1193323Sed//===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===// 2193323Sed// 3193323Sed// The LLVM Compiler Infrastructure 4193323Sed// 5193323Sed// This file is distributed under the University of Illinois Open Source 6193323Sed// License. See LICENSE.TXT for details. 7193323Sed// 8193323Sed//===----------------------------------------------------------------------===// 9193323Sed// 10193323Sed// This file promotes memory references to be register references. It promotes 11193323Sed// alloca instructions which only have loads and stores as uses. An alloca is 12218893Sdim// transformed by using iterated dominator frontiers to place PHI nodes, then 13218893Sdim// traversing the function in depth-first order to rewrite loads and stores as 14218893Sdim// appropriate. 15193323Sed// 16218893Sdim// The algorithm used here is based on: 17218893Sdim// 18218893Sdim// Sreedhar and Gao. A linear time algorithm for placing phi-nodes. 19218893Sdim// In Proceedings of the 22nd ACM SIGPLAN-SIGACT Symposium on Principles of 20218893Sdim// Programming Languages 21218893Sdim// POPL '95. ACM, New York, NY, 62-73. 22218893Sdim// 23218893Sdim// It has been modified to not explicitly use the DJ graph data structure and to 24218893Sdim// directly compute pruned SSA using per-variable liveness information. 25218893Sdim// 26193323Sed//===----------------------------------------------------------------------===// 27193323Sed 28193323Sed#define DEBUG_TYPE "mem2reg" 29193323Sed#include "llvm/Transforms/Utils/PromoteMemToReg.h" 30263508Sdim#include "llvm/ADT/ArrayRef.h" 31193323Sed#include "llvm/ADT/DenseMap.h" 32249423Sdim#include "llvm/ADT/STLExtras.h" 33193323Sed#include "llvm/ADT/SmallPtrSet.h" 34193323Sed#include "llvm/ADT/SmallVector.h" 35193323Sed#include "llvm/ADT/Statistic.h" 36249423Sdim#include "llvm/Analysis/AliasSetTracker.h" 37249423Sdim#include "llvm/Analysis/Dominators.h" 38249423Sdim#include "llvm/Analysis/InstructionSimplify.h" 39249423Sdim#include "llvm/Analysis/ValueTracking.h" 40249423Sdim#include "llvm/DIBuilder.h" 41249423Sdim#include "llvm/DebugInfo.h" 42249423Sdim#include "llvm/IR/Constants.h" 43249423Sdim#include "llvm/IR/DerivedTypes.h" 44249423Sdim#include "llvm/IR/Function.h" 45249423Sdim#include "llvm/IR/Instructions.h" 46249423Sdim#include "llvm/IR/IntrinsicInst.h" 47249423Sdim#include "llvm/IR/Metadata.h" 48193323Sed#include "llvm/Support/CFG.h" 49249423Sdim#include "llvm/Transforms/Utils/Local.h" 50193323Sed#include <algorithm> 51218893Sdim#include <queue> 52193323Sedusing namespace llvm; 53193323Sed 54193323SedSTATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block"); 55193323SedSTATISTIC(NumSingleStore, "Number of alloca's promoted with a single store"); 56193323SedSTATISTIC(NumDeadAlloca, "Number of dead alloca's removed"); 57193323SedSTATISTIC(NumPHIInsert, "Number of PHI nodes inserted"); 58193323Sed 59193323Sedbool llvm::isAllocaPromotable(const AllocaInst *AI) { 60193323Sed // FIXME: If the memory unit is of pointer or integer type, we can permit 61193323Sed // assignments to subsections of the memory unit. 62193323Sed 63193323Sed // Only allow direct and non-volatile loads and stores... 64206083Srdivacky for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end(); 65263508Sdim UI != UE; ++UI) { // Loop over all of the uses of the alloca 66210299Sed const User *U = *UI; 67210299Sed if (const LoadInst *LI = dyn_cast<LoadInst>(U)) { 68226633Sdim // Note that atomic loads can be transformed; atomic semantics do 69226633Sdim // not have any meaning for a local alloca. 70193323Sed if (LI->isVolatile()) 71193323Sed return false; 72210299Sed } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) { 73193323Sed if (SI->getOperand(0) == AI) 74263508Sdim return false; // Don't allow a store OF the AI, only INTO the AI. 75226633Sdim // Note that atomic stores can be transformed; atomic semantics do 76226633Sdim // not have any meaning for a local alloca. 77193323Sed if (SI->isVolatile()) 78193323Sed return false; 79224145Sdim } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 80224145Sdim if (II->getIntrinsicID() != Intrinsic::lifetime_start && 81224145Sdim II->getIntrinsicID() != Intrinsic::lifetime_end) 82224145Sdim return false; 83224145Sdim } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 84224145Sdim if (BCI->getType() != Type::getInt8PtrTy(U->getContext())) 85224145Sdim return false; 86224145Sdim if (!onlyUsedByLifetimeMarkers(BCI)) 87224145Sdim return false; 88224145Sdim } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { 89224145Sdim if (GEPI->getType() != Type::getInt8PtrTy(U->getContext())) 90224145Sdim return false; 91224145Sdim if (!GEPI->hasAllZeroIndices()) 92224145Sdim return false; 93224145Sdim if (!onlyUsedByLifetimeMarkers(GEPI)) 94224145Sdim return false; 95193323Sed } else { 96193323Sed return false; 97193323Sed } 98210299Sed } 99193323Sed 100193323Sed return true; 101193323Sed} 102193323Sed 103193323Sednamespace { 104193323Sed 105263508Sdimstruct AllocaInfo { 106263508Sdim SmallVector<BasicBlock *, 32> DefiningBlocks; 107263508Sdim SmallVector<BasicBlock *, 32> UsingBlocks; 108263508Sdim 109263508Sdim StoreInst *OnlyStore; 110263508Sdim BasicBlock *OnlyBlock; 111263508Sdim bool OnlyUsedInOneBlock; 112263508Sdim 113263508Sdim Value *AllocaPointerVal; 114263508Sdim DbgDeclareInst *DbgDeclare; 115263508Sdim 116263508Sdim void clear() { 117263508Sdim DefiningBlocks.clear(); 118263508Sdim UsingBlocks.clear(); 119263508Sdim OnlyStore = 0; 120263508Sdim OnlyBlock = 0; 121263508Sdim OnlyUsedInOneBlock = true; 122263508Sdim AllocaPointerVal = 0; 123263508Sdim DbgDeclare = 0; 124263508Sdim } 125263508Sdim 126263508Sdim /// Scan the uses of the specified alloca, filling in the AllocaInfo used 127263508Sdim /// by the rest of the pass to reason about the uses of this alloca. 128263508Sdim void AnalyzeAlloca(AllocaInst *AI) { 129263508Sdim clear(); 130263508Sdim 131263508Sdim // As we scan the uses of the alloca instruction, keep track of stores, 132263508Sdim // and decide whether all of the loads and stores to the alloca are within 133263508Sdim // the same basic block. 