1193323Sed//===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===// 2193323Sed// 3353358Sdim// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4353358Sdim// See https://llvm.org/LICENSE.txt for license information. 5353358Sdim// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6193323Sed// 7193323Sed//===----------------------------------------------------------------------===// 8193323Sed// 9193323Sed// This file promotes memory references to be register references. It promotes 10193323Sed// alloca instructions which only have loads and stores as uses. An alloca is 11218893Sdim// transformed by using iterated dominator frontiers to place PHI nodes, then 12218893Sdim// traversing the function in depth-first order to rewrite loads and stores as 13218893Sdim// appropriate. 14193323Sed// 15193323Sed//===----------------------------------------------------------------------===// 16193323Sed 17261991Sdim#include "llvm/ADT/ArrayRef.h" 18193323Sed#include "llvm/ADT/DenseMap.h" 19249423Sdim#include "llvm/ADT/STLExtras.h" 20193323Sed#include "llvm/ADT/SmallPtrSet.h" 21193323Sed#include "llvm/ADT/SmallVector.h" 22193323Sed#include "llvm/ADT/Statistic.h" 23327952Sdim#include "llvm/ADT/TinyPtrVector.h" 24327952Sdim#include "llvm/ADT/Twine.h" 25321369Sdim#include "llvm/Analysis/AssumptionCache.h" 26249423Sdim#include "llvm/Analysis/InstructionSimplify.h" 27288943Sdim#include "llvm/Analysis/IteratedDominanceFrontier.h" 28341825Sdim#include "llvm/Transforms/Utils/Local.h" 29249423Sdim#include "llvm/Analysis/ValueTracking.h" 30327952Sdim#include "llvm/IR/BasicBlock.h" 31276479Sdim#include "llvm/IR/CFG.h" 32327952Sdim#include "llvm/IR/Constant.h" 33249423Sdim#include "llvm/IR/Constants.h" 34276479Sdim#include "llvm/IR/DIBuilder.h" 35249423Sdim#include "llvm/IR/DerivedTypes.h" 36276479Sdim#include "llvm/IR/Dominators.h" 37249423Sdim#include "llvm/IR/Function.h" 38327952Sdim#include "llvm/IR/InstrTypes.h" 39327952Sdim#include "llvm/IR/Instruction.h" 40249423Sdim#include "llvm/IR/Instructions.h" 41249423Sdim#include "llvm/IR/IntrinsicInst.h" 42327952Sdim#include "llvm/IR/Intrinsics.h" 43327952Sdim#include "llvm/IR/LLVMContext.h" 44288943Sdim#include "llvm/IR/Module.h" 45327952Sdim#include "llvm/IR/Type.h" 46327952Sdim#include "llvm/IR/User.h" 47327952Sdim#include "llvm/Support/Casting.h" 48321369Sdim#include "llvm/Transforms/Utils/PromoteMemToReg.h" 49193323Sed#include <algorithm> 50327952Sdim#include <cassert> 51327952Sdim#include <iterator> 52327952Sdim#include <utility> 53327952Sdim#include <vector> 54327952Sdim 55193323Sedusing namespace llvm; 56193323Sed 57276479Sdim#define DEBUG_TYPE "mem2reg" 58276479Sdim 59193323SedSTATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block"); 60193323SedSTATISTIC(NumSingleStore, "Number of alloca's promoted with a single store"); 61193323SedSTATISTIC(NumDeadAlloca, "Number of dead alloca's removed"); 62193323SedSTATISTIC(NumPHIInsert, "Number of PHI nodes inserted"); 63193323Sed 64193323Sedbool llvm::isAllocaPromotable(const AllocaInst *AI) { 65193323Sed // FIXME: If the memory unit is of pointer or integer type, we can permit 66193323Sed // assignments to subsections of the memory unit. 67276479Sdim unsigned AS = AI->getType()->getAddressSpace(); 68193323Sed 69193323Sed // Only allow direct and non-volatile loads and stores... 70276479Sdim for (const User *U : AI->users()) { 71210299Sed if (const LoadInst *LI = dyn_cast<LoadInst>(U)) { 72226633Sdim // Note that atomic loads can be transformed; atomic semantics do 73226633Sdim // not have any meaning for a local alloca. 74193323Sed if (LI->isVolatile()) 75193323Sed return false; 76210299Sed } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) { 77193323Sed if (SI->getOperand(0) == AI) 78261991Sdim return false; // Don't allow a store OF the AI, only INTO the AI. 79226633Sdim // Note that atomic stores can be transformed; atomic semantics do 80226633Sdim // not have any meaning for a local alloca. 81193323Sed if (SI->isVolatile()) 82193323Sed return false; 83224145Sdim } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 84344779Sdim if (!II->isLifetimeStartOrEnd()) 85224145Sdim return false; 86224145Sdim } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 87276479Sdim if (BCI->getType() != Type::getInt8PtrTy(U->getContext(), AS)) 88224145Sdim return false; 89224145Sdim if (!onlyUsedByLifetimeMarkers(BCI)) 90224145Sdim return false; 91224145Sdim } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 92276479Sdim if (GEPI->getType() != Type::getInt8PtrTy(U->getContext(), AS)) 93224145Sdim return false; 94224145Sdim if (!GEPI->hasAllZeroIndices()) 95224145Sdim return false; 96224145Sdim if (!onlyUsedByLifetimeMarkers(GEPI)) 97224145Sdim return false; 98193323Sed } else { 99193323Sed return false; 100193323Sed } 101210299Sed } 102193323Sed 103193323Sed return true; 104193323Sed} 105193323Sed 106193323Sednamespace { 107193323Sed 108261991Sdimstruct AllocaInfo { 109261991Sdim SmallVector<BasicBlock *, 32> DefiningBlocks; 110261991Sdim SmallVector<BasicBlock *, 32> UsingBlocks; 111261991Sdim 112261991Sdim StoreInst *OnlyStore; 113261991Sdim BasicBlock *OnlyBlock; 114261991Sdim bool OnlyUsedInOneBlock; 115261991Sdim 116344779Sdim TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares; 117261991Sdim 118261991Sdim void clear() { 119261991Sdim DefiningBlocks.clear(); 120261991Sdim UsingBlocks.clear(); 121276479Sdim OnlyStore = nullptr; 122276479Sdim OnlyBlock = nullptr; 123261991Sdim OnlyUsedInOneBlock = true; 124327952Sdim DbgDeclares.clear(); 125261991Sdim } 126261991Sdim 127261991Sdim /// Scan the uses of the specified alloca, filling in the AllocaInfo used 128261991Sdim /// by the rest of the pass to reason about the uses of this alloca. 129261991Sdim void AnalyzeAlloca(AllocaInst *AI) { 130261991Sdim clear(); 131261991Sdim 132261991Sdim // As we scan the uses of the alloca instruction, keep track of stores, 133261991Sdim // and decide whether all of the loads and stores to the alloca are within 134261991Sdim // the same basic block. 