1 //===- MergedLoadStoreMotion.cpp - merge and hoist/sink load/stores -------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 //! \file 11 //! \brief This pass performs merges of loads and stores on both sides of a 12 // diamond (hammock). It hoists the loads and sinks the stores. 13 // 14 // The algorithm iteratively hoists two loads to the same address out of a 15 // diamond (hammock) and merges them into a single load in the header. Similar 16 // it sinks and merges two stores to the tail block (footer). The algorithm 17 // iterates over the instructions of one side of the diamond and attempts to 18 // find a matching load/store on the other side. It hoists / sinks when it 19 // thinks it safe to do so. This optimization helps with eg. hiding load 20 // latencies, triggering if-conversion, and reducing static code size. 21 // 22 //===----------------------------------------------------------------------===// 23 // 24 // 25 // Example: 26 // Diamond shaped code before merge: 27 // 28 // header: 29 // br %cond, label %if.then, label %if.else 30 // + + 31 // + + 32 // + + 33 // if.then: if.else: 34 // %lt = load %addr_l %le = load %addr_l 35 // <use %lt> <use %le> 36 // <...> <...> 37 // store %st, %addr_s store %se, %addr_s 38 // br label %if.end br label %if.end 39 // + + 40 // + + 41 // + + 42 // if.end ("footer"): 43 // <...> 44 // 45 // Diamond shaped code after merge: 46 // 47 // header: 48 // %l = load %addr_l 49 // br %cond, label %if.then, label %if.else 50 // + + 51 // + + 52 // + + 53 // if.then: if.else: 54 // <use %l> <use %l> 55 // <...> <...> 56 // br label %if.end br label %if.end 57 // + + 58 // + + 59 // + + 60 // if.end ("footer"): 61 // %s.sink = phi [%st, if.then], [%se, if.else] 62 // <...> 63 // store %s.sink, %addr_s 64 // <...> 65 // 66 // 67 //===----------------------- TODO -----------------------------------------===// 68 // 69 // 1) Generalize to regions other than diamonds 70 // 2) Be more aggressive merging memory operations 71 // Note that both changes require register pressure control 72 // 73 //===----------------------------------------------------------------------===// 74 75 #include "llvm/ADT/Statistic.h" 76 #include "llvm/Analysis/AliasAnalysis.h" 77 #include "llvm/Analysis/CFG.h" 78 #include "llvm/Analysis/GlobalsModRef.h" 79 #include "llvm/Analysis/Loads.h" 80 #include "llvm/Analysis/MemoryBuiltins.h" 81 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 82 #include "llvm/Analysis/TargetLibraryInfo.h" 83 #include "llvm/IR/Metadata.h" 84 #include "llvm/IR/PatternMatch.h" 85 #include "llvm/Support/Debug.h" 86 #include "llvm/Support/raw_ostream.h" 87 #include "llvm/Transforms/Scalar.h" 88 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 89 #include "llvm/Transforms/Utils/SSAUpdater.h" 90 91 using namespace llvm; 92 93 #define DEBUG_TYPE "mldst-motion" 94 95 //===----------------------------------------------------------------------===// 96 // MergedLoadStoreMotion Pass 97 //===----------------------------------------------------------------------===// 98 99 namespace { 100 class MergedLoadStoreMotion : public FunctionPass { 101 AliasAnalysis *AA; 102 MemoryDependenceResults *MD; 103 104 public: 105 static char ID; // Pass identification, replacement for typeid 106 MergedLoadStoreMotion() 107 : FunctionPass(ID), MD(nullptr), MagicCompileTimeControl(250) { 108 initializeMergedLoadStoreMotionPass(*PassRegistry::getPassRegistry()); 109 } 110 111 bool runOnFunction(Function &F) override; 112 113 private: 114 // This transformation requires dominator postdominator info 115 void getAnalysisUsage(AnalysisUsage &AU) const override { 116 AU.setPreservesCFG(); 117 AU.addRequired<TargetLibraryInfoWrapperPass>(); 118 AU.addRequired<AAResultsWrapperPass>(); 119 AU.addPreserved<GlobalsAAWrapperPass>(); 120 AU.addPreserved<MemoryDependenceWrapperPass>(); 121 } 122 123 // Helper routines 124 125 /// 126 /// \brief Remove instruction from parent and update memory dependence 127 /// analysis. 