LICM.cpp revision bdff548e4dd577a72094d57b282de4e765643b96
1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs loop invariant code motion, attempting to remove as much 11// code from the body of a loop as possible. It does this by either hoisting 12// code into the preheader block, or by sinking code to the exit blocks if it is 13// safe. This pass also promotes must-aliased memory locations in the loop to 14// live in registers, thus hoisting and sinking "invariant" loads and stores. 15// 16// This pass uses alias analysis for two purposes: 17// 18// 1. Moving loop invariant loads and calls out of loops. If we can determine 19// that a load or call inside of a loop never aliases anything stored to, 20// we can hoist it or sink it like any other instruction. 21// 2. Scalar Promotion of Memory - If there is a store instruction inside of 22// the loop, we try to move the store to happen AFTER the loop instead of 23// inside of the loop. This can only happen if a few conditions are true: 24// A. The pointer stored through is loop invariant 25// B. There are no stores or loads in the loop which _may_ alias the 26// pointer. There are no calls in the loop which mod/ref the pointer. 27// If these conditions are true, we can promote the loads and stores in the 28// loop of the pointer to use a temporary alloca'd variable. We then use 29// the mem2reg functionality to construct the appropriate SSA form for the 30// variable. 31// 32//===----------------------------------------------------------------------===// 33 34#define DEBUG_TYPE "licm" 35#include "llvm/Transforms/Scalar.h" 36#include "llvm/Constants.h" 37#include "llvm/DerivedTypes.h" 38#include "llvm/Instructions.h" 39#include "llvm/LLVMContext.h" 40#include "llvm/Target/TargetData.h" 41#include "llvm/Analysis/LoopInfo.h" 42#include "llvm/Analysis/LoopPass.h" 43#include "llvm/Analysis/AliasAnalysis.h" 44#include "llvm/Analysis/AliasSetTracker.h" 45#include "llvm/Analysis/Dominators.h" 46#include "llvm/Analysis/ScalarEvolution.h" 47#include "llvm/Transforms/Utils/PromoteMemToReg.h" 48#include "llvm/Support/CFG.h" 49#include "llvm/Support/Compiler.h" 50#include "llvm/Support/CommandLine.h" 51#include "llvm/Support/raw_ostream.h" 52#include "llvm/Support/Debug.h" 53#include "llvm/ADT/Statistic.h" 54#include <algorithm> 55using namespace llvm; 56 57STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 58STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 59STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 60STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 61STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 62 63static cl::opt<bool> 64DisablePromotion("disable-licm-promotion", cl::Hidden, 65 cl::desc("Disable memory promotion in LICM pass")); 66 67// This feature is currently disabled by default because CodeGen is not yet 68// capable of rematerializing these constants in PIC mode, so it can lead to 69// degraded performance. Compile test/CodeGen/X86/remat-constant.ll with 70// -relocation-model=pic to see an example of this. 71static cl::opt<bool> 72EnableLICMConstantMotion("enable-licm-constant-variables", cl::Hidden, 73 cl::desc("Enable hoisting/sinking of constant " 74 "global variables")); 75 76namespace { 77 struct VISIBILITY_HIDDEN LICM : public LoopPass { 78 static char ID; // Pass identification, replacement for typeid 79 LICM() : LoopPass(&ID) {} 80 81 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 82 83 /// This transformation requires natural loop information & requires that 84 /// loop preheaders be inserted into the CFG... 85 /// 86 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 87 AU.setPreservesCFG(); 88 AU.addRequiredID(LoopSimplifyID); 89 AU.addRequired<LoopInfo>(); 90 AU.addRequired<DominatorTree>(); 91 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 92 AU.addRequired<AliasAnalysis>(); 93 AU.addPreserved<ScalarEvolution>(); 94 AU.addPreserved<DominanceFrontier>(); 95 } 96 97 bool doFinalization() { 98 // Free the values stored in the map 99 for (std::map<Loop *, AliasSetTracker *>::iterator 100 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I) 101 delete I->second; 102 103 LoopToAliasMap.clear(); 104 return false; 105 } 106 107 private: 108 // Various analyses that we use... 109 AliasAnalysis *AA; // Current AliasAnalysis information 110 LoopInfo *LI; // Current LoopInfo 111 DominatorTree *DT; // Dominator Tree for the current Loop... 112 DominanceFrontier *DF; // Current Dominance Frontier 113 114 // State that is updated as we process loops 115 bool Changed; // Set to true when we change anything. 