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/Analysis/ValueTracking.h" 47#include "llvm/Transforms/Utils/Local.h" 48#include "llvm/Transforms/Utils/SSAUpdater.h" 49#include "llvm/Target/TargetData.h" 50#include "llvm/Target/TargetLibraryInfo.h" 51#include "llvm/Support/CFG.h" 52#include "llvm/Support/CommandLine.h" 53#include "llvm/Support/raw_ostream.h" 54#include "llvm/Support/Debug.h" 55#include "llvm/ADT/Statistic.h" 56#include <algorithm> 57using namespace llvm; 58 59STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 60STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 61STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 62STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 63STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 64 65static cl::opt<bool> 66DisablePromotion("disable-licm-promotion", cl::Hidden, 67 cl::desc("Disable memory promotion in LICM pass")); 68 69namespace { 70 struct LICM : public LoopPass { 71 static char ID; // Pass identification, replacement for typeid 72 LICM() : LoopPass(ID) { 73 initializeLICMPass(*PassRegistry::getPassRegistry()); 74 } 75 76 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 77 78 /// This transformation requires natural loop information & requires that 79 /// loop preheaders be inserted into the CFG... 80 /// 81 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 82 AU.setPreservesCFG(); 83 AU.addRequired<DominatorTree>(); 84 AU.addRequired<LoopInfo>(); 85 AU.addRequiredID(LoopSimplifyID); 86 AU.addRequired<AliasAnalysis>(); 87 AU.addPreserved<AliasAnalysis>(); 88 AU.addPreserved("scalar-evolution"); 89 AU.addPreservedID(LoopSimplifyID); 90 AU.addRequired<TargetLibraryInfo>(); 91 } 92 93 bool doFinalization() { 94 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets"); 95 return false; 96 } 97 98 private: 99 AliasAnalysis *AA; // Current AliasAnalysis information 100 LoopInfo *LI; // Current LoopInfo 101 DominatorTree *DT; // Dominator Tree for the current Loop. 102 103 TargetData *TD; // TargetData for constant folding. 104 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding. 105 106 // State that is updated as we process loops. 107 bool Changed; // Set to true when we change anything. 108 BasicBlock *Preheader; // The preheader block of the current loop... 109 Loop *CurLoop; // The current loop we are working on... 110 AliasSetTracker *CurAST; // AliasSet information for the current loop... 111 bool MayThrow; // The current loop contains an instruction which 112 // may throw, thus preventing code motion of 113 // instructions with side effects. 114 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap; 115 116 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 117 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 118 119 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 120 /// set. 121 void deleteAnalysisValue(Value *V, Loop *L); 122 123 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 124 /// dominated by the specified block, and that are in the current loop) in 125 /// reverse depth first order w.r.t the DominatorTree. This allows us to 126 /// visit uses before definitions, allowing us to sink a loop body in one 127 /// pass without iteration. 128 /// 129 void SinkRegion(DomTreeNode *N); 130 131 /// HoistRegion - Walk the specified region of the CFG (defined by all 132 /// blocks dominated by the specified block, and that are in the current 133 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 134 /// visit definitions before uses, allowing us to hoist a loop body in one 135 /// pass without iteration. 136 /// 137 void HoistRegion(DomTreeNode *N); 138 139 /// inSubLoop - Little predicate that returns true if the specified basic 140 /// block is in a subloop of the current one, not the current one itself. 141 /// 142 bool inSubLoop(BasicBlock *BB) { 143 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 144 return LI->getLoopFor(BB) != CurLoop; 145 } 146 147 /// sink - When an instruction is found to only be used outside of the loop, 148 /// this function moves it to the exit blocks and patches up SSA form as 149 /// needed. 150 /// 151 void sink(Instruction &I); 152 153 /// hoist - When an instruction is found to only use loop invariant operands 154 /// that is safe to hoist, this instruction is called to do the dirty work. 155 /// 156 void hoist(Instruction &I); 157 158 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 159 /// is not a trapping instruction or if it is a trapping instruction and is 160 /// guaranteed to execute. 