LICM.cpp revision c955d212a8c060d7082fe975fd65fcbc9263853a
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 SSAUpdater to construct the appropriate SSA form for the value. 30// 31//===----------------------------------------------------------------------===// 32 33#define DEBUG_TYPE "licm" 34#include "llvm/Transforms/Scalar.h" 35#include "llvm/Constants.h" 36#include "llvm/DerivedTypes.h" 37#include "llvm/IntrinsicInst.h" 38#include "llvm/Instructions.h" 39#include "llvm/LLVMContext.h" 40#include "llvm/Analysis/AliasAnalysis.h" 41#include "llvm/Analysis/AliasSetTracker.h" 42#include "llvm/Analysis/ConstantFolding.h" 43#include "llvm/Analysis/LoopInfo.h" 44#include "llvm/Analysis/LoopPass.h" 45#include "llvm/Analysis/Dominators.h" 46#include "llvm/Transforms/Utils/Local.h" 47#include "llvm/Transforms/Utils/SSAUpdater.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 initializeLICMPass(*PassRegistry::getPassRegistry()); 71 } 72 73 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 74 75 /// This transformation requires natural loop information & requires that 76 /// loop preheaders be inserted into the CFG... 77 /// 78 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 79 AU.setPreservesCFG(); 80 AU.addRequired<DominatorTree>(); 81 AU.addRequired<LoopInfo>(); 82 AU.addRequiredID(LoopSimplifyID); 83 AU.addRequired<AliasAnalysis>(); 84 AU.addPreserved<AliasAnalysis>(); 85 AU.addPreserved("scalar-evolution"); 86 AU.addPreservedID(LoopSimplifyID); 87 } 88 89 bool doFinalization() { 90 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets"); 91 return false; 92 } 93 94 private: 95 AliasAnalysis *AA; // Current AliasAnalysis information 96 LoopInfo *LI; // Current LoopInfo 97 DominatorTree *DT; // Dominator Tree for the current Loop. 98 99 // State that is updated as we process loops. 100 bool Changed; // Set to true when we change anything. 101 BasicBlock *Preheader; // The preheader block of the current loop... 102 Loop *CurLoop; // The current loop we are working on... 103 AliasSetTracker *CurAST; // AliasSet information for the current loop... 104 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap; 105 106 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 107 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 108 109 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 110 /// set. 111 void deleteAnalysisValue(Value *V, Loop *L); 112 113 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 114 /// dominated by the specified block, and that are in the current loop) in 115 /// reverse depth first order w.r.t the DominatorTree. This allows us to 116 /// visit uses before definitions, allowing us to sink a loop body in one 117 /// pass without iteration. 118 /// 119 void SinkRegion(DomTreeNode *N); 120 121 /// HoistRegion - Walk the specified region of the CFG (defined by all 122 /// blocks dominated by the specified block, and that are in the current 123 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 124 /// visit definitions before uses, allowing us to hoist a loop body in one 125 /// pass without iteration. 126 /// 127 void HoistRegion(DomTreeNode *N); 128 129 /// inSubLoop - Little predicate that returns true if the specified basic 130 /// block is in a subloop of the current one, not the current one itself. 131 /// 132 bool inSubLoop(BasicBlock *BB) { 133 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 134 return LI->getLoopFor(BB) != CurLoop; 135 } 136 137 /// sink - When an instruction is found to only be used outside of the loop, 138 /// this function moves it to the exit blocks and patches up SSA form as 139 /// needed. 140 /// 141 void sink(Instruction &I); 142 143 /// hoist - When an instruction is found to only use loop invariant operands 144 /// that is safe to hoist, this instruction is called to do the dirty work. 145 /// 146 void hoist(Instruction &I); 147 148 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 149 /// is not a trapping instruction or if it is a trapping instruction and is 150 /// guaranteed to execute. 