PromoteMemoryToRegister.cpp revision 5dd75b4ca7e582f44da2f50362e8ab4c59972b5f
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 the alloca is only read and written in one basic block, just perform a 277 // linear sweep over the block to eliminate it. 278 if (Info.OnlyUsedInOneBlock) { 279 LocallyUsedAllocas[Info.OnlyBlock].push_back(AI); 280 281 // Remove the alloca from the Allocas list, since it will be processed. 282 RemoveFromAllocasList(AllocaNum); 283 continue; 284 } 285 286 // If there is only a single store to this value, replace any loads of 287 // it that are directly dominated by the definition with the value stored. 288 if (Info.DefiningBlocks.size() == 1) { 289 RewriteSingleStoreAlloca(AI, Info); 290 291 // Finally, after the scan, check to see if the store is all that is left. 292 if (Info.UsingBlocks.empty()) { 293 ++NumSingleStore; 294 // The alloca has been processed, move on. 295 RemoveFromAllocasList(AllocaNum); 296 continue; 297 } 298 } 299 300 301 if (AST) 302 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal; 303 304 // If we haven't computed a numbering for the BB's in the function, do so 305 // now. 306 if (BBNumbers.empty()) { 307 unsigned ID = 0; 308 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) 309 BBNumbers[I] = ID++; 310 } 311 312 // Compute the locations where PhiNodes need to be inserted. Look at the 313 // dominance frontier of EACH basic-block we have a write in. 314 // 315 unsigned CurrentVersion = 0; 316 SmallPtrSet<PHINode*, 16> InsertedPHINodes; 317 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks; 318 while (!Info.DefiningBlocks.empty()) { 319 BasicBlock *BB = Info.DefiningBlocks.back(); 320 Info.DefiningBlocks.pop_back(); 321 322 // Look up the DF for this write, add it to PhiNodes 323 DominanceFrontier::const_iterator it = DF.find(BB); 324 if (it != DF.end()) { 325 const DominanceFrontier::DomSetType &S = it->second; 326 327 // In theory we don't need the indirection through the DFBlocks vector. 328 // In practice, the order of calling QueuePhiNode would depend on the 329 // (unspecified) ordering of basic blocks in the dominance frontier, 330 // which would give PHI nodes non-determinstic subscripts. Fix this by 331 // processing blocks in order of the occurance in the function. 332 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(), 333 PE = S.end(); P != PE; ++P) 334 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P)); 335 336 // Sort by which the block ordering in the function. 337 std::sort(DFBlocks.begin(), DFBlocks.end()); 338 339 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) { 340 BasicBlock *BB = DFBlocks[i].second; 341 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes)) 342 Info.DefiningBlocks.push_back(BB); 343 } 344 DFBlocks.clear(); 345 } 346 } 347 348 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier 349 // of the writes to the variable, scan through the reads of the variable, 350 // marking PHI nodes which are actually necessary as alive (by removing them 351 // from the InsertedPHINodes set). This is not perfect: there may PHI 352 // marked alive because of loads which are dominated by stores, but there 353 // will be no unmarked PHI nodes which are actually used. 354 // 355 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) 356 MarkDominatingPHILive(Info.UsingBlocks[i], AllocaNum, InsertedPHINodes); 357 Info.UsingBlocks.clear(); 358 359 // If there are any PHI nodes which are now known to be dead, remove them! 360 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(), 361 E = InsertedPHINodes.end(); I != E; ++I) { 362 PHINode *PN = *I; 363 bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum)); 364 Erased=Erased; 365 assert(Erased && "PHI already removed?"); 366 367 if (AST && isa<PointerType>(PN->getType())) 368 AST->deleteValue(PN); 369 PN->eraseFromParent(); 370 PhiToAllocaMap.erase(PN); 371 } 372 373 // Keep the reverse mapping of the 'Allocas' array. 374 AllocaLookup[Allocas[AllocaNum]] = AllocaNum; 375 } 376 377 // Process all allocas which are only used in a single basic block. 378 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I = 379 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){ 380 const std::vector<AllocaInst*> &LocAllocas = I->second; 381 assert(!LocAllocas.empty() && "empty alloca list??"); 382 383 // It's common for there to only be one alloca in the list. Handle it 384 // efficiently. 385 if (LocAllocas.size() == 1) { 386 // If we can do the quick promotion pass, do so now. 387 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0])) 388 RetryList.push_back(LocAllocas[0]); // Failed, retry later. 389 } else { 390 // Locally promote anything possible. Note that if this is unable to 391 // promote a particular alloca, it puts the alloca onto the Allocas vector 392 // for global processing. 