LLVM API Documentation
00001 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This pass performs loop invariant code motion, attempting to remove as much 00011 // code from the body of a loop as possible. It does this by either hoisting 00012 // code into the preheader block, or by sinking code to the exit blocks if it is 00013 // safe. This pass also promotes must-aliased memory locations in the loop to 00014 // live in registers, thus hoisting and sinking "invariant" loads and stores. 00015 // 00016 // This pass uses alias analysis for two purposes: 00017 // 00018 // 1. Moving loop invariant loads and calls out of loops. If we can determine 00019 // that a load or call inside of a loop never aliases anything stored to, 00020 // we can hoist it or sink it like any other instruction. 00021 // 2. Scalar Promotion of Memory - If there is a store instruction inside of 00022 // the loop, we try to move the store to happen AFTER the loop instead of 00023 // inside of the loop. This can only happen if a few conditions are true: 00024 // A. The pointer stored through is loop invariant 00025 // B. There are no stores or loads in the loop which _may_ alias the 00026 // pointer. There are no calls in the loop which mod/ref the pointer. 00027 // If these conditions are true, we can promote the loads and stores in the 00028 // loop of the pointer to use a temporary alloca'd variable. We then use 00029 // the mem2reg functionality to construct the appropriate SSA form for the 00030 // variable. 00031 // 00032 //===----------------------------------------------------------------------===// 00033 00034 #define DEBUG_TYPE "licm" 00035 #include "llvm/Transforms/Scalar.h" 00036 #include "llvm/Constants.h" 00037 #include "llvm/DerivedTypes.h" 00038 #include "llvm/Instructions.h" 00039 #include "llvm/LLVMContext.h" 00040 #include "llvm/Target/TargetData.h" 00041 #include "llvm/Analysis/LoopInfo.h" 00042 #include "llvm/Analysis/LoopPass.h" 00043 #include "llvm/Analysis/AliasAnalysis.h" 00044 #include "llvm/Analysis/AliasSetTracker.h" 00045 #include "llvm/Analysis/Dominators.h" 00046 #include "llvm/Analysis/ScalarEvolution.h" 00047 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 00048 #include "llvm/Support/CFG.h" 00049 #include "llvm/Support/Compiler.h" 00050 #include "llvm/Support/CommandLine.h" 00051 #include "llvm/Support/raw_ostream.h" 00052 #include "llvm/Support/Debug.h" 00053 #include "llvm/ADT/Statistic.h" 00054 #include <algorithm> 00055 using namespace llvm; 00056 00057 STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 00058 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 00059 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 00060 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 00061 STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 00062 00063 static cl::opt<bool> 00064 DisablePromotion("disable-licm-promotion", cl::Hidden, 00065 cl::desc("Disable memory promotion in LICM pass")); 00066 00067 // This feature is currently disabled by default because CodeGen is not yet 00068 // capable of rematerializing these constants in PIC mode, so it can lead to 00069 // degraded performance. Compile test/CodeGen/X86/remat-constant.ll with 00070 // -relocation-model=pic to see an example of this. 00071 static cl::opt<bool> 00072 EnableLICMConstantMotion("enable-licm-constant-variables", cl::Hidden, 00073 cl::desc("Enable hoisting/sinking of constant " 00074 "global variables")); 00075 00076 namespace { 00077 struct VISIBILITY_HIDDEN LICM : public LoopPass { 00078 static char ID; // Pass identification, replacement for typeid 00079 LICM() : LoopPass(&ID) {} 00080 00081 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 00082 00083 /// This transformation requires natural loop information & requires that 00084 /// loop preheaders be inserted into the CFG... 00085 /// 00086 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 00087 AU.setPreservesCFG(); 00088 AU.addRequiredID(LoopSimplifyID); 00089 AU.addRequired<LoopInfo>(); 00090 AU.addRequired<DominatorTree>(); 00091 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg) 00092 AU.addRequired<AliasAnalysis>(); 00093 AU.addPreserved<ScalarEvolution>(); 00094 AU.addPreserved<DominanceFrontier>(); 00095 } 00096 00097 bool doFinalization() { 00098 // Free the values stored in the map 00099 for (std::map<Loop *, AliasSetTracker *>::iterator 00100 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I) 00101 delete I->second; 00102 00103 LoopToAliasMap.clear(); 00104 return false; 00105 } 00106 00107 private: 00108 // Various analyses that we use... 00109 AliasAnalysis *AA; // Current AliasAnalysis information 00110 LoopInfo *LI; // Current LoopInfo 00111 DominatorTree *DT; // Dominator Tree for the current Loop... 00112 DominanceFrontier *DF; // Current Dominance Frontier 00113 00114 // State that is updated as we process loops 00115 bool Changed; // Set to true when we change anything. 