134263508Sdim for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 135263508Sdim UI != E;) { 136263508Sdim Instruction *User = cast<Instruction>(*UI++); 137263508Sdim 138263508Sdim if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 139263508Sdim // Remember the basic blocks which define new values for the alloca 140263508Sdim DefiningBlocks.push_back(SI->getParent()); 141263508Sdim AllocaPointerVal = SI->getOperand(0); 142263508Sdim OnlyStore = SI; 143263508Sdim } else { 144263508Sdim LoadInst *LI = cast<LoadInst>(User); 145263508Sdim // Otherwise it must be a load instruction, keep track of variable 146263508Sdim // reads. 147263508Sdim UsingBlocks.push_back(LI->getParent()); 148263508Sdim AllocaPointerVal = LI; 149263508Sdim } 150263508Sdim 151263508Sdim if (OnlyUsedInOneBlock) { 152263508Sdim if (OnlyBlock == 0) 153263508Sdim OnlyBlock = User->getParent(); 154263508Sdim else if (OnlyBlock != User->getParent()) 155263508Sdim OnlyUsedInOneBlock = false; 156263508Sdim } 157193323Sed } 158263508Sdim 159263508Sdim DbgDeclare = FindAllocaDbgDeclare(AI); 160263508Sdim } 161263508Sdim}; 162263508Sdim 163263508Sdim// Data package used by RenamePass() 164263508Sdimclass RenamePassData { 165263508Sdimpublic: 166263508Sdim typedef std::vector<Value *> ValVector; 167263508Sdim 168263508Sdim RenamePassData() : BB(NULL), Pred(NULL), Values() {} 169263508Sdim RenamePassData(BasicBlock *B, BasicBlock *P, const ValVector &V) 170263508Sdim : BB(B), Pred(P), Values(V) {} 171263508Sdim BasicBlock *BB; 172263508Sdim BasicBlock *Pred; 173263508Sdim ValVector Values; 174263508Sdim 175263508Sdim void swap(RenamePassData &RHS) { 176263508Sdim std::swap(BB, RHS.BB); 177263508Sdim std::swap(Pred, RHS.Pred); 178263508Sdim Values.swap(RHS.Values); 179263508Sdim } 180263508Sdim}; 181263508Sdim 182263508Sdim/// \brief This assigns and keeps a per-bb relative ordering of load/store 183263508Sdim/// instructions in the block that directly load or store an alloca. 184263508Sdim/// 185263508Sdim/// This functionality is important because it avoids scanning large basic 186263508Sdim/// blocks multiple times when promoting many allocas in the same block. 187263508Sdimclass LargeBlockInfo { 188263508Sdim /// \brief For each instruction that we track, keep the index of the 189263508Sdim /// instruction. 190193323Sed /// 191263508Sdim /// The index starts out as the number of the instruction from the start of 192263508Sdim /// the block. 193263508Sdim DenseMap<const Instruction *, unsigned> InstNumbers; 194263508Sdim 195263508Sdimpublic: 196263508Sdim 197263508Sdim /// This code only looks at accesses to allocas. 198263508Sdim static bool isInterestingInstruction(const Instruction *I) { 199263508Sdim return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) || 200263508Sdim (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1))); 201263508Sdim } 202263508Sdim 203263508Sdim /// Get or calculate the index of the specified instruction. 204263508Sdim unsigned getInstructionIndex(const Instruction *I) { 205263508Sdim assert(isInterestingInstruction(I) && 206263508Sdim "Not a load/store to/from an alloca?"); 207263508Sdim 208263508Sdim // If we already have this instruction number, return it. 209263508Sdim DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I); 210263508Sdim if (It != InstNumbers.end()) 211193323Sed return It->second; 212193323Sed 213263508Sdim // Scan the whole block to get the instruction. This accumulates 214263508Sdim // information for every interesting instruction in the block, in order to 215263508Sdim // avoid gratuitus rescans. 216263508Sdim const BasicBlock *BB = I->getParent(); 217263508Sdim unsigned InstNo = 0; 218263508Sdim for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end(); BBI != E; 219263508Sdim ++BBI) 220263508Sdim if (isInterestingInstruction(BBI)) 221263508Sdim InstNumbers[BBI] = InstNo++; 222263508Sdim It = InstNumbers.find(I); 223193323Sed 224263508Sdim assert(It != InstNumbers.end() && "Didn't insert instruction?"); 225263508Sdim return It->second; 226263508Sdim } 227193323Sed 228263508Sdim void deleteValue(const Instruction *I) { InstNumbers.erase(I); } 229193323Sed 230263508Sdim void clear() { InstNumbers.clear(); } 231263508Sdim}; 232203954Srdivacky 233263508Sdimstruct PromoteMem2Reg { 234263508Sdim /// The alloca instructions being promoted. 235263508Sdim std::vector<AllocaInst *> Allocas; 236263508Sdim DominatorTree &DT; 237263508Sdim DIBuilder DIB; 238193323Sed 239263508Sdim /// An AliasSetTracker object to update. If null, don't update it. 240263508Sdim AliasSetTracker *AST; 241193323Sed 242263508Sdim /// Reverse mapping of Allocas. 243263508Sdim DenseMap<AllocaInst *, unsigned> AllocaLookup; 244218893Sdim 245263508Sdim /// \brief The PhiNodes we're adding. 246263508Sdim /// 247263508Sdim /// That map is used to simplify some Phi nodes as we iterate over it, so 248263508Sdim /// it should have deterministic iterators. We could use a MapVector, but 249263508Sdim /// since we already maintain a map from BasicBlock* to a stable numbering 250263508Sdim /// (BBNumbers), the DenseMap is more efficient (also supports removal). 251263508Sdim DenseMap<std::pair<unsigned, unsigned>, PHINode *> NewPhiNodes; 252193323Sed 253263508Sdim /// For each PHI node, keep track of which entry in Allocas it corresponds 254263508Sdim /// to. 255263508Sdim DenseMap<PHINode *, unsigned> PhiToAllocaMap; 256193323Sed 257263508Sdim /// If we are updating an AliasSetTracker, then for each alloca that is of 258263508Sdim /// pointer type, we keep track of what to copyValue to the inserted PHI 259263508Sdim /// nodes here. 260263508Sdim std::vector<Value *> PointerAllocaValues; 261193323Sed 262263508Sdim /// For each alloca, we keep track of the dbg.declare intrinsic that 263263508Sdim /// describes it, if any, so that we can convert it to a dbg.value 264263508Sdim /// intrinsic if the alloca gets promoted. 265263508Sdim SmallVector<DbgDeclareInst *, 8> AllocaDbgDeclares; 266193323Sed 267263508Sdim /// The set of basic blocks the renamer has already visited. 