PromoteMemoryToRegister.cpp revision fb312c7e449bbb8b780603bf44620be91d6a65bb
1//===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file promote memory references to be register references. It promotes 11// alloca instructions which only have loads and stores as uses. An alloca is 12// transformed by using dominator frontiers to place PHI nodes, then traversing 13// the function in depth-first order to rewrite loads and stores as appropriate. 14// This is just the standard SSA construction algorithm to construct "pruned" 15// SSA form. 16// 17//===----------------------------------------------------------------------===// 18 19#define DEBUG_TYPE "mem2reg" 20#include "llvm/Transforms/Utils/PromoteMemToReg.h" 21#include "llvm/Constants.h" 22#include "llvm/DerivedTypes.h" 23#include "llvm/Function.h" 24#include "llvm/Instructions.h" 25#include "llvm/Analysis/Dominators.h" 26#include "llvm/Analysis/AliasSetTracker.h" 27#include "llvm/ADT/DenseMap.h" 28#include "llvm/ADT/SmallPtrSet.h" 29#include "llvm/ADT/SmallVector.h" 30#include "llvm/ADT/Statistic.h" 31#include "llvm/ADT/StringExtras.h" 32#include "llvm/Support/CFG.h" 33#include "llvm/Support/Compiler.h" 34#include <algorithm> 35using namespace llvm; 36 37STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block"); 38STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store"); 39STATISTIC(NumDeadAlloca, "Number of dead alloca's removed"); 40 41// Provide DenseMapKeyInfo for all pointers. 42namespace llvm { 43template<> 44struct DenseMapKeyInfo<std::pair<BasicBlock*, unsigned> > { 45 static inline std::pair<BasicBlock*, unsigned> getEmptyKey() { 46 return std::make_pair((BasicBlock*)-1, ~0U); 47 } 48 static inline std::pair<BasicBlock*, unsigned> getTombstoneKey() { 49 return std::make_pair((BasicBlock*)-2, 0U); 50 } 51 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) { 52 return DenseMapKeyInfo<void*>::getHashValue(Val.first) + Val.second*2; 53 } 54 static bool isPod() { return true; } 55}; 56} 57 58/// isAllocaPromotable - Return true if this alloca is legal for promotion. 59/// This is true if there are only loads and stores to the alloca. 60/// 61bool llvm::isAllocaPromotable(const AllocaInst *AI) { 62 // FIXME: If the memory unit is of pointer or integer type, we can permit 63 // assignments to subsections of the memory unit. 64 65 // Only allow direct loads and stores... 66 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end(); 67 UI != UE; ++UI) // Loop over all of the uses of the alloca 68 if (isa<LoadInst>(*UI)) { 69 // noop 70 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 71 if (SI->getOperand(0) == AI) 72 return false; // Don't allow a store OF the AI, only INTO the AI. 73 } else { 74 return false; // Not a load or store. 75 } 76 77 return true; 78} 79 80namespace { 81 struct AllocaInfo; 82 83 // Data package used by RenamePass() 84 class VISIBILITY_HIDDEN RenamePassData { 85 public: 86 typedef std::vector<Value *> ValVector; 87 88 RenamePassData() {} 89 RenamePassData(BasicBlock *B, BasicBlock *P, 90 const ValVector &V) : BB(B), Pred(P), Values(V) {} 91 BasicBlock *BB; 92 BasicBlock *Pred; 93 ValVector Values; 94 95 void swap(RenamePassData &RHS) { 96 std::swap(BB, RHS.BB); 97 std::swap(Pred, RHS.Pred); 98 Values.swap(RHS.Values); 99 } 100 }; 101 102 struct VISIBILITY_HIDDEN PromoteMem2Reg { 103 /// Allocas - The alloca instructions being promoted. 104 /// 105 std::vector<AllocaInst*> Allocas; 106 SmallVector<AllocaInst*, 16> &RetryList; 107 DominatorTree &DT; 108 DominanceFrontier &DF; 109 110 /// AST - An AliasSetTracker object to update. If null, don't update it. 111 /// 112 AliasSetTracker *AST; 113 114 /// AllocaLookup - Reverse mapping of Allocas. 115 /// 116 std::map<AllocaInst*, unsigned> AllocaLookup; 117 118 /// NewPhiNodes - The PhiNodes we're adding. 119 /// 120 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes; 121 122 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas 123 /// it corresponds to. 124 DenseMap<PHINode*, unsigned> PhiToAllocaMap; 125 126 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for 127 /// each alloca that is of pointer type, we keep track of what to copyValue 128 /// to the inserted PHI nodes here. 129 /// 130 std::vector<Value*> PointerAllocaValues; 131 132 /// Visited - The set of basic blocks the renamer has already visited. 