00116 BasicBlock *Preheader; // The preheader block of the current loop... 00117 Loop *CurLoop; // The current loop we are working on... 00118 AliasSetTracker *CurAST; // AliasSet information for the current loop... 00119 std::map<Loop *, AliasSetTracker *> LoopToAliasMap; 00120 00121 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 00122 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L); 00123 00124 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 00125 /// set. 00126 void deleteAnalysisValue(Value *V, Loop *L); 00127 00128 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 00129 /// dominated by the specified block, and that are in the current loop) in 00130 /// reverse depth first order w.r.t the DominatorTree. This allows us to 00131 /// visit uses before definitions, allowing us to sink a loop body in one 00132 /// pass without iteration. 00133 /// 00134 void SinkRegion(DomTreeNode *N); 00135 00136 /// HoistRegion - Walk the specified region of the CFG (defined by all 00137 /// blocks dominated by the specified block, and that are in the current 00138 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 00139 /// visit definitions before uses, allowing us to hoist a loop body in one 00140 /// pass without iteration. 00141 /// 00142 void HoistRegion(DomTreeNode *N); 00143 00144 /// inSubLoop - Little predicate that returns true if the specified basic 00145 /// block is in a subloop of the current one, not the current one itself. 00146 /// 00147 bool inSubLoop(BasicBlock *BB) { 00148 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 00149 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I) 00150 if ((*I)->contains(BB)) 00151 return true; // A subloop actually contains this block! 00152 return false; 00153 } 00154 00155 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the 00156 /// specified exit block of the loop is dominated by the specified block 00157 /// that is in the body of the loop. We use these constraints to 00158 /// dramatically limit the amount of the dominator tree that needs to be 00159 /// searched. 00160 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock, 00161 BasicBlock *BlockInLoop) const { 00162 // If the block in the loop is the loop header, it must be dominated! 00163 BasicBlock *LoopHeader = CurLoop->getHeader(); 00164 if (BlockInLoop == LoopHeader) 00165 return true; 00166 00167 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop); 00168 DomTreeNode *IDom = DT->getNode(ExitBlock); 00169 00170 // Because the exit block is not in the loop, we know we have to get _at 00171 // least_ its immediate dominator. 00172 do { 00173 // Get next Immediate Dominator. 00174 IDom = IDom->getIDom(); 00175 00176 // If we have got to the header of the loop, then the instructions block 00177 // did not dominate the exit node, so we can't hoist it. 00178 if (IDom->getBlock() == LoopHeader) 00179 return false; 00180 00181 } while (IDom != BlockInLoopNode); 00182 00183 return true; 00184 } 00185 00186 /// sink - When an instruction is found to only be used outside of the loop, 00187 /// this function moves it to the exit blocks and patches up SSA form as 00188 /// needed. 00189 /// 00190 void sink(Instruction &I); 00191 00192 /// hoist - When an instruction is found to only use loop invariant operands 00193 /// that is safe to hoist, this instruction is called to do the dirty work. 00194 /// 00195 void hoist(Instruction &I); 00196 00197 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 00198 /// is not a trapping instruction or if it is a trapping instruction and is 00199 /// guaranteed to execute. 