151 /// 152 bool isSafeToExecuteUnconditionally(Instruction &I); 153 154 /// pointerInvalidatedByLoop - Return true if the body of this loop may 155 /// store into the memory location pointed to by V. 156 /// 157 bool pointerInvalidatedByLoop(Value *V, uint64_t Size, 158 const MDNode *TBAAInfo) { 159 // Check to see if any of the basic blocks in CurLoop invalidate *V. 160 return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod(); 161 } 162 163 bool canSinkOrHoistInst(Instruction &I); 164 bool isNotUsedInLoop(Instruction &I); 165 166 void PromoteAliasSet(AliasSet &AS); 167 }; 168} 169 170char LICM::ID = 0; 171INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false) 172INITIALIZE_PASS_DEPENDENCY(DominatorTree) 173INITIALIZE_PASS_DEPENDENCY(LoopInfo) 174INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 175INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 176INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false) 177 178Pass *llvm::createLICMPass() { return new LICM(); } 179 180/// Hoist expressions out of the specified loop. Note, alias info for inner 181/// loop is not preserved so it is not a good idea to run LICM multiple 182/// times on one loop. 183/// 184bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 185 Changed = false; 186 187 // Get our Loop and Alias Analysis information... 188 LI = &getAnalysis<LoopInfo>(); 189 AA = &getAnalysis<AliasAnalysis>(); 190 DT = &getAnalysis<DominatorTree>(); 191 192 CurAST = new AliasSetTracker(*AA); 193 // Collect Alias info from subloops. 194 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 195 LoopItr != LoopItrE; ++LoopItr) { 196 Loop *InnerL = *LoopItr; 197 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL]; 198 assert(InnerAST && "Where is my AST?"); 199 200 // What if InnerLoop was modified by other passes ? 201 CurAST->add(*InnerAST); 202 203 // Once we've incorporated the inner loop's AST into ours, we don't need the 204 // subloop's anymore. 205 delete InnerAST; 206 LoopToAliasSetMap.erase(InnerL); 207 } 208 209 CurLoop = L; 210 211 // Get the preheader block to move instructions into... 212 Preheader = L->getLoopPreheader(); 213 214 // Loop over the body of this loop, looking for calls, invokes, and stores. 215 // Because subloops have already been incorporated into AST, we skip blocks in 216 // subloops. 217 // 218 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 219 I != E; ++I) { 220 BasicBlock *BB = *I; 221 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops. 222 CurAST->add(*BB); // Incorporate the specified basic block 223 } 224 225 // We want to visit all of the instructions in this loop... that are not parts 226 // of our subloops (they have already had their invariants hoisted out of 227 // their loop, into this loop, so there is no need to process the BODIES of 228 // the subloops). 229 // 230 // Traverse the body of the loop in depth first order on the dominator tree so 231 // that we are guaranteed to see definitions before we see uses. This allows 232 // us to sink instructions in one pass, without iteration. After sinking 233 // instructions, we perform another pass to hoist them out of the loop. 234 // 235 if (L->hasDedicatedExits()) 236 SinkRegion(DT->getNode(L->getHeader())); 237 if (Preheader) 238 HoistRegion(DT->getNode(L->getHeader())); 239 240 // Now that all loop invariants have been removed from the loop, promote any 241 // memory references to scalars that we can. 242 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) { 243 // Loop over all of the alias sets in the tracker object. 244 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 245 I != E; ++I) 246 PromoteAliasSet(*I); 247 } 248 249 // Clear out loops state information for the next iteration 250 CurLoop = 0; 251 Preheader = 0; 252 253 // If this loop is nested inside of another one, save the alias information 254 // for when we process the outer loop. 255 if (L->getParentLoop()) 256 LoopToAliasSetMap[L] = CurAST; 257 else 258 delete CurAST; 259 return Changed; 260} 261 262/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 263/// dominated by the specified block, and that are in the current loop) in 264/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 265/// uses before definitions, allowing us to sink a loop body in one pass without 266/// iteration. 267/// 268void LICM::SinkRegion(DomTreeNode *N) { 269 assert(N != 0 && "Null dominator tree node?"); 270 BasicBlock *BB = N->getBlock(); 271 272 // If this subregion is not in the top level loop at all, exit. 273 if (!CurLoop->contains(BB)) return; 274 275 // We are processing blocks in reverse dfo, so process children first. 276 const std::vector<DomTreeNode*> &Children = N->getChildren(); 277 for (unsigned i = 0, e = Children.size(); i != e; ++i) 278 SinkRegion(Children[i]); 279 280 // Only need to process the contents of this block if it is not part of a 281 // subloop (which would already have been processed). 