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/Transforms/Utils/Local.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/CFG.h"
33 #include "llvm/IR/Constant.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DIBuilder.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/InstrTypes.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/IR/User.h"
48 #include "llvm/Support/Casting.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->isLifetimeStartOrEnd())
87 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
88 if (BCI->getType() != Type::getInt8PtrTy(U->getContext(), AS))
90 if (!onlyUsedByLifetimeMarkers(BCI))
92 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
93 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext(), AS))
95 if (!GEPI->hasAllZeroIndices())
97 if (!onlyUsedByLifetimeMarkers(GEPI))
110 SmallVector<BasicBlock *, 32> DefiningBlocks;
111 SmallVector<BasicBlock *, 32> UsingBlocks;
113 StoreInst *OnlyStore;
114 BasicBlock *OnlyBlock;
115 bool OnlyUsedInOneBlock;
117 Value *AllocaPointerVal;
118 TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares;
121 DefiningBlocks.clear();
125 OnlyUsedInOneBlock = true;
126 AllocaPointerVal = nullptr;
130 /// Scan the uses of the specified alloca, filling in the AllocaInfo used
131 /// by the rest of the pass to reason about the uses of this alloca.
132 void AnalyzeAlloca(AllocaInst *AI) {
135 // As we scan the uses of the alloca instruction, keep track of stores,
136 // and decide whether all of the loads and stores to the alloca are within
137 // the same basic block.
138 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
139 Instruction *User = cast<Instruction>(*UI++);
141 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
142 // Remember the basic blocks which define new values for the alloca
143 DefiningBlocks.push_back(SI->getParent());
144 AllocaPointerVal = SI->getOperand(0);
147 LoadInst *LI = cast<LoadInst>(User);
148 // Otherwise it must be a load instruction, keep track of variable
150 UsingBlocks.push_back(LI->getParent());
151 AllocaPointerVal = LI;
154 if (OnlyUsedInOneBlock) {
156 OnlyBlock = User->getParent();
157 else if (OnlyBlock != User->getParent())
158 OnlyUsedInOneBlock = false;
162 DbgDeclares = FindDbgAddrUses(AI);
166 /// Data package used by RenamePass().
167 struct RenamePassData {
168 using ValVector = std::vector<Value *>;
169 using LocationVector = std::vector<DebugLoc>;
171 RenamePassData(BasicBlock *B, BasicBlock *P, ValVector V, LocationVector L)
172 : BB(B), Pred(P), Values(std::move(V)), Locations(std::move(L)) {}
177 LocationVector Locations;
180 /// 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 /// 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 /// 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<DbgVariableIntrinsic *>, 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 = pred_size(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 RenamePassData::LocationVector &IncLocs,
307 std::vector<RenamePassData> &Worklist);
308 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
311 } // end anonymous namespace
313 /// Given a LoadInst LI this adds assume(LI != null) after it.
314 static void addAssumeNonNull(AssumptionCache *AC, LoadInst *LI) {
315 Function *AssumeIntrinsic =
316 Intrinsic::getDeclaration(LI->getModule(), Intrinsic::assume);
317 ICmpInst *LoadNotNull = new ICmpInst(ICmpInst::ICMP_NE, LI,
318 Constant::getNullValue(LI->getType()));
319 LoadNotNull->insertAfter(LI);
320 CallInst *CI = CallInst::Create(AssumeIntrinsic, {LoadNotNull});
321 CI->insertAfter(LoadNotNull);
322 AC->registerAssumption(CI);
325 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
326 // Knowing that this alloca is promotable, we know that it's safe to kill all
327 // instructions except for load and store.
329 for (auto UI = AI->user_begin(), UE = AI->user_end(); UI != UE;) {
330 Instruction *I = cast<Instruction>(*UI);
332 if (isa<LoadInst>(I) || isa<StoreInst>(I))
335 if (!I->getType()->isVoidTy()) {
336 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
337 // Follow the use/def chain to erase them now instead of leaving it for
338 // dead code elimination later.
339 for (auto UUI = I->user_begin(), UUE = I->user_end(); UUI != UUE;) {
340 Instruction *Inst = cast<Instruction>(*UUI);
342 Inst->eraseFromParent();
345 I->eraseFromParent();
349 /// Rewrite as many loads as possible given a single store.