268263508Sdim /// 269263508Sdim SmallPtrSet<BasicBlock *, 16> Visited; 270193323Sed 271263508Sdim /// Contains a stable numbering of basic blocks to avoid non-determinstic 272263508Sdim /// behavior. 273263508Sdim DenseMap<BasicBlock *, unsigned> BBNumbers; 274193323Sed 275263508Sdim /// Maps DomTreeNodes to their level in the dominator tree. 276263508Sdim DenseMap<DomTreeNode *, unsigned> DomLevels; 277202878Srdivacky 278263508Sdim /// Lazily compute the number of predecessors a block has. 279263508Sdim DenseMap<const BasicBlock *, unsigned> BBNumPreds; 280218893Sdim 281263508Sdimpublic: 282263508Sdim PromoteMem2Reg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT, 283263508Sdim AliasSetTracker *AST) 284263508Sdim : Allocas(Allocas.begin(), Allocas.end()), DT(DT), 285263508Sdim DIB(*DT.getRoot()->getParent()->getParent()), AST(AST) {} 286218893Sdim 287263508Sdim void run(); 288193323Sed 289263508Sdimprivate: 290263508Sdim void RemoveFromAllocasList(unsigned &AllocaIdx) { 291263508Sdim Allocas[AllocaIdx] = Allocas.back(); 292263508Sdim Allocas.pop_back(); 293263508Sdim --AllocaIdx; 294263508Sdim } 295263508Sdim 296263508Sdim unsigned getNumPreds(const BasicBlock *BB) { 297263508Sdim unsigned &NP = BBNumPreds[BB]; 298263508Sdim if (NP == 0) 299263508Sdim NP = std::distance(pred_begin(BB), pred_end(BB)) + 1; 300263508Sdim return NP - 1; 301263508Sdim } 302263508Sdim 303263508Sdim void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum, 304263508Sdim AllocaInfo &Info); 305263508Sdim void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info, 306263508Sdim const SmallPtrSet<BasicBlock *, 32> &DefBlocks, 307263508Sdim SmallPtrSet<BasicBlock *, 32> &LiveInBlocks); 308263508Sdim void RenamePass(BasicBlock *BB, BasicBlock *Pred, 309263508Sdim RenamePassData::ValVector &IncVals, 310263508Sdim std::vector<RenamePassData> &Worklist); 311263508Sdim bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version); 312263508Sdim}; 313263508Sdim 314263508Sdim} // end of anonymous namespace 315263508Sdim 316224145Sdimstatic void removeLifetimeIntrinsicUsers(AllocaInst *AI) { 317224145Sdim // Knowing that this alloca is promotable, we know that it's safe to kill all 318224145Sdim // instructions except for load and store. 319193323Sed 320224145Sdim for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 321224145Sdim UI != UE;) { 322224145Sdim Instruction *I = cast<Instruction>(*UI); 323224145Sdim ++UI; 324224145Sdim if (isa<LoadInst>(I) || isa<StoreInst>(I)) 325224145Sdim continue; 326224145Sdim 327224145Sdim if (!I->getType()->isVoidTy()) { 328224145Sdim // The only users of this bitcast/GEP instruction are lifetime intrinsics. 329224145Sdim // Follow the use/def chain to erase them now instead of leaving it for 330224145Sdim // dead code elimination later. 331224145Sdim for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); 332224145Sdim UI != UE;) { 333224145Sdim Instruction *Inst = cast<Instruction>(*UI); 334224145Sdim ++UI; 335224145Sdim Inst->eraseFromParent(); 336224145Sdim } 337224145Sdim } 338224145Sdim I->eraseFromParent(); 339224145Sdim } 340224145Sdim} 341224145Sdim 342263508Sdim/// \brief Rewrite as many loads as possible given a single store. 343263508Sdim/// 344263508Sdim/// When there is only a single store, we can use the domtree to trivially 345263508Sdim/// replace all of the dominated loads with the stored value. Do so, and return 346263508Sdim/// true if this has successfully promoted the alloca entirely. If this returns 347263508Sdim/// false there were some loads which were not dominated by the single store 348263508Sdim/// and thus must be phi-ed with undef. We fall back to the standard alloca 349263508Sdim/// promotion algorithm in that case. 350263508Sdimstatic bool rewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info, 351263508Sdim LargeBlockInfo &LBI, 352263508Sdim DominatorTree &DT, 353263508Sdim AliasSetTracker *AST) { 354263508Sdim StoreInst *OnlyStore = Info.OnlyStore; 355263508Sdim bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0)); 356263508Sdim BasicBlock *StoreBB = OnlyStore->getParent(); 357263508Sdim int StoreIndex = -1; 358263508Sdim 359263508Sdim // Clear out UsingBlocks. We will reconstruct it here if needed. 360263508Sdim Info.UsingBlocks.clear(); 361263508Sdim 362263508Sdim for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) { 363263508Sdim Instruction *UserInst = cast<Instruction>(*UI++); 364263508Sdim if (!isa<LoadInst>(UserInst)) { 365263508Sdim assert(UserInst == OnlyStore && "Should only have load/stores"); 366263508Sdim continue; 367263508Sdim } 368263508Sdim LoadInst *LI = cast<LoadInst>(UserInst); 369263508Sdim 370263508Sdim // Okay, if we have a load from the alloca, we want to replace it with the 371263508Sdim // only value stored to the alloca. We can do this if the value is 372263508Sdim // dominated by the store. If not, we use the rest of the mem2reg machinery 373263508Sdim // to insert the phi nodes as needed. 374263508Sdim if (!StoringGlobalVal) { // Non-instructions are always dominated. 375263508Sdim if (LI->getParent() == StoreBB) { 376263508Sdim // If we have a use that is in the same block as the store, compare the 377263508Sdim // indices of the two instructions to see which one came first. If the 378263508Sdim // load came before the store, we can't handle it. 379263508Sdim if (StoreIndex == -1) 380263508Sdim StoreIndex = LBI.getInstructionIndex(OnlyStore); 381263508Sdim 382263508Sdim if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) { 383263508Sdim // Can't handle this load, bail out. 384263508Sdim Info.UsingBlocks.push_back(StoreBB); 385263508Sdim continue; 386263508Sdim } 387263508Sdim 388263508Sdim } else if (LI->getParent() != StoreBB && 389263508Sdim !DT.dominates(StoreBB, LI->getParent())) { 390263508Sdim // If the load and store are in different blocks, use BB dominance to 391263508Sdim // check their relationships. If the store doesn't dom the use, bail 392263508Sdim // out. 393263508Sdim Info.UsingBlocks.push_back(LI->getParent()); 394263508Sdim continue; 395263508Sdim } 396263508Sdim } 397263508Sdim 398263508Sdim // Otherwise, we *can* safely rewrite this load. 399263508Sdim Value *ReplVal = OnlyStore->getOperand(0); 400263508Sdim // If the replacement value is the load, this must occur in unreachable 401263508Sdim // code. 402263508Sdim if (ReplVal == LI) 403263508Sdim ReplVal = UndefValue::get(LI->getType()); 404263508Sdim LI->replaceAllUsesWith(ReplVal); 405263508Sdim if (AST && LI->getType()->isPointerTy()) 406263508Sdim AST->deleteValue(LI); 407263508Sdim LI->eraseFromParent(); 408263508Sdim LBI.deleteValue(LI); 409263508Sdim } 410263508Sdim 411263508Sdim // Finally, after the scan, check to see if the store is all that is left. 412263508Sdim if (!Info.UsingBlocks.empty()) 413263508Sdim return false; // If not, we'll have to fall back for the remainder. 