282 if (inSubLoop(BB)) return; 283 284 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 285 Instruction &I = *--II; 286 287 // If the instruction is dead, we would try to sink it because it isn't used 288 // in the loop, instead, just delete it. 289 if (isInstructionTriviallyDead(&I)) { 290 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); 291 ++II; 292 CurAST->deleteValue(&I); 293 I.eraseFromParent(); 294 Changed = true; 295 continue; 296 } 297 298 // Check to see if we can sink this instruction to the exit blocks 299 // of the loop. We can do this if the all users of the instruction are 300 // outside of the loop. In this case, it doesn't even matter if the 301 // operands of the instruction are loop invariant. 302 // 303 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 304 ++II; 305 sink(I); 306 } 307 } 308} 309 310/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 311/// dominated by the specified block, and that are in the current loop) in depth 312/// first order w.r.t the DominatorTree. This allows us to visit definitions 313/// before uses, allowing us to hoist a loop body in one pass without iteration. 314/// 315void LICM::HoistRegion(DomTreeNode *N) { 316 assert(N != 0 && "Null dominator tree node?"); 317 BasicBlock *BB = N->getBlock(); 318 319 // If this subregion is not in the top level loop at all, exit. 320 if (!CurLoop->contains(BB)) return; 321 322 // Only need to process the contents of this block if it is not part of a 323 // subloop (which would already have been processed). 324 if (!inSubLoop(BB)) 325 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 326 Instruction &I = *II++; 327 328 // Try constant folding this instruction. If all the operands are 329 // constants, it is technically hoistable, but it would be better to just 330 // fold it. 331 if (Constant *C = ConstantFoldInstruction(&I)) { 332 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'); 333 CurAST->copyValue(&I, C); 334 CurAST->deleteValue(&I); 335 I.replaceAllUsesWith(C); 336 I.eraseFromParent(); 337 continue; 338 } 339 340 // Try hoisting the instruction out to the preheader. We can only do this 341 // if all of the operands of the instruction are loop invariant and if it 342 // is safe to hoist the instruction. 343 // 344 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) && 345 isSafeToExecuteUnconditionally(I)) 346 hoist(I); 347 } 348 349 const std::vector<DomTreeNode*> &Children = N->getChildren(); 350 for (unsigned i = 0, e = Children.size(); i != e; ++i) 351 HoistRegion(Children[i]); 352} 353 354/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 355/// instruction. 356/// 357bool LICM::canSinkOrHoistInst(Instruction &I) { 358 // Loads have extra constraints we have to verify before we can hoist them. 359 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 360 if (LI->isVolatile()) 361 return false; // Don't hoist volatile loads! 362 363 // Loads from constant memory are always safe to move, even if they end up 364 // in the same alias set as something that ends up being modified. 365 if (AA->pointsToConstantMemory(LI->getOperand(0))) 366 return true; 367 368 // Don't hoist loads which have may-aliased stores in loop. 369 uint64_t Size = 0; 370 if (LI->getType()->isSized()) 371 Size = AA->getTypeStoreSize(LI->getType()); 372 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, 373 LI->getMetadata(LLVMContext::MD_tbaa)); 374 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 375 // Handle obvious cases efficiently. 376 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 377 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 378 return true; 379 if (AliasAnalysis::onlyReadsMemory(Behavior)) { 380 // If this call only reads from memory and there are no writes to memory 381 // in the loop, we can hoist or sink the call as appropriate. 382 bool FoundMod = false; 383 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 384 I != E; ++I) { 385 AliasSet &AS = *I; 386 if (!AS.isForwardingAliasSet() && AS.isMod()) { 387 FoundMod = true; 388 break; 389 } 390 } 391 if (!FoundMod) return true; 392 } 393 394 // FIXME: This should use mod/ref information to see if we can hoist or sink 395 // the call. 396 397 return false; 398 } 399 400 // Otherwise these instructions are hoistable/sinkable 401 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 402 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 403 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 404 isa<ShuffleVectorInst>(I); 405} 406 407/// isNotUsedInLoop - Return true if the only users of this instruction are 408/// outside of the loop. If this is true, we can sink the instruction to the 409/// exit blocks of the loop. 410/// 411bool LICM::isNotUsedInLoop(Instruction &I) { 412 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 413 Instruction *User = cast<Instruction>(*UI); 414 if (PHINode *PN = dyn_cast<PHINode>(User)) { 415 // PHI node uses occur in predecessor blocks! 