PromoteMemoryToRegister.cpp revision c837615cf0bfef743f98bb7101f27c23f6f21ba1
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/SmallPtrSet.h" 27#include "llvm/ADT/SmallVector.h" 28#include "llvm/ADT/StringExtras.h" 29#include "llvm/Support/CFG.h" 30#include "llvm/Support/StableBasicBlockNumbering.h" 31#include "llvm/Support/Compiler.h" 32#include <algorithm> 33using namespace llvm; 34 35/// isAllocaPromotable - Return true if this alloca is legal for promotion. 36/// This is true if there are only loads and stores to the alloca. 37/// 38bool llvm::isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) { 39 // FIXME: If the memory unit is of pointer or integer type, we can permit 40 // assignments to subsections of the memory unit. 41 42 // Only allow direct loads and stores... 43 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end(); 44 UI != UE; ++UI) // Loop over all of the uses of the alloca 45 if (isa<LoadInst>(*UI)) { 46 // noop 47 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 48 if (SI->getOperand(0) == AI) 49 return false; // Don't allow a store OF the AI, only INTO the AI. 50 } else { 51 return false; // Not a load or store. 52 } 53 54 return true; 55} 56 57namespace { 58 struct VISIBILITY_HIDDEN PromoteMem2Reg { 59 /// Allocas - The alloca instructions being promoted. 60 /// 61 std::vector<AllocaInst*> Allocas; 62 SmallVector<AllocaInst*, 16> &RetryList; 63 DominatorTree &DT; 64 DominanceFrontier &DF; 65 const TargetData &TD; 66 67 /// AST - An AliasSetTracker object to update. If null, don't update it. 68 /// 69 AliasSetTracker *AST; 70 71 /// AllocaLookup - Reverse mapping of Allocas. 72 /// 73 std::map<AllocaInst*, unsigned> AllocaLookup; 74 75 /// NewPhiNodes - The PhiNodes we're adding. 76 /// 77 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes; 78 79 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for 80 /// each alloca that is of pointer type, we keep track of what to copyValue 81 /// to the inserted PHI nodes here. 82 /// 83 std::vector<Value*> PointerAllocaValues; 84 85 /// Visited - The set of basic blocks the renamer has already visited. 86 /// 87 std::set<BasicBlock*> Visited; 88 89 /// BBNumbers - Contains a stable numbering of basic blocks to avoid 90 /// non-determinstic behavior. 91 StableBasicBlockNumbering BBNumbers; 92 93 public: 94 PromoteMem2Reg(const std::vector<AllocaInst*> &A, 95 SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt, 96 DominanceFrontier &df, const TargetData &td, 97 AliasSetTracker *ast) 98 : Allocas(A), RetryList(Retry), DT(dt), DF(df), TD(td), AST(ast) {} 99 100 void run(); 101 102 /// properlyDominates - Return true if I1 properly dominates I2. 103 /// 104 bool properlyDominates(Instruction *I1, Instruction *I2) const { 105 if (InvokeInst *II = dyn_cast<InvokeInst>(I1)) 106 I1 = II->getNormalDest()->begin(); 107 return DT[I1->getParent()]->properlyDominates(DT[I2->getParent()]); 108 } 109 110 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree. 111 /// 112 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const { 113 return DT[BB1]->dominates(DT[BB2]); 114 } 115 116 private: 117 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum, 118 SmallPtrSet<PHINode*, 16> &DeadPHINodes); 119 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI); 120 void PromoteLocallyUsedAllocas(BasicBlock *BB, 121 const std::vector<AllocaInst*> &AIs); 122 123 void RenamePass(BasicBlock *BB, BasicBlock *Pred, 124 std::vector<Value*> &IncVals); 125 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version, 126 SmallPtrSet<PHINode*, 16> &InsertedPHINodes); 127 }; 128} // end of anonymous namespace 129 130void PromoteMem2Reg::run() { 131 Function &F = *DF.getRoot()->getParent(); 132 133 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are 134 // only used in a single basic block. These instructions can be efficiently 135 // promoted by performing a single linear scan over that one block. Since 136 // individual basic