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/ArrayRef.h"
31 #include "llvm/ADT/DenseMap.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(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
283 AliasSetTracker *AST)
284 : Allocas(Allocas.begin(), Allocas.end()), DT(DT),
285 DIB(*DT.getRoot()->getParent()->getParent()), AST(AST) {}
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();
421 LBI.deleteValue(DDI);
423 // Remove the (now dead) store and alloca.
424 Info.OnlyStore->eraseFromParent();
425 LBI.deleteValue(Info.OnlyStore);
428 AST->deleteValue(AI);
429 AI->eraseFromParent();
435 /// This is a helper predicate used to search by the first element of a pair.
436 struct StoreIndexSearchPredicate {
437 bool operator()(const std::pair<unsigned, StoreInst *> &LHS,
438 const std::pair<unsigned, StoreInst *> &RHS) {
439 return LHS.first < RHS.first;
444 /// Many allocas are only used within a single basic block. If this is the
445 /// case, avoid traversing the CFG and inserting a lot of potentially useless
446 /// PHI nodes by just performing a single linear pass over the basic block
447 /// using the Alloca.
449 /// If we cannot promote this alloca (because it is read before it is written),
450 /// return true. This is necessary in cases where, due to control flow, the
451 /// alloca is potentially undefined on some control flow paths. e.g. code like
452 /// this is potentially correct:
454 /// for (...) { if (c) { A = undef; undef = B; } }
456 /// ... so long as A is not used before undef is set.
457 static void promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
459 AliasSetTracker *AST) {
460 // The trickiest case to handle is when we have large blocks. Because of this,
461 // this code is optimized assuming that large blocks happen. This does not
462 // significantly pessimize the small block case. This uses LargeBlockInfo to
463 // make it efficient to get the index of various operations in the block.
465 // Walk the use-def list of the alloca, getting the locations of all stores.
466 typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy;
467 StoresByIndexTy StoresByIndex;
469 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;
471 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
472 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
474 // Sort the stores by their index, making it efficient to do a lookup with a
476 std::sort(StoresByIndex.begin(), StoresByIndex.end(),
477 StoreIndexSearchPredicate());
479 // Walk all of the loads from this alloca, replacing them with the nearest
480 // store above them, if any.
481 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
482 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
486 unsigned LoadIdx = LBI.getInstructionIndex(LI);
488 // Find the nearest store that has a lower index than this load.
489 StoresByIndexTy::iterator I =
490 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
491 std::make_pair(LoadIdx, static_cast<StoreInst *>(0)),
492 StoreIndexSearchPredicate());
494 if (I == StoresByIndex.begin())
495 // If there is no store before this load, the load takes the undef value.
496 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
498 // Otherwise, there was a store before this load, the load takes its value.
499 LI->replaceAllUsesWith(llvm::prior(I)->second->getOperand(0));
501 if (AST && LI->getType()->isPointerTy())
502 AST->deleteValue(LI);
503 LI->eraseFromParent();
507 // Remove the (now dead) stores and alloca.
508 while (!AI->use_empty()) {
509 StoreInst *SI = cast<StoreInst>(AI->use_back());
510 // Record debuginfo for the store before removing it.
511 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
512 DIBuilder DIB(*AI->getParent()->getParent()->getParent());
513 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
515 SI->eraseFromParent();
520 AST->deleteValue(AI);
521 AI->eraseFromParent();
524 // The alloca's debuginfo can be removed as well.
525 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
526 DDI->eraseFromParent();
527 LBI.deleteValue(DDI);
533 void PromoteMem2Reg::run() {
534 Function &F = *DT.getRoot()->getParent();
537 PointerAllocaValues.resize(Allocas.size());
538 AllocaDbgDeclares.resize(Allocas.size());
543 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
544 AllocaInst *AI = Allocas[AllocaNum];
546 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
547 assert(AI->getParent()->getParent() == &F &&
548 "All allocas should be in the same function, which is same as DF!");
550 removeLifetimeIntrinsicUsers(AI);
552 if (AI->use_empty()) {
553 // If there are no uses of the alloca, just delete it now.
555 AST->deleteValue(AI);
556 AI->eraseFromParent();
558 // Remove the alloca from the Allocas list, since it has been processed
559 RemoveFromAllocasList(AllocaNum);
564 // Calculate the set of read and write-locations for each alloca. This is
565 // analogous to finding the 'uses' and 'definitions' of each variable.
566 Info.AnalyzeAlloca(AI);
568 // If there is only a single store to this value, replace any loads of
569 // it that are directly dominated by the definition with the value stored.
570 if (Info.DefiningBlocks.size() == 1) {
571 if (rewriteSingleStoreAlloca(AI, Info, LBI, DT, AST)) {
572 // The alloca has been processed, move on.