414263508Sdim 415263508Sdim // Record debuginfo for the store and remove the declaration's 416263508Sdim // debuginfo. 417263508Sdim if (DbgDeclareInst *DDI = Info.DbgDeclare) { 418263508Sdim DIBuilder DIB(*AI->getParent()->getParent()->getParent()); 419263508Sdim ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, DIB); 420263508Sdim DDI->eraseFromParent(); 421263508Sdim LBI.deleteValue(DDI); 422263508Sdim } 423263508Sdim // Remove the (now dead) store and alloca. 424263508Sdim Info.OnlyStore->eraseFromParent(); 425263508Sdim LBI.deleteValue(Info.OnlyStore); 426263508Sdim 427263508Sdim if (AST) 428263508Sdim AST->deleteValue(AI); 429263508Sdim AI->eraseFromParent(); 430263508Sdim LBI.deleteValue(AI); 431263508Sdim return true; 432263508Sdim} 433263508Sdim 434263508Sdim/// Many allocas are only used within a single basic block. If this is the 435263508Sdim/// case, avoid traversing the CFG and inserting a lot of potentially useless 436263508Sdim/// PHI nodes by just performing a single linear pass over the basic block 437263508Sdim/// using the Alloca. 438263508Sdim/// 439263508Sdim/// If we cannot promote this alloca (because it is read before it is written), 440263508Sdim/// return true. This is necessary in cases where, due to control flow, the 441263508Sdim/// alloca is potentially undefined on some control flow paths. e.g. code like 442263508Sdim/// this is potentially correct: 443263508Sdim/// 444263508Sdim/// for (...) { if (c) { A = undef; undef = B; } } 445263508Sdim/// 446263508Sdim/// ... so long as A is not used before undef is set. 447263508Sdimstatic void promoteSingleBlockAlloca(AllocaInst *AI, const AllocaInfo &Info, 448263508Sdim LargeBlockInfo &LBI, 449263508Sdim AliasSetTracker *AST) { 450263508Sdim // The trickiest case to handle is when we have large blocks. Because of this, 451263508Sdim // this code is optimized assuming that large blocks happen. This does not 452263508Sdim // significantly pessimize the small block case. This uses LargeBlockInfo to 453263508Sdim // make it efficient to get the index of various operations in the block. 454263508Sdim 455263508Sdim // Walk the use-def list of the alloca, getting the locations of all stores. 456263508Sdim typedef SmallVector<std::pair<unsigned, StoreInst *>, 64> StoresByIndexTy; 457263508Sdim StoresByIndexTy StoresByIndex; 458263508Sdim 459263508Sdim for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; 460263508Sdim ++UI) 461263508Sdim if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) 462263508Sdim StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI)); 463263508Sdim 464263508Sdim // Sort the stores by their index, making it efficient to do a lookup with a 465263508Sdim // binary search. 466263508Sdim std::sort(StoresByIndex.begin(), StoresByIndex.end(), less_first()); 467263508Sdim 468263508Sdim // Walk all of the loads from this alloca, replacing them with the nearest 469263508Sdim // store above them, if any. 470263508Sdim for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) { 471263508Sdim LoadInst *LI = dyn_cast<LoadInst>(*UI++); 472263508Sdim if (!LI) 473263508Sdim continue; 474263508Sdim 475263508Sdim unsigned LoadIdx = LBI.getInstructionIndex(LI); 476263508Sdim 477263508Sdim // Find the nearest store that has a lower index than this load. 478263508Sdim StoresByIndexTy::iterator I = 479263508Sdim std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(), 480263508Sdim std::make_pair(LoadIdx, static_cast<StoreInst *>(0)), 481263508Sdim less_first()); 482263508Sdim 483263508Sdim if (I == StoresByIndex.begin()) 484263508Sdim // If there is no store before this load, the load takes the undef value. 485263508Sdim LI->replaceAllUsesWith(UndefValue::get(LI->getType())); 486263508Sdim else 487263508Sdim // Otherwise, there was a store before this load, the load takes its value. 488263508Sdim LI->replaceAllUsesWith(llvm::prior(I)->second->getOperand(0)); 489263508Sdim 490263508Sdim if (AST && LI->getType()->isPointerTy()) 491263508Sdim AST->deleteValue(LI); 492263508Sdim LI->eraseFromParent(); 493263508Sdim LBI.deleteValue(LI); 494263508Sdim } 495263508Sdim 496263508Sdim // Remove the (now dead) stores and alloca. 497263508Sdim while (!AI->use_empty()) { 498263508Sdim StoreInst *SI = cast<StoreInst>(AI->use_back()); 499263508Sdim // Record debuginfo for the store before removing it. 500263508Sdim if (DbgDeclareInst *DDI = Info.DbgDeclare) { 501263508Sdim DIBuilder DIB(*AI->getParent()->getParent()->getParent()); 502263508Sdim ConvertDebugDeclareToDebugValue(DDI, SI, DIB); 503263508Sdim } 504263508Sdim SI->eraseFromParent(); 505263508Sdim LBI.deleteValue(SI); 506263508Sdim } 507263508Sdim 508263508Sdim if (AST) 509263508Sdim AST->deleteValue(AI); 510263508Sdim AI->eraseFromParent(); 511263508Sdim LBI.deleteValue(AI); 512263508Sdim 513263508Sdim // The alloca's debuginfo can be removed as well. 514263508Sdim if (DbgDeclareInst *DDI = Info.DbgDeclare) { 515263508Sdim DDI->eraseFromParent(); 516263508Sdim LBI.deleteValue(DDI); 517263508Sdim } 518263508Sdim 519263508Sdim ++NumLocalPromoted; 520263508Sdim} 521263508Sdim 522193323Sedvoid PromoteMem2Reg::run() { 523218893Sdim Function &F = *DT.getRoot()->getParent(); 524193323Sed 525263508Sdim if (AST) 526263508Sdim PointerAllocaValues.resize(Allocas.size()); 