LICM.cpp revision 91d22c8b1ec2ad8f2f29804b729473ccf720fb3e
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... 222/// 223bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 224 Changed = false; 225 226 // Get our Loop and Alias Analysis information... 227 LI = &getAnalysis<LoopInfo>(); 228 AA = &getAnalysis<AliasAnalysis>(); 229 DF = &getAnalysis<DominanceFrontier>(); 230 DT = &getAnalysis<DominatorTree>(); 231 232 CurAST = new AliasSetTracker(*AA); 233 // Collect Alias info from subloops 234 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 235 LoopItr != LoopItrE; ++LoopItr) { 236 Loop *InnerL = *LoopItr; 237 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 238 assert (InnerAST && "Where is my AST?"); 239 240 // What if InnerLoop was modified by other passes ? 241 CurAST->add(*InnerAST); 242 } 243 244 CurLoop = L; 245 246 // Get the preheader block to move instructions into... 247 Preheader = L->getLoopPreheader(); 248 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 249 250 // Loop over the body of this loop, looking for calls, invokes, and stores. 251 // Because subloops have already been incorporated into AST, we skip blocks in 252 // subloops. 253 // 254 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(), 255 E = L->getBlocks().end(); I != E; ++I) 256 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops... 257 CurAST->add(**I); // Incorporate the specified basic block 258 259 // We want to visit all of the instructions in this loop... that are not parts 260 // of our subloops (they have already had their invariants hoisted out of 261 // their loop, into this loop, so there is no need to process the BODIES of 262 // the subloops). 263 // 264 // Traverse the body of the loop in depth first order on the dominator tree so 265 // that we are guaranteed to see definitions before we see uses. This allows 266 // us to sink instructions in one pass, without iteration. AFter sinking 267 // instructions, we perform another pass to hoist them out of the loop. 268 // 269 SinkRegion(DT->getNode(L->getHeader())); 270 HoistRegion(DT->getNode(L->getHeader())); 271 272 // Now that all loop invariants have been removed from the loop, promote any 273 // memory references to scalars that we can... 274 if (!DisablePromotion) 275 PromoteValuesInLoop(); 276 277 // Clear out loops state information for the next iteration 278 CurLoop = 0; 279 Preheader = 0; 280 281 LoopToAliasMap[L] = CurAST; 282 return Changed; 283} 284 285/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 286/// dominated by the specified block, and that are in the current loop) in 287/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 288/// uses before definitions, allowing us to sink a loop body in one pass without 289/// iteration. 290/// 291void LICM::SinkRegion(DomTreeNode *N) { 292 assert(N != 0 && "Null dominator tree node?"); 293 BasicBlock *BB = N->getBlock(); 294 295 // If this subregion is not in the top level loop at all, exit. 296 if (!CurLoop->contains(BB)) return; 297 298 // We are processing blocks in reverse dfo, so process children first... 299 const std::vector<DomTreeNode*> &Children = N->getChildren(); 300 for (unsigned i = 0, e = Children.size(); i != e; ++i) 301 SinkRegion(Children[i]); 302 303 // Only need to process the contents of this block if it is not part of a 304 // subloop (which would already have been processed). 305 if (inSubLoop(BB)) return; 306 307 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 308 Instruction &I = *--II; 309 310 // Check to see if we can sink this instruction to the exit blocks 311 // of the loop. We can do this if the all users of the instruction are 312 // outside of the loop. In this case, it doesn't even matter if the 313 // operands of the instruction are loop invariant. 314 // 315 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 316 ++II; 317 sink(I); 318 } 319 } 320} 321 322 323/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 324/// dominated by the specified block, and that are in the current loop) in depth 325/// first order w.r.t the DominatorTree. This allows us to visit definitions 326/// before uses, allowing us to hoist a loop body in one pass without iteration. 327/// 328void LICM::HoistRegion(DomTreeNode *N) { 329 assert(N != 0 && "Null dominator tree node?"); 330 BasicBlock *BB = N->getBlock(); 331 332 // If this subregion is not in the top level loop at all, exit. 333 if (!CurLoop->contains(BB)) return; 334 335 // Only need to process the contents of this block if it is not part of a 336 // subloop (which would already have been processed). 