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))
106 SmallVector<BasicBlock *, 32> DefiningBlocks;
107 SmallVector<BasicBlock *, 32> UsingBlocks;
109 StoreInst *OnlyStore;
110 BasicBlock *OnlyBlock;
111 bool OnlyUsedInOneBlock;
113 Value *AllocaPointerVal;
114 DbgDeclareInst *DbgDeclare;
117 DefiningBlocks.clear();
121 OnlyUsedInOneBlock = true;
122 AllocaPointerVal = 0;
126 /// Scan the uses of the specified alloca, filling in the AllocaInfo used
127 /// by the rest of the pass to reason about the uses of this alloca.
128 void AnalyzeAlloca(AllocaInst *AI) {
131 // As we scan the uses of the alloca instruction, keep track of stores,
132 // and decide whether all of the loads and stores to the alloca are within
133 // the same basic block.
134 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
136 Instruction *User = cast<Instruction>(*UI++);
138 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
139 // Remember the basic blocks which define new values for the alloca
140 DefiningBlocks.push_back(SI->getParent());
141 AllocaPointerVal = SI->getOperand(0);
144 LoadInst *LI = cast<LoadInst>(User);
145 // Otherwise it must be a load instruction, keep track of variable
147 UsingBlocks.push_back(LI->getParent());
148 AllocaPointerVal = LI;
151 if (OnlyUsedInOneBlock) {
153 OnlyBlock = User->getParent();
154 else if (OnlyBlock != User->getParent())
155 OnlyUsedInOneBlock = false;
159 DbgDeclare = FindAllocaDbgDeclare(AI);
163 // Data package used by RenamePass()
164 class RenamePassData {
166 typedef std::vector<Value *> ValVector;
168 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
169 RenamePassData(BasicBlock *B, BasicBlock *P, const ValVector &V)
170 : BB(B), Pred(P), Values(V) {}
175 void swap(RenamePassData &RHS) {
176 std::swap(BB, RHS.BB);
177 std::swap(Pred, RHS.Pred);
178 Values.swap(RHS.Values);
182 /// \brief This assigns and keeps a per-bb relative ordering of load/store
183 /// instructions in the block that directly load or store an alloca.
185 /// This functionality is important because it avoids scanning large basic
186 /// blocks multiple times when promoting many allocas in the same block.
187 class LargeBlockInfo {
188 /// \brief For each instruction that we track, keep the index of the
191 /// The index starts out as the number of the instruction from the start of
193 DenseMap<const Instruction *, unsigned> InstNumbers;
197 /// This code only looks at accesses to allocas.
198 static bool isInterestingInstruction(const Instruction *I) {
199 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
200 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
203 /// Get or calculate the index of the specified instruction.
204 unsigned getInstructionIndex(const Instruction *I) {
205 assert(isInterestingInstruction(I) &&
206 "Not a load/store to/from an alloca?");
208 // If we already have this instruction number, return it.
209 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
210 if (It != InstNumbers.end())
213 // Scan the whole block to get the instruction. This accumulates
214 // information for every interesting instruction in the block, in order to
215 // avoid gratuitus rescans.
216 const BasicBlock *BB = I->getParent();
218 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end(); BBI != E;
220 if (isInterestingInstruction(BBI))
221 InstNumbers[BBI] = InstNo++;
222 It = InstNumbers.find(I);
224 assert(It != InstNumbers.end() && "Didn't insert instruction?");
228 void deleteValue(const Instruction *I) { InstNumbers.erase(I); }
230 void clear() { InstNumbers.clear(); }
233 struct PromoteMem2Reg {
234 /// The alloca instructions being promoted.
235 std::vector<AllocaInst *> Allocas;
239 /// An AliasSetTracker object to update. If null, don't update it.
240 AliasSetTracker *AST;
242 /// Reverse mapping of Allocas.
243 DenseMap<AllocaInst *, unsigned> AllocaLookup;
245 /// \brief The PhiNodes we're adding.
247 /// That map is used to simplify some Phi nodes as we iterate over it, so
248 /// it should have deterministic iterators. We could use a MapVector, but
249 /// since we already maintain a map from BasicBlock* to a stable numbering
250 /// (BBNumbers), the DenseMap is more efficient (also supports removal).
