LICM.cpp revision 1bf5ebc7be0d5b05e4175c7adb767b38896adef1
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 mem2reg functionality to construct the appropriate SSA form for the 30// variable. 31// 32//===----------------------------------------------------------------------===// 33 34#define DEBUG_TYPE "licm" 35#include "llvm/Transforms/Scalar.h" 36#include "llvm/Constants.h" 37#include "llvm/DerivedTypes.h" 38#include "llvm/IntrinsicInst.h" 39#include "llvm/Instructions.h" 40#include "llvm/Target/TargetData.h" 41#include "llvm/Analysis/LoopInfo.h" 42#include "llvm/Analysis/LoopPass.h" 43#include "llvm/Analysis/AliasAnalysis.h" 44#include "llvm/Analysis/AliasSetTracker.h" 45#include "llvm/Analysis/Dominators.h" 46#include "llvm/Analysis/ScalarEvolution.h" 47#include "llvm/Transforms/Utils/PromoteMemToReg.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 66// This feature is currently disabled by default because CodeGen is not yet 67// capable of rematerializing these constants in PIC mode, so it can lead to 68// degraded performance. Compile test/CodeGen/X86/remat-constant.ll with 69// -relocation-model=pic to see an example of this. 70static cl::opt<bool> 71EnableLICMConstantMotion("enable-licm-constant-variables", cl::Hidden, 72 cl::desc("Enable hoisting/sinking of constant " 73 "global variables")); 74 75namespace { 76 struct LICM : public LoopPass { 77 static char ID; // Pass identification, replacement for typeid 78 LICM() : LoopPass(&ID) {} 79 80 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 81 82 /// This transformation requires natural loop information & requires that 83 /// loop preheaders be inserted into the CFG... 84 /// 85 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 86 AU.setPreservesCFG(); 87 AU.addRequiredID(LoopSimplifyID); 88 AU.addRequired<LoopInfo>(); 89 AU.addRequired<DominatorTree>(); 90 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 91 AU.addRequired<AliasAnalysis>(); 92 AU.addPreserved<ScalarEvolution>(); 93 AU.addPreserved<DominanceFrontier>(); 94 AU.addPreservedID(LoopSimplifyID); 95 } 96 97 bool doFinalization() { 98 // Free the values stored in the map 99 for (std::map<Loop *, AliasSetTracker *>::iterator 100 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I) 101 delete I->second; 102 103 LoopToAliasMap.clear(); 104 return false; 105 } 106 107 private: 108 // Various analyses that we use... 109 AliasAnalysis *AA; // Current AliasAnalysis information 110 LoopInfo *LI; // Current LoopInfo 111 DominatorTree *DT; // Dominator Tree for the current Loop... 112 DominanceFrontier *DF; // Current Dominance Frontier 113 114 // State that is updated as we process loops 115 bool Changed; // Set to true when we change anything. 116 BasicBlock *Preheader; // The preheader block of the current loop... 117 Loop *CurLoop; // The current loop we are working on... 118 AliasSetTracker *CurAST; // AliasSet information for the current loop... 119 std::map<Loop *, AliasSetTracker *> LoopToAliasMap; 120 121 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 122 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 123 124 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 125 /// set. 126 void deleteAnalysisValue(Value *V, Loop *L); 127 128 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 129 /// dominated by the specified block, and that are in the current loop) in 130 /// reverse depth first order w.r.t the DominatorTree. This allows us to 131 /// visit uses before definitions, allowing us to sink a loop body in one 132 /// pass without iteration. 133 /// 134 void SinkRegion(DomTreeNode *N); 135 136 /// HoistRegion - Walk the specified region of the CFG (defined by all 137 /// blocks dominated by the specified block, and that are in the current 138 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 139 /// visit definitions before uses, allowing us to hoist a loop body in one 140 /// pass without iteration. 141 /// 142 void HoistRegion(DomTreeNode *N); 143 144 /// inSubLoop - Little predicate that returns true if the specified basic 145 /// block is in a subloop of the current one, not the current one itself. 146 /// 147 bool inSubLoop(BasicBlock *BB) { 148 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 149 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 150 if ((*I)->contains(BB)) 151 return true; // A subloop actually contains this block! 152 return false; 153 } 154 155 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 156 /// specified exit block of the loop is dominated by the specified block 157 /// that is in the body of the loop. We use these constraints to 158 /// dramatically limit the amount of the dominator tree that needs to be 159 /// searched. 