135276479Sdim for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) { 136261991Sdim Instruction *User = cast<Instruction>(*UI++); 137261991Sdim 138261991Sdim if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 139261991Sdim // Remember the basic blocks which define new values for the alloca 140261991Sdim DefiningBlocks.push_back(SI->getParent()); 141261991Sdim OnlyStore = SI; 142261991Sdim } else { 143261991Sdim LoadInst *LI = cast<LoadInst>(User); 144261991Sdim // Otherwise it must be a load instruction, keep track of variable 145261991Sdim // reads. 146261991Sdim UsingBlocks.push_back(LI->getParent()); 147261991Sdim } 148261991Sdim 149261991Sdim if (OnlyUsedInOneBlock) { 150276479Sdim if (!OnlyBlock) 151261991Sdim OnlyBlock = User->getParent(); 152261991Sdim else if (OnlyBlock != User->getParent()) 153261991Sdim OnlyUsedInOneBlock = false; 154261991Sdim } 155193323Sed } 156261991Sdim 157327952Sdim DbgDeclares = FindDbgAddrUses(AI); 158261991Sdim } 159261991Sdim}; 160261991Sdim 161341825Sdim/// Data package used by RenamePass(). 162341825Sdimstruct RenamePassData { 163327952Sdim using ValVector = std::vector<Value *>; 164341825Sdim using LocationVector = std::vector<DebugLoc>; 165261991Sdim 166341825Sdim RenamePassData(BasicBlock *B, BasicBlock *P, ValVector V, LocationVector L) 167341825Sdim : BB(B), Pred(P), Values(std::move(V)), Locations(std::move(L)) {} 168327952Sdim 169261991Sdim BasicBlock *BB; 170261991Sdim BasicBlock *Pred; 171261991Sdim ValVector Values; 172341825Sdim LocationVector Locations; 173261991Sdim}; 174261991Sdim 175341825Sdim/// This assigns and keeps a per-bb relative ordering of load/store 176261991Sdim/// instructions in the block that directly load or store an alloca. 177261991Sdim/// 178261991Sdim/// This functionality is important because it avoids scanning large basic 179261991Sdim/// blocks multiple times when promoting many allocas in the same block. 180261991Sdimclass LargeBlockInfo { 181341825Sdim /// For each instruction that we track, keep the index of the 182261991Sdim /// instruction. 183193323Sed /// 184261991Sdim /// The index starts out as the number of the instruction from the start of 185261991Sdim /// the block. 186261991Sdim DenseMap<const Instruction *, unsigned> InstNumbers; 187261991Sdim 188261991Sdimpublic: 189261991Sdim 190261991Sdim /// This code only looks at accesses to allocas. 191261991Sdim static bool isInterestingInstruction(const Instruction *I) { 192261991Sdim return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) || 193261991Sdim (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1))); 194261991Sdim } 195261991Sdim 196261991Sdim /// Get or calculate the index of the specified instruction. 197261991Sdim unsigned getInstructionIndex(const Instruction *I) { 198261991Sdim assert(isInterestingInstruction(I) && 199261991Sdim "Not a load/store to/from an alloca?"); 200261991Sdim 201261991Sdim // If we already have this instruction number, return it. 202261991Sdim DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I); 203261991Sdim if (It != InstNumbers.end()) 204193323Sed return It->second; 205193323Sed 206261991Sdim // Scan the whole block to get the instruction. This accumulates 207261991Sdim // information for every interesting instruction in the block, in order to 208261991Sdim // avoid gratuitus rescans. 209261991Sdim const BasicBlock *BB = I->getParent(); 210261991Sdim unsigned InstNo = 0; 211296417Sdim for (const Instruction &BBI : *BB) 212296417Sdim if (isInterestingInstruction(&BBI)) 213296417Sdim InstNumbers[&BBI] = InstNo++; 214261991Sdim It = InstNumbers.find(I); 215193323Sed 216261991Sdim assert(It != InstNumbers.end() && "Didn't insert instruction?"); 217261991Sdim return It->second; 218261991Sdim } 219193323Sed 220261991Sdim void deleteValue(const Instruction *I) { InstNumbers.erase(I); } 221193323Sed 222261991Sdim void clear() { InstNumbers.clear(); } 223261991Sdim}; 224203954Srdivacky 225261991Sdimstruct PromoteMem2Reg { 226261991Sdim /// The alloca instructions being promoted. 227261991Sdim std::vector<AllocaInst *> Allocas; 228327952Sdim 229261991Sdim DominatorTree &DT; 230261991Sdim DIBuilder DIB; 231327952Sdim 232280031Sdim /// A cache of @llvm.assume intrinsics used by SimplifyInstruction. 233280031Sdim AssumptionCache *AC; 234280031Sdim 235321369Sdim const SimplifyQuery SQ; 236327952Sdim 237261991Sdim /// Reverse mapping of Allocas. 238261991Sdim DenseMap<AllocaInst *, unsigned> AllocaLookup; 239218893Sdim 240341825Sdim /// The PhiNodes we're adding. 241261991Sdim /// 242261991Sdim /// That map is used to simplify some Phi nodes as we iterate over it, so 243261991Sdim /// it should have deterministic iterators. We could use a MapVector, but 244261991Sdim /// since we already maintain a map from BasicBlock* to a stable numbering 245261991Sdim /// (BBNumbers), the DenseMap is more efficient (also supports removal). 246261991Sdim DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes; 247193323Sed 248261991Sdim /// For each PHI node, keep track of which entry in Allocas it corresponds 249261991Sdim /// to. 250261991Sdim DenseMap<PHINode *, unsigned> PhiToAllocaMap; 251193323Sed 252261991Sdim /// For each alloca, we keep track of the dbg.declare intrinsic that 253261991Sdim /// describes it, if any, so that we can convert it to a dbg.value 254261991Sdim /// intrinsic if the alloca gets promoted. 255344779Sdim SmallVector<TinyPtrVector<DbgVariableIntrinsic *>, 8> AllocaDbgDeclares; 256193323Sed 257261991Sdim /// The set of basic blocks the renamer has already visited. 258261991Sdim SmallPtrSet<BasicBlock *, 16> Visited; 259193323Sed 260261991Sdim /// Contains a stable numbering of basic blocks to avoid non-determinstic 261261991Sdim /// behavior. 