LICM.cpp revision 8e0d1c03ca7fd86e6879b4e37d0d7f0e982feef6
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 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap; 112 113 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 114 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 115 116 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 117 /// set. 118 void deleteAnalysisValue(Value *V, Loop *L); 119 120 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 121 /// dominated by the specified block, and that are in the current loop) in 122 /// reverse depth first order w.r.t the DominatorTree. This allows us to 123 /// visit uses before definitions, allowing us to sink a loop body in one 124 /// pass without iteration. 125 /// 126 void SinkRegion(DomTreeNode *N); 127 128 /// HoistRegion - Walk the specified region of the CFG (defined by all 129 /// blocks dominated by the specified block, and that are in the current 130 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 131 /// visit definitions before uses, allowing us to hoist a loop body in one 132 /// pass without iteration. 133 /// 134 void HoistRegion(DomTreeNode *N); 135 136 /// inSubLoop - Little predicate that returns true if the specified basic 137 /// block is in a subloop of the current one, not the current one itself. 138 /// 139 bool inSubLoop(BasicBlock *BB) { 140 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 141 return LI->getLoopFor(BB) != CurLoop; 142 } 143 144 /// sink - When an instruction is found to only be used outside of the loop, 145 /// this function moves it to the exit blocks and patches up SSA form as 146 /// needed. 147 /// 148 void sink(Instruction &I); 149 150 /// hoist - When an instruction is found to only use loop invariant operands 151 /// that is safe to hoist, this instruction is called to do the dirty work. 152 /// 153 void hoist(Instruction &I); 154 155 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 156 /// is not a trapping instruction or if it is a trapping instruction and is 157 /// guaranteed to execute. 158 /// 159 bool isSafeToExecuteUnconditionally(Instruction &I); 160 161 /// isGuaranteedToExecute - Check that the instruction is guaranteed to 162 /// execute. 163 /// 164 bool isGuaranteedToExecute(Instruction &I); 165 166 /// pointerInvalidatedByLoop - Return true if the body of this loop may 167 /// store into the memory location pointed to by V. 168 /// 169 bool pointerInvalidatedByLoop(Value *V, uint64_t Size, 170 const MDNode *TBAAInfo) { 171 // Check to see if any of the basic blocks in CurLoop invalidate *V. 172 return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod(); 173 } 174 175 bool canSinkOrHoistInst(Instruction &I); 176 bool isNotUsedInLoop(Instruction &I); 177 178 void PromoteAliasSet(AliasSet &AS, 179 SmallVectorImpl<BasicBlock*> &ExitBlocks, 180 SmallVectorImpl<Instruction*> &InsertPts); 181 }; 182} 183 184char LICM::ID = 0; 185INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false) 186INITIALIZE_PASS_DEPENDENCY(DominatorTree) 187INITIALIZE_PASS_DEPENDENCY(LoopInfo) 188INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 189INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 190INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 191INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false) 192 193Pass *llvm::createLICMPass() { return new LICM(); } 194 195/// Hoist expressions out of the specified loop. Note, alias info for inner 196/// loop is not preserved so it is not a good idea to run LICM multiple 197/// times on one loop. 198/// 199bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 200 Changed = false; 201 202 // Get our Loop and Alias Analysis information... 203 LI = &getAnalysis<LoopInfo>(); 204 AA = &getAnalysis<AliasAnalysis>(); 205 DT = &getAnalysis<DominatorTree>(); 206 207 TD = getAnalysisIfAvailable<TargetData>(); 208 TLI = &getAnalysis<TargetLibraryInfo>(); 209 210 CurAST = new AliasSetTracker(*AA); 211 // Collect Alias info from subloops. 