160 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 161 BasicBlock *BlockInLoop) const { 162 // If the block in the loop is the loop header, it must be dominated! 163 BasicBlock *LoopHeader = CurLoop->getHeader(); 164 if (BlockInLoop == LoopHeader) 165 return true; 166 167 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 168 DomTreeNode *IDom = DT->getNode(ExitBlock); 169 170 // Because the exit block is not in the loop, we know we have to get _at 171 // least_ its immediate dominator. 172 do { 173 // Get next Immediate Dominator. 174 IDom = IDom->getIDom(); 175 176 // If we have got to the header of the loop, then the instructions block 177 // did not dominate the exit node, so we can't hoist it. 178 if (IDom->getBlock() == LoopHeader) 179 return false; 180 181 } while (IDom != BlockInLoopNode); 182 183 return true; 184 } 185 186 /// sink - When an instruction is found to only be used outside of the loop, 187 /// this function moves it to the exit blocks and patches up SSA form as 188 /// needed. 189 /// 190 void sink(Instruction &I); 191 192 /// hoist - When an instruction is found to only use loop invariant operands 193 /// that is safe to hoist, this instruction is called to do the dirty work. 194 /// 195 void hoist(Instruction &I); 196 197 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 198 /// is not a trapping instruction or if it is a trapping instruction and is 199 /// guaranteed to execute. 200 /// 201 bool isSafeToExecuteUnconditionally(Instruction &I); 202 203 /// pointerInvalidatedByLoop - Return true if the body of this loop may 204 /// store into the memory location pointed to by V. 205 /// 206 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 207 // Check to see if any of the basic blocks in CurLoop invalidate *V. 208 return CurAST->getAliasSetForPointer(V, Size).isMod(); 209 } 210 211 bool canSinkOrHoistInst(Instruction &I); 212 bool isLoopInvariantInst(Instruction &I); 213 bool isNotUsedInLoop(Instruction &I); 214 215 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 216 /// to scalars as we can. 217 /// 218 void PromoteValuesInLoop(); 219 220 /// FindPromotableValuesInLoop - Check the current loop for stores to 221 /// definite pointers, which are not loaded and stored through may aliases. 222 /// If these are found, create an alloca for the value, add it to the 223 /// PromotedValues list, and keep track of the mapping from value to 224 /// alloca... 225 /// 226 void FindPromotableValuesInLoop( 227 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 228 std::map<Value*, AllocaInst*> &Val2AlMap); 229 }; 230} 231 232char LICM::ID = 0; 233static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 234 235Pass *llvm::createLICMPass() { return new LICM(); } 236 237/// Hoist expressions out of the specified loop. Note, alias info for inner 238/// loop is not preserved so it is not a good idea to run LICM multiple 239/// times on one loop. 240/// 241bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 242 Changed = false; 243 244 // Get our Loop and Alias Analysis information... 245 LI = &getAnalysis<LoopInfo>(); 246 AA = &getAnalysis<AliasAnalysis>(); 247 DF = &getAnalysis<DominanceFrontier>(); 248 DT = &getAnalysis<DominatorTree>(); 249 250 CurAST = new AliasSetTracker(*AA); 251 // Collect Alias info from subloops 252 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 253 LoopItr != LoopItrE; ++LoopItr) { 254 Loop *InnerL = *LoopItr; 255 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 256 assert (InnerAST && "Where is my AST?"); 257 258 // What if InnerLoop was modified by other passes ? 259 CurAST->add(*InnerAST); 260 } 261 262 CurLoop = L; 263 264 // Get the preheader block to move instructions into... 265 Preheader = L->getLoopPreheader(); 266 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 267 268 // Loop over the body of this loop, looking for calls, invokes, and stores. 269 // Because subloops have already been incorporated into AST, we skip blocks in 270 // subloops. 271 // 272 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 273 I != E; ++I) { 274 BasicBlock *BB = *I; 275 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops... 276 CurAST->add(*BB); // Incorporate the specified basic block 277 } 278 279 // We want to visit all of the instructions in this loop... that are not parts 280 // of our subloops (they have already had their invariants hoisted out of 281 // their loop, into this loop, so there is no need to process the BODIES of 282 // the subloops). 