LICM.cpp revision 3eee6542f5c65dce299361fa5435340513cf3fe4
1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// 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 61static cl::opt<bool> 62DisablePromotion("disable-licm-promotion", cl::Hidden, 63 cl::desc("Disable memory promotion in LICM pass")); 64 65// This feature is currently disabled by default because CodeGen is not yet capable 66// of rematerializing these constants in PIC mode, so it can lead to degraded 67// performance. Compile test/CodeGen/X86/remat-constant.ll with 68// -relocation-model=pic to see an example of this. 69static cl::opt<bool> 70EnableLICMConstantMotion("enable-licm-constant-variables", cl::Hidden, 71 cl::desc("Enable hoisting/sinking of constant " 72 "global variables")); 73 74namespace { 75 struct VISIBILITY_HIDDEN LICM : public LoopPass { 76 static char ID; // Pass identification, replacement for typeid 77 LICM() : LoopPass((intptr_t)&ID) {} 78 79 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 80 81 /// This transformation requires natural loop information & requires that 82 /// loop preheaders be inserted into the CFG... 83 /// 84 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 85 AU.setPreservesCFG(); 86 AU.addRequiredID(LoopSimplifyID); 87 AU.addRequired<LoopInfo>(); 88 AU.addRequired<DominatorTree>(); 89 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 90 AU.addRequired<AliasAnalysis>(); 91 AU.addPreserved<ScalarEvolution>(); 92 AU.addPreserved<DominanceFrontier>(); 93 } 94 95 bool doFinalization() { 96 // Free the values stored in the map 97 for (std::map<Loop *, AliasSetTracker *>::iterator 98 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I) 99 delete I->second; 100 101 LoopToAliasMap.clear(); 102 return false; 103 } 104 105 private: 106 // Various analyses that we use... 107 AliasAnalysis *AA; // Current AliasAnalysis information 108 LoopInfo *LI; // Current LoopInfo 109 DominatorTree *DT; // Dominator Tree for the current Loop... 110 DominanceFrontier *DF; // Current Dominance Frontier 111 112 // State that is updated as we process loops 113 bool Changed; // Set to true when we change anything. 114 BasicBlock *Preheader; // The preheader block of the current loop... 115 Loop *CurLoop; // The current loop we are working on... 116 AliasSetTracker *CurAST; // AliasSet information for the current loop... 117 std::map<Loop *, AliasSetTracker *> LoopToAliasMap; 118 119 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 120 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 121 122 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 123 /// set. 124 void deleteAnalysisValue(Value *V, Loop *L); 125 126 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 127 /// dominated by the specified block, and that are in the current loop) in 128 /// reverse depth first order w.r.t the DominatorTree. This allows us to 129 /// visit uses before definitions, allowing us to sink a loop body in one 130 /// pass without iteration. 131 /// 132 void SinkRegion(DomTreeNode *N); 133 134 /// HoistRegion - Walk the specified region of the CFG (defined by all 135 /// blocks dominated by the specified block, and that are in the current 136 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 137 /// visit definitions before uses, allowing us to hoist a loop body in one 138 /// pass without iteration. 139 /// 140 void HoistRegion(DomTreeNode *N); 141 142 /// inSubLoop - Little predicate that returns true if the specified basic 143 /// block is in a subloop of the current one, not the current one itself. 144 /// 145 bool inSubLoop(BasicBlock *BB) { 146 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 147 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 148 if ((*I)->contains(BB)) 149 return true; // A subloop actually contains this block! 150 return false; 151 } 152 153 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 154 /// specified exit block of the loop is dominated by the specified block 155 /// that is in the body of the loop. We use these constraints to 156 /// dramatically limit the amount of the dominator tree that needs to be 157 /// searched. 158 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 159 BasicBlock *BlockInLoop) const { 160 // If the block in the loop is the loop header, it must be dominated! 161 BasicBlock *LoopHeader = CurLoop->getHeader(); 162 if (BlockInLoop == LoopHeader) 163 return true; 164 165 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 166 DomTreeNode *IDom = DT->getNode(ExitBlock); 167 168 // Because the exit block is not in the loop, we know we have to get _at 169 // least_ its immediate dominator. 