212 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 213 LoopItr != LoopItrE; ++LoopItr) { 214 Loop *InnerL = *LoopItr; 215 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL]; 216 assert(InnerAST && "Where is my AST?"); 217 218 // What if InnerLoop was modified by other passes ? 219 CurAST->add(*InnerAST); 220 221 // Once we've incorporated the inner loop's AST into ours, we don't need the 222 // subloop's anymore. 223 delete InnerAST; 224 LoopToAliasSetMap.erase(InnerL); 225 } 226 227 CurLoop = L; 228 229 // Get the preheader block to move instructions into... 230 Preheader = L->getLoopPreheader(); 231 232 // Loop over the body of this loop, looking for calls, invokes, and stores. 233 // Because subloops have already been incorporated into AST, we skip blocks in 234 // subloops. 235 // 236 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 237 I != E; ++I) { 238 BasicBlock *BB = *I; 239 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops. 240 CurAST->add(*BB); // Incorporate the specified basic block 241 } 242 243 // We want to visit all of the instructions in this loop... that are not parts 244 // of our subloops (they have already had their invariants hoisted out of 245 // their loop, into this loop, so there is no need to process the BODIES of 246 // the subloops). 247 // 248 // Traverse the body of the loop in depth first order on the dominator tree so 249 // that we are guaranteed to see definitions before we see uses. This allows 250 // us to sink instructions in one pass, without iteration. After sinking 251 // instructions, we perform another pass to hoist them out of the loop. 252 // 253 if (L->hasDedicatedExits()) 254 SinkRegion(DT->getNode(L->getHeader())); 255 if (Preheader) 256 HoistRegion(DT->getNode(L->getHeader())); 257 258 // Now that all loop invariants have been removed from the loop, promote any 259 // memory references to scalars that we can. 260 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) { 261 SmallVector<BasicBlock *, 8> ExitBlocks; 262 SmallVector<Instruction *, 8> InsertPts; 263 264 // Loop over all of the alias sets in the tracker object. 265 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 266 I != E; ++I) 267 PromoteAliasSet(*I, ExitBlocks, InsertPts); 268 } 269 270 // Clear out loops state information for the next iteration 271 CurLoop = 0; 272 Preheader = 0; 273 274 // If this loop is nested inside of another one, save the alias information 275 // for when we process the outer loop. 276 if (L->getParentLoop()) 277 LoopToAliasSetMap[L] = CurAST; 278 else 279 delete CurAST; 280 return Changed; 281} 282 283/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 284/// dominated by the specified block, and that are in the current loop) in 285/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 286/// uses before definitions, allowing us to sink a loop body in one pass without 287/// iteration. 288/// 289void LICM::SinkRegion(DomTreeNode *N) { 290 assert(N != 0 && "Null dominator tree node?"); 291 BasicBlock *BB = N->getBlock(); 292 293 // If this subregion is not in the top level loop at all, exit. 294 if (!CurLoop->contains(BB)) return; 295 296 // We are processing blocks in reverse dfo, so process children first. 297 const std::vector<DomTreeNode*> &Children = N->getChildren(); 298 for (unsigned i = 0, e = Children.size(); i != e; ++i) 299 SinkRegion(Children[i]); 300 301 // Only need to process the contents of this block if it is not part of a 302 // subloop (which would already have been processed). 303 if (inSubLoop(BB)) return; 304 305 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 306 Instruction &I = *--II; 307 308 // If the instruction is dead, we would try to sink it because it isn't used 309 // in the loop, instead, just delete it. 310 if (isInstructionTriviallyDead(&I, TLI)) { 311 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); 312 ++II; 313 CurAST->deleteValue(&I); 314 I.eraseFromParent(); 315 Changed = true; 316 continue; 317 } 318 319 // Check to see if we can sink this instruction to the exit blocks 320 // of the loop. We can do this if the all users of the instruction are 321 // outside of the loop. In this case, it doesn't even matter if the 322 // operands of the instruction are loop invariant. 