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/IR/Module.h"
49 #include "llvm/Transforms/Utils/Local.h"
54 #define DEBUG_TYPE "mem2reg"
56 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
57 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
58 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
59 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
61 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
62 // FIXME: If the memory unit is of pointer or integer type, we can permit
63 // assignments to subsections of the memory unit.
64 unsigned AS = AI->getType()->getAddressSpace();
66 // Only allow direct and non-volatile loads and stores...
67 for (const User *U : AI->users()) {
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(), AS))
87 if (!onlyUsedByLifetimeMarkers(BCI))
89 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
90 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext(), AS))
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 = nullptr;
124 DbgDeclare = nullptr;
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 (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
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(nullptr), Pred(nullptr), 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 /// A cache of @llvm.assume intrinsics used by SimplifyInstruction.
245 /// Reverse mapping of Allocas.
246 DenseMap<AllocaInst *, unsigned> AllocaLookup;
248 /// \brief The PhiNodes we're adding.
250 /// That map is used to simplify some Phi nodes as we iterate over it, so
251 /// it should have deterministic iterators. We could use a MapVector, but
252 /// since we already maintain a map from BasicBlock* to a stable numbering
253 /// (BBNumbers), the DenseMap is more efficient (also supports removal).
254 DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes;
256 /// For each PHI node, keep track of which entry in Allocas it corresponds
258 DenseMap<PHINode *, unsigned> PhiToAllocaMap;
260 /// If we are updating an AliasSetTracker, then for each alloca that is of
261 /// pointer type, we keep track of what to copyValue to the inserted PHI
263 std::vector<Value *> PointerAllocaValues;
265 /// For each alloca, we keep track of the dbg.declare intrinsic that
266 /// describes it, if any, so that we can convert it to a dbg.value
267 /// intrinsic if the alloca gets promoted.
268 SmallVector<DbgDeclareInst *, 8> AllocaDbgDeclares;
270 /// The set of basic blocks the renamer has already visited.
272 SmallPtrSet<BasicBlock *, 16> Visited;
274 /// Contains a stable numbering of basic blocks to avoid non-determinstic
276 DenseMap<BasicBlock *, unsigned> BBNumbers;
278 /// Maps DomTreeNodes to their level in the dominator tree.
279 DenseMap<DomTreeNode *, unsigned> DomLevels;
281 /// Lazily compute the number of predecessors a block has.
282 DenseMap<const BasicBlock *, unsigned> BBNumPreds;
285 PromoteMem2Reg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
286 AliasSetTracker *AST, AssumptionCache *AC)
287 : Allocas(Allocas.begin(), Allocas.end()), DT(DT),
288 DIB(*DT.getRoot()->getParent()->getParent(), /*AllowUnresolved*/ false),
294 void RemoveFromAllocasList(unsigned &AllocaIdx) {
295 Allocas[AllocaIdx] = Allocas.back();
300 unsigned getNumPreds(const BasicBlock *BB) {
301 unsigned &NP = BBNumPreds[BB];
303 NP = std::distance(pred_begin(BB), pred_end(BB)) + 1;
307 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
309 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
310 const SmallPtrSetImpl<BasicBlock *> &DefBlocks,
311 SmallPtrSetImpl<BasicBlock *> &LiveInBlocks);
312 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
313 RenamePassData::ValVector &IncVals,
314 std::vector<RenamePassData> &Worklist);
315 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
318 } // end of anonymous namespace
320 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
321 // Knowing that this alloca is promotable, we know that it's safe to kill all
322 // instructions except for load and store.
324 for (auto UI = AI->user_begin(), UE = AI->user_end(); UI != UE;) {
325 Instruction *I = cast<Instruction>(*UI);
327 if (isa<LoadInst>(I) || isa<StoreInst>(I))
330 if (!I->getType()->isVoidTy()) {
331 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
332 // Follow the use/def chain to erase them now instead of leaving it for
333 // dead code elimination later.
334 for (auto UUI = I->user_begin(), UUE = I->user_end(); UUI != UUE;) {
335 Instruction *Inst = cast<Instruction>(*UUI);
337 Inst->eraseFromParent();
340 I->eraseFromParent();
344 /// \brief Rewrite as many loads as possible given a single store.
