LICM.cpp revision 844731a7f1909f55935e3514c9e713a62d67662e
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/Target/TargetData.h" 40#include "llvm/Analysis/LoopInfo.h" 41#include "llvm/Analysis/LoopPass.h" 42#include "llvm/Analysis/AliasAnalysis.h" 43#include "llvm/Analysis/AliasSetTracker.h" 44#include "llvm/Analysis/Dominators.h" 45#include "llvm/Analysis/ScalarEvolution.h" 46#include "llvm/Transforms/Utils/PromoteMemToReg.h" 47#include "llvm/Support/CFG.h" 48#include "llvm/Support/Compiler.h" 49#include "llvm/Support/CommandLine.h" 50#include "llvm/Support/Debug.h" 51#include "llvm/ADT/Statistic.h" 52#include <algorithm> 53using namespace llvm; 54 55STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 56STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 57STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 58STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 59STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 60 61static cl::opt<bool> 62DisablePromotion("disable-licm-promotion", cl::Hidden, 63 cl::desc("Disable memory promotion in LICM pass")); 64 65namespace { 66 struct VISIBILITY_HIDDEN LICM : public LoopPass { 67 static char ID; // Pass identification, replacement for typeid 68 LICM() : LoopPass((intptr_t)&ID) {} 69 70 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 71 72 /// This transformation requires natural loop information & requires that 73 /// loop preheaders be inserted into the CFG... 74 /// 75 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 76 AU.setPreservesCFG(); 77 AU.addRequiredID(LoopSimplifyID); 78 AU.addRequired<LoopInfo>(); 79 AU.addRequired<DominatorTree>(); 80 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 81 AU.addRequired<AliasAnalysis>(); 82 AU.addPreserved<ScalarEvolution>(); 83 AU.addPreserved<DominanceFrontier>(); 84 } 85 86 bool doFinalization() { 87 // Free the values stored in the map 88 for (std::map<Loop *, AliasSetTracker *>::iterator 89 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I) 90 delete I->second; 91 92 LoopToAliasMap.clear(); 93 return false; 94 } 95 96 private: 97 // Various analyses that we use... 98 AliasAnalysis *AA; // Current AliasAnalysis information 99 LoopInfo *LI; // Current LoopInfo 100 DominatorTree *DT; // Dominator Tree for the current Loop... 101 DominanceFrontier *DF; // Current Dominance Frontier 102 103 // State that is updated as we process loops 104 bool Changed; // Set to true when we change anything. 105 BasicBlock *Preheader; // The preheader block of the current loop... 106 Loop *CurLoop; // The current loop we are working on... 107 AliasSetTracker *CurAST; // AliasSet information for the current loop... 108 std::map<Loop *, AliasSetTracker *> LoopToAliasMap; 109 110 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 111 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 112 113 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 114 /// set. 115 void deleteAnalysisValue(Value *V, Loop *L); 116 117 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 118 /// dominated by the specified block, and that are in the current loop) in 119 /// reverse depth first order w.r.t the DominatorTree. This allows us to 120 /// visit uses before definitions, allowing us to sink a loop body in one 121 /// pass without iteration. 122 /// 123 void SinkRegion(DomTreeNode *N); 124 125 /// HoistRegion - Walk the specified region of the CFG (defined by all 126 /// blocks dominated by the specified block, and that are in the current 127 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 128 /// visit definitions before uses, allowing us to hoist a loop body in one 129 /// pass without iteration. 130 /// 131 void HoistRegion(DomTreeNode *N); 132 133 /// inSubLoop - Little predicate that returns true if the specified basic 134 /// block is in a subloop of the current one, not the current one itself. 135 /// 136 bool inSubLoop(BasicBlock *BB) { 137 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 138 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 139 if ((*I)->contains(BB)) 140 return true; // A subloop actually contains this block! 141 return false; 142 } 143 144 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 145 /// specified exit block of the loop is dominated by the specified block 146 /// that is in the body of the loop. We use these constraints to 147 /// dramatically limit the amount of the dominator tree that needs to be 148 /// searched. 