393 PromoteLocallyUsedAllocas(I->first, LocAllocas); 394 } 395 } 396 397 if (Allocas.empty()) 398 return; // All of the allocas must have been trivial! 399 400 // Set the incoming values for the basic block to be null values for all of 401 // the alloca's. We do this in case there is a load of a value that has not 402 // been stored yet. In this case, it will get this null value. 403 // 404 RenamePassData::ValVector Values(Allocas.size()); 405 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) 406 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType()); 407 408 // Walks all basic blocks in the function performing the SSA rename algorithm 409 // and inserting the phi nodes we marked as necessary 410 // 411 std::vector<RenamePassData> RenamePassWorkList; 412 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values)); 413 while (!RenamePassWorkList.empty()) { 414 RenamePassData RPD; 415 RPD.swap(RenamePassWorkList.back()); 416 RenamePassWorkList.pop_back(); 417 // RenamePass may add new worklist entries. 418 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList); 419 } 420 421 // The renamer uses the Visited set to avoid infinite loops. Clear it now. 422 Visited.clear(); 423 424 // Remove the allocas themselves from the function. 425 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 426 Instruction *A = Allocas[i]; 427 428 // If there are any uses of the alloca instructions left, they must be in 429 // sections of dead code that were not processed on the dominance frontier. 430 // Just delete the users now. 431 // 432 if (!A->use_empty()) 433 A->replaceAllUsesWith(UndefValue::get(A->getType())); 434 if (AST) AST->deleteValue(A); 435 A->eraseFromParent(); 436 } 437 438 439 // Loop over all of the PHI nodes and see if there are any that we can get 440 // rid of because they merge all of the same incoming values. This can 441 // happen due to undef values coming into the PHI nodes. This process is 442 // iterative, because eliminating one PHI node can cause others to be removed. 443 bool EliminatedAPHI = true; 444 while (EliminatedAPHI) { 445 EliminatedAPHI = false; 446 447 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I = 448 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) { 449 PHINode *PN = I->second; 450 451 // If this PHI node merges one value and/or undefs, get the value. 452 if (Value *V = PN->hasConstantValue(true)) { 453 if (!isa<Instruction>(V) || 454 properlyDominates(cast<Instruction>(V), PN)) { 455 if (AST && isa<PointerType>(PN->getType())) 456 AST->deleteValue(PN); 457 PN->replaceAllUsesWith(V); 458 PN->eraseFromParent(); 459 NewPhiNodes.erase(I++); 460 EliminatedAPHI = true; 461 continue; 462 } 463 } 464 ++I; 465 } 466 } 467 468 // At this point, the renamer has added entries to PHI nodes for all reachable 469 // code. Unfortunately, there may be unreachable blocks which the renamer 470 // hasn't traversed. If this is the case, the PHI nodes may not 471 // have incoming values for all predecessors. Loop over all PHI nodes we have 472 // created, inserting undef values if they are missing any incoming values. 473 // 474 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I = 475 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) { 476 // We want to do this once per basic block. As such, only process a block 477 // when we find the PHI that is the first entry in the block. 478 PHINode *SomePHI = I->second; 479 BasicBlock *BB = SomePHI->getParent(); 480 if (&BB->front() != SomePHI) 481 continue; 482 483 // Count the number of preds for BB. 484 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 485 486 // Only do work here if there the PHI nodes are missing incoming values. We 487 // know that all PHI nodes that were inserted in a block will have the same 488 // number of incoming values, so we can just check any of them. 489 if (SomePHI->getNumIncomingValues() == Preds.size()) 490 continue; 491 492 // Ok, now we know that all of the PHI nodes are missing entries for some 493 // basic blocks. Start by sorting the incoming predecessors for efficient 494 // access. 495 std::sort(Preds.begin(), Preds.end()); 496 497 // Now we loop through all BB's which have entries in SomePHI and remove 498 // them from the Preds list. 499 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) { 500 // Do a log(n) search of the Preds list for the entry we want. 501 SmallVector<BasicBlock*, 16>::iterator EntIt = 502 std::lower_bound(Preds.begin(), Preds.end(), 503 SomePHI->getIncomingBlock(i)); 504 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&& 505 "PHI node has entry for a block which is not a predecessor!"); 506 507 // Remove the entry 508 Preds.erase(EntIt); 509 } 510 511 // At this point, the blocks left in the preds list must have dummy 512 // entries inserted into every PHI nodes for the block. Update all the phi 513 // nodes in this block that we are inserting (there could be phis before 514 // mem2reg runs). 