251 DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes;
253 /// For each PHI node, keep track of which entry in Allocas it corresponds
255 DenseMap<PHINode *, unsigned> PhiToAllocaMap;
257 /// If we are updating an AliasSetTracker, then for each alloca that is of
258 /// pointer type, we keep track of what to copyValue to the inserted PHI
260 std::vector<Value *> PointerAllocaValues;
262 /// For each alloca, we keep track of the dbg.declare intrinsic that
263 /// describes it, if any, so that we can convert it to a dbg.value
264 /// intrinsic if the alloca gets promoted.
265 SmallVector<DbgDeclareInst *, 8> AllocaDbgDeclares;
267 /// The set of basic blocks the renamer has already visited.
269 SmallPtrSet<BasicBlock *, 16> Visited;
271 /// Contains a stable numbering of basic blocks to avoid non-determinstic
273 DenseMap<BasicBlock *, unsigned> BBNumbers;
275 /// Maps DomTreeNodes to their level in the dominator tree.
276 DenseMap<DomTreeNode *, unsigned> DomLevels;
278 /// Lazily compute the number of predecessors a block has.
279 DenseMap<const BasicBlock *, unsigned> BBNumPreds;
282 PromoteMem2Reg(const std::vector<AllocaInst *> &Allocas, DominatorTree &DT,
283 AliasSetTracker *AST)
284 : Allocas(Allocas), DT(DT), DIB(*DT.getRoot()->getParent()->getParent()),
290 void RemoveFromAllocasList(unsigned &AllocaIdx) {
291 Allocas[AllocaIdx] = Allocas.back();
296 unsigned getNumPreds(const BasicBlock *BB) {
297 unsigned &NP = BBNumPreds[BB];
299 NP = std::distance(pred_begin(BB), pred_end(BB)) + 1;
303 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
305 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
306 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
307 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks);
308 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
309 RenamePassData::ValVector &IncVals,
310 std::vector<RenamePassData> &Worklist);
311 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
314 } // end of anonymous namespace
316 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
317 // Knowing that this alloca is promotable, we know that it's safe to kill all
318 // instructions except for load and store.
320 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
322 Instruction *I = cast<Instruction>(*UI);
324 if (isa<LoadInst>(I) || isa<StoreInst>(I))
327 if (!I->getType()->isVoidTy()) {
328 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
329 // Follow the use/def chain to erase them now instead of leaving it for
330 // dead code elimination later.
331 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
333 Instruction *Inst = cast<Instruction>(*UI);
335 Inst->eraseFromParent();
338 I->eraseFromParent();
342 /// \brief Rewrite as many loads as possible given a single store.
344 /// When there is only a single store, we can use the domtree to trivially
345 /// replace all of the dominated loads with the stored value. Do so, and return
346 /// true if this has successfully promoted the alloca entirely. If this returns
347 /// false there were some loads which were not dominated by the single store
348 /// and thus must be phi-ed with undef. We fall back to the standard alloca
349 /// promotion algorithm in that case.
350 static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
353 AliasSetTracker *AST) {
354 StoreInst *OnlyStore = Info.OnlyStore;
355 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
356 BasicBlock *StoreBB = OnlyStore->getParent();
359 // Clear out UsingBlocks. We will reconstruct it here if needed.
360 Info.UsingBlocks.clear();
362 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
363 Instruction *UserInst = cast<Instruction>(*UI++);
364 if (!isa<LoadInst>(UserInst)) {
365 assert(UserInst == OnlyStore && "Should only have load/stores");
368 LoadInst *LI = cast<LoadInst>(UserInst);
370 // Okay, if we have a load from the alloca, we want to replace it with the
371 // only value stored to the alloca. We can do this if the value is
372 // dominated by the store. If not, we use the rest of the mem2reg machinery
373 // to insert the phi nodes as needed.
374 if (!StoringGlobalVal) { // Non-instructions are always dominated.
375 if (LI->getParent() == StoreBB) {
376 // If we have a use that is in the same block as the store, compare the
377 // indices of the two instructions to see which one came first. If the
378 // load came before the store, we can't handle it.
