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. New tail/footer block may be 18 // insterted if the tail/footer block has more predecessors (not only the two 19 // predecessors that are forming the diamond). It hoists / sinks when it thinks 20 // it safe to do so. This optimization helps with eg. hiding load latencies, 21 // triggering if-conversion, and reducing static code size. 22 // 23 // NOTE: This code no longer performs load hoisting, it is subsumed by GVNHoist. 24 // 25 //===----------------------------------------------------------------------===// 26 // 27 // 28 // Example: 29 // Diamond shaped code before merge: 30 // 31 // header: 32 // br %cond, label %if.then, label %if.else 33 // + + 34 // + + 35 // + + 36 // if.then: if.else: 37 // %lt = load %addr_l %le = load %addr_l 38 // <use %lt> <use %le> 39 // <...> <...> 40 // store %st, %addr_s store %se, %addr_s 41 // br label %if.end br label %if.end 42 // + + 43 // + + 44 // + + 45 // if.end ("footer"): 46 // <...> 47 // 48 // Diamond shaped code after merge: 49 // 50 // header: 51 // %l = load %addr_l 52 // br %cond, label %if.then, label %if.else 53 // + + 54 // + + 55 // + + 56 // if.then: if.else: 57 // <use %l> <use %l> 58 // <...> <...> 59 // br label %if.end br label %if.end 60 // + + 61 // + + 62 // + + 63 // if.end ("footer"): 64 // %s.sink = phi [%st, if.then], [%se, if.else] 65 // <...> 66 // store %s.sink, %addr_s 67 // <...> 68 // 69 // 70 //===----------------------- TODO -----------------------------------------===// 71 // 72 // 1) Generalize to regions other than diamonds 73 // 2) Be more aggressive merging memory operations 74 // Note that both changes require register pressure control 75 // 76 //===----------------------------------------------------------------------===// 77 78 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h" 79 #include "llvm/Analysis/AliasAnalysis.h" 80 #include "llvm/Analysis/GlobalsModRef.h" 81 #include "llvm/IR/IRBuilder.h" 82 #include "llvm/IR/Instructions.h" 83 #include "llvm/Support/Debug.h" 84 #include "llvm/Support/raw_ostream.h" 85 #include "llvm/Transforms/Scalar.h" 86 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 87 88 using namespace llvm; 89 90 #define DEBUG_TYPE "mldst-motion" 91 92 namespace { 93 //===----------------------------------------------------------------------===// 94 // MergedLoadStoreMotion Pass 95 //===----------------------------------------------------------------------===// 96 class MergedLoadStoreMotion { 97 AliasAnalysis *AA = nullptr; 98 99 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity, 100 // where Size0 and Size1 are the #instructions on the two sides of 101 // the diamond. The constant chosen here is arbitrary. Compiler Time 102 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl. 103 const int MagicCompileTimeControl = 250; 104 105 const bool SplitFooterBB; 106 public: 107 MergedLoadStoreMotion(bool SplitFooterBB) : SplitFooterBB(SplitFooterBB) {} 108 bool run(Function &F, AliasAnalysis &AA); 109 110 private: 111 BasicBlock *getDiamondTail(BasicBlock *BB); 112 bool isDiamondHead(BasicBlock *BB); 113 // Routines for sinking stores 114 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 115 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 116 bool isStoreSinkBarrierInRange(const Instruction &Start, 117 const Instruction &End, MemoryLocation Loc); 118 bool canSinkStoresAndGEPs(StoreInst *S0, StoreInst *S1) const; 119 void sinkStoresAndGEPs(BasicBlock *BB, StoreInst *SinkCand, 120 StoreInst *ElseInst); 121 bool mergeStores(BasicBlock *BB); 122 }; 123 } // end anonymous namespace 124 125 /// 126 /// Return tail block of a diamond. 127 /// 128 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 129 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 130 return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor(); 131 } 132 133 /// 134 /// True when BB is the head of a diamond (hammock) 135 /// 136 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 137 if (!BB) 138 return false; 139 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 140 if (!BI || !BI->isConditional()) 141 return false; 142 143 BasicBlock *Succ0 = BI->getSuccessor(0); 144 BasicBlock *Succ1 = BI->getSuccessor(1); 145 146 if (!Succ0->getSinglePredecessor()) 147 return false; 148 if (!Succ1->getSinglePredecessor()) 149 return false; 150 151 BasicBlock *Succ0Succ = Succ0->getSingleSuccessor(); 152 BasicBlock *Succ1Succ = Succ1->getSingleSuccessor(); 153 // Ignore triangles. 