161 /// 162 bool isSafeToExecuteUnconditionally(Instruction &I); 163 164 /// isGuaranteedToExecute - Check that the instruction is guaranteed to 165 /// execute. 166 /// 167 bool isGuaranteedToExecute(Instruction &I); 168 169 /// pointerInvalidatedByLoop - Return true if the body of this loop may 170 /// store into the memory location pointed to by V. 171 /// 172 bool pointerInvalidatedByLoop(Value *V, uint64_t Size, 173 const MDNode *TBAAInfo) { 174 // Check to see if any of the basic blocks in CurLoop invalidate *V. 175 return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod(); 176 } 177 178 bool canSinkOrHoistInst(Instruction &I); 179 bool isNotUsedInLoop(Instruction &I); 180 181 void PromoteAliasSet(AliasSet &AS, 182 SmallVectorImpl<BasicBlock*> &ExitBlocks, 183 SmallVectorImpl<Instruction*> &InsertPts); 184 }; 185} 186 187char LICM::ID = 0; 188INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false) 189INITIALIZE_PASS_DEPENDENCY(DominatorTree) 190INITIALIZE_PASS_DEPENDENCY(LoopInfo) 191INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 192INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 193INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 194INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false) 195 196Pass *llvm::createLICMPass() { return new LICM(); } 197 198/// Hoist expressions out of the specified loop. Note, alias info for inner 199/// loop is not preserved so it is not a good idea to run LICM multiple 200/// times on one loop. 201/// 202bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 203 Changed = false; 204 205 // Get our Loop and Alias Analysis information... 206 LI = &getAnalysis<LoopInfo>(); 207 AA = &getAnalysis<AliasAnalysis>(); 208 DT = &getAnalysis<DominatorTree>(); 209 210 TD = getAnalysisIfAvailable<TargetData>(); 211 TLI = &getAnalysis<TargetLibraryInfo>(); 212 213 CurAST = new AliasSetTracker(*AA); 214 // Collect Alias info from subloops. 215 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 216 LoopItr != LoopItrE; ++LoopItr) { 217 Loop *InnerL = *LoopItr; 218 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL]; 219 assert(InnerAST && "Where is my AST?"); 220 221 // What if InnerLoop was modified by other passes ? 222 CurAST->add(*InnerAST); 223 224 // Once we've incorporated the inner loop's AST into ours, we don't need the 225 // subloop's anymore. 226 delete InnerAST; 227 LoopToAliasSetMap.erase(InnerL); 228 } 229 230 CurLoop = L; 231 232 // Get the preheader block to move instructions into... 233 Preheader = L->getLoopPreheader(); 234 235 // Loop over the body of this loop, looking for calls, invokes, and stores. 236 // Because subloops have already been incorporated into AST, we skip blocks in 237 // subloops. 238 // 239 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 240 I != E; ++I) { 241 BasicBlock *BB = *I; 242 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops. 243 CurAST->add(*BB); // Incorporate the specified basic block 244 } 245 246 MayThrow = false; 247 // TODO: We've already searched for instructions which may throw in subloops. 248 // We may want to reuse this information. 249 for (Loop::block_iterator BB = L->block_begin(), BBE = L->block_end(); 250 (BB != BBE) && !MayThrow ; ++BB) 251 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); 252 (I != E) && !MayThrow; ++I) 253 MayThrow |= I->mayThrow(); 254 255 // We want to visit all of the instructions in this loop... that are not parts 256 // of our subloops (they have already had their invariants hoisted out of 257 // their loop, into this loop, so there is no need to process the BODIES of 258 // the subloops). 259 // 260 // Traverse the body of the loop in depth first order on the dominator tree so 261 // that we are guaranteed to see definitions before we see uses. This allows 262 // us to sink instructions in one pass, without iteration. After sinking 263 // instructions, we perform another pass to hoist them out of the loop. 264 // 265 if (L->hasDedicatedExits()) 266 SinkRegion(DT->getNode(L->getHeader())); 267 if (Preheader) 268 HoistRegion(DT->getNode(L->getHeader())); 269 270 // Now that all loop invariants have been removed from the loop, promote any 271 // memory references to scalars that we can. 272 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) { 273 SmallVector<BasicBlock *, 8> ExitBlocks; 274 SmallVector<Instruction *, 8> InsertPts; 275 276 // Loop over all of the alias sets in the tracker object. 