BBVectorize.cpp revision 65309660fa61a837cc05323f69c618a7d8134d56
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 102static cl::opt<bool> 103NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden, 104 cl::desc("Don't try to vectorize pointer values")); 105 106static cl::opt<bool> 107NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 108 cl::desc("Don't try to vectorize casting (conversion) operations")); 109 110static cl::opt<bool> 111NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 112 cl::desc("Don't try to vectorize floating-point math intrinsics")); 113 114static cl::opt<bool> 115NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 116 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 117 118static cl::opt<bool> 119NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 120 cl::desc("Don't try to vectorize select instructions")); 121 122static cl::opt<bool> 123NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 124 cl::desc("Don't try to vectorize comparison instructions")); 125 126static cl::opt<bool> 127NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 128 cl::desc("Don't try to vectorize getelementptr instructions")); 129 130static cl::opt<bool> 131NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 132 cl::desc("Don't try to vectorize loads and stores")); 133 134static cl::opt<bool> 135AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 136 cl::desc("Only generate aligned loads and stores")); 137 138static cl::opt<bool> 139NoMemOpBoost("bb-vectorize-no-mem-op-boost", 140 cl::init(false), cl::Hidden, 141 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 142 143static cl::opt<bool> 144FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 145 cl::desc("Use a fast instruction dependency analysis")); 146 147#ifndef NDEBUG 148static cl::opt<bool> 149DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 150 cl::init(false), cl::Hidden, 151 cl::desc("When debugging is enabled, output information on the" 152 " instruction-examination process")); 153static cl::opt<bool> 154DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 155 cl::init(false), cl::Hidden, 156 cl::desc("When debugging is enabled, output information on the" 157 " candidate-selection process")); 158static cl::opt<bool> 159DebugPairSelection("bb-vectorize-debug-pair-selection", 160 cl::init(false), cl::Hidden, 161 cl::desc("When debugging is enabled, output information on the" 162 " pair-selection process")); 163static cl::opt<bool> 164DebugCycleCheck("bb-vectorize-debug-cycle-check", 165 cl::init(false), cl::Hidden, 166 cl::desc("When debugging is enabled, output information on the" 167 " cycle-checking process")); 168#endif 169 170STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 171 172namespace { 173 struct BBVectorize : public BasicBlockPass { 174 static char ID; // Pass identification, replacement for typeid 175 176 const VectorizeConfig Config; 177 178 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 179 : BasicBlockPass(ID), Config(C) { 180 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 181 } 182 183 BBVectorize(Pass *P, const VectorizeConfig &C) 184 : BasicBlockPass(ID), Config(C) { 185 AA = &P->getAnalysis<AliasAnalysis>(); 186 DT = &P->getAnalysis<DominatorTree>(); 187 SE = &P->getAnalysis<ScalarEvolution>(); 188 TD = P->getAnalysisIfAvailable<DataLayout>(); 189 TTI = IgnoreTargetInfo ? 0 : 190 P->getAnalysisIfAvailable<TargetTransformInfo>(); 191 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 192 } 193 194 typedef std::pair<Value *, Value *> ValuePair; 195 typedef std::pair<ValuePair, int> ValuePairWithCost; 196 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 197 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 198 typedef std::pair<std::multimap<Value *, Value *>::iterator, 199 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 200 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 201 std::multimap<ValuePair, ValuePair>::iterator> 202 VPPIteratorPair; 203 204 AliasAnalysis *AA; 205 DominatorTree *DT; 206 ScalarEvolution *SE; 207 DataLayout *TD; 208 TargetTransformInfo *TTI; 209 const VectorTargetTransformInfo *VTTI; 210 211 // FIXME: const correct? 212 213 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 214 215 bool getCandidatePairs(BasicBlock &BB, 216 BasicBlock::iterator &Start, 217 std::multimap<Value *, Value *> &CandidatePairs, 218 DenseMap<ValuePair, int> &CandidatePairCostSavings, 219 std::vector<Value *> &PairableInsts, bool NonPow2Len); 220 221 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 222 std::vector<Value *> &PairableInsts, 223 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 224 225 void buildDepMap(BasicBlock &BB, 226 std::multimap<Value *, Value *> &CandidatePairs, 227 std::vector<Value *> &PairableInsts, 228 DenseSet<ValuePair> &PairableInstUsers); 229 230 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 231 DenseMap<ValuePair, int> &CandidatePairCostSavings, 232 std::vector<Value *> &PairableInsts, 233 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 234 DenseSet<ValuePair> &PairableInstUsers, 235 DenseMap<Value *, Value *>& ChosenPairs); 236 237 void fuseChosenPairs(BasicBlock &BB, 238 std::vector<Value *> &PairableInsts, 239 DenseMap<Value *, Value *>& ChosenPairs); 240 241 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 242 243 bool areInstsCompatible(Instruction *I, Instruction *J, 244 bool IsSimpleLoadStore, bool NonPow2Len, 245 int &CostSavings); 246 247 bool trackUsesOfI(DenseSet<Value *> &Users, 248 AliasSetTracker &WriteSet, Instruction *I, 249 Instruction *J, bool UpdateUsers = true, 250 std::multimap<Value *, Value *> *LoadMoveSet = 0); 251 252 void computePairsConnectedTo( 253 std::multimap<Value *, Value *> &CandidatePairs, 254 std::vector<Value *> &PairableInsts, 255 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 256 ValuePair P); 257 258 bool pairsConflict(ValuePair P, ValuePair Q, 259 DenseSet<ValuePair> &PairableInstUsers, 260 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 261 262 bool pairWillFormCycle(ValuePair P, 263 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 264 DenseSet<ValuePair> &CurrentPairs); 265 266 void pruneTreeFor( 267 std::multimap<Value *, Value *> &CandidatePairs, 268 std::vector<Value *> &PairableInsts, 269 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 270 DenseSet<ValuePair> &PairableInstUsers, 271 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 272 DenseMap<Value *, Value *> &ChosenPairs, 273 DenseMap<ValuePair, size_t> &Tree, 274 DenseSet<ValuePair> &PrunedTree, ValuePair J, 275 bool UseCycleCheck); 276 277 void buildInitialTreeFor( 278 std::multimap<Value *, Value *> &CandidatePairs, 279 std::vector<Value *> &PairableInsts, 280 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 281 DenseSet<ValuePair> &PairableInstUsers, 282 DenseMap<Value *, Value *> &ChosenPairs, 283 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 284 285 void findBestTreeFor( 286 std::multimap<Value *, Value *> &CandidatePairs, 287 DenseMap<ValuePair, int> &CandidatePairCostSavings, 288 std::vector<Value *> &PairableInsts, 289 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 290 DenseSet<ValuePair> &PairableInstUsers, 291 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 292 DenseMap<Value *, Value *> &ChosenPairs, 293 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 294 int &BestEffSize, VPIteratorPair ChoiceRange, 295 bool UseCycleCheck); 296 297 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 298 Instruction *J, unsigned o, bool FlipMemInputs); 299 300 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 301 unsigned MaskOffset, unsigned NumInElem, 302 unsigned NumInElem1, unsigned IdxOffset, 303 std::vector<Constant*> &Mask); 304 305 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 306 Instruction *J); 307 308 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 309 unsigned o, Value *&LOp, unsigned numElemL, 310 Type *ArgTypeL, Type *ArgTypeR, 311 unsigned IdxOff = 0); 312 313 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 314 Instruction *J, unsigned o, bool FlipMemInputs); 315 316 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 317 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 318 bool FlipMemInputs); 319 320 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 321 Instruction *J, Instruction *K, 322 Instruction *&InsertionPt, Instruction *&K1, 323 Instruction *&K2, bool FlipMemInputs); 324 325 void collectPairLoadMoveSet(BasicBlock &BB, 326 DenseMap<Value *, Value *> &ChosenPairs, 327 std::multimap<Value *, Value *> &LoadMoveSet, 328 Instruction *I); 329 330 void collectLoadMoveSet(BasicBlock &BB, 331 std::vector<Value *> &PairableInsts, 332 DenseMap<Value *, Value *> &ChosenPairs, 333 std::multimap<Value *, Value *> &LoadMoveSet); 334 335 void collectPtrInfo(std::vector<Value *> &PairableInsts, 336 DenseMap<Value *, Value *> &ChosenPairs, 337 DenseSet<Value *> &LowPtrInsts); 338 339 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 340 std::multimap<Value *, Value *> &LoadMoveSet, 341 Instruction *I, Instruction *J); 342 343 void moveUsesOfIAfterJ(BasicBlock &BB, 344 std::multimap<Value *, Value *> &LoadMoveSet, 345 Instruction *&InsertionPt, 346 Instruction *I, Instruction *J); 347 348 void combineMetadata(Instruction *K, const Instruction *J); 349 350 bool vectorizeBB(BasicBlock &BB) { 351 if (!DT->isReachableFromEntry(&BB)) { 352 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 353 " in " << BB.getParent()->getName() << "\n"); 354 return false; 355 } 356 357 DEBUG(if (VTTI) dbgs() << "BBV: using target information\n"); 358 359 bool changed = false; 360 // Iterate a sufficient number of times to merge types of size 1 bit, 361 // then 2 bits, then 4, etc. up to half of the target vector width of the 362 // target vector register. 363 unsigned n = 1; 364 for (unsigned v = 2; 365 (VTTI || v <= Config.VectorBits) && 366 (!Config.MaxIter || n <= Config.MaxIter); 367 v *= 2, ++n) { 368 DEBUG(dbgs() << "BBV: fusing loop #" << n << 369 " for " << BB.getName() << " in " << 370 BB.getParent()->getName() << "...\n"); 371 if (vectorizePairs(BB)) 372 changed = true; 373 else 374 break; 375 } 376 377 if (changed && !Pow2LenOnly) { 378 ++n; 379 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 380 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 381 n << " for " << BB.getName() << " in " << 382 BB.getParent()->getName() << "...\n"); 383 if (!vectorizePairs(BB, true)) break; 384 } 385 } 386 387 DEBUG(dbgs() << "BBV: done!\n"); 388 return changed; 389 } 390 391 virtual bool runOnBasicBlock(BasicBlock &BB) { 392 AA = &getAnalysis<AliasAnalysis>(); 393 DT = &getAnalysis<DominatorTree>(); 394 SE = &getAnalysis<ScalarEvolution>(); 395 TD = getAnalysisIfAvailable<DataLayout>(); 396 TTI = IgnoreTargetInfo ? 