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> 57 using namespace llvm; 58 59 STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 60 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 61 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 62 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 63 STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 64 65 static cl::opt<bool> 66 DisablePromotion("disable-licm-promotion", cl::Hidden, 67 cl::desc("Disable memory promotion in LICM pass")); 68 69 namespace { 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 }; 180 } 181 182 char LICM::ID = 0; 183 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false) 184 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 185 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 186 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 187 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 188 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 189 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false) 190 191 Pass *llvm::createLICMPass() { return new LICM(); } 192 193 /// Hoist expressions out of the specified loop. Note, alias info for inner 194 /// loop is not preserved so it is not a good idea to run LICM multiple 195 /// times on one loop. 196 /// 197 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 198 Changed = false; 199 200 // Get our Loop and Alias Analysis information... 201 LI = &getAnalysis<LoopInfo>(); 202 AA = &getAnalysis<AliasAnalysis>(); 203 DT = &getAnalysis<DominatorTree>(); 204 205 TD = getAnalysisIfAvailable<TargetData>(); 206 TLI = &getAnalysis<TargetLibraryInfo>(); 207 208 CurAST = new AliasSetTracker(*AA); 209 // Collect Alias info from subloops. 210 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 211 LoopItr != LoopItrE; ++LoopItr) { 212 Loop *InnerL = *LoopItr; 213 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL]; 214 assert(InnerAST && "Where is my AST?"); 215 216 // What if InnerLoop was modified by other passes ? 217 CurAST->add(*InnerAST); 218 219 // Once we've incorporated the inner loop's AST into ours, we don't need the 220 // subloop's anymore. 221 delete InnerAST; 222 LoopToAliasSetMap.erase(InnerL); 223 } 224 225 CurLoop = L; 226 227 // Get the preheader block to move instructions into... 228 Preheader = L->getLoopPreheader(); 229 230 // Loop over the body of this loop, looking for calls, invokes, and stores. 231 // Because subloops have already been incorporated into AST, we skip blocks in 232 // subloops. 233 // 234 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 235 I != E; ++I) { 236 BasicBlock *BB = *I; 237 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops. 238 CurAST->add(*BB); // Incorporate the specified basic block 239 } 240 241 // We want to visit all of the instructions in this loop... that are not parts 242 // of our subloops (they have already had their invariants hoisted out of 243 // their loop, into this loop, so there is no need to process the BODIES of 244 // the subloops). 245 // 246 // Traverse the body of the loop in depth first order on the dominator tree so 247 // that we are guaranteed to see definitions before we see uses. This allows 248 // us to sink instructions in one pass, without iteration. After sinking 249 // instructions, we perform another pass to hoist them out of the loop. 250 // 251 if (L->hasDedicatedExits()) 252 SinkRegion(DT->getNode(L->getHeader())); 253 if (Preheader) 254 HoistRegion(DT->getNode(L->getHeader())); 255 256 // Now that all loop invariants have been removed from the loop, promote any 257 // memory references to scalars that we can. 258 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) { 259 // Loop over all of the alias sets in the tracker object. 