323 // 324 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 325 ++II; 326 sink(I); 327 } 328 } 329} 330 331/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 332/// dominated by the specified block, and that are in the current loop) in depth 333/// first order w.r.t the DominatorTree. This allows us to visit definitions 334/// before uses, allowing us to hoist a loop body in one pass without iteration. 335/// 336void LICM::HoistRegion(DomTreeNode *N) { 337 assert(N != 0 && "Null dominator tree node?"); 338 BasicBlock *BB = N->getBlock(); 339 340 // If this subregion is not in the top level loop at all, exit. 341 if (!CurLoop->contains(BB)) return; 342 343 // Only need to process the contents of this block if it is not part of a 344 // subloop (which would already have been processed). 345 if (!inSubLoop(BB)) 346 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 347 Instruction &I = *II++; 348 349 // Try constant folding this instruction. If all the operands are 350 // constants, it is technically hoistable, but it would be better to just 351 // fold it. 352 if (Constant *C = ConstantFoldInstruction(&I, TD, TLI)) { 353 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'); 354 CurAST->copyValue(&I, C); 355 CurAST->deleteValue(&I); 356 I.replaceAllUsesWith(C); 357 I.eraseFromParent(); 358 continue; 359 } 360 361 // Try hoisting the instruction out to the preheader. We can only do this 362 // if all of the operands of the instruction are loop invariant and if it 363 // is safe to hoist the instruction. 364 // 365 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) && 366 isSafeToExecuteUnconditionally(I)) 367 hoist(I); 368 } 369 370 const std::vector<DomTreeNode*> &Children = N->getChildren(); 371 for (unsigned i = 0, e = Children.size(); i != e; ++i) 372 HoistRegion(Children[i]); 373} 374 375/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 376/// instruction. 377/// 378bool LICM::canSinkOrHoistInst(Instruction &I) { 379 // Loads have extra constraints we have to verify before we can hoist them. 380 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 381 if (!LI->isUnordered()) 382 return false; // Don't hoist volatile/atomic loads! 383 384 // Loads from constant memory are always safe to move, even if they end up 385 // in the same alias set as something that ends up being modified. 386 if (AA->pointsToConstantMemory(LI->getOperand(0))) 387 return true; 388 if (LI->getMetadata("invariant.load")) 389 return true; 390 391 // Don't hoist loads which have may-aliased stores in loop. 392 uint64_t Size = 0; 393 if (LI->getType()->isSized()) 394 Size = AA->getTypeStoreSize(LI->getType()); 395 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, 396 LI->getMetadata(LLVMContext::MD_tbaa)); 397 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 398 // Don't sink or hoist dbg info; it's legal, but not useful. 399 if (isa<DbgInfoIntrinsic>(I)) 400 return false; 401 402 // Handle simple cases by querying alias analysis. 403 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 404 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 405 return true; 406 if (AliasAnalysis::onlyReadsMemory(Behavior)) { 407 // If this call only reads from memory and there are no writes to memory 408 // in the loop, we can hoist or sink the call as appropriate. 