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 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
19 #include "llvm/ADT/ArrayRef.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/AliasSetTracker.h"
26 #include "llvm/Analysis/InstructionSimplify.h"
27 #include "llvm/Analysis/IteratedDominanceFrontier.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/IR/CFG.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DIBuilder.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/Metadata.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/Transforms/Utils/Local.h"
44 #define DEBUG_TYPE "mem2reg"
46 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
47 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
48 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
49 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
51 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
52 // FIXME: If the memory unit is of pointer or integer type, we can permit
53 // assignments to subsections of the memory unit.
54 unsigned AS = AI->getType()->getAddressSpace();
56 // Only allow direct and non-volatile loads and stores...
57 for (const User *U : AI->users()) {
58 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
59 // Note that atomic loads can be transformed; atomic semantics do
60 // not have any meaning for a local alloca.
63 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
64 if (SI->getOperand(0) == AI)
65 return false; // Don't allow a store OF the AI, only INTO the AI.
66 // Note that atomic stores can be transformed; atomic semantics do
67 // not have any meaning for a local alloca.
70 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
71 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
72 II->getIntrinsicID() != Intrinsic::lifetime_end)
74 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
75 if (BCI->getType() != Type::getInt8PtrTy(U->getContext(), AS))
77 if (!onlyUsedByLifetimeMarkers(BCI))
79 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
80 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext(), AS))
82 if (!GEPI->hasAllZeroIndices())
84 if (!onlyUsedByLifetimeMarkers(GEPI))
97 SmallVector<BasicBlock *, 32> DefiningBlocks;
98 SmallVector<BasicBlock *, 32> UsingBlocks;
100 StoreInst *OnlyStore;
101 BasicBlock *OnlyBlock;
102 bool OnlyUsedInOneBlock;
104 Value *AllocaPointerVal;
105 DbgDeclareInst *DbgDeclare;
108 DefiningBlocks.clear();
112 OnlyUsedInOneBlock = true;
113 AllocaPointerVal = nullptr;
114 DbgDeclare = nullptr;
117 /// Scan the uses of the specified alloca, filling in the AllocaInfo used
118 /// by the rest of the pass to reason about the uses of this alloca.
119 void AnalyzeAlloca(AllocaInst *AI) {
122 // As we scan the uses of the alloca instruction, keep track of stores,
123 // and decide whether all of the loads and stores to the alloca are within
124 // the same basic block.
125 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
126 Instruction *User = cast<Instruction>(*UI++);
128 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
129 // Remember the basic blocks which define new values for the alloca
130 DefiningBlocks.push_back(SI->getParent());
131 AllocaPointerVal = SI->getOperand(0);
134 LoadInst *LI = cast<LoadInst>(User);
135 // Otherwise it must be a load instruction, keep track of variable
137 UsingBlocks.push_back(LI->getParent());
138 AllocaPointerVal = LI;
141 if (OnlyUsedInOneBlock) {
143 OnlyBlock = User->getParent();
144 else if (OnlyBlock != User->getParent())
145 OnlyUsedInOneBlock = false;
149 DbgDeclare = FindAllocaDbgDeclare(AI);
153 // Data package used by RenamePass()
154 class RenamePassData {
156 typedef std::vector<Value *> ValVector;
158 RenamePassData() : BB(nullptr), Pred(nullptr), Values() {}
159 RenamePassData(BasicBlock *B, BasicBlock *P, const ValVector &V)
160 : BB(B), Pred(P), Values(V) {}
165 void swap(RenamePassData &RHS) {
166 std::swap(BB, RHS.BB);
167 std::swap(Pred, RHS.Pred);
168 Values.swap(RHS.Values);
172 /// \brief This assigns and keeps a per-bb relative ordering of load/store
173 /// instructions in the block that directly load or store an alloca.
175 /// This functionality is important because it avoids scanning large basic
176 /// blocks multiple times when promoting many allocas in the same block.