170 do { 171 // Get next Immediate Dominator. 172 IDom = IDom->getIDom(); 173 174 // If we have got to the header of the loop, then the instructions block 175 // did not dominate the exit node, so we can't hoist it. 176 if (IDom->getBlock() == LoopHeader) 177 return false; 178 179 } while (IDom != BlockInLoopNode); 180 181 return true; 182 } 183 184 /// sink - When an instruction is found to only be used outside of the loop, 185 /// this function moves it to the exit blocks and patches up SSA form as 186 /// needed. 187 /// 188 void sink(Instruction &I); 189 190 /// hoist - When an instruction is found to only use loop invariant operands 191 /// that is safe to hoist, this instruction is called to do the dirty work. 192 /// 193 void hoist(Instruction &I); 194 195 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 196 /// is not a trapping instruction or if it is a trapping instruction and is 197 /// guaranteed to execute. 198 /// 199 bool isSafeToExecuteUnconditionally(Instruction &I); 200 201 /// pointerInvalidatedByLoop - Return true if the body of this loop may 202 /// store into the memory location pointed to by V. 203 /// 204 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 205 // Check to see if any of the basic blocks in CurLoop invalidate *V. 206 return CurAST->getAliasSetForPointer(V, Size).isMod(); 207 } 208 209 bool canSinkOrHoistInst(Instruction &I); 210 bool isLoopInvariantInst(Instruction &I); 211 bool isNotUsedInLoop(Instruction &I); 212 213 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 214 /// to scalars as we can. 215 /// 216 void PromoteValuesInLoop(); 217 218 /// FindPromotableValuesInLoop - Check the current loop for stores to 219 /// definite pointers, which are not loaded and stored through may aliases. 220 /// If these are found, create an alloca for the value, add it to the 221 /// PromotedValues list, and keep track of the mapping from value to 222 /// alloca... 223 /// 224 void FindPromotableValuesInLoop( 225 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 226 std::map<Value*, AllocaInst*> &Val2AlMap); 227 }; 228} 229 230char LICM::ID = 0; 231static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 232 233LoopPass *llvm::createLICMPass() { return new LICM(); } 234 235/// Hoist expressions out of the specified loop. Note, alias info for inner 236/// loop is not preserved so it is not a good idea to run LICM multiple 237/// times on one loop. 238/// 239bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 240 Changed = false; 241 242 // Get our Loop and Alias Analysis information... 243 LI = &getAnalysis<LoopInfo>(); 244 AA = &getAnalysis<AliasAnalysis>(); 245 DF = &getAnalysis<DominanceFrontier>(); 246 DT = &getAnalysis<DominatorTree>(); 247 248 CurAST = new AliasSetTracker(*AA); 249 // Collect Alias info from subloops 250 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 251 LoopItr != LoopItrE; ++LoopItr) { 252 Loop *InnerL = *LoopItr; 253 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 254 assert (InnerAST && "Where is my AST?"); 255 256 // What if InnerLoop was modified by other passes ? 257 CurAST->add(*InnerAST); 258 } 259 260 CurLoop = L; 261 262 // Get the preheader block to move instructions into... 263 Preheader = L->getLoopPreheader(); 264 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 265 266 // Loop over the body of this loop, looking for calls, invokes, and stores. 267 // Because subloops have already been incorporated into AST, we skip blocks in 268 // subloops. 269 // 270 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 271 I != E; ++I) { 272 BasicBlock *BB = *I; 273 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops... 274 CurAST->add(*BB); // Incorporate the specified basic block 275 } 276 277 // We want to visit all of the instructions in this loop... that are not parts 278 // of our subloops (they have already had their invariants hoisted out of 279 // their loop, into this loop, so there is no need to process the BODIES of 280 // the subloops). 