283 // 284 // Traverse the body of the loop in depth first order on the dominator tree so 285 // that we are guaranteed to see definitions before we see uses. This allows 286 // us to sink instructions in one pass, without iteration. After sinking 287 // instructions, we perform another pass to hoist them out of the loop. 288 // 289 SinkRegion(DT->getNode(L->getHeader())); 290 HoistRegion(DT->getNode(L->getHeader())); 291 292 // Now that all loop invariants have been removed from the loop, promote any 293 // memory references to scalars that we can... 294 if (!DisablePromotion) 295 PromoteValuesInLoop(); 296 297 // Clear out loops state information for the next iteration 298 CurLoop = 0; 299 Preheader = 0; 300 301 LoopToAliasMap[L] = CurAST; 302 return Changed; 303} 304 305/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 306/// dominated by the specified block, and that are in the current loop) in 307/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 308/// uses before definitions, allowing us to sink a loop body in one pass without 309/// iteration. 310/// 311void LICM::SinkRegion(DomTreeNode *N) { 312 assert(N != 0 && "Null dominator tree node?"); 313 BasicBlock *BB = N->getBlock(); 314 315 // If this subregion is not in the top level loop at all, exit. 316 if (!CurLoop->contains(BB)) return; 317 318 // We are processing blocks in reverse dfo, so process children first... 319 const std::vector<DomTreeNode*> &Children = N->getChildren(); 320 for (unsigned i = 0, e = Children.size(); i != e; ++i) 321 SinkRegion(Children[i]); 322 323 // Only need to process the contents of this block if it is not part of a 324 // subloop (which would already have been processed). 325 if (inSubLoop(BB)) return; 326 327 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 328 Instruction &I = *--II; 329 330 // Check to see if we can sink this instruction to the exit blocks 331 // of the loop. We can do this if the all users of the instruction are 332 // outside of the loop. In this case, it doesn't even matter if the 333 // operands of the instruction are loop invariant. 334 // 335 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 336 ++II; 337 sink(I); 338 } 339 } 340} 341 342/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 343/// dominated by the specified block, and that are in the current loop) in depth 344/// first order w.r.t the DominatorTree. This allows us to visit definitions 345/// before uses, allowing us to hoist a loop body in one pass without iteration. 346/// 347void LICM::HoistRegion(DomTreeNode *N) { 348 assert(N != 0 && "Null dominator tree node?"); 349 BasicBlock *BB = N->getBlock(); 350 351 // If this subregion is not in the top level loop at all, exit. 352 if (!CurLoop->contains(BB)) return; 353 354 // Only need to process the contents of this block if it is not part of a 355 // subloop (which would already have been processed). 356 if (!inSubLoop(BB)) 357 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 358 Instruction &I = *II++; 359 360 // Try hoisting the instruction out to the preheader. We can only do this 361 // if all of the operands of the instruction are loop invariant and if it 362 // is safe to hoist the instruction. 363 // 364 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 365 isSafeToExecuteUnconditionally(I)) 366 hoist(I); 367 } 368 369 const std::vector<DomTreeNode*> &Children = N->getChildren(); 370 for (unsigned i = 0, e = Children.size(); i != e; ++i) 371 HoistRegion(Children[i]); 372} 373 374/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 375/// instruction. 376/// 377bool LICM::canSinkOrHoistInst(Instruction &I) { 378 // Loads have extra constraints we have to verify before we can hoist them. 379 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 380 if (LI->isVolatile()) 381 return false; // Don't hoist volatile loads! 382 383 // Loads from constant memory are always safe to move, even if they end up 384 // in the same alias set as something that ends up being modified. 385 if (EnableLICMConstantMotion && 386 AA->pointsToConstantMemory(LI->getOperand(0))) 387 return true; 388 389 // Don't hoist loads which have may-aliased stores in loop. 390 unsigned Size = 0; 391 if (LI->getType()->isSized()) 392 Size = AA->getTypeStoreSize(LI->getType()); 393 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 394 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 395 if (isa<DbgStopPointInst>(CI)) { 396 // Don't hoist/sink dbgstoppoints, we handle them separately 397 return false; 398 } 399 // Handle obvious cases efficiently. 