527203954Srdivacky AllocaDbgDeclares.resize(Allocas.size()); 528193323Sed 529193323Sed AllocaInfo Info; 530193323Sed LargeBlockInfo LBI; 531193323Sed 532193323Sed for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) { 533193323Sed AllocaInst *AI = Allocas[AllocaNum]; 534193323Sed 535263508Sdim assert(isAllocaPromotable(AI) && "Cannot promote non-promotable alloca!"); 536193323Sed assert(AI->getParent()->getParent() == &F && 537193323Sed "All allocas should be in the same function, which is same as DF!"); 538193323Sed 539224145Sdim removeLifetimeIntrinsicUsers(AI); 540224145Sdim 541193323Sed if (AI->use_empty()) { 542193323Sed // If there are no uses of the alloca, just delete it now. 543263508Sdim if (AST) 544263508Sdim AST->deleteValue(AI); 545193323Sed AI->eraseFromParent(); 546193323Sed 547193323Sed // Remove the alloca from the Allocas list, since it has been processed 548193323Sed RemoveFromAllocasList(AllocaNum); 549193323Sed ++NumDeadAlloca; 550193323Sed continue; 551193323Sed } 552263508Sdim 553193323Sed // Calculate the set of read and write-locations for each alloca. This is 554193323Sed // analogous to finding the 'uses' and 'definitions' of each variable. 555193323Sed Info.AnalyzeAlloca(AI); 556193323Sed 557193323Sed // If there is only a single store to this value, replace any loads of 558193323Sed // it that are directly dominated by the definition with the value stored. 559193323Sed if (Info.DefiningBlocks.size() == 1) { 560263508Sdim if (rewriteSingleStoreAlloca(AI, Info, LBI, DT, AST)) { 561193323Sed // The alloca has been processed, move on. 562193323Sed RemoveFromAllocasList(AllocaNum); 563193323Sed ++NumSingleStore; 564193323Sed continue; 565193323Sed } 566193323Sed } 567263508Sdim 568193323Sed // If the alloca is only read and written in one basic block, just perform a 569193323Sed // linear sweep over the block to eliminate it. 570193323Sed if (Info.OnlyUsedInOneBlock) { 571263508Sdim promoteSingleBlockAlloca(AI, Info, LBI, AST); 572203954Srdivacky 573263508Sdim // The alloca has been processed, move on. 574263508Sdim RemoveFromAllocasList(AllocaNum); 575263508Sdim continue; 576193323Sed } 577218893Sdim 578218893Sdim // If we haven't computed dominator tree levels, do so now. 579218893Sdim if (DomLevels.empty()) { 580263508Sdim SmallVector<DomTreeNode *, 32> Worklist; 581218893Sdim 582218893Sdim DomTreeNode *Root = DT.getRootNode(); 583218893Sdim DomLevels[Root] = 0; 584218893Sdim Worklist.push_back(Root); 585218893Sdim 586218893Sdim while (!Worklist.empty()) { 587218893Sdim DomTreeNode *Node = Worklist.pop_back_val(); 588218893Sdim unsigned ChildLevel = DomLevels[Node] + 1; 589218893Sdim for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); 590218893Sdim CI != CE; ++CI) { 591218893Sdim DomLevels[*CI] = ChildLevel; 592218893Sdim Worklist.push_back(*CI); 593218893Sdim } 594218893Sdim } 595218893Sdim } 596218893Sdim 597193323Sed // If we haven't computed a numbering for the BB's in the function, do so 598193323Sed // now. 599193323Sed if (BBNumbers.empty()) { 600193323Sed unsigned ID = 0; 601193323Sed for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) 602193323Sed BBNumbers[I] = ID++; 603193323Sed } 604193323Sed 605193323Sed // If we have an AST to keep updated, remember some pointer value that is 606193323Sed // stored into the alloca. 607193323Sed if (AST) 608193323Sed PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal; 609263508Sdim 610203954Srdivacky // Remember the dbg.declare intrinsic describing this alloca, if any. 611263508Sdim if (Info.DbgDeclare) 612263508Sdim AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare; 613263508Sdim 614193323Sed // Keep the reverse mapping of the 'Allocas' array for the rename pass. 615193323Sed AllocaLookup[Allocas[AllocaNum]] = AllocaNum; 616193323Sed 617193323Sed // At this point, we're committed to promoting the alloca using IDF's, and 618193323Sed // the standard SSA construction algorithm. Determine which blocks need PHI 619193323Sed // nodes and see if we can optimize out some work by avoiding insertion of 620193323Sed // dead phi nodes. 621193323Sed DetermineInsertionPoint(AI, AllocaNum, Info); 622193323Sed } 623193323Sed 624193323Sed if (Allocas.empty()) 625193323Sed return; // All of the allocas must have been trivial! 626193323Sed 627193323Sed LBI.clear(); 628263508Sdim 629193323Sed // Set the incoming values for the basic block to be null values for all of 630193323Sed // the alloca's. We do this in case there is a load of a value that has not 631193323Sed // been stored yet. In this case, it will get this null value. 632193323Sed // 633193323Sed RenamePassData::ValVector Values(Allocas.size()); 634193323Sed for (unsigned i = 0, e = Allocas.size(); i != e; ++i) 635193323Sed Values[i] = UndefValue::get(Allocas[i]->getAllocatedType()); 636193323Sed 637193323Sed // Walks all basic blocks in the function performing the SSA rename algorithm 638193323Sed // and inserting the phi nodes we marked as necessary 639193323Sed // 640193323Sed std::vector<RenamePassData> RenamePassWorkList; 641193323Sed RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values)); 642202375Srdivacky do { 643193323Sed RenamePassData RPD; 644193323Sed RPD.swap(RenamePassWorkList.back()); 645193323Sed RenamePassWorkList.pop_back(); 646193323Sed // RenamePass may add new worklist entries. 647193323Sed RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList); 648202375Srdivacky } while (!RenamePassWorkList.empty()); 649263508Sdim 650193323Sed // The renamer uses the Visited set to avoid infinite loops. Clear it now. 651193323Sed Visited.clear(); 652193323Sed 653193323Sed // Remove the allocas themselves from the function. 654193323Sed for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 655193323Sed Instruction *A = Allocas[i]; 656193323Sed 657193323Sed // If there are any uses of the alloca instructions left, they must be in 658218893Sdim // unreachable basic blocks that were not processed by walking the dominator 659218893Sdim // tree. Just delete the users now. 660193323Sed if (!A->use_empty()) 661193323Sed A->replaceAllUsesWith(UndefValue::get(A->getType())); 662263508Sdim if (AST) 663263508Sdim AST->deleteValue(A); 664193323Sed A->eraseFromParent(); 665193323Sed } 666193323Sed 667203954Srdivacky // Remove alloca's dbg.declare instrinsics from the function. 