both uses and defines the value. If the first
619 // reference to the alloca is a def (store), then we know it isn't live-in.
620 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
621 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
622 if (SI->getOperand(1) != AI) continue;
624 // We found a store to the alloca before a load. The alloca is not
625 // actually live-in here.
626 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
627 LiveInBlockWorklist.pop_back();
632 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
633 if (LI->getOperand(0) != AI) continue;
635 // Okay, we found a load before a store to the alloca. It is actually
636 // live into this block.
642 // Now that we have a set of blocks where the phi is live-in, recursively add
643 // their predecessors until we find the full region the value is live.
644 while (!LiveInBlockWorklist.empty()) {
645 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
647 // The block really is live in here, insert it into the set. If already in
648 // the set, then it has already been processed.
649 if (!LiveInBlocks.insert(BB))
652 // Since the value is live into BB, it is either defined in a predecessor or
653 // live into it to. Add the preds to the worklist unless they are a
655 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
658 // The value is not live into a predecessor if it defines the value.
659 if (DefBlocks.count(P))
662 // Otherwise it is, add to the worklist.
663 LiveInBlockWorklist.push_back(P);
668 /// DetermineInsertionPoint - At this point, we're committed to promoting the
669 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
670 /// which blocks need phi nodes and see if we can optimize out some work by
671 /// avoiding insertion of dead phi nodes.
672 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
675 // Unique the set of defining blocks for efficient lookup.
676 SmallPtrSet<BasicBlock*, 32> DefBlocks;
677 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
679 // Determine which blocks the value is live in. These are blocks which lead
681 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
682 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
684 // Compute the locations where PhiNodes need to be inserted. Look at the
685 // dominance frontier of EACH basic-block we have a write in.
686 unsigned CurrentVersion = 0;
687 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
688 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
689 while (!Info.DefiningBlocks.empty()) {
690 BasicBlock *BB = Info.DefiningBlocks.back();
691 Info.DefiningBlocks.pop_back();
693 // Look up the DF for this write, add it to defining blocks.
694 DominanceFrontier::const_iterator it = DF.find(BB);
695 if (it == DF.end()) continue;
697 const DominanceFrontier::DomSetType &S = it->second;
699 // In theory we don't need the indirection through the DFBlocks vector.
700 // In practice, the order of calling QueuePhiNode would depend on the
701 // (unspecified) ordering of basic blocks in the dominance frontier,
702 // which would give PHI nodes non-determinstic subscripts. Fix this by
703 // processing blocks in order of the occurance in the function.
704 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
705 PE = S.end(); P != PE; ++P) {
706 // If the frontier block is not in the live-in set for the alloca, don't
707 // bother processing it.
708 if (!LiveInBlocks.count(*P))
711 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
714 // Sort by which the block ordering in the function.
715 if (DFBlocks.size() > 1)
716 std::sort(DFBlocks.begin(), DFBlocks.end());
718 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
719 BasicBlock *BB = DFBlocks[i].second;
720 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
721 Info.DefiningBlocks.push_back(BB);
727 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
728 /// replace any loads of it that are directly dominated by the definition with
729 /// the value stored.
730 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
732 LargeBlockInfo &LBI) {
733 StoreInst *OnlyStore = Info.OnlyStore;
734 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
735 BasicBlock *StoreBB = OnlyStore->getParent();
738 // Clear out UsingBlocks. We will reconstruct it here if needed.
739 Info.UsingBlocks.clear();
741 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
742 Instruction *UserInst = cast<Instruction>(*UI++);
743 if (!isa<LoadInst>(UserInst)) {
744 assert(UserInst == OnlyStore && "Should only have load/stores");
747 LoadInst *LI = cast<LoadInst>(UserInst);
749 // Okay, if we have a load from the alloca, we want to replace it with the
750 // only value stored to the alloca. We can do this if the value is
751 // dominated by the store. If not, we use the rest of the mem2reg machinery
752 // to insert the phi nodes as needed.
753 if (!StoringGlobalVal) { // Non-instructions are always dominated.
754 if (LI->getParent() == StoreBB) {
755 // If we have a use that is in the same block as the store, compare the
756 // indices of the two instructions to see which one came first. If the
757 // load came before the store, we can't handle it.