573 RemoveFromAllocasList(AllocaNum);
579 // If the alloca is only read and written in one basic block, just perform a
580 // linear sweep over the block to eliminate it.
581 if (Info.OnlyUsedInOneBlock) {
582 promoteSingleBlockAlloca(AI, Info, LBI, AST);
584 // The alloca has been processed, move on.
585 RemoveFromAllocasList(AllocaNum);
589 // If we haven't computed dominator tree levels, do so now.
590 if (DomLevels.empty()) {
591 SmallVector<DomTreeNode *, 32> Worklist;
593 DomTreeNode *Root = DT.getRootNode();
595 Worklist.push_back(Root);
597 while (!Worklist.empty()) {
598 DomTreeNode *Node = Worklist.pop_back_val();
599 unsigned ChildLevel = DomLevels[Node] + 1;
600 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
602 DomLevels[*CI] = ChildLevel;
603 Worklist.push_back(*CI);
608 // If we haven't computed a numbering for the BB's in the function, do so
610 if (BBNumbers.empty()) {
612 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
616 // If we have an AST to keep updated, remember some pointer value that is
617 // stored into the alloca.
619 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
621 // Remember the dbg.declare intrinsic describing this alloca, if any.
623 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
625 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
626 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
628 // At this point, we're committed to promoting the alloca using IDF's, and
629 // the standard SSA construction algorithm. Determine which blocks need PHI
630 // nodes and see if we can optimize out some work by avoiding insertion of
632 DetermineInsertionPoint(AI, AllocaNum, Info);
636 return; // All of the allocas must have been trivial!
640 // Set the incoming values for the basic block to be null values for all of
641 // the alloca's. We do this in case there is a load of a value that has not
642 // been stored yet. In this case, it will get this null value.
644 RenamePassData::ValVector Values(Allocas.size());
645 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
646 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
648 // Walks all basic blocks in the function performing the SSA rename algorithm
649 // and inserting the phi nodes we marked as necessary
651 std::vector<RenamePassData> RenamePassWorkList;
652 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
655 RPD.swap(RenamePassWorkList.back());
656 RenamePassWorkList.pop_back();
657 // RenamePass may add new worklist entries.
658 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
659 } while (!RenamePassWorkList.empty());
661 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
664 // Remove the allocas themselves from the function.
665 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
666 Instruction *A = Allocas[i];
668 // If there are any uses of the alloca instructions left, they must be in
669 // unreachable basic blocks that were not processed by walking the dominator
670 // tree. Just delete the users now.
672 A->replaceAllUsesWith(UndefValue::get(A->getType()));
675 A->eraseFromParent();
678 // Remove alloca's dbg.declare instrinsics from the function.
679 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
680 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
681 DDI->eraseFromParent();
683 // Loop over all of the PHI nodes and see if there are any that we can get
684 // rid of because they merge all of the same incoming values. This can
685 // happen due to undef values coming into the PHI nodes. This process is
686 // iterative, because eliminating one PHI node can cause others to be removed.
687 bool EliminatedAPHI = true;
688 while (EliminatedAPHI) {
689 EliminatedAPHI = false;
691 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
692 // simplify and RAUW them as we go. If it was not, we could add uses to
693 // the values we replace with in a non deterministic order, thus creating
694 // non deterministic def->use chains.
695 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
696 I = NewPhiNodes.begin(),
697 E = NewPhiNodes.end();
699 PHINode *PN = I->second;
701 // If this PHI node merges one value and/or undefs, get the value.
702 if (Value *V = SimplifyInstruction(PN, 0, 0, &DT)) {
703 if (AST && PN->getType()->isPointerTy())
704 AST->deleteValue(PN);
705 PN->replaceAllUsesWith(V);
706 PN->eraseFromParent();
707 NewPhiNodes.erase(I++);
708 EliminatedAPHI = true;
715 // At this point, the renamer has added entries to PHI nodes for all reachable
716 // code. Unfortunately, there may be unreachable blocks which the renamer
717 // hasn't traversed. If this is the case, the PHI nodes may not
718 // have incoming values for all predecessors. Loop over all PHI nodes we have
719 // created, inserting undef values if they are missing any incoming values.
721 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
722 I = NewPhiNodes.begin(),
723 E = NewPhiNodes.end();
725 // We want to do this once per basic block. As such, only process a block
726 // when we find the PHI that is the first entry in the block.
727 PHINode *SomePHI = I->second;
728 BasicBlock *BB = SomePHI->getParent();
729 if (&BB->front() != SomePHI)
732 // Only do work here if there the PHI nodes are missing incoming values. We
733 // know that all PHI nodes that were inserted in a block will have the same
734 // number of incoming values, so we can just check any of them.
735 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
738 // Get the preds for BB.
