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/Hashing.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/SmallVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/Analysis/AliasSetTracker.h"
38 #include "llvm/Analysis/Dominators.h"
39 #include "llvm/Analysis/InstructionSimplify.h"
40 #include "llvm/Analysis/ValueTracking.h"
41 #include "llvm/DIBuilder.h"
42 #include "llvm/DebugInfo.h"
43 #include "llvm/IR/Constants.h"
44 #include "llvm/IR/DerivedTypes.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/Instructions.h"
47 #include "llvm/IR/IntrinsicInst.h"
48 #include "llvm/IR/Metadata.h"
49 #include "llvm/Support/CFG.h"
50 #include "llvm/Transforms/Utils/Local.h"
55 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
56 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
57 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
58 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
60 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
61 // FIXME: If the memory unit is of pointer or integer type, we can permit
62 // assignments to subsections of the memory unit.
64 // Only allow direct and non-volatile loads and stores...
65 for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end();
66 UI != UE; ++UI) { // Loop over all of the uses of the alloca
68 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
69 // Note that atomic loads can be transformed; atomic semantics do
70 // not have any meaning for a local alloca.
73 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
74 if (SI->getOperand(0) == AI)
75 return false; // Don't allow a store OF the AI, only INTO the AI.
76 // Note that atomic stores can be transformed; atomic semantics do
77 // not have any meaning for a local alloca.
80 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
81 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
82 II->getIntrinsicID() != Intrinsic::lifetime_end)
84 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
85 if (BCI->getType() != Type::getInt8PtrTy(U->getContext()))
87 if (!onlyUsedByLifetimeMarkers(BCI))
89 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
90 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext()))
92 if (!GEPI->hasAllZeroIndices())
94 if (!onlyUsedByLifetimeMarkers(GEPI))
107 SmallVector<BasicBlock *, 32> DefiningBlocks;
108 SmallVector<BasicBlock *, 32> UsingBlocks;
110 StoreInst *OnlyStore;
111 BasicBlock *OnlyBlock;
112 bool OnlyUsedInOneBlock;
114 Value *AllocaPointerVal;
115 DbgDeclareInst *DbgDeclare;
118 DefiningBlocks.clear();
122 OnlyUsedInOneBlock = true;
123 AllocaPointerVal = 0;
127 /// Scan the uses of the specified alloca, filling in the AllocaInfo used
128 /// by the rest of the pass to reason about the uses of this alloca.
129 void AnalyzeAlloca(AllocaInst *AI) {
132 // As we scan the uses of the alloca instruction, keep track of stores,
133 // and decide whether all of the loads and stores to the alloca are within
134 // the same basic block.
135 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
137 Instruction *User = cast<Instruction>(*UI++);
139 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
140 // Remember the basic blocks which define new values for the alloca
141 DefiningBlocks.push_back(SI->getParent());
142 AllocaPointerVal = SI->getOperand(0);
145 LoadInst *LI = cast<LoadInst>(User);
146 // Otherwise it must be a load instruction, keep track of variable
148 UsingBlocks.push_back(LI->getParent());
149 AllocaPointerVal = LI;
152 if (OnlyUsedInOneBlock) {
154 OnlyBlock = User->getParent();
155 else if (OnlyBlock != User->getParent())
156 OnlyUsedInOneBlock = false;
160 DbgDeclare = FindAllocaDbgDeclare(AI);
164 // Data package used by RenamePass()
165 class RenamePassData {
167 typedef std::vector<Value *> ValVector;
169 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
170 RenamePassData(BasicBlock *B, BasicBlock *P, const ValVector &V)
171 : BB(B), Pred(P), Values(V) {}
176 void swap(RenamePassData &RHS) {
177 std::swap(BB, RHS.BB);
178 std::swap(Pred, RHS.Pred);
179 Values.swap(RHS.Values);
183 /// \brief This assigns and keeps a per-bb relative ordering of load/store
184 /// instructions in the block that directly load or store an alloca.
186 /// This functionality is important because it avoids scanning large basic
187 /// blocks multiple times when promoting many allocas in the same block.
188 class LargeBlockInfo {
189 /// \brief For each instruction that we track, keep the index of the
192 /// The index starts out as the number of the instruction from the start of
194 DenseMap<const Instruction *, unsigned> InstNumbers;
198 /// This code only looks at accesses to allocas.
199 static bool isInterestingInstruction(const Instruction *I) {
200 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
201 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
204 /// Get or calculate the index of the specified instruction.