400 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 401 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 402 return true; 403 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 404 // If this call only reads from memory and there are no writes to memory 405 // in the loop, we can hoist or sink the call as appropriate. 406 bool FoundMod = false; 407 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 408 I != E; ++I) { 409 AliasSet &AS = *I; 410 if (!AS.isForwardingAliasSet() && AS.isMod()) { 411 FoundMod = true; 412 break; 413 } 414 } 415 if (!FoundMod) return true; 416 } 417 418 // FIXME: This should use mod/ref information to see if we can hoist or sink 419 // the call. 420 421 return false; 422 } 423 424 // Otherwise these instructions are hoistable/sinkable 425 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 426 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 427 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 428 isa<ShuffleVectorInst>(I); 429} 430 431/// isNotUsedInLoop - Return true if the only users of this instruction are 432/// outside of the loop. If this is true, we can sink the instruction to the 433/// exit blocks of the loop. 434/// 435bool LICM::isNotUsedInLoop(Instruction &I) { 436 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 437 Instruction *User = cast<Instruction>(*UI); 438 if (PHINode *PN = dyn_cast<PHINode>(User)) { 439 // PHI node uses occur in predecessor blocks! 440 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 441 if (PN->getIncomingValue(i) == &I) 442 if (CurLoop->contains(PN->getIncomingBlock(i))) 443 return false; 444 } else if (CurLoop->contains(User->getParent())) { 445 return false; 446 } 447 } 448 return true; 449} 450 451 452/// isLoopInvariantInst - Return true if all operands of this instruction are 453/// loop invariant. We also filter out non-hoistable instructions here just for 454/// efficiency. 455/// 456bool LICM::isLoopInvariantInst(Instruction &I) { 457 // The instruction is loop invariant if all of its operands are loop-invariant 458 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 459 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 460 return false; 461 462 // If we got this far, the instruction is loop invariant! 463 return true; 464} 465 466/// sink - When an instruction is found to only be used outside of the loop, 467/// this function moves it to the exit blocks and patches up SSA form as needed. 468/// This method is guaranteed to remove the original instruction from its 469/// position, and may either delete it or move it to outside of the loop. 470/// 471void LICM::sink(Instruction &I) { 472 DEBUG(errs() << "LICM sinking instruction: " << I); 473 474 SmallVector<BasicBlock*, 8> ExitBlocks; 475 CurLoop->getExitBlocks(ExitBlocks); 476 477 if (isa<LoadInst>(I)) ++NumMovedLoads; 478 else if (isa<CallInst>(I)) ++NumMovedCalls; 479 ++NumSunk; 480 Changed = true; 481 482 // The case where there is only a single exit node of this loop is common 483 // enough that we handle it as a special (more efficient) case. It is more 484 // efficient to handle because there are no PHI nodes that need to be placed. 485 if (ExitBlocks.size() == 1) { 486 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 487 // Instruction is not used, just delete it. 488 CurAST->deleteValue(&I); 489 // If I has users in unreachable blocks, eliminate. 490 I.replaceAllUsesWith(UndefValue::get(I.getType())); 491 I.eraseFromParent(); 492 } else { 493 // Move the instruction to the start of the exit block, after any PHI 494 // nodes in it. 495 I.removeFromParent(); 496 BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI(); 497 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 498 } 499 } else if (ExitBlocks.empty()) { 500 // The instruction is actually dead if there ARE NO exit blocks. 501 CurAST->deleteValue(&I); 502 // If I has users in unreachable blocks, eliminate. 503 I.replaceAllUsesWith(UndefValue::get(I.getType())); 504 I.eraseFromParent(); 505 } else { 506 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 507 // do all of the hard work of inserting PHI nodes as necessary. We convert 508 // the value into a stack object to get it to do this. 509 510 // Firstly, we create a stack object to hold the value... 