262261991Sdim DenseMap<BasicBlock *, unsigned> BBNumbers; 263193323Sed 264261991Sdim /// Lazily compute the number of predecessors a block has. 265261991Sdim DenseMap<const BasicBlock *, unsigned> BBNumPreds; 266218893Sdim 267261991Sdimpublic: 268261991Sdim PromoteMem2Reg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT, 269321369Sdim AssumptionCache *AC) 270261991Sdim : Allocas(Allocas.begin(), Allocas.end()), DT(DT), 271280031Sdim DIB(*DT.getRoot()->getParent()->getParent(), /*AllowUnresolved*/ false), 272321369Sdim AC(AC), SQ(DT.getRoot()->getParent()->getParent()->getDataLayout(), 273321369Sdim nullptr, &DT, AC) {} 274218893Sdim 275261991Sdim void run(); 276193323Sed 277261991Sdimprivate: 278261991Sdim void RemoveFromAllocasList(unsigned &AllocaIdx) { 279261991Sdim Allocas[AllocaIdx] = Allocas.back(); 280261991Sdim Allocas.pop_back(); 281261991Sdim --AllocaIdx; 282261991Sdim } 283261991Sdim 284261991Sdim unsigned getNumPreds(const BasicBlock *BB) { 285261991Sdim unsigned &NP = BBNumPreds[BB]; 286261991Sdim if (NP == 0) 287341825Sdim NP = pred_size(BB) + 1; 288261991Sdim return NP - 1; 289261991Sdim } 290261991Sdim 291261991Sdim void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info, 292280031Sdim const SmallPtrSetImpl<BasicBlock *> &DefBlocks, 293280031Sdim SmallPtrSetImpl<BasicBlock *> &LiveInBlocks); 294261991Sdim void RenamePass(BasicBlock *BB, BasicBlock *Pred, 295261991Sdim RenamePassData::ValVector &IncVals, 296341825Sdim RenamePassData::LocationVector &IncLocs, 297261991Sdim std::vector<RenamePassData> &Worklist); 298261991Sdim bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version); 299261991Sdim}; 300261991Sdim 301327952Sdim} // end anonymous namespace 302261991Sdim 303321369Sdim/// Given a LoadInst LI this adds assume(LI != null) after it. 304321369Sdimstatic void addAssumeNonNull(AssumptionCache *AC, LoadInst *LI) { 305321369Sdim Function *AssumeIntrinsic = 306321369Sdim Intrinsic::getDeclaration(LI->getModule(), Intrinsic::assume); 307321369Sdim ICmpInst *LoadNotNull = new ICmpInst(ICmpInst::ICMP_NE, LI, 308321369Sdim Constant::getNullValue(LI->getType())); 309321369Sdim LoadNotNull->insertAfter(LI); 310321369Sdim CallInst *CI = CallInst::Create(AssumeIntrinsic, {LoadNotNull}); 311321369Sdim CI->insertAfter(LoadNotNull); 312321369Sdim AC->registerAssumption(CI); 313321369Sdim} 314321369Sdim 315224145Sdimstatic void removeLifetimeIntrinsicUsers(AllocaInst *AI) { 316224145Sdim // Knowing that this alloca is promotable, we know that it's safe to kill all 317224145Sdim // instructions except for load and store. 318193323Sed 319276479Sdim for (auto UI = AI->user_begin(), UE = AI->user_end(); UI != UE;) { 320224145Sdim Instruction *I = cast<Instruction>(*UI); 321224145Sdim ++UI; 322224145Sdim if (isa<LoadInst>(I) || isa<StoreInst>(I)) 323224145Sdim continue; 324224145Sdim 325224145Sdim if (!I->getType()->isVoidTy()) { 326224145Sdim // The only users of this bitcast/GEP instruction are lifetime intrinsics. 327224145Sdim // Follow the use/def chain to erase them now instead of leaving it for 328224145Sdim // dead code elimination later. 329276479Sdim for (auto UUI = I->user_begin(), UUE = I->user_end(); UUI != UUE;) { 330276479Sdim Instruction *Inst = cast<Instruction>(*UUI); 331276479Sdim ++UUI; 332224145Sdim Inst->eraseFromParent(); 333224145Sdim } 334224145Sdim } 335224145Sdim I->eraseFromParent(); 336224145Sdim } 337224145Sdim} 338224145Sdim 339341825Sdim/// Rewrite as many loads as possible given a single store. 340261991Sdim/// 341261991Sdim/// When there is only a single store, we can use the domtree to trivially 342261991Sdim/// replace all of the dominated loads with the stored value. Do so, and return 343261991Sdim/// true if this has successfully promoted the alloca entirely. If this returns 344261991Sdim/// false there were some loads which were not dominated by the single store 345261991Sdim/// and thus must be phi-ed with undef. We fall back to the standard alloca 346261991Sdim/// promotion algorithm in that case. 347261991Sdimstatic bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info, 348327952Sdim LargeBlockInfo &LBI, const DataLayout &DL, 349327952Sdim DominatorTree &DT, AssumptionCache *AC) { 350261991Sdim StoreInst *OnlyStore = Info.OnlyStore; 351261991Sdim bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0)); 352261991Sdim BasicBlock *StoreBB = OnlyStore->getParent(); 353261991Sdim int StoreIndex = -1; 354261991Sdim 355261991Sdim // Clear out UsingBlocks. We will reconstruct it here if needed. 356261991Sdim Info.UsingBlocks.clear(); 357261991Sdim 358276479Sdim for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) { 359261991Sdim Instruction *UserInst = cast<Instruction>(*UI++); 360353358Sdim if (UserInst == OnlyStore) 361261991Sdim continue; 362261991Sdim LoadInst *LI = cast<LoadInst>(UserInst); 363261991Sdim 364261991Sdim // Okay, if we have a load from the alloca, we want to replace it with the 365261991Sdim // only value stored to the alloca. We can do this if the value is 366261991Sdim // dominated by the store. If not, we use the rest of the mem2reg machinery 367261991Sdim // to insert the phi nodes as needed. 368261991Sdim if (!StoringGlobalVal) { // Non-instructions are always dominated. 369261991Sdim if (LI->getParent() == StoreBB) { 370261991Sdim // If we have a use that is in the same block as the store, compare the 371261991Sdim // indices of the two instructions to see which one came first. If the 372261991Sdim // load came before the store, we can't handle it. 373261991Sdim if (StoreIndex == -1) 374261991Sdim StoreIndex = LBI.getInstructionIndex(OnlyStore); 375261991Sdim 376261991Sdim if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) { 377261991Sdim // Can't handle this load, bail out. 378261991Sdim Info.UsingBlocks.push_back(StoreBB); 379261991Sdim continue; 380261991Sdim } 381353358Sdim } else if (!DT.dominates(StoreBB, LI->getParent())) { 382261991Sdim // If the load and store are in different blocks, use BB dominance to 383261991Sdim // check their relationships. If the store doesn't dom the use, bail 384261991Sdim // out. 385261991Sdim Info.UsingBlocks.push_back(LI->getParent()); 386261991Sdim continue; 387261991Sdim } 388261991Sdim } 389261991Sdim 390261991Sdim // Otherwise, we *can* safely rewrite this load. 391261991Sdim Value *ReplVal = OnlyStore->getOperand(0); 392261991Sdim // If the replacement value is the load, this must occur in unreachable 393261991Sdim // code. 394261991Sdim if (ReplVal == LI) 395261991Sdim ReplVal = UndefValue::get(LI->getType()); 396321369Sdim 397321369Sdim // If the load was marked as nonnull we don't want to lose 398321369Sdim // that information when we erase this Load. So we preserve 399321369Sdim // it with an assume. 