0 : 397 getAnalysisIfAvailable<TargetTransformInfo>(); 398 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 399 400 return vectorizeBB(BB); 401 } 402 403 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 404 BasicBlockPass::getAnalysisUsage(AU); 405 AU.addRequired<AliasAnalysis>(); 406 AU.addRequired<DominatorTree>(); 407 AU.addRequired<ScalarEvolution>(); 408 AU.addPreserved<AliasAnalysis>(); 409 AU.addPreserved<DominatorTree>(); 410 AU.addPreserved<ScalarEvolution>(); 411 AU.setPreservesCFG(); 412 } 413 414 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 415 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 416 "Cannot form vector from incompatible scalar types"); 417 Type *STy = ElemTy->getScalarType(); 418 419 unsigned numElem; 420 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 421 numElem = VTy->getNumElements(); 422 } else { 423 numElem = 1; 424 } 425 426 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 427 numElem += VTy->getNumElements(); 428 } else { 429 numElem += 1; 430 } 431 432 return VectorType::get(STy, numElem); 433 } 434 435 static inline void getInstructionTypes(Instruction *I, 436 Type *&T1, Type *&T2) { 437 if (isa<StoreInst>(I)) { 438 // For stores, it is the value type, not the pointer type that matters 439 // because the value is what will come from a vector register. 440 441 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 442 T1 = IVal->getType(); 443 } else { 444 T1 = I->getType(); 445 } 446 447 if (I->isCast()) 448 T2 = cast<CastInst>(I)->getSrcTy(); 449 else 450 T2 = T1; 451 452 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 453 T2 = SI->getCondition()->getType(); 454 } 455 } 456 457 // Returns the weight associated with the provided value. A chain of 458 // candidate pairs has a length given by the sum of the weights of its 459 // members (one weight per pair; the weight of each member of the pair 460 // is assumed to be the same). This length is then compared to the 461 // chain-length threshold to determine if a given chain is significant 462 // enough to be vectorized. The length is also used in comparing 463 // candidate chains where longer chains are considered to be better. 464 // Note: when this function returns 0, the resulting instructions are 465 // not actually fused. 466 inline size_t getDepthFactor(Value *V) { 467 // InsertElement and ExtractElement have a depth factor of zero. This is 468 // for two reasons: First, they cannot be usefully fused. Second, because 469 // the pass generates a lot of these, they can confuse the simple metric 470 // used to compare the trees in the next iteration. Thus, giving them a 471 // weight of zero allows the pass to essentially ignore them in 472 // subsequent iterations when looking for vectorization opportunities 473 // while still tracking dependency chains that flow through those 474 // instructions. 475 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 476 return 0; 477 478 // Give a load or store half of the required depth so that load/store 479 // pairs will vectorize. 480 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 481 return Config.ReqChainDepth/2; 482 483 return 1; 484 } 485 486 // This determines the relative offset of two loads or stores, returning 487 // true if the offset could be determined to be some constant value. 488 // For example, if OffsetInElmts == 1, then J accesses the memory directly 489 // after I; if OffsetInElmts == -1 then I accesses the memory 490 // directly after J. 491 bool getPairPtrInfo(Instruction *I, Instruction *J, 492 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 493 unsigned &IAddressSpace, unsigned &JAddressSpace, 494 int64_t &OffsetInElmts) { 495 OffsetInElmts = 0; 496 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 497 LoadInst *LJ = cast<LoadInst>(J); 498 IPtr = LI->getPointerOperand(); 499 JPtr = LJ->getPointerOperand(); 500 IAlignment = LI->getAlignment(); 501 JAlignment = LJ->getAlignment(); 502 IAddressSpace = LI->getPointerAddressSpace(); 503 JAddressSpace = LJ->getPointerAddressSpace(); 504 } else { 505 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 506 IPtr = SI->getPointerOperand(); 507 JPtr = SJ->getPointerOperand(); 508 IAlignment = SI->getAlignment(); 509 JAlignment = SJ->getAlignment(); 510 IAddressSpace = SI->getPointerAddressSpace(); 511 JAddressSpace = SJ->getPointerAddressSpace(); 512 } 513 514 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 515 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 516 517 // If this is a trivial offset, then we'll get something like 518 // 1*sizeof(type). With target data, which we need anyway, this will get 519 // constant folded into a number. 520 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 521 if (const SCEVConstant *ConstOffSCEV = 522 dyn_cast<SCEVConstant>(OffsetSCEV)) { 523 ConstantInt *IntOff = ConstOffSCEV->getValue(); 524 int64_t Offset = IntOff->getSExtValue(); 525 526 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 527 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 528 529 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 530 if (VTy != VTy2 && Offset < 0) { 531 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 532 OffsetInElmts = Offset/VTy2TSS; 533 return (abs64(Offset) % VTy2TSS) == 0; 534 } 535 536 OffsetInElmts = Offset/VTyTSS; 537 return (abs64(Offset) % VTyTSS) == 0; 538 } 539 540 return false; 541 } 542 543 // Returns true if the provided CallInst represents an intrinsic that can 544 // be vectorized. 545 bool isVectorizableIntrinsic(CallInst* I) { 546 Function *F = I->getCalledFunction(); 547 if (!F) return false; 548 549 unsigned IID = F->getIntrinsicID(); 550 if (!IID) return false; 551 552 switch(IID) { 553 default: 554 return false; 555 case Intrinsic::sqrt: 556 case Intrinsic::powi: 557 case Intrinsic::sin: 558 case Intrinsic::cos: 559 case Intrinsic::log: 560 case Intrinsic::log2: 561 case Intrinsic::log10: 562 case Intrinsic::exp: 563 case Intrinsic::exp2: 564 case Intrinsic::pow: 565 return Config.VectorizeMath; 566 case Intrinsic::fma: 567 return Config.VectorizeFMA; 568 } 569 } 570 571 // Returns true if J is the second element in some pair referenced by 572 // some multimap pair iterator pair. 573 template <typename V> 574 bool isSecondInIteratorPair(V J, std::pair< 575 typename std::multimap<V, V>::iterator, 576 typename std::multimap<V, V>::iterator> PairRange) { 577 for (typename std::multimap<V, V>::iterator K = PairRange.first; 578 K != PairRange.second; ++K) 579 if (K->second == J) return true; 580 581 return false; 582 } 583 }; 584 585 // This function implements one vectorization iteration on the provided 586 // basic block. It returns true if the block is changed. 587 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 588 bool ShouldContinue; 589 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 590 591 std::vector<Value *> AllPairableInsts; 592 DenseMap<Value *, Value *> AllChosenPairs; 593 594 do { 595 std::vector<Value *> PairableInsts; 596 std::multimap<Value *, Value *> CandidatePairs; 597 DenseMap<ValuePair, int> CandidatePairCostSavings; 598 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 599 CandidatePairCostSavings, 600 PairableInsts, NonPow2Len); 601 if (PairableInsts.empty()) continue; 602 603 // Now we have a map of all of the pairable instructions and we need to 604 // select the best possible pairing. A good pairing is one such that the 605 // users of the pair are also paired. This defines a (directed) forest 606 // over the pairs such that two pairs are connected iff the second pair 607 // uses the first. 608 609 // Note that it only matters that both members of the second pair use some 610 // element of the first pair (to allow for splatting). 611 612 std::multimap<ValuePair, ValuePair> ConnectedPairs; 613 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 614 if (ConnectedPairs.empty()) continue; 615 616 // Build the pairable-instruction dependency map 617 DenseSet<ValuePair> PairableInstUsers; 618 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 619 620 // There is now a graph of the connected pairs. For each variable, pick 621 // the pairing with the largest tree meeting the depth requirement on at 622 // least one branch. Then select all pairings that are part of that tree 623 // and remove them from the list of available pairings and pairable 624 // variables. 625 626 DenseMap<Value *, Value *> ChosenPairs; 627 choosePairs(CandidatePairs, CandidatePairCostSavings, 628 PairableInsts, ConnectedPairs, 629 PairableInstUsers, ChosenPairs); 630 631 if (ChosenPairs.empty()) continue; 632 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 633 PairableInsts.end()); 634 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 635 } while (ShouldContinue); 636 637 if (AllChosenPairs.empty()) return false; 638 NumFusedOps += AllChosenPairs.size(); 639 640 // A set of pairs has now been selected. It is now necessary to replace the 641 // paired instructions with vector instructions. For this procedure each 642 // operand must be replaced with a vector operand. This vector is formed 643 // by using build_vector on the old operands. The replaced values are then 644 // replaced with a vector_extract on the result. Subsequent optimization 645 // passes should coalesce the build/extract combinations. 646 647 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 648 649 // It is important to cleanup here so that future iterations of this 650 // function have less work to do. 651 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 652 return true; 653 } 654 655 // This function returns true if the provided instruction is capable of being 656 // fused into a vector instruction. This determination is based only on the 657 // type and other attributes of the instruction. 658 bool BBVectorize::isInstVectorizable(Instruction *I, 659 bool &IsSimpleLoadStore) { 660 IsSimpleLoadStore = false; 661 662 if (CallInst *C = dyn_cast<CallInst>(I)) { 663 if (!isVectorizableIntrinsic(C)) 664 return false; 665 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 666 // Vectorize simple loads if possbile: 667 IsSimpleLoadStore = L->isSimple(); 668 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 669 return false; 670 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 671 // Vectorize simple stores if possbile: 672 IsSimpleLoadStore = S->isSimple(); 673 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 674 return false; 675 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 676 // We can vectorize casts, but not casts of pointer types, etc. 677 if (!Config.VectorizeCasts) 678 return false; 679 680 Type *SrcTy = C->getSrcTy(); 681 if (!SrcTy->isSingleValueType()) 682 return false; 683 684 Type *DestTy = C->getDestTy(); 685 if (!DestTy->isSingleValueType()) 686 return false; 687 } else if (isa<SelectInst>(I)) { 688 if (!Config.VectorizeSelect) 689 return false; 690 } else if (isa<CmpInst>(I)) { 691 if (!Config.VectorizeCmp) 692 return false; 693 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 694 if (!Config.VectorizeGEP) 695 return false; 696 697 // Currently, vector GEPs exist only with one index. 