177 class LargeBlockInfo {
178 /// \brief For each instruction that we track, keep the index of the
181 /// The index starts out as the number of the instruction from the start of
183 DenseMap<const Instruction *, unsigned> InstNumbers;
187 /// This code only looks at accesses to allocas.
188 static bool isInterestingInstruction(const Instruction *I) {
189 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
190 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
193 /// Get or calculate the index of the specified instruction.
194 unsigned getInstructionIndex(const Instruction *I) {
195 assert(isInterestingInstruction(I) &&
196 "Not a load/store to/from an alloca?");
198 // If we already have this instruction number, return it.
199 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
200 if (It != InstNumbers.end())
203 // Scan the whole block to get the instruction. This accumulates
204 // information for every interesting instruction in the block, in order to
205 // avoid gratuitus rescans.
206 const BasicBlock *BB = I->getParent();
208 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end(); BBI != E;
210 if (isInterestingInstruction(BBI))
211 InstNumbers[BBI] = InstNo++;
212 It = InstNumbers.find(I);
214 assert(It != InstNumbers.end() && "Didn't insert instruction?");
218 void deleteValue(const Instruction *I) { InstNumbers.erase(I); }
220 void clear() { InstNumbers.clear(); }
223 struct PromoteMem2Reg {
224 /// The alloca instructions being promoted.
225 std::vector<AllocaInst *> Allocas;
229 /// An AliasSetTracker object to update. If null, don't update it.
230 AliasSetTracker *AST;
232 /// A cache of @llvm.assume intrinsics used by SimplifyInstruction.
235 /// Reverse mapping of Allocas.
236 DenseMap<AllocaInst *, unsigned> AllocaLookup;
238 /// \brief The PhiNodes we're adding.
240 /// That map is used to simplify some Phi nodes as we iterate over it, so
241 /// it should have deterministic iterators. We could use a MapVector, but
242 /// since we already maintain a map from BasicBlock* to a stable numbering
243 /// (BBNumbers), the DenseMap is more efficient (also supports removal).
244 DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes;
246 /// For each PHI node, keep track of which entry in Allocas it corresponds
248 DenseMap<PHINode *, unsigned> PhiToAllocaMap;
250 /// If we are updating an AliasSetTracker, then for each alloca that is of
251 /// pointer type, we keep track of what to copyValue to the inserted PHI
253 std::vector<Value *> PointerAllocaValues;
255 /// For each alloca, we keep track of the dbg.declare intrinsic that
256 /// describes it, if any, so that we can convert it to a dbg.value
257 /// intrinsic if the alloca gets promoted.
258 SmallVector<DbgDeclareInst *, 8> AllocaDbgDeclares;
260 /// The set of basic blocks the renamer has already visited.
262 SmallPtrSet<BasicBlock *, 16> Visited;
264 /// Contains a stable numbering of basic blocks to avoid non-determinstic
266 DenseMap<BasicBlock *, unsigned> BBNumbers;
268 /// Lazily compute the number of predecessors a block has.
269 DenseMap<const BasicBlock *, unsigned> BBNumPreds;
272 PromoteMem2Reg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
273 AliasSetTracker *AST, AssumptionCache *AC)
274 : Allocas(Allocas.begin(), Allocas.end()), DT(DT),
275 DIB(*DT.getRoot()->getParent()->getParent(), /*AllowUnresolved*/ false),
281 void RemoveFromAllocasList(unsigned &AllocaIdx) {
282 Allocas[AllocaIdx] = Allocas.back();
287 unsigned getNumPreds(const BasicBlock *BB) {
288 unsigned &NP = BBNumPreds[BB];
290 NP = std::distance(pred_begin(BB), pred_end(BB)) + 1;
294 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
295 const SmallPtrSetImpl<BasicBlock *> &DefBlocks,
296 SmallPtrSetImpl<BasicBlock *> &LiveInBlocks);
297 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
298 RenamePassData::ValVector &IncVals,
299 std::vector<RenamePassData> &Worklist);
300 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
303 } // end of anonymous namespace
305 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
306 // Knowing that this alloca is promotable, we know that it's safe to kill all
307 // instructions except for load and store.