416 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 417 if (PN->getIncomingValue(i) == &I) 418 if (CurLoop->contains(PN->getIncomingBlock(i))) 419 return false; 420 } else if (CurLoop->contains(User)) { 421 return false; 422 } 423 } 424 return true; 425} 426 427 428/// sink - When an instruction is found to only be used outside of the loop, 429/// this function moves it to the exit blocks and patches up SSA form as needed. 430/// This method is guaranteed to remove the original instruction from its 431/// position, and may either delete it or move it to outside of the loop. 432/// 433void LICM::sink(Instruction &I) { 434 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 435 436 SmallVector<BasicBlock*, 8> ExitBlocks; 437 CurLoop->getUniqueExitBlocks(ExitBlocks); 438 439 if (isa<LoadInst>(I)) ++NumMovedLoads; 440 else if (isa<CallInst>(I)) ++NumMovedCalls; 441 ++NumSunk; 442 Changed = true; 443 444 // The case where there is only a single exit node of this loop is common 445 // enough that we handle it as a special (more efficient) case. It is more 446 // efficient to handle because there are no PHI nodes that need to be placed. 447 if (ExitBlocks.size() == 1) { 448 if (!isa<DbgInfoIntrinsic>(I) && 449 !DT->dominates(I.getParent(), ExitBlocks[0])) { 450 // Instruction is not used, just delete it. 451 CurAST->deleteValue(&I); 452 // If I has users in unreachable blocks, eliminate. 453 // If I is not void type then replaceAllUsesWith undef. 454 // This allows ValueHandlers and custom metadata to adjust itself. 455 if (!I.use_empty()) 456 I.replaceAllUsesWith(UndefValue::get(I.getType())); 457 I.eraseFromParent(); 458 } else { 459 // Move the instruction to the start of the exit block, after any PHI 460 // nodes in it. 461 I.moveBefore(ExitBlocks[0]->getFirstNonPHI()); 462 463 // This instruction is no longer in the AST for the current loop, because 464 // we just sunk it out of the loop. If we just sunk it into an outer 465 // loop, we will rediscover the operation when we process it. 466 CurAST->deleteValue(&I); 467 } 468 return; 469 } 470 471 if (ExitBlocks.empty()) { 472 // The instruction is actually dead if there ARE NO exit blocks. 473 CurAST->deleteValue(&I); 474 // If I has users in unreachable blocks, eliminate. 475 // If I is not void type then replaceAllUsesWith undef. 476 // This allows ValueHandlers and custom metadata to adjust itself. 477 if (!I.use_empty()) 478 I.replaceAllUsesWith(UndefValue::get(I.getType())); 479 I.eraseFromParent(); 480 return; 481 } 482 483 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the 484 // hard work of inserting PHI nodes as necessary. 485 SmallVector<PHINode*, 8> NewPHIs; 486 SSAUpdater SSA(&NewPHIs); 487 488 if (!I.use_empty()) 489 SSA.Initialize(I.getType(), I.getName()); 490 491 // Insert a copy of the instruction in each exit block of the loop that is 492 // dominated by the instruction. Each exit block is known to only be in the 493 // ExitBlocks list once. 494 BasicBlock *InstOrigBB = I.getParent(); 495 unsigned NumInserted = 0; 496 497 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 498 BasicBlock *ExitBlock = ExitBlocks[i]; 499 500 if (!DT->dominates(InstOrigBB, ExitBlock)) 501 continue; 502 503 // Insert the code after the last PHI node. 504 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 505 506 // If this is the first exit block processed, just move the original 507 // instruction, otherwise clone the original instruction and insert 508 // the copy. 509 Instruction *New; 510 if (NumInserted++ == 0) { 511 I.moveBefore(InsertPt); 512 New = &I; 513 } else { 514 New = I.clone(); 515 if (!I.getName().empty()) 516 New->setName(I.getName()+".le"); 517 ExitBlock->getInstList().insert(InsertPt, New); 518 } 519 520 // Now that we have inserted the instruction, inform SSAUpdater. 