LICM.cpp revision 529429224c9334ca308e1a8563d03f52196fc342
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 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 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 96 /// dominated by the specified block, and that are in the current loop) in 97 /// reverse depth first order w.r.t the DominatorTree. This allows us to 98 /// visit uses before definitions, allowing us to sink a loop body in one 99 /// pass without iteration. 100 /// 101 void SinkRegion(DominatorTree::Node *N); 102 103 /// HoistRegion - Walk the specified region of the CFG (defined by all 104 /// blocks dominated by the specified block, and that are in the current 105 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 106 /// visit definitions before uses, allowing us to hoist a loop body in one 107 /// pass without iteration. 108 /// 109 void HoistRegion(DominatorTree::Node *N); 110 111 /// inSubLoop - Little predicate that returns true if the specified basic 112 /// block is in a subloop of the current one, not the current one itself. 113 /// 114 bool inSubLoop(BasicBlock *BB) { 115 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 116 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 117 if ((*I)->contains(BB)) 118 return true; // A subloop actually contains this block! 119 return false; 120 } 121 122 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 123 /// specified exit block of the loop is dominated by the specified block 124 /// that is in the body of the loop. We use these constraints to 125 /// dramatically limit the amount of the dominator tree that needs to be 126 /// searched. 127 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 128 BasicBlock *BlockInLoop) const { 129 // If the block in the loop is the loop header, it must be dominated! 130 BasicBlock *LoopHeader = CurLoop->getHeader(); 131 if (BlockInLoop == LoopHeader) 132 return true; 133 134 DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop); 135 DominatorTree::Node *IDom = DT->getNode(ExitBlock); 136 137 // Because the exit block is not in the loop, we know we have to get _at 138 // least_ it's immediate dominator. 139 do { 140 // Get next Immediate Dominator. 141 IDom = IDom->getIDom(); 142 143 // If we have got to the header of the loop, then the instructions block 144 // did not dominate the exit node, so we can't hoist it. 145 if (IDom->getBlock() == LoopHeader) 146 return false; 147 148 } while (IDom != BlockInLoopNode); 149 150 return true; 151 } 152 153 /// sink - When an instruction is found to only be used outside of the loop, 154 /// this function moves it to the exit blocks and patches up SSA form as 155 /// needed. 156 /// 157 void sink(Instruction &I); 158 159 /// hoist - When an instruction is found to only use loop invariant operands 160 /// that is safe to hoist, this instruction is called to do the dirty work. 161 /// 162 void hoist(Instruction &I); 163 164 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 165 /// is not a trapping instruction or if it is a trapping instruction and is 166 /// guaranteed to execute. 167 /// 168 bool isSafeToExecuteUnconditionally(Instruction &I); 169 170 /// pointerInvalidatedByLoop - Return true if the body of this loop may 171 /// store into the memory location pointed to by V. 172 /// 173 bool pointerInvalidatedByLoop(Value *V) { 174 // Check to see if any of the basic blocks in CurLoop invalidate *V. 175 return CurAST->getAliasSetForPointer(V, 0).isMod(); 176 } 177 178 /// isLoopInvariant - Return true if the specified value is loop invariant 179 /// 180 inline bool isLoopInvariant(Value *V) { 181 if (Instruction *I = dyn_cast<Instruction>(V)) 182 return !CurLoop->contains(I->getParent()); 183 return true; // All non-instructions are loop invariant 184 } 185 186 bool canSinkOrHoistInst(Instruction &I); 187 bool isLoopInvariantInst(Instruction &I); 188 bool isNotUsedInLoop(Instruction &I); 189 190 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 191 /// to scalars as we can. 192 /// 193 void PromoteValuesInLoop(); 194 195 /// findPromotableValuesInLoop - Check the current loop for stores to 196 /// definite pointers, which are not loaded and stored through may aliases. 197 /// If these are found, create an alloca for the value, add it to the 198 /// PromotedValues list, and keep track of the mapping from value to 199 /// alloca... 200 /// 201 void findPromotableValuesInLoop( 202 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 203 std::map<Value*, AllocaInst*> &Val2AlMap); 204 }; 205 206 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion"); 207} 208 209FunctionPass *llvm::createLICMPass() { return new LICM(); } 210 211/// runOnFunction - For LICM, this simply traverses the loop structure of the 212/// function, hoisting expressions out of loops if possible. 