309 for (auto UI = AI->user_begin(), UE = AI->user_end(); UI != UE;) {
310 Instruction *I = cast<Instruction>(*UI);
312 if (isa<LoadInst>(I) || isa<StoreInst>(I))
315 if (!I->getType()->isVoidTy()) {
316 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
317 // Follow the use/def chain to erase them now instead of leaving it for
318 // dead code elimination later.
319 for (auto UUI = I->user_begin(), UUE = I->user_end(); UUI != UUE;) {
320 Instruction *Inst = cast<Instruction>(*UUI);
322 Inst->eraseFromParent();
325 I->eraseFromParent();
329 /// \brief Rewrite as many loads as possible given a single store.
331 /// When there is only a single store, we can use the domtree to trivially
332 /// replace all of the dominated loads with the stored value. Do so, and return
333 /// true if this has successfully promoted the alloca entirely. If this returns
334 /// false there were some loads which were not dominated by the single store
335 /// and thus must be phi-ed with undef. We fall back to the standard alloca
336 /// promotion algorithm in that case.
337 static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
340 AliasSetTracker *AST) {
341 StoreInst *OnlyStore = Info.OnlyStore;
342 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
343 BasicBlock *StoreBB = OnlyStore->getParent();
346 // Clear out UsingBlocks. We will reconstruct it here if needed.
347 Info.UsingBlocks.clear();
349 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
350 Instruction *UserInst = cast<Instruction>(*UI++);
351 if (!isa<LoadInst>(UserInst)) {
352 assert(UserInst == OnlyStore && "Should only have load/stores");
355 LoadInst *LI = cast<LoadInst>(UserInst);
357 // Okay, if we have a load from the alloca, we want to replace it with the
358 // only value stored to the alloca. We can do this if the value is
359 // dominated by the store. If not, we use the rest of the mem2reg machinery
360 // to insert the phi nodes as needed.
361 if (!StoringGlobalVal) { // Non-instructions are always dominated.
362 if (LI->getParent() == StoreBB) {
363 // If we have a use that is in the same block as the store, compare the
364 // indices of the two instructions to see which one came first. If the
365 // load came before the store, we can't handle it.
366 if (StoreIndex == -1)
367 StoreIndex = LBI.getInstructionIndex(OnlyStore);
369 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
370 // Can't handle this load, bail out.
371 Info.UsingBlocks.push_back(StoreBB);
375 } else if (LI->getParent() != StoreBB &&
376 !DT.dominates(StoreBB, LI->getParent())) {
377 // If the load and store are in different blocks, use BB dominance to
378 // check their relationships. If the store doesn't dom the use, bail
380 Info.UsingBlocks.push_back(LI->getParent());
385 // Otherwise, we *can* safely rewrite this load.
386 Value *ReplVal = OnlyStore->getOperand(0);
387 // If the replacement value is the load, this must occur in unreachable
390 ReplVal = UndefValue::get(LI->getType());
391 LI->replaceAllUsesWith(ReplVal);
392 if (AST && LI->getType()->isPointerTy())
393 AST->deleteValue(LI);
394 LI->eraseFromParent();
398 // Finally, after the scan, check to see if the store is all that is left.
399 if (!Info.UsingBlocks.empty())
400 return false; // If not, we'll have to fall back for the remainder.
402 // Record debuginfo for the store and remove the declaration's
404 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
405 DIBuilder DIB(*AI->getParent()->getParent()->getParent(),
406 /*AllowUnresolved*/ false);
407 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB);
408 DDI->eraseFromParent();
409 LBI.deleteValue(DDI);
411 // Remove the (now dead) store and alloca.
412 Info.OnlyStore->eraseFromParent();
413 LBI.deleteValue(Info.OnlyStore);
416 AST->deleteValue(AI);
417 AI->eraseFromParent();
422 /// Many allocas are only used within a single basic block. If this is the
423 /// case, avoid traversing the CFG and inserting a lot of potentially useless
424 /// PHI nodes by just performing a single linear pass over the basic block
425 /// using the Alloca.
