BBVectorize.cpp revision dc330f75b732b4ce1beace69ae7ed8e19d89bd9f
1//===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements a basic-block vectorization pass. The algorithm was 11// inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral, 12// et al. It works by looking for chains of pairable operations and then 13// pairing them. 14// 15//===----------------------------------------------------------------------===// 16 17#define BBV_NAME "bb-vectorize" 18#define DEBUG_TYPE BBV_NAME 19#include "llvm/Constants.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Function.h" 22#include "llvm/Instructions.h" 23#include "llvm/IntrinsicInst.h" 24#include "llvm/Intrinsics.h" 25#include "llvm/LLVMContext.h" 26#include "llvm/Metadata.h" 27#include "llvm/Pass.h" 28#include "llvm/Type.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/DenseSet.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/ADT/STLExtras.h" 34#include "llvm/ADT/StringExtras.h" 35#include "llvm/Analysis/AliasAnalysis.h" 36#include "llvm/Analysis/AliasSetTracker.h" 37#include "llvm/Analysis/Dominators.h" 38#include "llvm/Analysis/ScalarEvolution.h" 39#include "llvm/Analysis/ScalarEvolutionExpressions.h" 40#include "llvm/Analysis/ValueTracking.h" 41#include "llvm/Support/CommandLine.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/raw_ostream.h" 44#include "llvm/Support/ValueHandle.h" 45#include "llvm/DataLayout.h" 46#include "llvm/TargetTransformInfo.h" 47#include "llvm/Transforms/Utils/Local.h" 48#include "llvm/Transforms/Vectorize.h" 49#include <algorithm> 50#include <map> 51using namespace llvm; 52 53static cl::opt<bool> 54IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false), 55 cl::Hidden, cl::desc("Ignore target information")); 56 57static cl::opt<unsigned> 58ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 59 cl::desc("The required chain depth for vectorization")); 60 61static cl::opt<unsigned> 62SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 63 cl::desc("The maximum search distance for instruction pairs")); 64 65static cl::opt<bool> 66SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 67 cl::desc("Replicating one element to a pair breaks the chain")); 68 69static cl::opt<unsigned> 70VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 71 cl::desc("The size of the native vector registers")); 72 73static cl::opt<unsigned> 74MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 75 cl::desc("The maximum number of pairing iterations")); 76 77static cl::opt<bool> 78Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden, 79 cl::desc("Don't try to form non-2^n-length vectors")); 80 81static cl::opt<unsigned> 82MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 83 cl::desc("The maximum number of pairable instructions per group")); 84 85static cl::opt<unsigned> 86MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 87 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 88 " a full cycle check")); 89 90static cl::opt<bool> 91NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden, 92 cl::desc("Don't try to vectorize boolean (i1) values")); 93 94static cl::opt<bool> 95NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 96 cl::desc("Don't try to vectorize integer values")); 97 98static cl::opt<bool> 99NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 100 cl::desc("Don't try to vectorize floating-point values")); 101 102// FIXME: This should default to false once pointer vector support works. 103static cl::opt<bool> 104NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden, 105 cl::desc("Don't try to vectorize pointer values")); 106 107static cl::opt<bool> 108NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 109 cl::desc("Don't try to vectorize casting (conversion) operations")); 110 111static cl::opt<bool> 112NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 113 cl::desc("Don't try to vectorize floating-point math intrinsics")); 114 115static cl::opt<bool> 116NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 117 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 118 119static cl::opt<bool> 120NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 121 cl::desc("Don't try to vectorize select instructions")); 122 123static cl::opt<bool> 124NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 125 cl::desc("Don't try to vectorize comparison instructions")); 126 127static cl::opt<bool> 128NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 129 cl::desc("Don't try to vectorize getelementptr instructions")); 130 131static cl::opt<bool> 132NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 133 cl::desc("Don't try to vectorize loads and stores")); 134 135static cl::opt<bool> 136AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 137 cl::desc("Only generate aligned loads and stores")); 138 139static cl::opt<bool> 140NoMemOpBoost("bb-vectorize-no-mem-op-boost", 141 cl::init(false), cl::Hidden, 142 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 143 144static cl::opt<bool> 145FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 146 cl::desc("Use a fast instruction dependency analysis")); 147 148#ifndef NDEBUG 149static cl::opt<bool> 150DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 151 cl::init(false), cl::Hidden, 152 cl::desc("When debugging is enabled, output information on the" 153 " instruction-examination process")); 154static cl::opt<bool> 155DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 156 cl::init(false), cl::Hidden, 157 cl::desc("When debugging is enabled, output information on the" 158 " candidate-selection process")); 159static cl::opt<bool> 160DebugPairSelection("bb-vectorize-debug-pair-selection", 161 cl::init(false), cl::Hidden, 162 cl::desc("When debugging is enabled, output information on the" 163 " pair-selection process")); 164static cl::opt<bool> 165DebugCycleCheck("bb-vectorize-debug-cycle-check", 166 cl::init(false), cl::Hidden, 167 cl::desc("When debugging is enabled, output information on the" 168 " cycle-checking process")); 169#endif 170 171STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 172 173namespace { 174 struct BBVectorize : public BasicBlockPass { 175 static char ID; // Pass identification, replacement for typeid 176 177 const VectorizeConfig Config; 178 179 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 180 : BasicBlockPass(ID), Config(C) { 181 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 182 } 183 184 BBVectorize(Pass *P, const VectorizeConfig &C) 185 : BasicBlockPass(ID), Config(C) { 186 AA = &P->getAnalysis<AliasAnalysis>(); 187 DT = &P->getAnalysis<DominatorTree>(); 188 SE = &P->getAnalysis<ScalarEvolution>(); 189 TD = P->getAnalysisIfAvailable<DataLayout>(); 190 TTI = IgnoreTargetInfo ? 0 : 191 P->getAnalysisIfAvailable<TargetTransformInfo>(); 192 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 193 } 194 195 typedef std::pair<Value *, Value *> ValuePair; 196 typedef std::pair<ValuePair, int> ValuePairWithCost; 197 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 198 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 199 typedef std::pair<std::multimap<Value *, Value *>::iterator, 200 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 201 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 202 std::multimap<ValuePair, ValuePair>::iterator> 203 VPPIteratorPair; 204 205 AliasAnalysis *AA; 206 DominatorTree *DT; 207 ScalarEvolution *SE; 208 DataLayout *TD; 209 TargetTransformInfo *TTI; 210 const VectorTargetTransformInfo *VTTI; 211 212 // FIXME: const correct? 213 214 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 215 216 bool getCandidatePairs(BasicBlock &BB, 217 BasicBlock::iterator &Start, 218 std::multimap<Value *, Value *> &CandidatePairs, 219 DenseMap<ValuePair, int> &CandidatePairCostSavings, 220 std::vector<Value *> &PairableInsts, bool NonPow2Len); 221 222 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 223 std::vector<Value *> &PairableInsts, 224 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 225 226 void buildDepMap(BasicBlock &BB, 227 std::multimap<Value *, Value *> &CandidatePairs, 228 std::vector<Value *> &PairableInsts, 229 DenseSet<ValuePair> &PairableInstUsers); 230 231 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 232 DenseMap<ValuePair, int> &CandidatePairCostSavings, 233 std::vector<Value *> &PairableInsts, 234 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 235 DenseSet<ValuePair> &PairableInstUsers, 236 DenseMap<Value *, Value *>& ChosenPairs); 237 238 void fuseChosenPairs(BasicBlock &BB, 239 std::vector<Value *> &PairableInsts, 240 DenseMap<Value *, Value *>& ChosenPairs); 241 242 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 243 244 bool areInstsCompatible(Instruction *I, Instruction *J, 245 bool IsSimpleLoadStore, bool NonPow2Len, 246 int &CostSavings); 247 248 bool trackUsesOfI(DenseSet<Value *> &Users, 249 AliasSetTracker &WriteSet, Instruction *I, 250 Instruction *J, bool UpdateUsers = true, 251 std::multimap<Value *, Value *> *LoadMoveSet = 0); 252 253 void computePairsConnectedTo( 254 std::multimap<Value *, Value *> &CandidatePairs, 255 std::vector<Value *> &PairableInsts, 256 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 257 ValuePair P); 258 259 bool pairsConflict(ValuePair P, ValuePair Q, 260 DenseSet<ValuePair> &PairableInstUsers, 261 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 262 263 bool pairWillFormCycle(ValuePair P, 264 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 265 DenseSet<ValuePair> &CurrentPairs); 266 267 void pruneTreeFor( 268 std::multimap<Value *, Value *> &CandidatePairs, 269 std::vector<Value *> &PairableInsts, 270 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 271 DenseSet<ValuePair> &PairableInstUsers, 272 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 273 DenseMap<Value *, Value *> &ChosenPairs, 274 DenseMap<ValuePair, size_t> &Tree, 275 DenseSet<ValuePair> &PrunedTree, ValuePair J, 276 bool UseCycleCheck); 277 278 void buildInitialTreeFor( 279 std::multimap<Value *, Value *> &CandidatePairs, 280 std::vector<Value *> &PairableInsts, 281 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 282 DenseSet<ValuePair> &PairableInstUsers, 283 DenseMap<Value *, Value *> &ChosenPairs, 284 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 285 286 void findBestTreeFor( 287 std::multimap<Value *, Value *> &CandidatePairs, 288 DenseMap<ValuePair, int> &CandidatePairCostSavings, 289 std::vector<Value *> &PairableInsts, 290 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 291 DenseSet<ValuePair> &PairableInstUsers, 292 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 293 DenseMap<Value *, Value *> &ChosenPairs, 294 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 295 int &BestEffSize, VPIteratorPair ChoiceRange, 296 bool UseCycleCheck); 297 298 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 299 Instruction *J, unsigned o, bool FlipMemInputs); 300 301 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 302 unsigned MaskOffset, unsigned NumInElem, 303 unsigned NumInElem1, unsigned IdxOffset, 304 std::vector<Constant*> &Mask); 305 306 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 307 Instruction *J); 308 309 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 310 unsigned o, Value *&LOp, unsigned numElemL, 311 Type *ArgTypeL, Type *ArgTypeR, 312 unsigned IdxOff = 0); 313 314 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 315 Instruction *J, unsigned o, bool FlipMemInputs); 316 317 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 318 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 319 bool FlipMemInputs); 320 321 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 322 Instruction *J, Instruction *K, 323 Instruction *&InsertionPt, Instruction *&K1, 324 Instruction *&K2, bool FlipMemInputs); 325 326 void collectPairLoadMoveSet(BasicBlock &BB, 327 DenseMap<Value *, Value *> &ChosenPairs, 328 std::multimap<Value *, Value *> &LoadMoveSet, 329 Instruction *I); 330 331 void collectLoadMoveSet(BasicBlock &BB, 332 std::vector<Value *> &PairableInsts, 333 DenseMap<Value *, Value *> &ChosenPairs, 334 std::multimap<Value *, Value *> &LoadMoveSet); 335 336 void collectPtrInfo(std::vector<Value *> &PairableInsts, 337 DenseMap<Value *, Value *> &ChosenPairs, 338 DenseSet<Value *> &LowPtrInsts); 339 340 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 341 std::multimap<Value *, Value *> &LoadMoveSet, 342 Instruction *I, Instruction *J); 343 344 void moveUsesOfIAfterJ(BasicBlock &BB, 345 std::multimap<Value *, Value *> &LoadMoveSet, 346 Instruction *&InsertionPt, 347 Instruction *I, Instruction *J); 348 349 void combineMetadata(Instruction *K, const Instruction *J); 350 351 bool vectorizeBB(BasicBlock &BB) { 352 if (!DT->isReachableFromEntry(&BB)) { 353 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 354 " in " << BB.getParent()->getName() << "\n"); 355 return false; 356 } 357 358 DEBUG(if (VTTI) dbgs() << "BBV: using target information\n"); 359 360 bool changed = false; 361 // Iterate a sufficient number of times to merge types of size 1 bit, 362 // then 2 bits, then 4, etc. up to half of the target vector width of the 363 // target vector register. 