LICM.cpp revision 9ccaf53ada99c63737547c0235baeb8454b04e80
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/IntrinsicInst.h" 39#include "llvm/Instructions.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/CommandLine.h" 50#include "llvm/Support/raw_ostream.h" 51#include "llvm/Support/Debug.h" 52#include "llvm/ADT/Statistic.h" 53#include <algorithm> 54using namespace llvm; 55 56STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 57STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 58STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 59STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 60STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 61 62static cl::opt<bool> 63DisablePromotion("disable-licm-promotion", cl::Hidden, 64 cl::desc("Disable memory promotion in LICM pass")); 65 66namespace { 67 struct LICM : public LoopPass { 68 static char ID; // Pass identification, replacement for typeid 69 LICM() : LoopPass(ID) {} 70 71 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 72 73 /// This transformation requires natural loop information & requires that 74 /// loop preheaders be inserted into the CFG... 75 /// 76 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 77 AU.setPreservesCFG(); 78 AU.addRequired<DominatorTree>(); 79 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 80 AU.addRequired<LoopInfo>(); 81 AU.addRequiredID(LoopSimplifyID); 82 AU.addRequired<AliasAnalysis>(); 83 AU.addPreserved<ScalarEvolution>(); 84 AU.addPreserved<DominanceFrontier>(); 85 AU.addPreservedID(LoopSimplifyID); 86 } 87 88 bool doFinalization() { 89 // Free the values stored in the map 90 for (std::map<Loop *, AliasSetTracker *>::iterator 91 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I) 92 delete I->second; 93 94 LoopToAliasMap.clear(); 95 return false; 96 } 97 98 private: 99 // Various analyses that we use... 100 AliasAnalysis *AA; // Current AliasAnalysis information 101 LoopInfo *LI; // Current LoopInfo 102 DominatorTree *DT; // Dominator Tree for the current Loop... 103 DominanceFrontier *DF; // Current Dominance Frontier 104 105 // State that is updated as we process loops 106 bool Changed; // Set to true when we change anything. 107 BasicBlock *Preheader; // The preheader block of the current loop... 108 Loop *CurLoop; // The current loop we are working on... 109 AliasSetTracker *CurAST; // AliasSet information for the current loop... 110 std::map<Loop *, AliasSetTracker *> LoopToAliasMap; 111 112 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 113 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 114 115 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 116 /// set. 117 void deleteAnalysisValue(Value *V, Loop *L); 118 119 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 120 /// dominated by the specified block, and that are in the current loop) in 121 /// reverse depth first order w.r.t the DominatorTree. This allows us to 122 /// visit uses before definitions, allowing us to sink a loop body in one 123 /// pass without iteration. 124 /// 125 void SinkRegion(DomTreeNode *N); 126 127 /// HoistRegion - Walk the specified region of the CFG (defined by all 128 /// blocks dominated by the specified block, and that are in the current 129 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 130 /// visit definitions before uses, allowing us to hoist a loop body in one 131 /// pass without iteration. 132 /// 133 void HoistRegion(DomTreeNode *N); 134 135 /// inSubLoop - Little predicate that returns true if the specified basic 136 /// block is in a subloop of the current one, not the current one itself. 137 /// 138 bool inSubLoop(BasicBlock *BB) { 139 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 140 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 141 if ((*I)->contains(BB)) 142 return true; // A subloop actually contains this block! 143 return false; 144 } 145 146 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 147 /// specified exit block of the loop is dominated by the specified block 148 /// that is in the body of the loop. We use these constraints to 149 /// dramatically limit the amount of the dominator tree that needs to be 150 /// searched. 151 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 152 BasicBlock *BlockInLoop) const { 153 // If the block in the loop is the loop header, it must be dominated! 