346 /// When there is only a single store, we can use the domtree to trivially
347 /// replace all of the dominated loads with the stored value. Do so, and return
348 /// true if this has successfully promoted the alloca entirely. If this returns
349 /// false there were some loads which were not dominated by the single store
350 /// and thus must be phi-ed with undef. We fall back to the standard alloca
351 /// promotion algorithm in that case.
352 static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
355 AliasSetTracker *AST) {
356 StoreInst *OnlyStore = Info.OnlyStore;
357 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
358 BasicBlock *StoreBB = OnlyStore->getParent();
361 // Clear out UsingBlocks. We will reconstruct it here if needed.
362 Info.UsingBlocks.clear();
364 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
365 Instruction *UserInst = cast<Instruction>(*UI++);
366 if (!isa<LoadInst>(UserInst)) {
367 assert(UserInst == OnlyStore && "Should only have load/stores");
370 LoadInst *LI = cast<LoadInst>(UserInst);
372 // Okay, if we have a load from the alloca, we want to replace it with the
373 // only value stored to the alloca. We can do this if the value is
374 // dominated by the store. If not, we use the rest of the mem2reg machinery
375 // to insert the phi nodes as needed.
376 if (!StoringGlobalVal) { // Non-instructions are always dominated.
377 if (LI->getParent() == StoreBB) {
378 // If we have a use that is in the same block as the store, compare the
379 // indices of the two instructions to see which one came first. If the
380 // load came before the store, we can't handle it.
381 if (StoreIndex == -1)
382 StoreIndex = LBI.getInstructionIndex(OnlyStore);
384 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
385 // Can't handle this load, bail out.
386 Info.UsingBlocks.push_back(StoreBB);
390 } else if (LI->getParent() != StoreBB &&
391 !DT.dominates(StoreBB, LI->getParent())) {
392 // If the load and store are in different blocks, use BB dominance to
393 // check their relationships. If the store doesn't dom the use, bail
395 Info.UsingBlocks.push_back(LI->getParent());
400 // Otherwise, we *can* safely rewrite this load.
401 Value *ReplVal = OnlyStore->getOperand(0);
402 // If the replacement value is the load, this must occur in unreachable
405 ReplVal = UndefValue::get(LI->getType());
406 LI->replaceAllUsesWith(ReplVal);
407 if (AST && LI->getType()->isPointerTy())
408 AST->deleteValue(LI);
409 LI->eraseFromParent();
413 // Finally, after the scan, check to see if the store is all that is left.
414 if (!Info.UsingBlocks.empty())
415 return false; // If not, we'll have to fall back for the remainder.
417 // Record debuginfo for the store and remove the declaration's
419 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
420 DIBuilder DIB(*AI->getParent()->getParent()->getParent(),
421 /*AllowUnresolved*/ false);
422 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB);
423 DDI->eraseFromParent();
424 LBI.deleteValue(DDI);
426 // Remove the (now dead) store and alloca.
427 Info.OnlyStore->eraseFromParent();
428 LBI.deleteValue(Info.OnlyStore);
431 AST->deleteValue(AI);
432 AI->eraseFromParent();
437 /// Many allocas are only used within a single basic block. If this is the
438 /// case, avoid traversing the CFG and inserting a lot of potentially useless
439 /// PHI nodes by just performing a single linear pass over the basic block
440 /// using the Alloca.
442 /// If we cannot promote this alloca (because it is read before it is written),
443 /// return true. This is necessary in cases where, due to control flow, the
444 /// alloca is potentially undefined on some control flow paths. e.g. code like
445 /// this is potentially correct:
447 /// for (...) { if (c) { A = undef; undef = B; } }
449 /// ... so long as A is not used before undef is set.
450 static void promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
452 AliasSetTracker *AST) {
453 // The trickiest case to handle is when we have large blocks. Because of this,
454 // this code is optimized assuming that large blocks happen. This does not
455 // significantly pessimize the small block case. This uses LargeBlockInfo to
456 // make it efficient to get the index of various operations in the block.
458 // Walk the use-def list of the alloca, getting the locations of all stores.
459 typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy;
460 StoresByIndexTy StoresByIndex;
462 for (User *U : AI->users())
463 if (StoreInst *SI = dyn_cast<StoreInst>(U))
464 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
466 // Sort the stores by their index, making it efficient to do a lookup with a
468 std::sort(StoresByIndex.begin(), StoresByIndex.end(), less_first());
470 // Walk all of the loads from this alloca, replacing them with the nearest
471 // store above them, if any.
