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 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
29 #include "llvm/ADT/ArrayRef.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Analysis/AliasSetTracker.h"
36 #include "llvm/Analysis/InstructionSimplify.h"
37 #include "llvm/Analysis/ValueTracking.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DIBuilder.h"
41 #include "llvm/IR/DebugInfo.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/Dominators.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/Transforms/Utils/Local.h"
53 #define DEBUG_TYPE "mem2reg"
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 (const User *U : AI->users()) {
66 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
67 // Note that atomic loads can be transformed; atomic semantics do
68 // not have any meaning for a local alloca.
71 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
72 if (SI->getOperand(0) == AI)
73 return false; // Don't allow a store OF the AI, only INTO the AI.
74 // Note that atomic stores can be transformed; atomic semantics do
75 // not have any meaning for a local alloca.
78 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
79 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
80 II->getIntrinsicID() != Intrinsic::lifetime_end)
82 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
83 if (BCI->getType() != Type::getInt8PtrTy(U->getContext()))
85 if (!onlyUsedByLifetimeMarkers(BCI))
87 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
88 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext()))
90 if (!GEPI->hasAllZeroIndices())
92 if (!onlyUsedByLifetimeMarkers(GEPI))
105 SmallVector<BasicBlock *, 32> DefiningBlocks;
106 SmallVector<BasicBlock *, 32> UsingBlocks;
108 StoreInst *OnlyStore;
109 BasicBlock *OnlyBlock;
110 bool OnlyUsedInOneBlock;
112 Value *AllocaPointerVal;
113 DbgDeclareInst *DbgDeclare;
116 DefiningBlocks.clear();
120 OnlyUsedInOneBlock = true;
121 AllocaPointerVal = 0;
125 /// Scan the uses of the specified alloca, filling in the AllocaInfo used
126 /// by the rest of the pass to reason about the uses of this alloca.
127 void AnalyzeAlloca(AllocaInst *AI) {
130 // As we scan the uses of the alloca instruction, keep track of stores,
131 // and decide whether all of the loads and stores to the alloca are within
132 // the same basic block.
133 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
134 Instruction *User = cast<Instruction>(*UI++);
136 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
137 // Remember the basic blocks which define new values for the alloca
138 DefiningBlocks.push_back(SI->getParent());
139 AllocaPointerVal = SI->getOperand(0);
142 LoadInst *LI = cast<LoadInst>(User);
143 // Otherwise it must be a load instruction, keep track of variable
145 UsingBlocks.push_back(LI->getParent());
146 AllocaPointerVal = LI;
149 if (OnlyUsedInOneBlock) {
151 OnlyBlock = User->getParent();
152 else if (OnlyBlock != User->getParent())
153 OnlyUsedInOneBlock = false;
157 DbgDeclare = FindAllocaDbgDeclare(AI);
161 // Data package used by RenamePass()
162 class RenamePassData {
164 typedef std::vector<Value *> ValVector;
166 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
167 RenamePassData(BasicBlock *B, BasicBlock *P, const ValVector &V)
168 : BB(B), Pred(P), Values(V) {}
173 void swap(RenamePassData &RHS) {
174 std::swap(BB, RHS.BB);
175 std::swap(Pred, RHS.Pred);
176 Values.swap(RHS.Values);
180 /// \brief This assigns and keeps a per-bb relative ordering of load/store
181 /// instructions in the block that directly load or store an alloca.
183 /// This functionality is important because it avoids scanning large basic
184 /// blocks multiple times when promoting many allocas in the same block.
185 class LargeBlockInfo {
186 /// \brief For each instruction that we track, keep the index of the
189 /// The index starts out as the number of the instruction from the start of
191 DenseMap<const Instruction *, unsigned> InstNumbers;
195 /// This code only looks at accesses to allocas.
196 static bool isInterestingInstruction(const Instruction *I) {
197 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
198 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
201 /// Get or calculate the index of the specified instruction.
202 unsigned getInstructionIndex(const Instruction *I) {
203 assert(isInterestingInstruction(I) &&
204 "Not a load/store to/from an alloca?");
206 // If we already have this instruction number, return it.
207 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
208 if (It != InstNumbers.end())
211 // Scan the whole block to get the instruction. This accumulates
212 // information for every interesting instruction in the block, in order to
213 // avoid gratuitus rescans.