00200 /// 00201 bool isSafeToExecuteUnconditionally(Instruction &I); 00202 00203 /// pointerInvalidatedByLoop - Return true if the body of this loop may 00204 /// store into the memory location pointed to by V. 00205 /// 00206 bool pointerInvalidatedByLoop(Value *V, unsigned Size) { 00207 // Check to see if any of the basic blocks in CurLoop invalidate *V. 00208 return CurAST->getAliasSetForPointer(V, Size).isMod(); 00209 } 00210 00211 bool canSinkOrHoistInst(Instruction &I); 00212 bool isLoopInvariantInst(Instruction &I); 00213 bool isNotUsedInLoop(Instruction &I); 00214 00215 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many 00216 /// to scalars as we can. 00217 /// 00218 void PromoteValuesInLoop(); 00219 00220 /// FindPromotableValuesInLoop - Check the current loop for stores to 00221 /// definite pointers, which are not loaded and stored through may aliases. 00222 /// If these are found, create an alloca for the value, add it to the 00223 /// PromotedValues list, and keep track of the mapping from value to 00224 /// alloca... 00225 /// 00226 void FindPromotableValuesInLoop( 00227 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 00228 std::map<Value*, AllocaInst*> &Val2AlMap); 00229 }; 00230 } 00231 00232 char LICM::ID = 0; 00233 static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion"); 00234 00235 Pass *llvm::createLICMPass() { return new LICM(); } 00236 00237 /// Hoist expressions out of the specified loop. Note, alias info for inner 00238 /// loop is not preserved so it is not a good idea to run LICM multiple 00239 /// times on one loop. 00240 /// 00241 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 00242 Changed = false; 00243 00244 // Get our Loop and Alias Analysis information... 00245 LI = &getAnalysis<LoopInfo>(); 00246 AA = &getAnalysis<AliasAnalysis>(); 00247 DF = &getAnalysis<DominanceFrontier>(); 00248 DT = &getAnalysis<DominatorTree>(); 00249 00250 CurAST = new AliasSetTracker(*AA); 00251 // Collect Alias info from subloops 00252 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 00253 LoopItr != LoopItrE; ++LoopItr) { 00254 Loop *InnerL = *LoopItr; 00255 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL]; 00256 assert (InnerAST && "Where is my AST?"); 00257 00258 // What if InnerLoop was modified by other passes ? 00259 CurAST->add(*InnerAST); 00260 } 00261 00262 CurLoop = L; 00263 00264 // Get the preheader block to move instructions into... 00265 Preheader = L->getLoopPreheader(); 00266 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!"); 00267 00268 // Loop over the body of this loop, looking for calls, invokes, and stores. 00269 // Because subloops have already been incorporated into AST, we skip blocks in 00270 // subloops. 00271 // 00272 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 00273 I != E; ++I) { 00274 BasicBlock *BB = *I; 00275 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops... 00276 CurAST->add(*BB); // Incorporate the specified basic block 00277 } 00278 00279 // We want to visit all of the instructions in this loop... that are not parts 00280 // of our subloops (they have already had their invariants hoisted out of 00281 // their loop, into this loop, so there is no need to process the BODIES of 00282 // the subloops). 00283 // 00284 // Traverse the body of the loop in depth first order on the dominator tree so 00285 // that we are guaranteed to see definitions before we see uses. This allows 00286 // us to sink instructions in one pass, without iteration. After sinking 00287 // instructions, we perform another pass to hoist them out of the loop. 00288 // 00289 SinkRegion(DT->getNode(L->getHeader())); 00290 HoistRegion(DT->getNode(L->getHeader())); 00291 00292 // Now that all loop invariants have been removed from the loop, promote any 00293 // memory references to scalars that we can... 00294 if (!DisablePromotion) 00295 PromoteValuesInLoop(); 00296 00297 // Clear out loops state information for the next iteration 00298 CurLoop = 0; 00299 Preheader = 0; 00300 00301 LoopToAliasMap[L] = CurAST; 00302 return Changed; 00303 } 00304 00305 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 00306 /// dominated by the specified block, and that are in the current loop) in 00307 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit 00308 /// uses before definitions, allowing us to sink a loop body in one pass without 00309 /// iteration. 00310 /// 00311 void LICM::SinkRegion(DomTreeNode *N) { 00312 assert(N != 0 && "Null dominator tree node?"); 00313 BasicBlock *BB = N->getBlock(); 00314 00315 // If this subregion is not in the top level loop at all, exit. 00316 if (!CurLoop->contains(BB)) return; 00317 00318 // We are processing blocks in reverse dfo, so process children first... 