409 bool FoundMod = false; 410 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 411 I != E; ++I) { 412 AliasSet &AS = *I; 413 if (!AS.isForwardingAliasSet() && AS.isMod()) { 414 FoundMod = true; 415 break; 416 } 417 } 418 if (!FoundMod) return true; 419 } 420 421 // FIXME: This should use mod/ref information to see if we can hoist or sink 422 // the call. 423 424 return false; 425 } 426 427 // Otherwise these instructions are hoistable/sinkable 428 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 429 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 430 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 431 isa<ShuffleVectorInst>(I); 432} 433 434/// isNotUsedInLoop - Return true if the only users of this instruction are 435/// outside of the loop. If this is true, we can sink the instruction to the 436/// exit blocks of the loop. 437/// 438bool LICM::isNotUsedInLoop(Instruction &I) { 439 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 440 Instruction *User = cast<Instruction>(*UI); 441 if (PHINode *PN = dyn_cast<PHINode>(User)) { 442 // PHI node uses occur in predecessor blocks! 443 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 444 if (PN->getIncomingValue(i) == &I) 445 if (CurLoop->contains(PN->getIncomingBlock(i))) 446 return false; 447 } else if (CurLoop->contains(User)) { 448 return false; 449 } 450 } 451 return true; 452} 453 454 455/// sink - When an instruction is found to only be used outside of the loop, 456/// this function moves it to the exit blocks and patches up SSA form as needed. 457/// This method is guaranteed to remove the original instruction from its 458/// position, and may either delete it or move it to outside of the loop. 459/// 460void LICM::sink(Instruction &I) { 461 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 462 463 SmallVector<BasicBlock*, 8> ExitBlocks; 464 CurLoop->getUniqueExitBlocks(ExitBlocks); 465 466 if (isa<LoadInst>(I)) ++NumMovedLoads; 467 else if (isa<CallInst>(I)) ++NumMovedCalls; 468 ++NumSunk; 469 Changed = true; 470 471 // The case where there is only a single exit node of this loop is common 472 // enough that we handle it as a special (more efficient) case. It is more 473 // efficient to handle because there are no PHI nodes that need to be placed. 474 if (ExitBlocks.size() == 1) { 475 if (!DT->dominates(I.getParent(), ExitBlocks[0])) { 476 // Instruction is not used, just delete it. 477 CurAST->deleteValue(&I); 478 // If I has users in unreachable blocks, eliminate. 479 // If I is not void type then replaceAllUsesWith undef. 480 // This allows ValueHandlers and custom metadata to adjust itself. 481 if (!I.use_empty()) 482 I.replaceAllUsesWith(UndefValue::get(I.getType())); 483 I.eraseFromParent(); 484 } else { 485 // Move the instruction to the start of the exit block, after any PHI 486 // nodes in it. 487 I.moveBefore(ExitBlocks[0]->getFirstInsertionPt()); 488 489 // This instruction is no longer in the AST for the current loop, because 490 // we just sunk it out of the loop. If we just sunk it into an outer 491 // loop, we will rediscover the operation when we process it. 492 CurAST->deleteValue(&I); 493 } 494 return; 495 } 496 497 if (ExitBlocks.empty()) { 498 // The instruction is actually dead if there ARE NO exit blocks. 499 CurAST->deleteValue(&I); 500 // If I has users in unreachable blocks, eliminate. 501 // If I is not void type then replaceAllUsesWith undef. 502 // This allows ValueHandlers and custom metadata to adjust itself. 503 if (!I.use_empty()) 504 I.replaceAllUsesWith(UndefValue::get(I.getType())); 505 I.eraseFromParent(); 506 return; 507 } 508 509 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the 510 // hard work of inserting PHI nodes as necessary. 511 SmallVector<PHINode*, 8> NewPHIs; 512 SSAUpdater SSA(&NewPHIs); 513 514 if (!I.use_empty()) 515 SSA.Initialize(I.getType(), I.getName()); 516 517 // Insert a copy of the instruction in each exit block of the loop that is 518 // dominated by the instruction. Each exit block is known to only be in the 519 // ExitBlocks list once. 520 BasicBlock *InstOrigBB = I.getParent(); 521 unsigned NumInserted = 0; 522 523 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 524 BasicBlock *ExitBlock = ExitBlocks[i]; 525 526 if (!DT->dominates(InstOrigBB, ExitBlock)) 527 continue; 528 529 // Insert the code after the last PHI node. 530 BasicBlock::iterator InsertPt = ExitBlock->getFirstInsertionPt(); 531 532 // If this is the first exit block processed, just move the original 533 // instruction, otherwise clone the original instruction and insert 534 // the copy. 535 Instruction *New; 536 if (NumInserted++ == 0) { 537 I.moveBefore(InsertPt); 538 New = &I; 539 } else { 540 New = I.clone(); 541 if (!I.getName().empty()) 542 New->setName(I.getName()+".le"); 543 ExitBlock->getInstList().insert(InsertPt, New); 544 } 545 546 // Now that we have inserted the instruction, inform SSAUpdater. 