116 BasicBlock *Preheader; // The preheader block of the current loop... 117 Loop *CurLoop; // The current loop we are working on... 118 AliasSetTracker *CurAST; // AliasSet information for the current loop... 119 std::map<Loop *, AliasSetTracker *> LoopToAliasMap; 120 121 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 122 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 123 124 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 125 /// set. 126 void deleteAnalysisValue(Value *V, Loop *L); 127 128 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 129 /// dominated by the specified block, and that are in the current loop) in 130 /// reverse depth first order w.r.t the DominatorTree. This allows us to 131 /// visit uses before definitions, allowing us to sink a loop body in one 132 /// pass without iteration. 133 /// 134 void SinkRegion(DomTreeNode *N); 135 136 /// HoistRegion - Walk the specified region of the CFG (defined by all 137 /// blocks dominated by the specified block, and that are in the current 138 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 139 /// visit definitions before uses, allowing us to hoist a loop body in one 140 /// pass without iteration. 141 /// 142 void HoistRegion(DomTreeNode *N); 143 144 /// inSubLoop - Little predicate that returns true if the specified basic 145 /// block is in a subloop of the current one, not the current one itself. 146 /// 147 bool inSubLoop(BasicBlock *BB) { 148 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 149 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 150 if ((*I)->contains(BB)) 151 return true; // A subloop actually contains this block! 152 return false; 153 } 154 155 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 156 /// specified exit block of the loop is dominated by the specified block 157 /// that is in the body of the loop. We use these constraints to 158 /// dramatically limit the amount of the dominator tree that needs to be 159 /// searched. 160 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 161 BasicBlock *BlockInLoop) const { 162 // If the block in the loop is the loop header, it must be dominated! 163 BasicBlock *LoopHeader = CurLoop->getHeader(); 164 if (BlockInLoop == LoopHeader) 165 return true; 166 167 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 168 DomTreeNode *IDom = DT->getNode(ExitBlock); 169 170 // Because the exit block is not in the loop, we know we have to get _at 171 // least_ its immediate dominator. 172 do { 173 // Get next Immediate Dominator. 174 IDom = IDom->getIDom(); 175 176 // If we have got to the header of the loop, then the instructions block 177 // did not dominate the exit node, so we can't hoist it. 178 if (IDom->getBlock() == LoopHeader) 179 return false; 180 181 } while (IDom != BlockInLoopNode); 182 183 return true; 184 } 185 186 /// sink - When an instruction is found to only be used outside of the loop, 187 /// this function moves it to the exit blocks and patches up SSA form as 188 /// needed. 189 /// 190 void sink(Instruction &I); 191 192 /// hoist - When an instruction is found to only use loop invariant operands 193 /// that is safe to hoist, this instruction is called to do the dirty work. 194 /// 195 void hoist(Instruction &I); 196 197 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 198 /// is not a trapping instruction or if it is a trapping instruction and is 199 /// guaranteed to execute. 200 /// 201 bool isSafeToExecuteUnconditionally(Instruction &I); 202 203 /// pointerInvalidatedByLoop - Return true if the body of this loop may 204 /// store into the memory location pointed to by V. 205 /// 206 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 207 // Check to see if any of the basic blocks in CurLoop invalidate *V. 208 return CurAST->getAliasSetForPointer(V, Size).isMod(); 209 } 210 211 bool canSinkOrHoistInst(Instruction &I); 212 bool isLoopInvariantInst(Instruction &I); 213 bool isNotUsedInLoop(Instruction &I); 214 215 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 216 /// to scalars as we can. 217 /// 218 void PromoteValuesInLoop(); 219 220 /// FindPromotableValuesInLoop - Check the current loop for stores to 221 /// definite pointers, which are not loaded and stored through may aliases. 222 /// If these are found, create an alloca for the value, add it to the 223 /// PromotedValues list, and keep track of the mapping from value to 224 /// alloca... 225 /// 226 void FindPromotableValuesInLoop( 227 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 228 std::map<Value*, AllocaInst*> &Val2AlMap); 229 }; 230} 231 232char LICM::ID = 0; 233static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 234 235Pass *llvm::createLICMPass() { return new LICM(); } 236 237/// Hoist expressions out of the specified loop. Note, alias info for inner 238/// loop is not preserved so it is not a good idea to run LICM multiple 239/// times on one loop. 240/// 241bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 242 Changed = false; 243 244 // Get our Loop and Alias Analysis information... 