00319 const std::vector<DomTreeNode*> &Children = N->getChildren(); 00320 for (unsigned i = 0, e = Children.size(); i != e; ++i) 00321 SinkRegion(Children[i]); 00322 00323 // Only need to process the contents of this block if it is not part of a 00324 // subloop (which would already have been processed). 00325 if (inSubLoop(BB)) return; 00326 00327 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 00328 Instruction &I = *--II; 00329 00330 // Check to see if we can sink this instruction to the exit blocks 00331 // of the loop. We can do this if the all users of the instruction are 00332 // outside of the loop. In this case, it doesn't even matter if the 00333 // operands of the instruction are loop invariant. 00334 // 00335 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 00336 ++II; 00337 sink(I); 00338 } 00339 } 00340 } 00341 00342 00343 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks 00344 /// dominated by the specified block, and that are in the current loop) in depth 00345 /// first order w.r.t the DominatorTree. This allows us to visit definitions 00346 /// before uses, allowing us to hoist a loop body in one pass without iteration. 00347 /// 00348 void LICM::HoistRegion(DomTreeNode *N) { 00349 assert(N != 0 && "Null dominator tree node?"); 00350 BasicBlock *BB = N->getBlock(); 00351 00352 // If this subregion is not in the top level loop at all, exit. 00353 if (!CurLoop->contains(BB)) return; 00354 00355 // Only need to process the contents of this block if it is not part of a 00356 // subloop (which would already have been processed). 00357 if (!inSubLoop(BB)) 00358 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 00359 Instruction &I = *II++; 00360 00361 // Try hoisting the instruction out to the preheader. We can only do this 00362 // if all of the operands of the instruction are loop invariant and if it 00363 // is safe to hoist the instruction. 00364 // 00365 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) && 00366 isSafeToExecuteUnconditionally(I)) 00367 hoist(I); 00368 } 00369 00370 const std::vector<DomTreeNode*> &Children = N->getChildren(); 00371 for (unsigned i = 0, e = Children.size(); i != e; ++i) 00372 HoistRegion(Children[i]); 00373 } 00374 00375 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 00376 /// instruction. 00377 /// 00378 bool LICM::canSinkOrHoistInst(Instruction &I) { 00379 // Loads have extra constraints we have to verify before we can hoist them. 00380 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 00381 if (LI->isVolatile()) 00382 return false; // Don't hoist volatile loads! 00383 00384 // Loads from constant memory are always safe to move, even if they end up 00385 // in the same alias set as something that ends up being modified. 00386 if (EnableLICMConstantMotion && 00387 AA->pointsToConstantMemory(LI->getOperand(0))) 00388 return true; 00389 00390 // Don't hoist loads which have may-aliased stores in loop. 00391 unsigned Size = 0; 00392 if (LI->getType()->isSized()) 00393 Size = AA->getTypeStoreSize(LI->getType()); 00394 return !pointerInvalidatedByLoop(LI->getOperand(0), Size); 00395 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 00396 // Handle obvious cases efficiently. 00397 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 00398 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 00399 return true; 00400 else if (Behavior == AliasAnalysis::OnlyReadsMemory) { 00401 // If this call only reads from memory and there are no writes to memory 00402 // in the loop, we can hoist or sink the call as appropriate. 00403 bool FoundMod = false; 00404 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 00405 I != E; ++I) { 00406 AliasSet &AS = *I; 00407 if (!AS.isForwardingAliasSet() && AS.isMod()) { 00408 FoundMod = true; 00409 break; 00410 } 00411 } 00412 if (!FoundMod) return true; 00413 } 00414 00415 // FIXME: This should use mod/ref information to see if we can hoist or sink 00416 // the call. 00417 00418 return false; 00419 } 00420 00421 // Otherwise these instructions are hoistable/sinkable 00422 return isa<BinaryOperator>(I) || isa<CastInst>(I) || 00423 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || 00424 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 00425 isa<ShuffleVectorInst>(I); 00426 } 00427 00428 /// isNotUsedInLoop - Return true if the only users of this instruction are 00429 /// outside of the loop. If this is true, we can sink the instruction to the 00430 /// exit blocks of the loop. 00431 /// 00432 bool LICM::isNotUsedInLoop(Instruction &I) { 00433 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) { 00434 Instruction *User = cast<Instruction>(*UI); 00435 if (PHINode *PN = dyn_cast<PHINode>(User)) { 00436 // PHI node uses occur in predecessor blocks! 