698 if (G->getNumIndices() != 1) 699 return false; 700 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 701 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 702 return false; 703 } 704 705 // We can't vectorize memory operations without target data 706 if (TD == 0 && IsSimpleLoadStore) 707 return false; 708 709 Type *T1, *T2; 710 getInstructionTypes(I, T1, T2); 711 712 // Not every type can be vectorized... 713 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 714 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 715 return false; 716 717 if (T1->getScalarSizeInBits() == 1) { 718 if (!Config.VectorizeBools) 719 return false; 720 } else { 721 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 722 return false; 723 } 724 725 if (T2->getScalarSizeInBits() == 1) { 726 if (!Config.VectorizeBools) 727 return false; 728 } else { 729 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 730 return false; 731 } 732 733 if (!Config.VectorizeFloats 734 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 735 return false; 736 737 // Don't vectorize target-specific types. 738 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 739 return false; 740 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 741 return false; 742 743 if ((!Config.VectorizePointers || TD == 0) && 744 (T1->getScalarType()->isPointerTy() || 745 T2->getScalarType()->isPointerTy())) 746 return false; 747 748 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 749 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 750 return false; 751 752 return true; 753 } 754 755 // This function returns true if the two provided instructions are compatible 756 // (meaning that they can be fused into a vector instruction). This assumes 757 // that I has already been determined to be vectorizable and that J is not 758 // in the use tree of I. 759 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 760 bool IsSimpleLoadStore, bool NonPow2Len, 761 int &CostSavings) { 762 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 763 " <-> " << *J << "\n"); 764 765 CostSavings = 0; 766 767 // Loads and stores can be merged if they have different alignments, 768 // but are otherwise the same. 769 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 770 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 771 return false; 772 773 Type *IT1, *IT2, *JT1, *JT2; 774 getInstructionTypes(I, IT1, IT2); 775 getInstructionTypes(J, JT1, JT2); 776 unsigned MaxTypeBits = std::max( 777 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 778 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 779 if (!VTTI && MaxTypeBits > Config.VectorBits) 780 return false; 781 782 // FIXME: handle addsub-type operations! 783 784 if (IsSimpleLoadStore) { 785 Value *IPtr, *JPtr; 786 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 787 int64_t OffsetInElmts = 0; 788 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 789 IAddressSpace, JAddressSpace, 790 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 791 unsigned BottomAlignment = IAlignment; 792 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 793 794 Type *aTypeI = isa<StoreInst>(I) ? 795 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 796 Type *aTypeJ = isa<StoreInst>(J) ? 797 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 798 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 799 800 if (Config.AlignedOnly) { 801 // An aligned load or store is possible only if the instruction 802 // with the lower offset has an alignment suitable for the 803 // vector type. 804 805 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 806 if (BottomAlignment < VecAlignment) 807 return false; 808 } 809 810 if (VTTI) { 811 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(), 812 IAlignment, IAddressSpace); 813 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(), 814 JAlignment, JAddressSpace); 815 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType, 816 BottomAlignment, 817 IAddressSpace); 818 if (VCost > ICost + JCost) 819 return false; 820 CostSavings = ICost + JCost - VCost; 821 } 822 } else { 823 return false; 824 } 825 } else if (VTTI) { 826 unsigned ICost = VTTI->getInstrCost(I->getOpcode(), IT1, IT2); 827 unsigned JCost = VTTI->getInstrCost(J->getOpcode(), JT1, JT2); 828 Type *VT1 = getVecTypeForPair(IT1, JT1), 829 *VT2 = getVecTypeForPair(IT2, JT2); 830 unsigned VCost = VTTI->getInstrCost(I->getOpcode(), VT1, VT2); 831 832 if (VCost > ICost + JCost) 833 return false; 834 CostSavings = ICost + JCost - VCost; 835 } 836 837 // The powi intrinsic is special because only the first argument is 838 // vectorized, the second arguments must be equal. 839 CallInst *CI = dyn_cast<CallInst>(I); 840 Function *FI; 841 if (CI && (FI = CI->getCalledFunction()) && 842 FI->getIntrinsicID() == Intrinsic::powi) { 843 844 Value *A1I = CI->getArgOperand(1), 845 *A1J = cast<CallInst>(J)->getArgOperand(1); 846 const SCEV *A1ISCEV = SE->getSCEV(A1I), 847 *A1JSCEV = SE->getSCEV(A1J); 848 return (A1ISCEV == A1JSCEV); 849 } 850 851 return true; 852 } 853 854 // Figure out whether or not J uses I and update the users and write-set 855 // structures associated with I. Specifically, Users represents the set of 856 // instructions that depend on I. WriteSet represents the set 857 // of memory locations that are dependent on I. If UpdateUsers is true, 858 // and J uses I, then Users is updated to contain J and WriteSet is updated 859 // to contain any memory locations to which J writes. The function returns 860 // true if J uses I. By default, alias analysis is used to determine 861 // whether J reads from memory that overlaps with a location in WriteSet. 862 // If LoadMoveSet is not null, then it is a previously-computed multimap 863 // where the key is the memory-based user instruction and the value is 864 // the instruction to be compared with I. So, if LoadMoveSet is provided, 865 // then the alias analysis is not used. This is necessary because this 866 // function is called during the process of moving instructions during 867 // vectorization and the results of the alias analysis are not stable during 868 // that process. 869 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 870 AliasSetTracker &WriteSet, Instruction *I, 871 Instruction *J, bool UpdateUsers, 872 std::multimap<Value *, Value *> *LoadMoveSet) { 873 bool UsesI = false; 874 875 // This instruction may already be marked as a user due, for example, to 876 // being a member of a selected pair. 877 if (Users.count(J)) 878 UsesI = true; 879 880 if (!UsesI) 881 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 882 JU != JE; ++JU) { 883 Value *V = *JU; 884 if (I == V || Users.count(V)) { 885 UsesI = true; 886 break; 887 } 888 } 889 if (!UsesI && J->mayReadFromMemory()) { 890 if (LoadMoveSet) { 891 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 892 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 893 } else { 894 for (AliasSetTracker::iterator W = WriteSet.begin(), 895 WE = WriteSet.end(); W != WE; ++W) { 896 if (W->aliasesUnknownInst(J, *AA)) { 897 UsesI = true; 898 break; 899 } 900 } 901 } 902 } 903 904 if (UsesI && UpdateUsers) { 905 if (J->mayWriteToMemory()) WriteSet.add(J); 906 Users.insert(J); 907 } 908 909 return UsesI; 910 } 911 912 // This function iterates over all instruction pairs in the provided 913 // basic block and collects all candidate pairs for vectorization. 914 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 915 BasicBlock::iterator &Start, 916 std::multimap<Value *, Value *> &CandidatePairs, 917 DenseMap<ValuePair, int> &CandidatePairCostSavings, 918 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 919 BasicBlock::iterator E = BB.end(); 920 if (Start == E) return false; 921 922 bool ShouldContinue = false, IAfterStart = false; 923 for (BasicBlock::iterator I = Start++; I != E; ++I) { 924 if (I == Start) IAfterStart = true; 925 926 bool IsSimpleLoadStore; 927 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 928 929 // Look for an instruction with which to pair instruction *I... 930 DenseSet<Value *> Users; 931 AliasSetTracker WriteSet(*AA); 932 bool JAfterStart = IAfterStart; 933 BasicBlock::iterator J = llvm::next(I); 934 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 935 if (J == Start) JAfterStart = true; 936 937 // Determine if J uses I, if so, exit the loop. 938 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 939 if (Config.FastDep) { 940 // Note: For this heuristic to be effective, independent operations 941 // must tend to be intermixed. This is likely to be true from some 942 // kinds of grouped loop unrolling (but not the generic LLVM pass), 943 // but otherwise may require some kind of reordering pass. 944 945 // When using fast dependency analysis, 946 // stop searching after first use: 947 if (UsesI) break; 948 } else { 949 if (UsesI) continue; 950 } 951 952 // J does not use I, and comes before the first use of I, so it can be 953 // merged with I if the instructions are compatible. 954 int CostSavings; 955 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 956 CostSavings)) continue; 957 958 // J is a candidate for merging with I. 959 if (!PairableInsts.size() || 960 PairableInsts[PairableInsts.size()-1] != I) { 961 PairableInsts.push_back(I); 962 } 963 964 CandidatePairs.insert(ValuePair(I, J)); 965 if (VTTI) 966 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 967 CostSavings)); 968 969 // The next call to this function must start after the last instruction 970 // selected during this invocation. 971 if (JAfterStart) { 972 Start = llvm::next(J); 973 IAfterStart = JAfterStart = false; 974 } 975 976 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 977 << *I << " <-> " << *J << " (cost savings: " << 978 CostSavings << ")\n"); 979 980 // If we have already found too many pairs, break here and this function 981 // will be called again starting after the last instruction selected 982 // during this invocation. 983 if (PairableInsts.size() >= Config.MaxInsts) { 984 ShouldContinue = true; 985 break; 986 } 987 } 988 989 if (ShouldContinue) 990 break; 991 } 992 993 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 994 << " instructions with candidate pairs\n"); 995 996 return ShouldContinue; 997 } 998 999 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1000 // it looks for pairs such that both members have an input which is an 1001 // output of PI or PJ. 1002 void BBVectorize::computePairsConnectedTo( 1003 std::multimap<Value *, Value *> &CandidatePairs, 1004 std::vector<Value *> &PairableInsts, 1005 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1006 ValuePair P) { 1007 StoreInst *SI, *SJ; 1008 1009 // For each possible pairing for this variable, look at the uses of 1010 // the first value... 1011 for (Value::use_iterator I = P.first->use_begin(), 1012 E = P.first->use_end(); I != E; ++I) { 1013 if (isa<LoadInst>(*I)) { 1014 // A pair cannot be connected to a load because the load only takes one 1015 // operand (the address) and it is a scalar even after vectorization. 1016 continue; 1017 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1018 P.first == SI->getPointerOperand()) { 1019 // Similarly, a pair cannot be connected to a store through its 1020 // pointer operand. 