351 /// When there is only a single store, we can use the domtree to trivially
352 /// replace all of the dominated loads with the stored value. Do so, and return
353 /// true if this has successfully promoted the alloca entirely. If this returns
354 /// false there were some loads which were not dominated by the single store
355 /// and thus must be phi-ed with undef. We fall back to the standard alloca
356 /// promotion algorithm in that case.
357 static bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
358 LargeBlockInfo &LBI, const DataLayout &DL,
359 DominatorTree &DT, AssumptionCache *AC) {
360 StoreInst *OnlyStore = Info.OnlyStore;
361 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
362 BasicBlock *StoreBB = OnlyStore->getParent();
365 // Clear out UsingBlocks. We will reconstruct it here if needed.
366 Info.UsingBlocks.clear();
368 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
369 Instruction *UserInst = cast<Instruction>(*UI++);
370 if (!isa<LoadInst>(UserInst)) {
371 assert(UserInst == OnlyStore && "Should only have load/stores");
374 LoadInst *LI = cast<LoadInst>(UserInst);
376 // Okay, if we have a load from the alloca, we want to replace it with the
377 // only value stored to the alloca. We can do this if the value is
378 // dominated by the store. If not, we use the rest of the mem2reg machinery
379 // to insert the phi nodes as needed.
380 if (!StoringGlobalVal) { // Non-instructions are always dominated.
381 if (LI->getParent() == StoreBB) {
382 // If we have a use that is in the same block as the store, compare the
383 // indices of the two instructions to see which one came first. If the
384 // load came before the store, we can't handle it.
385 if (StoreIndex == -1)
386 StoreIndex = LBI.getInstructionIndex(OnlyStore);
388 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
389 // Can't handle this load, bail out.
390 Info.UsingBlocks.push_back(StoreBB);
393 } else if (LI->getParent() != StoreBB &&
394 !DT.dominates(StoreBB, LI->getParent())) {
395 // If the load and store are in different blocks, use BB dominance to
396 // check their relationships. If the store doesn't dom the use, bail
398 Info.UsingBlocks.push_back(LI->getParent());
403 // Otherwise, we *can* safely rewrite this load.
404 Value *ReplVal = OnlyStore->getOperand(0);
405 // If the replacement value is the load, this must occur in unreachable
408 ReplVal = UndefValue::get(LI->getType());
410 // If the load was marked as nonnull we don't want to lose
411 // that information when we erase this Load. So we preserve
412 // it with an assume.
413 if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
414 !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
415 addAssumeNonNull(AC, LI);
417 LI->replaceAllUsesWith(ReplVal);
418 LI->eraseFromParent();
422 // Finally, after the scan, check to see if the store is all that is left.
423 if (!Info.UsingBlocks.empty())
424 return false; // If not, we'll have to fall back for the remainder.
426 // Record debuginfo for the store and remove the declaration's
428 for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
429 DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
430 ConvertDebugDeclareToDebugValue(DII, Info.OnlyStore, DIB);
431 DII->eraseFromParent();
432 LBI.deleteValue(DII);
434 // Remove the (now dead) store and alloca.
435 Info.OnlyStore->eraseFromParent();
436 LBI.deleteValue(Info.OnlyStore);
438 AI->eraseFromParent();
443 /// Many allocas are only used within a single basic block. If this is the
444 /// case, avoid traversing the CFG and inserting a lot of potentially useless
445 /// PHI nodes by just performing a single linear pass over the basic block
446 /// using the Alloca.
448 /// If we cannot promote this alloca (because it is read before it is written),
449 /// return false. This is necessary in cases where, due to control flow, the
450 /// alloca is undefined only on some control flow paths. e.g. code like
451 /// this is correct in LLVM IR:
452 /// // A is an alloca with no stores so far
455 /// if (!first_iteration)
459 static bool promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info,
461 const DataLayout &DL,
463 AssumptionCache *AC) {
464 // The trickiest case to handle is when we have large blocks. Because of this,
465 // this code is optimized assuming that large blocks happen. This does not
466 // significantly pessimize the small block case. This uses LargeBlockInfo to
467 // make it efficient to get the index of various operations in the block.
469 // Walk the use-def list of the alloca, getting the locations of all stores.
