PromoteMemoryToRegister.cpp revision 634c76c08c34e5fb475527a4e91927a142dd9c46
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#include "llvm/Transforms/Utils/PromoteMemToReg.h" 20#include "llvm/Constants.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/Function.h" 23#include "llvm/Instructions.h" 24#include "llvm/Analysis/Dominators.h" 25#include "llvm/Analysis/AliasSetTracker.h" 26#include "llvm/ADT/StringExtras.h" 27#include "llvm/Support/CFG.h" 28#include "llvm/Support/StableBasicBlockNumbering.h" 29#include <algorithm> 30using namespace llvm; 31 32/// isAllocaPromotable - Return true if this alloca is legal for promotion. 33/// This is true if there are only loads and stores to the alloca. 34/// 35bool llvm::isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) { 36 // FIXME: If the memory unit is of pointer or integer type, we can permit 37 // assignments to subsections of the memory unit. 38 39 // Only allow direct loads and stores... 40 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end(); 41 UI != UE; ++UI) // Loop over all of the uses of the alloca 42 if (isa<LoadInst>(*UI)) { 43 // noop 44 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 45 if (SI->getOperand(0) == AI) 46 return false; // Don't allow a store OF the AI, only INTO the AI. 47 } else { 48 return false; // Not a load or store. 49 } 50 51 return true; 52} 53 54namespace { 55 struct PromoteMem2Reg { 56 /// Allocas - The alloca instructions being promoted. 57 /// 58 std::vector<AllocaInst*> Allocas; 59 std::vector<AllocaInst*> &RetryList; 60 DominatorTree &DT; 61 DominanceFrontier &DF; 62 const TargetData &TD; 63 64 /// AST - An AliasSetTracker object to update. If null, don't update it. 65 /// 66 AliasSetTracker *AST; 67 68 /// AllocaLookup - Reverse mapping of Allocas. 69 /// 70 std::map<AllocaInst*, unsigned> AllocaLookup; 71 72 /// NewPhiNodes - The PhiNodes we're adding. 73 /// 74 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes; 75 76 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for 77 /// each alloca that is of pointer type, we keep track of what to copyValue 78 /// to the inserted PHI nodes here. 79 /// 80 std::vector<Value*> PointerAllocaValues; 81 82 /// Visited - The set of basic blocks the renamer has already visited. 83 /// 84 std::set<BasicBlock*> Visited; 85 86 /// BBNumbers - Contains a stable numbering of basic blocks to avoid 87 /// non-determinstic behavior. 88 StableBasicBlockNumbering BBNumbers; 89 90 public: 91 PromoteMem2Reg(const std::vector<AllocaInst*> &A, 92 std::vector<AllocaInst*> &Retry, DominatorTree &dt, 93 DominanceFrontier &df, const TargetData &td, 94 AliasSetTracker *ast) 95 : Allocas(A), RetryList(Retry), DT(dt), DF(df), TD(td), AST(ast) {} 96 97 void run(); 98 99 /// properlyDominates - Return true if I1 properly dominates I2. 100 /// 101 bool properlyDominates(Instruction *I1, Instruction *I2) const { 102 if (InvokeInst *II = dyn_cast<InvokeInst>(I1)) 103 I1 = II->getNormalDest()->begin(); 104 return DT[I1->getParent()]->properlyDominates(DT[I2->getParent()]); 105 } 106 107 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree. 108 /// 109 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const { 110 return DT[BB1]->dominates(DT[BB2]); 111 } 112 113 private: 114 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum, 115 std::set<PHINode*> &DeadPHINodes); 116 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI); 117 void PromoteLocallyUsedAllocas(BasicBlock *BB, 118 const std::vector<AllocaInst*> &AIs); 119 120 void RenamePass(BasicBlock *BB, BasicBlock *Pred, 121 std::vector<Value*> &IncVals); 122 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version, 123 std::set<PHINode*> &InsertedPHINodes); 124 }; 125} // end of anonymous namespace 126 127void PromoteMem2Reg::run() { 128 Function &F = *DF.getRoot()->getParent(); 129 130 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are 131 // only used in a single basic block. These instructions can be efficiently 132 // promoted by performing a single linear scan over