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/InitializePasses.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 const bool SplitFooterBB; 107 public: 108 MergedLoadStoreMotion(bool SplitFooterBB) : SplitFooterBB(SplitFooterBB) {} 109 bool run(Function &F, AliasAnalysis &AA); 110 111 private: 112 BasicBlock *getDiamondTail(BasicBlock *BB); 113 bool isDiamondHead(BasicBlock *BB); 114 // Routines for sinking stores 115 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 116 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 117 bool isStoreSinkBarrierInRange(const Instruction &Start, 118 const Instruction &End, MemoryLocation Loc); 119 bool canSinkStoresAndGEPs(StoreInst *S0, StoreInst *S1) const; 120 void sinkStoresAndGEPs(BasicBlock *BB, StoreInst *SinkCand, 121 StoreInst *ElseInst); 122 bool mergeStores(BasicBlock *BB); 123 }; 124 } // end anonymous namespace 125 126 /// 127 /// Return tail block of a diamond. 128 /// 129 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 130 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 131 return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor(); 132 } 133 134 /// 135 /// True when BB is the head of a diamond (hammock) 136 /// 137 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 138 if (!BB) 139 return false; 140 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 141 if (!BI || !BI->isConditional()) 142 return false; 143 144 BasicBlock *Succ0 = BI->getSuccessor(0); 145 BasicBlock *Succ1 = BI->getSuccessor(1); 146 147 if (!Succ0->getSinglePredecessor()) 148 return false; 149 if (!Succ1->getSinglePredecessor()) 150 return false; 151 152 BasicBlock *Succ0Succ = Succ0->getSingleSuccessor(); 153 BasicBlock *Succ1Succ = Succ1->getSingleSuccessor(); 154 // Ignore triangles. 155 if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ) 156 return false; 157 return true; 158 } 159 160 161 /// 162 /// True when instruction is a sink barrier for a store 163 /// located in Loc 164 /// 165 /// Whenever an instruction could possibly read or modify the 166 /// value being stored or protect against the store from 167 /// happening it is considered a sink barrier. 168 /// 169 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 170 const Instruction &End, 171 MemoryLocation Loc) { 172 for (const Instruction &Inst : 173 make_range(Start.getIterator(), End.getIterator())) 174 if (Inst.mayThrow()) 175 return true; 176 return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef); 177 } 178 179 /// 180 /// Check if \p BB contains a store to the same address as \p SI 181 /// 182 /// \return The store in \p when it is safe to sink. Otherwise return Null. 183 /// 184 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 185 StoreInst *Store0) { 186 LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 187 BasicBlock *BB0 = Store0->getParent(); 188 for (Instruction &Inst : reverse(*BB1)) { 189 auto *Store1 = dyn_cast<StoreInst>(&Inst); 190 if (!Store1) 191 continue; 192 193 MemoryLocation Loc0 = MemoryLocation::get(Store0); 194 MemoryLocation Loc1 = MemoryLocation::get(Store1); 195 196 if (AA->isMustAlias(Loc0, Loc1) && 197 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) && 198 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0) && 199 Store0->hasSameSpecialState(Store1) && 200 CastInst::isBitOrNoopPointerCastable( 201 Store0->getValueOperand()->getType(), 202 Store1->getValueOperand()->getType(), 203 Store0->getModule()->getDataLayout())) 204 return Store1; 205 } 206 return nullptr; 207 } 208 209 /// 210 /// Create a PHI node in BB for the operands of S0 and S1 211 /// 212 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 213 StoreInst *S1) { 214 // Create a phi if the values mismatch. 