1021 continue; 1022 } 1023 1024 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1025 1026 // For each use of the first variable, look for uses of the second 1027 // variable... 1028 for (Value::use_iterator J = P.second->use_begin(), 1029 E2 = P.second->use_end(); J != E2; ++J) { 1030 if ((SJ = dyn_cast<StoreInst>(*J)) && 1031 P.second == SJ->getPointerOperand()) 1032 continue; 1033 1034 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 1035 1036 // Look for <I, J>: 1037 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 1038 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 1039 1040 // Look for <J, I>: 1041 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 1042 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 1043 } 1044 1045 if (Config.SplatBreaksChain) continue; 1046 // Look for cases where just the first value in the pair is used by 1047 // both members of another pair (splatting). 1048 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1049 if ((SJ = dyn_cast<StoreInst>(*J)) && 1050 P.first == SJ->getPointerOperand()) 1051 continue; 1052 1053 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 1054 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 1055 } 1056 } 1057 1058 if (Config.SplatBreaksChain) return; 1059 // Look for cases where just the second value in the pair is used by 1060 // both members of another pair (splatting). 1061 for (Value::use_iterator I = P.second->use_begin(), 1062 E = P.second->use_end(); I != E; ++I) { 1063 if (isa<LoadInst>(*I)) 1064 continue; 1065 else if ((SI = dyn_cast<StoreInst>(*I)) && 1066 P.second == SI->getPointerOperand()) 1067 continue; 1068 1069 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1070 1071 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1072 if ((SJ = dyn_cast<StoreInst>(*J)) && 1073 P.second == SJ->getPointerOperand()) 1074 continue; 1075 1076 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 1077 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 1078 } 1079 } 1080 } 1081 1082 // This function figures out which pairs are connected. Two pairs are 1083 // connected if some output of the first pair forms an input to both members 1084 // of the second pair. 1085 void BBVectorize::computeConnectedPairs( 1086 std::multimap<Value *, Value *> &CandidatePairs, 1087 std::vector<Value *> &PairableInsts, 1088 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 1089 1090 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1091 PE = PairableInsts.end(); PI != PE; ++PI) { 1092 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1093 1094 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1095 P != choiceRange.second; ++P) 1096 computePairsConnectedTo(CandidatePairs, PairableInsts, 1097 ConnectedPairs, *P); 1098 } 1099 1100 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1101 << " pair connections.\n"); 1102 } 1103 1104 // This function builds a set of use tuples such that <A, B> is in the set 1105 // if B is in the use tree of A. If B is in the use tree of A, then B 1106 // depends on the output of A. 1107 void BBVectorize::buildDepMap( 1108 BasicBlock &BB, 1109 std::multimap<Value *, Value *> &CandidatePairs, 1110 std::vector<Value *> &PairableInsts, 1111 DenseSet<ValuePair> &PairableInstUsers) { 1112 DenseSet<Value *> IsInPair; 1113 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1114 E = CandidatePairs.end(); C != E; ++C) { 1115 IsInPair.insert(C->first); 1116 IsInPair.insert(C->second); 1117 } 1118 1119 // Iterate through the basic block, recording all Users of each 1120 // pairable instruction. 1121 1122 BasicBlock::iterator E = BB.end(); 1123 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1124 if (IsInPair.find(I) == IsInPair.end()) continue; 1125 1126 DenseSet<Value *> Users; 1127 AliasSetTracker WriteSet(*AA); 1128 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1129 (void) trackUsesOfI(Users, WriteSet, I, J); 1130 1131 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1132 U != E; ++U) 1133 PairableInstUsers.insert(ValuePair(I, *U)); 1134 } 1135 } 1136 1137 // Returns true if an input to pair P is an output of pair Q and also an 1138 // input of pair Q is an output of pair P. If this is the case, then these 1139 // two pairs cannot be simultaneously fused. 1140 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1141 DenseSet<ValuePair> &PairableInstUsers, 1142 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1143 // Two pairs are in conflict if they are mutual Users of eachother. 1144 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1145 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1146 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1147 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1148 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1149 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1150 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1151 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1152 if (PairableInstUserMap) { 1153 // FIXME: The expensive part of the cycle check is not so much the cycle 1154 // check itself but this edge insertion procedure. This needs some 1155 // profiling and probably a different data structure (same is true of 1156 // most uses of std::multimap). 1157 if (PUsesQ) { 1158 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1159 if (!isSecondInIteratorPair(P, QPairRange)) 1160 PairableInstUserMap->insert(VPPair(Q, P)); 1161 } 1162 if (QUsesP) { 1163 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1164 if (!isSecondInIteratorPair(Q, PPairRange)) 1165 PairableInstUserMap->insert(VPPair(P, Q)); 1166 } 1167 } 1168 1169 return (QUsesP && PUsesQ); 1170 } 1171 1172 // This function walks the use graph of current pairs to see if, starting 1173 // from P, the walk returns to P. 1174 bool BBVectorize::pairWillFormCycle(ValuePair P, 1175 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1176 DenseSet<ValuePair> &CurrentPairs) { 1177 DEBUG(if (DebugCycleCheck) 1178 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1179 << *P.second << "\n"); 1180 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1181 // contains non-direct associations. 1182 DenseSet<ValuePair> Visited; 1183 SmallVector<ValuePair, 32> Q; 1184 // General depth-first post-order traversal: 1185 Q.push_back(P); 1186 do { 1187 ValuePair QTop = Q.pop_back_val(); 1188 Visited.insert(QTop); 1189 1190 DEBUG(if (DebugCycleCheck) 1191 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1192 << *QTop.second << "\n"); 1193 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1194 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1195 C != QPairRange.second; ++C) { 1196 if (C->second == P) { 1197 DEBUG(dbgs() 1198 << "BBV: rejected to prevent non-trivial cycle formation: " 1199 << *C->first.first << " <-> " << *C->first.second << "\n"); 1200 return true; 1201 } 1202 1203 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1204 Q.push_back(C->second); 1205 } 1206 } while (!Q.empty()); 1207 1208 return false; 1209 } 1210 1211 // This function builds the initial tree of connected pairs with the 1212 // pair J at the root. 1213 void BBVectorize::buildInitialTreeFor( 1214 std::multimap<Value *, Value *> &CandidatePairs, 1215 std::vector<Value *> &PairableInsts, 1216 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1217 DenseSet<ValuePair> &PairableInstUsers, 1218 DenseMap<Value *, Value *> &ChosenPairs, 1219 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1220 // Each of these pairs is viewed as the root node of a Tree. The Tree 1221 // is then walked (depth-first). As this happens, we keep track of 1222 // the pairs that compose the Tree and the maximum depth of the Tree. 1223 SmallVector<ValuePairWithDepth, 32> Q; 1224 // General depth-first post-order traversal: 1225 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1226 do { 1227 ValuePairWithDepth QTop = Q.back(); 1228 1229 // Push each child onto the queue: 1230 bool MoreChildren = false; 1231 size_t MaxChildDepth = QTop.second; 1232 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1233 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1234 k != qtRange.second; ++k) { 1235 // Make sure that this child pair is still a candidate: 1236 bool IsStillCand = false; 1237 VPIteratorPair checkRange = 1238 CandidatePairs.equal_range(k->second.first); 1239 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1240 m != checkRange.second; ++m) { 1241 if (m->second == k->second.second) { 1242 IsStillCand = true; 1243 break; 1244 } 1245 } 1246 1247 if (IsStillCand) { 1248 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1249 if (C == Tree.end()) { 1250 size_t d = getDepthFactor(k->second.first); 1251 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1252 MoreChildren = true; 1253 } else { 1254 MaxChildDepth = std::max(MaxChildDepth, C->second); 1255 } 1256 } 1257 } 1258 1259 if (!MoreChildren) { 1260 // Record the current pair as part of the Tree: 1261 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1262 Q.pop_back(); 1263 } 1264 } while (!Q.empty()); 1265 } 1266 1267 // Given some initial tree, prune it by removing conflicting pairs (pairs 1268 // that cannot be simultaneously chosen for vectorization). 1269 void BBVectorize::pruneTreeFor( 1270 std::multimap<Value *, Value *> &CandidatePairs, 1271 std::vector<Value *> &PairableInsts, 1272 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1273 DenseSet<ValuePair> &PairableInstUsers, 1274 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1275 DenseMap<Value *, Value *> &ChosenPairs, 1276 DenseMap<ValuePair, size_t> &Tree, 1277 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1278 bool UseCycleCheck) { 1279 SmallVector<ValuePairWithDepth, 32> Q; 1280 // General depth-first post-order traversal: 1281 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1282 do { 1283 ValuePairWithDepth QTop = Q.pop_back_val(); 1284 PrunedTree.insert(QTop.first); 1285 1286 // Visit each child, pruning as necessary... 1287 DenseMap<ValuePair, size_t> BestChildren; 1288 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1289 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1290 K != QTopRange.second; ++K) { 1291 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1292 if (C == Tree.end()) continue; 1293 1294 // This child is in the Tree, now we need to make sure it is the 1295 // best of any conflicting children. There could be multiple 1296 // conflicting children, so first, determine if we're keeping 1297 // this child, then delete conflicting children as necessary. 1298 1299 // It is also necessary to guard against pairing-induced 1300 // dependencies. Consider instructions a .. x .. y .. b 1301 // such that (a,b) are to be fused and (x,y) are to be fused 1302 // but a is an input to x and b is an output from y. This 1303 // means that y cannot be moved after b but x must be moved 1304 // after b for (a,b) to be fused. In other words, after 1305 // fusing (a,b) we have y .. a/b .. x where y is an input 1306 // to a/b and x is an output to a/b: x and y can no longer 1307 // be legally fused. To prevent this condition, we must 1308 // make sure that a child pair added to the Tree is not 1309 // both an input and output of an already-selected pair. 1310 1311 // Pairing-induced dependencies can also form from more complicated 1312 // cycles. The pair vs. pair conflicts are easy to check, and so 1313 // that is done explicitly for "fast rejection", and because for 1314 // child vs. child conflicts, we may prefer to keep the current 1315 // pair in preference to the already-selected child. 