472 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
473 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
477 unsigned LoadIdx = LBI.getInstructionIndex(LI);
479 // Find the nearest store that has a lower index than this load.
480 StoresByIndexTy::iterator I =
481 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
482 std::make_pair(LoadIdx,
483 static_cast<StoreInst *>(nullptr)),
486 if (I == StoresByIndex.begin())
487 // If there is no store before this load, the load takes the undef value.
488 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
490 // Otherwise, there was a store before this load, the load takes its value.
491 LI->replaceAllUsesWith(std::prev(I)->second->getOperand(0));
493 if (AST && LI->getType()->isPointerTy())
494 AST->deleteValue(LI);
495 LI->eraseFromParent();
499 // Remove the (now dead) stores and alloca.
500 while (!AI->use_empty()) {
501 StoreInst *SI = cast<StoreInst>(AI->user_back());
502 // Record debuginfo for the store before removing it.
503 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
504 DIBuilder DIB(*AI->getParent()->getParent()->getParent(),
505 /*AllowUnresolved*/ false);
506 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
508 SI->eraseFromParent();
513 AST->deleteValue(AI);
514 AI->eraseFromParent();
517 // The alloca's debuginfo can be removed as well.
518 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
519 DDI->eraseFromParent();
520 LBI.deleteValue(DDI);
526 void PromoteMem2Reg::run() {
527 Function &F = *DT.getRoot()->getParent();
530 PointerAllocaValues.resize(Allocas.size());
531 AllocaDbgDeclares.resize(Allocas.size());
536 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
537 AllocaInst *AI = Allocas[AllocaNum];
539 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
540 assert(AI->getParent()->getParent() == &F &&
541 "All allocas should be in the same function, which is same as DF!");
543 removeLifetimeIntrinsicUsers(AI);
545 if (AI->use_empty()) {
546 // If there are no uses of the alloca, just delete it now.
548 AST->deleteValue(AI);
549 AI->eraseFromParent();
551 // Remove the alloca from the Allocas list, since it has been processed
552 RemoveFromAllocasList(AllocaNum);
557 // Calculate the set of read and write-locations for each alloca. This is
558 // analogous to finding the 'uses' and 'definitions' of each variable.
559 Info.AnalyzeAlloca(AI);
561 // If there is only a single store to this value, replace any loads of
562 // it that are directly dominated by the definition with the value stored.
563 if (Info.DefiningBlocks.size() == 1) {
564 if (rewriteSingleStoreAlloca(AI, Info, LBI, DT, AST)) {
565 // The alloca has been processed, move on.
566 RemoveFromAllocasList(AllocaNum);
572 // If the alloca is only read and written in one basic block, just perform a
573 // linear sweep over the block to eliminate it.
574 if (Info.OnlyUsedInOneBlock) {
575 promoteSingleBlockAlloca(AI, Info, LBI, AST);
577 // The alloca has been processed, move on.
578 RemoveFromAllocasList(AllocaNum);
582 // If we haven't computed dominator tree levels, do so now.
583 if (DomLevels.empty()) {
584 SmallVector<DomTreeNode *, 32> Worklist;
586 DomTreeNode *Root = DT.getRootNode();
588 Worklist.push_back(Root);
590 while (!Worklist.empty()) {
591 DomTreeNode *Node = Worklist.pop_back_val();
592 unsigned ChildLevel = DomLevels[Node] + 1;
593 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
595 DomLevels[*CI] = ChildLevel;
596 Worklist.push_back(*CI);
601 // If we haven't computed a numbering for the BB's in the function, do so
603 if (BBNumbers.empty()) {
605 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
609 // If we have an AST to keep updated, remember some pointer value that is
610 // stored into the alloca.
612 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
614 // Remember the dbg.declare intrinsic describing this alloca, if any.
616 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
618 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
619 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
621 // At this point, we're committed to promoting the alloca using IDF's, and
622 // the standard SSA construction algorithm. Determine which blocks need PHI
623 // nodes and see if we can optimize out some work by avoiding insertion of
625 DetermineInsertionPoint(AI, AllocaNum, Info);
629 return; // All of the allocas must have been trivial!
633 // Set the incoming values for the basic block to be null values for all of
634 // the alloca's. We do this in case there is a load of a value that has not
635 // been stored yet. In this case, it will get this null value.
