LICM.cpp revision 30a121b4e83c2bed062cc2d89f4b8fc354ac370e
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 // Don't sink or hoist dbg info; it's legal, but not useful. 376 if (isa<DbgInfoIntrinsic>(I)) 377 return false; 378 379 // Handle simple cases by querying alias analysis. 380 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 381 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 382 return true; 383 if (AliasAnalysis::onlyReadsMemory(Behavior)) { 384 // If this call only reads from memory and there are no writes to memory 385 // in the loop, we can hoist or sink the call as appropriate. 386 bool FoundMod = false; 387 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 388 I != E; ++I) { 389 AliasSet &AS = *I; 390 if (!AS.isForwardingAliasSet() && AS.isMod()) { 391 FoundMod = true; 392 break; 393 } 394 } 395 if (!FoundMod) return true; 396 } 397 398 // FIXME: This should use mod/ref information to see if we can hoist or sink 399 // the call. 400 401 return false; 402 } 403 404 // Otherwise these instructions are hoistable/sinkable 405 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 406 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 407 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 408 isa<ShuffleVectorInst>(I); 409} 410 411/// isNotUsedInLoop - Return true if the only users of this instruction are 412/// outside of the loop. If this is true, we can sink the instruction to the 413/// exit blocks of the loop. 414/// 415bool LICM::isNotUsedInLoop(Instruction &I) { 416 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 417 Instruction *User = cast<Instruction>(*UI); 418 if (PHINode *PN = dyn_cast<PHINode>(User)) { 419 // PHI node uses occur in predecessor blocks! 420 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 421 if (PN->getIncomingValue(i) == &I) 422 if (CurLoop->contains(PN->getIncomingBlock(i))) 423 return false; 424 } else if (CurLoop->contains(User)) { 425 return false; 426 } 427 } 428 return true; 429} 430 431 432/// sink - When an instruction is found to only be used outside of the loop, 433/// this function moves it to the exit blocks and patches up SSA form as needed. 434/// This method is guaranteed to remove the original instruction from its 435/// position, and may either delete it or move it to outside of the loop. 436/// 437void LICM::sink(Instruction &I) { 438 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 439 440 SmallVector<BasicBlock*, 8> ExitBlocks; 441 CurLoop->getUniqueExitBlocks(ExitBlocks); 442 443 if (isa<LoadInst>(I)) ++NumMovedLoads; 444 else if (isa<CallInst>(I)) ++NumMovedCalls; 445 ++NumSunk; 446 Changed = true; 447 448 // The case where there is only a single exit node of this loop is common 449 // enough that we handle it as a special (more efficient) case. It is more 450 // efficient to handle because there are no PHI nodes that need to be placed. 451 if (ExitBlocks.size() == 1) { 452 if (!DT->dominates(I.getParent(), ExitBlocks[0])) { 453 // Instruction is not used, just delete it. 454 CurAST->deleteValue(&I); 455 // If I has users in unreachable blocks, eliminate. 456 // If I is not void type then replaceAllUsesWith undef. 457 // This allows ValueHandlers and custom metadata to adjust itself. 458 if (!I.use_empty()) 459 I.replaceAllUsesWith(UndefValue::get(I.getType())); 460 I.eraseFromParent(); 461 } else { 462 // Move the instruction to the start of the exit block, after any PHI 463 // nodes in it. 464 I.moveBefore(ExitBlocks[0]->getFirstNonPHI()); 465 466 // This instruction is no longer in the AST for the current loop, because 467 // we just sunk it out of the loop. If we just sunk it into an outer 468 // loop, we will rediscover the operation when we process it. 469 CurAST->deleteValue(&I); 470 } 471 return; 472 } 473 474 if (ExitBlocks.empty()) { 475 // The instruction is actually dead if there ARE NO exit blocks. 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.use_empty()) 481 I.replaceAllUsesWith(UndefValue::get(I.getType())); 482 I.eraseFromParent(); 483 return; 484 } 485 486 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the 487 // hard work of inserting PHI nodes as necessary. 488 SmallVector<PHINode*, 8> NewPHIs; 489 SSAUpdater SSA(&NewPHIs); 490 491 if (!I.use_empty()) 492 SSA.Initialize(I.getType(), I.getName()); 493 494 // Insert a copy of the instruction in each exit block of the loop that is 495 // dominated by the instruction. Each exit block is known to only be in the 496 // ExitBlocks list once. 497 BasicBlock *InstOrigBB = I.getParent(); 498 unsigned NumInserted = 0; 499 500 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 501 BasicBlock *ExitBlock = ExitBlocks[i]; 502 503 if (!DT->dominates(InstOrigBB, ExitBlock)) 504 continue; 505 506 // Insert the code after the last PHI node. 507 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 508 509 // If this is the first exit block processed, just move the original 510 // instruction, otherwise clone the original instruction and insert 511 // the copy. 512 Instruction *New; 513 if (NumInserted++ == 0) { 514 I.moveBefore(InsertPt); 515 New = &I; 516 } else { 517 New = I.clone(); 518 if (!I.getName().empty()) 519 New->setName(I.getName()+".le"); 520 ExitBlock->getInstList().insert(InsertPt, New); 521 } 522 523 // Now that we have inserted the instruction, inform SSAUpdater. 