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 iterated dominator frontiers to place PHI nodes, then
13 // traversing the function in depth-first order to rewrite loads and stores as
16 // The algorithm used here is based on:
18 // Sreedhar and Gao. A linear time algorithm for placing phi-nodes.
19 // In Proceedings of the 22nd ACM SIGPLAN-SIGACT Symposium on Principles of
20 // Programming Languages
21 // POPL '95. ACM, New York, NY, 62-73.
23 // It has been modified to not explicitly use the DJ graph data structure and to
24 // directly compute pruned SSA using per-variable liveness information.
26 //===----------------------------------------------------------------------===//
28 #define DEBUG_TYPE "mem2reg"
29 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/Hashing.h"
32 #include "llvm/ADT/STLExtras.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Analysis/AliasSetTracker.h"
37 #include "llvm/Analysis/Dominators.h"
38 #include "llvm/Analysis/InstructionSimplify.h"
39 #include "llvm/Analysis/ValueTracking.h"
40 #include "llvm/DIBuilder.h"
41 #include "llvm/DebugInfo.h"
42 #include "llvm/IR/Constants.h"
43 #include "llvm/IR/DerivedTypes.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Metadata.h"
48 #include "llvm/Support/CFG.h"
49 #include "llvm/Transforms/Utils/Local.h"
54 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
55 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
56 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
57 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
59 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
60 // FIXME: If the memory unit is of pointer or integer type, we can permit
61 // assignments to subsections of the memory unit.
63 // Only allow direct and non-volatile loads and stores...
64 for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end();
65 UI != UE; ++UI) { // Loop over all of the uses of the alloca
67 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
68 // Note that atomic loads can be transformed; atomic semantics do
69 // not have any meaning for a local alloca.
72 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
73 if (SI->getOperand(0) == AI)
74 return false; // Don't allow a store OF the AI, only INTO the AI.
75 // Note that atomic stores can be transformed; atomic semantics do
76 // not have any meaning for a local alloca.
79 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
80 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
81 II->getIntrinsicID() != Intrinsic::lifetime_end)
83 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
84 if (BCI->getType() != Type::getInt8PtrTy(U->getContext()))
86 if (!onlyUsedByLifetimeMarkers(BCI))
88 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
89 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext()))
91 if (!GEPI->hasAllZeroIndices())
93 if (!onlyUsedByLifetimeMarkers(GEPI))
107 // Data package used by RenamePass()
108 class RenamePassData {
110 typedef std::vector<Value *> ValVector;
112 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
113 RenamePassData(BasicBlock *B, BasicBlock *P, const ValVector &V)
114 : BB(B), Pred(P), Values(V) {}
119 void swap(RenamePassData &RHS) {
120 std::swap(BB, RHS.BB);
121 std::swap(Pred, RHS.Pred);
122 Values.swap(RHS.Values);
126 /// \brief This assigns and keeps a per-bb relative ordering of load/store
127 /// instructions in the block that directly load or store an alloca.
129 /// This functionality is important because it avoids scanning large basic
130 /// blocks multiple times when promoting many allocas in the same block.
131 class LargeBlockInfo {
132 /// \brief For each instruction that we track, keep the index of the
135 /// The index starts out as the number of the instruction from the start of
137 DenseMap<const Instruction *, unsigned> InstNumbers;
141 /// This code only looks at accesses to allocas.
142 static bool isInterestingInstruction(const Instruction *I) {
143 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
144 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
147 /// Get or calculate the index of the specified instruction.
148 unsigned getInstructionIndex(const Instruction *I) {
149 assert(isInterestingInstruction(I) &&
150 "Not a load/store to/from an alloca?");
152 // If we already have this instruction number, return it.
153 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
154 if (It != InstNumbers.end())
157 // Scan the whole block to get the instruction. This accumulates
158 // information for every interesting instruction in the block, in order to
159 // avoid gratuitus rescans.
