LICM.cpp revision bc2265abe1d1b830995bb2f6322409bc7d7c5740
1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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/Instructions.h" 39#include "llvm/Target/TargetData.h" 40#include "llvm/Analysis/LoopInfo.h" 41#include "llvm/Analysis/LoopPass.h" 42#include "llvm/Analysis/AliasAnalysis.h" 43#include "llvm/Analysis/AliasSetTracker.h" 44#include "llvm/Analysis/Dominators.h" 45#include "llvm/Analysis/ScalarEvolution.h" 46#include "llvm/Transforms/Utils/PromoteMemToReg.h" 47#include "llvm/Support/CFG.h" 48#include "llvm/Support/Compiler.h" 49#include "llvm/Support/CommandLine.h" 50#include "llvm/Support/Debug.h" 51#include "llvm/ADT/Statistic.h" 52#include <algorithm> 53using namespace llvm; 54 55STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 56STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 57STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 58STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 59STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 60 61namespace { 62 cl::opt<bool> 63 DisablePromotion("disable-licm-promotion", cl::Hidden, 64 cl::desc("Disable memory promotion in LICM pass")); 65 66 struct VISIBILITY_HIDDEN LICM : public LoopPass { 67 static char ID; // Pass identification, replacement for typeid 68 LICM() : LoopPass((intptr_t)&ID) {} 69 70 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 71 72 /// This transformation requires natural loop information & requires that 73 /// loop preheaders be inserted into the CFG... 74 /// 75 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 76 AU.setPreservesCFG(); 77 AU.addRequiredID(LoopSimplifyID); 78 AU.addRequired<LoopInfo>(); 79 AU.addRequired<DominatorTree>(); 80 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 81 AU.addRequired<AliasAnalysis>(); 82 AU.addPreserved<ScalarEvolution>(); 83 AU.addPreserved<DominanceFrontier>(); 84 } 85 86 bool doFinalization() { 87 LoopToAliasMap.clear(); 88 return false; 89 } 90 91 private: 92 // Various analyses that we use... 93 AliasAnalysis *AA; // Current AliasAnalysis information 94 LoopInfo *LI; // Current LoopInfo 95 DominatorTree *DT; // Dominator Tree for the current Loop... 96 DominanceFrontier *DF; // Current Dominance Frontier 97 98 // State that is updated as we process loops 99 bool Changed; // Set to true when we change anything. 100 BasicBlock *Preheader; // The preheader block of the current loop... 101 Loop *CurLoop; // The current loop we are working on... 102 AliasSetTracker *CurAST; // AliasSet information for the current loop... 103 std::map<Loop *, AliasSetTracker *> LoopToAliasMap; 104 105 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 106 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 107 108 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 109 /// set. 110 void deleteAnalysisValue(Value *V, Loop *L); 111 112 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 113 /// dominated by the specified block, and that are in the current loop) in 114 /// reverse depth first order w.r.t the DominatorTree. This allows us to 115 /// visit uses before definitions, allowing us to sink a loop body in one 116 /// pass without iteration. 117 /// 118 void SinkRegion(DomTreeNode *N); 119 120 /// HoistRegion - Walk the specified region of the CFG (defined by all 121 /// blocks dominated by the specified block, and that are in the current 122 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 123 /// visit definitions before uses, allowing us to hoist a loop body in one 124 /// pass without iteration. 125 /// 126 void HoistRegion(DomTreeNode *N); 127 128 /// inSubLoop - Little predicate that returns true if the specified basic 129 /// block is in a subloop of the current one, not the current one itself. 130 /// 131 bool inSubLoop(BasicBlock *BB) { 132 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 133 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 134 if ((*I)->contains(BB)) 135 return true; // A subloop actually contains this block! 136 return false; 137 } 138 139 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 140 /// specified exit block of the loop is dominated by the specified block 141 /// that is in the body of the loop. We use these constraints to 142 /// dramatically limit the amount of the dominator tree that needs to be 143 /// searched. 