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/ADT/Statistic.h" 80 #include "llvm/Analysis/AliasAnalysis.h" 81 #include "llvm/Analysis/CFG.h" 82 #include "llvm/Analysis/GlobalsModRef.h" 83 #include "llvm/Analysis/Loads.h" 84 #include "llvm/Analysis/ValueTracking.h" 85 #include "llvm/IR/Instructions.h" 86 #include "llvm/IR/Metadata.h" 87 #include "llvm/InitializePasses.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 AliasAnalysis *AA = nullptr; 103 104 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity, 105 // where Size0 and Size1 are the #instructions on the two sides of 106 // the diamond. The constant chosen here is arbitrary. Compiler Time 107 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl. 108 const int MagicCompileTimeControl = 250; 109 110 const bool SplitFooterBB; 111 public: 112 MergedLoadStoreMotion(bool SplitFooterBB) : SplitFooterBB(SplitFooterBB) {} 113 bool run(Function &F, AliasAnalysis &AA); 114 115 private: 116 BasicBlock *getDiamondTail(BasicBlock *BB); 117 bool isDiamondHead(BasicBlock *BB); 118 // Routines for sinking stores 119 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 120 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 121 bool isStoreSinkBarrierInRange(const Instruction &Start, 122 const Instruction &End, MemoryLocation Loc); 123 bool canSinkStoresAndGEPs(StoreInst *S0, StoreInst *S1) const; 124 void sinkStoresAndGEPs(BasicBlock *BB, StoreInst *SinkCand, 125 StoreInst *ElseInst); 126 bool mergeStores(BasicBlock *BB); 127 }; 128 } // end anonymous namespace 129 130 /// 131 /// Return tail block of a diamond. 132 /// 133 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 134 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 135 return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor(); 136 } 137 138 /// 139 /// True when BB is the head of a diamond (hammock) 140 /// 141 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 142 if (!BB) 143 return false; 144 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 145 if (!BI || !BI->isConditional()) 146 return false; 147 148 BasicBlock *Succ0 = BI->getSuccessor(0); 149 BasicBlock *Succ1 = BI->getSuccessor(1); 150 151 if (!Succ0->getSinglePredecessor()) 152 return false; 153 if (!Succ1->getSinglePredecessor()) 154 return false; 155 156 BasicBlock *Succ0Succ = Succ0->getSingleSuccessor(); 157 BasicBlock *Succ1Succ = Succ1->getSingleSuccessor(); 158 // Ignore triangles. 159 if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ) 160 return false; 161 return true; 162 } 163 164 165 /// 166 /// True when instruction is a sink barrier for a store 167 /// located in Loc 168 /// 169 /// Whenever an instruction could possibly read or modify the 170 /// value being stored or protect against the store from 171 /// happening it is considered a sink barrier. 172 /// 173 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 174 const Instruction &End, 175 MemoryLocation Loc) { 176 for (const Instruction &Inst : 177 make_range(Start.getIterator(), End.getIterator())) 178 if (Inst.mayThrow()) 179 return true; 180 return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef); 181 } 182 183 /// 184 /// Check if \p BB contains a store to the same address as \p SI 185 /// 186 /// \return The store in \p when it is safe to sink. Otherwise return Null. 187 /// 188 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 189 StoreInst *Store0) { 190 LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 191 BasicBlock *BB0 = Store0->getParent(); 192 for (Instruction &Inst : reverse(*BB1)) { 193 auto *Store1 = dyn_cast<StoreInst>(&Inst); 194 if (!Store1) 195 continue; 196 197 MemoryLocation Loc0 = MemoryLocation::get(Store0); 198 MemoryLocation Loc1 = MemoryLocation::get(Store1); 199 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) && 200 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) && 201 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0)) { 202 return Store1; 203 } 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 &BB->front()); 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 together with corresponding GEPs 229 /// 230 bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0, 231 StoreInst *S1) const { 232 auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 233 auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 234 return A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() && 235 (A0->getParent() == S0->getParent()) && A1->hasOneUse() && 236 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0); 237 } 238 239 /// 240 /// Merge two stores to same address and sink into \p BB 241 /// 242 /// Also sinks GEP instruction computing the store address 243 /// 244 void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0, 245 StoreInst *S1) { 246 // Only one definition? 247 auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 248 auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 249 LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 250 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 251 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 252 // Hoist the instruction. 253 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 254 // Intersect optional metadata. 255 S0->andIRFlags(S1); 256 S0->dropUnknownNonDebugMetadata(); 257 258 // Create the new store to be inserted at the join point. 