149 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 150 BasicBlock *BlockInLoop) const { 151 // If the block in the loop is the loop header, it must be dominated! 152 BasicBlock *LoopHeader = CurLoop->getHeader(); 153 if (BlockInLoop == LoopHeader) 154 return true; 155 156 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 157 DomTreeNode *IDom = DT->getNode(ExitBlock); 158 159 // Because the exit block is not in the loop, we know we have to get _at 160 // least_ its immediate dominator. 161 do { 162 // Get next Immediate Dominator. 163 IDom = IDom->getIDom(); 164 165 // If we have got to the header of the loop, then the instructions block 166 // did not dominate the exit node, so we can't hoist it. 167 if (IDom->getBlock() == LoopHeader) 168 return false; 169 170 } while (IDom != BlockInLoopNode); 171 172 return true; 173 } 174 175 /// sink - When an instruction is found to only be used outside of the loop, 176 /// this function moves it to the exit blocks and patches up SSA form as 177 /// needed. 178 /// 179 void sink(Instruction &I); 180 181 /// hoist - When an instruction is found to only use loop invariant operands 182 /// that is safe to hoist, this instruction is called to do the dirty work. 183 /// 184 void hoist(Instruction &I); 185 186 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 187 /// is not a trapping instruction or if it is a trapping instruction and is 188 /// guaranteed to execute. 189 /// 190 bool isSafeToExecuteUnconditionally(Instruction &I); 191 192 /// pointerInvalidatedByLoop - Return true if the body of this loop may 193 /// store into the memory location pointed to by V. 194 /// 195 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 196 // Check to see if any of the basic blocks in CurLoop invalidate *V. 197 return CurAST->getAliasSetForPointer(V, Size).isMod(); 198 } 199 200 bool canSinkOrHoistInst(Instruction &I); 201 bool isLoopInvariantInst(Instruction &I); 202 bool isNotUsedInLoop(Instruction &I); 203 204 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 205 /// to scalars as we can. 206 /// 207 void PromoteValuesInLoop(); 208 209 /// FindPromotableValuesInLoop - Check the current loop for stores to 210 /// definite pointers, which are not loaded and stored through may aliases. 211 /// If these are found, create an alloca for the value, add it to the 212 /// PromotedValues list, and keep track of the mapping from value to 213 /// alloca... 214 /// 215 void FindPromotableValuesInLoop( 216 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 217 std::map<Value*, AllocaInst*> &Val2AlMap); 218 }; 219} 220 221char LICM::ID = 0; 222static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 223 224LoopPass *llvm::createLICMPass() { return new LICM(); } 225 226/// Hoist expressions out of the specified loop. Note, alias info for inner 227/// loop is not preserved so it is not a good idea to run LICM multiple 228/// times on one loop. 229/// 230bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 231 Changed = false; 232 233 // Get our Loop and Alias Analysis information... 234 LI = &getAnalysis<LoopInfo>(); 235 AA = &getAnalysis<AliasAnalysis>(); 236 DF = &getAnalysis<DominanceFrontier>(); 237 DT = &getAnalysis<DominatorTree>(); 238 239 CurAST = new AliasSetTracker(*AA); 240 // Collect Alias info from subloops 241 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 242 LoopItr != LoopItrE; ++LoopItr) { 243 Loop *InnerL = *LoopItr; 244 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 245 assert (InnerAST && "Where is my AST?"); 246 247 // What if InnerLoop was modified by other passes ? 248 CurAST->add(*InnerAST); 249 } 250 251 CurLoop = L; 252 253 // Get the preheader block to move instructions into... 254 Preheader = L->getLoopPreheader(); 255 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 256 257 // Loop over the body of this loop, looking for calls, invokes, and stores. 258 // Because subloops have already been incorporated into AST, we skip blocks in 259 // subloops. 260 // 261 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(), 262 E = L->getBlocks().end(); I != E; ++I) 263 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops... 264 CurAST->add(**I); // Incorporate the specified basic block 265 266 // We want to visit all of the instructions in this loop... that are not parts 267 // of our subloops (they have already had their invariants hoisted out of 268 // their loop, into this loop, so there is no need to process the BODIES of 269 // the subloops). 