337 if (!inSubLoop(BB)) 338 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 339 Instruction &I = *II++; 340 341 // Try hoisting the instruction out to the preheader. We can only do this 342 // if all of the operands of the instruction are loop invariant and if it 343 // is safe to hoist the instruction. 344 // 345 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 346 isSafeToExecuteUnconditionally(I)) 347 hoist(I); 348 } 349 350 const std::vector<DomTreeNode*> &Children = N->getChildren(); 351 for (unsigned i = 0, e = Children.size(); i != e; ++i) 352 HoistRegion(Children[i]); 353} 354 355/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 356/// instruction. 357/// 358bool LICM::canSinkOrHoistInst(Instruction &I) { 359 // Loads have extra constraints we have to verify before we can hoist them. 360 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 361 if (LI->isVolatile()) 362 return false; // Don't hoist volatile loads! 363 364 // Don't hoist loads which have may-aliased stores in loop. 365 unsigned Size = 0; 366 if (LI->getType()->isSized()) 367 Size = AA->getTargetData().getTypeSize(LI->getType()); 368 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 369 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 370 // Handle obvious cases efficiently. 371 if (Function *Callee = CI->getCalledFunction()) { 372 AliasAnalysis::ModRefBehavior Behavior =AA->getModRefBehavior(Callee, CI); 373 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 374 return true; 375 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 376 // If this call only reads from memory and there are no writes to memory 377 // in the loop, we can hoist or sink the call as appropriate. 378 bool FoundMod = false; 379 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 380 I != E; ++I) { 381 AliasSet &AS = *I; 382 if (!AS.isForwardingAliasSet() && AS.isMod()) { 383 FoundMod = true; 384 break; 385 } 386 } 387 if (!FoundMod) return true; 388 } 389 } 390 391 // FIXME: This should use mod/ref information to see if we can hoist or sink 392 // the call. 393 394 return false; 395 } 396 397 // Otherwise these instructions are hoistable/sinkable 398 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 399 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 400 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 401 isa<ShuffleVectorInst>(I); 402} 403 404/// isNotUsedInLoop - Return true if the only users of this instruction are 405/// outside of the loop. If this is true, we can sink the instruction to the 406/// exit blocks of the loop. 407/// 408bool LICM::isNotUsedInLoop(Instruction &I) { 409 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 410 Instruction *User = cast<Instruction>(*UI); 411 if (PHINode *PN = dyn_cast<PHINode>(User)) { 412 // PHI node uses occur in predecessor blocks! 413 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 414 if (PN->getIncomingValue(i) == &I) 415 if (CurLoop->contains(PN->getIncomingBlock(i))) 416 return false; 417 } else if (CurLoop->contains(User->getParent())) { 418 return false; 419 } 420 } 421 return true; 422} 423 424 425/// isLoopInvariantInst - Return true if all operands of this instruction are 426/// loop invariant. We also filter out non-hoistable instructions here just for 427/// efficiency. 428/// 429bool LICM::isLoopInvariantInst(Instruction &I) { 430 // The instruction is loop invariant if all of its operands are loop-invariant 431 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 432 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 433 return false; 434 435 // If we got this far, the instruction is loop invariant! 436 return true; 437} 438 439/// sink - When an instruction is found to only be used outside of the loop, 440/// this function moves it to the exit blocks and patches up SSA form as needed. 441/// This method is guaranteed to remove the original instruction from its 442/// position, and may either delete it or move it to outside of the loop. 443/// 444void LICM::sink(Instruction &I) { 445 DOUT << "LICM sinking instruction: " << I; 446 447 std::vector<BasicBlock*> ExitBlocks; 448 CurLoop->getExitBlocks(ExitBlocks); 449 450 if (isa<LoadInst>(I)) ++NumMovedLoads; 451 else if (isa<CallInst>(I)) ++NumMovedCalls; 452 ++NumSunk; 453 Changed = true; 454 455 // The case where there is only a single exit node of this loop is common 456 // enough that we handle it as a special (more efficient) case. It is more 457 // efficient to handle because there are no PHI nodes that need to be placed. 