205 unsigned getInstructionIndex(const Instruction *I) {
206 assert(isInterestingInstruction(I) &&
207 "Not a load/store to/from an alloca?");
209 // If we already have this instruction number, return it.
210 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
211 if (It != InstNumbers.end())
214 // Scan the whole block to get the instruction. This accumulates
215 // information for every interesting instruction in the block, in order to
216 // avoid gratuitus rescans.
217 const BasicBlock *BB = I->getParent();
219 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end(); BBI != E;
221 if (isInterestingInstruction(BBI))
222 InstNumbers[BBI] = InstNo++;
223 It = InstNumbers.find(I);
225 assert(It != InstNumbers.end() && "Didn't insert instruction?");
229 void deleteValue(const Instruction *I) { InstNumbers.erase(I); }
231 void clear() { InstNumbers.clear(); }
234 struct PromoteMem2Reg {
235 /// The alloca instructions being promoted.
236 std::vector<AllocaInst *> Allocas;
240 /// An AliasSetTracker object to update. If null, don't update it.
241 AliasSetTracker *AST;
243 /// Reverse mapping of Allocas.
244 DenseMap<AllocaInst *, unsigned> AllocaLookup;
246 /// \brief The PhiNodes we're adding.
248 /// That map is used to simplify some Phi nodes as we iterate over it, so
249 /// it should have deterministic iterators. We could use a MapVector, but
250 /// since we already maintain a map from BasicBlock* to a stable numbering
251 /// (BBNumbers), the DenseMap is more efficient (also supports removal).
252 DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes;
254 /// For each PHI node, keep track of which entry in Allocas it corresponds
256 DenseMap<PHINode *, unsigned> PhiToAllocaMap;
258 /// If we are updating an AliasSetTracker, then for each alloca that is of
259 /// pointer type, we keep track of what to copyValue to the inserted PHI
261 std::vector<Value *> PointerAllocaValues;
263 /// For each alloca, we keep track of the dbg.declare intrinsic that
264 /// describes it, if any, so that we can convert it to a dbg.value
265 /// intrinsic if the alloca gets promoted.
266 SmallVector<DbgDeclareInst *, 8> AllocaDbgDeclares;
268 /// The set of basic blocks the renamer has already visited.
270 SmallPtrSet<BasicBlock *, 16> Visited;
272 /// Contains a stable numbering of basic blocks to avoid non-determinstic
274 DenseMap<BasicBlock *, unsigned> BBNumbers;
276 /// Maps DomTreeNodes to their level in the dominator tree.
277 DenseMap<DomTreeNode *, unsigned> DomLevels;
279 /// Lazily compute the number of predecessors a block has.
280 DenseMap<const BasicBlock *, unsigned> BBNumPreds;
283 PromoteMem2Reg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
284 AliasSetTracker *AST)
285 : Allocas(Allocas.begin(), Allocas.end()), DT(DT),
286 DIB(*DT.getRoot()->getParent()->getParent()), AST(AST) {}
291 void RemoveFromAllocasList(unsigned &AllocaIdx) {
292 Allocas[AllocaIdx] = Allocas.back();
297 unsigned getNumPreds(const BasicBlock *BB) {
298 unsigned &NP = BBNumPreds[BB];
300 NP = std::distance(pred_begin(BB), pred_end(BB)) + 1;
304 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
306 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
307 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
308 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks);
309 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
310 RenamePassData::ValVector &IncVals,
311 std::vector<RenamePassData> &Worklist);
312 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
315 } // end of anonymous namespace
317 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
318 // Knowing that this alloca is promotable, we know that it's safe to kill all
319 // instructions except for load and store.
321 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
323 Instruction *I = cast<Instruction>(*UI);
325 if (isa<LoadInst>(I) || isa<StoreInst>(I))
328 if (!I->getType()->isVoidTy()) {
329 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
330 // Follow the use/def chain to erase them now instead of leaving it for
331 // dead code elimination later.
332 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
334 Instruction *Inst = cast<Instruction>(*UI);
336 Inst->eraseFromParent();
339 I->eraseFromParent();
343 /// \brief Rewrite as many loads as possible given a single store.
345 /// When there is only a single store, we can use the domtree to trivially
346 /// replace all of the dominated loads with the stored value. Do so, and return
347 /// true if this has successfully promoted the alloca entirely. If this returns
348 /// false there were some loads which were not dominated by the single store
349 /// and thus must be phi-ed with undef. We fall back to the standard alloca
350 /// promotion algorithm in that case.