547 if (!I.use_empty()) 548 SSA.AddAvailableValue(ExitBlock, New); 549 } 550 551 // If the instruction doesn't dominate any exit blocks, it must be dead. 552 if (NumInserted == 0) { 553 CurAST->deleteValue(&I); 554 if (!I.use_empty()) 555 I.replaceAllUsesWith(UndefValue::get(I.getType())); 556 I.eraseFromParent(); 557 return; 558 } 559 560 // Next, rewrite uses of the instruction, inserting PHI nodes as needed. 561 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) { 562 // Grab the use before incrementing the iterator. 563 Use &U = UI.getUse(); 564 // Increment the iterator before removing the use from the list. 565 ++UI; 566 SSA.RewriteUseAfterInsertions(U); 567 } 568 569 // Update CurAST for NewPHIs if I had pointer type. 570 if (I.getType()->isPointerTy()) 571 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 572 CurAST->copyValue(&I, NewPHIs[i]); 573 574 // Finally, remove the instruction from CurAST. It is no longer in the loop. 575 CurAST->deleteValue(&I); 576} 577 578/// hoist - When an instruction is found to only use loop invariant operands 579/// that is safe to hoist, this instruction is called to do the dirty work. 580/// 581void LICM::hoist(Instruction &I) { 582 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 583 << I << "\n"); 584 585 // Move the new node to the Preheader, before its terminator. 586 I.moveBefore(Preheader->getTerminator()); 587 588 if (isa<LoadInst>(I)) ++NumMovedLoads; 589 else if (isa<CallInst>(I)) ++NumMovedCalls; 590 ++NumHoisted; 591 Changed = true; 592} 593 594/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 595/// not a trapping instruction or if it is a trapping instruction and is 596/// guaranteed to execute. 597/// 598bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 599 // If it is not a trapping instruction, it is always safe to hoist. 600 if (isSafeToSpeculativelyExecute(&Inst)) 601 return true; 602 603 return isGuaranteedToExecute(Inst); 604} 605 606bool LICM::isGuaranteedToExecute(Instruction &Inst) { 607 // Otherwise we have to check to make sure that the instruction dominates all 608 // of the exit blocks. If it doesn't, then there is a path out of the loop 609 // which does not execute this instruction, so we can't hoist it. 610 611 // If the instruction is in the header block for the loop (which is very 612 // common), it is always guaranteed to dominate the exit blocks. Since this 613 // is a common case, and can save some work, check it now. 614 if (Inst.getParent() == CurLoop->getHeader()) 615 return true; 616 617 // Get the exit blocks for the current loop. 618 SmallVector<BasicBlock*, 8> ExitBlocks; 619 CurLoop->getExitBlocks(ExitBlocks); 620 621 // Verify that the block dominates each of the exit blocks of the loop. 622 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 623 if (!DT->dominates(Inst.getParent(), ExitBlocks[i])) 624 return false; 625 626 // As a degenerate case, if the loop is statically infinite then we haven't 627 // proven anything since there are no exit blocks. 