1 //===- MergedLoadStoreMotion.cpp - merge and hoist/sink load/stores -------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 //! \file 11 //! \brief This pass performs merges of loads and stores on both sides of a 12 // diamond (hammock). It hoists the loads and sinks the stores. 13 // 14 // The algorithm iteratively hoists two loads to the same address out of a 15 // diamond (hammock) and merges them into a single load in the header. Similar 16 // it sinks and merges two stores to the tail block (footer). The algorithm 17 // iterates over the instructions of one side of the diamond and attempts to 18 // find a matching load/store on the other side. It hoists / sinks when it 19 // thinks it safe to do so. This optimization helps with eg. hiding load 20 // latencies, triggering if-conversion, and reducing static code size. 21 // 22 // NOTE: This code no longer performs load hoisting, it is subsumed by GVNHoist. 23 // 24 //===----------------------------------------------------------------------===// 25 // 26 // 27 // Example: 28 // Diamond shaped code before merge: 29 // 30 // header: 31 // br %cond, label %if.then, label %if.else 32 // + + 33 // + + 34 // + + 35 // if.then: if.else: 36 // %lt = load %addr_l %le = load %addr_l 37 // <use %lt> <use %le> 38 // <...> <...> 39 // store %st, %addr_s store %se, %addr_s 40 // br label %if.end br label %if.end 41 // + + 42 // + + 43 // + + 44 // if.end ("footer"): 45 // <...> 46 // 47 // Diamond shaped code after merge: 48 // 49 // header: 50 // %l = load %addr_l 51 // br %cond, label %if.then, label %if.else 52 // + + 53 // + + 54 // + + 55 // if.then: if.else: 56 // <use %l> <use %l> 57 // <...> <...> 58 // br label %if.end br label %if.end 59 // + + 60 // + + 61 // + + 62 // if.end ("footer"): 63 // %s.sink = phi [%st, if.then], [%se, if.else] 64 // <...> 65 // store %s.sink, %addr_s 66 // <...> 67 // 68 // 69 //===----------------------- TODO -----------------------------------------===// 70 // 71 // 1) Generalize to regions other than diamonds 72 // 2) Be more aggressive merging memory operations 73 // Note that both changes require register pressure control 74 // 75 //===----------------------------------------------------------------------===// 76 77 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h" 78 #include "llvm/ADT/Statistic.h" 79 #include "llvm/Analysis/AliasAnalysis.h" 80 #include "llvm/Analysis/CFG.h" 81 #include "llvm/Analysis/GlobalsModRef.h" 82 #include "llvm/Analysis/Loads.h" 83 #include "llvm/Analysis/MemoryBuiltins.h" 84 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 85 #include "llvm/Analysis/ValueTracking.h" 86 #include "llvm/IR/Metadata.h" 87 #include "llvm/IR/PatternMatch.h" 88 #include "llvm/Support/Debug.h" 89 #include "llvm/Support/raw_ostream.h" 90 #include "llvm/Transforms/Scalar.h" 91 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 92 93 using namespace llvm; 94 95 #define DEBUG_TYPE "mldst-motion" 96 97 namespace { 98 //===----------------------------------------------------------------------===// 99 // MergedLoadStoreMotion Pass 100 //===----------------------------------------------------------------------===// 101 class MergedLoadStoreMotion { 102 MemoryDependenceResults *MD = nullptr; 103 AliasAnalysis *AA = nullptr; 104 105 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity, 106 // where Size0 and Size1 are the #instructions on the two sides of 107 // the diamond. The constant chosen here is arbitrary. Compiler Time 108 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl. 109 const int MagicCompileTimeControl = 250; 110 111 public: 112 bool run(Function &F, MemoryDependenceResults *MD, AliasAnalysis &AA); 113 114 private: 115 /// 116 /// \brief Remove instruction from parent and update memory dependence 117 /// analysis. 118 /// 119 void removeInstruction(Instruction *Inst); 120 BasicBlock *getDiamondTail(BasicBlock *BB); 121 bool isDiamondHead(BasicBlock *BB); 122 // Routines for sinking stores 123 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 124 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 125 bool isStoreSinkBarrierInRange(const Instruction &Start, 126 const Instruction &End, MemoryLocation Loc); 127 bool sinkStore(BasicBlock *BB, StoreInst *SinkCand, StoreInst *ElseInst); 128 bool mergeStores(BasicBlock *BB); 129 }; 130 } // end anonymous namespace 131 132 /// 133 /// \brief Remove instruction from parent and update memory dependence analysis. 134 /// 135 void MergedLoadStoreMotion::removeInstruction(Instruction *Inst) { 136 // Notify the memory dependence analysis. 