524 if (!I.use_empty()) 525 SSA.AddAvailableValue(ExitBlock, New); 526 } 527 528 // If the instruction doesn't dominate any exit blocks, it must be dead. 529 if (NumInserted == 0) { 530 CurAST->deleteValue(&I); 531 if (!I.use_empty()) 532 I.replaceAllUsesWith(UndefValue::get(I.getType())); 533 I.eraseFromParent(); 534 return; 535 } 536 537 // Next, rewrite uses of the instruction, inserting PHI nodes as needed. 538 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) { 539 // Grab the use before incrementing the iterator. 540 Use &U = UI.getUse(); 541 // Increment the iterator before removing the use from the list. 542 ++UI; 543 SSA.RewriteUseAfterInsertions(U); 544 } 545 546 // Update CurAST for NewPHIs if I had pointer type. 547 if (I.getType()->isPointerTy()) 548 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 549 CurAST->copyValue(&I, NewPHIs[i]); 550 551 // Finally, remove the instruction from CurAST. It is no longer in the loop. 552 CurAST->deleteValue(&I); 553} 554 555/// hoist - When an instruction is found to only use loop invariant operands 556/// that is safe to hoist, this instruction is called to do the dirty work. 557/// 558void LICM::hoist(Instruction &I) { 559 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 560 << I << "\n"); 561 562 // Move the new node to the Preheader, before its terminator. 563 I.moveBefore(Preheader->getTerminator()); 564 565 if (isa<LoadInst>(I)) ++NumMovedLoads; 566 else if (isa<CallInst>(I)) ++NumMovedCalls; 567 ++NumHoisted; 568 Changed = true; 569} 570 571/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 572/// not a trapping instruction or if it is a trapping instruction and is 573/// guaranteed to execute. 574/// 575bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 576 // If it is not a trapping instruction, it is always safe to hoist. 577 if (Inst.isSafeToSpeculativelyExecute()) 578 return true; 579 580 // Otherwise we have to check to make sure that the instruction dominates all 581 // of the exit blocks. If it doesn't, then there is a path out of the loop 582 // which does not execute this instruction, so we can't hoist it. 583 584 // If the instruction is in the header block for the loop (which is very 585 // common), it is always guaranteed to dominate the exit blocks. Since this 586 // is a common case, and can save some work, check it now. 587 if (Inst.getParent() == CurLoop->getHeader()) 588 return true; 589 590 // Get the exit blocks for the current loop. 591 SmallVector<BasicBlock*, 8> ExitBlocks; 592 CurLoop->getExitBlocks(ExitBlocks); 593 594 // Verify that the block dominates each of the exit blocks of the loop. 595 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 596 if (!DT->dominates(Inst.getParent(), ExitBlocks[i])) 597 return false; 598 599 return true; 600} 601 602namespace { 603 class LoopPromoter : public LoadAndStorePromoter { 604 Value *SomePtr; // Designated pointer to store to. 605 SmallPtrSet<Value*, 4> &PointerMustAliases; 606 SmallVectorImpl<BasicBlock*> &LoopExitBlocks; 607 AliasSetTracker &AST; 608 public: 609 LoopPromoter(Value *SP, 610 const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 611 SmallPtrSet<Value*, 4> &PMA, 612 SmallVectorImpl<BasicBlock*> &LEB, AliasSetTracker &ast) 613 : LoadAndStorePromoter(Insts, S, 0, 0), SomePtr(SP), 614 PointerMustAliases(PMA), LoopExitBlocks(LEB), AST(ast) {} 615 616 virtual bool isInstInList(Instruction *I, 617 const SmallVectorImpl<Instruction*> &) const { 618 Value *Ptr; 619 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 620 Ptr = LI->getOperand(0); 621 else 622 Ptr = cast<StoreInst>(I)->getPointerOperand(); 623 return PointerMustAliases.count(Ptr); 624 } 625 626 virtual void doExtraRewritesBeforeFinalDeletion() const { 627 // Insert stores after in the loop exit blocks. Each exit block gets a 628 // store of the live-out values that feed them. Since we've already told 629 // the SSA updater about the defs in the loop and the preheader 630 // definition, it is all set and we can start using it. 631 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { 632 BasicBlock *ExitBlock = LoopExitBlocks[i]; 633 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); 634 Instruction *InsertPos = ExitBlock->getFirstNonPHI(); 635 new StoreInst(LiveInValue, SomePtr, InsertPos); 636 } 637 } 638 639 virtual void replaceLoadWithValue(LoadInst *LI, Value *V) const { 640 // Update alias analysis. 