133 /// 134 SmallPtrSet<BasicBlock*, 16> Visited; 135 136 /// BBNumbers - Contains a stable numbering of basic blocks to avoid 137 /// non-determinstic behavior. 138 DenseMap<BasicBlock*, unsigned> BBNumbers; 139 140 public: 141 PromoteMem2Reg(const std::vector<AllocaInst*> &A, 142 SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt, 143 DominanceFrontier &df, AliasSetTracker *ast) 144 : Allocas(A), RetryList(Retry), DT(dt), DF(df), AST(ast) {} 145 146 void run(); 147 148 /// properlyDominates - Return true if I1 properly dominates I2. 149 /// 150 bool properlyDominates(Instruction *I1, Instruction *I2) const { 151 if (InvokeInst *II = dyn_cast<InvokeInst>(I1)) 152 I1 = II->getNormalDest()->begin(); 153 return DT.properlyDominates(I1->getParent(), I2->getParent()); 154 } 155 156 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree. 157 /// 158 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const { 159 return DT.dominates(BB1, BB2); 160 } 161 162 private: 163 void RemoveFromAllocasList(unsigned &AllocaIdx) { 164 Allocas[AllocaIdx] = Allocas.back(); 165 Allocas.pop_back(); 166 --AllocaIdx; 167 } 168 169 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info); 170 171 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum, 172 SmallPtrSet<PHINode*, 16> &DeadPHINodes); 173 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI); 174 void PromoteLocallyUsedAllocas(BasicBlock *BB, 175 const std::vector<AllocaInst*> &AIs); 176 177 void RenamePass(BasicBlock *BB, BasicBlock *Pred, 178 RenamePassData::ValVector &IncVals, 179 std::vector<RenamePassData> &Worklist); 180 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version, 181 SmallPtrSet<PHINode*, 16> &InsertedPHINodes); 182 }; 183 184 struct AllocaInfo { 185 std::vector<BasicBlock*> DefiningBlocks; 186 std::vector<BasicBlock*> UsingBlocks; 187 188 StoreInst *OnlyStore; 189 BasicBlock *OnlyBlock; 190 bool OnlyUsedInOneBlock; 191 192 Value *AllocaPointerVal; 193 194 void clear() { 195 DefiningBlocks.clear(); 196 UsingBlocks.clear(); 197 OnlyStore = 0; 198 OnlyBlock = 0; 199 OnlyUsedInOneBlock = true; 200 AllocaPointerVal = 0; 201 } 202 203 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our 204 /// ivars. 205 void AnalyzeAlloca(AllocaInst *AI) { 206 clear(); 207 208 // As we scan the uses of the alloca instruction, keep track of stores, 209 // and decide whether all of the loads and stores to the alloca are within 210 // the same basic block. 211 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end(); 212 U != E; ++U){ 213 Instruction *User = cast<Instruction>(*U); 214 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 215 // Remember the basic blocks which define new values for the alloca 216 DefiningBlocks.push_back(SI->getParent()); 217 AllocaPointerVal = SI->getOperand(0); 218 OnlyStore = SI; 219 } else { 220 LoadInst *LI = cast<LoadInst>(User); 221 // Otherwise it must be a load instruction, keep track of variable reads 222 UsingBlocks.push_back(LI->getParent()); 223 AllocaPointerVal = LI; 224 } 225 226 if (OnlyUsedInOneBlock) { 227 if (OnlyBlock == 0) 228 OnlyBlock = User->getParent(); 229 else if (OnlyBlock != User->getParent()) 230 OnlyUsedInOneBlock = false; 231 } 232 } 233 } 234 }; 235 236} // end of anonymous namespace 237 238 239void PromoteMem2Reg::run() { 240 Function &F = *DF.getRoot()->getParent(); 241 242 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are 243 // only used in a single basic block. These instructions can be efficiently 244 // promoted by performing a single linear scan over that one block. Since 245 // individual basic blocks are sometimes large, we group together all allocas 246 // that are live in a single basic block by the basic block they are live in. 247 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas; 248 249 if (AST) PointerAllocaValues.resize(Allocas.size()); 250 251 AllocaInfo Info; 252 253 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) { 254 AllocaInst *AI = Allocas[AllocaNum]; 255 256 assert(isAllocaPromotable(AI) && 257 "Cannot promote non-promotable alloca!"); 258 assert(AI->getParent()->getParent() == &F && 259 "All allocas should be in the same function, which is same as DF!"); 260 261 if (AI->use_empty()) { 262 // If there are no uses of the alloca, just delete it now. 263 if (AST) AST->deleteValue(AI); 264 AI->eraseFromParent(); 265 266 // Remove the alloca from the Allocas list, since it has been processed 267 RemoveFromAllocasList(AllocaNum); 268 ++NumDeadAlloca; 269 continue; 270 } 271 272 // Calculate the set of read and write-locations for each alloca. This is 273 // analogous to finding the 'uses' and 'definitions' of each variable. 