154 BasicBlock *LoopHeader = CurLoop->getHeader(); 155 if (BlockInLoop == LoopHeader) 156 return true; 157 158 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 159 DomTreeNode *IDom = DT->getNode(ExitBlock); 160 161 // Because the exit block is not in the loop, we know we have to get _at 162 // least_ its immediate dominator. 163 IDom = IDom->getIDom(); 164 165 while (IDom && IDom != BlockInLoopNode) { 166 // If we have got to the header of the loop, then the instructions block 167 // did not dominate the exit node, so we can't hoist it. 168 if (IDom->getBlock() == LoopHeader) 169 return false; 170 171 // Get next Immediate Dominator. 172 IDom = IDom->getIDom(); 173 }; 174 175 return true; 176 } 177 178 /// sink - When an instruction is found to only be used outside of the loop, 179 /// this function moves it to the exit blocks and patches up SSA form as 180 /// needed. 181 /// 182 void sink(Instruction &I); 183 184 /// hoist - When an instruction is found to only use loop invariant operands 185 /// that is safe to hoist, this instruction is called to do the dirty work. 186 /// 187 void hoist(Instruction &I); 188 189 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 190 /// is not a trapping instruction or if it is a trapping instruction and is 191 /// guaranteed to execute. 192 /// 193 bool isSafeToExecuteUnconditionally(Instruction &I); 194 195 /// pointerInvalidatedByLoop - Return true if the body of this loop may 196 /// store into the memory location pointed to by V. 197 /// 198 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 199 // Check to see if any of the basic blocks in CurLoop invalidate *V. 200 return CurAST->getAliasSetForPointer(V, Size).isMod(); 201 } 202 203 bool canSinkOrHoistInst(Instruction &I); 204 bool isLoopInvariantInst(Instruction &I); 205 bool isNotUsedInLoop(Instruction &I); 206 207 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 208 /// to scalars as we can. 209 /// 210 void PromoteValuesInLoop(); 211 212 /// FindPromotableValuesInLoop - Check the current loop for stores to 213 /// definite pointers, which are not loaded and stored through may aliases. 214 /// If these are found, create an alloca for the value, add it to the 215 /// PromotedValues list, and keep track of the mapping from value to 216 /// alloca... 217 /// 218 void FindPromotableValuesInLoop( 219 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 220 std::map<Value*, AllocaInst*> &Val2AlMap); 221 }; 222} 223 224char LICM::ID = 0; 225INITIALIZE_PASS(LICM, "licm", "Loop Invariant Code Motion", false, false); 226 227Pass *llvm::createLICMPass() { return new LICM(); } 228 229/// Hoist expressions out of the specified loop. Note, alias info for inner 230/// loop is not preserved so it is not a good idea to run LICM multiple 231/// times on one loop. 232/// 233bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 234 Changed = false; 235 236 // Get our Loop and Alias Analysis information... 237 LI = &getAnalysis<LoopInfo>(); 238 AA = &getAnalysis<AliasAnalysis>(); 239 DF = &getAnalysis<DominanceFrontier>(); 240 DT = &getAnalysis<DominatorTree>(); 241 242 CurAST = new AliasSetTracker(*AA); 243 // Collect Alias info from subloops 244 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 245 LoopItr != LoopItrE; ++LoopItr) { 246 Loop *InnerL = *LoopItr; 247 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 248 assert (InnerAST && "Where is my AST?"); 249 250 // What if InnerLoop was modified by other passes ? 251 CurAST->add(*InnerAST); 252 } 253 254 CurLoop = L; 255 256 // Get the preheader block to move instructions into... 257 Preheader = L->getLoopPreheader(); 258 259 // Loop over the body of this loop, looking for calls, invokes, and stores. 260 // Because subloops have already been incorporated into AST, we skip blocks in 261 // subloops. 262 // 263 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 264 I != E; ++I) { 265 BasicBlock *BB = *I; 266 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops... 267 CurAST->add(*BB); // Incorporate the specified basic block 268 } 269 270 // We want to visit all of the instructions in this loop... that are not parts 271 // of our subloops (they have already had their invariants hoisted out of 272 // their loop, into this loop, so there is no need to process the BODIES of 273 // the subloops). 