160 const BasicBlock *BB = I->getParent();
162 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end(); BBI != E;
164 if (isInterestingInstruction(BBI))
165 InstNumbers[BBI] = InstNo++;
166 It = InstNumbers.find(I);
168 assert(It != InstNumbers.end() && "Didn't insert instruction?");
172 void deleteValue(const Instruction *I) { InstNumbers.erase(I); }
174 void clear() { InstNumbers.clear(); }
177 struct PromoteMem2Reg {
178 /// The alloca instructions being promoted.
179 std::vector<AllocaInst *> Allocas;
183 /// An AliasSetTracker object to update. If null, don't update it.
184 AliasSetTracker *AST;
186 /// Reverse mapping of Allocas.
187 DenseMap<AllocaInst *, unsigned> AllocaLookup;
189 /// \brief The PhiNodes we're adding.
191 /// That map is used to simplify some Phi nodes as we iterate over it, so
192 /// it should have deterministic iterators. We could use a MapVector, but
193 /// since we already maintain a map from BasicBlock* to a stable numbering
194 /// (BBNumbers), the DenseMap is more efficient (also supports removal).
195 DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes;
197 /// For each PHI node, keep track of which entry in Allocas it corresponds
199 DenseMap<PHINode *, unsigned> PhiToAllocaMap;
201 /// If we are updating an AliasSetTracker, then for each alloca that is of
202 /// pointer type, we keep track of what to copyValue to the inserted PHI
204 std::vector<Value *> PointerAllocaValues;
206 /// For each alloca, we keep track of the dbg.declare intrinsic that
207 /// describes it, if any, so that we can convert it to a dbg.value
208 /// intrinsic if the alloca gets promoted.
209 SmallVector<DbgDeclareInst *, 8> AllocaDbgDeclares;
211 /// The set of basic blocks the renamer has already visited.
213 SmallPtrSet<BasicBlock *, 16> Visited;
215 /// Contains a stable numbering of basic blocks to avoid non-determinstic
217 DenseMap<BasicBlock *, unsigned> BBNumbers;
219 /// Maps DomTreeNodes to their level in the dominator tree.
220 DenseMap<DomTreeNode *, unsigned> DomLevels;
222 /// Lazily compute the number of predecessors a block has.
223 DenseMap<const BasicBlock *, unsigned> BBNumPreds;
226 PromoteMem2Reg(const std::vector<AllocaInst *> &Allocas, DominatorTree &DT,
227 AliasSetTracker *AST)
228 : Allocas(Allocas), DT(DT), DIB(*DT.getRoot()->getParent()->getParent()),
233 /// Return true if BB1 dominates BB2 using the DominatorTree.
234 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
235 return DT.dominates(BB1, BB2);
239 void RemoveFromAllocasList(unsigned &AllocaIdx) {
240 Allocas[AllocaIdx] = Allocas.back();
245 unsigned getNumPreds(const BasicBlock *BB) {
246 unsigned &NP = BBNumPreds[BB];
248 NP = std::distance(pred_begin(BB), pred_end(BB)) + 1;
252 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
254 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
255 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
256 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks);
258 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
259 LargeBlockInfo &LBI);
260 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
261 LargeBlockInfo &LBI);
263 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
264 RenamePassData::ValVector &IncVals,
265 std::vector<RenamePassData> &Worklist);
266 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
270 SmallVector<BasicBlock *, 32> DefiningBlocks;
271 SmallVector<BasicBlock *, 32> UsingBlocks;
273 StoreInst *OnlyStore;
274 BasicBlock *OnlyBlock;
275 bool OnlyUsedInOneBlock;
277 Value *AllocaPointerVal;
278 DbgDeclareInst *DbgDeclare;
281 DefiningBlocks.clear();
285 OnlyUsedInOneBlock = true;
286 AllocaPointerVal = 0;
290 /// Scan the uses of the specified alloca, filling in the AllocaInfo used
291 /// by the rest of the pass to reason about the uses of this alloca.