470 using StoresByIndexTy = SmallVector<std::pair<unsigned, StoreInst *>, 64>;
471 StoresByIndexTy StoresByIndex;
473 for (User *U : AI->users())
474 if (StoreInst *SI = dyn_cast<StoreInst>(U))
475 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
477 // Sort the stores by their index, making it efficient to do a lookup with a
479 llvm::sort(StoresByIndex, less_first());
481 // Walk all of the loads from this alloca, replacing them with the nearest
482 // store above them, if any.
483 for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) {
484 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
488 unsigned LoadIdx = LBI.getInstructionIndex(LI);
490 // Find the nearest store that has a lower index than this load.
491 StoresByIndexTy::iterator I =
492 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
493 std::make_pair(LoadIdx,
494 static_cast<StoreInst *>(nullptr)),
496 if (I == StoresByIndex.begin()) {
497 if (StoresByIndex.empty())
498 // If there are no stores, the load takes the undef value.
499 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
501 // There is no store before this load, bail out (load may be affected
502 // by the following stores - see main comment).
505 // Otherwise, there was a store before this load, the load takes its value.
506 // Note, if the load was marked as nonnull we don't want to lose that
507 // information when we erase it. So we preserve it with an assume.
508 Value *ReplVal = std::prev(I)->second->getOperand(0);
509 if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
510 !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT))
511 addAssumeNonNull(AC, LI);
513 // If the replacement value is the load, this must occur in unreachable
516 ReplVal = UndefValue::get(LI->getType());
518 LI->replaceAllUsesWith(ReplVal);
521 LI->eraseFromParent();
525 // Remove the (now dead) stores and alloca.
526 while (!AI->use_empty()) {
527 StoreInst *SI = cast<StoreInst>(AI->user_back());
528 // Record debuginfo for the store before removing it.
529 for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
530 DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false);
531 ConvertDebugDeclareToDebugValue(DII, SI, DIB);
533 SI->eraseFromParent();
537 AI->eraseFromParent();
540 // The alloca's debuginfo can be removed as well.
541 for (DbgVariableIntrinsic *DII : Info.DbgDeclares) {
542 DII->eraseFromParent();
543 LBI.deleteValue(DII);
550 void PromoteMem2Reg::run() {
551 Function &F = *DT.getRoot()->getParent();
553 AllocaDbgDeclares.resize(Allocas.size());
557 ForwardIDFCalculator IDF(DT);
559 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
560 AllocaInst *AI = Allocas[AllocaNum];
562 assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!");
563 assert(AI->getParent()->getParent() == &F &&
564 "All allocas should be in the same function, which is same as DF!");
566 removeLifetimeIntrinsicUsers(AI);
568 if (AI->use_empty()) {
569 // If there are no uses of the alloca, just delete it now.
570 AI->eraseFromParent();
572 // Remove the alloca from the Allocas list, since it has been processed
573 RemoveFromAllocasList(AllocaNum);
578 // Calculate the set of read and write-locations for each alloca. This is
579 // analogous to finding the 'uses' and 'definitions' of each variable.
580 Info.AnalyzeAlloca(AI);
582 // If there is only a single store to this value, replace any loads of
583 // it that are directly dominated by the definition with the value stored.
584 if (Info.DefiningBlocks.size() == 1) {
585 if (rewriteSingleStoreAlloca(AI, Info, LBI, SQ.DL, DT, AC)) {
586 // The alloca has been processed, move on.
587 RemoveFromAllocasList(AllocaNum);
593 // If the alloca is only read and written in one basic block, just perform a
594 // linear sweep over the block to eliminate it.
595 if (Info.OnlyUsedInOneBlock &&
596 promoteSingleBlockAlloca(AI, Info, LBI, SQ.DL, DT, AC)) {
597 // The alloca has been processed, move on.
598 RemoveFromAllocasList(AllocaNum);
602 // If we haven't computed a numbering for the BB's in the function, do so
604 if (BBNumbers.empty()) {
607 BBNumbers[&BB] = ID++;
610 // Remember the dbg.declare intrinsic describing this alloca, if any.
611 if (!Info.DbgDeclares.empty())
612 AllocaDbgDeclares[AllocaNum] = Info.DbgDeclares;
614 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
615 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
617 // At this point, we're committed to promoting the alloca using IDF's, and
618 // the standard SSA construction algorithm. Determine which blocks need PHI
619 // nodes and see if we can optimize out some work by avoiding insertion of
622 // Unique the set of defining blocks for efficient lookup.
