LICM.cpp revision 7d3ced934f1bb2c6845676c7333ef879d5219e88
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. 15// 16// This pass uses alias analysis for two purposes: 17// 18// 1. Moving loop invariant loads out of loops. If we can determine that a 19// load inside of a loop never aliases anything stored to, we can hoist it 20// 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#include "llvm/Transforms/Scalar.h" 35#include "llvm/Transforms/Utils/PromoteMemToReg.h" 36#include "llvm/Transforms/Utils/Local.h" 37#include "llvm/Analysis/LoopInfo.h" 38#include "llvm/Analysis/AliasAnalysis.h" 39#include "llvm/Analysis/AliasSetTracker.h" 40#include "llvm/Analysis/Dominators.h" 41#include "llvm/Instructions.h" 42#include "llvm/DerivedTypes.h" 43#include "llvm/Target/TargetData.h" 44#include "llvm/Support/CFG.h" 45#include "Support/CommandLine.h" 46#include "Support/Debug.h" 47#include "Support/Statistic.h" 48#include "llvm/Assembly/Writer.h" 49#include <algorithm> 50using namespace llvm; 51 52namespace { 53 cl::opt<bool> 54 DisablePromotion("disable-licm-promotion", cl::Hidden, 55 cl::desc("Disable memory promotion in LICM pass")); 56 57 Statistic<> NumSunk("licm", "Number of instructions sunk out of loop"); 58 Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop"); 59 Statistic<> NumMovedLoads("licm", "Number of load insts hoisted or sunk"); 60 Statistic<> NumPromoted("licm", 61 "Number of memory locations promoted to registers"); 62 63 struct LICM : public FunctionPass { 64 virtual bool runOnFunction(Function &F); 65 66 /// This transformation requires natural loop information & requires that 67 /// loop preheaders be inserted into the CFG... 68 /// 69 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 70 AU.setPreservesCFG(); 71 AU.addRequiredID(LoopSimplifyID); 72 AU.addRequired<LoopInfo>(); 73 AU.addRequired<DominatorTree>(); 74 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 75 AU.addRequired<AliasAnalysis>(); 76 } 77 78 private: 79 // Various analyses that we use... 80 AliasAnalysis *AA; // Current AliasAnalysis information 81 LoopInfo *LI; // Current LoopInfo 82 DominatorTree *DT; // Dominator Tree for the current Loop... 83 DominanceFrontier *DF; // Current Dominance Frontier 84 85 // State that is updated as we process loops 86 bool Changed; // Set to true when we change anything. 87 BasicBlock *Preheader; // The preheader block of the current loop... 88 Loop *CurLoop; // The current loop we are working on... 89 AliasSetTracker *CurAST; // AliasSet information for the current loop... 90 91 /// visitLoop - Hoist expressions out of the specified loop... 92 /// 93 void visitLoop(Loop *L, AliasSetTracker &AST); 94 95 /// HoistRegion - Walk the specified region of the CFG (defined by all 96 /// blocks dominated by the specified block, and that are in the current 97 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 98 /// visit definitions before uses, allowing us to hoist a loop body in one 99 /// pass without iteration. 100 /// 101 void HoistRegion(DominatorTree::Node *N); 102 103 /// inSubLoop - Little predicate that returns true if the specified basic 104 /// block is in a subloop of the current one, not the current one itself. 105 /// 106 bool inSubLoop(BasicBlock *BB) { 107 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 108 for (unsigned i = 0, e = CurLoop->getSubLoops().size(); i != e; ++i) 109 if (CurLoop->getSubLoops()[i]->contains(BB)) 110 return true; // A subloop actually contains this block! 111 return false; 112 } 113 114 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 115 /// specified exit block of the loop is dominated by the specified block 116 /// that is in the body of the loop. We use these constraints to 117 /// dramatically limit the amount of the dominator tree that needs to be 118 /// searched. 119 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 120 BasicBlock *BlockInLoop) const { 121 // If the block in the loop is the loop header, it must be dominated! 122 BasicBlock *LoopHeader = CurLoop->getHeader(); 123 if (BlockInLoop == LoopHeader) 124 return true; 125 126 DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop); 127 DominatorTree::Node *IDom = DT->getNode(ExitBlock); 128 129 // Because the exit block is not in the loop, we know we have to get _at 130 // least_ it's immediate dominator. 