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 96using namespace llvm; 97 98#define DEBUG_TYPE "mldst-motion" 99 100//===----------------------------------------------------------------------===// 101// MergedLoadStoreMotion Pass 102//===----------------------------------------------------------------------===// 103 104namespace { 105class MergedLoadStoreMotion : public FunctionPass { 106 AliasAnalysis *AA; 107 MemoryDependenceAnalysis *MD; 108 109public: 110 static char ID; // Pass identification, replacement for typeid 111 MergedLoadStoreMotion() 112 : FunctionPass(ID), MD(nullptr), MagicCompileTimeControl(250) { 113 initializeMergedLoadStoreMotionPass(*PassRegistry::getPassRegistry()); 114 } 115 116 bool runOnFunction(Function &F) override; 117 118private: 119 // This transformation requires dominator postdominator info 120 void getAnalysisUsage(AnalysisUsage &AU) const override { 121 AU.setPreservesCFG(); 122 AU.addRequired<TargetLibraryInfoWrapperPass>(); 123 AU.addRequired<AAResultsWrapperPass>(); 124 AU.addPreserved<GlobalsAAWrapperPass>(); 125 AU.addPreserved<MemoryDependenceAnalysis>(); 126 } 127 128 // Helper routines 129 130 /// 131 /// \brief Remove instruction from parent and update memory dependence 132 /// analysis. 133 /// 134 void removeInstruction(Instruction *Inst); 135 BasicBlock *getDiamondTail(BasicBlock *BB); 136 bool isDiamondHead(BasicBlock *BB); 137 // Routines for hoisting loads 138 bool isLoadHoistBarrierInRange(const Instruction& Start, 139 const Instruction& End, 140 LoadInst* LI); 141 LoadInst *canHoistFromBlock(BasicBlock *BB, LoadInst *LI); 142 void hoistInstruction(BasicBlock *BB, Instruction *HoistCand, 143 Instruction *ElseInst); 144 bool isSafeToHoist(Instruction *I) const; 145 bool hoistLoad(BasicBlock *BB, LoadInst *HoistCand, LoadInst *ElseInst); 146 bool mergeLoads(BasicBlock *BB); 147 // Routines for sinking stores 148 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 149 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 150 bool isStoreSinkBarrierInRange(const Instruction &Start, 151 const Instruction &End, MemoryLocation Loc); 152 bool sinkStore(BasicBlock *BB, StoreInst *SinkCand, StoreInst *ElseInst); 153 bool mergeStores(BasicBlock *BB); 154 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity, 155 // where Size0 and Size1 are the #instructions on the two sides of 156 // the diamond. The constant chosen here is arbitrary. Compiler Time 157 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl. 158 const int MagicCompileTimeControl; 159}; 160 161char MergedLoadStoreMotion::ID = 0; 162} // anonymous namespace 163 164/// 165/// \brief createMergedLoadStoreMotionPass - The public interface to this file. 166/// 167FunctionPass *llvm::createMergedLoadStoreMotionPass() { 168 return new MergedLoadStoreMotion(); 169} 170 171INITIALIZE_PASS_BEGIN(MergedLoadStoreMotion, "mldst-motion", 172 "MergedLoadStoreMotion", false, false) 173INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 174INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 175INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 176INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) 177INITIALIZE_PASS_END(MergedLoadStoreMotion, "mldst-motion", 178 "MergedLoadStoreMotion", false, false) 179 180/// 181/// \brief Remove instruction from parent and update memory dependence analysis. 182/// 183void MergedLoadStoreMotion::removeInstruction(Instruction *Inst) { 184 // Notify the memory dependence analysis. 185 if (MD) { 186 MD->removeInstruction(Inst); 187 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 188 MD->invalidateCachedPointerInfo(LI->getPointerOperand()); 189 if (Inst->getType()->getScalarType()->isPointerTy()) { 190 MD->invalidateCachedPointerInfo(Inst); 191 } 192 } 193 Inst->eraseFromParent(); 194} 195 196/// 197/// \brief Return tail block of a diamond. 