245 LI = &getAnalysis<LoopInfo>(); 246 AA = &getAnalysis<AliasAnalysis>(); 247 DF = &getAnalysis<DominanceFrontier>(); 248 DT = &getAnalysis<DominatorTree>(); 249 250 CurAST = new AliasSetTracker(*AA); 251 // Collect Alias info from subloops 252 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 253 LoopItr != LoopItrE; ++LoopItr) { 254 Loop *InnerL = *LoopItr; 255 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 256 assert (InnerAST && "Where is my AST?"); 257 258 // What if InnerLoop was modified by other passes ? 259 CurAST->add(*InnerAST); 260 } 261 262 CurLoop = L; 263 264 // Get the preheader block to move instructions into... 265 Preheader = L->getLoopPreheader(); 266 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 267 268 // Loop over the body of this loop, looking for calls, invokes, and stores. 269 // Because subloops have already been incorporated into AST, we skip blocks in 270 // subloops. 271 // 272 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 273 I != E; ++I) { 274 BasicBlock *BB = *I; 275 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops... 276 CurAST->add(*BB); // Incorporate the specified basic block 277 } 278 279 // We want to visit all of the instructions in this loop... that are not parts 280 // of our subloops (they have already had their invariants hoisted out of 281 // their loop, into this loop, so there is no need to process the BODIES of 282 // the subloops). 283 // 284 // Traverse the body of the loop in depth first order on the dominator tree so 285 // that we are guaranteed to see definitions before we see uses. This allows 286 // us to sink instructions in one pass, without iteration. After sinking 287 // instructions, we perform another pass to hoist them out of the loop. 288 // 289 SinkRegion(DT->getNode(L->getHeader())); 290 HoistRegion(DT->getNode(L->getHeader())); 291 292 // Now that all loop invariants have been removed from the loop, promote any 293 // memory references to scalars that we can... 294 if (!DisablePromotion) 295 PromoteValuesInLoop(); 296 297 // Clear out loops state information for the next iteration 298 CurLoop = 0; 299 Preheader = 0; 300 301 LoopToAliasMap[L] = CurAST; 302 return Changed; 303} 304 305/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 306/// dominated by the specified block, and that are in the current loop) in 307/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 308/// uses before definitions, allowing us to sink a loop body in one pass without 309/// iteration. 310/// 311void LICM::SinkRegion(DomTreeNode *N) { 312 assert(N != 0 && "Null dominator tree node?"); 313 BasicBlock *BB = N->getBlock(); 314 315 // If this subregion is not in the top level loop at all, exit. 316 if (!CurLoop->contains(BB)) return; 317 318 // We are processing blocks in reverse dfo, so process children first... 319 const std::vector<DomTreeNode*> &Children = N->getChildren(); 320 for (unsigned i = 0, e = Children.size(); i != e; ++i) 321 SinkRegion(Children[i]); 322 323 // Only need to process the contents of this block if it is not part of a 324 // subloop (which would already have been processed). 325 if (inSubLoop(BB)) return; 326 327 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 328 Instruction &I = *--II; 329 330 // Check to see if we can sink this instruction to the exit blocks 331 // of the loop. We can do this if the all users of the instruction are 332 // outside of the loop. In this case, it doesn't even matter if the 333 // operands of the instruction are loop invariant. 334 // 335 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 336 ++II; 337 sink(I); 338 } 339 } 340} 341 342 343/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 344/// dominated by the specified block, and that are in the current loop) in depth 345/// first order w.r.t the DominatorTree. This allows us to visit definitions 346/// before uses, allowing us to hoist a loop body in one pass without iteration. 347/// 348void LICM::HoistRegion(DomTreeNode *N) { 349 assert(N != 0 && "Null dominator tree node?"); 350 BasicBlock *BB = N->getBlock(); 351 352 // If this subregion is not in the top level loop at all, exit. 353 if (!CurLoop->contains(BB)) return; 354 355 // Only need to process the contents of this block if it is not part of a 356 // subloop (which would already have been processed). 357 if (!inSubLoop(BB)) 358 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 359 Instruction &I = *II++; 360 361 // Try hoisting the instruction out to the preheader. We can only do this 362 // if all of the operands of the instruction are loop invariant and if it 363 // is safe to hoist the instruction. 