1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===// 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 // This file implements a basic-block vectorization pass. The algorithm was 11 // inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral, 12 // et al. It works by looking for chains of pairable operations and then 13 // pairing them. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #define BBV_NAME "bb-vectorize" 18 #include "llvm/Transforms/Vectorize.h" 19 #include "llvm/ADT/DenseMap.h" 20 #include "llvm/ADT/DenseSet.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/SmallSet.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/ADT/Statistic.h" 25 #include "llvm/ADT/StringExtras.h" 26 #include "llvm/Analysis/AliasAnalysis.h" 27 #include "llvm/Analysis/AliasSetTracker.h" 28 #include "llvm/Analysis/ScalarEvolution.h" 29 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 30 #include "llvm/Analysis/TargetTransformInfo.h" 31 #include "llvm/Analysis/ValueTracking.h" 32 #include "llvm/IR/Constants.h" 33 #include "llvm/IR/DataLayout.h" 34 #include "llvm/IR/DerivedTypes.h" 35 #include "llvm/IR/Dominators.h" 36 #include "llvm/IR/Function.h" 37 #include "llvm/IR/Instructions.h" 38 #include "llvm/IR/IntrinsicInst.h" 39 #include "llvm/IR/Intrinsics.h" 40 #include "llvm/IR/LLVMContext.h" 41 #include "llvm/IR/Metadata.h" 42 #include "llvm/IR/Module.h" 43 #include "llvm/IR/Type.h" 44 #include "llvm/IR/ValueHandle.h" 45 #include "llvm/Pass.h" 46 #include "llvm/Support/CommandLine.h" 47 #include "llvm/Support/Debug.h" 48 #include "llvm/Support/raw_ostream.h" 49 #include "llvm/Transforms/Utils/Local.h" 50 #include <algorithm> 51 using namespace llvm; 52 53 #define DEBUG_TYPE BBV_NAME 54 55 static cl::opt<bool> 56 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false), 57 cl::Hidden, cl::desc("Ignore target information")); 58 59 static cl::opt<unsigned> 60 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 61 cl::desc("The required chain depth for vectorization")); 62 63 static cl::opt<bool> 64 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false), 65 cl::Hidden, cl::desc("Use the chain depth requirement with" 66 " target information")); 67 68 static cl::opt<unsigned> 69 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 70 cl::desc("The maximum search distance for instruction pairs")); 71 72 static cl::opt<bool> 73 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 74 cl::desc("Replicating one element to a pair breaks the chain")); 75 76 static cl::opt<unsigned> 77 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 78 cl::desc("The size of the native vector registers")); 79 80 static cl::opt<unsigned> 81 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 82 cl::desc("The maximum number of pairing iterations")); 83 84 static cl::opt<bool> 85 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden, 86 cl::desc("Don't try to form non-2^n-length vectors")); 87 88 static cl::opt<unsigned> 89 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 90 cl::desc("The maximum number of pairable instructions per group")); 91 92 static cl::opt<unsigned> 93 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden, 94 cl::desc("The maximum number of candidate instruction pairs per group")); 95 96 static cl::opt<unsigned> 97 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 98 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 99 " a full cycle check")); 100 101 static cl::opt<bool> 102 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden, 103 cl::desc("Don't try to vectorize boolean (i1) values")); 104 105 static cl::opt<bool> 106 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 107 cl::desc("Don't try to vectorize integer values")); 108 109 static cl::opt<bool> 110 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 111 cl::desc("Don't try to vectorize floating-point values")); 112 113 // FIXME: This should default to false once pointer vector support works. 114 static cl::opt<bool> 115 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden, 116 cl::desc("Don't try to vectorize pointer values")); 117 118 static cl::opt<bool> 119 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 120 cl::desc("Don't try to vectorize casting (conversion) operations")); 121 122 static cl::opt<bool> 123 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 124 cl::desc("Don't try to vectorize floating-point math intrinsics")); 125 126 static cl::opt<bool> 127 NoBitManipulation("bb-vectorize-no-bitmanip", cl::init(false), cl::Hidden, 128 cl::desc("Don't try to vectorize BitManipulation intrinsics")); 129 130 static cl::opt<bool> 131 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 132 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 133 134 static cl::opt<bool> 135 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 136 cl::desc("Don't try to vectorize select instructions")); 137 138 static cl::opt<bool> 139 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 140 cl::desc("Don't try to vectorize comparison instructions")); 141 142 static cl::opt<bool> 143 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 144 cl::desc("Don't try to vectorize getelementptr instructions")); 145 146 static cl::opt<bool> 147 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 148 cl::desc("Don't try to vectorize loads and stores")); 149 150 static cl::opt<bool> 151 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 152 cl::desc("Only generate aligned loads and stores")); 153 154 static cl::opt<bool> 155 NoMemOpBoost("bb-vectorize-no-mem-op-boost", 156 cl::init(false), cl::Hidden, 157 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 158 159 static cl::opt<bool> 160 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 161 cl::desc("Use a fast instruction dependency analysis")); 162 163 #ifndef NDEBUG 164 static cl::opt<bool> 165 DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 166 cl::init(false), cl::Hidden, 167 cl::desc("When debugging is enabled, output information on the" 168 " instruction-examination process")); 169 static cl::opt<bool> 170 DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 171 cl::init(false), cl::Hidden, 172 cl::desc("When debugging is enabled, output information on the" 173 " candidate-selection process")); 174 static cl::opt<bool> 175 DebugPairSelection("bb-vectorize-debug-pair-selection", 176 cl::init(false), cl::Hidden, 177 cl::desc("When debugging is enabled, output information on the" 178 " pair-selection process")); 179 static cl::opt<bool> 180 DebugCycleCheck("bb-vectorize-debug-cycle-check", 181 cl::init(false), cl::Hidden, 182 cl::desc("When debugging is enabled, output information on the" 183 " cycle-checking process")); 184 185 static cl::opt<bool> 186 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair", 187 cl::init(false), cl::Hidden, 188 cl::desc("When debugging is enabled, dump the basic block after" 189 " every pair is fused")); 190 #endif 191 192 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 193 194 namespace { 195 struct BBVectorize : public BasicBlockPass { 196 static char ID; // Pass identification, replacement for typeid 197 198 const VectorizeConfig Config; 199 200 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 201 : BasicBlockPass(ID), Config(C) { 202 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 203 } 204 205 BBVectorize(Pass *P, Function &F, const VectorizeConfig &C) 206 : BasicBlockPass(ID), Config(C) { 207 AA = &P->getAnalysis<AliasAnalysis>(); 208 DT = &P->getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 209 SE = &P->getAnalysis<ScalarEvolution>(); 210 TTI = IgnoreTargetInfo 211 ? nullptr 212 : &P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 213 } 214 215 typedef std::pair<Value *, Value *> ValuePair; 216 typedef std::pair<ValuePair, int> ValuePairWithCost; 217 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 218 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 219 typedef std::pair<VPPair, unsigned> VPPairWithType; 220 221 AliasAnalysis *AA; 222 DominatorTree *DT; 223 ScalarEvolution *SE; 224 const TargetTransformInfo *TTI; 225 226 // FIXME: const correct? 227 228 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 229 230 bool getCandidatePairs(BasicBlock &BB, 231 BasicBlock::iterator &Start, 232 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 233 DenseSet<ValuePair> &FixedOrderPairs, 234 DenseMap<ValuePair, int> &CandidatePairCostSavings, 235 std::vector<Value *> &PairableInsts, bool NonPow2Len); 236 237 // FIXME: The current implementation does not account for pairs that 238 // are connected in multiple ways. For example: 239 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap) 240 enum PairConnectionType { 241 PairConnectionDirect, 242 PairConnectionSwap, 243 PairConnectionSplat 244 }; 245 246 void computeConnectedPairs( 247 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 248 DenseSet<ValuePair> &CandidatePairsSet, 249 std::vector<Value *> &PairableInsts, 250 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 251 DenseMap<VPPair, unsigned> &PairConnectionTypes); 252 253 void buildDepMap(BasicBlock &BB, 254 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 255 std::vector<Value *> &PairableInsts, 256 DenseSet<ValuePair> &PairableInstUsers); 257 258 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 259 DenseSet<ValuePair> &CandidatePairsSet, 260 DenseMap<ValuePair, int> &CandidatePairCostSavings, 261 std::vector<Value *> &PairableInsts, 262 DenseSet<ValuePair> &FixedOrderPairs, 263 DenseMap<VPPair, unsigned> &PairConnectionTypes, 264 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 265 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 266 DenseSet<ValuePair> &PairableInstUsers, 267 DenseMap<Value *, Value *>& ChosenPairs); 268 269 void fuseChosenPairs(BasicBlock &BB, 270 std::vector<Value *> &PairableInsts, 271 DenseMap<Value *, Value *>& ChosenPairs, 272 DenseSet<ValuePair> &FixedOrderPairs, 273 DenseMap<VPPair, unsigned> &PairConnectionTypes, 274 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 275 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps); 276 277 278 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 279 280 bool areInstsCompatible(Instruction *I, Instruction *J, 281 bool IsSimpleLoadStore, bool NonPow2Len, 282 int &CostSavings, int &FixedOrder); 283 284 bool trackUsesOfI(DenseSet<Value *> &Users, 285 AliasSetTracker &WriteSet, Instruction *I, 286 Instruction *J, bool UpdateUsers = true, 287 DenseSet<ValuePair> *LoadMoveSetPairs = nullptr); 288 289 void computePairsConnectedTo( 290 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 291 DenseSet<ValuePair> &CandidatePairsSet, 292 std::vector<Value *> &PairableInsts, 293 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 294 DenseMap<VPPair, unsigned> &PairConnectionTypes, 295 ValuePair P); 296 297 bool pairsConflict(ValuePair P, ValuePair Q, 298 DenseSet<ValuePair> &PairableInstUsers, 299 DenseMap<ValuePair, std::vector<ValuePair> > 300 *PairableInstUserMap = nullptr, 301 DenseSet<VPPair> *PairableInstUserPairSet = nullptr); 302 303 bool pairWillFormCycle(ValuePair P, 304 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers, 305 DenseSet<ValuePair> &CurrentPairs); 306 307 void pruneDAGFor( 308 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 309 std::vector<Value *> &PairableInsts, 310 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 311 DenseSet<ValuePair> &PairableInstUsers, 312 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 313 DenseSet<VPPair> &PairableInstUserPairSet, 314 DenseMap<Value *, Value *> &ChosenPairs, 315 DenseMap<ValuePair, size_t> &DAG, 316 DenseSet<ValuePair> &PrunedDAG, ValuePair J, 317 bool UseCycleCheck); 318 319 void buildInitialDAGFor( 320 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 321 DenseSet<ValuePair> &CandidatePairsSet, 322 std::vector<Value *> &PairableInsts, 323 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 324 DenseSet<ValuePair> &PairableInstUsers, 325 DenseMap<Value *, Value *> &ChosenPairs, 326 DenseMap<ValuePair, size_t> &DAG, ValuePair J); 327 328 void findBestDAGFor( 329 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 330 DenseSet<ValuePair> &CandidatePairsSet, 331 DenseMap<ValuePair, int> &CandidatePairCostSavings, 332 std::vector<Value *> &PairableInsts, 333 DenseSet<ValuePair> &FixedOrderPairs, 334 DenseMap<VPPair, unsigned> &PairConnectionTypes, 335 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 336 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 337 DenseSet<ValuePair> &PairableInstUsers, 338 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 339 DenseSet<VPPair> &PairableInstUserPairSet, 340 DenseMap<Value *, Value *> &ChosenPairs, 341 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth, 342 int &BestEffSize, Value *II, std::vector<Value *>&JJ, 343 bool UseCycleCheck); 344 345 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 346 Instruction *J, unsigned o); 347 348 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 349 unsigned MaskOffset, unsigned NumInElem, 350 unsigned NumInElem1, unsigned IdxOffset, 351 std::vector<Constant*> &Mask); 352 353 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 354 Instruction *J); 355 356 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 357 unsigned o, Value *&LOp, unsigned numElemL, 358 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, 359 unsigned IdxOff = 0); 360 361 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 362 Instruction *J, unsigned o, bool IBeforeJ); 363 364 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 365 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands, 366 bool IBeforeJ); 367 368 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 369 Instruction *J, Instruction *K, 370 Instruction *&InsertionPt, Instruction *&K1, 371 Instruction *&K2); 372 373 void collectPairLoadMoveSet(BasicBlock &BB, 374 DenseMap<Value *, Value *> &ChosenPairs, 375 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 376 DenseSet<ValuePair> &LoadMoveSetPairs, 377 Instruction *I); 378 379 void collectLoadMoveSet(BasicBlock &BB, 380 std::vector<Value *> &PairableInsts, 381 DenseMap<Value *, Value *> &ChosenPairs, 382 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 383 DenseSet<ValuePair> &LoadMoveSetPairs); 384 385 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 386 DenseSet<ValuePair> &LoadMoveSetPairs, 387 Instruction *I, Instruction *J); 388 389 void moveUsesOfIAfterJ(BasicBlock &BB, 390 DenseSet<ValuePair> &LoadMoveSetPairs, 391 Instruction *&InsertionPt, 392 Instruction *I, Instruction *J); 393 394 bool vectorizeBB(BasicBlock &BB) { 395 if (skipOptnoneFunction(BB)) 396 return false; 397 if (!DT->isReachableFromEntry(&BB)) { 398 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 399 " in " << BB.getParent()->getName() << "\n"); 400 return false; 401 } 402 403 DEBUG(if (TTI) dbgs() << "BBV: using target information\n"); 404 405 bool changed = false; 406 // Iterate a sufficient number of times to merge types of size 1 bit, 407 // then 2 bits, then 4, etc. up to half of the target vector width of the 408 // target vector register. 