515 unsigned NumBadPreds = SomePHI->getNumIncomingValues(); 516 BasicBlock::iterator BBI = BB->begin(); 517 while ((SomePHI = dyn_cast<PHINode>(BBI++)) && 518 SomePHI->getNumIncomingValues() == NumBadPreds) { 519 Value *UndefVal = UndefValue::get(SomePHI->getType()); 520 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred) 521 SomePHI->addIncoming(UndefVal, Preds[pred]); 522 } 523 } 524 525 NewPhiNodes.clear(); 526} 527 528 529/// RewriteSingleStoreAlloca - If there is only a single store to this value, 530/// replace any loads of it that are directly dominated by the definition with 531/// the value stored. 532void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI, 533 AllocaInfo &Info) { 534 // Be aware of loads before the store. 535 std::set<BasicBlock*> ProcessedBlocks; 536 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) { 537 // If the store dominates the block and if we haven't processed it yet, 538 // do so now. 539 if (!dominates(Info.OnlyStore->getParent(), Info.UsingBlocks[i])) 540 continue; 541 542 if (!ProcessedBlocks.insert(Info.UsingBlocks[i]).second) 543 continue; 544 545 BasicBlock *UseBlock = Info.UsingBlocks[i]; 546 547 // If the use and store are in the same block, do a quick scan to 548 // verify that there are no uses before the store. 549 if (UseBlock == Info.OnlyStore->getParent()) { 550 BasicBlock::iterator I = UseBlock->begin(); 551 for (; &*I != Info.OnlyStore; ++I) { // scan block for store. 552 if (isa<LoadInst>(I) && I->getOperand(0) == AI) 553 break; 554 } 555 if (&*I != Info.OnlyStore) break; // Do not handle this case. 556 } 557 558 // Otherwise, if this is a different block or if all uses happen 559 // after the store, do a simple linear scan to replace loads with 560 // the stored value. 561 for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end(); 562 I != E; ) { 563 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) { 564 if (LI->getOperand(0) == AI) { 565 LI->replaceAllUsesWith(Info.OnlyStore->getOperand(0)); 566 if (AST && isa<PointerType>(LI->getType())) 567 AST->deleteValue(LI); 568 LI->eraseFromParent(); 569 } 570 } 571 } 572 573 // Finally, remove this block from the UsingBlock set. 574 Info.UsingBlocks[i] = Info.UsingBlocks.back(); 575 --i; --e; 576 } 577} 578 579 580// MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not 581// "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF 582// as usual (inserting the PHI nodes in the DeadPHINodes set), then processes 583// each read of the variable. For each block that reads the variable, this 584// function is called, which removes used PHI nodes from the DeadPHINodes set. 585// After all of the reads have been processed, any PHI nodes left in the 586// DeadPHINodes set are removed. 587// 588void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum, 589 SmallPtrSet<PHINode*, 16> &DeadPHINodes) { 590 // Scan the immediate dominators of this block looking for a block which has a 591 // PHI node for Alloca num. If we find it, mark the PHI node as being alive! 592 DomTreeNode *IDomNode = DT.getNode(BB); 593 for (DomTreeNode *IDom = IDomNode; IDom; IDom = IDom->getIDom()) { 594 BasicBlock *DomBB = IDom->getBlock(); 595 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator 596 I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum)); 597 if (I != NewPhiNodes.end()) { 598 // Ok, we found an inserted PHI node which dominates this value. 599 PHINode *DominatingPHI = I->second; 600 601 // Find out if we previously thought it was dead. If so, mark it as being 602 // live by removing it from the DeadPHINodes set. 603 if (DeadPHINodes.erase(DominatingPHI)) { 604 // Now that we have marked the PHI node alive, also mark any PHI nodes 605 // which it might use as being alive as well. 606 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB); 607 PI != PE; ++PI) 608 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes); 609 } 610 } 611 } 612} 613 614/// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic 615/// block. If this is the case, avoid traversing the CFG and inserting a lot of 616/// potentially useless PHI nodes by just performing a single linear pass over 617/// the basic block using the Alloca. 618/// 619/// If we cannot promote this alloca (because it is read before it is written), 620/// return true. This is necessary in cases where, due to control flow, the 621/// alloca is potentially undefined on some control flow paths. e.g. code like 622/// this is potentially correct: 623/// 624/// for (...) { if (c) { A = undef; undef = B; } } 625/// 626/// ... so long as A is not used before undef is set. 627/// 628bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) { 629 assert(!AI->use_empty() && "There are no uses of the alloca!"); 630 631 // Handle degenerate cases quickly. 