292 void AnalyzeAlloca(AllocaInst *AI) {
295 // As we scan the uses of the alloca instruction, keep track of stores,
296 // and decide whether all of the loads and stores to the alloca are within
297 // the same basic block.
298 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
300 Instruction *User = cast<Instruction>(*UI++);
302 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
303 // Remember the basic blocks which define new values for the alloca
304 DefiningBlocks.push_back(SI->getParent());
305 AllocaPointerVal = SI->getOperand(0);
308 LoadInst *LI = cast<LoadInst>(User);
309 // Otherwise it must be a load instruction, keep track of variable
311 UsingBlocks.push_back(LI->getParent());
312 AllocaPointerVal = LI;
315 if (OnlyUsedInOneBlock) {
317 OnlyBlock = User->getParent();
318 else if (OnlyBlock != User->getParent())
319 OnlyUsedInOneBlock = false;
323 DbgDeclare = FindAllocaDbgDeclare(AI);
327 } // end of anonymous namespace
329 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
330 // Knowing that this alloca is promotable, we know that it's safe to kill all
331 // instructions except for load and store.
333 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
335 Instruction *I = cast<Instruction>(*UI);
337 if (isa<LoadInst>(I) || isa<StoreInst>(I))
340 if (!I->getType()->isVoidTy()) {
341 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
342 // Follow the use/def chain to erase them now instead of leaving it for
343 // dead code elimination later.
344 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
346 Instruction *Inst = cast<Instruction>(*UI);
348 Inst->eraseFromParent();
351 I->eraseFromParent();
355 void PromoteMem2Reg::run() {
356 Function &F = *DT.getRoot()->getParent();
359 PointerAllocaValues.resize(Allocas.size());
360 AllocaDbgDeclares.resize(Allocas.size());
365 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
366 AllocaInst *AI = Allocas[AllocaNum];
368 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
369 assert(AI->getParent()->getParent() == &F &&
370 "All allocas should be in the same function, which is same as DF!");
372 removeLifetimeIntrinsicUsers(AI);
374 if (AI->use_empty()) {
375 // If there are no uses of the alloca, just delete it now.
377 AST->deleteValue(AI);
378 AI->eraseFromParent();
380 // Remove the alloca from the Allocas list, since it has been processed
381 RemoveFromAllocasList(AllocaNum);
386 // Calculate the set of read and write-locations for each alloca. This is
387 // analogous to finding the 'uses' and 'definitions' of each variable.
388 Info.AnalyzeAlloca(AI);
390 // If there is only a single store to this value, replace any loads of
391 // it that are directly dominated by the definition with the value stored.
392 if (Info.DefiningBlocks.size() == 1) {
393 RewriteSingleStoreAlloca(AI, Info, LBI);
395 // Finally, after the scan, check to see if the store is all that is left.
396 if (Info.UsingBlocks.empty()) {
397 // Record debuginfo for the store and remove the declaration's
399 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
400 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB);
401 DDI->eraseFromParent();
403 // Remove the (now dead) store and alloca.
404 Info.OnlyStore->eraseFromParent();
405 LBI.deleteValue(Info.OnlyStore);
408 AST->deleteValue(AI);
409 AI->eraseFromParent();
412 // The alloca has been processed, move on.
413 RemoveFromAllocasList(AllocaNum);
420 // If the alloca is only read and written in one basic block, just perform a
421 // linear sweep over the block to eliminate it.
422 if (Info.OnlyUsedInOneBlock) {
423 PromoteSingleBlockAlloca(AI, Info, LBI);
425 // Finally, after the scan, check to see if the stores are all that is
427 if (Info.UsingBlocks.empty()) {
429 // Remove the (now dead) stores and alloca.
430 while (!AI->use_empty()) {
431 StoreInst *SI = cast<StoreInst>(AI->use_back());
432 // Record debuginfo for the store before removing it.