400321369Sdim if (AC && LI->getMetadata(LLVMContext::MD_nonnull) && 401327952Sdim !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT)) 402321369Sdim addAssumeNonNull(AC, LI); 403321369Sdim 404261991Sdim LI->replaceAllUsesWith(ReplVal); 405261991Sdim LI->eraseFromParent(); 406261991Sdim LBI.deleteValue(LI); 407261991Sdim } 408261991Sdim 409261991Sdim // Finally, after the scan, check to see if the store is all that is left. 410261991Sdim if (!Info.UsingBlocks.empty()) 411261991Sdim return false; // If not, we'll have to fall back for the remainder. 412261991Sdim 413261991Sdim // Record debuginfo for the store and remove the declaration's 414261991Sdim // debuginfo. 415344779Sdim for (DbgVariableIntrinsic *DII : Info.DbgDeclares) { 416296417Sdim DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false); 417327952Sdim ConvertDebugDeclareToDebugValue(DII, Info.OnlyStore, DIB); 418327952Sdim DII->eraseFromParent(); 419261991Sdim } 420261991Sdim // Remove the (now dead) store and alloca. 421261991Sdim Info.OnlyStore->eraseFromParent(); 422261991Sdim LBI.deleteValue(Info.OnlyStore); 423261991Sdim 424261991Sdim AI->eraseFromParent(); 425261991Sdim return true; 426261991Sdim} 427261991Sdim 428261991Sdim/// Many allocas are only used within a single basic block. If this is the 429261991Sdim/// case, avoid traversing the CFG and inserting a lot of potentially useless 430261991Sdim/// PHI nodes by just performing a single linear pass over the basic block 431261991Sdim/// using the Alloca. 432261991Sdim/// 433261991Sdim/// If we cannot promote this alloca (because it is read before it is written), 434296417Sdim/// return false. This is necessary in cases where, due to control flow, the 435296417Sdim/// alloca is undefined only on some control flow paths. e.g. code like 436296417Sdim/// this is correct in LLVM IR: 437296417Sdim/// // A is an alloca with no stores so far 438296417Sdim/// for (...) { 439296417Sdim/// int t = *A; 440296417Sdim/// if (!first_iteration) 441296417Sdim/// use(t); 442296417Sdim/// *A = 42; 443296417Sdim/// } 444296417Sdimstatic bool promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info, 445261991Sdim LargeBlockInfo &LBI, 446327952Sdim const DataLayout &DL, 447321369Sdim DominatorTree &DT, 448321369Sdim AssumptionCache *AC) { 449261991Sdim // The trickiest case to handle is when we have large blocks. Because of this, 450261991Sdim // this code is optimized assuming that large blocks happen. This does not 451261991Sdim // significantly pessimize the small block case. This uses LargeBlockInfo to 452261991Sdim // make it efficient to get the index of various operations in the block. 453261991Sdim 454261991Sdim // Walk the use-def list of the alloca, getting the locations of all stores. 455327952Sdim using StoresByIndexTy = SmallVector<std::pair<unsigned, StoreInst *>, 64>; 456261991Sdim StoresByIndexTy StoresByIndex; 457261991Sdim 458276479Sdim for (User *U : AI->users()) 459276479Sdim if (StoreInst *SI = dyn_cast<StoreInst>(U)) 460261991Sdim StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI)); 461261991Sdim 462261991Sdim // Sort the stores by their index, making it efficient to do a lookup with a 463261991Sdim // binary search. 464344779Sdim llvm::sort(StoresByIndex, less_first()); 465261991Sdim 466261991Sdim // Walk all of the loads from this alloca, replacing them with the nearest 467261991Sdim // store above them, if any. 468276479Sdim for (auto UI = AI->user_begin(), E = AI->user_end(); UI != E;) { 469261991Sdim LoadInst *LI = dyn_cast<LoadInst>(*UI++); 470261991Sdim if (!LI) 471261991Sdim continue; 472261991Sdim 473261991Sdim unsigned LoadIdx = LBI.getInstructionIndex(LI); 474261991Sdim 475261991Sdim // Find the nearest store that has a lower index than this load. 476353358Sdim StoresByIndexTy::iterator I = llvm::lower_bound( 477353358Sdim StoresByIndex, 478353358Sdim std::make_pair(LoadIdx, static_cast<StoreInst *>(nullptr)), 479353358Sdim less_first()); 480296417Sdim if (I == StoresByIndex.begin()) { 481296417Sdim if (StoresByIndex.empty()) 482296417Sdim // If there are no stores, the load takes the undef value. 483296417Sdim LI->replaceAllUsesWith(UndefValue::get(LI->getType())); 484296417Sdim else 485296417Sdim // There is no store before this load, bail out (load may be affected 486296417Sdim // by the following stores - see main comment). 487296417Sdim return false; 488321369Sdim } else { 489261991Sdim // Otherwise, there was a store before this load, the load takes its value. 490321369Sdim // Note, if the load was marked as nonnull we don't want to lose that 491321369Sdim // information when we erase it. So we preserve it with an assume. 492321369Sdim Value *ReplVal = std::prev(I)->second->getOperand(0); 493321369Sdim if (AC && LI->getMetadata(LLVMContext::MD_nonnull) && 494327952Sdim !isKnownNonZero(ReplVal, DL, 0, AC, LI, &DT)) 495321369Sdim addAssumeNonNull(AC, LI); 496261991Sdim 497341825Sdim // If the replacement value is the load, this must occur in unreachable 498341825Sdim // code. 499341825Sdim if (ReplVal == LI) 500341825Sdim ReplVal = UndefValue::get(LI->getType()); 501341825Sdim 502321369Sdim LI->replaceAllUsesWith(ReplVal); 503321369Sdim } 504321369Sdim 505261991Sdim LI->eraseFromParent(); 506261991Sdim LBI.deleteValue(LI); 507261991Sdim } 508261991Sdim 509261991Sdim // Remove the (now dead) stores and alloca. 