LICM.cpp revision 199989
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.addRequiredID(LoopSimplifyID); 79 AU.addRequired<LoopInfo>(); 80 AU.addRequired<DominatorTree>(); 81 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 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 do { 164 // Get next Immediate Dominator. 165 IDom = IDom->getIDom(); 166 167 // If we have got to the header of the loop, then the instructions block 168 // did not dominate the exit node, so we can't hoist it. 169 if (IDom->getBlock() == LoopHeader) 170 return false; 171 172 } while (IDom != BlockInLoopNode); 173 174 return true; 175 } 176 177 /// sink - When an instruction is found to only be used outside of the loop, 178 /// this function moves it to the exit blocks and patches up SSA form as 179 /// needed. 180 /// 181 void sink(Instruction &I); 182 183 /// hoist - When an instruction is found to only use loop invariant operands 184 /// that is safe to hoist, this instruction is called to do the dirty work. 185 /// 186 void hoist(Instruction &I); 187 188 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 189 /// is not a trapping instruction or if it is a trapping instruction and is 190 /// guaranteed to execute. 191 /// 192 bool isSafeToExecuteUnconditionally(Instruction &I); 193 194 /// pointerInvalidatedByLoop - Return true if the body of this loop may 195 /// store into the memory location pointed to by V. 196 /// 197 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 198 // Check to see if any of the basic blocks in CurLoop invalidate *V. 199 return CurAST->getAliasSetForPointer(V, Size).isMod(); 200 } 201 202 bool canSinkOrHoistInst(Instruction &I); 203 bool isLoopInvariantInst(Instruction &I); 204 bool isNotUsedInLoop(Instruction &I); 205 206 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 207 /// to scalars as we can. 208 /// 209 void PromoteValuesInLoop(); 210 211 /// FindPromotableValuesInLoop - Check the current loop for stores to 212 /// definite pointers, which are not loaded and stored through may aliases. 213 /// If these are found, create an alloca for the value, add it to the 214 /// PromotedValues list, and keep track of the mapping from value to 215 /// alloca... 216 /// 217 void FindPromotableValuesInLoop( 218 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 219 std::map<Value*, AllocaInst*> &Val2AlMap); 220 }; 221} 222 223char LICM::ID = 0; 224static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 225 226Pass *llvm::createLICMPass() { return new LICM(); } 227 228/// Hoist expressions out of the specified loop. Note, alias info for inner 229/// loop is not preserved so it is not a good idea to run LICM multiple 230/// times on one loop. 231/// 232bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 233 Changed = false; 234 235 // Get our Loop and Alias Analysis information... 236 LI = &getAnalysis<LoopInfo>(); 237 AA = &getAnalysis<AliasAnalysis>(); 238 DF = &getAnalysis<DominanceFrontier>(); 239 DT = &getAnalysis<DominatorTree>(); 240 241 CurAST = new AliasSetTracker(*AA); 242 // Collect Alias info from subloops 243 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 244 LoopItr != LoopItrE; ++LoopItr) { 245 Loop *InnerL = *LoopItr; 246 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 247 assert (InnerAST && "Where is my AST?"); 248 249 // What if InnerLoop was modified by other passes ? 250 CurAST->add(*InnerAST); 251 } 252 253 CurLoop = L; 254 255 // Get the preheader block to move instructions into... 256 Preheader = L->getLoopPreheader(); 257 258 // Loop over the body of this loop, looking for calls, invokes, and stores. 259 // Because subloops have already been incorporated into AST, we skip blocks in 260 // subloops. 261 // 262 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 263 I != E; ++I) { 264 BasicBlock *BB = *I; 265 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops... 266 CurAST->add(*BB); // Incorporate the specified basic block 267 } 268 269 // We want to visit all of the instructions in this loop... that are not parts 270 // of our subloops (they have already had their invariants hoisted out of 271 // their loop, into this loop, so there is no need to process the BODIES of 272 // the subloops). 