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/ADT/ArrayRef.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/TinyPtrVector.h"
25 #include "llvm/ADT/Twine.h"
26 #include "llvm/Analysis/AssumptionCache.h"
27 #include "llvm/Analysis/InstructionSimplify.h"
28 #include "llvm/Analysis/IteratedDominanceFrontier.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/CFG.h"
32 #include "llvm/IR/Constant.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DIBuilder.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Instruction.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/IR/Type.h"
46 #include "llvm/IR/User.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #define DEBUG_TYPE "mem2reg"
60 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
61 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
62 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
63 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
65 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
66 // FIXME: If the memory unit is of pointer or integer type, we can permit
67 // assignments to subsections of the memory unit.
68 unsigned AS = AI->getType()->getAddressSpace();
70 // Only allow direct and non-volatile loads and stores...
71 for (const User *U : AI->users()) {
72 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
73 // Note that atomic loads can be transformed; atomic semantics do
74 // not have any meaning for a local alloca.
77 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
78 if (SI->getOperand(0) == AI)
79 return false; // Don't allow a store OF the AI, only INTO the AI.
80 // Note that atomic stores can be transformed; atomic semantics do
81 // not have any meaning for a local alloca.
84 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
85 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
86 II->getIntrinsicID() != Intrinsic::lifetime_end)
88 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
89 if (BCI->getType() != Type::getInt8PtrTy(U->getContext(), AS))
91 if (!onlyUsedByLifetimeMarkers(BCI))
93 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
94 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext(), AS))
96 if (!GEPI->hasAllZeroIndices())
98 if (!onlyUsedByLifetimeMarkers(GEPI))
111 SmallVector<BasicBlock *, 32> DefiningBlocks;
112 SmallVector<BasicBlock *, 32> UsingBlocks;
114 StoreInst *OnlyStore;
115 BasicBlock *OnlyBlock;
116 bool OnlyUsedInOneBlock;
118 Value *AllocaPointerVal;
119 TinyPtrVector<DbgInfoIntrinsic *> DbgDeclares;
122 DefiningBlocks.clear();
126 OnlyUsedInOneBlock = true;
127 AllocaPointerVal = nullptr;
131 /// Scan the uses of the specified alloca, filling in the AllocaInfo used
132 /// by the rest of the pass to reason about the uses of this alloca.
133 void AnalyzeAlloca(AllocaInst *AI) {
136 // As we scan the uses of the alloca instruction, keep track of stores,
137 // and decide whether all of the loads and stores to the alloca are within
138 // the same basic block.
139 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
140 Instruction *User = cast<Instruction>(*UI++);
142 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
143 // Remember the basic blocks which define new values for the alloca
144 DefiningBlocks.push_back(SI->getParent());
145 AllocaPointerVal = SI->getOperand(0);
148 LoadInst *LI = cast<LoadInst>(User);
149 // Otherwise it must be a load instruction, keep track of variable
151 UsingBlocks.push_back(LI->getParent());
152 AllocaPointerVal = LI;
155 if (OnlyUsedInOneBlock) {
157 OnlyBlock = User->getParent();
158 else if (OnlyBlock != User->getParent())
159 OnlyUsedInOneBlock = false;
163 DbgDeclares = FindDbgAddrUses(AI);
167 // Data package used by RenamePass()
168 class RenamePassData {
170 using ValVector = std::vector<Value *>;
172 RenamePassData(BasicBlock *B, BasicBlock *P, ValVector V)
173 : BB(B), Pred(P), Values(std::move(V)) {}
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 (const Instruction &BBI : *BB)
217 if (isInterestingInstruction(&BBI))
218 InstNumbers[&BBI] = InstNo++;
219 It = InstNumbers.find(I);
221 assert(It != InstNumbers.end() && "Didn't insert instruction?");
225 void deleteValue(const Instruction *I) { InstNumbers.erase(I); }
227 void clear() { InstNumbers.clear(); }
230 struct PromoteMem2Reg {
231 /// The alloca instructions being promoted.