131 do { 132 // Get next Immediate Dominator. 133 IDom = IDom->getIDom(); 134 135 // If we have got to the header of the loop, then the instructions block 136 // did not dominate the exit node, so we can't hoist it. 137 if (IDom->getBlock() == LoopHeader) 138 return false; 139 140 } while (IDom != BlockInLoopNode); 141 142 return true; 143 } 144 145 /// sink - When an instruction is found to only be used outside of the loop, 146 /// this function moves it to the exit blocks and patches up SSA form as 147 /// needed. 148 /// 149 void sink(Instruction &I); 150 151 /// hoist - When an instruction is found to only use loop invariant operands 152 /// that is safe to hoist, this instruction is called to do the dirty work. 153 /// 154 void hoist(Instruction &I); 155 156 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 157 /// is not a trapping instruction or if it is a trapping instruction and is 158 /// guaranteed to execute. 159 /// 160 bool isSafeToExecuteUnconditionally(Instruction &I); 161 162 /// pointerInvalidatedByLoop - Return true if the body of this loop may 163 /// store into the memory location pointed to by V. 164 /// 165 bool pointerInvalidatedByLoop(Value *V) { 166 // Check to see if any of the basic blocks in CurLoop invalidate *V. 167 return CurAST->getAliasSetForPointer(V, 0).isMod(); 168 } 169 170 /// isLoopInvariant - Return true if the specified value is loop invariant 171 /// 172 inline bool isLoopInvariant(Value *V) { 173 if (Instruction *I = dyn_cast<Instruction>(V)) 174 return !CurLoop->contains(I->getParent()); 175 return true; // All non-instructions are loop invariant 176 } 177 178 bool canSinkOrHoistInst(Instruction &I); 179 bool isLoopInvariantInst(Instruction &I); 180 bool isNotUsedInLoop(Instruction &I); 181 182 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 183 /// to scalars as we can. 184 /// 185 void PromoteValuesInLoop(); 186 187 /// findPromotableValuesInLoop - Check the current loop for stores to 188 /// definite pointers, which are not loaded and stored through may aliases. 189 /// If these are found, create an alloca for the value, add it to the 190 /// PromotedValues list, and keep track of the mapping from value to 191 /// alloca... 192 /// 193 void findPromotableValuesInLoop( 194 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 195 std::map<Value*, AllocaInst*> &Val2AlMap); 196 }; 197 198 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion"); 199} 200 201FunctionPass *llvm::createLICMPass() { return new LICM(); } 202 203/// runOnFunction - For LICM, this simply traverses the loop structure of the 204/// function, hoisting expressions out of loops if possible. 205/// 206bool LICM::runOnFunction(Function &) { 207 Changed = false; 208 209 // Get our Loop and Alias Analysis information... 210 LI = &getAnalysis<LoopInfo>(); 211 AA = &getAnalysis<AliasAnalysis>(); 212 DF = &getAnalysis<DominanceFrontier>(); 213 DT = &getAnalysis<DominatorTree>(); 214 215 // Hoist expressions out of all of the top-level loops. 216 const std::vector<Loop*> &TopLevelLoops = LI->getTopLevelLoops(); 217 for (std::vector<Loop*>::const_iterator I = TopLevelLoops.begin(), 218 E = TopLevelLoops.end(); I != E; ++I) { 219 AliasSetTracker AST(*AA); 220 visitLoop(*I, AST); 221 } 222 return Changed; 223} 224 225 226/// visitLoop - Hoist expressions out of the specified loop... 227/// 228void LICM::visitLoop(Loop *L, AliasSetTracker &AST) { 229 // Recurse through all subloops before we process this loop... 230 for (std::vector<Loop*>::const_iterator I = L->getSubLoops().begin(), 231 E = L->getSubLoops().end(); I != E; ++I) { 232 AliasSetTracker SubAST(*AA); 233 visitLoop(*I, SubAST); 234 235 // Incorporate information about the subloops into this loop... 236 AST.add(SubAST); 237 } 238 CurLoop = L; 239 CurAST = &AST; 240 241 // Get the preheader block to move instructions into... 242 Preheader = L->getLoopPreheader(); 243 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 244 245 // Loop over the body of this loop, looking for calls, invokes, and stores. 246 // Because subloops have already been incorporated into AST, we skip blocks in 247 // subloops. 