274 // 275 // Traverse the body of the loop in depth first order on the dominator tree so 276 // that we are guaranteed to see definitions before we see uses. This allows 277 // us to sink instructions in one pass, without iteration. After sinking 278 // instructions, we perform another pass to hoist them out of the loop. 279 // 280 if (L->hasDedicatedExits()) 281 SinkRegion(DT->getNode(L->getHeader())); 282 if (Preheader) 283 HoistRegion(DT->getNode(L->getHeader())); 284 285 // Now that all loop invariants have been removed from the loop, promote any 286 // memory references to scalars that we can... 287 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) 288 PromoteValuesInLoop(); 289 290 // Clear out loops state information for the next iteration 291 CurLoop = 0; 292 Preheader = 0; 293 294 LoopToAliasMap[L] = CurAST; 295 return Changed; 296} 297 298/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 299/// dominated by the specified block, and that are in the current loop) in 300/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 301/// uses before definitions, allowing us to sink a loop body in one pass without 302/// iteration. 303/// 304void LICM::SinkRegion(DomTreeNode *N) { 305 assert(N != 0 && "Null dominator tree node?"); 306 BasicBlock *BB = N->getBlock(); 307 308 // If this subregion is not in the top level loop at all, exit. 309 if (!CurLoop->contains(BB)) return; 310 311 // We are processing blocks in reverse dfo, so process children first... 312 const std::vector<DomTreeNode*> &Children = N->getChildren(); 313 for (unsigned i = 0, e = Children.size(); i != e; ++i) 314 SinkRegion(Children[i]); 315 316 // Only need to process the contents of this block if it is not part of a 317 // subloop (which would already have been processed). 318 if (inSubLoop(BB)) return; 319 320 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 321 Instruction &I = *--II; 322 323 // Check to see if we can sink this instruction to the exit blocks 324 // of the loop. We can do this if the all users of the instruction are 325 // outside of the loop. In this case, it doesn't even matter if the 326 // operands of the instruction are loop invariant. 327 // 328 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 329 ++II; 330 sink(I); 331 } 332 } 333} 334 335/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 336/// dominated by the specified block, and that are in the current loop) in depth 337/// first order w.r.t the DominatorTree. This allows us to visit definitions 338/// before uses, allowing us to hoist a loop body in one pass without iteration. 339/// 340void LICM::HoistRegion(DomTreeNode *N) { 341 assert(N != 0 && "Null dominator tree node?"); 342 BasicBlock *BB = N->getBlock(); 343 344 // If this subregion is not in the top level loop at all, exit. 345 if (!CurLoop->contains(BB)) return; 346 347 // Only need to process the contents of this block if it is not part of a 348 // subloop (which would already have been processed). 349 if (!inSubLoop(BB)) 350 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 351 Instruction &I = *II++; 352 353 // Try hoisting the instruction out to the preheader. We can only do this 354 // if all of the operands of the instruction are loop invariant and if it 355 // is safe to hoist the instruction. 356 // 357 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 358 isSafeToExecuteUnconditionally(I)) 359 hoist(I); 360 } 361 362 const std::vector<DomTreeNode*> &Children = N->getChildren(); 363 for (unsigned i = 0, e = Children.size(); i != e; ++i) 364 HoistRegion(Children[i]); 365} 366 367/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 368/// instruction. 369/// 370bool LICM::canSinkOrHoistInst(Instruction &I) { 371 // Loads have extra constraints we have to verify before we can hoist them. 372 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 373 if (LI->isVolatile()) 374 return false; // Don't hoist volatile loads! 375 376 // Loads from constant memory are always safe to move, even if they end up 377 // in the same alias set as something that ends up being modified. 378 if (AA->pointsToConstantMemory(LI->getOperand(0))) 379 return true; 380 381 // Don't hoist loads which have may-aliased stores in loop. 382 unsigned Size = 0; 383 if (LI->getType()->isSized()) 384 Size = AA->getTypeStoreSize(LI->getType()); 385 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 386 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 387 // Handle obvious cases efficiently. 388 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 389 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 390 return true; 391 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 392 // If this call only reads from memory and there are no writes to memory 393 // in the loop, we can hoist or sink the call as appropriate. 