blocks are sometimes large, we group together all allocas 137 // that are live in a single basic block by the basic block they are live in. 138 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas; 139 140 if (AST) PointerAllocaValues.resize(Allocas.size()); 141 142 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) { 143 AllocaInst *AI = Allocas[AllocaNum]; 144 145 assert(isAllocaPromotable(AI, TD) && 146 "Cannot promote non-promotable alloca!"); 147 assert(AI->getParent()->getParent() == &F && 148 "All allocas should be in the same function, which is same as DF!"); 149 150 if (AI->use_empty()) { 151 // If there are no uses of the alloca, just delete it now. 152 if (AST) AST->deleteValue(AI); 153 AI->eraseFromParent(); 154 155 // Remove the alloca from the Allocas list, since it has been processed 156 Allocas[AllocaNum] = Allocas.back(); 157 Allocas.pop_back(); 158 --AllocaNum; 159 continue; 160 } 161 162 // Calculate the set of read and write-locations for each alloca. This is 163 // analogous to finding the 'uses' and 'definitions' of each variable. 164 std::vector<BasicBlock*> DefiningBlocks; 165 std::vector<BasicBlock*> UsingBlocks; 166 167 StoreInst *OnlyStore = 0; 168 BasicBlock *OnlyBlock = 0; 169 bool OnlyUsedInOneBlock = true; 170 171 // As we scan the uses of the alloca instruction, keep track of stores, and 172 // decide whether all of the loads and stores to the alloca are within the 173 // same basic block. 174 Value *AllocaPointerVal = 0; 175 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){ 176 Instruction *User = cast<Instruction>(*U); 177 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 178 // Remember the basic blocks which define new values for the alloca 179 DefiningBlocks.push_back(SI->getParent()); 180 AllocaPointerVal = SI->getOperand(0); 181 OnlyStore = SI; 182 } else { 183 LoadInst *LI = cast<LoadInst>(User); 184 // Otherwise it must be a load instruction, keep track of variable reads 185 UsingBlocks.push_back(LI->getParent()); 186 AllocaPointerVal = LI; 187 } 188 189 if (OnlyUsedInOneBlock) { 190 if (OnlyBlock == 0) 191 OnlyBlock = User->getParent(); 192 else if (OnlyBlock != User->getParent()) 193 OnlyUsedInOneBlock = false; 194 } 195 } 196 197 // If the alloca is only read and written in one basic block, just perform a 198 // linear sweep over the block to eliminate it. 199 if (OnlyUsedInOneBlock) { 200 LocallyUsedAllocas[OnlyBlock].push_back(AI); 201 202 // Remove the alloca from the Allocas list, since it will be processed. 203 Allocas[AllocaNum] = Allocas.back(); 204 Allocas.pop_back(); 205 --AllocaNum; 206 continue; 207 } 208 209 // If there is only a single store to this value, replace any loads of 210 // it that are directly dominated by the definition with the value stored. 211 if (DefiningBlocks.size() == 1) { 212 // Be aware of loads before the store. 213 std::set<BasicBlock*> ProcessedBlocks; 214 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i) 215 // If the store dominates the block and if we haven't processed it yet, 216 // do so now. 217 if (dominates(OnlyStore->getParent(), UsingBlocks[i])) 218 if (ProcessedBlocks.insert(UsingBlocks[i]).second) { 219 BasicBlock *UseBlock = UsingBlocks[i]; 220 221 // If the use and store are in the same block, do a quick scan to 222 // verify that there are no uses before the store. 223 if (UseBlock == OnlyStore->getParent()) { 224 BasicBlock::iterator I = UseBlock->begin(); 225 for (; &*I != OnlyStore; ++I) { // scan block for store. 226 if (isa<LoadInst>(I) && I->getOperand(0) == AI) 227 break; 228 } 229 if (&*I != OnlyStore) break; // Do not handle this case. 230 } 231 232 // Otherwise, if this is a different block or if all uses happen 233 // after the store, do a simple linear scan to replace loads with 234 // the stored value. 