232 std::vector<AllocaInst *> Allocas;
237 /// A cache of @llvm.assume intrinsics used by SimplifyInstruction.
240 const SimplifyQuery SQ;
242 /// Reverse mapping of Allocas.
243 DenseMap<AllocaInst *, unsigned> AllocaLookup;
245 /// \brief The PhiNodes we're adding.
247 /// That map is used to simplify some Phi nodes as we iterate over it, so
248 /// it should have deterministic iterators. We could use a MapVector, but
249 /// since we already maintain a map from BasicBlock* to a stable numbering
250 /// (BBNumbers), the DenseMap is more efficient (also supports removal).
251 DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes;
253 /// For each PHI node, keep track of which entry in Allocas it corresponds
255 DenseMap<PHINode *, unsigned> PhiToAllocaMap;
257 /// If we are updating an AliasSetTracker, then for each alloca that is of
258 /// pointer type, we keep track of what to copyValue to the inserted PHI
260 std::vector<Value *> PointerAllocaValues;
262 /// For each alloca, we keep track of the dbg.declare intrinsic that
263 /// describes it, if any, so that we can convert it to a dbg.value
264 /// intrinsic if the alloca gets promoted.
265 SmallVector<TinyPtrVector<DbgInfoIntrinsic *>, 8> AllocaDbgDeclares;
267 /// The set of basic blocks the renamer has already visited.
268 SmallPtrSet<BasicBlock *, 16> Visited;
270 /// Contains a stable numbering of basic blocks to avoid non-determinstic
272 DenseMap<BasicBlock *, unsigned> BBNumbers;
274 /// Lazily compute the number of predecessors a block has.
275 DenseMap<const BasicBlock *, unsigned> BBNumPreds;
278 PromoteMem2Reg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
280 : Allocas(Allocas.begin(), Allocas.end()), DT(DT),
281 DIB(*DT.getRoot()->getParent()->getParent(), /*AllowUnresolved*/ false),
282 AC(AC), SQ(DT.getRoot()->getParent()->getParent()->getDataLayout(),
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 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
302 const SmallPtrSetImpl<BasicBlock *> &DefBlocks,
303 SmallPtrSetImpl<BasicBlock *> &LiveInBlocks);
304 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
305 RenamePassData::ValVector &IncVals,
306 std::vector<RenamePassData> &Worklist);
307 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
310 } // end anonymous namespace
312 /// Given a LoadInst LI this adds assume(LI != null) after it.
313 static void addAssumeNonNull(AssumptionCache *AC, LoadInst *LI) {
314 Function *AssumeIntrinsic =
315 Intrinsic::getDeclaration(LI->getModule(), Intrinsic::assume);
316 ICmpInst *LoadNotNull = new ICmpInst(ICmpInst::ICMP_NE, LI,
317 Constant::getNullValue(LI->getType()));
318 LoadNotNull->insertAfter(LI);
319 CallInst *CI = CallInst::Create(AssumeIntrinsic, {LoadNotNull});
320 CI->insertAfter(LoadNotNull);
321 AC->registerAssumption(CI);
324 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
325 // Knowing that this alloca is promotable, we know that it's safe to kill all
326 // instructions except for load and store.
328 for (auto UI = AI->user_begin(), UE = AI->user_end(); UI != UE;) {
329 Instruction *I = cast<Instruction>(*UI);
331 if (isa<LoadInst>(I) || isa<StoreInst>(I))
334 if (!I->getType()->isVoidTy()) {
335 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
336 // Follow the use/def chain to erase them now instead of leaving it for
337 // dead code elimination later.
338 for (auto UUI = I->user_begin(), UUE = I->user_end(); UUI != UUE;) {
339 Instruction *Inst = cast<Instruction>(*UUI);
341 Inst->eraseFromParent();
344 I->eraseFromParent();
348 /// \brief Rewrite as many loads as possible given a single store.
350 /// When there is only a single store, we can use the domtree to trivially
351 /// replace all of the dominated loads with the stored value. Do so, and return
352 /// true if this has successfully promoted the alloca entirely. If this returns
353 /// false there were some loads which were not dominated by the single store
354 /// and thus must be phi-ed with undef. We fall back to the standard alloca
355 /// promotion algorithm in that case.