214 const BasicBlock *BB = I->getParent();
216 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end(); BBI != E;
218 if (isInterestingInstruction(BBI))
219 InstNumbers[BBI] = InstNo++;
220 It = InstNumbers.find(I);
222 assert(It != InstNumbers.end() && "Didn't insert instruction?");
226 void deleteValue(const Instruction *I) { InstNumbers.erase(I); }
228 void clear() { InstNumbers.clear(); }
231 struct PromoteMem2Reg {
232 /// The alloca instructions being promoted.
233 std::vector<AllocaInst *> Allocas;
237 /// An AliasSetTracker object to update. If null, don't update it.
238 AliasSetTracker *AST;
240 /// Reverse mapping of Allocas.
241 DenseMap<AllocaInst *, unsigned> AllocaLookup;
243 /// \brief The PhiNodes we're adding.
245 /// That map is used to simplify some Phi nodes as we iterate over it, so
246 /// it should have deterministic iterators. We could use a MapVector, but
247 /// since we already maintain a map from BasicBlock* to a stable numbering
248 /// (BBNumbers), the DenseMap is more efficient (also supports removal).
249 DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes;
251 /// For each PHI node, keep track of which entry in Allocas it corresponds
253 DenseMap<PHINode *, unsigned> PhiToAllocaMap;
255 /// If we are updating an AliasSetTracker, then for each alloca that is of
256 /// pointer type, we keep track of what to copyValue to the inserted PHI
258 std::vector<Value *> PointerAllocaValues;
260 /// For each alloca, we keep track of the dbg.declare intrinsic that
261 /// describes it, if any, so that we can convert it to a dbg.value
262 /// intrinsic if the alloca gets promoted.
263 SmallVector<DbgDeclareInst *, 8> AllocaDbgDeclares;
265 /// The set of basic blocks the renamer has already visited.
267 SmallPtrSet<BasicBlock *, 16> Visited;
269 /// Contains a stable numbering of basic blocks to avoid non-determinstic
271 DenseMap<BasicBlock *, unsigned> BBNumbers;
273 /// Maps DomTreeNodes to their level in the dominator tree.
274 DenseMap<DomTreeNode *, unsigned> DomLevels;
276 /// Lazily compute the number of predecessors a block has.
277 DenseMap<const BasicBlock *, unsigned> BBNumPreds;
280 PromoteMem2Reg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
281 AliasSetTracker *AST)
282 : Allocas(Allocas.begin(), Allocas.end()), DT(DT),
283 DIB(*DT.getRoot()->getParent()->getParent()), AST(AST) {}
288 void RemoveFromAllocasList(unsigned &AllocaIdx) {
289 Allocas[AllocaIdx] = Allocas.back();
294 unsigned getNumPreds(const BasicBlock *BB) {
295 unsigned &NP = BBNumPreds[BB];
297 NP = std::distance(pred_begin(BB), pred_end(BB)) + 1;
301 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
303 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
304 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
305 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks);
306 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
307 RenamePassData::ValVector &IncVals,
308 std::vector<RenamePassData> &Worklist);
309 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
312 } // end of anonymous namespace
314 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
315 // Knowing that this alloca is promotable, we know that it's safe to kill all
316 // instructions except for load and store.
318 for (auto UI = AI->user_begin(), UE = AI->user_end(); UI != UE;) {
319 Instruction *I = cast<Instruction>(*UI);
321 if (isa<LoadInst>(I) || isa<StoreInst>(I))
324 if (!I->getType()->isVoidTy()) {
325 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
326 // Follow the use/def chain to erase them now instead of leaving it for
327 // dead code elimination later.
328 for (auto UUI = I->user_begin(), UUE = I->user_end(); UUI != UUE;) {
329 Instruction *Inst = cast<Instruction>(*UUI);
331 Inst->eraseFromParent();
334 I->eraseFromParent();
338 /// \brief Rewrite as many loads as possible given a single store.
340 /// When there is only a single store, we can use the domtree to trivially
341 /// replace all of the dominated loads with the stored value. Do so, and return
342 /// true if this has successfully promoted the alloca entirely. If this returns
343 /// false there were some loads which were not dominated by the single store
344 /// and thus must be phi-ed with undef. We fall back to the standard alloca
345 /// promotion algorithm in that case.