433 if (DbgDeclareInst *DDI = Info.DbgDeclare)
434 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
435 SI->eraseFromParent();
440 AST->deleteValue(AI);
441 AI->eraseFromParent();
444 // The alloca has been processed, move on.
445 RemoveFromAllocasList(AllocaNum);
447 // The alloca's debuginfo can be removed as well.
448 if (DbgDeclareInst *DDI = Info.DbgDeclare)
449 DDI->eraseFromParent();
456 // If we haven't computed dominator tree levels, do so now.
457 if (DomLevels.empty()) {
458 SmallVector<DomTreeNode *, 32> Worklist;
460 DomTreeNode *Root = DT.getRootNode();
462 Worklist.push_back(Root);
464 while (!Worklist.empty()) {
465 DomTreeNode *Node = Worklist.pop_back_val();
466 unsigned ChildLevel = DomLevels[Node] + 1;
467 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
469 DomLevels[*CI] = ChildLevel;
470 Worklist.push_back(*CI);
475 // If we haven't computed a numbering for the BB's in the function, do so
477 if (BBNumbers.empty()) {
479 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
483 // If we have an AST to keep updated, remember some pointer value that is
484 // stored into the alloca.
486 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
488 // Remember the dbg.declare intrinsic describing this alloca, if any.
490 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
492 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
493 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
495 // At this point, we're committed to promoting the alloca using IDF's, and
496 // the standard SSA construction algorithm. Determine which blocks need PHI
497 // nodes and see if we can optimize out some work by avoiding insertion of
499 DetermineInsertionPoint(AI, AllocaNum, Info);
503 return; // All of the allocas must have been trivial!
507 // Set the incoming values for the basic block to be null values for all of
508 // the alloca's. We do this in case there is a load of a value that has not
509 // been stored yet. In this case, it will get this null value.
511 RenamePassData::ValVector Values(Allocas.size());
512 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
513 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
515 // Walks all basic blocks in the function performing the SSA rename algorithm
516 // and inserting the phi nodes we marked as necessary
518 std::vector<RenamePassData> RenamePassWorkList;
519 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
522 RPD.swap(RenamePassWorkList.back());
523 RenamePassWorkList.pop_back();
524 // RenamePass may add new worklist entries.
525 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
526 } while (!RenamePassWorkList.empty());
528 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
531 // Remove the allocas themselves from the function.
532 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
533 Instruction *A = Allocas[i];
535 // If there are any uses of the alloca instructions left, they must be in
536 // unreachable basic blocks that were not processed by walking the dominator
537 // tree. Just delete the users now.
539 A->replaceAllUsesWith(UndefValue::get(A->getType()));
542 A->eraseFromParent();
545 // Remove alloca's dbg.declare instrinsics from the function.
546 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
547 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
548 DDI->eraseFromParent();
550 // Loop over all of the PHI nodes and see if there are any that we can get
551 // rid of because they merge all of the same incoming values. This can
552 // happen due to undef values coming into the PHI nodes. This process is
553 // iterative, because eliminating one PHI node can cause others to be removed.
554 bool EliminatedAPHI = true;
555 while (EliminatedAPHI) {
556 EliminatedAPHI = false;
558 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
559 // simplify and RAUW them as we go. If it was not, we could add uses to
560 // the values we replace with in a non deterministic order, thus creating
561 // non deterministic def->use chains.
562 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
563 I = NewPhiNodes.begin(),
564 E = NewPhiNodes.end();
566 PHINode *PN = I->second;
568 // If this PHI node merges one value and/or undefs, get the value.
569 if (Value *V = SimplifyInstruction(PN, 0, 0, &DT)) {
570 if (AST && PN->getType()->isPointerTy())
571 AST->deleteValue(PN);
572 PN->replaceAllUsesWith(V);
573 PN->eraseFromParent();
574 NewPhiNodes.erase(I++);
575 EliminatedAPHI = true;
582 // At this point, the renamer has added entries to PHI nodes for all reachable
583 // code. Unfortunately, there may be unreachable blocks which the renamer
584 // hasn't traversed. If this is the case, the PHI nodes may not
585 // have incoming values for all predecessors. Loop over all PHI nodes we have
586 // created, inserting undef values if they are missing any incoming values.