356 static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
357 LargeBlockInfo &LBI, const DataLayout &DL,
358 DominatorTree &DT, AssumptionCache *AC) {
359 StoreInst *OnlyStore = Info.OnlyStore;
360 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
361 BasicBlock *StoreBB = OnlyStore->getParent();
364 // Clear out UsingBlocks. We will reconstruct it here if needed.
365 Info.UsingBlocks.clear();
367 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
368 Instruction *UserInst = cast<Instruction>(*UI++);
369 if (!isa<LoadInst>(UserInst)) {
370 assert(UserInst == OnlyStore && "Should only have load/stores");
373 LoadInst *LI = cast<LoadInst>(UserInst);
375 // Okay, if we have a load from the alloca, we want to replace it with the
376 // only value stored to the alloca. We can do this if the value is
377 // dominated by the store. If not, we use the rest of the mem2reg machinery
378 // to insert the phi nodes as needed.
379 if (!StoringGlobalVal) { // Non-instructions are always dominated.
380 if (LI->getParent() == StoreBB) {
381 // If we have a use that is in the same block as the store, compare the
382 // indices of the two instructions to see which one came first. If the
383 // load came before the store, we can't handle it.
384 if (StoreIndex == -1)
385 StoreIndex = LBI.getInstructionIndex(OnlyStore);
387 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
388 // Can't handle this load, bail out.
389 Info.UsingBlocks.push_back(StoreBB);
392 } else if (LI->getParent() != StoreBB &&
393 !DT.dominates(StoreBB, LI->getParent())) {
394 // If the load and store are in different blocks, use BB dominance to
395 // check their relationships. If the store doesn't dom the use, bail
397 Info.UsingBlocks.push_back(LI->getParent());
402 // Otherwise, we *can* safely rewrite this load.
403 Value *ReplVal = OnlyStore->getOperand(0);
404 // If the replacement value is the load, this must occur in unreachable
407 ReplVal = UndefValue::get(LI->getType());
409 // If the load was marked as nonnull we don't want to lose
410 // that information when we erase this Load. So we preserve
411 // it with an assume.
412 if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
413 !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
414 addAssumeNonNull(AC, LI);
416 LI->replaceAllUsesWith(ReplVal);
417 LI->eraseFromParent();
421 // Finally, after the scan, check to see if the store is all that is left.
422 if (!Info.UsingBlocks.empty())
423 return false; // If not, we'll have to fall back for the remainder.
425 // Record debuginfo for the store and remove the declaration's
427 for (DbgInfoIntrinsic *DII : Info.DbgDeclares) {
428 DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
429 ConvertDebugDeclareToDebugValue(DII, Info.OnlyStore, DIB);
430 DII->eraseFromParent();
431 LBI.deleteValue(DII);
433 // Remove the (now dead) store and alloca.
434 Info.OnlyStore->eraseFromParent();
435 LBI.deleteValue(Info.OnlyStore);
437 AI->eraseFromParent();
442 /// Many allocas are only used within a single basic block. If this is the
443 /// case, avoid traversing the CFG and inserting a lot of potentially useless
444 /// PHI nodes by just performing a single linear pass over the basic block
445 /// using the Alloca.
447 /// If we cannot promote this alloca (because it is read before it is written),
448 /// return false. This is necessary in cases where, due to control flow, the
449 /// alloca is undefined only on some control flow paths. e.g. code like
450 /// this is correct in LLVM IR:
451 /// // A is an alloca with no stores so far
454 /// if (!first_iteration)
458 static bool promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
460 const DataLayout &DL,
462 AssumptionCache *AC) {
463 // The trickiest case to handle is when we have large blocks. Because of this,
464 // this code is optimized assuming that large blocks happen. This does not
465 // significantly pessimize the small block case. This uses LargeBlockInfo to
466 // make it efficient to get the index of various operations in the block.
