1 //===- MergedLoadStoreMotion.cpp - merge and hoist/sink load/stores -------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 //! \file 10 //! This pass performs merges of loads and stores on both sides of a 11 // diamond (hammock). It hoists the loads and sinks the stores. 12 // 13 // The algorithm iteratively hoists two loads to the same address out of a 14 // diamond (hammock) and merges them into a single load in the header. Similar 15 // it sinks and merges two stores to the tail block (footer). The algorithm 16 // iterates over the instructions of one side of the diamond and attempts to 17 // find a matching load/store on the other side. It hoists / sinks when it 18 // thinks it safe to do so. This optimization helps with eg. hiding load 19 // latencies, triggering if-conversion, and reducing static code size. 20 // 21 // NOTE: This code no longer performs load hoisting, it is subsumed by GVNHoist. 22 // 23 //===----------------------------------------------------------------------===// 24 // 25 // 26 // Example: 27 // Diamond shaped code before merge: 28 // 29 // header: 30 // br %cond, label %if.then, label %if.else 31 // + + 32 // + + 33 // + + 34 // if.then: if.else: 35 // %lt = load %addr_l %le = load %addr_l 36 // <use %lt> <use %le> 37 // <...> <...> 38 // store %st, %addr_s store %se, %addr_s 39 // br label %if.end br label %if.end 40 // + + 41 // + + 42 // + + 43 // if.end ("footer"): 44 // <...> 45 // 46 // Diamond shaped code after merge: 47 // 48 // header: 49 // %l = load %addr_l 50 // br %cond, label %if.then, label %if.else 51 // + + 52 // + + 53 // + + 54 // if.then: if.else: 55 // <use %l> <use %l> 56 // <...> <...> 57 // br label %if.end br label %if.end 58 // + + 59 // + + 60 // + + 61 // if.end ("footer"): 62 // %s.sink = phi [%st, if.then], [%se, if.else] 63 // <...> 64 // store %s.sink, %addr_s 65 // <...> 66 // 67 // 68 //===----------------------- TODO -----------------------------------------===// 69 // 70 // 1) Generalize to regions other than diamonds 71 // 2) Be more aggressive merging memory operations 72 // Note that both changes require register pressure control 73 // 74 //===----------------------------------------------------------------------===// 75 76 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h" 77 #include "llvm/ADT/Statistic.h" 78 #include "llvm/Analysis/AliasAnalysis.h" 79 #include "llvm/Analysis/CFG.h" 80 #include "llvm/Analysis/GlobalsModRef.h" 81 #include "llvm/Analysis/Loads.h" 82 #include "llvm/Analysis/ValueTracking.h" 83 #include "llvm/IR/Metadata.h" 84 #include "llvm/Support/Debug.h" 85 #include "llvm/Support/raw_ostream.h" 86 #include "llvm/Transforms/Scalar.h" 87 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 88 89 using namespace llvm; 90 91 #define DEBUG_TYPE "mldst-motion" 92 93 namespace { 94 //===----------------------------------------------------------------------===// 95 // MergedLoadStoreMotion Pass 96 //===----------------------------------------------------------------------===// 97 class MergedLoadStoreMotion { 98 AliasAnalysis *AA = nullptr; 99 100 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity, 101 // where Size0 and Size1 are the #instructions on the two sides of 102 // the diamond. The constant chosen here is arbitrary. Compiler Time 103 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl. 104 const int MagicCompileTimeControl = 250; 105 106 public: 107 bool run(Function &F, AliasAnalysis &AA); 108 109 private: 110 BasicBlock *getDiamondTail(BasicBlock *BB); 111 bool isDiamondHead(BasicBlock *BB); 112 // Routines for sinking stores 113 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 114 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 115 bool isStoreSinkBarrierInRange(const Instruction &Start, 116 const Instruction &End, MemoryLocation Loc); 117 bool sinkStore(BasicBlock *BB, StoreInst *SinkCand, StoreInst *ElseInst); 118 bool mergeStores(BasicBlock *BB); 119 }; 120 } // end anonymous namespace 121 122 /// 123 /// Return tail block of a diamond. 