409 unsigned n = 1; 410 for (unsigned v = 2; 411 (TTI || v <= Config.VectorBits) && 412 (!Config.MaxIter || n <= Config.MaxIter); 413 v *= 2, ++n) { 414 DEBUG(dbgs() << "BBV: fusing loop #" << n << 415 " for " << BB.getName() << " in " << 416 BB.getParent()->getName() << "...\n"); 417 if (vectorizePairs(BB)) 418 changed = true; 419 else 420 break; 421 } 422 423 if (changed && !Pow2LenOnly) { 424 ++n; 425 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 426 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 427 n << " for " << BB.getName() << " in " << 428 BB.getParent()->getName() << "...\n"); 429 if (!vectorizePairs(BB, true)) break; 430 } 431 } 432 433 DEBUG(dbgs() << "BBV: done!\n"); 434 return changed; 435 } 436 437 bool runOnBasicBlock(BasicBlock &BB) override { 438 // OptimizeNone check deferred to vectorizeBB(). 439 440 AA = &getAnalysis<AliasAnalysis>(); 441 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 442 SE = &getAnalysis<ScalarEvolution>(); 443 TTI = IgnoreTargetInfo 444 ? nullptr 445 : &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( 446 *BB.getParent()); 447 448 return vectorizeBB(BB); 449 } 450 451 void getAnalysisUsage(AnalysisUsage &AU) const override { 452 BasicBlockPass::getAnalysisUsage(AU); 453 AU.addRequired<AliasAnalysis>(); 454 AU.addRequired<DominatorTreeWrapperPass>(); 455 AU.addRequired<ScalarEvolution>(); 456 AU.addRequired<TargetTransformInfoWrapperPass>(); 457 AU.addPreserved<AliasAnalysis>(); 458 AU.addPreserved<DominatorTreeWrapperPass>(); 459 AU.addPreserved<ScalarEvolution>(); 460 AU.setPreservesCFG(); 461 } 462 463 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 464 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 465 "Cannot form vector from incompatible scalar types"); 466 Type *STy = ElemTy->getScalarType(); 467 468 unsigned numElem; 469 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 470 numElem = VTy->getNumElements(); 471 } else { 472 numElem = 1; 473 } 474 475 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 476 numElem += VTy->getNumElements(); 477 } else { 478 numElem += 1; 479 } 480 481 return VectorType::get(STy, numElem); 482 } 483 484 static inline void getInstructionTypes(Instruction *I, 485 Type *&T1, Type *&T2) { 486 if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 487 // For stores, it is the value type, not the pointer type that matters 488 // because the value is what will come from a vector register. 489 490 Value *IVal = SI->getValueOperand(); 491 T1 = IVal->getType(); 492 } else { 493 T1 = I->getType(); 494 } 495 496 if (CastInst *CI = dyn_cast<CastInst>(I)) 497 T2 = CI->getSrcTy(); 498 else 499 T2 = T1; 500 501 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 502 T2 = SI->getCondition()->getType(); 503 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { 504 T2 = SI->getOperand(0)->getType(); 505 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 506 T2 = CI->getOperand(0)->getType(); 507 } 508 } 509 510 // Returns the weight associated with the provided value. A chain of 511 // candidate pairs has a length given by the sum of the weights of its 512 // members (one weight per pair; the weight of each member of the pair 513 // is assumed to be the same). This length is then compared to the 514 // chain-length threshold to determine if a given chain is significant 515 // enough to be vectorized. The length is also used in comparing 516 // candidate chains where longer chains are considered to be better. 517 // Note: when this function returns 0, the resulting instructions are 518 // not actually fused. 519 inline size_t getDepthFactor(Value *V) { 520 // InsertElement and ExtractElement have a depth factor of zero. This is 521 // for two reasons: First, they cannot be usefully fused. Second, because 522 // the pass generates a lot of these, they can confuse the simple metric 523 // used to compare the dags in the next iteration. Thus, giving them a 524 // weight of zero allows the pass to essentially ignore them in 525 // subsequent iterations when looking for vectorization opportunities 526 // while still tracking dependency chains that flow through those 527 // instructions. 528 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 529 return 0; 530 531 // Give a load or store half of the required depth so that load/store 532 // pairs will vectorize. 533 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 534 return Config.ReqChainDepth/2; 535 536 return 1; 537 } 538 539 // Returns the cost of the provided instruction using TTI. 540 // This does not handle loads and stores. 541 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2, 542 TargetTransformInfo::OperandValueKind Op1VK = 543 TargetTransformInfo::OK_AnyValue, 544 TargetTransformInfo::OperandValueKind Op2VK = 545 TargetTransformInfo::OK_AnyValue) { 546 switch (Opcode) { 547 default: break; 548 case Instruction::GetElementPtr: 549 // We mark this instruction as zero-cost because scalar GEPs are usually 550 // lowered to the instruction addressing mode. At the moment we don't 551 // generate vector GEPs. 552 return 0; 553 case Instruction::Br: 554 return TTI->getCFInstrCost(Opcode); 555 case Instruction::PHI: 556 return 0; 557 case Instruction::Add: 558 case Instruction::FAdd: 559 case Instruction::Sub: 560 case Instruction::FSub: 561 case Instruction::Mul: 562 case Instruction::FMul: 563 case Instruction::UDiv: 564 case Instruction::SDiv: 565 case Instruction::FDiv: 566 case Instruction::URem: 567 case Instruction::SRem: 568 case Instruction::FRem: 569 case Instruction::Shl: 570 case Instruction::LShr: 571 case Instruction::AShr: 572 case Instruction::And: 573 case Instruction::Or: 574 case Instruction::Xor: 575 return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK); 576 case Instruction::Select: 577 case Instruction::ICmp: 578 case Instruction::FCmp: 579 return TTI->getCmpSelInstrCost(Opcode, T1, T2); 580 case Instruction::ZExt: 581 case Instruction::SExt: 582 case Instruction::FPToUI: 583 case Instruction::FPToSI: 584 case Instruction::FPExt: 585 case Instruction::PtrToInt: 586 case Instruction::IntToPtr: 587 case Instruction::SIToFP: 588 case Instruction::UIToFP: 589 case Instruction::Trunc: 590 case Instruction::FPTrunc: 591 case Instruction::BitCast: 592 case Instruction::ShuffleVector: 593 return TTI->getCastInstrCost(Opcode, T1, T2); 594 } 595 596 return 1; 597 } 598 599 // This determines the relative offset of two loads or stores, returning 600 // true if the offset could be determined to be some constant value. 601 // For example, if OffsetInElmts == 1, then J accesses the memory directly 602 // after I; if OffsetInElmts == -1 then I accesses the memory 603 // directly after J. 604 bool getPairPtrInfo(Instruction *I, Instruction *J, 605 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 606 unsigned &IAddressSpace, unsigned &JAddressSpace, 607 int64_t &OffsetInElmts, bool ComputeOffset = true) { 608 OffsetInElmts = 0; 609 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 610 LoadInst *LJ = cast<LoadInst>(J); 611 IPtr = LI->getPointerOperand(); 612 JPtr = LJ->getPointerOperand(); 613 IAlignment = LI->getAlignment(); 614 JAlignment = LJ->getAlignment(); 615 IAddressSpace = LI->getPointerAddressSpace(); 616 JAddressSpace = LJ->getPointerAddressSpace(); 617 } else { 618 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 619 IPtr = SI->getPointerOperand(); 620 JPtr = SJ->getPointerOperand(); 621 IAlignment = SI->getAlignment(); 622 JAlignment = SJ->getAlignment(); 623 IAddressSpace = SI->getPointerAddressSpace(); 624 JAddressSpace = SJ->getPointerAddressSpace(); 625 } 626 627 if (!ComputeOffset) 628 return true; 629 630 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 631 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 632 633 // If this is a trivial offset, then we'll get something like 634 // 1*sizeof(type). With target data, which we need anyway, this will get 635 // constant folded into a number. 636 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 637 if (const SCEVConstant *ConstOffSCEV = 638 dyn_cast<SCEVConstant>(OffsetSCEV)) { 639 ConstantInt *IntOff = ConstOffSCEV->getValue(); 640 int64_t Offset = IntOff->getSExtValue(); 641 const DataLayout &DL = I->getModule()->getDataLayout(); 642 Type *VTy = IPtr->getType()->getPointerElementType(); 643 int64_t VTyTSS = (int64_t)DL.getTypeStoreSize(VTy); 644 645 Type *VTy2 = JPtr->getType()->getPointerElementType(); 646 if (VTy != VTy2 && Offset < 0) { 647 int64_t VTy2TSS = (int64_t)DL.getTypeStoreSize(VTy2); 648 OffsetInElmts = Offset/VTy2TSS; 649 return (std::abs(Offset) % VTy2TSS) == 0; 650 } 651 652 OffsetInElmts = Offset/VTyTSS; 653 return (std::abs(Offset) % VTyTSS) == 0; 654 } 655 656 return false; 657 } 658 659 // Returns true if the provided CallInst represents an intrinsic that can 660 // be vectorized. 661 bool isVectorizableIntrinsic(CallInst* I) { 662 Function *F = I->getCalledFunction(); 663 if (!F) return false; 664 665 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 666 if (!IID) return false; 667 668 switch(IID) { 669 default: 670 return false; 671 case Intrinsic::sqrt: 672 case Intrinsic::powi: 673 case Intrinsic::sin: 674 case Intrinsic::cos: 675 case Intrinsic::log: 676 case Intrinsic::log2: 677 case Intrinsic::log10: 678 case Intrinsic::exp: 679 case Intrinsic::exp2: 680 case Intrinsic::pow: 681 case Intrinsic::round: 682 case Intrinsic::copysign: 683 case Intrinsic::ceil: 684 case Intrinsic::nearbyint: 685 case Intrinsic::rint: 686 case Intrinsic::trunc: 687 case Intrinsic::floor: 688 case Intrinsic::fabs: 689 case Intrinsic::minnum: 690 case Intrinsic::maxnum: 691 return Config.VectorizeMath; 692 case Intrinsic::bswap: 693 case Intrinsic::ctpop: 694 case Intrinsic::ctlz: 695 case Intrinsic::cttz: 696 return Config.VectorizeBitManipulations; 697 case Intrinsic::fma: 698 case Intrinsic::fmuladd: 699 return Config.VectorizeFMA; 700 } 701 } 702 703 bool isPureIEChain(InsertElementInst *IE) { 704 InsertElementInst *IENext = IE; 705 do { 706 if (!isa<UndefValue>(IENext->getOperand(0)) && 707 !isa<InsertElementInst>(IENext->getOperand(0))) { 708 return false; 709 } 710 } while ((IENext = 711 dyn_cast<InsertElementInst>(IENext->getOperand(0)))); 712 713 return true; 714 } 715 }; 716 717 // This function implements one vectorization iteration on the provided 718 // basic block. It returns true if the block is changed. 719 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 720 bool ShouldContinue; 721 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 722 723 std::vector<Value *> AllPairableInsts; 724 DenseMap<Value *, Value *> AllChosenPairs; 725 DenseSet<ValuePair> AllFixedOrderPairs; 726 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 727 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs, 728 AllConnectedPairDeps; 729 730 do { 731 std::vector<Value *> PairableInsts; 732 DenseMap<Value *, std::vector<Value *> > CandidatePairs; 733 DenseSet<ValuePair> FixedOrderPairs; 734 DenseMap<ValuePair, int> CandidatePairCostSavings; 735 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 736 FixedOrderPairs, 737 CandidatePairCostSavings, 738 PairableInsts, NonPow2Len); 739 if (PairableInsts.empty()) continue; 740 741 // Build the candidate pair set for faster lookups. 742 DenseSet<ValuePair> CandidatePairsSet; 743 for (DenseMap<Value *, std::vector<Value *> >::iterator I = 744 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I) 745 for (std::vector<Value *>::iterator J = I->second.begin(), 746 JE = I->second.end(); J != JE; ++J) 747 CandidatePairsSet.insert(ValuePair(I->first, *J)); 748 749 // Now we have a map of all of the pairable instructions and we need to 750 // select the best possible pairing. A good pairing is one such that the 751 // users of the pair are also paired. This defines a (directed) forest 752 // over the pairs such that two pairs are connected iff the second pair 753 // uses the first. 754 755 // Note that it only matters that both members of the second pair use some 756 // element of the first pair (to allow for splatting). 757 758 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs, 759 ConnectedPairDeps; 760 DenseMap<VPPair, unsigned> PairConnectionTypes; 761 computeConnectedPairs(CandidatePairs, CandidatePairsSet, 762 PairableInsts, ConnectedPairs, PairConnectionTypes); 763 if (ConnectedPairs.empty()) continue; 764 765 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator 766 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 767 I != IE; ++I) 768 for (std::vector<ValuePair>::iterator J = I->second.begin(), 769 JE = I->second.end(); J != JE; ++J) 770 ConnectedPairDeps[*J].push_back(I->first); 771 772 // Build the pairable-instruction dependency map 773 DenseSet<ValuePair> PairableInstUsers; 774 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 775 776 // There is now a graph of the connected pairs. For each variable, pick 777 // the pairing with the largest dag meeting the depth requirement on at 778 // least one branch. Then select all pairings that are part of that dag 779 // and remove them from the list of available pairings and pairable 780 // variables. 781 782 DenseMap<Value *, Value *> ChosenPairs; 783 choosePairs(CandidatePairs, CandidatePairsSet, 784 CandidatePairCostSavings, 785 PairableInsts, FixedOrderPairs, PairConnectionTypes, 786 ConnectedPairs, ConnectedPairDeps, 787 PairableInstUsers, ChosenPairs); 788 789 if (ChosenPairs.empty()) continue; 790 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 791 PairableInsts.end()); 792 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 793 794 // Only for the chosen pairs, propagate information on fixed-order pairs, 795 // pair connections, and their types to the data structures used by the 796 // pair fusion procedures. 