LICM.cpp revision 9c282011e6b7a6e23a3a163ec9c865890449e8dc
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/Analysis/AliasAnalysis.h" 40#include "llvm/Analysis/AliasSetTracker.h" 41#include "llvm/Analysis/ConstantFolding.h" 42#include "llvm/Analysis/LoopInfo.h" 43#include "llvm/Analysis/LoopPass.h" 44#include "llvm/Analysis/Dominators.h" 45#include "llvm/Analysis/ScalarEvolution.h" 46#include "llvm/Transforms/Utils/Local.h" 47#include "llvm/Transforms/Utils/SSAUpdater.h" 48#include "llvm/Support/CFG.h" 49#include "llvm/Support/CommandLine.h" 50#include "llvm/Support/raw_ostream.h" 51#include "llvm/Support/Debug.h" 52#include "llvm/ADT/Statistic.h" 53#include <algorithm> 54using namespace llvm; 55 56STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 57STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 58STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 59STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 60STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 61 62static cl::opt<bool> 63DisablePromotion("disable-licm-promotion", cl::Hidden, 64 cl::desc("Disable memory promotion in LICM pass")); 65 66namespace { 67 struct LICM : public LoopPass { 68 static char ID; // Pass identification, replacement for typeid 69 LICM() : LoopPass(ID) {} 70 71 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 72 73 /// This transformation requires natural loop information & requires that 74 /// loop preheaders be inserted into the CFG... 75 /// 76 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 77 AU.setPreservesCFG(); 78 AU.addRequired<DominatorTree>(); 79 AU.addRequired<LoopInfo>(); 80 AU.addRequiredID(LoopSimplifyID); 81 AU.addRequired<AliasAnalysis>(); 82 AU.addPreserved<AliasAnalysis>(); 83 AU.addPreserved<ScalarEvolution>(); 84 AU.addPreservedID(LoopSimplifyID); 85 } 86 87 bool doFinalization() { 88 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets"); 89 return false; 90 } 91 92 private: 93 AliasAnalysis *AA; // Current AliasAnalysis information 94 LoopInfo *LI; // Current LoopInfo 95 DominatorTree *DT; // Dominator Tree for the current Loop. 96 97 // State that is updated as we process loops. 98 bool Changed; // Set to true when we change anything. 99 BasicBlock *Preheader; // The preheader block of the current loop... 100 Loop *CurLoop; // The current loop we are working on... 101 AliasSetTracker *CurAST; // AliasSet information for the current loop... 102 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap; 103 104 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 105 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 106 107 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 108 /// set. 109 void deleteAnalysisValue(Value *V, Loop *L); 110 111 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 112 /// dominated by the specified block, and that are in the current loop) in 113 /// reverse depth first order w.r.t the DominatorTree. This allows us to 114 /// visit uses before definitions, allowing us to sink a loop body in one 115 /// pass without iteration. 116 /// 117 void SinkRegion(DomTreeNode *N); 118 119 /// HoistRegion - Walk the specified region of the CFG (defined by all 120 /// blocks dominated by the specified block, and that are in the current 121 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 122 /// visit definitions before uses, allowing us to hoist a loop body in one 123 /// pass without iteration. 124 /// 125 void HoistRegion(DomTreeNode *N); 126 127 /// inSubLoop - Little predicate that returns true if the specified basic 128 /// block is in a subloop of the current one, not the current one itself. 129 /// 130 bool inSubLoop(BasicBlock *BB) { 131 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 132 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 133 if ((*I)->contains(BB)) 134 return true; // A subloop actually contains this block! 135 return false; 136 } 137 138 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 139 /// specified exit block of the loop is dominated by the specified block 140 /// that is in the body of the loop. We use these constraints to 141 /// dramatically limit the amount of the dominator tree that needs to be 142 /// searched. 143 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 144 BasicBlock *BlockInLoop) const { 145 // If the block in the loop is the loop header, it must be dominated! 