BBVectorize.cpp revision e29c19091cca58db668407dfc5dd86c70e8b3d49
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/Transforms/Utils/Local.h" 47#include "llvm/Transforms/Vectorize.h" 48#include <algorithm> 49#include <map> 50using namespace llvm; 51 52static cl::opt<unsigned> 53ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 54 cl::desc("The required chain depth for vectorization")); 55 56static cl::opt<unsigned> 57SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 58 cl::desc("The maximum search distance for instruction pairs")); 59 60static cl::opt<bool> 61SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 62 cl::desc("Replicating one element to a pair breaks the chain")); 63 64static cl::opt<unsigned> 65VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 66 cl::desc("The size of the native vector registers")); 67 68static cl::opt<unsigned> 69MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 70 cl::desc("The maximum number of pairing iterations")); 71 72static cl::opt<bool> 73Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden, 74 cl::desc("Don't try to form non-2^n-length vectors")); 75 76static cl::opt<unsigned> 77MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 78 cl::desc("The maximum number of pairable instructions per group")); 79 80static cl::opt<unsigned> 81MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 82 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 83 " a full cycle check")); 84 85static cl::opt<bool> 86NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden, 87 cl::desc("Don't try to vectorize boolean (i1) values")); 88 89static cl::opt<bool> 90NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 91 cl::desc("Don't try to vectorize integer values")); 92 93static cl::opt<bool> 94NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 95 cl::desc("Don't try to vectorize floating-point values")); 96 97static cl::opt<bool> 98NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden, 99 cl::desc("Don't try to vectorize pointer values")); 100 101static cl::opt<bool> 102NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 103 cl::desc("Don't try to vectorize casting (conversion) operations")); 104 105static cl::opt<bool> 106NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 107 cl::desc("Don't try to vectorize floating-point math intrinsics")); 108 109static cl::opt<bool> 110NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 111 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 112 113static cl::opt<bool> 114NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 115 cl::desc("Don't try to vectorize select instructions")); 116 117static cl::opt<bool> 118NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 119 cl::desc("Don't try to vectorize comparison instructions")); 120 121static cl::opt<bool> 122NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 123 cl::desc("Don't try to vectorize getelementptr instructions")); 124 125static cl::opt<bool> 126NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 127 cl::desc("Don't try to vectorize loads and stores")); 128 129static cl::opt<bool> 130AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 131 cl::desc("Only generate aligned loads and stores")); 132 133static cl::opt<bool> 134NoMemOpBoost("bb-vectorize-no-mem-op-boost", 135 cl::init(false), cl::Hidden, 136 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 137 138static cl::opt<bool> 139FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 140 cl::desc("Use a fast instruction dependency analysis")); 141 142#ifndef NDEBUG 143static cl::opt<bool> 144DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 145 cl::init(false), cl::Hidden, 146 cl::desc("When debugging is enabled, output information on the" 147 " instruction-examination process")); 148static cl::opt<bool> 149DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 150 cl::init(false), cl::Hidden, 151 cl::desc("When debugging is enabled, output information on the" 152 " candidate-selection process")); 153static cl::opt<bool> 154DebugPairSelection("bb-vectorize-debug-pair-selection", 155 cl::init(false), cl::Hidden, 156 cl::desc("When debugging is enabled, output information on the" 157 " pair-selection process")); 158static cl::opt<bool> 159DebugCycleCheck("bb-vectorize-debug-cycle-check", 160 cl::init(false), cl::Hidden, 161 cl::desc("When debugging is enabled, output information on the" 162 " cycle-checking process")); 163#endif 164 165STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 166 167namespace { 168 struct BBVectorize : public BasicBlockPass { 169 static char ID; // Pass identification, replacement for typeid 170 171 const VectorizeConfig Config; 172 173 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 174 : BasicBlockPass(ID), Config(C) { 175 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 176 } 177 178 BBVectorize(Pass *P, const VectorizeConfig &C) 179 : BasicBlockPass(ID), Config(C) { 180 AA = &P->getAnalysis<AliasAnalysis>(); 181 DT = &P->getAnalysis<DominatorTree>(); 182 SE = &P->getAnalysis<ScalarEvolution>(); 183 TD = P->getAnalysisIfAvailable<DataLayout>(); 184 } 185 186 typedef std::pair<Value *, Value *> ValuePair; 187 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 188 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 189 typedef std::pair<std::multimap<Value *, Value *>::iterator, 190 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 191 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 192 std::multimap<ValuePair, ValuePair>::iterator> 193 VPPIteratorPair; 194 195 AliasAnalysis *AA; 196 DominatorTree *DT; 197 ScalarEvolution *SE; 198 DataLayout *TD; 199 200 // FIXME: const correct? 201 202 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 203 204 bool getCandidatePairs(BasicBlock &BB, 205 BasicBlock::iterator &Start, 206 std::multimap<Value *, Value *> &CandidatePairs, 207 std::vector<Value *> &PairableInsts, bool NonPow2Len); 208 209 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 210 std::vector<Value *> &PairableInsts, 211 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 212 213 void buildDepMap(BasicBlock &BB, 214 std::multimap<Value *, Value *> &CandidatePairs, 215 std::vector<Value *> &PairableInsts, 216 DenseSet<ValuePair> &PairableInstUsers); 217 218 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 219 std::vector<Value *> &PairableInsts, 220 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 221 DenseSet<ValuePair> &PairableInstUsers, 222 DenseMap<Value *, Value *>& ChosenPairs); 223 224 void fuseChosenPairs(BasicBlock &BB, 225 std::vector<Value *> &PairableInsts, 226 DenseMap<Value *, Value *>& ChosenPairs); 227 228 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 229 230 bool areInstsCompatible(Instruction *I, Instruction *J, 231 bool IsSimpleLoadStore, bool NonPow2Len); 232 233 bool trackUsesOfI(DenseSet<Value *> &Users, 234 AliasSetTracker &WriteSet, Instruction *I, 235 Instruction *J, bool UpdateUsers = true, 236 std::multimap<Value *, Value *> *LoadMoveSet = 0); 237 238 void computePairsConnectedTo( 239 std::multimap<Value *, Value *> &CandidatePairs, 240 std::vector<Value *> &PairableInsts, 241 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 242 ValuePair P); 243 244 bool pairsConflict(ValuePair P, ValuePair Q, 245 DenseSet<ValuePair> &PairableInstUsers, 246 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 247 248 bool pairWillFormCycle(ValuePair P, 249 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 250 DenseSet<ValuePair> &CurrentPairs); 251 252 void pruneTreeFor( 253 std::multimap<Value *, Value *> &CandidatePairs, 254 std::vector<Value *> &PairableInsts, 255 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 256 DenseSet<ValuePair> &PairableInstUsers, 257 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 258 DenseMap<Value *, Value *> &ChosenPairs, 259 DenseMap<ValuePair, size_t> &Tree, 260 DenseSet<ValuePair> &PrunedTree, ValuePair J, 261 bool UseCycleCheck); 262 263 void buildInitialTreeFor( 264 std::multimap<Value *, Value *> &CandidatePairs, 265 std::vector<Value *> &PairableInsts, 266 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 267 DenseSet<ValuePair> &PairableInstUsers, 268 DenseMap<Value *, Value *> &ChosenPairs, 269 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 270 271 void findBestTreeFor( 272 std::multimap<Value *, Value *> &CandidatePairs, 273 std::vector<Value *> &PairableInsts, 274 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 275 DenseSet<ValuePair> &PairableInstUsers, 276 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 277 DenseMap<Value *, Value *> &ChosenPairs, 278 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 279 size_t &BestEffSize, VPIteratorPair ChoiceRange, 280 bool UseCycleCheck); 281 282 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 283 Instruction *J, unsigned o, bool FlipMemInputs); 284 285 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 286 unsigned MaskOffset, unsigned NumInElem, 287 unsigned NumInElem1, unsigned IdxOffset, 288 std::vector<Constant*> &Mask); 289 290 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 291 Instruction *J); 292 293 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 294 unsigned o, Value *&LOp, unsigned numElemL, 295 Type *ArgTypeL, Type *ArgTypeR, 296 unsigned IdxOff = 0); 297 298 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 299 Instruction *J, unsigned o, bool FlipMemInputs); 300 301 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 302 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 303 bool FlipMemInputs); 304 305 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 306 Instruction *J, Instruction *K, 307 Instruction *&InsertionPt, Instruction *&K1, 308 Instruction *&K2, bool FlipMemInputs); 309 310 void collectPairLoadMoveSet(BasicBlock &BB, 311 DenseMap<Value *, Value *> &ChosenPairs, 312 std::multimap<Value *, Value *> &LoadMoveSet, 313 Instruction *I); 314 315 void collectLoadMoveSet(BasicBlock &BB, 316 std::vector<Value *> &PairableInsts, 317 DenseMap<Value *, Value *> &ChosenPairs, 318 std::multimap<Value *, Value *> &LoadMoveSet); 319 320 void collectPtrInfo(std::vector<Value *> &PairableInsts, 321 DenseMap<Value *, Value *> &ChosenPairs, 322 DenseSet<Value *> &LowPtrInsts); 323 324 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 325 std::multimap<Value *, Value *> &LoadMoveSet, 326 Instruction *I, Instruction *J); 327 328 void moveUsesOfIAfterJ(BasicBlock &BB, 329 std::multimap<Value *, Value *> &LoadMoveSet, 330 Instruction *&InsertionPt, 331 Instruction *I, Instruction *J); 332 333 void combineMetadata(Instruction *K, const Instruction *J); 334 335 bool vectorizeBB(BasicBlock &BB) { 336 if (!DT->isReachableFromEntry(&BB)) { 337 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 338 " in " << BB.getParent()->getName() << "\n"); 339 return false; 340 } 341 342 bool changed = false; 343 // Iterate a sufficient number of times to merge types of size 1 bit, 344 // then 2 bits, then 4, etc. up to half of the target vector width of the 345 // target vector register. 346 unsigned n = 1; 347 for (unsigned v = 2; 348 v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter); 349 v *= 2, ++n) { 350 DEBUG(dbgs() << "BBV: fusing loop #" << n << 351 " for " << BB.getName() << " in " << 352 BB.getParent()->getName() << "...\n"); 353 if (vectorizePairs(BB)) 354 changed = true; 355 else 356 break; 357 } 358 359 if (changed && !Pow2LenOnly) { 360 ++n; 361 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 362 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 363 n << " for " << BB.getName() << " in " << 364 BB.getParent()->getName() << "...\n"); 365 if (!vectorizePairs(BB, true)) break; 366 } 367 } 368 369 DEBUG(dbgs() << "BBV: done!\n"); 370 return changed; 371 } 372 373 virtual bool runOnBasicBlock(BasicBlock &BB) { 374 AA = &getAnalysis<AliasAnalysis>(); 375 DT = &getAnalysis<DominatorTree>(); 376 SE = &getAnalysis<ScalarEvolution>(); 377 TD = getAnalysisIfAvailable<DataLayout>(); 378 379 return vectorizeBB(BB); 380 } 381 382 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 383 BasicBlockPass::getAnalysisUsage(AU); 384 AU.addRequired<AliasAnalysis>(); 385 AU.addRequired<DominatorTree>(); 386 AU.addRequired<ScalarEvolution>(); 387 AU.addPreserved<AliasAnalysis>(); 388 AU.addPreserved<DominatorTree>(); 389 AU.addPreserved<ScalarEvolution>(); 390 AU.setPreservesCFG(); 391 } 392 393 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 394 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 395 "Cannot form vector from incompatible scalar types"); 396 Type *STy = ElemTy->getScalarType(); 397 398 unsigned numElem; 399 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 400 numElem = VTy->getNumElements(); 401 } else { 402 numElem = 1; 403 } 404 405 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 406 numElem += VTy->getNumElements(); 407 } else { 408 numElem += 1; 409 } 410 411 return VectorType::get(STy, numElem); 412 } 413 414 static inline void getInstructionTypes(Instruction *I, 415 Type *&T1, Type *&T2) { 416 if (isa<StoreInst>(I)) { 417 // For stores, it is the value type, not the pointer type that matters 418 // because the value is what will come from a vector register. 419 420 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 421 T1 = IVal->getType(); 422 } else { 423 T1 = I->getType(); 424 } 425 426 if (I->isCast()) 427 T2 = cast<CastInst>(I)->getSrcTy(); 428 else 429 T2 = T1; 430 } 431 432 // Returns the weight associated with the provided value. A chain of 433 // candidate pairs has a length given by the sum of the weights of its 434 // members (one weight per pair; the weight of each member of the pair 435 // is assumed to be the same). This length is then compared to the 436 // chain-length threshold to determine if a given chain is significant 437 // enough to be vectorized. The length is also used in comparing 438 // candidate chains where longer chains are considered to be better. 439 // Note: when this function returns 0, the resulting instructions are 440 // not actually fused. 441 inline size_t getDepthFactor(Value *V) { 442 // InsertElement and ExtractElement have a depth factor of zero. This is 443 // for two reasons: First, they cannot be usefully fused. Second, because 444 // the pass generates a lot of these, they can confuse the simple metric 445 // used to compare the trees in the next iteration. Thus, giving them a 446 // weight of zero allows the pass to essentially ignore them in 447 // subsequent iterations when looking for vectorization opportunities 448 // while still tracking dependency chains that flow through those 449 // instructions. 450 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 451 return 0; 452 453 // Give a load or store half of the required depth so that load/store 454 // pairs will vectorize. 455 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 456 return Config.ReqChainDepth/2; 457 458 return 1; 459 } 460 461 // This determines the relative offset of two loads or stores, returning 462 // true if the offset could be determined to be some constant value. 463 // For example, if OffsetInElmts == 1, then J accesses the memory directly 464 // after I; if OffsetInElmts == -1 then I accesses the memory 465 // directly after J. 466 bool getPairPtrInfo(Instruction *I, Instruction *J, 467 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 468 int64_t &OffsetInElmts) { 469 OffsetInElmts = 0; 470 if (isa<LoadInst>(I)) { 471 IPtr = cast<LoadInst>(I)->getPointerOperand(); 472 JPtr = cast<LoadInst>(J)->getPointerOperand(); 473 IAlignment = cast<LoadInst>(I)->getAlignment(); 474 JAlignment = cast<LoadInst>(J)->getAlignment(); 475 } else { 476 IPtr = cast<StoreInst>(I)->getPointerOperand(); 477 JPtr = cast<StoreInst>(J)->getPointerOperand(); 478 IAlignment = cast<StoreInst>(I)->getAlignment(); 479 JAlignment = cast<StoreInst>(J)->getAlignment(); 480 } 481 482 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 483 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 484 485 // If this is a trivial offset, then we'll get something like 486 // 1*sizeof(type). With target data, which we need anyway, this will get 487 // constant folded into a number. 488 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 489 if (const SCEVConstant *ConstOffSCEV = 490 dyn_cast<SCEVConstant>(OffsetSCEV)) { 491 ConstantInt *IntOff = ConstOffSCEV->getValue(); 492 int64_t Offset = IntOff->getSExtValue(); 493 494 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 495 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 496 497 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 498 if (VTy != VTy2 && Offset < 0) { 499 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 500 OffsetInElmts = Offset/VTy2TSS; 501 return (abs64(Offset) % VTy2TSS) == 0; 502 } 503 504 OffsetInElmts = Offset/VTyTSS; 505 return (abs64(Offset) % VTyTSS) == 0; 506 } 507 508 return false; 509 } 510 511 // Returns true if the provided CallInst represents an intrinsic that can 512 // be vectorized. 