281 // 282 // Traverse the body of the loop in depth first order on the dominator tree so 283 // that we are guaranteed to see definitions before we see uses. This allows 284 // us to sink instructions in one pass, without iteration. After sinking 285 // instructions, we perform another pass to hoist them out of the loop. 286 // 287 SinkRegion(DT->getNode(L->getHeader())); 288 HoistRegion(DT->getNode(L->getHeader())); 289 290 // Now that all loop invariants have been removed from the loop, promote any 291 // memory references to scalars that we can... 292 if (!DisablePromotion) 293 PromoteValuesInLoop(); 294 295 // Clear out loops state information for the next iteration 296 CurLoop = 0; 297 Preheader = 0; 298 299 LoopToAliasMap[L] = CurAST; 300 return Changed; 301} 302 303/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 304/// dominated by the specified block, and that are in the current loop) in 305/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 306/// uses before definitions, allowing us to sink a loop body in one pass without 307/// iteration. 308/// 309void LICM::SinkRegion(DomTreeNode *N) { 310 assert(N != 0 && "Null dominator tree node?"); 311 BasicBlock *BB = N->getBlock(); 312 313 // If this subregion is not in the top level loop at all, exit. 314 if (!CurLoop->contains(BB)) return; 315 316 // We are processing blocks in reverse dfo, so process children first... 317 const std::vector<DomTreeNode*> &Children = N->getChildren(); 318 for (unsigned i = 0, e = Children.size(); i != e; ++i) 319 SinkRegion(Children[i]); 320 321 // Only need to process the contents of this block if it is not part of a 322 // subloop (which would already have been processed). 323 if (inSubLoop(BB)) return; 324 325 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 326 Instruction &I = *--II; 327 328 // Check to see if we can sink this instruction to the exit blocks 329 // of the loop. We can do this if the all users of the instruction are 330 // outside of the loop. In this case, it doesn't even matter if the 331 // operands of the instruction are loop invariant. 332 // 333 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 334 ++II; 335 sink(I); 336 } 337 } 338} 339 340 341/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 342/// dominated by the specified block, and that are in the current loop) in depth 343/// first order w.r.t the DominatorTree. This allows us to visit definitions 344/// before uses, allowing us to hoist a loop body in one pass without iteration. 345/// 346void LICM::HoistRegion(DomTreeNode *N) { 347 assert(N != 0 && "Null dominator tree node?"); 348 BasicBlock *BB = N->getBlock(); 349 350 // If this subregion is not in the top level loop at all, exit. 351 if (!CurLoop->contains(BB)) return; 352 353 // Only need to process the contents of this block if it is not part of a 354 // subloop (which would already have been processed). 355 if (!inSubLoop(BB)) 356 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 357 Instruction &I = *II++; 358 359 // Try hoisting the instruction out to the preheader. We can only do this 360 // if all of the operands of the instruction are loop invariant and if it 361 // is safe to hoist the instruction. 362 // 363 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 364 isSafeToExecuteUnconditionally(I)) 365 hoist(I); 366 } 367 368 const std::vector<DomTreeNode*> &Children = N->getChildren(); 369 for (unsigned i = 0, e = Children.size(); i != e; ++i) 370 HoistRegion(Children[i]); 371} 372 373/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 374/// instruction. 375/// 376bool LICM::canSinkOrHoistInst(Instruction &I) { 377 // Loads have extra constraints we have to verify before we can hoist them. 378 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 379 if (LI->isVolatile()) 380 return false; // Don't hoist volatile loads! 381 382 // Loads from constant memory are always safe to move, even if they end up 383 // in the same alias set as something that ends up being modified. 384 if (EnableLICMConstantMotion && 385 AA->pointsToConstantMemory(LI->getOperand(0))) 386 return true; 387 388 // Don't hoist loads which have may-aliased stores in loop. 389 unsigned Size = 0; 390 if (LI->getType()->isSized()) 391 Size = AA->getTargetData().getTypeStoreSize(LI->getType()); 392 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 393 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 394 // Handle obvious cases efficiently. 395 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 396 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 397 return true; 398 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 399 // If this call only reads from memory and there are no writes to memory 400 // in the loop, we can hoist or sink the call as appropriate. 