146 BasicBlock *LoopHeader = CurLoop->getHeader(); 147 if (BlockInLoop == LoopHeader) 148 return true; 149 150 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 151 DomTreeNode *IDom = DT->getNode(ExitBlock); 152 153 // Because the exit block is not in the loop, we know we have to get _at 154 // least_ its immediate dominator. 155 IDom = IDom->getIDom(); 156 157 while (IDom && IDom != BlockInLoopNode) { 158 // If we have got to the header of the loop, then the instructions block 159 // did not dominate the exit node, so we can't hoist it. 160 if (IDom->getBlock() == LoopHeader) 161 return false; 162 163 // Get next Immediate Dominator. 164 IDom = IDom->getIDom(); 165 }; 166 167 return true; 168 } 169 170 /// sink - When an instruction is found to only be used outside of the loop, 171 /// this function moves it to the exit blocks and patches up SSA form as 172 /// needed. 173 /// 174 void sink(Instruction &I); 175 176 /// hoist - When an instruction is found to only use loop invariant operands 177 /// that is safe to hoist, this instruction is called to do the dirty work. 178 /// 179 void hoist(Instruction &I); 180 181 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 182 /// is not a trapping instruction or if it is a trapping instruction and is 183 /// guaranteed to execute. 184 /// 185 bool isSafeToExecuteUnconditionally(Instruction &I); 186 187 /// pointerInvalidatedByLoop - Return true if the body of this loop may 188 /// store into the memory location pointed to by V. 189 /// 190 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 191 // Check to see if any of the basic blocks in CurLoop invalidate *V. 192 return CurAST->getAliasSetForPointer(V, Size).isMod(); 193 } 194 195 bool canSinkOrHoistInst(Instruction &I); 196 bool isLoopInvariantInst(Instruction &I); 197 bool isNotUsedInLoop(Instruction &I); 198 199 void PromoteAliasSet(AliasSet &AS); 200 }; 201} 202 203char LICM::ID = 0; 204INITIALIZE_PASS(LICM, "licm", "Loop Invariant Code Motion", false, false); 205 206Pass *llvm::createLICMPass() { return new LICM(); } 207 208/// Hoist expressions out of the specified loop. Note, alias info for inner 209/// loop is not preserved so it is not a good idea to run LICM multiple 210/// times on one loop. 211/// 212bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 213 Changed = false; 214 215 // Get our Loop and Alias Analysis information... 216 LI = &getAnalysis<LoopInfo>(); 217 AA = &getAnalysis<AliasAnalysis>(); 218 DT = &getAnalysis<DominatorTree>(); 219 220 CurAST = new AliasSetTracker(*AA); 221 // Collect Alias info from subloops. 222 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 223 LoopItr != LoopItrE; ++LoopItr) { 224 Loop *InnerL = *LoopItr; 225 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL]; 226 assert(InnerAST && "Where is my AST?"); 227 228 // What if InnerLoop was modified by other passes ? 229 CurAST->add(*InnerAST); 230 231 // Once we've incorporated the inner loop's AST into ours, we don't need the 232 // subloop's anymore. 233 delete InnerAST; 234 LoopToAliasSetMap.erase(InnerL); 235 } 236 237 CurLoop = L; 238 239 // Get the preheader block to move instructions into... 240 Preheader = L->getLoopPreheader(); 241 242 // Loop over the body of this loop, looking for calls, invokes, and stores. 243 // Because subloops have already been incorporated into AST, we skip blocks in 244 // subloops. 245 // 246 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 247 I != E; ++I) { 248 BasicBlock *BB = *I; 249 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops. 250 CurAST->add(*BB); // Incorporate the specified basic block 251 } 252 253 // We want to visit all of the instructions in this loop... that are not parts 254 // of our subloops (they have already had their invariants hoisted out of 255 // their loop, into this loop, so there is no need to process the BODIES of 256 // the subloops). 257 // 258 // Traverse the body of the loop in depth first order on the dominator tree so 259 // that we are guaranteed to see definitions before we see uses. This allows 260 // us to sink instructions in one pass, without iteration. After sinking 261 // instructions, we perform another pass to hoist them out of the loop. 262 // 263 if (L->hasDedicatedExits()) 264 SinkRegion(DT->getNode(L->getHeader())); 265 if (Preheader) 266 HoistRegion(DT->getNode(L->getHeader())); 267 268 // Now that all loop invariants have been removed from the loop, promote any 269 // memory references to scalars that we can. 270 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) { 271 // Loop over all of the alias sets in the tracker object. 