248 // 249 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(), 250 E = L->getBlocks().end(); I != E; ++I) 251 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops... 252 AST.add(**I); // Incorporate the specified basic block 253 254 // We want to visit all of the instructions in this loop... that are not parts 255 // of our subloops (they have already had their invariants hoisted out of 256 // their loop, into this loop, so there is no need to process the BODIES of 257 // the subloops). 258 // 259 // Traverse the body of the loop in depth first order on the dominator tree so 260 // that we are guaranteed to see definitions before we see uses. This allows 261 // us to perform the LICM transformation in one pass, without iteration. 262 // 263 HoistRegion(DT->getNode(L->getHeader())); 264 265 // Now that all loop invariants have been removed from the loop, promote any 266 // memory references to scalars that we can... 267 if (!DisablePromotion) 268 PromoteValuesInLoop(); 269 270 // Clear out loops state information for the next iteration 271 CurLoop = 0; 272 Preheader = 0; 273} 274 275/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 276/// dominated by the specified block, and that are in the current loop) in depth 277/// first order w.r.t the DominatorTree. This allows us to visit definitions 278/// before uses, allowing us to hoist a loop body in one pass without iteration. 279/// 280void LICM::HoistRegion(DominatorTree::Node *N) { 281 assert(N != 0 && "Null dominator tree node?"); 282 BasicBlock *BB = N->getBlock(); 283 284 // If this subregion is not in the top level loop at all, exit. 285 if (!CurLoop->contains(BB)) return; 286 287 // Only need to process the contents of this block if it is not part of a 288 // subloop (which would already have been processed). 289 if (!inSubLoop(BB)) 290 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 291 Instruction &I = *II++; 292 293 // We can only handle simple expressions and loads with this code. 294 if (canSinkOrHoistInst(I)) { 295 // First check to see if we can sink this instruction to the exit blocks 296 // of the loop. We can do this if the only users of the instruction are 297 // outside of the loop. In this case, it doesn't even matter if the 298 // operands of the instruction are loop invariant. 299 // 300 if (isNotUsedInLoop(I)) 301 sink(I); 302 303 // If we can't sink the instruction, try hoisting it out to the 304 // preheader. We can only do this if all of the operands of the 305 // instruction are loop invariant and if it is safe to hoist the 306 // instruction. 307 // 308 else if (isLoopInvariantInst(I) && isSafeToExecuteUnconditionally(I)) 309 hoist(I); 310 } 311 } 312 313 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 314 for (unsigned i = 0, e = Children.size(); i != e; ++i) 315 HoistRegion(Children[i]); 316} 317 318/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 319/// instruction. 320/// 321bool LICM::canSinkOrHoistInst(Instruction &I) { 322 // Loads have extra constraints we have to verify before we can hoist them. 323 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 324 if (LI->isVolatile()) 325 return false; // Don't hoist volatile loads! 326 327 // Don't hoist loads which have may-aliased stores in loop. 328 return !pointerInvalidatedByLoop(LI->getOperand(0)); 329 } 330 331 return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) || 332 isa<GetElementPtrInst>(I) || isa<VANextInst>(I) || isa<VAArgInst>(I); 333} 334 335/// isNotUsedInLoop - Return true if the only users of this instruction are 336/// outside of the loop. If this is true, we can sink the instruction to the 337/// exit blocks of the loop. 338/// 339bool LICM::isNotUsedInLoop(Instruction &I) { 340 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) 341 if (CurLoop->contains(cast<Instruction>(*UI)->getParent())) 342 return false; 343 return true; 344} 345 346 347/// isLoopInvariantInst - Return true if all operands of this instruction are 348/// loop invariant. We also filter out non-hoistable instructions here just for 349/// efficiency. 350/// 351bool LICM::isLoopInvariantInst(Instruction &I) { 352 // The instruction is loop invariant if all of its operands are loop-invariant 353 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 354 if (!isLoopInvariant(I.getOperand(i))) 355 return false; 356 357 // If we got this far, the instruction is loop invariant! 