215 Value *Opd1 = S0->getValueOperand(); 216 Value *Opd2 = S1->getValueOperand(); 217 if (Opd1 == Opd2) 218 return nullptr; 219 220 auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink", 221 &BB->front()); 222 NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 223 NewPN->addIncoming(Opd1, S0->getParent()); 224 NewPN->addIncoming(Opd2, S1->getParent()); 225 return NewPN; 226 } 227 228 /// 229 /// Check if 2 stores can be sunk, optionally together with corresponding GEPs. 230 /// 231 bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0, 232 StoreInst *S1) const { 233 if (S0->getPointerOperand() == S1->getPointerOperand()) 234 return true; 235 auto *GEP0 = dyn_cast<GetElementPtrInst>(S0->getPointerOperand()); 236 auto *GEP1 = dyn_cast<GetElementPtrInst>(S1->getPointerOperand()); 237 return GEP0 && GEP1 && GEP0->isIdenticalTo(GEP1) && GEP0->hasOneUse() && 238 (GEP0->getParent() == S0->getParent()) && GEP1->hasOneUse() && 239 (GEP1->getParent() == S1->getParent()); 240 } 241 242 /// 243 /// Merge two stores to same address and sink into \p BB 244 /// 245 /// Optionally also sinks GEP instruction computing the store address 246 /// 247 void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0, 248 StoreInst *S1) { 249 Value *Ptr0 = S0->getPointerOperand(); 250 Value *Ptr1 = S1->getPointerOperand(); 251 // Only one definition? 252 LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 253 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 254 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 255 // Hoist the instruction. 256 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 257 // Intersect optional metadata. 258 S0->andIRFlags(S1); 259 S0->dropUnknownNonDebugMetadata(); 260 S0->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 261 S0->mergeDIAssignID(S1); 262 263 // Insert bitcast for conflicting typed stores (or just use original value if 264 // same type). 265 IRBuilder<> Builder(S0); 266 auto Cast = Builder.CreateBitOrPointerCast(S0->getValueOperand(), 267 S1->getValueOperand()->getType()); 268 S0->setOperand(0, Cast); 269 270 // Create the new store to be inserted at the join point. 271 StoreInst *SNew = cast<StoreInst>(S0->clone()); 272 SNew->insertBefore(&*InsertPt); 273 // New PHI operand? Use it. 274 if (PHINode *NewPN = getPHIOperand(BB, S0, S1)) 275 SNew->setOperand(0, NewPN); 276 S0->eraseFromParent(); 277 S1->eraseFromParent(); 278 279 if (Ptr0 != Ptr1) { 280 auto *GEP0 = cast<GetElementPtrInst>(Ptr0); 281 auto *GEP1 = cast<GetElementPtrInst>(Ptr1); 282 Instruction *GEPNew = GEP0->clone(); 283 GEPNew->insertBefore(SNew); 284 GEPNew->applyMergedLocation(GEP0->getDebugLoc(), GEP1->getDebugLoc()); 285 SNew->setOperand(1, GEPNew); 286 GEP0->replaceAllUsesWith(GEPNew); 287 GEP0->eraseFromParent(); 288 GEP1->replaceAllUsesWith(GEPNew); 289 GEP1->eraseFromParent(); 290 } 291 } 292 293 /// 294 /// True when two stores are equivalent and can sink into the footer 295 /// 296 /// Starting from a diamond head block, iterate over the instructions in one 297 /// successor block and try to match a store in the second successor. 298 /// 299 bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) { 300 301 bool MergedStores = false; 302 BasicBlock *TailBB = getDiamondTail(HeadBB); 303 BasicBlock *SinkBB = TailBB; 304 assert(SinkBB && "Footer of a diamond cannot be empty"); 305 306 succ_iterator SI = succ_begin(HeadBB); 307 assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors"); 308 BasicBlock *Pred0 = *SI; 309 ++SI; 310 assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor"); 311 BasicBlock *Pred1 = *SI; 312 // tail block of a diamond/hammock? 313 if (Pred0 == Pred1) 314 return false; // No. 315 // bail out early if we can not merge into the footer BB 316 if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3)) 317 return false; 318 // #Instructions in Pred1 for Compile Time Control 319 auto InstsNoDbg = Pred1->instructionsWithoutDebug(); 320 int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end()); 321 int NStores = 0; 322 323 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 324 RBI != RBE;) { 325 326 Instruction *I = &*RBI; 327 ++RBI; 328 329 // Don't sink non-simple (atomic, volatile) stores. 