668203954Srdivacky for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i) 669203954Srdivacky if (DbgDeclareInst *DDI = AllocaDbgDeclares[i]) 670203954Srdivacky DDI->eraseFromParent(); 671203954Srdivacky 672193323Sed // Loop over all of the PHI nodes and see if there are any that we can get 673193323Sed // rid of because they merge all of the same incoming values. This can 674193323Sed // happen due to undef values coming into the PHI nodes. This process is 675193323Sed // iterative, because eliminating one PHI node can cause others to be removed. 676193323Sed bool EliminatedAPHI = true; 677193323Sed while (EliminatedAPHI) { 678193323Sed EliminatedAPHI = false; 679263508Sdim 680243830Sdim // Iterating over NewPhiNodes is deterministic, so it is safe to try to 681243830Sdim // simplify and RAUW them as we go. If it was not, we could add uses to 682243830Sdim // the values we replace with in a non deterministic order, thus creating 683243830Sdim // non deterministic def->use chains. 684263508Sdim for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator 685263508Sdim I = NewPhiNodes.begin(), 686263508Sdim E = NewPhiNodes.end(); 687263508Sdim I != E;) { 688193323Sed PHINode *PN = I->second; 689218893Sdim 690193323Sed // If this PHI node merges one value and/or undefs, get the value. 691234353Sdim if (Value *V = SimplifyInstruction(PN, 0, 0, &DT)) { 692204642Srdivacky if (AST && PN->getType()->isPointerTy()) 693198090Srdivacky AST->deleteValue(PN); 694198090Srdivacky PN->replaceAllUsesWith(V); 695198090Srdivacky PN->eraseFromParent(); 696198090Srdivacky NewPhiNodes.erase(I++); 697198090Srdivacky EliminatedAPHI = true; 698198090Srdivacky continue; 699193323Sed } 700193323Sed ++I; 701193323Sed } 702193323Sed } 703263508Sdim 704193323Sed // At this point, the renamer has added entries to PHI nodes for all reachable 705193323Sed // code. Unfortunately, there may be unreachable blocks which the renamer 706193323Sed // hasn't traversed. If this is the case, the PHI nodes may not 707193323Sed // have incoming values for all predecessors. Loop over all PHI nodes we have 708193323Sed // created, inserting undef values if they are missing any incoming values. 709193323Sed // 710263508Sdim for (DenseMap<std::pair<unsigned, unsigned>, PHINode *>::iterator 711263508Sdim I = NewPhiNodes.begin(), 712263508Sdim E = NewPhiNodes.end(); 713263508Sdim 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. 728263508Sdim SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB)); 729263508Sdim 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. 733193323Sed std::sort(Preds.begin(), Preds.end()); 734263508Sdim 735193323Sed // Now we loop through all BB's which have entries in SomePHI and remove 736193323Sed // them from the Preds list. 737193323Sed for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) { 738193323Sed // Do a log(n) search of the Preds list for the entry we want. 739263508Sdim SmallVectorImpl<BasicBlock *>::iterator EntIt = std::lower_bound( 740263508Sdim Preds.begin(), Preds.end(), SomePHI->getIncomingBlock(i)); 741263508Sdim assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i) && 742193323Sed "PHI node has entry for a block which is not a predecessor!"); 743193323Sed 744193323Sed // Remove the entry 745193323Sed Preds.erase(EntIt); 746193323Sed } 747193323Sed 748193323Sed // At this point, the blocks left in the preds list must have dummy 749193323Sed // entries inserted into every PHI nodes for the block. Update all the phi 750193323Sed // nodes in this block that we are inserting (there could be phis before 751193323Sed // mem2reg runs). 752193323Sed unsigned NumBadPreds = SomePHI->getNumIncomingValues(); 753193323Sed BasicBlock::iterator BBI = BB->begin(); 754193323Sed while ((SomePHI = dyn_cast<PHINode>(BBI++)) && 755193323Sed SomePHI->getNumIncomingValues() == NumBadPreds) { 756193323Sed Value *UndefVal = UndefValue::get(SomePHI->getType()); 757193323Sed for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred) 758193323Sed SomePHI->addIncoming(UndefVal, Preds[pred]); 759193323Sed } 760193323Sed } 761263508Sdim 762193323Sed NewPhiNodes.clear(); 763193323Sed} 764193323Sed 765263508Sdim/// \brief Determine which blocks the value is live in. 766263508Sdim/// 767263508Sdim/// These are blocks which lead to uses. Knowing this allows us to avoid 768263508Sdim/// inserting PHI nodes into blocks which don't lead to uses (thus, the 769263508Sdim/// inserted phi nodes would be dead). 770263508Sdimvoid PromoteMem2Reg::ComputeLiveInBlocks( 771263508Sdim AllocaInst *AI, AllocaInfo &Info, 772263508Sdim const SmallPtrSet<BasicBlock *, 32> &DefBlocks, 773263508Sdim SmallPtrSet<BasicBlock *, 32> &LiveInBlocks) { 774193323Sed 775193323Sed // To determine liveness, we must iterate through the predecessors of blocks 776193323Sed // where the def is live. Blocks are added to the worklist if we need to 777193323Sed // check their predecessors. Start with all the using blocks. 778263508Sdim SmallVector<BasicBlock *, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(), 779263508Sdim Info.UsingBlocks.end()); 780263508Sdim 781193323Sed // If any of the using blocks is also a definition block, check to see if the 782193323Sed // definition occurs before or after the use. If it happens before the use, 783193323Sed // the value isn't really live-in. 784193323Sed for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) { 785193323Sed BasicBlock *BB = LiveInBlockWorklist[i]; 786263508Sdim if (!DefBlocks.count(BB)) 787263508Sdim continue; 788263508Sdim 789193323Sed // Okay, this is a block that both uses and defines the value. If the first 790193323Sed // reference to the alloca is a def (store), then we know it isn't live-in. 791263508Sdim for (BasicBlock::iterator I = BB->begin();; ++I) { 792193323Sed if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 793263508Sdim if (SI->getOperand(1) != AI) 794263508Sdim continue; 795263508Sdim 796193323Sed // We found a store to the alloca before a load. The alloca is not 797193323Sed // actually live-in here. 798193323Sed LiveInBlockWorklist[i] = LiveInBlockWorklist.back(); 799193323Sed LiveInBlockWorklist.pop_back(); 800193323Sed --i, --e; 801193323Sed break; 802198090Srdivacky } 803263508Sdim 804198090Srdivacky if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 805263508Sdim if (LI->getOperand(0) != AI) 806263508Sdim continue; 807263508Sdim 808193323Sed // Okay, we found a load before a store to the alloca. It is actually 809193323Sed // live into this block. 