128 /// 129 void removeInstruction(Instruction *Inst); 130 BasicBlock *getDiamondTail(BasicBlock *BB); 131 bool isDiamondHead(BasicBlock *BB); 132 // Routines for hoisting loads 133 bool isLoadHoistBarrierInRange(const Instruction& Start, 134 const Instruction& End, 135 LoadInst* LI); 136 LoadInst *canHoistFromBlock(BasicBlock *BB, LoadInst *LI); 137 void hoistInstruction(BasicBlock *BB, Instruction *HoistCand, 138 Instruction *ElseInst); 139 bool isSafeToHoist(Instruction *I) const; 140 bool hoistLoad(BasicBlock *BB, LoadInst *HoistCand, LoadInst *ElseInst); 141 bool mergeLoads(BasicBlock *BB); 142 // Routines for sinking stores 143 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 144 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 145 bool isStoreSinkBarrierInRange(const Instruction &Start, 146 const Instruction &End, MemoryLocation Loc); 147 bool sinkStore(BasicBlock *BB, StoreInst *SinkCand, StoreInst *ElseInst); 148 bool mergeStores(BasicBlock *BB); 149 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity, 150 // where Size0 and Size1 are the #instructions on the two sides of 151 // the diamond. The constant chosen here is arbitrary. Compiler Time 152 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl. 153 const int MagicCompileTimeControl; 154 }; 155 156 char MergedLoadStoreMotion::ID = 0; 157 } // anonymous namespace 158 159 /// 160 /// \brief createMergedLoadStoreMotionPass - The public interface to this file. 161 /// 162 FunctionPass *llvm::createMergedLoadStoreMotionPass() { 163 return new MergedLoadStoreMotion(); 164 } 165 166 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotion, "mldst-motion", 167 "MergedLoadStoreMotion", false, false) 168 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass) 169 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 170 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 171 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) 172 INITIALIZE_PASS_END(MergedLoadStoreMotion, "mldst-motion", 173 "MergedLoadStoreMotion", false, false) 174 175 /// 176 /// \brief Remove instruction from parent and update memory dependence analysis. 177 /// 178 void MergedLoadStoreMotion::removeInstruction(Instruction *Inst) { 179 // Notify the memory dependence analysis. 180 if (MD) { 181 MD->removeInstruction(Inst); 182 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 183 MD->invalidateCachedPointerInfo(LI->getPointerOperand()); 184 if (Inst->getType()->getScalarType()->isPointerTy()) { 185 MD->invalidateCachedPointerInfo(Inst); 186 } 187 } 188 Inst->eraseFromParent(); 189 } 190 191 /// 192 /// \brief Return tail block of a diamond. 193 /// 194 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 195 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 196 BranchInst *BI = (BranchInst *)(BB->getTerminator()); 197 BasicBlock *Succ0 = BI->getSuccessor(0); 198 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0); 199 return Tail; 200 } 201 202 /// 203 /// \brief True when BB is the head of a diamond (hammock) 204 /// 205 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 206 if (!BB) 207 return false; 208 if (!isa<BranchInst>(BB->getTerminator())) 209 return false; 210 if (BB->getTerminator()->getNumSuccessors() != 2) 211 return false; 212 213 BranchInst *BI = (BranchInst *)(BB->getTerminator()); 214 BasicBlock *Succ0 = BI->getSuccessor(0); 215 BasicBlock *Succ1 = BI->getSuccessor(1); 216 217 if (!Succ0->getSinglePredecessor() || 218 Succ0->getTerminator()->getNumSuccessors() != 1) 219 return false; 220 if (!Succ1->getSinglePredecessor() || 221 Succ1->getTerminator()->getNumSuccessors() != 1) 222 return false; 223 224 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0); 225 // Ignore triangles. 