458 if (ExitBlocks.size() == 1) { 459 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 460 // Instruction is not used, just delete it. 461 CurAST->deleteValue(&I); 462 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 463 I.replaceAllUsesWith(UndefValue::get(I.getType())); 464 I.eraseFromParent(); 465 } else { 466 // Move the instruction to the start of the exit block, after any PHI 467 // nodes in it. 468 I.removeFromParent(); 469 470 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin(); 471 while (isa<PHINode>(InsertPt)) ++InsertPt; 472 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 473 } 474 } else if (ExitBlocks.size() == 0) { 475 // The instruction is actually dead if there ARE NO exit blocks. 476 CurAST->deleteValue(&I); 477 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 478 I.replaceAllUsesWith(UndefValue::get(I.getType())); 479 I.eraseFromParent(); 480 } else { 481 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 482 // do all of the hard work of inserting PHI nodes as necessary. We convert 483 // the value into a stack object to get it to do this. 484 485 // Firstly, we create a stack object to hold the value... 486 AllocaInst *AI = 0; 487 488 if (I.getType() != Type::VoidTy) { 489 AI = new AllocaInst(I.getType(), 0, I.getName(), 490 I.getParent()->getParent()->getEntryBlock().begin()); 491 CurAST->add(AI); 492 } 493 494 // Secondly, insert load instructions for each use of the instruction 495 // outside of the loop. 496 while (!I.use_empty()) { 497 Instruction *U = cast<Instruction>(I.use_back()); 498 499 // If the user is a PHI Node, we actually have to insert load instructions 500 // in all predecessor blocks, not in the PHI block itself! 501 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 502 // Only insert into each predecessor once, so that we don't have 503 // different incoming values from the same block! 504 std::map<BasicBlock*, Value*> InsertedBlocks; 505 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 506 if (UPN->getIncomingValue(i) == &I) { 507 BasicBlock *Pred = UPN->getIncomingBlock(i); 508 Value *&PredVal = InsertedBlocks[Pred]; 509 if (!PredVal) { 510 // Insert a new load instruction right before the terminator in 511 // the predecessor block. 512 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 513 CurAST->add(cast<LoadInst>(PredVal)); 514 } 515 516 UPN->setIncomingValue(i, PredVal); 517 } 518 519 } else { 520 LoadInst *L = new LoadInst(AI, "", U); 521 U->replaceUsesOfWith(&I, L); 522 CurAST->add(L); 523 } 524 } 525 526 // Thirdly, insert a copy of the instruction in each exit block of the loop 527 // that is dominated by the instruction, storing the result into the memory 528 // location. Be careful not to insert the instruction into any particular 529 // basic block more than once. 530 std::set<BasicBlock*> InsertedBlocks; 531 BasicBlock *InstOrigBB = I.getParent(); 532 533 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 534 BasicBlock *ExitBlock = ExitBlocks[i]; 535 536 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 537 // If we haven't already processed this exit block, do so now. 538 if (InsertedBlocks.insert(ExitBlock).second) { 539 // Insert the code after the last PHI node... 540 BasicBlock::iterator InsertPt = ExitBlock->begin(); 541 while (isa<PHINode>(InsertPt)) ++InsertPt; 542 543 // If this is the first exit block processed, just move the original 544 // instruction, otherwise clone the original instruction and insert 545 // the copy. 546 Instruction *New; 547 if (InsertedBlocks.size() == 1) { 548 I.removeFromParent(); 549 ExitBlock->getInstList().insert(InsertPt, &I); 550 New = &I; 551 } else { 552 New = I.clone(); 553 CurAST->copyValue(&I, New); 554 if (!I.getName().empty()) 555 New->setName(I.getName()+".le"); 556 ExitBlock->getInstList().insert(InsertPt, New); 557 } 558 559 // Now that we have inserted the instruction, store it into the alloca 560 if (AI) new StoreInst(New, AI, InsertPt); 561 } 562 } 563 } 564 565 // If the instruction doesn't dominate any exit blocks, it must be dead. 566 if (InsertedBlocks.empty()) { 567 CurAST->deleteValue(&I); 568 I.eraseFromParent(); 569 } 570 571 // Finally, promote the fine value to SSA form. 572 if (AI) { 573 std::vector<AllocaInst*> Allocas; 574 Allocas.push_back(AI); 575 PromoteMemToReg(Allocas, *DT, *DF, CurAST); 576 } 577 } 578} 579 580/// hoist - When an instruction is found to only use loop invariant operands 581/// that is safe to hoist, this instruction is called to do the dirty work. 582/// 583void LICM::hoist(Instruction &I) { 584 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I; 585 586 // Remove the instruction from its current basic block... but don't delete the 587 // instruction. 