144 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 145 BasicBlock *BlockInLoop) const { 146 // If the block in the loop is the loop header, it must be dominated! 147 BasicBlock *LoopHeader = CurLoop->getHeader(); 148 if (BlockInLoop == LoopHeader) 149 return true; 150 151 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 152 DomTreeNode *IDom = DT->getNode(ExitBlock); 153 154 // Because the exit block is not in the loop, we know we have to get _at 155 // least_ its immediate dominator. 156 do { 157 // Get next Immediate Dominator. 158 IDom = IDom->getIDom(); 159 160 // If we have got to the header of the loop, then the instructions block 161 // did not dominate the exit node, so we can't hoist it. 162 if (IDom->getBlock() == LoopHeader) 163 return false; 164 165 } while (IDom != BlockInLoopNode); 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 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 200 /// to scalars as we can. 201 /// 202 void PromoteValuesInLoop(); 203 204 /// FindPromotableValuesInLoop - Check the current loop for stores to 205 /// definite pointers, which are not loaded and stored through may aliases. 206 /// If these are found, create an alloca for the value, add it to the 207 /// PromotedValues list, and keep track of the mapping from value to 208 /// alloca... 209 /// 210 void FindPromotableValuesInLoop( 211 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 212 std::map<Value*, AllocaInst*> &Val2AlMap); 213 }; 214 215 char LICM::ID = 0; 216 RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 217} 218 219LoopPass *llvm::createLICMPass() { return new LICM(); } 220 221/// Hoist expressions out of the specified loop. Note, alias info for inner 222/// loop is not preserved so it is not a good idea to run LICM multiple 223/// times on one loop. 224/// 225bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 226 Changed = false; 227 228 // Get our Loop and Alias Analysis information... 229 LI = &getAnalysis<LoopInfo>(); 230 AA = &getAnalysis<AliasAnalysis>(); 231 DF = &getAnalysis<DominanceFrontier>(); 232 DT = &getAnalysis<DominatorTree>(); 233 234 CurAST = new AliasSetTracker(*AA); 235 // Collect Alias info from subloops 236 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 237 LoopItr != LoopItrE; ++LoopItr) { 238 Loop *InnerL = *LoopItr; 239 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 240 assert (InnerAST && "Where is my AST?"); 241 242 // What if InnerLoop was modified by other passes ? 243 CurAST->add(*InnerAST); 244 } 245 246 CurLoop = L; 247 248 // Get the preheader block to move instructions into... 249 Preheader = L->getLoopPreheader(); 250 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 251 252 // Loop over the body of this loop, looking for calls, invokes, and stores. 253 // Because subloops have already been incorporated into AST, we skip blocks in 254 // subloops. 255 // 256 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(), 257 E = L->getBlocks().end(); I != E; ++I) 258 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops... 259 CurAST->add(**I); // Incorporate the specified basic block 260 261 // We want to visit all of the instructions in this loop... that are not parts 262 // of our subloops (they have already had their invariants hoisted out of 263 // their loop, into this loop, so there is no need to process the BODIES of 264 // the subloops). 265 // 266 // Traverse the body of the loop in depth first order on the dominator tree so 267 // that we are guaranteed to see definitions before we see uses. This allows 268 // us to sink instructions in one pass, without iteration. After sinking 269 // instructions, we perform another pass to hoist them out of the loop. 270 // 271 SinkRegion(DT->getNode(L->getHeader())); 272 HoistRegion(DT->getNode(L->getHeader())); 273 274 // Now that all loop invariants have been removed from the loop, promote any 275 // memory references to scalars that we can... 276 if (!DisablePromotion) 277 PromoteValuesInLoop(); 278 279 // Clear out loops state information for the next iteration 280 CurLoop = 0; 281 Preheader = 0; 282 283 LoopToAliasMap[L] = CurAST; 284 return Changed; 285} 286 287/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 288/// dominated by the specified block, and that are in the current loop) in 289/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 290/// uses before definitions, allowing us to sink a loop body in one pass without 291/// iteration. 