LICM.cpp revision 118dd0ce3d8e4b0a945afc95c9538d5005abacde
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<> NumMovedCalls("licm", "Number of call insts hoisted or sunk"); 61 Statistic<> NumPromoted("licm", 62 "Number of memory locations promoted to registers"); 63 64 struct LICM : public FunctionPass { 65 virtual bool runOnFunction(Function &F); 66 67 /// This transformation requires natural loop information & requires that 68 /// loop preheaders be inserted into the CFG... 69 /// 70 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 71 AU.setPreservesCFG(); 72 AU.addRequiredID(LoopSimplifyID); 73 AU.addRequired<LoopInfo>(); 74 AU.addRequired<DominatorTree>(); 75 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 76 AU.addRequired<AliasAnalysis>(); 77 } 78 79 private: 80 // Various analyses that we use... 81 AliasAnalysis *AA; // Current AliasAnalysis information 82 LoopInfo *LI; // Current LoopInfo 83 DominatorTree *DT; // Dominator Tree for the current Loop... 84 DominanceFrontier *DF; // Current Dominance Frontier 85 86 // State that is updated as we process loops 87 bool Changed; // Set to true when we change anything. 88 BasicBlock *Preheader; // The preheader block of the current loop... 89 Loop *CurLoop; // The current loop we are working on... 90 AliasSetTracker *CurAST; // AliasSet information for the current loop... 91 92 /// visitLoop - Hoist expressions out of the specified loop... 93 /// 94 void visitLoop(Loop *L, AliasSetTracker &AST); 95 96 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 97 /// dominated by the specified block, and that are in the current loop) in 98 /// reverse depth first order w.r.t the DominatorTree. This allows us to 99 /// visit uses before definitions, allowing us to sink a loop body in one 100 /// pass without iteration. 101 /// 102 void SinkRegion(DominatorTree::Node *N); 103 104 /// HoistRegion - Walk the specified region of the CFG (defined by all 105 /// blocks dominated by the specified block, and that are in the current 106 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 107 /// visit definitions before uses, allowing us to hoist a loop body in one 108 /// pass without iteration. 109 /// 110 void HoistRegion(DominatorTree::Node *N); 111 112 /// inSubLoop - Little predicate that returns true if the specified basic 113 /// block is in a subloop of the current one, not the current one itself. 114 /// 115 bool inSubLoop(BasicBlock *BB) { 116 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 117 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 118 if ((*I)->contains(BB)) 119 return true; // A subloop actually contains this block! 120 return false; 121 } 122 123 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 124 /// specified exit block of the loop is dominated by the specified block 125 /// that is in the body of the loop. We use these constraints to 126 /// dramatically limit the amount of the dominator tree that needs to be 127 /// searched. 128 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 129 BasicBlock *BlockInLoop) const { 130 // If the block in the loop is the loop header, it must be dominated! 131 BasicBlock *LoopHeader = CurLoop->getHeader(); 132 if (BlockInLoop == LoopHeader) 133 return true; 134 135 DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop); 136 DominatorTree::Node *IDom = DT->getNode(ExitBlock); 137 138 // Because the exit block is not in the loop, we know we have to get _at 139 // least_ it's immediate dominator. 140 do { 141 // Get next Immediate Dominator. 142 IDom = IDom->getIDom(); 143 144 // If we have got to the header of the loop, then the instructions block 145 // did not dominate the exit node, so we can't hoist it. 146 if (IDom->getBlock() == LoopHeader) 147 return false; 148 149 } while (IDom != BlockInLoopNode); 150 151 return true; 152 } 153 154 /// sink - When an instruction is found to only be used outside of the loop, 155 /// this function moves it to the exit blocks and patches up SSA form as 156 /// needed. 157 /// 158 void sink(Instruction &I); 159 160 /// hoist - When an instruction is found to only use loop invariant operands 161 /// that is safe to hoist, this instruction is called to do the dirty work. 162 /// 163 void hoist(Instruction &I); 164 165 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 166 /// is not a trapping instruction or if it is a trapping instruction and is 167 /// guaranteed to execute. 