510261991Sdim while (!AI->use_empty()) { 511276479Sdim StoreInst *SI = cast<StoreInst>(AI->user_back()); 512261991Sdim // Record debuginfo for the store before removing it. 513344779Sdim for (DbgVariableIntrinsic *DII : Info.DbgDeclares) { 514296417Sdim DIBuilder DIB(*AI->getModule(), /*AllowUnresolved*/ false); 515327952Sdim ConvertDebugDeclareToDebugValue(DII, SI, DIB); 516261991Sdim } 517261991Sdim SI->eraseFromParent(); 518261991Sdim LBI.deleteValue(SI); 519261991Sdim } 520261991Sdim 521261991Sdim AI->eraseFromParent(); 522261991Sdim 523261991Sdim // The alloca's debuginfo can be removed as well. 524353358Sdim for (DbgVariableIntrinsic *DII : Info.DbgDeclares) 525327952Sdim DII->eraseFromParent(); 526261991Sdim 527261991Sdim ++NumLocalPromoted; 528296417Sdim return true; 529261991Sdim} 530261991Sdim 531193323Sedvoid PromoteMem2Reg::run() { 532218893Sdim Function &F = *DT.getRoot()->getParent(); 533193323Sed 534203954Srdivacky AllocaDbgDeclares.resize(Allocas.size()); 535193323Sed 536193323Sed AllocaInfo Info; 537193323Sed LargeBlockInfo LBI; 538309124Sdim ForwardIDFCalculator IDF(DT); 539193323Sed 540193323Sed for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) { 541193323Sed AllocaInst *AI = Allocas[AllocaNum]; 542193323Sed 543261991Sdim assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!"); 544193323Sed assert(AI->getParent()->getParent() == &F && 545193323Sed "All allocas should be in the same function, which is same as DF!"); 546193323Sed 547224145Sdim removeLifetimeIntrinsicUsers(AI); 548224145Sdim 549193323Sed if (AI->use_empty()) { 550193323Sed // If there are no uses of the alloca, just delete it now. 551193323Sed AI->eraseFromParent(); 552193323Sed 553193323Sed // Remove the alloca from the Allocas list, since it has been processed 554193323Sed RemoveFromAllocasList(AllocaNum); 555193323Sed ++NumDeadAlloca; 556193323Sed continue; 557193323Sed } 558261991Sdim 559193323Sed // Calculate the set of read and write-locations for each alloca. This is 560193323Sed // analogous to finding the 'uses' and 'definitions' of each variable. 561193323Sed Info.AnalyzeAlloca(AI); 562193323Sed 563193323Sed // If there is only a single store to this value, replace any loads of 564193323Sed // it that are directly dominated by the definition with the value stored. 565193323Sed if (Info.DefiningBlocks.size() == 1) { 566327952Sdim if (rewriteSingleStoreAlloca(AI, Info, LBI, SQ.DL, DT, AC)) { 567193323Sed // The alloca has been processed, move on. 568193323Sed RemoveFromAllocasList(AllocaNum); 569193323Sed ++NumSingleStore; 570193323Sed continue; 571193323Sed } 572193323Sed } 573261991Sdim 574193323Sed // If the alloca is only read and written in one basic block, just perform a 575193323Sed // linear sweep over the block to eliminate it. 576296417Sdim if (Info.OnlyUsedInOneBlock && 577327952Sdim promoteSingleBlockAlloca(AI, Info, LBI, SQ.DL, DT, AC)) { 578261991Sdim // The alloca has been processed, move on. 579261991Sdim RemoveFromAllocasList(AllocaNum); 580261991Sdim continue; 581193323Sed } 582218893Sdim 583193323Sed // If we haven't computed a numbering for the BB's in the function, do so 584193323Sed // now. 585193323Sed if (BBNumbers.empty()) { 586193323Sed unsigned ID = 0; 587288943Sdim for (auto &BB : F) 588288943Sdim BBNumbers[&BB] = ID++; 589193323Sed } 590193323Sed 591203954Srdivacky // Remember the dbg.declare intrinsic describing this alloca, if any. 592327952Sdim if (!Info.DbgDeclares.empty()) 593327952Sdim AllocaDbgDeclares[AllocaNum] = Info.DbgDeclares; 594261991Sdim 595193323Sed // Keep the reverse mapping of the 'Allocas' array for the rename pass. 596193323Sed AllocaLookup[Allocas[AllocaNum]] = AllocaNum; 597193323Sed 598193323Sed // At this point, we're committed to promoting the alloca using IDF's, and 599193323Sed // the standard SSA construction algorithm. Determine which blocks need PHI 600193323Sed // nodes and see if we can optimize out some work by avoiding insertion of 601193323Sed // dead phi nodes. 602288943Sdim 603288943Sdim // Unique the set of defining blocks for efficient lookup. 604353358Sdim SmallPtrSet<BasicBlock *, 32> DefBlocks(Info.DefiningBlocks.begin(), 605353358Sdim Info.DefiningBlocks.end()); 606288943Sdim 607288943Sdim // Determine which blocks the value is live in. These are blocks which lead 608288943Sdim // to uses. 609288943Sdim SmallPtrSet<BasicBlock *, 32> LiveInBlocks; 610288943Sdim ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks); 611288943Sdim 612288943Sdim // At this point, we're committed to promoting the alloca using IDF's, and 613288943Sdim // the standard SSA construction algorithm. Determine which blocks need phi 614288943Sdim // nodes and see if we can optimize out some work by avoiding insertion of 615288943Sdim // dead phi nodes. 616288943Sdim IDF.setLiveInBlocks(LiveInBlocks); 617288943Sdim IDF.setDefiningBlocks(DefBlocks); 618288943Sdim SmallVector<BasicBlock *, 32> PHIBlocks; 619288943Sdim IDF.calculate(PHIBlocks); 620353358Sdim llvm::sort(PHIBlocks, [this](BasicBlock *A, BasicBlock *B) { 621353358Sdim return BBNumbers.find(A)->second < BBNumbers.find(B)->second; 622353358Sdim }); 623288943Sdim 624288943Sdim unsigned CurrentVersion = 0; 625327952Sdim for (BasicBlock *BB : PHIBlocks) 626327952Sdim QueuePhiNode(BB, AllocaNum, CurrentVersion); 627193323Sed } 628193323Sed 629193323Sed if (Allocas.empty()) 630193323Sed return; // All of the allocas must have been trivial! 631193323Sed 632193323Sed LBI.clear(); 633261991Sdim 634193323Sed // Set the incoming values for the basic block to be null values for all of 635193323Sed // the alloca's. We do this in case there is a load of a value that has not 636193323Sed // been stored yet. In this case, it will get this null value. 