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