260 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 261 I != E; ++I) 262 PromoteAliasSet(*I); 263 } 264 265 // Clear out loops state information for the next iteration 266 CurLoop = 0; 267 Preheader = 0; 268 269 // If this loop is nested inside of another one, save the alias information 270 // for when we process the outer loop. 271 if (L->getParentLoop()) 272 LoopToAliasSetMap[L] = CurAST; 273 else 274 delete CurAST; 275 return Changed; 276 } 277 278 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 279 /// dominated by the specified block, and that are in the current loop) in 280 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit 281 /// uses before definitions, allowing us to sink a loop body in one pass without 282 /// iteration. 283 /// 284 void LICM::SinkRegion(DomTreeNode *N) { 285 assert(N != 0 && "Null dominator tree node?"); 286 BasicBlock *BB = N->getBlock(); 287 288 // If this subregion is not in the top level loop at all, exit. 289 if (!CurLoop->contains(BB)) return; 290 291 // We are processing blocks in reverse dfo, so process children first. 292 const std::vector<DomTreeNode*> &Children = N->getChildren(); 293 for (unsigned i = 0, e = Children.size(); i != e; ++i) 294 SinkRegion(Children[i]); 295 296 // Only need to process the contents of this block if it is not part of a 297 // subloop (which would already have been processed). 298 if (inSubLoop(BB)) return; 299 300 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 301 Instruction &I = *--II; 302 303 // If the instruction is dead, we would try to sink it because it isn't used 304 // in the loop, instead, just delete it. 305 if (isInstructionTriviallyDead(&I)) { 306 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); 307 ++II; 308 CurAST->deleteValue(&I); 309 I.eraseFromParent(); 310 Changed = true; 311 continue; 312 } 313 314 // Check to see if we can sink this instruction to the exit blocks 315 // of the loop. We can do this if the all users of the instruction are 316 // outside of the loop. In this case, it doesn't even matter if the 317 // operands of the instruction are loop invariant. 318 // 319 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 320 ++II; 321 sink(I); 322 } 323 } 324 } 325 326 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks 327 /// dominated by the specified block, and that are in the current loop) in depth 328 /// first order w.r.t the DominatorTree. This allows us to visit definitions 329 /// before uses, allowing us to hoist a loop body in one pass without iteration. 330 /// 331 void LICM::HoistRegion(DomTreeNode *N) { 332 assert(N != 0 && "Null dominator tree node?"); 333 BasicBlock *BB = N->getBlock(); 334 335 // If this subregion is not in the top level loop at all, exit. 336 if (!CurLoop->contains(BB)) return; 337 338 // Only need to process the contents of this block if it is not part of a 339 // subloop (which would already have been processed). 340 if (!inSubLoop(BB)) 341 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 342 Instruction &I = *II++; 343 344 // Try constant folding this instruction. If all the operands are 345 // constants, it is technically hoistable, but it would be better to just 346 // fold it. 347 if (Constant *C = ConstantFoldInstruction(&I, TD, TLI)) { 348 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'); 349 CurAST->copyValue(&I, C); 350 CurAST->deleteValue(&I); 351 I.replaceAllUsesWith(C); 352 I.eraseFromParent(); 353 continue; 354 } 355 356 // Try hoisting the instruction out to the preheader. We can only do this 357 // if all of the operands of the instruction are loop invariant and if it 358 // is safe to hoist the instruction. 359 // 360 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) && 361 isSafeToExecuteUnconditionally(I)) 362 hoist(I); 363 } 364 365 const std::vector<DomTreeNode*> &Children = N->getChildren(); 366 for (unsigned i = 0, e = Children.size(); i != e; ++i) 367 HoistRegion(Children[i]); 368 } 369 370 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 371 /// instruction. 372 /// 373 bool LICM::canSinkOrHoistInst(Instruction &I) { 374 // Loads have extra constraints we have to verify before we can hoist them. 375 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 376 if (!LI->isUnordered()) 377 return false; // Don't hoist volatile/atomic loads! 378 379 // Loads from constant memory are always safe to move, even if they end up 380 // in the same alias set as something that ends up being modified. 