623 SmallPtrSet<BasicBlock *, 32> DefBlocks;
624 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
626 // Determine which blocks the value is live in. These are blocks which lead
628 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
629 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
631 // At this point, we're committed to promoting the alloca using IDF's, and
632 // the standard SSA construction algorithm. Determine which blocks need phi
633 // nodes and see if we can optimize out some work by avoiding insertion of
635 IDF.setLiveInBlocks(LiveInBlocks);
636 IDF.setDefiningBlocks(DefBlocks);
637 SmallVector<BasicBlock *, 32> PHIBlocks;
638 IDF.calculate(PHIBlocks);
639 if (PHIBlocks.size() > 1)
640 llvm::sort(PHIBlocks, [this](BasicBlock *A, BasicBlock *B) {
641 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
644 unsigned CurrentVersion = 0;
645 for (BasicBlock *BB : PHIBlocks)
646 QueuePhiNode(BB, AllocaNum, CurrentVersion);
650 return; // All of the allocas must have been trivial!
654 // Set the incoming values for the basic block to be null values for all of
655 // the alloca's. We do this in case there is a load of a value that has not
656 // been stored yet. In this case, it will get this null value.
657 RenamePassData::ValVector Values(Allocas.size());
658 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
659 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
661 // When handling debug info, treat all incoming values as if they have unknown
662 // locations until proven otherwise.
663 RenamePassData::LocationVector Locations(Allocas.size());
665 // Walks all basic blocks in the function performing the SSA rename algorithm
666 // and inserting the phi nodes we marked as necessary
667 std::vector<RenamePassData> RenamePassWorkList;
668 RenamePassWorkList.emplace_back(&F.front(), nullptr, std::move(Values),
669 std::move(Locations));
671 RenamePassData RPD = std::move(RenamePassWorkList.back());
672 RenamePassWorkList.pop_back();
673 // RenamePass may add new worklist entries.
674 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RPD.Locations, RenamePassWorkList);
675 } while (!RenamePassWorkList.empty());
677 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
680 // Remove the allocas themselves from the function.
681 for (Instruction *A : Allocas) {
682 // If there are any uses of the alloca instructions left, they must be in
683 // unreachable basic blocks that were not processed by walking the dominator
684 // tree. Just delete the users now.
686 A->replaceAllUsesWith(UndefValue::get(A->getType()));
687 A->eraseFromParent();
690 // Remove alloca's dbg.declare instrinsics from the function.
691 for (auto &Declares : AllocaDbgDeclares)
692 for (auto *DII : Declares)
693 DII->eraseFromParent();
695 // Loop over all of the PHI nodes and see if there are any that we can get
696 // rid of because they merge all of the same incoming values. This can
697 // happen due to undef values coming into the PHI nodes. This process is
698 // iterative, because eliminating one PHI node can cause others to be removed.
699 bool EliminatedAPHI = true;
700 while (EliminatedAPHI) {
701 EliminatedAPHI = false;
703 // Iterating over NewPhiNodes is deterministic, so it is safe to try to
704 // simplify and RAUW them as we go. If it was not, we could add uses to
705 // the values we replace with in a non-deterministic order, thus creating
706 // non-deterministic def->use chains.
707 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
708 I = NewPhiNodes.begin(),
709 E = NewPhiNodes.end();
711 PHINode *PN = I->second;
713 // If this PHI node merges one value and/or undefs, get the value.
714 if (Value *V = SimplifyInstruction(PN, SQ)) {
715 PN->replaceAllUsesWith(V);
716 PN->eraseFromParent();
717 NewPhiNodes.erase(I++);
718 EliminatedAPHI = true;
725 // At this point, the renamer has added entries to PHI nodes for all reachable
726 // code. Unfortunately, there may be unreachable blocks which the renamer
727 // hasn't traversed. If this is the case, the PHI nodes may not
728 // have incoming values for all predecessors. Loop over all PHI nodes we have
729 // created, inserting undef values if they are missing any incoming values.
730 for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator
731 I = NewPhiNodes.begin(),
732 E = NewPhiNodes.end();
734 // We want to do this once per basic block. As such, only process a block
735 // when we find the PHI that is the first entry in the block.
736 PHINode *SomePHI = I->second;
737 BasicBlock *BB = SomePHI->getParent();
738 if (&BB->front() != SomePHI)
741 // Only do work here if there the PHI nodes are missing incoming values. We
742 // know that all PHI nodes that were inserted in a block will have the same
743 // number of incoming values, so we can just check any of them.