274 Info.AnalyzeAlloca(AI); 275 276 // If there is only a single store to this value, replace any loads of 277 // it that are directly dominated by the definition with the value stored. 278 if (Info.DefiningBlocks.size() == 1) { 279 RewriteSingleStoreAlloca(AI, Info); 280 281 // Finally, after the scan, check to see if the store is all that is left. 282 if (Info.UsingBlocks.empty()) { 283 // Remove the (now dead) store and alloca. 284 Info.OnlyStore->eraseFromParent(); 285 if (AST) AST->deleteValue(AI); 286 AI->eraseFromParent(); 287 288 // The alloca has been processed, move on. 289 RemoveFromAllocasList(AllocaNum); 290 291 ++NumSingleStore; 292 continue; 293 } 294 } 295 296 // If the alloca is only read and written in one basic block, just perform a 297 // linear sweep over the block to eliminate it. 298 if (Info.OnlyUsedInOneBlock) { 299 LocallyUsedAllocas[Info.OnlyBlock].push_back(AI); 300 301 // Remove the alloca from the Allocas list, since it will be processed. 302 RemoveFromAllocasList(AllocaNum); 303 continue; 304 } 305 306 if (AST) 307 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal; 308 309 // If we haven't computed a numbering for the BB's in the function, do so 310 // now. 311 if (BBNumbers.empty()) { 312 unsigned ID = 0; 313 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) 314 BBNumbers[I] = ID++; 315 } 316 317 // Compute the locations where PhiNodes need to be inserted. Look at the 318 // dominance frontier of EACH basic-block we have a write in. 319 // 320 unsigned CurrentVersion = 0; 321 SmallPtrSet<PHINode*, 16> InsertedPHINodes; 322 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks; 323 while (!Info.DefiningBlocks.empty()) { 324 BasicBlock *BB = Info.DefiningBlocks.back(); 325 Info.DefiningBlocks.pop_back(); 326 327 // Look up the DF for this write, add it to PhiNodes 328 DominanceFrontier::const_iterator it = DF.find(BB); 329 if (it != DF.end()) { 330 const DominanceFrontier::DomSetType &S = it->second; 331 332 // In theory we don't need the indirection through the DFBlocks vector. 333 // In practice, the order of calling QueuePhiNode would depend on the 334 // (unspecified) ordering of basic blocks in the dominance frontier, 335 // which would give PHI nodes non-determinstic subscripts. Fix this by 336 // processing blocks in order of the occurance in the function. 337 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(), 338 PE = S.end(); P != PE; ++P) 339 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P)); 340 341 // Sort by which the block ordering in the function. 342 std::sort(DFBlocks.begin(), DFBlocks.end()); 343 344 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) { 345 BasicBlock *BB = DFBlocks[i].second; 346 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes)) 347 Info.DefiningBlocks.push_back(BB); 348 } 349 DFBlocks.clear(); 350 } 351 } 352 353 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier 354 // of the writes to the variable, scan through the reads of the variable, 355 // marking PHI nodes which are actually necessary as alive (by removing them 356 // from the InsertedPHINodes set). This is not perfect: there may PHI 357 // marked alive because of loads which are dominated by stores, but there 358 // will be no unmarked PHI nodes which are actually used. 359 // 360 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) 361 MarkDominatingPHILive(Info.UsingBlocks[i], AllocaNum, InsertedPHINodes); 362 Info.UsingBlocks.clear(); 363 364 // If there are any PHI nodes which are now known to be dead, remove them! 365 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(), 366 E = InsertedPHINodes.end(); I != E; ++I) { 367 PHINode *PN = *I; 368 bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum)); 369 Erased=Erased; 370 assert(Erased && "PHI already removed?"); 371 372 if (AST && isa<PointerType>(PN->getType())) 373 AST->deleteValue(PN); 374 PN->eraseFromParent(); 375 PhiToAllocaMap.erase(PN); 376 } 377 378 // Keep the reverse mapping of the 'Allocas' array. 