632 if (AI->hasOneUse()) { 633 Instruction *U = cast<Instruction>(AI->use_back()); 634 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 635 // Must be a load of uninitialized value. 636 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType())); 637 if (AST && isa<PointerType>(LI->getType())) 638 AST->deleteValue(LI); 639 } else { 640 // Otherwise it must be a store which is never read. 641 assert(isa<StoreInst>(U)); 642 } 643 BB->getInstList().erase(U); 644 } else { 645 // Uses of the uninitialized memory location shall get undef. 646 Value *CurVal = 0; 647 648 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 649 Instruction *Inst = I++; 650 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 651 if (LI->getOperand(0) == AI) { 652 if (!CurVal) return true; // Could not locally promote! 653 654 // Loads just returns the "current value"... 655 LI->replaceAllUsesWith(CurVal); 656 if (AST && isa<PointerType>(LI->getType())) 657 AST->deleteValue(LI); 658 BB->getInstList().erase(LI); 659 } 660 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 661 if (SI->getOperand(1) == AI) { 662 // Store updates the "current value"... 663 CurVal = SI->getOperand(0); 664 BB->getInstList().erase(SI); 665 } 666 } 667 } 668 } 669 670 // After traversing the basic block, there should be no more uses of the 671 // alloca, remove it now. 672 assert(AI->use_empty() && "Uses of alloca from more than one BB??"); 673 if (AST) AST->deleteValue(AI); 674 AI->getParent()->getInstList().erase(AI); 675 676 ++NumLocalPromoted; 677 return false; 678} 679 680/// PromoteLocallyUsedAllocas - This method is just like 681/// PromoteLocallyUsedAlloca, except that it processes multiple alloca 682/// instructions in parallel. This is important in cases where we have large 683/// basic blocks, as we don't want to rescan the entire basic block for each 684/// alloca which is locally used in it (which might be a lot). 685void PromoteMem2Reg:: 686PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) { 687 std::map<AllocaInst*, Value*> CurValues; 688 for (unsigned i = 0, e = AIs.size(); i != e; ++i) 689 CurValues[AIs[i]] = 0; // Insert with null value 690 691 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 692 Instruction *Inst = I++; 693 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 694 // Is this a load of an alloca we are tracking? 695 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) { 696 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 697 if (AIt != CurValues.end()) { 698 // If loading an uninitialized value, allow the inter-block case to 699 // handle it. Due to control flow, this might actually be ok. 700 if (AIt->second == 0) { // Use of locally uninitialized value?? 701 RetryList.push_back(AI); // Retry elsewhere. 702 CurValues.erase(AIt); // Stop tracking this here. 703 if (CurValues.empty()) return; 704 } else { 705 // Loads just returns the "current value"... 706 LI->replaceAllUsesWith(AIt->second); 707 if (AST && isa<PointerType>(LI->getType())) 708 AST->deleteValue(LI); 709 BB->getInstList().erase(LI); 710 } 711 } 712 } 713 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 714 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) { 715 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 716 if (AIt != CurValues.end()) { 717 // Store updates the "current value"... 718 AIt->second = SI->getOperand(0); 719 BB->getInstList().erase(SI); 720 } 721 } 722 } 723 } 724} 725 726 727 728// QueuePhiNode - queues a phi-node to be added to a basic-block for a specific 729// Alloca returns true if there wasn't already a phi-node for that variable 730// 731bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo, 732 unsigned &Version, 733 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) { 734 // Look up the basic-block in question. 735 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)]; 736 737 // If the BB already has a phi node added for the i'th alloca then we're done! 738 if (PN) return false; 739 740 // Create a PhiNode using the dereferenced type... and add the phi-node to the 741 // BasicBlock. 742 PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(), 743 Allocas[AllocaNo]->getName() + "." + 744 utostr(Version++), BB->begin()); 745 PhiToAllocaMap[PN] = AllocaNo; 746 747 InsertedPHINodes.insert(PN); 748 749 if (AST && isa<PointerType>(PN->getType())) 750 AST->copyValue(PointerAllocaValues[AllocaNo], PN); 751 752 return true; 753} 754 755 756// RenamePass - Recursively traverse the CFG of the function, renaming loads and 757// stores to the allocas which we are promoting. IncomingVals indicates what 758// value each Alloca contains on exit from the predecessor block Pred. 759// 760void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred, 761 RenamePassData::ValVector &IncomingVals, 762 std::vector<RenamePassData> &Worklist) { 763 // If we are inserting any phi nodes into this BB, they will already be in the 764 // block. 