270 // 271 // Traverse the body of the loop in depth first order on the dominator tree so 272 // that we are guaranteed to see definitions before we see uses. This allows 273 // us to sink instructions in one pass, without iteration. After sinking 274 // instructions, we perform another pass to hoist them out of the loop. 275 // 276 SinkRegion(DT->getNode(L->getHeader())); 277 HoistRegion(DT->getNode(L->getHeader())); 278 279 // Now that all loop invariants have been removed from the loop, promote any 280 // memory references to scalars that we can... 281 if (!DisablePromotion) 282 PromoteValuesInLoop(); 283 284 // Clear out loops state information for the next iteration 285 CurLoop = 0; 286 Preheader = 0; 287 288 LoopToAliasMap[L] = CurAST; 289 return Changed; 290} 291 292/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 293/// dominated by the specified block, and that are in the current loop) in 294/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 295/// uses before definitions, allowing us to sink a loop body in one pass without 296/// iteration. 297/// 298void LICM::SinkRegion(DomTreeNode *N) { 299 assert(N != 0 && "Null dominator tree node?"); 300 BasicBlock *BB = N->getBlock(); 301 302 // If this subregion is not in the top level loop at all, exit. 303 if (!CurLoop->contains(BB)) return; 304 305 // We are processing blocks in reverse dfo, so process children first... 306 const std::vector<DomTreeNode*> &Children = N->getChildren(); 307 for (unsigned i = 0, e = Children.size(); i != e; ++i) 308 SinkRegion(Children[i]); 309 310 // Only need to process the contents of this block if it is not part of a 311 // subloop (which would already have been processed). 312 if (inSubLoop(BB)) return; 313 314 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 315 Instruction &I = *--II; 316 317 // Check to see if we can sink this instruction to the exit blocks 318 // of the loop. We can do this if the all users of the instruction are 319 // outside of the loop. In this case, it doesn't even matter if the 320 // operands of the instruction are loop invariant. 321 // 322 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 323 ++II; 324 sink(I); 325 } 326 } 327} 328 329 330/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 331/// dominated by the specified block, and that are in the current loop) in depth 332/// first order w.r.t the DominatorTree. This allows us to visit definitions 333/// before uses, allowing us to hoist a loop body in one pass without iteration. 334/// 335void LICM::HoistRegion(DomTreeNode *N) { 336 assert(N != 0 && "Null dominator tree node?"); 337 BasicBlock *BB = N->getBlock(); 338 339 // If this subregion is not in the top level loop at all, exit. 340 if (!CurLoop->contains(BB)) return; 341 342 // Only need to process the contents of this block if it is not part of a 343 // subloop (which would already have been processed). 344 if (!inSubLoop(BB)) 345 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 346 Instruction &I = *II++; 347 348 // Try hoisting the instruction out to the preheader. We can only do this 349 // if all of the operands of the instruction are loop invariant and if it 350 // is safe to hoist the instruction. 351 // 352 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 353 isSafeToExecuteUnconditionally(I)) 354 hoist(I); 355 } 356 357 const std::vector<DomTreeNode*> &Children = N->getChildren(); 358 for (unsigned i = 0, e = Children.size(); i != e; ++i) 359 HoistRegion(Children[i]); 360} 361 362/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 363/// instruction. 364/// 365bool LICM::canSinkOrHoistInst(Instruction &I) { 366 // Loads have extra constraints we have to verify before we can hoist them. 367 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 368 if (LI->isVolatile()) 369 return false; // Don't hoist volatile loads! 370 371 // Don't hoist loads which have may-aliased stores in loop. 372 unsigned Size = 0; 373 if (LI->getType()->isSized()) 374 Size = AA->getTargetData().getTypeStoreSize(LI->getType()); 375 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 376 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 377 // Handle obvious cases efficiently. 378 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 379 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 380 return true; 381 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 382 // If this call only reads from memory and there are no writes to memory 383 // in the loop, we can hoist or sink the call as appropriate. 