379 AllocaLookup[Allocas[AllocaNum]] = AllocaNum; 380 } 381 382 // Process all allocas which are only used in a single basic block. 383 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I = 384 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){ 385 const std::vector<AllocaInst*> &LocAllocas = I->second; 386 assert(!LocAllocas.empty() && "empty alloca list??"); 387 388 // It's common for there to only be one alloca in the list. Handle it 389 // efficiently. 390 if (LocAllocas.size() == 1) { 391 // If we can do the quick promotion pass, do so now. 392 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0])) 393 RetryList.push_back(LocAllocas[0]); // Failed, retry later. 394 } else { 395 // Locally promote anything possible. Note that if this is unable to 396 // promote a particular alloca, it puts the alloca onto the Allocas vector 397 // for global processing. 398 PromoteLocallyUsedAllocas(I->first, LocAllocas); 399 } 400 } 401 402 if (Allocas.empty()) 403 return; // All of the allocas must have been trivial! 404 405 // Set the incoming values for the basic block to be null values for all of 406 // the alloca's. We do this in case there is a load of a value that has not 407 // been stored yet. In this case, it will get this null value. 408 // 409 RenamePassData::ValVector Values(Allocas.size()); 410 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) 411 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType()); 412 413 // Walks all basic blocks in the function performing the SSA rename algorithm 414 // and inserting the phi nodes we marked as necessary 415 // 416 std::vector<RenamePassData> RenamePassWorkList; 417 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values)); 418 while (!RenamePassWorkList.empty()) { 419 RenamePassData RPD; 420 RPD.swap(RenamePassWorkList.back()); 421 RenamePassWorkList.pop_back(); 422 // RenamePass may add new worklist entries. 423 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList); 424 } 425 426 // The renamer uses the Visited set to avoid infinite loops. Clear it now. 427 Visited.clear(); 428 429 // Remove the allocas themselves from the function. 430 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 431 Instruction *A = Allocas[i]; 432 433 // If there are any uses of the alloca instructions left, they must be in 434 // sections of dead code that were not processed on the dominance frontier. 435 // Just delete the users now. 436 // 437 if (!A->use_empty()) 438 A->replaceAllUsesWith(UndefValue::get(A->getType())); 439 if (AST) AST->deleteValue(A); 440 A->eraseFromParent(); 441 } 442 443 444 // Loop over all of the PHI nodes and see if there are any that we can get 445 // rid of because they merge all of the same incoming values. This can 446 // happen due to undef values coming into the PHI nodes. This process is 447 // iterative, because eliminating one PHI node can cause others to be removed. 448 bool EliminatedAPHI = true; 449 while (EliminatedAPHI) { 450 EliminatedAPHI = false; 451 452 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I = 453 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) { 454 PHINode *PN = I->second; 455 456 // If this PHI node merges one value and/or undefs, get the value. 457 if (Value *V = PN->hasConstantValue(true)) { 458 if (!isa<Instruction>(V) || 459 properlyDominates(cast<Instruction>(V), PN)) { 460 if (AST && isa<PointerType>(PN->getType())) 461 AST->deleteValue(PN); 462 PN->replaceAllUsesWith(V); 463 PN->eraseFromParent(); 464 NewPhiNodes.erase(I++); 465 EliminatedAPHI = true; 466 continue; 467 } 468 } 469 ++I; 470 } 471 } 472 473 // At this point, the renamer has added entries to PHI nodes for all reachable 474 // code. Unfortunately, there may be unreachable blocks which the renamer 475 // hasn't traversed. If this is the case, the PHI nodes may not 476 // have incoming values for all predecessors. Loop over all PHI nodes we have 477 // created, inserting undef values if they are missing any incoming values. 