797 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 798 IE = ChosenPairs.end(); I != IE; ++I) { 799 if (FixedOrderPairs.count(*I)) 800 AllFixedOrderPairs.insert(*I); 801 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 802 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 803 804 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 805 J != IE; ++J) { 806 DenseMap<VPPair, unsigned>::iterator K = 807 PairConnectionTypes.find(VPPair(*I, *J)); 808 if (K != PairConnectionTypes.end()) { 809 AllPairConnectionTypes.insert(*K); 810 } else { 811 K = PairConnectionTypes.find(VPPair(*J, *I)); 812 if (K != PairConnectionTypes.end()) 813 AllPairConnectionTypes.insert(*K); 814 } 815 } 816 } 817 818 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator 819 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 820 I != IE; ++I) 821 for (std::vector<ValuePair>::iterator J = I->second.begin(), 822 JE = I->second.end(); J != JE; ++J) 823 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) { 824 AllConnectedPairs[I->first].push_back(*J); 825 AllConnectedPairDeps[*J].push_back(I->first); 826 } 827 } while (ShouldContinue); 828 829 if (AllChosenPairs.empty()) return false; 830 NumFusedOps += AllChosenPairs.size(); 831 832 // A set of pairs has now been selected. It is now necessary to replace the 833 // paired instructions with vector instructions. For this procedure each 834 // operand must be replaced with a vector operand. This vector is formed 835 // by using build_vector on the old operands. The replaced values are then 836 // replaced with a vector_extract on the result. Subsequent optimization 837 // passes should coalesce the build/extract combinations. 838 839 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 840 AllPairConnectionTypes, 841 AllConnectedPairs, AllConnectedPairDeps); 842 843 // It is important to cleanup here so that future iterations of this 844 // function have less work to do. 845 (void)SimplifyInstructionsInBlock(&BB, AA->getTargetLibraryInfo()); 846 return true; 847 } 848 849 // This function returns true if the provided instruction is capable of being 850 // fused into a vector instruction. This determination is based only on the 851 // type and other attributes of the instruction. 852 bool BBVectorize::isInstVectorizable(Instruction *I, 853 bool &IsSimpleLoadStore) { 854 IsSimpleLoadStore = false; 855 856 if (CallInst *C = dyn_cast<CallInst>(I)) { 857 if (!isVectorizableIntrinsic(C)) 858 return false; 859 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 860 // Vectorize simple loads if possbile: 861 IsSimpleLoadStore = L->isSimple(); 862 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 863 return false; 864 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 865 // Vectorize simple stores if possbile: 866 IsSimpleLoadStore = S->isSimple(); 867 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 868 return false; 869 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 870 // We can vectorize casts, but not casts of pointer types, etc. 871 if (!Config.VectorizeCasts) 872 return false; 873 874 Type *SrcTy = C->getSrcTy(); 875 if (!SrcTy->isSingleValueType()) 876 return false; 877 878 Type *DestTy = C->getDestTy(); 879 if (!DestTy->isSingleValueType()) 880 return false; 881 } else if (isa<SelectInst>(I)) { 882 if (!Config.VectorizeSelect) 883 return false; 884 } else if (isa<CmpInst>(I)) { 885 if (!Config.VectorizeCmp) 886 return false; 887 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 888 if (!Config.VectorizeGEP) 889 return false; 890 891 // Currently, vector GEPs exist only with one index. 892 if (G->getNumIndices() != 1) 893 return false; 894 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 895 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 896 return false; 897 } 898 899 Type *T1, *T2; 900 getInstructionTypes(I, T1, T2); 901 902 // Not every type can be vectorized... 903 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 904 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 905 return false; 906 907 if (T1->getScalarSizeInBits() == 1) { 908 if (!Config.VectorizeBools) 909 return false; 910 } else { 911 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 912 return false; 913 } 914 915 if (T2->getScalarSizeInBits() == 1) { 916 if (!Config.VectorizeBools) 917 return false; 918 } else { 919 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 920 return false; 921 } 922 923 if (!Config.VectorizeFloats 924 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 925 return false; 926 927 // Don't vectorize target-specific types. 928 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 929 return false; 930 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 931 return false; 932 933 if (!Config.VectorizePointers && (T1->getScalarType()->isPointerTy() || 934 T2->getScalarType()->isPointerTy())) 935 return false; 936 937 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 938 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 939 return false; 940 941 return true; 942 } 943 944 // This function returns true if the two provided instructions are compatible 945 // (meaning that they can be fused into a vector instruction). This assumes 946 // that I has already been determined to be vectorizable and that J is not 947 // in the use dag of I. 948 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 949 bool IsSimpleLoadStore, bool NonPow2Len, 950 int &CostSavings, int &FixedOrder) { 951 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 952 " <-> " << *J << "\n"); 953 954 CostSavings = 0; 955 FixedOrder = 0; 956 957 // Loads and stores can be merged if they have different alignments, 958 // but are otherwise the same. 959 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 960 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 961 return false; 962 963 Type *IT1, *IT2, *JT1, *JT2; 964 getInstructionTypes(I, IT1, IT2); 965 getInstructionTypes(J, JT1, JT2); 966 unsigned MaxTypeBits = std::max( 967 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 968 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 969 if (!TTI && MaxTypeBits > Config.VectorBits) 970 return false; 971 972 // FIXME: handle addsub-type operations! 973 974 if (IsSimpleLoadStore) { 975 Value *IPtr, *JPtr; 976 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 977 int64_t OffsetInElmts = 0; 978 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 979 IAddressSpace, JAddressSpace, OffsetInElmts) && 980 std::abs(OffsetInElmts) == 1) { 981 FixedOrder = (int) OffsetInElmts; 982 unsigned BottomAlignment = IAlignment; 983 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 984 985 Type *aTypeI = isa<StoreInst>(I) ? 986 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 987 Type *aTypeJ = isa<StoreInst>(J) ? 988 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 989 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 990 991 if (Config.AlignedOnly) { 992 // An aligned load or store is possible only if the instruction 993 // with the lower offset has an alignment suitable for the 994 // vector type. 995 const DataLayout &DL = I->getModule()->getDataLayout(); 996 unsigned VecAlignment = DL.getPrefTypeAlignment(VType); 997 if (BottomAlignment < VecAlignment) 998 return false; 999 } 1000 1001 if (TTI) { 1002 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI, 1003 IAlignment, IAddressSpace); 1004 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ, 1005 JAlignment, JAddressSpace); 1006 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType, 1007 BottomAlignment, 1008 IAddressSpace); 1009 1010 ICost += TTI->getAddressComputationCost(aTypeI); 1011 JCost += TTI->getAddressComputationCost(aTypeJ); 1012 VCost += TTI->getAddressComputationCost(VType); 1013 1014 if (VCost > ICost + JCost) 1015 return false; 1016 1017 // We don't want to fuse to a type that will be split, even 1018 // if the two input types will also be split and there is no other 1019 // associated cost. 1020 unsigned VParts = TTI->getNumberOfParts(VType); 1021 if (VParts > 1) 1022 return false; 1023 else if (!VParts && VCost == ICost + JCost) 1024 return false; 1025 1026 CostSavings = ICost + JCost - VCost; 1027 } 1028 } else { 1029 return false; 1030 } 1031 } else if (TTI) { 1032 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 1033 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 1034 Type *VT1 = getVecTypeForPair(IT1, JT1), 1035 *VT2 = getVecTypeForPair(IT2, JT2); 1036 TargetTransformInfo::OperandValueKind Op1VK = 1037 TargetTransformInfo::OK_AnyValue; 1038 TargetTransformInfo::OperandValueKind Op2VK = 1039 TargetTransformInfo::OK_AnyValue; 1040 1041 // On some targets (example X86) the cost of a vector shift may vary 1042 // depending on whether the second operand is a Uniform or 1043 // NonUniform Constant. 1044 switch (I->getOpcode()) { 1045 default : break; 1046 case Instruction::Shl: 1047 case Instruction::LShr: 1048 case Instruction::AShr: 1049 1050 // If both I and J are scalar shifts by constant, then the 1051 // merged vector shift count would be either a constant splat value 1052 // or a non-uniform vector of constants. 1053 if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) { 1054 if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1))) 1055 Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue : 1056 TargetTransformInfo::OK_NonUniformConstantValue; 1057 } else { 1058 // Check for a splat of a constant or for a non uniform vector 1059 // of constants. 1060 Value *IOp = I->getOperand(1); 1061 Value *JOp = J->getOperand(1); 1062 if ((isa<ConstantVector>(IOp) || isa<ConstantDataVector>(IOp)) && 1063 (isa<ConstantVector>(JOp) || isa<ConstantDataVector>(JOp))) { 1064 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; 1065 Constant *SplatValue = cast<Constant>(IOp)->getSplatValue(); 1066 if (SplatValue != nullptr && 1067 SplatValue == cast<Constant>(JOp)->getSplatValue()) 1068 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 1069 } 1070 } 1071 } 1072 1073 // Note that this procedure is incorrect for insert and extract element 1074 // instructions (because combining these often results in a shuffle), 1075 // but this cost is ignored (because insert and extract element 1076 // instructions are assigned a zero depth factor and are not really 1077 // fused in general). 1078 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK); 1079 1080 if (VCost > ICost + JCost) 1081 return false; 1082 1083 // We don't want to fuse to a type that will be split, even 1084 // if the two input types will also be split and there is no other 1085 // associated cost. 1086 unsigned VParts1 = TTI->getNumberOfParts(VT1), 1087 VParts2 = TTI->getNumberOfParts(VT2); 1088 if (VParts1 > 1 || VParts2 > 1) 1089 return false; 1090 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) 1091 return false; 1092 1093 CostSavings = ICost + JCost - VCost; 1094 } 1095 1096 // The powi,ctlz,cttz intrinsics are special because only the first 1097 // argument is vectorized, the second arguments must be equal. 1098 CallInst *CI = dyn_cast<CallInst>(I); 1099 Function *FI; 1100 if (CI && (FI = CI->getCalledFunction())) { 1101 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID(); 1102 if (IID == Intrinsic::powi || IID == Intrinsic::ctlz || 1103 IID == Intrinsic::cttz) { 1104 Value *A1I = CI->getArgOperand(1), 1105 *A1J = cast<CallInst>(J)->getArgOperand(1); 1106 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1107 *A1JSCEV = SE->getSCEV(A1J); 1108 return (A1ISCEV == A1JSCEV); 1109 } 1110 1111 if (IID && TTI) { 1112 SmallVector<Type*, 4> Tys; 1113 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 1114 Tys.push_back(CI->getArgOperand(i)->getType()); 1115 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys); 1116 1117 Tys.clear(); 1118 CallInst *CJ = cast<CallInst>(J); 1119 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i) 1120 Tys.push_back(CJ->getArgOperand(i)->getType()); 1121 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys); 1122 1123 Tys.clear(); 1124 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() && 1125 "Intrinsic argument counts differ"); 1126 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1127 if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz || 1128 IID == Intrinsic::cttz) && i == 1) 1129 Tys.push_back(CI->getArgOperand(i)->getType()); 1130 else 1131 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(), 1132 CJ->getArgOperand(i)->getType())); 1133 } 1134 1135 Type *RetTy = getVecTypeForPair(IT1, JT1); 1136 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys); 1137 1138 if (VCost > ICost + JCost) 1139 return false; 1140 1141 // We don't want to fuse to a type that will be split, even 1142 // if the two input types will also be split and there is no other 1143 // associated cost. 1144 unsigned RetParts = TTI->getNumberOfParts(RetTy); 1145 if (RetParts > 1) 1146 return false; 1147 else if (!RetParts && VCost == ICost + JCost) 1148 return false; 1149 1150 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1151 if (!Tys[i]->isVectorTy()) 1152 continue; 1153 1154 unsigned NumParts = TTI->getNumberOfParts(Tys[i]); 1155 if (NumParts > 1) 1156 return false; 1157 else if (!NumParts && VCost == ICost + JCost) 1158 return false; 1159 } 1160 1161 CostSavings = ICost + JCost - VCost; 1162 } 1163 } 1164 1165 return true; 1166 } 1167 1168 // Figure out whether or not J uses I and update the users and write-set 1169 // structures associated with I. Specifically, Users represents the set of 1170 // instructions that depend on I. WriteSet represents the set 1171 // of memory locations that are dependent on I. If UpdateUsers is true, 1172 // and J uses I, then Users is updated to contain J and WriteSet is updated 1173 // to contain any memory locations to which J writes. The function returns 1174 // true if J uses I. By default, alias analysis is used to determine 1175 // whether J reads from memory that overlaps with a location in WriteSet. 1176 // If LoadMoveSet is not null, then it is a previously-computed map 1177 // where the key is the memory-based user instruction and the value is 1178 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1179 // then the alias analysis is not used. This is necessary because this 1180 // function is called during the process of moving instructions during 1181 // vectorization and the results of the alias analysis are not stable during 1182 // that process. 1183 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1184 AliasSetTracker &WriteSet, Instruction *I, 1185 Instruction *J, bool UpdateUsers, 1186 DenseSet<ValuePair> *LoadMoveSetPairs) { 1187 bool UsesI = false; 1188 1189 // This instruction may already be marked as a user due, for example, to 1190 // being a member of a selected pair. 1191 if (Users.count(J)) 1192 UsesI = true; 1193 1194 if (!UsesI) 1195 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1196 JU != JE; ++JU) { 1197 Value *V = *JU; 1198 if (I == V || Users.count(V)) { 1199 UsesI = true; 1200 break; 1201 } 1202 } 1203 if (!UsesI && J->mayReadFromMemory()) { 1204 if (LoadMoveSetPairs) { 1205 UsesI = LoadMoveSetPairs->count(ValuePair(J, I)); 1206 } else { 1207 for (AliasSetTracker::iterator W = WriteSet.begin(), 1208 WE = WriteSet.end(); W != WE; ++W) { 1209 if (W->aliasesUnknownInst(J, *AA)) { 1210 UsesI = true; 1211 break; 1212 } 1213 } 1214 } 1215 } 1216 1217 if (UsesI && UpdateUsers) { 1218 if (J->mayWriteToMemory()) WriteSet.add(J); 1219 Users.insert(J); 1220 } 1221 1222 return UsesI; 1223 } 1224 1225 // This function iterates over all instruction pairs in the provided 1226 // basic block and collects all candidate pairs for vectorization. 