292/// 293void LICM::SinkRegion(DomTreeNode *N) { 294 assert(N != 0 && "Null dominator tree node?"); 295 BasicBlock *BB = N->getBlock(); 296 297 // If this subregion is not in the top level loop at all, exit. 298 if (!CurLoop->contains(BB)) return; 299 300 // We are processing blocks in reverse dfo, so process children first... 301 const std::vector<DomTreeNode*> &Children = N->getChildren(); 302 for (unsigned i = 0, e = Children.size(); i != e; ++i) 303 SinkRegion(Children[i]); 304 305 // Only need to process the contents of this block if it is not part of a 306 // subloop (which would already have been processed). 307 if (inSubLoop(BB)) return; 308 309 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 310 Instruction &I = *--II; 311 312 // Check to see if we can sink this instruction to the exit blocks 313 // of the loop. We can do this if the all users of the instruction are 314 // outside of the loop. In this case, it doesn't even matter if the 315 // operands of the instruction are loop invariant. 316 // 317 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 318 ++II; 319 sink(I); 320 } 321 } 322} 323 324 325/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 326/// dominated by the specified block, and that are in the current loop) in depth 327/// first order w.r.t the DominatorTree. This allows us to visit definitions 328/// before uses, allowing us to hoist a loop body in one pass without iteration. 329/// 330void LICM::HoistRegion(DomTreeNode *N) { 331 assert(N != 0 && "Null dominator tree node?"); 332 BasicBlock *BB = N->getBlock(); 333 334 // If this subregion is not in the top level loop at all, exit. 335 if (!CurLoop->contains(BB)) return; 336 337 // Only need to process the contents of this block if it is not part of a 338 // subloop (which would already have been processed). 339 if (!inSubLoop(BB)) 340 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 341 Instruction &I = *II++; 342 343 // Try hoisting the instruction out to the preheader. We can only do this 344 // if all of the operands of the instruction are loop invariant and if it 345 // is safe to hoist the instruction. 346 // 347 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 348 isSafeToExecuteUnconditionally(I)) 349 hoist(I); 350 } 351 352 const std::vector<DomTreeNode*> &Children = N->getChildren(); 353 for (unsigned i = 0, e = Children.size(); i != e; ++i) 354 HoistRegion(Children[i]); 355} 356 357/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 358/// instruction. 359/// 360bool LICM::canSinkOrHoistInst(Instruction &I) { 361 // Loads have extra constraints we have to verify before we can hoist them. 362 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 363 if (LI->isVolatile()) 364 return false; // Don't hoist volatile loads! 365 366 // Don't hoist loads which have may-aliased stores in loop. 367 unsigned Size = 0; 368 if (LI->getType()->isSized()) 369 Size = AA->getTargetData().getTypeSize(LI->getType()); 370 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 371 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 372 // Handle obvious cases efficiently. 373 if (Function *Callee = CI->getCalledFunction()) { 374 AliasAnalysis::ModRefBehavior Behavior =AA->getModRefBehavior(Callee, CI); 375 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 376 return true; 377 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 378 // If this call only reads from memory and there are no writes to memory 379 // in the loop, we can hoist or sink the call as appropriate. 380 bool FoundMod = false; 381 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 382 I != E; ++I) { 383 AliasSet &AS = *I; 384 if (!AS.isForwardingAliasSet() && AS.isMod()) { 385 FoundMod = true; 386 break; 387 } 388 } 389 if (!FoundMod) return true; 390 } 391 } 392 393 // FIXME: This should use mod/ref information to see if we can hoist or sink 394 // the call. 395 396 return false; 397 } 398 399 // Otherwise these instructions are hoistable/sinkable 400 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 401 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 402 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 403 isa<ShuffleVectorInst>(I); 404} 405 406/// isNotUsedInLoop - Return true if the only users of this instruction are 407/// outside of the loop. If this is true, we can sink the instruction to the 408/// exit blocks of the loop. 409/// 410bool LICM::isNotUsedInLoop(Instruction &I) { 411 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 412 Instruction *User = cast<Instruction>(*UI); 413 if (PHINode *PN = dyn_cast<PHINode>(User)) { 414 // PHI node uses occur in predecessor blocks! 