379 if (StoreIndex == -1)
380 StoreIndex = LBI.getInstructionIndex(OnlyStore);
382 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
383 // Can't handle this load, bail out.
384 Info.UsingBlocks.push_back(StoreBB);
388 } else if (LI->getParent() != StoreBB &&
389 !DT.dominates(StoreBB, LI->getParent())) {
390 // If the load and store are in different blocks, use BB dominance to
391 // check their relationships. If the store doesn't dom the use, bail
393 Info.UsingBlocks.push_back(LI->getParent());
398 // Otherwise, we *can* safely rewrite this load.
399 Value *ReplVal = OnlyStore->getOperand(0);
400 // If the replacement value is the load, this must occur in unreachable
403 ReplVal = UndefValue::get(LI->getType());
404 LI->replaceAllUsesWith(ReplVal);
405 if (AST && LI->getType()->isPointerTy())
406 AST->deleteValue(LI);
407 LI->eraseFromParent();
411 // Finally, after the scan, check to see if the store is all that is left.
412 if (!Info.UsingBlocks.empty())
413 return false; // If not, we'll have to fall back for the remainder.
415 // Record debuginfo for the store and remove the declaration's
417 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
418 DIBuilder DIB(*AI->getParent()->getParent()->getParent());
419 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB);
420 DDI->eraseFromParent();
422 // Remove the (now dead) store and alloca.
423 Info.OnlyStore->eraseFromParent();
424 LBI.deleteValue(Info.OnlyStore);
427 AST->deleteValue(AI);
428 AI->eraseFromParent();
434 /// This is a helper predicate used to search by the first element of a pair.
435 struct StoreIndexSearchPredicate {
436 bool operator()(const std::pair<unsigned, StoreInst *> &LHS,
437 const std::pair<unsigned, StoreInst *> &RHS) {
438 return LHS.first < RHS.first;
443 /// Many allocas are only used within a single basic block. If this is the
444 /// case, avoid traversing the CFG and inserting a lot of potentially useless
445 /// PHI nodes by just performing a single linear pass over the basic block
446 /// using the Alloca.
448 /// If we cannot promote this alloca (because it is read before it is written),
449 /// return true. This is necessary in cases where, due to control flow, the
450 /// alloca is potentially undefined on some control flow paths. e.g. code like
451 /// this is potentially correct:
453 /// for (...) { if (c) { A = undef; undef = B; } }
455 /// ... so long as A is not used before undef is set.
456 static void promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
458 AliasSetTracker *AST) {
459 // The trickiest case to handle is when we have large blocks. Because of this,
460 // this code is optimized assuming that large blocks happen. This does not
461 // significantly pessimize the small block case. This uses LargeBlockInfo to
462 // make it efficient to get the index of various operations in the block.
464 // Walk the use-def list of the alloca, getting the locations of all stores.
465 typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy;
466 StoresByIndexTy StoresByIndex;
468 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;
470 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
471 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
473 // Sort the stores by their index, making it efficient to do a lookup with a
475 std::sort(StoresByIndex.begin(), StoresByIndex.end());
477 // Walk all of the loads from this alloca, replacing them with the nearest
478 // store above them, if any.
479 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
480 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
484 unsigned LoadIdx = LBI.getInstructionIndex(LI);
486 // Find the nearest store that has a lower than this load.
487 StoresByIndexTy::iterator I = std::lower_bound(
488 StoresByIndex.begin(), StoresByIndex.end(),
489 std::pair<unsigned, StoreInst *>(LoadIdx, static_cast<StoreInst *>(0)),
490 StoreIndexSearchPredicate());
492 if (I == StoresByIndex.begin())
493 // If there is no store before this load, the load takes the undef value.
494 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
496 // Otherwise, there was a store before this load, the load takes its value.
497 LI->replaceAllUsesWith(llvm::prior(I)->second->getOperand(0));
499 if (AST && LI->getType()->isPointerTy())
500 AST->deleteValue(LI);
501 LI->eraseFromParent();
505 // Remove the (now dead) stores and alloca.