468 // Walk the use-def list of the alloca, getting the locations of all stores.
469 using StoresByIndexTy = SmallVector<std::pair<unsigned, StoreInst *>, 64>;
470 StoresByIndexTy StoresByIndex;
472 for (User *U : AI->users())
473 if (StoreInst *SI = dyn_cast<StoreInst>(U))
474 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
476 // Sort the stores by their index, making it efficient to do a lookup with a
478 std::sort(StoresByIndex.begin(), StoresByIndex.end(), less_first());
480 // Walk all of the loads from this alloca, replacing them with the nearest
481 // store above them, if any.
482 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
483 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
487 unsigned LoadIdx = LBI.getInstructionIndex(LI);
489 // Find the nearest store that has a lower index than this load.
490 StoresByIndexTy::iterator I =
491 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
492 std::make_pair(LoadIdx,
493 static_cast<StoreInst *>(nullptr)),
495 if (I == StoresByIndex.begin()) {
496 if (StoresByIndex.empty())
497 // If there are no stores, the load takes the undef value.
498 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
500 // There is no store before this load, bail out (load may be affected
501 // by the following stores - see main comment).
504 // Otherwise, there was a store before this load, the load takes its value.
505 // Note, if the load was marked as nonnull we don't want to lose that
506 // information when we erase it. So we preserve it with an assume.
507 Value *ReplVal = std::prev(I)->second->getOperand(0);
508 if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
509 !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
510 addAssumeNonNull(AC, LI);
512 LI->replaceAllUsesWith(ReplVal);
515 LI->eraseFromParent();
519 // Remove the (now dead) stores and alloca.
520 while (!AI->use_empty()) {
521 StoreInst *SI = cast<StoreInst>(AI->user_back());
522 // Record debuginfo for the store before removing it.
523 for (DbgInfoIntrinsic *DII : Info.DbgDeclares) {
524 DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
525 ConvertDebugDeclareToDebugValue(DII, SI, DIB);
527 SI->eraseFromParent();
531 AI->eraseFromParent();
534 // The alloca's debuginfo can be removed as well.
535 for (DbgInfoIntrinsic *DII : Info.DbgDeclares) {
536 DII->eraseFromParent();
537 LBI.deleteValue(DII);
544 void PromoteMem2Reg::run() {
545 Function &F = *DT.getRoot()->getParent();
547 AllocaDbgDeclares.resize(Allocas.size());
551 ForwardIDFCalculator IDF(DT);
553 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
554 AllocaInst *AI = Allocas[AllocaNum];
556 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
557 assert(AI->getParent()->getParent() == &F &&
558 "All allocas should be in the same function, which is same as DF!");
560 removeLifetimeIntrinsicUsers(AI);
562 if (AI->use_empty()) {
563 // If there are no uses of the alloca, just delete it now.
564 AI->eraseFromParent();
566 // Remove the alloca from the Allocas list, since it has been processed
567 RemoveFromAllocasList(AllocaNum);
572 // Calculate the set of read and write-locations for each alloca. This is
573 // analogous to finding the 'uses' and 'definitions' of each variable.
574 Info.AnalyzeAlloca(AI);
576 // If there is only a single store to this value, replace any loads of
577 // it that are directly dominated by the definition with the value stored.
578 if (Info.DefiningBlocks.size() == 1) {
579 if (rewriteSingleStoreAlloca(AI, Info, LBI, SQ.DL, DT, AC)) {
580 // The alloca has been processed, move on.
581 RemoveFromAllocasList(AllocaNum);
587 // If the alloca is only read and written in one basic block, just perform a
588 // linear sweep over the block to eliminate it.
589 if (Info.OnlyUsedInOneBlock &&
590 promoteSingleBlockAlloca(AI, Info, LBI, SQ.DL, DT, AC)) {
591 // The alloca has been processed, move on.
592 RemoveFromAllocasList(AllocaNum);
596 // If we haven't computed a numbering for the BB's in the function, do so
598 if (BBNumbers.empty()) {
601 BBNumbers[&BB] = ID++;
604 // Remember the dbg.declare intrinsic describing this alloca, if any.