1227 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1228 BasicBlock::iterator &Start, 1229 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1230 DenseSet<ValuePair> &FixedOrderPairs, 1231 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1232 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1233 size_t TotalPairs = 0; 1234 BasicBlock::iterator E = BB.end(); 1235 if (Start == E) return false; 1236 1237 bool ShouldContinue = false, IAfterStart = false; 1238 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1239 if (I == Start) IAfterStart = true; 1240 1241 bool IsSimpleLoadStore; 1242 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1243 1244 // Look for an instruction with which to pair instruction *I... 1245 DenseSet<Value *> Users; 1246 AliasSetTracker WriteSet(*AA); 1247 if (I->mayWriteToMemory()) WriteSet.add(I); 1248 1249 bool JAfterStart = IAfterStart; 1250 BasicBlock::iterator J = std::next(I); 1251 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1252 if (J == Start) JAfterStart = true; 1253 1254 // Determine if J uses I, if so, exit the loop. 1255 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1256 if (Config.FastDep) { 1257 // Note: For this heuristic to be effective, independent operations 1258 // must tend to be intermixed. This is likely to be true from some 1259 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1260 // but otherwise may require some kind of reordering pass. 1261 1262 // When using fast dependency analysis, 1263 // stop searching after first use: 1264 if (UsesI) break; 1265 } else { 1266 if (UsesI) continue; 1267 } 1268 1269 // J does not use I, and comes before the first use of I, so it can be 1270 // merged with I if the instructions are compatible. 1271 int CostSavings, FixedOrder; 1272 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1273 CostSavings, FixedOrder)) continue; 1274 1275 // J is a candidate for merging with I. 1276 if (PairableInsts.empty() || 1277 PairableInsts[PairableInsts.size()-1] != I) { 1278 PairableInsts.push_back(I); 1279 } 1280 1281 CandidatePairs[I].push_back(J); 1282 ++TotalPairs; 1283 if (TTI) 1284 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1285 CostSavings)); 1286 1287 if (FixedOrder == 1) 1288 FixedOrderPairs.insert(ValuePair(I, J)); 1289 else if (FixedOrder == -1) 1290 FixedOrderPairs.insert(ValuePair(J, I)); 1291 1292 // The next call to this function must start after the last instruction 1293 // selected during this invocation. 1294 if (JAfterStart) { 1295 Start = std::next(J); 1296 IAfterStart = JAfterStart = false; 1297 } 1298 1299 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1300 << *I << " <-> " << *J << " (cost savings: " << 1301 CostSavings << ")\n"); 1302 1303 // If we have already found too many pairs, break here and this function 1304 // will be called again starting after the last instruction selected 1305 // during this invocation. 1306 if (PairableInsts.size() >= Config.MaxInsts || 1307 TotalPairs >= Config.MaxPairs) { 1308 ShouldContinue = true; 1309 break; 1310 } 1311 } 1312 1313 if (ShouldContinue) 1314 break; 1315 } 1316 1317 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1318 << " instructions with candidate pairs\n"); 1319 1320 return ShouldContinue; 1321 } 1322 1323 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1324 // it looks for pairs such that both members have an input which is an 1325 // output of PI or PJ. 1326 void BBVectorize::computePairsConnectedTo( 1327 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1328 DenseSet<ValuePair> &CandidatePairsSet, 1329 std::vector<Value *> &PairableInsts, 1330 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1331 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1332 ValuePair P) { 1333 StoreInst *SI, *SJ; 1334 1335 // For each possible pairing for this variable, look at the uses of 1336 // the first value... 1337 for (Value::user_iterator I = P.first->user_begin(), 1338 E = P.first->user_end(); 1339 I != E; ++I) { 1340 User *UI = *I; 1341 if (isa<LoadInst>(UI)) { 1342 // A pair cannot be connected to a load because the load only takes one 1343 // operand (the address) and it is a scalar even after vectorization. 1344 continue; 1345 } else if ((SI = dyn_cast<StoreInst>(UI)) && 1346 P.first == SI->getPointerOperand()) { 1347 // Similarly, a pair cannot be connected to a store through its 1348 // pointer operand. 1349 continue; 1350 } 1351 1352 // For each use of the first variable, look for uses of the second 1353 // variable... 1354 for (User *UJ : P.second->users()) { 1355 if ((SJ = dyn_cast<StoreInst>(UJ)) && 1356 P.second == SJ->getPointerOperand()) 1357 continue; 1358 1359 // Look for <I, J>: 1360 if (CandidatePairsSet.count(ValuePair(UI, UJ))) { 1361 VPPair VP(P, ValuePair(UI, UJ)); 1362 ConnectedPairs[VP.first].push_back(VP.second); 1363 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1364 } 1365 1366 // Look for <J, I>: 1367 if (CandidatePairsSet.count(ValuePair(UJ, UI))) { 1368 VPPair VP(P, ValuePair(UJ, UI)); 1369 ConnectedPairs[VP.first].push_back(VP.second); 1370 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1371 } 1372 } 1373 1374 if (Config.SplatBreaksChain) continue; 1375 // Look for cases where just the first value in the pair is used by 1376 // both members of another pair (splatting). 1377 for (Value::user_iterator J = P.first->user_begin(); J != E; ++J) { 1378 User *UJ = *J; 1379 if ((SJ = dyn_cast<StoreInst>(UJ)) && 1380 P.first == SJ->getPointerOperand()) 1381 continue; 1382 1383 if (CandidatePairsSet.count(ValuePair(UI, UJ))) { 1384 VPPair VP(P, ValuePair(UI, UJ)); 1385 ConnectedPairs[VP.first].push_back(VP.second); 1386 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1387 } 1388 } 1389 } 1390 1391 if (Config.SplatBreaksChain) return; 1392 // Look for cases where just the second value in the pair is used by 1393 // both members of another pair (splatting). 1394 for (Value::user_iterator I = P.second->user_begin(), 1395 E = P.second->user_end(); 1396 I != E; ++I) { 1397 User *UI = *I; 1398 if (isa<LoadInst>(UI)) 1399 continue; 1400 else if ((SI = dyn_cast<StoreInst>(UI)) && 1401 P.second == SI->getPointerOperand()) 1402 continue; 1403 1404 for (Value::user_iterator J = P.second->user_begin(); J != E; ++J) { 1405 User *UJ = *J; 1406 if ((SJ = dyn_cast<StoreInst>(UJ)) && 1407 P.second == SJ->getPointerOperand()) 1408 continue; 1409 1410 if (CandidatePairsSet.count(ValuePair(UI, UJ))) { 1411 VPPair VP(P, ValuePair(UI, UJ)); 1412 ConnectedPairs[VP.first].push_back(VP.second); 1413 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1414 } 1415 } 1416 } 1417 } 1418 1419 // This function figures out which pairs are connected. Two pairs are 1420 // connected if some output of the first pair forms an input to both members 1421 // of the second pair. 1422 void BBVectorize::computeConnectedPairs( 1423 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1424 DenseSet<ValuePair> &CandidatePairsSet, 1425 std::vector<Value *> &PairableInsts, 1426 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1427 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1428 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1429 PE = PairableInsts.end(); PI != PE; ++PI) { 1430 DenseMap<Value *, std::vector<Value *> >::iterator PP = 1431 CandidatePairs.find(*PI); 1432 if (PP == CandidatePairs.end()) 1433 continue; 1434 1435 for (std::vector<Value *>::iterator P = PP->second.begin(), 1436 E = PP->second.end(); P != E; ++P) 1437 computePairsConnectedTo(CandidatePairs, CandidatePairsSet, 1438 PairableInsts, ConnectedPairs, 1439 PairConnectionTypes, ValuePair(*PI, *P)); 1440 } 1441 1442 DEBUG(size_t TotalPairs = 0; 1443 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I = 1444 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I) 1445 TotalPairs += I->second.size(); 1446 dbgs() << "BBV: found " << TotalPairs 1447 << " pair connections.\n"); 1448 } 1449 1450 // This function builds a set of use tuples such that <A, B> is in the set 1451 // if B is in the use dag of A. If B is in the use dag of A, then B 1452 // depends on the output of A. 1453 void BBVectorize::buildDepMap( 1454 BasicBlock &BB, 1455 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1456 std::vector<Value *> &PairableInsts, 1457 DenseSet<ValuePair> &PairableInstUsers) { 1458 DenseSet<Value *> IsInPair; 1459 for (DenseMap<Value *, std::vector<Value *> >::iterator C = 1460 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) { 1461 IsInPair.insert(C->first); 1462 IsInPair.insert(C->second.begin(), C->second.end()); 1463 } 1464 1465 // Iterate through the basic block, recording all users of each 1466 // pairable instruction. 1467 1468 BasicBlock::iterator E = BB.end(), EL = 1469 BasicBlock::iterator(cast<Instruction>(PairableInsts.back())); 1470 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1471 if (IsInPair.find(I) == IsInPair.end()) continue; 1472 1473 DenseSet<Value *> Users; 1474 AliasSetTracker WriteSet(*AA); 1475 if (I->mayWriteToMemory()) WriteSet.add(I); 1476 1477 for (BasicBlock::iterator J = std::next(I); J != E; ++J) { 1478 (void) trackUsesOfI(Users, WriteSet, I, J); 1479 1480 if (J == EL) 1481 break; 1482 } 1483 1484 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1485 U != E; ++U) { 1486 if (IsInPair.find(*U) == IsInPair.end()) continue; 1487 PairableInstUsers.insert(ValuePair(I, *U)); 1488 } 1489 1490 if (I == EL) 1491 break; 1492 } 1493 } 1494 1495 // Returns true if an input to pair P is an output of pair Q and also an 1496 // input of pair Q is an output of pair P. If this is the case, then these 1497 // two pairs cannot be simultaneously fused. 1498 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1499 DenseSet<ValuePair> &PairableInstUsers, 1500 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap, 1501 DenseSet<VPPair> *PairableInstUserPairSet) { 1502 // Two pairs are in conflict if they are mutual Users of eachother. 1503 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1504 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1505 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1506 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1507 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1508 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1509 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1510 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1511 if (PairableInstUserMap) { 1512 // FIXME: The expensive part of the cycle check is not so much the cycle 1513 // check itself but this edge insertion procedure. This needs some 1514 // profiling and probably a different data structure. 1515 if (PUsesQ) { 1516 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second) 1517 (*PairableInstUserMap)[Q].push_back(P); 1518 } 1519 if (QUsesP) { 1520 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second) 1521 (*PairableInstUserMap)[P].push_back(Q); 1522 } 1523 } 1524 1525 return (QUsesP && PUsesQ); 1526 } 1527 1528 // This function walks the use graph of current pairs to see if, starting 1529 // from P, the walk returns to P. 1530 bool BBVectorize::pairWillFormCycle(ValuePair P, 1531 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 1532 DenseSet<ValuePair> &CurrentPairs) { 1533 DEBUG(if (DebugCycleCheck) 1534 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1535 << *P.second << "\n"); 1536 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1537 // contains non-direct associations. 1538 DenseSet<ValuePair> Visited; 1539 SmallVector<ValuePair, 32> Q; 1540 // General depth-first post-order traversal: 1541 Q.push_back(P); 1542 do { 1543 ValuePair QTop = Q.pop_back_val(); 1544 Visited.insert(QTop); 1545 1546 DEBUG(if (DebugCycleCheck) 1547 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1548 << *QTop.second << "\n"); 1549 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ = 1550 PairableInstUserMap.find(QTop); 1551 if (QQ == PairableInstUserMap.end()) 1552 continue; 1553 1554 for (std::vector<ValuePair>::iterator C = QQ->second.begin(), 1555 CE = QQ->second.end(); C != CE; ++C) { 1556 if (*C == P) { 1557 DEBUG(dbgs() 1558 << "BBV: rejected to prevent non-trivial cycle formation: " 1559 << QTop.first << " <-> " << C->second << "\n"); 1560 return true; 1561 } 1562 1563 if (CurrentPairs.count(*C) && !Visited.count(*C)) 1564 Q.push_back(*C); 1565 } 1566 } while (!Q.empty()); 1567 1568 return false; 1569 } 1570 1571 // This function builds the initial dag of connected pairs with the 1572 // pair J at the root. 1573 void BBVectorize::buildInitialDAGFor( 1574 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1575 DenseSet<ValuePair> &CandidatePairsSet, 1576 std::vector<Value *> &PairableInsts, 1577 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1578 DenseSet<ValuePair> &PairableInstUsers, 1579 DenseMap<Value *, Value *> &ChosenPairs, 1580 DenseMap<ValuePair, size_t> &DAG, ValuePair J) { 1581 // Each of these pairs is viewed as the root node of a DAG. The DAG 1582 // is then walked (depth-first). As this happens, we keep track of 1583 // the pairs that compose the DAG and the maximum depth of the DAG. 1584 SmallVector<ValuePairWithDepth, 32> Q; 1585 // General depth-first post-order traversal: 1586 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1587 do { 1588 ValuePairWithDepth QTop = Q.back(); 1589 1590 // Push each child onto the queue: 1591 bool MoreChildren = false; 1592 size_t MaxChildDepth = QTop.second; 1593 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ = 1594 ConnectedPairs.find(QTop.first); 1595 if (QQ != ConnectedPairs.end()) 1596 for (std::vector<ValuePair>::iterator k = QQ->second.begin(), 1597 ke = QQ->second.end(); k != ke; ++k) { 1598 // Make sure that this child pair is still a candidate: 1599 if (CandidatePairsSet.count(*k)) { 1600 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k); 1601 if (C == DAG.end()) { 1602 size_t d = getDepthFactor(k->first); 1603 Q.push_back(ValuePairWithDepth(*k, QTop.second+d)); 1604 MoreChildren = true; 1605 } else { 1606 MaxChildDepth = std::max(MaxChildDepth, C->second); 1607 } 1608 } 1609 } 1610 1611 if (!MoreChildren) { 1612 // Record the current pair as part of the DAG: 1613 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1614 Q.pop_back(); 1615 } 1616 } while (!Q.empty()); 1617 } 1618 1619 // Given some initial dag, prune it by removing conflicting pairs (pairs 1620 // that cannot be simultaneously chosen for vectorization). 