that one block. Since 133 // individual basic blocks are sometimes large, we group together all allocas 134 // that are live in a single basic block by the basic block they are live in. 135 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas; 136 137 if (AST) PointerAllocaValues.resize(Allocas.size()); 138 139 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) { 140 AllocaInst *AI = Allocas[AllocaNum]; 141 142 assert(isAllocaPromotable(AI, TD) && 143 "Cannot promote non-promotable alloca!"); 144 assert(AI->getParent()->getParent() == &F && 145 "All allocas should be in the same function, which is same as DF!"); 146 147 if (AI->use_empty()) { 148 // If there are no uses of the alloca, just delete it now. 149 if (AST) AST->deleteValue(AI); 150 AI->eraseFromParent(); 151 152 // Remove the alloca from the Allocas list, since it has been processed 153 Allocas[AllocaNum] = Allocas.back(); 154 Allocas.pop_back(); 155 --AllocaNum; 156 continue; 157 } 158 159 // Calculate the set of read and write-locations for each alloca. This is 160 // analogous to finding the 'uses' and 'definitions' of each variable. 161 std::vector<BasicBlock*> DefiningBlocks; 162 std::vector<BasicBlock*> UsingBlocks; 163 164 StoreInst *OnlyStore = 0; 165 BasicBlock *OnlyBlock = 0; 166 bool OnlyUsedInOneBlock = true; 167 168 // As we scan the uses of the alloca instruction, keep track of stores, and 169 // decide whether all of the loads and stores to the alloca are within the 170 // same basic block. 171 Value *AllocaPointerVal = 0; 172 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){ 173 Instruction *User = cast<Instruction>(*U); 174 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 175 // Remember the basic blocks which define new values for the alloca 176 DefiningBlocks.push_back(SI->getParent()); 177 AllocaPointerVal = SI->getOperand(0); 178 OnlyStore = SI; 179 } else { 180 LoadInst *LI = cast<LoadInst>(User); 181 // Otherwise it must be a load instruction, keep track of variable reads 182 UsingBlocks.push_back(LI->getParent()); 183 AllocaPointerVal = LI; 184 } 185 186 if (OnlyUsedInOneBlock) { 187 if (OnlyBlock == 0) 188 OnlyBlock = User->getParent(); 189 else if (OnlyBlock != User->getParent()) 190 OnlyUsedInOneBlock = false; 191 } 192 } 193 194 // If the alloca is only read and written in one basic block, just perform a 195 // linear sweep over the block to eliminate it. 196 if (OnlyUsedInOneBlock) { 197 LocallyUsedAllocas[OnlyBlock].push_back(AI); 198 199 // Remove the alloca from the Allocas list, since it will be processed. 200 Allocas[AllocaNum] = Allocas.back(); 201 Allocas.pop_back(); 202 --AllocaNum; 203 continue; 204 } 205 206 // If there is only a single store to this value, replace any loads of 207 // it that are directly dominated by the definition with the value stored. 208 if (DefiningBlocks.size() == 1) { 209 // Be aware of loads before the store. 210 std::set<BasicBlock*> ProcessedBlocks; 211 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i) 212 // If the store dominates the block and if we haven't processed it yet, 213 // do so now. 214 if (dominates(OnlyStore->getParent(), UsingBlocks[i])) 215 if (ProcessedBlocks.insert(UsingBlocks[i]).second) { 216 BasicBlock *UseBlock = UsingBlocks[i]; 217 218 // If the use and store are in the same block, do a quick scan to 219 // verify that there are no uses before the store. 220 if (UseBlock == OnlyStore->getParent()) { 221 BasicBlock::iterator I = UseBlock->begin(); 222 for (; &*I != OnlyStore; ++I) { // scan block for store. 223 if (isa<LoadInst>(I) && I->getOperand(0) == AI) 224 break; 225 } 226 if (&*I != OnlyStore) break; // Do not handle this case. 227 } 228 229 // Otherwise, if this is a different block or if all uses happen 230 // after the store, do a simple linear scan to replace loads with 231 // the stored value. 