588 I.removeFromParent(); 589 590 // Insert the new node in Preheader, before the terminator. 591 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 592 593 if (isa<LoadInst>(I)) ++NumMovedLoads; 594 else if (isa<CallInst>(I)) ++NumMovedCalls; 595 ++NumHoisted; 596 Changed = true; 597} 598 599/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 600/// not a trapping instruction or if it is a trapping instruction and is 601/// guaranteed to execute. 602/// 603bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 604 // If it is not a trapping instruction, it is always safe to hoist. 605 if (!Inst.isTrapping()) return true; 606 607 // Otherwise we have to check to make sure that the instruction dominates all 608 // of the exit blocks. If it doesn't, then there is a path out of the loop 609 // which does not execute this instruction, so we can't hoist it. 610 611 // If the instruction is in the header block for the loop (which is very 612 // common), it is always guaranteed to dominate the exit blocks. Since this 613 // is a common case, and can save some work, check it now. 614 if (Inst.getParent() == CurLoop->getHeader()) 615 return true; 616 617 // It's always safe to load from a global or alloca. 618 if (isa<LoadInst>(Inst)) 619 if (isa<AllocationInst>(Inst.getOperand(0)) || 620 isa<GlobalVariable>(Inst.getOperand(0))) 621 return true; 622 623 // Get the exit blocks for the current loop. 624 std::vector<BasicBlock*> ExitBlocks; 625 CurLoop->getExitBlocks(ExitBlocks); 626 627 // For each exit block, get the DT node and walk up the DT until the 628 // instruction's basic block is found or we exit the loop. 629 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 630 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 631 return false; 632 633 return true; 634} 635 636 637/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 638/// stores out of the loop and moving loads to before the loop. We do this by 639/// looping over the stores in the loop, looking for stores to Must pointers 640/// which are loop invariant. We promote these memory locations to use allocas 641/// instead. These allocas can easily be raised to register values by the 642/// PromoteMem2Reg functionality. 643/// 644void LICM::PromoteValuesInLoop() { 645 // PromotedValues - List of values that are promoted out of the loop. Each 646 // value has an alloca instruction for it, and a canonical version of the 647 // pointer. 648 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 649 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 650 651 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 652 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 653 654 Changed = true; 655 NumPromoted += PromotedValues.size(); 656 657 std::vector<Value*> PointerValueNumbers; 658 659 // Emit a copy from the value into the alloca'd value in the loop preheader 660 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 661 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 662 Value *Ptr = PromotedValues[i].second; 663 664 // If we are promoting a pointer value, update alias information for the 665 // inserted load. 666 Value *LoadValue = 0; 667 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 668 // Locate a load or store through the pointer, and assign the same value 669 // to LI as we are loading or storing. Since we know that the value is 670 // stored in this loop, this will always succeed. 671 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 672 UI != E; ++UI) 673 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 674 LoadValue = LI; 675 break; 676 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 677 if (SI->getOperand(1) == Ptr) { 678 LoadValue = SI->getOperand(0); 679 break; 680 } 681 } 682 assert(LoadValue && "No store through the pointer found!"); 683 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 684 } 685 686 // Load from the memory we are promoting. 687 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 688 689 if (LoadValue) CurAST->copyValue(LoadValue, LI); 690 691 // Store into the temporary alloca. 692 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 693 } 694 695 // Scan the basic blocks in the loop, replacing uses of our pointers with 696 // uses of the allocas in question. 697 // 698 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 699 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 700 E = LoopBBs.end(); I != E; ++I) { 701 // Rewrite all loads and stores in the block of the pointer... 