BBVectorize.cpp revision 86ccc55c82651f91fd6a312c5f6a4b511bcd1aec
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#define DEBUG_TYPE BBV_NAME 19#include "llvm/Constants.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Function.h" 22#include "llvm/Instructions.h" 23#include "llvm/IntrinsicInst.h" 24#include "llvm/Intrinsics.h" 25#include "llvm/LLVMContext.h" 26#include "llvm/Metadata.h" 27#include "llvm/Pass.h" 28#include "llvm/Type.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/DenseSet.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/ADT/STLExtras.h" 34#include "llvm/ADT/StringExtras.h" 35#include "llvm/Analysis/AliasAnalysis.h" 36#include "llvm/Analysis/AliasSetTracker.h" 37#include "llvm/Analysis/Dominators.h" 38#include "llvm/Analysis/ScalarEvolution.h" 39#include "llvm/Analysis/ScalarEvolutionExpressions.h" 40#include "llvm/Analysis/ValueTracking.h" 41#include "llvm/Support/CommandLine.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/raw_ostream.h" 44#include "llvm/Support/ValueHandle.h" 45#include "llvm/DataLayout.h" 46#include "llvm/TargetTransformInfo.h" 47#include "llvm/Transforms/Utils/Local.h" 48#include "llvm/Transforms/Vectorize.h" 49#include <algorithm> 50#include <map> 51using namespace llvm; 52 53static cl::opt<bool> 54IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false), 55 cl::Hidden, cl::desc("Ignore target information")); 56 57static cl::opt<unsigned> 58ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 59 cl::desc("The required chain depth for vectorization")); 60 61static cl::opt<unsigned> 62SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 63 cl::desc("The maximum search distance for instruction pairs")); 64 65static cl::opt<bool> 66SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 67 cl::desc("Replicating one element to a pair breaks the chain")); 68 69static cl::opt<unsigned> 70VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 71 cl::desc("The size of the native vector registers")); 72 73static cl::opt<unsigned> 74MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 75 cl::desc("The maximum number of pairing iterations")); 76 77static cl::opt<bool> 78Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden, 79 cl::desc("Don't try to form non-2^n-length vectors")); 80 81static cl::opt<unsigned> 82MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 83 cl::desc("The maximum number of pairable instructions per group")); 84 85static cl::opt<unsigned> 86MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 87 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 88 " a full cycle check")); 89 90static cl::opt<bool> 91NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden, 92 cl::desc("Don't try to vectorize boolean (i1) values")); 93 94static cl::opt<bool> 95NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 96 cl::desc("Don't try to vectorize integer values")); 97 98static cl::opt<bool> 99NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 100 cl::desc("Don't try to vectorize floating-point values")); 101 102// FIXME: This should default to false once pointer vector support works. 103static cl::opt<bool> 104NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden, 105 cl::desc("Don't try to vectorize pointer values")); 106 107static cl::opt<bool> 108NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 109 cl::desc("Don't try to vectorize casting (conversion) operations")); 110 111static cl::opt<bool> 112NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 113 cl::desc("Don't try to vectorize floating-point math intrinsics")); 114 115static cl::opt<bool> 116NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 117 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 118 119static cl::opt<bool> 120NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 121 cl::desc("Don't try to vectorize select instructions")); 122 123static cl::opt<bool> 124NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 125 cl::desc("Don't try to vectorize comparison instructions")); 126 127static cl::opt<bool> 128NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 129 cl::desc("Don't try to vectorize getelementptr instructions")); 130 131static cl::opt<bool> 132NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 133 cl::desc("Don't try to vectorize loads and stores")); 134 135static cl::opt<bool> 136AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 137 cl::desc("Only generate aligned loads and stores")); 138 139static cl::opt<bool> 140NoMemOpBoost("bb-vectorize-no-mem-op-boost", 141 cl::init(false), cl::Hidden, 142 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 143 144static cl::opt<bool> 145FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 146 cl::desc("Use a fast instruction dependency analysis")); 147 148#ifndef NDEBUG 149static cl::opt<bool> 150DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 151 cl::init(false), cl::Hidden, 152 cl::desc("When debugging is enabled, output information on the" 153 " instruction-examination process")); 154static cl::opt<bool> 155DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 156 cl::init(false), cl::Hidden, 157 cl::desc("When debugging is enabled, output information on the" 158 " candidate-selection process")); 159static cl::opt<bool> 160DebugPairSelection("bb-vectorize-debug-pair-selection", 161 cl::init(false), cl::Hidden, 162 cl::desc("When debugging is enabled, output information on the" 163 " pair-selection process")); 164static cl::opt<bool> 165DebugCycleCheck("bb-vectorize-debug-cycle-check", 166 cl::init(false), cl::Hidden, 167 cl::desc("When debugging is enabled, output information on the" 168 " cycle-checking process")); 169 170static cl::opt<bool> 171PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair", 172 cl::init(false), cl::Hidden, 173 cl::desc("When debugging is enabled, dump the basic block after" 174 " every pair is fused")); 175#endif 176 177STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 178 179namespace { 180 struct BBVectorize : public BasicBlockPass { 181 static char ID; // Pass identification, replacement for typeid 182 183 const VectorizeConfig Config; 184 185 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 186 : BasicBlockPass(ID), Config(C) { 187 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 188 } 189 190 BBVectorize(Pass *P, const VectorizeConfig &C) 191 : BasicBlockPass(ID), Config(C) { 192 AA = &P->getAnalysis<AliasAnalysis>(); 193 DT = &P->getAnalysis<DominatorTree>(); 194 SE = &P->getAnalysis<ScalarEvolution>(); 195 TD = P->getAnalysisIfAvailable<DataLayout>(); 196 TTI = IgnoreTargetInfo ? 0 : 197 P->getAnalysisIfAvailable<TargetTransformInfo>(); 198 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 199 } 200 201 typedef std::pair<Value *, Value *> ValuePair; 202 typedef std::pair<ValuePair, int> ValuePairWithCost; 203 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 204 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 205 typedef std::pair<VPPair, unsigned> VPPairWithType; 206 typedef std::pair<std::multimap<Value *, Value *>::iterator, 207 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 208 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 209 std::multimap<ValuePair, ValuePair>::iterator> 210 VPPIteratorPair; 211 212 AliasAnalysis *AA; 213 DominatorTree *DT; 214 ScalarEvolution *SE; 215 DataLayout *TD; 216 TargetTransformInfo *TTI; 217 const VectorTargetTransformInfo *VTTI; 218 219 // FIXME: const correct? 220 221 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 222 223 bool getCandidatePairs(BasicBlock &BB, 224 BasicBlock::iterator &Start, 225 std::multimap<Value *, Value *> &CandidatePairs, 226 DenseSet<ValuePair> &FixedOrderPairs, 227 DenseMap<ValuePair, int> &CandidatePairCostSavings, 228 std::vector<Value *> &PairableInsts, bool NonPow2Len); 229 230 enum PairConnectionType { 231 PairConnectionDirect, 232 PairConnectionSwap, 233 PairConnectionSplat 234 }; 235 236 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 237 std::vector<Value *> &PairableInsts, 238 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 239 DenseMap<VPPair, unsigned> &PairConnectionTypes); 240 241 void buildDepMap(BasicBlock &BB, 242 std::multimap<Value *, Value *> &CandidatePairs, 243 std::vector<Value *> &PairableInsts, 244 DenseSet<ValuePair> &PairableInstUsers); 245 246 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 247 DenseMap<ValuePair, int> &CandidatePairCostSavings, 248 std::vector<Value *> &PairableInsts, 249 DenseSet<ValuePair> &FixedOrderPairs, 250 DenseMap<VPPair, unsigned> &PairConnectionTypes, 251 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 252 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 253 DenseSet<ValuePair> &PairableInstUsers, 254 DenseMap<Value *, Value *>& ChosenPairs); 255 256 void fuseChosenPairs(BasicBlock &BB, 257 std::vector<Value *> &PairableInsts, 258 DenseMap<Value *, Value *>& ChosenPairs, 259 DenseSet<ValuePair> &FixedOrderPairs, 260 DenseMap<VPPair, unsigned> &PairConnectionTypes, 261 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 262 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps); 263 264 265 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 266 267 bool areInstsCompatible(Instruction *I, Instruction *J, 268 bool IsSimpleLoadStore, bool NonPow2Len, 269 int &CostSavings, int &FixedOrder); 270 271 bool trackUsesOfI(DenseSet<Value *> &Users, 272 AliasSetTracker &WriteSet, Instruction *I, 273 Instruction *J, bool UpdateUsers = true, 274 std::multimap<Value *, Value *> *LoadMoveSet = 0); 275 276 void computePairsConnectedTo( 277 std::multimap<Value *, Value *> &CandidatePairs, 278 std::vector<Value *> &PairableInsts, 279 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 280 DenseMap<VPPair, unsigned> &PairConnectionTypes, 281 ValuePair P); 282 283 bool pairsConflict(ValuePair P, ValuePair Q, 284 DenseSet<ValuePair> &PairableInstUsers, 285 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 286 287 bool pairWillFormCycle(ValuePair P, 288 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 289 DenseSet<ValuePair> &CurrentPairs); 290 291 void pruneTreeFor( 292 std::multimap<Value *, Value *> &CandidatePairs, 293 std::vector<Value *> &PairableInsts, 294 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 295 DenseSet<ValuePair> &PairableInstUsers, 296 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 297 DenseMap<Value *, Value *> &ChosenPairs, 298 DenseMap<ValuePair, size_t> &Tree, 299 DenseSet<ValuePair> &PrunedTree, ValuePair J, 300 bool UseCycleCheck); 301 302 void buildInitialTreeFor( 303 std::multimap<Value *, Value *> &CandidatePairs, 304 std::vector<Value *> &PairableInsts, 305 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 306 DenseSet<ValuePair> &PairableInstUsers, 307 DenseMap<Value *, Value *> &ChosenPairs, 308 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 309 310 void findBestTreeFor( 311 std::multimap<Value *, Value *> &CandidatePairs, 312 DenseMap<ValuePair, int> &CandidatePairCostSavings, 313 std::vector<Value *> &PairableInsts, 314 DenseSet<ValuePair> &FixedOrderPairs, 315 DenseMap<VPPair, unsigned> &PairConnectionTypes, 316 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 317 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 318 DenseSet<ValuePair> &PairableInstUsers, 319 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 320 DenseMap<Value *, Value *> &ChosenPairs, 321 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 322 int &BestEffSize, VPIteratorPair ChoiceRange, 323 bool UseCycleCheck); 324 325 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 326 Instruction *J, unsigned o); 327 328 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 329 unsigned MaskOffset, unsigned NumInElem, 330 unsigned NumInElem1, unsigned IdxOffset, 331 std::vector<Constant*> &Mask); 332 333 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 334 Instruction *J); 335 336 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 337 unsigned o, Value *&LOp, unsigned numElemL, 338 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, 339 unsigned IdxOff = 0); 340 341 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 342 Instruction *J, unsigned o, bool IBeforeJ); 343 344 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 345 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 346 bool IBeforeJ); 347 348 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 349 Instruction *J, Instruction *K, 350 Instruction *&InsertionPt, Instruction *&K1, 351 Instruction *&K2); 352 353 void collectPairLoadMoveSet(BasicBlock &BB, 354 DenseMap<Value *, Value *> &ChosenPairs, 355 std::multimap<Value *, Value *> &LoadMoveSet, 356 Instruction *I); 357 358 void collectLoadMoveSet(BasicBlock &BB, 359 std::vector<Value *> &PairableInsts, 360 DenseMap<Value *, Value *> &ChosenPairs, 361 std::multimap<Value *, Value *> &LoadMoveSet); 362 363 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 364 std::multimap<Value *, Value *> &LoadMoveSet, 365 Instruction *I, Instruction *J); 366 367 void moveUsesOfIAfterJ(BasicBlock &BB, 368 std::multimap<Value *, Value *> &LoadMoveSet, 369 Instruction *&InsertionPt, 370 Instruction *I, Instruction *J); 371 372 void combineMetadata(Instruction *K, const Instruction *J); 373 374 bool vectorizeBB(BasicBlock &BB) { 375 if (!DT->isReachableFromEntry(&BB)) { 376 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 377 " in " << BB.getParent()->getName() << "\n"); 378 return false; 379 } 380 381 DEBUG(if (VTTI) dbgs() << "BBV: using target information\n"); 382 383 bool changed = false; 384 // Iterate a sufficient number of times to merge types of size 1 bit, 385 // then 2 bits, then 4, etc. up to half of the target vector width of the 386 // target vector register. 