272 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 273 I != E; ++I) 274 PromoteAliasSet(*I); 275 } 276 277 // Clear out loops state information for the next iteration 278 CurLoop = 0; 279 Preheader = 0; 280 281 // If this loop is nested inside of another one, save the alias information 282 // for when we process the outer loop. 283 if (L->getParentLoop()) 284 LoopToAliasSetMap[L] = CurAST; 285 else 286 delete CurAST; 287 return Changed; 288} 289 290/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 291/// dominated by the specified block, and that are in the current loop) in 292/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 293/// uses before definitions, allowing us to sink a loop body in one pass without 294/// iteration. 295/// 296void LICM::SinkRegion(DomTreeNode *N) { 297 assert(N != 0 && "Null dominator tree node?"); 298 BasicBlock *BB = N->getBlock(); 299 300 // If this subregion is not in the top level loop at all, exit. 301 if (!CurLoop->contains(BB)) return; 302 303 // We are processing blocks in reverse dfo, so process children first. 304 const std::vector<DomTreeNode*> &Children = N->getChildren(); 305 for (unsigned i = 0, e = Children.size(); i != e; ++i) 306 SinkRegion(Children[i]); 307 308 // Only need to process the contents of this block if it is not part of a 309 // subloop (which would already have been processed). 310 if (inSubLoop(BB)) return; 311 312 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 313 Instruction &I = *--II; 314 315 // If the instruction is dead, we would try to sink it because it isn't used 316 // in the loop, instead, just delete it. 317 if (isInstructionTriviallyDead(&I)) { 318 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); 319 ++II; 320 CurAST->deleteValue(&I); 321 I.eraseFromParent(); 322 Changed = true; 323 continue; 324 } 325 326 // Check to see if we can sink this instruction to the exit blocks 327 // of the loop. We can do this if the all users of the instruction are 328 // outside of the loop. In this case, it doesn't even matter if the 329 // operands of the instruction are loop invariant. 330 // 331 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 332 ++II; 333 sink(I); 334 } 335 } 336} 337 338/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 339/// dominated by the specified block, and that are in the current loop) in depth 340/// first order w.r.t the DominatorTree. This allows us to visit definitions 341/// before uses, allowing us to hoist a loop body in one pass without iteration. 342/// 343void LICM::HoistRegion(DomTreeNode *N) { 344 assert(N != 0 && "Null dominator tree node?"); 345 BasicBlock *BB = N->getBlock(); 346 347 // If this subregion is not in the top level loop at all, exit. 348 if (!CurLoop->contains(BB)) return; 349 350 // Only need to process the contents of this block if it is not part of a 351 // subloop (which would already have been processed). 352 if (!inSubLoop(BB)) 353 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 354 Instruction &I = *II++; 355 356 // Try constant folding this instruction. If all the operands are 357 // constants, it is technically hoistable, but it would be better to just 358 // fold it. 359 if (Constant *C = ConstantFoldInstruction(&I)) { 360 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'); 361 CurAST->copyValue(&I, C); 362 CurAST->deleteValue(&I); 363 I.replaceAllUsesWith(C); 364 I.eraseFromParent(); 365 continue; 366 } 367 368 // Try hoisting the instruction out to the preheader. We can only do this 369 // if all of the operands of the instruction are loop invariant and if it 370 // is safe to hoist the instruction. 371 // 372 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 373 isSafeToExecuteUnconditionally(I)) 374 hoist(I); 375 } 376 377 const std::vector<DomTreeNode*> &Children = N->getChildren(); 378 for (unsigned i = 0, e = Children.size(); i != e; ++i) 379 HoistRegion(Children[i]); 380} 381 382/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 383/// instruction. 384/// 385bool LICM::canSinkOrHoistInst(Instruction &I) { 386 // Loads have extra constraints we have to verify before we can hoist them. 387 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 388 if (LI->isVolatile()) 389 return false; // Don't hoist volatile loads! 