513 bool isVectorizableIntrinsic(CallInst* I) { 514 Function *F = I->getCalledFunction(); 515 if (!F) return false; 516 517 unsigned IID = F->getIntrinsicID(); 518 if (!IID) return false; 519 520 switch(IID) { 521 default: 522 return false; 523 case Intrinsic::sqrt: 524 case Intrinsic::powi: 525 case Intrinsic::sin: 526 case Intrinsic::cos: 527 case Intrinsic::log: 528 case Intrinsic::log2: 529 case Intrinsic::log10: 530 case Intrinsic::exp: 531 case Intrinsic::exp2: 532 case Intrinsic::pow: 533 return Config.VectorizeMath; 534 case Intrinsic::fma: 535 return Config.VectorizeFMA; 536 } 537 } 538 539 // Returns true if J is the second element in some pair referenced by 540 // some multimap pair iterator pair. 541 template <typename V> 542 bool isSecondInIteratorPair(V J, std::pair< 543 typename std::multimap<V, V>::iterator, 544 typename std::multimap<V, V>::iterator> PairRange) { 545 for (typename std::multimap<V, V>::iterator K = PairRange.first; 546 K != PairRange.second; ++K) 547 if (K->second == J) return true; 548 549 return false; 550 } 551 }; 552 553 // This function implements one vectorization iteration on the provided 554 // basic block. It returns true if the block is changed. 555 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 556 bool ShouldContinue; 557 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 558 559 std::vector<Value *> AllPairableInsts; 560 DenseMap<Value *, Value *> AllChosenPairs; 561 562 do { 563 std::vector<Value *> PairableInsts; 564 std::multimap<Value *, Value *> CandidatePairs; 565 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 566 PairableInsts, NonPow2Len); 567 if (PairableInsts.empty()) continue; 568 569 // Now we have a map of all of the pairable instructions and we need to 570 // select the best possible pairing. A good pairing is one such that the 571 // users of the pair are also paired. This defines a (directed) forest 572 // over the pairs such that two pairs are connected iff the second pair 573 // uses the first. 574 575 // Note that it only matters that both members of the second pair use some 576 // element of the first pair (to allow for splatting). 577 578 std::multimap<ValuePair, ValuePair> ConnectedPairs; 579 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 580 if (ConnectedPairs.empty()) continue; 581 582 // Build the pairable-instruction dependency map 583 DenseSet<ValuePair> PairableInstUsers; 584 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 585 586 // There is now a graph of the connected pairs. For each variable, pick 587 // the pairing with the largest tree meeting the depth requirement on at 588 // least one branch. Then select all pairings that are part of that tree 589 // and remove them from the list of available pairings and pairable 590 // variables. 591 592 DenseMap<Value *, Value *> ChosenPairs; 593 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, 594 PairableInstUsers, ChosenPairs); 595 596 if (ChosenPairs.empty()) continue; 597 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 598 PairableInsts.end()); 599 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 600 } while (ShouldContinue); 601 602 if (AllChosenPairs.empty()) return false; 603 NumFusedOps += AllChosenPairs.size(); 604 605 // A set of pairs has now been selected. It is now necessary to replace the 606 // paired instructions with vector instructions. For this procedure each 607 // operand must be replaced with a vector operand. This vector is formed 608 // by using build_vector on the old operands. The replaced values are then 609 // replaced with a vector_extract on the result. Subsequent optimization 610 // passes should coalesce the build/extract combinations. 611 612 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 613 614 // It is important to cleanup here so that future iterations of this 615 // function have less work to do. 616 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 617 return true; 618 } 619 620 // This function returns true if the provided instruction is capable of being 621 // fused into a vector instruction. This determination is based only on the 622 // type and other attributes of the instruction. 623 bool BBVectorize::isInstVectorizable(Instruction *I, 624 bool &IsSimpleLoadStore) { 625 IsSimpleLoadStore = false; 626 627 if (CallInst *C = dyn_cast<CallInst>(I)) { 628 if (!isVectorizableIntrinsic(C)) 629 return false; 630 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 631 // Vectorize simple loads if possbile: 632 IsSimpleLoadStore = L->isSimple(); 633 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 634 return false; 635 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 636 // Vectorize simple stores if possbile: 637 IsSimpleLoadStore = S->isSimple(); 638 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 639 return false; 640 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 641 // We can vectorize casts, but not casts of pointer types, etc. 642 if (!Config.VectorizeCasts) 643 return false; 644 645 Type *SrcTy = C->getSrcTy(); 646 if (!SrcTy->isSingleValueType()) 647 return false; 648 649 Type *DestTy = C->getDestTy(); 650 if (!DestTy->isSingleValueType()) 651 return false; 652 } else if (isa<SelectInst>(I)) { 653 if (!Config.VectorizeSelect) 654 return false; 655 } else if (isa<CmpInst>(I)) { 656 if (!Config.VectorizeCmp) 657 return false; 658 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 659 if (!Config.VectorizeGEP) 660 return false; 661 662 // Currently, vector GEPs exist only with one index. 663 if (G->getNumIndices() != 1) 664 return false; 665 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 666 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 667 return false; 668 } 669 670 // We can't vectorize memory operations without target data 671 if (TD == 0 && IsSimpleLoadStore) 672 return false; 673 674 Type *T1, *T2; 675 getInstructionTypes(I, T1, T2); 676 677 // Not every type can be vectorized... 678 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 679 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 680 return false; 681 682 if (T1->getScalarSizeInBits() == 1 && T2->getScalarSizeInBits() == 1) { 683 if (!Config.VectorizeBools) 684 return false; 685 } else { 686 if (!Config.VectorizeInts 687 && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) 688 return false; 689 } 690 691 if (!Config.VectorizeFloats 692 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 693 return false; 694 695 // Don't vectorize target-specific types. 696 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 697 return false; 698 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 699 return false; 700 701 if ((!Config.VectorizePointers || TD == 0) && 702 (T1->getScalarType()->isPointerTy() || 703 T2->getScalarType()->isPointerTy())) 704 return false; 705 706 if (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 707 T2->getPrimitiveSizeInBits() >= Config.VectorBits) 708 return false; 709 710 return true; 711 } 712 713 // This function returns true if the two provided instructions are compatible 714 // (meaning that they can be fused into a vector instruction). This assumes 715 // that I has already been determined to be vectorizable and that J is not 716 // in the use tree of I. 717 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 718 bool IsSimpleLoadStore, bool NonPow2Len) { 719 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 720 " <-> " << *J << "\n"); 721 722 // Loads and stores can be merged if they have different alignments, 723 // but are otherwise the same. 724 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 725 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 726 return false; 727 728 Type *IT1, *IT2, *JT1, *JT2; 729 getInstructionTypes(I, IT1, IT2); 730 getInstructionTypes(J, JT1, JT2); 731 unsigned MaxTypeBits = std::max( 732 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 733 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 734 if (MaxTypeBits > Config.VectorBits) 735 return false; 736 737 // FIXME: handle addsub-type operations! 