213/// 214bool LICM::runOnFunction(Function &) { 215 Changed = false; 216 217 // Get our Loop and Alias Analysis information... 218 LI = &getAnalysis<LoopInfo>(); 219 AA = &getAnalysis<AliasAnalysis>(); 220 DF = &getAnalysis<DominanceFrontier>(); 221 DT = &getAnalysis<DominatorTree>(); 222 223 // Hoist expressions out of all of the top-level loops. 224 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) { 225 AliasSetTracker AST(*AA); 226 visitLoop(*I, AST); 227 } 228 return Changed; 229} 230 231 232/// visitLoop - Hoist expressions out of the specified loop... 233/// 234void LICM::visitLoop(Loop *L, AliasSetTracker &AST) { 235 // Recurse through all subloops before we process this loop... 236 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) { 237 AliasSetTracker SubAST(*AA); 238 visitLoop(*I, SubAST); 239 240 // Incorporate information about the subloops into this loop... 241 AST.add(SubAST); 242 } 243 CurLoop = L; 244 CurAST = &AST; 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 AST.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} 281 282/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 283/// dominated by the specified block, and that are in the current loop) in 284/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 285/// uses before definitions, allowing us to sink a loop body in one pass without 286/// iteration. 287/// 288void LICM::SinkRegion(DominatorTree::Node *N) { 289 assert(N != 0 && "Null dominator tree node?"); 290 BasicBlock *BB = N->getBlock(); 291 292 // If this subregion is not in the top level loop at all, exit. 293 if (!CurLoop->contains(BB)) return; 294 295 // We are processing blocks in reverse dfo, so process children first... 296 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 297 for (unsigned i = 0, e = Children.size(); i != e; ++i) 298 SinkRegion(Children[i]); 299 300 // Only need to process the contents of this block if it is not part of a 301 // subloop (which would already have been processed). 302 if (inSubLoop(BB)) return; 303 304 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 305 Instruction &I = *--II; 306 307 // Check to see if we can sink this instruction to the exit blocks 308 // of the loop. We can do this if the all users of the instruction are 309 // outside of the loop. In this case, it doesn't even matter if the 310 // operands of the instruction are loop invariant. 311 // 312 if (canSinkOrHoistInst(I) && isNotUsedInLoop(I)) { 313 ++II; 314 sink(I); 315 } 316 } 317} 318 319 320/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 321/// dominated by the specified block, and that are in the current loop) in depth 322/// first order w.r.t the DominatorTree. This allows us to visit definitions 323/// before uses, allowing us to hoist a loop body in one pass without iteration. 324/// 325void LICM::HoistRegion(DominatorTree::Node *N) { 326 assert(N != 0 && "Null dominator tree node?"); 327 BasicBlock *BB = N->getBlock(); 328 329 // If this subregion is not in the top level loop at all, exit. 330 if (!CurLoop->contains(BB)) return; 331 332 // Only need to process the contents of this block if it is not part of a 333 // subloop (which would already have been processed). 334 if (!inSubLoop(BB)) 335 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 336 Instruction &I = *II++; 337 338 // Try hoisting the instruction out to the preheader. We can only do this 339 // if all of the operands of the instruction are loop invariant and if it 340 // is safe to hoist the instruction. 341 // 342 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 343 isSafeToExecuteUnconditionally(I)) 344 hoist(I); 345 } 346 347 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 348 for (unsigned i = 0, e = Children.size(); i != e; ++i) 349 HoistRegion(Children[i]); 350} 351 352/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 353/// instruction. 354/// 355bool LICM::canSinkOrHoistInst(Instruction &I) { 356 // Loads have extra constraints we have to verify before we can hoist them. 357 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 358 if (LI->isVolatile()) 359 return false; // Don't hoist volatile loads! 