628 if (ExitBlocks.empty()) 629 return false; 630 631 return true; 632} 633 634namespace { 635 class LoopPromoter : public LoadAndStorePromoter { 636 Value *SomePtr; // Designated pointer to store to. 637 SmallPtrSet<Value*, 4> &PointerMustAliases; 638 SmallVectorImpl<BasicBlock*> &LoopExitBlocks; 639 SmallVectorImpl<Instruction*> &LoopInsertPts; 640 AliasSetTracker &AST; 641 DebugLoc DL; 642 int Alignment; 643 public: 644 LoopPromoter(Value *SP, 645 const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 646 SmallPtrSet<Value*, 4> &PMA, 647 SmallVectorImpl<BasicBlock*> &LEB, 648 SmallVectorImpl<Instruction*> &LIP, 649 AliasSetTracker &ast, DebugLoc dl, int alignment) 650 : LoadAndStorePromoter(Insts, S), SomePtr(SP), 651 PointerMustAliases(PMA), LoopExitBlocks(LEB), LoopInsertPts(LIP), 652 AST(ast), DL(dl), Alignment(alignment) {} 653 654 virtual bool isInstInList(Instruction *I, 655 const SmallVectorImpl<Instruction*> &) const { 656 Value *Ptr; 657 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 658 Ptr = LI->getOperand(0); 659 else 660 Ptr = cast<StoreInst>(I)->getPointerOperand(); 661 return PointerMustAliases.count(Ptr); 662 } 663 664 virtual void doExtraRewritesBeforeFinalDeletion() const { 665 // Insert stores after in the loop exit blocks. Each exit block gets a 666 // store of the live-out values that feed them. Since we've already told 667 // the SSA updater about the defs in the loop and the preheader 668 // definition, it is all set and we can start using it. 669 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { 670 BasicBlock *ExitBlock = LoopExitBlocks[i]; 671 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); 672 Instruction *InsertPos = LoopInsertPts[i]; 673 StoreInst *NewSI = new StoreInst(LiveInValue, SomePtr, InsertPos); 674 NewSI->setAlignment(Alignment); 675 NewSI->setDebugLoc(DL); 676 } 677 } 678 679 virtual void replaceLoadWithValue(LoadInst *LI, Value *V) const { 680 // Update alias analysis. 681 AST.copyValue(LI, V); 682 } 683 virtual void instructionDeleted(Instruction *I) const { 684 AST.deleteValue(I); 685 } 686 }; 687} // end anon namespace 688 689/// PromoteAliasSet - Try to promote memory values to scalars by sinking 690/// stores out of the loop and moving loads to before the loop. We do this by 691/// looping over the stores in the loop, looking for stores to Must pointers 692/// which are loop invariant. 693/// 694void LICM::PromoteAliasSet(AliasSet &AS, 695 SmallVectorImpl<BasicBlock*> &ExitBlocks, 696 SmallVectorImpl<Instruction*> &InsertPts) { 697 // We can promote this alias set if it has a store, if it is a "Must" alias 698 // set, if the pointer is loop invariant, and if we are not eliminating any 699 // volatile loads or stores. 700 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 701 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 702 return; 703 704 assert(!AS.empty() && 705 "Must alias set should have at least one pointer element in it!"); 706 Value *SomePtr = AS.begin()->getValue(); 707 708 // It isn't safe to promote a load/store from the loop if the load/store is 709 // conditional. For example, turning: 710 // 711 // for () { if (c) *P += 1; } 712 // 713 // into: 714 // 715 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 716 // 717 // is not safe, because *P may only be valid to access if 'c' is true. 718 // 719 // It is safe to promote P if all uses are direct load/stores and if at 720 // least one is guaranteed to be executed. 721 bool GuaranteedToExecute = false; 722 723 SmallVector<Instruction*, 64> LoopUses; 724 SmallPtrSet<Value*, 4> PointerMustAliases; 725 726 // We start with an alignment of one and try to find instructions that allow 727 // us to prove better alignment. 728 unsigned Alignment = 1; 729 730 // Check that all of the pointers in the alias set have the same type. We 731 // cannot (yet) promote a memory location that is loaded and stored in 732 // different sizes. 733 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) { 734 Value *ASIV = ASI->getValue(); 735 PointerMustAliases.insert(ASIV); 736 737 // Check that all of the pointers in the alias set have the same type. We 738 // cannot (yet) promote a memory location that is loaded and stored in 739 // different sizes. 