351 static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
354 AliasSetTracker *AST) {
355 StoreInst *OnlyStore = Info.OnlyStore;
356 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
357 BasicBlock *StoreBB = OnlyStore->getParent();
360 // Clear out UsingBlocks. We will reconstruct it here if needed.
361 Info.UsingBlocks.clear();
363 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
364 Instruction *UserInst = cast<Instruction>(*UI++);
365 if (!isa<LoadInst>(UserInst)) {
366 assert(UserInst == OnlyStore && "Should only have load/stores");
369 LoadInst *LI = cast<LoadInst>(UserInst);
371 // Okay, if we have a load from the alloca, we want to replace it with the
372 // only value stored to the alloca. We can do this if the value is
373 // dominated by the store. If not, we use the rest of the mem2reg machinery
374 // to insert the phi nodes as needed.
375 if (!StoringGlobalVal) { // Non-instructions are always dominated.
376 if (LI->getParent() == StoreBB) {
377 // If we have a use that is in the same block as the store, compare the
378 // indices of the two instructions to see which one came first. If the
379 // load came before the store, we can't handle it.
380 if (StoreIndex == -1)
381 StoreIndex = LBI.getInstructionIndex(OnlyStore);
383 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
384 // Can't handle this load, bail out.
385 Info.UsingBlocks.push_back(StoreBB);
389 } else if (LI->getParent() != StoreBB &&
390 !DT.dominates(StoreBB, LI->getParent())) {
391 // If the load and store are in different blocks, use BB dominance to
392 // check their relationships. If the store doesn't dom the use, bail
394 Info.UsingBlocks.push_back(LI->getParent());
399 // Otherwise, we *can* safely rewrite this load.
400 Value *ReplVal = OnlyStore->getOperand(0);
401 // If the replacement value is the load, this must occur in unreachable
404 ReplVal = UndefValue::get(LI->getType());
405 LI->replaceAllUsesWith(ReplVal);
406 if (AST && LI->getType()->isPointerTy())
407 AST->deleteValue(LI);
408 LI->eraseFromParent();
412 // Finally, after the scan, check to see if the store is all that is left.
413 if (!Info.UsingBlocks.empty())
414 return false; // If not, we'll have to fall back for the remainder.
416 // Record debuginfo for the store and remove the declaration's
418 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
419 DIBuilder DIB(*AI->getParent()->getParent()->getParent());
420 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB);
421 DDI->eraseFromParent();
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());
478 // Walk all of the loads from this alloca, replacing them with the nearest
479 // store above them, if any.
480 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
481 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
485 unsigned LoadIdx = LBI.getInstructionIndex(LI);
487 // Find the nearest store that has a lower than this load.
488 StoresByIndexTy::iterator I = std::lower_bound(
489 StoresByIndex.begin(), StoresByIndex.end(),
490 std::pair<unsigned, StoreInst *>(LoadIdx, static_cast<StoreInst *>(0)),
491 StoreIndexSearchPredicate());
493 if (I == StoresByIndex.begin())
494 // If there is no store before this load, the load takes the undef value.
495 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
497 // Otherwise, there was a store before this load, the load takes its value.
498 LI->replaceAllUsesWith(llvm::prior(I)->second->getOperand(0));
500 if (AST && LI->getType()->isPointerTy())
501 AST->deleteValue(LI);
502 LI->eraseFromParent();
506 // Remove the (now dead) stores and alloca.
507 while (!AI->use_empty()) {
508 StoreInst *SI = cast<StoreInst>(AI->use_back());
509 // Record debuginfo for the store before removing it.
510 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
511 DIBuilder DIB(*AI->getParent()->getParent()->getParent());
512 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
514 SI->eraseFromParent();
519 AST->deleteValue(AI);
520 AI->eraseFromParent();
523 // The alloca's debuginfo can be removed as well.
524 if (DbgDeclareInst *DDI = Info.DbgDeclare)
525 DDI->eraseFromParent();
530 void PromoteMem2Reg::run() {
531 Function &F = *DT.getRoot()->getParent();
534 PointerAllocaValues.resize(Allocas.size());
535 AllocaDbgDeclares.resize(Allocas.size());
540 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
541 AllocaInst *AI = Allocas[AllocaNum];
543 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
544 assert(AI->getParent()->getParent() == &F &&
545 "All allocas should be in the same function, which is same as DF!");
547 removeLifetimeIntrinsicUsers(AI);
549 if (AI->use_empty()) {
550 // If there are no uses of the alloca, just delete it now.