390 391 // Loads from constant memory are always safe to move, even if they end up 392 // in the same alias set as something that ends up being modified. 393 if (AA->pointsToConstantMemory(LI->getOperand(0))) 394 return true; 395 396 // Don't hoist loads which have may-aliased stores in loop. 397 unsigned Size = 0; 398 if (LI->getType()->isSized()) 399 Size = AA->getTypeStoreSize(LI->getType()); 400 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 401 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 402 // Handle obvious cases efficiently. 403 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 404 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 405 return true; 406 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 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/// isLoopInvariantInst - Return true if all operands of this instruction are 456/// loop invariant. We also filter out non-hoistable instructions here just for 457/// efficiency. 458/// 459bool LICM::isLoopInvariantInst(Instruction &I) { 460 // The instruction is loop invariant if all of its operands are loop-invariant 461 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 462 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 463 return false; 464 465 // If we got this far, the instruction is loop invariant! 466 return true; 467} 468 469/// sink - When an instruction is found to only be used outside of the loop, 470/// this function moves it to the exit blocks and patches up SSA form as needed. 471/// This method is guaranteed to remove the original instruction from its 472/// position, and may either delete it or move it to outside of the loop. 473/// 474void LICM::sink(Instruction &I) { 475 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 476 477 SmallVector<BasicBlock*, 8> ExitBlocks; 478 CurLoop->getUniqueExitBlocks(ExitBlocks); 479 480 if (isa<LoadInst>(I)) ++NumMovedLoads; 481 else if (isa<CallInst>(I)) ++NumMovedCalls; 482 ++NumSunk; 483 Changed = true; 484 485 // The case where there is only a single exit node of this loop is common 486 // enough that we handle it as a special (more efficient) case. It is more 487 // efficient to handle because there are no PHI nodes that need to be placed. 488 if (ExitBlocks.size() == 1) { 489 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 490 // Instruction is not used, just delete it. 491 CurAST->deleteValue(&I); 492 // If I has users in unreachable blocks, eliminate. 493 // If I is not void type then replaceAllUsesWith undef. 494 // This allows ValueHandlers and custom metadata to adjust itself. 495 if (!I.use_empty()) 496 I.replaceAllUsesWith(UndefValue::get(I.getType())); 497 I.eraseFromParent(); 498 } else { 499 // Move the instruction to the start of the exit block, after any PHI 500 // nodes in it. 501 I.moveBefore(ExitBlocks[0]->getFirstNonPHI()); 502 503 // This instruction is no longer in the AST for the current loop, because 504 // we just sunk it out of the loop. If we just sunk it into an outer 505 // loop, we will rediscover the operation when we process it. 506 CurAST->deleteValue(&I); 507 } 508 return; 509 } 510 511 if (ExitBlocks.empty()) { 512 // The instruction is actually dead if there ARE NO exit blocks. 513 CurAST->deleteValue(&I); 514 // If I has users in unreachable blocks, eliminate. 515 // If I is not void type then replaceAllUsesWith undef. 516 // This allows ValueHandlers and custom metadata to adjust itself. 517 if (!I.use_empty()) 518 I.replaceAllUsesWith(UndefValue::get(I.getType())); 519 I.eraseFromParent(); 520 return; 521 } 522 523 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the 524 // hard work of inserting PHI nodes as necessary. 