364 unsigned n = 1; 365 for (unsigned v = 2; 366 (VTTI || v <= Config.VectorBits) && 367 (!Config.MaxIter || n <= Config.MaxIter); 368 v *= 2, ++n) { 369 DEBUG(dbgs() << "BBV: fusing loop #" << n << 370 " for " << BB.getName() << " in " << 371 BB.getParent()->getName() << "...\n"); 372 if (vectorizePairs(BB)) 373 changed = true; 374 else 375 break; 376 } 377 378 if (changed && !Pow2LenOnly) { 379 ++n; 380 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 381 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 382 n << " for " << BB.getName() << " in " << 383 BB.getParent()->getName() << "...\n"); 384 if (!vectorizePairs(BB, true)) break; 385 } 386 } 387 388 DEBUG(dbgs() << "BBV: done!\n"); 389 return changed; 390 } 391 392 virtual bool runOnBasicBlock(BasicBlock &BB) { 393 AA = &getAnalysis<AliasAnalysis>(); 394 DT = &getAnalysis<DominatorTree>(); 395 SE = &getAnalysis<ScalarEvolution>(); 396 TD = getAnalysisIfAvailable<DataLayout>(); 397 TTI = IgnoreTargetInfo ? 0 : 398 getAnalysisIfAvailable<TargetTransformInfo>(); 399 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 400 401 return vectorizeBB(BB); 402 } 403 404 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 405 BasicBlockPass::getAnalysisUsage(AU); 406 AU.addRequired<AliasAnalysis>(); 407 AU.addRequired<DominatorTree>(); 408 AU.addRequired<ScalarEvolution>(); 409 AU.addPreserved<AliasAnalysis>(); 410 AU.addPreserved<DominatorTree>(); 411 AU.addPreserved<ScalarEvolution>(); 412 AU.setPreservesCFG(); 413 } 414 415 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 416 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 417 "Cannot form vector from incompatible scalar types"); 418 Type *STy = ElemTy->getScalarType(); 419 420 unsigned numElem; 421 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 422 numElem = VTy->getNumElements(); 423 } else { 424 numElem = 1; 425 } 426 427 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 428 numElem += VTy->getNumElements(); 429 } else { 430 numElem += 1; 431 } 432 433 return VectorType::get(STy, numElem); 434 } 435 436 static inline void getInstructionTypes(Instruction *I, 437 Type *&T1, Type *&T2) { 438 if (isa<StoreInst>(I)) { 439 // For stores, it is the value type, not the pointer type that matters 440 // because the value is what will come from a vector register. 441 442 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 443 T1 = IVal->getType(); 444 } else { 445 T1 = I->getType(); 446 } 447 448 if (I->isCast()) 449 T2 = cast<CastInst>(I)->getSrcTy(); 450 else 451 T2 = T1; 452 453 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 454 T2 = SI->getCondition()->getType(); 455 } 456 } 457 458 // Returns the weight associated with the provided value. A chain of 459 // candidate pairs has a length given by the sum of the weights of its 460 // members (one weight per pair; the weight of each member of the pair 461 // is assumed to be the same). This length is then compared to the 462 // chain-length threshold to determine if a given chain is significant 463 // enough to be vectorized. The length is also used in comparing 464 // candidate chains where longer chains are considered to be better. 465 // Note: when this function returns 0, the resulting instructions are 466 // not actually fused. 467 inline size_t getDepthFactor(Value *V) { 468 // InsertElement and ExtractElement have a depth factor of zero. This is 469 // for two reasons: First, they cannot be usefully fused. Second, because 470 // the pass generates a lot of these, they can confuse the simple metric 471 // used to compare the trees in the next iteration. Thus, giving them a 472 // weight of zero allows the pass to essentially ignore them in 473 // subsequent iterations when looking for vectorization opportunities 474 // while still tracking dependency chains that flow through those 475 // instructions. 476 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 477 return 0; 478 479 // Give a load or store half of the required depth so that load/store 480 // pairs will vectorize. 481 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 482 return Config.ReqChainDepth/2; 483 484 return 1; 485 } 486 487 // This determines the relative offset of two loads or stores, returning 488 // true if the offset could be determined to be some constant value. 489 // For example, if OffsetInElmts == 1, then J accesses the memory directly 490 // after I; if OffsetInElmts == -1 then I accesses the memory 491 // directly after J. 492 bool getPairPtrInfo(Instruction *I, Instruction *J, 493 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 494 unsigned &IAddressSpace, unsigned &JAddressSpace, 495 int64_t &OffsetInElmts) { 496 OffsetInElmts = 0; 497 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 498 LoadInst *LJ = cast<LoadInst>(J); 499 IPtr = LI->getPointerOperand(); 500 JPtr = LJ->getPointerOperand(); 501 IAlignment = LI->getAlignment(); 502 JAlignment = LJ->getAlignment(); 503 IAddressSpace = LI->getPointerAddressSpace(); 504 JAddressSpace = LJ->getPointerAddressSpace(); 505 } else { 506 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 507 IPtr = SI->getPointerOperand(); 508 JPtr = SJ->getPointerOperand(); 509 IAlignment = SI->getAlignment(); 510 JAlignment = SJ->getAlignment(); 511 IAddressSpace = SI->getPointerAddressSpace(); 512 JAddressSpace = SJ->getPointerAddressSpace(); 513 } 514 515 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 516 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 517 518 // If this is a trivial offset, then we'll get something like 519 // 1*sizeof(type). With target data, which we need anyway, this will get 520 // constant folded into a number. 521 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 522 if (const SCEVConstant *ConstOffSCEV = 523 dyn_cast<SCEVConstant>(OffsetSCEV)) { 524 ConstantInt *IntOff = ConstOffSCEV->getValue(); 525 int64_t Offset = IntOff->getSExtValue(); 526 527 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 528 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 529 530 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 531 if (VTy != VTy2 && Offset < 0) { 532 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 533 OffsetInElmts = Offset/VTy2TSS; 534 return (abs64(Offset) % VTy2TSS) == 0; 535 } 536 537 OffsetInElmts = Offset/VTyTSS; 538 return (abs64(Offset) % VTyTSS) == 0; 539 } 540 541 return false; 542 } 543 544 // Returns true if the provided CallInst represents an intrinsic that can 545 // be vectorized. 546 bool isVectorizableIntrinsic(CallInst* I) { 547 Function *F = I->getCalledFunction(); 548 if (!F) return false; 549 550 unsigned IID = F->getIntrinsicID(); 551 if (!IID) return false; 552 553 switch(IID) { 554 default: 555 return false; 556 case Intrinsic::sqrt: 557 case Intrinsic::powi: 558 case Intrinsic::sin: 559 case Intrinsic::cos: 560 case Intrinsic::log: 561 case Intrinsic::log2: 562 case Intrinsic::log10: 563 case Intrinsic::exp: 564 case Intrinsic::exp2: 565 case Intrinsic::pow: 566 return Config.VectorizeMath; 567 case Intrinsic::fma: 568 return Config.VectorizeFMA; 569 } 570 } 571 572 // Returns true if J is the second element in some pair referenced by 573 // some multimap pair iterator pair. 574 template <typename V> 575 bool isSecondInIteratorPair(V J, std::pair< 576 typename std::multimap<V, V>::iterator, 577 typename std::multimap<V, V>::iterator> PairRange) { 578 for (typename std::multimap<V, V>::iterator K = PairRange.first; 579 K != PairRange.second; ++K) 580 if (K->second == J) return true; 581 582 return false; 583 } 584 }; 585 586 // This function implements one vectorization iteration on the provided 587 // basic block. It returns true if the block is changed. 588 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 589 bool ShouldContinue; 590 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 591 592 std::vector<Value *> AllPairableInsts; 593 DenseMap<Value *, Value *> AllChosenPairs; 594 595 do { 596 std::vector<Value *> PairableInsts; 597 std::multimap<Value *, Value *> CandidatePairs; 598 DenseMap<ValuePair, int> CandidatePairCostSavings; 599 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 600 CandidatePairCostSavings, 601 PairableInsts, NonPow2Len); 602 if (PairableInsts.empty()) continue; 603 604 // Now we have a map of all of the pairable instructions and we need to 605 // select the best possible pairing. A good pairing is one such that the 606 // users of the pair are also paired. This defines a (directed) forest 607 // over the pairs such that two pairs are connected iff the second pair 608 // uses the first. 609 610 // Note that it only matters that both members of the second pair use some 611 // element of the first pair (to allow for splatting). 612 613 std::multimap<ValuePair, ValuePair> ConnectedPairs; 614 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 615 if (ConnectedPairs.empty()) continue; 616 617 // Build the pairable-instruction dependency map 618 DenseSet<ValuePair> PairableInstUsers; 619 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 620 621 // There is now a graph of the connected pairs. For each variable, pick 622 // the pairing with the largest tree meeting the depth requirement on at 623 // least one branch. Then select all pairings that are part of that tree 624 // and remove them from the list of available pairings and pairable 625 // variables. 626 627 DenseMap<Value *, Value *> ChosenPairs; 628 choosePairs(CandidatePairs, CandidatePairCostSavings, 629 PairableInsts, ConnectedPairs, 630 PairableInstUsers, ChosenPairs); 631 632 if (ChosenPairs.empty()) continue; 633 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 634 PairableInsts.end()); 635 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 636 } while (ShouldContinue); 637 638 if (AllChosenPairs.empty()) return false; 639 NumFusedOps += AllChosenPairs.size(); 640 641 // A set of pairs has now been selected. It is now necessary to replace the 642 // paired instructions with vector instructions. For this procedure each 643 // operand must be replaced with a vector operand. This vector is formed 644 // by using build_vector on the old operands. The replaced values are then 645 // replaced with a vector_extract on the result. Subsequent optimization 646 // passes should coalesce the build/extract combinations. 