232 for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end(); 233 I != E; ) { 234 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) { 235 if (LI->getOperand(0) == AI) { 236 LI->replaceAllUsesWith(OnlyStore->getOperand(0)); 237 if (AST && isa<PointerType>(LI->getType())) 238 AST->deleteValue(LI); 239 LI->eraseFromParent(); 240 } 241 } 242 } 243 244 // Finally, remove this block from the UsingBlock set. 245 UsingBlocks[i] = UsingBlocks.back(); 246 --i; --e; 247 } 248 249 // Finally, after the scan, check to see if the store is all that is left. 250 if (UsingBlocks.empty()) { 251 // The alloca has been processed, move on. 252 Allocas[AllocaNum] = Allocas.back(); 253 Allocas.pop_back(); 254 --AllocaNum; 255 continue; 256 } 257 } 258 259 260 if (AST) 261 PointerAllocaValues[AllocaNum] = AllocaPointerVal; 262 263 // If we haven't computed a numbering for the BB's in the function, do so 264 // now. 265 BBNumbers.compute(F); 266 267 // Compute the locations where PhiNodes need to be inserted. Look at the 268 // dominance frontier of EACH basic-block we have a write in. 269 // 270 unsigned CurrentVersion = 0; 271 std::set<PHINode*> InsertedPHINodes; 272 std::vector<unsigned> DFBlocks; 273 while (!DefiningBlocks.empty()) { 274 BasicBlock *BB = DefiningBlocks.back(); 275 DefiningBlocks.pop_back(); 276 277 // Look up the DF for this write, add it to PhiNodes 278 DominanceFrontier::const_iterator it = DF.find(BB); 279 if (it != DF.end()) { 280 const DominanceFrontier::DomSetType &S = it->second; 281 282 // In theory we don't need the indirection through the DFBlocks vector. 283 // In practice, the order of calling QueuePhiNode would depend on the 284 // (unspecified) ordering of basic blocks in the dominance frontier, 285 // which would give PHI nodes non-determinstic subscripts. Fix this by 286 // processing blocks in order of the occurance in the function. 287 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(), 288 PE = S.end(); P != PE; ++P) 289 DFBlocks.push_back(BBNumbers.getNumber(*P)); 290 291 // Sort by which the block ordering in the function. 292 std::sort(DFBlocks.begin(), DFBlocks.end()); 293 294 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) { 295 BasicBlock *BB = BBNumbers.getBlock(DFBlocks[i]); 296 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes)) 297 DefiningBlocks.push_back(BB); 298 } 299 DFBlocks.clear(); 300 } 301 } 302 303 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier 304 // of the writes to the variable, scan through the reads of the variable, 305 // marking PHI nodes which are actually necessary as alive (by removing them 306 // from the InsertedPHINodes set). This is not perfect: there may PHI 307 // marked alive because of loads which are dominated by stores, but there 308 // will be no unmarked PHI nodes which are actually used. 309 // 310 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i) 311 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes); 312 UsingBlocks.clear(); 313 314 // If there are any PHI nodes which are now known to be dead, remove them! 315 for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(), 316 E = InsertedPHINodes.end(); I != E; ++I) { 317 PHINode *PN = *I; 318 std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()]; 319 BBPNs[AllocaNum] = 0; 320 321 // Check to see if we just removed the last inserted PHI node from this 322 // basic block. If so, remove the entry for the basic block. 323 bool HasOtherPHIs = false; 324 for (unsigned i = 0, e = BBPNs.size(); i != e; ++i) 325 if (BBPNs[i]) { 326 HasOtherPHIs = true; 327 break; 328 } 329 if (!HasOtherPHIs) 330 NewPhiNodes.erase(PN->getParent()); 331 332 if (AST && isa<PointerType>(PN->getType())) 333 AST->deleteValue(PN); 334 PN->eraseFromParent(); 335 } 336 337 // Keep the reverse mapping of the 'Allocas' array. 338 AllocaLookup[Allocas[AllocaNum]] = AllocaNum; 339 } 340 341 // Process all allocas which are only used in a single basic block. 342 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I = 343 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){ 344 const std::vector<AllocaInst*> &LocAllocas = I->second; 345 assert(!LocAllocas.empty() && "empty alloca list??"); 346 347 // It's common for there to only be one alloca in the list. Handle it 348 // efficiently. 