525 SmallVector<PHINode*, 8> NewPHIs; 526 SSAUpdater SSA(&NewPHIs); 527 528 if (!I.use_empty()) 529 SSA.Initialize(I.getType(), I.getName()); 530 531 // Insert a copy of the instruction in each exit block of the loop that is 532 // dominated by the instruction. Each exit block is known to only be in the 533 // ExitBlocks list once. 534 BasicBlock *InstOrigBB = I.getParent(); 535 unsigned NumInserted = 0; 536 537 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 538 BasicBlock *ExitBlock = ExitBlocks[i]; 539 540 if (!isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) 541 continue; 542 543 // Insert the code after the last PHI node. 544 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 545 546 // If this is the first exit block processed, just move the original 547 // instruction, otherwise clone the original instruction and insert 548 // the copy. 549 Instruction *New; 550 if (NumInserted++ == 0) { 551 I.moveBefore(InsertPt); 552 New = &I; 553 } else { 554 New = I.clone(); 555 if (!I.getName().empty()) 556 New->setName(I.getName()+".le"); 557 ExitBlock->getInstList().insert(InsertPt, New); 558 } 559 560 // Now that we have inserted the instruction, inform SSAUpdater. 561 if (!I.use_empty()) 562 SSA.AddAvailableValue(ExitBlock, New); 563 } 564 565 // If the instruction doesn't dominate any exit blocks, it must be dead. 566 if (NumInserted == 0) { 567 CurAST->deleteValue(&I); 568 if (!I.use_empty()) 569 I.replaceAllUsesWith(UndefValue::get(I.getType())); 570 I.eraseFromParent(); 571 return; 572 } 573 574 // Next, rewrite uses of the instruction, inserting PHI nodes as needed. 575 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) { 576 // Grab the use before incrementing the iterator. 577 Use &U = UI.getUse(); 578 // Increment the iterator before removing the use from the list. 579 ++UI; 580 SSA.RewriteUseAfterInsertions(U); 581 } 582 583 // Update CurAST for NewPHIs if I had pointer type. 584 if (I.getType()->isPointerTy()) 585 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 586 CurAST->copyValue(&I, NewPHIs[i]); 587 588 // Finally, remove the instruction from CurAST. It is no longer in the loop. 589 CurAST->deleteValue(&I); 590} 591 592/// hoist - When an instruction is found to only use loop invariant operands 593/// that is safe to hoist, this instruction is called to do the dirty work. 594/// 595void LICM::hoist(Instruction &I) { 596 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 597 << I << "\n"); 598 599 // Move the new node to the Preheader, before its terminator. 600 I.moveBefore(Preheader->getTerminator()); 601 602 if (isa<LoadInst>(I)) ++NumMovedLoads; 603 else if (isa<CallInst>(I)) ++NumMovedCalls; 604 ++NumHoisted; 605 Changed = true; 606} 607 608/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 609/// not a trapping instruction or if it is a trapping instruction and is 610/// guaranteed to execute. 611/// 612bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 613 // If it is not a trapping instruction, it is always safe to hoist. 614 if (Inst.isSafeToSpeculativelyExecute()) 615 return true; 616 617 // Otherwise we have to check to make sure that the instruction dominates all 618 // of the exit blocks. If it doesn't, then there is a path out of the loop 619 // which does not execute this instruction, so we can't hoist it. 620 621 // If the instruction is in the header block for the loop (which is very 622 // common), it is always guaranteed to dominate the exit blocks. Since this 623 // is a common case, and can save some work, check it now. 624 if (Inst.getParent() == CurLoop->getHeader()) 625 return true; 626 627 // Get the exit blocks for the current loop. 628 SmallVector<BasicBlock*, 8> ExitBlocks; 629 CurLoop->getExitBlocks(ExitBlocks); 630 631 // For each exit block, get the DT node and walk up the DT until the 632 // instruction's basic block is found or we exit the loop. 633 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 634 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 635 return false; 636 637 return true; 638} 639 640/// PromoteAliasSet - Try to promote memory values to scalars by sinking 641/// stores out of the loop and moving loads to before the loop. We do this by 642/// looping over the stores in the loop, looking for stores to Must pointers 643/// which are loop invariant. 644/// 645void LICM::PromoteAliasSet(AliasSet &AS) { 646 // We can promote this alias set if it has a store, if it is a "Must" alias 647 // set, if the pointer is loop invariant, and if we are not eliminating any 648 // volatile loads or stores. 649 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 650 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 651 return; 652 653 assert(!AS.empty() && 654 "Must alias set should have at least one pointer element in it!"); 655 Value *SomePtr = AS.begin()->getValue(); 656 657 // It isn't safe to promote a load/store from the loop if the load/store is 658 // conditional. For example, turning: 659 // 660 // for () { if (c) *P += 1; } 661 // 662 // into: 663 // 664 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 665 // 666 // is not safe, because *P may only be valid to access if 'c' is true. 667 // 668 // It is safe to promote P if all uses are direct load/stores and if at 669 // least one is guaranteed to be executed. 