1621 void BBVectorize::pruneDAGFor( 1622 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1623 std::vector<Value *> &PairableInsts, 1624 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1625 DenseSet<ValuePair> &PairableInstUsers, 1626 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 1627 DenseSet<VPPair> &PairableInstUserPairSet, 1628 DenseMap<Value *, Value *> &ChosenPairs, 1629 DenseMap<ValuePair, size_t> &DAG, 1630 DenseSet<ValuePair> &PrunedDAG, ValuePair J, 1631 bool UseCycleCheck) { 1632 SmallVector<ValuePairWithDepth, 32> Q; 1633 // General depth-first post-order traversal: 1634 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1635 do { 1636 ValuePairWithDepth QTop = Q.pop_back_val(); 1637 PrunedDAG.insert(QTop.first); 1638 1639 // Visit each child, pruning as necessary... 1640 SmallVector<ValuePairWithDepth, 8> BestChildren; 1641 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ = 1642 ConnectedPairs.find(QTop.first); 1643 if (QQ == ConnectedPairs.end()) 1644 continue; 1645 1646 for (std::vector<ValuePair>::iterator K = QQ->second.begin(), 1647 KE = QQ->second.end(); K != KE; ++K) { 1648 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K); 1649 if (C == DAG.end()) continue; 1650 1651 // This child is in the DAG, now we need to make sure it is the 1652 // best of any conflicting children. There could be multiple 1653 // conflicting children, so first, determine if we're keeping 1654 // this child, then delete conflicting children as necessary. 1655 1656 // It is also necessary to guard against pairing-induced 1657 // dependencies. Consider instructions a .. x .. y .. b 1658 // such that (a,b) are to be fused and (x,y) are to be fused 1659 // but a is an input to x and b is an output from y. This 1660 // means that y cannot be moved after b but x must be moved 1661 // after b for (a,b) to be fused. In other words, after 1662 // fusing (a,b) we have y .. a/b .. x where y is an input 1663 // to a/b and x is an output to a/b: x and y can no longer 1664 // be legally fused. To prevent this condition, we must 1665 // make sure that a child pair added to the DAG is not 1666 // both an input and output of an already-selected pair. 1667 1668 // Pairing-induced dependencies can also form from more complicated 1669 // cycles. The pair vs. pair conflicts are easy to check, and so 1670 // that is done explicitly for "fast rejection", and because for 1671 // child vs. child conflicts, we may prefer to keep the current 1672 // pair in preference to the already-selected child. 1673 DenseSet<ValuePair> CurrentPairs; 1674 1675 bool CanAdd = true; 1676 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 1677 = BestChildren.begin(), E2 = BestChildren.end(); 1678 C2 != E2; ++C2) { 1679 if (C2->first.first == C->first.first || 1680 C2->first.first == C->first.second || 1681 C2->first.second == C->first.first || 1682 C2->first.second == C->first.second || 1683 pairsConflict(C2->first, C->first, PairableInstUsers, 1684 UseCycleCheck ? &PairableInstUserMap : nullptr, 1685 UseCycleCheck ? &PairableInstUserPairSet 1686 : nullptr)) { 1687 if (C2->second >= C->second) { 1688 CanAdd = false; 1689 break; 1690 } 1691 1692 CurrentPairs.insert(C2->first); 1693 } 1694 } 1695 if (!CanAdd) continue; 1696 1697 // Even worse, this child could conflict with another node already 1698 // selected for the DAG. If that is the case, ignore this child. 1699 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(), 1700 E2 = PrunedDAG.end(); T != E2; ++T) { 1701 if (T->first == C->first.first || 1702 T->first == C->first.second || 1703 T->second == C->first.first || 1704 T->second == C->first.second || 1705 pairsConflict(*T, C->first, PairableInstUsers, 1706 UseCycleCheck ? &PairableInstUserMap : nullptr, 1707 UseCycleCheck ? &PairableInstUserPairSet 1708 : nullptr)) { 1709 CanAdd = false; 1710 break; 1711 } 1712 1713 CurrentPairs.insert(*T); 1714 } 1715 if (!CanAdd) continue; 1716 1717 // And check the queue too... 1718 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(), 1719 E2 = Q.end(); C2 != E2; ++C2) { 1720 if (C2->first.first == C->first.first || 1721 C2->first.first == C->first.second || 1722 C2->first.second == C->first.first || 1723 C2->first.second == C->first.second || 1724 pairsConflict(C2->first, C->first, PairableInstUsers, 1725 UseCycleCheck ? &PairableInstUserMap : nullptr, 1726 UseCycleCheck ? &PairableInstUserPairSet 1727 : nullptr)) { 1728 CanAdd = false; 1729 break; 1730 } 1731 1732 CurrentPairs.insert(C2->first); 1733 } 1734 if (!CanAdd) continue; 1735 1736 // Last but not least, check for a conflict with any of the 1737 // already-chosen pairs. 1738 for (DenseMap<Value *, Value *>::iterator C2 = 1739 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1740 C2 != E2; ++C2) { 1741 if (pairsConflict(*C2, C->first, PairableInstUsers, 1742 UseCycleCheck ? &PairableInstUserMap : nullptr, 1743 UseCycleCheck ? &PairableInstUserPairSet 1744 : nullptr)) { 1745 CanAdd = false; 1746 break; 1747 } 1748 1749 CurrentPairs.insert(*C2); 1750 } 1751 if (!CanAdd) continue; 1752 1753 // To check for non-trivial cycles formed by the addition of the 1754 // current pair we've formed a list of all relevant pairs, now use a 1755 // graph walk to check for a cycle. We start from the current pair and 1756 // walk the use dag to see if we again reach the current pair. If we 1757 // do, then the current pair is rejected. 1758 1759 // FIXME: It may be more efficient to use a topological-ordering 1760 // algorithm to improve the cycle check. This should be investigated. 1761 if (UseCycleCheck && 1762 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1763 continue; 1764 1765 // This child can be added, but we may have chosen it in preference 1766 // to an already-selected child. Check for this here, and if a 1767 // conflict is found, then remove the previously-selected child 1768 // before adding this one in its place. 1769 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 1770 = BestChildren.begin(); C2 != BestChildren.end();) { 1771 if (C2->first.first == C->first.first || 1772 C2->first.first == C->first.second || 1773 C2->first.second == C->first.first || 1774 C2->first.second == C->first.second || 1775 pairsConflict(C2->first, C->first, PairableInstUsers)) 1776 C2 = BestChildren.erase(C2); 1777 else 1778 ++C2; 1779 } 1780 1781 BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); 1782 } 1783 1784 for (SmallVectorImpl<ValuePairWithDepth>::iterator C 1785 = BestChildren.begin(), E2 = BestChildren.end(); 1786 C != E2; ++C) { 1787 size_t DepthF = getDepthFactor(C->first.first); 1788 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1789 } 1790 } while (!Q.empty()); 1791 } 1792 1793 // This function finds the best dag of mututally-compatible connected 1794 // pairs, given the choice of root pairs as an iterator range. 1795 void BBVectorize::findBestDAGFor( 1796 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1797 DenseSet<ValuePair> &CandidatePairsSet, 1798 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1799 std::vector<Value *> &PairableInsts, 1800 DenseSet<ValuePair> &FixedOrderPairs, 1801 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1802 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1803 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 1804 DenseSet<ValuePair> &PairableInstUsers, 1805 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 1806 DenseSet<VPPair> &PairableInstUserPairSet, 1807 DenseMap<Value *, Value *> &ChosenPairs, 1808 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth, 1809 int &BestEffSize, Value *II, std::vector<Value *>&JJ, 1810 bool UseCycleCheck) { 1811 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end(); 1812 J != JE; ++J) { 1813 ValuePair IJ(II, *J); 1814 if (!CandidatePairsSet.count(IJ)) 1815 continue; 1816 1817 // Before going any further, make sure that this pair does not 1818 // conflict with any already-selected pairs (see comment below 1819 // near the DAG pruning for more details). 1820 DenseSet<ValuePair> ChosenPairSet; 1821 bool DoesConflict = false; 1822 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1823 E = ChosenPairs.end(); C != E; ++C) { 1824 if (pairsConflict(*C, IJ, PairableInstUsers, 1825 UseCycleCheck ? &PairableInstUserMap : nullptr, 1826 UseCycleCheck ? &PairableInstUserPairSet : nullptr)) { 1827 DoesConflict = true; 1828 break; 1829 } 1830 1831 ChosenPairSet.insert(*C); 1832 } 1833 if (DoesConflict) continue; 1834 1835 if (UseCycleCheck && 1836 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet)) 1837 continue; 1838 1839 DenseMap<ValuePair, size_t> DAG; 1840 buildInitialDAGFor(CandidatePairs, CandidatePairsSet, 1841 PairableInsts, ConnectedPairs, 1842 PairableInstUsers, ChosenPairs, DAG, IJ); 1843 1844 // Because we'll keep the child with the largest depth, the largest 1845 // depth is still the same in the unpruned DAG. 1846 size_t MaxDepth = DAG.lookup(IJ); 1847 1848 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {" 1849 << *IJ.first << " <-> " << *IJ.second << "} of depth " << 1850 MaxDepth << " and size " << DAG.size() << "\n"); 1851 1852 // At this point the DAG has been constructed, but, may contain 1853 // contradictory children (meaning that different children of 1854 // some dag node may be attempting to fuse the same instruction). 1855 // So now we walk the dag again, in the case of a conflict, 1856 // keep only the child with the largest depth. To break a tie, 1857 // favor the first child. 1858 1859 DenseSet<ValuePair> PrunedDAG; 1860 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs, 1861 PairableInstUsers, PairableInstUserMap, 1862 PairableInstUserPairSet, 1863 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck); 1864 1865 int EffSize = 0; 1866 if (TTI) { 1867 DenseSet<Value *> PrunedDAGInstrs; 1868 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(), 1869 E = PrunedDAG.end(); S != E; ++S) { 1870 PrunedDAGInstrs.insert(S->first); 1871 PrunedDAGInstrs.insert(S->second); 1872 } 1873 1874 // The set of pairs that have already contributed to the total cost. 1875 DenseSet<ValuePair> IncomingPairs; 1876 1877 // If the cost model were perfect, this might not be necessary; but we 1878 // need to make sure that we don't get stuck vectorizing our own 1879 // shuffle chains. 1880 bool HasNontrivialInsts = false; 1881 1882 // The node weights represent the cost savings associated with 1883 // fusing the pair of instructions. 1884 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(), 1885 E = PrunedDAG.end(); S != E; ++S) { 1886 if (!isa<ShuffleVectorInst>(S->first) && 1887 !isa<InsertElementInst>(S->first) && 1888 !isa<ExtractElementInst>(S->first)) 1889 HasNontrivialInsts = true; 1890 1891 bool FlipOrder = false; 1892 1893 if (getDepthFactor(S->first)) { 1894 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1895 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1896 << *S->first << " <-> " << *S->second << "} = " << 1897 ESContrib << "\n"); 1898 EffSize += ESContrib; 1899 } 1900 1901 // The edge weights contribute in a negative sense: they represent 1902 // the cost of shuffles. 1903 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS = 1904 ConnectedPairDeps.find(*S); 1905 if (SS != ConnectedPairDeps.end()) { 1906 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1907 for (std::vector<ValuePair>::iterator T = SS->second.begin(), 1908 TE = SS->second.end(); T != TE; ++T) { 1909 VPPair Q(*S, *T); 1910 if (!PrunedDAG.count(Q.second)) 1911 continue; 1912 DenseMap<VPPair, unsigned>::iterator R = 1913 PairConnectionTypes.find(VPPair(Q.second, Q.first)); 1914 assert(R != PairConnectionTypes.end() && 1915 "Cannot find pair connection type"); 1916 if (R->second == PairConnectionDirect) 1917 ++NumDepsDirect; 1918 else if (R->second == PairConnectionSwap) 1919 ++NumDepsSwap; 1920 } 1921 1922 // If there are more swaps than direct connections, then 1923 // the pair order will be flipped during fusion. So the real 1924 // number of swaps is the minimum number. 1925 FlipOrder = !FixedOrderPairs.count(*S) && 1926 ((NumDepsSwap > NumDepsDirect) || 1927 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1928 1929 for (std::vector<ValuePair>::iterator T = SS->second.begin(), 1930 TE = SS->second.end(); T != TE; ++T) { 1931 VPPair Q(*S, *T); 1932 if (!PrunedDAG.count(Q.second)) 1933 continue; 1934 DenseMap<VPPair, unsigned>::iterator R = 1935 PairConnectionTypes.find(VPPair(Q.second, Q.first)); 1936 assert(R != PairConnectionTypes.end() && 1937 "Cannot find pair connection type"); 1938 Type *Ty1 = Q.second.first->getType(), 1939 *Ty2 = Q.second.second->getType(); 1940 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1941 if ((R->second == PairConnectionDirect && FlipOrder) || 1942 (R->second == PairConnectionSwap && !FlipOrder) || 1943 R->second == PairConnectionSplat) { 1944 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1945 VTy, VTy); 1946 1947 if (VTy->getVectorNumElements() == 2) { 1948 if (R->second == PairConnectionSplat) 1949 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1950 TargetTransformInfo::SK_Broadcast, VTy)); 1951 else 1952 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1953 TargetTransformInfo::SK_Reverse, VTy)); 1954 } 1955 1956 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1957 *Q.second.first << " <-> " << *Q.second.second << 1958 "} -> {" << 1959 *S->first << " <-> " << *S->second << "} = " << 1960 ESContrib << "\n"); 1961 EffSize -= ESContrib; 1962 } 1963 } 1964 } 1965 1966 // Compute the cost of outgoing edges. We assume that edges outgoing 1967 // to shuffles, inserts or extracts can be merged, and so contribute 1968 // no additional cost. 1969 if (!S->first->getType()->isVoidTy()) { 1970 Type *Ty1 = S->first->getType(), 1971 *Ty2 = S->second->getType(); 1972 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1973 1974 bool NeedsExtraction = false; 1975 for (User *U : S->first->users()) { 1976 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) { 1977 // Shuffle can be folded if it has no other input 1978 if (isa<UndefValue>(SI->getOperand(1))) 1979 continue; 1980 } 1981 if (isa<ExtractElementInst>(U)) 1982 continue; 1983 if (PrunedDAGInstrs.count(U)) 1984 continue; 1985 NeedsExtraction = true; 1986 break; 1987 } 1988 1989 if (NeedsExtraction) { 1990 int ESContrib; 1991 if (Ty1->isVectorTy()) { 1992 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1993 Ty1, VTy); 1994 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1995 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1)); 1996 } else 1997 ESContrib = (int) TTI->getVectorInstrCost( 1998 Instruction::ExtractElement, VTy, 0); 1999 2000 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 2001 *S->first << "} = " << ESContrib << "\n"); 2002 EffSize -= ESContrib; 2003 } 2004 2005 NeedsExtraction = false; 2006 for (User *U : S->second->users()) { 2007 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) { 2008 // Shuffle can be folded if it has no other input 2009 if (isa<UndefValue>(SI->getOperand(1))) 2010 continue; 2011 } 2012 if (isa<ExtractElementInst>(U)) 2013 continue; 2014 if (PrunedDAGInstrs.count(U)) 2015 continue; 2016 NeedsExtraction = true; 2017 break; 2018 } 2019 2020 if (NeedsExtraction) { 2021 int ESContrib; 2022 if (Ty2->isVectorTy()) { 2023 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2024 Ty2, VTy); 2025 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 2026 TargetTransformInfo::SK_ExtractSubvector, VTy, 2027 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2)); 2028 } else 2029 ESContrib = (int) TTI->getVectorInstrCost( 2030 Instruction::ExtractElement, VTy, 1); 2031 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 2032 *S->second << "} = " << ESContrib << "\n"); 2033 EffSize -= ESContrib; 2034 } 2035 } 2036 2037 // Compute the cost of incoming edges. 