364 // 365 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 366 isSafeToExecuteUnconditionally(I)) 367 hoist(I); 368 } 369 370 const std::vector<DomTreeNode*> &Children = N->getChildren(); 371 for (unsigned i = 0, e = Children.size(); i != e; ++i) 372 HoistRegion(Children[i]); 373} 374 375/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 376/// instruction. 377/// 378bool LICM::canSinkOrHoistInst(Instruction &I) { 379 // Loads have extra constraints we have to verify before we can hoist them. 380 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 381 if (LI->isVolatile()) 382 return false; // Don't hoist volatile loads! 383 384 // Loads from constant memory are always safe to move, even if they end up 385 // in the same alias set as something that ends up being modified. 386 if (EnableLICMConstantMotion && 387 AA->pointsToConstantMemory(LI->getOperand(0))) 388 return true; 389 390 // Don't hoist loads which have may-aliased stores in loop. 391 unsigned Size = 0; 392 if (LI->getType()->isSized()) 393 Size = AA->getTypeStoreSize(LI->getType()); 394 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 395 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 396 // Handle obvious cases efficiently. 397 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 398 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 399 return true; 400 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 401 // If this call only reads from memory and there are no writes to memory 402 // in the loop, we can hoist or sink the call as appropriate. 403 bool FoundMod = false; 404 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 405 I != E; ++I) { 406 AliasSet &AS = *I; 407 if (!AS.isForwardingAliasSet() && AS.isMod()) { 408 FoundMod = true; 409 break; 410 } 411 } 412 if (!FoundMod) return true; 413 } 414 415 // FIXME: This should use mod/ref information to see if we can hoist or sink 416 // the call. 417 418 return false; 419 } 420 421 // Otherwise these instructions are hoistable/sinkable 422 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 423 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 424 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 425 isa<ShuffleVectorInst>(I); 426} 427 428/// isNotUsedInLoop - Return true if the only users of this instruction are 429/// outside of the loop. If this is true, we can sink the instruction to the 430/// exit blocks of the loop. 431/// 432bool LICM::isNotUsedInLoop(Instruction &I) { 433 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 434 Instruction *User = cast<Instruction>(*UI); 435 if (PHINode *PN = dyn_cast<PHINode>(User)) { 436 // PHI node uses occur in predecessor blocks! 437 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 438 if (PN->getIncomingValue(i) == &I) 439 if (CurLoop->contains(PN->getIncomingBlock(i))) 440 return false; 441 } else if (CurLoop->contains(User->getParent())) { 442 return false; 443 } 444 } 445 return true; 446} 447 448 449/// isLoopInvariantInst - Return true if all operands of this instruction are 450/// loop invariant. We also filter out non-hoistable instructions here just for 451/// efficiency. 452/// 453bool LICM::isLoopInvariantInst(Instruction &I) { 454 // The instruction is loop invariant if all of its operands are loop-invariant 455 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 456 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 457 return false; 458 459 // If we got this far, the instruction is loop invariant! 460 return true; 461} 462 463/// sink - When an instruction is found to only be used outside of the loop, 464/// this function moves it to the exit blocks and patches up SSA form as needed. 465/// This method is guaranteed to remove the original instruction from its 466/// position, and may either delete it or move it to outside of the loop. 467/// 468void LICM::sink(Instruction &I) { 469 DEBUG(errs() << "LICM sinking instruction: " << I); 470 471 SmallVector<BasicBlock*, 8> ExitBlocks; 472 CurLoop->getExitBlocks(ExitBlocks); 473 474 if (isa<LoadInst>(I)) ++NumMovedLoads; 475 else if (isa<CallInst>(I)) ++NumMovedCalls; 476 ++NumSunk; 477 Changed = true; 478 479 LLVMContext &Context = I.getContext(); 480 481 // The case where there is only a single exit node of this loop is common 482 // enough that we handle it as a special (more efficient) case. It is more 483 // efficient to handle because there are no PHI nodes that need to be placed. 484 if (ExitBlocks.size() == 1) { 485 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 486 // Instruction is not used, just delete it. 487 CurAST->deleteValue(&I); 488 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 489 I.replaceAllUsesWith(UndefValue::get(I.getType())); 490 I.eraseFromParent(); 491 } else { 492 // Move the instruction to the start of the exit block, after any PHI 493 // nodes in it. 494 I.removeFromParent(); 495 496 BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI(); 497 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 498 } 499 } else if (ExitBlocks.empty()) { 500 // The instruction is actually dead if there ARE NO exit blocks. 