00437 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 00438 if (PN->getIncomingValue(i) == &I) 00439 if (CurLoop->contains(PN->getIncomingBlock(i))) 00440 return false; 00441 } else if (CurLoop->contains(User->getParent())) { 00442 return false; 00443 } 00444 } 00445 return true; 00446 } 00447 00448 00449 /// isLoopInvariantInst - Return true if all operands of this instruction are 00450 /// loop invariant. We also filter out non-hoistable instructions here just for 00451 /// efficiency. 00452 /// 00453 bool LICM::isLoopInvariantInst(Instruction &I) { 00454 // The instruction is loop invariant if all of its operands are loop-invariant 00455 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) 00456 if (!CurLoop->isLoopInvariant(I.getOperand(i))) 00457 return false; 00458 00459 // If we got this far, the instruction is loop invariant! 00460 return true; 00461 } 00462 00463 /// sink - When an instruction is found to only be used outside of the loop, 00464 /// this function moves it to the exit blocks and patches up SSA form as needed. 00465 /// This method is guaranteed to remove the original instruction from its 00466 /// position, and may either delete it or move it to outside of the loop. 00467 /// 00468 void LICM::sink(Instruction &I) { 00469 DOUT << "LICM sinking instruction: " << I; 00470 00471 SmallVector<BasicBlock*, 8> ExitBlocks; 00472 CurLoop->getExitBlocks(ExitBlocks); 00473 00474 if (isa<LoadInst>(I)) ++NumMovedLoads; 00475 else if (isa<CallInst>(I)) ++NumMovedCalls; 00476 ++NumSunk; 00477 Changed = true; 00478 00479 LLVMContext &Context = I.getContext(); 00480 00481 // The case where there is only a single exit node of this loop is common 00482 // enough that we handle it as a special (more efficient) case. It is more 00483 // efficient to handle because there are no PHI nodes that need to be placed. 00484 if (ExitBlocks.size() == 1) { 00485 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) { 00486 // Instruction is not used, just delete it. 00487 CurAST->deleteValue(&I); 00488 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 00489 I.replaceAllUsesWith(UndefValue::get(I.getType())); 00490 I.eraseFromParent(); 00491 } else { 00492 // Move the instruction to the start of the exit block, after any PHI 00493 // nodes in it. 00494 I.removeFromParent(); 00495 00496 BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI(); 00497 ExitBlocks[0]->getInstList().insert(InsertPt, &I); 00498 } 00499 } else if (ExitBlocks.empty()) { 00500 // The instruction is actually dead if there ARE NO exit blocks. 00501 CurAST->deleteValue(&I); 00502 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate. 00503 I.replaceAllUsesWith(UndefValue::get(I.getType())); 00504 I.eraseFromParent(); 00505 } else { 00506 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to 00507 // do all of the hard work of inserting PHI nodes as necessary. We convert 00508 // the value into a stack object to get it to do this. 00509 00510 // Firstly, we create a stack object to hold the value... 00511 AllocaInst *AI = 0; 00512 00513 if (I.getType() != Type::getVoidTy(I.getContext())) { 00514 AI = new AllocaInst(I.getType(), 0, I.getName(), 00515 I.getParent()->getParent()->getEntryBlock().begin()); 00516 CurAST->add(AI); 00517 } 00518 00519 // Secondly, insert load instructions for each use of the instruction 00520 // outside of the loop. 00521 while (!I.use_empty()) { 00522 Instruction *U = cast<Instruction>(I.use_back()); 00523 00524 // If the user is a PHI Node, we actually have to insert load instructions 00525 // in all predecessor blocks, not in the PHI block itself! 00526 if (PHINode *UPN = dyn_cast<PHINode>(U)) { 00527 // Only insert into each predecessor once, so that we don't have 00528 // different incoming values from the same block! 00529 std::map<BasicBlock*, Value*> InsertedBlocks; 00530 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i) 00531 if (UPN->getIncomingValue(i) == &I) { 00532 BasicBlock *Pred = UPN->getIncomingBlock(i); 00533 Value *&PredVal = InsertedBlocks[Pred]; 00534 if (!PredVal) { 00535 // Insert a new load instruction right before the terminator in 00536 // the predecessor block. 