226 if (Succ1->getTerminator()->getSuccessor(0) != Tail) 227 return false; 228 return true; 229 } 230 231 /// 232 /// \brief True when instruction is a hoist barrier for a load 233 /// 234 /// Whenever an instruction could possibly modify the value 235 /// being loaded or protect against the load from happening 236 /// it is considered a hoist barrier. 237 /// 238 bool MergedLoadStoreMotion::isLoadHoistBarrierInRange(const Instruction& Start, 239 const Instruction& End, 240 LoadInst* LI) { 241 MemoryLocation Loc = MemoryLocation::get(LI); 242 return AA->canInstructionRangeModRef(Start, End, Loc, MRI_Mod); 243 } 244 245 /// 246 /// \brief Decide if a load can be hoisted 247 /// 248 /// When there is a load in \p BB to the same address as \p LI 249 /// and it can be hoisted from \p BB, return that load. 250 /// Otherwise return Null. 251 /// 252 LoadInst *MergedLoadStoreMotion::canHoistFromBlock(BasicBlock *BB1, 253 LoadInst *Load0) { 254 255 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); BBI != BBE; 256 ++BBI) { 257 Instruction *Inst = &*BBI; 258 259 // Only merge and hoist loads when their result in used only in BB 260 if (!isa<LoadInst>(Inst) || Inst->isUsedOutsideOfBlock(BB1)) 261 continue; 262 263 LoadInst *Load1 = dyn_cast<LoadInst>(Inst); 264 BasicBlock *BB0 = Load0->getParent(); 265 266 MemoryLocation Loc0 = MemoryLocation::get(Load0); 267 MemoryLocation Loc1 = MemoryLocation::get(Load1); 268 if (AA->isMustAlias(Loc0, Loc1) && Load0->isSameOperationAs(Load1) && 269 !isLoadHoistBarrierInRange(BB1->front(), *Load1, Load1) && 270 !isLoadHoistBarrierInRange(BB0->front(), *Load0, Load0)) { 271 return Load1; 272 } 273 } 274 return nullptr; 275 } 276 277 /// 278 /// \brief Merge two equivalent instructions \p HoistCand and \p ElseInst into 279 /// \p BB 280 /// 281 /// BB is the head of a diamond 282 /// 283 void MergedLoadStoreMotion::hoistInstruction(BasicBlock *BB, 284 Instruction *HoistCand, 285 Instruction *ElseInst) { 286 DEBUG(dbgs() << " Hoist Instruction into BB \n"; BB->dump(); 287 dbgs() << "Instruction Left\n"; HoistCand->dump(); dbgs() << "\n"; 288 dbgs() << "Instruction Right\n"; ElseInst->dump(); dbgs() << "\n"); 289 // Hoist the instruction. 290 assert(HoistCand->getParent() != BB); 291 292 // Intersect optional metadata. 293 HoistCand->intersectOptionalDataWith(ElseInst); 294 HoistCand->dropUnknownNonDebugMetadata(); 295 296 // Prepend point for instruction insert 297 Instruction *HoistPt = BB->getTerminator(); 298 299 // Merged instruction 300 Instruction *HoistedInst = HoistCand->clone(); 301 302 // Hoist instruction. 303 HoistedInst->insertBefore(HoistPt); 304 305 HoistCand->replaceAllUsesWith(HoistedInst); 306 removeInstruction(HoistCand); 307 // Replace the else block instruction. 308 ElseInst->replaceAllUsesWith(HoistedInst); 309 removeInstruction(ElseInst); 310 } 311 312 /// 313 /// \brief Return true if no operand of \p I is defined in I's parent block 314 /// 315 bool MergedLoadStoreMotion::isSafeToHoist(Instruction *I) const { 316 BasicBlock *Parent = I->getParent(); 317 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 318 Instruction *Instr = dyn_cast<Instruction>(I->getOperand(i)); 319 if (Instr && Instr->getParent() == Parent) 320 return false; 321 } 322 return true; 323 } 324 325 /// 326 /// \brief Merge two equivalent loads and GEPs and hoist into diamond head 327 /// 328 bool MergedLoadStoreMotion::hoistLoad(BasicBlock *BB, LoadInst *L0, 329 LoadInst *L1) { 330 // Only one definition? 331 Instruction *A0 = dyn_cast<Instruction>(L0->getPointerOperand()); 332 Instruction *A1 = dyn_cast<Instruction>(L1->getPointerOperand()); 333 if (A0 && A1 && A0->isIdenticalTo(A1) && isSafeToHoist(A0) && 334 A0->hasOneUse() && (A0->getParent() == L0->getParent()) && 335 A1->hasOneUse() && (A1->getParent() == L1->getParent()) && 336 isa<GetElementPtrInst>(A0)) { 337 DEBUG(dbgs() << "Hoist Instruction into BB \n"; BB->dump(); 338 dbgs() << "Instruction Left\n"; L0->dump(); dbgs() << "\n"; 339 dbgs() << "Instruction Right\n"; L1->dump(); dbgs() << "\n"); 340 hoistInstruction(BB, A0, A1); 341 hoistInstruction(BB, L0, L1); 342 return true; 343 } else 344 return false; 345 } 346 347 /// 348 /// \brief Try to hoist two loads to same address into diamond header 349 /// 350 /// Starting from a diamond head block, iterate over the instructions in one 351 /// successor block and try to match a load in the second successor. 