670 bool GuaranteedToExecute = false; 671 672 SmallVector<Instruction*, 64> LoopUses; 673 SmallPtrSet<Value*, 4> PointerMustAliases; 674 675 // Check that all of the pointers in the alias set have the same type. We 676 // cannot (yet) promote a memory location that is loaded and stored in 677 // different sizes. 678 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) { 679 Value *ASIV = ASI->getValue(); 680 PointerMustAliases.insert(ASIV); 681 682 // Check that all of the pointers in the alias set have the same type. We 683 // cannot (yet) promote a memory location that is loaded and stored in 684 // different sizes. 685 if (SomePtr->getType() != ASIV->getType()) 686 return; 687 688 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end(); 689 UI != UE; ++UI) { 690 // Ignore instructions that are outside the loop. 691 Instruction *Use = dyn_cast<Instruction>(*UI); 692 if (!Use || !CurLoop->contains(Use)) 693 continue; 694 695 // If there is an non-load/store instruction in the loop, we can't promote 696 // it. 697 if (isa<LoadInst>(Use)) 698 assert(!cast<LoadInst>(Use)->isVolatile() && "AST broken"); 699 else if (isa<StoreInst>(Use)) 700 assert(!cast<StoreInst>(Use)->isVolatile() && 701 Use->getOperand(0) != ASIV && "AST broken"); 702 else 703 return; // Not a load or store. 704 705 if (!GuaranteedToExecute) 706 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 707 708 LoopUses.push_back(Use); 709 } 710 } 711 712 // If there isn't a guaranteed-to-execute instruction, we can't promote. 713 if (!GuaranteedToExecute) 714 return; 715 716 // Otherwise, this is safe to promote, lets do it! 717 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n'); 718 Changed = true; 719 ++NumPromoted; 720 721 // We use the SSAUpdater interface to insert phi nodes as required. 722 SmallVector<PHINode*, 16> NewPHIs; 723 SSAUpdater SSA(&NewPHIs); 724 725 // It wants to know some value of the same type as what we'll be inserting. 726 Value *SomeValue; 727 if (isa<LoadInst>(LoopUses[0])) 728 SomeValue = LoopUses[0]; 729 else 730 SomeValue = cast<StoreInst>(LoopUses[0])->getOperand(0); 731 SSA.Initialize(SomeValue->getType(), SomeValue->getName()); 732 733 // First step: bucket up uses of the pointers by the block they occur in. 734 // This is important because we have to handle multiple defs/uses in a block 735 // ourselves: SSAUpdater is purely for cross-block references. 736 // FIXME: Want a TinyVector<Instruction*> since there is usually 0/1 element. 737 DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock; 738 for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) { 739 Instruction *User = LoopUses[i]; 740 UsesByBlock[User->getParent()].push_back(User); 741 } 742 743 // Okay, now we can iterate over all the blocks in the loop with uses, 744 // processing them. Keep track of which loads are loading a live-in value. 745 SmallVector<LoadInst*, 32> LiveInLoads; 746 747 for (unsigned LoopUse = 0, e = LoopUses.size(); LoopUse != e; ++LoopUse) { 748 Instruction *User = LoopUses[LoopUse]; 749 std::vector<Instruction*> &BlockUses = UsesByBlock[User->getParent()]; 750 751 // If this block has already been processed, ignore this repeat use. 752 if (BlockUses.empty()) continue; 753 754 // Okay, this is the first use in the block. If this block just has a 755 // single user in it, we can rewrite it trivially. 756 if (BlockUses.size() == 1) { 757 // If it is a store, it is a trivial def of the value in the block. 758 if (isa<StoreInst>(User)) { 759 SSA.AddAvailableValue(User->getParent(), 760 cast<StoreInst>(User)->getOperand(0)); 761 } else { 762 // Otherwise it is a load, queue it to rewrite as a live-in load. 763 LiveInLoads.push_back(cast<LoadInst>(User)); 764 } 765 BlockUses.clear(); 766 continue; 767 } 768 769 // Otherwise, check to see if this block is all loads. If so, we can queue 770 // them all as live in loads. 