330 auto *S0 = dyn_cast<StoreInst>(I); 331 if (!S0 || !S0->isSimple()) 332 continue; 333 334 ++NStores; 335 if (NStores * Size1 >= MagicCompileTimeControl) 336 break; 337 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 338 if (!canSinkStoresAndGEPs(S0, S1)) 339 // Don't attempt to sink below stores that had to stick around 340 // But after removal of a store and some of its feeding 341 // instruction search again from the beginning since the iterator 342 // is likely stale at this point. 343 break; 344 345 if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) { 346 // We have more than 2 predecessors. Insert a new block 347 // postdominating 2 predecessors we're going to sink from. 348 SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split"); 349 if (!SinkBB) 350 break; 351 } 352 353 MergedStores = true; 354 sinkStoresAndGEPs(SinkBB, S0, S1); 355 RBI = Pred0->rbegin(); 356 RBE = Pred0->rend(); 357 LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 358 } 359 } 360 return MergedStores; 361 } 362 363 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) { 364 this->AA = &AA; 365 366 bool Changed = false; 367 LLVM_DEBUG(dbgs() << "Instruction Merger\n"); 368 369 // Merge unconditional branches, allowing PRE to catch more 370 // optimization opportunities. 371 // This loop doesn't care about newly inserted/split blocks 372 // since they never will be diamond heads. 373 for (BasicBlock &BB : make_early_inc_range(F)) 374 // Hoist equivalent loads and sink stores 375 // outside diamonds when possible 376 if (isDiamondHead(&BB)) 377 Changed |= mergeStores(&BB); 378 return Changed; 379 } 380 381 namespace { 382 class MergedLoadStoreMotionLegacyPass : public FunctionPass { 383 const bool SplitFooterBB; 384 public: 385 static char ID; // Pass identification, replacement for typeid 386 MergedLoadStoreMotionLegacyPass(bool SplitFooterBB = false) 387 : FunctionPass(ID), SplitFooterBB(SplitFooterBB) { 388 initializeMergedLoadStoreMotionLegacyPassPass( 389 *PassRegistry::getPassRegistry()); 390 } 391 392 /// 393 /// Run the transformation for each function 394 /// 395 bool runOnFunction(Function &F) override { 396 if (skipFunction(F)) 397 return false; 398 MergedLoadStoreMotion Impl(SplitFooterBB); 399 return Impl.run(F, getAnalysis<AAResultsWrapperPass>().getAAResults()); 400 } 401 402 private: 403 void getAnalysisUsage(AnalysisUsage &AU) const override { 404 if (!SplitFooterBB) 405 AU.setPreservesCFG(); 406 AU.addRequired<AAResultsWrapperPass>(); 407 AU.addPreserved<GlobalsAAWrapperPass>(); 408 } 409 }; 410 411 char MergedLoadStoreMotionLegacyPass::ID = 0; 412 } // anonymous namespace 413 414 /// 415 /// createMergedLoadStoreMotionPass - The public interface to this file. 416 /// 417 FunctionPass *llvm::createMergedLoadStoreMotionPass(bool SplitFooterBB) { 418 return new MergedLoadStoreMotionLegacyPass(SplitFooterBB); 419 } 420 421 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion", 422 "MergedLoadStoreMotion", false, false) 423 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 424 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion", 425 "MergedLoadStoreMotion", false, false) 426 427 PreservedAnalyses 428 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) { 429 MergedLoadStoreMotion Impl(Options.SplitFooterBB); 430 auto &AA = AM.getResult<AAManager>(F); 431 if (!Impl.run(F, AA)) 432 return PreservedAnalyses::all(); 433 434 PreservedAnalyses PA; 435 if (!Options.SplitFooterBB) 436 PA.preserveSet<CFGAnalyses>(); 437 return PA; 438 } 439 440 void MergedLoadStoreMotionPass::printPipeline( 441 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 442 static_cast<PassInfoMixin<MergedLoadStoreMotionPass> *>(this)->printPipeline( 443 OS, MapClassName2PassName); 444 OS << '<'; 445 OS << (Options.SplitFooterBB ? "" : "no-") << "split-footer-bb"; 446 OS << '>'; 447 } 448