LICM.cpp revision caadc93f7b3e950c16ff6cd497baffe7cadc1477
1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// 2// 3// This pass is a simple loop invariant code motion pass. An interesting aspect 4// of this pass is that it uses alias analysis for two purposes: 5// 6// 1. Moving loop invariant loads out of loops. If we can determine that a 7// load inside of a loop never aliases anything stored to, we can hoist it 8// like any other instruction. 9// 2. Scalar Promotion of Memory - If there is a store instruction inside of 10// the loop, we try to move the store to happen AFTER the loop instead of 11// inside of the loop. This can only happen if a few conditions are true: 12// A. The pointer stored through is loop invariant 13// B. There are no stores or loads in the loop which _may_ alias the 14// pointer. There are no calls in the loop which mod/ref the pointer. 15// If these conditions are true, we can promote the loads and stores in the 16// loop of the pointer to use a temporary alloca'd variable. We then use 17// the mem2reg functionality to construct the appropriate SSA form for the 18// variable. 19// 20//===----------------------------------------------------------------------===// 21 22#include "llvm/Transforms/Scalar.h" 23#include "llvm/Transforms/Utils/PromoteMemToReg.h" 24#include "llvm/Transforms/Utils/Local.h" 25#include "llvm/Analysis/LoopInfo.h" 26#include "llvm/Analysis/AliasAnalysis.h" 27#include "llvm/Analysis/Dominators.h" 28#include "llvm/Instructions.h" 29#include "llvm/DerivedTypes.h" 30#include "llvm/Support/InstVisitor.h" 31#include "llvm/Support/CFG.h" 32#include "Support/Statistic.h" 33#include "Support/CommandLine.h" 34#include "llvm/Assembly/Writer.h" 35#include <algorithm> 36 37namespace { 38 cl::opt<bool> DisablePromotion("disable-licm-promotion", cl::Hidden, 39 cl::desc("Disable memory promotion in LICM pass")); 40 41 Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop"); 42 Statistic<> NumHoistedLoads("licm", "Number of load insts hoisted"); 43 Statistic<> NumPromoted("licm", "Number of memory locations promoted to registers"); 44 45 /// LoopBodyInfo - We recursively traverse loops from most-deeply-nested to 46 /// least-deeply-nested. For all of the loops nested within the current one, 47 /// we keep track of information so that we don't have to repeat queries. 48 /// 49 struct LoopBodyInfo { 50 std::vector<CallInst*> Calls; // Call instructions in loop 51 std::vector<InvokeInst*> Invokes; // Invoke instructions in loop 52 53 // StoredPointers - Targets of store instructions... 54 std::set<Value*> StoredPointers; 55 56 // LoadedPointers - Source pointers for load instructions... 57 std::set<Value*> LoadedPointers; 58 59 enum PointerClass { 60 PointerUnknown = 0, // Nothing is known about this pointer yet 61 PointerMustStore, // Memory is stored to ONLY through this pointer 62 PointerMayStore, // Memory is stored to through this or other pointers 63 PointerNoStore // Memory is not modified in this loop 64 }; 65 66 // PointerIsModified - Keep track of information as we find out about it in 67 // the loop body... 68 // 69 std::map<Value*, enum PointerClass> PointerIsModified; 70 71 /// CantModifyAnyPointers - Return true if no memory modifying instructions 72 /// occur in this loop. This is just a conservative approximation, because 73 /// a call may not actually store anything. 74 bool CantModifyAnyPointers() const { 75 return Calls.empty() && Invokes.empty() && StoredPointers.empty(); 76 } 77 78 /// incorporate - Incorporate information about a subloop into the current 79 /// loop. 80 void incorporate(const LoopBodyInfo &OtherLBI); 81 void incorporate(BasicBlock &BB); // do the same for a basic block 82 83 PointerClass getPointerInfo(Value *V, AliasAnalysis &AA) { 84 PointerClass &VInfo = PointerIsModified[V]; 85 if (VInfo == PointerUnknown) 86 VInfo = calculatePointerInfo(V, AA); 87 return VInfo; 88 } 89 private: 90 /// calculatePointerInfo - Calculate information about the specified 91 /// pointer. 92 PointerClass calculatePointerInfo(Value *V, AliasAnalysis &AA) const; 93 }; 94} 95 96/// incorporate - Incorporate information about a subloop into the current loop. 97void LoopBodyInfo::incorporate(const LoopBodyInfo &OtherLBI) { 98 // Do not incorporate NonModifiedPointers (which is just a cache) because it 99 // is too much trouble to make sure it's still valid. 