738 739 if (IsSimpleLoadStore) { 740 Value *IPtr, *JPtr; 741 unsigned IAlignment, JAlignment; 742 int64_t OffsetInElmts = 0; 743 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 744 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 745 if (Config.AlignedOnly) { 746 Type *aTypeI = isa<StoreInst>(I) ? 747 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 748 Type *aTypeJ = isa<StoreInst>(J) ? 749 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 750 751 // An aligned load or store is possible only if the instruction 752 // with the lower offset has an alignment suitable for the 753 // vector type. 754 755 unsigned BottomAlignment = IAlignment; 756 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 757 758 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 759 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 760 if (BottomAlignment < VecAlignment) 761 return false; 762 } 763 } else { 764 return false; 765 } 766 } 767 768 // The powi intrinsic is special because only the first argument is 769 // vectorized, the second arguments must be equal. 770 CallInst *CI = dyn_cast<CallInst>(I); 771 Function *FI; 772 if (CI && (FI = CI->getCalledFunction()) && 773 FI->getIntrinsicID() == Intrinsic::powi) { 774 775 Value *A1I = CI->getArgOperand(1), 776 *A1J = cast<CallInst>(J)->getArgOperand(1); 777 const SCEV *A1ISCEV = SE->getSCEV(A1I), 778 *A1JSCEV = SE->getSCEV(A1J); 779 return (A1ISCEV == A1JSCEV); 780 } 781 782 return true; 783 } 784 785 // Figure out whether or not J uses I and update the users and write-set 786 // structures associated with I. Specifically, Users represents the set of 787 // instructions that depend on I. WriteSet represents the set 788 // of memory locations that are dependent on I. If UpdateUsers is true, 789 // and J uses I, then Users is updated to contain J and WriteSet is updated 790 // to contain any memory locations to which J writes. The function returns 791 // true if J uses I. By default, alias analysis is used to determine 792 // whether J reads from memory that overlaps with a location in WriteSet. 793 // If LoadMoveSet is not null, then it is a previously-computed multimap 794 // where the key is the memory-based user instruction and the value is 795 // the instruction to be compared with I. So, if LoadMoveSet is provided, 796 // then the alias analysis is not used. This is necessary because this 797 // function is called during the process of moving instructions during 798 // vectorization and the results of the alias analysis are not stable during 799 // that process. 800 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 801 AliasSetTracker &WriteSet, Instruction *I, 802 Instruction *J, bool UpdateUsers, 803 std::multimap<Value *, Value *> *LoadMoveSet) { 804 bool UsesI = false; 805 806 // This instruction may already be marked as a user due, for example, to 807 // being a member of a selected pair. 808 if (Users.count(J)) 809 UsesI = true; 810 811 if (!UsesI) 812 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 813 JU != JE; ++JU) { 814 Value *V = *JU; 815 if (I == V || Users.count(V)) { 816 UsesI = true; 817 break; 818 } 819 } 820 if (!UsesI && J->mayReadFromMemory()) { 821 if (LoadMoveSet) { 822 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 823 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 824 } else { 825 for (AliasSetTracker::iterator W = WriteSet.begin(), 826 WE = WriteSet.end(); W != WE; ++W) { 827 if (W->aliasesUnknownInst(J, *AA)) { 828 UsesI = true; 829 break; 830 } 831 } 832 } 833 } 834 835 if (UsesI && UpdateUsers) { 836 if (J->mayWriteToMemory()) WriteSet.add(J); 837 Users.insert(J); 838 } 839 840 return UsesI; 841 } 842 843 // This function iterates over all instruction pairs in the provided 844 // basic block and collects all candidate pairs for vectorization. 845 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 846 BasicBlock::iterator &Start, 847 std::multimap<Value *, Value *> &CandidatePairs, 848 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 849 BasicBlock::iterator E = BB.end(); 850 if (Start == E) return false; 851 852 bool ShouldContinue = false, IAfterStart = false; 853 for (BasicBlock::iterator I = Start++; I != E; ++I) { 854 if (I == Start) IAfterStart = true; 855 856 bool IsSimpleLoadStore; 857 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 858 859 // Look for an instruction with which to pair instruction *I... 860 DenseSet<Value *> Users; 861 AliasSetTracker WriteSet(*AA); 862 bool JAfterStart = IAfterStart; 863 BasicBlock::iterator J = llvm::next(I); 864 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 865 if (J == Start) JAfterStart = true; 866 867 // Determine if J uses I, if so, exit the loop. 868 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 869 if (Config.FastDep) { 870 // Note: For this heuristic to be effective, independent operations 871 // must tend to be intermixed. This is likely to be true from some 872 // kinds of grouped loop unrolling (but not the generic LLVM pass), 873 // but otherwise may require some kind of reordering pass. 874 875 // When using fast dependency analysis, 876 // stop searching after first use: 877 if (UsesI) break; 878 } else { 879 if (UsesI) continue; 880 } 881 882 // J does not use I, and comes before the first use of I, so it can be 883 // merged with I if the instructions are compatible. 884 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len)) continue; 885 886 // J is a candidate for merging with I. 887 if (!PairableInsts.size() || 888 PairableInsts[PairableInsts.size()-1] != I) { 889 PairableInsts.push_back(I); 890 } 891 892 CandidatePairs.insert(ValuePair(I, J)); 893 894 // The next call to this function must start after the last instruction 895 // selected during this invocation. 896 if (JAfterStart) { 897 Start = llvm::next(J); 898 IAfterStart = JAfterStart = false; 899 } 900 901 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 902 << *I << " <-> " << *J << "\n"); 903 904 // If we have already found too many pairs, break here and this function 905 // will be called again starting after the last instruction selected 906 // during this invocation. 907 if (PairableInsts.size() >= Config.MaxInsts) { 908 ShouldContinue = true; 909 break; 910 } 911 } 912 913 if (ShouldContinue) 914 break; 915 } 916 917 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 918 << " instructions with candidate pairs\n"); 919 920 return ShouldContinue; 921 } 922 923 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 924 // it looks for pairs such that both members have an input which is an 925 // output of PI or PJ. 926 void BBVectorize::computePairsConnectedTo( 927 std::multimap<Value *, Value *> &CandidatePairs, 928 std::vector<Value *> &PairableInsts, 929 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 930 ValuePair P) { 931 StoreInst *SI, *SJ; 932 933 // For each possible pairing for this variable, look at the uses of 934 // the first value... 935 for (Value::use_iterator I = P.first->use_begin(), 936 E = P.first->use_end(); I != E; ++I) { 937 if (isa<LoadInst>(*I)) { 938 // A pair cannot be connected to a load because the load only takes one 939 // operand (the address) and it is a scalar even after vectorization. 940 continue; 941 } else if ((SI = dyn_cast<StoreInst>(*I)) && 942 P.first == SI->getPointerOperand()) { 943 // Similarly, a pair cannot be connected to a store through its 944 // pointer operand. 945 continue; 946 } 947 948 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 949 950 // For each use of the first variable, look for uses of the second 951 // variable... 952 for (Value::use_iterator J = P.second->use_begin(), 953 E2 = P.second->use_end(); J != E2; ++J) { 954 if ((SJ = dyn_cast<StoreInst>(*J)) && 955 P.second == SJ->getPointerOperand()) 956 continue; 957 958 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 959 960 // Look for <I, J>: 961 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 962 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 963 964 // Look for <J, I>: 965 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 966 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 967 } 968 969 if (Config.SplatBreaksChain) continue; 970 // Look for cases where just the first value in the pair is used by 971 // both members of another pair (splatting). 