415 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 416 if (PN->getIncomingValue(i) == &I) 417 if (CurLoop->contains(PN->getIncomingBlock(i))) 418 return false; 419 } else if (CurLoop->contains(User->getParent())) { 420 return false; 421 } 422 } 423 return true; 424} 425 426 427/// isLoopInvariantInst - Return true if all operands of this instruction are 428/// loop invariant. We also filter out non-hoistable instructions here just for 429/// efficiency. 430/// 431bool LICM::isLoopInvariantInst(Instruction &I) { 432 // The instruction is loop invariant if all of its operands are loop-invariant 433 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 434 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 435 return false; 436 437 // If we got this far, the instruction is loop invariant! 438 return true; 439} 440 441/// sink - When an instruction is found to only be used outside of the loop, 442/// this function moves it to the exit blocks and patches up SSA form as needed. 443/// This method is guaranteed to remove the original instruction from its 444/// position, and may either delete it or move it to outside of the loop. 445/// 446void LICM::sink(Instruction &I) { 447 DOUT << "LICM sinking instruction: " << I; 448 449 SmallVector<BasicBlock*, 8> ExitBlocks; 450 CurLoop->getExitBlocks(ExitBlocks); 451 452 if (isa<LoadInst>(I)) ++NumMovedLoads; 453 else if (isa<CallInst>(I)) ++NumMovedCalls; 454 ++NumSunk; 455 Changed = true; 456 457 // The case where there is only a single exit node of this loop is common 458 // enough that we handle it as a special (more efficient) case. It is more 459 // efficient to handle because there are no PHI nodes that need to be placed. 460 if (ExitBlocks.size() == 1) { 461 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 462 // Instruction is not used, just delete it. 463 CurAST->deleteValue(&I); 464 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 465 I.replaceAllUsesWith(UndefValue::get(I.getType())); 466 I.eraseFromParent(); 467 } else { 468 // Move the instruction to the start of the exit block, after any PHI 469 // nodes in it. 470 I.removeFromParent(); 471 472 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin(); 473 while (isa<PHINode>(InsertPt)) ++InsertPt; 474 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 475 } 476 } else if (ExitBlocks.size() == 0) { 477 // The instruction is actually dead if there ARE NO exit blocks. 478 CurAST->deleteValue(&I); 479 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 480 I.replaceAllUsesWith(UndefValue::get(I.getType())); 481 I.eraseFromParent(); 482 } else { 483 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 484 // do all of the hard work of inserting PHI nodes as necessary. We convert 485 // the value into a stack object to get it to do this. 486 487 // Firstly, we create a stack object to hold the value... 488 AllocaInst *AI = 0; 489 490 if (I.getType() != Type::VoidTy) { 491 AI = new AllocaInst(I.getType(), 0, I.getName(), 492 I.getParent()->getParent()->getEntryBlock().begin()); 493 CurAST->add(AI); 494 } 495 496 // Secondly, insert load instructions for each use of the instruction 497 // outside of the loop. 498 while (!I.use_empty()) { 499 Instruction *U = cast<Instruction>(I.use_back()); 500 501 // If the user is a PHI Node, we actually have to insert load instructions 502 // in all predecessor blocks, not in the PHI block itself! 503 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 504 // Only insert into each predecessor once, so that we don't have 505 // different incoming values from the same block! 506 std::map<BasicBlock*, Value*> InsertedBlocks; 507 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 508 if (UPN->getIncomingValue(i) == &I) { 509 BasicBlock *Pred = UPN->getIncomingBlock(i); 510 Value *&PredVal = InsertedBlocks[Pred]; 511 if (!PredVal) { 512 // Insert a new load instruction right before the terminator in 513 // the predecessor block. 