605 if (!Info.DbgDeclares.empty())
606 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclares;
608 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
609 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
611 // At this point, we're committed to promoting the alloca using IDF's, and
612 // the standard SSA construction algorithm. Determine which blocks need PHI
613 // nodes and see if we can optimize out some work by avoiding insertion of
616 // Unique the set of defining blocks for efficient lookup.
617 SmallPtrSet<BasicBlock *, 32> DefBlocks;
618 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
620 // Determine which blocks the value is live in. These are blocks which lead
622 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
623 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
625 // At this point, we're committed to promoting the alloca using IDF's, and
626 // the standard SSA construction algorithm. Determine which blocks need phi
627 // nodes and see if we can optimize out some work by avoiding insertion of
629 IDF.setLiveInBlocks(LiveInBlocks);
630 IDF.setDefiningBlocks(DefBlocks);
631 SmallVector<BasicBlock *, 32> PHIBlocks;
632 IDF.calculate(PHIBlocks);
633 if (PHIBlocks.size() > 1)
634 std::sort(PHIBlocks.begin(), PHIBlocks.end(),
635 [this](BasicBlock *A, BasicBlock *B) {
636 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
639 unsigned CurrentVersion = 0;
640 for (BasicBlock *BB : PHIBlocks)
641 QueuePhiNode(BB, AllocaNum, CurrentVersion);
645 return; // All of the allocas must have been trivial!
649 // Set the incoming values for the basic block to be null values for all of
650 // the alloca's. We do this in case there is a load of a value that has not
651 // been stored yet. In this case, it will get this null value.
652 RenamePassData::ValVector Values(Allocas.size());
653 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
654 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
656 // Walks all basic blocks in the function performing the SSA rename algorithm
657 // and inserting the phi nodes we marked as necessary
658 std::vector<RenamePassData> RenamePassWorkList;
659 RenamePassWorkList.emplace_back(&F.front(), nullptr, std::move(Values));
661 RenamePassData RPD = std::move(RenamePassWorkList.back());
662 RenamePassWorkList.pop_back();
663 // RenamePass may add new worklist entries.
664 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
665 } while (!RenamePassWorkList.empty());
667 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
670 // Remove the allocas themselves from the function.
671 for (Instruction *A : Allocas) {
672 // If there are any uses of the alloca instructions left, they must be in
673 // unreachable basic blocks that were not processed by walking the dominator
674 // tree. Just delete the users now.
676 A->replaceAllUsesWith(UndefValue::get(A->getType()));
677 A->eraseFromParent();
680 // Remove alloca's dbg.declare instrinsics from the function.
681 for (auto &Declares : AllocaDbgDeclares)
682 for (auto *DII : Declares)
683 DII->eraseFromParent();
685 // Loop over all of the PHI nodes and see if there are any that we can get
686 // rid of because they merge all of the same incoming values. This can
687 // happen due to undef values coming into the PHI nodes. This process is
688 // iterative, because eliminating one PHI node can cause others to be removed.
689 bool EliminatedAPHI = true;
690 while (EliminatedAPHI) {
691 EliminatedAPHI = false;
693 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
694 // simplify and RAUW them as we go. If it was not, we could add uses to
695 // the values we replace with in a non-deterministic order, thus creating
696 // non-deterministic def->use chains.
697 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
698 I = NewPhiNodes.begin(),
699 E = NewPhiNodes.end();
701 PHINode *PN = I->second;
703 // If this PHI node merges one value and/or undefs, get the value.
704 if (Value *V = SimplifyInstruction(PN, SQ)) {
705 PN->replaceAllUsesWith(V);
706 PN->eraseFromParent();
707 NewPhiNodes.erase(I++);
708 EliminatedAPHI = true;
715 // At this point, the renamer has added entries to PHI nodes for all reachable
716 // code. Unfortunately, there may be unreachable blocks which the renamer
717 // hasn't traversed. If this is the case, the PHI nodes may not
718 // have incoming values for all predecessors. Loop over all PHI nodes we have
719 // created, inserting undef values if they are missing any incoming values.