360 361 // Don't hoist loads which have may-aliased stores in loop. 362 return !pointerInvalidatedByLoop(LI->getOperand(0)); 363 } 364 365 return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) || 366 isa<GetElementPtrInst>(I) || isa<VANextInst>(I) || isa<VAArgInst>(I); 367} 368 369/// isNotUsedInLoop - Return true if the only users of this instruction are 370/// outside of the loop. If this is true, we can sink the instruction to the 371/// exit blocks of the loop. 372/// 373bool LICM::isNotUsedInLoop(Instruction &I) { 374 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 375 Instruction *User = cast<Instruction>(*UI); 376 if (PHINode *PN = dyn_cast<PHINode>(User)) { 377 // PHI node uses occur in predecessor blocks! 378 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 379 if (PN->getIncomingValue(i) == &I) 380 if (CurLoop->contains(PN->getIncomingBlock(i))) 381 return false; 382 } else if (CurLoop->contains(User->getParent())) { 383 return false; 384 } 385 } 386 return true; 387} 388 389 390/// isLoopInvariantInst - Return true if all operands of this instruction are 391/// loop invariant. We also filter out non-hoistable instructions here just for 392/// efficiency. 393/// 394bool LICM::isLoopInvariantInst(Instruction &I) { 395 // The instruction is loop invariant if all of its operands are loop-invariant 396 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 397 if (!isLoopInvariant(I.getOperand(i))) 398 return false; 399 400 // If we got this far, the instruction is loop invariant! 401 return true; 402} 403 404/// sink - When an instruction is found to only be used outside of the loop, 405/// this function moves it to the exit blocks and patches up SSA form as needed. 406/// This method is guaranteed to remove the original instruction from its 407/// position, and may either delete it or move it to outside of the loop. 408/// 409void LICM::sink(Instruction &I) { 410 DEBUG(std::cerr << "LICM sinking instruction: " << I); 411 412 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 413 414 if (isa<LoadInst>(I)) ++NumMovedLoads; 415 ++NumSunk; 416 Changed = true; 417 418 // The case where there is only a single exit node of this loop is common 419 // enough that we handle it as a special (more efficient) case. It is more 420 // efficient to handle because there are no PHI nodes that need to be placed. 421 if (ExitBlocks.size() == 1) { 422 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 423 // Instruction is not used, just delete it. 424 CurAST->remove(&I); 425 I.getParent()->getInstList().erase(&I); 426 } else { 427 // Move the instruction to the start of the exit block, after any PHI 428 // nodes in it. 429 I.getParent()->getInstList().remove(&I); 430 431 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin(); 432 while (isa<PHINode>(InsertPt)) ++InsertPt; 433 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 434 } 435 } else if (ExitBlocks.size() == 0) { 436 // The instruction is actually dead if there ARE NO exit blocks. 437 CurAST->remove(&I); 438 I.getParent()->getInstList().erase(&I); 439 } else { 440 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 441 // do all of the hard work of inserting PHI nodes as necessary. We convert 442 // the value into a stack object to get it to do this. 443 444 // Firstly, we create a stack object to hold the value... 445 AllocaInst *AI = new AllocaInst(I.getType(), 0, I.getName(), 446 I.getParent()->getParent()->front().begin()); 447 448 // Secondly, insert load instructions for each use of the instruction 449 // outside of the loop. 450 while (!I.use_empty()) { 451 Instruction *U = cast<Instruction>(I.use_back()); 452 453 // If the user is a PHI Node, we actually have to insert load instructions 454 // in all predecessor blocks, not in the PHI block itself! 455 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 456 // Only insert into each predecessor once, so that we don't have 457 // different incoming values from the same block! 458 std::map<BasicBlock*, Value*> InsertedBlocks; 459 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 460 if (UPN->getIncomingValue(i) == &I) { 461 BasicBlock *Pred = UPN->getIncomingBlock(i); 462 Value *&PredVal = InsertedBlocks[Pred]; 463 if (!PredVal) { 464 // Insert a new load instruction right before the terminator in 465 // the predecessor block. 