972 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 973 if ((SJ = dyn_cast<StoreInst>(*J)) && 974 P.first == SJ->getPointerOperand()) 975 continue; 976 977 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 978 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 979 } 980 } 981 982 if (Config.SplatBreaksChain) return; 983 // Look for cases where just the second value in the pair is used by 984 // both members of another pair (splatting). 985 for (Value::use_iterator I = P.second->use_begin(), 986 E = P.second->use_end(); I != E; ++I) { 987 if (isa<LoadInst>(*I)) 988 continue; 989 else if ((SI = dyn_cast<StoreInst>(*I)) && 990 P.second == SI->getPointerOperand()) 991 continue; 992 993 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 994 995 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 996 if ((SJ = dyn_cast<StoreInst>(*J)) && 997 P.second == SJ->getPointerOperand()) 998 continue; 999 1000 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 1001 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 1002 } 1003 } 1004 } 1005 1006 // This function figures out which pairs are connected. Two pairs are 1007 // connected if some output of the first pair forms an input to both members 1008 // of the second pair. 1009 void BBVectorize::computeConnectedPairs( 1010 std::multimap<Value *, Value *> &CandidatePairs, 1011 std::vector<Value *> &PairableInsts, 1012 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 1013 1014 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1015 PE = PairableInsts.end(); PI != PE; ++PI) { 1016 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1017 1018 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1019 P != choiceRange.second; ++P) 1020 computePairsConnectedTo(CandidatePairs, PairableInsts, 1021 ConnectedPairs, *P); 1022 } 1023 1024 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1025 << " pair connections.\n"); 1026 } 1027 1028 // This function builds a set of use tuples such that <A, B> is in the set 1029 // if B is in the use tree of A. If B is in the use tree of A, then B 1030 // depends on the output of A. 1031 void BBVectorize::buildDepMap( 1032 BasicBlock &BB, 1033 std::multimap<Value *, Value *> &CandidatePairs, 1034 std::vector<Value *> &PairableInsts, 1035 DenseSet<ValuePair> &PairableInstUsers) { 1036 DenseSet<Value *> IsInPair; 1037 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1038 E = CandidatePairs.end(); C != E; ++C) { 1039 IsInPair.insert(C->first); 1040 IsInPair.insert(C->second); 1041 } 1042 1043 // Iterate through the basic block, recording all Users of each 1044 // pairable instruction. 1045 1046 BasicBlock::iterator E = BB.end(); 1047 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1048 if (IsInPair.find(I) == IsInPair.end()) continue; 1049 1050 DenseSet<Value *> Users; 1051 AliasSetTracker WriteSet(*AA); 1052 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1053 (void) trackUsesOfI(Users, WriteSet, I, J); 1054 1055 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1056 U != E; ++U) 1057 PairableInstUsers.insert(ValuePair(I, *U)); 1058 } 1059 } 1060 1061 // Returns true if an input to pair P is an output of pair Q and also an 1062 // input of pair Q is an output of pair P. If this is the case, then these 1063 // two pairs cannot be simultaneously fused. 1064 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1065 DenseSet<ValuePair> &PairableInstUsers, 1066 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1067 // Two pairs are in conflict if they are mutual Users of eachother. 1068 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1069 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1070 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1071 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1072 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1073 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1074 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1075 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1076 if (PairableInstUserMap) { 1077 // FIXME: The expensive part of the cycle check is not so much the cycle 1078 // check itself but this edge insertion procedure. This needs some 1079 // profiling and probably a different data structure (same is true of 1080 // most uses of std::multimap). 1081 if (PUsesQ) { 1082 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1083 if (!isSecondInIteratorPair(P, QPairRange)) 1084 PairableInstUserMap->insert(VPPair(Q, P)); 1085 } 1086 if (QUsesP) { 1087 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1088 if (!isSecondInIteratorPair(Q, PPairRange)) 1089 PairableInstUserMap->insert(VPPair(P, Q)); 1090 } 1091 } 1092 1093 return (QUsesP && PUsesQ); 1094 } 1095 1096 // This function walks the use graph of current pairs to see if, starting 1097 // from P, the walk returns to P. 1098 bool BBVectorize::pairWillFormCycle(ValuePair P, 1099 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1100 DenseSet<ValuePair> &CurrentPairs) { 1101 DEBUG(if (DebugCycleCheck) 1102 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1103 << *P.second << "\n"); 1104 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1105 // contains non-direct associations. 1106 DenseSet<ValuePair> Visited; 1107 SmallVector<ValuePair, 32> Q; 1108 // General depth-first post-order traversal: 1109 Q.push_back(P); 1110 do { 1111 ValuePair QTop = Q.pop_back_val(); 1112 Visited.insert(QTop); 1113 1114 DEBUG(if (DebugCycleCheck) 1115 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1116 << *QTop.second << "\n"); 1117 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1118 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1119 C != QPairRange.second; ++C) { 1120 if (C->second == P) { 1121 DEBUG(dbgs() 1122 << "BBV: rejected to prevent non-trivial cycle formation: " 1123 << *C->first.first << " <-> " << *C->first.second << "\n"); 1124 return true; 1125 } 1126 1127 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1128 Q.push_back(C->second); 1129 } 1130 } while (!Q.empty()); 1131 1132 return false; 1133 } 1134 1135 // This function builds the initial tree of connected pairs with the 1136 // pair J at the root. 1137 void BBVectorize::buildInitialTreeFor( 1138 std::multimap<Value *, Value *> &CandidatePairs, 1139 std::vector<Value *> &PairableInsts, 1140 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1141 DenseSet<ValuePair> &PairableInstUsers, 1142 DenseMap<Value *, Value *> &ChosenPairs, 1143 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1144 // Each of these pairs is viewed as the root node of a Tree. The Tree 1145 // is then walked (depth-first). As this happens, we keep track of 1146 // the pairs that compose the Tree and the maximum depth of the Tree. 1147 SmallVector<ValuePairWithDepth, 32> Q; 1148 // General depth-first post-order traversal: 1149 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1150 do { 1151 ValuePairWithDepth QTop = Q.back(); 1152 1153 // Push each child onto the queue: 1154 bool MoreChildren = false; 1155 size_t MaxChildDepth = QTop.second; 1156 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1157 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1158 k != qtRange.second; ++k) { 1159 // Make sure that this child pair is still a candidate: 1160 bool IsStillCand = false; 1161 VPIteratorPair checkRange = 1162 CandidatePairs.equal_range(k->second.first); 1163 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1164 m != checkRange.second; ++m) { 1165 if (m->second == k->second.second) { 1166 IsStillCand = true; 1167 break; 1168 } 1169 } 1170 1171 if (IsStillCand) { 1172 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1173 if (C == Tree.end()) { 1174 size_t d = getDepthFactor(k->second.first); 1175 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1176 MoreChildren = true; 1177 } else { 1178 MaxChildDepth = std::max(MaxChildDepth, C->second); 1179 } 1180 } 1181 } 1182 1183 if (!MoreChildren) { 1184 // Record the current pair as part of the Tree: 1185 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1186 Q.pop_back(); 1187 } 1188 } while (!Q.empty()); 1189 } 1190 1191 // Given some initial tree, prune it by removing conflicting pairs (pairs 1192 // that cannot be simultaneously chosen for vectorization). 