168 /// 169 bool isSafeToExecuteUnconditionally(Instruction &I); 170 171 /// pointerInvalidatedByLoop - Return true if the body of this loop may 172 /// store into the memory location pointed to by V. 173 /// 174 bool pointerInvalidatedByLoop(Value *V) { 175 // Check to see if any of the basic blocks in CurLoop invalidate *V. 176 return CurAST->getAliasSetForPointer(V, 0).isMod(); 177 } 178 179 /// isLoopInvariant - Return true if the specified value is loop invariant 180 /// 181 inline bool isLoopInvariant(Value *V) { 182 if (Instruction *I = dyn_cast<Instruction>(V)) 183 return !CurLoop->contains(I->getParent()); 184 return true; // All non-instructions are loop invariant 185 } 186 187 bool canSinkOrHoistInst(Instruction &I); 188 bool isLoopInvariantInst(Instruction &I); 189 bool isNotUsedInLoop(Instruction &I); 190 191 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 192 /// to scalars as we can. 193 /// 194 void PromoteValuesInLoop(); 195 196 /// findPromotableValuesInLoop - Check the current loop for stores to 197 /// definite pointers, which are not loaded and stored through may aliases. 198 /// If these are found, create an alloca for the value, add it to the 199 /// PromotedValues list, and keep track of the mapping from value to 200 /// alloca... 201 /// 202 void findPromotableValuesInLoop( 203 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 204 std::map<Value*, AllocaInst*> &Val2AlMap); 205 }; 206 207 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion"); 208} 209 210FunctionPass *llvm::createLICMPass() { return new LICM(); } 211 212/// runOnFunction - For LICM, this simply traverses the loop structure of the 213/// function, hoisting expressions out of loops if possible. 214/// 215bool LICM::runOnFunction(Function &) { 216 Changed = false; 217 218 // Get our Loop and Alias Analysis information... 219 LI = &getAnalysis<LoopInfo>(); 220 AA = &getAnalysis<AliasAnalysis>(); 221 DF = &getAnalysis<DominanceFrontier>(); 222 DT = &getAnalysis<DominatorTree>(); 223 224 // Hoist expressions out of all of the top-level loops. 225 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) { 226 AliasSetTracker AST(*AA); 227 visitLoop(*I, AST); 228 } 229 return Changed; 230} 231 232 233/// visitLoop - Hoist expressions out of the specified loop... 234/// 235void LICM::visitLoop(Loop *L, AliasSetTracker &AST) { 236 // Recurse through all subloops before we process this loop... 237 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) { 238 AliasSetTracker SubAST(*AA); 239 visitLoop(*I, SubAST); 240 241 // Incorporate information about the subloops into this loop... 242 AST.add(SubAST); 243 } 244 CurLoop = L; 245 CurAST = &AST; 246 247 // Get the preheader block to move instructions into... 248 Preheader = L->getLoopPreheader(); 249 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 250 251 // Loop over the body of this loop, looking for calls, invokes, and stores. 252 // Because subloops have already been incorporated into AST, we skip blocks in 253 // subloops. 254 // 255 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(), 256 E = L->getBlocks().end(); I != E; ++I) 257 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops... 258 AST.add(**I); // Incorporate the specified basic block 259 260 // We want to visit all of the instructions in this loop... that are not parts 261 // of our subloops (they have already had their invariants hoisted out of 262 // their loop, into this loop, so there is no need to process the BODIES of 263 // the subloops). 264 // 265 // Traverse the body of the loop in depth first order on the dominator tree so 266 // that we are guaranteed to see definitions before we see uses. This allows 267 // us to sink instructions in one pass, without iteration. AFter sinking 268 // instructions, we perform another pass to hoist them out of the loop. 269 // 270 SinkRegion(DT->getNode(L->getHeader())); 271 HoistRegion(DT->getNode(L->getHeader())); 272 273 // Now that all loop invariants have been removed from the loop, promote any 274 // memory references to scalars that we can... 275 if (!DisablePromotion) 276 PromoteValuesInLoop(); 277 278 // Clear out loops state information for the next iteration 279 CurLoop = 0; 280 Preheader = 0; 281} 282 283/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 284/// dominated by the specified block, and that are in the current loop) in 285/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 286/// uses before definitions, allowing us to sink a loop body in one pass without 287/// iteration. 