739 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
741 // Ok, now we know that all of the PHI nodes are missing entries for some
742 // basic blocks. Start by sorting the incoming predecessors for efficient
744 std::sort(Preds.begin(), Preds.end());
746 // Now we loop through all BB's which have entries in SomePHI and remove
747 // them from the Preds list.
748 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
749 // Do a log(n) search of the Preds list for the entry we want.
750 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
751 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
752 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
753 "PHI node has entry for a block which is not a predecessor!");
759 // At this point, the blocks left in the preds list must have dummy
760 // entries inserted into every PHI nodes for the block. Update all the phi
761 // nodes in this block that we are inserting (there could be phis before
763 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
764 BasicBlock::iterator BBI = BB->begin();
765 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
766 SomePHI->getNumIncomingValues() == NumBadPreds) {
767 Value *UndefVal = UndefValue::get(SomePHI->getType());
768 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
769 SomePHI->addIncoming(UndefVal, Preds[pred]);
776 /// \brief Determine which blocks the value is live in.
778 /// These are blocks which lead to uses. Knowing this allows us to avoid
779 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
780 /// inserted phi nodes would be dead).
781 void PromoteMem2Reg::ComputeLiveInBlocks(
782 AllocaInst *AI, AllocaInfo &Info,
783 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
784 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks) {
786 // To determine liveness, we must iterate through the predecessors of blocks
787 // where the def is live. Blocks are added to the worklist if we need to
788 // check their predecessors. Start with all the using blocks.
789 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
790 Info.UsingBlocks.end());
792 // If any of the using blocks is also a definition block, check to see if the
793 // definition occurs before or after the use. If it happens before the use,
794 // the value isn't really live-in.
795 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
796 BasicBlock *BB = LiveInBlockWorklist[i];
797 if (!DefBlocks.count(BB))
800 // Okay, this is a block that both uses and defines the value. If the first
801 // reference to the alloca is a def (store), then we know it isn't live-in.
802 for (BasicBlock::iterator I = BB->begin();; ++I) {
803 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
804 if (SI->getOperand(1) != AI)
807 // We found a store to the alloca before a load. The alloca is not
808 // actually live-in here.
809 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
810 LiveInBlockWorklist.pop_back();
815 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
816 if (LI->getOperand(0) != AI)
819 // Okay, we found a load before a store to the alloca. It is actually
820 // live into this block.
826 // Now that we have a set of blocks where the phi is live-in, recursively add
827 // their predecessors until we find the full region the value is live.
828 while (!LiveInBlockWorklist.empty()) {
829 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
831 // The block really is live in here, insert it into the set. If already in
832 // the set, then it has already been processed.
833 if (!LiveInBlocks.insert(BB))
836 // Since the value is live into BB, it is either defined in a predecessor or
837 // live into it to. Add the preds to the worklist unless they are a
839 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
842 // The value is not live into a predecessor if it defines the value.
843 if (DefBlocks.count(P))
846 // Otherwise it is, add to the worklist.
847 LiveInBlockWorklist.push_back(P);
853 typedef std::pair<DomTreeNode *, unsigned> DomTreeNodePair;
855 struct DomTreeNodeCompare {
856 bool operator()(const DomTreeNodePair &LHS, const DomTreeNodePair &RHS) {
857 return LHS.second < RHS.second;
860 } // end anonymous namespace
862 /// At this point, we're committed to promoting the alloca using IDF's, and the
863 /// standard SSA construction algorithm. Determine which blocks need phi nodes
864 /// and see if we can optimize out some work by avoiding insertion of dead phi
866 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
868 // Unique the set of defining blocks for efficient lookup.
869 SmallPtrSet<BasicBlock *, 32> DefBlocks;
870 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
872 // Determine which blocks the value is live in. These are blocks which lead
874 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
875 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
877 // Use a priority queue keyed on dominator tree level so that inserted nodes
878 // are handled from the bottom of the dominator tree upwards.
879 typedef std::priority_queue<DomTreeNodePair,
880 SmallVector<DomTreeNodePair, 32>,
881 DomTreeNodeCompare> IDFPriorityQueue;
884 for (SmallPtrSet<BasicBlock *, 32>::const_iterator I = DefBlocks.begin(),
887 if (DomTreeNode *Node = DT.getNode(*I))
888 PQ.push(std::make_pair(Node, DomLevels[Node]));
891 SmallVector<std::pair<unsigned, BasicBlock *>, 32> DFBlocks;
892 SmallPtrSet<DomTreeNode *, 32> Visited;
893 SmallVector<DomTreeNode *, 32> Worklist;
894 while (!PQ.empty()) {
895 DomTreeNodePair RootPair = PQ.top();
897 DomTreeNode *Root = RootPair.first;
898 unsigned RootLevel = RootPair.second;
900 // Walk all dominator tree children of Root, inspecting their CFG edges with
901 // targets elsewhere on the dominator tree. Only targets whose level is at
902 // most Root's level are added to the iterated dominance frontier of the
906 Worklist.push_back(Root);
908 while (!Worklist.empty()) {
909 DomTreeNode *Node = Worklist.pop_back_val();
910 BasicBlock *BB = Node->getBlock();
912 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
914 DomTreeNode *SuccNode = DT.getNode(*SI);
916 // Quickly skip all CFG edges that are also dominator tree edges instead
917 // of catching them below.