346 static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
349 AliasSetTracker *AST) {
350 StoreInst *OnlyStore = Info.OnlyStore;
351 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
352 BasicBlock *StoreBB = OnlyStore->getParent();
355 // Clear out UsingBlocks. We will reconstruct it here if needed.
356 Info.UsingBlocks.clear();
358 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
359 Instruction *UserInst = cast<Instruction>(*UI++);
360 if (!isa<LoadInst>(UserInst)) {
361 assert(UserInst == OnlyStore && "Should only have load/stores");
364 LoadInst *LI = cast<LoadInst>(UserInst);
366 // Okay, if we have a load from the alloca, we want to replace it with the
367 // only value stored to the alloca. We can do this if the value is
368 // dominated by the store. If not, we use the rest of the mem2reg machinery
369 // to insert the phi nodes as needed.
370 if (!StoringGlobalVal) { // Non-instructions are always dominated.
371 if (LI->getParent() == StoreBB) {
372 // If we have a use that is in the same block as the store, compare the
373 // indices of the two instructions to see which one came first. If the
374 // load came before the store, we can't handle it.
375 if (StoreIndex == -1)
376 StoreIndex = LBI.getInstructionIndex(OnlyStore);
378 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
379 // Can't handle this load, bail out.
380 Info.UsingBlocks.push_back(StoreBB);
384 } else if (LI->getParent() != StoreBB &&
385 !DT.dominates(StoreBB, LI->getParent())) {
386 // If the load and store are in different blocks, use BB dominance to
387 // check their relationships. If the store doesn't dom the use, bail
389 Info.UsingBlocks.push_back(LI->getParent());
394 // Otherwise, we *can* safely rewrite this load.
395 Value *ReplVal = OnlyStore->getOperand(0);
396 // If the replacement value is the load, this must occur in unreachable
399 ReplVal = UndefValue::get(LI->getType());
400 LI->replaceAllUsesWith(ReplVal);
401 if (AST && LI->getType()->isPointerTy())
402 AST->deleteValue(LI);
403 LI->eraseFromParent();
407 // Finally, after the scan, check to see if the store is all that is left.
408 if (!Info.UsingBlocks.empty())
409 return false; // If not, we'll have to fall back for the remainder.
411 // Record debuginfo for the store and remove the declaration's
413 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
414 DIBuilder DIB(*AI->getParent()->getParent()->getParent());
415 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB);
416 DDI->eraseFromParent();
417 LBI.deleteValue(DDI);
419 // Remove the (now dead) store and alloca.
420 Info.OnlyStore->eraseFromParent();
421 LBI.deleteValue(Info.OnlyStore);
424 AST->deleteValue(AI);
425 AI->eraseFromParent();
430 /// Many allocas are only used within a single basic block. If this is the
431 /// case, avoid traversing the CFG and inserting a lot of potentially useless
432 /// PHI nodes by just performing a single linear pass over the basic block
433 /// using the Alloca.
435 /// If we cannot promote this alloca (because it is read before it is written),
436 /// return true. This is necessary in cases where, due to control flow, the
437 /// alloca is potentially undefined on some control flow paths. e.g. code like
438 /// this is potentially correct:
440 /// for (...) { if (c) { A = undef; undef = B; } }
442 /// ... so long as A is not used before undef is set.
443 static void promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
445 AliasSetTracker *AST) {
446 // The trickiest case to handle is when we have large blocks. Because of this,
447 // this code is optimized assuming that large blocks happen. This does not
448 // significantly pessimize the small block case. This uses LargeBlockInfo to
449 // make it efficient to get the index of various operations in the block.
451 // Walk the use-def list of the alloca, getting the locations of all stores.
452 typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy;
453 StoresByIndexTy StoresByIndex;
455 for (User *U : AI->users())
456 if (StoreInst *SI = dyn_cast<StoreInst>(U))
457 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
459 // Sort the stores by their index, making it efficient to do a lookup with a
461 std::sort(StoresByIndex.begin(), StoresByIndex.end(), less_first());
463 // Walk all of the loads from this alloca, replacing them with the nearest
464 // store above them, if any.
