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/Instructions.h" 82 #include "llvm/InitializePasses.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 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) && 195 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) && 196 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0)) { 197 return Store1; 198 } 199 } 200 return nullptr; 201 } 202 203 /// 204 /// Create a PHI node in BB for the operands of S0 and S1 205 /// 206 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 207 StoreInst *S1) { 208 // Create a phi if the values mismatch. 209 Value *Opd1 = S0->getValueOperand(); 210 Value *Opd2 = S1->getValueOperand(); 211 if (Opd1 == Opd2) 212 return nullptr; 213 214 auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink", 215 &BB->front()); 216 NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 217 NewPN->addIncoming(Opd1, S0->getParent()); 218 NewPN->addIncoming(Opd2, S1->getParent()); 219 return NewPN; 220 } 221 222 /// 223 /// Check if 2 stores can be sunk, optionally together with corresponding GEPs. 224 /// 225 bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0, 226 StoreInst *S1) const { 227 if (S0->getPointerOperand() == S1->getPointerOperand()) 228 return true; 229 auto *GEP0 = dyn_cast<GetElementPtrInst>(S0->getPointerOperand()); 230 auto *GEP1 = dyn_cast<GetElementPtrInst>(S1->getPointerOperand()); 231 return GEP0 && GEP1 && GEP0->isIdenticalTo(GEP1) && GEP0->hasOneUse() && 232 (GEP0->getParent() == S0->getParent()) && GEP1->hasOneUse() && 233 (GEP1->getParent() == S1->getParent()); 234 } 235 236 /// 237 /// Merge two stores to same address and sink into \p BB 238 /// 239 /// Optionally also sinks GEP instruction computing the store address 240 /// 241 void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0, 242 StoreInst *S1) { 243 Value *Ptr0 = S0->getPointerOperand(); 244 Value *Ptr1 = S1->getPointerOperand(); 245 // Only one definition? 246 LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 247 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 248 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 249 // Hoist the instruction. 250 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 251 // Intersect optional metadata. 252 S0->andIRFlags(S1); 253 S0->dropUnknownNonDebugMetadata(); 254 S0->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 255 S0->mergeDIAssignID(S1); 256 257 // Create the new store to be inserted at the join point. 258 StoreInst *SNew = cast<StoreInst>(S0->clone()); 259 SNew->insertBefore(&*InsertPt); 260 // New PHI operand? Use it. 261 if (PHINode *NewPN = getPHIOperand(BB, S0, S1)) 262 SNew->setOperand(0, NewPN); 263 S0->eraseFromParent(); 264 S1->eraseFromParent(); 265 266 if (Ptr0 != Ptr1) { 267 auto *GEP0 = cast<GetElementPtrInst>(Ptr0); 268 auto *GEP1 = cast<GetElementPtrInst>(Ptr1); 269 Instruction *GEPNew = GEP0->clone(); 270 GEPNew->insertBefore(SNew); 271 GEPNew->applyMergedLocation(GEP0->getDebugLoc(), GEP1->getDebugLoc()); 272 SNew->setOperand(1, GEPNew); 273 GEP0->replaceAllUsesWith(GEPNew); 274 GEP0->eraseFromParent(); 275 GEP1->replaceAllUsesWith(GEPNew); 276 GEP1->eraseFromParent(); 277 } 278 } 279 280 /// 281 /// True when two stores are equivalent and can sink into the footer 282 /// 283 /// Starting from a diamond head block, iterate over the instructions in one 284 /// successor block and try to match a store in the second successor. 285 /// 286 bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) { 287 288 bool MergedStores = false; 289 BasicBlock *TailBB = getDiamondTail(HeadBB); 290 BasicBlock *SinkBB = TailBB; 291 assert(SinkBB && "Footer of a diamond cannot be empty"); 292 293 succ_iterator SI = succ_begin(HeadBB); 294 assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors"); 295 BasicBlock *Pred0 = *SI; 296 ++SI; 297 assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor"); 298 BasicBlock *Pred1 = *SI; 299 // tail block of a diamond/hammock? 