478 // 479 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I = 480 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) { 481 // We want to do this once per basic block. As such, only process a block 482 // when we find the PHI that is the first entry in the block. 483 PHINode *SomePHI = I->second; 484 BasicBlock *BB = SomePHI->getParent(); 485 if (&BB->front() != SomePHI) 486 continue; 487 488 // Count the number of preds for BB. 489 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 490 491 // Only do work here if there the PHI nodes are missing incoming values. We 492 // know that all PHI nodes that were inserted in a block will have the same 493 // number of incoming values, so we can just check any of them. 494 if (SomePHI->getNumIncomingValues() == Preds.size()) 495 continue; 496 497 // Ok, now we know that all of the PHI nodes are missing entries for some 498 // basic blocks. Start by sorting the incoming predecessors for efficient 499 // access. 500 std::sort(Preds.begin(), Preds.end()); 501 502 // Now we loop through all BB's which have entries in SomePHI and remove 503 // them from the Preds list. 504 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) { 505 // Do a log(n) search of the Preds list for the entry we want. 506 SmallVector<BasicBlock*, 16>::iterator EntIt = 507 std::lower_bound(Preds.begin(), Preds.end(), 508 SomePHI->getIncomingBlock(i)); 509 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&& 510 "PHI node has entry for a block which is not a predecessor!"); 511 512 // Remove the entry 513 Preds.erase(EntIt); 514 } 515 516 // At this point, the blocks left in the preds list must have dummy 517 // entries inserted into every PHI nodes for the block. Update all the phi 518 // nodes in this block that we are inserting (there could be phis before 519 // mem2reg runs). 520 unsigned NumBadPreds = SomePHI->getNumIncomingValues(); 521 BasicBlock::iterator BBI = BB->begin(); 522 while ((SomePHI = dyn_cast<PHINode>(BBI++)) && 523 SomePHI->getNumIncomingValues() == NumBadPreds) { 524 Value *UndefVal = UndefValue::get(SomePHI->getType()); 525 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred) 526 SomePHI->addIncoming(UndefVal, Preds[pred]); 527 } 528 } 529 530 NewPhiNodes.clear(); 531} 532 533 534/// RewriteSingleStoreAlloca - If there is only a single store to this value, 535/// replace any loads of it that are directly dominated by the definition with 536/// the value stored. 537void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI, 538 AllocaInfo &Info) { 539 StoreInst *OnlyStore = Info.OnlyStore; 540 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0)); 541 542 // Be aware of loads before the store. 543 SmallPtrSet<BasicBlock*, 32> ProcessedBlocks; 544 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) { 545 BasicBlock *UseBlock = Info.UsingBlocks[i]; 546 547 // If we already processed this block, don't reprocess it. 548 if (!ProcessedBlocks.insert(UseBlock)) { 549 Info.UsingBlocks[i] = Info.UsingBlocks.back(); 550 Info.UsingBlocks.pop_back(); 551 --i; --e; 552 continue; 553 } 554 555 // If the store dominates the block and if we haven't processed it yet, 556 // do so now. We can't handle the case where the store doesn't dominate a 557 // block because there may be a path between the store and the use, but we 558 // may need to insert phi nodes to handle dominance properly. 559 if (!StoringGlobalVal && !dominates(OnlyStore->getParent(), UseBlock)) 560 continue; 561 562 // If the use and store are in the same block, do a quick scan to 563 // verify that there are no uses before the store. 564 if (UseBlock == OnlyStore->getParent()) { 565 BasicBlock::iterator I = UseBlock->begin(); 566 for (; &*I != OnlyStore; ++I) { // scan block for store. 567 if (isa<LoadInst>(I) && I->getOperand(0) == AI) 568 break; 569 } 570 if (&*I != OnlyStore) 571 continue; // Do not promote the uses of this in this block. 572 } 573 574 // Otherwise, if this is a different block or if all uses happen 575 // after the store, do a simple linear scan to replace loads with 576 // the stored value. 