381 if (AA->pointsToConstantMemory(LI->getOperand(0))) 382 return true; 383 if (LI->getMetadata("invariant.load")) 384 return true; 385 386 // Don't hoist loads which have may-aliased stores in loop. 387 uint64_t Size = 0; 388 if (LI->getType()->isSized()) 389 Size = AA->getTypeStoreSize(LI->getType()); 390 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, 391 LI->getMetadata(LLVMContext::MD_tbaa)); 392 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 393 // Don't sink or hoist dbg info; it's legal, but not useful. 394 if (isa<DbgInfoIntrinsic>(I)) 395 return false; 396 397 // Handle simple cases by querying alias analysis. 398 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 399 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 400 return true; 401 if (AliasAnalysis::onlyReadsMemory(Behavior)) { 402 // If this call only reads from memory and there are no writes to memory 403 // in the loop, we can hoist or sink the call as appropriate. 404 bool FoundMod = false; 405 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 406 I != E; ++I) { 407 AliasSet &AS = *I; 408 if (!AS.isForwardingAliasSet() && AS.isMod()) { 409 FoundMod = true; 410 break; 411 } 412 } 413 if (!FoundMod) return true; 414 } 415 416 // FIXME: This should use mod/ref information to see if we can hoist or sink 417 // the call. 418 419 return false; 420 } 421 422 // Otherwise these instructions are hoistable/sinkable 423 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 424 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 425 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 426 isa<ShuffleVectorInst>(I); 427 } 428 429 /// isNotUsedInLoop - Return true if the only users of this instruction are 430 /// outside of the loop. If this is true, we can sink the instruction to the 431 /// exit blocks of the loop. 432 /// 433 bool LICM::isNotUsedInLoop(Instruction &I) { 434 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 435 Instruction *User = cast<Instruction>(*UI); 436 if (PHINode *PN = dyn_cast<PHINode>(User)) { 437 // PHI node uses occur in predecessor blocks! 438 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 439 if (PN->getIncomingValue(i) == &I) 440 if (CurLoop->contains(PN->getIncomingBlock(i))) 441 return false; 442 } else if (CurLoop->contains(User)) { 443 return false; 444 } 445 } 446 return true; 447 } 448 449 450 /// sink - When an instruction is found to only be used outside of the loop, 451 /// this function moves it to the exit blocks and patches up SSA form as needed. 452 /// This method is guaranteed to remove the original instruction from its 453 /// position, and may either delete it or move it to outside of the loop. 454 /// 455 void LICM::sink(Instruction &I) { 456 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 457 458 SmallVector<BasicBlock*, 8> ExitBlocks; 459 CurLoop->getUniqueExitBlocks(ExitBlocks); 460 461 if (isa<LoadInst>(I)) ++NumMovedLoads; 462 else if (isa<CallInst>(I)) ++NumMovedCalls; 463 ++NumSunk; 464 Changed = true; 465 466 // The case where there is only a single exit node of this loop is common 467 // enough that we handle it as a special (more efficient) case. It is more 468 // efficient to handle because there are no PHI nodes that need to be placed. 469 if (ExitBlocks.size() == 1) { 470 if (!DT->dominates(I.getParent(), ExitBlocks[0])) { 471 // Instruction is not used, just delete it. 472 CurAST->deleteValue(&I); 473 // If I has users in unreachable blocks, eliminate. 474 // If I is not void type then replaceAllUsesWith undef. 475 // This allows ValueHandlers and custom metadata to adjust itself. 476 if (!I.use_empty()) 477 I.replaceAllUsesWith(UndefValue::get(I.getType())); 478 I.eraseFromParent(); 479 } else { 480 // Move the instruction to the start of the exit block, after any PHI 481 // nodes in it. 482 I.moveBefore(ExitBlocks[0]->getFirstInsertionPt()); 483 484 // This instruction is no longer in the AST for the current loop, because 485 // we just sunk it out of the loop. If we just sunk it into an outer 486 // loop, we will rediscover the operation when we process it. 487 CurAST->deleteValue(&I); 488 } 489 return; 490 } 491 492 if (ExitBlocks.empty()) { 493 // The instruction is actually dead if there ARE NO exit blocks. 