1193 void BBVectorize::pruneTreeFor( 1194 std::multimap<Value *, Value *> &CandidatePairs, 1195 std::vector<Value *> &PairableInsts, 1196 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1197 DenseSet<ValuePair> &PairableInstUsers, 1198 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1199 DenseMap<Value *, Value *> &ChosenPairs, 1200 DenseMap<ValuePair, size_t> &Tree, 1201 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1202 bool UseCycleCheck) { 1203 SmallVector<ValuePairWithDepth, 32> Q; 1204 // General depth-first post-order traversal: 1205 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1206 do { 1207 ValuePairWithDepth QTop = Q.pop_back_val(); 1208 PrunedTree.insert(QTop.first); 1209 1210 // Visit each child, pruning as necessary... 1211 DenseMap<ValuePair, size_t> BestChildren; 1212 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1213 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1214 K != QTopRange.second; ++K) { 1215 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1216 if (C == Tree.end()) continue; 1217 1218 // This child is in the Tree, now we need to make sure it is the 1219 // best of any conflicting children. There could be multiple 1220 // conflicting children, so first, determine if we're keeping 1221 // this child, then delete conflicting children as necessary. 1222 1223 // It is also necessary to guard against pairing-induced 1224 // dependencies. Consider instructions a .. x .. y .. b 1225 // such that (a,b) are to be fused and (x,y) are to be fused 1226 // but a is an input to x and b is an output from y. This 1227 // means that y cannot be moved after b but x must be moved 1228 // after b for (a,b) to be fused. In other words, after 1229 // fusing (a,b) we have y .. a/b .. x where y is an input 1230 // to a/b and x is an output to a/b: x and y can no longer 1231 // be legally fused. To prevent this condition, we must 1232 // make sure that a child pair added to the Tree is not 1233 // both an input and output of an already-selected pair. 1234 1235 // Pairing-induced dependencies can also form from more complicated 1236 // cycles. The pair vs. pair conflicts are easy to check, and so 1237 // that is done explicitly for "fast rejection", and because for 1238 // child vs. child conflicts, we may prefer to keep the current 1239 // pair in preference to the already-selected child. 1240 DenseSet<ValuePair> CurrentPairs; 1241 1242 bool CanAdd = true; 1243 for (DenseMap<ValuePair, size_t>::iterator C2 1244 = BestChildren.begin(), E2 = BestChildren.end(); 1245 C2 != E2; ++C2) { 1246 if (C2->first.first == C->first.first || 1247 C2->first.first == C->first.second || 1248 C2->first.second == C->first.first || 1249 C2->first.second == C->first.second || 1250 pairsConflict(C2->first, C->first, PairableInstUsers, 1251 UseCycleCheck ? &PairableInstUserMap : 0)) { 1252 if (C2->second >= C->second) { 1253 CanAdd = false; 1254 break; 1255 } 1256 1257 CurrentPairs.insert(C2->first); 1258 } 1259 } 1260 if (!CanAdd) continue; 1261 1262 // Even worse, this child could conflict with another node already 1263 // selected for the Tree. If that is the case, ignore this child. 1264 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1265 E2 = PrunedTree.end(); T != E2; ++T) { 1266 if (T->first == C->first.first || 1267 T->first == C->first.second || 1268 T->second == C->first.first || 1269 T->second == C->first.second || 1270 pairsConflict(*T, C->first, PairableInstUsers, 1271 UseCycleCheck ? &PairableInstUserMap : 0)) { 1272 CanAdd = false; 1273 break; 1274 } 1275 1276 CurrentPairs.insert(*T); 1277 } 1278 if (!CanAdd) continue; 1279 1280 // And check the queue too... 1281 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1282 E2 = Q.end(); C2 != E2; ++C2) { 1283 if (C2->first.first == C->first.first || 1284 C2->first.first == C->first.second || 1285 C2->first.second == C->first.first || 1286 C2->first.second == C->first.second || 1287 pairsConflict(C2->first, C->first, PairableInstUsers, 1288 UseCycleCheck ? &PairableInstUserMap : 0)) { 1289 CanAdd = false; 1290 break; 1291 } 1292 1293 CurrentPairs.insert(C2->first); 1294 } 1295 if (!CanAdd) continue; 1296 1297 // Last but not least, check for a conflict with any of the 1298 // already-chosen pairs. 1299 for (DenseMap<Value *, Value *>::iterator C2 = 1300 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1301 C2 != E2; ++C2) { 1302 if (pairsConflict(*C2, C->first, PairableInstUsers, 1303 UseCycleCheck ? &PairableInstUserMap : 0)) { 1304 CanAdd = false; 1305 break; 1306 } 1307 1308 CurrentPairs.insert(*C2); 1309 } 1310 if (!CanAdd) continue; 1311 1312 // To check for non-trivial cycles formed by the addition of the 1313 // current pair we've formed a list of all relevant pairs, now use a 1314 // graph walk to check for a cycle. We start from the current pair and 1315 // walk the use tree to see if we again reach the current pair. If we 1316 // do, then the current pair is rejected. 1317 1318 // FIXME: It may be more efficient to use a topological-ordering 1319 // algorithm to improve the cycle check. This should be investigated. 1320 if (UseCycleCheck && 1321 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1322 continue; 1323 1324 // This child can be added, but we may have chosen it in preference 1325 // to an already-selected child. Check for this here, and if a 1326 // conflict is found, then remove the previously-selected child 1327 // before adding this one in its place. 1328 for (DenseMap<ValuePair, size_t>::iterator C2 1329 = BestChildren.begin(); C2 != BestChildren.end();) { 1330 if (C2->first.first == C->first.first || 1331 C2->first.first == C->first.second || 1332 C2->first.second == C->first.first || 1333 C2->first.second == C->first.second || 1334 pairsConflict(C2->first, C->first, PairableInstUsers)) 1335 BestChildren.erase(C2++); 1336 else 1337 ++C2; 1338 } 1339 1340 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1341 } 1342 1343 for (DenseMap<ValuePair, size_t>::iterator C 1344 = BestChildren.begin(), E2 = BestChildren.end(); 1345 C != E2; ++C) { 1346 size_t DepthF = getDepthFactor(C->first.first); 1347 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1348 } 1349 } while (!Q.empty()); 1350 } 1351 1352 // This function finds the best tree of mututally-compatible connected 1353 // pairs, given the choice of root pairs as an iterator range. 1354 void BBVectorize::findBestTreeFor( 1355 std::multimap<Value *, Value *> &CandidatePairs, 1356 std::vector<Value *> &PairableInsts, 1357 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1358 DenseSet<ValuePair> &PairableInstUsers, 1359 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1360 DenseMap<Value *, Value *> &ChosenPairs, 1361 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1362 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1363 bool UseCycleCheck) { 1364 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1365 J != ChoiceRange.second; ++J) { 1366 1367 // Before going any further, make sure that this pair does not 1368 // conflict with any already-selected pairs (see comment below 1369 // near the Tree pruning for more details). 1370 DenseSet<ValuePair> ChosenPairSet; 1371 bool DoesConflict = false; 1372 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1373 E = ChosenPairs.end(); C != E; ++C) { 1374 if (pairsConflict(*C, *J, PairableInstUsers, 1375 UseCycleCheck ? &PairableInstUserMap : 0)) { 1376 DoesConflict = true; 1377 break; 1378 } 1379 1380 ChosenPairSet.insert(*C); 1381 } 1382 if (DoesConflict) continue; 1383 1384 if (UseCycleCheck && 1385 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1386 continue; 1387 1388 DenseMap<ValuePair, size_t> Tree; 1389 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1390 PairableInstUsers, ChosenPairs, Tree, *J); 1391 1392 // Because we'll keep the child with the largest depth, the largest 1393 // depth is still the same in the unpruned Tree. 1394 size_t MaxDepth = Tree.lookup(*J); 1395 1396 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1397 << *J->first << " <-> " << *J->second << "} of depth " << 1398 MaxDepth << " and size " << Tree.size() << "\n"); 1399 1400 // At this point the Tree has been constructed, but, may contain 1401 // contradictory children (meaning that different children of 1402 // some tree node may be attempting to fuse the same instruction). 1403 // So now we walk the tree again, in the case of a conflict, 1404 // keep only the child with the largest depth. To break a tie, 1405 // favor the first child. 