387 unsigned n = 1; 388 for (unsigned v = 2; 389 (VTTI || v <= Config.VectorBits) && 390 (!Config.MaxIter || n <= Config.MaxIter); 391 v *= 2, ++n) { 392 DEBUG(dbgs() << "BBV: fusing loop #" << n << 393 " for " << BB.getName() << " in " << 394 BB.getParent()->getName() << "...\n"); 395 if (vectorizePairs(BB)) 396 changed = true; 397 else 398 break; 399 } 400 401 if (changed && !Pow2LenOnly) { 402 ++n; 403 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 404 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 405 n << " for " << BB.getName() << " in " << 406 BB.getParent()->getName() << "...\n"); 407 if (!vectorizePairs(BB, true)) break; 408 } 409 } 410 411 DEBUG(dbgs() << "BBV: done!\n"); 412 return changed; 413 } 414 415 virtual bool runOnBasicBlock(BasicBlock &BB) { 416 AA = &getAnalysis<AliasAnalysis>(); 417 DT = &getAnalysis<DominatorTree>(); 418 SE = &getAnalysis<ScalarEvolution>(); 419 TD = getAnalysisIfAvailable<DataLayout>(); 420 TTI = IgnoreTargetInfo ? 0 : 421 getAnalysisIfAvailable<TargetTransformInfo>(); 422 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 423 424 return vectorizeBB(BB); 425 } 426 427 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 428 BasicBlockPass::getAnalysisUsage(AU); 429 AU.addRequired<AliasAnalysis>(); 430 AU.addRequired<DominatorTree>(); 431 AU.addRequired<ScalarEvolution>(); 432 AU.addPreserved<AliasAnalysis>(); 433 AU.addPreserved<DominatorTree>(); 434 AU.addPreserved<ScalarEvolution>(); 435 AU.setPreservesCFG(); 436 } 437 438 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 439 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 440 "Cannot form vector from incompatible scalar types"); 441 Type *STy = ElemTy->getScalarType(); 442 443 unsigned numElem; 444 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 445 numElem = VTy->getNumElements(); 446 } else { 447 numElem = 1; 448 } 449 450 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 451 numElem += VTy->getNumElements(); 452 } else { 453 numElem += 1; 454 } 455 456 return VectorType::get(STy, numElem); 457 } 458 459 static inline void getInstructionTypes(Instruction *I, 460 Type *&T1, Type *&T2) { 461 if (isa<StoreInst>(I)) { 462 // For stores, it is the value type, not the pointer type that matters 463 // because the value is what will come from a vector register. 464 465 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 466 T1 = IVal->getType(); 467 } else { 468 T1 = I->getType(); 469 } 470 471 if (I->isCast()) 472 T2 = cast<CastInst>(I)->getSrcTy(); 473 else 474 T2 = T1; 475 476 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 477 T2 = SI->getCondition()->getType(); 478 } 479 } 480 481 // Returns the weight associated with the provided value. A chain of 482 // candidate pairs has a length given by the sum of the weights of its 483 // members (one weight per pair; the weight of each member of the pair 484 // is assumed to be the same). This length is then compared to the 485 // chain-length threshold to determine if a given chain is significant 486 // enough to be vectorized. The length is also used in comparing 487 // candidate chains where longer chains are considered to be better. 488 // Note: when this function returns 0, the resulting instructions are 489 // not actually fused. 490 inline size_t getDepthFactor(Value *V) { 491 // InsertElement and ExtractElement have a depth factor of zero. This is 492 // for two reasons: First, they cannot be usefully fused. Second, because 493 // the pass generates a lot of these, they can confuse the simple metric 494 // used to compare the trees in the next iteration. Thus, giving them a 495 // weight of zero allows the pass to essentially ignore them in 496 // subsequent iterations when looking for vectorization opportunities 497 // while still tracking dependency chains that flow through those 498 // instructions. 499 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 500 return 0; 501 502 // Give a load or store half of the required depth so that load/store 503 // pairs will vectorize. 504 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 505 return Config.ReqChainDepth/2; 506 507 return 1; 508 } 509 510 // Returns the cost of the provided instruction using VTTI. 511 // This does not handle loads and stores. 512 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { 513 switch (Opcode) { 514 default: break; 515 case Instruction::GetElementPtr: 516 // We mark this instruction as zero-cost because scalar GEPs are usually 517 // lowered to the intruction addressing mode. At the moment we don't 518 // generate vector GEPs. 519 return 0; 520 case Instruction::Br: 521 return VTTI->getCFInstrCost(Opcode); 522 case Instruction::PHI: 523 return 0; 524 case Instruction::Add: 525 case Instruction::FAdd: 526 case Instruction::Sub: 527 case Instruction::FSub: 528 case Instruction::Mul: 529 case Instruction::FMul: 530 case Instruction::UDiv: 531 case Instruction::SDiv: 532 case Instruction::FDiv: 533 case Instruction::URem: 534 case Instruction::SRem: 535 case Instruction::FRem: 536 case Instruction::Shl: 537 case Instruction::LShr: 538 case Instruction::AShr: 539 case Instruction::And: 540 case Instruction::Or: 541 case Instruction::Xor: 542 return VTTI->getArithmeticInstrCost(Opcode, T1); 543 case Instruction::Select: 544 case Instruction::ICmp: 545 case Instruction::FCmp: 546 return VTTI->getCmpSelInstrCost(Opcode, T1, T2); 547 case Instruction::ZExt: 548 case Instruction::SExt: 549 case Instruction::FPToUI: 550 case Instruction::FPToSI: 551 case Instruction::FPExt: 552 case Instruction::PtrToInt: 553 case Instruction::IntToPtr: 554 case Instruction::SIToFP: 555 case Instruction::UIToFP: 556 case Instruction::Trunc: 557 case Instruction::FPTrunc: 558 case Instruction::BitCast: 559 case Instruction::ShuffleVector: 560 return VTTI->getCastInstrCost(Opcode, T1, T2); 561 } 562 563 return 1; 564 } 565 566 // This determines the relative offset of two loads or stores, returning 567 // true if the offset could be determined to be some constant value. 568 // For example, if OffsetInElmts == 1, then J accesses the memory directly 569 // after I; if OffsetInElmts == -1 then I accesses the memory 570 // directly after J. 571 bool getPairPtrInfo(Instruction *I, Instruction *J, 572 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 573 unsigned &IAddressSpace, unsigned &JAddressSpace, 574 int64_t &OffsetInElmts, bool ComputeOffset = true) { 575 OffsetInElmts = 0; 576 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 577 LoadInst *LJ = cast<LoadInst>(J); 578 IPtr = LI->getPointerOperand(); 579 JPtr = LJ->getPointerOperand(); 580 IAlignment = LI->getAlignment(); 581 JAlignment = LJ->getAlignment(); 582 IAddressSpace = LI->getPointerAddressSpace(); 583 JAddressSpace = LJ->getPointerAddressSpace(); 584 } else { 585 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 586 IPtr = SI->getPointerOperand(); 587 JPtr = SJ->getPointerOperand(); 588 IAlignment = SI->getAlignment(); 589 JAlignment = SJ->getAlignment(); 590 IAddressSpace = SI->getPointerAddressSpace(); 591 JAddressSpace = SJ->getPointerAddressSpace(); 592 } 593 594 if (!ComputeOffset) 595 return true; 596 597 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 598 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 599 600 // If this is a trivial offset, then we'll get something like 601 // 1*sizeof(type). With target data, which we need anyway, this will get 602 // constant folded into a number. 603 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 604 if (const SCEVConstant *ConstOffSCEV = 605 dyn_cast<SCEVConstant>(OffsetSCEV)) { 606 ConstantInt *IntOff = ConstOffSCEV->getValue(); 607 int64_t Offset = IntOff->getSExtValue(); 608 609 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 610 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 611 612 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 613 if (VTy != VTy2 && Offset < 0) { 614 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 615 OffsetInElmts = Offset/VTy2TSS; 616 return (abs64(Offset) % VTy2TSS) == 0; 617 } 618 619 OffsetInElmts = Offset/VTyTSS; 620 return (abs64(Offset) % VTyTSS) == 0; 621 } 622 623 return false; 624 } 625 626 // Returns true if the provided CallInst represents an intrinsic that can 627 // be vectorized. 628 bool isVectorizableIntrinsic(CallInst* I) { 629 Function *F = I->getCalledFunction(); 630 if (!F) return false; 631 632 unsigned IID = F->getIntrinsicID(); 633 if (!IID) return false; 634 635 switch(IID) { 636 default: 637 return false; 638 case Intrinsic::sqrt: 639 case Intrinsic::powi: 640 case Intrinsic::sin: 641 case Intrinsic::cos: 642 case Intrinsic::log: 643 case Intrinsic::log2: 644 case Intrinsic::log10: 645 case Intrinsic::exp: 646 case Intrinsic::exp2: 647 case Intrinsic::pow: 648 return Config.VectorizeMath; 649 case Intrinsic::fma: 650 return Config.VectorizeFMA; 651 } 652 } 653 654 // Returns true if J is the second element in some pair referenced by 655 // some multimap pair iterator pair. 656 template <typename V> 657 bool isSecondInIteratorPair(V J, std::pair< 658 typename std::multimap<V, V>::iterator, 659 typename std::multimap<V, V>::iterator> PairRange) { 660 for (typename std::multimap<V, V>::iterator K = PairRange.first; 661 K != PairRange.second; ++K) 662 if (K->second == J) return true; 663 664 return false; 665 } 666 }; 667 668 // This function implements one vectorization iteration on the provided 669 // basic block. It returns true if the block is changed. 670 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 671 bool ShouldContinue; 672 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 673 674 std::vector<Value *> AllPairableInsts; 675 DenseMap<Value *, Value *> AllChosenPairs; 676 DenseSet<ValuePair> AllFixedOrderPairs; 677 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 678 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; 679 680 do { 681 std::vector<Value *> PairableInsts; 682 std::multimap<Value *, Value *> CandidatePairs; 683 DenseSet<ValuePair> FixedOrderPairs; 684 DenseMap<ValuePair, int> CandidatePairCostSavings; 685 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 686 FixedOrderPairs, 687 CandidatePairCostSavings, 688 PairableInsts, NonPow2Len); 689 if (PairableInsts.empty()) continue; 690 691 // Now we have a map of all of the pairable instructions and we need to 692 // select the best possible pairing. A good pairing is one such that the 693 // users of the pair are also paired. This defines a (directed) forest 694 // over the pairs such that two pairs are connected iff the second pair 695 // uses the first. 696 697 // Note that it only matters that both members of the second pair use some 698 // element of the first pair (to allow for splatting). 699 700 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; 701 DenseMap<VPPair, unsigned> PairConnectionTypes; 702 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs, 703 PairConnectionTypes); 704 if (ConnectedPairs.empty()) continue; 705 706 for (std::multimap<ValuePair, ValuePair>::iterator 707 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 708 I != IE; ++I) { 709 ConnectedPairDeps.insert(VPPair(I->second, I->first)); 710 } 711 712 // Build the pairable-instruction dependency map 713 DenseSet<ValuePair> PairableInstUsers; 714 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 715 716 // There is now a graph of the connected pairs. For each variable, pick 717 // the pairing with the largest tree meeting the depth requirement on at 718 // least one branch. Then select all pairings that are part of that tree 719 // and remove them from the list of available pairings and pairable 720 // variables. 