506 while (!AI->use_empty()) {
507 StoreInst *SI = cast<StoreInst>(AI->use_back());
508 // Record debuginfo for the store before removing it.
509 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
510 DIBuilder DIB(*AI->getParent()->getParent()->getParent());
511 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
513 SI->eraseFromParent();
518 AST->deleteValue(AI);
519 AI->eraseFromParent();
522 // The alloca's debuginfo can be removed as well.
523 if (DbgDeclareInst *DDI = Info.DbgDeclare)
524 DDI->eraseFromParent();
529 void PromoteMem2Reg::run() {
530 Function &F = *DT.getRoot()->getParent();
533 PointerAllocaValues.resize(Allocas.size());
534 AllocaDbgDeclares.resize(Allocas.size());
539 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
540 AllocaInst *AI = Allocas[AllocaNum];
542 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
543 assert(AI->getParent()->getParent() == &F &&
544 "All allocas should be in the same function, which is same as DF!");
546 removeLifetimeIntrinsicUsers(AI);
548 if (AI->use_empty()) {
549 // If there are no uses of the alloca, just delete it now.
551 AST->deleteValue(AI);
552 AI->eraseFromParent();
554 // Remove the alloca from the Allocas list, since it has been processed
555 RemoveFromAllocasList(AllocaNum);
560 // Calculate the set of read and write-locations for each alloca. This is
561 // analogous to finding the 'uses' and 'definitions' of each variable.
562 Info.AnalyzeAlloca(AI);
564 // If there is only a single store to this value, replace any loads of
565 // it that are directly dominated by the definition with the value stored.
566 if (Info.DefiningBlocks.size() == 1) {
567 if (rewriteSingleStoreAlloca(AI, Info, LBI, DT, AST)) {
568 // The alloca has been processed, move on.
569 RemoveFromAllocasList(AllocaNum);
575 // If the alloca is only read and written in one basic block, just perform a
576 // linear sweep over the block to eliminate it.
577 if (Info.OnlyUsedInOneBlock) {
578 promoteSingleBlockAlloca(AI, Info, LBI, AST);
580 // The alloca has been processed, move on.
581 RemoveFromAllocasList(AllocaNum);
585 // If we haven't computed dominator tree levels, do so now.
586 if (DomLevels.empty()) {
587 SmallVector<DomTreeNode *, 32> Worklist;
589 DomTreeNode *Root = DT.getRootNode();
591 Worklist.push_back(Root);
593 while (!Worklist.empty()) {
594 DomTreeNode *Node = Worklist.pop_back_val();
595 unsigned ChildLevel = DomLevels[Node] + 1;
596 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
598 DomLevels[*CI] = ChildLevel;
599 Worklist.push_back(*CI);
604 // If we haven't computed a numbering for the BB's in the function, do so
606 if (BBNumbers.empty()) {
608 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
612 // If we have an AST to keep updated, remember some pointer value that is
613 // stored into the alloca.
615 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
617 // Remember the dbg.declare intrinsic describing this alloca, if any.
619 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
621 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
622 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
624 // At this point, we're committed to promoting the alloca using IDF's, and
625 // the standard SSA construction algorithm. Determine which blocks need PHI
626 // nodes and see if we can optimize out some work by avoiding insertion of
628 DetermineInsertionPoint(AI, AllocaNum, Info);
632 return; // All of the allocas must have been trivial!
636 // Set the incoming values for the basic block to be null values for all of
637 // the alloca's. We do this in case there is a load of a value that has not
638 // been stored yet. In this case, it will get this null value.
640 RenamePassData::ValVector Values(Allocas.size());
641 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
642 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
644 // Walks all basic blocks in the function performing the SSA rename algorithm
645 // and inserting the phi nodes we marked as necessary
647 std::vector<RenamePassData> RenamePassWorkList;
648 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
651 RPD.swap(RenamePassWorkList.back());
652 RenamePassWorkList.pop_back();
653 // RenamePass may add new worklist entries.
654 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
655 } while (!RenamePassWorkList.empty());
657 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
660 // Remove the allocas themselves from the function.