641 AST.copyValue(LI, V); 642 } 643 virtual void instructionDeleted(Instruction *I) const { 644 AST.deleteValue(I); 645 } 646 }; 647} // end anon namespace 648 649/// PromoteAliasSet - Try to promote memory values to scalars by sinking 650/// stores out of the loop and moving loads to before the loop. We do this by 651/// looping over the stores in the loop, looking for stores to Must pointers 652/// which are loop invariant. 653/// 654void LICM::PromoteAliasSet(AliasSet &AS) { 655 // We can promote this alias set if it has a store, if it is a "Must" alias 656 // set, if the pointer is loop invariant, and if we are not eliminating any 657 // volatile loads or stores. 658 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 659 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 660 return; 661 662 assert(!AS.empty() && 663 "Must alias set should have at least one pointer element in it!"); 664 Value *SomePtr = AS.begin()->getValue(); 665 666 // It isn't safe to promote a load/store from the loop if the load/store is 667 // conditional. For example, turning: 668 // 669 // for () { if (c) *P += 1; } 670 // 671 // into: 672 // 673 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 674 // 675 // is not safe, because *P may only be valid to access if 'c' is true. 676 // 677 // It is safe to promote P if all uses are direct load/stores and if at 678 // least one is guaranteed to be executed. 679 bool GuaranteedToExecute = false; 680 681 SmallVector<Instruction*, 64> LoopUses; 682 SmallPtrSet<Value*, 4> PointerMustAliases; 683 684 // Check that all of the pointers in the alias set have the same type. We 685 // cannot (yet) promote a memory location that is loaded and stored in 686 // different sizes. 687 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) { 688 Value *ASIV = ASI->getValue(); 689 PointerMustAliases.insert(ASIV); 690 691 // Check that all of the pointers in the alias set have the same type. We 692 // cannot (yet) promote a memory location that is loaded and stored in 693 // different sizes. 694 if (SomePtr->getType() != ASIV->getType()) 695 return; 696 697 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end(); 698 UI != UE; ++UI) { 699 // Ignore instructions that are outside the loop. 700 Instruction *Use = dyn_cast<Instruction>(*UI); 701 if (!Use || !CurLoop->contains(Use)) 702 continue; 703 704 // If there is an non-load/store instruction in the loop, we can't promote 705 // it. 706 if (isa<LoadInst>(Use)) 707 assert(!cast<LoadInst>(Use)->isVolatile() && "AST broken"); 708 else if (isa<StoreInst>(Use)) { 709 // Stores *of* the pointer are not interesting, only stores *to* the 710 // pointer. 711 if (Use->getOperand(1) != ASIV) 712 continue; 713 assert(!cast<StoreInst>(Use)->isVolatile() && "AST broken"); 714 } else 715 return; // Not a load or store. 716 717 if (!GuaranteedToExecute) 718 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 719 720 LoopUses.push_back(Use); 721 } 722 } 723 724 // If there isn't a guaranteed-to-execute instruction, we can't promote. 725 if (!GuaranteedToExecute) 726 return; 727 728 // Otherwise, this is safe to promote, lets do it! 729 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n'); 730 Changed = true; 731 ++NumPromoted; 732 733 SmallVector<BasicBlock*, 8> ExitBlocks; 734 CurLoop->getUniqueExitBlocks(ExitBlocks); 735 736 // We use the SSAUpdater interface to insert phi nodes as required. 737 SmallVector<PHINode*, 16> NewPHIs; 738 SSAUpdater SSA(&NewPHIs); 739 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, 740 *CurAST); 741 742 // Set up the preheader to have a definition of the value. It is the live-out 743 // value from the preheader that uses in the loop will use. 744 LoadInst *PreheaderLoad = 745 new LoadInst(SomePtr, SomePtr->getName()+".promoted", 746 Preheader->getTerminator()); 747 SSA.AddAvailableValue(Preheader, PreheaderLoad); 748 749 // Rewrite all the loads in the loop and remember all the definitions from 750 // stores in the loop. 751 Promoter.run(LoopUses); 752 753 // If the SSAUpdater didn't use the load in the preheader, just zap it now. 754 if (PreheaderLoad->use_empty()) 755 PreheaderLoad->eraseFromParent(); 756} 757 758 759/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 760void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 761 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 762 if (!AST) 763 return; 764 765 AST->copyValue(From, To); 766} 767 768/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 769/// set. 770void LICM::deleteAnalysisValue(Value *V, Loop *L) { 771 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 772 if (!AST) 773 return; 774 775 AST->deleteValue(V); 776} 777