LICM.cpp revision 79066fa6acce02c97c714a5a6e151ed8a249721c
1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs loop invariant code motion, attempting to remove as much 11// code from the body of a loop as possible. It does this by either hoisting 12// code into the preheader block, or by sinking code to the exit blocks if it is 13// safe. This pass also promotes must-aliased memory locations in the loop to 14// live in registers, thus hoisting and sinking "invariant" loads and stores. 15// 16// This pass uses alias analysis for two purposes: 17// 18// 1. Moving loop invariant loads and calls out of loops. If we can determine 19// that a load or call inside of a loop never aliases anything stored to, 20// we can hoist it or sink it like any other instruction. 21// 2. Scalar Promotion of Memory - If there is a store instruction inside of 22// the loop, we try to move the store to happen AFTER the loop instead of 23// inside of the loop. This can only happen if a few conditions are true: 24// A. The pointer stored through is loop invariant 25// B. There are no stores or loads in the loop which _may_ alias the 26// pointer. There are no calls in the loop which mod/ref the pointer. 27// If these conditions are true, we can promote the loads and stores in the 28// loop of the pointer to use a temporary alloca'd variable. We then use 29// the mem2reg functionality to construct the appropriate SSA form for the 30// variable. 31// 32//===----------------------------------------------------------------------===// 33 34#define DEBUG_TYPE "licm" 35#include "llvm/Transforms/Scalar.h" 36#include "llvm/Constants.h" 37#include "llvm/DerivedTypes.h" 38#include "llvm/Instructions.h" 39#include "llvm/Target/TargetData.h" 40#include "llvm/Analysis/LoopInfo.h" 41#include "llvm/Analysis/AliasAnalysis.h" 42#include "llvm/Analysis/AliasSetTracker.h" 43#include "llvm/Analysis/Dominators.h" 44#include "llvm/Support/CFG.h" 45#include "llvm/Transforms/Utils/PromoteMemToReg.h" 46#include "llvm/Support/CommandLine.h" 47#include "llvm/Support/Debug.h" 48#include "llvm/ADT/Statistic.h" 49#include <algorithm> 50using namespace llvm; 51 52STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 53STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 54STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 55STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 56STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 57 58namespace { 59 cl::opt<bool> 60 DisablePromotion("disable-licm-promotion", cl::Hidden, 61 cl::desc("Disable memory promotion in LICM pass")); 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_ its 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, unsigned Size) { 174 // Check to see if any of the basic blocks in CurLoop invalidate *V. 175 return CurAST->getAliasSetForPointer(V, Size).isMod(); 176 } 177 178 bool canSinkOrHoistInst(Instruction &I); 179 bool isLoopInvariantInst(Instruction &I); 180 bool isNotUsedInLoop(Instruction &I); 181 182 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 183 /// to scalars as we can. 184 /// 185 void PromoteValuesInLoop(); 186 187 /// FindPromotableValuesInLoop - Check the current loop for stores to 188 /// definite pointers, which are not loaded and stored through may aliases. 189 /// If these are found, create an alloca for the value, add it to the 190 /// PromotedValues list, and keep track of the mapping from value to 191 /// alloca... 192 /// 193 void FindPromotableValuesInLoop( 194 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 195 std::map<Value*, AllocaInst*> &Val2AlMap); 196 }; 197 198 RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 199} 200 201FunctionPass *llvm::createLICMPass() { return new LICM(); } 202 203/// runOnFunction - For LICM, this simply traverses the loop structure of the 204/// function, hoisting expressions out of loops if possible. 205/// 206bool LICM::runOnFunction(Function &) { 207 Changed = false; 208 209 // Get our Loop and Alias Analysis information... 210 LI = &getAnalysis<LoopInfo>(); 211 AA = &getAnalysis<AliasAnalysis>(); 212 DF = &getAnalysis<DominanceFrontier>(); 213 DT = &getAnalysis<DominatorTree>(); 214 215 // Hoist expressions out of all of the top-level loops. 