401 bool FoundMod = false; 402 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 403 I != E; ++I) { 404 AliasSet &AS = *I; 405 if (!AS.isForwardingAliasSet() && AS.isMod()) { 406 FoundMod = true; 407 break; 408 } 409 } 410 if (!FoundMod) return true; 411 } 412 413 // FIXME: This should use mod/ref information to see if we can hoist or sink 414 // the call. 415 416 return false; 417 } 418 419 // Otherwise these instructions are hoistable/sinkable 420 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 421 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 422 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 423 isa<ShuffleVectorInst>(I); 424} 425 426/// isNotUsedInLoop - Return true if the only users of this instruction are 427/// outside of the loop. If this is true, we can sink the instruction to the 428/// exit blocks of the loop. 429/// 430bool LICM::isNotUsedInLoop(Instruction &I) { 431 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 432 Instruction *User = cast<Instruction>(*UI); 433 if (PHINode *PN = dyn_cast<PHINode>(User)) { 434 // PHI node uses occur in predecessor blocks! 435 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 436 if (PN->getIncomingValue(i) == &I) 437 if (CurLoop->contains(PN->getIncomingBlock(i))) 438 return false; 439 } else if (CurLoop->contains(User->getParent())) { 440 return false; 441 } 442 } 443 return true; 444} 445 446 447/// isLoopInvariantInst - Return true if all operands of this instruction are 448/// loop invariant. We also filter out non-hoistable instructions here just for 449/// efficiency. 450/// 451bool LICM::isLoopInvariantInst(Instruction &I) { 452 // The instruction is loop invariant if all of its operands are loop-invariant 453 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 454 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 455 return false; 456 457 // If we got this far, the instruction is loop invariant! 458 return true; 459} 460 461/// sink - When an instruction is found to only be used outside of the loop, 462/// this function moves it to the exit blocks and patches up SSA form as needed. 463/// This method is guaranteed to remove the original instruction from its 464/// position, and may either delete it or move it to outside of the loop. 465/// 466void LICM::sink(Instruction &I) { 467 DOUT << "LICM sinking instruction: " << I; 468 469 SmallVector<BasicBlock*, 8> ExitBlocks; 470 CurLoop->getExitBlocks(ExitBlocks); 471 472 if (isa<LoadInst>(I)) ++NumMovedLoads; 473 else if (isa<CallInst>(I)) ++NumMovedCalls; 474 ++NumSunk; 475 Changed = true; 476 477 // The case where there is only a single exit node of this loop is common 478 // enough that we handle it as a special (more efficient) case. It is more 479 // efficient to handle because there are no PHI nodes that need to be placed. 480 if (ExitBlocks.size() == 1) { 481 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 482 // Instruction is not used, just delete it. 483 CurAST->deleteValue(&I); 484 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 485 I.replaceAllUsesWith(UndefValue::get(I.getType())); 486 I.eraseFromParent(); 487 } else { 488 // Move the instruction to the start of the exit block, after any PHI 489 // nodes in it. 490 I.removeFromParent(); 491 492 BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI(); 493 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 494 } 495 } else if (ExitBlocks.empty()) { 496 // The instruction is actually dead if there ARE NO exit blocks. 497 CurAST->deleteValue(&I); 498 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 499 I.replaceAllUsesWith(UndefValue::get(I.getType())); 500 I.eraseFromParent(); 501 } else { 502 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 503 // do all of the hard work of inserting PHI nodes as necessary. We convert 504 // the value into a stack object to get it to do this. 505 506 // Firstly, we create a stack object to hold the value... 507 AllocaInst *AI = 0; 508 509 if (I.getType() != Type::VoidTy) { 510 AI = new AllocaInst(I.getType(), 0, I.getName(), 511 I.getParent()->getParent()->getEntryBlock().begin()); 512 CurAST->add(AI); 513 } 514 515 // Secondly, insert load instructions for each use of the instruction 516 // outside of the loop. 517 while (!I.use_empty()) { 518 Instruction *U = cast<Instruction>(I.use_back()); 519 520 // If the user is a PHI Node, we actually have to insert load instructions 521 // in all predecessor blocks, not in the PHI block itself! 