1316 DenseSet<ValuePair> CurrentPairs; 1317 1318 bool CanAdd = true; 1319 for (DenseMap<ValuePair, size_t>::iterator C2 1320 = BestChildren.begin(), E2 = BestChildren.end(); 1321 C2 != E2; ++C2) { 1322 if (C2->first.first == C->first.first || 1323 C2->first.first == C->first.second || 1324 C2->first.second == C->first.first || 1325 C2->first.second == C->first.second || 1326 pairsConflict(C2->first, C->first, PairableInstUsers, 1327 UseCycleCheck ? &PairableInstUserMap : 0)) { 1328 if (C2->second >= C->second) { 1329 CanAdd = false; 1330 break; 1331 } 1332 1333 CurrentPairs.insert(C2->first); 1334 } 1335 } 1336 if (!CanAdd) continue; 1337 1338 // Even worse, this child could conflict with another node already 1339 // selected for the Tree. If that is the case, ignore this child. 1340 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1341 E2 = PrunedTree.end(); T != E2; ++T) { 1342 if (T->first == C->first.first || 1343 T->first == C->first.second || 1344 T->second == C->first.first || 1345 T->second == C->first.second || 1346 pairsConflict(*T, C->first, PairableInstUsers, 1347 UseCycleCheck ? &PairableInstUserMap : 0)) { 1348 CanAdd = false; 1349 break; 1350 } 1351 1352 CurrentPairs.insert(*T); 1353 } 1354 if (!CanAdd) continue; 1355 1356 // And check the queue too... 1357 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1358 E2 = Q.end(); C2 != E2; ++C2) { 1359 if (C2->first.first == C->first.first || 1360 C2->first.first == C->first.second || 1361 C2->first.second == C->first.first || 1362 C2->first.second == C->first.second || 1363 pairsConflict(C2->first, C->first, PairableInstUsers, 1364 UseCycleCheck ? &PairableInstUserMap : 0)) { 1365 CanAdd = false; 1366 break; 1367 } 1368 1369 CurrentPairs.insert(C2->first); 1370 } 1371 if (!CanAdd) continue; 1372 1373 // Last but not least, check for a conflict with any of the 1374 // already-chosen pairs. 1375 for (DenseMap<Value *, Value *>::iterator C2 = 1376 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1377 C2 != E2; ++C2) { 1378 if (pairsConflict(*C2, C->first, PairableInstUsers, 1379 UseCycleCheck ? &PairableInstUserMap : 0)) { 1380 CanAdd = false; 1381 break; 1382 } 1383 1384 CurrentPairs.insert(*C2); 1385 } 1386 if (!CanAdd) continue; 1387 1388 // To check for non-trivial cycles formed by the addition of the 1389 // current pair we've formed a list of all relevant pairs, now use a 1390 // graph walk to check for a cycle. We start from the current pair and 1391 // walk the use tree to see if we again reach the current pair. If we 1392 // do, then the current pair is rejected. 1393 1394 // FIXME: It may be more efficient to use a topological-ordering 1395 // algorithm to improve the cycle check. This should be investigated. 1396 if (UseCycleCheck && 1397 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1398 continue; 1399 1400 // This child can be added, but we may have chosen it in preference 1401 // to an already-selected child. Check for this here, and if a 1402 // conflict is found, then remove the previously-selected child 1403 // before adding this one in its place. 1404 for (DenseMap<ValuePair, size_t>::iterator C2 1405 = BestChildren.begin(); C2 != BestChildren.end();) { 1406 if (C2->first.first == C->first.first || 1407 C2->first.first == C->first.second || 1408 C2->first.second == C->first.first || 1409 C2->first.second == C->first.second || 1410 pairsConflict(C2->first, C->first, PairableInstUsers)) 1411 BestChildren.erase(C2++); 1412 else 1413 ++C2; 1414 } 1415 1416 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1417 } 1418 1419 for (DenseMap<ValuePair, size_t>::iterator C 1420 = BestChildren.begin(), E2 = BestChildren.end(); 1421 C != E2; ++C) { 1422 size_t DepthF = getDepthFactor(C->first.first); 1423 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1424 } 1425 } while (!Q.empty()); 1426 } 1427 1428 // This function finds the best tree of mututally-compatible connected 1429 // pairs, given the choice of root pairs as an iterator range. 1430 void BBVectorize::findBestTreeFor( 1431 std::multimap<Value *, Value *> &CandidatePairs, 1432 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1433 std::vector<Value *> &PairableInsts, 1434 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1435 DenseSet<ValuePair> &PairableInstUsers, 1436 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1437 DenseMap<Value *, Value *> &ChosenPairs, 1438 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1439 int &BestEffSize, VPIteratorPair ChoiceRange, 1440 bool UseCycleCheck) { 1441 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1442 J != ChoiceRange.second; ++J) { 1443 1444 // Before going any further, make sure that this pair does not 1445 // conflict with any already-selected pairs (see comment below 1446 // near the Tree pruning for more details). 1447 DenseSet<ValuePair> ChosenPairSet; 1448 bool DoesConflict = false; 1449 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1450 E = ChosenPairs.end(); C != E; ++C) { 1451 if (pairsConflict(*C, *J, PairableInstUsers, 1452 UseCycleCheck ? &PairableInstUserMap : 0)) { 1453 DoesConflict = true; 1454 break; 1455 } 1456 1457 ChosenPairSet.insert(*C); 1458 } 1459 if (DoesConflict) continue; 1460 1461 if (UseCycleCheck && 1462 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1463 continue; 1464 1465 DenseMap<ValuePair, size_t> Tree; 1466 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1467 PairableInstUsers, ChosenPairs, Tree, *J); 1468 1469 // Because we'll keep the child with the largest depth, the largest 1470 // depth is still the same in the unpruned Tree. 1471 size_t MaxDepth = Tree.lookup(*J); 1472 1473 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1474 << *J->first << " <-> " << *J->second << "} of depth " << 1475 MaxDepth << " and size " << Tree.size() << "\n"); 1476 1477 // At this point the Tree has been constructed, but, may contain 1478 // contradictory children (meaning that different children of 1479 // some tree node may be attempting to fuse the same instruction). 1480 // So now we walk the tree again, in the case of a conflict, 1481 // keep only the child with the largest depth. To break a tie, 1482 // favor the first child. 1483 1484 DenseSet<ValuePair> PrunedTree; 1485 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1486 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1487 PrunedTree, *J, UseCycleCheck); 1488 1489 int EffSize = 0; 1490 if (VTTI) { 1491 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1492 E = PrunedTree.end(); S != E; ++S) { 1493 if (getDepthFactor(S->first)) 1494 EffSize += CandidatePairCostSavings.find(*S)->second; 1495 } 1496 } else { 1497 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1498 E = PrunedTree.end(); S != E; ++S) 1499 EffSize += (int) getDepthFactor(S->first); 1500 } 1501 1502 DEBUG(if (DebugPairSelection) 1503 dbgs() << "BBV: found pruned Tree for pair {" 1504 << *J->first << " <-> " << *J->second << "} of depth " << 1505 MaxDepth << " and size " << PrunedTree.size() << 1506 " (effective size: " << EffSize << ")\n"); 1507 if (MaxDepth >= Config.ReqChainDepth && 1508 EffSize > 0 && EffSize > BestEffSize) { 1509 BestMaxDepth = MaxDepth; 1510 BestEffSize = EffSize; 1511 BestTree = PrunedTree; 1512 } 1513 } 1514 } 1515 1516 // Given the list of candidate pairs, this function selects those 1517 // that will be fused into vector instructions. 1518 void BBVectorize::choosePairs( 1519 std::multimap<Value *, Value *> &CandidatePairs, 1520 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1521 std::vector<Value *> &PairableInsts, 1522 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1523 DenseSet<ValuePair> &PairableInstUsers, 1524 DenseMap<Value *, Value *>& ChosenPairs) { 1525 bool UseCycleCheck = 1526 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1527 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1528 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1529 E = PairableInsts.end(); I != E; ++I) { 1530 // The number of possible pairings for this variable: 1531 size_t NumChoices = CandidatePairs.count(*I); 1532 if (!NumChoices) continue; 1533 1534 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1535 1536 // The best pair to choose and its tree: 1537 size_t BestMaxDepth = 0; 1538 int BestEffSize = 0; 1539 DenseSet<ValuePair> BestTree; 1540 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 1541 PairableInsts, ConnectedPairs, 1542 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1543 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1544 UseCycleCheck); 1545 1546 // A tree has been chosen (or not) at this point. If no tree was 1547 // chosen, then this instruction, I, cannot be paired (and is no longer 1548 // considered). 1549 1550 DEBUG(if (BestTree.size() > 0) 1551 dbgs() << "BBV: selected pairs in the best tree for: " 1552 << *cast<Instruction>(*I) << "\n"); 1553 1554 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1555 SE2 = BestTree.end(); S != SE2; ++S) { 1556 // Insert the members of this tree into the list of chosen pairs. 1557 ChosenPairs.insert(ValuePair(S->first, S->second)); 1558 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1559 *S->second << "\n"); 1560 1561 // Remove all candidate pairs that have values in the chosen tree. 1562 for (std::multimap<Value *, Value *>::iterator K = 1563 CandidatePairs.begin(); K != CandidatePairs.end();) { 1564 if (K->first == S->first || K->second == S->first || 1565 K->second == S->second || K->first == S->second) { 1566 // Don't remove the actual pair chosen so that it can be used 1567 // in subsequent tree selections. 1568 if (!(K->first == S->first && K->second == S->second)) 1569 CandidatePairs.erase(K++); 1570 else 1571 ++K; 1572 } else { 1573 ++K; 1574 } 1575 } 1576 } 1577 } 1578 1579 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1580 } 1581 1582 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1583 unsigned n = 0) { 1584 if (!I->hasName()) 1585 return ""; 1586 1587 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1588 (n > 0 ? "." + utostr(n) : "")).str(); 1589 } 1590 1591 // Returns the value that is to be used as the pointer input to the vector 1592 // instruction that fuses I with J. 1593 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1594 Instruction *I, Instruction *J, unsigned o, 1595 bool FlipMemInputs) { 1596 Value *IPtr, *JPtr; 1597 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 1598 int64_t OffsetInElmts; 1599 1600 // Note: the analysis might fail here, that is why FlipMemInputs has 1601 // been precomputed (OffsetInElmts must be unused here). 