771 bool HasStore = false; 772 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) { 773 if (isa<StoreInst>(BlockUses[i])) { 774 HasStore = true; 775 break; 776 } 777 } 778 779 if (!HasStore) { 780 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) 781 LiveInLoads.push_back(cast<LoadInst>(BlockUses[i])); 782 BlockUses.clear(); 783 continue; 784 } 785 786 // Otherwise, we have mixed loads and stores (or just a bunch of stores). 787 // Since SSAUpdater is purely for cross-block values, we need to determine 788 // the order of these instructions in the block. If the first use in the 789 // block is a load, then it uses the live in value. The last store defines 790 // the live out value. We handle this by doing a linear scan of the block. 791 BasicBlock *BB = User->getParent(); 792 Value *StoredValue = 0; 793 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 794 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 795 // If this is a load to an unrelated pointer, ignore it. 796 if (!PointerMustAliases.count(L->getOperand(0))) continue; 797 798 // If we haven't seen a store yet, this is a live in use, otherwise 799 // use the stored value. 800 if (StoredValue) 801 L->replaceAllUsesWith(StoredValue); 802 else 803 LiveInLoads.push_back(L); 804 continue; 805 } 806 807 if (StoreInst *S = dyn_cast<StoreInst>(II)) { 808 // If this is a store to an unrelated pointer, ignore it. 809 if (!PointerMustAliases.count(S->getOperand(1))) continue; 810 811 // Remember that this is the active value in the block. 812 StoredValue = S->getOperand(0); 813 } 814 } 815 816 // The last stored value that happened is the live-out for the block. 817 assert(StoredValue && "Already checked that there is a store in block"); 818 SSA.AddAvailableValue(BB, StoredValue); 819 BlockUses.clear(); 820 } 821 822 // Now that all the intra-loop values are classified, set up the preheader. 823 // It gets a load of the pointer we're promoting, and it is the live-out value 824 // from the preheader. 825 LoadInst *PreheaderLoad = new LoadInst(SomePtr,SomePtr->getName()+".promoted", 826 Preheader->getTerminator()); 827 SSA.AddAvailableValue(Preheader, PreheaderLoad); 828 829 // Now that the preheader is good to go, set up the exit blocks. Each exit 830 // block gets a store of the live-out values that feed them. Since we've 831 // already told the SSA updater about the defs in the loop and the preheader 832 // definition, it is all set and we can start using it. 833 SmallVector<BasicBlock*, 8> ExitBlocks; 834 CurLoop->getUniqueExitBlocks(ExitBlocks); 835 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 836 BasicBlock *ExitBlock = ExitBlocks[i]; 837 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); 838 Instruction *InsertPos = ExitBlock->getFirstNonPHI(); 839 new StoreInst(LiveInValue, SomePtr, InsertPos); 840 } 841 842 // Okay, now we rewrite all loads that use live-in values in the loop, 843 // inserting PHI nodes as necessary. 844 for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) { 845 LoadInst *ALoad = LiveInLoads[i]; 846 Value *NewVal = SSA.GetValueInMiddleOfBlock(ALoad->getParent()); 847 ALoad->replaceAllUsesWith(NewVal); 848 CurAST->copyValue(ALoad, NewVal); 849 } 850 851 // Now that everything is rewritten, delete the old instructions from the body 852 // of the loop. They should all be dead now. 853 for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) { 854 Instruction *User = LoopUses[i]; 855 CurAST->deleteValue(User); 856 User->eraseFromParent(); 857 } 858 859 // If the preheader load is itself a pointer, we need to tell alias analysis 860 // about the new pointer we created in the preheader block and about any PHI 861 // nodes that just got inserted. 862 if (PreheaderLoad->getType()->isPointerTy()) { 863 // Copy any value stored to or loaded from a must-alias of the pointer. 864 CurAST->copyValue(SomeValue, PreheaderLoad); 865 866 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 867 CurAST->copyValue(SomeValue, NewPHIs[i]); 868 } 869 870 // fwew, we're done! 871} 872 873 874/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 875void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 876 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 877 if (!AST) 878 return; 879 880 AST->copyValue(From, To); 881} 882 883/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 884/// set. 885void LICM::deleteAnalysisValue(Value *V, Loop *L) { 886 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 887 if (!AST) 888 return; 889 890 AST->deleteValue(V); 891} 892