288/// 289void LICM::SinkRegion(DominatorTree::Node *N) { 290 assert(N != 0 && "Null dominator tree node?"); 291 BasicBlock *BB = N->getBlock(); 292 293 // If this subregion is not in the top level loop at all, exit. 294 if (!CurLoop->contains(BB)) return; 295 296 // We are processing blocks in reverse dfo, so process children first... 297 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 298 for (unsigned i = 0, e = Children.size(); i != e; ++i) 299 SinkRegion(Children[i]); 300 301 // Only need to process the contents of this block if it is not part of a 302 // subloop (which would already have been processed). 303 if (inSubLoop(BB)) return; 304 305 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 306 Instruction &I = *--II; 307 308 // Check to see if we can sink this instruction to the exit blocks 309 // of the loop. We can do this if the all users of the instruction are 310 // outside of the loop. In this case, it doesn't even matter if the 311 // operands of the instruction are loop invariant. 312 // 313 if (canSinkOrHoistInst(I) && isNotUsedInLoop(I)) { 314 ++II; 315 sink(I); 316 } 317 } 318} 319 320 321/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 322/// dominated by the specified block, and that are in the current loop) in depth 323/// first order w.r.t the DominatorTree. This allows us to visit definitions 324/// before uses, allowing us to hoist a loop body in one pass without iteration. 325/// 326void LICM::HoistRegion(DominatorTree::Node *N) { 327 assert(N != 0 && "Null dominator tree node?"); 328 BasicBlock *BB = N->getBlock(); 329 330 // If this subregion is not in the top level loop at all, exit. 331 if (!CurLoop->contains(BB)) return; 332 333 // Only need to process the contents of this block if it is not part of a 334 // subloop (which would already have been processed). 335 if (!inSubLoop(BB)) 336 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 337 Instruction &I = *II++; 338 339 // Try hoisting the instruction out to the preheader. We can only do this 340 // if all of the operands of the instruction are loop invariant and if it 341 // is safe to hoist the instruction. 342 // 343 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 344 isSafeToExecuteUnconditionally(I)) 345 hoist(I); 346 } 347 348 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 349 for (unsigned i = 0, e = Children.size(); i != e; ++i) 350 HoistRegion(Children[i]); 351} 352 353/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 354/// instruction. 355/// 356bool LICM::canSinkOrHoistInst(Instruction &I) { 357 // Loads have extra constraints we have to verify before we can hoist them. 358 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 359 if (LI->isVolatile()) 360 return false; // Don't hoist volatile loads! 361 362 // Don't hoist loads which have may-aliased stores in loop. 363 return !pointerInvalidatedByLoop(LI->getOperand(0)); 364 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 365 // Handle obvious cases efficiently. 366 if (Function *Callee = CI->getCalledFunction()) { 367 if (AA->doesNotAccessMemory(Callee)) 368 return true; 369 else if (AA->onlyReadsMemory(Callee)) { 370 // If this call only reads from memory and there are no writes to memory 371 // in the loop, we can hoist or sink the call as appropriate. 372 bool FoundMod = false; 373 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 374 I != E; ++I) { 375 AliasSet &AS = *I; 376 if (!AS.isForwardingAliasSet() && AS.isMod()) { 377 FoundMod = true; 378 break; 379 } 380 } 381 if (!FoundMod) return true; 382 } 383 } 384 385 // FIXME: This should use mod/ref information to see if we can hoist or sink 386 // the call. 387 388 return false; 389 } 390 391 return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) || 392 isa<SelectInst>(I) || 393 isa<GetElementPtrInst>(I) || isa<VANextInst>(I) || isa<VAArgInst>(I); 394} 395 396/// isNotUsedInLoop - Return true if the only users of this instruction are 397/// outside of the loop. If this is true, we can sink the instruction to the 398/// exit blocks of the loop. 399/// 400bool LICM::isNotUsedInLoop(Instruction &I) { 401 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 402 Instruction *User = cast<Instruction>(*UI); 403 if (PHINode *PN = dyn_cast<PHINode>(User)) { 404 // PHI node uses occur in predecessor blocks! 