744 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
747 // Get the preds for BB.
748 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
750 // Ok, now we know that all of the PHI nodes are missing entries for some
751 // basic blocks. Start by sorting the incoming predecessors for efficient
753 auto CompareBBNumbers = [this](BasicBlock *A, BasicBlock *B) {
754 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
756 llvm::sort(Preds, CompareBBNumbers);
758 // Now we loop through all BB's which have entries in SomePHI and remove
759 // them from the Preds list.
760 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
761 // Do a log(n) search of the Preds list for the entry we want.
762 SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound(
763 Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i),
765 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) &&
766 "PHI node has entry for a block which is not a predecessor!");
772 // At this point, the blocks left in the preds list must have dummy
773 // entries inserted into every PHI nodes for the block. Update all the phi
774 // nodes in this block that we are inserting (there could be phis before
776 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
777 BasicBlock::iterator BBI = BB->begin();
778 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
779 SomePHI->getNumIncomingValues() == NumBadPreds) {
780 Value *UndefVal = UndefValue::get(SomePHI->getType());
781 for (BasicBlock *Pred : Preds)
782 SomePHI->addIncoming(UndefVal, Pred);
789 /// Determine which blocks the value is live in.
791 /// These are blocks which lead to uses. Knowing this allows us to avoid
792 /// inserting PHI nodes into blocks which don't lead to uses (thus, the
793 /// inserted phi nodes would be dead).
794 void PromoteMem2Reg::ComputeLiveInBlocks(
795 AllocaInst *AI, AllocaInfo &Info,
796 const SmallPtrSetImpl<BasicBlock *> &DefBlocks,
797 SmallPtrSetImpl<BasicBlock *> &LiveInBlocks) {
798 // To determine liveness, we must iterate through the predecessors of blocks
799 // where the def is live. Blocks are added to the worklist if we need to
800 // check their predecessors. Start with all the using blocks.
801 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
802 Info.UsingBlocks.end());
804 // If any of the using blocks is also a definition block, check to see if the
805 // definition occurs before or after the use. If it happens before the use,
806 // the value isn't really live-in.
807 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
808 BasicBlock *BB = LiveInBlockWorklist[i];
809 if (!DefBlocks.count(BB))
812 // Okay, this is a block that both uses and defines the value. If the first
813 // reference to the alloca is a def (store), then we know it isn't live-in.
814 for (BasicBlock::iterator I = BB->begin();; ++I) {
815 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
816 if (SI->getOperand(1) != AI)
819 // We found a store to the alloca before a load. The alloca is not
820 // actually live-in here.
821 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
822 LiveInBlockWorklist.pop_back();
828 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
829 if (LI->getOperand(0) != AI)
832 // Okay, we found a load before a store to the alloca. It is actually
833 // live into this block.
839 // Now that we have a set of blocks where the phi is live-in, recursively add
840 // their predecessors until we find the full region the value is live.
841 while (!LiveInBlockWorklist.empty()) {
842 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
844 // The block really is live in here, insert it into the set. If already in
845 // the set, then it has already been processed.
846 if (!LiveInBlocks.insert(BB).second)
849 // Since the value is live into BB, it is either defined in a predecessor or
850 // live into it to. Add the preds to the worklist unless they are a
852 for (BasicBlock *P : predecessors(BB)) {
853 // The value is not live into a predecessor if it defines the value.
854 if (DefBlocks.count(P))
857 // Otherwise it is, add to the worklist.
858 LiveInBlockWorklist.push_back(P);
863 /// Queue a phi-node to be added to a basic-block for a specific Alloca.
865 /// Returns true if there wasn't already a phi-node for that variable
866 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
868 // Look up the basic-block in question.
869 PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)];
871 // If the BB already has a phi node added for the i'th alloca then we're done!
875 // Create a PhiNode using the dereferenced type... and add the phi-node to the
877 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
878 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
881 PhiToAllocaMap[PN] = AllocaNo;
885 /// Update the debug location of a phi. \p ApplyMergedLoc indicates whether to
886 /// create a merged location incorporating \p DL, or to set \p DL directly.
887 static void updateForIncomingValueLocation(PHINode *PN, DebugLoc DL,
888 bool ApplyMergedLoc) {
890 PN->applyMergedLocation(PN->getDebugLoc(), DL);
895 /// Recursively traverse the CFG of the function, renaming loads and
896 /// stores to the allocas which we are promoting.