1602 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1603 IAddressSpace, JAddressSpace, 1604 OffsetInElmts); 1605 1606 // The pointer value is taken to be the one with the lowest offset. 1607 Value *VPtr; 1608 if (!FlipMemInputs) { 1609 VPtr = IPtr; 1610 } else { 1611 VPtr = JPtr; 1612 } 1613 1614 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 1615 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 1616 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1617 Type *VArgPtrType = PointerType::get(VArgType, 1618 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1619 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1620 /* insert before */ FlipMemInputs ? J : I); 1621 } 1622 1623 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1624 unsigned MaskOffset, unsigned NumInElem, 1625 unsigned NumInElem1, unsigned IdxOffset, 1626 std::vector<Constant*> &Mask) { 1627 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 1628 for (unsigned v = 0; v < NumElem1; ++v) { 1629 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1630 if (m < 0) { 1631 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1632 } else { 1633 unsigned mm = m + (int) IdxOffset; 1634 if (m >= (int) NumInElem1) 1635 mm += (int) NumInElem; 1636 1637 Mask[v+MaskOffset] = 1638 ConstantInt::get(Type::getInt32Ty(Context), mm); 1639 } 1640 } 1641 } 1642 1643 // Returns the value that is to be used as the vector-shuffle mask to the 1644 // vector instruction that fuses I with J. 1645 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1646 Instruction *I, Instruction *J) { 1647 // This is the shuffle mask. We need to append the second 1648 // mask to the first, and the numbers need to be adjusted. 1649 1650 Type *ArgTypeI = I->getType(); 1651 Type *ArgTypeJ = J->getType(); 1652 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1653 1654 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 1655 1656 // Get the total number of elements in the fused vector type. 1657 // By definition, this must equal the number of elements in 1658 // the final mask. 1659 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1660 std::vector<Constant*> Mask(NumElem); 1661 1662 Type *OpTypeI = I->getOperand(0)->getType(); 1663 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 1664 Type *OpTypeJ = J->getOperand(0)->getType(); 1665 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 1666 1667 // The fused vector will be: 1668 // ----------------------------------------------------- 1669 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 1670 // ----------------------------------------------------- 1671 // from which we'll extract NumElem total elements (where the first NumElemI 1672 // of them come from the mask in I and the remainder come from the mask 1673 // in J. 1674 1675 // For the mask from the first pair... 1676 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 1677 0, Mask); 1678 1679 // For the mask from the second pair... 1680 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 1681 NumInElemI, Mask); 1682 1683 return ConstantVector::get(Mask); 1684 } 1685 1686 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 1687 Instruction *J, unsigned o, Value *&LOp, 1688 unsigned numElemL, 1689 Type *ArgTypeL, Type *ArgTypeH, 1690 unsigned IdxOff) { 1691 bool ExpandedIEChain = false; 1692 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 1693 // If we have a pure insertelement chain, then this can be rewritten 1694 // into a chain that directly builds the larger type. 1695 bool PureChain = true; 1696 InsertElementInst *LIENext = LIE; 1697 do { 1698 if (!isa<UndefValue>(LIENext->getOperand(0)) && 1699 !isa<InsertElementInst>(LIENext->getOperand(0))) { 1700 PureChain = false; 1701 break; 1702 } 1703 } while ((LIENext = 1704 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1705 1706 if (PureChain) { 1707 SmallVector<Value *, 8> VectElemts(numElemL, 1708 UndefValue::get(ArgTypeL->getScalarType())); 1709 InsertElementInst *LIENext = LIE; 1710 do { 1711 unsigned Idx = 1712 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 1713 VectElemts[Idx] = LIENext->getOperand(1); 1714 } while ((LIENext = 1715 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1716 1717 LIENext = 0; 1718 Value *LIEPrev = UndefValue::get(ArgTypeH); 1719 for (unsigned i = 0; i < numElemL; ++i) { 1720 if (isa<UndefValue>(VectElemts[i])) continue; 1721 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 1722 ConstantInt::get(Type::getInt32Ty(Context), 1723 i + IdxOff), 1724 getReplacementName(I, true, o, i+1)); 1725 LIENext->insertBefore(J); 1726 LIEPrev = LIENext; 1727 } 1728 1729 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 1730 ExpandedIEChain = true; 1731 } 1732 } 1733 1734 return ExpandedIEChain; 1735 } 1736 1737 // Returns the value to be used as the specified operand of the vector 1738 // instruction that fuses I with J. 1739 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1740 Instruction *J, unsigned o, bool FlipMemInputs) { 1741 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1742 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1743 1744 // Compute the fused vector type for this operand 1745 Type *ArgTypeI = I->getOperand(o)->getType(); 1746 Type *ArgTypeJ = J->getOperand(o)->getType(); 1747 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1748 1749 Instruction *L = I, *H = J; 1750 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 1751 if (FlipMemInputs) { 1752 L = J; 1753 H = I; 1754 ArgTypeL = ArgTypeJ; 1755 ArgTypeH = ArgTypeI; 1756 } 1757 1758 unsigned numElemL; 1759 if (ArgTypeL->isVectorTy()) 1760 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 1761 else 1762 numElemL = 1; 1763 1764 unsigned numElemH; 1765 if (ArgTypeH->isVectorTy()) 1766 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 1767 else 1768 numElemH = 1; 1769 1770 Value *LOp = L->getOperand(o); 1771 Value *HOp = H->getOperand(o); 1772 unsigned numElem = VArgType->getNumElements(); 1773 1774 // First, we check if we can reuse the "original" vector outputs (if these 1775 // exist). We might need a shuffle. 1776 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 1777 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 1778 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 1779 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 1780 1781 // FIXME: If we're fusing shuffle instructions, then we can't apply this 1782 // optimization. The input vectors to the shuffle might be a different 1783 // length from the shuffle outputs. Unfortunately, the replacement 1784 // shuffle mask has already been formed, and the mask entries are sensitive 1785 // to the sizes of the inputs. 1786 bool IsSizeChangeShuffle = 1787 isa<ShuffleVectorInst>(L) && 1788 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 1789 1790 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 1791 // We can have at most two unique vector inputs. 1792 bool CanUseInputs = true; 1793 Value *I1, *I2 = 0; 1794 if (LEE) { 1795 I1 = LEE->getOperand(0); 1796 } else { 1797 I1 = LSV->getOperand(0); 1798 I2 = LSV->getOperand(1); 1799 if (I2 == I1 || isa<UndefValue>(I2)) 1800 I2 = 0; 1801 } 1802 1803 if (HEE) { 1804 Value *I3 = HEE->getOperand(0); 1805 if (!I2 && I3 != I1) 1806 I2 = I3; 1807 else if (I3 != I1 && I3 != I2) 1808 CanUseInputs = false; 1809 } else { 1810 Value *I3 = HSV->getOperand(0); 1811 if (!I2 && I3 != I1) 1812 I2 = I3; 1813 else if (I3 != I1 && I3 != I2) 1814 CanUseInputs = false; 1815 1816 if (CanUseInputs) { 1817 Value *I4 = HSV->getOperand(1); 1818 if (!isa<UndefValue>(I4)) { 1819 if (!I2 && I4 != I1) 1820 I2 = I4; 1821 else if (I4 != I1 && I4 != I2) 1822 CanUseInputs = false; 1823 } 1824 } 1825 } 1826 1827 if (CanUseInputs) { 1828 unsigned LOpElem = 1829 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 1830 ->getNumElements(); 1831 unsigned HOpElem = 1832 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 1833 ->getNumElements(); 1834 1835 // We have one or two input vectors. We need to map each index of the 1836 // operands to the index of the original vector. 1837 SmallVector<std::pair<int, int>, 8> II(numElem); 1838 for (unsigned i = 0; i < numElemL; ++i) { 1839 int Idx, INum; 1840 if (LEE) { 1841 Idx = 1842 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 1843 INum = LEE->getOperand(0) == I1 ? 0 : 1; 1844 } else { 1845 Idx = LSV->getMaskValue(i); 1846 if (Idx < (int) LOpElem) { 1847 INum = LSV->getOperand(0) == I1 ? 0 : 1; 1848 } else { 1849 Idx -= LOpElem; 1850 INum = LSV->getOperand(1) == I1 ? 0 : 1; 1851 } 1852 } 1853 1854 II[i] = std::pair<int, int>(Idx, INum); 1855 } 1856 for (unsigned i = 0; i < numElemH; ++i) { 1857 int Idx, INum; 1858 if (HEE) { 1859 Idx = 1860 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 1861 INum = HEE->getOperand(0) == I1 ? 0 : 1; 1862 } else { 1863 Idx = HSV->getMaskValue(i); 1864 if (Idx < (int) HOpElem) { 1865 INum = HSV->getOperand(0) == I1 ? 0 : 1; 1866 } else { 1867 Idx -= HOpElem; 1868 INum = HSV->getOperand(1) == I1 ? 0 : 1; 1869 } 1870 } 1871 1872 II[i + numElemL] = std::pair<int, int>(Idx, INum); 1873 } 1874 1875 // We now have an array which tells us from which index of which 1876 // input vector each element of the operand comes. 1877 VectorType *I1T = cast<VectorType>(I1->getType()); 1878 unsigned I1Elem = I1T->getNumElements(); 1879 1880 if (!I2) { 1881 // In this case there is only one underlying vector input. Check for 1882 // the trivial case where we can use the input directly. 1883 if (I1Elem == numElem) { 1884 bool ElemInOrder = true; 1885 for (unsigned i = 0; i < numElem; ++i) { 1886 if (II[i].first != (int) i && II[i].first != -1) { 1887 ElemInOrder = false; 1888 break; 1889 } 1890 } 1891 1892 if (ElemInOrder) 1893 return I1; 1894 } 1895 1896 // A shuffle is needed. 1897 std::vector<Constant *> Mask(numElem); 1898 for (unsigned i = 0; i < numElem; ++i) { 1899 int Idx = II[i].first; 1900 if (Idx == -1) 1901 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 1902 else 1903 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 1904 } 1905 1906 Instruction *S = 1907 new ShuffleVectorInst(I1, UndefValue::get(I1T), 1908 ConstantVector::get(Mask), 1909 getReplacementName(I, true, o)); 1910 S->insertBefore(J); 1911 return S; 1912 } 1913 1914 VectorType *I2T = cast<VectorType>(I2->getType()); 1915 unsigned I2Elem = I2T->getNumElements(); 1916 1917 // This input comes from two distinct vectors. The first step is to 1918 // make sure that both vectors are the same length. If not, the 1919 // smaller one will need to grow before they can be shuffled together. 