637193323Sed RenamePassData::ValVector Values(Allocas.size()); 638193323Sed for (unsigned i = 0, e = Allocas.size(); i != e; ++i) 639193323Sed Values[i] = UndefValue::get(Allocas[i]->getAllocatedType()); 640193323Sed 641341825Sdim // When handling debug info, treat all incoming values as if they have unknown 642341825Sdim // locations until proven otherwise. 643341825Sdim RenamePassData::LocationVector Locations(Allocas.size()); 644341825Sdim 645193323Sed // Walks all basic blocks in the function performing the SSA rename algorithm 646193323Sed // and inserting the phi nodes we marked as necessary 647193323Sed std::vector<RenamePassData> RenamePassWorkList; 648341825Sdim RenamePassWorkList.emplace_back(&F.front(), nullptr, std::move(Values), 649341825Sdim std::move(Locations)); 650202375Srdivacky do { 651327952Sdim RenamePassData RPD = std::move(RenamePassWorkList.back()); 652193323Sed RenamePassWorkList.pop_back(); 653193323Sed // RenamePass may add new worklist entries. 654341825Sdim RenamePass(RPD.BB, RPD.Pred, RPD.Values, RPD.Locations, RenamePassWorkList); 655202375Srdivacky } while (!RenamePassWorkList.empty()); 656261991Sdim 657193323Sed // The renamer uses the Visited set to avoid infinite loops. Clear it now. 658193323Sed Visited.clear(); 659193323Sed 660193323Sed // Remove the allocas themselves from the function. 661327952Sdim for (Instruction *A : Allocas) { 662193323Sed // If there are any uses of the alloca instructions left, they must be in 663218893Sdim // unreachable basic blocks that were not processed by walking the dominator 664218893Sdim // tree. Just delete the users now. 665193323Sed if (!A->use_empty()) 666193323Sed A->replaceAllUsesWith(UndefValue::get(A->getType())); 667193323Sed A->eraseFromParent(); 668193323Sed } 669193323Sed 670203954Srdivacky // Remove alloca's dbg.declare instrinsics from the function. 671327952Sdim for (auto &Declares : AllocaDbgDeclares) 672327952Sdim for (auto *DII : Declares) 673327952Sdim DII->eraseFromParent(); 674203954Srdivacky 675193323Sed // Loop over all of the PHI nodes and see if there are any that we can get 676193323Sed // rid of because they merge all of the same incoming values. This can 677193323Sed // happen due to undef values coming into the PHI nodes. This process is 678193323Sed // iterative, because eliminating one PHI node can cause others to be removed. 679193323Sed bool EliminatedAPHI = true; 680193323Sed while (EliminatedAPHI) { 681193323Sed EliminatedAPHI = false; 682261991Sdim 683243830Sdim // Iterating over NewPhiNodes is deterministic, so it is safe to try to 684243830Sdim // simplify and RAUW them as we go. If it was not, we could add uses to 685276479Sdim // the values we replace with in a non-deterministic order, thus creating 686276479Sdim // non-deterministic def->use chains. 687261991Sdim for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator 688261991Sdim I = NewPhiNodes.begin(), 689261991Sdim E = NewPhiNodes.end(); 690261991Sdim I != E;) { 691193323Sed PHINode *PN = I->second; 692218893Sdim 693193323Sed // If this PHI node merges one value and/or undefs, get the value. 694321369Sdim if (Value *V = SimplifyInstruction(PN, SQ)) { 695198090Srdivacky PN->replaceAllUsesWith(V); 696198090Srdivacky PN->eraseFromParent(); 697198090Srdivacky NewPhiNodes.erase(I++); 698198090Srdivacky EliminatedAPHI = true; 699198090Srdivacky continue; 700193323Sed } 701193323Sed ++I; 702193323Sed } 703193323Sed } 704261991Sdim 705193323Sed // At this point, the renamer has added entries to PHI nodes for all reachable 706193323Sed // code. Unfortunately, there may be unreachable blocks which the renamer 707193323Sed // hasn't traversed. If this is the case, the PHI nodes may not 708193323Sed // have incoming values for all predecessors. Loop over all PHI nodes we have 709193323Sed // created, inserting undef values if they are missing any incoming values. 710261991Sdim for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator 711261991Sdim I = NewPhiNodes.begin(), 712261991Sdim E = NewPhiNodes.end(); 713261991Sdim I != E; ++I) { 714193323Sed // We want to do this once per basic block. As such, only process a block 715193323Sed // when we find the PHI that is the first entry in the block. 716193323Sed PHINode *SomePHI = I->second; 717193323Sed BasicBlock *BB = SomePHI->getParent(); 718193323Sed if (&BB->front() != SomePHI) 719193323Sed continue; 720193323Sed 721193323Sed // Only do work here if there the PHI nodes are missing incoming values. We 722193323Sed // know that all PHI nodes that were inserted in a block will have the same 723193323Sed // number of incoming values, so we can just check any of them. 724193323Sed if (SomePHI->getNumIncomingValues() == getNumPreds(BB)) 725193323Sed continue; 726193323Sed 727193323Sed // Get the preds for BB. 728261991Sdim SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB)); 729261991Sdim 730193323Sed // Ok, now we know that all of the PHI nodes are missing entries for some 731193323Sed // basic blocks. Start by sorting the incoming predecessors for efficient 732193323Sed // access. 733344779Sdim auto CompareBBNumbers = [this](BasicBlock *A, BasicBlock *B) { 734353358Sdim return BBNumbers.find(A)->second < BBNumbers.find(B)->second; 735344779Sdim }; 736344779Sdim llvm::sort(Preds, CompareBBNumbers); 737261991Sdim 738193323Sed // Now we loop through all BB's which have entries in SomePHI and remove 739193323Sed // them from the Preds list. 740193323Sed for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) { 741193323Sed // Do a log(n) search of the Preds list for the entry we want. 742353358Sdim SmallVectorImpl<BasicBlock *>::iterator EntIt = llvm::lower_bound( 743353358Sdim Preds, SomePHI->getIncomingBlock(i), CompareBBNumbers); 744261991Sdim assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) && 745193323Sed "PHI node has entry for a block which is not a predecessor!"); 746193323Sed 747193323Sed // Remove the entry 748193323Sed Preds.erase(EntIt); 749193323Sed } 750193323Sed 751193323Sed // At this point, the blocks left in the preds list must have dummy 752193323Sed // entries inserted into every PHI nodes for the block. Update all the phi 753193323Sed // nodes in this block that we are inserting (there could be phis before 754193323Sed // mem2reg runs). 