100 Calls.insert (Calls.end(), OtherLBI.Calls.begin(), OtherLBI.Calls.end()); 101 Invokes.insert(Invokes.end(),OtherLBI.Invokes.begin(),OtherLBI.Invokes.end()); 102 StoredPointers.insert(OtherLBI.StoredPointers.begin(), 103 OtherLBI.StoredPointers.end()); 104 LoadedPointers.insert(OtherLBI.LoadedPointers.begin(), 105 OtherLBI.LoadedPointers.end()); 106} 107 108void LoopBodyInfo::incorporate(BasicBlock &BB) { 109 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 110 if (CallInst *CI = dyn_cast<CallInst>(&*I)) 111 Calls.push_back(CI); 112 else if (StoreInst *SI = dyn_cast<StoreInst>(&*I)) 113 StoredPointers.insert(SI->getOperand(1)); 114 else if (LoadInst *LI = dyn_cast<LoadInst>(&*I)) 115 LoadedPointers.insert(LI->getOperand(0)); 116 117 if (InvokeInst *II = dyn_cast<InvokeInst>(BB.getTerminator())) 118 Invokes.push_back(II); 119} 120 121 122// calculatePointerInfo - Calculate information about the specified pointer. 123LoopBodyInfo::PointerClass LoopBodyInfo::calculatePointerInfo(Value *V, 124 AliasAnalysis &AA) const { 125 for (unsigned i = 0, e = Calls.size(); i != e; ++i) 126 if (AA.getModRefInfo(Calls[i], V, ~0)) 127 return PointerMayStore; 128 129 for (unsigned i = 0, e = Invokes.size(); i != e; ++i) 130 if (AA.getModRefInfo(Invokes[i], V, ~0)) 131 return PointerMayStore; 132 133 PointerClass Result = PointerNoStore; 134 for (std::set<Value*>::const_iterator I = StoredPointers.begin(), 135 E = StoredPointers.end(); I != E; ++I) 136 if (AA.alias(V, ~0, *I, ~0)) 137 if (V == *I) 138 Result = PointerMustStore; // If this is the only alias, return must 139 else 140 return PointerMayStore; // We have to return may 141 return Result; 142} 143 144namespace { 145 struct LICM : public FunctionPass, public InstVisitor<LICM> { 146 virtual bool runOnFunction(Function &F); 147 148 /// This transformation requires natural loop information & requires that 149 /// loop preheaders be inserted into the CFG... 150 /// 151 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 152 AU.setPreservesCFG(); 153 AU.addRequiredID(LoopPreheadersID); 154 AU.addRequired<LoopInfo>(); 155 AU.addRequired<DominatorTree>(); 156 AU.addRequired<DominanceFrontier>(); 157 AU.addRequired<AliasAnalysis>(); 158 } 159 160 private: 161 LoopInfo *LI; // Current LoopInfo 162 AliasAnalysis *AA; // Current AliasAnalysis information 163 bool Changed; // Set to true when we change anything. 164 BasicBlock *Preheader; // The preheader block of the current loop... 165 Loop *CurLoop; // The current loop we are working on... 166 LoopBodyInfo *CurLBI; // Information about the current loop... 167 168 /// visitLoop - Hoist expressions out of the specified loop... 169 /// 170 void visitLoop(Loop *L, LoopBodyInfo &LBI); 171 172 /// HoistRegion - Walk the specified region of the CFG (defined by all 173 /// blocks dominated by the specified block, and that are in the current 174 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 175 /// visit defintions before uses, allowing us to hoist a loop body in one 176 /// pass without iteration. 177 /// 178 void HoistRegion(DominatorTree::Node *N); 179 180 /// inSubLoop - Little predicate that returns true if the specified basic 181 /// block is in a subloop of the current one, not the current one itself. 182 /// 183 bool inSubLoop(BasicBlock *BB) { 184 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 185 for (unsigned i = 0, e = CurLoop->getSubLoops().size(); i != e; ++i) 186 if (CurLoop->getSubLoops()[i]->contains(BB)) 187 return true; // A subloop actually contains this block! 