1920 if (I1Elem < I2Elem) { 1921 std::vector<Constant *> Mask(I2Elem); 1922 unsigned v = 0; 1923 for (; v < I1Elem; ++v) 1924 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1925 for (; v < I2Elem; ++v) 1926 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1927 1928 Instruction *NewI1 = 1929 new ShuffleVectorInst(I1, UndefValue::get(I1T), 1930 ConstantVector::get(Mask), 1931 getReplacementName(I, true, o, 1)); 1932 NewI1->insertBefore(J); 1933 I1 = NewI1; 1934 I1T = I2T; 1935 I1Elem = I2Elem; 1936 } else if (I1Elem > I2Elem) { 1937 std::vector<Constant *> Mask(I1Elem); 1938 unsigned v = 0; 1939 for (; v < I2Elem; ++v) 1940 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1941 for (; v < I1Elem; ++v) 1942 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1943 1944 Instruction *NewI2 = 1945 new ShuffleVectorInst(I2, UndefValue::get(I2T), 1946 ConstantVector::get(Mask), 1947 getReplacementName(I, true, o, 1)); 1948 NewI2->insertBefore(J); 1949 I2 = NewI2; 1950 I2T = I1T; 1951 I2Elem = I1Elem; 1952 } 1953 1954 // Now that both I1 and I2 are the same length we can shuffle them 1955 // together (and use the result). 1956 std::vector<Constant *> Mask(numElem); 1957 for (unsigned v = 0; v < numElem; ++v) { 1958 if (II[v].first == -1) { 1959 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1960 } else { 1961 int Idx = II[v].first + II[v].second * I1Elem; 1962 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 1963 } 1964 } 1965 1966 Instruction *NewOp = 1967 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 1968 getReplacementName(I, true, o)); 1969 NewOp->insertBefore(J); 1970 return NewOp; 1971 } 1972 } 1973 1974 Type *ArgType = ArgTypeL; 1975 if (numElemL < numElemH) { 1976 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 1977 ArgTypeL, VArgType, 1)) { 1978 // This is another short-circuit case: we're combining a scalar into 1979 // a vector that is formed by an IE chain. We've just expanded the IE 1980 // chain, now insert the scalar and we're done. 1981 1982 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 1983 getReplacementName(I, true, o)); 1984 S->insertBefore(J); 1985 return S; 1986 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 1987 ArgTypeH)) { 1988 // The two vector inputs to the shuffle must be the same length, 1989 // so extend the smaller vector to be the same length as the larger one. 1990 Instruction *NLOp; 1991 if (numElemL > 1) { 1992 1993 std::vector<Constant *> Mask(numElemH); 1994 unsigned v = 0; 1995 for (; v < numElemL; ++v) 1996 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1997 for (; v < numElemH; ++v) 1998 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1999 2000 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2001 ConstantVector::get(Mask), 2002 getReplacementName(I, true, o, 1)); 2003 } else { 2004 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2005 getReplacementName(I, true, o, 1)); 2006 } 2007 2008 NLOp->insertBefore(J); 2009 LOp = NLOp; 2010 } 2011 2012 ArgType = ArgTypeH; 2013 } else if (numElemL > numElemH) { 2014 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2015 ArgTypeH, VArgType)) { 2016 Instruction *S = 2017 InsertElementInst::Create(LOp, HOp, 2018 ConstantInt::get(Type::getInt32Ty(Context), 2019 numElemL), 2020 getReplacementName(I, true, o)); 2021 S->insertBefore(J); 2022 return S; 2023 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2024 ArgTypeL)) { 2025 Instruction *NHOp; 2026 if (numElemH > 1) { 2027 std::vector<Constant *> Mask(numElemL); 2028 unsigned v = 0; 2029 for (; v < numElemH; ++v) 2030 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2031 for (; v < numElemL; ++v) 2032 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2033 2034 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2035 ConstantVector::get(Mask), 2036 getReplacementName(I, true, o, 1)); 2037 } else { 2038 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2039 getReplacementName(I, true, o, 1)); 2040 } 2041 2042 NHOp->insertBefore(J); 2043 HOp = NHOp; 2044 } 2045 } 2046 2047 if (ArgType->isVectorTy()) { 2048 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2049 std::vector<Constant*> Mask(numElem); 2050 for (unsigned v = 0; v < numElem; ++v) { 2051 unsigned Idx = v; 2052 // If the low vector was expanded, we need to skip the extra 2053 // undefined entries. 2054 if (v >= numElemL && numElemH > numElemL) 2055 Idx += (numElemH - numElemL); 2056 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2057 } 2058 2059 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2060 ConstantVector::get(Mask), 2061 getReplacementName(I, true, o)); 2062 BV->insertBefore(J); 2063 return BV; 2064 } 2065 2066 Instruction *BV1 = InsertElementInst::Create( 2067 UndefValue::get(VArgType), LOp, CV0, 2068 getReplacementName(I, true, o, 1)); 2069 BV1->insertBefore(I); 2070 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2071 getReplacementName(I, true, o, 2)); 2072 BV2->insertBefore(J); 2073 return BV2; 2074 } 2075 2076 // This function creates an array of values that will be used as the inputs 2077 // to the vector instruction that fuses I with J. 2078 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2079 Instruction *I, Instruction *J, 2080 SmallVector<Value *, 3> &ReplacedOperands, 2081 bool FlipMemInputs) { 2082 unsigned NumOperands = I->getNumOperands(); 2083 2084 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2085 // Iterate backward so that we look at the store pointer 2086 // first and know whether or not we need to flip the inputs. 2087 2088 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2089 // This is the pointer for a load/store instruction. 2090 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 2091 FlipMemInputs); 2092 continue; 2093 } else if (isa<CallInst>(I)) { 2094 Function *F = cast<CallInst>(I)->getCalledFunction(); 2095 unsigned IID = F->getIntrinsicID(); 2096 if (o == NumOperands-1) { 2097 BasicBlock &BB = *I->getParent(); 2098 2099 Module *M = BB.getParent()->getParent(); 2100 Type *ArgTypeI = I->getType(); 2101 Type *ArgTypeJ = J->getType(); 2102 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2103 2104 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 2105 (Intrinsic::ID) IID, VArgType); 2106 continue; 2107 } else if (IID == Intrinsic::powi && o == 1) { 2108 // The second argument of powi is a single integer and we've already 2109 // checked that both arguments are equal. As a result, we just keep 2110 // I's second argument. 2111 ReplacedOperands[o] = I->getOperand(o); 2112 continue; 2113 } 2114 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2115 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2116 continue; 2117 } 2118 2119 ReplacedOperands[o] = 2120 getReplacementInput(Context, I, J, o, FlipMemInputs); 2121 } 2122 } 2123 2124 // This function creates two values that represent the outputs of the 2125 // original I and J instructions. These are generally vector shuffles 2126 // or extracts. In many cases, these will end up being unused and, thus, 2127 // eliminated by later passes. 2128 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2129 Instruction *J, Instruction *K, 2130 Instruction *&InsertionPt, 2131 Instruction *&K1, Instruction *&K2, 2132 bool FlipMemInputs) { 2133 if (isa<StoreInst>(I)) { 2134 AA->replaceWithNewValue(I, K); 2135 AA->replaceWithNewValue(J, K); 2136 } else { 2137 Type *IType = I->getType(); 2138 Type *JType = J->getType(); 2139 2140 VectorType *VType = getVecTypeForPair(IType, JType); 2141 unsigned numElem = VType->getNumElements(); 2142 2143 unsigned numElemI, numElemJ; 2144 if (IType->isVectorTy()) 2145 numElemI = cast<VectorType>(IType)->getNumElements(); 2146 else 2147 numElemI = 1; 2148 2149 if (JType->isVectorTy()) 2150 numElemJ = cast<VectorType>(JType)->getNumElements(); 2151 else 2152 numElemJ = 1; 2153 2154 if (IType->isVectorTy()) { 2155 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2156 for (unsigned v = 0; v < numElemI; ++v) { 2157 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2158 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2159 } 2160 2161 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2162 ConstantVector::get( 2163 FlipMemInputs ? Mask2 : Mask1), 2164 getReplacementName(K, false, 1)); 2165 } else { 2166 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2167 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2168 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 2169 getReplacementName(K, false, 1)); 2170 } 2171 2172 if (JType->isVectorTy()) { 2173 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2174 for (unsigned v = 0; v < numElemJ; ++v) { 2175 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2176 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2177 } 2178 2179 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2180 ConstantVector::get( 2181 FlipMemInputs ? Mask1 : Mask2), 2182 getReplacementName(K, false, 2)); 2183 } else { 2184 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2185 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2186 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 2187 getReplacementName(K, false, 2)); 2188 } 2189 2190 K1->insertAfter(K); 2191 K2->insertAfter(K1); 2192 InsertionPt = K2; 2193 } 2194 } 2195 2196 // Move all uses of the function I (including pairing-induced uses) after J. 2197 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2198 std::multimap<Value *, Value *> &LoadMoveSet, 2199 Instruction *I, Instruction *J) { 2200 // Skip to the first instruction past I. 2201 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2202 2203 DenseSet<Value *> Users; 2204 AliasSetTracker WriteSet(*AA); 2205 for (; cast<Instruction>(L) != J; ++L) 2206 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 2207 2208 assert(cast<Instruction>(L) == J && 2209 "Tracking has not proceeded far enough to check for dependencies"); 2210 // If J is now in the use set of I, then trackUsesOfI will return true 2211 // and we have a dependency cycle (and the fusing operation must abort). 2212 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 2213 } 2214 2215 // Move all uses of the function I (including pairing-induced uses) after J. 2216 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2217 std::multimap<Value *, Value *> &LoadMoveSet, 2218 Instruction *&InsertionPt, 2219 Instruction *I, Instruction *J) { 2220 // Skip to the first instruction past I. 2221 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2222 2223 DenseSet<Value *> Users; 2224 AliasSetTracker WriteSet(*AA); 2225 for (; cast<Instruction>(L) != J;) { 2226 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 2227 // Move this instruction 2228 Instruction *InstToMove = L; ++L; 2229 2230 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2231 " to after " << *InsertionPt << "\n"); 2232 InstToMove->removeFromParent(); 2233 InstToMove->insertAfter(InsertionPt); 2234 InsertionPt = InstToMove; 2235 } else { 2236 ++L; 2237 } 2238 } 2239 } 2240 2241 // Collect all load instruction that are in the move set of a given first 2242 // pair member. These loads depend on the first instruction, I, and so need 2243 // to be moved after J (the second instruction) when the pair is fused. 2244 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2245 DenseMap<Value *, Value *> &ChosenPairs, 2246 std::multimap<Value *, Value *> &LoadMoveSet, 2247 Instruction *I) { 2248 // Skip to the first instruction past I. 2249 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2250 2251 DenseSet<Value *> Users; 2252 AliasSetTracker WriteSet(*AA); 2253 2254 // Note: We cannot end the loop when we reach J because J could be moved 2255 // farther down the use chain by another instruction pairing. Also, J 2256 // could be before I if this is an inverted input. 