137 if (MD) { 138 MD->removeInstruction(Inst); 139 if (auto *LI = dyn_cast<LoadInst>(Inst)) 140 MD->invalidateCachedPointerInfo(LI->getPointerOperand()); 141 if (Inst->getType()->isPtrOrPtrVectorTy()) { 142 MD->invalidateCachedPointerInfo(Inst); 143 } 144 } 145 Inst->eraseFromParent(); 146 } 147 148 /// 149 /// \brief Return tail block of a diamond. 150 /// 151 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 152 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 153 return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor(); 154 } 155 156 /// 157 /// \brief True when BB is the head of a diamond (hammock) 158 /// 159 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 160 if (!BB) 161 return false; 162 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 163 if (!BI || !BI->isConditional()) 164 return false; 165 166 BasicBlock *Succ0 = BI->getSuccessor(0); 167 BasicBlock *Succ1 = BI->getSuccessor(1); 168 169 if (!Succ0->getSinglePredecessor()) 170 return false; 171 if (!Succ1->getSinglePredecessor()) 172 return false; 173 174 BasicBlock *Succ0Succ = Succ0->getSingleSuccessor(); 175 BasicBlock *Succ1Succ = Succ1->getSingleSuccessor(); 176 // Ignore triangles. 177 if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ) 178 return false; 179 return true; 180 } 181 182 183 /// 184 /// \brief True when instruction is a sink barrier for a store 185 /// located in Loc 186 /// 187 /// Whenever an instruction could possibly read or modify the 188 /// value being stored or protect against the store from 189 /// happening it is considered a sink barrier. 190 /// 191 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 192 const Instruction &End, 193 MemoryLocation Loc) { 194 for (const Instruction &Inst : 195 make_range(Start.getIterator(), End.getIterator())) 196 if (Inst.mayThrow()) 197 return true; 198 return AA->canInstructionRangeModRef(Start, End, Loc, MRI_ModRef); 199 } 200 201 /// 202 /// \brief Check if \p BB contains a store to the same address as \p SI 203 /// 204 /// \return The store in \p when it is safe to sink. Otherwise return Null. 205 /// 206 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 207 StoreInst *Store0) { 208 DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 209 BasicBlock *BB0 = Store0->getParent(); 210 for (Instruction &Inst : reverse(*BB1)) { 211 auto *Store1 = dyn_cast<StoreInst>(&Inst); 212 if (!Store1) 213 continue; 214 215 MemoryLocation Loc0 = MemoryLocation::get(Store0); 216 MemoryLocation Loc1 = MemoryLocation::get(Store1); 217 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) && 218 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) && 219 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0)) { 220 return Store1; 221 } 222 } 223 return nullptr; 224 } 225 226 /// 227 /// \brief Create a PHI node in BB for the operands of S0 and S1 228 /// 229 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 230 StoreInst *S1) { 231 // Create a phi if the values mismatch. 232 Value *Opd1 = S0->getValueOperand(); 233 Value *Opd2 = S1->getValueOperand(); 234 if (Opd1 == Opd2) 235 return nullptr; 236 237 auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink", 238 &BB->front()); 239 NewPN->addIncoming(Opd1, S0->getParent()); 240 NewPN->addIncoming(Opd2, S1->getParent()); 241 if (MD && NewPN->getType()->isPtrOrPtrVectorTy()) 242 MD->invalidateCachedPointerInfo(NewPN); 243 return NewPN; 244 } 245 246 /// 247 /// \brief Merge two stores to same address and sink into \p BB 248 /// 249 /// Also sinks GEP instruction computing the store address 250 /// 251 bool MergedLoadStoreMotion::sinkStore(BasicBlock *BB, StoreInst *S0, 252 StoreInst *S1) { 253 // Only one definition? 254 auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 255 auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 256 if (A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() && 257 (A0->getParent() == S0->getParent()) && A1->hasOneUse() && 258 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0)) { 259 DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 260 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 261 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 262 // Hoist the instruction. 263 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 264 // Intersect optional metadata. 265 S0->andIRFlags(S1); 266 S0->dropUnknownNonDebugMetadata(); 267 268 // Create the new store to be inserted at the join point. 