235 for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end(); 236 I != E; ) { 237 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) { 238 if (LI->getOperand(0) == AI) { 239 LI->replaceAllUsesWith(OnlyStore->getOperand(0)); 240 if (AST && isa<PointerType>(LI->getType())) 241 AST->deleteValue(LI); 242 LI->eraseFromParent(); 243 } 244 } 245 } 246 247 // Finally, remove this block from the UsingBlock set. 248 UsingBlocks[i] = UsingBlocks.back(); 249 --i; --e; 250 } 251 252 // Finally, after the scan, check to see if the store is all that is left. 253 if (UsingBlocks.empty()) { 254 // The alloca has been processed, move on. 255 Allocas[AllocaNum] = Allocas.back(); 256 Allocas.pop_back(); 257 --AllocaNum; 258 continue; 259 } 260 } 261 262 263 if (AST) 264 PointerAllocaValues[AllocaNum] = AllocaPointerVal; 265 266 // If we haven't computed a numbering for the BB's in the function, do so 267 // now. 268 BBNumbers.compute(F); 269 270 // Compute the locations where PhiNodes need to be inserted. Look at the 271 // dominance frontier of EACH basic-block we have a write in. 272 // 273 unsigned CurrentVersion = 0; 274 SmallPtrSet<PHINode*, 16> InsertedPHINodes; 275 std::vector<unsigned> DFBlocks; 276 while (!DefiningBlocks.empty()) { 277 BasicBlock *BB = DefiningBlocks.back(); 278 DefiningBlocks.pop_back(); 279 280 // Look up the DF for this write, add it to PhiNodes 281 DominanceFrontier::const_iterator it = DF.find(BB); 282 if (it != DF.end()) { 283 const DominanceFrontier::DomSetType &S = it->second; 284 285 // In theory we don't need the indirection through the DFBlocks vector. 286 // In practice, the order of calling QueuePhiNode would depend on the 287 // (unspecified) ordering of basic blocks in the dominance frontier, 288 // which would give PHI nodes non-determinstic subscripts. Fix this by 289 // processing blocks in order of the occurance in the function. 290 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(), 291 PE = S.end(); P != PE; ++P) 292 DFBlocks.push_back(BBNumbers.getNumber(*P)); 293 294 // Sort by which the block ordering in the function. 295 std::sort(DFBlocks.begin(), DFBlocks.end()); 296 297 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) { 298 BasicBlock *BB = BBNumbers.getBlock(DFBlocks[i]); 299 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes)) 300 DefiningBlocks.push_back(BB); 301 } 302 DFBlocks.clear(); 303 } 304 } 305 306 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier 307 // of the writes to the variable, scan through the reads of the variable, 308 // marking PHI nodes which are actually necessary as alive (by removing them 309 // from the InsertedPHINodes set). This is not perfect: there may PHI 310 // marked alive because of loads which are dominated by stores, but there 311 // will be no unmarked PHI nodes which are actually used. 312 // 313 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i) 314 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes); 315 UsingBlocks.clear(); 316 317 // If there are any PHI nodes which are now known to be dead, remove them! 318 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(), 319 E = InsertedPHINodes.end(); I != E; ++I) { 320 PHINode *PN = *I; 321 std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()]; 322 BBPNs[AllocaNum] = 0; 323 324 // Check to see if we just removed the last inserted PHI node from this 325 // basic block. If so, remove the entry for the basic block. 326 bool HasOtherPHIs = false; 327 for (unsigned i = 0, e = BBPNs.size(); i != e; ++i) 328 if (BBPNs[i]) { 329 HasOtherPHIs = true; 330 break; 331 } 332 if (!HasOtherPHIs) 333 NewPhiNodes.erase(PN->getParent()); 334 335 if (AST && isa<PointerType>(PN->getType())) 336 AST->deleteValue(PN); 337 PN->eraseFromParent(); 338 } 339 340 // Keep the reverse mapping of the 'Allocas' array. 341 AllocaLookup[Allocas[AllocaNum]] = AllocaNum; 342 } 343 344 // Process all allocas which are only used in a single basic block. 345 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I = 346 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){ 347 const std::vector<AllocaInst*> &LocAllocas = I->second; 348 assert(!LocAllocas.empty() && "empty alloca list??"); 349 350 // It's common for there to only be one alloca in the list. Handle it 351 // efficiently. 