188 return false; 189 } 190 191 /// hoist - When an instruction is found to only use loop invariant operands 192 /// that is safe to hoist, this instruction is called to do the dirty work. 193 /// 194 void hoist(Instruction &I); 195 196 /// pointerInvalidatedByLoop - Return true if the body of this loop may 197 /// store into the memory location pointed to by V. 198 /// 199 bool pointerInvalidatedByLoop(Value *V) { 200 // Check to see if any of the basic blocks in CurLoop invalidate V. 201 return CurLBI->getPointerInfo(V, *AA) != LoopBodyInfo::PointerNoStore; 202 } 203 204 /// isLoopInvariant - Return true if the specified value is loop invariant 205 /// 206 inline bool isLoopInvariant(Value *V) { 207 if (Instruction *I = dyn_cast<Instruction>(V)) 208 return !CurLoop->contains(I->getParent()); 209 return true; // All non-instructions are loop invariant 210 } 211 212 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 213 /// to scalars as we can. 214 /// 215 void PromoteValuesInLoop(); 216 217 /// findPromotableValuesInLoop - Check the current loop for stores to 218 /// definate pointers, which are not loaded and stored through may aliases. 219 /// If these are found, create an alloca for the value, add it to the 220 /// PromotedValues list, and keep track of the mapping from value to 221 /// alloca... 222 /// 223 void findPromotableValuesInLoop( 224 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 225 std::map<Value*, AllocaInst*> &Val2AlMap); 226 227 228 /// Instruction visitation handlers... these basically control whether or 229 /// not the specified instruction types are hoisted. 230 /// 231 friend class InstVisitor<LICM>; 232 void visitBinaryOperator(Instruction &I) { 233 if (isLoopInvariant(I.getOperand(0)) && isLoopInvariant(I.getOperand(1))) 234 hoist(I); 235 } 236 void visitCastInst(CastInst &CI) { 237 Instruction &I = (Instruction&)CI; 238 if (isLoopInvariant(I.getOperand(0))) hoist(I); 239 } 240 void visitShiftInst(ShiftInst &I) { visitBinaryOperator((Instruction&)I); } 241 242 void visitLoadInst(LoadInst &LI); 243 244 void visitGetElementPtrInst(GetElementPtrInst &GEPI) { 245 Instruction &I = (Instruction&)GEPI; 246 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 247 if (!isLoopInvariant(I.getOperand(i))) return; 248 hoist(I); 249 } 250 }; 251 252 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion"); 253} 254 255Pass *createLICMPass() { return new LICM(); } 256 257/// runOnFunction - For LICM, this simply traverses the loop structure of the 258/// function, hoisting expressions out of loops if possible. 259/// 260bool LICM::runOnFunction(Function &) { 261 Changed = false; 262 263 // Get our Loop and Alias Analysis information... 264 LI = &getAnalysis<LoopInfo>(); 265 AA = &getAnalysis<AliasAnalysis>(); 266 267 // Hoist expressions out of all of the top-level loops. 268 const std::vector<Loop*> &TopLevelLoops = LI->getTopLevelLoops(); 269 for (std::vector<Loop*>::const_iterator I = TopLevelLoops.begin(), 270 E = TopLevelLoops.end(); I != E; ++I) { 271 LoopBodyInfo LBI; 272 LICM::visitLoop(*I, LBI); 273 } 274 return Changed; 275} 276 277 278/// visitLoop - Hoist expressions out of the specified loop... 279/// 280void LICM::visitLoop(Loop *L, LoopBodyInfo &LBI) { 281 // Recurse through all subloops before we process this loop... 282 for (std::vector<Loop*>::const_iterator I = L->getSubLoops().begin(), 283 E = L->getSubLoops().end(); I != E; ++I) { 284 LoopBodyInfo SubLBI; 285 LICM::visitLoop(*I, SubLBI); 286 287 // Incorporate information about the subloops into this loop... 288 LBI.incorporate(SubLBI); 289 } 290 CurLoop = L; 291 CurLBI = &LBI; 292 293 // Get the preheader block to move instructions into... 294 Preheader = L->getLoopPreheader(); 295 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 296 297 // Loop over the body of this loop, looking for calls, invokes, and stores. 298 // Because subloops have already been incorporated into LBI, we skip blocks in 299 // subloops. 300 // 301 const std::vector<BasicBlock*> &LoopBBs = L->getBlocks(); 302 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 303 E = LoopBBs.end(); I != E; ++I) 304 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops... 305 LBI.incorporate(**I); // Incorporate the specified basic block 306 307 // We want to visit all of the instructions in this loop... that are not parts 308 // of our subloops (they have already had their invariants hoisted out of 309 // their loop, into this loop, so there is no need to process the BODIES of 310 // the subloops). 311 // 312 // Traverse the body of the loop in depth first order on the dominator tree so 313 // that we are guaranteed to see definitions before we see uses. This allows 314 // us to perform the LICM transformation in one pass, without iteration. 