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 return; 502 } 503 504 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the 505 // hard work of inserting PHI nodes as necessary. 506 SmallVector<PHINode*, 8> NewPHIs; 507 SSAUpdater SSA(&NewPHIs); 508 509 if (!I.use_empty()) 510 SSA.Initialize(I.getType(), I.getName()); 511 512 // Insert a copy of the instruction in each exit block of the loop that is 513 // dominated by the instruction. Each exit block is known to only be in the 514 // ExitBlocks list once. 515 BasicBlock *InstOrigBB = I.getParent(); 516 unsigned NumInserted = 0; 517 518 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 519 BasicBlock *ExitBlock = ExitBlocks[i]; 520 521 if (!DT->dominates(InstOrigBB, ExitBlock)) 522 continue; 523 524 // Insert the code after the last PHI node. 525 BasicBlock::iterator InsertPt = ExitBlock->getFirstInsertionPt(); 526 527 // If this is the first exit block processed, just move the original 528 // instruction, otherwise clone the original instruction and insert 529 // the copy. 530 Instruction *New; 531 if (NumInserted++ == 0) { 532 I.moveBefore(InsertPt); 533 New = &I; 534 } else { 535 New = I.clone(); 536 if (!I.getName().empty()) 537 New->setName(I.getName()+".le"); 538 ExitBlock->getInstList().insert(InsertPt, New); 539 } 540 541 // Now that we have inserted the instruction, inform SSAUpdater. 542 if (!I.use_empty()) 543 SSA.AddAvailableValue(ExitBlock, New); 544 } 545 546 // If the instruction doesn't dominate any exit blocks, it must be dead. 547 if (NumInserted == 0) { 548 CurAST->deleteValue(&I); 549 if (!I.use_empty()) 550 I.replaceAllUsesWith(UndefValue::get(I.getType())); 551 I.eraseFromParent(); 552 return; 553 } 554 555 // Next, rewrite uses of the instruction, inserting PHI nodes as needed. 556 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) { 557 // Grab the use before incrementing the iterator. 558 Use &U = UI.getUse(); 559 // Increment the iterator before removing the use from the list. 560 ++UI; 561 SSA.RewriteUseAfterInsertions(U); 562 } 563 564 // Update CurAST for NewPHIs if I had pointer type. 565 if (I.getType()->isPointerTy()) 566 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 567 CurAST->copyValue(&I, NewPHIs[i]); 568 569 // Finally, remove the instruction from CurAST. It is no longer in the loop. 570 CurAST->deleteValue(&I); 571 } 572 573 /// hoist - When an instruction is found to only use loop invariant operands 574 /// that is safe to hoist, this instruction is called to do the dirty work. 575 /// 576 void LICM::hoist(Instruction &I) { 577 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 578 << I << "\n"); 579 580 // Move the new node to the Preheader, before its terminator. 581 I.moveBefore(Preheader->getTerminator()); 582 583 if (isa<LoadInst>(I)) ++NumMovedLoads; 584 else if (isa<CallInst>(I)) ++NumMovedCalls; 585 ++NumHoisted; 586 Changed = true; 587 } 588 589 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 590 /// not a trapping instruction or if it is a trapping instruction and is 591 /// guaranteed to execute. 592 /// 593 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 594 // If it is not a trapping instruction, it is always safe to hoist. 595 if (isSafeToSpeculativelyExecute(&Inst)) 596 return true; 597 598 return isGuaranteedToExecute(Inst); 599 } 600 601 bool LICM::isGuaranteedToExecute(Instruction &Inst) { 602 // Otherwise we have to check to make sure that the instruction dominates all 603 // of the exit blocks. If it doesn't, then there is a path out of the loop 604 // which does not execute this instruction, so we can't hoist it. 605 606 // If the instruction is in the header block for the loop (which is very 607 // common), it is always guaranteed to dominate the exit blocks. Since this 608 // is a common case, and can save some work, check it now. 609 if (Inst.getParent() == CurLoop->getHeader()) 610 return true; 611 612 // Get the exit blocks for the