465 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
466 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
470 unsigned LoadIdx = LBI.getInstructionIndex(LI);
472 // Find the nearest store that has a lower index than this load.
473 StoresByIndexTy::iterator I =
474 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
475 std::make_pair(LoadIdx, static_cast<StoreInst *>(0)),
478 if (I == StoresByIndex.begin())
479 // If there is no store before this load, the load takes the undef value.
480 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
482 // Otherwise, there was a store before this load, the load takes its value.
483 LI->replaceAllUsesWith(std::prev(I)->second->getOperand(0));
485 if (AST && LI->getType()->isPointerTy())
486 AST->deleteValue(LI);
487 LI->eraseFromParent();
491 // Remove the (now dead) stores and alloca.
492 while (!AI->use_empty()) {
493 StoreInst *SI = cast<StoreInst>(AI->user_back());
494 // Record debuginfo for the store before removing it.
495 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
496 DIBuilder DIB(*AI->getParent()->getParent()->getParent());
497 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
499 SI->eraseFromParent();
504 AST->deleteValue(AI);
505 AI->eraseFromParent();
508 // The alloca's debuginfo can be removed as well.
509 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
510 DDI->eraseFromParent();
511 LBI.deleteValue(DDI);
517 void PromoteMem2Reg::run() {
518 Function &F = *DT.getRoot()->getParent();
521 PointerAllocaValues.resize(Allocas.size());
522 AllocaDbgDeclares.resize(Allocas.size());
527 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
528 AllocaInst *AI = Allocas[AllocaNum];
530 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
531 assert(AI->getParent()->getParent() == &F &&
532 "All allocas should be in the same function, which is same as DF!");
534 removeLifetimeIntrinsicUsers(AI);
536 if (AI->use_empty()) {
537 // If there are no uses of the alloca, just delete it now.
539 AST->deleteValue(AI);
540 AI->eraseFromParent();
542 // Remove the alloca from the Allocas list, since it has been processed
543 RemoveFromAllocasList(AllocaNum);
548 // Calculate the set of read and write-locations for each alloca. This is
549 // analogous to finding the 'uses' and 'definitions' of each variable.
550 Info.AnalyzeAlloca(AI);
552 // If there is only a single store to this value, replace any loads of
553 // it that are directly dominated by the definition with the value stored.
554 if (Info.DefiningBlocks.size() == 1) {
555 if (rewriteSingleStoreAlloca(AI, Info, LBI, DT, AST)) {
556 // The alloca has been processed, move on.
557 RemoveFromAllocasList(AllocaNum);
563 // If the alloca is only read and written in one basic block, just perform a
564 // linear sweep over the block to eliminate it.
565 if (Info.OnlyUsedInOneBlock) {
566 promoteSingleBlockAlloca(AI, Info, LBI, AST);
568 // The alloca has been processed, move on.
569 RemoveFromAllocasList(AllocaNum);
573 // If we haven't computed dominator tree levels, do so now.
574 if (DomLevels.empty()) {
575 SmallVector<DomTreeNode *, 32> Worklist;
577 DomTreeNode *Root = DT.getRootNode();
579 Worklist.push_back(Root);
581 while (!Worklist.empty()) {
582 DomTreeNode *Node = Worklist.pop_back_val();
583 unsigned ChildLevel = DomLevels[Node] + 1;
584 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
586 DomLevels[*CI] = ChildLevel;
587 Worklist.push_back(*CI);
592 // If we haven't computed a numbering for the BB's in the function, do so
594 if (BBNumbers.empty()) {
596 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
600 // If we have an AST to keep updated, remember some pointer value that is
601 // stored into the alloca.
603 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
605 // Remember the dbg.declare intrinsic describing this alloca, if any.
607 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
609 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
610 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
612 // At this point, we're committed to promoting the alloca using IDF's, and
613 // the standard SSA construction algorithm. Determine which blocks need PHI
614 // nodes and see if we can optimize out some work by avoiding insertion of
616 DetermineInsertionPoint(AI, AllocaNum, Info);
620 return; // All of the allocas must have been trivial!
624 // Set the incoming values for the basic block to be null values for all of
625 // the alloca's. We do this in case there is a load of a value that has not
626 // been stored yet. In this case, it will get this null value.