315 // 316 HoistRegion(getAnalysis<DominatorTree>()[L->getHeader()]); 317 318 // Now that all loop invariants have been removed from the loop, promote any 319 // memory references to scalars that we can... 320 if (!DisablePromotion) 321 PromoteValuesInLoop(); 322 323 // Clear out loops state information for the next iteration 324 CurLoop = 0; 325 Preheader = 0; 326} 327 328/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 329/// dominated by the specified block, and that are in the current loop) in depth 330/// first order w.r.t the DominatorTree. This allows us to visit defintions 331/// before uses, allowing us to hoist a loop body in one pass without iteration. 332/// 333void LICM::HoistRegion(DominatorTree::Node *N) { 334 assert(N != 0 && "Null dominator tree node?"); 335 336 // If this subregion is not in the top level loop at all, exit. 337 if (!CurLoop->contains(N->getNode())) return; 338 339 // Only need to hoist the contents of this block if it is not part of a 340 // subloop (which would already have been hoisted) 341 if (!inSubLoop(N->getNode())) 342 visit(*N->getNode()); 343 344 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 345 for (unsigned i = 0, e = Children.size(); i != e; ++i) 346 HoistRegion(Children[i]); 347} 348 349 350/// hoist - When an instruction is found to only use loop invariant operands 351/// that is safe to hoist, this instruction is called to do the dirty work. 352/// 353void LICM::hoist(Instruction &Inst) { 354 DEBUG(std::cerr << "LICM hoisting to"; 355 WriteAsOperand(std::cerr, Preheader, false); 356 std::cerr << ": " << Inst); 357 358 // Remove the instruction from its current basic block... but don't delete the 359 // instruction. 360 Inst.getParent()->getInstList().remove(&Inst); 361 362 // Insert the new node in Preheader, before the terminator. 363 Preheader->getInstList().insert(Preheader->getTerminator(), &Inst); 364 365 ++NumHoisted; 366 Changed = true; 367} 368 369 370void LICM::visitLoadInst(LoadInst &LI) { 371 if (isLoopInvariant(LI.getOperand(0)) && 372 !pointerInvalidatedByLoop(LI.getOperand(0))) { 373 hoist(LI); 374 ++NumHoistedLoads; 375 } 376} 377 378/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 379/// stores out of the loop and moving loads to before the loop. We do this by 380/// looping over the stores in the loop, looking for stores to Must pointers 381/// which are loop invariant. We promote these memory locations to use allocas 382/// instead. These allocas can easily be raised to register values by the 383/// PromoteMem2Reg functionality. 384/// 385void LICM::PromoteValuesInLoop() { 386 // PromotedValues - List of values that are promoted out of the loop. Each 387 // value has an alloca instruction for it, and a cannonical version of the 388 // pointer. 389 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 390 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 391 392 findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 393 if (ValueToAllocaMap.empty()) return; // If there are values to promote... 394 395 Changed = true; 396 NumPromoted += PromotedValues.size(); 397 398 // Emit a copy from the value into the alloca'd value in the loop preheader 399 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 400 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 401 // Load from the memory we are promoting... 402 LoadInst *LI = new LoadInst(PromotedValues[i].second, 403 PromotedValues[i].second->getName()+".promoted", 404 LoopPredInst); 405 // Store into the temporary alloca... 406 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 407 } 408 409 // Scan the basic blocks in the loop, replacing uses of our pointers with 410 // uses of the allocas in question. If we find a branch that exits the 411 // loop, make sure to put reload code into all of the successors of the 412 // loop. 413 // 414 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 415 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 416 E = LoopBBs.end(); I != E; ++I) { 417 // Rewrite all loads and stores in the block of the pointer... 418 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 419 II != E; ++II) { 420 if (LoadInst *L = dyn_cast<LoadInst>(&*II)) { 421 std::map<Value*, AllocaInst*>::iterator 422 I = ValueToAllocaMap.find(L->getOperand(0)); 423 if (I != ValueToAllocaMap.end()) 424 L->setOperand(0, I->second); // Rewrite load instruction... 