2038 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 2039 Instruction *S1 = cast<Instruction>(S->first), 2040 *S2 = cast<Instruction>(S->second); 2041 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 2042 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 2043 2044 // Combining constants into vector constants (or small vector 2045 // constants into larger ones are assumed free). 2046 if (isa<Constant>(O1) && isa<Constant>(O2)) 2047 continue; 2048 2049 if (FlipOrder) 2050 std::swap(O1, O2); 2051 2052 ValuePair VP = ValuePair(O1, O2); 2053 ValuePair VPR = ValuePair(O2, O1); 2054 2055 // Internal edges are not handled here. 2056 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR)) 2057 continue; 2058 2059 Type *Ty1 = O1->getType(), 2060 *Ty2 = O2->getType(); 2061 Type *VTy = getVecTypeForPair(Ty1, Ty2); 2062 2063 // Combining vector operations of the same type is also assumed 2064 // folded with other operations. 2065 if (Ty1 == Ty2) { 2066 // If both are insert elements, then both can be widened. 2067 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), 2068 *IEO2 = dyn_cast<InsertElementInst>(O2); 2069 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) 2070 continue; 2071 // If both are extract elements, and both have the same input 2072 // type, then they can be replaced with a shuffle 2073 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 2074 *EIO2 = dyn_cast<ExtractElementInst>(O2); 2075 if (EIO1 && EIO2 && 2076 EIO1->getOperand(0)->getType() == 2077 EIO2->getOperand(0)->getType()) 2078 continue; 2079 // If both are a shuffle with equal operand types and only two 2080 // unqiue operands, then they can be replaced with a single 2081 // shuffle 2082 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 2083 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 2084 if (SIO1 && SIO2 && 2085 SIO1->getOperand(0)->getType() == 2086 SIO2->getOperand(0)->getType()) { 2087 SmallSet<Value *, 4> SIOps; 2088 SIOps.insert(SIO1->getOperand(0)); 2089 SIOps.insert(SIO1->getOperand(1)); 2090 SIOps.insert(SIO2->getOperand(0)); 2091 SIOps.insert(SIO2->getOperand(1)); 2092 if (SIOps.size() <= 2) 2093 continue; 2094 } 2095 } 2096 2097 int ESContrib; 2098 // This pair has already been formed. 2099 if (IncomingPairs.count(VP)) { 2100 continue; 2101 } else if (IncomingPairs.count(VPR)) { 2102 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2103 VTy, VTy); 2104 2105 if (VTy->getVectorNumElements() == 2) 2106 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 2107 TargetTransformInfo::SK_Reverse, VTy)); 2108 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 2109 ESContrib = (int) TTI->getVectorInstrCost( 2110 Instruction::InsertElement, VTy, 0); 2111 ESContrib += (int) TTI->getVectorInstrCost( 2112 Instruction::InsertElement, VTy, 1); 2113 } else if (!Ty1->isVectorTy()) { 2114 // O1 needs to be inserted into a vector of size O2, and then 2115 // both need to be shuffled together. 2116 ESContrib = (int) TTI->getVectorInstrCost( 2117 Instruction::InsertElement, Ty2, 0); 2118 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2119 VTy, Ty2); 2120 } else if (!Ty2->isVectorTy()) { 2121 // O2 needs to be inserted into a vector of size O1, and then 2122 // both need to be shuffled together. 2123 ESContrib = (int) TTI->getVectorInstrCost( 2124 Instruction::InsertElement, Ty1, 0); 2125 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2126 VTy, Ty1); 2127 } else { 2128 Type *TyBig = Ty1, *TySmall = Ty2; 2129 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 2130 std::swap(TyBig, TySmall); 2131 2132 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2133 VTy, TyBig); 2134 if (TyBig != TySmall) 2135 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2136 TyBig, TySmall); 2137 } 2138 2139 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 2140 << *O1 << " <-> " << *O2 << "} = " << 2141 ESContrib << "\n"); 2142 EffSize -= ESContrib; 2143 IncomingPairs.insert(VP); 2144 } 2145 } 2146 } 2147 2148 if (!HasNontrivialInsts) { 2149 DEBUG(if (DebugPairSelection) dbgs() << 2150 "\tNo non-trivial instructions in DAG;" 2151 " override to zero effective size\n"); 2152 EffSize = 0; 2153 } 2154 } else { 2155 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(), 2156 E = PrunedDAG.end(); S != E; ++S) 2157 EffSize += (int) getDepthFactor(S->first); 2158 } 2159 2160 DEBUG(if (DebugPairSelection) 2161 dbgs() << "BBV: found pruned DAG for pair {" 2162 << *IJ.first << " <-> " << *IJ.second << "} of depth " << 2163 MaxDepth << " and size " << PrunedDAG.size() << 2164 " (effective size: " << EffSize << ")\n"); 2165 if (((TTI && !UseChainDepthWithTI) || 2166 MaxDepth >= Config.ReqChainDepth) && 2167 EffSize > 0 && EffSize > BestEffSize) { 2168 BestMaxDepth = MaxDepth; 2169 BestEffSize = EffSize; 2170 BestDAG = PrunedDAG; 2171 } 2172 } 2173 } 2174 2175 // Given the list of candidate pairs, this function selects those 2176 // that will be fused into vector instructions. 2177 void BBVectorize::choosePairs( 2178 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 2179 DenseSet<ValuePair> &CandidatePairsSet, 2180 DenseMap<ValuePair, int> &CandidatePairCostSavings, 2181 std::vector<Value *> &PairableInsts, 2182 DenseSet<ValuePair> &FixedOrderPairs, 2183 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2184 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 2185 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 2186 DenseSet<ValuePair> &PairableInstUsers, 2187 DenseMap<Value *, Value *>& ChosenPairs) { 2188 bool UseCycleCheck = 2189 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck; 2190 2191 DenseMap<Value *, std::vector<Value *> > CandidatePairs2; 2192 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(), 2193 E = CandidatePairsSet.end(); I != E; ++I) { 2194 std::vector<Value *> &JJ = CandidatePairs2[I->second]; 2195 if (JJ.empty()) JJ.reserve(32); 2196 JJ.push_back(I->first); 2197 } 2198 2199 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap; 2200 DenseSet<VPPair> PairableInstUserPairSet; 2201 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 2202 E = PairableInsts.end(); I != E; ++I) { 2203 // The number of possible pairings for this variable: 2204 size_t NumChoices = CandidatePairs.lookup(*I).size(); 2205 if (!NumChoices) continue; 2206 2207 std::vector<Value *> &JJ = CandidatePairs[*I]; 2208 2209 // The best pair to choose and its dag: 2210 size_t BestMaxDepth = 0; 2211 int BestEffSize = 0; 2212 DenseSet<ValuePair> BestDAG; 2213 findBestDAGFor(CandidatePairs, CandidatePairsSet, 2214 CandidatePairCostSavings, 2215 PairableInsts, FixedOrderPairs, PairConnectionTypes, 2216 ConnectedPairs, ConnectedPairDeps, 2217 PairableInstUsers, PairableInstUserMap, 2218 PairableInstUserPairSet, ChosenPairs, 2219 BestDAG, BestMaxDepth, BestEffSize, *I, JJ, 2220 UseCycleCheck); 2221 2222 if (BestDAG.empty()) 2223 continue; 2224 2225 // A dag has been chosen (or not) at this point. If no dag was 2226 // chosen, then this instruction, I, cannot be paired (and is no longer 2227 // considered). 2228 2229 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: " 2230 << *cast<Instruction>(*I) << "\n"); 2231 2232 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(), 2233 SE2 = BestDAG.end(); S != SE2; ++S) { 2234 // Insert the members of this dag into the list of chosen pairs. 2235 ChosenPairs.insert(ValuePair(S->first, S->second)); 2236 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 2237 *S->second << "\n"); 2238 2239 // Remove all candidate pairs that have values in the chosen dag. 2240 std::vector<Value *> &KK = CandidatePairs[S->first]; 2241 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end(); 2242 K != KE; ++K) { 2243 if (*K == S->second) 2244 continue; 2245 2246 CandidatePairsSet.erase(ValuePair(S->first, *K)); 2247 } 2248 2249 std::vector<Value *> &LL = CandidatePairs2[S->second]; 2250 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end(); 2251 L != LE; ++L) { 2252 if (*L == S->first) 2253 continue; 2254 2255 CandidatePairsSet.erase(ValuePair(*L, S->second)); 2256 } 2257 2258 std::vector<Value *> &MM = CandidatePairs[S->second]; 2259 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end(); 2260 M != ME; ++M) { 2261 assert(*M != S->first && "Flipped pair in candidate list?"); 2262 CandidatePairsSet.erase(ValuePair(S->second, *M)); 2263 } 2264 2265 std::vector<Value *> &NN = CandidatePairs2[S->first]; 2266 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end(); 2267 N != NE; ++N) { 2268 assert(*N != S->second && "Flipped pair in candidate list?"); 2269 CandidatePairsSet.erase(ValuePair(*N, S->first)); 2270 } 2271 } 2272 } 2273 2274 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2275 } 2276 2277 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2278 unsigned n = 0) { 2279 if (!I->hasName()) 2280 return ""; 2281 2282 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2283 (n > 0 ? "." + utostr(n) : "")).str(); 2284 } 2285 2286 // Returns the value that is to be used as the pointer input to the vector 2287 // instruction that fuses I with J. 2288 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2289 Instruction *I, Instruction *J, unsigned o) { 2290 Value *IPtr, *JPtr; 2291 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2292 int64_t OffsetInElmts; 2293 2294 // Note: the analysis might fail here, that is why the pair order has 2295 // been precomputed (OffsetInElmts must be unused here). 2296 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2297 IAddressSpace, JAddressSpace, 2298 OffsetInElmts, false); 2299 2300 // The pointer value is taken to be the one with the lowest offset. 2301 Value *VPtr = IPtr; 2302 2303 Type *ArgTypeI = IPtr->getType()->getPointerElementType(); 2304 Type *ArgTypeJ = JPtr->getType()->getPointerElementType(); 2305 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2306 Type *VArgPtrType 2307 = PointerType::get(VArgType, 2308 IPtr->getType()->getPointerAddressSpace()); 2309 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2310 /* insert before */ I); 2311 } 2312 2313 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2314 unsigned MaskOffset, unsigned NumInElem, 2315 unsigned NumInElem1, unsigned IdxOffset, 2316 std::vector<Constant*> &Mask) { 2317 unsigned NumElem1 = J->getType()->getVectorNumElements(); 2318 for (unsigned v = 0; v < NumElem1; ++v) { 2319 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2320 if (m < 0) { 2321 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2322 } else { 2323 unsigned mm = m + (int) IdxOffset; 2324 if (m >= (int) NumInElem1) 2325 mm += (int) NumInElem; 2326 2327 Mask[v+MaskOffset] = 2328 ConstantInt::get(Type::getInt32Ty(Context), mm); 2329 } 2330 } 2331 } 2332 2333 // Returns the value that is to be used as the vector-shuffle mask to the 2334 // vector instruction that fuses I with J. 2335 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2336 Instruction *I, Instruction *J) { 2337 // This is the shuffle mask. We need to append the second 2338 // mask to the first, and the numbers need to be adjusted. 2339 2340 Type *ArgTypeI = I->getType(); 2341 Type *ArgTypeJ = J->getType(); 2342 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2343 2344 unsigned NumElemI = ArgTypeI->getVectorNumElements(); 2345 2346 // Get the total number of elements in the fused vector type. 2347 // By definition, this must equal the number of elements in 2348 // the final mask. 2349 unsigned NumElem = VArgType->getVectorNumElements(); 2350 std::vector<Constant*> Mask(NumElem); 2351 2352 Type *OpTypeI = I->getOperand(0)->getType(); 2353 unsigned NumInElemI = OpTypeI->getVectorNumElements(); 2354 Type *OpTypeJ = J->getOperand(0)->getType(); 2355 unsigned NumInElemJ = OpTypeJ->getVectorNumElements(); 2356 2357 // The fused vector will be: 2358 // ----------------------------------------------------- 2359 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2360 // ----------------------------------------------------- 2361 // from which we'll extract NumElem total elements (where the first NumElemI 2362 // of them come from the mask in I and the remainder come from the mask 2363 // in J. 2364 2365 // For the mask from the first pair... 2366 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2367 0, Mask); 2368 2369 // For the mask from the second pair... 2370 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2371 NumInElemI, Mask); 2372 2373 return ConstantVector::get(Mask); 2374 } 2375 2376 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2377 Instruction *J, unsigned o, Value *&LOp, 2378 unsigned numElemL, 2379 Type *ArgTypeL, Type *ArgTypeH, 2380 bool IBeforeJ, unsigned IdxOff) { 2381 bool ExpandedIEChain = false; 2382 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2383 // If we have a pure insertelement chain, then this can be rewritten 2384 // into a chain that directly builds the larger type. 2385 if (isPureIEChain(LIE)) { 2386 SmallVector<Value *, 8> VectElemts(numElemL, 2387 UndefValue::get(ArgTypeL->getScalarType())); 2388 InsertElementInst *LIENext = LIE; 2389 do { 2390 unsigned Idx = 2391 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2392 VectElemts[Idx] = LIENext->getOperand(1); 2393 } while ((LIENext = 2394 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2395 2396 LIENext = nullptr; 2397 Value *LIEPrev = UndefValue::get(ArgTypeH); 2398 for (unsigned i = 0; i < numElemL; ++i) { 2399 if (isa<UndefValue>(VectElemts[i])) continue; 2400 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2401 ConstantInt::get(Type::getInt32Ty(Context), 2402 i + IdxOff), 2403 getReplacementName(IBeforeJ ? I : J, 2404 true, o, i+1)); 2405 LIENext->insertBefore(IBeforeJ ? J : I); 2406 LIEPrev = LIENext; 2407 } 2408 2409 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2410 ExpandedIEChain = true; 2411 } 2412 } 2413 2414 return ExpandedIEChain; 2415 } 2416 2417 static unsigned getNumScalarElements(Type *Ty) { 2418 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) 2419 return VecTy->getNumElements(); 2420 return 1; 2421 } 2422 2423 // Returns the value to be used as the specified operand of the vector 2424 // instruction that fuses I with J. 2425 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2426 Instruction *J, unsigned o, bool IBeforeJ) { 2427 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2428 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2429 2430 // Compute the fused vector type for this operand 2431 Type *ArgTypeI = I->getOperand(o)->getType(); 2432 Type *ArgTypeJ = J->getOperand(o)->getType(); 2433 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2434 2435 Instruction *L = I, *H = J; 2436 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2437 2438 unsigned numElemL = getNumScalarElements(ArgTypeL); 2439 unsigned numElemH = getNumScalarElements(ArgTypeH); 2440 2441 Value *LOp = L->getOperand(o); 2442 Value *HOp = H->getOperand(o); 2443 unsigned numElem = VArgType->getNumElements(); 2444 2445 // First, we check if we can reuse the "original" vector outputs (if these 2446 // exist). We might need a shuffle. 