647 648 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 649 650 // It is important to cleanup here so that future iterations of this 651 // function have less work to do. 652 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 653 return true; 654 } 655 656 // This function returns true if the provided instruction is capable of being 657 // fused into a vector instruction. This determination is based only on the 658 // type and other attributes of the instruction. 659 bool BBVectorize::isInstVectorizable(Instruction *I, 660 bool &IsSimpleLoadStore) { 661 IsSimpleLoadStore = false; 662 663 if (CallInst *C = dyn_cast<CallInst>(I)) { 664 if (!isVectorizableIntrinsic(C)) 665 return false; 666 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 667 // Vectorize simple loads if possbile: 668 IsSimpleLoadStore = L->isSimple(); 669 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 670 return false; 671 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 672 // Vectorize simple stores if possbile: 673 IsSimpleLoadStore = S->isSimple(); 674 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 675 return false; 676 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 677 // We can vectorize casts, but not casts of pointer types, etc. 678 if (!Config.VectorizeCasts) 679 return false; 680 681 Type *SrcTy = C->getSrcTy(); 682 if (!SrcTy->isSingleValueType()) 683 return false; 684 685 Type *DestTy = C->getDestTy(); 686 if (!DestTy->isSingleValueType()) 687 return false; 688 } else if (isa<SelectInst>(I)) { 689 if (!Config.VectorizeSelect) 690 return false; 691 } else if (isa<CmpInst>(I)) { 692 if (!Config.VectorizeCmp) 693 return false; 694 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 695 if (!Config.VectorizeGEP) 696 return false; 697 698 // Currently, vector GEPs exist only with one index. 699 if (G->getNumIndices() != 1) 700 return false; 701 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 702 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 703 return false; 704 } 705 706 // We can't vectorize memory operations without target data 707 if (TD == 0 && IsSimpleLoadStore) 708 return false; 709 710 Type *T1, *T2; 711 getInstructionTypes(I, T1, T2); 712 713 // Not every type can be vectorized... 714 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 715 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 716 return false; 717 718 if (T1->getScalarSizeInBits() == 1) { 719 if (!Config.VectorizeBools) 720 return false; 721 } else { 722 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 723 return false; 724 } 725 726 if (T2->getScalarSizeInBits() == 1) { 727 if (!Config.VectorizeBools) 728 return false; 729 } else { 730 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 731 return false; 732 } 733 734 if (!Config.VectorizeFloats 735 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 736 return false; 737 738 // Don't vectorize target-specific types. 739 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 740 return false; 741 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 742 return false; 743 744 if ((!Config.VectorizePointers || TD == 0) && 745 (T1->getScalarType()->isPointerTy() || 746 T2->getScalarType()->isPointerTy())) 747 return false; 748 749 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 750 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 751 return false; 752 753 return true; 754 } 755 756 // This function returns true if the two provided instructions are compatible 757 // (meaning that they can be fused into a vector instruction). This assumes 758 // that I has already been determined to be vectorizable and that J is not 759 // in the use tree of I. 760 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 761 bool IsSimpleLoadStore, bool NonPow2Len, 762 int &CostSavings) { 763 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 764 " <-> " << *J << "\n"); 765 766 CostSavings = 0; 767 768 // Loads and stores can be merged if they have different alignments, 769 // but are otherwise the same. 770 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 771 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 772 return false; 773 774 Type *IT1, *IT2, *JT1, *JT2; 775 getInstructionTypes(I, IT1, IT2); 776 getInstructionTypes(J, JT1, JT2); 777 unsigned MaxTypeBits = std::max( 778 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 779 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 780 if (!VTTI && MaxTypeBits > Config.VectorBits) 781 return false; 782 783 // FIXME: handle addsub-type operations! 784 785 if (IsSimpleLoadStore) { 786 Value *IPtr, *JPtr; 787 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 788 int64_t OffsetInElmts = 0; 789 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 790 IAddressSpace, JAddressSpace, 791 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 792 unsigned BottomAlignment = IAlignment; 793 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 794 795 Type *aTypeI = isa<StoreInst>(I) ? 796 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 797 Type *aTypeJ = isa<StoreInst>(J) ? 798 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 799 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 800 801 if (Config.AlignedOnly) { 802 // An aligned load or store is possible only if the instruction 803 // with the lower offset has an alignment suitable for the 804 // vector type. 805 806 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 807 if (BottomAlignment < VecAlignment) 808 return false; 809 } 810 811 if (VTTI) { 812 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(), 813 IAlignment, IAddressSpace); 814 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(), 815 JAlignment, JAddressSpace); 816 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType, 817 BottomAlignment, 818 IAddressSpace); 819 if (VCost > ICost + JCost) 820 return false; 821 822 // We don't want to fuse to a type that will be split, even 823 // if the two input types will also be split and there is no other 824 // associated cost. 825 unsigned VParts = VTTI->getNumberOfParts(VType); 826 if (VParts > 1) 827 return false; 828 else if (!VParts && VCost == ICost + JCost) 829 return false; 830 831 CostSavings = ICost + JCost - VCost; 832 } 833 } else { 834 return false; 835 } 836 } else if (VTTI) { 837 unsigned ICost = VTTI->getInstrCost(I->getOpcode(), IT1, IT2); 838 unsigned JCost = VTTI->getInstrCost(J->getOpcode(), JT1, JT2); 839 Type *VT1 = getVecTypeForPair(IT1, JT1), 840 *VT2 = getVecTypeForPair(IT2, JT2); 841 unsigned VCost = VTTI->getInstrCost(I->getOpcode(), VT1, VT2); 842 843 if (VCost > ICost + JCost) 844 return false; 845 846 // We don't want to fuse to a type that will be split, even 847 // if the two input types will also be split and there is no other 848 // associated cost. 849 unsigned VParts = VTTI->getNumberOfParts(VT1); 850 if (VParts > 1) 851 return false; 852 else if (!VParts && VCost == ICost + JCost) 853 return false; 854 855 CostSavings = ICost + JCost - VCost; 856 } 857 858 // The powi intrinsic is special because only the first argument is 859 // vectorized, the second arguments must be equal. 860 CallInst *CI = dyn_cast<CallInst>(I); 861 Function *FI; 862 if (CI && (FI = CI->getCalledFunction()) && 863 FI->getIntrinsicID() == Intrinsic::powi) { 864 865 Value *A1I = CI->getArgOperand(1), 866 *A1J = cast<CallInst>(J)->getArgOperand(1); 867 const SCEV *A1ISCEV = SE->getSCEV(A1I), 868 *A1JSCEV = SE->getSCEV(A1J); 869 return (A1ISCEV == A1JSCEV); 870 } 871 872 return true; 873 } 874 875 // Figure out whether or not J uses I and update the users and write-set 876 // structures associated with I. Specifically, Users represents the set of 877 // instructions that depend on I. WriteSet represents the set 878 // of memory locations that are dependent on I. If UpdateUsers is true, 879 // and J uses I, then Users is updated to contain J and WriteSet is updated 880 // to contain any memory locations to which J writes. The function returns 881 // true if J uses I. By default, alias analysis is used to determine 882 // whether J reads from memory that overlaps with a location in WriteSet. 883 // If LoadMoveSet is not null, then it is a previously-computed multimap 884 // where the key is the memory-based user instruction and the value is 885 // the instruction to be compared with I. So, if LoadMoveSet is provided, 886 // then the alias analysis is not used. This is necessary because this 887 // function is called during the process of moving instructions during 888 // vectorization and the results of the alias analysis are not stable during 889 // that process. 890 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 891 AliasSetTracker &WriteSet, Instruction *I, 892 Instruction *J, bool UpdateUsers, 893 std::multimap<Value *, Value *> *LoadMoveSet) { 894 bool UsesI = false; 895 896 // This instruction may already be marked as a user due, for example, to 897 // being a member of a selected pair. 898 if (Users.count(J)) 899 UsesI = true; 900 901 if (!UsesI) 902 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 903 JU != JE; ++JU) { 904 Value *V = *JU; 905 if (I == V || Users.count(V)) { 906 UsesI = true; 907 break; 908 } 909 } 910 if (!UsesI && J->mayReadFromMemory()) { 911 if (LoadMoveSet) { 912 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 913 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 914 } else { 915 for (AliasSetTracker::iterator W = WriteSet.begin(), 916 WE = WriteSet.end(); W != WE; ++W) { 917 if (W->aliasesUnknownInst(J, *AA)) { 918 UsesI = true; 919 break; 920 } 921 } 922 } 923 } 924 925 if (UsesI && UpdateUsers) { 926 if (J->mayWriteToMemory()) WriteSet.add(J); 927 Users.insert(J); 928 } 929 930 return UsesI; 931 } 932 933 // This function iterates over all instruction pairs in the provided 934 // basic block and collects all candidate pairs for vectorization. 935 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 936 BasicBlock::iterator &Start, 937 std::multimap<Value *, Value *> &CandidatePairs, 938 DenseMap<ValuePair, int> &CandidatePairCostSavings, 939 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 940 BasicBlock::iterator E = BB.end(); 941 if (Start == E) return false; 942 943 bool ShouldContinue = false, IAfterStart = false; 944 for (BasicBlock::iterator I = Start++; I != E; ++I) { 945 if (I == Start) IAfterStart = true; 946 947 bool IsSimpleLoadStore; 948 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 949 950 // Look for an instruction with which to pair instruction *I... 951 DenseSet<Value *> Users; 952 AliasSetTracker WriteSet(*AA); 953 bool JAfterStart = IAfterStart; 954 BasicBlock::iterator J = llvm::next(I); 955 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 956 if (J == Start) JAfterStart = true; 957 958 // Determine if J uses I, if so, exit the loop. 959 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 960 if (Config.FastDep) { 961 // Note: For this heuristic to be effective, independent operations 962 // must tend to be intermixed. This is likely to be true from some 963 // kinds of grouped loop unrolling (but not the generic LLVM pass), 964 // but otherwise may require some kind of reordering pass. 965 966 // When using fast dependency analysis, 967 // stop searching after first use: 968 if (UsesI) break; 969 } else { 970 if (UsesI) continue; 971 } 972 973 // J does not use I, and comes before the first use of I, so it can be 974 // merged with I if the instructions are compatible. 