394 bool FoundMod = false; 395 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 396 I != E; ++I) { 397 AliasSet &AS = *I; 398 if (!AS.isForwardingAliasSet() && AS.isMod()) { 399 FoundMod = true; 400 break; 401 } 402 } 403 if (!FoundMod) return true; 404 } 405 406 // FIXME: This should use mod/ref information to see if we can hoist or sink 407 // the call. 408 409 return false; 410 } 411 412 // Otherwise these instructions are hoistable/sinkable 413 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 414 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 415 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 416 isa<ShuffleVectorInst>(I); 417} 418 419/// isNotUsedInLoop - Return true if the only users of this instruction are 420/// outside of the loop. If this is true, we can sink the instruction to the 421/// exit blocks of the loop. 422/// 423bool LICM::isNotUsedInLoop(Instruction &I) { 424 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 425 Instruction *User = cast<Instruction>(*UI); 426 if (PHINode *PN = dyn_cast<PHINode>(User)) { 427 // PHI node uses occur in predecessor blocks! 428 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 429 if (PN->getIncomingValue(i) == &I) 430 if (CurLoop->contains(PN->getIncomingBlock(i))) 431 return false; 432 } else if (CurLoop->contains(User)) { 433 return false; 434 } 435 } 436 return true; 437} 438 439 440/// isLoopInvariantInst - Return true if all operands of this instruction are 441/// loop invariant. We also filter out non-hoistable instructions here just for 442/// efficiency. 443/// 444bool LICM::isLoopInvariantInst(Instruction &I) { 445 // The instruction is loop invariant if all of its operands are loop-invariant 446 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 447 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 448 return false; 449 450 // If we got this far, the instruction is loop invariant! 451 return true; 452} 453 454/// sink - When an instruction is found to only be used outside of the loop, 455/// this function moves it to the exit blocks and patches up SSA form as needed. 456/// This method is guaranteed to remove the original instruction from its 457/// position, and may either delete it or move it to outside of the loop. 458/// 459void LICM::sink(Instruction &I) { 460 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 461 462 SmallVector<BasicBlock*, 8> ExitBlocks; 463 CurLoop->getExitBlocks(ExitBlocks); 464 465 if (isa<LoadInst>(I)) ++NumMovedLoads; 466 else if (isa<CallInst>(I)) ++NumMovedCalls; 467 ++NumSunk; 468 Changed = true; 469 470 // The case where there is only a single exit node of this loop is common 471 // enough that we handle it as a special (more efficient) case. It is more 472 // efficient to handle because there are no PHI nodes that need to be placed. 473 if (ExitBlocks.size() == 1) { 474 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 475 // Instruction is not used, just delete it. 476 CurAST->deleteValue(&I); 477 // If I has users in unreachable blocks, eliminate. 478 // If I is not void type then replaceAllUsesWith undef. 479 // This allows ValueHandlers and custom metadata to adjust itself. 480 if (!I.getType()->isVoidTy()) 481 I.replaceAllUsesWith(UndefValue::get(I.getType())); 482 I.eraseFromParent(); 483 } else { 484 // Move the instruction to the start of the exit block, after any PHI 485 // nodes in it. 486 I.removeFromParent(); 487 BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI(); 488 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 489 } 490 } else if (ExitBlocks.empty()) { 491 // The instruction is actually dead if there ARE NO exit blocks. 492 CurAST->deleteValue(&I); 493 // If I has users in unreachable blocks, eliminate. 494 // If I is not void type then replaceAllUsesWith undef. 495 // This allows ValueHandlers and custom metadata to adjust itself. 496 if (!I.getType()->isVoidTy()) 497 I.replaceAllUsesWith(UndefValue::get(I.getType())); 498 I.eraseFromParent(); 499 } else { 500 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 501 // do all of the hard work of inserting PHI nodes as necessary. We convert 502 // the value into a stack object to get it to do this. 503 504 // Firstly, we create a stack object to hold the value... 505 AllocaInst *AI = 0; 506 507 if (!I.getType()->isVoidTy()) { 508 AI = new AllocaInst(I.getType(), 0, I.getName(), 509 I.getParent()->getParent()->getEntryBlock().begin()); 510 CurAST->add(AI); 511 } 512 513 // Secondly, insert load instructions for each use of the instruction 514 // outside of the loop. 515 while (!I.use_empty()) { 516 Instruction *U = cast<Instruction>(I.use_back()); 517 518 // If the user is a PHI Node, we actually have to insert load instructions 519 // in all predecessor blocks, not in the PHI block itself! 