740 if (SomePtr->getType() != ASIV->getType()) 741 return; 742 743 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end(); 744 UI != UE; ++UI) { 745 // Ignore instructions that are outside the loop. 746 Instruction *Use = dyn_cast<Instruction>(*UI); 747 if (!Use || !CurLoop->contains(Use)) 748 continue; 749 750 // If there is an non-load/store instruction in the loop, we can't promote 751 // it. 752 if (LoadInst *load = dyn_cast<LoadInst>(Use)) { 753 assert(!load->isVolatile() && "AST broken"); 754 if (!load->isSimple()) 755 return; 756 } else if (StoreInst *store = dyn_cast<StoreInst>(Use)) { 757 // Stores *of* the pointer are not interesting, only stores *to* the 758 // pointer. 759 if (Use->getOperand(1) != ASIV) 760 continue; 761 assert(!store->isVolatile() && "AST broken"); 762 if (!store->isSimple()) 763 return; 764 765 // Note that we only check GuaranteedToExecute inside the store case 766 // so that we do not introduce stores where they did not exist before 767 // (which would break the LLVM concurrency model). 768 769 // If the alignment of this instruction allows us to specify a more 770 // restrictive (and performant) alignment and if we are sure this 771 // instruction will be executed, update the alignment. 772 // Larger is better, with the exception of 0 being the best alignment. 773 unsigned InstAlignment = store->getAlignment(); 774 if ((InstAlignment > Alignment || InstAlignment == 0) 775 && (Alignment != 0)) 776 if (isGuaranteedToExecute(*Use)) { 777 GuaranteedToExecute = true; 778 Alignment = InstAlignment; 779 } 780 781 if (!GuaranteedToExecute) 782 GuaranteedToExecute = isGuaranteedToExecute(*Use); 783 784 } else 785 return; // Not a load or store. 786 787 LoopUses.push_back(Use); 788 } 789 } 790 791 // If there isn't a guaranteed-to-execute instruction, we can't promote. 792 if (!GuaranteedToExecute) 793 return; 794 795 // Otherwise, this is safe to promote, lets do it! 796 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n'); 797 Changed = true; 798 ++NumPromoted; 799 800 // Grab a debug location for the inserted loads/stores; given that the 801 // inserted loads/stores have little relation to the original loads/stores, 802 // this code just arbitrarily picks a location from one, since any debug 803 // location is better than none. 804 DebugLoc DL = LoopUses[0]->getDebugLoc(); 805 806 // Figure out the loop exits and their insertion points, if this is the 807 // first promotion. 808 if (ExitBlocks.empty()) { 809 CurLoop->getUniqueExitBlocks(ExitBlocks); 810 InsertPts.resize(ExitBlocks.size()); 811 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 812 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt(); 813 } 814 815 // We use the SSAUpdater interface to insert phi nodes as required. 816 SmallVector<PHINode*, 16> NewPHIs; 817 SSAUpdater SSA(&NewPHIs); 818 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, 819 InsertPts, *CurAST, DL, Alignment); 820 821 // Set up the preheader to have a definition of the value. It is the live-out 822 // value from the preheader that uses in the loop will use. 823 LoadInst *PreheaderLoad = 824 new LoadInst(SomePtr, SomePtr->getName()+".promoted", 825 Preheader->getTerminator()); 826 PreheaderLoad->setAlignment(Alignment); 827 PreheaderLoad->setDebugLoc(DL); 828 SSA.AddAvailableValue(Preheader, PreheaderLoad); 829 830 // Rewrite all the loads in the loop and remember all the definitions from 831 // stores in the loop. 832 Promoter.run(LoopUses); 833 834 // If the SSAUpdater didn't use the load in the preheader, just zap it now. 835 if (PreheaderLoad->use_empty()) 836 PreheaderLoad->eraseFromParent(); 837} 838 839 840/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 841void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 842 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 843 if (!AST) 844 return; 845 846 AST->copyValue(From, To); 847} 848 849/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 850/// set. 851void LICM::deleteAnalysisValue(Value *V, Loop *L) { 852 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 853 if (!AST) 854 return; 855 856 AST->deleteValue(V); 857} 858