720 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
721 I = NewPhiNodes.begin(),
722 E = NewPhiNodes.end();
724 // We want to do this once per basic block. As such, only process a block
725 // when we find the PHI that is the first entry in the block.
726 PHINode *SomePHI = I->second;
727 BasicBlock *BB = SomePHI->getParent();
728 if (&BB->front() != SomePHI)
731 // Only do work here if there the PHI nodes are missing incoming values. We
732 // know that all PHI nodes that were inserted in a block will have the same
733 // number of incoming values, so we can just check any of them.
734 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
737 // Get the preds for BB.
738 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
740 // Ok, now we know that all of the PHI nodes are missing entries for some
741 // basic blocks. Start by sorting the incoming predecessors for efficient
743 std::sort(Preds.begin(), Preds.end());
745 // Now we loop through all BB's which have entries in SomePHI and remove
746 // them from the Preds list.
747 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
748 // Do a log(n) search of the Preds list for the entry we want.
749 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
750 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i));
751 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
752 "PHI node has entry for a block which is not a predecessor!");
758 // At this point, the blocks left in the preds list must have dummy
759 // entries inserted into every PHI nodes for the block. Update all the phi
760 // nodes in this block that we are inserting (there could be phis before
762 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
763 BasicBlock::iterator BBI = BB->begin();
764 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
765 SomePHI->getNumIncomingValues() == NumBadPreds) {
766 Value *UndefVal = UndefValue::get(SomePHI->getType());
767 for (BasicBlock *Pred : Preds)
768 SomePHI->addIncoming(UndefVal, Pred);
775 /// \brief Determine which blocks the value is live in.
777 /// These are blocks which lead to uses. Knowing this allows us to avoid
778 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
779 /// inserted phi nodes would be dead).
780 void PromoteMem2Reg::ComputeLiveInBlocks(
781 AllocaInst *AI, AllocaInfo &Info,
782 const SmallPtrSetImpl<BasicBlock *> &DefBlocks,
783 SmallPtrSetImpl<BasicBlock *> &LiveInBlocks) {
784 // To determine liveness, we must iterate through the predecessors of blocks
785 // where the def is live. Blocks are added to the worklist if we need to
786 // check their predecessors. Start with all the using blocks.
787 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
788 Info.UsingBlocks.end());
790 // If any of the using blocks is also a definition block, check to see if the
791 // definition occurs before or after the use. If it happens before the use,
792 // the value isn't really live-in.
793 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
794 BasicBlock *BB = LiveInBlockWorklist[i];
795 if (!DefBlocks.count(BB))
798 // Okay, this is a block that both uses and defines the value. If the first
799 // reference to the alloca is a def (store), then we know it isn't live-in.
800 for (BasicBlock::iterator I = BB->begin();; ++I) {
801 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
802 if (SI->getOperand(1) != AI)
805 // We found a store to the alloca before a load. The alloca is not
806 // actually live-in here.
807 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
808 LiveInBlockWorklist.pop_back();
814 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
815 if (LI->getOperand(0) != AI)
818 // Okay, we found a load before a store to the alloca. It is actually
819 // live into this block.
825 // Now that we have a set of blocks where the phi is live-in, recursively add
826 // their predecessors until we find the full region the value is live.
827 while (!LiveInBlockWorklist.empty()) {
828 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
830 // The block really is live in here, insert it into the set. If already in
831 // the set, then it has already been processed.
832 if (!LiveInBlocks.insert(BB).second)
835 // Since the value is live into BB, it is either defined in a predecessor or
836 // live into it to. Add the preds to the worklist unless they are a
838 for (BasicBlock *P : predecessors(BB)) {
839 // The value is not live into a predecessor if it defines the value.
840 if (DefBlocks.count(P))
843 // Otherwise it is, add to the worklist.
844 LiveInBlockWorklist.push_back(P);
849 /// \brief Queue a phi-node to be added to a basic-block for a specific Alloca.
851 /// Returns true if there wasn't already a phi-node for that variable
852 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
854 // Look up the basic-block in question.