661 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
662 Instruction *A = Allocas[i];
664 // If there are any uses of the alloca instructions left, they must be in
665 // unreachable basic blocks that were not processed by walking the dominator
666 // tree. Just delete the users now.
668 A->replaceAllUsesWith(UndefValue::get(A->getType()));
671 A->eraseFromParent();
674 // Remove alloca's dbg.declare instrinsics from the function.
675 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
676 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
677 DDI->eraseFromParent();
679 // Loop over all of the PHI nodes and see if there are any that we can get
680 // rid of because they merge all of the same incoming values. This can
681 // happen due to undef values coming into the PHI nodes. This process is
682 // iterative, because eliminating one PHI node can cause others to be removed.
683 bool EliminatedAPHI = true;
684 while (EliminatedAPHI) {
685 EliminatedAPHI = false;
687 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
688 // simplify and RAUW them as we go. If it was not, we could add uses to
689 // the values we replace with in a non deterministic order, thus creating
690 // non deterministic def->use chains.
691 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
692 I = NewPhiNodes.begin(),
693 E = NewPhiNodes.end();
695 PHINode *PN = I->second;
697 // If this PHI node merges one value and/or undefs, get the value.
698 if (Value *V = SimplifyInstruction(PN, 0, 0, &DT)) {
699 if (AST && PN->getType()->isPointerTy())
700 AST->deleteValue(PN);
701 PN->replaceAllUsesWith(V);
702 PN->eraseFromParent();
703 NewPhiNodes.erase(I++);
704 EliminatedAPHI = true;
711 // At this point, the renamer has added entries to PHI nodes for all reachable
712 // code. Unfortunately, there may be unreachable blocks which the renamer
713 // hasn't traversed. If this is the case, the PHI nodes may not
714 // have incoming values for all predecessors. Loop over all PHI nodes we have
715 // created, inserting undef values if they are missing any incoming values.
717 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
718 I = NewPhiNodes.begin(),
719 E = NewPhiNodes.end();
721 // We want to do this once per basic block. As such, only process a block
722 // when we find the PHI that is the first entry in the block.
723 PHINode *SomePHI = I->second;
724 BasicBlock *BB = SomePHI->getParent();
725 if (&BB->front() != SomePHI)
728 // Only do work here if there the PHI nodes are missing incoming values. We
729 // know that all PHI nodes that were inserted in a block will have the same
730 // number of incoming values, so we can just check any of them.
731 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
734 // Get the preds for BB.
735 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
737 // Ok, now we know that all of the PHI nodes are missing entries for some
738 // basic blocks. Start by sorting the incoming predecessors for efficient
740 std::sort(Preds.begin(), Preds.end());
742 // Now we loop through all BB's which have entries in SomePHI and remove
743 // them from the Preds list.
744 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
745 // Do a log(n) search of the Preds list for the entry we want.
746 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
747 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
748 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
749 "PHI node has entry for a block which is not a predecessor!");
755 // At this point, the blocks left in the preds list must have dummy
756 // entries inserted into every PHI nodes for the block. Update all the phi
757 // nodes in this block that we are inserting (there could be phis before
759 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
760 BasicBlock::iterator BBI = BB->begin();
761 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
762 SomePHI->getNumIncomingValues() == NumBadPreds) {
763 Value *UndefVal = UndefValue::get(SomePHI->getType());
764 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
765 SomePHI->addIncoming(UndefVal, Preds[pred]);
772 /// \brief Determine which blocks the value is live in.
774 /// These are blocks which lead to uses. Knowing this allows us to avoid
775 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
776 /// inserted phi nodes would be dead).
777 void PromoteMem2Reg::ComputeLiveInBlocks(
778 AllocaInst *AI, AllocaInfo &Info,
779 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
780 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks) {
782 // To determine liveness, we must iterate through the predecessors of blocks
783 // where the def is live. Blocks are added to the worklist if we need to
784 // check their predecessors. Start with all the using blocks.
785 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
786 Info.UsingBlocks.end());
788 // If any of the using blocks is also a definition block, check to see if the
789 // definition occurs before or after the use. If it happens before the use,
790 // the value isn't really live-in.