975 int CostSavings; 976 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 977 CostSavings)) continue; 978 979 // J is a candidate for merging with I. 980 if (!PairableInsts.size() || 981 PairableInsts[PairableInsts.size()-1] != I) { 982 PairableInsts.push_back(I); 983 } 984 985 CandidatePairs.insert(ValuePair(I, J)); 986 if (VTTI) 987 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 988 CostSavings)); 989 990 // The next call to this function must start after the last instruction 991 // selected during this invocation. 992 if (JAfterStart) { 993 Start = llvm::next(J); 994 IAfterStart = JAfterStart = false; 995 } 996 997 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 998 << *I << " <-> " << *J << " (cost savings: " << 999 CostSavings << ")\n"); 1000 1001 // If we have already found too many pairs, break here and this function 1002 // will be called again starting after the last instruction selected 1003 // during this invocation. 1004 if (PairableInsts.size() >= Config.MaxInsts) { 1005 ShouldContinue = true; 1006 break; 1007 } 1008 } 1009 1010 if (ShouldContinue) 1011 break; 1012 } 1013 1014 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1015 << " instructions with candidate pairs\n"); 1016 1017 return ShouldContinue; 1018 } 1019 1020 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1021 // it looks for pairs such that both members have an input which is an 1022 // output of PI or PJ. 1023 void BBVectorize::computePairsConnectedTo( 1024 std::multimap<Value *, Value *> &CandidatePairs, 1025 std::vector<Value *> &PairableInsts, 1026 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1027 ValuePair P) { 1028 StoreInst *SI, *SJ; 1029 1030 // For each possible pairing for this variable, look at the uses of 1031 // the first value... 1032 for (Value::use_iterator I = P.first->use_begin(), 1033 E = P.first->use_end(); I != E; ++I) { 1034 if (isa<LoadInst>(*I)) { 1035 // A pair cannot be connected to a load because the load only takes one 1036 // operand (the address) and it is a scalar even after vectorization. 1037 continue; 1038 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1039 P.first == SI->getPointerOperand()) { 1040 // Similarly, a pair cannot be connected to a store through its 1041 // pointer operand. 1042 continue; 1043 } 1044 1045 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1046 1047 // For each use of the first variable, look for uses of the second 1048 // variable... 1049 for (Value::use_iterator J = P.second->use_begin(), 1050 E2 = P.second->use_end(); J != E2; ++J) { 1051 if ((SJ = dyn_cast<StoreInst>(*J)) && 1052 P.second == SJ->getPointerOperand()) 1053 continue; 1054 1055 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 1056 1057 // Look for <I, J>: 1058 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 1059 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 1060 1061 // Look for <J, I>: 1062 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 1063 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 1064 } 1065 1066 if (Config.SplatBreaksChain) continue; 1067 // Look for cases where just the first value in the pair is used by 1068 // both members of another pair (splatting). 1069 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1070 if ((SJ = dyn_cast<StoreInst>(*J)) && 1071 P.first == SJ->getPointerOperand()) 1072 continue; 1073 1074 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 1075 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 1076 } 1077 } 1078 1079 if (Config.SplatBreaksChain) return; 1080 // Look for cases where just the second value in the pair is used by 1081 // both members of another pair (splatting). 1082 for (Value::use_iterator I = P.second->use_begin(), 1083 E = P.second->use_end(); I != E; ++I) { 1084 if (isa<LoadInst>(*I)) 1085 continue; 1086 else if ((SI = dyn_cast<StoreInst>(*I)) && 1087 P.second == SI->getPointerOperand()) 1088 continue; 1089 1090 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1091 1092 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1093 if ((SJ = dyn_cast<StoreInst>(*J)) && 1094 P.second == SJ->getPointerOperand()) 1095 continue; 1096 1097 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 1098 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 1099 } 1100 } 1101 } 1102 1103 // This function figures out which pairs are connected. Two pairs are 1104 // connected if some output of the first pair forms an input to both members 1105 // of the second pair. 1106 void BBVectorize::computeConnectedPairs( 1107 std::multimap<Value *, Value *> &CandidatePairs, 1108 std::vector<Value *> &PairableInsts, 1109 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 1110 1111 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1112 PE = PairableInsts.end(); PI != PE; ++PI) { 1113 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1114 1115 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1116 P != choiceRange.second; ++P) 1117 computePairsConnectedTo(CandidatePairs, PairableInsts, 1118 ConnectedPairs, *P); 1119 } 1120 1121 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1122 << " pair connections.\n"); 1123 } 1124 1125 // This function builds a set of use tuples such that <A, B> is in the set 1126 // if B is in the use tree of A. If B is in the use tree of A, then B 1127 // depends on the output of A. 1128 void BBVectorize::buildDepMap( 1129 BasicBlock &BB, 1130 std::multimap<Value *, Value *> &CandidatePairs, 1131 std::vector<Value *> &PairableInsts, 1132 DenseSet<ValuePair> &PairableInstUsers) { 1133 DenseSet<Value *> IsInPair; 1134 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1135 E = CandidatePairs.end(); C != E; ++C) { 1136 IsInPair.insert(C->first); 1137 IsInPair.insert(C->second); 1138 } 1139 1140 // Iterate through the basic block, recording all Users of each 1141 // pairable instruction. 1142 1143 BasicBlock::iterator E = BB.end(); 1144 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1145 if (IsInPair.find(I) == IsInPair.end()) continue; 1146 1147 DenseSet<Value *> Users; 1148 AliasSetTracker WriteSet(*AA); 1149 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1150 (void) trackUsesOfI(Users, WriteSet, I, J); 1151 1152 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1153 U != E; ++U) 1154 PairableInstUsers.insert(ValuePair(I, *U)); 1155 } 1156 } 1157 1158 // Returns true if an input to pair P is an output of pair Q and also an 1159 // input of pair Q is an output of pair P. If this is the case, then these 1160 // two pairs cannot be simultaneously fused. 1161 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1162 DenseSet<ValuePair> &PairableInstUsers, 1163 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1164 // Two pairs are in conflict if they are mutual Users of eachother. 1165 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1166 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1167 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1168 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1169 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1170 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1171 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1172 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1173 if (PairableInstUserMap) { 1174 // FIXME: The expensive part of the cycle check is not so much the cycle 1175 // check itself but this edge insertion procedure. This needs some 1176 // profiling and probably a different data structure (same is true of 1177 // most uses of std::multimap). 1178 if (PUsesQ) { 1179 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1180 if (!isSecondInIteratorPair(P, QPairRange)) 1181 PairableInstUserMap->insert(VPPair(Q, P)); 1182 } 1183 if (QUsesP) { 1184 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1185 if (!isSecondInIteratorPair(Q, PPairRange)) 1186 PairableInstUserMap->insert(VPPair(P, Q)); 1187 } 1188 } 1189 1190 return (QUsesP && PUsesQ); 1191 } 1192 1193 // This function walks the use graph of current pairs to see if, starting 1194 // from P, the walk returns to P. 1195 bool BBVectorize::pairWillFormCycle(ValuePair P, 1196 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1197 DenseSet<ValuePair> &CurrentPairs) { 1198 DEBUG(if (DebugCycleCheck) 1199 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1200 << *P.second << "\n"); 1201 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1202 // contains non-direct associations. 1203 DenseSet<ValuePair> Visited; 1204 SmallVector<ValuePair, 32> Q; 1205 // General depth-first post-order traversal: 1206 Q.push_back(P); 1207 do { 1208 ValuePair QTop = Q.pop_back_val(); 1209 Visited.insert(QTop); 1210 1211 DEBUG(if (DebugCycleCheck) 1212 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1213 << *QTop.second << "\n"); 1214 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1215 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1216 C != QPairRange.second; ++C) { 1217 if (C->second == P) { 1218 DEBUG(dbgs() 1219 << "BBV: rejected to prevent non-trivial cycle formation: " 1220 << *C->first.first << " <-> " << *C->first.second << "\n"); 1221 return true; 1222 } 1223 1224 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1225 Q.push_back(C->second); 1226 } 1227 } while (!Q.empty()); 1228 1229 return false; 1230 } 1231 1232 // This function builds the initial tree of connected pairs with the 1233 // pair J at the root. 1234 void BBVectorize::buildInitialTreeFor( 1235 std::multimap<Value *, Value *> &CandidatePairs, 1236 std::vector<Value *> &PairableInsts, 1237 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1238 DenseSet<ValuePair> &PairableInstUsers, 1239 DenseMap<Value *, Value *> &ChosenPairs, 1240 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1241 // Each of these pairs is viewed as the root node of a Tree. The Tree 1242 // is then walked (depth-first). As this happens, we keep track of 1243 // the pairs that compose the Tree and the maximum depth of the Tree. 1244 SmallVector<ValuePairWithDepth, 32> Q; 1245 // General depth-first post-order traversal: 1246 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1247 do { 1248 ValuePairWithDepth QTop = Q.back(); 1249 1250 // Push each child onto the queue: 1251 bool MoreChildren = false; 1252 size_t MaxChildDepth = QTop.second; 1253 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1254 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1255 k != qtRange.second; ++k) { 1256 // Make sure that this child pair is still a candidate: 1257 bool IsStillCand = false; 1258 VPIteratorPair checkRange = 1259 CandidatePairs.equal_range(k->second.first); 1260 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1261 m != checkRange.second; ++m) { 1262 if (m->second == k->second.second) { 1263 IsStillCand = true; 1264 break; 1265 } 1266 } 1267 1268 if (IsStillCand) { 1269 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1270 if (C == Tree.end()) { 1271 size_t d = getDepthFactor(k->second.first); 1272 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1273 MoreChildren = true; 1274 } else { 1275 MaxChildDepth = std::max(MaxChildDepth, C->second); 1276 } 1277 } 1278 } 1279 1280 if (!MoreChildren) { 1281 // Record the current pair as part of the Tree: 1282 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1283 Q.pop_back(); 1284 } 1285 } while (!Q.empty()); 1286 } 1287 1288 // Given some initial tree, prune it by removing conflicting pairs (pairs 1289 // that cannot be simultaneously chosen for vectorization). 