273 // 274 // Traverse the body of the loop in depth first order on the dominator tree so 275 // that we are guaranteed to see definitions before we see uses. This allows 276 // us to sink instructions in one pass, without iteration. After sinking 277 // instructions, we perform another pass to hoist them out of the loop. 278 // 279 if (L->hasDedicatedExits()) 280 SinkRegion(DT->getNode(L->getHeader())); 281 if (Preheader) 282 HoistRegion(DT->getNode(L->getHeader())); 283 284 // Now that all loop invariants have been removed from the loop, promote any 285 // memory references to scalars that we can... 286 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) 287 PromoteValuesInLoop(); 288 289 // Clear out loops state information for the next iteration 290 CurLoop = 0; 291 Preheader = 0; 292 293 LoopToAliasMap[L] = CurAST; 294 return Changed; 295} 296 297/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 298/// dominated by the specified block, and that are in the current loop) in 299/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 300/// uses before definitions, allowing us to sink a loop body in one pass without 301/// iteration. 302/// 303void LICM::SinkRegion(DomTreeNode *N) { 304 assert(N != 0 && "Null dominator tree node?"); 305 BasicBlock *BB = N->getBlock(); 306 307 // If this subregion is not in the top level loop at all, exit. 308 if (!CurLoop->contains(BB)) return; 309 310 // We are processing blocks in reverse dfo, so process children first... 311 const std::vector<DomTreeNode*> &Children = N->getChildren(); 312 for (unsigned i = 0, e = Children.size(); i != e; ++i) 313 SinkRegion(Children[i]); 314 315 // Only need to process the contents of this block if it is not part of a 316 // subloop (which would already have been processed). 317 if (inSubLoop(BB)) return; 318 319 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 320 Instruction &I = *--II; 321 322 // Check to see if we can sink this instruction to the exit blocks 323 // of the loop. We can do this if the all users of the instruction are 324 // outside of the loop. In this case, it doesn't even matter if the 325 // operands of the instruction are loop invariant. 326 // 327 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 328 ++II; 329 sink(I); 330 } 331 } 332} 333 334/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 335/// dominated by the specified block, and that are in the current loop) in depth 336/// first order w.r.t the DominatorTree. This allows us to visit definitions 337/// before uses, allowing us to hoist a loop body in one pass without iteration. 338/// 339void LICM::HoistRegion(DomTreeNode *N) { 340 assert(N != 0 && "Null dominator tree node?"); 341 BasicBlock *BB = N->getBlock(); 342 343 // If this subregion is not in the top level loop at all, exit. 344 if (!CurLoop->contains(BB)) return; 345 346 // Only need to process the contents of this block if it is not part of a 347 // subloop (which would already have been processed). 348 if (!inSubLoop(BB)) 349 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 350 Instruction &I = *II++; 351 352 // Try hoisting the instruction out to the preheader. We can only do this 353 // if all of the operands of the instruction are loop invariant and if it 354 // is safe to hoist the instruction. 355 // 356 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 357 isSafeToExecuteUnconditionally(I)) 358 hoist(I); 359 } 360 361 const std::vector<DomTreeNode*> &Children = N->getChildren(); 362 for (unsigned i = 0, e = Children.size(); i != e; ++i) 363 HoistRegion(Children[i]); 364} 365 366/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 367/// instruction. 368/// 369bool LICM::canSinkOrHoistInst(Instruction &I) { 370 // Loads have extra constraints we have to verify before we can hoist them. 371 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 372 if (LI->isVolatile()) 373 return false; // Don't hoist volatile loads! 374 375 // Loads from constant memory are always safe to move, even if they end up 376 // in the same alias set as something that ends up being modified. 377 if (AA->pointsToConstantMemory(LI->getOperand(0))) 378 return true; 379 380 // Don't hoist loads which have may-aliased stores in loop. 381 unsigned Size = 0; 382 if (LI->getType()->isSized()) 383 Size = AA->getTypeStoreSize(LI->getType()); 384 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 385 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 386 if (isa<DbgStopPointInst>(CI)) { 387 // Don't hoist/sink dbgstoppoints, we handle them separately 388 return false; 389 } 390 // Handle obvious cases efficiently. 391 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 392 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 393 return true; 394 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 395 // If this call only reads from memory and there are no writes to memory 396 // in the loop, we can hoist or sink the call as appropriate. 