755193323Sed unsigned NumBadPreds = SomePHI->getNumIncomingValues(); 756193323Sed BasicBlock::iterator BBI = BB->begin(); 757193323Sed while ((SomePHI = dyn_cast<PHINode>(BBI++)) && 758193323Sed SomePHI->getNumIncomingValues() == NumBadPreds) { 759193323Sed Value *UndefVal = UndefValue::get(SomePHI->getType()); 760327952Sdim for (BasicBlock *Pred : Preds) 761327952Sdim SomePHI->addIncoming(UndefVal, Pred); 762193323Sed } 763193323Sed } 764261991Sdim 765193323Sed NewPhiNodes.clear(); 766193323Sed} 767193323Sed 768341825Sdim/// Determine which blocks the value is live in. 769261991Sdim/// 770261991Sdim/// These are blocks which lead to uses. Knowing this allows us to avoid 771261991Sdim/// inserting PHI nodes into blocks which don't lead to uses (thus, the 772261991Sdim/// inserted phi nodes would be dead). 773261991Sdimvoid PromoteMem2Reg::ComputeLiveInBlocks( 774261991Sdim AllocaInst *AI, AllocaInfo &Info, 775280031Sdim const SmallPtrSetImpl<BasicBlock *> &DefBlocks, 776280031Sdim SmallPtrSetImpl<BasicBlock *> &LiveInBlocks) { 777193323Sed // To determine liveness, we must iterate through the predecessors of blocks 778193323Sed // where the def is live. Blocks are added to the worklist if we need to 779193323Sed // check their predecessors. Start with all the using blocks. 780261991Sdim SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(), 781261991Sdim Info.UsingBlocks.end()); 782261991Sdim 783193323Sed // If any of the using blocks is also a definition block, check to see if the 784193323Sed // definition occurs before or after the use. If it happens before the use, 785193323Sed // the value isn't really live-in. 786193323Sed for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) { 787193323Sed BasicBlock *BB = LiveInBlockWorklist[i]; 788261991Sdim if (!DefBlocks.count(BB)) 789261991Sdim continue; 790261991Sdim 791193323Sed // Okay, this is a block that both uses and defines the value. If the first 792193323Sed // reference to the alloca is a def (store), then we know it isn't live-in. 793261991Sdim for (BasicBlock::iterator I = BB->begin();; ++I) { 794193323Sed if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 795261991Sdim if (SI->getOperand(1) != AI) 796261991Sdim continue; 797261991Sdim 798193323Sed // We found a store to the alloca before a load. The alloca is not 799193323Sed // actually live-in here. 800193323Sed LiveInBlockWorklist[i] = LiveInBlockWorklist.back(); 801193323Sed LiveInBlockWorklist.pop_back(); 802309124Sdim --i; 803309124Sdim --e; 804193323Sed break; 805198090Srdivacky } 806261991Sdim 807353358Sdim if (LoadInst *LI = dyn_cast<LoadInst>(I)) 808193323Sed // Okay, we found a load before a store to the alloca. It is actually 809193323Sed // live into this block. 810353358Sdim if (LI->getOperand(0) == AI) 811353358Sdim break; 812193323Sed } 813193323Sed } 814261991Sdim 815193323Sed // Now that we have a set of blocks where the phi is live-in, recursively add 816193323Sed // their predecessors until we find the full region the value is live. 817193323Sed while (!LiveInBlockWorklist.empty()) { 818193323Sed BasicBlock *BB = LiveInBlockWorklist.pop_back_val(); 819261991Sdim 820193323Sed // The block really is live in here, insert it into the set. If already in 821193323Sed // the set, then it has already been processed. 822280031Sdim if (!LiveInBlocks.insert(BB).second) 823193323Sed continue; 824261991Sdim 825193323Sed // Since the value is live into BB, it is either defined in a predecessor or 826193323Sed // live into it to. Add the preds to the worklist unless they are a 827193323Sed // defining block. 828327952Sdim for (BasicBlock *P : predecessors(BB)) { 829193323Sed // The value is not live into a predecessor if it defines the value. 830193323Sed if (DefBlocks.count(P)) 831193323Sed continue; 832261991Sdim 833193323Sed // Otherwise it is, add to the worklist. 834193323Sed LiveInBlockWorklist.push_back(P); 835193323Sed } 836193323Sed } 837193323Sed} 838193323Sed 839341825Sdim/// Queue a phi-node to be added to a basic-block for a specific Alloca. 840193323Sed/// 841261991Sdim/// Returns true if there wasn't already a phi-node for that variable 842193323Sedbool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo, 843218893Sdim unsigned &Version) { 844193323Sed // Look up the basic-block in question. 845243830Sdim PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)]; 846193323Sed 847193323Sed // If the BB already has a phi node added for the i'th alloca then we're done! 848261991Sdim if (PN) 849261991Sdim return false; 850193323Sed 851193323Sed // Create a PhiNode using the dereferenced type... and add the phi-node to the 852193323Sed // BasicBlock. 853221345Sdim PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB), 854261991Sdim Allocas[AllocaNo]->getName() + "." + Twine(Version++), 855296417Sdim &BB->front()); 856193323Sed ++NumPHIInsert; 857193323Sed PhiToAllocaMap[PN] = AllocaNo; 858193323Sed return true; 859193323Sed} 860193323Sed 861341825Sdim/// Update the debug location of a phi. \p ApplyMergedLoc indicates whether to 862341825Sdim/// create a merged location incorporating \p DL, or to set \p DL directly. 863341825Sdimstatic void updateForIncomingValueLocation(PHINode *PN, DebugLoc DL, 864341825Sdim bool ApplyMergedLoc) { 865341825Sdim if (ApplyMergedLoc) 866341825Sdim PN->applyMergedLocation(PN->getDebugLoc(), DL); 867341825Sdim else 868341825Sdim PN->setDebugLoc(DL); 869341825Sdim} 870341825Sdim 871341825Sdim/// Recursively traverse the CFG of the function, renaming loads and 872261991Sdim/// stores to the allocas which we are promoting. 