521 if (!I.use_empty()) 522 SSA.AddAvailableValue(ExitBlock, New); 523 } 524 525 // If the instruction doesn't dominate any exit blocks, it must be dead. 526 if (NumInserted == 0) { 527 CurAST->deleteValue(&I); 528 if (!I.use_empty()) 529 I.replaceAllUsesWith(UndefValue::get(I.getType())); 530 I.eraseFromParent(); 531 return; 532 } 533 534 // Next, rewrite uses of the instruction, inserting PHI nodes as needed. 535 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) { 536 // Grab the use before incrementing the iterator. 537 Use &U = UI.getUse(); 538 // Increment the iterator before removing the use from the list. 539 ++UI; 540 SSA.RewriteUseAfterInsertions(U); 541 } 542 543 // Update CurAST for NewPHIs if I had pointer type. 544 if (I.getType()->isPointerTy()) 545 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 546 CurAST->copyValue(&I, NewPHIs[i]); 547 548 // Finally, remove the instruction from CurAST. It is no longer in the loop. 549 CurAST->deleteValue(&I); 550} 551 552/// hoist - When an instruction is found to only use loop invariant operands 553/// that is safe to hoist, this instruction is called to do the dirty work. 554/// 555void LICM::hoist(Instruction &I) { 556 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 557 << I << "\n"); 558 559 // Move the new node to the Preheader, before its terminator. 560 I.moveBefore(Preheader->getTerminator()); 561 562 if (isa<LoadInst>(I)) ++NumMovedLoads; 563 else if (isa<CallInst>(I)) ++NumMovedCalls; 564 ++NumHoisted; 565 Changed = true; 566} 567 568/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 569/// not a trapping instruction or if it is a trapping instruction and is 570/// guaranteed to execute. 571/// 572bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 573 // If it is not a trapping instruction, it is always safe to hoist. 574 if (Inst.isSafeToSpeculativelyExecute()) 575 return true; 576 577 // Otherwise we have to check to make sure that the instruction dominates all 578 // of the exit blocks. If it doesn't, then there is a path out of the loop 579 // which does not execute this instruction, so we can't hoist it. 580 581 // If the instruction is in the header block for the loop (which is very 582 // common), it is always guaranteed to dominate the exit blocks. Since this 583 // is a common case, and can save some work, check it now. 584 if (Inst.getParent() == CurLoop->getHeader()) 585 return true; 586 587 // Get the exit blocks for the current loop. 588 SmallVector<BasicBlock*, 8> ExitBlocks; 589 CurLoop->getExitBlocks(ExitBlocks); 590 591 // Verify that the block dominates each of the exit blocks of the loop. 592 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 593 if (!DT->dominates(Inst.getParent(), ExitBlocks[i])) 594 return false; 595 596 return true; 597} 598 599namespace { 600 class LoopPromoter : public LoadAndStorePromoter { 601 Value *SomePtr; // Designated pointer to store to. 602 SmallPtrSet<Value*, 4> &PointerMustAliases; 603 SmallVectorImpl<BasicBlock*> &LoopExitBlocks; 604 AliasSetTracker &AST; 605 public: 606 LoopPromoter(Value *SP, 607 const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 608 SmallPtrSet<Value*, 4> &PMA, 609 SmallVectorImpl<BasicBlock*> &LEB, AliasSetTracker &ast) 610 : LoadAndStorePromoter(Insts, S, 0, 0), SomePtr(SP), 611 PointerMustAliases(PMA), LoopExitBlocks(LEB), AST(ast) {} 612 613 virtual bool isInstInList(Instruction *I, 614 const SmallVectorImpl<Instruction*> &) const { 615 Value *Ptr; 616 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 617 Ptr = LI->getOperand(0); 618 else 619 Ptr = cast<StoreInst>(I)->getPointerOperand(); 620 return PointerMustAliases.count(Ptr); 621 } 622 623 virtual void doExtraRewritesBeforeFinalDeletion() const { 624 // Insert stores after in the loop exit blocks. Each exit block gets a 625 // store of the live-out values that feed them. Since we've already told 626 // the SSA updater about the defs in the loop and the preheader 627 // definition, it is all set and we can start using it. 628 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { 629 BasicBlock *ExitBlock = LoopExitBlocks[i]; 630 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); 631 Instruction *InsertPos = ExitBlock->getFirstNonPHI(); 632 new StoreInst(LiveInValue, SomePtr, InsertPos); 633 } 634 } 635 636 virtual void replaceLoadWithValue(LoadInst *LI, Value *V) const { 637 // Update alias analysis. 