2257 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2258 if (trackUsesOfI(Users, WriteSet, I, L)) { 2259 if (L->mayReadFromMemory()) 2260 LoadMoveSet.insert(ValuePair(L, I)); 2261 } 2262 } 2263 } 2264 2265 // In cases where both load/stores and the computation of their pointers 2266 // are chosen for vectorization, we can end up in a situation where the 2267 // aliasing analysis starts returning different query results as the 2268 // process of fusing instruction pairs continues. Because the algorithm 2269 // relies on finding the same use trees here as were found earlier, we'll 2270 // need to precompute the necessary aliasing information here and then 2271 // manually update it during the fusion process. 2272 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2273 std::vector<Value *> &PairableInsts, 2274 DenseMap<Value *, Value *> &ChosenPairs, 2275 std::multimap<Value *, Value *> &LoadMoveSet) { 2276 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2277 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2278 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2279 if (P == ChosenPairs.end()) continue; 2280 2281 Instruction *I = cast<Instruction>(P->first); 2282 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 2283 } 2284 } 2285 2286 // As with the aliasing information, SCEV can also change because of 2287 // vectorization. This information is used to compute relative pointer 2288 // offsets; the necessary information will be cached here prior to 2289 // fusion. 2290 void BBVectorize::collectPtrInfo(std::vector<Value *> &PairableInsts, 2291 DenseMap<Value *, Value *> &ChosenPairs, 2292 DenseSet<Value *> &LowPtrInsts) { 2293 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2294 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2295 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2296 if (P == ChosenPairs.end()) continue; 2297 2298 Instruction *I = cast<Instruction>(P->first); 2299 Instruction *J = cast<Instruction>(P->second); 2300 2301 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 2302 continue; 2303 2304 Value *IPtr, *JPtr; 2305 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2306 int64_t OffsetInElmts; 2307 if (!getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2308 IAddressSpace, JAddressSpace, 2309 OffsetInElmts) || abs64(OffsetInElmts) != 1) 2310 llvm_unreachable("Pre-fusion pointer analysis failed"); 2311 2312 Value *LowPI = (OffsetInElmts > 0) ? I : J; 2313 LowPtrInsts.insert(LowPI); 2314 } 2315 } 2316 2317 // When the first instruction in each pair is cloned, it will inherit its 2318 // parent's metadata. This metadata must be combined with that of the other 2319 // instruction in a safe way. 2320 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2321 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2322 K->getAllMetadataOtherThanDebugLoc(Metadata); 2323 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2324 unsigned Kind = Metadata[i].first; 2325 MDNode *JMD = J->getMetadata(Kind); 2326 MDNode *KMD = Metadata[i].second; 2327 2328 switch (Kind) { 2329 default: 2330 K->setMetadata(Kind, 0); // Remove unknown metadata 2331 break; 2332 case LLVMContext::MD_tbaa: 2333 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2334 break; 2335 case LLVMContext::MD_fpmath: 2336 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2337 break; 2338 } 2339 } 2340 } 2341 2342 // This function fuses the chosen instruction pairs into vector instructions, 2343 // taking care preserve any needed scalar outputs and, then, it reorders the 2344 // remaining instructions as needed (users of the first member of the pair 2345 // need to be moved to after the location of the second member of the pair 2346 // because the vector instruction is inserted in the location of the pair's 2347 // second member). 2348 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2349 std::vector<Value *> &PairableInsts, 2350 DenseMap<Value *, Value *> &ChosenPairs) { 2351 LLVMContext& Context = BB.getContext(); 2352 2353 // During the vectorization process, the order of the pairs to be fused 2354 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2355 // list. After a pair is fused, the flipped pair is removed from the list. 2356 std::vector<ValuePair> FlippedPairs; 2357 FlippedPairs.reserve(ChosenPairs.size()); 2358 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2359 E = ChosenPairs.end(); P != E; ++P) 2360 FlippedPairs.push_back(ValuePair(P->second, P->first)); 2361 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 2362 E = FlippedPairs.end(); P != E; ++P) 2363 ChosenPairs.insert(*P); 2364 2365 std::multimap<Value *, Value *> LoadMoveSet; 2366 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 2367 2368 DenseSet<Value *> LowPtrInsts; 2369 collectPtrInfo(PairableInsts, ChosenPairs, LowPtrInsts); 2370 2371 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2372 2373 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2374 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2375 if (P == ChosenPairs.end()) { 2376 ++PI; 2377 continue; 2378 } 2379 2380 if (getDepthFactor(P->first) == 0) { 2381 // These instructions are not really fused, but are tracked as though 2382 // they are. Any case in which it would be interesting to fuse them 2383 // will be taken care of by InstCombine. 2384 --NumFusedOps; 2385 ++PI; 2386 continue; 2387 } 2388 2389 Instruction *I = cast<Instruction>(P->first), 2390 *J = cast<Instruction>(P->second); 2391 2392 DEBUG(dbgs() << "BBV: fusing: " << *I << 2393 " <-> " << *J << "\n"); 2394 2395 // Remove the pair and flipped pair from the list. 2396 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2397 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2398 ChosenPairs.erase(FP); 2399 ChosenPairs.erase(P); 2400 2401 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 2402 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2403 " <-> " << *J << 2404 " aborted because of non-trivial dependency cycle\n"); 2405 --NumFusedOps; 2406 ++PI; 2407 continue; 2408 } 2409 2410 bool FlipMemInputs = false; 2411 if (isa<LoadInst>(I) || isa<StoreInst>(I)) 2412 FlipMemInputs = (LowPtrInsts.find(I) == LowPtrInsts.end()); 2413 2414 unsigned NumOperands = I->getNumOperands(); 2415 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2416 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 2417 FlipMemInputs); 2418 2419 // Make a copy of the original operation, change its type to the vector 2420 // type and replace its operands with the vector operands. 2421 Instruction *K = I->clone(); 2422 if (I->hasName()) K->takeName(I); 2423 2424 if (!isa<StoreInst>(K)) 2425 K->mutateType(getVecTypeForPair(I->getType(), J->getType())); 2426 2427 combineMetadata(K, J); 2428 2429 for (unsigned o = 0; o < NumOperands; ++o) 2430 K->setOperand(o, ReplacedOperands[o]); 2431 2432 // If we've flipped the memory inputs, make sure that we take the correct 2433 // alignment. 2434 if (FlipMemInputs) { 2435 if (isa<StoreInst>(K)) 2436 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 2437 else 2438 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 2439 } 2440 2441 K->insertAfter(J); 2442 2443 // Instruction insertion point: 2444 Instruction *InsertionPt = K; 2445 Instruction *K1 = 0, *K2 = 0; 2446 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 2447 FlipMemInputs); 2448 2449 // The use tree of the first original instruction must be moved to after 2450 // the location of the second instruction. The entire use tree of the 2451 // first instruction is disjoint from the input tree of the second 2452 // (by definition), and so commutes with it. 2453 2454 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 2455 2456 if (!isa<StoreInst>(I)) { 2457 I->replaceAllUsesWith(K1); 2458 J->replaceAllUsesWith(K2); 2459 AA->replaceWithNewValue(I, K1); 2460 AA->replaceWithNewValue(J, K2); 2461 } 2462 2463 // Instructions that may read from memory may be in the load move set. 2464 // Once an instruction is fused, we no longer need its move set, and so 2465 // the values of the map never need to be updated. However, when a load 2466 // is fused, we need to merge the entries from both instructions in the 2467 // pair in case those instructions were in the move set of some other 2468 // yet-to-be-fused pair. The loads in question are the keys of the map. 2469 if (I->mayReadFromMemory()) { 2470 std::vector<ValuePair> NewSetMembers; 2471 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 2472 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 2473 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 2474 N != IPairRange.second; ++N) 2475 NewSetMembers.push_back(ValuePair(K, N->second)); 2476 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 2477 N != JPairRange.second; ++N) 2478 NewSetMembers.push_back(ValuePair(K, N->second)); 2479 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 2480 AE = NewSetMembers.end(); A != AE; ++A) 2481 LoadMoveSet.insert(*A); 2482 } 2483 2484 // Before removing I, set the iterator to the next instruction. 2485 PI = llvm::next(BasicBlock::iterator(I)); 2486 if (cast<Instruction>(PI) == J) 2487 ++PI; 2488 2489 SE->forgetValue(I); 2490 SE->forgetValue(J); 2491 I->eraseFromParent(); 2492 J->eraseFromParent(); 2493 } 2494 2495 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 2496 } 2497} 2498 2499char BBVectorize::ID = 0; 2500static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 2501INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2502INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2503INITIALIZE_PASS_DEPENDENCY(DominatorTree) 2504INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2505INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2506 2507BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 2508 return new BBVectorize(C); 2509} 2510 2511bool 2512llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 2513 BBVectorize BBVectorizer(P, C); 2514 return BBVectorizer.vectorizeBB(BB); 2515} 2516 2517//===----------------------------------------------------------------------===// 2518VectorizeConfig::VectorizeConfig() { 2519 VectorBits = ::VectorBits; 2520 VectorizeBools = !::NoBools; 2521 VectorizeInts = !::NoInts; 2522 VectorizeFloats = !::NoFloats; 2523 VectorizePointers = !::NoPointers; 2524 VectorizeCasts = !::NoCasts; 2525 VectorizeMath = !::NoMath; 2526 VectorizeFMA = !::NoFMA; 2527 VectorizeSelect = !::NoSelect; 2528 VectorizeCmp = !::NoCmp; 2529 VectorizeGEP = !::NoGEP; 2530 VectorizeMemOps = !::NoMemOps; 2531 AlignedOnly = ::AlignedOnly; 2532 ReqChainDepth= ::ReqChainDepth; 2533 SearchLimit = ::SearchLimit; 2534 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 2535 SplatBreaksChain = ::SplatBreaksChain; 2536 MaxInsts = ::MaxInsts; 2537 MaxIter = ::MaxIter; 2538 Pow2LenOnly = ::Pow2LenOnly; 2539 NoMemOpBoost = ::NoMemOpBoost; 2540 FastDep = ::FastDep; 2541} 2542