349 if (LocAllocas.size() == 1) { 350 // If we can do the quick promotion pass, do so now. 351 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0])) 352 RetryList.push_back(LocAllocas[0]); // Failed, retry later. 353 } else { 354 // Locally promote anything possible. Note that if this is unable to 355 // promote a particular alloca, it puts the alloca onto the Allocas vector 356 // for global processing. 357 PromoteLocallyUsedAllocas(I->first, LocAllocas); 358 } 359 } 360 361 if (Allocas.empty()) 362 return; // All of the allocas must have been trivial! 363 364 // Set the incoming values for the basic block to be null values for all of 365 // the alloca's. We do this in case there is a load of a value that has not 366 // been stored yet. In this case, it will get this null value. 367 // 368 std::vector<Value *> Values(Allocas.size()); 369 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) 370 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType()); 371 372 // Walks all basic blocks in the function performing the SSA rename algorithm 373 // and inserting the phi nodes we marked as necessary 374 // 375 RenamePass(F.begin(), 0, Values); 376 377 // The renamer uses the Visited set to avoid infinite loops. Clear it now. 378 Visited.clear(); 379 380 // Remove the allocas themselves from the function. 381 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 382 Instruction *A = Allocas[i]; 383 384 // If there are any uses of the alloca instructions left, they must be in 385 // sections of dead code that were not processed on the dominance frontier. 386 // Just delete the users now. 387 // 388 if (!A->use_empty()) 389 A->replaceAllUsesWith(UndefValue::get(A->getType())); 390 if (AST) AST->deleteValue(A); 391 A->eraseFromParent(); 392 } 393 394 395 // Loop over all of the PHI nodes and see if there are any that we can get 396 // rid of because they merge all of the same incoming values. This can 397 // happen due to undef values coming into the PHI nodes. This process is 398 // iterative, because eliminating one PHI node can cause others to be removed. 399 bool EliminatedAPHI = true; 400 while (EliminatedAPHI) { 401 EliminatedAPHI = false; 402 403 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I = 404 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) { 405 std::vector<PHINode*> &PNs = I->second; 406 for (unsigned i = 0, e = PNs.size(); i != e; ++i) { 407 if (!PNs[i]) continue; 408 409 // If this PHI node merges one value and/or undefs, get the value. 410 if (Value *V = PNs[i]->hasConstantValue(true)) { 411 if (!isa<Instruction>(V) || 412 properlyDominates(cast<Instruction>(V), PNs[i])) { 413 if (AST && isa<PointerType>(PNs[i]->getType())) 414 AST->deleteValue(PNs[i]); 415 PNs[i]->replaceAllUsesWith(V); 416 PNs[i]->eraseFromParent(); 417 PNs[i] = 0; 418 EliminatedAPHI = true; 419 continue; 420 } 421 } 422 } 423 } 424 } 425 426 // At this point, the renamer has added entries to PHI nodes for all reachable 427 // code. Unfortunately, there may be blocks which are not reachable, which 428 // the renamer hasn't traversed. If this is the case, the PHI nodes may not 429 // have incoming values for all predecessors. Loop over all PHI nodes we have 430 // created, inserting undef values if they are missing any incoming values. 431 // 432 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I = 433 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) { 434 435 std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first)); 436 std::vector<PHINode*> &PNs = I->second; 437 assert(!PNs.empty() && "Empty PHI node list??"); 438 PHINode *SomePHI = 0; 439 for (unsigned i = 0, e = PNs.size(); i != e; ++i) 440 if (PNs[i]) { 441 SomePHI = PNs[i]; 442 break; 443 } 444 445 // Only do work here if there the PHI nodes are missing incoming values. We 446 // know that all PHI nodes that were inserted in a block will have the same 447 // number of incoming values, so we can just check any PHI node. 448 if (SomePHI && Preds.size() != SomePHI->getNumIncomingValues()) { 449 // Ok, now we know that all of the PHI nodes are missing entries for some 450 // basic blocks. Start by sorting the incoming predecessors for efficient 451 // access. 