638 AST.copyValue(LI, V); 639 } 640 virtual void instructionDeleted(Instruction *I) const { 641 AST.deleteValue(I); 642 } 643 }; 644} // end anon namespace 645 646/// PromoteAliasSet - Try to promote memory values to scalars by sinking 647/// stores out of the loop and moving loads to before the loop. We do this by 648/// looping over the stores in the loop, looking for stores to Must pointers 649/// which are loop invariant. 650/// 651void LICM::PromoteAliasSet(AliasSet &AS) { 652 // We can promote this alias set if it has a store, if it is a "Must" alias 653 // set, if the pointer is loop invariant, and if we are not eliminating any 654 // volatile loads or stores. 655 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 656 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 657 return; 658 659 assert(!AS.empty() && 660 "Must alias set should have at least one pointer element in it!"); 661 Value *SomePtr = AS.begin()->getValue(); 662 663 // It isn't safe to promote a load/store from the loop if the load/store is 664 // conditional. For example, turning: 665 // 666 // for () { if (c) *P += 1; } 667 // 668 // into: 669 // 670 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 671 // 672 // is not safe, because *P may only be valid to access if 'c' is true. 673 // 674 // It is safe to promote P if all uses are direct load/stores and if at 675 // least one is guaranteed to be executed. 676 bool GuaranteedToExecute = false; 677 678 SmallVector<Instruction*, 64> LoopUses; 679 SmallPtrSet<Value*, 4> PointerMustAliases; 680 681 // Check that all of the pointers in the alias set have the same type. We 682 // cannot (yet) promote a memory location that is loaded and stored in 683 // different sizes. 684 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) { 685 Value *ASIV = ASI->getValue(); 686 PointerMustAliases.insert(ASIV); 687 688 // Check that all of the pointers in the alias set have the same type. We 689 // cannot (yet) promote a memory location that is loaded and stored in 690 // different sizes. 691 if (SomePtr->getType() != ASIV->getType()) 692 return; 693 694 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end(); 695 UI != UE; ++UI) { 696 // Ignore instructions that are outside the loop. 697 Instruction *Use = dyn_cast<Instruction>(*UI); 698 if (!Use || !CurLoop->contains(Use)) 699 continue; 700 701 // If there is an non-load/store instruction in the loop, we can't promote 702 // it. 703 if (isa<LoadInst>(Use)) 704 assert(!cast<LoadInst>(Use)->isVolatile() && "AST broken"); 705 else if (isa<StoreInst>(Use)) { 706 // Stores *of* the pointer are not interesting, only stores *to* the 707 // pointer. 708 if (Use->getOperand(1) != ASIV) 709 continue; 710 assert(!cast<StoreInst>(Use)->isVolatile() && "AST broken"); 711 } else 712 return; // Not a load or store. 713 714 if (!GuaranteedToExecute) 715 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 716 717 LoopUses.push_back(Use); 718 } 719 } 720 721 // If there isn't a guaranteed-to-execute instruction, we can't promote. 722 if (!GuaranteedToExecute) 723 return; 724 725 // Otherwise, this is safe to promote, lets do it! 726 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n'); 727 Changed = true; 728 ++NumPromoted; 729 730 SmallVector<BasicBlock*, 8> ExitBlocks; 731 CurLoop->getUniqueExitBlocks(ExitBlocks); 732 733 // We use the SSAUpdater interface to insert phi nodes as required. 734 SmallVector<PHINode*, 16> NewPHIs; 735 SSAUpdater SSA(&NewPHIs); 736 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, 737 *CurAST); 738 739 // Set up the preheader to have a definition of the value. It is the live-out 740 // value from the preheader that uses in the loop will use. 741 LoadInst *PreheaderLoad = 742 new LoadInst(SomePtr, SomePtr->getName()+".promoted", 743 Preheader->getTerminator()); 744 SSA.AddAvailableValue(Preheader, PreheaderLoad); 745 746 // Rewrite all the loads in the loop and remember all the definitions from 747 // stores in the loop. 748 Promoter.run(LoopUses); 749 750 // If the SSAUpdater didn't use the load in the preheader, just zap it now. 751 if (PreheaderLoad->use_empty()) 752 PreheaderLoad->eraseFromParent(); 753} 754 755 756/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 757void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 758 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 759 if (!AST) 760 return; 761 762 AST->copyValue(From, To); 763} 764 765/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 766/// set. 767void LICM::deleteAnalysisValue(Value *V, Loop *L) { 768 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 769 if (!AST) 770 return; 771 772 AST->deleteValue(V); 773} 774