758 if (StoreIndex == -1)
759 StoreIndex = LBI.getInstructionIndex(OnlyStore);
761 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
762 // Can't handle this load, bail out.
763 Info.UsingBlocks.push_back(StoreBB);
767 } else if (LI->getParent() != StoreBB &&
768 !dominates(StoreBB, LI->getParent())) {
769 // If the load and store are in different blocks, use BB dominance to
770 // check their relationships. If the store doesn't dom the use, bail
772 Info.UsingBlocks.push_back(LI->getParent());
777 // Otherwise, we *can* safely rewrite this load.
778 Value *ReplVal = OnlyStore->getOperand(0);
779 // If the replacement value is the load, this must occur in unreachable
782 ReplVal = UndefValue::get(LI->getType());
783 LI->replaceAllUsesWith(ReplVal);
784 if (AST && LI->getType()->isPointerTy())
785 AST->deleteValue(LI);
786 LI->eraseFromParent();
793 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
794 /// first element of a pair.
795 struct StoreIndexSearchPredicate {
796 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
797 const std::pair<unsigned, StoreInst*> &RHS) {
798 return LHS.first < RHS.first;
804 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
805 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
806 /// potentially useless PHI nodes by just performing a single linear pass over
807 /// the basic block using the Alloca.
809 /// If we cannot promote this alloca (because it is read before it is written),
810 /// return true. This is necessary in cases where, due to control flow, the
811 /// alloca is potentially undefined on some control flow paths. e.g. code like
812 /// this is potentially correct:
814 /// for (...) { if (c) { A = undef; undef = B; } }
816 /// ... so long as A is not used before undef is set.
818 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
819 LargeBlockInfo &LBI) {
820 // The trickiest case to handle is when we have large blocks. Because of this,
821 // this code is optimized assuming that large blocks happen. This does not
822 // significantly pessimize the small block case. This uses LargeBlockInfo to
823 // make it efficient to get the index of various operations in the block.
825 // Clear out UsingBlocks. We will reconstruct it here if needed.
826 Info.UsingBlocks.clear();
828 // Walk the use-def list of the alloca, getting the locations of all stores.
829 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
830 StoresByIndexTy StoresByIndex;
832 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
834 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
835 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
837 // If there are no stores to the alloca, just replace any loads with undef.
838 if (StoresByIndex.empty()) {
839 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
840 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
841 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
842 if (AST && LI->getType()->isPointerTy())
843 AST->deleteValue(LI);
845 LI->eraseFromParent();
850 // Sort the stores by their index, making it efficient to do a lookup with a
852 std::sort(StoresByIndex.begin(), StoresByIndex.end());
854 // Walk all of the loads from this alloca, replacing them with the nearest
855 // store above them, if any.
856 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
857 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
860 unsigned LoadIdx = LBI.getInstructionIndex(LI);
862 // Find the nearest store that has a lower than this load.
863 StoresByIndexTy::iterator I =
864 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
865 std::pair<unsigned, StoreInst*>(LoadIdx, static_cast<StoreInst*>(0)),
866 StoreIndexSearchPredicate());
868 // If there is no store before this load, then we can't promote this load.
869 if (I == StoresByIndex.begin()) {
870 // Can't handle this load, bail out.
871 Info.UsingBlocks.push_back(LI->getParent());
875 // Otherwise, there was a store before this load, the load takes its value.
877 LI->replaceAllUsesWith(I->second->getOperand(0));
878 if (AST && LI->getType()->isPointerTy())
879 AST->deleteValue(LI);
880 LI->eraseFromParent();
885 // Inserts a llvm.dbg.value instrinsic before the stores to an alloca'd value
886 // that has an associated llvm.dbg.decl intrinsic.
887 void PromoteMem2Reg::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
889 DIVariable DIVar(DDI->getVariable());
894 DIF = new DIFactory(*SI->getParent()->getParent()->getParent());
895 Instruction *DbgVal = DIF->InsertDbgValueIntrinsic(SI->getOperand(0), 0,
898 // Propagate any debug metadata from the store onto the dbg.value.
899 DebugLoc SIDL = SI->getDebugLoc();
900 if (!SIDL.isUnknown())
901 DbgVal->setDebugLoc(SIDL);
902 // Otherwise propagate debug metadata from dbg.declare.