810193323Sed break; 811193323Sed } 812193323Sed } 813193323Sed } 814263508Sdim 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(); 819263508Sdim 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. 822193323Sed if (!LiveInBlocks.insert(BB)) 823193323Sed continue; 824263508Sdim 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. 828193323Sed for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 829193323Sed BasicBlock *P = *PI; 830263508Sdim 831193323Sed // The value is not live into a predecessor if it defines the value. 832193323Sed if (DefBlocks.count(P)) 833193323Sed continue; 834263508Sdim 835193323Sed // Otherwise it is, add to the worklist. 836193323Sed LiveInBlockWorklist.push_back(P); 837193323Sed } 838193323Sed } 839193323Sed} 840193323Sed 841263508Sdim/// At this point, we're committed to promoting the alloca using IDF's, and the 842263508Sdim/// standard SSA construction algorithm. Determine which blocks need phi nodes 843263508Sdim/// and see if we can optimize out some work by avoiding insertion of dead phi 844263508Sdim/// nodes. 845193323Sedvoid PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum, 846193323Sed AllocaInfo &Info) { 847193323Sed // Unique the set of defining blocks for efficient lookup. 848263508Sdim SmallPtrSet<BasicBlock *, 32> DefBlocks; 849193323Sed DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end()); 850193323Sed 851193323Sed // Determine which blocks the value is live in. These are blocks which lead 852193323Sed // to uses. 853263508Sdim SmallPtrSet<BasicBlock *, 32> LiveInBlocks; 854193323Sed ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks); 855193323Sed 856218893Sdim // Use a priority queue keyed on dominator tree level so that inserted nodes 857218893Sdim // are handled from the bottom of the dominator tree upwards. 858263508Sdim typedef std::pair<DomTreeNode *, unsigned> DomTreeNodePair; 859218893Sdim typedef std::priority_queue<DomTreeNodePair, SmallVector<DomTreeNodePair, 32>, 860263508Sdim less_second> IDFPriorityQueue; 861218893Sdim IDFPriorityQueue PQ; 862218893Sdim 863263508Sdim for (SmallPtrSet<BasicBlock *, 32>::const_iterator I = DefBlocks.begin(), 864263508Sdim E = DefBlocks.end(); 865263508Sdim I != E; ++I) { 866218893Sdim if (DomTreeNode *Node = DT.getNode(*I)) 867218893Sdim PQ.push(std::make_pair(Node, DomLevels[Node])); 868218893Sdim } 869218893Sdim 870263508Sdim SmallVector<std::pair<unsigned, BasicBlock *>, 32> DFBlocks; 871263508Sdim SmallPtrSet<DomTreeNode *, 32> Visited; 872263508Sdim SmallVector<DomTreeNode *, 32> Worklist; 873218893Sdim while (!PQ.empty()) { 874218893Sdim DomTreeNodePair RootPair = PQ.top(); 875218893Sdim PQ.pop(); 876218893Sdim DomTreeNode *Root = RootPair.first; 877218893Sdim unsigned RootLevel = RootPair.second; 878218893Sdim 879218893Sdim // Walk all dominator tree children of Root, inspecting their CFG edges with 880218893Sdim // targets elsewhere on the dominator tree. Only targets whose level is at 881218893Sdim // most Root's level are added to the iterated dominance frontier of the 882218893Sdim // definition set. 883218893Sdim 884218893Sdim Worklist.clear(); 885218893Sdim Worklist.push_back(Root); 886218893Sdim 887218893Sdim while (!Worklist.empty()) { 888218893Sdim DomTreeNode *Node = Worklist.pop_back_val(); 889218893Sdim BasicBlock *BB = Node->getBlock(); 890218893Sdim 891218893Sdim for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; 892218893Sdim ++SI) { 893218893Sdim DomTreeNode *SuccNode = DT.getNode(*SI); 894218893Sdim 895218893Sdim // Quickly skip all CFG edges that are also dominator tree edges instead 896218893Sdim // of catching them below. 897218893Sdim if (SuccNode->getIDom() == Node) 898218893Sdim continue; 899218893Sdim 900218893Sdim unsigned SuccLevel = DomLevels[SuccNode]; 901218893Sdim if (SuccLevel > RootLevel) 902218893Sdim continue; 903218893Sdim 904218893Sdim if (!Visited.insert(SuccNode)) 905218893Sdim continue; 906218893Sdim 907218893Sdim BasicBlock *SuccBB = SuccNode->getBlock(); 908218893Sdim if (!LiveInBlocks.count(SuccBB)) 909218893Sdim continue; 910218893Sdim 911218893Sdim DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB)); 912218893Sdim if (!DefBlocks.count(SuccBB)) 913218893Sdim PQ.push(std::make_pair(SuccNode, SuccLevel)); 914218893Sdim } 915218893Sdim 916218893Sdim for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE; 917218893Sdim ++CI) { 918218893Sdim if (!Visited.count(*CI)) 919218893Sdim Worklist.push_back(*CI); 920218893Sdim } 921193323Sed } 922193323Sed } 923218893Sdim 924218893Sdim if (DFBlocks.size() > 1) 925218893Sdim std::sort(DFBlocks.begin(), DFBlocks.end()); 926218893Sdim 927218893Sdim unsigned CurrentVersion = 0; 928218893Sdim for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) 929218893Sdim QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion); 930193323Sed} 931193323Sed 932263508Sdim/// \brief Queue a phi-node to be added to a basic-block for a specific Alloca. 933193323Sed/// 934263508Sdim/// Returns true if there wasn't already a phi-node for that variable 935193323Sedbool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo, 936218893Sdim unsigned &Version) { 937193323Sed // Look up the basic-block in question. 938243830Sdim PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)]; 939193323Sed 940193323Sed // If the BB already has a phi node added for the i'th alloca then we're done! 941263508Sdim if (PN) 942263508Sdim return false; 943193323Sed 944193323Sed // Create a PhiNode using the dereferenced type... and add the phi-node to the 945193323Sed // BasicBlock. 