514 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 515 CurAST->add(cast<LoadInst>(PredVal)); 516 } 517 518 UPN->setIncomingValue(i, PredVal); 519 } 520 521 } else { 522 LoadInst *L = new LoadInst(AI, "", U); 523 U->replaceUsesOfWith(&I, L); 524 CurAST->add(L); 525 } 526 } 527 528 // Thirdly, insert a copy of the instruction in each exit block of the loop 529 // that is dominated by the instruction, storing the result into the memory 530 // location. Be careful not to insert the instruction into any particular 531 // basic block more than once. 532 std::set<BasicBlock*> InsertedBlocks; 533 BasicBlock *InstOrigBB = I.getParent(); 534 535 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 536 BasicBlock *ExitBlock = ExitBlocks[i]; 537 538 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 539 // If we haven't already processed this exit block, do so now. 540 if (InsertedBlocks.insert(ExitBlock).second) { 541 // Insert the code after the last PHI node... 542 BasicBlock::iterator InsertPt = ExitBlock->begin(); 543 while (isa<PHINode>(InsertPt)) ++InsertPt; 544 545 // If this is the first exit block processed, just move the original 546 // instruction, otherwise clone the original instruction and insert 547 // the copy. 548 Instruction *New; 549 if (InsertedBlocks.size() == 1) { 550 I.removeFromParent(); 551 ExitBlock->getInstList().insert(InsertPt, &I); 552 New = &I; 553 } else { 554 New = I.clone(); 555 CurAST->copyValue(&I, New); 556 if (!I.getName().empty()) 557 New->setName(I.getName()+".le"); 558 ExitBlock->getInstList().insert(InsertPt, New); 559 } 560 561 // Now that we have inserted the instruction, store it into the alloca 562 if (AI) new StoreInst(New, AI, InsertPt); 563 } 564 } 565 } 566 567 // If the instruction doesn't dominate any exit blocks, it must be dead. 568 if (InsertedBlocks.empty()) { 569 CurAST->deleteValue(&I); 570 I.eraseFromParent(); 571 } 572 573 // Finally, promote the fine value to SSA form. 574 if (AI) { 575 std::vector<AllocaInst*> Allocas; 576 Allocas.push_back(AI); 577 PromoteMemToReg(Allocas, *DT, *DF, CurAST); 578 } 579 } 580} 581 582/// hoist - When an instruction is found to only use loop invariant operands 583/// that is safe to hoist, this instruction is called to do the dirty work. 584/// 585void LICM::hoist(Instruction &I) { 586 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I; 587 588 // Remove the instruction from its current basic block... but don't delete the 589 // instruction. 590 I.removeFromParent(); 591 592 // Insert the new node in Preheader, before the terminator. 593 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 594 595 if (isa<LoadInst>(I)) ++NumMovedLoads; 596 else if (isa<CallInst>(I)) ++NumMovedCalls; 597 ++NumHoisted; 598 Changed = true; 599} 600 601/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 602/// not a trapping instruction or if it is a trapping instruction and is 603/// guaranteed to execute. 604/// 605bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 606 // If it is not a trapping instruction, it is always safe to hoist. 607 if (!Inst.isTrapping()) return true; 608 609 // Otherwise we have to check to make sure that the instruction dominates all 610 // of the exit blocks. If it doesn't, then there is a path out of the loop 611 // which does not execute this instruction, so we can't hoist it. 612 613 // If the instruction is in the header block for the loop (which is very 614 // common), it is always guaranteed to dominate the exit blocks. Since this 615 // is a common case, and can save some work, check it now. 616 if (Inst.getParent() == CurLoop->getHeader()) 617 return true; 618 619 // It's always safe to load from a global or alloca. 620 if (isa<LoadInst>(Inst)) 621 if (isa<AllocationInst>(Inst.getOperand(0)) || 622 isa<GlobalVariable>(Inst.getOperand(0))) 623 return true; 624 625 // Get the exit blocks for the current loop. 626 SmallVector<BasicBlock*, 8> ExitBlocks; 627 CurLoop->getExitBlocks(ExitBlocks); 628 629 // For each exit block, get the DT node and walk up the DT until the 630 // instruction's basic block is found or we exit the loop. 631 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 632 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 633 return false; 634 635 return true; 636} 637 638 639/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 640/// stores out of the loop and moving loads to before the loop. We do this by 641/// looping over the stores in the loop, looking for stores to Must pointers 642/// which are loop invariant. We promote these memory locations to use allocas 643/// instead. These allocas can easily be raised to register values by the 644/// PromoteMem2Reg functionality. 