1290 void BBVectorize::pruneTreeFor( 1291 std::multimap<Value *, Value *> &CandidatePairs, 1292 std::vector<Value *> &PairableInsts, 1293 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1294 DenseSet<ValuePair> &PairableInstUsers, 1295 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1296 DenseMap<Value *, Value *> &ChosenPairs, 1297 DenseMap<ValuePair, size_t> &Tree, 1298 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1299 bool UseCycleCheck) { 1300 SmallVector<ValuePairWithDepth, 32> Q; 1301 // General depth-first post-order traversal: 1302 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1303 do { 1304 ValuePairWithDepth QTop = Q.pop_back_val(); 1305 PrunedTree.insert(QTop.first); 1306 1307 // Visit each child, pruning as necessary... 1308 DenseMap<ValuePair, size_t> BestChildren; 1309 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1310 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1311 K != QTopRange.second; ++K) { 1312 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1313 if (C == Tree.end()) continue; 1314 1315 // This child is in the Tree, now we need to make sure it is the 1316 // best of any conflicting children. There could be multiple 1317 // conflicting children, so first, determine if we're keeping 1318 // this child, then delete conflicting children as necessary. 1319 1320 // It is also necessary to guard against pairing-induced 1321 // dependencies. Consider instructions a .. x .. y .. b 1322 // such that (a,b) are to be fused and (x,y) are to be fused 1323 // but a is an input to x and b is an output from y. This 1324 // means that y cannot be moved after b but x must be moved 1325 // after b for (a,b) to be fused. In other words, after 1326 // fusing (a,b) we have y .. a/b .. x where y is an input 1327 // to a/b and x is an output to a/b: x and y can no longer 1328 // be legally fused. To prevent this condition, we must 1329 // make sure that a child pair added to the Tree is not 1330 // both an input and output of an already-selected pair. 1331 1332 // Pairing-induced dependencies can also form from more complicated 1333 // cycles. The pair vs. pair conflicts are easy to check, and so 1334 // that is done explicitly for "fast rejection", and because for 1335 // child vs. child conflicts, we may prefer to keep the current 1336 // pair in preference to the already-selected child. 1337 DenseSet<ValuePair> CurrentPairs; 1338 1339 bool CanAdd = true; 1340 for (DenseMap<ValuePair, size_t>::iterator C2 1341 = BestChildren.begin(), E2 = BestChildren.end(); 1342 C2 != E2; ++C2) { 1343 if (C2->first.first == C->first.first || 1344 C2->first.first == C->first.second || 1345 C2->first.second == C->first.first || 1346 C2->first.second == C->first.second || 1347 pairsConflict(C2->first, C->first, PairableInstUsers, 1348 UseCycleCheck ? &PairableInstUserMap : 0)) { 1349 if (C2->second >= C->second) { 1350 CanAdd = false; 1351 break; 1352 } 1353 1354 CurrentPairs.insert(C2->first); 1355 } 1356 } 1357 if (!CanAdd) continue; 1358 1359 // Even worse, this child could conflict with another node already 1360 // selected for the Tree. If that is the case, ignore this child. 1361 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1362 E2 = PrunedTree.end(); T != E2; ++T) { 1363 if (T->first == C->first.first || 1364 T->first == C->first.second || 1365 T->second == C->first.first || 1366 T->second == C->first.second || 1367 pairsConflict(*T, C->first, PairableInstUsers, 1368 UseCycleCheck ? &PairableInstUserMap : 0)) { 1369 CanAdd = false; 1370 break; 1371 } 1372 1373 CurrentPairs.insert(*T); 1374 } 1375 if (!CanAdd) continue; 1376 1377 // And check the queue too... 1378 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1379 E2 = Q.end(); C2 != E2; ++C2) { 1380 if (C2->first.first == C->first.first || 1381 C2->first.first == C->first.second || 1382 C2->first.second == C->first.first || 1383 C2->first.second == C->first.second || 1384 pairsConflict(C2->first, C->first, PairableInstUsers, 1385 UseCycleCheck ? &PairableInstUserMap : 0)) { 1386 CanAdd = false; 1387 break; 1388 } 1389 1390 CurrentPairs.insert(C2->first); 1391 } 1392 if (!CanAdd) continue; 1393 1394 // Last but not least, check for a conflict with any of the 1395 // already-chosen pairs. 1396 for (DenseMap<Value *, Value *>::iterator C2 = 1397 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1398 C2 != E2; ++C2) { 1399 if (pairsConflict(*C2, C->first, PairableInstUsers, 1400 UseCycleCheck ? &PairableInstUserMap : 0)) { 1401 CanAdd = false; 1402 break; 1403 } 1404 1405 CurrentPairs.insert(*C2); 1406 } 1407 if (!CanAdd) continue; 1408 1409 // To check for non-trivial cycles formed by the addition of the 1410 // current pair we've formed a list of all relevant pairs, now use a 1411 // graph walk to check for a cycle. We start from the current pair and 1412 // walk the use tree to see if we again reach the current pair. If we 1413 // do, then the current pair is rejected. 1414 1415 // FIXME: It may be more efficient to use a topological-ordering 1416 // algorithm to improve the cycle check. This should be investigated. 1417 if (UseCycleCheck && 1418 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1419 continue; 1420 1421 // This child can be added, but we may have chosen it in preference 1422 // to an already-selected child. Check for this here, and if a 1423 // conflict is found, then remove the previously-selected child 1424 // before adding this one in its place. 1425 for (DenseMap<ValuePair, size_t>::iterator C2 1426 = BestChildren.begin(); C2 != BestChildren.end();) { 1427 if (C2->first.first == C->first.first || 1428 C2->first.first == C->first.second || 1429 C2->first.second == C->first.first || 1430 C2->first.second == C->first.second || 1431 pairsConflict(C2->first, C->first, PairableInstUsers)) 1432 BestChildren.erase(C2++); 1433 else 1434 ++C2; 1435 } 1436 1437 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1438 } 1439 1440 for (DenseMap<ValuePair, size_t>::iterator C 1441 = BestChildren.begin(), E2 = BestChildren.end(); 1442 C != E2; ++C) { 1443 size_t DepthF = getDepthFactor(C->first.first); 1444 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1445 } 1446 } while (!Q.empty()); 1447 } 1448 1449 // This function finds the best tree of mututally-compatible connected 1450 // pairs, given the choice of root pairs as an iterator range. 1451 void BBVectorize::findBestTreeFor( 1452 std::multimap<Value *, Value *> &CandidatePairs, 1453 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1454 std::vector<Value *> &PairableInsts, 1455 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1456 DenseSet<ValuePair> &PairableInstUsers, 1457 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1458 DenseMap<Value *, Value *> &ChosenPairs, 1459 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1460 int &BestEffSize, VPIteratorPair ChoiceRange, 1461 bool UseCycleCheck) { 1462 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1463 J != ChoiceRange.second; ++J) { 1464 1465 // Before going any further, make sure that this pair does not 1466 // conflict with any already-selected pairs (see comment below 1467 // near the Tree pruning for more details). 1468 DenseSet<ValuePair> ChosenPairSet; 1469 bool DoesConflict = false; 1470 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1471 E = ChosenPairs.end(); C != E; ++C) { 1472 if (pairsConflict(*C, *J, PairableInstUsers, 1473 UseCycleCheck ? &PairableInstUserMap : 0)) { 1474 DoesConflict = true; 1475 break; 1476 } 1477 1478 ChosenPairSet.insert(*C); 1479 } 1480 if (DoesConflict) continue; 1481 1482 if (UseCycleCheck && 1483 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1484 continue; 1485 1486 DenseMap<ValuePair, size_t> Tree; 1487 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1488 PairableInstUsers, ChosenPairs, Tree, *J); 1489 1490 // Because we'll keep the child with the largest depth, the largest 1491 // depth is still the same in the unpruned Tree. 1492 size_t MaxDepth = Tree.lookup(*J); 1493 1494 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1495 << *J->first << " <-> " << *J->second << "} of depth " << 1496 MaxDepth << " and size " << Tree.size() << "\n"); 1497 1498 // At this point the Tree has been constructed, but, may contain 1499 // contradictory children (meaning that different children of 1500 // some tree node may be attempting to fuse the same instruction). 1501 // So now we walk the tree again, in the case of a conflict, 1502 // keep only the child with the largest depth. To break a tie, 1503 // favor the first child. 1504 1505 DenseSet<ValuePair> PrunedTree; 1506 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1507 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1508 PrunedTree, *J, UseCycleCheck); 1509 1510 int EffSize = 0; 1511 if (VTTI) { 1512 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1513 E = PrunedTree.end(); S != E; ++S) { 1514 if (getDepthFactor(S->first)) 1515 EffSize += CandidatePairCostSavings.find(*S)->second; 1516 } 1517 } else { 1518 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1519 E = PrunedTree.end(); S != E; ++S) 1520 EffSize += (int) getDepthFactor(S->first); 1521 } 1522 1523 DEBUG(if (DebugPairSelection) 1524 dbgs() << "BBV: found pruned Tree for pair {" 1525 << *J->first << " <-> " << *J->second << "} of depth " << 1526 MaxDepth << " and size " << PrunedTree.size() << 1527 " (effective size: " << EffSize << ")\n"); 1528 if (MaxDepth >= Config.ReqChainDepth && 1529 EffSize > 0 && EffSize > BestEffSize) { 1530 BestMaxDepth = MaxDepth; 1531 BestEffSize = EffSize; 1532 BestTree = PrunedTree; 1533 } 1534 } 1535 } 1536 1537 // Given the list of candidate pairs, this function selects those 1538 // that will be fused into vector instructions. 1539 void BBVectorize::choosePairs( 1540 std::multimap<Value *, Value *> &CandidatePairs, 1541 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1542 std::vector<Value *> &PairableInsts, 1543 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1544 DenseSet<ValuePair> &PairableInstUsers, 1545 DenseMap<Value *, Value *>& ChosenPairs) { 1546 bool UseCycleCheck = 1547 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1548 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1549 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1550 E = PairableInsts.end(); I != E; ++I) { 1551 // The number of possible pairings for this variable: 1552 size_t NumChoices = CandidatePairs.count(*I); 1553 if (!NumChoices) continue; 1554 1555 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1556 1557 // The best pair to choose and its tree: 1558 size_t BestMaxDepth = 0; 1559 int BestEffSize = 0; 1560 DenseSet<ValuePair> BestTree; 1561 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 1562 PairableInsts, ConnectedPairs, 1563 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1564 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1565 UseCycleCheck); 1566 1567 // A tree has been chosen (or not) at this point. If no tree was 1568 // chosen, then this instruction, I, cannot be paired (and is no longer 1569 // considered). 1570 1571 DEBUG(if (BestTree.size() > 0) 1572 dbgs() << "BBV: selected pairs in the best tree for: " 1573 << *cast<Instruction>(*I) << "\n"); 1574 1575 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1576 SE2 = BestTree.end(); S != SE2; ++S) { 1577 // Insert the members of this tree into the list of chosen pairs. 