384 bool FoundMod = false; 385 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 386 I != E; ++I) { 387 AliasSet &AS = *I; 388 if (!AS.isForwardingAliasSet() && AS.isMod()) { 389 FoundMod = true; 390 break; 391 } 392 } 393 if (!FoundMod) return true; 394 } 395 396 // FIXME: This should use mod/ref information to see if we can hoist or sink 397 // the call. 398 399 return false; 400 } 401 402 // Otherwise these instructions are hoistable/sinkable 403 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 404 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 405 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 406 isa<ShuffleVectorInst>(I); 407} 408 409/// isNotUsedInLoop - Return true if the only users of this instruction are 410/// outside of the loop. If this is true, we can sink the instruction to the 411/// exit blocks of the loop. 412/// 413bool LICM::isNotUsedInLoop(Instruction &I) { 414 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 415 Instruction *User = cast<Instruction>(*UI); 416 if (PHINode *PN = dyn_cast<PHINode>(User)) { 417 // PHI node uses occur in predecessor blocks! 418 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 419 if (PN->getIncomingValue(i) == &I) 420 if (CurLoop->contains(PN->getIncomingBlock(i))) 421 return false; 422 } else if (CurLoop->contains(User->getParent())) { 423 return false; 424 } 425 } 426 return true; 427} 428 429 430/// isLoopInvariantInst - Return true if all operands of this instruction are 431/// loop invariant. We also filter out non-hoistable instructions here just for 432/// efficiency. 433/// 434bool LICM::isLoopInvariantInst(Instruction &I) { 435 // The instruction is loop invariant if all of its operands are loop-invariant 436 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 437 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 438 return false; 439 440 // If we got this far, the instruction is loop invariant! 441 return true; 442} 443 444/// sink - When an instruction is found to only be used outside of the loop, 445/// this function moves it to the exit blocks and patches up SSA form as needed. 446/// This method is guaranteed to remove the original instruction from its 447/// position, and may either delete it or move it to outside of the loop. 448/// 449void LICM::sink(Instruction &I) { 450 DOUT << "LICM sinking instruction: " << I; 451 452 SmallVector<BasicBlock*, 8> ExitBlocks; 453 CurLoop->getExitBlocks(ExitBlocks); 454 455 if (isa<LoadInst>(I)) ++NumMovedLoads; 456 else if (isa<CallInst>(I)) ++NumMovedCalls; 457 ++NumSunk; 458 Changed = true; 459 460 // The case where there is only a single exit node of this loop is common 461 // enough that we handle it as a special (more efficient) case. It is more 462 // efficient to handle because there are no PHI nodes that need to be placed. 463 if (ExitBlocks.size() == 1) { 464 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 465 // Instruction is not used, just delete it. 466 CurAST->deleteValue(&I); 467 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 468 I.replaceAllUsesWith(UndefValue::get(I.getType())); 469 I.eraseFromParent(); 470 } else { 471 // Move the instruction to the start of the exit block, after any PHI 472 // nodes in it. 473 I.removeFromParent(); 474 475 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin(); 476 while (isa<PHINode>(InsertPt)) ++InsertPt; 477 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 478 } 479 } else if (ExitBlocks.empty()) { 480 // The instruction is actually dead if there ARE NO exit blocks. 481 CurAST->deleteValue(&I); 482 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 483 I.replaceAllUsesWith(UndefValue::get(I.getType())); 484 I.eraseFromParent(); 485 } else { 486 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 487 // do all of the hard work of inserting PHI nodes as necessary. We convert 488 // the value into a stack object to get it to do this. 489 490 // Firstly, we create a stack object to hold the value... 491 AllocaInst *AI = 0; 492 493 if (I.getType() != Type::VoidTy) { 494 AI = new AllocaInst(I.getType(), 0, I.getName(), 495 I.getParent()->getParent()->getEntryBlock().begin()); 496 CurAST->add(AI); 497 } 498 499 // Secondly, insert load instructions for each use of the instruction 500 // outside of the loop. 501 while (!I.use_empty()) { 502 Instruction *U = cast<Instruction>(I.use_back()); 503 504 // If the user is a PHI Node, we actually have to insert load instructions 505 // in all predecessor blocks, not in the PHI block itself! 