721 722 DenseMap<Value *, Value *> ChosenPairs; 723 choosePairs(CandidatePairs, CandidatePairCostSavings, 724 PairableInsts, FixedOrderPairs, PairConnectionTypes, 725 ConnectedPairs, ConnectedPairDeps, 726 PairableInstUsers, ChosenPairs); 727 728 if (ChosenPairs.empty()) continue; 729 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 730 PairableInsts.end()); 731 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 732 733 // Only for the chosen pairs, propagate information on fixed-order pairs, 734 // pair connections, and their types to the data structures used by the 735 // pair fusion procedures. 736 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 737 IE = ChosenPairs.end(); I != IE; ++I) { 738 if (FixedOrderPairs.count(*I)) 739 AllFixedOrderPairs.insert(*I); 740 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 741 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 742 743 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 744 J != IE; ++J) { 745 DenseMap<VPPair, unsigned>::iterator K = 746 PairConnectionTypes.find(VPPair(*I, *J)); 747 if (K != PairConnectionTypes.end()) { 748 AllPairConnectionTypes.insert(*K); 749 } else { 750 K = PairConnectionTypes.find(VPPair(*J, *I)); 751 if (K != PairConnectionTypes.end()) 752 AllPairConnectionTypes.insert(*K); 753 } 754 } 755 } 756 757 for (std::multimap<ValuePair, ValuePair>::iterator 758 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 759 I != IE; ++I) { 760 if (AllPairConnectionTypes.count(*I)) { 761 AllConnectedPairs.insert(*I); 762 AllConnectedPairDeps.insert(VPPair(I->second, I->first)); 763 } 764 } 765 } while (ShouldContinue); 766 767 if (AllChosenPairs.empty()) return false; 768 NumFusedOps += AllChosenPairs.size(); 769 770 // A set of pairs has now been selected. It is now necessary to replace the 771 // paired instructions with vector instructions. For this procedure each 772 // operand must be replaced with a vector operand. This vector is formed 773 // by using build_vector on the old operands. The replaced values are then 774 // replaced with a vector_extract on the result. Subsequent optimization 775 // passes should coalesce the build/extract combinations. 776 777 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 778 AllPairConnectionTypes, 779 AllConnectedPairs, AllConnectedPairDeps); 780 781 // It is important to cleanup here so that future iterations of this 782 // function have less work to do. 783 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 784 return true; 785 } 786 787 // This function returns true if the provided instruction is capable of being 788 // fused into a vector instruction. This determination is based only on the 789 // type and other attributes of the instruction. 790 bool BBVectorize::isInstVectorizable(Instruction *I, 791 bool &IsSimpleLoadStore) { 792 IsSimpleLoadStore = false; 793 794 if (CallInst *C = dyn_cast<CallInst>(I)) { 795 if (!isVectorizableIntrinsic(C)) 796 return false; 797 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 798 // Vectorize simple loads if possbile: 799 IsSimpleLoadStore = L->isSimple(); 800 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 801 return false; 802 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 803 // Vectorize simple stores if possbile: 804 IsSimpleLoadStore = S->isSimple(); 805 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 806 return false; 807 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 808 // We can vectorize casts, but not casts of pointer types, etc. 809 if (!Config.VectorizeCasts) 810 return false; 811 812 Type *SrcTy = C->getSrcTy(); 813 if (!SrcTy->isSingleValueType()) 814 return false; 815 816 Type *DestTy = C->getDestTy(); 817 if (!DestTy->isSingleValueType()) 818 return false; 819 } else if (isa<SelectInst>(I)) { 820 if (!Config.VectorizeSelect) 821 return false; 822 } else if (isa<CmpInst>(I)) { 823 if (!Config.VectorizeCmp) 824 return false; 825 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 826 if (!Config.VectorizeGEP) 827 return false; 828 829 // Currently, vector GEPs exist only with one index. 830 if (G->getNumIndices() != 1) 831 return false; 832 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 833 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 834 return false; 835 } 836 837 // We can't vectorize memory operations without target data 838 if (TD == 0 && IsSimpleLoadStore) 839 return false; 840 841 Type *T1, *T2; 842 getInstructionTypes(I, T1, T2); 843 844 // Not every type can be vectorized... 845 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 846 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 847 return false; 848 849 if (T1->getScalarSizeInBits() == 1) { 850 if (!Config.VectorizeBools) 851 return false; 852 } else { 853 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 854 return false; 855 } 856 857 if (T2->getScalarSizeInBits() == 1) { 858 if (!Config.VectorizeBools) 859 return false; 860 } else { 861 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 862 return false; 863 } 864 865 if (!Config.VectorizeFloats 866 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 867 return false; 868 869 // Don't vectorize target-specific types. 870 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 871 return false; 872 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 873 return false; 874 875 if ((!Config.VectorizePointers || TD == 0) && 876 (T1->getScalarType()->isPointerTy() || 877 T2->getScalarType()->isPointerTy())) 878 return false; 879 880 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 881 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 882 return false; 883 884 return true; 885 } 886 887 // This function returns true if the two provided instructions are compatible 888 // (meaning that they can be fused into a vector instruction). This assumes 889 // that I has already been determined to be vectorizable and that J is not 890 // in the use tree of I. 891 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 892 bool IsSimpleLoadStore, bool NonPow2Len, 893 int &CostSavings, int &FixedOrder) { 894 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 895 " <-> " << *J << "\n"); 896 897 CostSavings = 0; 898 FixedOrder = 0; 899 900 // Loads and stores can be merged if they have different alignments, 901 // but are otherwise the same. 902 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 903 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 904 return false; 905 906 Type *IT1, *IT2, *JT1, *JT2; 907 getInstructionTypes(I, IT1, IT2); 908 getInstructionTypes(J, JT1, JT2); 909 unsigned MaxTypeBits = std::max( 910 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 911 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 912 if (!VTTI && MaxTypeBits > Config.VectorBits) 913 return false; 914 915 // FIXME: handle addsub-type operations! 916 917 if (IsSimpleLoadStore) { 918 Value *IPtr, *JPtr; 919 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 920 int64_t OffsetInElmts = 0; 921 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 922 IAddressSpace, JAddressSpace, 923 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 924 FixedOrder = (int) OffsetInElmts; 925 unsigned BottomAlignment = IAlignment; 926 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 927 928 Type *aTypeI = isa<StoreInst>(I) ? 929 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 930 Type *aTypeJ = isa<StoreInst>(J) ? 931 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 932 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 933 934 if (Config.AlignedOnly) { 935 // An aligned load or store is possible only if the instruction 936 // with the lower offset has an alignment suitable for the 937 // vector type. 938 939 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 940 if (BottomAlignment < VecAlignment) 941 return false; 942 } 943 944 if (VTTI) { 945 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(), 946 IAlignment, IAddressSpace); 947 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(), 948 JAlignment, JAddressSpace); 949 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType, 950 BottomAlignment, 951 IAddressSpace); 952 if (VCost > ICost + JCost) 953 return false; 954 955 // We don't want to fuse to a type that will be split, even 956 // if the two input types will also be split and there is no other 957 // associated cost. 958 unsigned VParts = VTTI->getNumberOfParts(VType); 959 if (VParts > 1) 960 return false; 961 else if (!VParts && VCost == ICost + JCost) 962 return false; 963 964 CostSavings = ICost + JCost - VCost; 965 } 966 } else { 967 return false; 968 } 969 } else if (VTTI) { 970 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 971 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 972 Type *VT1 = getVecTypeForPair(IT1, JT1), 973 *VT2 = getVecTypeForPair(IT2, JT2); 974 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 975 976 if (VCost > ICost + JCost) 977 return false; 978 979 // We don't want to fuse to a type that will be split, even 980 // if the two input types will also be split and there is no other 981 // associated cost. 982 unsigned VParts = VTTI->getNumberOfParts(VT1); 983 if (VParts > 1) 984 return false; 985 else if (!VParts && VCost == ICost + JCost) 986 return false; 987 988 CostSavings = ICost + JCost - VCost; 989 } 990 991 // The powi intrinsic is special because only the first argument is 992 // vectorized, the second arguments must be equal. 993 CallInst *CI = dyn_cast<CallInst>(I); 994 Function *FI; 995 if (CI && (FI = CI->getCalledFunction()) && 996 FI->getIntrinsicID() == Intrinsic::powi) { 997 998 Value *A1I = CI->getArgOperand(1), 999 *A1J = cast<CallInst>(J)->getArgOperand(1); 1000 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1001 *A1JSCEV = SE->getSCEV(A1J); 1002 return (A1ISCEV == A1JSCEV); 1003 } 1004 1005 return true; 1006 } 1007 1008 // Figure out whether or not J uses I and update the users and write-set 1009 // structures associated with I. Specifically, Users represents the set of 1010 // instructions that depend on I. WriteSet represents the set 1011 // of memory locations that are dependent on I. If UpdateUsers is true, 1012 // and J uses I, then Users is updated to contain J and WriteSet is updated 1013 // to contain any memory locations to which J writes. The function returns 1014 // true if J uses I. By default, alias analysis is used to determine 1015 // whether J reads from memory that overlaps with a location in WriteSet. 1016 // If LoadMoveSet is not null, then it is a previously-computed multimap 1017 // where the key is the memory-based user instruction and the value is 1018 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1019 // then the alias analysis is not used. This is necessary because this 1020 // function is called during the process of moving instructions during 1021 // vectorization and the results of the alias analysis are not stable during 1022 // that process. 1023 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1024 AliasSetTracker &WriteSet, Instruction *I, 1025 Instruction *J, bool UpdateUsers, 1026 std::multimap<Value *, Value *> *LoadMoveSet) { 1027 bool UsesI = false; 1028 1029 // This instruction may already be marked as a user due, for example, to 1030 // being a member of a selected pair. 1031 if (Users.count(J)) 1032 UsesI = true; 1033 1034 if (!UsesI) 1035 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1036 JU != JE; ++JU) { 1037 Value *V = *JU; 1038 if (I == V || Users.count(V)) { 1039 UsesI = true; 1040 break; 1041 } 1042 } 1043 if (!UsesI && J->mayReadFromMemory()) { 1044 if (LoadMoveSet) { 1045 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 1046 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 1047 } else { 1048 for (AliasSetTracker::iterator W = WriteSet.begin(), 1049 WE = WriteSet.end(); W != WE; ++W) { 1050 if (W->aliasesUnknownInst(J, *AA)) { 1051 UsesI = true; 1052 break; 1053 } 1054 } 1055 } 1056 } 1057 1058 if (UsesI && UpdateUsers) { 1059 if (J->mayWriteToMemory()) WriteSet.add(J); 1060 Users.insert(J); 1061 } 1062 1063 return UsesI; 1064 } 1065 1066 // This function iterates over all instruction pairs in the provided 1067 // basic block and collects all candidate pairs for vectorization. 