637 RenamePassData::ValVector Values(Allocas.size());
638 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
639 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
641 // Walks all basic blocks in the function performing the SSA rename algorithm
642 // and inserting the phi nodes we marked as necessary
644 std::vector<RenamePassData> RenamePassWorkList;
645 RenamePassWorkList.push_back(RenamePassData(F.begin(), nullptr, Values));
648 RPD.swap(RenamePassWorkList.back());
649 RenamePassWorkList.pop_back();
650 // RenamePass may add new worklist entries.
651 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
652 } while (!RenamePassWorkList.empty());
654 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
657 // Remove the allocas themselves from the function.
658 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
659 Instruction *A = Allocas[i];
661 // If there are any uses of the alloca instructions left, they must be in
662 // unreachable basic blocks that were not processed by walking the dominator
663 // tree. Just delete the users now.
665 A->replaceAllUsesWith(UndefValue::get(A->getType()));
668 A->eraseFromParent();
671 const DataLayout &DL = F.getParent()->getDataLayout();
673 // Remove alloca's dbg.declare instrinsics from the function.
674 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
675 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
676 DDI->eraseFromParent();
678 // Loop over all of the PHI nodes and see if there are any that we can get
679 // rid of because they merge all of the same incoming values. This can
680 // happen due to undef values coming into the PHI nodes. This process is
681 // iterative, because eliminating one PHI node can cause others to be removed.
682 bool EliminatedAPHI = true;
683 while (EliminatedAPHI) {
684 EliminatedAPHI = false;
686 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
687 // simplify and RAUW them as we go. If it was not, we could add uses to
688 // the values we replace with in a non-deterministic order, thus creating
689 // non-deterministic def->use chains.
690 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
691 I = NewPhiNodes.begin(),
692 E = NewPhiNodes.end();
694 PHINode *PN = I->second;
696 // If this PHI node merges one value and/or undefs, get the value.
697 if (Value *V = SimplifyInstruction(PN, DL, nullptr, &DT, AC)) {
698 if (AST && PN->getType()->isPointerTy())
699 AST->deleteValue(PN);
700 PN->replaceAllUsesWith(V);
701 PN->eraseFromParent();
702 NewPhiNodes.erase(I++);
703 EliminatedAPHI = true;
710 // At this point, the renamer has added entries to PHI nodes for all reachable
711 // code. Unfortunately, there may be unreachable blocks which the renamer
712 // hasn't traversed. If this is the case, the PHI nodes may not
713 // have incoming values for all predecessors. Loop over all PHI nodes we have
714 // created, inserting undef values if they are missing any incoming values.
716 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
717 I = NewPhiNodes.begin(),
718 E = NewPhiNodes.end();
720 // We want to do this once per basic block. As such, only process a block
721 // when we find the PHI that is the first entry in the block.
722 PHINode *SomePHI = I->second;
723 BasicBlock *BB = SomePHI->getParent();
724 if (&BB->front() != SomePHI)
727 // Only do work here if there the PHI nodes are missing incoming values. We
728 // know that all PHI nodes that were inserted in a block will have the same
729 // number of incoming values, so we can just check any of them.
730 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
733 // Get the preds for BB.
734 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
736 // Ok, now we know that all of the PHI nodes are missing entries for some
737 // basic blocks. Start by sorting the incoming predecessors for efficient
739 std::sort(Preds.begin(), Preds.end());
741 // Now we loop through all BB's which have entries in SomePHI and remove
742 // them from the Preds list.
743 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
744 // Do a log(n) search of the Preds list for the entry we want.
745 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
746 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
747 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
748 "PHI node has entry for a block which is not a predecessor!");
754 // At this point, the blocks left in the preds list must have dummy
755 // entries inserted into every PHI nodes for the block. Update all the phi
756 // nodes in this block that we are inserting (there could be phis before
758 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
759 BasicBlock::iterator BBI = BB->begin();
760 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
761 SomePHI->getNumIncomingValues() == NumBadPreds) {
762 Value *UndefVal = UndefValue::get(SomePHI->getType());
763 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
764 SomePHI->addIncoming(UndefVal, Preds[pred]);
771 /// \brief Determine which blocks the value is live in.
773 /// These are blocks which lead to uses. Knowing this allows us to avoid
774 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
775 /// inserted phi nodes would be dead).