522 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 523 // Only insert into each predecessor once, so that we don't have 524 // different incoming values from the same block! 525 std::map<BasicBlock*, Value*> InsertedBlocks; 526 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 527 if (UPN->getIncomingValue(i) == &I) { 528 BasicBlock *Pred = UPN->getIncomingBlock(i); 529 Value *&PredVal = InsertedBlocks[Pred]; 530 if (!PredVal) { 531 // Insert a new load instruction right before the terminator in 532 // the predecessor block. 533 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 534 CurAST->add(cast<LoadInst>(PredVal)); 535 } 536 537 UPN->setIncomingValue(i, PredVal); 538 } 539 540 } else { 541 LoadInst *L = new LoadInst(AI, "", U); 542 U->replaceUsesOfWith(&I, L); 543 CurAST->add(L); 544 } 545 } 546 547 // Thirdly, insert a copy of the instruction in each exit block of the loop 548 // that is dominated by the instruction, storing the result into the memory 549 // location. Be careful not to insert the instruction into any particular 550 // basic block more than once. 551 std::set<BasicBlock*> InsertedBlocks; 552 BasicBlock *InstOrigBB = I.getParent(); 553 554 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 555 BasicBlock *ExitBlock = ExitBlocks[i]; 556 557 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 558 // If we haven't already processed this exit block, do so now. 559 if (InsertedBlocks.insert(ExitBlock).second) { 560 // Insert the code after the last PHI node... 561 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 562 563 // If this is the first exit block processed, just move the original 564 // instruction, otherwise clone the original instruction and insert 565 // the copy. 566 Instruction *New; 567 if (InsertedBlocks.size() == 1) { 568 I.removeFromParent(); 569 ExitBlock->getInstList().insert(InsertPt, &I); 570 New = &I; 571 } else { 572 New = I.clone(); 573 CurAST->copyValue(&I, New); 574 if (!I.getName().empty()) 575 New->setName(I.getName()+".le"); 576 ExitBlock->getInstList().insert(InsertPt, New); 577 } 578 579 // Now that we have inserted the instruction, store it into the alloca 580 if (AI) new StoreInst(New, AI, InsertPt); 581 } 582 } 583 } 584 585 // If the instruction doesn't dominate any exit blocks, it must be dead. 586 if (InsertedBlocks.empty()) { 587 CurAST->deleteValue(&I); 588 I.eraseFromParent(); 589 } 590 591 // Finally, promote the fine value to SSA form. 592 if (AI) { 593 std::vector<AllocaInst*> Allocas; 594 Allocas.push_back(AI); 595 PromoteMemToReg(Allocas, *DT, *DF, CurAST); 596 } 597 } 598} 599 600/// hoist - When an instruction is found to only use loop invariant operands 601/// that is safe to hoist, this instruction is called to do the dirty work. 602/// 603void LICM::hoist(Instruction &I) { 604 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I; 605 606 // Remove the instruction from its current basic block... but don't delete the 607 // instruction. 608 I.removeFromParent(); 609 610 // Insert the new node in Preheader, before the terminator. 611 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 612 613 if (isa<LoadInst>(I)) ++NumMovedLoads; 614 else if (isa<CallInst>(I)) ++NumMovedCalls; 615 ++NumHoisted; 616 Changed = true; 617} 618 619/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 620/// not a trapping instruction or if it is a trapping instruction and is 621/// guaranteed to execute. 622/// 623bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 624 // If it is not a trapping instruction, it is always safe to hoist. 625 if (!Inst.isTrapping()) return true; 626 627 // Otherwise we have to check to make sure that the instruction dominates all 628 // of the exit blocks. If it doesn't, then there is a path out of the loop 629 // which does not execute this instruction, so we can't hoist it. 630 631 // If the instruction is in the header block for the loop (which is very 632 // common), it is always guaranteed to dominate the exit blocks. Since this 633 // is a common case, and can save some work, check it now. 634 if (Inst.getParent() == CurLoop->getHeader()) 635 return true; 636 637 // It's always safe to load from a global or alloca. 