405 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 406 if (PN->getIncomingValue(i) == &I) 407 if (CurLoop->contains(PN->getIncomingBlock(i))) 408 return false; 409 } else if (CurLoop->contains(User->getParent())) { 410 return false; 411 } 412 } 413 return true; 414} 415 416 417/// isLoopInvariantInst - Return true if all operands of this instruction are 418/// loop invariant. We also filter out non-hoistable instructions here just for 419/// efficiency. 420/// 421bool LICM::isLoopInvariantInst(Instruction &I) { 422 // The instruction is loop invariant if all of its operands are loop-invariant 423 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 424 if (!isLoopInvariant(I.getOperand(i))) 425 return false; 426 427 // If we got this far, the instruction is loop invariant! 428 return true; 429} 430 431/// sink - When an instruction is found to only be used outside of the loop, 432/// this function moves it to the exit blocks and patches up SSA form as needed. 433/// This method is guaranteed to remove the original instruction from its 434/// position, and may either delete it or move it to outside of the loop. 435/// 436void LICM::sink(Instruction &I) { 437 DEBUG(std::cerr << "LICM sinking instruction: " << I); 438 439 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 440 441 if (isa<LoadInst>(I)) ++NumMovedLoads; 442 else if (isa<CallInst>(I)) ++NumMovedCalls; 443 ++NumSunk; 444 Changed = true; 445 446 // The case where there is only a single exit node of this loop is common 447 // enough that we handle it as a special (more efficient) case. It is more 448 // efficient to handle because there are no PHI nodes that need to be placed. 449 if (ExitBlocks.size() == 1) { 450 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 451 // Instruction is not used, just delete it. 452 CurAST->remove(&I); 453 I.getParent()->getInstList().erase(&I); 454 } else { 455 // Move the instruction to the start of the exit block, after any PHI 456 // nodes in it. 457 I.getParent()->getInstList().remove(&I); 458 459 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin(); 460 while (isa<PHINode>(InsertPt)) ++InsertPt; 461 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 462 } 463 } else if (ExitBlocks.size() == 0) { 464 // The instruction is actually dead if there ARE NO exit blocks. 465 CurAST->remove(&I); 466 I.getParent()->getInstList().erase(&I); 467 } else { 468 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 469 // do all of the hard work of inserting PHI nodes as necessary. We convert 470 // the value into a stack object to get it to do this. 471 472 // Firstly, we create a stack object to hold the value... 473 AllocaInst *AI = new AllocaInst(I.getType(), 0, I.getName(), 474 I.getParent()->getParent()->front().begin()); 475 476 // Secondly, insert load instructions for each use of the instruction 477 // outside of the loop. 478 while (!I.use_empty()) { 479 Instruction *U = cast<Instruction>(I.use_back()); 480 481 // If the user is a PHI Node, we actually have to insert load instructions 482 // in all predecessor blocks, not in the PHI block itself! 483 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 484 // Only insert into each predecessor once, so that we don't have 485 // different incoming values from the same block! 486 std::map<BasicBlock*, Value*> InsertedBlocks; 487 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 488 if (UPN->getIncomingValue(i) == &I) { 489 BasicBlock *Pred = UPN->getIncomingBlock(i); 490 Value *&PredVal = InsertedBlocks[Pred]; 491 if (!PredVal) { 492 // Insert a new load instruction right before the terminator in 493 // the predecessor block. 494 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 495 } 496 497 UPN->setIncomingValue(i, PredVal); 498 } 499 500 } else { 501 LoadInst *L = new LoadInst(AI, "", U); 502 U->replaceUsesOfWith(&I, L); 503 } 504 } 505 506 // Thirdly, insert a copy of the instruction in each exit block of the loop 507 // that is dominated by the instruction, storing the result into the memory 508 // location. Be careful not to insert the instruction into any particular 509 // basic block more than once. 510 std::set<BasicBlock*> InsertedBlocks; 511 BasicBlock *InstOrigBB = I.getParent(); 512 513 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 514 BasicBlock *ExitBlock = ExitBlocks[i]; 515 516 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 517 // If we haven't already processed this exit block, do so now. 518 if (InsertedBlocks.insert(ExitBlock).second) { 519 // Insert the code after the last PHI node... 520 BasicBlock::iterator InsertPt = ExitBlock->begin(); 521 while (isa<PHINode>(InsertPt)) ++InsertPt; 522 523 // If this is the first exit block processed, just move the original 524 // instruction, otherwise clone the original instruction and insert 525 // the copy. 526 Instruction *New; 527 if (InsertedBlocks.size() == 1) { 528 I.getParent()->getInstList().remove(&I); 529 ExitBlock->getInstList().insert(InsertPt, &I); 530 New = &I; 531 } else { 532 New = I.clone(); 533 New->setName(I.getName()+".le"); 534 ExitBlock->getInstList().insert(InsertPt, New); 535 } 536 537 // Now that we have inserted the instruction, store it into the alloca 538 new StoreInst(New, AI, InsertPt); 539 } 540 } 541 } 542 543 // If the instruction doesn't dominate any exit blocks, it must be dead. 544 if (InsertedBlocks.empty()) { 545 CurAST->remove(&I); 546 I.getParent()->getInstList().erase(&I); 547 } 548 549 // Finally, promote the fine value to SSA form. 550 std::vector<AllocaInst*> Allocas; 551 Allocas.push_back(AI); 552 PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData()); 553 } 554} 555 556/// hoist - When an instruction is found to only use loop invariant operands 557/// that is safe to hoist, this instruction is called to do the dirty work. 558/// 559void LICM::hoist(Instruction &I) { 560 DEBUG(std::cerr << "LICM hoisting to"; 561 WriteAsOperand(std::cerr, Preheader, false); 562 std::cerr << ": " << I); 563 564 // Remove the instruction from its current basic block... but don't delete the 565 // instruction. 566 I.getParent()->getInstList().remove(&I); 567 568 // Insert the new node in Preheader, before the terminator. 569 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 570 571 if (isa<LoadInst>(I)) ++NumMovedLoads; 572 else if (isa<CallInst>(I)) ++NumMovedCalls; 573 ++NumHoisted; 574 Changed = true; 575} 576 577/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 578/// not a trapping instruction or if it is a trapping instruction and is 579/// guaranteed to execute. 580/// 581bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 582 // If it is not a trapping instruction, it is always safe to hoist. 583 if (!Inst.isTrapping()) return true; 584 585 // Otherwise we have to check to make sure that the instruction dominates all 586 // of the exit blocks. If it doesn't, then there is a path out of the loop 587 // which does not execute this instruction, so we can't hoist it. 588 589 // If the instruction is in the header block for the loop (which is very 590 // common), it is always guaranteed to dominate the exit blocks. Since this 591 // is a common case, and can save some work, check it now. 592 if (Inst.getParent() == CurLoop->getHeader()) 593 return true; 594 595 // Get the exit blocks for the current loop. 596 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 597 598 // For each exit block, get the DT node and walk up the DT until the 599 // instruction's basic block is found or we exit the loop. 600 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 601 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 602 return false; 603 604 return true; 605} 606 607 608/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 609/// stores out of the loop and moving loads to before the loop. We do this by 610/// looping over the stores in the loop, looking for stores to Must pointers 611/// which are loop invariant. We promote these memory locations to use allocas 612/// instead. These allocas can easily be raised to register values by the 613/// PromoteMem2Reg functionality. 614/// 615void LICM::PromoteValuesInLoop() { 616 // PromotedValues - List of values that are promoted out of the loop. Each 617 // value has an alloca instruction for it, and a canonical version of the 618 // pointer. 619 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 620 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 621 622 findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 623 if (ValueToAllocaMap.empty()) return; // If there are values to promote... 624 625 Changed = true; 626 NumPromoted += PromotedValues.size(); 627 628 // Emit a copy from the value into the alloca'd value in the loop preheader 629 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 630 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 631 // Load from the memory we are promoting... 632 LoadInst *LI = new LoadInst(PromotedValues[i].second, 633 PromotedValues[i].second->getName()+".promoted", 634 LoopPredInst); 635 // Store into the temporary alloca... 