898 /// IncomingVals indicates what value each Alloca contains on exit from the
899 /// predecessor block Pred.
900 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
901 RenamePassData::ValVector &IncomingVals,
902 RenamePassData::LocationVector &IncomingLocs,
903 std::vector<RenamePassData> &Worklist) {
905 // If we are inserting any phi nodes into this BB, they will already be in the
907 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
908 // If we have PHI nodes to update, compute the number of edges from Pred to
910 if (PhiToAllocaMap.count(APN)) {
911 // We want to be able to distinguish between PHI nodes being inserted by
912 // this invocation of mem2reg from those phi nodes that already existed in
913 // the IR before mem2reg was run. We determine that APN is being inserted
914 // because it is missing incoming edges. All other PHI nodes being
915 // inserted by this pass of mem2reg will have the same number of incoming
916 // operands so far. Remember this count.
917 unsigned NewPHINumOperands = APN->getNumOperands();
919 unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB);
920 assert(NumEdges && "Must be at least one edge from Pred to BB!");
922 // Add entries for all the phis.
923 BasicBlock::iterator PNI = BB->begin();
925 unsigned AllocaNo = PhiToAllocaMap[APN];
927 // Update the location of the phi node.
928 updateForIncomingValueLocation(APN, IncomingLocs[AllocaNo],
929 APN->getNumIncomingValues() > 0);
931 // Add N incoming values to the PHI node.
932 for (unsigned i = 0; i != NumEdges; ++i)
933 APN->addIncoming(IncomingVals[AllocaNo], Pred);
935 // The currently active variable for this block is now the PHI.
936 IncomingVals[AllocaNo] = APN;
937 for (DbgVariableIntrinsic *DII : AllocaDbgDeclares[AllocaNo])
938 ConvertDebugDeclareToDebugValue(DII, APN, DIB);
940 // Get the next phi node.
942 APN = dyn_cast<PHINode>(PNI);
946 // Verify that it is missing entries. If not, it is not being inserted
947 // by this mem2reg invocation so we want to ignore it.
948 } while (APN->getNumOperands() == NewPHINumOperands);
952 // Don't revisit blocks.
953 if (!Visited.insert(BB).second)
956 for (BasicBlock::iterator II = BB->begin(); !II->isTerminator();) {
957 Instruction *I = &*II++; // get the instruction, increment iterator
959 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
960 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
964 DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src);
965 if (AI == AllocaLookup.end())
968 Value *V = IncomingVals[AI->second];
970 // If the load was marked as nonnull we don't want to lose
971 // that information when we erase this Load. So we preserve
972 // it with an assume.
973 if (AC && LI->getMetadata(LLVMContext::MD_nonnull) &&
974 !isKnownNonZero(V, SQ.DL, 0, AC, LI, &DT))
975 addAssumeNonNull(AC, LI);
977 // Anything using the load now uses the current value.
978 LI->replaceAllUsesWith(V);
979 BB->getInstList().erase(LI);
980 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
981 // Delete this instruction and mark the name as the current holder of the
983 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
987 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
988 if (ai == AllocaLookup.end())
991 // what value were we writing?
992 unsigned AllocaNo = ai->second;
993 IncomingVals[AllocaNo] = SI->getOperand(0);
995 // Record debuginfo for the store before removing it.
996 IncomingLocs[AllocaNo] = SI->getDebugLoc();
997 for (DbgVariableIntrinsic *DII : AllocaDbgDeclares[ai->second])
998 ConvertDebugDeclareToDebugValue(DII, SI, DIB);
999 BB->getInstList().erase(SI);
1003 // 'Recurse' to our successors.
1004 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1008 // Keep track of the successors so we don't visit the same successor twice
1009 SmallPtrSet<BasicBlock *, 8> VisitedSuccs;
1011 // Handle the first successor without using the worklist.
1012 VisitedSuccs.insert(*I);
1018 if (VisitedSuccs.insert(*I).second)
1019 Worklist.emplace_back(*I, Pred, IncomingVals, IncomingLocs);
1024 void llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT,
1025 AssumptionCache *AC) {
1026 // If there is nothing to do, bail out...
1027 if (Allocas.empty())
1030 PromoteMem2Reg(Allocas, DT, AC).run();