588 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
589 I = NewPhiNodes.begin(),
590 E = NewPhiNodes.end();
592 // We want to do this once per basic block. As such, only process a block
593 // when we find the PHI that is the first entry in the block.
594 PHINode *SomePHI = I->second;
595 BasicBlock *BB = SomePHI->getParent();
596 if (&BB->front() != SomePHI)
599 // Only do work here if there the PHI nodes are missing incoming values. We
600 // know that all PHI nodes that were inserted in a block will have the same
601 // number of incoming values, so we can just check any of them.
602 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
605 // Get the preds for BB.
606 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
608 // Ok, now we know that all of the PHI nodes are missing entries for some
609 // basic blocks. Start by sorting the incoming predecessors for efficient
611 std::sort(Preds.begin(), Preds.end());
613 // Now we loop through all BB's which have entries in SomePHI and remove
614 // them from the Preds list.
615 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
616 // Do a log(n) search of the Preds list for the entry we want.
617 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
618 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
619 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
620 "PHI node has entry for a block which is not a predecessor!");
626 // At this point, the blocks left in the preds list must have dummy
627 // entries inserted into every PHI nodes for the block. Update all the phi
628 // nodes in this block that we are inserting (there could be phis before
630 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
631 BasicBlock::iterator BBI = BB->begin();
632 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
633 SomePHI->getNumIncomingValues() == NumBadPreds) {
634 Value *UndefVal = UndefValue::get(SomePHI->getType());
635 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
636 SomePHI->addIncoming(UndefVal, Preds[pred]);
643 /// \brief Determine which blocks the value is live in.
645 /// These are blocks which lead to uses. Knowing this allows us to avoid
646 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
647 /// inserted phi nodes would be dead).
648 void PromoteMem2Reg::ComputeLiveInBlocks(
649 AllocaInst *AI, AllocaInfo &Info,
650 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
651 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks) {
653 // To determine liveness, we must iterate through the predecessors of blocks
654 // where the def is live. Blocks are added to the worklist if we need to
655 // check their predecessors. Start with all the using blocks.
656 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
657 Info.UsingBlocks.end());
659 // If any of the using blocks is also a definition block, check to see if the
660 // definition occurs before or after the use. If it happens before the use,
661 // the value isn't really live-in.
662 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
663 BasicBlock *BB = LiveInBlockWorklist[i];
664 if (!DefBlocks.count(BB))
667 // Okay, this is a block that both uses and defines the value. If the first
668 // reference to the alloca is a def (store), then we know it isn't live-in.
669 for (BasicBlock::iterator I = BB->begin();; ++I) {
670 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
671 if (SI->getOperand(1) != AI)
674 // We found a store to the alloca before a load. The alloca is not
675 // actually live-in here.
676 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
677 LiveInBlockWorklist.pop_back();
682 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
683 if (LI->getOperand(0) != AI)
686 // Okay, we found a load before a store to the alloca. It is actually
687 // live into this block.
693 // Now that we have a set of blocks where the phi is live-in, recursively add
694 // their predecessors until we find the full region the value is live.
695 while (!LiveInBlockWorklist.empty()) {
696 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
698 // The block really is live in here, insert it into the set. If already in
699 // the set, then it has already been processed.
700 if (!LiveInBlocks.insert(BB))
703 // Since the value is live into BB, it is either defined in a predecessor or
704 // live into it to. Add the preds to the worklist unless they are a
706 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
709 // The value is not live into a predecessor if it defines the value.
710 if (DefBlocks.count(P))
713 // Otherwise it is, add to the worklist.