577 for (BasicBlock::iterator I = UseBlock->begin(), E = UseBlock->end(); 578 I != E; ) { 579 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) { 580 if (LI->getOperand(0) == AI) { 581 LI->replaceAllUsesWith(OnlyStore->getOperand(0)); 582 if (AST && isa<PointerType>(LI->getType())) 583 AST->deleteValue(LI); 584 LI->eraseFromParent(); 585 } 586 } 587 } 588 589 // Finally, remove this block from the UsingBlock set. 590 Info.UsingBlocks[i] = Info.UsingBlocks.back(); 591 Info.UsingBlocks.pop_back(); 592 --i; --e; 593 } 594} 595 596 597// MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not 598// "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF 599// as usual (inserting the PHI nodes in the DeadPHINodes set), then processes 600// each read of the variable. For each block that reads the variable, this 601// function is called, which removes used PHI nodes from the DeadPHINodes set. 602// After all of the reads have been processed, any PHI nodes left in the 603// DeadPHINodes set are removed. 604// 605void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum, 606 SmallPtrSet<PHINode*, 16> &DeadPHINodes) { 607 // Scan the immediate dominators of this block looking for a block which has a 608 // PHI node for Alloca num. If we find it, mark the PHI node as being alive! 609 DomTreeNode *IDomNode = DT.getNode(BB); 610 for (DomTreeNode *IDom = IDomNode; IDom; IDom = IDom->getIDom()) { 611 BasicBlock *DomBB = IDom->getBlock(); 612 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator 613 I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum)); 614 if (I != NewPhiNodes.end()) { 615 // Ok, we found an inserted PHI node which dominates this value. 616 PHINode *DominatingPHI = I->second; 617 618 // Find out if we previously thought it was dead. If so, mark it as being 619 // live by removing it from the DeadPHINodes set. 620 if (DeadPHINodes.erase(DominatingPHI)) { 621 // Now that we have marked the PHI node alive, also mark any PHI nodes 622 // which it might use as being alive as well. 623 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB); 624 PI != PE; ++PI) 625 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes); 626 } 627 } 628 } 629} 630 631/// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic 632/// block. If this is the case, avoid traversing the CFG and inserting a lot of 633/// potentially useless PHI nodes by just performing a single linear pass over 634/// the basic block using the Alloca. 635/// 636/// If we cannot promote this alloca (because it is read before it is written), 637/// return true. This is necessary in cases where, due to control flow, the 638/// alloca is potentially undefined on some control flow paths. e.g. code like 639/// this is potentially correct: 640/// 641/// for (...) { if (c) { A = undef; undef = B; } } 642/// 643/// ... so long as A is not used before undef is set. 644/// 645bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) { 646 assert(!AI->use_empty() && "There are no uses of the alloca!"); 647 648 // Handle degenerate cases quickly. 649 if (AI->hasOneUse()) { 650 Instruction *U = cast<Instruction>(AI->use_back()); 651 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 652 // Must be a load of uninitialized value. 653 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType())); 654 if (AST && isa<PointerType>(LI->getType())) 655 AST->deleteValue(LI); 656 } else { 657 // Otherwise it must be a store which is never read. 658 assert(isa<StoreInst>(U)); 659 } 660 BB->getInstList().erase(U); 661 } else { 662 // Uses of the uninitialized memory location shall get undef. 663 Value *CurVal = 0; 664 665 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 666 Instruction *Inst = I++; 667 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 668 if (LI->getOperand(0) == AI) { 669 if (!CurVal) return true; // Could not locally promote! 670 671 // Loads just returns the "current value"... 672 LI->replaceAllUsesWith(CurVal); 673 if (AST && isa<PointerType>(LI->getType())) 674 AST->deleteValue(LI); 675 BB->getInstList().erase(LI); 676 } 677 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 678 if (SI->getOperand(1) == AI) { 679 // Store updates the "current value"... 