216 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) { 217 AliasSetTracker AST(*AA); 218 visitLoop(*I, AST); 219 } 220 return Changed; 221} 222 223 224/// visitLoop - Hoist expressions out of the specified loop... 225/// 226void LICM::visitLoop(Loop *L, AliasSetTracker &AST) { 227 // Recurse through all subloops before we process this loop... 228 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) { 229 AliasSetTracker SubAST(*AA); 230 visitLoop(*I, SubAST); 231 232 // Incorporate information about the subloops into this loop... 233 AST.add(SubAST); 234 } 235 CurLoop = L; 236 CurAST = &AST; 237 238 // Get the preheader block to move instructions into... 239 Preheader = L->getLoopPreheader(); 240 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 241 242 // Loop over the body of this loop, looking for calls, invokes, and stores. 243 // Because subloops have already been incorporated into AST, we skip blocks in 244 // subloops. 245 // 246 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(), 247 E = L->getBlocks().end(); I != E; ++I) 248 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops... 249 AST.add(**I); // Incorporate the specified basic block 250 251 // We want to visit all of the instructions in this loop... that are not parts 252 // of our subloops (they have already had their invariants hoisted out of 253 // their loop, into this loop, so there is no need to process the BODIES of 254 // the subloops). 255 // 256 // Traverse the body of the loop in depth first order on the dominator tree so 257 // that we are guaranteed to see definitions before we see uses. This allows 258 // us to sink instructions in one pass, without iteration. AFter sinking 259 // instructions, we perform another pass to hoist them out of the loop. 260 // 261 SinkRegion(DT->getNode(L->getHeader())); 262 HoistRegion(DT->getNode(L->getHeader())); 263 264 // Now that all loop invariants have been removed from the loop, promote any 265 // memory references to scalars that we can... 266 if (!DisablePromotion) 267 PromoteValuesInLoop(); 268 269 // Clear out loops state information for the next iteration 270 CurLoop = 0; 271 Preheader = 0; 272} 273 274/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 275/// dominated by the specified block, and that are in the current loop) in 276/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 277/// uses before definitions, allowing us to sink a loop body in one pass without 278/// iteration. 279/// 280void LICM::SinkRegion(DominatorTree::Node *N) { 281 assert(N != 0 && "Null dominator tree node?"); 282 BasicBlock *BB = N->getBlock(); 283 284 // If this subregion is not in the top level loop at all, exit. 285 if (!CurLoop->contains(BB)) return; 286 287 // We are processing blocks in reverse dfo, so process children first... 288 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 289 for (unsigned i = 0, e = Children.size(); i != e; ++i) 290 SinkRegion(Children[i]); 291 292 // Only need to process the contents of this block if it is not part of a 293 // subloop (which would already have been processed). 294 if (inSubLoop(BB)) return; 295 296 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 297 Instruction &I = *--II; 298 299 // Check to see if we can sink this instruction to the exit blocks 300 // of the loop. We can do this if the all users of the instruction are 301 // outside of the loop. In this case, it doesn't even matter if the 302 // operands of the instruction are loop invariant. 303 // 304 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 305 ++II; 306 sink(I); 307 } 308 } 309} 310 311 312/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 313/// dominated by the specified block, and that are in the current loop) in depth 314/// first order w.r.t the DominatorTree. This allows us to visit definitions 315/// before uses, allowing us to hoist a loop body in one pass without iteration. 316/// 317void LICM::HoistRegion(DominatorTree::Node *N) { 318 assert(N != 0 && "Null dominator tree node?"); 319 BasicBlock *BB = N->getBlock(); 320 321 // If this subregion is not in the top level loop at all, exit. 322 if (!CurLoop->contains(BB)) return; 323 324 // Only need to process the contents of this block if it is not part of a 325 // subloop (which would already have been processed). 326 if (!inSubLoop(BB)) 327 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 328 Instruction &I = *II++; 329 330 // Try hoisting the instruction out to the preheader. We can only do this 331 // if all of the operands of the instruction are loop invariant and if it 332 // is safe to hoist the instruction. 