501 CurAST->deleteValue(&I); 502 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 503 I.replaceAllUsesWith(UndefValue::get(I.getType())); 504 I.eraseFromParent(); 505 } else { 506 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 507 // do all of the hard work of inserting PHI nodes as necessary. We convert 508 // the value into a stack object to get it to do this. 509 510 // Firstly, we create a stack object to hold the value... 511 AllocaInst *AI = 0; 512 513 if (I.getType() != Type::getVoidTy(I.getContext())) { 514 AI = new AllocaInst(I.getType(), 0, I.getName(), 515 I.getParent()->getParent()->getEntryBlock().begin()); 516 CurAST->add(AI); 517 } 518 519 // Secondly, insert load instructions for each use of the instruction 520 // outside of the loop. 521 while (!I.use_empty()) { 522 Instruction *U = cast<Instruction>(I.use_back()); 523 524 // If the user is a PHI Node, we actually have to insert load instructions 525 // in all predecessor blocks, not in the PHI block itself! 526 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 527 // Only insert into each predecessor once, so that we don't have 528 // different incoming values from the same block! 529 std::map<BasicBlock*, Value*> InsertedBlocks; 530 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 531 if (UPN->getIncomingValue(i) == &I) { 532 BasicBlock *Pred = UPN->getIncomingBlock(i); 533 Value *&PredVal = InsertedBlocks[Pred]; 534 if (!PredVal) { 535 // Insert a new load instruction right before the terminator in 536 // the predecessor block. 537 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 538 CurAST->add(cast<LoadInst>(PredVal)); 539 } 540 541 UPN->setIncomingValue(i, PredVal); 542 } 543 544 } else { 545 LoadInst *L = new LoadInst(AI, "", U); 546 U->replaceUsesOfWith(&I, L); 547 CurAST->add(L); 548 } 549 } 550 551 // Thirdly, insert a copy of the instruction in each exit block of the loop 552 // that is dominated by the instruction, storing the result into the memory 553 // location. Be careful not to insert the instruction into any particular 554 // basic block more than once. 555 std::set<BasicBlock*> InsertedBlocks; 556 BasicBlock *InstOrigBB = I.getParent(); 557 558 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 559 BasicBlock *ExitBlock = ExitBlocks[i]; 560 561 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 562 // If we haven't already processed this exit block, do so now. 563 if (InsertedBlocks.insert(ExitBlock).second) { 564 // Insert the code after the last PHI node... 565 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 566 567 // If this is the first exit block processed, just move the original 568 // instruction, otherwise clone the original instruction and insert 569 // the copy. 570 Instruction *New; 571 if (InsertedBlocks.size() == 1) { 572 I.removeFromParent(); 573 ExitBlock->getInstList().insert(InsertPt, &I); 574 New = &I; 575 } else { 576 New = I.clone(Context); 577 CurAST->copyValue(&I, New); 578 if (!I.getName().empty()) 579 New->setName(I.getName()+".le"); 580 ExitBlock->getInstList().insert(InsertPt, New); 581 } 582 583 // Now that we have inserted the instruction, store it into the alloca 584 if (AI) new StoreInst(New, AI, InsertPt); 585 } 586 } 587 } 588 589 // If the instruction doesn't dominate any exit blocks, it must be dead. 590 if (InsertedBlocks.empty()) { 591 CurAST->deleteValue(&I); 592 I.eraseFromParent(); 593 } 594 595 // Finally, promote the fine value to SSA form. 596 if (AI) { 597 std::vector<AllocaInst*> Allocas; 598 Allocas.push_back(AI); 599 PromoteMemToReg(Allocas, *DT, *DF, Context, CurAST); 600 } 601 } 602} 603 604/// hoist - When an instruction is found to only use loop invariant operands 605/// that is safe to hoist, this instruction is called to do the dirty work. 606/// 607void LICM::hoist(Instruction &I) { 608 DEBUG(errs() << "LICM hoisting to " << Preheader->getName() << ": " << I); 609 610 // Remove the instruction from its current basic block... but don't delete the 611 // instruction. 612 I.removeFromParent(); 613 614 // Insert the new node in Preheader, before the terminator. 615 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 616 617 if (isa<LoadInst>(I)) ++NumMovedLoads; 618 else if (isa<CallInst>(I)) ++NumMovedCalls; 619 ++NumHoisted; 620 Changed = true; 621} 622 623/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 624/// not a trapping instruction or if it is a trapping instruction and is 625/// guaranteed to execute. 626/// 627bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 628 // If it is not a trapping instruction, it is always safe to hoist. 