1406 1407 DenseSet<ValuePair> PrunedTree; 1408 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1409 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1410 PrunedTree, *J, UseCycleCheck); 1411 1412 size_t EffSize = 0; 1413 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1414 E = PrunedTree.end(); S != E; ++S) 1415 EffSize += getDepthFactor(S->first); 1416 1417 DEBUG(if (DebugPairSelection) 1418 dbgs() << "BBV: found pruned Tree for pair {" 1419 << *J->first << " <-> " << *J->second << "} of depth " << 1420 MaxDepth << " and size " << PrunedTree.size() << 1421 " (effective size: " << EffSize << ")\n"); 1422 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) { 1423 BestMaxDepth = MaxDepth; 1424 BestEffSize = EffSize; 1425 BestTree = PrunedTree; 1426 } 1427 } 1428 } 1429 1430 // Given the list of candidate pairs, this function selects those 1431 // that will be fused into vector instructions. 1432 void BBVectorize::choosePairs( 1433 std::multimap<Value *, Value *> &CandidatePairs, 1434 std::vector<Value *> &PairableInsts, 1435 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1436 DenseSet<ValuePair> &PairableInstUsers, 1437 DenseMap<Value *, Value *>& ChosenPairs) { 1438 bool UseCycleCheck = 1439 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1440 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1441 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1442 E = PairableInsts.end(); I != E; ++I) { 1443 // The number of possible pairings for this variable: 1444 size_t NumChoices = CandidatePairs.count(*I); 1445 if (!NumChoices) continue; 1446 1447 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1448 1449 // The best pair to choose and its tree: 1450 size_t BestMaxDepth = 0, BestEffSize = 0; 1451 DenseSet<ValuePair> BestTree; 1452 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1453 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1454 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1455 UseCycleCheck); 1456 1457 // A tree has been chosen (or not) at this point. If no tree was 1458 // chosen, then this instruction, I, cannot be paired (and is no longer 1459 // considered). 1460 1461 DEBUG(if (BestTree.size() > 0) 1462 dbgs() << "BBV: selected pairs in the best tree for: " 1463 << *cast<Instruction>(*I) << "\n"); 1464 1465 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1466 SE2 = BestTree.end(); S != SE2; ++S) { 1467 // Insert the members of this tree into the list of chosen pairs. 1468 ChosenPairs.insert(ValuePair(S->first, S->second)); 1469 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1470 *S->second << "\n"); 1471 1472 // Remove all candidate pairs that have values in the chosen tree. 1473 for (std::multimap<Value *, Value *>::iterator K = 1474 CandidatePairs.begin(); K != CandidatePairs.end();) { 1475 if (K->first == S->first || K->second == S->first || 1476 K->second == S->second || K->first == S->second) { 1477 // Don't remove the actual pair chosen so that it can be used 1478 // in subsequent tree selections. 1479 if (!(K->first == S->first && K->second == S->second)) 1480 CandidatePairs.erase(K++); 1481 else 1482 ++K; 1483 } else { 1484 ++K; 1485 } 1486 } 1487 } 1488 } 1489 1490 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1491 } 1492 1493 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1494 unsigned n = 0) { 1495 if (!I->hasName()) 1496 return ""; 1497 1498 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1499 (n > 0 ? "." + utostr(n) : "")).str(); 1500 } 1501 1502 // Returns the value that is to be used as the pointer input to the vector 1503 // instruction that fuses I with J. 1504 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1505 Instruction *I, Instruction *J, unsigned o, 1506 bool FlipMemInputs) { 1507 Value *IPtr, *JPtr; 1508 unsigned IAlignment, JAlignment; 1509 int64_t OffsetInElmts; 1510 1511 // Note: the analysis might fail here, that is why FlipMemInputs has 1512 // been precomputed (OffsetInElmts must be unused here). 1513 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1514 OffsetInElmts); 1515 1516 // The pointer value is taken to be the one with the lowest offset. 1517 Value *VPtr; 1518 if (!FlipMemInputs) { 1519 VPtr = IPtr; 1520 } else { 1521 VPtr = JPtr; 1522 } 1523 1524 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 1525 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 1526 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1527 Type *VArgPtrType = PointerType::get(VArgType, 1528 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1529 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1530 /* insert before */ FlipMemInputs ? J : I); 1531 } 1532 1533 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1534 unsigned MaskOffset, unsigned NumInElem, 1535 unsigned NumInElem1, unsigned IdxOffset, 1536 std::vector<Constant*> &Mask) { 1537 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 1538 for (unsigned v = 0; v < NumElem1; ++v) { 1539 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1540 if (m < 0) { 1541 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1542 } else { 1543 unsigned mm = m + (int) IdxOffset; 1544 if (m >= (int) NumInElem1) 1545 mm += (int) NumInElem; 1546 1547 Mask[v+MaskOffset] = 1548 ConstantInt::get(Type::getInt32Ty(Context), mm); 1549 } 1550 } 1551 } 1552 1553 // Returns the value that is to be used as the vector-shuffle mask to the 1554 // vector instruction that fuses I with J. 1555 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1556 Instruction *I, Instruction *J) { 1557 // This is the shuffle mask. We need to append the second 1558 // mask to the first, and the numbers need to be adjusted. 1559 1560 Type *ArgTypeI = I->getType(); 1561 Type *ArgTypeJ = J->getType(); 1562 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1563 1564 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 1565 1566 // Get the total number of elements in the fused vector type. 1567 // By definition, this must equal the number of elements in 1568 // the final mask. 1569 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1570 std::vector<Constant*> Mask(NumElem); 1571 1572 Type *OpTypeI = I->getOperand(0)->getType(); 1573 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 1574 Type *OpTypeJ = J->getOperand(0)->getType(); 1575 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 1576 1577 // The fused vector will be: 1578 // ----------------------------------------------------- 1579 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 1580 // ----------------------------------------------------- 1581 // from which we'll extract NumElem total elements (where the first NumElemI 1582 // of them come from the mask in I and the remainder come from the mask 1583 // in J. 1584 1585 // For the mask from the first pair... 1586 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 1587 0, Mask); 1588 1589 // For the mask from the second pair... 1590 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 1591 NumInElemI, Mask); 1592 1593 return ConstantVector::get(Mask); 1594 } 1595 1596 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 1597 Instruction *J, unsigned o, Value *&LOp, 1598 unsigned numElemL, 1599 Type *ArgTypeL, Type *ArgTypeH, 1600 unsigned IdxOff) { 1601 bool ExpandedIEChain = false; 1602 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 1603 // If we have a pure insertelement chain, then this can be rewritten 1604 // into a chain that directly builds the larger type. 1605 bool PureChain = true; 1606 InsertElementInst *LIENext = LIE; 1607 do { 1608 if (!isa<UndefValue>(LIENext->getOperand(0)) && 1609 !isa<InsertElementInst>(LIENext->getOperand(0))) { 1610 PureChain = false; 1611 break; 1612 } 1613 } while ((LIENext = 1614 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1615 1616 if (PureChain) { 1617 SmallVector<Value *, 8> VectElemts(numElemL, 1618 UndefValue::get(ArgTypeL->getScalarType())); 1619 InsertElementInst *LIENext = LIE; 1620 do { 1621 unsigned Idx = 1622 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 1623 VectElemts[Idx] = LIENext->getOperand(1); 1624 } while ((LIENext = 1625 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1626 1627 LIENext = 0; 1628 Value *LIEPrev = UndefValue::get(ArgTypeH); 1629 for (unsigned i = 0; i < numElemL; ++i) { 1630 if (isa<UndefValue>(VectElemts[i])) continue; 1631 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 