427 /// If we cannot promote this alloca (because it is read before it is written),
428 /// return true. This is necessary in cases where, due to control flow, the
429 /// alloca is potentially undefined on some control flow paths. e.g. code like
430 /// this is potentially correct:
432 /// for (...) { if (c) { A = undef; undef = B; } }
434 /// ... so long as A is not used before undef is set.
435 static void promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
437 AliasSetTracker *AST) {
438 // The trickiest case to handle is when we have large blocks. Because of this,
439 // this code is optimized assuming that large blocks happen. This does not
440 // significantly pessimize the small block case. This uses LargeBlockInfo to
441 // make it efficient to get the index of various operations in the block.
443 // Walk the use-def list of the alloca, getting the locations of all stores.
444 typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy;
445 StoresByIndexTy StoresByIndex;
447 for (User *U : AI->users())
448 if (StoreInst *SI = dyn_cast<StoreInst>(U))
449 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
451 // Sort the stores by their index, making it efficient to do a lookup with a
453 std::sort(StoresByIndex.begin(), StoresByIndex.end(), less_first());
455 // Walk all of the loads from this alloca, replacing them with the nearest
456 // store above them, if any.
457 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
458 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
462 unsigned LoadIdx = LBI.getInstructionIndex(LI);
464 // Find the nearest store that has a lower index than this load.
465 StoresByIndexTy::iterator I =
466 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
467 std::make_pair(LoadIdx,
468 static_cast<StoreInst *>(nullptr)),
471 if (I == StoresByIndex.begin())
472 // If there is no store before this load, the load takes the undef value.
473 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
475 // Otherwise, there was a store before this load, the load takes its value.
476 LI->replaceAllUsesWith(std::prev(I)->second->getOperand(0));
478 if (AST && LI->getType()->isPointerTy())
479 AST->deleteValue(LI);
480 LI->eraseFromParent();
484 // Remove the (now dead) stores and alloca.
485 while (!AI->use_empty()) {
486 StoreInst *SI = cast<StoreInst>(AI->user_back());
487 // Record debuginfo for the store before removing it.
488 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
489 DIBuilder DIB(*AI->getParent()->getParent()->getParent(),
490 /*AllowUnresolved*/ false);
491 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
493 SI->eraseFromParent();
498 AST->deleteValue(AI);
499 AI->eraseFromParent();
502 // The alloca's debuginfo can be removed as well.
503 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
504 DDI->eraseFromParent();
505 LBI.deleteValue(DDI);
511 void PromoteMem2Reg::run() {
512 Function &F = *DT.getRoot()->getParent();
515 PointerAllocaValues.resize(Allocas.size());
516 AllocaDbgDeclares.resize(Allocas.size());
520 IDFCalculator IDF(DT);
522 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
523 AllocaInst *AI = Allocas[AllocaNum];
525 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
526 assert(AI->getParent()->getParent() == &F &&
527 "All allocas should be in the same function, which is same as DF!");
529 removeLifetimeIntrinsicUsers(AI);
531 if (AI->use_empty()) {
532 // If there are no uses of the alloca, just delete it now.
534 AST->deleteValue(AI);
535 AI->eraseFromParent();
537 // Remove the alloca from the Allocas list, since it has been processed
538 RemoveFromAllocasList(AllocaNum);
543 // Calculate the set of read and write-locations for each alloca. This is
544 // analogous to finding the 'uses' and 'definitions' of each variable.
545 Info.AnalyzeAlloca(AI);
547 // If there is only a single store to this value, replace any loads of
548 // it that are directly dominated by the definition with the value stored.
549 if (Info.DefiningBlocks.size() == 1) {
550 if (rewriteSingleStoreAlloca(AI, Info, LBI, DT, AST)) {
551 // The alloca has been processed, move on.
552 RemoveFromAllocasList(AllocaNum);
558 // If the alloca is only read and written in one basic block, just perform a
559 // linear sweep over the block to eliminate it.