552 AST->deleteValue(AI);
553 AI->eraseFromParent();
555 // Remove the alloca from the Allocas list, since it has been processed
556 RemoveFromAllocasList(AllocaNum);
561 // Calculate the set of read and write-locations for each alloca. This is
562 // analogous to finding the 'uses' and 'definitions' of each variable.
563 Info.AnalyzeAlloca(AI);
565 // If there is only a single store to this value, replace any loads of
566 // it that are directly dominated by the definition with the value stored.
567 if (Info.DefiningBlocks.size() == 1) {
568 if (rewriteSingleStoreAlloca(AI, Info, LBI, DT, AST)) {
569 // The alloca has been processed, move on.
570 RemoveFromAllocasList(AllocaNum);
576 // If the alloca is only read and written in one basic block, just perform a
577 // linear sweep over the block to eliminate it.
578 if (Info.OnlyUsedInOneBlock) {
579 promoteSingleBlockAlloca(AI, Info, LBI, AST);
581 // The alloca has been processed, move on.
582 RemoveFromAllocasList(AllocaNum);
586 // If we haven't computed dominator tree levels, do so now.
587 if (DomLevels.empty()) {
588 SmallVector<DomTreeNode *, 32> Worklist;
590 DomTreeNode *Root = DT.getRootNode();
592 Worklist.push_back(Root);
594 while (!Worklist.empty()) {
595 DomTreeNode *Node = Worklist.pop_back_val();
596 unsigned ChildLevel = DomLevels[Node] + 1;
597 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
599 DomLevels[*CI] = ChildLevel;
600 Worklist.push_back(*CI);
605 // If we haven't computed a numbering for the BB's in the function, do so
607 if (BBNumbers.empty()) {
609 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
613 // If we have an AST to keep updated, remember some pointer value that is
614 // stored into the alloca.
616 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
618 // Remember the dbg.declare intrinsic describing this alloca, if any.
620 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
622 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
623 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
625 // At this point, we're committed to promoting the alloca using IDF's, and
626 // the standard SSA construction algorithm. Determine which blocks need PHI
627 // nodes and see if we can optimize out some work by avoiding insertion of
629 DetermineInsertionPoint(AI, AllocaNum, Info);
633 return; // All of the allocas must have been trivial!
637 // Set the incoming values for the basic block to be null values for all of
638 // the alloca's. We do this in case there is a load of a value that has not
639 // been stored yet. In this case, it will get this null value.
641 RenamePassData::ValVector Values(Allocas.size());
642 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
643 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
645 // Walks all basic blocks in the function performing the SSA rename algorithm
646 // and inserting the phi nodes we marked as necessary
648 std::vector<RenamePassData> RenamePassWorkList;
649 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
652 RPD.swap(RenamePassWorkList.back());
653 RenamePassWorkList.pop_back();
654 // RenamePass may add new worklist entries.
655 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
656 } while (!RenamePassWorkList.empty());
658 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
661 // Remove the allocas themselves from the function.
662 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
663 Instruction *A = Allocas[i];
665 // If there are any uses of the alloca instructions left, they must be in
666 // unreachable basic blocks that were not processed by walking the dominator
667 // tree. Just delete the users now.
669 A->replaceAllUsesWith(UndefValue::get(A->getType()));
672 A->eraseFromParent();
675 // Remove alloca's dbg.declare instrinsics from the function.
676 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
677 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
678 DDI->eraseFromParent();
680 // Loop over all of the PHI nodes and see if there are any that we can get
681 // rid of because they merge all of the same incoming values. This can
682 // happen due to undef values coming into the PHI nodes. This process is
683 // iterative, because eliminating one PHI node can cause others to be removed.
684 bool EliminatedAPHI = true;
685 while (EliminatedAPHI) {
686 EliminatedAPHI = false;
688 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
689 // simplify and RAUW them as we go. If it was not, we could add uses to
690 // the values we replace with in a non deterministic order, thus creating
691 // non deterministic def->use chains.
692 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
693 I = NewPhiNodes.begin(),
694 E = NewPhiNodes.end();
696 PHINode *PN = I->second;
698 // If this PHI node merges one value and/or undefs, get the value.