714 LiveInBlockWorklist.push_back(P);
720 typedef std::pair<DomTreeNode *, unsigned> DomTreeNodePair;
722 struct DomTreeNodeCompare {
723 bool operator()(const DomTreeNodePair &LHS, const DomTreeNodePair &RHS) {
724 return LHS.second < RHS.second;
727 } // end anonymous namespace
729 /// At this point, we're committed to promoting the alloca using IDF's, and the
730 /// standard SSA construction algorithm. Determine which blocks need phi nodes
731 /// and see if we can optimize out some work by avoiding insertion of dead phi
733 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
735 // Unique the set of defining blocks for efficient lookup.
736 SmallPtrSet<BasicBlock *, 32> DefBlocks;
737 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
739 // Determine which blocks the value is live in. These are blocks which lead
741 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
742 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
744 // Use a priority queue keyed on dominator tree level so that inserted nodes
745 // are handled from the bottom of the dominator tree upwards.
746 typedef std::priority_queue<DomTreeNodePair,
747 SmallVector<DomTreeNodePair, 32>,
748 DomTreeNodeCompare> IDFPriorityQueue;
751 for (SmallPtrSet<BasicBlock *, 32>::const_iterator I = DefBlocks.begin(),
754 if (DomTreeNode *Node = DT.getNode(*I))
755 PQ.push(std::make_pair(Node, DomLevels[Node]));
758 SmallVector<std::pair<unsigned, BasicBlock *>, 32> DFBlocks;
759 SmallPtrSet<DomTreeNode *, 32> Visited;
760 SmallVector<DomTreeNode *, 32> Worklist;
761 while (!PQ.empty()) {
762 DomTreeNodePair RootPair = PQ.top();
764 DomTreeNode *Root = RootPair.first;
765 unsigned RootLevel = RootPair.second;
767 // Walk all dominator tree children of Root, inspecting their CFG edges with
768 // targets elsewhere on the dominator tree. Only targets whose level is at
769 // most Root's level are added to the iterated dominance frontier of the
773 Worklist.push_back(Root);
775 while (!Worklist.empty()) {
776 DomTreeNode *Node = Worklist.pop_back_val();
777 BasicBlock *BB = Node->getBlock();
779 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
781 DomTreeNode *SuccNode = DT.getNode(*SI);
783 // Quickly skip all CFG edges that are also dominator tree edges instead
784 // of catching them below.
785 if (SuccNode->getIDom() == Node)
788 unsigned SuccLevel = DomLevels[SuccNode];
789 if (SuccLevel > RootLevel)
792 if (!Visited.insert(SuccNode))
795 BasicBlock *SuccBB = SuccNode->getBlock();
796 if (!LiveInBlocks.count(SuccBB))
799 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
800 if (!DefBlocks.count(SuccBB))
801 PQ.push(std::make_pair(SuccNode, SuccLevel));
804 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
806 if (!Visited.count(*CI))
807 Worklist.push_back(*CI);
812 if (DFBlocks.size() > 1)
813 std::sort(DFBlocks.begin(), DFBlocks.end());
815 unsigned CurrentVersion = 0;
816 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
817 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
820 /// If there is only a single store to this value, replace any loads of it that
821 /// are directly dominated by the definition with the value stored.
822 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
823 LargeBlockInfo &LBI) {
824 StoreInst *OnlyStore = Info.OnlyStore;
825 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
826 BasicBlock *StoreBB = OnlyStore->getParent();
829 // Clear out UsingBlocks. We will reconstruct it here if needed.
830 Info.UsingBlocks.clear();
832 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
833 Instruction *UserInst = cast<Instruction>(*UI++);
834 if (!isa<LoadInst>(UserInst)) {
835 assert(UserInst == OnlyStore && "Should only have load/stores");
838 LoadInst *LI = cast<LoadInst>(UserInst);
840 // Okay, if we have a load from the alloca, we want to replace it with the
841 // only value stored to the alloca. We can do this if the value is
842 // dominated by the store. If not, we use the rest of the mem2reg machinery
843 // to insert the phi nodes as needed.
844 if (!StoringGlobalVal) { // Non-instructions are always dominated.