352 if (LocAllocas.size() == 1) { 353 // If we can do the quick promotion pass, do so now. 354 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0])) 355 RetryList.push_back(LocAllocas[0]); // Failed, retry later. 356 } else { 357 // Locally promote anything possible. Note that if this is unable to 358 // promote a particular alloca, it puts the alloca onto the Allocas vector 359 // for global processing. 360 PromoteLocallyUsedAllocas(I->first, LocAllocas); 361 } 362 } 363 364 if (Allocas.empty()) 365 return; // All of the allocas must have been trivial! 366 367 // Set the incoming values for the basic block to be null values for all of 368 // the alloca's. We do this in case there is a load of a value that has not 369 // been stored yet. In this case, it will get this null value. 370 // 371 std::vector<Value *> Values(Allocas.size()); 372 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) 373 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType()); 374 375 // Walks all basic blocks in the function performing the SSA rename algorithm 376 // and inserting the phi nodes we marked as necessary 377 // 378 RenamePass(F.begin(), 0, Values); 379 380 // The renamer uses the Visited set to avoid infinite loops. Clear it now. 381 Visited.clear(); 382 383 // Remove the allocas themselves from the function. 384 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 385 Instruction *A = Allocas[i]; 386 387 // If there are any uses of the alloca instructions left, they must be in 388 // sections of dead code that were not processed on the dominance frontier. 389 // Just delete the users now. 390 // 391 if (!A->use_empty()) 392 A->replaceAllUsesWith(UndefValue::get(A->getType())); 393 if (AST) AST->deleteValue(A); 394 A->eraseFromParent(); 395 } 396 397 398 // Loop over all of the PHI nodes and see if there are any that we can get 399 // rid of because they merge all of the same incoming values. This can 400 // happen due to undef values coming into the PHI nodes. This process is 401 // iterative, because eliminating one PHI node can cause others to be removed. 402 bool EliminatedAPHI = true; 403 while (EliminatedAPHI) { 404 EliminatedAPHI = false; 405 406 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I = 407 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) { 408 std::vector<PHINode*> &PNs = I->second; 409 for (unsigned i = 0, e = PNs.size(); i != e; ++i) { 410 if (!PNs[i]) continue; 411 412 // If this PHI node merges one value and/or undefs, get the value. 413 if (Value *V = PNs[i]->hasConstantValue(true)) { 414 if (!isa<Instruction>(V) || 415 properlyDominates(cast<Instruction>(V), PNs[i])) { 416 if (AST && isa<PointerType>(PNs[i]->getType())) 417 AST->deleteValue(PNs[i]); 418 PNs[i]->replaceAllUsesWith(V); 419 PNs[i]->eraseFromParent(); 420 PNs[i] = 0; 421 EliminatedAPHI = true; 422 continue; 423 } 424 } 425 } 426 } 427 } 428 429 // At this point, the renamer has added entries to PHI nodes for all reachable 430 // code. Unfortunately, there may be unreachable blocks which the renamer 431 // hasn't traversed. If this is the case, the PHI nodes may not 432 // have incoming values for all predecessors. Loop over all PHI nodes we have 433 // created, inserting undef values if they are missing any incoming values. 434 // 435 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I = 436 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) { 437 438 std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first)); 439 std::vector<PHINode*> &PNs = I->second; 440 assert(!PNs.empty() && "Empty PHI node list??"); 441 PHINode *SomePHI = 0; 442 for (unsigned i = 0, e = PNs.size(); i != e; ++i) 443 if (PNs[i]) { 444 SomePHI = PNs[i]; 445 break; 446 } 447 448 // Only do work here if there the PHI nodes are missing incoming values. We 449 // know that all PHI nodes that were inserted in a block will have the same 450 // number of incoming values, so we can just check any PHI node. 451 if (SomePHI && Preds.size() != SomePHI->getNumIncomingValues()) { 452 // Ok, now we know that all of the PHI nodes are missing entries for some 453 // basic blocks. Start by sorting the incoming predecessors for efficient 454 // access. 