1068 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1069 BasicBlock::iterator &Start, 1070 std::multimap<Value *, Value *> &CandidatePairs, 1071 DenseSet<ValuePair> &FixedOrderPairs, 1072 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1073 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1074 BasicBlock::iterator E = BB.end(); 1075 if (Start == E) return false; 1076 1077 bool ShouldContinue = false, IAfterStart = false; 1078 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1079 if (I == Start) IAfterStart = true; 1080 1081 bool IsSimpleLoadStore; 1082 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1083 1084 // Look for an instruction with which to pair instruction *I... 1085 DenseSet<Value *> Users; 1086 AliasSetTracker WriteSet(*AA); 1087 bool JAfterStart = IAfterStart; 1088 BasicBlock::iterator J = llvm::next(I); 1089 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1090 if (J == Start) JAfterStart = true; 1091 1092 // Determine if J uses I, if so, exit the loop. 1093 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1094 if (Config.FastDep) { 1095 // Note: For this heuristic to be effective, independent operations 1096 // must tend to be intermixed. This is likely to be true from some 1097 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1098 // but otherwise may require some kind of reordering pass. 1099 1100 // When using fast dependency analysis, 1101 // stop searching after first use: 1102 if (UsesI) break; 1103 } else { 1104 if (UsesI) continue; 1105 } 1106 1107 // J does not use I, and comes before the first use of I, so it can be 1108 // merged with I if the instructions are compatible. 1109 int CostSavings, FixedOrder; 1110 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1111 CostSavings, FixedOrder)) continue; 1112 1113 // J is a candidate for merging with I. 1114 if (!PairableInsts.size() || 1115 PairableInsts[PairableInsts.size()-1] != I) { 1116 PairableInsts.push_back(I); 1117 } 1118 1119 CandidatePairs.insert(ValuePair(I, J)); 1120 if (VTTI) 1121 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1122 CostSavings)); 1123 1124 if (FixedOrder == 1) 1125 FixedOrderPairs.insert(ValuePair(I, J)); 1126 else if (FixedOrder == -1) 1127 FixedOrderPairs.insert(ValuePair(J, I)); 1128 1129 // The next call to this function must start after the last instruction 1130 // selected during this invocation. 1131 if (JAfterStart) { 1132 Start = llvm::next(J); 1133 IAfterStart = JAfterStart = false; 1134 } 1135 1136 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1137 << *I << " <-> " << *J << " (cost savings: " << 1138 CostSavings << ")\n"); 1139 1140 // If we have already found too many pairs, break here and this function 1141 // will be called again starting after the last instruction selected 1142 // during this invocation. 1143 if (PairableInsts.size() >= Config.MaxInsts) { 1144 ShouldContinue = true; 1145 break; 1146 } 1147 } 1148 1149 if (ShouldContinue) 1150 break; 1151 } 1152 1153 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1154 << " instructions with candidate pairs\n"); 1155 1156 return ShouldContinue; 1157 } 1158 1159 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1160 // it looks for pairs such that both members have an input which is an 1161 // output of PI or PJ. 1162 void BBVectorize::computePairsConnectedTo( 1163 std::multimap<Value *, Value *> &CandidatePairs, 1164 std::vector<Value *> &PairableInsts, 1165 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1166 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1167 ValuePair P) { 1168 StoreInst *SI, *SJ; 1169 1170 // For each possible pairing for this variable, look at the uses of 1171 // the first value... 1172 for (Value::use_iterator I = P.first->use_begin(), 1173 E = P.first->use_end(); I != E; ++I) { 1174 if (isa<LoadInst>(*I)) { 1175 // A pair cannot be connected to a load because the load only takes one 1176 // operand (the address) and it is a scalar even after vectorization. 1177 continue; 1178 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1179 P.first == SI->getPointerOperand()) { 1180 // Similarly, a pair cannot be connected to a store through its 1181 // pointer operand. 1182 continue; 1183 } 1184 1185 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1186 1187 // For each use of the first variable, look for uses of the second 1188 // variable... 1189 for (Value::use_iterator J = P.second->use_begin(), 1190 E2 = P.second->use_end(); J != E2; ++J) { 1191 if ((SJ = dyn_cast<StoreInst>(*J)) && 1192 P.second == SJ->getPointerOperand()) 1193 continue; 1194 1195 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 1196 1197 // Look for <I, J>: 1198 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1199 VPPair VP(P, ValuePair(*I, *J)); 1200 ConnectedPairs.insert(VP); 1201 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1202 } 1203 1204 // Look for <J, I>: 1205 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) { 1206 VPPair VP(P, ValuePair(*J, *I)); 1207 ConnectedPairs.insert(VP); 1208 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1209 } 1210 } 1211 1212 if (Config.SplatBreaksChain) continue; 1213 // Look for cases where just the first value in the pair is used by 1214 // both members of another pair (splatting). 1215 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1216 if ((SJ = dyn_cast<StoreInst>(*J)) && 1217 P.first == SJ->getPointerOperand()) 1218 continue; 1219 1220 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1221 VPPair VP(P, ValuePair(*I, *J)); 1222 ConnectedPairs.insert(VP); 1223 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1224 } 1225 } 1226 } 1227 1228 if (Config.SplatBreaksChain) return; 1229 // Look for cases where just the second value in the pair is used by 1230 // both members of another pair (splatting). 1231 for (Value::use_iterator I = P.second->use_begin(), 1232 E = P.second->use_end(); I != E; ++I) { 1233 if (isa<LoadInst>(*I)) 1234 continue; 1235 else if ((SI = dyn_cast<StoreInst>(*I)) && 1236 P.second == SI->getPointerOperand()) 1237 continue; 1238 1239 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1240 1241 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1242 if ((SJ = dyn_cast<StoreInst>(*J)) && 1243 P.second == SJ->getPointerOperand()) 1244 continue; 1245 1246 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1247 VPPair VP(P, ValuePair(*I, *J)); 1248 ConnectedPairs.insert(VP); 1249 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1250 } 1251 } 1252 } 1253 } 1254 1255 // This function figures out which pairs are connected. Two pairs are 1256 // connected if some output of the first pair forms an input to both members 1257 // of the second pair. 1258 void BBVectorize::computeConnectedPairs( 1259 std::multimap<Value *, Value *> &CandidatePairs, 1260 std::vector<Value *> &PairableInsts, 1261 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1262 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1263 1264 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1265 PE = PairableInsts.end(); PI != PE; ++PI) { 1266 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1267 1268 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1269 P != choiceRange.second; ++P) 1270 computePairsConnectedTo(CandidatePairs, PairableInsts, 1271 ConnectedPairs, PairConnectionTypes, *P); 1272 } 1273 1274 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1275 << " pair connections.\n"); 1276 } 1277 1278 // This function builds a set of use tuples such that <A, B> is in the set 1279 // if B is in the use tree of A. If B is in the use tree of A, then B 1280 // depends on the output of A. 1281 void BBVectorize::buildDepMap( 1282 BasicBlock &BB, 1283 std::multimap<Value *, Value *> &CandidatePairs, 1284 std::vector<Value *> &PairableInsts, 1285 DenseSet<ValuePair> &PairableInstUsers) { 1286 DenseSet<Value *> IsInPair; 1287 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1288 E = CandidatePairs.end(); C != E; ++C) { 1289 IsInPair.insert(C->first); 1290 IsInPair.insert(C->second); 1291 } 1292 1293 // Iterate through the basic block, recording all Users of each 1294 // pairable instruction. 1295 1296 BasicBlock::iterator E = BB.end(); 1297 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1298 if (IsInPair.find(I) == IsInPair.end()) continue; 1299 1300 DenseSet<Value *> Users; 1301 AliasSetTracker WriteSet(*AA); 1302 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1303 (void) trackUsesOfI(Users, WriteSet, I, J); 1304 1305 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1306 U != E; ++U) 1307 PairableInstUsers.insert(ValuePair(I, *U)); 1308 } 1309 } 1310 1311 // Returns true if an input to pair P is an output of pair Q and also an 1312 // input of pair Q is an output of pair P. If this is the case, then these 1313 // two pairs cannot be simultaneously fused. 1314 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1315 DenseSet<ValuePair> &PairableInstUsers, 1316 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1317 // Two pairs are in conflict if they are mutual Users of eachother. 1318 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1319 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1320 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1321 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1322 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1323 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1324 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1325 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1326 if (PairableInstUserMap) { 1327 // FIXME: The expensive part of the cycle check is not so much the cycle 1328 // check itself but this edge insertion procedure. This needs some 1329 // profiling and probably a different data structure (same is true of 1330 // most uses of std::multimap). 1331 if (PUsesQ) { 1332 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1333 if (!isSecondInIteratorPair(P, QPairRange)) 1334 PairableInstUserMap->insert(VPPair(Q, P)); 1335 } 1336 if (QUsesP) { 1337 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1338 if (!isSecondInIteratorPair(Q, PPairRange)) 1339 PairableInstUserMap->insert(VPPair(P, Q)); 1340 } 1341 } 1342 1343 return (QUsesP && PUsesQ); 1344 } 1345 1346 // This function walks the use graph of current pairs to see if, starting 1347 // from P, the walk returns to P. 1348 bool BBVectorize::pairWillFormCycle(ValuePair P, 1349 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1350 DenseSet<ValuePair> &CurrentPairs) { 1351 DEBUG(if (DebugCycleCheck) 1352 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1353 << *P.second << "\n"); 1354 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1355 // contains non-direct associations. 1356 DenseSet<ValuePair> Visited; 1357 SmallVector<ValuePair, 32> Q; 1358 // General depth-first post-order traversal: 1359 Q.push_back(P); 1360 do { 1361 ValuePair QTop = Q.pop_back_val(); 1362 Visited.insert(QTop); 1363 1364 DEBUG(if (DebugCycleCheck) 1365 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1366 << *QTop.second << "\n"); 1367 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1368 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1369 C != QPairRange.second; ++C) { 1370 if (C->second == P) { 1371 DEBUG(dbgs() 1372 << "BBV: rejected to prevent non-trivial cycle formation: " 1373 << *C->first.first << " <-> " << *C->first.second << "\n"); 1374 return true; 1375 } 1376 1377 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1378 Q.push_back(C->second); 1379 } 1380 } while (!Q.empty()); 1381 1382 return false; 1383 } 1384 1385 // This function builds the initial tree of connected pairs with the 1386 // pair J at the root. 1387 void BBVectorize::buildInitialTreeFor( 1388 std::multimap<Value *, Value *> &CandidatePairs, 1389 std::vector<Value *> &PairableInsts, 1390 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1391 DenseSet<ValuePair> &PairableInstUsers, 1392 DenseMap<Value *, Value *> &ChosenPairs, 1393 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1394 // Each of these pairs is viewed as the root node of a Tree. The Tree 1395 // is then walked (depth-first). As this happens, we keep track of 1396 // the pairs that compose the Tree and the maximum depth of the Tree. 1397 SmallVector<ValuePairWithDepth, 32> Q; 1398 // General depth-first post-order traversal: 1399 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1400 do { 1401 ValuePairWithDepth QTop = Q.back(); 1402 1403 // Push each child onto the queue: 1404 bool MoreChildren = false; 1405 size_t MaxChildDepth = QTop.second; 1406 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1407 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1408 k != qtRange.second; ++k) { 1409 // Make sure that this child pair is still a candidate: 1410 bool IsStillCand = false; 1411 VPIteratorPair checkRange = 1412 CandidatePairs.equal_range(k->second.first); 1413 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1414 m != checkRange.second; ++m) { 1415 if (m->second == k->second.second) { 1416 IsStillCand = true; 1417 break; 1418 } 1419 } 1420 1421 if (IsStillCand) { 1422 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1423 if (C == Tree.end()) { 1424 size_t d = getDepthFactor(k->second.first); 1425 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1426 MoreChildren = true; 1427 } else { 1428 MaxChildDepth = std::max(MaxChildDepth, C->second); 1429 } 1430 } 1431 } 1432 1433 if (!MoreChildren) { 1434 // Record the current pair as part of the Tree: 1435 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1436 Q.pop_back(); 1437 } 1438 } while (!Q.empty()); 1439 } 1440 1441 // Given some initial tree, prune it by removing conflicting pairs (pairs 1442 // that cannot be simultaneously chosen for vectorization). 