645/// 646void LICM::PromoteValuesInLoop() { 647 // PromotedValues - List of values that are promoted out of the loop. Each 648 // value has an alloca instruction for it, and a canonical version of the 649 // pointer. 650 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 651 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 652 653 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 654 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 655 656 Changed = true; 657 NumPromoted += PromotedValues.size(); 658 659 std::vector<Value*> PointerValueNumbers; 660 661 // Emit a copy from the value into the alloca'd value in the loop preheader 662 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 663 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 664 Value *Ptr = PromotedValues[i].second; 665 666 // If we are promoting a pointer value, update alias information for the 667 // inserted load. 668 Value *LoadValue = 0; 669 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 670 // Locate a load or store through the pointer, and assign the same value 671 // to LI as we are loading or storing. Since we know that the value is 672 // stored in this loop, this will always succeed. 673 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 674 UI != E; ++UI) 675 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 676 LoadValue = LI; 677 break; 678 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 679 if (SI->getOperand(1) == Ptr) { 680 LoadValue = SI->getOperand(0); 681 break; 682 } 683 } 684 assert(LoadValue && "No store through the pointer found!"); 685 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 686 } 687 688 // Load from the memory we are promoting. 689 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 690 691 if (LoadValue) CurAST->copyValue(LoadValue, LI); 692 693 // Store into the temporary alloca. 694 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 695 } 696 697 // Scan the basic blocks in the loop, replacing uses of our pointers with 698 // uses of the allocas in question. 699 // 700 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 701 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 702 E = LoopBBs.end(); I != E; ++I) { 703 // Rewrite all loads and stores in the block of the pointer... 704 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 705 II != E; ++II) { 706 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 707 std::map<Value*, AllocaInst*>::iterator 708 I = ValueToAllocaMap.find(L->getOperand(0)); 709 if (I != ValueToAllocaMap.end()) 710 L->setOperand(0, I->second); // Rewrite load instruction... 711 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 712 std::map<Value*, AllocaInst*>::iterator 713 I = ValueToAllocaMap.find(S->getOperand(1)); 714 if (I != ValueToAllocaMap.end()) 715 S->setOperand(1, I->second); // Rewrite store instruction... 716 } 717 } 718 } 719 720 // Now that the body of the loop uses the allocas instead of the original 721 // memory locations, insert code to copy the alloca value back into the 722 // original memory location on all exits from the loop. Note that we only 723 // want to insert one copy of the code in each exit block, though the loop may 724 // exit to the same block more than once. 725 // 726 std::set<BasicBlock*> ProcessedBlocks; 727 728 SmallVector<BasicBlock*, 8> ExitBlocks; 729 CurLoop->getExitBlocks(ExitBlocks); 730 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 731 if (ProcessedBlocks.insert(ExitBlocks[i]).second) { 732 // Copy all of the allocas into their memory locations. 733 BasicBlock::iterator BI = ExitBlocks[i]->begin(); 734 while (isa<PHINode>(*BI)) 735 ++BI; // Skip over all of the phi nodes in the block. 736 Instruction *InsertPos = BI; 737 unsigned PVN = 0; 738 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 739 // Load from the alloca. 740 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 741 742 // If this is a pointer type, update alias info appropriately. 743 if (isa<PointerType>(LI->getType())) 744 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 745 746 // Store into the memory we promoted. 