455 std::sort(Preds.begin(), Preds.end()); 456 457 // Now we loop through all BB's which have entries in SomePHI and remove 458 // them from the Preds list. 459 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) { 460 // Do a log(n) search of the Preds list for the entry we want. 461 std::vector<BasicBlock*>::iterator EntIt = 462 std::lower_bound(Preds.begin(), Preds.end(), 463 SomePHI->getIncomingBlock(i)); 464 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&& 465 "PHI node has entry for a block which is not a predecessor!"); 466 467 // Remove the entry 468 Preds.erase(EntIt); 469 } 470 471 // At this point, the blocks left in the preds list must have dummy 472 // entries inserted into every PHI nodes for the block. 473 for (unsigned i = 0, e = PNs.size(); i != e; ++i) 474 if (PHINode *PN = PNs[i]) { 475 Value *UndefVal = UndefValue::get(PN->getType()); 476 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred) 477 PN->addIncoming(UndefVal, Preds[pred]); 478 } 479 } 480 } 481} 482 483// MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not 484// "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF 485// as usual (inserting the PHI nodes in the DeadPHINodes set), then processes 486// each read of the variable. For each block that reads the variable, this 487// function is called, which removes used PHI nodes from the DeadPHINodes set. 488// After all of the reads have been processed, any PHI nodes left in the 489// DeadPHINodes set are removed. 490// 491void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum, 492 SmallPtrSet<PHINode*, 16> &DeadPHINodes) { 493 // Scan the immediate dominators of this block looking for a block which has a 494 // PHI node for Alloca num. If we find it, mark the PHI node as being alive! 495 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) { 496 BasicBlock *DomBB = N->getBlock(); 497 std::map<BasicBlock*, std::vector<PHINode*> >::iterator 498 I = NewPhiNodes.find(DomBB); 499 if (I != NewPhiNodes.end() && I->second[AllocaNum]) { 500 // Ok, we found an inserted PHI node which dominates this value. 501 PHINode *DominatingPHI = I->second[AllocaNum]; 502 503 // Find out if we previously thought it was dead. If so, mark it as being 504 // live by removing it from the DeadPHINodes set. 505 if (DeadPHINodes.erase(DominatingPHI)) { 506 // Now that we have marked the PHI node alive, also mark any PHI nodes 507 // which it might use as being alive as well. 508 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB); 509 PI != PE; ++PI) 510 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes); 511 } 512 } 513 } 514} 515 516/// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic 517/// block. If this is the case, avoid traversing the CFG and inserting a lot of 518/// potentially useless PHI nodes by just performing a single linear pass over 519/// the basic block using the Alloca. 520/// 521/// If we cannot promote this alloca (because it is read before it is written), 522/// return true. This is necessary in cases where, due to control flow, the 523/// alloca is potentially undefined on some control flow paths. e.g. code like 524/// this is potentially correct: 525/// 526/// for (...) { if (c) { A = undef; undef = B; } } 527/// 528/// ... so long as A is not used before undef is set. 529/// 530bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) { 531 assert(!AI->use_empty() && "There are no uses of the alloca!"); 532 533 // Handle degenerate cases quickly. 534 if (AI->hasOneUse()) { 535 Instruction *U = cast<Instruction>(AI->use_back()); 536 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 537 // Must be a load of uninitialized value. 