638 if (isa<LoadInst>(Inst)) 639 if (isa<AllocationInst>(Inst.getOperand(0)) || 640 isa<GlobalVariable>(Inst.getOperand(0))) 641 return true; 642 643 // Get the exit blocks for the current loop. 644 SmallVector<BasicBlock*, 8> ExitBlocks; 645 CurLoop->getExitBlocks(ExitBlocks); 646 647 // For each exit block, get the DT node and walk up the DT until the 648 // instruction's basic block is found or we exit the loop. 649 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 650 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 651 return false; 652 653 return true; 654} 655 656 657/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 658/// stores out of the loop and moving loads to before the loop. We do this by 659/// looping over the stores in the loop, looking for stores to Must pointers 660/// which are loop invariant. We promote these memory locations to use allocas 661/// instead. These allocas can easily be raised to register values by the 662/// PromoteMem2Reg functionality. 663/// 664void LICM::PromoteValuesInLoop() { 665 // PromotedValues - List of values that are promoted out of the loop. Each 666 // value has an alloca instruction for it, and a canonical version of the 667 // pointer. 668 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 669 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 670 671 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 672 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 673 674 Changed = true; 675 NumPromoted += PromotedValues.size(); 676 677 std::vector<Value*> PointerValueNumbers; 678 679 // Emit a copy from the value into the alloca'd value in the loop preheader 680 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 681 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 682 Value *Ptr = PromotedValues[i].second; 683 684 // If we are promoting a pointer value, update alias information for the 685 // inserted load. 686 Value *LoadValue = 0; 687 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 688 // Locate a load or store through the pointer, and assign the same value 689 // to LI as we are loading or storing. Since we know that the value is 690 // stored in this loop, this will always succeed. 691 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 692 UI != E; ++UI) 693 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 694 LoadValue = LI; 695 break; 696 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 697 if (SI->getOperand(1) == Ptr) { 698 LoadValue = SI->getOperand(0); 699 break; 700 } 701 } 702 assert(LoadValue && "No store through the pointer found!"); 703 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 704 } 705 706 // Load from the memory we are promoting. 707 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 708 709 if (LoadValue) CurAST->copyValue(LoadValue, LI); 710 711 // Store into the temporary alloca. 712 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 713 } 714 715 // Scan the basic blocks in the loop, replacing uses of our pointers with 716 // uses of the allocas in question. 717 // 718 for (Loop::block_iterator I = CurLoop->block_begin(), 719 E = CurLoop->block_end(); I != E; ++I) { 720 BasicBlock *BB = *I; 721 // Rewrite all loads and stores in the block of the pointer... 722 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 723 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 724 std::map<Value*, AllocaInst*>::iterator 725 I = ValueToAllocaMap.find(L->getOperand(0)); 726 if (I != ValueToAllocaMap.end()) 727 L->setOperand(0, I->second); // Rewrite load instruction... 728 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 729 std::map<Value*, AllocaInst*>::iterator 730 I = ValueToAllocaMap.find(S->getOperand(1)); 731 if (I != ValueToAllocaMap.end()) 732 S->setOperand(1, I->second); // Rewrite store instruction... 733 } 734 } 735 } 736 737 // Now that the body of the loop uses the allocas instead of the original 738 // memory locations, insert code to copy the alloca value back into the 739 // original memory location on all exits from the loop. Note that we only 740 // want to insert one copy of the code in each exit block, though the loop may 741 // exit to the same block more than once. 742 // 743 SmallPtrSet<BasicBlock*, 16> ProcessedBlocks; 744 745 SmallVector<BasicBlock*, 8> ExitBlocks; 746 CurLoop->getExitBlocks(ExitBlocks); 747 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 748 if (!ProcessedBlocks.insert(ExitBlocks[i])) 749 continue; 750 751 // Copy all of the allocas into their memory locations. 752 BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI(); 753 Instruction *InsertPos = BI; 754 unsigned PVN = 0; 755 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 756 // Load from the alloca. 757 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 758 759 // If this is a pointer type, update alias info appropriately. 