628 RenamePassData::ValVector Values(Allocas.size());
629 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
630 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
632 // Walks all basic blocks in the function performing the SSA rename algorithm
633 // and inserting the phi nodes we marked as necessary
635 std::vector<RenamePassData> RenamePassWorkList;
636 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
639 RPD.swap(RenamePassWorkList.back());
640 RenamePassWorkList.pop_back();
641 // RenamePass may add new worklist entries.
642 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
643 } while (!RenamePassWorkList.empty());
645 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
648 // Remove the allocas themselves from the function.
649 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
650 Instruction *A = Allocas[i];
652 // If there are any uses of the alloca instructions left, they must be in
653 // unreachable basic blocks that were not processed by walking the dominator
654 // tree. Just delete the users now.
656 A->replaceAllUsesWith(UndefValue::get(A->getType()));
659 A->eraseFromParent();
662 // Remove alloca's dbg.declare instrinsics from the function.
663 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
664 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
665 DDI->eraseFromParent();
667 // Loop over all of the PHI nodes and see if there are any that we can get
668 // rid of because they merge all of the same incoming values. This can
669 // happen due to undef values coming into the PHI nodes. This process is
670 // iterative, because eliminating one PHI node can cause others to be removed.
671 bool EliminatedAPHI = true;
672 while (EliminatedAPHI) {
673 EliminatedAPHI = false;
675 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
676 // simplify and RAUW them as we go. If it was not, we could add uses to
677 // the values we replace with in a non-deterministic order, thus creating
678 // non-deterministic def->use chains.
679 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
680 I = NewPhiNodes.begin(),
681 E = NewPhiNodes.end();
683 PHINode *PN = I->second;
685 // If this PHI node merges one value and/or undefs, get the value.
686 if (Value *V = SimplifyInstruction(PN, 0, 0, &DT)) {
687 if (AST && PN->getType()->isPointerTy())
688 AST->deleteValue(PN);
689 PN->replaceAllUsesWith(V);
690 PN->eraseFromParent();
691 NewPhiNodes.erase(I++);
692 EliminatedAPHI = true;
699 // At this point, the renamer has added entries to PHI nodes for all reachable
700 // code. Unfortunately, there may be unreachable blocks which the renamer
701 // hasn't traversed. If this is the case, the PHI nodes may not
702 // have incoming values for all predecessors. Loop over all PHI nodes we have
703 // created, inserting undef values if they are missing any incoming values.
705 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
706 I = NewPhiNodes.begin(),
707 E = NewPhiNodes.end();
709 // We want to do this once per basic block. As such, only process a block
710 // when we find the PHI that is the first entry in the block.
711 PHINode *SomePHI = I->second;
712 BasicBlock *BB = SomePHI->getParent();
713 if (&BB->front() != SomePHI)
716 // Only do work here if there the PHI nodes are missing incoming values. We
717 // know that all PHI nodes that were inserted in a block will have the same
718 // number of incoming values, so we can just check any of them.
719 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
722 // Get the preds for BB.
723 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
725 // Ok, now we know that all of the PHI nodes are missing entries for some
726 // basic blocks. Start by sorting the incoming predecessors for efficient
728 std::sort(Preds.begin(), Preds.end());
730 // Now we loop through all BB's which have entries in SomePHI and remove
731 // them from the Preds list.
732 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
733 // Do a log(n) search of the Preds list for the entry we want.
734 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
735 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
736 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
737 "PHI node has entry for a block which is not a predecessor!");
743 // At this point, the blocks left in the preds list must have dummy
744 // entries inserted into every PHI nodes for the block. Update all the phi
745 // nodes in this block that we are inserting (there could be phis before
747 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
748 BasicBlock::iterator BBI = BB->begin();
749 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
750 SomePHI->getNumIncomingValues() == NumBadPreds) {
751 Value *UndefVal = UndefValue::get(SomePHI->getType());
752 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
753 SomePHI->addIncoming(UndefVal, Preds[pred]);
760 /// \brief Determine which blocks the value is live in.
762 /// These are blocks which lead to uses. Knowing this allows us to avoid
763 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
764 /// inserted phi nodes would be dead).