560 if (Info.OnlyUsedInOneBlock) {
561 promoteSingleBlockAlloca(AI, Info, LBI, AST);
563 // The alloca has been processed, move on.
564 RemoveFromAllocasList(AllocaNum);
568 // If we haven't computed a numbering for the BB's in the function, do so
570 if (BBNumbers.empty()) {
573 BBNumbers[&BB] = ID++;
576 // If we have an AST to keep updated, remember some pointer value that is
577 // stored into the alloca.
579 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
581 // Remember the dbg.declare intrinsic describing this alloca, if any.
583 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
585 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
586 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
588 // At this point, we're committed to promoting the alloca using IDF's, and
589 // the standard SSA construction algorithm. Determine which blocks need PHI
590 // nodes and see if we can optimize out some work by avoiding insertion of
594 // Unique the set of defining blocks for efficient lookup.
595 SmallPtrSet<BasicBlock *, 32> DefBlocks;
596 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
598 // Determine which blocks the value is live in. These are blocks which lead
600 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
601 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
603 // At this point, we're committed to promoting the alloca using IDF's, and
604 // the standard SSA construction algorithm. Determine which blocks need phi
605 // nodes and see if we can optimize out some work by avoiding insertion of
607 IDF.setLiveInBlocks(LiveInBlocks);
608 IDF.setDefiningBlocks(DefBlocks);
609 SmallVector<BasicBlock *, 32> PHIBlocks;
610 IDF.calculate(PHIBlocks);
611 if (PHIBlocks.size() > 1)
612 std::sort(PHIBlocks.begin(), PHIBlocks.end(),
613 [this](BasicBlock *A, BasicBlock *B) {
614 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
617 unsigned CurrentVersion = 0;
618 for (unsigned i = 0, e = PHIBlocks.size(); i != e; ++i)
619 QueuePhiNode(PHIBlocks[i], AllocaNum, CurrentVersion);
623 return; // All of the allocas must have been trivial!
627 // Set the incoming values for the basic block to be null values for all of
628 // the alloca's. We do this in case there is a load of a value that has not
629 // been stored yet. In this case, it will get this null value.
631 RenamePassData::ValVector Values(Allocas.size());
632 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
633 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
635 // Walks all basic blocks in the function performing the SSA rename algorithm
636 // and inserting the phi nodes we marked as necessary
638 std::vector<RenamePassData> RenamePassWorkList;
639 RenamePassWorkList.emplace_back(F.begin(), nullptr, std::move(Values));
642 RPD.swap(RenamePassWorkList.back());
643 RenamePassWorkList.pop_back();
644 // RenamePass may add new worklist entries.
645 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
646 } while (!RenamePassWorkList.empty());
648 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
651 // Remove the allocas themselves from the function.
652 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
653 Instruction *A = Allocas[i];
655 // If there are any uses of the alloca instructions left, they must be in
656 // unreachable basic blocks that were not processed by walking the dominator
657 // tree. Just delete the users now.
659 A->replaceAllUsesWith(UndefValue::get(A->getType()));
662 A->eraseFromParent();
665 const DataLayout &DL = F.getParent()->getDataLayout();
667 // Remove alloca's dbg.declare instrinsics from the function.
668 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
669 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
670 DDI->eraseFromParent();
672 // Loop over all of the PHI nodes and see if there are any that we can get
673 // rid of because they merge all of the same incoming values. This can
674 // happen due to undef values coming into the PHI nodes. This process is
675 // iterative, because eliminating one PHI node can cause others to be removed.
676 bool EliminatedAPHI = true;
677 while (EliminatedAPHI) {
678 EliminatedAPHI = false;
680 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
681 // simplify and RAUW them as we go. If it was not, we could add uses to
682 // the values we replace with in a non-deterministic order, thus creating
683 // non-deterministic def->use chains.
684 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
685 I = NewPhiNodes.begin(),
686 E = NewPhiNodes.end();
688 PHINode *PN = I->second;
690 // If this PHI node merges one value and/or undefs, get the value.