198/// 199BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 200 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 201 BranchInst *BI = (BranchInst *)(BB->getTerminator()); 202 BasicBlock *Succ0 = BI->getSuccessor(0); 203 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0); 204 return Tail; 205} 206 207/// 208/// \brief True when BB is the head of a diamond (hammock) 209/// 210bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 211 if (!BB) 212 return false; 213 if (!isa<BranchInst>(BB->getTerminator())) 214 return false; 215 if (BB->getTerminator()->getNumSuccessors() != 2) 216 return false; 217 218 BranchInst *BI = (BranchInst *)(BB->getTerminator()); 219 BasicBlock *Succ0 = BI->getSuccessor(0); 220 BasicBlock *Succ1 = BI->getSuccessor(1); 221 222 if (!Succ0->getSinglePredecessor() || 223 Succ0->getTerminator()->getNumSuccessors() != 1) 224 return false; 225 if (!Succ1->getSinglePredecessor() || 226 Succ1->getTerminator()->getNumSuccessors() != 1) 227 return false; 228 229 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0); 230 // Ignore triangles. 231 if (Succ1->getTerminator()->getSuccessor(0) != Tail) 232 return false; 233 return true; 234} 235 236/// 237/// \brief True when instruction is a hoist barrier for a load 238/// 239/// Whenever an instruction could possibly modify the value 240/// being loaded or protect against the load from happening 241/// it is considered a hoist barrier. 242/// 243bool MergedLoadStoreMotion::isLoadHoistBarrierInRange(const Instruction& Start, 244 const Instruction& End, 245 LoadInst* LI) { 246 MemoryLocation Loc = MemoryLocation::get(LI); 247 return AA->canInstructionRangeModRef(Start, End, Loc, MRI_Mod); 248} 249 250/// 251/// \brief Decide if a load can be hoisted 252/// 253/// When there is a load in \p BB to the same address as \p LI 254/// and it can be hoisted from \p BB, return that load. 255/// Otherwise return Null. 256/// 257LoadInst *MergedLoadStoreMotion::canHoistFromBlock(BasicBlock *BB1, 258 LoadInst *Load0) { 259 260 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); BBI != BBE; 261 ++BBI) { 262 Instruction *Inst = &*BBI; 263 264 // Only merge and hoist loads when their result in used only in BB 265 if (!isa<LoadInst>(Inst) || Inst->isUsedOutsideOfBlock(BB1)) 266 continue; 267 268 LoadInst *Load1 = dyn_cast<LoadInst>(Inst); 269 BasicBlock *BB0 = Load0->getParent(); 270 271 MemoryLocation Loc0 = MemoryLocation::get(Load0); 272 MemoryLocation Loc1 = MemoryLocation::get(Load1); 273 if (AA->isMustAlias(Loc0, Loc1) && Load0->isSameOperationAs(Load1) && 274 !isLoadHoistBarrierInRange(BB1->front(), *Load1, Load1) && 275 !isLoadHoistBarrierInRange(BB0->front(), *Load0, Load0)) { 276 return Load1; 277 } 278 } 279 return nullptr; 280} 281 282/// 283/// \brief Merge two equivalent instructions \p HoistCand and \p ElseInst into 284/// \p BB 285/// 286/// BB is the head of a diamond 287/// 288void MergedLoadStoreMotion::hoistInstruction(BasicBlock *BB, 289 Instruction *HoistCand, 290 Instruction *ElseInst) { 291 DEBUG(dbgs() << " Hoist Instruction into BB \n"; BB->dump(); 292 dbgs() << "Instruction Left\n"; HoistCand->dump(); dbgs() << "\n"; 293 dbgs() << "Instruction Right\n"; ElseInst->dump(); dbgs() << "\n"); 294 // Hoist the instruction. 295 assert(HoistCand->getParent() != BB); 296 297 // Intersect optional metadata. 298 HoistCand->intersectOptionalDataWith(ElseInst); 299 HoistCand->dropUnknownNonDebugMetadata(); 300 301 // Prepend point for instruction insert 302 Instruction *HoistPt = BB->getTerminator(); 303 304 // Merged instruction 305 Instruction *HoistedInst = HoistCand->clone(); 306 307 // Hoist instruction. 308 HoistedInst->insertBefore(HoistPt); 309 310 HoistCand->replaceAllUsesWith(HoistedInst); 311 removeInstruction(HoistCand); 312 // Replace the else block instruction. 313 ElseInst->replaceAllUsesWith(HoistedInst); 314 removeInstruction(ElseInst); 315} 316 317/// 318/// \brief Return true if no operand of \p I is defined in I's parent block 319/// 320bool MergedLoadStoreMotion::isSafeToHoist(Instruction *I) const { 321 BasicBlock *Parent = I->getParent(); 322 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 323 Instruction *Instr = dyn_cast<Instruction>(I->getOperand(i)); 324 if (Instr && Instr->getParent() == Parent) 325 return false; 326 } 327 return true; 328} 329 330/// 331/// \brief Merge two equivalent loads and GEPs and hoist into diamond head 332/// 333bool MergedLoadStoreMotion::hoistLoad(BasicBlock *BB, LoadInst *L0, 334 LoadInst *L1) { 335 // Only one definition? 