333 // 334 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 335 isSafeToExecuteUnconditionally(I)) 336 hoist(I); 337 } 338 339 const std::vector<DominatorTree::Node*> &Children = N->getChildren(); 340 for (unsigned i = 0, e = Children.size(); i != e; ++i) 341 HoistRegion(Children[i]); 342} 343 344/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 345/// instruction. 346/// 347bool LICM::canSinkOrHoistInst(Instruction &I) { 348 // Loads have extra constraints we have to verify before we can hoist them. 349 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 350 if (LI->isVolatile()) 351 return false; // Don't hoist volatile loads! 352 353 // Don't hoist loads which have may-aliased stores in loop. 354 unsigned Size = 0; 355 if (LI->getType()->isSized()) 356 Size = AA->getTargetData().getTypeSize(LI->getType()); 357 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 358 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 359 // Handle obvious cases efficiently. 360 if (Function *Callee = CI->getCalledFunction()) { 361 AliasAnalysis::ModRefBehavior Behavior =AA->getModRefBehavior(Callee, CI); 362 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 363 return true; 364 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 365 // If this call only reads from memory and there are no writes to memory 366 // in the loop, we can hoist or sink the call as appropriate. 367 bool FoundMod = false; 368 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 369 I != E; ++I) { 370 AliasSet &AS = *I; 371 if (!AS.isForwardingAliasSet() && AS.isMod()) { 372 FoundMod = true; 373 break; 374 } 375 } 376 if (!FoundMod) return true; 377 } 378 } 379 380 // FIXME: This should use mod/ref information to see if we can hoist or sink 381 // the call. 382 383 return false; 384 } 385 386 // Otherwise these instructions are hoistable/sinkable 387 return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) || 388 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I); 389} 390 391/// isNotUsedInLoop - Return true if the only users of this instruction are 392/// outside of the loop. If this is true, we can sink the instruction to the 393/// exit blocks of the loop. 394/// 395bool LICM::isNotUsedInLoop(Instruction &I) { 396 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 397 Instruction *User = cast<Instruction>(*UI); 398 if (PHINode *PN = dyn_cast<PHINode>(User)) { 399 // PHI node uses occur in predecessor blocks! 400 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 401 if (PN->getIncomingValue(i) == &I) 402 if (CurLoop->contains(PN->getIncomingBlock(i))) 403 return false; 404 } else if (CurLoop->contains(User->getParent())) { 405 return false; 406 } 407 } 408 return true; 409} 410 411 412/// isLoopInvariantInst - Return true if all operands of this instruction are 413/// loop invariant. We also filter out non-hoistable instructions here just for 414/// efficiency. 415/// 416bool LICM::isLoopInvariantInst(Instruction &I) { 417 // The instruction is loop invariant if all of its operands are loop-invariant 418 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 419 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 420 return false; 421 422 // If we got this far, the instruction is loop invariant! 423 return true; 424} 425 426/// sink - When an instruction is found to only be used outside of the loop, 427/// this function moves it to the exit blocks and patches up SSA form as needed. 428/// This method is guaranteed to remove the original instruction from its 429/// position, and may either delete it or move it to outside of the loop. 430/// 431void LICM::sink(Instruction &I) { 432 DOUT << "LICM sinking instruction: " << I; 433 434 std::vector<BasicBlock*> ExitBlocks; 435 CurLoop->getExitBlocks(ExitBlocks); 436 437 if (isa<LoadInst>(I)) ++NumMovedLoads; 438 else if (isa<CallInst>(I)) ++NumMovedCalls; 439 ++NumSunk; 440 Changed = true; 441 442 // The case where there is only a single exit node of this loop is common 443 // enough that we handle it as a special (more efficient) case. It is more 444 // efficient to handle because there are no PHI nodes that need to be placed. 