791 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
792 BasicBlock *BB = LiveInBlockWorklist[i];
793 if (!DefBlocks.count(BB))
796 // Okay, this is a block that both uses and defines the value. If the first
797 // reference to the alloca is a def (store), then we know it isn't live-in.
798 for (BasicBlock::iterator I = BB->begin();; ++I) {
799 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
800 if (SI->getOperand(1) != AI)
803 // We found a store to the alloca before a load. The alloca is not
804 // actually live-in here.
805 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
806 LiveInBlockWorklist.pop_back();
811 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
812 if (LI->getOperand(0) != AI)
815 // Okay, we found a load before a store to the alloca. It is actually
816 // live into this block.
822 // Now that we have a set of blocks where the phi is live-in, recursively add
823 // their predecessors until we find the full region the value is live.
824 while (!LiveInBlockWorklist.empty()) {
825 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
827 // The block really is live in here, insert it into the set. If already in
828 // the set, then it has already been processed.
829 if (!LiveInBlocks.insert(BB))
832 // Since the value is live into BB, it is either defined in a predecessor or
833 // live into it to. Add the preds to the worklist unless they are a
835 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
838 // The value is not live into a predecessor if it defines the value.
839 if (DefBlocks.count(P))
842 // Otherwise it is, add to the worklist.
843 LiveInBlockWorklist.push_back(P);
849 typedef std::pair<DomTreeNode *, unsigned> DomTreeNodePair;
851 struct DomTreeNodeCompare {
852 bool operator()(const DomTreeNodePair &LHS, const DomTreeNodePair &RHS) {
853 return LHS.second < RHS.second;
856 } // end anonymous namespace
858 /// At this point, we're committed to promoting the alloca using IDF's, and the
859 /// standard SSA construction algorithm. Determine which blocks need phi nodes
860 /// and see if we can optimize out some work by avoiding insertion of dead phi
862 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
864 // Unique the set of defining blocks for efficient lookup.
865 SmallPtrSet<BasicBlock *, 32> DefBlocks;
866 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
868 // Determine which blocks the value is live in. These are blocks which lead
870 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
871 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
873 // Use a priority queue keyed on dominator tree level so that inserted nodes
874 // are handled from the bottom of the dominator tree upwards.
875 typedef std::priority_queue<DomTreeNodePair,
876 SmallVector<DomTreeNodePair, 32>,
877 DomTreeNodeCompare> IDFPriorityQueue;
880 for (SmallPtrSet<BasicBlock *, 32>::const_iterator I = DefBlocks.begin(),
883 if (DomTreeNode *Node = DT.getNode(*I))
884 PQ.push(std::make_pair(Node, DomLevels[Node]));
887 SmallVector<std::pair<unsigned, BasicBlock *>, 32> DFBlocks;
888 SmallPtrSet<DomTreeNode *, 32> Visited;
889 SmallVector<DomTreeNode *, 32> Worklist;
890 while (!PQ.empty()) {
891 DomTreeNodePair RootPair = PQ.top();
893 DomTreeNode *Root = RootPair.first;
894 unsigned RootLevel = RootPair.second;
896 // Walk all dominator tree children of Root, inspecting their CFG edges with
897 // targets elsewhere on the dominator tree. Only targets whose level is at
898 // most Root's level are added to the iterated dominance frontier of the
902 Worklist.push_back(Root);
904 while (!Worklist.empty()) {
905 DomTreeNode *Node = Worklist.pop_back_val();
906 BasicBlock *BB = Node->getBlock();
908 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
910 DomTreeNode *SuccNode = DT.getNode(*SI);
912 // Quickly skip all CFG edges that are also dominator tree edges instead
913 // of catching them below.