506 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 507 // Only insert into each predecessor once, so that we don't have 508 // different incoming values from the same block! 509 std::map<BasicBlock*, Value*> InsertedBlocks; 510 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 511 if (UPN->getIncomingValue(i) == &I) { 512 BasicBlock *Pred = UPN->getIncomingBlock(i); 513 Value *&PredVal = InsertedBlocks[Pred]; 514 if (!PredVal) { 515 // Insert a new load instruction right before the terminator in 516 // the predecessor block. 517 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 518 CurAST->add(cast<LoadInst>(PredVal)); 519 } 520 521 UPN->setIncomingValue(i, PredVal); 522 } 523 524 } else { 525 LoadInst *L = new LoadInst(AI, "", U); 526 U->replaceUsesOfWith(&I, L); 527 CurAST->add(L); 528 } 529 } 530 531 // Thirdly, insert a copy of the instruction in each exit block of the loop 532 // that is dominated by the instruction, storing the result into the memory 533 // location. Be careful not to insert the instruction into any particular 534 // basic block more than once. 535 std::set<BasicBlock*> InsertedBlocks; 536 BasicBlock *InstOrigBB = I.getParent(); 537 538 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 539 BasicBlock *ExitBlock = ExitBlocks[i]; 540 541 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 542 // If we haven't already processed this exit block, do so now. 543 if (InsertedBlocks.insert(ExitBlock).second) { 544 // Insert the code after the last PHI node... 545 BasicBlock::iterator InsertPt = ExitBlock->begin(); 546 while (isa<PHINode>(InsertPt)) ++InsertPt; 547 548 // If this is the first exit block processed, just move the original 549 // instruction, otherwise clone the original instruction and insert 550 // the copy. 551 Instruction *New; 552 if (InsertedBlocks.size() == 1) { 553 I.removeFromParent(); 554 ExitBlock->getInstList().insert(InsertPt, &I); 555 New = &I; 556 } else { 557 New = I.clone(); 558 CurAST->copyValue(&I, New); 559 if (!I.getName().empty()) 560 New->setName(I.getName()+".le"); 561 ExitBlock->getInstList().insert(InsertPt, New); 562 } 563 564 // Now that we have inserted the instruction, store it into the alloca 565 if (AI) new StoreInst(New, AI, InsertPt); 566 } 567 } 568 } 569 570 // If the instruction doesn't dominate any exit blocks, it must be dead. 571 if (InsertedBlocks.empty()) { 572 CurAST->deleteValue(&I); 573 I.eraseFromParent(); 574 } 575 576 // Finally, promote the fine value to SSA form. 577 if (AI) { 578 std::vector<AllocaInst*> Allocas; 579 Allocas.push_back(AI); 580 PromoteMemToReg(Allocas, *DT, *DF, CurAST); 581 } 582 } 583} 584 585/// hoist - When an instruction is found to only use loop invariant operands 586/// that is safe to hoist, this instruction is called to do the dirty work. 587/// 588void LICM::hoist(Instruction &I) { 589 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I; 590 591 // Remove the instruction from its current basic block... but don't delete the 592 // instruction. 593 I.removeFromParent(); 594 595 // Insert the new node in Preheader, before the terminator. 596 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 597 598 if (isa<LoadInst>(I)) ++NumMovedLoads; 599 else if (isa<CallInst>(I)) ++NumMovedCalls; 600 ++NumHoisted; 601 Changed = true; 602} 603 604/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 605/// not a trapping instruction or if it is a trapping instruction and is 606/// guaranteed to execute. 607/// 608bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 609 // If it is not a trapping instruction, it is always safe to hoist. 610 if (!Inst.isTrapping()) return true; 611 612 // Otherwise we have to check to make sure that the instruction dominates all 613 // of the exit blocks. If it doesn't, then there is a path out of the loop 614 // which does not execute this instruction, so we can't hoist it. 615 616 // If the instruction is in the header block for the loop (which is very 617 // common), it is always guaranteed to dominate the exit blocks. Since this 618 // is a common case, and can save some work, check it now. 619 if (Inst.getParent() == CurLoop->getHeader()) 620 return true; 621 622 // It's always safe to load from a global or alloca. 