1578 ChosenPairs.insert(ValuePair(S->first, S->second)); 1579 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1580 *S->second << "\n"); 1581 1582 // Remove all candidate pairs that have values in the chosen tree. 1583 for (std::multimap<Value *, Value *>::iterator K = 1584 CandidatePairs.begin(); K != CandidatePairs.end();) { 1585 if (K->first == S->first || K->second == S->first || 1586 K->second == S->second || K->first == S->second) { 1587 // Don't remove the actual pair chosen so that it can be used 1588 // in subsequent tree selections. 1589 if (!(K->first == S->first && K->second == S->second)) 1590 CandidatePairs.erase(K++); 1591 else 1592 ++K; 1593 } else { 1594 ++K; 1595 } 1596 } 1597 } 1598 } 1599 1600 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1601 } 1602 1603 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1604 unsigned n = 0) { 1605 if (!I->hasName()) 1606 return ""; 1607 1608 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1609 (n > 0 ? "." + utostr(n) : "")).str(); 1610 } 1611 1612 // Returns the value that is to be used as the pointer input to the vector 1613 // instruction that fuses I with J. 1614 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1615 Instruction *I, Instruction *J, unsigned o, 1616 bool FlipMemInputs) { 1617 Value *IPtr, *JPtr; 1618 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 1619 int64_t OffsetInElmts; 1620 1621 // Note: the analysis might fail here, that is why FlipMemInputs has 1622 // been precomputed (OffsetInElmts must be unused here). 1623 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1624 IAddressSpace, JAddressSpace, 1625 OffsetInElmts); 1626 1627 // The pointer value is taken to be the one with the lowest offset. 1628 Value *VPtr; 1629 if (!FlipMemInputs) { 1630 VPtr = IPtr; 1631 } else { 1632 VPtr = JPtr; 1633 } 1634 1635 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 1636 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 1637 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1638 Type *VArgPtrType = PointerType::get(VArgType, 1639 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1640 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1641 /* insert before */ FlipMemInputs ? J : I); 1642 } 1643 1644 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1645 unsigned MaskOffset, unsigned NumInElem, 1646 unsigned NumInElem1, unsigned IdxOffset, 1647 std::vector<Constant*> &Mask) { 1648 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 1649 for (unsigned v = 0; v < NumElem1; ++v) { 1650 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1651 if (m < 0) { 1652 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1653 } else { 1654 unsigned mm = m + (int) IdxOffset; 1655 if (m >= (int) NumInElem1) 1656 mm += (int) NumInElem; 1657 1658 Mask[v+MaskOffset] = 1659 ConstantInt::get(Type::getInt32Ty(Context), mm); 1660 } 1661 } 1662 } 1663 1664 // Returns the value that is to be used as the vector-shuffle mask to the 1665 // vector instruction that fuses I with J. 1666 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1667 Instruction *I, Instruction *J) { 1668 // This is the shuffle mask. We need to append the second 1669 // mask to the first, and the numbers need to be adjusted. 1670 1671 Type *ArgTypeI = I->getType(); 1672 Type *ArgTypeJ = J->getType(); 1673 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1674 1675 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 1676 1677 // Get the total number of elements in the fused vector type. 1678 // By definition, this must equal the number of elements in 1679 // the final mask. 1680 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1681 std::vector<Constant*> Mask(NumElem); 1682 1683 Type *OpTypeI = I->getOperand(0)->getType(); 1684 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 1685 Type *OpTypeJ = J->getOperand(0)->getType(); 1686 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 1687 1688 // The fused vector will be: 1689 // ----------------------------------------------------- 1690 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 1691 // ----------------------------------------------------- 1692 // from which we'll extract NumElem total elements (where the first NumElemI 1693 // of them come from the mask in I and the remainder come from the mask 1694 // in J. 1695 1696 // For the mask from the first pair... 1697 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 1698 0, Mask); 1699 1700 // For the mask from the second pair... 1701 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 1702 NumInElemI, Mask); 1703 1704 return ConstantVector::get(Mask); 1705 } 1706 1707 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 1708 Instruction *J, unsigned o, Value *&LOp, 1709 unsigned numElemL, 1710 Type *ArgTypeL, Type *ArgTypeH, 1711 unsigned IdxOff) { 1712 bool ExpandedIEChain = false; 1713 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 1714 // If we have a pure insertelement chain, then this can be rewritten 1715 // into a chain that directly builds the larger type. 1716 bool PureChain = true; 1717 InsertElementInst *LIENext = LIE; 1718 do { 1719 if (!isa<UndefValue>(LIENext->getOperand(0)) && 1720 !isa<InsertElementInst>(LIENext->getOperand(0))) { 1721 PureChain = false; 1722 break; 1723 } 1724 } while ((LIENext = 1725 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1726 1727 if (PureChain) { 1728 SmallVector<Value *, 8> VectElemts(numElemL, 1729 UndefValue::get(ArgTypeL->getScalarType())); 1730 InsertElementInst *LIENext = LIE; 1731 do { 1732 unsigned Idx = 1733 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 1734 VectElemts[Idx] = LIENext->getOperand(1); 1735 } while ((LIENext = 1736 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1737 1738 LIENext = 0; 1739 Value *LIEPrev = UndefValue::get(ArgTypeH); 1740 for (unsigned i = 0; i < numElemL; ++i) { 1741 if (isa<UndefValue>(VectElemts[i])) continue; 1742 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 1743 ConstantInt::get(Type::getInt32Ty(Context), 1744 i + IdxOff), 1745 getReplacementName(I, true, o, i+1)); 1746 LIENext->insertBefore(J); 1747 LIEPrev = LIENext; 1748 } 1749 1750 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 1751 ExpandedIEChain = true; 1752 } 1753 } 1754 1755 return ExpandedIEChain; 1756 } 1757 1758 // Returns the value to be used as the specified operand of the vector 1759 // instruction that fuses I with J. 1760 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1761 Instruction *J, unsigned o, bool FlipMemInputs) { 1762 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1763 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1764 1765 // Compute the fused vector type for this operand 1766 Type *ArgTypeI = I->getOperand(o)->getType(); 1767 Type *ArgTypeJ = J->getOperand(o)->getType(); 1768 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1769 1770 Instruction *L = I, *H = J; 1771 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 1772 if (FlipMemInputs) { 1773 L = J; 1774 H = I; 1775 ArgTypeL = ArgTypeJ; 1776 ArgTypeH = ArgTypeI; 1777 } 1778 1779 unsigned numElemL; 1780 if (ArgTypeL->isVectorTy()) 1781 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 1782 else 1783 numElemL = 1; 1784 1785 unsigned numElemH; 1786 if (ArgTypeH->isVectorTy()) 1787 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 1788 else 1789 numElemH = 1; 1790 1791 Value *LOp = L->getOperand(o); 1792 Value *HOp = H->getOperand(o); 1793 unsigned numElem = VArgType->getNumElements(); 1794 1795 // First, we check if we can reuse the "original" vector outputs (if these 1796 // exist). We might need a shuffle. 1797 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 1798 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 1799 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 1800 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 1801 1802 // FIXME: If we're fusing shuffle instructions, then we can't apply this 1803 // optimization. The input vectors to the shuffle might be a different 1804 // length from the shuffle outputs. Unfortunately, the replacement 1805 // shuffle mask has already been formed, and the mask entries are sensitive 1806 // to the sizes of the inputs. 1807 bool IsSizeChangeShuffle = 1808 isa<ShuffleVectorInst>(L) && 1809 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 1810 1811 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 1812 // We can have at most two unique vector inputs. 1813 bool CanUseInputs = true; 1814 Value *I1, *I2 = 0; 1815 if (LEE) { 1816 I1 = LEE->getOperand(0); 1817 } else { 1818 I1 = LSV->getOperand(0); 1819 I2 = LSV->getOperand(1); 1820 if (I2 == I1 || isa<UndefValue>(I2)) 1821 I2 = 0; 1822 } 1823 1824 if (HEE) { 1825 Value *I3 = HEE->getOperand(0); 1826 if (!I2 && I3 != I1) 1827 I2 = I3; 1828 else if (I3 != I1 && I3 != I2) 1829 CanUseInputs = false; 1830 } else { 1831 Value *I3 = HSV->getOperand(0); 1832 if (!I2 && I3 != I1) 1833 I2 = I3; 1834 else if (I3 != I1 && I3 != I2) 1835 CanUseInputs = false; 1836 1837 if (CanUseInputs) { 1838 Value *I4 = HSV->getOperand(1); 1839 if (!isa<UndefValue>(I4)) { 1840 if (!I2 && I4 != I1) 1841 I2 = I4; 1842 else if (I4 != I1 && I4 != I2) 1843 CanUseInputs = false; 1844 } 1845 } 1846 } 1847 1848 if (CanUseInputs) { 1849 unsigned LOpElem = 1850 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 1851 ->getNumElements(); 1852 unsigned HOpElem = 1853 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 1854 ->getNumElements(); 1855 1856 // We have one or two input vectors. We need to map each index of the 1857 // operands to the index of the original vector. 1858 SmallVector<std::pair<int, int>, 8> II(numElem); 1859 for (unsigned i = 0; i < numElemL; ++i) { 1860 int Idx, INum; 1861 if (LEE) { 1862 Idx = 1863 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 1864 INum = LEE->getOperand(0) == I1 ? 0 : 1; 1865 } else { 1866 Idx = LSV->getMaskValue(i); 1867 if (Idx < (int) LOpElem) { 1868 INum = LSV->getOperand(0) == I1 ? 0 : 1; 1869 } else { 1870 Idx -= LOpElem; 1871 INum = LSV->getOperand(1) == I1 ? 0 : 1; 1872 } 1873 } 1874 1875 II[i] = std::pair<int, int>(Idx, INum); 1876 } 1877 for (unsigned i = 0; i < numElemH; ++i) { 1878 int Idx, INum; 1879 if (HEE) { 1880 Idx = 1881 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 1882 INum = HEE->getOperand(0) == I1 ? 0 : 1; 1883 } else { 1884 Idx = HSV->getMaskValue(i); 1885 if (Idx < (int) HOpElem) { 1886 INum = HSV->getOperand(0) == I1 ? 0 : 1; 1887 } else { 1888 Idx -= HOpElem; 1889 INum = HSV->getOperand(1) == I1 ? 0 : 1; 1890 } 1891 } 1892 1893 II[i + numElemL] = std::pair<int, int>(Idx, INum); 1894 } 1895 1896 // We now have an array which tells us from which index of which 1897 // input vector each element of the operand comes. 1898 VectorType *I1T = cast<VectorType>(I1->getType()); 1899 unsigned I1Elem = I1T->getNumElements(); 1900 1901 if (!I2) { 1902 // In this case there is only one underlying vector input. Check for 1903 // the trivial case where we can use the input directly. 1904 if (I1Elem == numElem) { 1905 bool ElemInOrder = true; 1906 for (unsigned i = 0; i < numElem; ++i) { 1907 if (II[i].first != (int) i && II[i].first != -1) { 1908 ElemInOrder = false; 1909 break; 1910 } 1911 } 1912 1913 if (ElemInOrder) 1914 return I1; 1915 } 1916 1917 // A shuffle is needed. 