760 if (isa<PointerType>(LI->getType())) 761 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 762 763 // Store into the memory we promoted. 764 new StoreInst(LI, PromotedValues[i].second, InsertPos); 765 } 766 } 767 768 // Now that we have done the deed, use the mem2reg functionality to promote 769 // all of the new allocas we just created into real SSA registers. 770 // 771 std::vector<AllocaInst*> PromotedAllocas; 772 PromotedAllocas.reserve(PromotedValues.size()); 773 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 774 PromotedAllocas.push_back(PromotedValues[i].first); 775 PromoteMemToReg(PromotedAllocas, *DT, *DF, CurAST); 776} 777 778/// FindPromotableValuesInLoop - Check the current loop for stores to definite 779/// pointers, which are not loaded and stored through may aliases and are safe 780/// for promotion. If these are found, create an alloca for the value, add it 781/// to the PromotedValues list, and keep track of the mapping from value to 782/// alloca. 783void LICM::FindPromotableValuesInLoop( 784 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 785 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 786 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 787 788 SmallVector<BasicBlock*, 4> ExitingBlocks; 789 CurLoop->getExitingBlocks(ExitingBlocks); 790 791 // Loop over all of the alias sets in the tracker object. 792 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 793 I != E; ++I) { 794 AliasSet &AS = *I; 795 // We can promote this alias set if it has a store, if it is a "Must" alias 796 // set, if the pointer is loop invariant, and if we are not eliminating any 797 // volatile loads or stores. 798 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 799 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->first)) 800 continue; 801 802 assert(!AS.empty() && 803 "Must alias set should have at least one pointer element in it!"); 804 Value *V = AS.begin()->first; 805 806 // Check that all of the pointers in the alias set have the same type. We 807 // cannot (yet) promote a memory location that is loaded and stored in 808 // different sizes. 809 { 810 bool PointerOk = true; 811 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 812 if (V->getType() != I->first->getType()) { 813 PointerOk = false; 814 break; 815 } 816 if (!PointerOk) 817 continue; 818 } 819 820 // It isn't safe to promote a load/store from the loop if the load/store is 821 // conditional. For example, turning: 822 // 823 // for () { if (c) *P += 1; } 824 // 825 // into: 826 // 827 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 828 // 829 // is not safe, because *P may only be valid to access if 'c' is true. 830 // 831 // It is safe to promote P if all uses are direct load/stores and if at 832 // least one is guaranteed to be executed. 833 bool GuaranteedToExecute = false; 834 bool InvalidInst = false; 835 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 836 UI != UE; ++UI) { 837 // Ignore instructions not in this loop. 838 Instruction *Use = dyn_cast<Instruction>(*UI); 839 if (!Use || !CurLoop->contains(Use->getParent())) 840 continue; 841 842 if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) { 843 InvalidInst = true; 844 break; 845 } 846 847 if (!GuaranteedToExecute) 848 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 849 } 850 851 // If there is an non-load/store instruction in the loop, we can't promote 852 // it. If there isn't a guaranteed-to-execute instruction, we can't 853 // promote. 854 if (InvalidInst || !GuaranteedToExecute) 855 continue; 856 857 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 858 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 859 PromotedValues.push_back(std::make_pair(AI, V)); 860 861 // Update the AST and alias analysis. 862 CurAST->copyValue(V, AI); 863 864 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 865 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 866 867 DOUT << "LICM: Promoting value: " << *V << "\n"; 868 } 869} 870 871/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 872void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 873 AliasSetTracker *AST = LoopToAliasMap[L]; 874 if (!AST) 875 return; 876 877 AST->copyValue(From, To); 878} 879 880/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 881/// set. 882void LICM::deleteAnalysisValue(Value *V, Loop *L) { 883 AliasSetTracker *AST = LoopToAliasMap[L]; 884 if (!AST) 885 return; 886 887 AST->deleteValue(V); 888} 889