2447 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2448 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2449 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2450 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2451 2452 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2453 // optimization. The input vectors to the shuffle might be a different 2454 // length from the shuffle outputs. Unfortunately, the replacement 2455 // shuffle mask has already been formed, and the mask entries are sensitive 2456 // to the sizes of the inputs. 2457 bool IsSizeChangeShuffle = 2458 isa<ShuffleVectorInst>(L) && 2459 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2460 2461 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2462 // We can have at most two unique vector inputs. 2463 bool CanUseInputs = true; 2464 Value *I1, *I2 = nullptr; 2465 if (LEE) { 2466 I1 = LEE->getOperand(0); 2467 } else { 2468 I1 = LSV->getOperand(0); 2469 I2 = LSV->getOperand(1); 2470 if (I2 == I1 || isa<UndefValue>(I2)) 2471 I2 = nullptr; 2472 } 2473 2474 if (HEE) { 2475 Value *I3 = HEE->getOperand(0); 2476 if (!I2 && I3 != I1) 2477 I2 = I3; 2478 else if (I3 != I1 && I3 != I2) 2479 CanUseInputs = false; 2480 } else { 2481 Value *I3 = HSV->getOperand(0); 2482 if (!I2 && I3 != I1) 2483 I2 = I3; 2484 else if (I3 != I1 && I3 != I2) 2485 CanUseInputs = false; 2486 2487 if (CanUseInputs) { 2488 Value *I4 = HSV->getOperand(1); 2489 if (!isa<UndefValue>(I4)) { 2490 if (!I2 && I4 != I1) 2491 I2 = I4; 2492 else if (I4 != I1 && I4 != I2) 2493 CanUseInputs = false; 2494 } 2495 } 2496 } 2497 2498 if (CanUseInputs) { 2499 unsigned LOpElem = 2500 cast<Instruction>(LOp)->getOperand(0)->getType() 2501 ->getVectorNumElements(); 2502 2503 unsigned HOpElem = 2504 cast<Instruction>(HOp)->getOperand(0)->getType() 2505 ->getVectorNumElements(); 2506 2507 // We have one or two input vectors. We need to map each index of the 2508 // operands to the index of the original vector. 2509 SmallVector<std::pair<int, int>, 8> II(numElem); 2510 for (unsigned i = 0; i < numElemL; ++i) { 2511 int Idx, INum; 2512 if (LEE) { 2513 Idx = 2514 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2515 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2516 } else { 2517 Idx = LSV->getMaskValue(i); 2518 if (Idx < (int) LOpElem) { 2519 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2520 } else { 2521 Idx -= LOpElem; 2522 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2523 } 2524 } 2525 2526 II[i] = std::pair<int, int>(Idx, INum); 2527 } 2528 for (unsigned i = 0; i < numElemH; ++i) { 2529 int Idx, INum; 2530 if (HEE) { 2531 Idx = 2532 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2533 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2534 } else { 2535 Idx = HSV->getMaskValue(i); 2536 if (Idx < (int) HOpElem) { 2537 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2538 } else { 2539 Idx -= HOpElem; 2540 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2541 } 2542 } 2543 2544 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2545 } 2546 2547 // We now have an array which tells us from which index of which 2548 // input vector each element of the operand comes. 2549 VectorType *I1T = cast<VectorType>(I1->getType()); 2550 unsigned I1Elem = I1T->getNumElements(); 2551 2552 if (!I2) { 2553 // In this case there is only one underlying vector input. Check for 2554 // the trivial case where we can use the input directly. 2555 if (I1Elem == numElem) { 2556 bool ElemInOrder = true; 2557 for (unsigned i = 0; i < numElem; ++i) { 2558 if (II[i].first != (int) i && II[i].first != -1) { 2559 ElemInOrder = false; 2560 break; 2561 } 2562 } 2563 2564 if (ElemInOrder) 2565 return I1; 2566 } 2567 2568 // A shuffle is needed. 2569 std::vector<Constant *> Mask(numElem); 2570 for (unsigned i = 0; i < numElem; ++i) { 2571 int Idx = II[i].first; 2572 if (Idx == -1) 2573 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2574 else 2575 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2576 } 2577 2578 Instruction *S = 2579 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2580 ConstantVector::get(Mask), 2581 getReplacementName(IBeforeJ ? I : J, 2582 true, o)); 2583 S->insertBefore(IBeforeJ ? J : I); 2584 return S; 2585 } 2586 2587 VectorType *I2T = cast<VectorType>(I2->getType()); 2588 unsigned I2Elem = I2T->getNumElements(); 2589 2590 // This input comes from two distinct vectors. The first step is to 2591 // make sure that both vectors are the same length. If not, the 2592 // smaller one will need to grow before they can be shuffled together. 2593 if (I1Elem < I2Elem) { 2594 std::vector<Constant *> Mask(I2Elem); 2595 unsigned v = 0; 2596 for (; v < I1Elem; ++v) 2597 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2598 for (; v < I2Elem; ++v) 2599 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2600 2601 Instruction *NewI1 = 2602 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2603 ConstantVector::get(Mask), 2604 getReplacementName(IBeforeJ ? I : J, 2605 true, o, 1)); 2606 NewI1->insertBefore(IBeforeJ ? J : I); 2607 I1 = NewI1; 2608 I1Elem = I2Elem; 2609 } else if (I1Elem > I2Elem) { 2610 std::vector<Constant *> Mask(I1Elem); 2611 unsigned v = 0; 2612 for (; v < I2Elem; ++v) 2613 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2614 for (; v < I1Elem; ++v) 2615 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2616 2617 Instruction *NewI2 = 2618 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2619 ConstantVector::get(Mask), 2620 getReplacementName(IBeforeJ ? I : J, 2621 true, o, 1)); 2622 NewI2->insertBefore(IBeforeJ ? J : I); 2623 I2 = NewI2; 2624 } 2625 2626 // Now that both I1 and I2 are the same length we can shuffle them 2627 // together (and use the result). 2628 std::vector<Constant *> Mask(numElem); 2629 for (unsigned v = 0; v < numElem; ++v) { 2630 if (II[v].first == -1) { 2631 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2632 } else { 2633 int Idx = II[v].first + II[v].second * I1Elem; 2634 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2635 } 2636 } 2637 2638 Instruction *NewOp = 2639 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2640 getReplacementName(IBeforeJ ? I : J, true, o)); 2641 NewOp->insertBefore(IBeforeJ ? J : I); 2642 return NewOp; 2643 } 2644 } 2645 2646 Type *ArgType = ArgTypeL; 2647 if (numElemL < numElemH) { 2648 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2649 ArgTypeL, VArgType, IBeforeJ, 1)) { 2650 // This is another short-circuit case: we're combining a scalar into 2651 // a vector that is formed by an IE chain. We've just expanded the IE 2652 // chain, now insert the scalar and we're done. 2653 2654 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2655 getReplacementName(IBeforeJ ? I : J, true, o)); 2656 S->insertBefore(IBeforeJ ? J : I); 2657 return S; 2658 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2659 ArgTypeH, IBeforeJ)) { 2660 // The two vector inputs to the shuffle must be the same length, 2661 // so extend the smaller vector to be the same length as the larger one. 2662 Instruction *NLOp; 2663 if (numElemL > 1) { 2664 2665 std::vector<Constant *> Mask(numElemH); 2666 unsigned v = 0; 2667 for (; v < numElemL; ++v) 2668 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2669 for (; v < numElemH; ++v) 2670 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2671 2672 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2673 ConstantVector::get(Mask), 2674 getReplacementName(IBeforeJ ? I : J, 2675 true, o, 1)); 2676 } else { 2677 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2678 getReplacementName(IBeforeJ ? I : J, 2679 true, o, 1)); 2680 } 2681 2682 NLOp->insertBefore(IBeforeJ ? J : I); 2683 LOp = NLOp; 2684 } 2685 2686 ArgType = ArgTypeH; 2687 } else if (numElemL > numElemH) { 2688 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2689 ArgTypeH, VArgType, IBeforeJ)) { 2690 Instruction *S = 2691 InsertElementInst::Create(LOp, HOp, 2692 ConstantInt::get(Type::getInt32Ty(Context), 2693 numElemL), 2694 getReplacementName(IBeforeJ ? I : J, 2695 true, o)); 2696 S->insertBefore(IBeforeJ ? J : I); 2697 return S; 2698 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2699 ArgTypeL, IBeforeJ)) { 2700 Instruction *NHOp; 2701 if (numElemH > 1) { 2702 std::vector<Constant *> Mask(numElemL); 2703 unsigned v = 0; 2704 for (; v < numElemH; ++v) 2705 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2706 for (; v < numElemL; ++v) 2707 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2708 2709 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2710 ConstantVector::get(Mask), 2711 getReplacementName(IBeforeJ ? I : J, 2712 true, o, 1)); 2713 } else { 2714 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2715 getReplacementName(IBeforeJ ? I : J, 2716 true, o, 1)); 2717 } 2718 2719 NHOp->insertBefore(IBeforeJ ? J : I); 2720 HOp = NHOp; 2721 } 2722 } 2723 2724 if (ArgType->isVectorTy()) { 2725 unsigned numElem = VArgType->getVectorNumElements(); 2726 std::vector<Constant*> Mask(numElem); 2727 for (unsigned v = 0; v < numElem; ++v) { 2728 unsigned Idx = v; 2729 // If the low vector was expanded, we need to skip the extra 2730 // undefined entries. 2731 if (v >= numElemL && numElemH > numElemL) 2732 Idx += (numElemH - numElemL); 2733 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2734 } 2735 2736 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2737 ConstantVector::get(Mask), 2738 getReplacementName(IBeforeJ ? I : J, true, o)); 2739 BV->insertBefore(IBeforeJ ? J : I); 2740 return BV; 2741 } 2742 2743 Instruction *BV1 = InsertElementInst::Create( 2744 UndefValue::get(VArgType), LOp, CV0, 2745 getReplacementName(IBeforeJ ? I : J, 2746 true, o, 1)); 2747 BV1->insertBefore(IBeforeJ ? J : I); 2748 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2749 getReplacementName(IBeforeJ ? I : J, 2750 true, o, 2)); 2751 BV2->insertBefore(IBeforeJ ? J : I); 2752 return BV2; 2753 } 2754 2755 // This function creates an array of values that will be used as the inputs 2756 // to the vector instruction that fuses I with J. 2757 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2758 Instruction *I, Instruction *J, 2759 SmallVectorImpl<Value *> &ReplacedOperands, 2760 bool IBeforeJ) { 2761 unsigned NumOperands = I->getNumOperands(); 2762 2763 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2764 // Iterate backward so that we look at the store pointer 2765 // first and know whether or not we need to flip the inputs. 2766 2767 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2768 // This is the pointer for a load/store instruction. 2769 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2770 continue; 2771 } else if (isa<CallInst>(I)) { 2772 Function *F = cast<CallInst>(I)->getCalledFunction(); 2773 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 2774 if (o == NumOperands-1) { 2775 BasicBlock &BB = *I->getParent(); 2776 2777 Module *M = BB.getParent()->getParent(); 2778 Type *ArgTypeI = I->getType(); 2779 Type *ArgTypeJ = J->getType(); 2780 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2781 2782 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType); 2783 continue; 2784 } else if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz || 2785 IID == Intrinsic::cttz) && o == 1) { 2786 // The second argument of powi/ctlz/cttz is a single integer/constant 2787 // and we've already checked that both arguments are equal. 2788 // As a result, we just keep I's second argument. 2789 ReplacedOperands[o] = I->getOperand(o); 2790 continue; 2791 } 2792 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2793 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2794 continue; 2795 } 2796 2797 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2798 } 2799 } 2800 2801 // This function creates two values that represent the outputs of the 2802 // original I and J instructions. These are generally vector shuffles 2803 // or extracts. In many cases, these will end up being unused and, thus, 2804 // eliminated by later passes. 2805 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2806 Instruction *J, Instruction *K, 2807 Instruction *&InsertionPt, 2808 Instruction *&K1, Instruction *&K2) { 2809 if (isa<StoreInst>(I)) { 2810 AA->replaceWithNewValue(I, K); 2811 AA->replaceWithNewValue(J, K); 2812 } else { 2813 Type *IType = I->getType(); 2814 Type *JType = J->getType(); 2815 2816 VectorType *VType = getVecTypeForPair(IType, JType); 2817 unsigned numElem = VType->getNumElements(); 2818 2819 unsigned numElemI = getNumScalarElements(IType); 2820 unsigned numElemJ = getNumScalarElements(JType); 2821 2822 if (IType->isVectorTy()) { 2823 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2824 for (unsigned v = 0; v < numElemI; ++v) { 2825 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2826 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2827 } 2828 2829 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2830 ConstantVector::get( Mask1), 2831 getReplacementName(K, false, 1)); 2832 } else { 2833 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2834 K1 = ExtractElementInst::Create(K, CV0, 2835 getReplacementName(K, false, 1)); 2836 } 2837 2838 if (JType->isVectorTy()) { 2839 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2840 for (unsigned v = 0; v < numElemJ; ++v) { 2841 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2842 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2843 } 2844 2845 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2846 ConstantVector::get( Mask2), 2847 getReplacementName(K, false, 2)); 2848 } else { 2849 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2850 K2 = ExtractElementInst::Create(K, CV1, 2851 getReplacementName(K, false, 2)); 2852 } 2853 2854 K1->insertAfter(K); 2855 K2->insertAfter(K1); 2856 InsertionPt = K2; 2857 } 2858 } 2859 2860 // Move all uses of the function I (including pairing-induced uses) after J. 2861 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2862 DenseSet<ValuePair> &LoadMoveSetPairs, 2863 Instruction *I, Instruction *J) { 2864 // Skip to the first instruction past I. 2865 BasicBlock::iterator L = std::next(BasicBlock::iterator(I)); 2866 2867 DenseSet<Value *> Users; 2868 AliasSetTracker WriteSet(*AA); 2869 if (I->mayWriteToMemory()) WriteSet.add(I); 2870 2871 for (; cast<Instruction>(L) != J; ++L) 2872 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs); 2873 2874 assert(cast<Instruction>(L) == J && 2875 "Tracking has not proceeded far enough to check for dependencies"); 2876 // If J is now in the use set of I, then trackUsesOfI will return true 2877 // and we have a dependency cycle (and the fusing operation must abort). 