397 bool FoundMod = false; 398 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 399 I != E; ++I) { 400 AliasSet &AS = *I; 401 if (!AS.isForwardingAliasSet() && AS.isMod()) { 402 FoundMod = true; 403 break; 404 } 405 } 406 if (!FoundMod) return true; 407 } 408 409 // FIXME: This should use mod/ref information to see if we can hoist or sink 410 // the call. 411 412 return false; 413 } 414 415 // Otherwise these instructions are hoistable/sinkable 416 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 417 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 418 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 419 isa<ShuffleVectorInst>(I); 420} 421 422/// isNotUsedInLoop - Return true if the only users of this instruction are 423/// outside of the loop. If this is true, we can sink the instruction to the 424/// exit blocks of the loop. 425/// 426bool LICM::isNotUsedInLoop(Instruction &I) { 427 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 428 Instruction *User = cast<Instruction>(*UI); 429 if (PHINode *PN = dyn_cast<PHINode>(User)) { 430 // PHI node uses occur in predecessor blocks! 431 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 432 if (PN->getIncomingValue(i) == &I) 433 if (CurLoop->contains(PN->getIncomingBlock(i))) 434 return false; 435 } else if (CurLoop->contains(User->getParent())) { 436 return false; 437 } 438 } 439 return true; 440} 441 442 443/// isLoopInvariantInst - Return true if all operands of this instruction are 444/// loop invariant. We also filter out non-hoistable instructions here just for 445/// efficiency. 446/// 447bool LICM::isLoopInvariantInst(Instruction &I) { 448 // The instruction is loop invariant if all of its operands are loop-invariant 449 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 450 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 451 return false; 452 453 // If we got this far, the instruction is loop invariant! 454 return true; 455} 456 457/// sink - When an instruction is found to only be used outside of the loop, 458/// this function moves it to the exit blocks and patches up SSA form as needed. 459/// This method is guaranteed to remove the original instruction from its 460/// position, and may either delete it or move it to outside of the loop. 461/// 462void LICM::sink(Instruction &I) { 463 DEBUG(errs() << "LICM sinking instruction: " << I); 464 465 SmallVector<BasicBlock*, 8> ExitBlocks; 466 CurLoop->getExitBlocks(ExitBlocks); 467 468 if (isa<LoadInst>(I)) ++NumMovedLoads; 469 else if (isa<CallInst>(I)) ++NumMovedCalls; 470 ++NumSunk; 471 Changed = true; 472 473 // The case where there is only a single exit node of this loop is common 474 // enough that we handle it as a special (more efficient) case. It is more 475 // efficient to handle because there are no PHI nodes that need to be placed. 476 if (ExitBlocks.size() == 1) { 477 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 478 // Instruction is not used, just delete it. 479 CurAST->deleteValue(&I); 480 // If I has users in unreachable blocks, eliminate. 481 // If I is not void type then replaceAllUsesWith undef. 482 // This allows ValueHandlers and custom metadata to adjust itself. 483 if (!I.getType()->isVoidTy()) 484 I.replaceAllUsesWith(UndefValue::get(I.getType())); 485 I.eraseFromParent(); 486 } else { 487 // Move the instruction to the start of the exit block, after any PHI 488 // nodes in it. 489 I.removeFromParent(); 490 BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI(); 491 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 492 } 493 } else if (ExitBlocks.empty()) { 494 // The instruction is actually dead if there ARE NO exit blocks. 495 CurAST->deleteValue(&I); 496 // If I has users in unreachable blocks, eliminate. 497 // If I is not void type then replaceAllUsesWith undef. 498 // This allows ValueHandlers and custom metadata to adjust itself. 499 if (!I.getType()->isVoidTy()) 500 I.replaceAllUsesWith(UndefValue::get(I.getType())); 501 I.eraseFromParent(); 502 } else { 503 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 504 // do all of the hard work of inserting PHI nodes as necessary. We convert 505 // the value into a stack object to get it to do this. 506 507 // Firstly, we create a stack object to hold the value... 