511 AllocaInst *AI = 0; 512 513 if (I.getType() != Type::getVoidTy(I.getContext())) { 514 AI = new AllocaInst(I.getType(), 0, I.getName(), 515 I.getParent()->getParent()->getEntryBlock().begin()); 516 CurAST->add(AI); 517 } 518 519 // Secondly, insert load instructions for each use of the instruction 520 // outside of the loop. 521 while (!I.use_empty()) { 522 Instruction *U = cast<Instruction>(I.use_back()); 523 524 // If the user is a PHI Node, we actually have to insert load instructions 525 // in all predecessor blocks, not in the PHI block itself! 526 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 527 // Only insert into each predecessor once, so that we don't have 528 // different incoming values from the same block! 529 std::map<BasicBlock*, Value*> InsertedBlocks; 530 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 531 if (UPN->getIncomingValue(i) == &I) { 532 BasicBlock *Pred = UPN->getIncomingBlock(i); 533 Value *&PredVal = InsertedBlocks[Pred]; 534 if (!PredVal) { 535 // Insert a new load instruction right before the terminator in 536 // the predecessor block. 537 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 538 CurAST->add(cast<LoadInst>(PredVal)); 539 } 540 541 UPN->setIncomingValue(i, PredVal); 542 } 543 544 } else { 545 LoadInst *L = new LoadInst(AI, "", U); 546 U->replaceUsesOfWith(&I, L); 547 CurAST->add(L); 548 } 549 } 550 551 // Thirdly, insert a copy of the instruction in each exit block of the loop 552 // that is dominated by the instruction, storing the result into the memory 553 // location. Be careful not to insert the instruction into any particular 554 // basic block more than once. 555 std::set<BasicBlock*> InsertedBlocks; 556 BasicBlock *InstOrigBB = I.getParent(); 557 558 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 559 BasicBlock *ExitBlock = ExitBlocks[i]; 560 561 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 562 // If we haven't already processed this exit block, do so now. 563 if (InsertedBlocks.insert(ExitBlock).second) { 564 // Insert the code after the last PHI node... 565 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 566 567 // If this is the first exit block processed, just move the original 568 // instruction, otherwise clone the original instruction and insert 569 // the copy. 570 Instruction *New; 571 if (InsertedBlocks.size() == 1) { 572 I.removeFromParent(); 573 ExitBlock->getInstList().insert(InsertPt, &I); 574 New = &I; 575 } else { 576 New = I.clone(); 577 CurAST->copyValue(&I, New); 578 if (!I.getName().empty()) 579 New->setName(I.getName()+".le"); 580 ExitBlock->getInstList().insert(InsertPt, New); 581 } 582 583 // Now that we have inserted the instruction, store it into the alloca 584 if (AI) new StoreInst(New, AI, InsertPt); 585 } 586 } 587 } 588 589 // If the instruction doesn't dominate any exit blocks, it must be dead. 590 if (InsertedBlocks.empty()) { 591 CurAST->deleteValue(&I); 592 I.eraseFromParent(); 593 } 594 595 // Finally, promote the fine value to SSA form. 596 if (AI) { 597 std::vector<AllocaInst*> Allocas; 598 Allocas.push_back(AI); 599 PromoteMemToReg(Allocas, *DT, *DF, AI->getContext(), CurAST); 600 } 601 } 602} 603 604/// hoist - When an instruction is found to only use loop invariant operands 605/// that is safe to hoist, this instruction is called to do the dirty work. 606/// 607void LICM::hoist(Instruction &I) { 608 DEBUG(errs() << "LICM hoisting to " << Preheader->getName() << ": " 609 << I << "\n"); 610 611 // Remove the instruction from its current basic block... but don't delete the 612 // instruction. 613 I.removeFromParent(); 614 615 // Insert the new node in Preheader, before the terminator. 616 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 617 618 if (isa<LoadInst>(I)) ++NumMovedLoads; 619 else if (isa<CallInst>(I)) ++NumMovedCalls; 620 ++NumHoisted; 621 Changed = true; 622} 623 624/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 625/// not a trapping instruction or if it is a trapping instruction and is 626/// guaranteed to execute. 627/// 628bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 629 // If it is not a trapping instruction, it is always safe to hoist. 