538 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType())); 539 if (AST && isa<PointerType>(LI->getType())) 540 AST->deleteValue(LI); 541 } else { 542 // Otherwise it must be a store which is never read. 543 assert(isa<StoreInst>(U)); 544 } 545 BB->getInstList().erase(U); 546 } else { 547 // Uses of the uninitialized memory location shall get undef. 548 Value *CurVal = 0; 549 550 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 551 Instruction *Inst = I++; 552 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 553 if (LI->getOperand(0) == AI) { 554 if (!CurVal) return true; // Could not locally promote! 555 556 // Loads just returns the "current value"... 557 LI->replaceAllUsesWith(CurVal); 558 if (AST && isa<PointerType>(LI->getType())) 559 AST->deleteValue(LI); 560 BB->getInstList().erase(LI); 561 } 562 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 563 if (SI->getOperand(1) == AI) { 564 // Store updates the "current value"... 565 CurVal = SI->getOperand(0); 566 BB->getInstList().erase(SI); 567 } 568 } 569 } 570 } 571 572 // After traversing the basic block, there should be no more uses of the 573 // alloca, remove it now. 574 assert(AI->use_empty() && "Uses of alloca from more than one BB??"); 575 if (AST) AST->deleteValue(AI); 576 AI->getParent()->getInstList().erase(AI); 577 return false; 578} 579 580/// PromoteLocallyUsedAllocas - This method is just like 581/// PromoteLocallyUsedAlloca, except that it processes multiple alloca 582/// instructions in parallel. This is important in cases where we have large 583/// basic blocks, as we don't want to rescan the entire basic block for each 584/// alloca which is locally used in it (which might be a lot). 585void PromoteMem2Reg:: 586PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) { 587 std::map<AllocaInst*, Value*> CurValues; 588 for (unsigned i = 0, e = AIs.size(); i != e; ++i) 589 CurValues[AIs[i]] = 0; // Insert with null value 590 591 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 592 Instruction *Inst = I++; 593 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 594 // Is this a load of an alloca we are tracking? 595 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) { 596 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 597 if (AIt != CurValues.end()) { 598 // If loading an uninitialized value, allow the inter-block case to 599 // handle it. Due to control flow, this might actually be ok. 600 if (AIt->second == 0) { // Use of locally uninitialized value?? 601 RetryList.push_back(AI); // Retry elsewhere. 602 CurValues.erase(AIt); // Stop tracking this here. 603 if (CurValues.empty()) return; 604 } else { 605 // Loads just returns the "current value"... 606 LI->replaceAllUsesWith(AIt->second); 607 if (AST && isa<PointerType>(LI->getType())) 608 AST->deleteValue(LI); 609 BB->getInstList().erase(LI); 610 } 611 } 612 } 613 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 614 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) { 615 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI); 616 if (AIt != CurValues.end()) { 617 // Store updates the "current value"... 618 AIt->second = SI->getOperand(0); 619 BB->getInstList().erase(SI); 620 } 621 } 622 } 623 } 624} 625 626 627 628// QueuePhiNode - queues a phi-node to be added to a basic-block for a specific 629// Alloca returns true if there wasn't already a phi-node for that variable 630// 631bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo, 632 unsigned &Version, 633 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) { 634 // Look up the basic-block in question. 635 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB]; 636 if (BBPNs.empty()) BBPNs.resize(Allocas.size()); 637 638 // If the BB already has a phi node added for the i'th alloca then we're done! 639 if (BBPNs[AllocaNo]) return false; 640 641 // Create a PhiNode using the dereferenced type... and add the phi-node to the 642 // BasicBlock. 