358 return true; 359} 360 361/// sink - When an instruction is found to only be used outside of the loop, 362/// this function moves it to the exit blocks and patches up SSA form as 363/// needed. 364/// 365void LICM::sink(Instruction &I) { 366 DEBUG(std::cerr << "LICM sinking instruction: " << I); 367 368 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 369 std::vector<Value*> Operands(I.op_begin(), I.op_end()); 370 371 if (isa<LoadInst>(I)) ++NumMovedLoads; 372 ++NumSunk; 373 Changed = true; 374 375 // The case where there is only a single exit node of this loop is common 376 // enough that we handle it as a special (more efficient) case. It is more 377 // efficient to handle because there are no PHI nodes that need to be placed. 378 if (ExitBlocks.size() == 1) { 379 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 380 // Instruction is not used, just delete it. 381 I.getParent()->getInstList().erase(&I); 382 } else { 383 // Move the instruction to the start of the exit block, after any PHI 384 // nodes in it. 385 I.getParent()->getInstList().remove(&I); 386 387 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin(); 388 while (isa<PHINode>(InsertPt)) ++InsertPt; 389 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 390 } 391 } else if (ExitBlocks.size() == 0) { 392 // The instruction is actually dead if there ARE NO exit blocks. 393 I.getParent()->getInstList().erase(&I); 394 } else { 395 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 396 // do all of the hard work of inserting PHI nodes as necessary. We convert 397 // the value into a stack object to get it to do this. 398 399 // Firstly, we create a stack object to hold the value... 400 AllocaInst *AI = new AllocaInst(I.getType(), 0, I.getName(), 401 I.getParent()->getParent()->front().begin()); 402 403 // Secondly, insert load instructions for each use of the instruction 404 // outside of the loop. 405 while (!I.use_empty()) { 406 Instruction *U = cast<Instruction>(I.use_back()); 407 408 // If the user is a PHI Node, we actually have to insert load instructions 409 // in all predecessor blocks, not in the PHI block itself! 410 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 411 // Only insert into each predecessor once, so that we don't have 412 // different incoming values from the same block! 413 std::map<BasicBlock*, Value*> InsertedBlocks; 414 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 415 if (UPN->getIncomingValue(i) == &I) { 416 BasicBlock *Pred = UPN->getIncomingBlock(i); 417 Value *&PredVal = InsertedBlocks[Pred]; 418 if (!PredVal) { 419 // Insert a new load instruction right before the terminator in 420 // the predecessor block. 421 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 422 } 423 424 UPN->setIncomingValue(i, PredVal); 425 } 426 427 } else { 428 LoadInst *L = new LoadInst(AI, "", U); 429 U->replaceUsesOfWith(&I, L); 430 } 431 } 432 433 // Thirdly, insert a copy of the instruction in each exit block of the loop 434 // that is dominated by the instruction, storing the result into the memory 435 // location. Be careful not to insert the instruction into any particular 436 // basic block more than once. 437 std::set<BasicBlock*> InsertedBlocks; 438 BasicBlock *InstOrigBB = I.getParent(); 439 440 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 441 BasicBlock *ExitBlock = ExitBlocks[i]; 442 443 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 444 // If we haven't already processed this exit block, do so now. 445 if (InsertedBlocks.insert(ExitBlock).second) { 446 // Insert the code after the last PHI node... 447 BasicBlock::iterator InsertPt = ExitBlock->begin(); 448 while (isa<PHINode>(InsertPt)) ++InsertPt; 449 450 // If this is the first exit block processed, just move the original 451 // instruction, otherwise clone the original instruction and insert 452 // the copy. 453 Instruction *New; 454 if (InsertedBlocks.size() == 1) { 455 I.getParent()->getInstList().remove(&I); 456 ExitBlock->getInstList().insert(InsertPt, &I); 457 New = &I; 458 } else { 459 New = I.clone(); 460 New->setName(I.getName()+".le"); 461 ExitBlock->getInstList().insert(InsertPt, New); 462 } 