336 Instruction *A0 = dyn_cast<Instruction>(L0->getPointerOperand()); 337 Instruction *A1 = dyn_cast<Instruction>(L1->getPointerOperand()); 338 if (A0 && A1 && A0->isIdenticalTo(A1) && isSafeToHoist(A0) && 339 A0->hasOneUse() && (A0->getParent() == L0->getParent()) && 340 A1->hasOneUse() && (A1->getParent() == L1->getParent()) && 341 isa<GetElementPtrInst>(A0)) { 342 DEBUG(dbgs() << "Hoist Instruction into BB \n"; BB->dump(); 343 dbgs() << "Instruction Left\n"; L0->dump(); dbgs() << "\n"; 344 dbgs() << "Instruction Right\n"; L1->dump(); dbgs() << "\n"); 345 hoistInstruction(BB, A0, A1); 346 hoistInstruction(BB, L0, L1); 347 return true; 348 } else 349 return false; 350} 351 352/// 353/// \brief Try to hoist two loads to same address into diamond header 354/// 355/// Starting from a diamond head block, iterate over the instructions in one 356/// successor block and try to match a load in the second successor. 357/// 358bool MergedLoadStoreMotion::mergeLoads(BasicBlock *BB) { 359 bool MergedLoads = false; 360 assert(isDiamondHead(BB)); 361 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 362 BasicBlock *Succ0 = BI->getSuccessor(0); 363 BasicBlock *Succ1 = BI->getSuccessor(1); 364 // #Instructions in Succ1 for Compile Time Control 365 int Size1 = Succ1->size(); 366 int NLoads = 0; 367 for (BasicBlock::iterator BBI = Succ0->begin(), BBE = Succ0->end(); 368 BBI != BBE;) { 369 Instruction *I = &*BBI; 370 ++BBI; 371 372 // Only move non-simple (atomic, volatile) loads. 373 LoadInst *L0 = dyn_cast<LoadInst>(I); 374 if (!L0 || !L0->isSimple() || L0->isUsedOutsideOfBlock(Succ0)) 375 continue; 376 377 ++NLoads; 378 if (NLoads * Size1 >= MagicCompileTimeControl) 379 break; 380 if (LoadInst *L1 = canHoistFromBlock(Succ1, L0)) { 381 bool Res = hoistLoad(BB, L0, L1); 382 MergedLoads |= Res; 383 // Don't attempt to hoist above loads that had not been hoisted. 384 if (!Res) 385 break; 386 } 387 } 388 return MergedLoads; 389} 390 391/// 392/// \brief True when instruction is a sink barrier for a store 393/// located in Loc 394/// 395/// Whenever an instruction could possibly read or modify the 396/// value being stored or protect against the store from 397/// happening it is considered a sink barrier. 398/// 399bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 400 const Instruction &End, 401 MemoryLocation Loc) { 402 return AA->canInstructionRangeModRef(Start, End, Loc, MRI_ModRef); 403} 404 405/// 406/// \brief Check if \p BB contains a store to the same address as \p SI 407/// 408/// \return The store in \p when it is safe to sink. Otherwise return Null. 409/// 410StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 411 StoreInst *Store0) { 412 DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 413 BasicBlock *BB0 = Store0->getParent(); 414 for (BasicBlock::reverse_iterator RBI = BB1->rbegin(), RBE = BB1->rend(); 415 RBI != RBE; ++RBI) { 416 Instruction *Inst = &*RBI; 417 418 if (!isa<StoreInst>(Inst)) 419 continue; 420 421 StoreInst *Store1 = cast<StoreInst>(Inst); 422 423 MemoryLocation Loc0 = MemoryLocation::get(Store0); 424 MemoryLocation Loc1 = MemoryLocation::get(Store1); 425 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) && 426 !isStoreSinkBarrierInRange(*(std::next(BasicBlock::iterator(Store1))), 427 BB1->back(), Loc1) && 428 !isStoreSinkBarrierInRange(*(std::next(BasicBlock::iterator(Store0))), 429 BB0->back(), Loc0)) { 430 return Store1; 431 } 432 } 433 return nullptr; 434} 435 436/// 437/// \brief Create a PHI node in BB for the operands of S0 and S1 438/// 439PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 440 StoreInst *S1) { 441 // Create a phi if the values mismatch. 442 PHINode *NewPN = nullptr; 443 Value *Opd1 = S0->getValueOperand(); 444 Value *Opd2 = S1->getValueOperand(); 445 if (Opd1 != Opd2) { 446 NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink", 447 &BB->front()); 448 NewPN->addIncoming(Opd1, S0->getParent()); 449 NewPN->addIncoming(Opd2, S1->getParent()); 450 if (MD && NewPN->getType()->getScalarType()->isPointerTy()) 451 MD->invalidateCachedPointerInfo(NewPN); 452 } 453 return NewPN; 454} 455 456/// 457/// \brief Merge two stores to same address and sink into \p BB 458/// 459/// Also sinks GEP instruction computing the store address 460/// 461bool MergedLoadStoreMotion::sinkStore(BasicBlock *BB, StoreInst *S0, 462 StoreInst *S1) { 463 // Only one definition? 