643 PHINode *PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(), 644 Allocas[AllocaNo]->getName() + "." + 645 utostr(Version++), BB->begin()); 646 BBPNs[AllocaNo] = PN; 647 InsertedPHINodes.insert(PN); 648 649 if (AST && isa<PointerType>(PN->getType())) 650 AST->copyValue(PointerAllocaValues[AllocaNo], PN); 651 652 return true; 653} 654 655 656// RenamePass - Recursively traverse the CFG of the function, renaming loads and 657// stores to the allocas which we are promoting. IncomingVals indicates what 658// value each Alloca contains on exit from the predecessor block Pred. 659// 660void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred, 661 std::vector<Value*> &IncomingVals) { 662 663 // If this BB needs a PHI node, update the PHI node for each variable we need 664 // PHI nodes for. 665 std::map<BasicBlock*, std::vector<PHINode *> >::iterator 666 BBPNI = NewPhiNodes.find(BB); 667 if (BBPNI != NewPhiNodes.end()) { 668 std::vector<PHINode *> &BBPNs = BBPNI->second; 669 for (unsigned k = 0; k != BBPNs.size(); ++k) 670 if (PHINode *PN = BBPNs[k]) { 671 // Add this incoming value to the PHI node. 672 PN->addIncoming(IncomingVals[k], Pred); 673 674 // The currently active variable for this block is now the PHI. 675 IncomingVals[k] = PN; 676 } 677 } 678 679 // don't revisit nodes 680 if (Visited.count(BB)) return; 681 682 // mark as visited 683 Visited.insert(BB); 684 685 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) { 686 Instruction *I = II++; // get the instruction, increment iterator 687 688 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 689 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) { 690 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src); 691 if (AI != AllocaLookup.end()) { 692 Value *V = IncomingVals[AI->second]; 693 694 // walk the use list of this load and replace all uses with r 695 LI->replaceAllUsesWith(V); 696 if (AST && isa<PointerType>(LI->getType())) 697 AST->deleteValue(LI); 698 BB->getInstList().erase(LI); 699 } 700 } 701 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 702 // Delete this instruction and mark the name as the current holder of the 703 // value 704 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) { 705 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest); 706 if (ai != AllocaLookup.end()) { 707 // what value were we writing? 708 IncomingVals[ai->second] = SI->getOperand(0); 709 BB->getInstList().erase(SI); 710 } 711 } 712 } 713 } 714 715 // Recurse to our successors. 716 TerminatorInst *TI = BB->getTerminator(); 717 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) { 718 std::vector<Value*> OutgoingVals(IncomingVals); 719 RenamePass(TI->getSuccessor(i), BB, OutgoingVals); 720 } 721} 722 723/// PromoteMemToReg - Promote the specified list of alloca instructions into 724/// scalar registers, inserting PHI nodes as appropriate. This function makes 725/// use of DominanceFrontier information. This function does not modify the CFG 726/// of the function at all. All allocas must be from the same function. 727/// 728/// If AST is specified, the specified tracker is updated to reflect changes 729/// made to the IR. 730/// 731void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas, 732 DominatorTree &DT, DominanceFrontier &DF, 733 const TargetData &TD, AliasSetTracker *AST) { 734 // If there is nothing to do, bail out... 735 if (Allocas.empty()) return; 736 737 SmallVector<AllocaInst*, 16> RetryList; 738 PromoteMem2Reg(Allocas, RetryList, DT, DF, TD, AST).run(); 739 740 // PromoteMem2Reg may not have been able to promote all of the allocas in one 741 // pass, run it again if needed. 742 std::vector<AllocaInst*> NewAllocas; 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 NewAllocas.assign(RetryList.begin(), RetryList.end()); 755 RetryList.clear(); 756 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, TD, AST).run(); 757 NewAllocas.clear(); 758 } 759} 760