277 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 278 I != E; ++I) 279 PromoteAliasSet(*I, ExitBlocks, InsertPts); 280 } 281 282 // Clear out loops state information for the next iteration 283 CurLoop = 0; 284 Preheader = 0; 285 286 // If this loop is nested inside of another one, save the alias information 287 // for when we process the outer loop. 288 if (L->getParentLoop()) 289 LoopToAliasSetMap[L] = CurAST; 290 else 291 delete CurAST; 292 return Changed; 293} 294 295/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 296/// dominated by the specified block, and that are in the current loop) in 297/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 298/// uses before definitions, allowing us to sink a loop body in one pass without 299/// iteration. 300/// 301void LICM::SinkRegion(DomTreeNode *N) { 302 assert(N != 0 && "Null dominator tree node?"); 303 BasicBlock *BB = N->getBlock(); 304 305 // If this subregion is not in the top level loop at all, exit. 306 if (!CurLoop->contains(BB)) return; 307 308 // We are processing blocks in reverse dfo, so process children first. 309 const std::vector<DomTreeNode*> &Children = N->getChildren(); 310 for (unsigned i = 0, e = Children.size(); i != e; ++i) 311 SinkRegion(Children[i]); 312 313 // Only need to process the contents of this block if it is not part of a 314 // subloop (which would already have been processed). 315 if (inSubLoop(BB)) return; 316 317 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 318 Instruction &I = *--II; 319 320 // If the instruction is dead, we would try to sink it because it isn't used 321 // in the loop, instead, just delete it. 322 if (isInstructionTriviallyDead(&I, TLI)) { 323 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); 324 ++II; 325 CurAST->deleteValue(&I); 326 I.eraseFromParent(); 327 Changed = true; 328 continue; 329 } 330 331 // Check to see if we can sink this instruction to the exit blocks 332 // of the loop. We can do this if the all users of the instruction are 333 // outside of the loop. In this case, it doesn't even matter if the 334 // operands of the instruction are loop invariant. 335 // 336 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 337 ++II; 338 sink(I); 339 } 340 } 341} 342 343/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 344/// dominated by the specified block, and that are in the current loop) in depth 345/// first order w.r.t the DominatorTree. This allows us to visit definitions 346/// before uses, allowing us to hoist a loop body in one pass without iteration. 347/// 348void LICM::HoistRegion(DomTreeNode *N) { 349 assert(N != 0 && "Null dominator tree node?"); 350 BasicBlock *BB = N->getBlock(); 351 352 // If this subregion is not in the top level loop at all, exit. 353 if (!CurLoop->contains(BB)) return; 354 355 // Only need to process the contents of this block if it is not part of a 356 // subloop (which would already have been processed). 357 if (!inSubLoop(BB)) 358 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 359 Instruction &I = *II++; 360 361 // Try constant folding this instruction. If all the operands are 362 // constants, it is technically hoistable, but it would be better to just 363 // fold it. 364 if (Constant *C = ConstantFoldInstruction(&I, TD, TLI)) { 365 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'); 366 CurAST->copyValue(&I, C); 367 CurAST->deleteValue(&I); 368 I.replaceAllUsesWith(C); 369 I.eraseFromParent(); 370 continue; 371 } 372 373 // Try hoisting the instruction out to the preheader. We can only do this 374 // if all of the operands of the instruction are loop invariant and if it 375 // is safe to hoist the instruction. 376 // 377 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) && 378 isSafeToExecuteUnconditionally(I)) 379 hoist(I); 380 } 381 382 const std::vector<DomTreeNode*> &Children = N->getChildren(); 383 for (unsigned i = 0, e = Children.size(); i != e; ++i) 384 HoistRegion(Children[i]); 385} 386 387/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 388/// instruction. 