00537 PredVal = new LoadInst(AI, "", Pred->getTerminator()); 00538 CurAST->add(cast<LoadInst>(PredVal)); 00539 } 00540 00541 UPN->setIncomingValue(i, PredVal); 00542 } 00543 00544 } else { 00545 LoadInst *L = new LoadInst(AI, "", U); 00546 U->replaceUsesOfWith(&I, L); 00547 CurAST->add(L); 00548 } 00549 } 00550 00551 // Thirdly, insert a copy of the instruction in each exit block of the loop 00552 // that is dominated by the instruction, storing the result into the memory 00553 // location. Be careful not to insert the instruction into any particular 00554 // basic block more than once. 00555 std::set<BasicBlock*> InsertedBlocks; 00556 BasicBlock *InstOrigBB = I.getParent(); 00557 00558 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 00559 BasicBlock *ExitBlock = ExitBlocks[i]; 00560 00561 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) { 00562 // If we haven't already processed this exit block, do so now. 00563 if (InsertedBlocks.insert(ExitBlock).second) { 00564 // Insert the code after the last PHI node... 00565 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI(); 00566 00567 // If this is the first exit block processed, just move the original 00568 // instruction, otherwise clone the original instruction and insert 00569 // the copy. 00570 Instruction *New; 00571 if (InsertedBlocks.size() == 1) { 00572 I.removeFromParent(); 00573 ExitBlock->getInstList().insert(InsertPt, &I); 00574 New = &I; 00575 } else { 00576 New = I.clone(Context); 00577 CurAST->copyValue(&I, New); 00578 if (!I.getName().empty()) 00579 New->setName(I.getName()+".le"); 00580 ExitBlock->getInstList().insert(InsertPt, New); 00581 } 00582 00583 // Now that we have inserted the instruction, store it into the alloca 00584 if (AI) new StoreInst(New, AI, InsertPt); 00585 } 00586 } 00587 } 00588 00589 // If the instruction doesn't dominate any exit blocks, it must be dead. 00590 if (InsertedBlocks.empty()) { 00591 CurAST->deleteValue(&I); 00592 I.eraseFromParent(); 00593 } 00594 00595 // Finally, promote the fine value to SSA form. 00596 if (AI) { 00597 std::vector<AllocaInst*> Allocas; 00598 Allocas.push_back(AI); 00599 PromoteMemToReg(Allocas, *DT, *DF, Context, CurAST); 00600 } 00601 } 00602 } 00603 00604 /// hoist - When an instruction is found to only use loop invariant operands 00605 /// that is safe to hoist, this instruction is called to do the dirty work. 00606 /// 00607 void LICM::hoist(Instruction &I) { 00608 DEBUG(errs() << "LICM hoisting to " << Preheader->getName() << ": " << I); 00609 00610 // Remove the instruction from its current basic block... but don't delete the 00611 // instruction. 00612 I.removeFromParent(); 00613 00614 // Insert the new node in Preheader, before the terminator. 00615 Preheader->getInstList().insert(Preheader->getTerminator(), &I); 00616 00617 if (isa<LoadInst>(I)) ++NumMovedLoads; 00618 else if (isa<CallInst>(I)) ++NumMovedCalls; 00619 ++NumHoisted; 00620 Changed = true; 00621 } 00622 00623 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 00624 /// not a trapping instruction or if it is a trapping instruction and is 00625 /// guaranteed to execute. 00626 /// 00627 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 00628 // If it is not a trapping instruction, it is always safe to hoist. 00629 if (Inst.isSafeToSpeculativelyExecute()) 00630 return true; 00631 00632 // Otherwise we have to check to make sure that the instruction dominates all 00633 // of the exit blocks. If it doesn't, then there is a path out of the loop 00634 // which does not execute this instruction, so we can't hoist it. 00635 00636 // If the instruction is in the header block for the loop (which is very 00637 // common), it is always guaranteed to dominate the exit blocks. Since this 00638 // is a common case, and can save some work, check it now. 00639 if (Inst.getParent() == CurLoop->getHeader()) 00640 return true; 00641 00642 // Get the exit blocks for the current loop. 00643 SmallVector<BasicBlock*, 8> ExitBlocks; 00644 CurLoop->getExitBlocks(ExitBlocks); 00645 00646 // For each exit block, get the DT node and walk up the DT until the 00647 // instruction's basic block is found or we exit the loop. 00648 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 00649 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent())) 00650 return false; 00651 00652 return true; 00653 } 00654 00655 00656 /// PromoteValuesInLoop - Try to promote memory values to scalars by sinking 00657 /// stores out of the loop and moving loads to before the loop. We do this by 00658 /// looping over the stores in the loop, looking for stores to Must pointers 00659 /// which are loop invariant. We promote these memory locations to use allocas 00660 /// instead. These allocas can easily be raised to register values by the 00661 /// PromoteMem2Reg functionality. 