914 if (SuccNode->getIDom() == Node)
917 unsigned SuccLevel = DomLevels[SuccNode];
918 if (SuccLevel > RootLevel)
921 if (!Visited.insert(SuccNode))
924 BasicBlock *SuccBB = SuccNode->getBlock();
925 if (!LiveInBlocks.count(SuccBB))
928 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
929 if (!DefBlocks.count(SuccBB))
930 PQ.push(std::make_pair(SuccNode, SuccLevel));
933 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
935 if (!Visited.count(*CI))
936 Worklist.push_back(*CI);
941 if (DFBlocks.size() > 1)
942 std::sort(DFBlocks.begin(), DFBlocks.end());
944 unsigned CurrentVersion = 0;
945 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
946 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
949 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
951 /// Returns true if there wasn't already a phi-node for that variable
952 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
954 // Look up the basic-block in question.
955 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
957 // If the BB already has a phi node added for the i'th alloca then we're done!
961 // Create a PhiNode using the dereferenced type... and add the phi-node to the
963 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
964 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
967 PhiToAllocaMap[PN] = AllocaNo;
969 if (AST && PN->getType()->isPointerTy())
970 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
975 /// \brief Recursively traverse the CFG of the function, renaming loads and
976 /// stores to the allocas which we are promoting.
978 /// IncomingVals indicates what value each Alloca contains on exit from the
979 /// predecessor block Pred.
980 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
981 RenamePassData::ValVector &IncomingVals,
982 std::vector<RenamePassData> &Worklist) {
984 // If we are inserting any phi nodes into this BB, they will already be in the
986 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
987 // If we have PHI nodes to update, compute the number of edges from Pred to
989 if (PhiToAllocaMap.count(APN)) {
990 // We want to be able to distinguish between PHI nodes being inserted by
991 // this invocation of mem2reg from those phi nodes that already existed in
992 // the IR before mem2reg was run. We determine that APN is being inserted
993 // because it is missing incoming edges. All other PHI nodes being
994 // inserted by this pass of mem2reg will have the same number of incoming
995 // operands so far. Remember this count.
996 unsigned NewPHINumOperands = APN->getNumOperands();
998 unsigned NumEdges = 0;
999 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
1002 assert(NumEdges && "Must be at least one edge from Pred to BB!");
1004 // Add entries for all the phis.
1005 BasicBlock::iterator PNI = BB->begin();
1007 unsigned AllocaNo = PhiToAllocaMap[APN];
1009 // Add N incoming values to the PHI node.
1010 for (unsigned i = 0; i != NumEdges; ++i)
1011 APN->addIncoming(IncomingVals[AllocaNo], Pred);
1013 // The currently active variable for this block is now the PHI.
1014 IncomingVals[AllocaNo] = APN;
1016 // Get the next phi node.
1018 APN = dyn_cast<PHINode>(PNI);
1022 // Verify that it is missing entries. If not, it is not being inserted
1023 // by this mem2reg invocation so we want to ignore it.
1024 } while (APN->getNumOperands() == NewPHINumOperands);
1028 // Don't revisit blocks.
1029 if (!Visited.insert(BB))
1032 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) {
1033 Instruction *I = II++; // get the instruction, increment iterator
1035 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1036 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1040 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
1041 if (AI == AllocaLookup.end())
1044 Value *V = IncomingVals[AI->second];
1046 // Anything using the load now uses the current value.
1047 LI->replaceAllUsesWith(V);
1048 if (AST && LI->getType()->isPointerTy())
1049 AST->deleteValue(LI);
1050 BB->getInstList().erase(LI);
1051 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1052 // Delete this instruction and mark the name as the current holder of the
1054 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1058 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1059 if (ai == AllocaLookup.end())
1062 // what value were we writing?
1063 IncomingVals[ai->second] = SI->getOperand(0);
1064 // Record debuginfo for the store before removing it.
1065 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1066 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1067 BB->getInstList().erase(SI);
1071 // 'Recurse' to our successors.
1072 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1076 // Keep track of the successors so we don't visit the same successor twice
1077 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
1079 // Handle the first successor without using the worklist.
1080 VisitedSuccs.insert(*I);
1086 if (VisitedSuccs.insert(*I))
1087 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1092 void llvm::PromoteMemToReg(const std::vector<AllocaInst *> &Allocas,
1093 DominatorTree &DT, AliasSetTracker *AST) {
1094 // If there is nothing to do, bail out...
1095 if (Allocas.empty())
1098 PromoteMem2Reg(Allocas, DT, AST).run();