918 if (SuccNode->getIDom() == Node)
921 unsigned SuccLevel = DomLevels[SuccNode];
922 if (SuccLevel > RootLevel)
925 if (!Visited.insert(SuccNode))
928 BasicBlock *SuccBB = SuccNode->getBlock();
929 if (!LiveInBlocks.count(SuccBB))
932 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
933 if (!DefBlocks.count(SuccBB))
934 PQ.push(std::make_pair(SuccNode, SuccLevel));
937 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
939 if (!Visited.count(*CI))
940 Worklist.push_back(*CI);
945 if (DFBlocks.size() > 1)
946 std::sort(DFBlocks.begin(), DFBlocks.end());
948 unsigned CurrentVersion = 0;
949 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
950 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
953 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
955 /// Returns true if there wasn't already a phi-node for that variable
956 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
958 // Look up the basic-block in question.
959 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
961 // If the BB already has a phi node added for the i'th alloca then we're done!
965 // Create a PhiNode using the dereferenced type... and add the phi-node to the
967 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
968 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
971 PhiToAllocaMap[PN] = AllocaNo;
973 if (AST && PN->getType()->isPointerTy())
974 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
979 /// \brief Recursively traverse the CFG of the function, renaming loads and
980 /// stores to the allocas which we are promoting.
982 /// IncomingVals indicates what value each Alloca contains on exit from the
983 /// predecessor block Pred.
984 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
985 RenamePassData::ValVector &IncomingVals,
986 std::vector<RenamePassData> &Worklist) {
988 // If we are inserting any phi nodes into this BB, they will already be in the
990 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
991 // If we have PHI nodes to update, compute the number of edges from Pred to
993 if (PhiToAllocaMap.count(APN)) {
994 // We want to be able to distinguish between PHI nodes being inserted by
995 // this invocation of mem2reg from those phi nodes that already existed in
996 // the IR before mem2reg was run. We determine that APN is being inserted
997 // because it is missing incoming edges. All other PHI nodes being
998 // inserted by this pass of mem2reg will have the same number of incoming
999 // operands so far. Remember this count.
1000 unsigned NewPHINumOperands = APN->getNumOperands();
1002 unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB);
1003 assert(NumEdges && "Must be at least one edge from Pred to BB!");
1005 // Add entries for all the phis.
1006 BasicBlock::iterator PNI = BB->begin();
1008 unsigned AllocaNo = PhiToAllocaMap[APN];
1010 // Add N incoming values to the PHI node.
1011 for (unsigned i = 0; i != NumEdges; ++i)
1012 APN->addIncoming(IncomingVals[AllocaNo], Pred);
1014 // The currently active variable for this block is now the PHI.
1015 IncomingVals[AllocaNo] = APN;
1017 // Get the next phi node.
1019 APN = dyn_cast<PHINode>(PNI);
1023 // Verify that it is missing entries. If not, it is not being inserted
1024 // by this mem2reg invocation so we want to ignore it.
1025 } while (APN->getNumOperands() == NewPHINumOperands);
1029 // Don't revisit blocks.
1030 if (!Visited.insert(BB))
1033 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) {
1034 Instruction *I = II++; // get the instruction, increment iterator
1036 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1037 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1041 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
1042 if (AI == AllocaLookup.end())
1045 Value *V = IncomingVals[AI->second];
1047 // Anything using the load now uses the current value.
1048 LI->replaceAllUsesWith(V);
1049 if (AST && LI->getType()->isPointerTy())
1050 AST->deleteValue(LI);
1051 BB->getInstList().erase(LI);
1052 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1053 // Delete this instruction and mark the name as the current holder of the
1055 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1059 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1060 if (ai == AllocaLookup.end())
1063 // what value were we writing?
1064 IncomingVals[ai->second] = SI->getOperand(0);
1065 // Record debuginfo for the store before removing it.
1066 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1067 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1068 BB->getInstList().erase(SI);
1072 // 'Recurse' to our successors.
1073 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1077 // Keep track of the successors so we don't visit the same successor twice
1078 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
1080 // Handle the first successor without using the worklist.
1081 VisitedSuccs.insert(*I);
1087 if (VisitedSuccs.insert(*I))
1088 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1093 void llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
1094 AliasSetTracker *AST) {
1095 // If there is nothing to do, bail out...
1096 if (Allocas.empty())
1099 PromoteMem2Reg(Allocas, DT, AST).run();