00662 /// 00663 void LICM::PromoteValuesInLoop() { 00664 // PromotedValues - List of values that are promoted out of the loop. Each 00665 // value has an alloca instruction for it, and a canonical version of the 00666 // pointer. 00667 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues; 00668 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca 00669 00670 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap); 00671 if (ValueToAllocaMap.empty()) return; // If there are values to promote. 00672 00673 Changed = true; 00674 NumPromoted += PromotedValues.size(); 00675 00676 std::vector<Value*> PointerValueNumbers; 00677 00678 // Emit a copy from the value into the alloca'd value in the loop preheader 00679 TerminatorInst *LoopPredInst = Preheader->getTerminator(); 00680 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 00681 Value *Ptr = PromotedValues[i].second; 00682 00683 // If we are promoting a pointer value, update alias information for the 00684 // inserted load. 00685 Value *LoadValue = 0; 00686 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) { 00687 // Locate a load or store through the pointer, and assign the same value 00688 // to LI as we are loading or storing. Since we know that the value is 00689 // stored in this loop, this will always succeed. 00690 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end(); 00691 UI != E; ++UI) 00692 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 00693 LoadValue = LI; 00694 break; 00695 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 00696 if (SI->getOperand(1) == Ptr) { 00697 LoadValue = SI->getOperand(0); 00698 break; 00699 } 00700 } 00701 assert(LoadValue && "No store through the pointer found!"); 00702 PointerValueNumbers.push_back(LoadValue); // Remember this for later. 00703 } 00704 00705 // Load from the memory we are promoting. 00706 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst); 00707 00708 if (LoadValue) CurAST->copyValue(LoadValue, LI); 00709 00710 // Store into the temporary alloca. 00711 new StoreInst(LI, PromotedValues[i].first, LoopPredInst); 00712 } 00713 00714 // Scan the basic blocks in the loop, replacing uses of our pointers with 00715 // uses of the allocas in question. 00716 // 00717 for (Loop::block_iterator I = CurLoop->block_begin(), 00718 E = CurLoop->block_end(); I != E; ++I) { 00719 BasicBlock *BB = *I; 00720 // Rewrite all loads and stores in the block of the pointer... 00721 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 00722 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 00723 std::map<Value*, AllocaInst*>::iterator 00724 I = ValueToAllocaMap.find(L->getOperand(0)); 00725 if (I != ValueToAllocaMap.end()) 00726 L->setOperand(0, I->second); // Rewrite load instruction... 00727 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) { 00728 std::map<Value*, AllocaInst*>::iterator 00729 I = ValueToAllocaMap.find(S->getOperand(1)); 00730 if (I != ValueToAllocaMap.end()) 00731 S->setOperand(1, I->second); // Rewrite store instruction... 00732 } 00733 } 00734 } 00735 00736 // Now that the body of the loop uses the allocas instead of the original 00737 // memory locations, insert code to copy the alloca value back into the 00738 // original memory location on all exits from the loop. Note that we only 00739 // want to insert one copy of the code in each exit block, though the loop may 00740 // exit to the same block more than once. 00741 // 00742 SmallPtrSet<BasicBlock*, 16> ProcessedBlocks; 00743 00744 SmallVector<BasicBlock*, 8> ExitBlocks; 00745 CurLoop->getExitBlocks(ExitBlocks); 00746 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 00747 if (!ProcessedBlocks.insert(ExitBlocks[i])) 00748 continue; 00749 00750 // Copy all of the allocas into their memory locations. 00751 BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI(); 00752 Instruction *InsertPos = BI; 00753 unsigned PVN = 0; 00754 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) { 00755 // Load from the alloca. 00756 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos); 00757 00758 // If this is a pointer type, update alias info appropriately. 00759 if (isa<PointerType>(LI->getType())) 00760 CurAST->copyValue(PointerValueNumbers[PVN++], LI); 00761 00762 // Store into the memory we promoted. 