520 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 521 // Only insert into each predecessor once, so that we don't have 522 // different incoming values from the same block! 523 std::map<BasicBlock*, Value*> InsertedBlocks; 524 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 525 if (UPN->getIncomingValue(i) == &I) { 526 BasicBlock *Pred = UPN->getIncomingBlock(i); 527 Value *&PredVal = InsertedBlocks[Pred]; 528 if (!PredVal) { 529 // Insert a new load instruction right before the terminator in 530 // the predecessor block. 531 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 532 CurAST->add(cast<LoadInst>(PredVal)); 533 } 534 535 UPN->setIncomingValue(i, PredVal); 536 } 537 538 } else { 539 LoadInst *L = new LoadInst(AI, "", U); 540 U->replaceUsesOfWith(&I, L); 541 CurAST->add(L); 542 } 543 } 544 545 // Thirdly, insert a copy of the instruction in each exit block of the loop 546 // that is dominated by the instruction, storing the result into the memory 547 // location. Be careful not to insert the instruction into any particular 548 // basic block more than once. 549 std::set<BasicBlock*> InsertedBlocks; 550 BasicBlock *InstOrigBB = I.getParent(); 551 552 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 553 BasicBlock *ExitBlock = ExitBlocks[i]; 554 555 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 556 // If we haven't already processed this exit block, do so now. 557 if (InsertedBlocks.insert(ExitBlock).second) { 558 // Insert the code after the last PHI node... 559 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 560 561 // If this is the first exit block processed, just move the original 562 // instruction, otherwise clone the original instruction and insert 563 // the copy. 564 Instruction *New; 565 if (InsertedBlocks.size() == 1) { 566 I.removeFromParent(); 567 ExitBlock->getInstList().insert(InsertPt, &I); 568 New = &I; 569 } else { 570 New = I.clone(); 571 CurAST->copyValue(&I, New); 572 if (!I.getName().empty()) 573 New->setName(I.getName()+".le"); 574 ExitBlock->getInstList().insert(InsertPt, New); 575 } 576 577 // Now that we have inserted the instruction, store it into the alloca 578 if (AI) new StoreInst(New, AI, InsertPt); 579 } 580 } 581 } 582 583 // If the instruction doesn't dominate any exit blocks, it must be dead. 584 if (InsertedBlocks.empty()) { 585 CurAST->deleteValue(&I); 586 I.eraseFromParent(); 587 } 588 589 // Finally, promote the fine value to SSA form. 590 if (AI) { 591 std::vector<AllocaInst*> Allocas; 592 Allocas.push_back(AI); 593 PromoteMemToReg(Allocas, *DT, *DF, CurAST); 594 } 595 } 596} 597 598/// hoist - When an instruction is found to only use loop invariant operands 599/// that is safe to hoist, this instruction is called to do the dirty work. 600/// 601void LICM::hoist(Instruction &I) { 602 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 603 << I << "\n"); 604 605 // Remove the instruction from its current basic block... but don't delete the 606 // instruction. 607 I.removeFromParent(); 608 609 // Insert the new node in Preheader, before the terminator. 610 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 611 612 if (isa<LoadInst>(I)) ++NumMovedLoads; 613 else if (isa<CallInst>(I)) ++NumMovedCalls; 614 ++NumHoisted; 615 Changed = true; 616} 617 618/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 619/// not a trapping instruction or if it is a trapping instruction and is 620/// guaranteed to execute. 621/// 622bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 623 // If it is not a trapping instruction, it is always safe to hoist. 