1918 std::vector<Constant *> Mask(numElem); 1919 for (unsigned i = 0; i < numElem; ++i) { 1920 int Idx = II[i].first; 1921 if (Idx == -1) 1922 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 1923 else 1924 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 1925 } 1926 1927 Instruction *S = 1928 new ShuffleVectorInst(I1, UndefValue::get(I1T), 1929 ConstantVector::get(Mask), 1930 getReplacementName(I, true, o)); 1931 S->insertBefore(J); 1932 return S; 1933 } 1934 1935 VectorType *I2T = cast<VectorType>(I2->getType()); 1936 unsigned I2Elem = I2T->getNumElements(); 1937 1938 // This input comes from two distinct vectors. The first step is to 1939 // make sure that both vectors are the same length. If not, the 1940 // smaller one will need to grow before they can be shuffled together. 1941 if (I1Elem < I2Elem) { 1942 std::vector<Constant *> Mask(I2Elem); 1943 unsigned v = 0; 1944 for (; v < I1Elem; ++v) 1945 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1946 for (; v < I2Elem; ++v) 1947 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1948 1949 Instruction *NewI1 = 1950 new ShuffleVectorInst(I1, UndefValue::get(I1T), 1951 ConstantVector::get(Mask), 1952 getReplacementName(I, true, o, 1)); 1953 NewI1->insertBefore(J); 1954 I1 = NewI1; 1955 I1T = I2T; 1956 I1Elem = I2Elem; 1957 } else if (I1Elem > I2Elem) { 1958 std::vector<Constant *> Mask(I1Elem); 1959 unsigned v = 0; 1960 for (; v < I2Elem; ++v) 1961 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1962 for (; v < I1Elem; ++v) 1963 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1964 1965 Instruction *NewI2 = 1966 new ShuffleVectorInst(I2, UndefValue::get(I2T), 1967 ConstantVector::get(Mask), 1968 getReplacementName(I, true, o, 1)); 1969 NewI2->insertBefore(J); 1970 I2 = NewI2; 1971 I2T = I1T; 1972 I2Elem = I1Elem; 1973 } 1974 1975 // Now that both I1 and I2 are the same length we can shuffle them 1976 // together (and use the result). 1977 std::vector<Constant *> Mask(numElem); 1978 for (unsigned v = 0; v < numElem; ++v) { 1979 if (II[v].first == -1) { 1980 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1981 } else { 1982 int Idx = II[v].first + II[v].second * I1Elem; 1983 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 1984 } 1985 } 1986 1987 Instruction *NewOp = 1988 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 1989 getReplacementName(I, true, o)); 1990 NewOp->insertBefore(J); 1991 return NewOp; 1992 } 1993 } 1994 1995 Type *ArgType = ArgTypeL; 1996 if (numElemL < numElemH) { 1997 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 1998 ArgTypeL, VArgType, 1)) { 1999 // This is another short-circuit case: we're combining a scalar into 2000 // a vector that is formed by an IE chain. We've just expanded the IE 2001 // chain, now insert the scalar and we're done. 2002 2003 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2004 getReplacementName(I, true, o)); 2005 S->insertBefore(J); 2006 return S; 2007 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2008 ArgTypeH)) { 2009 // The two vector inputs to the shuffle must be the same length, 2010 // so extend the smaller vector to be the same length as the larger one. 2011 Instruction *NLOp; 2012 if (numElemL > 1) { 2013 2014 std::vector<Constant *> Mask(numElemH); 2015 unsigned v = 0; 2016 for (; v < numElemL; ++v) 2017 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2018 for (; v < numElemH; ++v) 2019 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2020 2021 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2022 ConstantVector::get(Mask), 2023 getReplacementName(I, true, o, 1)); 2024 } else { 2025 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2026 getReplacementName(I, true, o, 1)); 2027 } 2028 2029 NLOp->insertBefore(J); 2030 LOp = NLOp; 2031 } 2032 2033 ArgType = ArgTypeH; 2034 } else if (numElemL > numElemH) { 2035 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2036 ArgTypeH, VArgType)) { 2037 Instruction *S = 2038 InsertElementInst::Create(LOp, HOp, 2039 ConstantInt::get(Type::getInt32Ty(Context), 2040 numElemL), 2041 getReplacementName(I, true, o)); 2042 S->insertBefore(J); 2043 return S; 2044 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2045 ArgTypeL)) { 2046 Instruction *NHOp; 2047 if (numElemH > 1) { 2048 std::vector<Constant *> Mask(numElemL); 2049 unsigned v = 0; 2050 for (; v < numElemH; ++v) 2051 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2052 for (; v < numElemL; ++v) 2053 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2054 2055 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2056 ConstantVector::get(Mask), 2057 getReplacementName(I, true, o, 1)); 2058 } else { 2059 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2060 getReplacementName(I, true, o, 1)); 2061 } 2062 2063 NHOp->insertBefore(J); 2064 HOp = NHOp; 2065 } 2066 } 2067 2068 if (ArgType->isVectorTy()) { 2069 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2070 std::vector<Constant*> Mask(numElem); 2071 for (unsigned v = 0; v < numElem; ++v) { 2072 unsigned Idx = v; 2073 // If the low vector was expanded, we need to skip the extra 2074 // undefined entries. 2075 if (v >= numElemL && numElemH > numElemL) 2076 Idx += (numElemH - numElemL); 2077 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2078 } 2079 2080 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2081 ConstantVector::get(Mask), 2082 getReplacementName(I, true, o)); 2083 BV->insertBefore(J); 2084 return BV; 2085 } 2086 2087 Instruction *BV1 = InsertElementInst::Create( 2088 UndefValue::get(VArgType), LOp, CV0, 2089 getReplacementName(I, true, o, 1)); 2090 BV1->insertBefore(I); 2091 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2092 getReplacementName(I, true, o, 2)); 2093 BV2->insertBefore(J); 2094 return BV2; 2095 } 2096 2097 // This function creates an array of values that will be used as the inputs 2098 // to the vector instruction that fuses I with J. 2099 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2100 Instruction *I, Instruction *J, 2101 SmallVector<Value *, 3> &ReplacedOperands, 2102 bool FlipMemInputs) { 2103 unsigned NumOperands = I->getNumOperands(); 2104 2105 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2106 // Iterate backward so that we look at the store pointer 2107 // first and know whether or not we need to flip the inputs. 2108 2109 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2110 // This is the pointer for a load/store instruction. 2111 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 2112 FlipMemInputs); 2113 continue; 2114 } else if (isa<CallInst>(I)) { 2115 Function *F = cast<CallInst>(I)->getCalledFunction(); 2116 unsigned IID = F->getIntrinsicID(); 2117 if (o == NumOperands-1) { 2118 BasicBlock &BB = *I->getParent(); 2119 2120 Module *M = BB.getParent()->getParent(); 2121 Type *ArgTypeI = I->getType(); 2122 Type *ArgTypeJ = J->getType(); 2123 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2124 2125 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 2126 (Intrinsic::ID) IID, VArgType); 2127 continue; 2128 } else if (IID == Intrinsic::powi && o == 1) { 2129 // The second argument of powi is a single integer and we've already 2130 // checked that both arguments are equal. As a result, we just keep 2131 // I's second argument. 2132 ReplacedOperands[o] = I->getOperand(o); 2133 continue; 2134 } 2135 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2136 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2137 continue; 2138 } 2139 2140 ReplacedOperands[o] = 2141 getReplacementInput(Context, I, J, o, FlipMemInputs); 2142 } 2143 } 2144 2145 // This function creates two values that represent the outputs of the 2146 // original I and J instructions. These are generally vector shuffles 2147 // or extracts. In many cases, these will end up being unused and, thus, 2148 // eliminated by later passes. 2149 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2150 Instruction *J, Instruction *K, 2151 Instruction *&InsertionPt, 2152 Instruction *&K1, Instruction *&K2, 2153 bool FlipMemInputs) { 2154 if (isa<StoreInst>(I)) { 2155 AA->replaceWithNewValue(I, K); 2156 AA->replaceWithNewValue(J, K); 2157 } else { 2158 Type *IType = I->getType(); 2159 Type *JType = J->getType(); 2160 2161 VectorType *VType = getVecTypeForPair(IType, JType); 2162 unsigned numElem = VType->getNumElements(); 2163 2164 unsigned numElemI, numElemJ; 2165 if (IType->isVectorTy()) 2166 numElemI = cast<VectorType>(IType)->getNumElements(); 2167 else 2168 numElemI = 1; 2169 2170 if (JType->isVectorTy()) 2171 numElemJ = cast<VectorType>(JType)->getNumElements(); 2172 else 2173 numElemJ = 1; 2174 2175 if (IType->isVectorTy()) { 2176 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2177 for (unsigned v = 0; v < numElemI; ++v) { 2178 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2179 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2180 } 2181 2182 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2183 ConstantVector::get( 2184 FlipMemInputs ? Mask2 : Mask1), 2185 getReplacementName(K, false, 1)); 2186 } else { 2187 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2188 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2189 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 2190 getReplacementName(K, false, 1)); 2191 } 2192 2193 if (JType->isVectorTy()) { 2194 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2195 for (unsigned v = 0; v < numElemJ; ++v) { 2196 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2197 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2198 } 2199 2200 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2201 ConstantVector::get( 2202 FlipMemInputs ? Mask1 : Mask2), 2203 getReplacementName(K, false, 2)); 2204 } else { 2205 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2206 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2207 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 2208 getReplacementName(K, false, 2)); 2209 } 2210 2211 K1->insertAfter(K); 2212 K2->insertAfter(K1); 2213 InsertionPt = K2; 2214 } 2215 } 2216 2217 // Move all uses of the function I (including pairing-induced uses) after J. 2218 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2219 std::multimap<Value *, Value *> &LoadMoveSet, 2220 Instruction *I, Instruction *J) { 2221 // Skip to the first instruction past I. 2222 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2223 2224 DenseSet<Value *> Users; 2225 AliasSetTracker WriteSet(*AA); 2226 for (; cast<Instruction>(L) != J; ++L) 2227 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 2228 2229 assert(cast<Instruction>(L) == J && 2230 "Tracking has not proceeded far enough to check for dependencies"); 2231 // If J is now in the use set of I, then trackUsesOfI will return true 2232 // and we have a dependency cycle (and the fusing operation must abort). 2233 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 2234 } 2235 2236 // Move all uses of the function I (including pairing-induced uses) after J. 2237 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2238 std::multimap<Value *, Value *> &LoadMoveSet, 2239 Instruction *&InsertionPt, 2240 Instruction *I, Instruction *J) { 2241 // Skip to the first instruction past I. 2242 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2243 2244 DenseSet<Value *> Users; 2245 AliasSetTracker WriteSet(*AA); 2246 for (; cast<Instruction>(L) != J;) { 2247 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 2248 // Move this instruction 2249 Instruction *InstToMove = L; ++L; 2250 2251 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2252 " to after " << *InsertionPt << "\n"); 2253 InstToMove->removeFromParent(); 2254 InstToMove->insertAfter(InsertionPt); 2255 InsertionPt = InstToMove; 2256 } else { 2257 ++L; 2258 } 2259 } 2260 } 2261 2262 // Collect all load instruction that are in the move set of a given first 2263 // pair member. These loads depend on the first instruction, I, and so need 2264 // to be moved after J (the second instruction) when the pair is fused. 