845 if (LI->getParent() == StoreBB) {
846 // If we have a use that is in the same block as the store, compare the
847 // indices of the two instructions to see which one came first. If the
848 // load came before the store, we can't handle it.
849 if (StoreIndex == -1)
850 StoreIndex = LBI.getInstructionIndex(OnlyStore);
852 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
853 // Can't handle this load, bail out.
854 Info.UsingBlocks.push_back(StoreBB);
858 } else if (LI->getParent() != StoreBB &&
859 !dominates(StoreBB, LI->getParent())) {
860 // If the load and store are in different blocks, use BB dominance to
861 // check their relationships. If the store doesn't dom the use, bail
863 Info.UsingBlocks.push_back(LI->getParent());
868 // Otherwise, we *can* safely rewrite this load.
869 Value *ReplVal = OnlyStore->getOperand(0);
870 // If the replacement value is the load, this must occur in unreachable
873 ReplVal = UndefValue::get(LI->getType());
874 LI->replaceAllUsesWith(ReplVal);
875 if (AST && LI->getType()->isPointerTy())
876 AST->deleteValue(LI);
877 LI->eraseFromParent();
883 /// This is a helper predicate used to search by the first element of a pair.
884 struct StoreIndexSearchPredicate {
885 bool operator()(const std::pair<unsigned, StoreInst *> &LHS,
886 const std::pair<unsigned, StoreInst *> &RHS) {
887 return LHS.first < RHS.first;
892 /// Many allocas are only used within a single basic block. If this is the
893 /// case, avoid traversing the CFG and inserting a lot of potentially useless
894 /// PHI nodes by just performing a single linear pass over the basic block
895 /// using the Alloca.
897 /// If we cannot promote this alloca (because it is read before it is written),
898 /// return true. This is necessary in cases where, due to control flow, the
899 /// alloca is potentially undefined on some control flow paths. e.g. code like
900 /// this is potentially correct:
902 /// for (...) { if (c) { A = undef; undef = B; } }
904 /// ... so long as A is not used before undef is set.
905 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
906 LargeBlockInfo &LBI) {
907 // The trickiest case to handle is when we have large blocks. Because of this,
908 // this code is optimized assuming that large blocks happen. This does not
909 // significantly pessimize the small block case. This uses LargeBlockInfo to
910 // make it efficient to get the index of various operations in the block.
912 // Clear out UsingBlocks. We will reconstruct it here if needed.
913 Info.UsingBlocks.clear();
915 // Walk the use-def list of the alloca, getting the locations of all stores.
916 typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy;
917 StoresByIndexTy StoresByIndex;
919 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;
921 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
922 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
924 // If there are no stores to the alloca, just replace any loads with undef.
925 if (StoresByIndex.empty()) {
926 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
927 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
928 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
929 if (AST && LI->getType()->isPointerTy())
930 AST->deleteValue(LI);
932 LI->eraseFromParent();
937 // Sort the stores by their index, making it efficient to do a lookup with a
939 std::sort(StoresByIndex.begin(), StoresByIndex.end());
941 // Walk all of the loads from this alloca, replacing them with the nearest
942 // store above them, if any.
943 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
944 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
948 unsigned LoadIdx = LBI.getInstructionIndex(LI);
950 // Find the nearest store that has a lower than this load.
951 StoresByIndexTy::iterator I = std::lower_bound(
952 StoresByIndex.begin(), StoresByIndex.end(),
953 std::pair<unsigned, StoreInst *>(LoadIdx, static_cast<StoreInst *>(0)),
954 StoreIndexSearchPredicate());
956 // If there is no store before this load, then we can't promote this load.
957 if (I == StoresByIndex.begin()) {
958 // Can't handle this load, bail out.
959 Info.UsingBlocks.push_back(LI->getParent());
963 // Otherwise, there was a store before this load, the load takes its value.