352 /// 353 bool MergedLoadStoreMotion::mergeLoads(BasicBlock *BB) { 354 bool MergedLoads = false; 355 assert(isDiamondHead(BB)); 356 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 357 BasicBlock *Succ0 = BI->getSuccessor(0); 358 BasicBlock *Succ1 = BI->getSuccessor(1); 359 // #Instructions in Succ1 for Compile Time Control 360 int Size1 = Succ1->size(); 361 int NLoads = 0; 362 for (BasicBlock::iterator BBI = Succ0->begin(), BBE = Succ0->end(); 363 BBI != BBE;) { 364 Instruction *I = &*BBI; 365 ++BBI; 366 367 // Only move non-simple (atomic, volatile) loads. 368 LoadInst *L0 = dyn_cast<LoadInst>(I); 369 if (!L0 || !L0->isSimple() || L0->isUsedOutsideOfBlock(Succ0)) 370 continue; 371 372 ++NLoads; 373 if (NLoads * Size1 >= MagicCompileTimeControl) 374 break; 375 if (LoadInst *L1 = canHoistFromBlock(Succ1, L0)) { 376 bool Res = hoistLoad(BB, L0, L1); 377 MergedLoads |= Res; 378 // Don't attempt to hoist above loads that had not been hoisted. 379 if (!Res) 380 break; 381 } 382 } 383 return MergedLoads; 384 } 385 386 /// 387 /// \brief True when instruction is a sink barrier for a store 388 /// located in Loc 389 /// 390 /// Whenever an instruction could possibly read or modify the 391 /// value being stored or protect against the store from 392 /// happening it is considered a sink barrier. 393 /// 394 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 395 const Instruction &End, 396 MemoryLocation Loc) { 397 return AA->canInstructionRangeModRef(Start, End, Loc, MRI_ModRef); 398 } 399 400 /// 401 /// \brief Check if \p BB contains a store to the same address as \p SI 402 /// 403 /// \return The store in \p when it is safe to sink. Otherwise return Null. 404 /// 405 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 406 StoreInst *Store0) { 407 DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 408 BasicBlock *BB0 = Store0->getParent(); 409 for (BasicBlock::reverse_iterator RBI = BB1->rbegin(), RBE = BB1->rend(); 410 RBI != RBE; ++RBI) { 411 Instruction *Inst = &*RBI; 412 413 if (!isa<StoreInst>(Inst)) 414 continue; 415 416 StoreInst *Store1 = cast<StoreInst>(Inst); 417 418 MemoryLocation Loc0 = MemoryLocation::get(Store0); 419 MemoryLocation Loc1 = MemoryLocation::get(Store1); 420 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) && 421 !isStoreSinkBarrierInRange(*(std::next(BasicBlock::iterator(Store1))), 422 BB1->back(), Loc1) && 423 !isStoreSinkBarrierInRange(*(std::next(BasicBlock::iterator(Store0))), 424 BB0->back(), Loc0)) { 425 return Store1; 426 } 427 } 428 return nullptr; 429 } 430 431 /// 432 /// \brief Create a PHI node in BB for the operands of S0 and S1 433 /// 434 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 435 StoreInst *S1) { 436 // Create a phi if the values mismatch. 437 PHINode *NewPN = nullptr; 438 Value *Opd1 = S0->getValueOperand(); 439 Value *Opd2 = S1->getValueOperand(); 440 if (Opd1 != Opd2) { 441 NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink", 442 &BB->front()); 443 NewPN->addIncoming(Opd1, S0->getParent()); 444 NewPN->addIncoming(Opd2, S1->getParent()); 445 if (MD && NewPN->getType()->getScalarType()->isPointerTy()) 446 MD->invalidateCachedPointerInfo(NewPN); 447 } 448 return NewPN; 449 } 450 451 /// 452 /// \brief Merge two stores to same address and sink into \p BB 453 /// 454 /// Also sinks GEP instruction computing the store address 455 /// 456 bool MergedLoadStoreMotion::sinkStore(BasicBlock *BB, StoreInst *S0, 457 StoreInst *S1) { 458 // Only one definition? 