765 void PromoteMem2Reg::ComputeLiveInBlocks(
766 AllocaInst *AI, AllocaInfo &Info,
767 const SmallPtrSet<BasicBlock *, 32> &DefBlocks,
768 SmallPtrSet<BasicBlock *, 32> &LiveInBlocks) {
770 // To determine liveness, we must iterate through the predecessors of blocks
771 // where the def is live. Blocks are added to the worklist if we need to
772 // check their predecessors. Start with all the using blocks.
773 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
774 Info.UsingBlocks.end());
776 // If any of the using blocks is also a definition block, check to see if the
777 // definition occurs before or after the use. If it happens before the use,
778 // the value isn't really live-in.
779 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
780 BasicBlock *BB = LiveInBlockWorklist[i];
781 if (!DefBlocks.count(BB))
784 // Okay, this is a block that both uses and defines the value. If the first
785 // reference to the alloca is a def (store), then we know it isn't live-in.
786 for (BasicBlock::iterator I = BB->begin();; ++I) {
787 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
788 if (SI->getOperand(1) != AI)
791 // We found a store to the alloca before a load. The alloca is not
792 // actually live-in here.
793 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
794 LiveInBlockWorklist.pop_back();
799 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
800 if (LI->getOperand(0) != AI)
803 // Okay, we found a load before a store to the alloca. It is actually
804 // live into this block.
810 // Now that we have a set of blocks where the phi is live-in, recursively add
811 // their predecessors until we find the full region the value is live.
812 while (!LiveInBlockWorklist.empty()) {
813 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
815 // The block really is live in here, insert it into the set. If already in
816 // the set, then it has already been processed.
817 if (!LiveInBlocks.insert(BB))
820 // Since the value is live into BB, it is either defined in a predecessor or
821 // live into it to. Add the preds to the worklist unless they are a
823 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
826 // The value is not live into a predecessor if it defines the value.
827 if (DefBlocks.count(P))
830 // Otherwise it is, add to the worklist.
831 LiveInBlockWorklist.push_back(P);
836 /// At this point, we're committed to promoting the alloca using IDF's, and the
837 /// standard SSA construction algorithm. Determine which blocks need phi nodes
838 /// and see if we can optimize out some work by avoiding insertion of dead phi
840 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
842 // Unique the set of defining blocks for efficient lookup.
843 SmallPtrSet<BasicBlock *, 32> DefBlocks;
844 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
846 // Determine which blocks the value is live in. These are blocks which lead
848 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
849 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
851 // Use a priority queue keyed on dominator tree level so that inserted nodes
852 // are handled from the bottom of the dominator tree upwards.
853 typedef std::pair<DomTreeNode *, unsigned> DomTreeNodePair;
854 typedef std::priority_queue<DomTreeNodePair, SmallVector<DomTreeNodePair, 32>,
855 less_second> IDFPriorityQueue;
858 for (SmallPtrSet<BasicBlock *, 32>::const_iterator I = DefBlocks.begin(),
861 if (DomTreeNode *Node = DT.getNode(*I))
862 PQ.push(std::make_pair(Node, DomLevels[Node]));
865 SmallVector<std::pair<unsigned, BasicBlock *>, 32> DFBlocks;
866 SmallPtrSet<DomTreeNode *, 32> Visited;
867 SmallVector<DomTreeNode *, 32> Worklist;
868 while (!PQ.empty()) {
869 DomTreeNodePair RootPair = PQ.top();
871 DomTreeNode *Root = RootPair.first;
872 unsigned RootLevel = RootPair.second;
874 // Walk all dominator tree children of Root, inspecting their CFG edges with
875 // targets elsewhere on the dominator tree. Only targets whose level is at
876 // most Root's level are added to the iterated dominance frontier of the
880 Worklist.push_back(Root);
882 while (!Worklist.empty()) {
883 DomTreeNode *Node = Worklist.pop_back_val();
884 BasicBlock *BB = Node->getBlock();
886 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
888 DomTreeNode *SuccNode = DT.getNode(*SI);
890 // Quickly skip all CFG edges that are also dominator tree edges instead
891 // of catching them below.