464 Instruction *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 465 Instruction *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 466 if (A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() && 467 (A0->getParent() == S0->getParent()) && A1->hasOneUse() && 468 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0)) { 469 DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 470 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 471 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 472 // Hoist the instruction. 473 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 474 // Intersect optional metadata. 475 S0->intersectOptionalDataWith(S1); 476 S0->dropUnknownNonDebugMetadata(); 477 478 // Create the new store to be inserted at the join point. 479 StoreInst *SNew = (StoreInst *)(S0->clone()); 480 Instruction *ANew = A0->clone(); 481 SNew->insertBefore(&*InsertPt); 482 ANew->insertBefore(SNew); 483 484 assert(S0->getParent() == A0->getParent()); 485 assert(S1->getParent() == A1->getParent()); 486 487 PHINode *NewPN = getPHIOperand(BB, S0, S1); 488 // New PHI operand? Use it. 489 if (NewPN) 490 SNew->setOperand(0, NewPN); 491 removeInstruction(S0); 492 removeInstruction(S1); 493 A0->replaceAllUsesWith(ANew); 494 removeInstruction(A0); 495 A1->replaceAllUsesWith(ANew); 496 removeInstruction(A1); 497 return true; 498 } 499 return false; 500} 501 502/// 503/// \brief True when two stores are equivalent and can sink into the footer 504/// 505/// Starting from a diamond tail block, iterate over the instructions in one 506/// predecessor block and try to match a store in the second predecessor. 507/// 508bool MergedLoadStoreMotion::mergeStores(BasicBlock *T) { 509 510 bool MergedStores = false; 511 assert(T && "Footer of a diamond cannot be empty"); 512 513 pred_iterator PI = pred_begin(T), E = pred_end(T); 514 assert(PI != E); 515 BasicBlock *Pred0 = *PI; 516 ++PI; 517 BasicBlock *Pred1 = *PI; 518 ++PI; 519 // tail block of a diamond/hammock? 520 if (Pred0 == Pred1) 521 return false; // No. 522 if (PI != E) 523 return false; // No. More than 2 predecessors. 524 525 // #Instructions in Succ1 for Compile Time Control 526 int Size1 = Pred1->size(); 527 int NStores = 0; 528 529 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 530 RBI != RBE;) { 531 532 Instruction *I = &*RBI; 533 ++RBI; 534 535 // Sink move non-simple (atomic, volatile) stores 536 if (!isa<StoreInst>(I)) 537 continue; 538 StoreInst *S0 = (StoreInst *)I; 539 if (!S0->isSimple()) 540 continue; 541 542 ++NStores; 543 if (NStores * Size1 >= MagicCompileTimeControl) 544 break; 545 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 546 bool Res = sinkStore(T, S0, S1); 547 MergedStores |= Res; 548 // Don't attempt to sink below stores that had to stick around 549 // But after removal of a store and some of its feeding 550 // instruction search again from the beginning since the iterator 551 // is likely stale at this point. 552 if (!Res) 553 break; 554 else { 555 RBI = Pred0->rbegin(); 556 RBE = Pred0->rend(); 557 DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 558 } 559 } 560 } 561 return MergedStores; 562} 563 564/// 565/// \brief Run the transformation for each function 566/// 567bool MergedLoadStoreMotion::runOnFunction(Function &F) { 568 MD = getAnalysisIfAvailable<MemoryDependenceAnalysis>(); 569 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 570 571 bool Changed = false; 572 DEBUG(dbgs() << "Instruction Merger\n"); 573 574 // Merge unconditional branches, allowing PRE to catch more 575 // optimization opportunities. 576 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) { 577 BasicBlock *BB = &*FI++; 578 579 // Hoist equivalent loads and sink stores 580 // outside diamonds when possible 581 if (isDiamondHead(BB)) { 582 Changed |= mergeLoads(BB); 583 Changed |= mergeStores(getDiamondTail(BB)); 584 } 585 } 586 return Changed; 587} 588