389/// 390bool LICM::canSinkOrHoistInst(Instruction &I) { 391 // Loads have extra constraints we have to verify before we can hoist them. 392 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 393 if (!LI->isUnordered()) 394 return false; // Don't hoist volatile/atomic loads! 395 396 // Loads from constant memory are always safe to move, even if they end up 397 // in the same alias set as something that ends up being modified. 398 if (AA->pointsToConstantMemory(LI->getOperand(0))) 399 return true; 400 if (LI->getMetadata("invariant.load")) 401 return true; 402 403 // Don't hoist loads which have may-aliased stores in loop. 404 uint64_t Size = 0; 405 if (LI->getType()->isSized()) 406 Size = AA->getTypeStoreSize(LI->getType()); 407 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, 408 LI->getMetadata(LLVMContext::MD_tbaa)); 409 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 410 // Don't sink or hoist dbg info; it's legal, but not useful. 411 if (isa<DbgInfoIntrinsic>(I)) 412 return false; 413 414 // Handle simple cases by querying alias analysis. 415 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 416 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 417 return true; 418 if (AliasAnalysis::onlyReadsMemory(Behavior)) { 419 // If this call only reads from memory and there are no writes to memory 420 // in the loop, we can hoist or sink the call as appropriate. 421 bool FoundMod = false; 422 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 423 I != E; ++I) { 424 AliasSet &AS = *I; 425 if (!AS.isForwardingAliasSet() && AS.isMod()) { 426 FoundMod = true; 427 break; 428 } 429 } 430 if (!FoundMod) return true; 431 } 432 433 // FIXME: This should use mod/ref information to see if we can hoist or 434 // sink the call. 435 436 return false; 437 } 438 439 // Only these instructions are hoistable/sinkable. 440 bool HoistableKind = (isa<BinaryOperator>(I) || isa<CastInst>(I) || 441 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || 442 isa<CmpInst>(I) || isa<InsertElementInst>(I) || 443 isa<ExtractElementInst>(I) || 444 isa<ShuffleVectorInst>(I)); 445 if (!HoistableKind) 446 return false; 447 448 return isSafeToExecuteUnconditionally(I); 449} 450 451/// isNotUsedInLoop - Return true if the only users of this instruction are 452/// outside of the loop. If this is true, we can sink the instruction to the 453/// exit blocks of the loop. 454/// 455bool LICM::isNotUsedInLoop(Instruction &I) { 456 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 457 Instruction *User = cast<Instruction>(*UI); 458 if (PHINode *PN = dyn_cast<PHINode>(User)) { 459 // PHI node uses occur in predecessor blocks! 460 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 461 if (PN->getIncomingValue(i) == &I) 462 if (CurLoop->contains(PN->getIncomingBlock(i))) 463 return false; 464 } else if (CurLoop->contains(User)) { 465 return false; 466 } 467 } 468 return true; 469} 470 471 472/// sink - When an instruction is found to only be used outside of the loop, 473/// this function moves it to the exit blocks and patches up SSA form as needed. 474/// This method is guaranteed to remove the original instruction from its 475/// position, and may either delete it or move it to outside of the loop. 476/// 477void LICM::sink(Instruction &I) { 478 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 479 480 SmallVector<BasicBlock*, 8> ExitBlocks; 481 CurLoop->getUniqueExitBlocks(ExitBlocks); 482 483 if (isa<LoadInst>(I)) ++NumMovedLoads; 484 else if (isa<CallInst>(I)) ++NumMovedCalls; 485 ++NumSunk; 486 Changed = true; 487 488 // The case where there is only a single exit node of this loop is common 489 // enough that we handle it as a special (more efficient) case. It is more 490 // efficient to handle because there are no PHI nodes that need to be placed. 491 if (ExitBlocks.size() == 1) { 492 if (!DT->dominates(I.getParent(), ExitBlocks[0])) { 493 // Instruction is not used, just delete it. 494 CurAST->deleteValue(&I); 495 // If I has users in unreachable blocks, eliminate. 496 // If I is not void type then replaceAllUsesWith undef. 497 // This allows ValueHandlers and custom metadata to adjust itself. 