463 464 // Now that we have inserted the instruction, store it into the alloca 465 new StoreInst(New, AI, InsertPt); 466 } 467 } 468 } 469 470 // Finally, promote the fine value to SSA form. 471 std::vector<AllocaInst*> Allocas; 472 Allocas.push_back(AI); 473 PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData()); 474 } 475 476 // Since we just sunk an instruction, check to see if any other instructions 477 // used by this instruction are now sinkable. If so, sink them too. 478 for (unsigned i = 0, e = Operands.size(); i != e; ++i) 479 if (Instruction *OpI = dyn_cast<Instruction>(Operands[i])) 480 if (CurLoop->contains(OpI->getParent()) && canSinkOrHoistInst(*OpI) && 481 isNotUsedInLoop(*OpI) && isSafeToExecuteUnconditionally(*OpI)) 482 sink(*OpI); 483} 484 485/// hoist - When an instruction is found to only use loop invariant operands 486/// that is safe to hoist, this instruction is called to do the dirty work. 487/// 488void LICM::hoist(Instruction &I) { 489 DEBUG(std::cerr << "LICM hoisting to"; 490 WriteAsOperand(std::cerr, Preheader, false); 491 std::cerr << ": " << I); 492 493 // Remove the instruction from its current basic block... but don't delete the 494 // instruction. 495 I.getParent()->getInstList().remove(&I); 496 497 // Insert the new node in Preheader, before the terminator. 498 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 499 500 if (isa<LoadInst>(I)) ++NumMovedLoads; 501 ++NumHoisted; 502 Changed = true; 503} 504 505/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 506/// not a trapping instruction or if it is a trapping instruction and is 507/// guaranteed to execute. 508/// 509bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 510 // If it is not a trapping instruction, it is always safe to hoist. 511 if (!Inst.isTrapping()) return true; 512 513 // Otherwise we have to check to make sure that the instruction dominates all 514 // of the exit blocks. If it doesn't, then there is a path out of the loop 515 // which does not execute this instruction, so we can't hoist it. 516 517 // If the instruction is in the header block for the loop (which is very 518 // common), it is always guaranteed to dominate the exit blocks. Since this 519 // is a common case, and can save some work, check it now. 520 if (Inst.getParent() == CurLoop->getHeader()) 521 return true; 522 523 // Get the exit blocks for the current loop. 524 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 525 526 // For each exit block, get the DT node and walk up the DT until the 527 // instruction's basic block is found or we exit the loop. 528 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 529 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 530 return false; 531 532 return true; 533} 534 535 536/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 537/// stores out of the loop and moving loads to before the loop. We do this by 538/// looping over the stores in the loop, looking for stores to Must pointers 539/// which are loop invariant. We promote these memory locations to use allocas 540/// instead. These allocas can easily be raised to register values by the 541/// PromoteMem2Reg functionality. 542/// 543void LICM::PromoteValuesInLoop() { 544 // PromotedValues - List of values that are promoted out of the loop. Each 545 // value has an alloca instruction for it, and a canonical version of the 546 // pointer. 547 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 548 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 549 550 findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 551 if (ValueToAllocaMap.empty()) return; // If there are values to promote... 552 553 Changed = true; 554 NumPromoted += PromotedValues.size(); 555 556 // Emit a copy from the value into the alloca'd value in the loop preheader 557 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 558 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 559 // Load from the memory we are promoting... 560 LoadInst *LI = new LoadInst(PromotedValues[i].second, 561 PromotedValues[i].second->getName()+".promoted", 562 LoopPredInst); 563 // Store into the temporary alloca... 564 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 565 } 566 567 // Scan the basic blocks in the loop, replacing uses of our pointers with 568 // uses of the allocas in question. 