699 if (Value *V = SimplifyInstruction(PN, 0, 0, &DT)) {
700 if (AST && PN->getType()->isPointerTy())
701 AST->deleteValue(PN);
702 PN->replaceAllUsesWith(V);
703 PN->eraseFromParent();
704 NewPhiNodes.erase(I++);
705 EliminatedAPHI = true;
712 // At this point, the renamer has added entries to PHI nodes for all reachable
713 // code. Unfortunately, there may be unreachable blocks which the renamer
714 // hasn't traversed. If this is the case, the PHI nodes may not
715 // have incoming values for all predecessors. Loop over all PHI nodes we have
716 // created, inserting undef values if they are missing any incoming values.
718 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
719 I = NewPhiNodes.begin(),
720 E = NewPhiNodes.end();
722 // We want to do this once per basic block. As such, only process a block
723 // when we find the PHI that is the first entry in the block.
724 PHINode *SomePHI = I->second;
725 BasicBlock *BB = SomePHI->getParent();
726 if (&BB->front() != SomePHI)
729 // Only do work here if there the PHI nodes are missing incoming values. We
730 // know that all PHI nodes that were inserted in a block will have the same
731 // number of incoming values, so we can just check any of them.
732 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
735 // Get the preds for BB.
736 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
738 // Ok, now we know that all of the PHI nodes are missing entries for some
739 // basic blocks. Start by sorting the incoming predecessors for efficient
741 std::sort(Preds.begin(), Preds.end());
743 // Now we loop through all BB's which have entries in SomePHI and remove
744 // them from the Preds list.
745 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
746 // Do a log(n) search of the Preds list for the entry we want.
747 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
748 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
749 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
750 "PHI node has entry for a block which is not a predecessor!");
756 // At this point, the blocks left in the preds list must have dummy
757 // entries inserted into every PHI nodes for the block. Update all the phi
758 // nodes in this block that we are inserting (there could be phis before
760 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
761 BasicBlock::iterator BBI = BB->begin();
762 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
763 SomePHI->getNumIncomingValues() == NumBadPreds) {
764 Value *UndefVal = UndefValue::get(SomePHI->getType());
765 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
766 SomePHI->addIncoming(UndefVal, Preds[pred]);
773 /// \brief Determine which blocks the value is live in.
775 /// These are blocks which lead to uses. Knowing this allows us to avoid
776 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
777 /// inserted phi nodes would be dead).
778 void PromoteMem2Reg::ComputeLiveInBlocks(
779 AllocaInst *AI, AllocaInfo &Info,
780 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
781 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks) {
783 // To determine liveness, we must iterate through the predecessors of blocks
784 // where the def is live. Blocks are added to the worklist if we need to
785 // check their predecessors. Start with all the using blocks.
786 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
787 Info.UsingBlocks.end());
789 // If any of the using blocks is also a definition block, check to see if the
790 // definition occurs before or after the use. If it happens before the use,
791 // the value isn't really live-in.
792 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
793 BasicBlock *BB = LiveInBlockWorklist[i];
794 if (!DefBlocks.count(BB))
797 // Okay, this is a block that both uses and defines the value. If the first
798 // reference to the alloca is a def (store), then we know it isn't live-in.
799 for (BasicBlock::iterator I = BB->begin();; ++I) {
800 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
801 if (SI->getOperand(1) != AI)
804 // We found a store to the alloca before a load. The alloca is not
805 // actually live-in here.
806 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
807 LiveInBlockWorklist.pop_back();
812 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
813 if (LI->getOperand(0) != AI)
816 // Okay, we found a load before a store to the alloca. It is actually
817 // live into this block.
823 // Now that we have a set of blocks where the phi is live-in, recursively add
824 // their predecessors until we find the full region the value is live.
825 while (!LiveInBlockWorklist.empty()) {
826 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
828 // The block really is live in here, insert it into the set. If already in
829 // the set, then it has already been processed.
830 if (!LiveInBlocks.insert(BB))
833 // Since the value is live into BB, it is either defined in a predecessor or
834 // live into it to. Add the preds to the worklist unless they are a
836 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
839 // The value is not live into a predecessor if it defines the value.
840 if (DefBlocks.count(P))
843 // Otherwise it is, add to the worklist.