630 if (Inst.isSafeToSpeculativelyExecute()) 631 return true; 632 633 // Otherwise we have to check to make sure that the instruction dominates all 634 // of the exit blocks. If it doesn't, then there is a path out of the loop 635 // which does not execute this instruction, so we can't hoist it. 636 637 // If the instruction is in the header block for the loop (which is very 638 // common), it is always guaranteed to dominate the exit blocks. Since this 639 // is a common case, and can save some work, check it now. 640 if (Inst.getParent() == CurLoop->getHeader()) 641 return true; 642 643 // Get the exit blocks for the current loop. 644 SmallVector<BasicBlock*, 8> ExitBlocks; 645 CurLoop->getExitBlocks(ExitBlocks); 646 647 // For each exit block, get the DT node and walk up the DT until the 648 // instruction's basic block is found or we exit the loop. 649 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 650 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 651 return false; 652 653 return true; 654} 655 656 657/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 658/// stores out of the loop and moving loads to before the loop. We do this by 659/// looping over the stores in the loop, looking for stores to Must pointers 660/// which are loop invariant. We promote these memory locations to use allocas 661/// instead. These allocas can easily be raised to register values by the 662/// PromoteMem2Reg functionality. 663/// 664void LICM::PromoteValuesInLoop() { 665 // PromotedValues - List of values that are promoted out of the loop. Each 666 // value has an alloca instruction for it, and a canonical version of the 667 // pointer. 668 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 669 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 670 671 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 672 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 673 674 Changed = true; 675 NumPromoted += PromotedValues.size(); 676 677 std::vector<Value*> PointerValueNumbers; 678 679 // Emit a copy from the value into the alloca'd value in the loop preheader 680 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 681 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 682 Value *Ptr = PromotedValues[i].second; 683 684 // If we are promoting a pointer value, update alias information for the 685 // inserted load. 686 Value *LoadValue = 0; 687 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 688 // Locate a load or store through the pointer, and assign the same value 689 // to LI as we are loading or storing. Since we know that the value is 690 // stored in this loop, this will always succeed. 691 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 692 UI != E; ++UI) 693 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 694 LoadValue = LI; 695 break; 696 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 697 if (SI->getOperand(1) == Ptr) { 698 LoadValue = SI->getOperand(0); 699 break; 700 } 701 } 702 assert(LoadValue && "No store through the pointer found!"); 703 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 704 } 705 706 // Load from the memory we are promoting. 707 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 708 709 if (LoadValue) CurAST->copyValue(LoadValue, LI); 710 711 // Store into the temporary alloca. 712 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 713 } 714 715 // Scan the basic blocks in the loop, replacing uses of our pointers with 716 // uses of the allocas in question. 717 // 718 for (Loop::block_iterator I = CurLoop->block_begin(), 719 E = CurLoop->block_end(); I != E; ++I) { 720 BasicBlock *BB = *I; 721 // Rewrite all loads and stores in the block of the pointer... 722 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 723 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 724 std::map<Value*, AllocaInst*>::iterator 725 I = ValueToAllocaMap.find(L->getOperand(0)); 726 if (I != ValueToAllocaMap.end()) 727 L->setOperand(0, I->second); // Rewrite load instruction... 728 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 729 std::map<Value*, AllocaInst*>::iterator 730 I = ValueToAllocaMap.find(S->getOperand(1)); 731 if (I != ValueToAllocaMap.end()) 732 S->setOperand(1, I->second); // Rewrite store instruction... 733 } 734 } 735 } 736 737 // Now that the body of the loop uses the allocas instead of the original 738 // memory locations, insert code to copy the alloca value back into the 739 // original memory location on all exits from the loop. Note that we only 740 // want to insert one copy of the code in each exit block, though the loop may 741 // exit to the same block more than once. 742 // 743 SmallPtrSet<BasicBlock*, 16> ProcessedBlocks; 744 745 SmallVector<BasicBlock*, 8> ExitBlocks; 746 CurLoop->getExitBlocks(ExitBlocks); 747 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 748 if (!ProcessedBlocks.insert(ExitBlocks[i])) 749 continue; 750 751 // Copy all of the allocas into their memory locations. 752 BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI(); 753 Instruction *InsertPos = BI; 754 unsigned PVN = 0; 755 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 756 // Load from the alloca. 