445 if (ExitBlocks.size() == 1) { 446 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 447 // Instruction is not used, just delete it. 448 CurAST->deleteValue(&I); 449 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 450 I.replaceAllUsesWith(UndefValue::get(I.getType())); 451 I.eraseFromParent(); 452 } else { 453 // Move the instruction to the start of the exit block, after any PHI 454 // nodes in it. 455 I.removeFromParent(); 456 457 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin(); 458 while (isa<PHINode>(InsertPt)) ++InsertPt; 459 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 460 } 461 } else if (ExitBlocks.size() == 0) { 462 // The instruction is actually dead if there ARE NO exit blocks. 463 CurAST->deleteValue(&I); 464 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 465 I.replaceAllUsesWith(UndefValue::get(I.getType())); 466 I.eraseFromParent(); 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 = 0; 474 475 if (I.getType() != Type::VoidTy) 476 AI = new AllocaInst(I.getType(), 0, I.getName(), 477 I.getParent()->getParent()->front().begin()); 478 479 // Secondly, insert load instructions for each use of the instruction 480 // outside of the loop. 481 while (!I.use_empty()) { 482 Instruction *U = cast<Instruction>(I.use_back()); 483 484 // If the user is a PHI Node, we actually have to insert load instructions 485 // in all predecessor blocks, not in the PHI block itself! 486 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 487 // Only insert into each predecessor once, so that we don't have 488 // different incoming values from the same block! 489 std::map<BasicBlock*, Value*> InsertedBlocks; 490 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 491 if (UPN->getIncomingValue(i) == &I) { 492 BasicBlock *Pred = UPN->getIncomingBlock(i); 493 Value *&PredVal = InsertedBlocks[Pred]; 494 if (!PredVal) { 495 // Insert a new load instruction right before the terminator in 496 // the predecessor block. 497 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 498 } 499 500 UPN->setIncomingValue(i, PredVal); 501 } 502 503 } else { 504 LoadInst *L = new LoadInst(AI, "", U); 505 U->replaceUsesOfWith(&I, L); 506 } 507 } 508 509 // Thirdly, insert a copy of the instruction in each exit block of the loop 510 // that is dominated by the instruction, storing the result into the memory 511 // location. Be careful not to insert the instruction into any particular 512 // basic block more than once. 513 std::set<BasicBlock*> InsertedBlocks; 514 BasicBlock *InstOrigBB = I.getParent(); 515 516 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 517 BasicBlock *ExitBlock = ExitBlocks[i]; 518 519 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 520 // If we haven't already processed this exit block, do so now. 521 if (InsertedBlocks.insert(ExitBlock).second) { 522 // Insert the code after the last PHI node... 523 BasicBlock::iterator InsertPt = ExitBlock->begin(); 524 while (isa<PHINode>(InsertPt)) ++InsertPt; 525 526 // If this is the first exit block processed, just move the original 527 // instruction, otherwise clone the original instruction and insert 528 // the copy. 529 Instruction *New; 530 if (InsertedBlocks.size() == 1) { 531 I.removeFromParent(); 532 ExitBlock->getInstList().insert(InsertPt, &I); 533 New = &I; 534 } else { 535 New = I.clone(); 536 CurAST->copyValue(&I, New); 537 if (!I.getName().empty()) 538 New->setName(I.getName()+".le"); 539 ExitBlock->getInstList().insert(InsertPt, New); 540 } 541 542 // Now that we have inserted the instruction, store it into the alloca 543 if (AI) new StoreInst(New, AI, InsertPt); 544 } 545 } 546 } 547 548 // If the instruction doesn't dominate any exit blocks, it must be dead. 549 if (InsertedBlocks.empty()) { 550 CurAST->deleteValue(&I); 551 I.eraseFromParent(); 552 } 553 554 // Finally, promote the fine value to SSA form. 555 if (AI) { 556 std::vector<AllocaInst*> Allocas; 557 Allocas.push_back(AI); 558 PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData(), CurAST); 559 } 560 } 561} 562 563/// hoist - When an instruction is found to only use loop invariant operands 564/// that is safe to hoist, this instruction is called to do the dirty work. 565/// 566void LICM::hoist(Instruction &I) { 567 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I; 568 569 // Remove the instruction from its current basic block... but don't delete the 570 // instruction. 571 I.removeFromParent(); 572 573 // Insert the new node in Preheader, before the terminator. 