702 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 703 II != E; ++II) { 704 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 705 std::map<Value*, AllocaInst*>::iterator 706 I = ValueToAllocaMap.find(L->getOperand(0)); 707 if (I != ValueToAllocaMap.end()) 708 L->setOperand(0, I->second); // Rewrite load instruction... 709 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 710 std::map<Value*, AllocaInst*>::iterator 711 I = ValueToAllocaMap.find(S->getOperand(1)); 712 if (I != ValueToAllocaMap.end()) 713 S->setOperand(1, I->second); // Rewrite store instruction... 714 } 715 } 716 } 717 718 // Now that the body of the loop uses the allocas instead of the original 719 // memory locations, insert code to copy the alloca value back into the 720 // original memory location on all exits from the loop. Note that we only 721 // want to insert one copy of the code in each exit block, though the loop may 722 // exit to the same block more than once. 723 // 724 std::set<BasicBlock*> ProcessedBlocks; 725 726 std::vector<BasicBlock*> ExitBlocks; 727 CurLoop->getExitBlocks(ExitBlocks); 728 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 729 if (ProcessedBlocks.insert(ExitBlocks[i]).second) { 730 // Copy all of the allocas into their memory locations. 731 BasicBlock::iterator BI = ExitBlocks[i]->begin(); 732 while (isa<PHINode>(*BI)) 733 ++BI; // Skip over all of the phi nodes in the block. 734 Instruction *InsertPos = BI; 735 unsigned PVN = 0; 736 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 737 // Load from the alloca. 738 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 739 740 // If this is a pointer type, update alias info appropriately. 741 if (isa<PointerType>(LI->getType())) 742 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 743 744 // Store into the memory we promoted. 745 new StoreInst(LI, PromotedValues[i].second, InsertPos); 746 } 747 } 748 749 // Now that we have done the deed, use the mem2reg functionality to promote 750 // all of the new allocas we just created into real SSA registers. 751 // 752 std::vector<AllocaInst*> PromotedAllocas; 753 PromotedAllocas.reserve(PromotedValues.size()); 754 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 755 PromotedAllocas.push_back(PromotedValues[i].first); 756 PromoteMemToReg(PromotedAllocas, *DT, *DF, CurAST); 757} 758 759/// FindPromotableValuesInLoop - Check the current loop for stores to definite 760/// pointers, which are not loaded and stored through may aliases. If these are 761/// found, create an alloca for the value, add it to the PromotedValues list, 762/// and keep track of the mapping from value to alloca. 763/// 764void LICM::FindPromotableValuesInLoop( 765 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 766 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 767 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 768 769 // Loop over all of the alias sets in the tracker object. 770 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 771 I != E; ++I) { 772 AliasSet &AS = *I; 773 // We can promote this alias set if it has a store, if it is a "Must" alias 774 // set, if the pointer is loop invariant, and if we are not eliminating any 775 // volatile loads or stores. 776 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() && 777 !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) { 778 assert(AS.begin() != AS.end() && 779 "Must alias set should have at least one pointer element in it!"); 780 Value *V = AS.begin()->first; 781 782 // Check that all of the pointers in the alias set have the same type. We 783 // cannot (yet) promote a memory location that is loaded and stored in 784 // different sizes. 785 bool PointerOk = true; 786 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 787 if (V->getType() != I->first->getType()) { 788 PointerOk = false; 789 break; 790 } 791 792 if (PointerOk) { 793 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 794 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 795 PromotedValues.push_back(std::make_pair(AI, V)); 796 797 // Update the AST and alias analysis. 798 CurAST->copyValue(V, AI); 799 800 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 801 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 802 803 DOUT << "LICM: Promoting value: " << *V << "\n"; 804 } 805 } 806 } 807} 808 809/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 810void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 811 AliasSetTracker *AST = LoopToAliasMap[L]; 812 if (!AST) 813 return; 814 815 AST->copyValue(From, To); 816} 817 818/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 819/// set. 820void LICM::deleteAnalysisValue(Value *V, Loop *L) { 821 AliasSetTracker *AST = LoopToAliasMap[L]; 822 if (!AST) 823 return; 824 825 AST->deleteValue(V); 826} 827