452 std::sort(Preds.begin(), Preds.end()); 453 454 // Now we loop through all BB's which have entries in SomePHI and remove 455 // them from the Preds list. 456 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) { 457 // Do a log(n) search of the Preds list for the entry we want. 458 std::vector<BasicBlock*>::iterator EntIt = 459 std::lower_bound(Preds.begin(), Preds.end(), 460 SomePHI->getIncomingBlock(i)); 461 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&& 462 "PHI node has entry for a block which is not a predecessor!"); 463 464 // Remove the entry 465 Preds.erase(EntIt); 466 } 467 468 // At this point, the blocks left in the preds list must have dummy 469 // entries inserted into every PHI nodes for the block. 470 for (unsigned i = 0, e = PNs.size(); i != e; ++i) 471 if (PHINode *PN = PNs[i]) { 472 Value *UndefVal = UndefValue::get(PN->getType()); 473 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred) 474 PN->addIncoming(UndefVal, Preds[pred]); 475 } 476 } 477 } 478} 479 480// MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not 481// "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF 482// as usual (inserting the PHI nodes in the DeadPHINodes set), then processes 483// each read of the variable. For each block that reads the variable, this 484// function is called, which removes used PHI nodes from the DeadPHINodes set. 485// After all of the reads have been processed, any PHI nodes left in the 486// DeadPHINodes set are removed. 487// 488void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum, 489 std::set<PHINode*> &DeadPHINodes) { 490 // Scan the immediate dominators of this block looking for a block which has a 491 // PHI node for Alloca num. If we find it, mark the PHI node as being alive! 492 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) { 493 BasicBlock *DomBB = N->getBlock(); 494 std::map<BasicBlock*, std::vector<PHINode*> >::iterator 495 I = NewPhiNodes.find(DomBB); 496 if (I != NewPhiNodes.end() && I->second[AllocaNum]) { 497 // Ok, we found an inserted PHI node which dominates this value. 498 PHINode *DominatingPHI = I->second[AllocaNum]; 499 500 // Find out if we previously thought it was dead. 501 std::set<PHINode*>::iterator DPNI = DeadPHINodes.find(DominatingPHI); 502 if (DPNI != DeadPHINodes.end()) { 503 // Ok, until now, we thought this PHI node was dead. Mark it as being 504 // alive/needed. 505 DeadPHINodes.erase(DPNI); 506 507 // Now that we have marked the PHI node alive, also mark any PHI nodes 508 // which it might use as being alive as well. 509 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB); 510 PI != PE; ++PI) 511 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes); 512 } 513 } 514 } 515} 516 517/// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic 518/// block. If this is the case, avoid traversing the CFG and inserting a lot of 519/// potentially useless PHI nodes by just performing a single linear pass over 520/// the basic block using the Alloca. 521/// 522/// If we cannot promote this alloca (because it is read before it is written), 523/// return true. This is necessary in cases where, due to control flow, the 524/// alloca is potentially undefined on some control flow paths. e.g. code like 525/// this is potentially correct: 526/// 527/// for (...) { if (c) { A = undef; undef = B; } } 528/// 529/// ... so long as A is not used before undef is set. 530/// 531bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) { 532 assert(!AI->use_empty() && "There are no uses of the alloca!"); 533 534 // Handle degenerate cases quickly. 535 if (AI->hasOneUse()) { 536 Instruction *U = cast<Instruction>(AI->use_back()); 537 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 538 // Must be a load of uninitialized value. 