124 /// 125 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 126 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 127 return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor(); 128 } 129 130 /// 131 /// True when BB is the head of a diamond (hammock) 132 /// 133 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 134 if (!BB) 135 return false; 136 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 137 if (!BI || !BI->isConditional()) 138 return false; 139 140 BasicBlock *Succ0 = BI->getSuccessor(0); 141 BasicBlock *Succ1 = BI->getSuccessor(1); 142 143 if (!Succ0->getSinglePredecessor()) 144 return false; 145 if (!Succ1->getSinglePredecessor()) 146 return false; 147 148 BasicBlock *Succ0Succ = Succ0->getSingleSuccessor(); 149 BasicBlock *Succ1Succ = Succ1->getSingleSuccessor(); 150 // Ignore triangles. 151 if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ) 152 return false; 153 return true; 154 } 155 156 157 /// 158 /// True when instruction is a sink barrier for a store 159 /// located in Loc 160 /// 161 /// Whenever an instruction could possibly read or modify the 162 /// value being stored or protect against the store from 163 /// happening it is considered a sink barrier. 164 /// 165 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 166 const Instruction &End, 167 MemoryLocation Loc) { 168 for (const Instruction &Inst : 169 make_range(Start.getIterator(), End.getIterator())) 170 if (Inst.mayThrow()) 171 return true; 172 return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef); 173 } 174 175 /// 176 /// Check if \p BB contains a store to the same address as \p SI 177 /// 178 /// \return The store in \p when it is safe to sink. Otherwise return Null. 179 /// 180 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 181 StoreInst *Store0) { 182 LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 183 BasicBlock *BB0 = Store0->getParent(); 184 for (Instruction &Inst : reverse(*BB1)) { 185 auto *Store1 = dyn_cast<StoreInst>(&Inst); 186 if (!Store1) 187 continue; 188 189 MemoryLocation Loc0 = MemoryLocation::get(Store0); 190 MemoryLocation Loc1 = MemoryLocation::get(Store1); 191 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) && 192 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) && 193 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0)) { 194 return Store1; 195 } 196 } 197 return nullptr; 198 } 199 200 /// 201 /// Create a PHI node in BB for the operands of S0 and S1 202 /// 203 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 204 StoreInst *S1) { 205 // Create a phi if the values mismatch. 206 Value *Opd1 = S0->getValueOperand(); 207 Value *Opd2 = S1->getValueOperand(); 208 if (Opd1 == Opd2) 209 return nullptr; 210 211 auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink", 212 &BB->front()); 213 NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 214 NewPN->addIncoming(Opd1, S0->getParent()); 215 NewPN->addIncoming(Opd2, S1->getParent()); 216 return NewPN; 217 } 218 219 /// 220 /// Merge two stores to same address and sink into \p BB 221 /// 222 /// Also sinks GEP instruction computing the store address 223 /// 224 bool MergedLoadStoreMotion::sinkStore(BasicBlock *BB, StoreInst *S0, 225 StoreInst *S1) { 226 // Only one definition? 227 auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 228 auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 229 if (A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() && 230 (A0->getParent() == S0->getParent()) && A1->hasOneUse() && 231 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0)) { 232 LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 233 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 234 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 235 // Hoist the instruction. 236 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 237 // Intersect optional metadata. 238 S0->andIRFlags(S1); 239 S0->dropUnknownNonDebugMetadata(); 240 241 // Create the new store to be inserted at the join point. 242 StoreInst *SNew = cast<StoreInst>(S0->clone()); 243 Instruction *ANew = A0->clone(); 244 SNew->insertBefore(&*InsertPt); 245 ANew->insertBefore(SNew); 246 247 assert(S0->getParent() == A0->getParent()); 248 assert(S1->getParent() == A1->getParent()); 249 250 // New PHI operand? Use it. 251 if (PHINode *NewPN = getPHIOperand(BB, S0, S1)) 252 SNew->setOperand(0, NewPN); 253 S0->eraseFromParent(); 254 S1->eraseFromParent(); 255 A0->replaceAllUsesWith(ANew); 256 A0->eraseFromParent(); 257 A1->replaceAllUsesWith(ANew); 258 A1->eraseFromParent(); 259 return true; 260 } 261 return false; 262 } 263 264 /// 265 /// True when two stores are equivalent and can sink into the footer 266 /// 267 /// Starting from a diamond tail block, iterate over the instructions in one 268 /// predecessor block and try to match a store in the second predecessor. 