300 if (Pred0 == Pred1) 301 return false; // No. 302 // bail out early if we can not merge into the footer BB 303 if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3)) 304 return false; 305 // #Instructions in Pred1 for Compile Time Control 306 auto InstsNoDbg = Pred1->instructionsWithoutDebug(); 307 int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end()); 308 int NStores = 0; 309 310 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 311 RBI != RBE;) { 312 313 Instruction *I = &*RBI; 314 ++RBI; 315 316 // Don't sink non-simple (atomic, volatile) stores. 317 auto *S0 = dyn_cast<StoreInst>(I); 318 if (!S0 || !S0->isSimple()) 319 continue; 320 321 ++NStores; 322 if (NStores * Size1 >= MagicCompileTimeControl) 323 break; 324 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 325 if (!canSinkStoresAndGEPs(S0, S1)) 326 // Don't attempt to sink below stores that had to stick around 327 // But after removal of a store and some of its feeding 328 // instruction search again from the beginning since the iterator 329 // is likely stale at this point. 330 break; 331 332 if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) { 333 // We have more than 2 predecessors. Insert a new block 334 // postdominating 2 predecessors we're going to sink from. 335 SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split"); 336 if (!SinkBB) 337 break; 338 } 339 340 MergedStores = true; 341 sinkStoresAndGEPs(SinkBB, S0, S1); 342 RBI = Pred0->rbegin(); 343 RBE = Pred0->rend(); 344 LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 345 } 346 } 347 return MergedStores; 348 } 349 350 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) { 351 this->AA = &AA; 352 353 bool Changed = false; 354 LLVM_DEBUG(dbgs() << "Instruction Merger\n"); 355 356 // Merge unconditional branches, allowing PRE to catch more 357 // optimization opportunities. 358 // This loop doesn't care about newly inserted/split blocks 359 // since they never will be diamond heads. 360 for (BasicBlock &BB : make_early_inc_range(F)) 361 // Hoist equivalent loads and sink stores 362 // outside diamonds when possible 363 if (isDiamondHead(&BB)) 364 Changed |= mergeStores(&BB); 365 return Changed; 366 } 367 368 namespace { 369 class MergedLoadStoreMotionLegacyPass : public FunctionPass { 370 const bool SplitFooterBB; 371 public: 372 static char ID; // Pass identification, replacement for typeid 373 MergedLoadStoreMotionLegacyPass(bool SplitFooterBB = false) 374 : FunctionPass(ID), SplitFooterBB(SplitFooterBB) { 375 initializeMergedLoadStoreMotionLegacyPassPass( 376 *PassRegistry::getPassRegistry()); 377 } 378 379 /// 380 /// Run the transformation for each function 381 /// 382 bool runOnFunction(Function &F) override { 383 if (skipFunction(F)) 384 return false; 385 MergedLoadStoreMotion Impl(SplitFooterBB); 386 return Impl.run(F, getAnalysis<AAResultsWrapperPass>().getAAResults()); 387 } 388 389 private: 390 void getAnalysisUsage(AnalysisUsage &AU) const override { 391 if (!SplitFooterBB) 392 AU.setPreservesCFG(); 393 AU.addRequired<AAResultsWrapperPass>(); 394 AU.addPreserved<GlobalsAAWrapperPass>(); 395 } 396 }; 397 398 char MergedLoadStoreMotionLegacyPass::ID = 0; 399 } // anonymous namespace 400 401 /// 402 /// createMergedLoadStoreMotionPass - The public interface to this file. 403 /// 404 FunctionPass *llvm::createMergedLoadStoreMotionPass(bool SplitFooterBB) { 405 return new MergedLoadStoreMotionLegacyPass(SplitFooterBB); 406 } 407 408 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion", 409 "MergedLoadStoreMotion", false, false) 410 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 411 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion", 412 "MergedLoadStoreMotion", false, false) 413 414 PreservedAnalyses 415 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) { 416 MergedLoadStoreMotion Impl(Options.SplitFooterBB); 417 auto &AA = AM.getResult<AAManager>(F); 418 if (!Impl.run(F, AA)) 419 return PreservedAnalyses::all(); 420 421 PreservedAnalyses PA; 422 if (!Options.SplitFooterBB) 423 PA.preserveSet<CFGAnalyses>(); 424 return PA; 425 } 426 427 void MergedLoadStoreMotionPass::printPipeline( 428 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 429 static_cast<PassInfoMixin<MergedLoadStoreMotionPass> *>(this)->printPipeline( 430 OS, MapClassName2PassName); 431 OS << "<"; 432 OS << (Options.SplitFooterBB ? "" : "no-") << "split-footer-bb"; 433 OS << ">"; 434 } 435