539 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType())); 540 if (AST && isa<PointerType>(LI->getType())) 541 AST->deleteValue(LI); 542 } else { 543 // Otherwise it must be a store which is never read. 544 assert(isa<StoreInst>(U)); 545 } 546 BB->getInstList().erase(U); 547 } else { 548 // Uses of the uninitialized memory location shall get undef. 549 Value *CurVal = 0; 550 551 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 552 Instruction *Inst = I++; 553 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 554 if (LI->getOperand(0) == AI) { 555 if (!CurVal) return true; // Could not locally promote! 556 557 // Loads just returns the "current value"... 558 LI->replaceAllUsesWith(CurVal); 559 if (AST && isa<PointerType>(LI->getType())) 560 AST->deleteValue(LI); 561 BB->getInstList().erase(LI); 562 } 563 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 564 if (SI->getOperand(1) == AI) { 565 // Store updates the "current value"... 566 CurVal = SI->getOperand(0); 567 BB->getInstList().erase(SI); 568 } 569 } 570 } 571 } 572 573 // After traversing the basic block, there should be no more uses of the 574 // alloca, remove it now. 575 assert(AI->use_empty() && "Uses of alloca from more than one BB??"); 576 if (AST) AST->deleteValue(AI); 577 AI->getParent()->getInstList().erase(AI); 578 return false; 579} 580 581/// PromoteLocallyUsedAllocas - This method is just like 582/// PromoteLocallyUsedAlloca, except that it processes multiple alloca 583/// instructions in parallel. This is important in cases where we have large 584/// basic blocks, as we don't want to rescan the entire basic block for each 585/// alloca which is locally used in it (which might be a lot). 586void PromoteMem2Reg:: 587PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) { 588 std::map<AllocaInst*, Value*> CurValues; 589 for (unsigned i = 0, e = AIs.size(); i != e; ++i) 590 CurValues[AIs[i]] = 0; // Insert with null value 591 592 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 593 Instruction *Inst = I++; 594 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 595 // Is this a load of an alloca we are tracking? 596 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) { 597 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 598 if (AIt != CurValues.end()) { 599 // If loading an uninitialized value, allow the inter-block case to 600 // handle it. Due to control flow, this might actually be ok. 601 if (AIt->second == 0) { // Use of locally uninitialized value?? 602 RetryList.push_back(AI); // Retry elsewhere. 603 CurValues.erase(AIt); // Stop tracking this here. 604 if (CurValues.empty()) return; 605 } else { 606 // Loads just returns the "current value"... 607 LI->replaceAllUsesWith(AIt->second); 608 if (AST && isa<PointerType>(LI->getType())) 609 AST->deleteValue(LI); 610 BB->getInstList().erase(LI); 611 } 612 } 613 } 614 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 615 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) { 616 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 617 if (AIt != CurValues.end()) { 618 // Store updates the "current value"... 619 AIt->second = SI->getOperand(0); 620 BB->getInstList().erase(SI); 621 } 622 } 623 } 624 } 625} 626 627 628 629// QueuePhiNode - queues a phi-node to be added to a basic-block for a specific 630// Alloca returns true if there wasn't already a phi-node for that variable 631// 632bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo, 633 unsigned &Version, 634 std::set<PHINode*> &InsertedPHINodes) { 635 // Look up the basic-block in question. 636 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB]; 637 if (BBPNs.empty()) BBPNs.resize(Allocas.size()); 638 639 // If the BB already has a phi node added for the i'th alloca then we're done! 640 if (BBPNs[AllocaNo]) return false; 641 642 // Create a PhiNode using the dereferenced type... and add the phi-node to the 643 // BasicBlock. 