00763 new StoreInst(LI, PromotedValues[i].second, InsertPos); 00764 } 00765 } 00766 00767 // Now that we have done the deed, use the mem2reg functionality to promote 00768 // all of the new allocas we just created into real SSA registers. 00769 // 00770 std::vector<AllocaInst*> PromotedAllocas; 00771 PromotedAllocas.reserve(PromotedValues.size()); 00772 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) 00773 PromotedAllocas.push_back(PromotedValues[i].first); 00774 PromoteMemToReg(PromotedAllocas, *DT, *DF, Preheader->getContext(), CurAST); 00775 } 00776 00777 /// FindPromotableValuesInLoop - Check the current loop for stores to definite 00778 /// pointers, which are not loaded and stored through may aliases and are safe 00779 /// for promotion. If these are found, create an alloca for the value, add it 00780 /// to the PromotedValues list, and keep track of the mapping from value to 00781 /// alloca. 00782 void LICM::FindPromotableValuesInLoop( 00783 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues, 00784 std::map<Value*, AllocaInst*> &ValueToAllocaMap) { 00785 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin(); 00786 00787 // Loop over all of the alias sets in the tracker object. 00788 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 00789 I != E; ++I) { 00790 AliasSet &AS = *I; 00791 // We can promote this alias set if it has a store, if it is a "Must" alias 00792 // set, if the pointer is loop invariant, and if we are not eliminating any 00793 // volatile loads or stores. 00794 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 00795 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 00796 continue; 00797 00798 assert(!AS.empty() && 00799 "Must alias set should have at least one pointer element in it!"); 00800 Value *V = AS.begin()->getValue(); 00801 00802 // Check that all of the pointers in the alias set have the same type. We 00803 // cannot (yet) promote a memory location that is loaded and stored in 00804 // different sizes. 00805 { 00806 bool PointerOk = true; 00807 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 00808 if (V->getType() != I->getValue()->getType()) { 00809 PointerOk = false; 00810 break; 00811 } 00812 if (!PointerOk) 00813 continue; 00814 } 00815 00816 // It isn't safe to promote a load/store from the loop if the load/store is 00817 // conditional. For example, turning: 00818 // 00819 // for () { if (c) *P += 1; } 00820 // 00821 // into: 00822 // 00823 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 00824 // 00825 // is not safe, because *P may only be valid to access if 'c' is true. 00826 // 00827 // It is safe to promote P if all uses are direct load/stores and if at 00828 // least one is guaranteed to be executed. 00829 bool GuaranteedToExecute = false; 00830 bool InvalidInst = false; 00831 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); 00832 UI != UE; ++UI) { 00833 // Ignore instructions not in this loop. 00834 Instruction *Use = dyn_cast<Instruction>(*UI); 00835 if (!Use || !CurLoop->contains(Use->getParent())) 00836 continue; 00837 00838 if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) { 00839 InvalidInst = true; 00840 break; 00841 } 00842 00843 if (!GuaranteedToExecute) 00844 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use); 00845 } 00846 00847 // If there is an non-load/store instruction in the loop, we can't promote 00848 // it. If there isn't a guaranteed-to-execute instruction, we can't 00849 // promote. 00850 if (InvalidInst || !GuaranteedToExecute) 00851 continue; 00852 00853 const Type *Ty = cast<PointerType>(V->getType())->getElementType(); 00854 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart); 00855 PromotedValues.push_back(std::make_pair(AI, V)); 00856 00857 // Update the AST and alias analysis. 00858 CurAST->copyValue(V, AI); 00859 00860 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I) 00861 ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI)); 00862 00863 DOUT << "LICM: Promoting value: " << *V << "\n"; 00864 } 00865 } 00866 00867 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 00868 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 00869 AliasSetTracker *AST = LoopToAliasMap[L]; 00870 if (!AST) 00871 return; 00872 00873 AST->copyValue(From, To); 00874 } 00875 00876 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 00877 /// set. 00878 void LICM::deleteAnalysisValue(Value *V, Loop *L) { 00879 AliasSetTracker *AST = LoopToAliasMap[L]; 00880 if (!AST) 00881 return; 00882 00883 AST->deleteValue(V); 00884 }