844 LiveInBlockWorklist.push_back(P);
850 typedef std::pair<DomTreeNode *, unsigned> DomTreeNodePair;
852 struct DomTreeNodeCompare {
853 bool operator()(const DomTreeNodePair &LHS, const DomTreeNodePair &RHS) {
854 return LHS.second < RHS.second;
857 } // end anonymous namespace
859 /// At this point, we're committed to promoting the alloca using IDF's, and the
860 /// standard SSA construction algorithm. Determine which blocks need phi nodes
861 /// and see if we can optimize out some work by avoiding insertion of dead phi
863 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
865 // Unique the set of defining blocks for efficient lookup.
866 SmallPtrSet<BasicBlock *, 32> DefBlocks;
867 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
869 // Determine which blocks the value is live in. These are blocks which lead
871 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
872 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
874 // Use a priority queue keyed on dominator tree level so that inserted nodes
875 // are handled from the bottom of the dominator tree upwards.
876 typedef std::priority_queue<DomTreeNodePair,
877 SmallVector<DomTreeNodePair, 32>,
878 DomTreeNodeCompare> IDFPriorityQueue;
881 for (SmallPtrSet<BasicBlock *, 32>::const_iterator I = DefBlocks.begin(),
884 if (DomTreeNode *Node = DT.getNode(*I))
885 PQ.push(std::make_pair(Node, DomLevels[Node]));
888 SmallVector<std::pair<unsigned, BasicBlock *>, 32> DFBlocks;
889 SmallPtrSet<DomTreeNode *, 32> Visited;
890 SmallVector<DomTreeNode *, 32> Worklist;
891 while (!PQ.empty()) {
892 DomTreeNodePair RootPair = PQ.top();
894 DomTreeNode *Root = RootPair.first;
895 unsigned RootLevel = RootPair.second;
897 // Walk all dominator tree children of Root, inspecting their CFG edges with
898 // targets elsewhere on the dominator tree. Only targets whose level is at
899 // most Root's level are added to the iterated dominance frontier of the
903 Worklist.push_back(Root);
905 while (!Worklist.empty()) {
906 DomTreeNode *Node = Worklist.pop_back_val();
907 BasicBlock *BB = Node->getBlock();
909 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
911 DomTreeNode *SuccNode = DT.getNode(*SI);
913 // Quickly skip all CFG edges that are also dominator tree edges instead
914 // of catching them below.
915 if (SuccNode->getIDom() == Node)
918 unsigned SuccLevel = DomLevels[SuccNode];
919 if (SuccLevel > RootLevel)
922 if (!Visited.insert(SuccNode))
925 BasicBlock *SuccBB = SuccNode->getBlock();
926 if (!LiveInBlocks.count(SuccBB))
929 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
930 if (!DefBlocks.count(SuccBB))
931 PQ.push(std::make_pair(SuccNode, SuccLevel));
934 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
936 if (!Visited.count(*CI))
937 Worklist.push_back(*CI);
942 if (DFBlocks.size() > 1)
943 std::sort(DFBlocks.begin(), DFBlocks.end());
945 unsigned CurrentVersion = 0;
946 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
947 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
950 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
952 /// Returns true if there wasn't already a phi-node for that variable
953 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
955 // Look up the basic-block in question.
956 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
958 // If the BB already has a phi node added for the i'th alloca then we're done!
962 // Create a PhiNode using the dereferenced type... and add the phi-node to the
964 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
965 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
968 PhiToAllocaMap[PN] = AllocaNo;
970 if (AST && PN->getType()->isPointerTy())
971 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
976 /// \brief Recursively traverse the CFG of the function, renaming loads and
977 /// stores to the allocas which we are promoting.
979 /// IncomingVals indicates what value each Alloca contains on exit from the
980 /// predecessor block Pred.
981 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
982 RenamePassData::ValVector &IncomingVals,
983 std::vector<RenamePassData> &Worklist) {
985 // If we are inserting any phi nodes into this BB, they will already be in the
987 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
988 // If we have PHI nodes to update, compute the number of edges from Pred to
990 if (PhiToAllocaMap.count(APN)) {
991 // We want to be able to distinguish between PHI nodes being inserted by
992 // this invocation of mem2reg from those phi nodes that already existed in
993 // the IR before mem2reg was run. We determine that APN is being inserted
994 // because it is missing incoming edges. All other PHI nodes being
995 // inserted by this pass of mem2reg will have the same number of incoming
996 // operands so far. Remember this count.
997 unsigned NewPHINumOperands = APN->getNumOperands();
999 unsigned NumEdges = 0;
1000 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
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();