629 if (Inst.isSafeToSpeculativelyExecute()) 630 return true; 631 632 // Otherwise we have to check to make sure that the instruction dominates all 633 // of the exit blocks. If it doesn't, then there is a path out of the loop 634 // which does not execute this instruction, so we can't hoist it. 635 636 // If the instruction is in the header block for the loop (which is very 637 // common), it is always guaranteed to dominate the exit blocks. Since this 638 // is a common case, and can save some work, check it now. 639 if (Inst.getParent() == CurLoop->getHeader()) 640 return true; 641 642 // Get the exit blocks for the current loop. 643 SmallVector<BasicBlock*, 8> ExitBlocks; 644 CurLoop->getExitBlocks(ExitBlocks); 645 646 // For each exit block, get the DT node and walk up the DT until the 647 // instruction's basic block is found or we exit the loop. 648 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 649 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 650 return false; 651 652 return true; 653} 654 655 656/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 657/// stores out of the loop and moving loads to before the loop. We do this by 658/// looping over the stores in the loop, looking for stores to Must pointers 659/// which are loop invariant. We promote these memory locations to use allocas 660/// instead. These allocas can easily be raised to register values by the 661/// PromoteMem2Reg functionality. 662/// 663void LICM::PromoteValuesInLoop() { 664 // PromotedValues - List of values that are promoted out of the loop. Each 665 // value has an alloca instruction for it, and a canonical version of the 666 // pointer. 667 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 668 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 669 670 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 671 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 672 673 Changed = true; 674 NumPromoted += PromotedValues.size(); 675 676 std::vector<Value*> PointerValueNumbers; 677 678 // Emit a copy from the value into the alloca'd value in the loop preheader 679 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 680 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 681 Value *Ptr = PromotedValues[i].second; 682 683 // If we are promoting a pointer value, update alias information for the 684 // inserted load. 685 Value *LoadValue = 0; 686 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 687 // Locate a load or store through the pointer, and assign the same value 688 // to LI as we are loading or storing. Since we know that the value is 689 // stored in this loop, this will always succeed. 690 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 691 UI != E; ++UI) 692 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 693 LoadValue = LI; 694 break; 695 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 696 if (SI->getOperand(1) == Ptr) { 697 LoadValue = SI->getOperand(0); 698 break; 699 } 700 } 701 assert(LoadValue && "No store through the pointer found!"); 702 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 703 } 704 705 // Load from the memory we are promoting. 706 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 707 708 if (LoadValue) CurAST->copyValue(LoadValue, LI); 709 710 // Store into the temporary alloca. 711 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 712 } 713 714 // Scan the basic blocks in the loop, replacing uses of our pointers with 715 // uses of the allocas in question. 716 // 717 for (Loop::block_iterator I = CurLoop->block_begin(), 718 E = CurLoop->block_end(); I != E; ++I) { 719 BasicBlock *BB = *I; 720 // Rewrite all loads and stores in the block of the pointer... 721 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 722 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 723 std::map<Value*, AllocaInst*>::iterator 724 I = ValueToAllocaMap.find(L->getOperand(0)); 725 if (I != ValueToAllocaMap.end()) 726 L->setOperand(0, I->second); // Rewrite load instruction... 727 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 728 std::map<Value*, AllocaInst*>::iterator 729 I = ValueToAllocaMap.find(S->getOperand(1)); 730 if (I != ValueToAllocaMap.end()) 731 S->setOperand(1, I->second); // Rewrite store instruction... 732 } 733 } 734 } 735 736 // Now that the body of the loop uses the allocas instead of the original 737 // memory locations, insert code to copy the alloca value back into the 738 // original memory location on all exits from the loop. Note that we only 739 // want to insert one copy of the code in each exit block, though the loop may 740 // exit to the same block more than once. 741 // 742 SmallPtrSet<BasicBlock*, 16> ProcessedBlocks; 743 744 SmallVector<BasicBlock*, 8> ExitBlocks; 745 CurLoop->getExitBlocks(ExitBlocks); 746 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 747 if (!ProcessedBlocks.insert(ExitBlocks[i])) 748 continue; 749 750 // Copy all of the allocas into their memory locations. 