873261991Sdim/// 874261991Sdim/// IncomingVals indicates what value each Alloca contains on exit from the 875261991Sdim/// predecessor block Pred. 876193323Sedvoid PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred, 877193323Sed RenamePassData::ValVector &IncomingVals, 878341825Sdim RenamePassData::LocationVector &IncomingLocs, 879193323Sed std::vector<RenamePassData> &Worklist) { 880193323SedNextIteration: 881193323Sed // If we are inserting any phi nodes into this BB, they will already be in the 882193323Sed // block. 883193323Sed if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) { 884193323Sed // If we have PHI nodes to update, compute the number of edges from Pred to 885193323Sed // BB. 886193323Sed if (PhiToAllocaMap.count(APN)) { 887193323Sed // We want to be able to distinguish between PHI nodes being inserted by 888193323Sed // this invocation of mem2reg from those phi nodes that already existed in 889193323Sed // the IR before mem2reg was run. We determine that APN is being inserted 890193323Sed // because it is missing incoming edges. All other PHI nodes being 891193323Sed // inserted by this pass of mem2reg will have the same number of incoming 892193323Sed // operands so far. Remember this count. 893193323Sed unsigned NewPHINumOperands = APN->getNumOperands(); 894261991Sdim 895261991Sdim unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB); 896193323Sed assert(NumEdges && "Must be at least one edge from Pred to BB!"); 897261991Sdim 898193323Sed // Add entries for all the phis. 899193323Sed BasicBlock::iterator PNI = BB->begin(); 900193323Sed do { 901193323Sed unsigned AllocaNo = PhiToAllocaMap[APN]; 902261991Sdim 903341825Sdim // Update the location of the phi node. 904341825Sdim updateForIncomingValueLocation(APN, IncomingLocs[AllocaNo], 905341825Sdim APN->getNumIncomingValues() > 0); 906341825Sdim 907193323Sed // Add N incoming values to the PHI node. 908193323Sed for (unsigned i = 0; i != NumEdges; ++i) 909193323Sed APN->addIncoming(IncomingVals[AllocaNo], Pred); 910261991Sdim 911193323Sed // The currently active variable for this block is now the PHI. 912193323Sed IncomingVals[AllocaNo] = APN; 913344779Sdim for (DbgVariableIntrinsic *DII : AllocaDbgDeclares[AllocaNo]) 914327952Sdim ConvertDebugDeclareToDebugValue(DII, APN, DIB); 915261991Sdim 916193323Sed // Get the next phi node. 917193323Sed ++PNI; 918193323Sed APN = dyn_cast<PHINode>(PNI); 919276479Sdim if (!APN) 920261991Sdim break; 921261991Sdim 922193323Sed // Verify that it is missing entries. If not, it is not being inserted 923193323Sed // by this mem2reg invocation so we want to ignore it. 924193323Sed } while (APN->getNumOperands() == NewPHINumOperands); 925193323Sed } 926193323Sed } 927261991Sdim 928193323Sed // Don't revisit blocks. 929280031Sdim if (!Visited.insert(BB).second) 930261991Sdim return; 931193323Sed 932344779Sdim for (BasicBlock::iterator II = BB->begin(); !II->isTerminator();) { 933296417Sdim Instruction *I = &*II++; // get the instruction, increment iterator 934193323Sed 935193323Sed if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 936193323Sed AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand()); 937261991Sdim if (!Src) 938261991Sdim continue; 939193323Sed 940261991Sdim DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src); 941261991Sdim if (AI == AllocaLookup.end()) 942261991Sdim continue; 943261991Sdim 944193323Sed Value *V = IncomingVals[AI->second]; 945193323Sed 946321369Sdim // If the load was marked as nonnull we don't want to lose 947321369Sdim // that information when we erase this Load. So we preserve 948321369Sdim // it with an assume. 949321369Sdim if (AC && LI->getMetadata(LLVMContext::MD_nonnull) && 950327952Sdim !isKnownNonZero(V, SQ.DL, 0, AC, LI, &DT)) 951321369Sdim addAssumeNonNull(AC, LI); 952321369Sdim 953193323Sed // Anything using the load now uses the current value. 954193323Sed LI->replaceAllUsesWith(V); 955193323Sed BB->getInstList().erase(LI); 956193323Sed } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 957193323Sed // Delete this instruction and mark the name as the current holder of the 958193323Sed // value 959193323Sed AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand()); 960261991Sdim if (!Dest) 961261991Sdim continue; 962261991Sdim 963218893Sdim DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest); 964193323Sed if (ai == AllocaLookup.end()) 965193323Sed continue; 966261991Sdim 967193323Sed // what value were we writing? 968341825Sdim unsigned AllocaNo = ai->second; 969341825Sdim IncomingVals[AllocaNo] = SI->getOperand(0); 970341825Sdim 971202878Srdivacky // Record debuginfo for the store before removing it. 972341825Sdim IncomingLocs[AllocaNo] = SI->getDebugLoc(); 973344779Sdim for (DbgVariableIntrinsic *DII : AllocaDbgDeclares[ai->second]) 974327952Sdim ConvertDebugDeclareToDebugValue(DII, SI, DIB); 975193323Sed BB->getInstList().erase(SI); 976193323Sed } 977193323Sed } 978193323Sed 979193323Sed // 'Recurse' to our successors. 980193323Sed succ_iterator I = succ_begin(BB), E = succ_end(BB); 981261991Sdim if (I == E) 982261991Sdim return; 983193323Sed 984193323Sed // Keep track of the successors so we don't visit the same successor twice 985261991Sdim SmallPtrSet<BasicBlock *, 8> VisitedSuccs; 986193323Sed 987193323Sed // Handle the first successor without using the worklist. 988193323Sed VisitedSuccs.insert(*I); 989193323Sed Pred = BB; 990193323Sed BB = *I; 991193323Sed ++I; 992193323Sed 993193323Sed for (; I != E; ++I) 994280031Sdim if (VisitedSuccs.insert(*I).second) 995341825Sdim Worklist.emplace_back(*I, Pred, IncomingVals, IncomingLocs); 996193323Sed 997193323Sed goto NextIteration; 998193323Sed} 999193323Sed 1000261991Sdimvoid llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT, 1001321369Sdim AssumptionCache *AC) { 1002193323Sed // If there is nothing to do, bail out... 1003261991Sdim if (Allocas.empty()) 1004261991Sdim return; 1005193323Sed 1006321369Sdim PromoteMem2Reg(Allocas, DT, AC).run(); 1007193323Sed} 1008