498 if (!I.use_empty()) 499 I.replaceAllUsesWith(UndefValue::get(I.getType())); 500 I.eraseFromParent(); 501 } else { 502 // Move the instruction to the start of the exit block, after any PHI 503 // nodes in it. 504 I.moveBefore(ExitBlocks[0]->getFirstInsertionPt()); 505 506 // This instruction is no longer in the AST for the current loop, because 507 // we just sunk it out of the loop. If we just sunk it into an outer 508 // loop, we will rediscover the operation when we process it. 509 CurAST->deleteValue(&I); 510 } 511 return; 512 } 513 514 if (ExitBlocks.empty()) { 515 // The instruction is actually dead if there ARE NO exit blocks. 516 CurAST->deleteValue(&I); 517 // If I has users in unreachable blocks, eliminate. 518 // If I is not void type then replaceAllUsesWith undef. 519 // This allows ValueHandlers and custom metadata to adjust itself. 520 if (!I.use_empty()) 521 I.replaceAllUsesWith(UndefValue::get(I.getType())); 522 I.eraseFromParent(); 523 return; 524 } 525 526 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the 527 // hard work of inserting PHI nodes as necessary. 528 SmallVector<PHINode*, 8> NewPHIs; 529 SSAUpdater SSA(&NewPHIs); 530 531 if (!I.use_empty()) 532 SSA.Initialize(I.getType(), I.getName()); 533 534 // Insert a copy of the instruction in each exit block of the loop that is 535 // dominated by the instruction. Each exit block is known to only be in the 536 // ExitBlocks list once. 537 BasicBlock *InstOrigBB = I.getParent(); 538 unsigned NumInserted = 0; 539 540 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 541 BasicBlock *ExitBlock = ExitBlocks[i]; 542 543 if (!DT->dominates(InstOrigBB, ExitBlock)) 544 continue; 545 546 // Insert the code after the last PHI node. 547 BasicBlock::iterator InsertPt = ExitBlock->getFirstInsertionPt(); 548 549 // If this is the first exit block processed, just move the original 550 // instruction, otherwise clone the original instruction and insert 551 // the copy. 552 Instruction *New; 553 if (NumInserted++ == 0) { 554 I.moveBefore(InsertPt); 555 New = &I; 556 } else { 557 New = I.clone(); 558 if (!I.getName().empty()) 559 New->setName(I.getName()+".le"); 560 ExitBlock->getInstList().insert(InsertPt, New); 561 } 562 563 // Now that we have inserted the instruction, inform SSAUpdater. 564 if (!I.use_empty()) 565 SSA.AddAvailableValue(ExitBlock, New); 566 } 567 568 // If the instruction doesn't dominate any exit blocks, it must be dead. 569 if (NumInserted == 0) { 570 CurAST->deleteValue(&I); 571 if (!I.use_empty()) 572 I.replaceAllUsesWith(UndefValue::get(I.getType())); 573 I.eraseFromParent(); 574 return; 575 } 576 577 // Next, rewrite uses of the instruction, inserting PHI nodes as needed. 578 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) { 579 // Grab the use before incrementing the iterator. 580 Use &U = UI.getUse(); 581 // Increment the iterator before removing the use from the list. 582 ++UI; 583 SSA.RewriteUseAfterInsertions(U); 584 } 585 586 // Update CurAST for NewPHIs if I had pointer type. 587 if (I.getType()->isPointerTy()) 588 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 589 CurAST->copyValue(&I, NewPHIs[i]); 590 591 // Finally, remove the instruction from CurAST. It is no longer in the loop. 592 CurAST->deleteValue(&I); 593} 594 595/// hoist - When an instruction is found to only use loop invariant operands 596/// that is safe to hoist, this instruction is called to do the dirty work. 597/// 598void LICM::hoist(Instruction &I) { 599 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 600 << I << "\n"); 601 602 // Move the new node to the Preheader, before its terminator. 603 I.moveBefore(Preheader->getTerminator()); 604 605 if (isa<LoadInst>(I)) ++NumMovedLoads; 606 else if (isa<CallInst>(I)) ++NumMovedCalls; 607 ++NumHoisted; 608 Changed = true; 609} 610 611/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 612/// not a trapping instruction or if it is a trapping instruction and is 613/// guaranteed to execute. 614/// 615bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 616 // If it is not a trapping instruction, it is always safe to hoist. 617 if (isSafeToSpeculativelyExecute(&Inst)) 618 return true; 619 620 return isGuaranteedToExecute(Inst); 621} 622 623bool LICM::isGuaranteedToExecute(Instruction &Inst) { 624 625 // Somewhere in this loop there is an instruction which may throw and make us 626 // exit the loop. 627 if (MayThrow) 628 return false; 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 // Verify that the block dominates each of the exit blocks of the loop. 645 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 646 if (!DT->dominates(Inst.getParent(), ExitBlocks[i])) 647 return false; 648 649 // As a degenerate case, if the loop is statically infinite then we haven't 650 // proven anything since there are no exit blocks. 