466 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 467 } 468 469 UPN->setIncomingValue(i, PredVal); 470 } 471 472 } else { 473 LoadInst *L = new LoadInst(AI, "", U); 474 U->replaceUsesOfWith(&I, L); 475 } 476 } 477 478 // Thirdly, insert a copy of the instruction in each exit block of the loop 479 // that is dominated by the instruction, storing the result into the memory 480 // location. Be careful not to insert the instruction into any particular 481 // basic block more than once. 482 std::set<BasicBlock*> InsertedBlocks; 483 BasicBlock *InstOrigBB = I.getParent(); 484 485 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 486 BasicBlock *ExitBlock = ExitBlocks[i]; 487 488 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 489 // If we haven't already processed this exit block, do so now. 490 if (InsertedBlocks.insert(ExitBlock).second) { 491 // Insert the code after the last PHI node... 492 BasicBlock::iterator InsertPt = ExitBlock->begin(); 493 while (isa<PHINode>(InsertPt)) ++InsertPt; 494 495 // If this is the first exit block processed, just move the original 496 // instruction, otherwise clone the original instruction and insert 497 // the copy. 498 Instruction *New; 499 if (InsertedBlocks.size() == 1) { 500 I.getParent()->getInstList().remove(&I); 501 ExitBlock->getInstList().insert(InsertPt, &I); 502 New = &I; 503 } else { 504 New = I.clone(); 505 New->setName(I.getName()+".le"); 506 ExitBlock->getInstList().insert(InsertPt, New); 507 } 508 509 // Now that we have inserted the instruction, store it into the alloca 510 new StoreInst(New, AI, InsertPt); 511 } 512 } 513 } 514 515 // If the instruction doesn't dominate any exit blocks, it must be dead. 516 if (InsertedBlocks.empty()) { 517 CurAST->remove(&I); 518 I.getParent()->getInstList().erase(&I); 519 } 520 521 // Finally, promote the fine value to SSA form. 522 std::vector<AllocaInst*> Allocas; 523 Allocas.push_back(AI); 524 PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData()); 525 } 526} 527 528/// hoist - When an instruction is found to only use loop invariant operands 529/// that is safe to hoist, this instruction is called to do the dirty work. 530/// 531void LICM::hoist(Instruction &I) { 532 DEBUG(std::cerr << "LICM hoisting to"; 533 WriteAsOperand(std::cerr, Preheader, false); 534 std::cerr << ": " << I); 535 536 // Remove the instruction from its current basic block... but don't delete the 537 // instruction. 538 I.getParent()->getInstList().remove(&I); 539 540 // Insert the new node in Preheader, before the terminator. 541 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 542 543 if (isa<LoadInst>(I)) ++NumMovedLoads; 544 ++NumHoisted; 545 Changed = true; 546} 547 548/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 549/// not a trapping instruction or if it is a trapping instruction and is 550/// guaranteed to execute. 551/// 552bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 553 // If it is not a trapping instruction, it is always safe to hoist. 554 if (!Inst.isTrapping()) return true; 555 556 // Otherwise we have to check to make sure that the instruction dominates all 557 // of the exit blocks. If it doesn't, then there is a path out of the loop 558 // which does not execute this instruction, so we can't hoist it. 559 560 // If the instruction is in the header block for the loop (which is very 561 // common), it is always guaranteed to dominate the exit blocks. Since this 562 // is a common case, and can save some work, check it now. 563 if (Inst.getParent() == CurLoop->getHeader()) 564 return true; 565 566 // Get the exit blocks for the current loop. 567 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 568 569 // For each exit block, get the DT node and walk up the DT until the 570 // instruction's basic block is found or we exit the loop. 571 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 572 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 573 return false; 574 575 return true; 576} 577 578 579/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 580/// stores out of the loop and moving loads to before the loop. We do this by 581/// looping over the stores in the loop, looking for stores to Must pointers 582/// which are loop invariant. We promote these memory locations to use allocas 583/// instead. These allocas can easily be raised to register values by the 584/// PromoteMem2Reg functionality. 585/// 586void LICM::PromoteValuesInLoop() { 587 // PromotedValues - List of values that are promoted out of the loop. Each 588 // value has an alloca instruction for it, and a canonical version of the 589 // pointer. 590 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 591 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 592 593 findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 594 if (ValueToAllocaMap.empty()) return; // If there are values to promote... 595 596 Changed = true; 597 NumPromoted += PromotedValues.size(); 598 599 // Emit a copy from the value into the alloca'd value in the loop preheader 600 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 601 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 602 // Load from the memory we are promoting... 