623 if (isa<LoadInst>(Inst)) 624 if (isa<AllocationInst>(Inst.getOperand(0)) || 625 isa<GlobalVariable>(Inst.getOperand(0))) 626 return true; 627 628 // Get the exit blocks for the current loop. 629 SmallVector<BasicBlock*, 8> ExitBlocks; 630 CurLoop->getExitBlocks(ExitBlocks); 631 632 // For each exit block, get the DT node and walk up the DT until the 633 // instruction's basic block is found or we exit the loop. 634 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 635 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 636 return false; 637 638 return true; 639} 640 641 642/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 643/// stores out of the loop and moving loads to before the loop. We do this by 644/// looping over the stores in the loop, looking for stores to Must pointers 645/// which are loop invariant. We promote these memory locations to use allocas 646/// instead. These allocas can easily be raised to register values by the 647/// PromoteMem2Reg functionality. 648/// 649void LICM::PromoteValuesInLoop() { 650 // PromotedValues - List of values that are promoted out of the loop. Each 651 // value has an alloca instruction for it, and a canonical version of the 652 // pointer. 653 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 654 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 655 656 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 657 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 658 659 Changed = true; 660 NumPromoted += PromotedValues.size(); 661 662 std::vector<Value*> PointerValueNumbers; 663 664 // Emit a copy from the value into the alloca'd value in the loop preheader 665 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 666 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 667 Value *Ptr = PromotedValues[i].second; 668 669 // If we are promoting a pointer value, update alias information for the 670 // inserted load. 671 Value *LoadValue = 0; 672 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 673 // Locate a load or store through the pointer, and assign the same value 674 // to LI as we are loading or storing. Since we know that the value is 675 // stored in this loop, this will always succeed. 676 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 677 UI != E; ++UI) 678 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 679 LoadValue = LI; 680 break; 681 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 682 if (SI->getOperand(1) == Ptr) { 683 LoadValue = SI->getOperand(0); 684 break; 685 } 686 } 687 assert(LoadValue && "No store through the pointer found!"); 688 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 689 } 690 691 // Load from the memory we are promoting. 692 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 693 694 if (LoadValue) CurAST->copyValue(LoadValue, LI); 695 696 // Store into the temporary alloca. 697 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 698 } 699 700 // Scan the basic blocks in the loop, replacing uses of our pointers with 701 // uses of the allocas in question. 702 // 703 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 704 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 705 E = LoopBBs.end(); I != E; ++I) { 706 // Rewrite all loads and stores in the block of the pointer... 707 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 708 II != E; ++II) { 709 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 710 std::map<Value*, AllocaInst*>::iterator 711 I = ValueToAllocaMap.find(L->getOperand(0)); 712 if (I != ValueToAllocaMap.end()) 713 L->setOperand(0, I->second); // Rewrite load instruction... 714 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 715 std::map<Value*, AllocaInst*>::iterator 716 I = ValueToAllocaMap.find(S->getOperand(1)); 717 if (I != ValueToAllocaMap.end()) 718 S->setOperand(1, I->second); // Rewrite store instruction... 719 } 720 } 721 } 722 723 // Now that the body of the loop uses the allocas instead of the original 724 // memory locations, insert code to copy the alloca value back into the 725 // original memory location on all exits from the loop. Note that we only 726 // want to insert one copy of the code in each exit block, though the loop may 727 // exit to the same block more than once. 728 // 729 std::set<BasicBlock*> ProcessedBlocks; 730 731 SmallVector<BasicBlock*, 8> ExitBlocks; 732 CurLoop->getExitBlocks(ExitBlocks); 733 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 734 if (ProcessedBlocks.insert(ExitBlocks[i]).second) { 735 // Copy all of the allocas into their memory locations. 736 BasicBlock::iterator BI = ExitBlocks[i]->begin(); 737 while (isa<PHINode>(*BI)) 738 ++BI; // Skip over all of the phi nodes in the block. 739 Instruction *InsertPos = BI; 740 unsigned PVN = 0; 741 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 742 // Load from the alloca. 