current loop. 613 SmallVector<BasicBlock*, 8> ExitBlocks; 614 CurLoop->getExitBlocks(ExitBlocks); 615 616 // Verify that the block dominates each of the exit blocks of the loop. 617 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 618 if (!DT->dominates(Inst.getParent(), ExitBlocks[i])) 619 return false; 620 621 return true; 622 } 623 624 namespace { 625 class LoopPromoter : public LoadAndStorePromoter { 626 Value *SomePtr; // Designated pointer to store to. 627 SmallPtrSet<Value*, 4> &PointerMustAliases; 628 SmallVectorImpl<BasicBlock*> &LoopExitBlocks; 629 AliasSetTracker &AST; 630 DebugLoc DL; 631 int Alignment; 632 public: 633 LoopPromoter(Value *SP, 634 const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 635 SmallPtrSet<Value*, 4> &PMA, 636 SmallVectorImpl<BasicBlock*> &LEB, AliasSetTracker &ast, 637 DebugLoc dl, int alignment) 638 : LoadAndStorePromoter(Insts, S), SomePtr(SP), 639 PointerMustAliases(PMA), LoopExitBlocks(LEB), AST(ast), DL(dl), 640 Alignment(alignment) {} 641 642 virtual bool isInstInList(Instruction *I, 643 const SmallVectorImpl<Instruction*> &) const { 644 Value *Ptr; 645 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 646 Ptr = LI->getOperand(0); 647 else 648 Ptr = cast<StoreInst>(I)->getPointerOperand(); 649 return PointerMustAliases.count(Ptr); 650 } 651 652 virtual void doExtraRewritesBeforeFinalDeletion() const { 653 // Insert stores after in the loop exit blocks. Each exit block gets a 654 // store of the live-out values that feed them. Since we've already told 655 // the SSA updater about the defs in the loop and the preheader 656 // definition, it is all set and we can start using it. 657 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { 658 BasicBlock *ExitBlock = LoopExitBlocks[i]; 659 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); 660 Instruction *InsertPos = ExitBlock->getFirstInsertionPt(); 661 StoreInst *NewSI = new StoreInst(LiveInValue, SomePtr, InsertPos); 662 NewSI->setAlignment(Alignment); 663 NewSI->setDebugLoc(DL); 664 } 665 } 666 667 virtual void replaceLoadWithValue(LoadInst *LI, Value *V) const { 668 // Update alias analysis. 669 AST.copyValue(LI, V); 670 } 671 virtual void instructionDeleted(Instruction *I) const { 672 AST.deleteValue(I); 673 } 674 }; 675 } // end anon namespace 676 677 /// PromoteAliasSet - Try to promote memory values to scalars by sinking 678 /// stores out of the loop and moving loads to before the loop. We do this by 679 /// looping over the stores in the loop, looking for stores to Must pointers 680 /// which are loop invariant. 681 /// 682 void LICM::PromoteAliasSet(AliasSet &AS) { 683 // We can promote this alias set if it has a store, if it is a "Must" alias 684 // set, if the pointer is loop invariant, and if we are not eliminating any 685 // volatile loads or stores. 686 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 687 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 688 return; 689 690 assert(!AS.empty() && 691 "Must alias set should have at least one pointer element in it!"); 692 Value *SomePtr = AS.begin()->getValue(); 693 694 // It isn't safe to promote a load/store from the loop if the load/store is 695 // conditional. For example, turning: 696 // 697 // for () { if (c) *P += 1; } 698 // 699 // into: 700 // 701 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 702 // 703 // is not safe, because *P may only be valid to access if 'c' is true. 704 // 705 // It is safe to promote P if all uses are direct load/stores and if at 706 // least one is guaranteed to be executed. 707 bool GuaranteedToExecute = false; 708 709 SmallVector<Instruction*, 64> LoopUses; 710 SmallPtrSet<Value*, 4> PointerMustAliases; 711 712 // We start with an alignment of one and try to find instructions that allow 713 // us to prove better alignment. 714 unsigned Alignment = 1; 715 716 // Check that all of the pointers in the alias set have the same type. We 717 // cannot (yet) promote a memory location that is loaded and stored in 718 // different sizes. 