624 if (Inst.isSafeToSpeculativelyExecute()) 625 return true; 626 627 // Otherwise we have to check to make sure that the instruction dominates all 628 // of the exit blocks. If it doesn't, then there is a path out of the loop 629 // which does not execute this instruction, so we can't hoist it. 630 631 // If the instruction is in the header block for the loop (which is very 632 // common), it is always guaranteed to dominate the exit blocks. Since this 633 // is a common case, and can save some work, check it now. 634 if (Inst.getParent() == CurLoop->getHeader()) 635 return true; 636 637 // Get the exit blocks for the current loop. 638 SmallVector<BasicBlock*, 8> ExitBlocks; 639 CurLoop->getExitBlocks(ExitBlocks); 640 641 // For each exit block, get the DT node and walk up the DT until the 642 // instruction's basic block is found or we exit the loop. 643 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 644 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 645 return false; 646 647 return true; 648} 649 650 651/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 652/// stores out of the loop and moving loads to before the loop. We do this by 653/// looping over the stores in the loop, looking for stores to Must pointers 654/// which are loop invariant. We promote these memory locations to use allocas 655/// instead. These allocas can easily be raised to register values by the 656/// PromoteMem2Reg functionality. 657/// 658void LICM::PromoteValuesInLoop() { 659 // PromotedValues - List of values that are promoted out of the loop. Each 660 // value has an alloca instruction for it, and a canonical version of the 661 // pointer. 662 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 663 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 664 665 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 666 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 667 668 Changed = true; 669 NumPromoted += PromotedValues.size(); 670 671 std::vector<Value*> PointerValueNumbers; 672 673 // Emit a copy from the value into the alloca'd value in the loop preheader 674 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 675 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 676 Value *Ptr = PromotedValues[i].second; 677 678 // If we are promoting a pointer value, update alias information for the 679 // inserted load. 680 Value *LoadValue = 0; 681 if (cast<PointerType>(Ptr->getType())->getElementType()->isPointerTy()) { 682 // Locate a load or store through the pointer, and assign the same value 683 // to LI as we are loading or storing. Since we know that the value is 684 // stored in this loop, this will always succeed. 685 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 686 UI != E; ++UI) { 687 User *U = *UI; 688 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 689 LoadValue = LI; 690 break; 691 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 692 if (SI->getOperand(1) == Ptr) { 693 LoadValue = SI->getOperand(0); 694 break; 695 } 696 } 697 } 698 assert(LoadValue && "No store through the pointer found!"); 699 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 700 } 701 702 // Load from the memory we are promoting. 703 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 704 705 if (LoadValue) CurAST->copyValue(LoadValue, LI); 706 707 // Store into the temporary alloca. 708 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 709 } 710 711 // Scan the basic blocks in the loop, replacing uses of our pointers with 712 // uses of the allocas in question. 713 // 714 for (Loop::block_iterator I = CurLoop->block_begin(), 715 E = CurLoop->block_end(); I != E; ++I) { 716 BasicBlock *BB = *I; 717 // Rewrite all loads and stores in the block of the pointer... 718 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 719 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 720 std::map<Value*, AllocaInst*>::iterator 721 I = ValueToAllocaMap.find(L->getOperand(0)); 722 if (I != ValueToAllocaMap.end()) 723 L->setOperand(0, I->second); // Rewrite load instruction... 724 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 725 std::map<Value*, AllocaInst*>::iterator 726 I = ValueToAllocaMap.find(S->getOperand(1)); 727 if (I != ValueToAllocaMap.end()) 728 S->setOperand(1, I->second); // Rewrite store instruction... 729 } 730 } 731 } 732 733 // Now that the body of the loop uses the allocas instead of the original 734 // memory locations, insert code to copy the alloca value back into the 735 // original memory location on all exits from the loop. Note that we only 736 // want to insert one copy of the code in each exit block, though the loop may 737 // exit to the same block more than once. 738 // 739 SmallPtrSet<BasicBlock*, 16> ProcessedBlocks; 740 741 SmallVector<BasicBlock*, 8> ExitBlocks; 742 CurLoop->getExitBlocks(ExitBlocks); 743 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 744 if (!ProcessedBlocks.insert(ExitBlocks[i])) 745 continue; 746 747 // Copy all of the allocas into their memory locations. 748 BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI(); 749 Instruction *InsertPos = BI; 750 unsigned PVN = 0; 751 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 752 // Load from the alloca. 