765 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) { 766 // Pred may have multiple edges to BB. If so, we want to add N incoming 767 // values to each PHI we are inserting on the first time we see the edge. 768 // Check to see if APN already has incoming values from Pred. This also 769 // prevents us from modifying PHI nodes that are not currently being 770 // inserted. 771 bool HasPredEntries = false; 772 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) { 773 if (APN->getIncomingBlock(i) == Pred) { 774 HasPredEntries = true; 775 break; 776 } 777 } 778 779 // If we have PHI nodes to update, compute the number of edges from Pred to 780 // BB. 781 if (!HasPredEntries) { 782 TerminatorInst *PredTerm = Pred->getTerminator(); 783 unsigned NumEdges = 0; 784 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) { 785 if (PredTerm->getSuccessor(i) == BB) 786 ++NumEdges; 787 } 788 assert(NumEdges && "Must be at least one edge from Pred to BB!"); 789 790 // Add entries for all the phis. 791 BasicBlock::iterator PNI = BB->begin(); 792 do { 793 unsigned AllocaNo = PhiToAllocaMap[APN]; 794 795 // Add N incoming values to the PHI node. 796 for (unsigned i = 0; i != NumEdges; ++i) 797 APN->addIncoming(IncomingVals[AllocaNo], Pred); 798 799 // The currently active variable for this block is now the PHI. 800 IncomingVals[AllocaNo] = APN; 801 802 // Get the next phi node. 803 ++PNI; 804 APN = dyn_cast<PHINode>(PNI); 805 if (APN == 0) break; 806 807 // Verify it doesn't already have entries for Pred. If it does, it is 808 // not being inserted by this mem2reg invocation. 809 HasPredEntries = false; 810 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) { 811 if (APN->getIncomingBlock(i) == Pred) { 812 HasPredEntries = true; 813 break; 814 } 815 } 816 } while (!HasPredEntries); 817 } 818 } 819 820 // Don't revisit blocks. 821 if (!Visited.insert(BB)) return; 822 823 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) { 824 Instruction *I = II++; // get the instruction, increment iterator 825 826 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 827 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) { 828 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src); 829 if (AI != AllocaLookup.end()) { 830 Value *V = IncomingVals[AI->second]; 831 832 // walk the use list of this load and replace all uses with r 833 LI->replaceAllUsesWith(V); 834 if (AST && isa<PointerType>(LI->getType())) 835 AST->deleteValue(LI); 836 BB->getInstList().erase(LI); 837 } 838 } 839 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 840 // Delete this instruction and mark the name as the current holder of the 841 // value 842 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) { 843 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest); 844 if (ai != AllocaLookup.end()) { 845 // what value were we writing? 846 IncomingVals[ai->second] = SI->getOperand(0); 847 BB->getInstList().erase(SI); 848 } 849 } 850 } 851 } 852 853 // Recurse to our successors. 854 TerminatorInst *TI = BB->getTerminator(); 855 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) 856 Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals)); 857} 858 859/// PromoteMemToReg - Promote the specified list of alloca instructions into 860/// scalar registers, inserting PHI nodes as appropriate. This function makes 861/// use of DominanceFrontier information. This function does not modify the CFG 862/// of the function at all. All allocas must be from the same function. 863/// 864/// If AST is specified, the specified tracker is updated to reflect changes 865/// made to the IR. 866/// 867void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas, 868 DominatorTree &DT, DominanceFrontier &DF, 869 AliasSetTracker *AST) { 870 // If there is nothing to do, bail out... 871 if (Allocas.empty()) return; 872 873 SmallVector<AllocaInst*, 16> RetryList; 874 PromoteMem2Reg(Allocas, RetryList, DT, DF, AST).run(); 875 876 // PromoteMem2Reg may not have been able to promote all of the allocas in one 877 // pass, run it again if needed. 878 std::vector<AllocaInst*> NewAllocas; 879 while (!RetryList.empty()) { 880 // If we need to retry some allocas, this is due to there being no store 881 // before a read in a local block. To counteract this, insert a store of 882 // undef into the alloca right after the alloca itself. 883 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) { 884 BasicBlock::iterator BBI = RetryList[i]; 885 886 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()), 887 RetryList[i], ++BBI); 888 } 889 890 NewAllocas.assign(RetryList.begin(), RetryList.end()); 891 RetryList.clear(); 892 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, AST).run(); 893 NewAllocas.clear(); 894 } 895} 896