1443 void BBVectorize::pruneTreeFor( 1444 std::multimap<Value *, Value *> &CandidatePairs, 1445 std::vector<Value *> &PairableInsts, 1446 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1447 DenseSet<ValuePair> &PairableInstUsers, 1448 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1449 DenseMap<Value *, Value *> &ChosenPairs, 1450 DenseMap<ValuePair, size_t> &Tree, 1451 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1452 bool UseCycleCheck) { 1453 SmallVector<ValuePairWithDepth, 32> Q; 1454 // General depth-first post-order traversal: 1455 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1456 do { 1457 ValuePairWithDepth QTop = Q.pop_back_val(); 1458 PrunedTree.insert(QTop.first); 1459 1460 // Visit each child, pruning as necessary... 1461 DenseMap<ValuePair, size_t> BestChildren; 1462 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1463 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1464 K != QTopRange.second; ++K) { 1465 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1466 if (C == Tree.end()) continue; 1467 1468 // This child is in the Tree, now we need to make sure it is the 1469 // best of any conflicting children. There could be multiple 1470 // conflicting children, so first, determine if we're keeping 1471 // this child, then delete conflicting children as necessary. 1472 1473 // It is also necessary to guard against pairing-induced 1474 // dependencies. Consider instructions a .. x .. y .. b 1475 // such that (a,b) are to be fused and (x,y) are to be fused 1476 // but a is an input to x and b is an output from y. This 1477 // means that y cannot be moved after b but x must be moved 1478 // after b for (a,b) to be fused. In other words, after 1479 // fusing (a,b) we have y .. a/b .. x where y is an input 1480 // to a/b and x is an output to a/b: x and y can no longer 1481 // be legally fused. To prevent this condition, we must 1482 // make sure that a child pair added to the Tree is not 1483 // both an input and output of an already-selected pair. 1484 1485 // Pairing-induced dependencies can also form from more complicated 1486 // cycles. The pair vs. pair conflicts are easy to check, and so 1487 // that is done explicitly for "fast rejection", and because for 1488 // child vs. child conflicts, we may prefer to keep the current 1489 // pair in preference to the already-selected child. 1490 DenseSet<ValuePair> CurrentPairs; 1491 1492 bool CanAdd = true; 1493 for (DenseMap<ValuePair, size_t>::iterator C2 1494 = BestChildren.begin(), E2 = BestChildren.end(); 1495 C2 != E2; ++C2) { 1496 if (C2->first.first == C->first.first || 1497 C2->first.first == C->first.second || 1498 C2->first.second == C->first.first || 1499 C2->first.second == C->first.second || 1500 pairsConflict(C2->first, C->first, PairableInstUsers, 1501 UseCycleCheck ? &PairableInstUserMap : 0)) { 1502 if (C2->second >= C->second) { 1503 CanAdd = false; 1504 break; 1505 } 1506 1507 CurrentPairs.insert(C2->first); 1508 } 1509 } 1510 if (!CanAdd) continue; 1511 1512 // Even worse, this child could conflict with another node already 1513 // selected for the Tree. If that is the case, ignore this child. 1514 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1515 E2 = PrunedTree.end(); T != E2; ++T) { 1516 if (T->first == C->first.first || 1517 T->first == C->first.second || 1518 T->second == C->first.first || 1519 T->second == C->first.second || 1520 pairsConflict(*T, C->first, PairableInstUsers, 1521 UseCycleCheck ? &PairableInstUserMap : 0)) { 1522 CanAdd = false; 1523 break; 1524 } 1525 1526 CurrentPairs.insert(*T); 1527 } 1528 if (!CanAdd) continue; 1529 1530 // And check the queue too... 1531 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1532 E2 = Q.end(); C2 != E2; ++C2) { 1533 if (C2->first.first == C->first.first || 1534 C2->first.first == C->first.second || 1535 C2->first.second == C->first.first || 1536 C2->first.second == C->first.second || 1537 pairsConflict(C2->first, C->first, PairableInstUsers, 1538 UseCycleCheck ? &PairableInstUserMap : 0)) { 1539 CanAdd = false; 1540 break; 1541 } 1542 1543 CurrentPairs.insert(C2->first); 1544 } 1545 if (!CanAdd) continue; 1546 1547 // Last but not least, check for a conflict with any of the 1548 // already-chosen pairs. 1549 for (DenseMap<Value *, Value *>::iterator C2 = 1550 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1551 C2 != E2; ++C2) { 1552 if (pairsConflict(*C2, C->first, PairableInstUsers, 1553 UseCycleCheck ? &PairableInstUserMap : 0)) { 1554 CanAdd = false; 1555 break; 1556 } 1557 1558 CurrentPairs.insert(*C2); 1559 } 1560 if (!CanAdd) continue; 1561 1562 // To check for non-trivial cycles formed by the addition of the 1563 // current pair we've formed a list of all relevant pairs, now use a 1564 // graph walk to check for a cycle. We start from the current pair and 1565 // walk the use tree to see if we again reach the current pair. If we 1566 // do, then the current pair is rejected. 1567 1568 // FIXME: It may be more efficient to use a topological-ordering 1569 // algorithm to improve the cycle check. This should be investigated. 1570 if (UseCycleCheck && 1571 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1572 continue; 1573 1574 // This child can be added, but we may have chosen it in preference 1575 // to an already-selected child. Check for this here, and if a 1576 // conflict is found, then remove the previously-selected child 1577 // before adding this one in its place. 1578 for (DenseMap<ValuePair, size_t>::iterator C2 1579 = BestChildren.begin(); C2 != BestChildren.end();) { 1580 if (C2->first.first == C->first.first || 1581 C2->first.first == C->first.second || 1582 C2->first.second == C->first.first || 1583 C2->first.second == C->first.second || 1584 pairsConflict(C2->first, C->first, PairableInstUsers)) 1585 BestChildren.erase(C2++); 1586 else 1587 ++C2; 1588 } 1589 1590 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1591 } 1592 1593 for (DenseMap<ValuePair, size_t>::iterator C 1594 = BestChildren.begin(), E2 = BestChildren.end(); 1595 C != E2; ++C) { 1596 size_t DepthF = getDepthFactor(C->first.first); 1597 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1598 } 1599 } while (!Q.empty()); 1600 } 1601 1602 // This function finds the best tree of mututally-compatible connected 1603 // pairs, given the choice of root pairs as an iterator range. 1604 void BBVectorize::findBestTreeFor( 1605 std::multimap<Value *, Value *> &CandidatePairs, 1606 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1607 std::vector<Value *> &PairableInsts, 1608 DenseSet<ValuePair> &FixedOrderPairs, 1609 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1610 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1611 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1612 DenseSet<ValuePair> &PairableInstUsers, 1613 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1614 DenseMap<Value *, Value *> &ChosenPairs, 1615 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1616 int &BestEffSize, VPIteratorPair ChoiceRange, 1617 bool UseCycleCheck) { 1618 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1619 J != ChoiceRange.second; ++J) { 1620 1621 // Before going any further, make sure that this pair does not 1622 // conflict with any already-selected pairs (see comment below 1623 // near the Tree pruning for more details). 1624 DenseSet<ValuePair> ChosenPairSet; 1625 bool DoesConflict = false; 1626 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1627 E = ChosenPairs.end(); C != E; ++C) { 1628 if (pairsConflict(*C, *J, PairableInstUsers, 1629 UseCycleCheck ? &PairableInstUserMap : 0)) { 1630 DoesConflict = true; 1631 break; 1632 } 1633 1634 ChosenPairSet.insert(*C); 1635 } 1636 if (DoesConflict) continue; 1637 1638 if (UseCycleCheck && 1639 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1640 continue; 1641 1642 DenseMap<ValuePair, size_t> Tree; 1643 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1644 PairableInstUsers, ChosenPairs, Tree, *J); 1645 1646 // Because we'll keep the child with the largest depth, the largest 1647 // depth is still the same in the unpruned Tree. 1648 size_t MaxDepth = Tree.lookup(*J); 1649 1650 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1651 << *J->first << " <-> " << *J->second << "} of depth " << 1652 MaxDepth << " and size " << Tree.size() << "\n"); 1653 1654 // At this point the Tree has been constructed, but, may contain 1655 // contradictory children (meaning that different children of 1656 // some tree node may be attempting to fuse the same instruction). 1657 // So now we walk the tree again, in the case of a conflict, 1658 // keep only the child with the largest depth. To break a tie, 1659 // favor the first child. 1660 1661 DenseSet<ValuePair> PrunedTree; 1662 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1663 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1664 PrunedTree, *J, UseCycleCheck); 1665 1666 int EffSize = 0; 1667 if (VTTI) { 1668 // The node weights represent the cost savings associated with 1669 // fusing the pair of instructions. 1670 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1671 E = PrunedTree.end(); S != E; ++S) { 1672 if (getDepthFactor(S->first)) 1673 EffSize += CandidatePairCostSavings.find(*S)->second; 1674 1675 // The edge weights contribute in a negative sense: they represent 1676 // the cost of shuffles. 1677 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); 1678 if (IP.first != ConnectedPairDeps.end()) { 1679 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1680 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1681 Q != IP.second; ++Q) { 1682 DenseMap<VPPair, unsigned>::iterator R = 1683 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1684 assert(R != PairConnectionTypes.end() && 1685 "Cannot find pair connection type"); 1686 if (R->second == PairConnectionDirect) 1687 ++NumDepsDirect; 1688 else if (R->second == PairConnectionSwap) 1689 ++NumDepsSwap; 1690 } 1691 1692 // If there are more swaps than direct connections, then 1693 // the pair order will be flipped during fusion. So the real 1694 // number of swaps is the minimum number. 1695 bool FlipOrder = !FixedOrderPairs.count(*S) && 1696 ((NumDepsSwap > NumDepsDirect) || 1697 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1698 1699 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1700 Q != IP.second; ++Q) { 1701 DenseMap<VPPair, unsigned>::iterator R = 1702 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1703 assert(R != PairConnectionTypes.end() && 1704 "Cannot find pair connection type"); 1705 Type *Ty1 = Q->second.first->getType(), 1706 *Ty2 = Q->second.second->getType(); 1707 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1708 if ((R->second == PairConnectionDirect && FlipOrder) || 1709 (R->second == PairConnectionSwap && !FlipOrder) || 1710 R->second == PairConnectionSplat) 1711 EffSize -= (int) getInstrCost(Instruction::ShuffleVector, 1712 VTy, VTy); 1713 } 1714 } 1715 } 1716 } else { 1717 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1718 E = PrunedTree.end(); S != E; ++S) 1719 EffSize += (int) getDepthFactor(S->first); 1720 } 1721 1722 DEBUG(if (DebugPairSelection) 1723 dbgs() << "BBV: found pruned Tree for pair {" 1724 << *J->first << " <-> " << *J->second << "} of depth " << 1725 MaxDepth << " and size " << PrunedTree.size() << 1726 " (effective size: " << EffSize << ")\n"); 1727 if (MaxDepth >= Config.ReqChainDepth && 1728 EffSize > 0 && EffSize > BestEffSize) { 1729 BestMaxDepth = MaxDepth; 1730 BestEffSize = EffSize; 1731 BestTree = PrunedTree; 1732 } 1733 } 1734 } 1735 1736 // Given the list of candidate pairs, this function selects those 1737 // that will be fused into vector instructions. 