776 void PromoteMem2Reg::ComputeLiveInBlocks(
777 AllocaInst *AI, AllocaInfo &Info,
778 const SmallPtrSetImpl<BasicBlock *> &DefBlocks,
779 SmallPtrSetImpl<BasicBlock *> &LiveInBlocks) {
781 // To determine liveness, we must iterate through the predecessors of blocks
782 // where the def is live. Blocks are added to the worklist if we need to
783 // check their predecessors. Start with all the using blocks.
784 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
785 Info.UsingBlocks.end());
787 // If any of the using blocks is also a definition block, check to see if the
788 // definition occurs before or after the use. If it happens before the use,
789 // the value isn't really live-in.
790 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
791 BasicBlock *BB = LiveInBlockWorklist[i];
792 if (!DefBlocks.count(BB))
795 // Okay, this is a block that both uses and defines the value. If the first
796 // reference to the alloca is a def (store), then we know it isn't live-in.
797 for (BasicBlock::iterator I = BB->begin();; ++I) {
798 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
799 if (SI->getOperand(1) != AI)
802 // We found a store to the alloca before a load. The alloca is not
803 // actually live-in here.
804 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
805 LiveInBlockWorklist.pop_back();
810 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
811 if (LI->getOperand(0) != AI)
814 // Okay, we found a load before a store to the alloca. It is actually
815 // live into this block.
821 // Now that we have a set of blocks where the phi is live-in, recursively add
822 // their predecessors until we find the full region the value is live.
823 while (!LiveInBlockWorklist.empty()) {
824 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
826 // The block really is live in here, insert it into the set. If already in
827 // the set, then it has already been processed.
828 if (!LiveInBlocks.insert(BB).second)
831 // Since the value is live into BB, it is either defined in a predecessor or
832 // live into it to. Add the preds to the worklist unless they are a
834 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
837 // The value is not live into a predecessor if it defines the value.
838 if (DefBlocks.count(P))
841 // Otherwise it is, add to the worklist.
842 LiveInBlockWorklist.push_back(P);
847 /// At this point, we're committed to promoting the alloca using IDF's, and the
848 /// standard SSA construction algorithm. Determine which blocks need phi nodes
849 /// and see if we can optimize out some work by avoiding insertion of dead phi
851 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
853 // Unique the set of defining blocks for efficient lookup.
854 SmallPtrSet<BasicBlock *, 32> DefBlocks;
855 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
857 // Determine which blocks the value is live in. These are blocks which lead
859 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
860 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
862 // Use a priority queue keyed on dominator tree level so that inserted nodes
863 // are handled from the bottom of the dominator tree upwards.
864 typedef std::pair<DomTreeNode *, unsigned> DomTreeNodePair;
865 typedef std::priority_queue<DomTreeNodePair, SmallVector<DomTreeNodePair, 32>,
866 less_second> IDFPriorityQueue;
869 for (BasicBlock *BB : DefBlocks) {
870 if (DomTreeNode *Node = DT.getNode(BB))
871 PQ.push(std::make_pair(Node, DomLevels[Node]));
874 SmallVector<std::pair<unsigned, BasicBlock *>, 32> DFBlocks;
875 SmallVector<DomTreeNode *, 32> Worklist;
876 SmallPtrSet<DomTreeNode *, 32> VisitedPQ;
877 SmallPtrSet<DomTreeNode *, 32> VisitedWorklist;
879 while (!PQ.empty()) {
880 DomTreeNodePair RootPair = PQ.top();
882 DomTreeNode *Root = RootPair.first;
883 unsigned RootLevel = RootPair.second;
885 // Walk all dominator tree children of Root, inspecting their CFG edges with
886 // targets elsewhere on the dominator tree. Only targets whose level is at
887 // most Root's level are added to the iterated dominance frontier of the
891 Worklist.push_back(Root);
892 VisitedWorklist.insert(Root);
894 while (!Worklist.empty()) {
895 DomTreeNode *Node = Worklist.pop_back_val();
896 BasicBlock *BB = Node->getBlock();
898 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
900 DomTreeNode *SuccNode = DT.getNode(*SI);
902 // Quickly skip all CFG edges that are also dominator tree edges instead
903 // of catching them below.