644 PHINode *PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(), 645 Allocas[AllocaNo]->getName() + "." + 646 utostr(Version++), BB->begin()); 647 BBPNs[AllocaNo] = PN; 648 InsertedPHINodes.insert(PN); 649 650 if (AST && isa<PointerType>(PN->getType())) 651 AST->copyValue(PointerAllocaValues[AllocaNo], PN); 652 653 return true; 654} 655 656 657// RenamePass - Recursively traverse the CFG of the function, renaming loads and 658// stores to the allocas which we are promoting. IncomingVals indicates what 659// value each Alloca contains on exit from the predecessor block Pred. 660// 661void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred, 662 std::vector<Value*> &IncomingVals) { 663 664 // If this BB needs a PHI node, update the PHI node for each variable we need 665 // PHI nodes for. 666 std::map<BasicBlock*, std::vector<PHINode *> >::iterator 667 BBPNI = NewPhiNodes.find(BB); 668 if (BBPNI != NewPhiNodes.end()) { 669 std::vector<PHINode *> &BBPNs = BBPNI->second; 670 for (unsigned k = 0; k != BBPNs.size(); ++k) 671 if (PHINode *PN = BBPNs[k]) { 672 // Add this incoming value to the PHI node. 673 PN->addIncoming(IncomingVals[k], Pred); 674 675 // The currently active variable for this block is now the PHI. 676 IncomingVals[k] = PN; 677 } 678 } 679 680 // don't revisit nodes 681 if (Visited.count(BB)) return; 682 683 // mark as visited 684 Visited.insert(BB); 685 686 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) { 687 Instruction *I = II++; // get the instruction, increment iterator 688 689 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 690 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) { 691 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src); 692 if (AI != AllocaLookup.end()) { 693 Value *V = IncomingVals[AI->second]; 694 695 // walk the use list of this load and replace all uses with r 696 LI->replaceAllUsesWith(V); 697 if (AST && isa<PointerType>(LI->getType())) 698 AST->deleteValue(LI); 699 BB->getInstList().erase(LI); 700 } 701 } 702 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 703 // Delete this instruction and mark the name as the current holder of the 704 // value 705 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) { 706 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest); 707 if (ai != AllocaLookup.end()) { 708 // what value were we writing? 709 IncomingVals[ai->second] = SI->getOperand(0); 710 BB->getInstList().erase(SI); 711 } 712 } 713 } 714 } 715 716 // Recurse to our successors. 717 TerminatorInst *TI = BB->getTerminator(); 718 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) { 719 std::vector<Value*> OutgoingVals(IncomingVals); 720 RenamePass(TI->getSuccessor(i), BB, OutgoingVals); 721 } 722} 723 724/// PromoteMemToReg - Promote the specified list of alloca instructions into 725/// scalar registers, inserting PHI nodes as appropriate. This function makes 726/// use of DominanceFrontier information. This function does not modify the CFG 727/// of the function at all. All allocas must be from the same function. 728/// 729/// If AST is specified, the specified tracker is updated to reflect changes 730/// made to the IR. 731/// 732void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas, 733 DominatorTree &DT, DominanceFrontier &DF, 734 const TargetData &TD, AliasSetTracker *AST) { 735 // If there is nothing to do, bail out... 736 if (Allocas.empty()) return; 737 738 std::vector<AllocaInst*> RetryList; 739 PromoteMem2Reg(Allocas, RetryList, DT, DF, TD, AST).run(); 740 741 // PromoteMem2Reg may not have been able to promote all of the allocas in one 742 // pass, run it again if needed. 743 while (!RetryList.empty()) { 744 // If we need to retry some allocas, this is due to there being no store 745 // before a read in a local block. To counteract this, insert a store of 746 // undef into the alloca right after the alloca itself. 747 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) { 748 BasicBlock::iterator BBI = RetryList[i]; 749 750 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()), 751 RetryList[i], ++BBI); 752 } 753 754 std::vector<AllocaInst*> NewAllocas; 755 std::swap(NewAllocas, RetryList); 756 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, TD, AST).run(); 757 } 758} 759