691 if (Value *V = SimplifyInstruction(PN, DL, nullptr, &DT, AC)) {
692 if (AST && PN->getType()->isPointerTy())
693 AST->deleteValue(PN);
694 PN->replaceAllUsesWith(V);
695 PN->eraseFromParent();
696 NewPhiNodes.erase(I++);
697 EliminatedAPHI = true;
704 // At this point, the renamer has added entries to PHI nodes for all reachable
705 // code. Unfortunately, there may be unreachable blocks which the renamer
706 // hasn't traversed. If this is the case, the PHI nodes may not
707 // have incoming values for all predecessors. Loop over all PHI nodes we have
708 // created, inserting undef values if they are missing any incoming values.
710 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
711 I = NewPhiNodes.begin(),
712 E = NewPhiNodes.end();
714 // We want to do this once per basic block. As such, only process a block
715 // when we find the PHI that is the first entry in the block.
716 PHINode *SomePHI = I->second;
717 BasicBlock *BB = SomePHI->getParent();
718 if (&BB->front() != SomePHI)
721 // Only do work here if there the PHI nodes are missing incoming values. We
722 // know that all PHI nodes that were inserted in a block will have the same
723 // number of incoming values, so we can just check any of them.
724 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
727 // Get the preds for BB.
728 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
730 // Ok, now we know that all of the PHI nodes are missing entries for some
731 // basic blocks. Start by sorting the incoming predecessors for efficient
733 std::sort(Preds.begin(), Preds.end());
735 // Now we loop through all BB's which have entries in SomePHI and remove
736 // them from the Preds list.
737 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
738 // Do a log(n) search of the Preds list for the entry we want.
739 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
740 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
741 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
742 "PHI node has entry for a block which is not a predecessor!");
748 // At this point, the blocks left in the preds list must have dummy
749 // entries inserted into every PHI nodes for the block. Update all the phi
750 // nodes in this block that we are inserting (there could be phis before
752 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
753 BasicBlock::iterator BBI = BB->begin();
754 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
755 SomePHI->getNumIncomingValues() == NumBadPreds) {
756 Value *UndefVal = UndefValue::get(SomePHI->getType());
757 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
758 SomePHI->addIncoming(UndefVal, Preds[pred]);
765 /// \brief Determine which blocks the value is live in.
767 /// These are blocks which lead to uses. Knowing this allows us to avoid
768 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
769 /// inserted phi nodes would be dead).
770 void PromoteMem2Reg::ComputeLiveInBlocks(
771 AllocaInst *AI, AllocaInfo &Info,
772 const SmallPtrSetImpl<BasicBlock *> &DefBlocks,
773 SmallPtrSetImpl<BasicBlock *> &LiveInBlocks) {
775 // To determine liveness, we must iterate through the predecessors of blocks
776 // where the def is live. Blocks are added to the worklist if we need to
777 // check their predecessors. Start with all the using blocks.
778 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
779 Info.UsingBlocks.end());
781 // If any of the using blocks is also a definition block, check to see if the
782 // definition occurs before or after the use. If it happens before the use,
783 // the value isn't really live-in.
784 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
785 BasicBlock *BB = LiveInBlockWorklist[i];
786 if (!DefBlocks.count(BB))
789 // Okay, this is a block that both uses and defines the value. If the first
790 // reference to the alloca is a def (store), then we know it isn't live-in.
791 for (BasicBlock::iterator I = BB->begin();; ++I) {
792 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
793 if (SI->getOperand(1) != AI)
796 // We found a store to the alloca before a load. The alloca is not
797 // actually live-in here.
798 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
799 LiveInBlockWorklist.pop_back();
804 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
805 if (LI->getOperand(0) != AI)
808 // Okay, we found a load before a store to the alloca. It is actually
809 // live into this block.
815 // Now that we have a set of blocks where the phi is live-in, recursively add
816 // their predecessors until we find the full region the value is live.
817 while (!LiveInBlockWorklist.empty()) {
818 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
820 // The block really is live in here, insert it into the set. If already in
821 // the set, then it has already been processed.