751 BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI(); 752 Instruction *InsertPos = BI; 753 unsigned PVN = 0; 754 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 755 // Load from the alloca. 756 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 757 758 // If this is a pointer type, update alias info appropriately. 759 if (isa<PointerType>(LI->getType())) 760 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 761 762 // Store into the memory we promoted. 763 new StoreInst(LI, PromotedValues[i].second, InsertPos); 764 } 765 } 766 767 // Now that we have done the deed, use the mem2reg functionality to promote 768 // all of the new allocas we just created into real SSA registers. 769 // 770 std::vector<AllocaInst*> PromotedAllocas; 771 PromotedAllocas.reserve(PromotedValues.size()); 772 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 773 PromotedAllocas.push_back(PromotedValues[i].first); 774 PromoteMemToReg(PromotedAllocas, *DT, *DF, Preheader->getContext(), CurAST); 775} 776 777/// FindPromotableValuesInLoop - Check the current loop for stores to definite 778/// pointers, which are not loaded and stored through may aliases and are safe 779/// for promotion. If these are found, create an alloca for the value, add it 780/// to the PromotedValues list, and keep track of the mapping from value to 781/// alloca. 782void LICM::FindPromotableValuesInLoop( 783 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 784 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 785 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 786 787 // Loop over all of the alias sets in the tracker object. 788 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 789 I != E; ++I) { 790 AliasSet &AS = *I; 791 // We can promote this alias set if it has a store, if it is a "Must" alias 792 // set, if the pointer is loop invariant, and if we are not eliminating any 793 // volatile loads or stores. 794 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 795 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 796 continue; 797 798 assert(!AS.empty() && 799 "Must alias set should have at least one pointer element in it!"); 800 Value *V = AS.begin()->getValue(); 801 802 // Check that all of the pointers in the alias set have the same type. We 803 // cannot (yet) promote a memory location that is loaded and stored in 804 // different sizes. 805 { 806 bool PointerOk = true; 807 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 808 if (V->getType() != I->getValue()->getType()) { 809 PointerOk = false; 810 break; 811 } 812 if (!PointerOk) 813 continue; 814 } 815 816 // It isn't safe to promote a load/store from the loop if the load/store is 817 // conditional. For example, turning: 818 // 819 // for () { if (c) *P += 1; } 820 // 821 // into: 822 // 823 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 824 // 825 // is not safe, because *P may only be valid to access if 'c' is true. 826 // 827 // It is safe to promote P if all uses are direct load/stores and if at 828 // least one is guaranteed to be executed. 829 bool GuaranteedToExecute = false; 830 bool InvalidInst = false; 831 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 832 UI != UE; ++UI) { 833 // Ignore instructions not in this loop. 834 Instruction *Use = dyn_cast<Instruction>(*UI); 835 if (!Use || !CurLoop->contains(Use->getParent())) 836 continue; 837 838 if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) { 839 InvalidInst = true; 840 break; 841 } 842 843 if (!GuaranteedToExecute) 844 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 845 } 846 847 // If there is an non-load/store instruction in the loop, we can't promote 848 // it. If there isn't a guaranteed-to-execute instruction, we can't 849 // promote. 850 if (InvalidInst || !GuaranteedToExecute) 851 continue; 852 853 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 854 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 855 PromotedValues.push_back(std::make_pair(AI, V)); 856 857 // Update the AST and alias analysis. 858 CurAST->copyValue(V, AI); 859 860 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 861 ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI)); 862 863 DEBUG(errs() << "LICM: Promoting value: " << *V << "\n"); 864 } 865} 866 867/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 868void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 869 AliasSetTracker *AST = LoopToAliasMap[L]; 870 if (!AST) 871 return; 872 873 AST->copyValue(From, To); 874} 875 876/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 877/// set. 878void LICM::deleteAnalysisValue(Value *V, Loop *L) { 879 AliasSetTracker *AST = LoopToAliasMap[L]; 880 if (!AST) 881 return; 882 883 AST->deleteValue(V); 884} 885