965 LI->replaceAllUsesWith(I->second->getOperand(0));
966 if (AST && LI->getType()->isPointerTy())
967 AST->deleteValue(LI);
968 LI->eraseFromParent();
973 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
975 /// Returns true if there wasn't already a phi-node for that variable
976 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
978 // Look up the basic-block in question.
979 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
981 // If the BB already has a phi node added for the i'th alloca then we're done!
985 // Create a PhiNode using the dereferenced type... and add the phi-node to the
987 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
988 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
991 PhiToAllocaMap[PN] = AllocaNo;
993 if (AST && PN->getType()->isPointerTy())
994 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
999 /// \brief Recursively traverse the CFG of the function, renaming loads and
1000 /// stores to the allocas which we are promoting.
1002 /// IncomingVals indicates what value each Alloca contains on exit from the
1003 /// predecessor block Pred.
1004 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
1005 RenamePassData::ValVector &IncomingVals,
1006 std::vector<RenamePassData> &Worklist) {
1008 // If we are inserting any phi nodes into this BB, they will already be in the
1010 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
1011 // If we have PHI nodes to update, compute the number of edges from Pred to
1013 if (PhiToAllocaMap.count(APN)) {
1014 // We want to be able to distinguish between PHI nodes being inserted by
1015 // this invocation of mem2reg from those phi nodes that already existed in
1016 // the IR before mem2reg was run. We determine that APN is being inserted
1017 // because it is missing incoming edges. All other PHI nodes being
1018 // inserted by this pass of mem2reg will have the same number of incoming
1019 // operands so far. Remember this count.
1020 unsigned NewPHINumOperands = APN->getNumOperands();
1022 unsigned NumEdges = 0;
1023 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
1026 assert(NumEdges && "Must be at least one edge from Pred to BB!");
1028 // Add entries for all the phis.
1029 BasicBlock::iterator PNI = BB->begin();
1031 unsigned AllocaNo = PhiToAllocaMap[APN];
1033 // Add N incoming values to the PHI node.
1034 for (unsigned i = 0; i != NumEdges; ++i)
1035 APN->addIncoming(IncomingVals[AllocaNo], Pred);
1037 // The currently active variable for this block is now the PHI.
1038 IncomingVals[AllocaNo] = APN;
1040 // Get the next phi node.
1042 APN = dyn_cast<PHINode>(PNI);
1046 // Verify that it is missing entries. If not, it is not being inserted
1047 // by this mem2reg invocation so we want to ignore it.
1048 } while (APN->getNumOperands() == NewPHINumOperands);
1052 // Don't revisit blocks.
1053 if (!Visited.insert(BB))
1056 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) {
1057 Instruction *I = II++; // get the instruction, increment iterator
1059 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1060 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1064 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
1065 if (AI == AllocaLookup.end())
1068 Value *V = IncomingVals[AI->second];
1070 // Anything using the load now uses the current value.
1071 LI->replaceAllUsesWith(V);
1072 if (AST && LI->getType()->isPointerTy())
1073 AST->deleteValue(LI);
1074 BB->getInstList().erase(LI);
1075 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1076 // Delete this instruction and mark the name as the current holder of the
1078 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1082 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1083 if (ai == AllocaLookup.end())
1086 // what value were we writing?
1087 IncomingVals[ai->second] = SI->getOperand(0);
1088 // Record debuginfo for the store before removing it.
1089 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1090 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1091 BB->getInstList().erase(SI);
1095 // 'Recurse' to our successors.
1096 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1100 // Keep track of the successors so we don't visit the same successor twice
1101 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
1103 // Handle the first successor without using the worklist.
1104 VisitedSuccs.insert(*I);
1110 if (VisitedSuccs.insert(*I))
1111 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1116 void llvm::PromoteMemToReg(const std::vector<AllocaInst *> &Allocas,
1117 DominatorTree &DT, AliasSetTracker *AST) {
1118 // If there is nothing to do, bail out...
1119 if (Allocas.empty())
1122 PromoteMem2Reg(Allocas, DT, AST).run();