269 StoreInst *SNew = cast<StoreInst>(S0->clone()); 270 Instruction *ANew = A0->clone(); 271 SNew->insertBefore(&*InsertPt); 272 ANew->insertBefore(SNew); 273 274 assert(S0->getParent() == A0->getParent()); 275 assert(S1->getParent() == A1->getParent()); 276 277 // New PHI operand? Use it. 278 if (PHINode *NewPN = getPHIOperand(BB, S0, S1)) 279 SNew->setOperand(0, NewPN); 280 removeInstruction(S0); 281 removeInstruction(S1); 282 A0->replaceAllUsesWith(ANew); 283 removeInstruction(A0); 284 A1->replaceAllUsesWith(ANew); 285 removeInstruction(A1); 286 return true; 287 } 288 return false; 289 } 290 291 /// 292 /// \brief True when two stores are equivalent and can sink into the footer 293 /// 294 /// Starting from a diamond tail block, iterate over the instructions in one 295 /// predecessor block and try to match a store in the second predecessor. 296 /// 297 bool MergedLoadStoreMotion::mergeStores(BasicBlock *T) { 298 299 bool MergedStores = false; 300 assert(T && "Footer of a diamond cannot be empty"); 301 302 pred_iterator PI = pred_begin(T), E = pred_end(T); 303 assert(PI != E); 304 BasicBlock *Pred0 = *PI; 305 ++PI; 306 BasicBlock *Pred1 = *PI; 307 ++PI; 308 // tail block of a diamond/hammock? 309 if (Pred0 == Pred1) 310 return false; // No. 311 if (PI != E) 312 return false; // No. More than 2 predecessors. 313 314 // #Instructions in Succ1 for Compile Time Control 315 int Size1 = Pred1->size(); 316 int NStores = 0; 317 318 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 319 RBI != RBE;) { 320 321 Instruction *I = &*RBI; 322 ++RBI; 323 324 // Don't sink non-simple (atomic, volatile) stores. 325 auto *S0 = dyn_cast<StoreInst>(I); 326 if (!S0 || !S0->isSimple()) 327 continue; 328 329 ++NStores; 330 if (NStores * Size1 >= MagicCompileTimeControl) 331 break; 332 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 333 bool Res = sinkStore(T, S0, S1); 334 MergedStores |= Res; 335 // Don't attempt to sink below stores that had to stick around 336 // But after removal of a store and some of its feeding 337 // instruction search again from the beginning since the iterator 338 // is likely stale at this point. 339 if (!Res) 340 break; 341 RBI = Pred0->rbegin(); 342 RBE = Pred0->rend(); 343 DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 344 } 345 } 346 return MergedStores; 347 } 348 349 bool MergedLoadStoreMotion::run(Function &F, MemoryDependenceResults *MD, 350 AliasAnalysis &AA) { 351 this->MD = MD; 352 this->AA = &AA; 353 354 bool Changed = false; 355 DEBUG(dbgs() << "Instruction Merger\n"); 356 357 // Merge unconditional branches, allowing PRE to catch more 358 // optimization opportunities. 359 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) { 360 BasicBlock *BB = &*FI++; 361 362 // Hoist equivalent loads and sink stores 363 // outside diamonds when possible 364 if (isDiamondHead(BB)) { 365 Changed |= mergeStores(getDiamondTail(BB)); 366 } 367 } 368 return Changed; 369 } 370 371 namespace { 372 class MergedLoadStoreMotionLegacyPass : public FunctionPass { 373 public: 374 static char ID; // Pass identification, replacement for typeid 375 MergedLoadStoreMotionLegacyPass() : FunctionPass(ID) { 376 initializeMergedLoadStoreMotionLegacyPassPass( 377 *PassRegistry::getPassRegistry()); 378 } 379 380 /// 381 /// \brief Run the transformation for each function 382 /// 383 bool runOnFunction(Function &F) override { 384 if (skipFunction(F)) 385 return false; 386 MergedLoadStoreMotion Impl; 387 auto *MDWP = getAnalysisIfAvailable<MemoryDependenceWrapperPass>(); 388 return Impl.run(F, MDWP ? &MDWP->getMemDep() : nullptr, 389 getAnalysis<AAResultsWrapperPass>().getAAResults()); 390 } 391 392 private: 393 void getAnalysisUsage(AnalysisUsage &AU) const override { 394 AU.setPreservesCFG(); 395 AU.addRequired<AAResultsWrapperPass>(); 396 AU.addPreserved<GlobalsAAWrapperPass>(); 397 AU.addPreserved<MemoryDependenceWrapperPass>(); 398 } 399 }; 400 401 char MergedLoadStoreMotionLegacyPass::ID = 0; 402 } // anonymous namespace 403 404 /// 405 /// \brief createMergedLoadStoreMotionPass - The public interface to this file. 406 /// 407 FunctionPass *llvm::createMergedLoadStoreMotionPass() { 408 return new MergedLoadStoreMotionLegacyPass(); 409 } 410 411 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion", 412 "MergedLoadStoreMotion", false, false) 413 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass) 414 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 415 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion", 416 "MergedLoadStoreMotion", false, false) 417 418 PreservedAnalyses 419 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) { 420 MergedLoadStoreMotion Impl; 421 auto *MD = AM.getCachedResult<MemoryDependenceAnalysis>(F); 422 auto &AA = AM.getResult<AAManager>(F); 423 if (!Impl.run(F, MD, AA)) 424 return PreservedAnalyses::all(); 425 426 PreservedAnalyses PA; 427 PA.preserveSet<CFGAnalyses>(); 428 PA.preserve<GlobalsAA>(); 429 PA.preserve<MemoryDependenceAnalysis>(); 430 return PA; 431 } 432