574 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 575 576 if (isa<LoadInst>(I)) ++NumMovedLoads; 577 else if (isa<CallInst>(I)) ++NumMovedCalls; 578 ++NumHoisted; 579 Changed = true; 580} 581 582/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 583/// not a trapping instruction or if it is a trapping instruction and is 584/// guaranteed to execute. 585/// 586bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 587 // If it is not a trapping instruction, it is always safe to hoist. 588 if (!Inst.isTrapping()) return true; 589 590 // Otherwise we have to check to make sure that the instruction dominates all 591 // of the exit blocks. If it doesn't, then there is a path out of the loop 592 // which does not execute this instruction, so we can't hoist it. 593 594 // If the instruction is in the header block for the loop (which is very 595 // common), it is always guaranteed to dominate the exit blocks. Since this 596 // is a common case, and can save some work, check it now. 597 if (Inst.getParent() == CurLoop->getHeader()) 598 return true; 599 600 // It's always safe to load from a global or alloca. 601 if (isa<LoadInst>(Inst)) 602 if (isa<AllocationInst>(Inst.getOperand(0)) || 603 isa<GlobalVariable>(Inst.getOperand(0))) 604 return true; 605 606 // Get the exit blocks for the current loop. 607 std::vector<BasicBlock*> ExitBlocks; 608 CurLoop->getExitBlocks(ExitBlocks); 609 610 // For each exit block, get the DT node and walk up the DT until the 611 // instruction's basic block is found or we exit the loop. 612 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 613 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 614 return false; 615 616 return true; 617} 618 619 620/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 621/// stores out of the loop and moving loads to before the loop. We do this by 622/// looping over the stores in the loop, looking for stores to Must pointers 623/// which are loop invariant. We promote these memory locations to use allocas 624/// instead. These allocas can easily be raised to register values by the 625/// PromoteMem2Reg functionality. 626/// 627void LICM::PromoteValuesInLoop() { 628 // PromotedValues - List of values that are promoted out of the loop. Each 629 // value has an alloca instruction for it, and a canonical version of the 630 // pointer. 631 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 632 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 633 634 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 635 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 636 637 Changed = true; 638 NumPromoted += PromotedValues.size(); 639 640 std::vector<Value*> PointerValueNumbers; 641 642 // Emit a copy from the value into the alloca'd value in the loop preheader 643 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 644 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 645 Value *Ptr = PromotedValues[i].second; 646 647 // If we are promoting a pointer value, update alias information for the 648 // inserted load. 649 Value *LoadValue = 0; 650 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 651 // Locate a load or store through the pointer, and assign the same value 652 // to LI as we are loading or storing. Since we know that the value is 653 // stored in this loop, this will always succeed. 654 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 655 UI != E; ++UI) 656 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 657 LoadValue = LI; 658 break; 659 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 660 if (SI->getOperand(1) == Ptr) { 661 LoadValue = SI->getOperand(0); 662 break; 663 } 664 } 665 assert(LoadValue && "No store through the pointer found!"); 666 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 667 } 668 669 // Load from the memory we are promoting. 670 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 671 672 if (LoadValue) CurAST->copyValue(LoadValue, LI); 673 674 // Store into the temporary alloca. 675 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 676 } 677 678 // Scan the basic blocks in the loop, replacing uses of our pointers with 679 // uses of the allocas in question. 680 // 681 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks(); 682 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(), 683 E = LoopBBs.end(); I != E; ++I) { 684 // Rewrite all loads and stores in the block of the pointer... 