680 CurVal = SI->getOperand(0); 681 BB->getInstList().erase(SI); 682 } 683 } 684 } 685 } 686 687 // After traversing the basic block, there should be no more uses of the 688 // alloca: remove it now. 689 assert(AI->use_empty() && "Uses of alloca from more than one BB??"); 690 if (AST) AST->deleteValue(AI); 691 AI->eraseFromParent(); 692 693 ++NumLocalPromoted; 694 return false; 695} 696 697/// PromoteLocallyUsedAllocas - This method is just like 698/// PromoteLocallyUsedAlloca, except that it processes multiple alloca 699/// instructions in parallel. This is important in cases where we have large 700/// basic blocks, as we don't want to rescan the entire basic block for each 701/// alloca which is locally used in it (which might be a lot). 702void PromoteMem2Reg:: 703PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) { 704 DenseMap<AllocaInst*, Value*> CurValues; 705 for (unsigned i = 0, e = AIs.size(); i != e; ++i) 706 CurValues[AIs[i]] = 0; // Insert with null value 707 708 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 709 Instruction *Inst = I++; 710 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 711 // Is this a load of an alloca we are tracking? 712 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) { 713 DenseMap<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 714 if (AIt != CurValues.end()) { 715 // If loading an uninitialized value, allow the inter-block case to 716 // handle it. Due to control flow, this might actually be ok. 717 if (AIt->second == 0) { // Use of locally uninitialized value?? 718 RetryList.push_back(AI); // Retry elsewhere. 719 CurValues.erase(AIt); // Stop tracking this here. 720 if (CurValues.empty()) return; 721 } else { 722 // Loads just returns the "current value"... 723 LI->replaceAllUsesWith(AIt->second); 724 if (AST && isa<PointerType>(LI->getType())) 725 AST->deleteValue(LI); 726 BB->getInstList().erase(LI); 727 } 728 } 729 } 730 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 731 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) { 732 DenseMap<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 733 if (AIt != CurValues.end()) { 734 // Store updates the "current value"... 735 AIt->second = SI->getOperand(0); 736 SI->eraseFromParent(); 737 } 738 } 739 } 740 } 741 742 // At the end of the block scan, all allocas in CurValues are dead. 743 for (DenseMap<AllocaInst*, Value*>::iterator I = CurValues.begin(), 744 E = CurValues.end(); I != E; ++I) { 745 AllocaInst *AI = I->first; 746 assert(AI->use_empty() && "Uses of alloca from more than one BB??"); 747 if (AST) AST->deleteValue(AI); 748 AI->eraseFromParent(); 749 } 750 751 NumLocalPromoted += CurValues.size(); 752} 753 754 755 756// QueuePhiNode - queues a phi-node to be added to a basic-block for a specific 757// Alloca returns true if there wasn't already a phi-node for that variable 758// 759bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo, 760 unsigned &Version, 761 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) { 762 // Look up the basic-block in question. 763 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)]; 764 765 // If the BB already has a phi node added for the i'th alloca then we're done! 766 if (PN) return false; 767 768 // Create a PhiNode using the dereferenced type... and add the phi-node to the 769 // BasicBlock. 770 PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(), 771 Allocas[AllocaNo]->getName() + "." + 772 utostr(Version++), BB->begin()); 773 PhiToAllocaMap[PN] = AllocaNo; 774 775 InsertedPHINodes.insert(PN); 776 777 if (AST && isa<PointerType>(PN->getType())) 778 AST->copyValue(PointerAllocaValues[AllocaNo], PN); 779 780 return true; 781} 782 783 784// RenamePass - Recursively traverse the CFG of the function, renaming loads and 785// stores to the allocas which we are promoting. IncomingVals indicates what 786// value each Alloca contains on exit from the predecessor block Pred. 787// 788void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred, 789 RenamePassData::ValVector &IncomingVals, 790 std::vector<RenamePassData> &Worklist) { 791 // If we are inserting any phi nodes into this BB, they will already be in the 792 // block. 793 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) { 794 // Pred may have multiple edges to BB. If so, we want to add N incoming 795 // values to each PHI we are inserting on the first time we see the edge. 796 // Check to see if APN already has incoming values from Pred. This also 797 // prevents us from modifying PHI nodes that are not currently being 798 // inserted. 