1738 void BBVectorize::choosePairs( 1739 std::multimap<Value *, Value *> &CandidatePairs, 1740 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1741 std::vector<Value *> &PairableInsts, 1742 DenseSet<ValuePair> &FixedOrderPairs, 1743 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1744 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1745 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1746 DenseSet<ValuePair> &PairableInstUsers, 1747 DenseMap<Value *, Value *>& ChosenPairs) { 1748 bool UseCycleCheck = 1749 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1750 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1751 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1752 E = PairableInsts.end(); I != E; ++I) { 1753 // The number of possible pairings for this variable: 1754 size_t NumChoices = CandidatePairs.count(*I); 1755 if (!NumChoices) continue; 1756 1757 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1758 1759 // The best pair to choose and its tree: 1760 size_t BestMaxDepth = 0; 1761 int BestEffSize = 0; 1762 DenseSet<ValuePair> BestTree; 1763 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 1764 PairableInsts, FixedOrderPairs, PairConnectionTypes, 1765 ConnectedPairs, ConnectedPairDeps, 1766 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1767 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1768 UseCycleCheck); 1769 1770 // A tree has been chosen (or not) at this point. If no tree was 1771 // chosen, then this instruction, I, cannot be paired (and is no longer 1772 // considered). 1773 1774 DEBUG(if (BestTree.size() > 0) 1775 dbgs() << "BBV: selected pairs in the best tree for: " 1776 << *cast<Instruction>(*I) << "\n"); 1777 1778 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1779 SE2 = BestTree.end(); S != SE2; ++S) { 1780 // Insert the members of this tree into the list of chosen pairs. 1781 ChosenPairs.insert(ValuePair(S->first, S->second)); 1782 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1783 *S->second << "\n"); 1784 1785 // Remove all candidate pairs that have values in the chosen tree. 1786 for (std::multimap<Value *, Value *>::iterator K = 1787 CandidatePairs.begin(); K != CandidatePairs.end();) { 1788 if (K->first == S->first || K->second == S->first || 1789 K->second == S->second || K->first == S->second) { 1790 // Don't remove the actual pair chosen so that it can be used 1791 // in subsequent tree selections. 1792 if (!(K->first == S->first && K->second == S->second)) 1793 CandidatePairs.erase(K++); 1794 else 1795 ++K; 1796 } else { 1797 ++K; 1798 } 1799 } 1800 } 1801 } 1802 1803 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1804 } 1805 1806 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1807 unsigned n = 0) { 1808 if (!I->hasName()) 1809 return ""; 1810 1811 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1812 (n > 0 ? "." + utostr(n) : "")).str(); 1813 } 1814 1815 // Returns the value that is to be used as the pointer input to the vector 1816 // instruction that fuses I with J. 1817 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1818 Instruction *I, Instruction *J, unsigned o) { 1819 Value *IPtr, *JPtr; 1820 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 1821 int64_t OffsetInElmts; 1822 1823 // Note: the analysis might fail here, that is why the pair order has 1824 // been precomputed (OffsetInElmts must be unused here). 1825 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1826 IAddressSpace, JAddressSpace, 1827 OffsetInElmts, false); 1828 1829 // The pointer value is taken to be the one with the lowest offset. 1830 Value *VPtr = IPtr; 1831 1832 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 1833 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 1834 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1835 Type *VArgPtrType = PointerType::get(VArgType, 1836 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1837 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1838 /* insert before */ I); 1839 } 1840 1841 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1842 unsigned MaskOffset, unsigned NumInElem, 1843 unsigned NumInElem1, unsigned IdxOffset, 1844 std::vector<Constant*> &Mask) { 1845 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 1846 for (unsigned v = 0; v < NumElem1; ++v) { 1847 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1848 if (m < 0) { 1849 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1850 } else { 1851 unsigned mm = m + (int) IdxOffset; 1852 if (m >= (int) NumInElem1) 1853 mm += (int) NumInElem; 1854 1855 Mask[v+MaskOffset] = 1856 ConstantInt::get(Type::getInt32Ty(Context), mm); 1857 } 1858 } 1859 } 1860 1861 // Returns the value that is to be used as the vector-shuffle mask to the 1862 // vector instruction that fuses I with J. 1863 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1864 Instruction *I, Instruction *J) { 1865 // This is the shuffle mask. We need to append the second 1866 // mask to the first, and the numbers need to be adjusted. 1867 1868 Type *ArgTypeI = I->getType(); 1869 Type *ArgTypeJ = J->getType(); 1870 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1871 1872 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 1873 1874 // Get the total number of elements in the fused vector type. 1875 // By definition, this must equal the number of elements in 1876 // the final mask. 1877 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1878 std::vector<Constant*> Mask(NumElem); 1879 1880 Type *OpTypeI = I->getOperand(0)->getType(); 1881 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 1882 Type *OpTypeJ = J->getOperand(0)->getType(); 1883 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 1884 1885 // The fused vector will be: 1886 // ----------------------------------------------------- 1887 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 1888 // ----------------------------------------------------- 1889 // from which we'll extract NumElem total elements (where the first NumElemI 1890 // of them come from the mask in I and the remainder come from the mask 1891 // in J. 1892 1893 // For the mask from the first pair... 1894 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 1895 0, Mask); 1896 1897 // For the mask from the second pair... 1898 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 1899 NumInElemI, Mask); 1900 1901 return ConstantVector::get(Mask); 1902 } 1903 1904 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 1905 Instruction *J, unsigned o, Value *&LOp, 1906 unsigned numElemL, 1907 Type *ArgTypeL, Type *ArgTypeH, 1908 bool IBeforeJ, unsigned IdxOff) { 1909 bool ExpandedIEChain = false; 1910 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 1911 // If we have a pure insertelement chain, then this can be rewritten 1912 // into a chain that directly builds the larger type. 1913 bool PureChain = true; 1914 InsertElementInst *LIENext = LIE; 1915 do { 1916 if (!isa<UndefValue>(LIENext->getOperand(0)) && 1917 !isa<InsertElementInst>(LIENext->getOperand(0))) { 1918 PureChain = false; 1919 break; 1920 } 1921 } while ((LIENext = 1922 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1923 1924 if (PureChain) { 1925 SmallVector<Value *, 8> VectElemts(numElemL, 1926 UndefValue::get(ArgTypeL->getScalarType())); 1927 InsertElementInst *LIENext = LIE; 1928 do { 1929 unsigned Idx = 1930 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 1931 VectElemts[Idx] = LIENext->getOperand(1); 1932 } while ((LIENext = 1933 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1934 1935 LIENext = 0; 1936 Value *LIEPrev = UndefValue::get(ArgTypeH); 1937 for (unsigned i = 0; i < numElemL; ++i) { 1938 if (isa<UndefValue>(VectElemts[i])) continue; 1939 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 1940 ConstantInt::get(Type::getInt32Ty(Context), 1941 i + IdxOff), 1942 getReplacementName(IBeforeJ ? I : J, 1943 true, o, i+1)); 1944 LIENext->insertBefore(IBeforeJ ? J : I); 1945 LIEPrev = LIENext; 1946 } 1947 1948 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 1949 ExpandedIEChain = true; 1950 } 1951 } 1952 1953 return ExpandedIEChain; 1954 } 1955 1956 // Returns the value to be used as the specified operand of the vector 1957 // instruction that fuses I with J. 1958 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1959 Instruction *J, unsigned o, bool IBeforeJ) { 1960 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1961 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1962 1963 // Compute the fused vector type for this operand 1964 Type *ArgTypeI = I->getOperand(o)->getType(); 1965 Type *ArgTypeJ = J->getOperand(o)->getType(); 1966 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1967 1968 Instruction *L = I, *H = J; 1969 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 1970 1971 unsigned numElemL; 1972 if (ArgTypeL->isVectorTy()) 1973 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 1974 else 1975 numElemL = 1; 1976 1977 unsigned numElemH; 1978 if (ArgTypeH->isVectorTy()) 1979 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 1980 else 1981 numElemH = 1; 1982 1983 Value *LOp = L->getOperand(o); 1984 Value *HOp = H->getOperand(o); 1985 unsigned numElem = VArgType->getNumElements(); 1986 1987 // First, we check if we can reuse the "original" vector outputs (if these 1988 // exist). We might need a shuffle. 1989 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 1990 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 1991 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 1992 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 1993 1994 // FIXME: If we're fusing shuffle instructions, then we can't apply this 1995 // optimization. The input vectors to the shuffle might be a different 1996 // length from the shuffle outputs. Unfortunately, the replacement 1997 // shuffle mask has already been formed, and the mask entries are sensitive 1998 // to the sizes of the inputs. 1999 bool IsSizeChangeShuffle = 2000 isa<ShuffleVectorInst>(L) && 2001 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2002 2003 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2004 // We can have at most two unique vector inputs. 2005 bool CanUseInputs = true; 2006 Value *I1, *I2 = 0; 2007 if (LEE) { 2008 I1 = LEE->getOperand(0); 2009 } else { 2010 I1 = LSV->getOperand(0); 2011 I2 = LSV->getOperand(1); 2012 if (I2 == I1 || isa<UndefValue>(I2)) 2013 I2 = 0; 2014 } 2015 2016 if (HEE) { 2017 Value *I3 = HEE->getOperand(0); 2018 if (!I2 && I3 != I1) 2019 I2 = I3; 2020 else if (I3 != I1 && I3 != I2) 2021 CanUseInputs = false; 2022 } else { 2023 Value *I3 = HSV->getOperand(0); 2024 if (!I2 && I3 != I1) 2025 I2 = I3; 2026 else if (I3 != I1 && I3 != I2) 2027 CanUseInputs = false; 2028 2029 if (CanUseInputs) { 2030 Value *I4 = HSV->getOperand(1); 2031 if (!isa<UndefValue>(I4)) { 2032 if (!I2 && I4 != I1) 2033 I2 = I4; 2034 else if (I4 != I1 && I4 != I2) 2035 CanUseInputs = false; 2036 } 2037 } 2038 } 2039 2040 if (CanUseInputs) { 2041 unsigned LOpElem = 2042 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2043 ->getNumElements(); 2044 unsigned HOpElem = 2045 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2046 ->getNumElements(); 2047 2048 // We have one or two input vectors. We need to map each index of the 2049 // operands to the index of the original vector. 2050 SmallVector<std::pair<int, int>, 8> II(numElem); 2051 for (unsigned i = 0; i < numElemL; ++i) { 2052 int Idx, INum; 2053 if (LEE) { 2054 Idx = 2055 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2056 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2057 } else { 2058 Idx = LSV->getMaskValue(i); 2059 if (Idx < (int) LOpElem) { 2060 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2061 } else { 2062 Idx -= LOpElem; 2063 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2064 } 2065 } 2066 2067 II[i] = std::pair<int, int>(Idx, INum); 2068 } 2069 for (unsigned i = 0; i < numElemH; ++i) { 2070 int Idx, INum; 2071 if (HEE) { 2072 Idx = 2073 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2074 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2075 } else { 2076 Idx = HSV->getMaskValue(i); 2077 if (Idx < (int) HOpElem) { 2078 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2079 } else { 2080 Idx -= HOpElem; 2081 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2082 } 2083 } 2084 2085 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2086 } 2087 2088 // We now have an array which tells us from which index of which 2089 // input vector each element of the operand comes. 