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