822 if (!LiveInBlocks.insert(BB).second)
825 // Since the value is live into BB, it is either defined in a predecessor or
826 // live into it to. Add the preds to the worklist unless they are a
828 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
831 // The value is not live into a predecessor if it defines the value.
832 if (DefBlocks.count(P))
835 // Otherwise it is, add to the worklist.
836 LiveInBlockWorklist.push_back(P);
841 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
843 /// Returns true if there wasn't already a phi-node for that variable
844 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
846 // Look up the basic-block in question.
847 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
849 // If the BB already has a phi node added for the i'th alloca then we're done!
853 // Create a PhiNode using the dereferenced type... and add the phi-node to the
855 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
856 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
859 PhiToAllocaMap[PN] = AllocaNo;
861 if (AST && PN->getType()->isPointerTy())
862 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
867 /// \brief Recursively traverse the CFG of the function, renaming loads and
868 /// stores to the allocas which we are promoting.
870 /// IncomingVals indicates what value each Alloca contains on exit from the
871 /// predecessor block Pred.
872 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
873 RenamePassData::ValVector &IncomingVals,
874 std::vector<RenamePassData> &Worklist) {
876 // If we are inserting any phi nodes into this BB, they will already be in the
878 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
879 // If we have PHI nodes to update, compute the number of edges from Pred to
881 if (PhiToAllocaMap.count(APN)) {
882 // We want to be able to distinguish between PHI nodes being inserted by
883 // this invocation of mem2reg from those phi nodes that already existed in
884 // the IR before mem2reg was run. We determine that APN is being inserted
885 // because it is missing incoming edges. All other PHI nodes being
886 // inserted by this pass of mem2reg will have the same number of incoming
887 // operands so far. Remember this count.
888 unsigned NewPHINumOperands = APN->getNumOperands();
890 unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB);
891 assert(NumEdges && "Must be at least one edge from Pred to BB!");
893 // Add entries for all the phis.
894 BasicBlock::iterator PNI = BB->begin();
896 unsigned AllocaNo = PhiToAllocaMap[APN];
898 // Add N incoming values to the PHI node.
899 for (unsigned i = 0; i != NumEdges; ++i)
900 APN->addIncoming(IncomingVals[AllocaNo], Pred);
902 // The currently active variable for this block is now the PHI.
903 IncomingVals[AllocaNo] = APN;
905 // Get the next phi node.
907 APN = dyn_cast<PHINode>(PNI);
911 // Verify that it is missing entries. If not, it is not being inserted
912 // by this mem2reg invocation so we want to ignore it.
913 } while (APN->getNumOperands() == NewPHINumOperands);
917 // Don't revisit blocks.
918 if (!Visited.insert(BB).second)
921 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) {
922 Instruction *I = II++; // get the instruction, increment iterator
924 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
925 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
929 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
930 if (AI == AllocaLookup.end())
933 Value *V = IncomingVals[AI->second];
935 // Anything using the load now uses the current value.
936 LI->replaceAllUsesWith(V);
937 if (AST && LI->getType()->isPointerTy())
938 AST->deleteValue(LI);
939 BB->getInstList().erase(LI);
940 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
941 // Delete this instruction and mark the name as the current holder of the
943 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
947 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
948 if (ai == AllocaLookup.end())
951 // what value were we writing?
952 IncomingVals[ai->second] = SI->getOperand(0);
953 // Record debuginfo for the store before removing it.
954 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
955 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
956 BB->getInstList().erase(SI);
960 // 'Recurse' to our successors.
961 succ_iterator I = succ_begin(BB), E = succ_end(BB);
965 // Keep track of the successors so we don't visit the same successor twice
966 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
968 // Handle the first successor without using the worklist.
969 VisitedSuccs.insert(*I);
975 if (VisitedSuccs.insert(*I).second)
976 Worklist.emplace_back(*I, Pred, IncomingVals);
981 void llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
982 AliasSetTracker *AST, AssumptionCache *AC) {
983 // If there is nothing to do, bail out...
987 PromoteMem2Reg(Allocas, DT, AST, AC).run();