269 /// 270 bool MergedLoadStoreMotion::mergeStores(BasicBlock *T) { 271 272 bool MergedStores = false; 273 assert(T && "Footer of a diamond cannot be empty"); 274 275 pred_iterator PI = pred_begin(T), E = pred_end(T); 276 assert(PI != E); 277 BasicBlock *Pred0 = *PI; 278 ++PI; 279 BasicBlock *Pred1 = *PI; 280 ++PI; 281 // tail block of a diamond/hammock? 282 if (Pred0 == Pred1) 283 return false; // No. 284 if (PI != E) 285 return false; // No. More than 2 predecessors. 286 287 // #Instructions in Succ1 for Compile Time Control 288 auto InstsNoDbg = Pred1->instructionsWithoutDebug(); 289 int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end()); 290 int NStores = 0; 291 292 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 293 RBI != RBE;) { 294 295 Instruction *I = &*RBI; 296 ++RBI; 297 298 // Don't sink non-simple (atomic, volatile) stores. 299 auto *S0 = dyn_cast<StoreInst>(I); 300 if (!S0 || !S0->isSimple()) 301 continue; 302 303 ++NStores; 304 if (NStores * Size1 >= MagicCompileTimeControl) 305 break; 306 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 307 bool Res = sinkStore(T, S0, S1); 308 MergedStores |= Res; 309 // Don't attempt to sink below stores that had to stick around 310 // But after removal of a store and some of its feeding 311 // instruction search again from the beginning since the iterator 312 // is likely stale at this point. 313 if (!Res) 314 break; 315 RBI = Pred0->rbegin(); 316 RBE = Pred0->rend(); 317 LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 318 } 319 } 320 return MergedStores; 321 } 322 323 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) { 324 this->AA = &AA; 325 326 bool Changed = false; 327 LLVM_DEBUG(dbgs() << "Instruction Merger\n"); 328 329 // Merge unconditional branches, allowing PRE to catch more 330 // optimization opportunities. 331 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) { 332 BasicBlock *BB = &*FI++; 333 334 // Hoist equivalent loads and sink stores 335 // outside diamonds when possible 336 if (isDiamondHead(BB)) { 337 Changed |= mergeStores(getDiamondTail(BB)); 338 } 339 } 340 return Changed; 341 } 342 343 namespace { 344 class MergedLoadStoreMotionLegacyPass : public FunctionPass { 345 public: 346 static char ID; // Pass identification, replacement for typeid 347 MergedLoadStoreMotionLegacyPass() : FunctionPass(ID) { 348 initializeMergedLoadStoreMotionLegacyPassPass( 349 *PassRegistry::getPassRegistry()); 350 } 351 352 /// 353 /// Run the transformation for each function 354 /// 355 bool runOnFunction(Function &F) override { 356 if (skipFunction(F)) 357 return false; 358 MergedLoadStoreMotion Impl; 359 return Impl.run(F, getAnalysis<AAResultsWrapperPass>().getAAResults()); 360 } 361 362 private: 363 void getAnalysisUsage(AnalysisUsage &AU) const override { 364 AU.setPreservesCFG(); 365 AU.addRequired<AAResultsWrapperPass>(); 366 AU.addPreserved<GlobalsAAWrapperPass>(); 367 } 368 }; 369 370 char MergedLoadStoreMotionLegacyPass::ID = 0; 371 } // anonymous namespace 372 373 /// 374 /// createMergedLoadStoreMotionPass - The public interface to this file. 375 /// 376 FunctionPass *llvm::createMergedLoadStoreMotionPass() { 377 return new MergedLoadStoreMotionLegacyPass(); 378 } 379 380 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion", 381 "MergedLoadStoreMotion", false, false) 382 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 383 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion", 384 "MergedLoadStoreMotion", false, false) 385 386 PreservedAnalyses 387 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) { 388 MergedLoadStoreMotion Impl; 389 auto &AA = AM.getResult<AAManager>(F); 390 if (!Impl.run(F, AA)) 391 return PreservedAnalyses::all(); 392 393 PreservedAnalyses PA; 394 PA.preserveSet<CFGAnalyses>(); 395 PA.preserve<GlobalsAA>(); 396 return PA; 397 } 398