904 DbgVal->setDebugLoc(DDI->getDebugLoc());
907 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
908 // Alloca returns true if there wasn't already a phi-node for that variable
910 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
912 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
913 // Look up the basic-block in question.
914 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
916 // If the BB already has a phi node added for the i'th alloca then we're done!
917 if (PN) return false;
919 // Create a PhiNode using the dereferenced type... and add the phi-node to the
921 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
922 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
925 PhiToAllocaMap[PN] = AllocaNo;
926 PN->reserveOperandSpace(getNumPreds(BB));
928 InsertedPHINodes.insert(PN);
930 if (AST && PN->getType()->isPointerTy())
931 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
936 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
937 // stores to the allocas which we are promoting. IncomingVals indicates what
938 // value each Alloca contains on exit from the predecessor block Pred.
940 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
941 RenamePassData::ValVector &IncomingVals,
942 std::vector<RenamePassData> &Worklist) {
944 // If we are inserting any phi nodes into this BB, they will already be in the
946 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
947 // If we have PHI nodes to update, compute the number of edges from Pred to
949 if (PhiToAllocaMap.count(APN)) {
950 // We want to be able to distinguish between PHI nodes being inserted by
951 // this invocation of mem2reg from those phi nodes that already existed in
952 // the IR before mem2reg was run. We determine that APN is being inserted
953 // because it is missing incoming edges. All other PHI nodes being
954 // inserted by this pass of mem2reg will have the same number of incoming
955 // operands so far. Remember this count.
956 unsigned NewPHINumOperands = APN->getNumOperands();
958 unsigned NumEdges = 0;
959 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
962 assert(NumEdges && "Must be at least one edge from Pred to BB!");
964 // Add entries for all the phis.
965 BasicBlock::iterator PNI = BB->begin();
967 unsigned AllocaNo = PhiToAllocaMap[APN];
969 // Add N incoming values to the PHI node.
970 for (unsigned i = 0; i != NumEdges; ++i)
971 APN->addIncoming(IncomingVals[AllocaNo], Pred);
973 // The currently active variable for this block is now the PHI.
974 IncomingVals[AllocaNo] = APN;
976 // Get the next phi node.
978 APN = dyn_cast<PHINode>(PNI);
981 // Verify that it is missing entries. If not, it is not being inserted
982 // by this mem2reg invocation so we want to ignore it.
983 } while (APN->getNumOperands() == NewPHINumOperands);
987 // Don't revisit blocks.
988 if (!Visited.insert(BB)) return;
990 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
991 Instruction *I = II++; // get the instruction, increment iterator
993 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
994 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
997 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
998 if (AI == AllocaLookup.end()) continue;
1000 Value *V = IncomingVals[AI->second];
1002 // Anything using the load now uses the current value.
1003 LI->replaceAllUsesWith(V);
1004 if (AST && LI->getType()->isPointerTy())
1005 AST->deleteValue(LI);
1006 BB->getInstList().erase(LI);
1007 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1008 // Delete this instruction and mark the name as the current holder of the
1010 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1011 if (!Dest) continue;
1013 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1014 if (ai == AllocaLookup.end())
1017 // what value were we writing?
1018 IncomingVals[ai->second] = SI->getOperand(0);
1019 // Record debuginfo for the store before removing it.
1020 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1021 ConvertDebugDeclareToDebugValue(DDI, SI);
1022 BB->getInstList().erase(SI);
1026 // 'Recurse' to our successors.
1027 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1030 // Keep track of the successors so we don't visit the same successor twice
1031 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
1033 // Handle the first successor without using the worklist.
1034 VisitedSuccs.insert(*I);
1040 if (VisitedSuccs.insert(*I))
1041 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1046 /// PromoteMemToReg - Promote the specified list of alloca instructions into
1047 /// scalar registers, inserting PHI nodes as appropriate. This function makes
1048 /// use of DominanceFrontier information. This function does not modify the CFG
1049 /// of the function at all. All allocas must be from the same function.
1051 /// If AST is specified, the specified tracker is updated to reflect changes
1054 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1055 DominatorTree &DT, DominanceFrontier &DF,
1056 AliasSetTracker *AST) {
1057 // If there is nothing to do, bail out...
1058 if (Allocas.empty()) return;
1060 PromoteMem2Reg(Allocas, DT, DF, AST).run();