603 LoadInst *LI = new LoadInst(PromotedValues[i].second, 604 PromotedValues[i].second->getName()+".promoted", 605 LoopPredInst); 606 // Store into the temporary alloca... 607 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 608 } 609 610 // Scan the basic blocks in the loop, replacing uses of our pointers with 611 // uses of the allocas in question. 612 // 613 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 614 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 615 E = LoopBBs.end(); I != E; ++I) { 616 // Rewrite all loads and stores in the block of the pointer... 617 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 618 II != E; ++II) { 619 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 620 std::map<Value*, AllocaInst*>::iterator 621 I = ValueToAllocaMap.find(L->getOperand(0)); 622 if (I != ValueToAllocaMap.end()) 623 L->setOperand(0, I->second); // Rewrite load instruction... 624 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 625 std::map<Value*, AllocaInst*>::iterator 626 I = ValueToAllocaMap.find(S->getOperand(1)); 627 if (I != ValueToAllocaMap.end()) 628 S->setOperand(1, I->second); // Rewrite store instruction... 629 } 630 } 631 } 632 633 // Now that the body of the loop uses the allocas instead of the original 634 // memory locations, insert code to copy the alloca value back into the 635 // original memory location on all exits from the loop. Note that we only 636 // want to insert one copy of the code in each exit block, though the loop may 637 // exit to the same block more than once. 638 // 639 std::set<BasicBlock*> ProcessedBlocks; 640 641 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 642 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 643 if (ProcessedBlocks.insert(ExitBlocks[i]).second) { 644 // Copy all of the allocas into their memory locations... 645 BasicBlock::iterator BI = ExitBlocks[i]->begin(); 646 while (isa<PHINode>(*BI)) 647 ++BI; // Skip over all of the phi nodes in the block... 648 Instruction *InsertPos = BI; 649 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 650 // Load from the alloca... 651 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 652 // Store into the memory we promoted... 653 new StoreInst(LI, PromotedValues[i].second, InsertPos); 654 } 655 } 656 657 // Now that we have done the deed, use the mem2reg functionality to promote 658 // all of the new allocas we just created into real SSA registers... 659 // 660 std::vector<AllocaInst*> PromotedAllocas; 661 PromotedAllocas.reserve(PromotedValues.size()); 662 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 663 PromotedAllocas.push_back(PromotedValues[i].first); 664 PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData()); 665} 666 667/// findPromotableValuesInLoop - Check the current loop for stores to definite 668/// pointers, which are not loaded and stored through may aliases. If these are 669/// found, create an alloca for the value, add it to the PromotedValues list, 670/// and keep track of the mapping from value to alloca... 671/// 672void LICM::findPromotableValuesInLoop( 673 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 674 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 675 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 676 677 // Loop over all of the alias sets in the tracker object... 678 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 679 I != E; ++I) { 680 AliasSet &AS = *I; 681 // We can promote this alias set if it has a store, if it is a "Must" alias 682 // set, if the pointer is loop invariant, if if we are not eliminating any volatile loads or stores. 683 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() && 684 !AS.isVolatile() && isLoopInvariant(AS.begin()->first)) { 685 assert(AS.begin() != AS.end() && 686 "Must alias set should have at least one pointer element in it!"); 687 Value *V = AS.begin()->first; 688 689 // Check that all of the pointers in the alias set have the same type. We 690 // cannot (yet) promote a memory location that is loaded and stored in 691 // different sizes. 692 bool PointerOk = true; 693 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 694 if (V->getType() != I->first->getType()) { 695 PointerOk = false; 696 break; 697 } 698 699 if (PointerOk) { 700 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 701 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 702 PromotedValues.push_back(std::make_pair(AI, V)); 703 704 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 705 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 706 707 DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n"); 708 } 709 } 710 } 711} 712