753 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 754 755 // If this is a pointer type, update alias info appropriately. 756 if (LI->getType()->isPointerTy()) 757 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 758 759 // Store into the memory we promoted. 760 new StoreInst(LI, PromotedValues[i].second, InsertPos); 761 } 762 } 763 764 // Now that we have done the deed, use the mem2reg functionality to promote 765 // all of the new allocas we just created into real SSA registers. 766 // 767 std::vector<AllocaInst*> PromotedAllocas; 768 PromotedAllocas.reserve(PromotedValues.size()); 769 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 770 PromotedAllocas.push_back(PromotedValues[i].first); 771 PromoteMemToReg(PromotedAllocas, *DT, *DF, CurAST); 772} 773 774/// FindPromotableValuesInLoop - Check the current loop for stores to definite 775/// pointers, which are not loaded and stored through may aliases and are safe 776/// for promotion. If these are found, create an alloca for the value, add it 777/// to the PromotedValues list, and keep track of the mapping from value to 778/// alloca. 779void LICM::FindPromotableValuesInLoop( 780 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 781 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 782 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 783 784 // Loop over all of the alias sets in the tracker object. 785 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 786 I != E; ++I) { 787 AliasSet &AS = *I; 788 // We can promote this alias set if it has a store, if it is a "Must" alias 789 // set, if the pointer is loop invariant, and if we are not eliminating any 790 // volatile loads or stores. 791 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 792 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 793 continue; 794 795 assert(!AS.empty() && 796 "Must alias set should have at least one pointer element in it!"); 797 Value *V = AS.begin()->getValue(); 798 799 // Check that all of the pointers in the alias set have the same type. We 800 // cannot (yet) promote a memory location that is loaded and stored in 801 // different sizes. 802 { 803 bool PointerOk = true; 804 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 805 if (V->getType() != I->getValue()->getType()) { 806 PointerOk = false; 807 break; 808 } 809 if (!PointerOk) 810 continue; 811 } 812 813 // It isn't safe to promote a load/store from the loop if the load/store is 814 // conditional. For example, turning: 815 // 816 // for () { if (c) *P += 1; } 817 // 818 // into: 819 // 820 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 821 // 822 // is not safe, because *P may only be valid to access if 'c' is true. 823 // 824 // It is safe to promote P if all uses are direct load/stores and if at 825 // least one is guaranteed to be executed. 826 bool GuaranteedToExecute = false; 827 bool InvalidInst = false; 828 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 829 UI != UE; ++UI) { 830 // Ignore instructions not in this loop. 831 Instruction *Use = dyn_cast<Instruction>(*UI); 832 if (!Use || !CurLoop->contains(Use)) 833 continue; 834 835 if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) { 836 InvalidInst = true; 837 break; 838 } 839 840 if (!GuaranteedToExecute) 841 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 842 } 843 844 // If there is an non-load/store instruction in the loop, we can't promote 845 // it. If there isn't a guaranteed-to-execute instruction, we can't 846 // promote. 847 if (InvalidInst || !GuaranteedToExecute) 848 continue; 849 850 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 851 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 852 PromotedValues.push_back(std::make_pair(AI, V)); 853 854 // Update the AST and alias analysis. 855 CurAST->copyValue(V, AI); 856 857 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 858 ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI)); 859 860 DEBUG(dbgs() << "LICM: Promoting value: " << *V << "\n"); 861 } 862} 863 864/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 865void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 866 AliasSetTracker *AST = LoopToAliasMap[L]; 867 if (!AST) 868 return; 869 870 AST->copyValue(From, To); 871} 872 873/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 874/// set. 875void LICM::deleteAnalysisValue(Value *V, Loop *L) { 876 AliasSetTracker *AST = LoopToAliasMap[L]; 877 if (!AST) 878 return; 879 880 AST->deleteValue(V); 881} 882