799 bool HasPredEntries = false; 800 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) { 801 if (APN->getIncomingBlock(i) == Pred) { 802 HasPredEntries = true; 803 break; 804 } 805 } 806 807 // If we have PHI nodes to update, compute the number of edges from Pred to 808 // BB. 809 if (!HasPredEntries) { 810 TerminatorInst *PredTerm = Pred->getTerminator(); 811 unsigned NumEdges = 0; 812 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) { 813 if (PredTerm->getSuccessor(i) == BB) 814 ++NumEdges; 815 } 816 assert(NumEdges && "Must be at least one edge from Pred to BB!"); 817 818 // Add entries for all the phis. 819 BasicBlock::iterator PNI = BB->begin(); 820 do { 821 unsigned AllocaNo = PhiToAllocaMap[APN]; 822 823 // Add N incoming values to the PHI node. 824 for (unsigned i = 0; i != NumEdges; ++i) 825 APN->addIncoming(IncomingVals[AllocaNo], Pred); 826 827 // The currently active variable for this block is now the PHI. 828 IncomingVals[AllocaNo] = APN; 829 830 // Get the next phi node. 831 ++PNI; 832 APN = dyn_cast<PHINode>(PNI); 833 if (APN == 0) break; 834 835 // Verify it doesn't already have entries for Pred. If it does, it is 836 // not being inserted by this mem2reg invocation. 837 HasPredEntries = false; 838 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) { 839 if (APN->getIncomingBlock(i) == Pred) { 840 HasPredEntries = true; 841 break; 842 } 843 } 844 } while (!HasPredEntries); 845 } 846 } 847 848 // Don't revisit blocks. 849 if (!Visited.insert(BB)) return; 850 851 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) { 852 Instruction *I = II++; // get the instruction, increment iterator 853 854 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 855 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) { 856 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src); 857 if (AI != AllocaLookup.end()) { 858 Value *V = IncomingVals[AI->second]; 859 860 // walk the use list of this load and replace all uses with r 861 LI->replaceAllUsesWith(V); 862 if (AST && isa<PointerType>(LI->getType())) 863 AST->deleteValue(LI); 864 BB->getInstList().erase(LI); 865 } 866 } 867 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 868 // Delete this instruction and mark the name as the current holder of the 869 // value 870 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) { 871 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest); 872 if (ai != AllocaLookup.end()) { 873 // what value were we writing? 874 IncomingVals[ai->second] = SI->getOperand(0); 875 BB->getInstList().erase(SI); 876 } 877 } 878 } 879 } 880 881 // Recurse to our successors. 882 TerminatorInst *TI = BB->getTerminator(); 883 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) 884 Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals)); 885} 886 887/// PromoteMemToReg - Promote the specified list of alloca instructions into 888/// scalar registers, inserting PHI nodes as appropriate. This function makes 889/// use of DominanceFrontier information. This function does not modify the CFG 890/// of the function at all. All allocas must be from the same function. 891/// 892/// If AST is specified, the specified tracker is updated to reflect changes 893/// made to the IR. 894/// 895void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas, 896 DominatorTree &DT, DominanceFrontier &DF, 897 AliasSetTracker *AST) { 898 // If there is nothing to do, bail out... 899 if (Allocas.empty()) return; 900 901 SmallVector<AllocaInst*, 16> RetryList; 902 PromoteMem2Reg(Allocas, RetryList, DT, DF, AST).run(); 903 904 // PromoteMem2Reg may not have been able to promote all of the allocas in one 905 // pass, run it again if needed. 906 std::vector<AllocaInst*> NewAllocas; 907 while (!RetryList.empty()) { 908 // If we need to retry some allocas, this is due to there being no store 909 // before a read in a local block. To counteract this, insert a store of 910 // undef into the alloca right after the alloca itself. 911 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) { 912 BasicBlock::iterator BBI = RetryList[i]; 913 914 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()), 915 RetryList[i], ++BBI); 916 } 917 918 NewAllocas.assign(RetryList.begin(), RetryList.end()); 919 RetryList.clear(); 920 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, AST).run(); 921 NewAllocas.clear(); 922 } 923} 924