2090 VectorType *I1T = cast<VectorType>(I1->getType()); 2091 unsigned I1Elem = I1T->getNumElements(); 2092 2093 if (!I2) { 2094 // In this case there is only one underlying vector input. Check for 2095 // the trivial case where we can use the input directly. 2096 if (I1Elem == numElem) { 2097 bool ElemInOrder = true; 2098 for (unsigned i = 0; i < numElem; ++i) { 2099 if (II[i].first != (int) i && II[i].first != -1) { 2100 ElemInOrder = false; 2101 break; 2102 } 2103 } 2104 2105 if (ElemInOrder) 2106 return I1; 2107 } 2108 2109 // A shuffle is needed. 2110 std::vector<Constant *> Mask(numElem); 2111 for (unsigned i = 0; i < numElem; ++i) { 2112 int Idx = II[i].first; 2113 if (Idx == -1) 2114 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2115 else 2116 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2117 } 2118 2119 Instruction *S = 2120 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2121 ConstantVector::get(Mask), 2122 getReplacementName(IBeforeJ ? I : J, 2123 true, o)); 2124 S->insertBefore(IBeforeJ ? J : I); 2125 return S; 2126 } 2127 2128 VectorType *I2T = cast<VectorType>(I2->getType()); 2129 unsigned I2Elem = I2T->getNumElements(); 2130 2131 // This input comes from two distinct vectors. The first step is to 2132 // make sure that both vectors are the same length. If not, the 2133 // smaller one will need to grow before they can be shuffled together. 2134 if (I1Elem < I2Elem) { 2135 std::vector<Constant *> Mask(I2Elem); 2136 unsigned v = 0; 2137 for (; v < I1Elem; ++v) 2138 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2139 for (; v < I2Elem; ++v) 2140 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2141 2142 Instruction *NewI1 = 2143 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2144 ConstantVector::get(Mask), 2145 getReplacementName(IBeforeJ ? I : J, 2146 true, o, 1)); 2147 NewI1->insertBefore(IBeforeJ ? J : I); 2148 I1 = NewI1; 2149 I1T = I2T; 2150 I1Elem = I2Elem; 2151 } else if (I1Elem > I2Elem) { 2152 std::vector<Constant *> Mask(I1Elem); 2153 unsigned v = 0; 2154 for (; v < I2Elem; ++v) 2155 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2156 for (; v < I1Elem; ++v) 2157 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2158 2159 Instruction *NewI2 = 2160 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2161 ConstantVector::get(Mask), 2162 getReplacementName(IBeforeJ ? I : J, 2163 true, o, 1)); 2164 NewI2->insertBefore(IBeforeJ ? J : I); 2165 I2 = NewI2; 2166 I2T = I1T; 2167 I2Elem = I1Elem; 2168 } 2169 2170 // Now that both I1 and I2 are the same length we can shuffle them 2171 // together (and use the result). 2172 std::vector<Constant *> Mask(numElem); 2173 for (unsigned v = 0; v < numElem; ++v) { 2174 if (II[v].first == -1) { 2175 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2176 } else { 2177 int Idx = II[v].first + II[v].second * I1Elem; 2178 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2179 } 2180 } 2181 2182 Instruction *NewOp = 2183 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2184 getReplacementName(IBeforeJ ? I : J, true, o)); 2185 NewOp->insertBefore(IBeforeJ ? J : I); 2186 return NewOp; 2187 } 2188 } 2189 2190 Type *ArgType = ArgTypeL; 2191 if (numElemL < numElemH) { 2192 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2193 ArgTypeL, VArgType, IBeforeJ, 1)) { 2194 // This is another short-circuit case: we're combining a scalar into 2195 // a vector that is formed by an IE chain. We've just expanded the IE 2196 // chain, now insert the scalar and we're done. 2197 2198 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2199 getReplacementName(IBeforeJ ? I : J, true, o)); 2200 S->insertBefore(IBeforeJ ? J : I); 2201 return S; 2202 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2203 ArgTypeH, IBeforeJ)) { 2204 // The two vector inputs to the shuffle must be the same length, 2205 // so extend the smaller vector to be the same length as the larger one. 2206 Instruction *NLOp; 2207 if (numElemL > 1) { 2208 2209 std::vector<Constant *> Mask(numElemH); 2210 unsigned v = 0; 2211 for (; v < numElemL; ++v) 2212 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2213 for (; v < numElemH; ++v) 2214 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2215 2216 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2217 ConstantVector::get(Mask), 2218 getReplacementName(IBeforeJ ? I : J, 2219 true, o, 1)); 2220 } else { 2221 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2222 getReplacementName(IBeforeJ ? I : J, 2223 true, o, 1)); 2224 } 2225 2226 NLOp->insertBefore(IBeforeJ ? J : I); 2227 LOp = NLOp; 2228 } 2229 2230 ArgType = ArgTypeH; 2231 } else if (numElemL > numElemH) { 2232 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2233 ArgTypeH, VArgType, IBeforeJ)) { 2234 Instruction *S = 2235 InsertElementInst::Create(LOp, HOp, 2236 ConstantInt::get(Type::getInt32Ty(Context), 2237 numElemL), 2238 getReplacementName(IBeforeJ ? I : J, 2239 true, o)); 2240 S->insertBefore(IBeforeJ ? J : I); 2241 return S; 2242 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2243 ArgTypeL, IBeforeJ)) { 2244 Instruction *NHOp; 2245 if (numElemH > 1) { 2246 std::vector<Constant *> Mask(numElemL); 2247 unsigned v = 0; 2248 for (; v < numElemH; ++v) 2249 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2250 for (; v < numElemL; ++v) 2251 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2252 2253 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2254 ConstantVector::get(Mask), 2255 getReplacementName(IBeforeJ ? I : J, 2256 true, o, 1)); 2257 } else { 2258 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2259 getReplacementName(IBeforeJ ? I : J, 2260 true, o, 1)); 2261 } 2262 2263 NHOp->insertBefore(IBeforeJ ? J : I); 2264 HOp = NHOp; 2265 } 2266 } 2267 2268 if (ArgType->isVectorTy()) { 2269 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2270 std::vector<Constant*> Mask(numElem); 2271 for (unsigned v = 0; v < numElem; ++v) { 2272 unsigned Idx = v; 2273 // If the low vector was expanded, we need to skip the extra 2274 // undefined entries. 2275 if (v >= numElemL && numElemH > numElemL) 2276 Idx += (numElemH - numElemL); 2277 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2278 } 2279 2280 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2281 ConstantVector::get(Mask), 2282 getReplacementName(IBeforeJ ? I : J, true, o)); 2283 BV->insertBefore(IBeforeJ ? J : I); 2284 return BV; 2285 } 2286 2287 Instruction *BV1 = InsertElementInst::Create( 2288 UndefValue::get(VArgType), LOp, CV0, 2289 getReplacementName(IBeforeJ ? I : J, 2290 true, o, 1)); 2291 BV1->insertBefore(IBeforeJ ? J : I); 2292 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2293 getReplacementName(IBeforeJ ? I : J, 2294 true, o, 2)); 2295 BV2->insertBefore(IBeforeJ ? J : I); 2296 return BV2; 2297 } 2298 2299 // This function creates an array of values that will be used as the inputs 2300 // to the vector instruction that fuses I with J. 2301 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2302 Instruction *I, Instruction *J, 2303 SmallVector<Value *, 3> &ReplacedOperands, 2304 bool IBeforeJ) { 2305 unsigned NumOperands = I->getNumOperands(); 2306 2307 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2308 // Iterate backward so that we look at the store pointer 2309 // first and know whether or not we need to flip the inputs. 2310 2311 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2312 // This is the pointer for a load/store instruction. 2313 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2314 continue; 2315 } else if (isa<CallInst>(I)) { 2316 Function *F = cast<CallInst>(I)->getCalledFunction(); 2317 unsigned IID = F->getIntrinsicID(); 2318 if (o == NumOperands-1) { 2319 BasicBlock &BB = *I->getParent(); 2320 2321 Module *M = BB.getParent()->getParent(); 2322 Type *ArgTypeI = I->getType(); 2323 Type *ArgTypeJ = J->getType(); 2324 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2325 2326 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 2327 (Intrinsic::ID) IID, VArgType); 2328 continue; 2329 } else if (IID == Intrinsic::powi && o == 1) { 2330 // The second argument of powi is a single integer and we've already 2331 // checked that both arguments are equal. As a result, we just keep 2332 // I's second argument. 2333 ReplacedOperands[o] = I->getOperand(o); 2334 continue; 2335 } 2336 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2337 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2338 continue; 2339 } 2340 2341 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2342 } 2343 } 2344 2345 // This function creates two values that represent the outputs of the 2346 // original I and J instructions. These are generally vector shuffles 2347 // or extracts. In many cases, these will end up being unused and, thus, 2348 // eliminated by later passes. 2349 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2350 Instruction *J, Instruction *K, 2351 Instruction *&InsertionPt, 2352 Instruction *&K1, Instruction *&K2) { 2353 if (isa<StoreInst>(I)) { 2354 AA->replaceWithNewValue(I, K); 2355 AA->replaceWithNewValue(J, K); 2356 } else { 2357 Type *IType = I->getType(); 2358 Type *JType = J->getType(); 2359 2360 VectorType *VType = getVecTypeForPair(IType, JType); 2361 unsigned numElem = VType->getNumElements(); 2362 2363 unsigned numElemI, numElemJ; 2364 if (IType->isVectorTy()) 2365 numElemI = cast<VectorType>(IType)->getNumElements(); 2366 else 2367 numElemI = 1; 2368 2369 if (JType->isVectorTy()) 2370 numElemJ = cast<VectorType>(JType)->getNumElements(); 2371 else 2372 numElemJ = 1; 2373 2374 if (IType->isVectorTy()) { 2375 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2376 for (unsigned v = 0; v < numElemI; ++v) { 2377 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2378 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2379 } 2380 2381 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2382 ConstantVector::get( Mask1), 2383 getReplacementName(K, false, 1)); 2384 } else { 2385 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2386 K1 = ExtractElementInst::Create(K, CV0, 2387 getReplacementName(K, false, 1)); 2388 } 2389 2390 if (JType->isVectorTy()) { 2391 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2392 for (unsigned v = 0; v < numElemJ; ++v) { 2393 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2394 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2395 } 2396 2397 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2398 ConstantVector::get( Mask2), 2399 getReplacementName(K, false, 2)); 2400 } else { 2401 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2402 K2 = ExtractElementInst::Create(K, CV1, 2403 getReplacementName(K, false, 2)); 2404 } 2405 2406 K1->insertAfter(K); 2407 K2->insertAfter(K1); 2408 InsertionPt = K2; 2409 } 2410 } 2411 2412 // Move all uses of the function I (including pairing-induced uses) after J. 2413 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2414 std::multimap<Value *, Value *> &LoadMoveSet, 2415 Instruction *I, Instruction *J) { 2416 // Skip to the first instruction past I. 2417 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2418 2419 DenseSet<Value *> Users; 2420 AliasSetTracker WriteSet(*AA); 2421 for (; cast<Instruction>(L) != J; ++L) 2422 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 2423 2424 assert(cast<Instruction>(L) == J && 2425 "Tracking has not proceeded far enough to check for dependencies"); 2426 // If J is now in the use set of I, then trackUsesOfI will return true 2427 // and we have a dependency cycle (and the fusing operation must abort). 2428 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 2429 } 2430 2431 // Move all uses of the function I (including pairing-induced uses) after J. 2432 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2433 std::multimap<Value *, Value *> &LoadMoveSet, 2434 Instruction *&InsertionPt, 2435 Instruction *I, Instruction *J) { 2436 // Skip to the first instruction past I. 2437 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2438 2439 DenseSet<Value *> Users; 2440 AliasSetTracker WriteSet(*AA); 2441 for (; cast<Instruction>(L) != J;) { 2442 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 2443 // Move this instruction 2444 Instruction *InstToMove = L; ++L; 2445 2446 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2447 " to after " << *InsertionPt << "\n"); 2448 InstToMove->removeFromParent(); 2449 InstToMove->insertAfter(InsertionPt); 2450 InsertionPt = InstToMove; 2451 } else { 2452 ++L; 2453 } 2454 } 2455 } 2456 2457 // Collect all load instruction that are in the move set of a given first 2458 // pair member. These loads depend on the first instruction, I, and so need 2459 // to be moved after J (the second instruction) when the pair is fused. 