904 if (SuccNode->getIDom() == Node)
907 unsigned SuccLevel = DomLevels[SuccNode];
908 if (SuccLevel > RootLevel)
911 if (!VisitedPQ.insert(SuccNode).second)
914 BasicBlock *SuccBB = SuccNode->getBlock();
915 if (!LiveInBlocks.count(SuccBB))
918 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
919 if (!DefBlocks.count(SuccBB))
920 PQ.push(std::make_pair(SuccNode, SuccLevel));
923 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
925 if (VisitedWorklist.insert(*CI).second)
926 Worklist.push_back(*CI);
931 if (DFBlocks.size() > 1)
932 std::sort(DFBlocks.begin(), DFBlocks.end());
934 unsigned CurrentVersion = 0;
935 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
936 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
939 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
941 /// Returns true if there wasn't already a phi-node for that variable
942 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
944 // Look up the basic-block in question.
945 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
947 // If the BB already has a phi node added for the i'th alloca then we're done!
951 // Create a PhiNode using the dereferenced type... and add the phi-node to the
953 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
954 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
957 PhiToAllocaMap[PN] = AllocaNo;
959 if (AST && PN->getType()->isPointerTy())
960 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
965 /// \brief Recursively traverse the CFG of the function, renaming loads and
966 /// stores to the allocas which we are promoting.
968 /// IncomingVals indicates what value each Alloca contains on exit from the
969 /// predecessor block Pred.
970 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
971 RenamePassData::ValVector &IncomingVals,
972 std::vector<RenamePassData> &Worklist) {
974 // If we are inserting any phi nodes into this BB, they will already be in the
976 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
977 // If we have PHI nodes to update, compute the number of edges from Pred to
979 if (PhiToAllocaMap.count(APN)) {
980 // We want to be able to distinguish between PHI nodes being inserted by
981 // this invocation of mem2reg from those phi nodes that already existed in
982 // the IR before mem2reg was run. We determine that APN is being inserted
983 // because it is missing incoming edges. All other PHI nodes being
984 // inserted by this pass of mem2reg will have the same number of incoming
985 // operands so far. Remember this count.
986 unsigned NewPHINumOperands = APN->getNumOperands();
988 unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB);
989 assert(NumEdges && "Must be at least one edge from Pred to BB!");
991 // Add entries for all the phis.
992 BasicBlock::iterator PNI = BB->begin();
994 unsigned AllocaNo = PhiToAllocaMap[APN];
996 // Add N incoming values to the PHI node.
997 for (unsigned i = 0; i != NumEdges; ++i)
998 APN->addIncoming(IncomingVals[AllocaNo], Pred);
1000 // The currently active variable for this block is now the PHI.
1001 IncomingVals[AllocaNo] = APN;
1003 // Get the next phi node.
1005 APN = dyn_cast<PHINode>(PNI);
1009 // Verify that it is missing entries. If not, it is not being inserted
1010 // by this mem2reg invocation so we want to ignore it.
1011 } while (APN->getNumOperands() == NewPHINumOperands);
1015 // Don't revisit blocks.
1016 if (!Visited.insert(BB).second)
1019 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) {
1020 Instruction *I = II++; // get the instruction, increment iterator
1022 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1023 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1027 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
1028 if (AI == AllocaLookup.end())
1031 Value *V = IncomingVals[AI->second];
1033 // Anything using the load now uses the current value.
1034 LI->replaceAllUsesWith(V);
1035 if (AST && LI->getType()->isPointerTy())
1036 AST->deleteValue(LI);
1037 BB->getInstList().erase(LI);
1038 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1039 // Delete this instruction and mark the name as the current holder of the
1041 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1045 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1046 if (ai == AllocaLookup.end())
1049 // what value were we writing?
1050 IncomingVals[ai->second] = SI->getOperand(0);
1051 // Record debuginfo for the store before removing it.
1052 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1053 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1054 BB->getInstList().erase(SI);
1058 // 'Recurse' to our successors.
1059 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1063 // Keep track of the successors so we don't visit the same successor twice
1064 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
1066 // Handle the first successor without using the worklist.
1067 VisitedSuccs.insert(*I);
1073 if (VisitedSuccs.insert(*I).second)
1074 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1079 void llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
1080 AliasSetTracker *AST, AssumptionCache *AC) {
1081 // If there is nothing to do, bail out...
1082 if (Allocas.empty())
1085 PromoteMem2Reg(Allocas, DT, AST, AC).run();