747 new StoreInst(LI, PromotedValues[i].second, InsertPos); 748 } 749 } 750 751 // Now that we have done the deed, use the mem2reg functionality to promote 752 // all of the new allocas we just created into real SSA registers. 753 // 754 std::vector<AllocaInst*> PromotedAllocas; 755 PromotedAllocas.reserve(PromotedValues.size()); 756 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 757 PromotedAllocas.push_back(PromotedValues[i].first); 758 PromoteMemToReg(PromotedAllocas, *DT, *DF, CurAST); 759} 760 761/// FindPromotableValuesInLoop - Check the current loop for stores to definite 762/// pointers, which are not loaded and stored through may aliases and are safe 763/// for promotion. If these are found, create an alloca for the value, add it 764/// to the PromotedValues list, and keep track of the mapping from value to 765/// alloca. 766void LICM::FindPromotableValuesInLoop( 767 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 768 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 769 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 770 771 SmallVector<Instruction *, 4> LoopExits; 772 SmallVector<BasicBlock *, 4> Blocks; 773 CurLoop->getExitingBlocks(Blocks); 774 for (SmallVector<BasicBlock *, 4>::iterator BI = Blocks.begin(), 775 BE = Blocks.end(); BI != BE; ++BI) { 776 BasicBlock *BB = *BI; 777 LoopExits.push_back(BB->getTerminator()); 778 } 779 780 // Loop over all of the alias sets in the tracker object. 781 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 782 I != E; ++I) { 783 AliasSet &AS = *I; 784 // We can promote this alias set if it has a store, if it is a "Must" alias 785 // set, if the pointer is loop invariant, and if we are not eliminating any 786 // volatile loads or stores. 787 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() && 788 !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) { 789 assert(AS.begin() != AS.end() && 790 "Must alias set should have at least one pointer element in it!"); 791 Value *V = AS.begin()->first; 792 793 // Check that all of the pointers in the alias set have the same type. We 794 // cannot (yet) promote a memory location that is loaded and stored in 795 // different sizes. 796 bool PointerOk = true; 797 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 798 if (V->getType() != I->first->getType()) { 799 PointerOk = false; 800 break; 801 } 802 803 // If one use of value V inside the loop is safe then it is OK to promote 804 // this value. On the otherside if there is not any unsafe use inside the 805 // looop then also it is OK to promote this value. Otherwise it is 806 // unsafe to promote this value. 807 if (PointerOk) { 808 bool oneSafeUse = false; 809 bool oneUnsafeUse = false; 810 for(Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 811 UI != UE; ++UI) { 812 Instruction *Use = dyn_cast<Instruction>(*UI); 813 if (!Use || !CurLoop->contains(Use->getParent())) 814 continue; 815 for (SmallVector<Instruction *, 4>::iterator 816 ExitI = LoopExits.begin(), ExitE = LoopExits.end(); 817 ExitI != ExitE; ++ExitI) { 818 Instruction *Ex = *ExitI; 819 if (!isa<PHINode>(Use) && DT->dominates(Use, Ex)) { 820 oneSafeUse = true; 821 break; 822 } 823 else 824 oneUnsafeUse = true; 825 } 826 827 if (oneSafeUse) 828 break; 829 } 830 831 if (oneSafeUse) 832 PointerOk = true; 833 else if (!oneUnsafeUse) 834 PointerOk = true; 835 else 836 PointerOk = false; 837 } 838 839 if (PointerOk) { 840 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 841 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 842 PromotedValues.push_back(std::make_pair(AI, V)); 843 844 // Update the AST and alias analysis. 845 CurAST->copyValue(V, AI); 846 847 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 848 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 849 850 DOUT << "LICM: Promoting value: " << *V << "\n"; 851 } 852 } 853 } 854} 855 856/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 857void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 858 AliasSetTracker *AST = LoopToAliasMap[L]; 859 if (!AST) 860 return; 861 862 AST->copyValue(From, To); 863} 864 865/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 866/// set. 867void LICM::deleteAnalysisValue(Value *V, Loop *L) { 868 AliasSetTracker *AST = LoopToAliasMap[L]; 869 if (!AST) 870 return; 871 872 AST->deleteValue(V); 873} 874