636 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 637 } 638 639 // Scan the basic blocks in the loop, replacing uses of our pointers with 640 // uses of the allocas in question. 641 // 642 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 643 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 644 E = LoopBBs.end(); I != E; ++I) { 645 // Rewrite all loads and stores in the block of the pointer... 646 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 647 II != E; ++II) { 648 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 649 std::map<Value*, AllocaInst*>::iterator 650 I = ValueToAllocaMap.find(L->getOperand(0)); 651 if (I != ValueToAllocaMap.end()) 652 L->setOperand(0, I->second); // Rewrite load instruction... 653 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 654 std::map<Value*, AllocaInst*>::iterator 655 I = ValueToAllocaMap.find(S->getOperand(1)); 656 if (I != ValueToAllocaMap.end()) 657 S->setOperand(1, I->second); // Rewrite store instruction... 658 } 659 } 660 } 661 662 // Now that the body of the loop uses the allocas instead of the original 663 // memory locations, insert code to copy the alloca value back into the 664 // original memory location on all exits from the loop. Note that we only 665 // want to insert one copy of the code in each exit block, though the loop may 666 // exit to the same block more than once. 667 // 668 std::set<BasicBlock*> ProcessedBlocks; 669 670 const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks(); 671 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 672 if (ProcessedBlocks.insert(ExitBlocks[i]).second) { 673 // Copy all of the allocas into their memory locations... 674 BasicBlock::iterator BI = ExitBlocks[i]->begin(); 675 while (isa<PHINode>(*BI)) 676 ++BI; // Skip over all of the phi nodes in the block... 677 Instruction *InsertPos = BI; 678 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 679 // Load from the alloca... 680 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 681 // Store into the memory we promoted... 682 new StoreInst(LI, PromotedValues[i].second, InsertPos); 683 } 684 } 685 686 // Now that we have done the deed, use the mem2reg functionality to promote 687 // all of the new allocas we just created into real SSA registers... 688 // 689 std::vector<AllocaInst*> PromotedAllocas; 690 PromotedAllocas.reserve(PromotedValues.size()); 691 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 692 PromotedAllocas.push_back(PromotedValues[i].first); 693 PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData()); 694} 695 696/// findPromotableValuesInLoop - Check the current loop for stores to definite 697/// pointers, which are not loaded and stored through may aliases. If these are 698/// found, create an alloca for the value, add it to the PromotedValues list, 699/// and keep track of the mapping from value to alloca... 700/// 701void LICM::findPromotableValuesInLoop( 702 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 703 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 704 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 705 706 // Loop over all of the alias sets in the tracker object... 707 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 708 I != E; ++I) { 709 AliasSet &AS = *I; 710 // We can promote this alias set if it has a store, if it is a "Must" alias 711 // set, if the pointer is loop invariant, if if we are not eliminating any 712 // volatile loads or stores. 713 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() && 714 !AS.isVolatile() && isLoopInvariant(AS.begin()->first)) { 715 assert(AS.begin() != AS.end() && 716 "Must alias set should have at least one pointer element in it!"); 717 Value *V = AS.begin()->first; 718 719 // Check that all of the pointers in the alias set have the same type. We 720 // cannot (yet) promote a memory location that is loaded and stored in 721 // different sizes. 722 bool PointerOk = true; 723 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 724 if (V->getType() != I->first->getType()) { 725 PointerOk = false; 726 break; 727 } 728 729 if (PointerOk) { 730 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 731 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 732 PromotedValues.push_back(std::make_pair(AI, V)); 733 734 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 735 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 736 737 DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n"); 738 } 739 } 740 } 741} 742