425 } else if (StoreInst *S = dyn_cast<StoreInst>(&*II)) { 426 std::map<Value*, AllocaInst*>::iterator 427 I = ValueToAllocaMap.find(S->getOperand(1)); 428 if (I != ValueToAllocaMap.end()) 429 S->setOperand(1, I->second); // Rewrite store instruction... 430 } 431 } 432 433 // Check to see if any successors of this block are outside of the loop. 434 // If so, we need to copy the value from the alloca back into the memory 435 // location... 436 // 437 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI) 438 if (!CurLoop->contains(*SI)) { 439 // Copy all of the allocas into their memory locations... 440 BasicBlock::iterator BI = (*SI)->begin(); 441 while (isa<PHINode>(*BI)) 442 ++BI; // Skip over all of the phi nodes in the block... 443 Instruction *InsertPos = BI; 444 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 445 // Load from the alloca... 446 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 447 // Store into the memory we promoted... 448 new StoreInst(LI, PromotedValues[i].second, InsertPos); 449 } 450 } 451 } 452 453 // Now that we have done the deed, use the mem2reg functionality to promote 454 // all of the new allocas we just created into real SSA registers... 455 // 456 std::vector<AllocaInst*> PromotedAllocas; 457 PromotedAllocas.reserve(PromotedValues.size()); 458 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 459 PromotedAllocas.push_back(PromotedValues[i].first); 460 PromoteMemToReg(PromotedAllocas, getAnalysis<DominanceFrontier>()); 461} 462 463/// findPromotableValuesInLoop - Check the current loop for stores to definate 464/// pointers, which are not loaded and stored through may aliases. If these are 465/// found, create an alloca for the value, add it to the PromotedValues list, 466/// and keep track of the mapping from value to alloca... 467/// 468void LICM::findPromotableValuesInLoop( 469 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 470 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 471 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 472 473 for (std::set<Value*>::iterator I = CurLBI->StoredPointers.begin(), 474 E = CurLBI->StoredPointers.end(); I != E; ++I) { 475 Value *V = *I; 476 if (isLoopInvariant(V) && 477 CurLBI->getPointerInfo(V, *AA) == LoopBodyInfo::PointerMustStore) { 478 479 // Don't add a new entry for this stored pointer if it aliases something 480 // we have already processed. 481 std::map<Value*, AllocaInst*>::iterator V2AMI = 482 ValueToAllocaMap.lower_bound(V); 483 if (V2AMI == ValueToAllocaMap.end() || V2AMI->first != V) { 484 // Check to make sure that any loads in the loop are either NO or MUST 485 // aliases. We cannot rewrite loads that _might_ come from this memory 486 // location. 487 488 bool PointerOk = true; 489 for (std::set<Value*>::const_iterator I =CurLBI->LoadedPointers.begin(), 490 E = CurLBI->LoadedPointers.end(); PointerOk && I != E; ++I) 491 switch (AA->alias(V, ~0, *I, ~0)) { 492 case AliasAnalysis::MustAlias: 493 if (V->getType() != (*I)->getType()) 494 PointerOk = false; 495 break; 496 case AliasAnalysis::MayAlias: 497 PointerOk = false; 498 case AliasAnalysis::NoAlias: 499 break; 500 } 501 502 if (PointerOk) { 503 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 504 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 505 PromotedValues.push_back(std::make_pair(AI, V)); 506 ValueToAllocaMap.insert(V2AMI, std::make_pair(V, AI)); 507 508 DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n"); 509 510 // Loop over all of the loads and stores that alias this pointer, 511 // adding them to the Value2AllocaMap as well... 512 for (std::set<Value*>::const_iterator 513 I = CurLBI->LoadedPointers.begin(), 514 E = CurLBI->LoadedPointers.end(); I != E; ++I) 515 if (AA->alias(V, ~0, *I, ~0) == AliasAnalysis::MustAlias) 516 ValueToAllocaMap[*I] = AI; 517 518 for (std::set<Value*>::const_iterator 519 I = CurLBI->StoredPointers.begin(), 520 E = CurLBI->StoredPointers.end(); I != E; ++I) 521 if (AA->alias(V, ~0, *I, ~0) == AliasAnalysis::MustAlias) 522 ValueToAllocaMap[*I] = AI; 523 } 524 } 525 } 526 } 527} 528