651 if (ExitBlocks.empty()) 652 return false; 653 654 return true; 655} 656 657namespace { 658 class LoopPromoter : public LoadAndStorePromoter { 659 Value *SomePtr; // Designated pointer to store to. 660 SmallPtrSet<Value*, 4> &PointerMustAliases; 661 SmallVectorImpl<BasicBlock*> &LoopExitBlocks; 662 SmallVectorImpl<Instruction*> &LoopInsertPts; 663 AliasSetTracker &AST; 664 DebugLoc DL; 665 int Alignment; 666 public: 667 LoopPromoter(Value *SP, 668 const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 669 SmallPtrSet<Value*, 4> &PMA, 670 SmallVectorImpl<BasicBlock*> &LEB, 671 SmallVectorImpl<Instruction*> &LIP, 672 AliasSetTracker &ast, DebugLoc dl, int alignment) 673 : LoadAndStorePromoter(Insts, S), SomePtr(SP), 674 PointerMustAliases(PMA), LoopExitBlocks(LEB), LoopInsertPts(LIP), 675 AST(ast), DL(dl), Alignment(alignment) {} 676 677 virtual bool isInstInList(Instruction *I, 678 const SmallVectorImpl<Instruction*> &) const { 679 Value *Ptr; 680 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 681 Ptr = LI->getOperand(0); 682 else 683 Ptr = cast<StoreInst>(I)->getPointerOperand(); 684 return PointerMustAliases.count(Ptr); 685 } 686 687 virtual void doExtraRewritesBeforeFinalDeletion() const { 688 // Insert stores after in the loop exit blocks. Each exit block gets a 689 // store of the live-out values that feed them. Since we've already told 690 // the SSA updater about the defs in the loop and the preheader 691 // definition, it is all set and we can start using it. 692 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { 693 BasicBlock *ExitBlock = LoopExitBlocks[i]; 694 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); 695 Instruction *InsertPos = LoopInsertPts[i]; 696 StoreInst *NewSI = new StoreInst(LiveInValue, SomePtr, InsertPos); 697 NewSI->setAlignment(Alignment); 698 NewSI->setDebugLoc(DL); 699 } 700 } 701 702 virtual void replaceLoadWithValue(LoadInst *LI, Value *V) const { 703 // Update alias analysis. 704 AST.copyValue(LI, V); 705 } 706 virtual void instructionDeleted(Instruction *I) const { 707 AST.deleteValue(I); 708 } 709 }; 710} // end anon namespace 711 712/// PromoteAliasSet - Try to promote memory values to scalars by sinking 713/// stores out of the loop and moving loads to before the loop. We do this by 714/// looping over the stores in the loop, looking for stores to Must pointers 715/// which are loop invariant. 716/// 717void LICM::PromoteAliasSet(AliasSet &AS, 718 SmallVectorImpl<BasicBlock*> &ExitBlocks, 719 SmallVectorImpl<Instruction*> &InsertPts) { 720 // We can promote this alias set if it has a store, if it is a "Must" alias 721 // set, if the pointer is loop invariant, and if we are not eliminating any 722 // volatile loads or stores. 723 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 724 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 725 return; 726 727 assert(!AS.empty() && 728 "Must alias set should have at least one pointer element in it!"); 729 Value *SomePtr = AS.begin()->getValue(); 730 731 // It isn't safe to promote a load/store from the loop if the load/store is 732 // conditional. For example, turning: 733 // 734 // for () { if (c) *P += 1; } 735 // 736 // into: 737 // 738 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 739 // 740 // is not safe, because *P may only be valid to access if 'c' is true. 741 // 742 // It is safe to promote P if all uses are direct load/stores and if at 743 // least one is guaranteed to be executed. 744 bool GuaranteedToExecute = false; 745 746 SmallVector<Instruction*, 64> LoopUses; 747 SmallPtrSet<Value*, 4> PointerMustAliases; 748 749 // We start with an alignment of one and try to find instructions that allow 750 // us to prove better alignment. 751 unsigned Alignment = 1; 752 753 // Check that all of the pointers in the alias set have the same type. We 754 // cannot (yet) promote a memory location that is loaded and stored in 755 // different sizes. 