855 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
857 // If the BB already has a phi node added for the i'th alloca then we're done!
861 // Create a PhiNode using the dereferenced type... and add the phi-node to the
863 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
864 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
867 PhiToAllocaMap[PN] = AllocaNo;
871 /// \brief Recursively traverse the CFG of the function, renaming loads and
872 /// stores to the allocas which we are promoting.
874 /// IncomingVals indicates what value each Alloca contains on exit from the
875 /// predecessor block Pred.
876 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
877 RenamePassData::ValVector &IncomingVals,
878 std::vector<RenamePassData> &Worklist) {
880 // If we are inserting any phi nodes into this BB, they will already be in the
882 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
883 // If we have PHI nodes to update, compute the number of edges from Pred to
885 if (PhiToAllocaMap.count(APN)) {
886 // We want to be able to distinguish between PHI nodes being inserted by
887 // this invocation of mem2reg from those phi nodes that already existed in
888 // the IR before mem2reg was run. We determine that APN is being inserted
889 // because it is missing incoming edges. All other PHI nodes being
890 // inserted by this pass of mem2reg will have the same number of incoming
891 // operands so far. Remember this count.
892 unsigned NewPHINumOperands = APN->getNumOperands();
894 unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB);
895 assert(NumEdges && "Must be at least one edge from Pred to BB!");
897 // Add entries for all the phis.
898 BasicBlock::iterator PNI = BB->begin();
900 unsigned AllocaNo = PhiToAllocaMap[APN];
902 // Add N incoming values to the PHI node.
903 for (unsigned i = 0; i != NumEdges; ++i)
904 APN->addIncoming(IncomingVals[AllocaNo], Pred);
906 // The currently active variable for this block is now the PHI.
907 IncomingVals[AllocaNo] = APN;
908 for (DbgInfoIntrinsic *DII : AllocaDbgDeclares[AllocaNo])
909 ConvertDebugDeclareToDebugValue(DII, APN, DIB);
911 // Get the next phi node.
913 APN = dyn_cast<PHINode>(PNI);
917 // Verify that it is missing entries. If not, it is not being inserted
918 // by this mem2reg invocation so we want to ignore it.
919 } while (APN->getNumOperands() == NewPHINumOperands);
923 // Don't revisit blocks.
924 if (!Visited.insert(BB).second)
927 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) {
928 Instruction *I = &*II++; // get the instruction, increment iterator
930 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
931 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
935 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
936 if (AI == AllocaLookup.end())
939 Value *V = IncomingVals[AI->second];
941 // If the load was marked as nonnull we don't want to lose
942 // that information when we erase this Load. So we preserve
943 // it with an assume.
944 if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
945 !isKnownNonZero(V, SQ.DL, 0, AC, LI, &DT))
946 addAssumeNonNull(AC, LI);
948 // Anything using the load now uses the current value.
949 LI->replaceAllUsesWith(V);
950 BB->getInstList().erase(LI);
951 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
952 // Delete this instruction and mark the name as the current holder of the
954 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
958 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
959 if (ai == AllocaLookup.end())
962 // what value were we writing?
963 IncomingVals[ai->second] = SI->getOperand(0);
964 // Record debuginfo for the store before removing it.
965 for (DbgInfoIntrinsic *DII : AllocaDbgDeclares[ai->second])
966 ConvertDebugDeclareToDebugValue(DII, SI, DIB);
967 BB->getInstList().erase(SI);
971 // 'Recurse' to our successors.
972 succ_iterator I = succ_begin(BB), E = succ_end(BB);
976 // Keep track of the successors so we don't visit the same successor twice
977 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
979 // Handle the first successor without using the worklist.
980 VisitedSuccs.insert(*I);
986 if (VisitedSuccs.insert(*I).second)
987 Worklist.emplace_back(*I, Pred, IncomingVals);
992 void llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
993 AssumptionCache *AC) {
994 // If there is nothing to do, bail out...
998 PromoteMem2Reg(Allocas, DT, AC).run();