259 StoreInst *SNew = cast<StoreInst>(S0->clone()); 260 Instruction *ANew = A0->clone(); 261 SNew->insertBefore(&*InsertPt); 262 ANew->insertBefore(SNew); 263 264 assert(S0->getParent() == A0->getParent()); 265 assert(S1->getParent() == A1->getParent()); 266 267 // New PHI operand? Use it. 268 if (PHINode *NewPN = getPHIOperand(BB, S0, S1)) 269 SNew->setOperand(0, NewPN); 270 S0->eraseFromParent(); 271 S1->eraseFromParent(); 272 A0->replaceAllUsesWith(ANew); 273 A0->eraseFromParent(); 274 A1->replaceAllUsesWith(ANew); 275 A1->eraseFromParent(); 276 } 277 278 /// 279 /// True when two stores are equivalent and can sink into the footer 280 /// 281 /// Starting from a diamond head block, iterate over the instructions in one 282 /// successor block and try to match a store in the second successor. 283 /// 284 bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) { 285 286 bool MergedStores = false; 287 BasicBlock *TailBB = getDiamondTail(HeadBB); 288 BasicBlock *SinkBB = TailBB; 289 assert(SinkBB && "Footer of a diamond cannot be empty"); 290 291 succ_iterator SI = succ_begin(HeadBB); 292 assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors"); 293 BasicBlock *Pred0 = *SI; 294 ++SI; 295 assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor"); 296 BasicBlock *Pred1 = *SI; 297 // tail block of a diamond/hammock? 298 if (Pred0 == Pred1) 299 return false; // No. 300 // bail out early if we can not merge into the footer BB 301 if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3)) 302 return false; 303 // #Instructions in Pred1 for Compile Time Control 304 auto InstsNoDbg = Pred1->instructionsWithoutDebug(); 305 int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end()); 306 int NStores = 0; 307 308 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 309 RBI != RBE;) { 310 311 Instruction *I = &*RBI; 312 ++RBI; 313 314 // Don't sink non-simple (atomic, volatile) stores. 315 auto *S0 = dyn_cast<StoreInst>(I); 316 if (!S0 || !S0->isSimple()) 317 continue; 318 319 ++NStores; 320 if (NStores * Size1 >= MagicCompileTimeControl) 321 break; 322 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 323 if (!canSinkStoresAndGEPs(S0, S1)) 324 // Don't attempt to sink below stores that had to stick around 325 // But after removal of a store and some of its feeding 326 // instruction search again from the beginning since the iterator 327 // is likely stale at this point. 328 break; 329 330 if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) { 331 // We have more than 2 predecessors. Insert a new block 332 // postdominating 2 predecessors we're going to sink from. 333 SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split"); 334 if (!SinkBB) 335 break; 336 } 337 338 MergedStores = true; 339 sinkStoresAndGEPs(SinkBB, S0, S1); 340 RBI = Pred0->rbegin(); 341 RBE = Pred0->rend(); 342 LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 343 } 344 } 345 return MergedStores; 346 } 347 348 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) { 349 this->AA = &AA; 350 351 bool Changed = false; 352 LLVM_DEBUG(dbgs() << "Instruction Merger\n"); 353 354 // Merge unconditional branches, allowing PRE to catch more 355 // optimization opportunities. 356 // This loop doesn't care about newly inserted/split blocks 357 // since they never will be diamond heads. 358 for (BasicBlock &BB : make_early_inc_range(F)) 359 // Hoist equivalent loads and sink stores 360 // outside diamonds when possible 361 if (isDiamondHead(&BB)) 362 Changed |= mergeStores(&BB); 363 return Changed; 364 } 365 366 namespace { 367 class MergedLoadStoreMotionLegacyPass : public FunctionPass { 368 const bool SplitFooterBB; 369 public: 370 static char ID; // Pass identification, replacement for typeid 371 MergedLoadStoreMotionLegacyPass(bool SplitFooterBB = false) 372 : FunctionPass(ID), SplitFooterBB(SplitFooterBB) { 373 initializeMergedLoadStoreMotionLegacyPassPass( 374 *PassRegistry::getPassRegistry()); 375 } 376 377 /// 378 /// Run the transformation for each function 379 /// 380 bool runOnFunction(Function &F) override { 381 if (skipFunction(F)) 382 return false; 383 MergedLoadStoreMotion Impl(SplitFooterBB); 384 return Impl.run(F, getAnalysis<AAResultsWrapperPass>().getAAResults()); 385 } 386 387 private: 388 void getAnalysisUsage(AnalysisUsage &AU) const override { 389 if (!SplitFooterBB) 390 AU.setPreservesCFG(); 391 AU.addRequired<AAResultsWrapperPass>(); 392 AU.addPreserved<GlobalsAAWrapperPass>(); 393 } 394 }; 395 396 char MergedLoadStoreMotionLegacyPass::ID = 0; 397 } // anonymous namespace 398 399 /// 400 /// createMergedLoadStoreMotionPass - The public interface to this file. 401 /// 402 FunctionPass *llvm::createMergedLoadStoreMotionPass(bool SplitFooterBB) { 403 return new MergedLoadStoreMotionLegacyPass(SplitFooterBB); 404 } 405 406 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion", 407 "MergedLoadStoreMotion", false, false) 408 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 409 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion", 410 "MergedLoadStoreMotion", false, false) 411 412 PreservedAnalyses 413 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) { 414 MergedLoadStoreMotion Impl(Options.SplitFooterBB); 415 auto &AA = AM.getResult<AAManager>(F); 416 if (!Impl.run(F, AA)) 417 return PreservedAnalyses::all(); 418 419 PreservedAnalyses PA; 420 if (!Options.SplitFooterBB) 421 PA.preserveSet<CFGAnalyses>(); 422 return PA; 423 } 424 425 void MergedLoadStoreMotionPass::printPipeline( 426 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 427 static_cast<PassInfoMixin<MergedLoadStoreMotionPass> *>(this)->printPipeline( 428 OS, MapClassName2PassName); 429 OS << "<"; 430 OS << (Options.SplitFooterBB ? "" : "no-") << "split-footer-bb"; 431 OS << ">"; 432 } 433