508 AllocaInst *AI = 0; 509 510 if (!I.getType()->isVoidTy()) { 511 AI = new AllocaInst(I.getType(), 0, I.getName(), 512 I.getParent()->getParent()->getEntryBlock().begin()); 513 CurAST->add(AI); 514 } 515 516 // Secondly, insert load instructions for each use of the instruction 517 // outside of the loop. 518 while (!I.use_empty()) { 519 Instruction *U = cast<Instruction>(I.use_back()); 520 521 // If the user is a PHI Node, we actually have to insert load instructions 522 // in all predecessor blocks, not in the PHI block itself! 523 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 524 // Only insert into each predecessor once, so that we don't have 525 // different incoming values from the same block! 526 std::map<BasicBlock*, Value*> InsertedBlocks; 527 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 528 if (UPN->getIncomingValue(i) == &I) { 529 BasicBlock *Pred = UPN->getIncomingBlock(i); 530 Value *&PredVal = InsertedBlocks[Pred]; 531 if (!PredVal) { 532 // Insert a new load instruction right before the terminator in 533 // the predecessor block. 534 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 535 CurAST->add(cast<LoadInst>(PredVal)); 536 } 537 538 UPN->setIncomingValue(i, PredVal); 539 } 540 541 } else { 542 LoadInst *L = new LoadInst(AI, "", U); 543 U->replaceUsesOfWith(&I, L); 544 CurAST->add(L); 545 } 546 } 547 548 // Thirdly, insert a copy of the instruction in each exit block of the loop 549 // that is dominated by the instruction, storing the result into the memory 550 // location. Be careful not to insert the instruction into any particular 551 // basic block more than once. 552 std::set<BasicBlock*> InsertedBlocks; 553 BasicBlock *InstOrigBB = I.getParent(); 554 555 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 556 BasicBlock *ExitBlock = ExitBlocks[i]; 557 558 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 559 // If we haven't already processed this exit block, do so now. 560 if (InsertedBlocks.insert(ExitBlock).second) { 561 // Insert the code after the last PHI node... 562 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 563 564 // If this is the first exit block processed, just move the original 565 // instruction, otherwise clone the original instruction and insert 566 // the copy. 567 Instruction *New; 568 if (InsertedBlocks.size() == 1) { 569 I.removeFromParent(); 570 ExitBlock->getInstList().insert(InsertPt, &I); 571 New = &I; 572 } else { 573 New = I.clone(); 574 CurAST->copyValue(&I, New); 575 if (!I.getName().empty()) 576 New->setName(I.getName()+".le"); 577 ExitBlock->getInstList().insert(InsertPt, New); 578 } 579 580 // Now that we have inserted the instruction, store it into the alloca 581 if (AI) new StoreInst(New, AI, InsertPt); 582 } 583 } 584 } 585 586 // If the instruction doesn't dominate any exit blocks, it must be dead. 587 if (InsertedBlocks.empty()) { 588 CurAST->deleteValue(&I); 589 I.eraseFromParent(); 590 } 591 592 // Finally, promote the fine value to SSA form. 593 if (AI) { 594 std::vector<AllocaInst*> Allocas; 595 Allocas.push_back(AI); 596 PromoteMemToReg(Allocas, *DT, *DF, CurAST); 597 } 598 } 599} 600 601/// hoist - When an instruction is found to only use loop invariant operands 602/// that is safe to hoist, this instruction is called to do the dirty work. 603/// 604void LICM::hoist(Instruction &I) { 605 DEBUG(errs() << "LICM hoisting to " << Preheader->getName() << ": " 606 << I << "\n"); 607 608 // Remove the instruction from its current basic block... but don't delete the 609 // instruction. 610 I.removeFromParent(); 611 612 // Insert the new node in Preheader, before the terminator. 613 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 614 615 if (isa<LoadInst>(I)) ++NumMovedLoads; 616 else if (isa<CallInst>(I)) ++NumMovedCalls; 617 ++NumHoisted; 618 Changed = true; 619} 620 621/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 622/// not a trapping instruction or if it is a trapping instruction and is 623/// guaranteed to execute. 624/// 625bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 626 // If it is not a trapping instruction, it is always safe to hoist. 