459 Instruction *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 460 Instruction *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 461 if (A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() && 462 (A0->getParent() == S0->getParent()) && A1->hasOneUse() && 463 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0)) { 464 DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 465 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 466 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 467 // Hoist the instruction. 468 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 469 // Intersect optional metadata. 470 S0->intersectOptionalDataWith(S1); 471 S0->dropUnknownNonDebugMetadata(); 472 473 // Create the new store to be inserted at the join point. 474 StoreInst *SNew = (StoreInst *)(S0->clone()); 475 Instruction *ANew = A0->clone(); 476 SNew->insertBefore(&*InsertPt); 477 ANew->insertBefore(SNew); 478 479 assert(S0->getParent() == A0->getParent()); 480 assert(S1->getParent() == A1->getParent()); 481 482 PHINode *NewPN = getPHIOperand(BB, S0, S1); 483 // New PHI operand? Use it. 484 if (NewPN) 485 SNew->setOperand(0, NewPN); 486 removeInstruction(S0); 487 removeInstruction(S1); 488 A0->replaceAllUsesWith(ANew); 489 removeInstruction(A0); 490 A1->replaceAllUsesWith(ANew); 491 removeInstruction(A1); 492 return true; 493 } 494 return false; 495 } 496 497 /// 498 /// \brief True when two stores are equivalent and can sink into the footer 499 /// 500 /// Starting from a diamond tail block, iterate over the instructions in one 501 /// predecessor block and try to match a store in the second predecessor. 502 /// 503 bool MergedLoadStoreMotion::mergeStores(BasicBlock *T) { 504 505 bool MergedStores = false; 506 assert(T && "Footer of a diamond cannot be empty"); 507 508 pred_iterator PI = pred_begin(T), E = pred_end(T); 509 assert(PI != E); 510 BasicBlock *Pred0 = *PI; 511 ++PI; 512 BasicBlock *Pred1 = *PI; 513 ++PI; 514 // tail block of a diamond/hammock? 515 if (Pred0 == Pred1) 516 return false; // No. 517 if (PI != E) 518 return false; // No. More than 2 predecessors. 519 520 // #Instructions in Succ1 for Compile Time Control 521 int Size1 = Pred1->size(); 522 int NStores = 0; 523 524 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 525 RBI != RBE;) { 526 527 Instruction *I = &*RBI; 528 ++RBI; 529 530 // Sink move non-simple (atomic, volatile) stores 531 if (!isa<StoreInst>(I)) 532 continue; 533 StoreInst *S0 = (StoreInst *)I; 534 if (!S0->isSimple()) 535 continue; 536 537 ++NStores; 538 if (NStores * Size1 >= MagicCompileTimeControl) 539 break; 540 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 541 bool Res = sinkStore(T, S0, S1); 542 MergedStores |= Res; 543 // Don't attempt to sink below stores that had to stick around 544 // But after removal of a store and some of its feeding 545 // instruction search again from the beginning since the iterator 546 // is likely stale at this point. 547 if (!Res) 548 break; 549 else { 550 RBI = Pred0->rbegin(); 551 RBE = Pred0->rend(); 552 DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 553 } 554 } 555 } 556 return MergedStores; 557 } 558 559 /// 560 /// \brief Run the transformation for each function 561 /// 562 bool MergedLoadStoreMotion::runOnFunction(Function &F) { 563 if (skipFunction(F)) 564 return false; 565 566 auto *MDWP = getAnalysisIfAvailable<MemoryDependenceWrapperPass>(); 567 MD = MDWP ? &MDWP->getMemDep() : nullptr; 568 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 569 570 bool Changed = false; 571 DEBUG(dbgs() << "Instruction Merger\n"); 572 573 // Merge unconditional branches, allowing PRE to catch more 574 // optimization opportunities. 575 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) { 576 BasicBlock *BB = &*FI++; 577 578 // Hoist equivalent loads and sink stores 579 // outside diamonds when possible 580 if (isDiamondHead(BB)) { 581 Changed |= mergeLoads(BB); 582 Changed |= mergeStores(getDiamondTail(BB)); 583 } 584 } 585 return Changed; 586 } 587