756 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) { 757 Value *ASIV = ASI->getValue(); 758 PointerMustAliases.insert(ASIV); 759 760 // Check that all of the pointers in the alias set have the same type. We 761 // cannot (yet) promote a memory location that is loaded and stored in 762 // different sizes. 763 if (SomePtr->getType() != ASIV->getType()) 764 return; 765 766 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end(); 767 UI != UE; ++UI) { 768 // Ignore instructions that are outside the loop. 769 Instruction *Use = dyn_cast<Instruction>(*UI); 770 if (!Use || !CurLoop->contains(Use)) 771 continue; 772 773 // If there is an non-load/store instruction in the loop, we can't promote 774 // it. 775 if (LoadInst *load = dyn_cast<LoadInst>(Use)) { 776 assert(!load->isVolatile() && "AST broken"); 777 if (!load->isSimple()) 778 return; 779 } else if (StoreInst *store = dyn_cast<StoreInst>(Use)) { 780 // Stores *of* the pointer are not interesting, only stores *to* the 781 // pointer. 782 if (Use->getOperand(1) != ASIV) 783 continue; 784 assert(!store->isVolatile() && "AST broken"); 785 if (!store->isSimple()) 786 return; 787 788 // Note that we only check GuaranteedToExecute inside the store case 789 // so that we do not introduce stores where they did not exist before 790 // (which would break the LLVM concurrency model). 791 792 // If the alignment of this instruction allows us to specify a more 793 // restrictive (and performant) alignment and if we are sure this 794 // instruction will be executed, update the alignment. 795 // Larger is better, with the exception of 0 being the best alignment. 796 unsigned InstAlignment = store->getAlignment(); 797 if ((InstAlignment > Alignment || InstAlignment == 0) 798 && (Alignment != 0)) 799 if (isGuaranteedToExecute(*Use)) { 800 GuaranteedToExecute = true; 801 Alignment = InstAlignment; 802 } 803 804 if (!GuaranteedToExecute) 805 GuaranteedToExecute = isGuaranteedToExecute(*Use); 806 807 } else 808 return; // Not a load or store. 809 810 LoopUses.push_back(Use); 811 } 812 } 813 814 // If there isn't a guaranteed-to-execute instruction, we can't promote. 815 if (!GuaranteedToExecute) 816 return; 817 818 // Otherwise, this is safe to promote, lets do it! 819 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n'); 820 Changed = true; 821 ++NumPromoted; 822 823 // Grab a debug location for the inserted loads/stores; given that the 824 // inserted loads/stores have little relation to the original loads/stores, 825 // this code just arbitrarily picks a location from one, since any debug 826 // location is better than none. 827 DebugLoc DL = LoopUses[0]->getDebugLoc(); 828 829 // Figure out the loop exits and their insertion points, if this is the 830 // first promotion. 831 if (ExitBlocks.empty()) { 832 CurLoop->getUniqueExitBlocks(ExitBlocks); 833 InsertPts.resize(ExitBlocks.size()); 834 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 835 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt(); 836 } 837 838 // We use the SSAUpdater interface to insert phi nodes as required. 839 SmallVector<PHINode*, 16> NewPHIs; 840 SSAUpdater SSA(&NewPHIs); 841 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, 842 InsertPts, *CurAST, DL, Alignment); 843 844 // Set up the preheader to have a definition of the value. It is the live-out 845 // value from the preheader that uses in the loop will use. 846 LoadInst *PreheaderLoad = 847 new LoadInst(SomePtr, SomePtr->getName()+".promoted", 848 Preheader->getTerminator()); 849 PreheaderLoad->setAlignment(Alignment); 850 PreheaderLoad->setDebugLoc(DL); 851 SSA.AddAvailableValue(Preheader, PreheaderLoad); 852 853 // Rewrite all the loads in the loop and remember all the definitions from 854 // stores in the loop. 855 Promoter.run(LoopUses); 856 857 // If the SSAUpdater didn't use the load in the preheader, just zap it now. 858 if (PreheaderLoad->use_empty()) 859 PreheaderLoad->eraseFromParent(); 860} 861 862 863/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 864void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 865 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 866 if (!AST) 867 return; 868 869 AST->copyValue(From, To); 870} 871 872/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 873/// set. 874void LICM::deleteAnalysisValue(Value *V, Loop *L) { 875 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 876 if (!AST) 877 return; 878 879 AST->deleteValue(V); 880} 881