757 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 758 759 // If this is a pointer type, update alias info appropriately. 760 if (isa<PointerType>(LI->getType())) 761 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 762 763 // Store into the memory we promoted. 764 new StoreInst(LI, PromotedValues[i].second, InsertPos); 765 } 766 } 767 768 // Now that we have done the deed, use the mem2reg functionality to promote 769 // all of the new allocas we just created into real SSA registers. 770 // 771 std::vector<AllocaInst*> PromotedAllocas; 772 PromotedAllocas.reserve(PromotedValues.size()); 773 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 774 PromotedAllocas.push_back(PromotedValues[i].first); 775 PromoteMemToReg(PromotedAllocas, *DT, *DF, Preheader->getContext(), CurAST); 776} 777 778/// FindPromotableValuesInLoop - Check the current loop for stores to definite 779/// pointers, which are not loaded and stored through may aliases and are safe 780/// for promotion. If these are found, create an alloca for the value, add it 781/// to the PromotedValues list, and keep track of the mapping from value to 782/// alloca. 783void LICM::FindPromotableValuesInLoop( 784 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 785 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 786 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 787 788 // Loop over all of the alias sets in the tracker object. 789 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 790 I != E; ++I) { 791 AliasSet &AS = *I; 792 // We can promote this alias set if it has a store, if it is a "Must" alias 793 // set, if the pointer is loop invariant, and if we are not eliminating any 794 // volatile loads or stores. 795 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 796 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 797 continue; 798 799 assert(!AS.empty() && 800 "Must alias set should have at least one pointer element in it!"); 801 Value *V = AS.begin()->getValue(); 802 803 // Check that all of the pointers in the alias set have the same type. We 804 // cannot (yet) promote a memory location that is loaded and stored in 805 // different sizes. 806 { 807 bool PointerOk = true; 808 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 809 if (V->getType() != I->getValue()->getType()) { 810 PointerOk = false; 811 break; 812 } 813 if (!PointerOk) 814 continue; 815 } 816 817 // It isn't safe to promote a load/store from the loop if the load/store is 818 // conditional. For example, turning: 819 // 820 // for () { if (c) *P += 1; } 821 // 822 // into: 823 // 824 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 825 // 826 // is not safe, because *P may only be valid to access if 'c' is true. 827 // 828 // It is safe to promote P if all uses are direct load/stores and if at 829 // least one is guaranteed to be executed. 830 bool GuaranteedToExecute = false; 831 bool InvalidInst = false; 832 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 833 UI != UE; ++UI) { 834 // Ignore instructions not in this loop. 835 Instruction *Use = dyn_cast<Instruction>(*UI); 836 if (!Use || !CurLoop->contains(Use->getParent())) 837 continue; 838 839 if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) { 840 InvalidInst = true; 841 break; 842 } 843 844 if (!GuaranteedToExecute) 845 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 846 } 847 848 // If there is an non-load/store instruction in the loop, we can't promote 849 // it. If there isn't a guaranteed-to-execute instruction, we can't 850 // promote. 851 if (InvalidInst || !GuaranteedToExecute) 852 continue; 853 854 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 855 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 856 PromotedValues.push_back(std::make_pair(AI, V)); 857 858 // Update the AST and alias analysis. 859 CurAST->copyValue(V, AI); 860 861 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 862 ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI)); 863 864 DEBUG(errs() << "LICM: Promoting value: " << *V << "\n"); 865 } 866} 867 868/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 869void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 870 AliasSetTracker *AST = LoopToAliasMap[L]; 871 if (!AST) 872 return; 873 874 AST->copyValue(From, To); 875} 876 877/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 878/// set. 879void LICM::deleteAnalysisValue(Value *V, Loop *L) { 880 AliasSetTracker *AST = LoopToAliasMap[L]; 881 if (!AST) 882 return; 883 884 AST->deleteValue(V); 885} 886