2265 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2266 DenseMap<Value *, Value *> &ChosenPairs, 2267 std::multimap<Value *, Value *> &LoadMoveSet, 2268 Instruction *I) { 2269 // Skip to the first instruction past I. 2270 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2271 2272 DenseSet<Value *> Users; 2273 AliasSetTracker WriteSet(*AA); 2274 2275 // Note: We cannot end the loop when we reach J because J could be moved 2276 // farther down the use chain by another instruction pairing. Also, J 2277 // could be before I if this is an inverted input. 2278 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2279 if (trackUsesOfI(Users, WriteSet, I, L)) { 2280 if (L->mayReadFromMemory()) 2281 LoadMoveSet.insert(ValuePair(L, I)); 2282 } 2283 } 2284 } 2285 2286 // In cases where both load/stores and the computation of their pointers 2287 // are chosen for vectorization, we can end up in a situation where the 2288 // aliasing analysis starts returning different query results as the 2289 // process of fusing instruction pairs continues. Because the algorithm 2290 // relies on finding the same use trees here as were found earlier, we'll 2291 // need to precompute the necessary aliasing information here and then 2292 // manually update it during the fusion process. 2293 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2294 std::vector<Value *> &PairableInsts, 2295 DenseMap<Value *, Value *> &ChosenPairs, 2296 std::multimap<Value *, Value *> &LoadMoveSet) { 2297 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2298 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2299 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2300 if (P == ChosenPairs.end()) continue; 2301 2302 Instruction *I = cast<Instruction>(P->first); 2303 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 2304 } 2305 } 2306 2307 // As with the aliasing information, SCEV can also change because of 2308 // vectorization. This information is used to compute relative pointer 2309 // offsets; the necessary information will be cached here prior to 2310 // fusion. 2311 void BBVectorize::collectPtrInfo(std::vector<Value *> &PairableInsts, 2312 DenseMap<Value *, Value *> &ChosenPairs, 2313 DenseSet<Value *> &LowPtrInsts) { 2314 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2315 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2316 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2317 if (P == ChosenPairs.end()) continue; 2318 2319 Instruction *I = cast<Instruction>(P->first); 2320 Instruction *J = cast<Instruction>(P->second); 2321 2322 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 2323 continue; 2324 2325 Value *IPtr, *JPtr; 2326 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2327 int64_t OffsetInElmts; 2328 if (!getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2329 IAddressSpace, JAddressSpace, 2330 OffsetInElmts) || abs64(OffsetInElmts) != 1) 2331 llvm_unreachable("Pre-fusion pointer analysis failed"); 2332 2333 Value *LowPI = (OffsetInElmts > 0) ? I : J; 2334 LowPtrInsts.insert(LowPI); 2335 } 2336 } 2337 2338 // When the first instruction in each pair is cloned, it will inherit its 2339 // parent's metadata. This metadata must be combined with that of the other 2340 // instruction in a safe way. 2341 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2342 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2343 K->getAllMetadataOtherThanDebugLoc(Metadata); 2344 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2345 unsigned Kind = Metadata[i].first; 2346 MDNode *JMD = J->getMetadata(Kind); 2347 MDNode *KMD = Metadata[i].second; 2348 2349 switch (Kind) { 2350 default: 2351 K->setMetadata(Kind, 0); // Remove unknown metadata 2352 break; 2353 case LLVMContext::MD_tbaa: 2354 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2355 break; 2356 case LLVMContext::MD_fpmath: 2357 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2358 break; 2359 } 2360 } 2361 } 2362 2363 // This function fuses the chosen instruction pairs into vector instructions, 2364 // taking care preserve any needed scalar outputs and, then, it reorders the 2365 // remaining instructions as needed (users of the first member of the pair 2366 // need to be moved to after the location of the second member of the pair 2367 // because the vector instruction is inserted in the location of the pair's 2368 // second member). 2369 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2370 std::vector<Value *> &PairableInsts, 2371 DenseMap<Value *, Value *> &ChosenPairs) { 2372 LLVMContext& Context = BB.getContext(); 2373 2374 // During the vectorization process, the order of the pairs to be fused 2375 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2376 // list. After a pair is fused, the flipped pair is removed from the list. 2377 std::vector<ValuePair> FlippedPairs; 2378 FlippedPairs.reserve(ChosenPairs.size()); 2379 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2380 E = ChosenPairs.end(); P != E; ++P) 2381 FlippedPairs.push_back(ValuePair(P->second, P->first)); 2382 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 2383 E = FlippedPairs.end(); P != E; ++P) 2384 ChosenPairs.insert(*P); 2385 2386 std::multimap<Value *, Value *> LoadMoveSet; 2387 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 2388 2389 DenseSet<Value *> LowPtrInsts; 2390 collectPtrInfo(PairableInsts, ChosenPairs, LowPtrInsts); 2391 2392 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2393 2394 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2395 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2396 if (P == ChosenPairs.end()) { 2397 ++PI; 2398 continue; 2399 } 2400 2401 if (getDepthFactor(P->first) == 0) { 2402 // These instructions are not really fused, but are tracked as though 2403 // they are. Any case in which it would be interesting to fuse them 2404 // will be taken care of by InstCombine. 2405 --NumFusedOps; 2406 ++PI; 2407 continue; 2408 } 2409 2410 Instruction *I = cast<Instruction>(P->first), 2411 *J = cast<Instruction>(P->second); 2412 2413 DEBUG(dbgs() << "BBV: fusing: " << *I << 2414 " <-> " << *J << "\n"); 2415 2416 // Remove the pair and flipped pair from the list. 2417 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2418 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2419 ChosenPairs.erase(FP); 2420 ChosenPairs.erase(P); 2421 2422 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 2423 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2424 " <-> " << *J << 2425 " aborted because of non-trivial dependency cycle\n"); 2426 --NumFusedOps; 2427 ++PI; 2428 continue; 2429 } 2430 2431 bool FlipMemInputs = false; 2432 if (isa<LoadInst>(I) || isa<StoreInst>(I)) 2433 FlipMemInputs = (LowPtrInsts.find(I) == LowPtrInsts.end()); 2434 2435 unsigned NumOperands = I->getNumOperands(); 2436 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2437 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 2438 FlipMemInputs); 2439 2440 // Make a copy of the original operation, change its type to the vector 2441 // type and replace its operands with the vector operands. 2442 Instruction *K = I->clone(); 2443 if (I->hasName()) K->takeName(I); 2444 2445 if (!isa<StoreInst>(K)) 2446 K->mutateType(getVecTypeForPair(I->getType(), J->getType())); 2447 2448 combineMetadata(K, J); 2449 2450 for (unsigned o = 0; o < NumOperands; ++o) 2451 K->setOperand(o, ReplacedOperands[o]); 2452 2453 // If we've flipped the memory inputs, make sure that we take the correct 2454 // alignment. 2455 if (FlipMemInputs) { 2456 if (isa<StoreInst>(K)) 2457 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 2458 else 2459 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 2460 } 2461 2462 K->insertAfter(J); 2463 2464 // Instruction insertion point: 2465 Instruction *InsertionPt = K; 2466 Instruction *K1 = 0, *K2 = 0; 2467 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 2468 FlipMemInputs); 2469 2470 // The use tree of the first original instruction must be moved to after 2471 // the location of the second instruction. The entire use tree of the 2472 // first instruction is disjoint from the input tree of the second 2473 // (by definition), and so commutes with it. 2474 2475 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 2476 2477 if (!isa<StoreInst>(I)) { 2478 I->replaceAllUsesWith(K1); 2479 J->replaceAllUsesWith(K2); 2480 AA->replaceWithNewValue(I, K1); 2481 AA->replaceWithNewValue(J, K2); 2482 } 2483 2484 // Instructions that may read from memory may be in the load move set. 2485 // Once an instruction is fused, we no longer need its move set, and so 2486 // the values of the map never need to be updated. However, when a load 2487 // is fused, we need to merge the entries from both instructions in the 2488 // pair in case those instructions were in the move set of some other 2489 // yet-to-be-fused pair. The loads in question are the keys of the map. 2490 if (I->mayReadFromMemory()) { 2491 std::vector<ValuePair> NewSetMembers; 2492 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 2493 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 2494 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 2495 N != IPairRange.second; ++N) 2496 NewSetMembers.push_back(ValuePair(K, N->second)); 2497 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 2498 N != JPairRange.second; ++N) 2499 NewSetMembers.push_back(ValuePair(K, N->second)); 2500 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 2501 AE = NewSetMembers.end(); A != AE; ++A) 2502 LoadMoveSet.insert(*A); 2503 } 2504 2505 // Before removing I, set the iterator to the next instruction. 2506 PI = llvm::next(BasicBlock::iterator(I)); 2507 if (cast<Instruction>(PI) == J) 2508 ++PI; 2509 2510 SE->forgetValue(I); 2511 SE->forgetValue(J); 2512 I->eraseFromParent(); 2513 J->eraseFromParent(); 2514 } 2515 2516 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 2517 } 2518} 2519 2520char BBVectorize::ID = 0; 2521static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 2522INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2523INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2524INITIALIZE_PASS_DEPENDENCY(DominatorTree) 2525INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2526INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2527 2528BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 2529 return new BBVectorize(C); 2530} 2531 2532bool 2533llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 2534 BBVectorize BBVectorizer(P, C); 2535 return BBVectorizer.vectorizeBB(BB); 2536} 2537 2538//===----------------------------------------------------------------------===// 2539VectorizeConfig::VectorizeConfig() { 2540 VectorBits = ::VectorBits; 2541 VectorizeBools = !::NoBools; 2542 VectorizeInts = !::NoInts; 2543 VectorizeFloats = !::NoFloats; 2544 VectorizePointers = !::NoPointers; 2545 VectorizeCasts = !::NoCasts; 2546 VectorizeMath = !::NoMath; 2547 VectorizeFMA = !::NoFMA; 2548 VectorizeSelect = !::NoSelect; 2549 VectorizeCmp = !::NoCmp; 2550 VectorizeGEP = !::NoGEP; 2551 VectorizeMemOps = !::NoMemOps; 2552 AlignedOnly = ::AlignedOnly; 2553 ReqChainDepth= ::ReqChainDepth; 2554 SearchLimit = ::SearchLimit; 2555 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 2556 SplatBreaksChain = ::SplatBreaksChain; 2557 MaxInsts = ::MaxInsts; 2558 MaxIter = ::MaxIter; 2559 Pow2LenOnly = ::Pow2LenOnly; 2560 NoMemOpBoost = ::NoMemOpBoost; 2561 FastDep = ::FastDep; 2562} 2563