154 if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ) 155 return false; 156 return true; 157 } 158 159 160 /// 161 /// True when instruction is a sink barrier for a store 162 /// located in Loc 163 /// 164 /// Whenever an instruction could possibly read or modify the 165 /// value being stored or protect against the store from 166 /// happening it is considered a sink barrier. 167 /// 168 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 169 const Instruction &End, 170 MemoryLocation Loc) { 171 for (const Instruction &Inst : 172 make_range(Start.getIterator(), End.getIterator())) 173 if (Inst.mayThrow()) 174 return true; 175 return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef); 176 } 177 178 /// 179 /// Check if \p BB contains a store to the same address as \p SI 180 /// 181 /// \return The store in \p when it is safe to sink. Otherwise return Null. 182 /// 183 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 184 StoreInst *Store0) { 185 LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 186 BasicBlock *BB0 = Store0->getParent(); 187 for (Instruction &Inst : reverse(*BB1)) { 188 auto *Store1 = dyn_cast<StoreInst>(&Inst); 189 if (!Store1) 190 continue; 191 192 MemoryLocation Loc0 = MemoryLocation::get(Store0); 193 MemoryLocation Loc1 = MemoryLocation::get(Store1); 194 195 if (AA->isMustAlias(Loc0, Loc1) && 196 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) && 197 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0) && 198 Store0->hasSameSpecialState(Store1) && 199 CastInst::isBitOrNoopPointerCastable( 200 Store0->getValueOperand()->getType(), 201 Store1->getValueOperand()->getType(), 202 Store0->getModule()->getDataLayout())) 203 return Store1; 204 } 205 return nullptr; 206 } 207 208 /// 209 /// Create a PHI node in BB for the operands of S0 and S1 210 /// 211 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 212 StoreInst *S1) { 213 // Create a phi if the values mismatch. 214 Value *Opd1 = S0->getValueOperand(); 215 Value *Opd2 = S1->getValueOperand(); 216 if (Opd1 == Opd2) 217 return nullptr; 218 219 auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink"); 220 NewPN->insertBefore(BB->begin()); 221 NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 222 NewPN->addIncoming(Opd1, S0->getParent()); 223 NewPN->addIncoming(Opd2, S1->getParent()); 224 return NewPN; 225 } 226 227 /// 228 /// Check if 2 stores can be sunk, optionally together with corresponding GEPs. 229 /// 230 bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0, 231 StoreInst *S1) const { 232 if (S0->getPointerOperand() == S1->getPointerOperand()) 233 return true; 234 auto *GEP0 = dyn_cast<GetElementPtrInst>(S0->getPointerOperand()); 235 auto *GEP1 = dyn_cast<GetElementPtrInst>(S1->getPointerOperand()); 236 return GEP0 && GEP1 && GEP0->isIdenticalTo(GEP1) && GEP0->hasOneUse() && 237 (GEP0->getParent() == S0->getParent()) && GEP1->hasOneUse() && 238 (GEP1->getParent() == S1->getParent()); 239 } 240 241 /// 242 /// Merge two stores to same address and sink into \p BB 243 /// 244 /// Optionally also sinks GEP instruction computing the store address 245 /// 246 void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0, 247 StoreInst *S1) { 248 Value *Ptr0 = S0->getPointerOperand(); 249 Value *Ptr1 = S1->getPointerOperand(); 250 // Only one definition? 251 LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 252 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 253 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 254 // Hoist the instruction. 255 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 256 // Intersect optional metadata. 257 S0->andIRFlags(S1); 258 S0->dropUnknownNonDebugMetadata(); 259 S0->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 260 S0->mergeDIAssignID(S1); 261 262 // Insert bitcast for conflicting typed stores (or just use original value if 263 // same type). 264 IRBuilder<> Builder(S0); 265 auto Cast = Builder.CreateBitOrPointerCast(S0->getValueOperand(), 266 S1->getValueOperand()->getType()); 267 S0->setOperand(0, Cast); 268 269 // Create the new store to be inserted at the join point. 