719 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) { 720 Value *ASIV = ASI->getValue(); 721 PointerMustAliases.insert(ASIV); 722 723 // Check that all of the pointers in the alias set have the same type. We 724 // cannot (yet) promote a memory location that is loaded and stored in 725 // different sizes. 726 if (SomePtr->getType() != ASIV->getType()) 727 return; 728 729 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end(); 730 UI != UE; ++UI) { 731 // Ignore instructions that are outside the loop. 732 Instruction *Use = dyn_cast<Instruction>(*UI); 733 if (!Use || !CurLoop->contains(Use)) 734 continue; 735 736 // If there is an non-load/store instruction in the loop, we can't promote 737 // it. 738 if (LoadInst *load = dyn_cast<LoadInst>(Use)) { 739 assert(!load->isVolatile() && "AST broken"); 740 if (!load->isSimple()) 741 return; 742 } else if (StoreInst *store = dyn_cast<StoreInst>(Use)) { 743 // Stores *of* the pointer are not interesting, only stores *to* the 744 // pointer. 745 if (Use->getOperand(1) != ASIV) 746 continue; 747 assert(!store->isVolatile() && "AST broken"); 748 if (!store->isSimple()) 749 return; 750 751 // Note that we only check GuaranteedToExecute inside the store case 752 // so that we do not introduce stores where they did not exist before 753 // (which would break the LLVM concurrency model). 754 755 // If the alignment of this instruction allows us to specify a more 756 // restrictive (and performant) alignment and if we are sure this 757 // instruction will be executed, update the alignment. 758 // Larger is better, with the exception of 0 being the best alignment. 759 unsigned InstAlignment = store->getAlignment(); 760 if ((InstAlignment > Alignment || InstAlignment == 0) 761 && (Alignment != 0)) 762 if (isGuaranteedToExecute(*Use)) { 763 GuaranteedToExecute = true; 764 Alignment = InstAlignment; 765 } 766 767 if (!GuaranteedToExecute) 768 GuaranteedToExecute = isGuaranteedToExecute(*Use); 769 770 } else 771 return; // Not a load or store. 772 773 LoopUses.push_back(Use); 774 } 775 } 776 777 // If there isn't a guaranteed-to-execute instruction, we can't promote. 778 if (!GuaranteedToExecute) 779 return; 780 781 // Otherwise, this is safe to promote, lets do it! 782 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n'); 783 Changed = true; 784 ++NumPromoted; 785 786 // Grab a debug location for the inserted loads/stores; given that the 787 // inserted loads/stores have little relation to the original loads/stores, 788 // this code just arbitrarily picks a location from one, since any debug 789 // location is better than none. 790 DebugLoc DL = LoopUses[0]->getDebugLoc(); 791 792 SmallVector<BasicBlock*, 8> ExitBlocks; 793 CurLoop->getUniqueExitBlocks(ExitBlocks); 794 795 // We use the SSAUpdater interface to insert phi nodes as required. 796 SmallVector<PHINode*, 16> NewPHIs; 797 SSAUpdater SSA(&NewPHIs); 798 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, 799 *CurAST, DL, Alignment); 800 801 // Set up the preheader to have a definition of the value. It is the live-out 802 // value from the preheader that uses in the loop will use. 803 LoadInst *PreheaderLoad = 804 new LoadInst(SomePtr, SomePtr->getName()+".promoted", 805 Preheader->getTerminator()); 806 PreheaderLoad->setAlignment(Alignment); 807 PreheaderLoad->setDebugLoc(DL); 808 SSA.AddAvailableValue(Preheader, PreheaderLoad); 809 810 // Rewrite all the loads in the loop and remember all the definitions from 811 // stores in the loop. 812 Promoter.run(LoopUses); 813 814 // If the SSAUpdater didn't use the load in the preheader, just zap it now. 815 if (PreheaderLoad->use_empty()) 816 PreheaderLoad->eraseFromParent(); 817 } 818 819 820 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 821 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 822 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 823 if (!AST) 824 return; 825 826 AST->copyValue(From, To); 827 } 828 829 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 830 /// set. 831 void LICM::deleteAnalysisValue(Value *V, Loop *L) { 832 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 833 if (!AST) 834 return; 835 836 AST->deleteValue(V); 837 } 838