569 // 570 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 571 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 572 E = LoopBBs.end(); I != E; ++I) { 573 // Rewrite all loads and stores in the block of the pointer... 574 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 575 II != E; ++II) { 576 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 577 std::map<Value*, AllocaInst*>::iterator 578 I = ValueToAllocaMap.find(L->getOperand(0)); 579 if (I != ValueToAllocaMap.end()) 580 L->setOperand(0, I->second); // Rewrite load instruction... 581 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 582 std::map<Value*, AllocaInst*>::iterator 583 I = ValueToAllocaMap.find(S->getOperand(1)); 584 if (I != ValueToAllocaMap.end()) 585 S->setOperand(1, I->second); // Rewrite store instruction... 586 } 587 } 588 } 589 590 // Now that the body of the loop uses the allocas instead of the original 591 // memory locations, insert code to copy the alloca value back into the 592 // original memory location on all exits from the loop. Note that we only 593 // want to insert one copy of the code in each exit block, though the loop may 594 // exit to the same block more than once. 595 // 596 std::set<BasicBlock*> ProcessedBlocks; 597 598 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 599 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 600 if (ProcessedBlocks.insert(ExitBlocks[i]).second) { 601 // Copy all of the allocas into their memory locations... 602 BasicBlock::iterator BI = ExitBlocks[i]->begin(); 603 while (isa<PHINode>(*BI)) 604 ++BI; // Skip over all of the phi nodes in the block... 605 Instruction *InsertPos = BI; 606 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 607 // Load from the alloca... 608 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 609 // Store into the memory we promoted... 610 new StoreInst(LI, PromotedValues[i].second, InsertPos); 611 } 612 } 613 614 // Now that we have done the deed, use the mem2reg functionality to promote 615 // all of the new allocas we just created into real SSA registers... 616 // 617 std::vector<AllocaInst*> PromotedAllocas; 618 PromotedAllocas.reserve(PromotedValues.size()); 619 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 620 PromotedAllocas.push_back(PromotedValues[i].first); 621 PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData()); 622} 623 624/// findPromotableValuesInLoop - Check the current loop for stores to definite 625/// pointers, which are not loaded and stored through may aliases. If these are 626/// found, create an alloca for the value, add it to the PromotedValues list, 627/// and keep track of the mapping from value to alloca... 628/// 629void LICM::findPromotableValuesInLoop( 630 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 631 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 632 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 633 634 // Loop over all of the alias sets in the tracker object... 635 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 636 I != E; ++I) { 637 AliasSet &AS = *I; 638 // We can promote this alias set if it has a store, if it is a "Must" alias 639 // set, and if the pointer is loop invariant. 640 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() && 641 isLoopInvariant(AS.begin()->first)) { 642 assert(AS.begin() != AS.end() && 643 "Must alias set should have at least one pointer element in it!"); 644 Value *V = AS.begin()->first; 645 646 // Check that all of the pointers in the alias set have the same type. We 647 // cannot (yet) promote a memory location that is loaded and stored in 648 // different sizes. 649 bool PointerOk = true; 650 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 651 if (V->getType() != I->first->getType()) { 652 PointerOk = false; 653 break; 654 } 655 656 if (PointerOk) { 657 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 658 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 659 PromotedValues.push_back(std::make_pair(AI, V)); 660 661 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 662 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 663 664 DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n"); 665 } 666 } 667 } 668} 669