2460 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2461 DenseMap<Value *, Value *> &ChosenPairs, 2462 std::multimap<Value *, Value *> &LoadMoveSet, 2463 Instruction *I) { 2464 // Skip to the first instruction past I. 2465 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2466 2467 DenseSet<Value *> Users; 2468 AliasSetTracker WriteSet(*AA); 2469 2470 // Note: We cannot end the loop when we reach J because J could be moved 2471 // farther down the use chain by another instruction pairing. Also, J 2472 // could be before I if this is an inverted input. 2473 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2474 if (trackUsesOfI(Users, WriteSet, I, L)) { 2475 if (L->mayReadFromMemory()) 2476 LoadMoveSet.insert(ValuePair(L, I)); 2477 } 2478 } 2479 } 2480 2481 // In cases where both load/stores and the computation of their pointers 2482 // are chosen for vectorization, we can end up in a situation where the 2483 // aliasing analysis starts returning different query results as the 2484 // process of fusing instruction pairs continues. Because the algorithm 2485 // relies on finding the same use trees here as were found earlier, we'll 2486 // need to precompute the necessary aliasing information here and then 2487 // manually update it during the fusion process. 2488 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2489 std::vector<Value *> &PairableInsts, 2490 DenseMap<Value *, Value *> &ChosenPairs, 2491 std::multimap<Value *, Value *> &LoadMoveSet) { 2492 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2493 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2494 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2495 if (P == ChosenPairs.end()) continue; 2496 2497 Instruction *I = cast<Instruction>(P->first); 2498 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 2499 } 2500 } 2501 2502 // When the first instruction in each pair is cloned, it will inherit its 2503 // parent's metadata. This metadata must be combined with that of the other 2504 // instruction in a safe way. 2505 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2506 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2507 K->getAllMetadataOtherThanDebugLoc(Metadata); 2508 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2509 unsigned Kind = Metadata[i].first; 2510 MDNode *JMD = J->getMetadata(Kind); 2511 MDNode *KMD = Metadata[i].second; 2512 2513 switch (Kind) { 2514 default: 2515 K->setMetadata(Kind, 0); // Remove unknown metadata 2516 break; 2517 case LLVMContext::MD_tbaa: 2518 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2519 break; 2520 case LLVMContext::MD_fpmath: 2521 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2522 break; 2523 } 2524 } 2525 } 2526 2527 // This function fuses the chosen instruction pairs into vector instructions, 2528 // taking care preserve any needed scalar outputs and, then, it reorders the 2529 // remaining instructions as needed (users of the first member of the pair 2530 // need to be moved to after the location of the second member of the pair 2531 // because the vector instruction is inserted in the location of the pair's 2532 // second member). 2533 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2534 std::vector<Value *> &PairableInsts, 2535 DenseMap<Value *, Value *> &ChosenPairs, 2536 DenseSet<ValuePair> &FixedOrderPairs, 2537 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2538 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2539 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2540 LLVMContext& Context = BB.getContext(); 2541 2542 // During the vectorization process, the order of the pairs to be fused 2543 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2544 // list. After a pair is fused, the flipped pair is removed from the list. 2545 DenseSet<ValuePair> FlippedPairs; 2546 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2547 E = ChosenPairs.end(); P != E; ++P) 2548 FlippedPairs.insert(ValuePair(P->second, P->first)); 2549 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2550 E = FlippedPairs.end(); P != E; ++P) 2551 ChosenPairs.insert(*P); 2552 2553 std::multimap<Value *, Value *> LoadMoveSet; 2554 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 2555 2556 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2557 2558 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2559 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2560 if (P == ChosenPairs.end()) { 2561 ++PI; 2562 continue; 2563 } 2564 2565 if (getDepthFactor(P->first) == 0) { 2566 // These instructions are not really fused, but are tracked as though 2567 // they are. Any case in which it would be interesting to fuse them 2568 // will be taken care of by InstCombine. 2569 --NumFusedOps; 2570 ++PI; 2571 continue; 2572 } 2573 2574 Instruction *I = cast<Instruction>(P->first), 2575 *J = cast<Instruction>(P->second); 2576 2577 DEBUG(dbgs() << "BBV: fusing: " << *I << 2578 " <-> " << *J << "\n"); 2579 2580 // Remove the pair and flipped pair from the list. 2581 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2582 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2583 ChosenPairs.erase(FP); 2584 ChosenPairs.erase(P); 2585 2586 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 2587 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2588 " <-> " << *J << 2589 " aborted because of non-trivial dependency cycle\n"); 2590 --NumFusedOps; 2591 ++PI; 2592 continue; 2593 } 2594 2595 // If the pair must have the other order, then flip it. 2596 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2597 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2598 // This pair does not have a fixed order, and so we might want to 2599 // flip it if that will yield fewer shuffles. We count the number 2600 // of dependencies connected via swaps, and those directly connected, 2601 // and flip the order if the number of swaps is greater. 2602 bool OrigOrder = true; 2603 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2604 if (IP.first == ConnectedPairDeps.end()) { 2605 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2606 OrigOrder = false; 2607 } 2608 2609 if (IP.first != ConnectedPairDeps.end()) { 2610 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2611 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2612 Q != IP.second; ++Q) { 2613 DenseMap<VPPair, unsigned>::iterator R = 2614 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2615 assert(R != PairConnectionTypes.end() && 2616 "Cannot find pair connection type"); 2617 if (R->second == PairConnectionDirect) 2618 ++NumDepsDirect; 2619 else if (R->second == PairConnectionSwap) 2620 ++NumDepsSwap; 2621 } 2622 2623 if (!OrigOrder) 2624 std::swap(NumDepsDirect, NumDepsSwap); 2625 2626 if (NumDepsSwap > NumDepsDirect) { 2627 FlipPairOrder = true; 2628 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2629 " <-> " << *J << "\n"); 2630 } 2631 } 2632 } 2633 2634 Instruction *L = I, *H = J; 2635 if (FlipPairOrder) 2636 std::swap(H, L); 2637 2638 // If the pair being fused uses the opposite order from that in the pair 2639 // connection map, then we need to flip the types. 2640 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2641 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2642 Q != IP.second; ++Q) { 2643 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2644 assert(R != PairConnectionTypes.end() && 2645 "Cannot find pair connection type"); 2646 if (R->second == PairConnectionDirect) 2647 R->second = PairConnectionSwap; 2648 else if (R->second == PairConnectionSwap) 2649 R->second = PairConnectionDirect; 2650 } 2651 2652 bool LBeforeH = !FlipPairOrder; 2653 unsigned NumOperands = I->getNumOperands(); 2654 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2655 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2656 LBeforeH); 2657 2658 // Make a copy of the original operation, change its type to the vector 2659 // type and replace its operands with the vector operands. 2660 Instruction *K = L->clone(); 2661 if (L->hasName()) 2662 K->takeName(L); 2663 else if (H->hasName()) 2664 K->takeName(H); 2665 2666 if (!isa<StoreInst>(K)) 2667 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 2668 2669 combineMetadata(K, H); 2670 2671 for (unsigned o = 0; o < NumOperands; ++o) 2672 K->setOperand(o, ReplacedOperands[o]); 2673 2674 K->insertAfter(J); 2675 2676 // Instruction insertion point: 2677 Instruction *InsertionPt = K; 2678 Instruction *K1 = 0, *K2 = 0; 2679 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 2680 2681 // The use tree of the first original instruction must be moved to after 2682 // the location of the second instruction. The entire use tree of the 2683 // first instruction is disjoint from the input tree of the second 2684 // (by definition), and so commutes with it. 2685 2686 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 2687 2688 if (!isa<StoreInst>(I)) { 2689 L->replaceAllUsesWith(K1); 2690 H->replaceAllUsesWith(K2); 2691 AA->replaceWithNewValue(L, K1); 2692 AA->replaceWithNewValue(H, K2); 2693 } 2694 2695 // Instructions that may read from memory may be in the load move set. 2696 // Once an instruction is fused, we no longer need its move set, and so 2697 // the values of the map never need to be updated. However, when a load 2698 // is fused, we need to merge the entries from both instructions in the 2699 // pair in case those instructions were in the move set of some other 2700 // yet-to-be-fused pair. The loads in question are the keys of the map. 2701 if (I->mayReadFromMemory()) { 2702 std::vector<ValuePair> NewSetMembers; 2703 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 2704 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 2705 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 2706 N != IPairRange.second; ++N) 2707 NewSetMembers.push_back(ValuePair(K, N->second)); 2708 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 2709 N != JPairRange.second; ++N) 2710 NewSetMembers.push_back(ValuePair(K, N->second)); 2711 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 2712 AE = NewSetMembers.end(); A != AE; ++A) 2713 LoadMoveSet.insert(*A); 2714 } 2715 2716 // Before removing I, set the iterator to the next instruction. 2717 PI = llvm::next(BasicBlock::iterator(I)); 2718 if (cast<Instruction>(PI) == J) 2719 ++PI; 2720 2721 SE->forgetValue(I); 2722 SE->forgetValue(J); 2723 I->eraseFromParent(); 2724 J->eraseFromParent(); 2725 2726 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 2727 BB << "\n"); 2728 } 2729 2730 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 2731 } 2732} 2733 2734char BBVectorize::ID = 0; 2735static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 2736INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2737INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2738INITIALIZE_PASS_DEPENDENCY(DominatorTree) 2739INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2740INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2741 2742BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 2743 return new BBVectorize(C); 2744} 2745 2746bool 2747llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 2748 BBVectorize BBVectorizer(P, C); 2749 return BBVectorizer.vectorizeBB(BB); 2750} 2751 2752//===----------------------------------------------------------------------===// 2753VectorizeConfig::VectorizeConfig() { 2754 VectorBits = ::VectorBits; 2755 VectorizeBools = !::NoBools; 2756 VectorizeInts = !::NoInts; 2757 VectorizeFloats = !::NoFloats; 2758 VectorizePointers = !::NoPointers; 2759 VectorizeCasts = !::NoCasts; 2760 VectorizeMath = !::NoMath; 2761 VectorizeFMA = !::NoFMA; 2762 VectorizeSelect = !::NoSelect; 2763 VectorizeCmp = !::NoCmp; 2764 VectorizeGEP = !::NoGEP; 2765 VectorizeMemOps = !::NoMemOps; 2766 AlignedOnly = ::AlignedOnly; 2767 ReqChainDepth= ::ReqChainDepth; 2768 SearchLimit = ::SearchLimit; 2769 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 2770 SplatBreaksChain = ::SplatBreaksChain; 2771 MaxInsts = ::MaxInsts; 2772 MaxIter = ::MaxIter; 2773 Pow2LenOnly = ::Pow2LenOnly; 2774 NoMemOpBoost = ::NoMemOpBoost; 2775 FastDep = ::FastDep; 2776} 2777