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