270 StoreInst *SNew = cast<StoreInst>(S0->clone()); 271 SNew->insertBefore(InsertPt); 272 // New PHI operand? Use it. 273 if (PHINode *NewPN = getPHIOperand(BB, S0, S1)) 274 SNew->setOperand(0, NewPN); 275 S0->eraseFromParent(); 276 S1->eraseFromParent(); 277 278 if (Ptr0 != Ptr1) { 279 auto *GEP0 = cast<GetElementPtrInst>(Ptr0); 280 auto *GEP1 = cast<GetElementPtrInst>(Ptr1); 281 Instruction *GEPNew = GEP0->clone(); 282 GEPNew->insertBefore(SNew); 283 GEPNew->applyMergedLocation(GEP0->getDebugLoc(), GEP1->getDebugLoc()); 284 SNew->setOperand(1, GEPNew); 285 GEP0->replaceAllUsesWith(GEPNew); 286 GEP0->eraseFromParent(); 287 GEP1->replaceAllUsesWith(GEPNew); 288 GEP1->eraseFromParent(); 289 } 290 } 291 292 /// 293 /// True when two stores are equivalent and can sink into the footer 294 /// 295 /// Starting from a diamond head block, iterate over the instructions in one 296 /// successor block and try to match a store in the second successor. 297 /// 298 bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) { 299 300 bool MergedStores = false; 301 BasicBlock *TailBB = getDiamondTail(HeadBB); 302 BasicBlock *SinkBB = TailBB; 303 assert(SinkBB && "Footer of a diamond cannot be empty"); 304 305 succ_iterator SI = succ_begin(HeadBB); 306 assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors"); 307 BasicBlock *Pred0 = *SI; 308 ++SI; 309 assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor"); 310 BasicBlock *Pred1 = *SI; 311 // tail block of a diamond/hammock? 312 if (Pred0 == Pred1) 313 return false; // No. 314 // bail out early if we can not merge into the footer BB 315 if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3)) 316 return false; 317 // #Instructions in Pred1 for Compile Time Control 318 auto InstsNoDbg = Pred1->instructionsWithoutDebug(); 319 int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end()); 320 int NStores = 0; 321 322 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 323 RBI != RBE;) { 324 325 Instruction *I = &*RBI; 326 ++RBI; 327 328 // Don't sink non-simple (atomic, volatile) stores. 329 auto *S0 = dyn_cast<StoreInst>(I); 330 if (!S0 || !S0->isSimple()) 331 continue; 332 333 ++NStores; 334 if (NStores * Size1 >= MagicCompileTimeControl) 335 break; 336 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 337 if (!canSinkStoresAndGEPs(S0, S1)) 338 // Don't attempt to sink below stores that had to stick around 339 // But after removal of a store and some of its feeding 340 // instruction search again from the beginning since the iterator 341 // is likely stale at this point. 342 break; 343 344 if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) { 345 // We have more than 2 predecessors. Insert a new block 346 // postdominating 2 predecessors we're going to sink from. 347 SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split"); 348 if (!SinkBB) 349 break; 350 } 351 352 MergedStores = true; 353 sinkStoresAndGEPs(SinkBB, S0, S1); 354 RBI = Pred0->rbegin(); 355 RBE = Pred0->rend(); 356 LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 357 } 358 } 359 return MergedStores; 360 } 361 362 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) { 363 this->AA = &AA; 364 365 bool Changed = false; 366 LLVM_DEBUG(dbgs() << "Instruction Merger\n"); 367 368 // Merge unconditional branches, allowing PRE to catch more 369 // optimization opportunities. 370 // This loop doesn't care about newly inserted/split blocks 371 // since they never will be diamond heads. 372 for (BasicBlock &BB : make_early_inc_range(F)) 373 // Hoist equivalent loads and sink stores 374 // outside diamonds when possible 375 if (isDiamondHead(&BB)) 376 Changed |= mergeStores(&BB); 377 return Changed; 378 } 379 380 PreservedAnalyses 381 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) { 382 MergedLoadStoreMotion Impl(Options.SplitFooterBB); 383 auto &AA = AM.getResult<AAManager>(F); 384 if (!Impl.run(F, AA)) 385 return PreservedAnalyses::all(); 386 387 PreservedAnalyses PA; 388 if (!Options.SplitFooterBB) 389 PA.preserveSet<CFGAnalyses>(); 390 return PA; 391 } 392 393 void MergedLoadStoreMotionPass::printPipeline( 394 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 395 static_cast<PassInfoMixin<MergedLoadStoreMotionPass> *>(this)->printPipeline( 396 OS, MapClassName2PassName); 397 OS << '<'; 398 OS << (Options.SplitFooterBB ? "" : "no-") << "split-footer-bb"; 399 OS << '>'; 400 } 401