2878 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs); 2879 } 2880 2881 // Move all uses of the function I (including pairing-induced uses) after J. 2882 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2883 DenseSet<ValuePair> &LoadMoveSetPairs, 2884 Instruction *&InsertionPt, 2885 Instruction *I, Instruction *J) { 2886 // Skip to the first instruction past I. 2887 BasicBlock::iterator L = std::next(BasicBlock::iterator(I)); 2888 2889 DenseSet<Value *> Users; 2890 AliasSetTracker WriteSet(*AA); 2891 if (I->mayWriteToMemory()) WriteSet.add(I); 2892 2893 for (; cast<Instruction>(L) != J;) { 2894 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) { 2895 // Move this instruction 2896 Instruction *InstToMove = L; ++L; 2897 2898 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2899 " to after " << *InsertionPt << "\n"); 2900 InstToMove->removeFromParent(); 2901 InstToMove->insertAfter(InsertionPt); 2902 InsertionPt = InstToMove; 2903 } else { 2904 ++L; 2905 } 2906 } 2907 } 2908 2909 // Collect all load instruction that are in the move set of a given first 2910 // pair member. These loads depend on the first instruction, I, and so need 2911 // to be moved after J (the second instruction) when the pair is fused. 2912 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2913 DenseMap<Value *, Value *> &ChosenPairs, 2914 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 2915 DenseSet<ValuePair> &LoadMoveSetPairs, 2916 Instruction *I) { 2917 // Skip to the first instruction past I. 2918 BasicBlock::iterator L = std::next(BasicBlock::iterator(I)); 2919 2920 DenseSet<Value *> Users; 2921 AliasSetTracker WriteSet(*AA); 2922 if (I->mayWriteToMemory()) WriteSet.add(I); 2923 2924 // Note: We cannot end the loop when we reach J because J could be moved 2925 // farther down the use chain by another instruction pairing. Also, J 2926 // could be before I if this is an inverted input. 2927 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2928 if (trackUsesOfI(Users, WriteSet, I, L)) { 2929 if (L->mayReadFromMemory()) { 2930 LoadMoveSet[L].push_back(I); 2931 LoadMoveSetPairs.insert(ValuePair(L, I)); 2932 } 2933 } 2934 } 2935 } 2936 2937 // In cases where both load/stores and the computation of their pointers 2938 // are chosen for vectorization, we can end up in a situation where the 2939 // aliasing analysis starts returning different query results as the 2940 // process of fusing instruction pairs continues. Because the algorithm 2941 // relies on finding the same use dags here as were found earlier, we'll 2942 // need to precompute the necessary aliasing information here and then 2943 // manually update it during the fusion process. 2944 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2945 std::vector<Value *> &PairableInsts, 2946 DenseMap<Value *, Value *> &ChosenPairs, 2947 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 2948 DenseSet<ValuePair> &LoadMoveSetPairs) { 2949 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2950 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2951 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2952 if (P == ChosenPairs.end()) continue; 2953 2954 Instruction *I = cast<Instruction>(P->first); 2955 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, 2956 LoadMoveSetPairs, I); 2957 } 2958 } 2959 2960 // This function fuses the chosen instruction pairs into vector instructions, 2961 // taking care preserve any needed scalar outputs and, then, it reorders the 2962 // remaining instructions as needed (users of the first member of the pair 2963 // need to be moved to after the location of the second member of the pair 2964 // because the vector instruction is inserted in the location of the pair's 2965 // second member). 2966 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2967 std::vector<Value *> &PairableInsts, 2968 DenseMap<Value *, Value *> &ChosenPairs, 2969 DenseSet<ValuePair> &FixedOrderPairs, 2970 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2971 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 2972 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) { 2973 LLVMContext& Context = BB.getContext(); 2974 2975 // During the vectorization process, the order of the pairs to be fused 2976 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2977 // list. After a pair is fused, the flipped pair is removed from the list. 2978 DenseSet<ValuePair> FlippedPairs; 2979 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2980 E = ChosenPairs.end(); P != E; ++P) 2981 FlippedPairs.insert(ValuePair(P->second, P->first)); 2982 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2983 E = FlippedPairs.end(); P != E; ++P) 2984 ChosenPairs.insert(*P); 2985 2986 DenseMap<Value *, std::vector<Value *> > LoadMoveSet; 2987 DenseSet<ValuePair> LoadMoveSetPairs; 2988 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, 2989 LoadMoveSet, LoadMoveSetPairs); 2990 2991 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2992 2993 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2994 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2995 if (P == ChosenPairs.end()) { 2996 ++PI; 2997 continue; 2998 } 2999 3000 if (getDepthFactor(P->first) == 0) { 3001 // These instructions are not really fused, but are tracked as though 3002 // they are. Any case in which it would be interesting to fuse them 3003 // will be taken care of by InstCombine. 3004 --NumFusedOps; 3005 ++PI; 3006 continue; 3007 } 3008 3009 Instruction *I = cast<Instruction>(P->first), 3010 *J = cast<Instruction>(P->second); 3011 3012 DEBUG(dbgs() << "BBV: fusing: " << *I << 3013 " <-> " << *J << "\n"); 3014 3015 // Remove the pair and flipped pair from the list. 3016 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 3017 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 3018 ChosenPairs.erase(FP); 3019 ChosenPairs.erase(P); 3020 3021 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) { 3022 DEBUG(dbgs() << "BBV: fusion of: " << *I << 3023 " <-> " << *J << 3024 " aborted because of non-trivial dependency cycle\n"); 3025 --NumFusedOps; 3026 ++PI; 3027 continue; 3028 } 3029 3030 // If the pair must have the other order, then flip it. 3031 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 3032 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 3033 // This pair does not have a fixed order, and so we might want to 3034 // flip it if that will yield fewer shuffles. We count the number 3035 // of dependencies connected via swaps, and those directly connected, 3036 // and flip the order if the number of swaps is greater. 3037 bool OrigOrder = true; 3038 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ = 3039 ConnectedPairDeps.find(ValuePair(I, J)); 3040 if (IJ == ConnectedPairDeps.end()) { 3041 IJ = ConnectedPairDeps.find(ValuePair(J, I)); 3042 OrigOrder = false; 3043 } 3044 3045 if (IJ != ConnectedPairDeps.end()) { 3046 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 3047 for (std::vector<ValuePair>::iterator T = IJ->second.begin(), 3048 TE = IJ->second.end(); T != TE; ++T) { 3049 VPPair Q(IJ->first, *T); 3050 DenseMap<VPPair, unsigned>::iterator R = 3051 PairConnectionTypes.find(VPPair(Q.second, Q.first)); 3052 assert(R != PairConnectionTypes.end() && 3053 "Cannot find pair connection type"); 3054 if (R->second == PairConnectionDirect) 3055 ++NumDepsDirect; 3056 else if (R->second == PairConnectionSwap) 3057 ++NumDepsSwap; 3058 } 3059 3060 if (!OrigOrder) 3061 std::swap(NumDepsDirect, NumDepsSwap); 3062 3063 if (NumDepsSwap > NumDepsDirect) { 3064 FlipPairOrder = true; 3065 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 3066 " <-> " << *J << "\n"); 3067 } 3068 } 3069 } 3070 3071 Instruction *L = I, *H = J; 3072 if (FlipPairOrder) 3073 std::swap(H, L); 3074 3075 // If the pair being fused uses the opposite order from that in the pair 3076 // connection map, then we need to flip the types. 3077 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL = 3078 ConnectedPairs.find(ValuePair(H, L)); 3079 if (HL != ConnectedPairs.end()) 3080 for (std::vector<ValuePair>::iterator T = HL->second.begin(), 3081 TE = HL->second.end(); T != TE; ++T) { 3082 VPPair Q(HL->first, *T); 3083 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q); 3084 assert(R != PairConnectionTypes.end() && 3085 "Cannot find pair connection type"); 3086 if (R->second == PairConnectionDirect) 3087 R->second = PairConnectionSwap; 3088 else if (R->second == PairConnectionSwap) 3089 R->second = PairConnectionDirect; 3090 } 3091 3092 bool LBeforeH = !FlipPairOrder; 3093 unsigned NumOperands = I->getNumOperands(); 3094 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 3095 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 3096 LBeforeH); 3097 3098 // Make a copy of the original operation, change its type to the vector 3099 // type and replace its operands with the vector operands. 3100 Instruction *K = L->clone(); 3101 if (L->hasName()) 3102 K->takeName(L); 3103 else if (H->hasName()) 3104 K->takeName(H); 3105 3106 if (!isa<StoreInst>(K)) 3107 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 3108 3109 unsigned KnownIDs[] = { 3110 LLVMContext::MD_tbaa, 3111 LLVMContext::MD_alias_scope, 3112 LLVMContext::MD_noalias, 3113 LLVMContext::MD_fpmath 3114 }; 3115 combineMetadata(K, H, KnownIDs); 3116 K->intersectOptionalDataWith(H); 3117 3118 for (unsigned o = 0; o < NumOperands; ++o) 3119 K->setOperand(o, ReplacedOperands[o]); 3120 3121 K->insertAfter(J); 3122 3123 // Instruction insertion point: 3124 Instruction *InsertionPt = K; 3125 Instruction *K1 = nullptr, *K2 = nullptr; 3126 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 3127 3128 // The use dag of the first original instruction must be moved to after 3129 // the location of the second instruction. The entire use dag of the 3130 // first instruction is disjoint from the input dag of the second 3131 // (by definition), and so commutes with it. 3132 3133 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J); 3134 3135 if (!isa<StoreInst>(I)) { 3136 L->replaceAllUsesWith(K1); 3137 H->replaceAllUsesWith(K2); 3138 AA->replaceWithNewValue(L, K1); 3139 AA->replaceWithNewValue(H, K2); 3140 } 3141 3142 // Instructions that may read from memory may be in the load move set. 3143 // Once an instruction is fused, we no longer need its move set, and so 3144 // the values of the map never need to be updated. However, when a load 3145 // is fused, we need to merge the entries from both instructions in the 3146 // pair in case those instructions were in the move set of some other 3147 // yet-to-be-fused pair. The loads in question are the keys of the map. 3148 if (I->mayReadFromMemory()) { 3149 std::vector<ValuePair> NewSetMembers; 3150 DenseMap<Value *, std::vector<Value *> >::iterator II = 3151 LoadMoveSet.find(I); 3152 if (II != LoadMoveSet.end()) 3153 for (std::vector<Value *>::iterator N = II->second.begin(), 3154 NE = II->second.end(); N != NE; ++N) 3155 NewSetMembers.push_back(ValuePair(K, *N)); 3156 DenseMap<Value *, std::vector<Value *> >::iterator JJ = 3157 LoadMoveSet.find(J); 3158 if (JJ != LoadMoveSet.end()) 3159 for (std::vector<Value *>::iterator N = JJ->second.begin(), 3160 NE = JJ->second.end(); N != NE; ++N) 3161 NewSetMembers.push_back(ValuePair(K, *N)); 3162 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 3163 AE = NewSetMembers.end(); A != AE; ++A) { 3164 LoadMoveSet[A->first].push_back(A->second); 3165 LoadMoveSetPairs.insert(*A); 3166 } 3167 } 3168 3169 // Before removing I, set the iterator to the next instruction. 3170 PI = std::next(BasicBlock::iterator(I)); 3171 if (cast<Instruction>(PI) == J) 3172 ++PI; 3173 3174 SE->forgetValue(I); 3175 SE->forgetValue(J); 3176 I->eraseFromParent(); 3177 J->eraseFromParent(); 3178 3179 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 3180 BB << "\n"); 3181 } 3182 3183 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 3184 } 3185 } 3186 3187 char BBVectorize::ID = 0; 3188 static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 3189 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3190 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3191 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 3192 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 3193 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3194 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3195 3196 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 3197 return new BBVectorize(C); 3198 } 3199 3200 bool 3201 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 3202 BBVectorize BBVectorizer(P, *BB.getParent(), C); 3203 return BBVectorizer.vectorizeBB(BB); 3204 } 3205 3206 //===----------------------------------------------------------------------===// 3207 VectorizeConfig::VectorizeConfig() { 3208 VectorBits = ::VectorBits; 3209 VectorizeBools = !::NoBools; 3210 VectorizeInts = !::NoInts; 3211 VectorizeFloats = !::NoFloats; 3212 VectorizePointers = !::NoPointers; 3213 VectorizeCasts = !::NoCasts; 3214 VectorizeMath = !::NoMath; 3215 VectorizeBitManipulations = !::NoBitManipulation; 3216 VectorizeFMA = !::NoFMA; 3217 VectorizeSelect = !::NoSelect; 3218 VectorizeCmp = !::NoCmp; 3219 VectorizeGEP = !::NoGEP; 3220 VectorizeMemOps = !::NoMemOps; 3221 AlignedOnly = ::AlignedOnly; 3222 ReqChainDepth= ::ReqChainDepth; 3223 SearchLimit = ::SearchLimit; 3224 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 3225 SplatBreaksChain = ::SplatBreaksChain; 3226 MaxInsts = ::MaxInsts; 3227 MaxPairs = ::MaxPairs; 3228 MaxIter = ::MaxIter; 3229 Pow2LenOnly = ::Pow2LenOnly; 3230 NoMemOpBoost = ::NoMemOpBoost; 3231 FastDep = ::FastDep; 3232 } 3233