627 if (Inst.isSafeToSpeculativelyExecute()) 628 return true; 629 630 // Otherwise we have to check to make sure that the instruction dominates all 631 // of the exit blocks. If it doesn't, then there is a path out of the loop 632 // which does not execute this instruction, so we can't hoist it. 633 634 // If the instruction is in the header block for the loop (which is very 635 // common), it is always guaranteed to dominate the exit blocks. Since this 636 // is a common case, and can save some work, check it now. 637 if (Inst.getParent() == CurLoop->getHeader()) 638 return true; 639 640 // Get the exit blocks for the current loop. 641 SmallVector<BasicBlock*, 8> ExitBlocks; 642 CurLoop->getExitBlocks(ExitBlocks); 643 644 // For each exit block, get the DT node and walk up the DT until the 645 // instruction's basic block is found or we exit the loop. 646 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 647 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 648 return false; 649 650 return true; 651} 652 653 654/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 655/// stores out of the loop and moving loads to before the loop. We do this by 656/// looping over the stores in the loop, looking for stores to Must pointers 657/// which are loop invariant. We promote these memory locations to use allocas 658/// instead. These allocas can easily be raised to register values by the 659/// PromoteMem2Reg functionality. 660/// 661void LICM::PromoteValuesInLoop() { 662 // PromotedValues - List of values that are promoted out of the loop. Each 663 // value has an alloca instruction for it, and a canonical version of the 664 // pointer. 665 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 666 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 667 668 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 669 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 670 671 Changed = true; 672 NumPromoted += PromotedValues.size(); 673 674 std::vector<Value*> PointerValueNumbers; 675 676 // Emit a copy from the value into the alloca'd value in the loop preheader 677 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 678 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 679 Value *Ptr = PromotedValues[i].second; 680 681 // If we are promoting a pointer value, update alias information for the 682 // inserted load. 683 Value *LoadValue = 0; 684 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 685 // Locate a load or store through the pointer, and assign the same value 686 // to LI as we are loading or storing. Since we know that the value is 687 // stored in this loop, this will always succeed. 688 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 689 UI != E; ++UI) 690 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 691 LoadValue = LI; 692 break; 693 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 694 if (SI->getOperand(1) == Ptr) { 695 LoadValue = SI->getOperand(0); 696 break; 697 } 698 } 699 assert(LoadValue && "No store through the pointer found!"); 700 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 701 } 702 703 // Load from the memory we are promoting. 704 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 705 706 if (LoadValue) CurAST->copyValue(LoadValue, LI); 707 708 // Store into the temporary alloca. 709 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 710 } 711 712 // Scan the basic blocks in the loop, replacing uses of our pointers with 713 // uses of the allocas in question. 714 // 715 for (Loop::block_iterator I = CurLoop->block_begin(), 716 E = CurLoop->block_end(); I != E; ++I) { 717 BasicBlock *BB = *I; 718 // Rewrite all loads and stores in the block of the pointer... 719 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 720 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 721 std::map<Value*, AllocaInst*>::iterator 722 I = ValueToAllocaMap.find(L->getOperand(0)); 723 if (I != ValueToAllocaMap.end()) 724 L->setOperand(0, I->second); // Rewrite load instruction... 725 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 726 std::map<Value*, AllocaInst*>::iterator 727 I = ValueToAllocaMap.find(S->getOperand(1)); 728 if (I != ValueToAllocaMap.end()) 729 S->setOperand(1, I->second); // Rewrite store instruction... 730 } 731 } 732 } 733 734 // Now that the body of the loop uses the allocas instead of the original 735 // memory locations, insert code to copy the alloca value back into the 736 // original memory location on all exits from the loop. Note that we only 737 // want to insert one copy of the code in each exit block, though the loop may 738 // exit to the same block more than once. 739 // 740 SmallPtrSet<BasicBlock*, 16> ProcessedBlocks; 741 742 SmallVector<BasicBlock*, 8> ExitBlocks; 743 CurLoop->getExitBlocks(ExitBlocks); 744 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 745 if (!ProcessedBlocks.insert(ExitBlocks[i])) 746 continue; 747 748 // Copy all of the allocas into their memory locations. 749 BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI(); 750 Instruction *InsertPos = BI; 751 unsigned PVN = 0; 752 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 753 // Load from the alloca. 