892 if (SuccNode->getIDom() == Node)
895 unsigned SuccLevel = DomLevels[SuccNode];
896 if (SuccLevel > RootLevel)
899 if (!Visited.insert(SuccNode))
902 BasicBlock *SuccBB = SuccNode->getBlock();
903 if (!LiveInBlocks.count(SuccBB))
906 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
907 if (!DefBlocks.count(SuccBB))
908 PQ.push(std::make_pair(SuccNode, SuccLevel));
911 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
913 if (!Visited.count(*CI))
914 Worklist.push_back(*CI);
919 if (DFBlocks.size() > 1)
920 std::sort(DFBlocks.begin(), DFBlocks.end());
922 unsigned CurrentVersion = 0;
923 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
924 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
927 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
929 /// Returns true if there wasn't already a phi-node for that variable
930 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
932 // Look up the basic-block in question.
933 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
935 // If the BB already has a phi node added for the i'th alloca then we're done!
939 // Create a PhiNode using the dereferenced type... and add the phi-node to the
941 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
942 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
945 PhiToAllocaMap[PN] = AllocaNo;
947 if (AST && PN->getType()->isPointerTy())
948 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
953 /// \brief Recursively traverse the CFG of the function, renaming loads and
954 /// stores to the allocas which we are promoting.
956 /// IncomingVals indicates what value each Alloca contains on exit from the
957 /// predecessor block Pred.
958 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
959 RenamePassData::ValVector &IncomingVals,
960 std::vector<RenamePassData> &Worklist) {
962 // If we are inserting any phi nodes into this BB, they will already be in the
964 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
965 // If we have PHI nodes to update, compute the number of edges from Pred to
967 if (PhiToAllocaMap.count(APN)) {
968 // We want to be able to distinguish between PHI nodes being inserted by
969 // this invocation of mem2reg from those phi nodes that already existed in
970 // the IR before mem2reg was run. We determine that APN is being inserted
971 // because it is missing incoming edges. All other PHI nodes being
972 // inserted by this pass of mem2reg will have the same number of incoming
973 // operands so far. Remember this count.
974 unsigned NewPHINumOperands = APN->getNumOperands();
976 unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB);
977 assert(NumEdges && "Must be at least one edge from Pred to BB!");
979 // Add entries for all the phis.
980 BasicBlock::iterator PNI = BB->begin();
982 unsigned AllocaNo = PhiToAllocaMap[APN];
984 // Add N incoming values to the PHI node.
985 for (unsigned i = 0; i != NumEdges; ++i)
986 APN->addIncoming(IncomingVals[AllocaNo], Pred);
988 // The currently active variable for this block is now the PHI.
989 IncomingVals[AllocaNo] = APN;
991 // Get the next phi node.
993 APN = dyn_cast<PHINode>(PNI);
997 // Verify that it is missing entries. If not, it is not being inserted
998 // by this mem2reg invocation so we want to ignore it.
999 } while (APN->getNumOperands() == NewPHINumOperands);
1003 // Don't revisit blocks.
1004 if (!Visited.insert(BB))
1007 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) {
1008 Instruction *I = II++; // get the instruction, increment iterator
1010 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1011 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1015 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
1016 if (AI == AllocaLookup.end())
1019 Value *V = IncomingVals[AI->second];
1021 // Anything using the load now uses the current value.
1022 LI->replaceAllUsesWith(V);
1023 if (AST && LI->getType()->isPointerTy())
1024 AST->deleteValue(LI);
1025 BB->getInstList().erase(LI);
1026 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1027 // Delete this instruction and mark the name as the current holder of the
1029 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1033 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1034 if (ai == AllocaLookup.end())
1037 // what value were we writing?
1038 IncomingVals[ai->second] = SI->getOperand(0);
1039 // Record debuginfo for the store before removing it.
1040 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1041 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1042 BB->getInstList().erase(SI);
1046 // 'Recurse' to our successors.
1047 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1051 // Keep track of the successors so we don't visit the same successor twice
1052 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
1054 // Handle the first successor without using the worklist.
1055 VisitedSuccs.insert(*I);
1061 if (VisitedSuccs.insert(*I))
1062 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1067 void llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
1068 AliasSetTracker *AST) {
1069 // If there is nothing to do, bail out...
1070 if (Allocas.empty())
1073 PromoteMem2Reg(Allocas, DT, AST).run();