946221345Sdim PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB), 947263508Sdim Allocas[AllocaNo]->getName() + "." + Twine(Version++), 948198090Srdivacky BB->begin()); 949193323Sed ++NumPHIInsert; 950193323Sed PhiToAllocaMap[PN] = AllocaNo; 951193323Sed 952204642Srdivacky if (AST && PN->getType()->isPointerTy()) 953193323Sed AST->copyValue(PointerAllocaValues[AllocaNo], PN); 954193323Sed 955193323Sed return true; 956193323Sed} 957193323Sed 958263508Sdim/// \brief Recursively traverse the CFG of the function, renaming loads and 959263508Sdim/// stores to the allocas which we are promoting. 960263508Sdim/// 961263508Sdim/// IncomingVals indicates what value each Alloca contains on exit from the 962263508Sdim/// predecessor block Pred. 963193323Sedvoid PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred, 964193323Sed RenamePassData::ValVector &IncomingVals, 965193323Sed std::vector<RenamePassData> &Worklist) { 966193323SedNextIteration: 967193323Sed // If we are inserting any phi nodes into this BB, they will already be in the 968193323Sed // block. 969193323Sed if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) { 970193323Sed // If we have PHI nodes to update, compute the number of edges from Pred to 971193323Sed // BB. 972193323Sed if (PhiToAllocaMap.count(APN)) { 973193323Sed // We want to be able to distinguish between PHI nodes being inserted by 974193323Sed // this invocation of mem2reg from those phi nodes that already existed in 975193323Sed // the IR before mem2reg was run. We determine that APN is being inserted 976193323Sed // because it is missing incoming edges. All other PHI nodes being 977193323Sed // inserted by this pass of mem2reg will have the same number of incoming 978193323Sed // operands so far. Remember this count. 979193323Sed unsigned NewPHINumOperands = APN->getNumOperands(); 980263508Sdim 981263508Sdim unsigned NumEdges = std::count(succ_begin(Pred), succ_end(Pred), BB); 982193323Sed assert(NumEdges && "Must be at least one edge from Pred to BB!"); 983263508Sdim 984193323Sed // Add entries for all the phis. 985193323Sed BasicBlock::iterator PNI = BB->begin(); 986193323Sed do { 987193323Sed unsigned AllocaNo = PhiToAllocaMap[APN]; 988263508Sdim 989193323Sed // Add N incoming values to the PHI node. 990193323Sed for (unsigned i = 0; i != NumEdges; ++i) 991193323Sed APN->addIncoming(IncomingVals[AllocaNo], Pred); 992263508Sdim 993193323Sed // The currently active variable for this block is now the PHI. 994193323Sed IncomingVals[AllocaNo] = APN; 995263508Sdim 996193323Sed // Get the next phi node. 997193323Sed ++PNI; 998193323Sed APN = dyn_cast<PHINode>(PNI); 999263508Sdim if (APN == 0) 1000263508Sdim break; 1001263508Sdim 1002193323Sed // Verify that it is missing entries. If not, it is not being inserted 1003193323Sed // by this mem2reg invocation so we want to ignore it. 1004193323Sed } while (APN->getNumOperands() == NewPHINumOperands); 1005193323Sed } 1006193323Sed } 1007263508Sdim 1008193323Sed // Don't revisit blocks. 1009263508Sdim if (!Visited.insert(BB)) 1010263508Sdim return; 1011193323Sed 1012263508Sdim for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II);) { 1013193323Sed Instruction *I = II++; // get the instruction, increment iterator 1014193323Sed 1015193323Sed if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1016193323Sed AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand()); 1017263508Sdim if (!Src) 1018263508Sdim continue; 1019193323Sed 1020263508Sdim DenseMap<AllocaInst *, unsigned>::iterator AI = AllocaLookup.find(Src); 1021263508Sdim if (AI == AllocaLookup.end()) 1022263508Sdim continue; 1023263508Sdim 1024193323Sed Value *V = IncomingVals[AI->second]; 1025193323Sed 1026193323Sed // Anything using the load now uses the current value. 1027193323Sed LI->replaceAllUsesWith(V); 1028204642Srdivacky if (AST && LI->getType()->isPointerTy()) 1029193323Sed AST->deleteValue(LI); 1030193323Sed BB->getInstList().erase(LI); 1031193323Sed } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1032193323Sed // Delete this instruction and mark the name as the current holder of the 1033193323Sed // value 1034193323Sed AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand()); 1035263508Sdim if (!Dest) 1036263508Sdim continue; 1037263508Sdim 1038218893Sdim DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest); 1039193323Sed if (ai == AllocaLookup.end()) 1040193323Sed continue; 1041263508Sdim 1042193323Sed // what value were we writing? 1043193323Sed IncomingVals[ai->second] = SI->getOperand(0); 1044202878Srdivacky // Record debuginfo for the store before removing it. 1045263508Sdim if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second]) 1046263508Sdim ConvertDebugDeclareToDebugValue(DDI, SI, DIB); 1047193323Sed BB->getInstList().erase(SI); 1048193323Sed } 1049193323Sed } 1050193323Sed 1051193323Sed // 'Recurse' to our successors. 1052193323Sed succ_iterator I = succ_begin(BB), E = succ_end(BB); 1053263508Sdim if (I == E) 1054263508Sdim return; 1055193323Sed 1056193323Sed // Keep track of the successors so we don't visit the same successor twice 1057263508Sdim SmallPtrSet<BasicBlock *, 8> VisitedSuccs; 1058193323Sed 1059193323Sed // Handle the first successor without using the worklist. 1060193323Sed VisitedSuccs.insert(*I); 1061193323Sed Pred = BB; 1062193323Sed BB = *I; 1063193323Sed ++I; 1064193323Sed 1065193323Sed for (; I != E; ++I) 1066193323Sed if (VisitedSuccs.insert(*I)) 1067193323Sed Worklist.push_back(RenamePassData(*I, Pred, IncomingVals)); 1068193323Sed 1069193323Sed goto NextIteration; 1070193323Sed} 1071193323Sed 1072263508Sdimvoid llvm::PromoteMemToReg(ArrayRef<AllocaInst *> Allocas, DominatorTree &DT, 1073263508Sdim AliasSetTracker *AST) { 1074193323Sed // If there is nothing to do, bail out... 1075263508Sdim if (Allocas.empty()) 1076263508Sdim return; 1077193323Sed 1078218893Sdim PromoteMem2Reg(Allocas, DT, AST).run(); 1079193323Sed} 1080