754 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 755 756 // If this is a pointer type, update alias info appropriately. 757 if (isa<PointerType>(LI->getType())) 758 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 759 760 // Store into the memory we promoted. 761 new StoreInst(LI, PromotedValues[i].second, InsertPos); 762 } 763 } 764 765 // Now that we have done the deed, use the mem2reg functionality to promote 766 // all of the new allocas we just created into real SSA registers. 767 // 768 std::vector<AllocaInst*> PromotedAllocas; 769 PromotedAllocas.reserve(PromotedValues.size()); 770 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 771 PromotedAllocas.push_back(PromotedValues[i].first); 772 PromoteMemToReg(PromotedAllocas, *DT, *DF, CurAST); 773} 774 775/// FindPromotableValuesInLoop - Check the current loop for stores to definite 776/// pointers, which are not loaded and stored through may aliases and are safe 777/// for promotion. If these are found, create an alloca for the value, add it 778/// to the PromotedValues list, and keep track of the mapping from value to 779/// alloca. 780void LICM::FindPromotableValuesInLoop( 781 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 782 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 783 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 784 785 // Loop over all of the alias sets in the tracker object. 786 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 787 I != E; ++I) { 788 AliasSet &AS = *I; 789 // We can promote this alias set if it has a store, if it is a "Must" alias 790 // set, if the pointer is loop invariant, and if we are not eliminating any 791 // volatile loads or stores. 792 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 793 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 794 continue; 795 796 assert(!AS.empty() && 797 "Must alias set should have at least one pointer element in it!"); 798 Value *V = AS.begin()->getValue(); 799 800 // Check that all of the pointers in the alias set have the same type. We 801 // cannot (yet) promote a memory location that is loaded and stored in 802 // different sizes. 803 { 804 bool PointerOk = true; 805 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 806 if (V->getType() != I->getValue()->getType()) { 807 PointerOk = false; 808 break; 809 } 810 if (!PointerOk) 811 continue; 812 } 813 814 // It isn't safe to promote a load/store from the loop if the load/store is 815 // conditional. For example, turning: 816 // 817 // for () { if (c) *P += 1; } 818 // 819 // into: 820 // 821 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 822 // 823 // is not safe, because *P may only be valid to access if 'c' is true. 824 // 825 // It is safe to promote P if all uses are direct load/stores and if at 826 // least one is guaranteed to be executed. 827 bool GuaranteedToExecute = false; 828 bool InvalidInst = false; 829 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 830 UI != UE; ++UI) { 831 // Ignore instructions not in this loop. 832 Instruction *Use = dyn_cast<Instruction>(*UI); 833 if (!Use || !CurLoop->contains(Use->getParent())) 834 continue; 835 836 if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) { 837 InvalidInst = true; 838 break; 839 } 840 841 if (!GuaranteedToExecute) 842 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 843 } 844 845 // If there is an non-load/store instruction in the loop, we can't promote 846 // it. If there isn't a guaranteed-to-execute instruction, we can't 847 // promote. 848 if (InvalidInst || !GuaranteedToExecute) 849 continue; 850 851 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 852 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 853 PromotedValues.push_back(std::make_pair(AI, V)); 854 855 // Update the AST and alias analysis. 856 CurAST->copyValue(V, AI); 857 858 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 859 ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI)); 860 861 DEBUG(errs() << "LICM: Promoting value: " << *V << "\n"); 862 } 863} 864 865/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 866void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 867 AliasSetTracker *AST = LoopToAliasMap[L]; 868 if (!AST) 869 return; 870 871 AST->copyValue(From, To); 872} 873 874/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 875/// set. 876void LICM::deleteAnalysisValue(Value *V, Loop *L) { 877 AliasSetTracker *AST = LoopToAliasMap[L]; 878 if (!AST) 879 return; 880 881 AST->deleteValue(V); 882} 883