685 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end(); 686 II != E; ++II) { 687 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 688 std::map<Value*, AllocaInst*>::iterator 689 I = ValueToAllocaMap.find(L->getOperand(0)); 690 if (I != ValueToAllocaMap.end()) 691 L->setOperand(0, I->second); // Rewrite load instruction... 692 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 693 std::map<Value*, AllocaInst*>::iterator 694 I = ValueToAllocaMap.find(S->getOperand(1)); 695 if (I != ValueToAllocaMap.end()) 696 S->setOperand(1, I->second); // Rewrite store instruction... 697 } 698 } 699 } 700 701 // Now that the body of the loop uses the allocas instead of the original 702 // memory locations, insert code to copy the alloca value back into the 703 // original memory location on all exits from the loop. Note that we only 704 // want to insert one copy of the code in each exit block, though the loop may 705 // exit to the same block more than once. 706 // 707 std::set<BasicBlock*> ProcessedBlocks; 708 709 std::vector<BasicBlock*> ExitBlocks; 710 CurLoop->getExitBlocks(ExitBlocks); 711 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 712 if (ProcessedBlocks.insert(ExitBlocks[i]).second) { 713 // Copy all of the allocas into their memory locations. 714 BasicBlock::iterator BI = ExitBlocks[i]->begin(); 715 while (isa<PHINode>(*BI)) 716 ++BI; // Skip over all of the phi nodes in the block. 717 Instruction *InsertPos = BI; 718 unsigned PVN = 0; 719 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 720 // Load from the alloca. 721 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 722 723 // If this is a pointer type, update alias info appropriately. 724 if (isa<PointerType>(LI->getType())) 725 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 726 727 // Store into the memory we promoted. 728 new StoreInst(LI, PromotedValues[i].second, InsertPos); 729 } 730 } 731 732 // Now that we have done the deed, use the mem2reg functionality to promote 733 // all of the new allocas we just created into real SSA registers. 734 // 735 std::vector<AllocaInst*> PromotedAllocas; 736 PromotedAllocas.reserve(PromotedValues.size()); 737 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 738 PromotedAllocas.push_back(PromotedValues[i].first); 739 PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData(), CurAST); 740} 741 742/// FindPromotableValuesInLoop - Check the current loop for stores to definite 743/// pointers, which are not loaded and stored through may aliases. If these are 744/// found, create an alloca for the value, add it to the PromotedValues list, 745/// and keep track of the mapping from value to alloca. 746/// 747void LICM::FindPromotableValuesInLoop( 748 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 749 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 750 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 751 752 // Loop over all of the alias sets in the tracker object. 753 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 754 I != E; ++I) { 755 AliasSet &AS = *I; 756 // We can promote this alias set if it has a store, if it is a "Must" alias 757 // set, if the pointer is loop invariant, and if we are not eliminating any 758 // volatile loads or stores. 759 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() && 760 !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) { 761 assert(AS.begin() != AS.end() && 762 "Must alias set should have at least one pointer element in it!"); 763 Value *V = AS.begin()->first; 764 765 // Check that all of the pointers in the alias set have the same type. We 766 // cannot (yet) promote a memory location that is loaded and stored in 767 // different sizes. 768 bool PointerOk = true; 769 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 770 if (V->getType() != I->first->getType()) { 771 PointerOk = false; 772 break; 773 } 774 775 if (PointerOk) { 776 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 777 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 778 PromotedValues.push_back(std::make_pair(AI, V)); 779 780 // Update the AST and alias analysis. 781 CurAST->copyValue(V, AI); 782 783 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 784 ValueToAllocaMap.insert(std::make_pair(I->first, AI)); 785 786 DOUT << "LICM: Promoting value: " << *V << "\n"; 787 } 788 } 789 } 790} 791