743 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 744 745 // If this is a pointer type, update alias info appropriately. 746 if (isa<PointerType>(LI->getType())) 747 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 748 749 // Store into the memory we promoted. 750 new StoreInst(LI, PromotedValues[i].second, InsertPos); 751 } 752 } 753 754 // Now that we have done the deed, use the mem2reg functionality to promote 755 // all of the new allocas we just created into real SSA registers. 756 // 757 std::vector<AllocaInst*> PromotedAllocas; 758 PromotedAllocas.reserve(PromotedValues.size()); 759 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 760 PromotedAllocas.push_back(PromotedValues[i].first); 761 PromoteMemToReg(PromotedAllocas, *DT, *DF, CurAST); 762} 763 764/// FindPromotableValuesInLoop - Check the current loop for stores to definite 765/// pointers, which are not loaded and stored through may aliases and are safe 766/// for promotion. If these are found, create an alloca for the value, add it 767/// to the PromotedValues list, and keep track of the mapping from value to 768/// alloca. 769void LICM::FindPromotableValuesInLoop( 770 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 771 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 772 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 773 774 SmallVector<Instruction *, 4> LoopExits; 775 SmallVector<BasicBlock *, 4> Blocks; 776 CurLoop->getExitingBlocks(Blocks); 777 for (SmallVector<BasicBlock *, 4>::iterator BI = Blocks.begin(), 778 BE = Blocks.end(); BI != BE; ++BI) { 779 BasicBlock *BB = *BI; 780 LoopExits.push_back(BB->getTerminator()); 781 } 782 783 // Loop over all of the alias sets in the tracker object. 784 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 785 I != E; ++I) { 786 AliasSet &AS = *I; 787 // We can promote this alias set if it has a store, if it is a "Must" alias 788 // set, if the pointer is loop invariant, and if we are not eliminating any 789 // volatile loads or stores. 790 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() && 791 !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) { 792 assert(!AS.empty() && 793 "Must alias set should have at least one pointer element in it!"); 794 Value *V = AS.begin()->first; 795 796 // Check that all of the pointers in the alias set have the same type. We 797 // cannot (yet) promote a memory location that is loaded and stored in 798 // different sizes. 799 bool PointerOk = true; 800 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 801 if (V->getType() != I->first->getType()) { 802 PointerOk = false; 803 break; 804 } 805 806 // If one use of value V inside the loop is safe then it is OK to promote 807 // this value. On the otherside if there is not any unsafe use inside the 808 // loop then also it is OK to promote this value. Otherwise it is 809 // unsafe to promote this value. 810 if (PointerOk) { 811 bool oneSafeUse = false; 812 bool oneUnsafeUse = false; 813 for(Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 814 UI != UE; ++UI) { 815 Instruction *Use = dyn_cast<Instruction>(*UI); 816 if (!Use || !CurLoop->contains(Use->getParent())) 817 continue; 818 for (SmallVector<Instruction *, 4>::iterator 819 ExitI = LoopExits.begin(), ExitE = LoopExits.end(); 820 ExitI != ExitE; ++ExitI) { 821 Instruction *Ex = *ExitI; 822 if (!isa<PHINode>(Use) && DT->dominates(Use, Ex)) { 823 oneSafeUse = true; 824 break; 825 } 826 else 827 oneUnsafeUse = true; 828 } 829 830 if (oneSafeUse) 831 break; 832 } 833 834 if (oneSafeUse) 835 PointerOk = true; 836 else if (!oneUnsafeUse) 837 PointerOk = true; 838 else 839 PointerOk = false; 840 } 841 842 if (PointerOk) { 843 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 844 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 845 PromotedValues.push_back(std::make_pair(AI, V)); 846 847 // Update the AST and alias analysis. 848 CurAST->copyValue(V, AI); 849 850 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 851 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 852 853 DOUT << "LICM: Promoting value: " << *V << "\n"; 854 } 855 } 856 } 857} 858 859/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 860void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 861 AliasSetTracker *AST = LoopToAliasMap[L]; 862 if (!AST) 863 return; 864 865 AST->copyValue(From, To); 866} 867 868/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 869/// set. 870void LICM::deleteAnalysisValue(Value *V, Loop *L) { 871 AliasSetTracker *AST = LoopToAliasMap[L]; 872 if (!AST) 873 return; 874 875 AST->deleteValue(V); 876} 877