1632 ConstantInt::get(Type::getInt32Ty(Context), 1633 i + IdxOff), 1634 getReplacementName(I, true, o, i+1)); 1635 LIENext->insertBefore(J); 1636 LIEPrev = LIENext; 1637 } 1638 1639 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 1640 ExpandedIEChain = true; 1641 } 1642 } 1643 1644 return ExpandedIEChain; 1645 } 1646 1647 // Returns the value to be used as the specified operand of the vector 1648 // instruction that fuses I with J. 1649 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1650 Instruction *J, unsigned o, bool FlipMemInputs) { 1651 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1652 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1653 1654 // Compute the fused vector type for this operand 1655 Type *ArgTypeI = I->getOperand(o)->getType(); 1656 Type *ArgTypeJ = J->getOperand(o)->getType(); 1657 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1658 1659 Instruction *L = I, *H = J; 1660 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 1661 if (FlipMemInputs) { 1662 L = J; 1663 H = I; 1664 ArgTypeL = ArgTypeJ; 1665 ArgTypeH = ArgTypeI; 1666 } 1667 1668 unsigned numElemL; 1669 if (ArgTypeL->isVectorTy()) 1670 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 1671 else 1672 numElemL = 1; 1673 1674 unsigned numElemH; 1675 if (ArgTypeH->isVectorTy()) 1676 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 1677 else 1678 numElemH = 1; 1679 1680 Value *LOp = L->getOperand(o); 1681 Value *HOp = H->getOperand(o); 1682 unsigned numElem = VArgType->getNumElements(); 1683 1684 // First, we check if we can reuse the "original" vector outputs (if these 1685 // exist). We might need a shuffle. 1686 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 1687 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 1688 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 1689 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 1690 1691 // FIXME: If we're fusing shuffle instructions, then we can't apply this 1692 // optimization. The input vectors to the shuffle might be a different 1693 // length from the shuffle outputs. Unfortunately, the replacement 1694 // shuffle mask has already been formed, and the mask entries are sensitive 1695 // to the sizes of the inputs. 1696 bool IsSizeChangeShuffle = 1697 isa<ShuffleVectorInst>(L) && 1698 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 1699 1700 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 1701 // We can have at most two unique vector inputs. 1702 bool CanUseInputs = true; 1703 Value *I1, *I2 = 0; 1704 if (LEE) { 1705 I1 = LEE->getOperand(0); 1706 } else { 1707 I1 = LSV->getOperand(0); 1708 I2 = LSV->getOperand(1); 1709 if (I2 == I1 || isa<UndefValue>(I2)) 1710 I2 = 0; 1711 } 1712 1713 if (HEE) { 1714 Value *I3 = HEE->getOperand(0); 1715 if (!I2 && I3 != I1) 1716 I2 = I3; 1717 else if (I3 != I1 && I3 != I2) 1718 CanUseInputs = false; 1719 } else { 1720 Value *I3 = HSV->getOperand(0); 1721 if (!I2 && I3 != I1) 1722 I2 = I3; 1723 else if (I3 != I1 && I3 != I2) 1724 CanUseInputs = false; 1725 1726 if (CanUseInputs) { 1727 Value *I4 = HSV->getOperand(1); 1728 if (!isa<UndefValue>(I4)) { 1729 if (!I2 && I4 != I1) 1730 I2 = I4; 1731 else if (I4 != I1 && I4 != I2) 1732 CanUseInputs = false; 1733 } 1734 } 1735 } 1736 1737 if (CanUseInputs) { 1738 unsigned LOpElem = 1739 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 1740 ->getNumElements(); 1741 unsigned HOpElem = 1742 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 1743 ->getNumElements(); 1744 1745 // We have one or two input vectors. We need to map each index of the 1746 // operands to the index of the original vector. 1747 SmallVector<std::pair<int, int>, 8> II(numElem); 1748 for (unsigned i = 0; i < numElemL; ++i) { 1749 int Idx, INum; 1750 if (LEE) { 1751 Idx = 1752 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 1753 INum = LEE->getOperand(0) == I1 ? 0 : 1; 1754 } else { 1755 Idx = LSV->getMaskValue(i); 1756 if (Idx < (int) LOpElem) { 1757 INum = LSV->getOperand(0) == I1 ? 0 : 1; 1758 } else { 1759 Idx -= LOpElem; 1760 INum = LSV->getOperand(1) == I1 ? 0 : 1; 1761 } 1762 } 1763 1764 II[i] = std::pair<int, int>(Idx, INum); 1765 } 1766 for (unsigned i = 0; i < numElemH; ++i) { 1767 int Idx, INum; 1768 if (HEE) { 1769 Idx = 1770 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 1771 INum = HEE->getOperand(0) == I1 ? 0 : 1; 1772 } else { 1773 Idx = HSV->getMaskValue(i); 1774 if (Idx < (int) HOpElem) { 1775 INum = HSV->getOperand(0) == I1 ? 0 : 1; 1776 } else { 1777 Idx -= HOpElem; 1778 INum = HSV->getOperand(1) == I1 ? 0 : 1; 1779 } 1780 } 1781 1782 II[i + numElemL] = std::pair<int, int>(Idx, INum); 1783 } 1784 1785 // We now have an array which tells us from which index of which 1786 // input vector each element of the operand comes. 1787 VectorType *I1T = cast<VectorType>(I1->getType()); 1788 unsigned I1Elem = I1T->getNumElements(); 1789 1790 if (!I2) { 1791 // In this case there is only one underlying vector input. Check for 1792 // the trivial case where we can use the input directly. 1793 if (I1Elem == numElem) { 1794 bool ElemInOrder = true; 1795 for (unsigned i = 0; i < numElem; ++i) { 1796 if (II[i].first != (int) i && II[i].first != -1) { 1797 ElemInOrder = false; 1798 break; 1799 } 1800 } 1801 1802 if (ElemInOrder) 1803 return I1; 1804 } 1805 1806 // A shuffle is needed. 1807 std::vector<Constant *> Mask(numElem); 1808 for (unsigned i = 0; i < numElem; ++i) { 1809 int Idx = II[i].first; 1810 if (Idx == -1) 1811 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 1812 else 1813 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 1814 } 1815 1816 Instruction *S = 1817 new ShuffleVectorInst(I1, UndefValue::get(I1T), 1818 ConstantVector::get(Mask), 1819 getReplacementName(I, true, o)); 1820 S->insertBefore(J); 1821 return S; 1822 } 1823 1824 VectorType *I2T = cast<VectorType>(I2->getType()); 1825 unsigned I2Elem = I2T->getNumElements(); 1826 1827 // This input comes from two distinct vectors. The first step is to 1828 // make sure that both vectors are the same length. If not, the 1829 // smaller one will need to grow before they can be shuffled together. 1830 if (I1Elem < I2Elem) { 1831 std::vector<Constant *> Mask(I2Elem); 1832 unsigned v = 0; 1833 for (; v < I1Elem; ++v) 1834 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1835 for (; v < I2Elem; ++v) 1836 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1837 1838 Instruction *NewI1 = 1839 new ShuffleVectorInst(I1, UndefValue::get(I1T), 1840 ConstantVector::get(Mask), 1841 getReplacementName(I, true, o, 1)); 1842 NewI1->insertBefore(J); 1843 I1 = NewI1; 1844 I1T = I2T; 1845 I1Elem = I2Elem; 1846 } else if (I1Elem > I2Elem) { 1847 std::vector<Constant *> Mask(I1Elem); 1848 unsigned v = 0; 1849 for (; v < I2Elem; ++v) 1850 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1851 for (; v < I1Elem; ++v) 1852 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1853 1854 Instruction *NewI2 = 1855 new ShuffleVectorInst(I2, UndefValue::get(I2T), 1856 ConstantVector::get(Mask), 1857 getReplacementName(I, true, o, 1)); 1858 NewI2->insertBefore(J); 1859 I2 = NewI2; 1860 I2T = I1T; 1861 I2Elem = I1Elem; 1862 } 1863 1864 // Now that both I1 and I2 are the same length we can shuffle them 1865 // together (and use the result). 1866 std::vector<Constant *> Mask(numElem); 1867 for (unsigned v = 0; v < numElem; ++v) { 1868 if (II[v].first == -1) { 1869 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1870 } else { 1871 int Idx = II[v].first + II[v].second * I1Elem; 1872 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 1873 } 1874 } 1875 1876 Instruction *NewOp = 1877 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 1878 getReplacementName(I, true, o)); 1879 NewOp->insertBefore(J); 1880 return NewOp; 1881 } 1882 } 1883 1884 Type *ArgType = ArgTypeL; 1885 if (numElemL < numElemH) { 1886 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 1887 ArgTypeL, VArgType, 1)) { 1888 // This is another short-circuit case: we're combining a scalar into 1889 // a vector that is formed by an IE chain. We've just expanded the IE 1890 // chain, now insert the scalar and we're done. 1891 1892 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 1893 getReplacementName(I, true, o)); 1894 S->insertBefore(J); 1895 return S; 1896 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 1897 ArgTypeH)) { 1898 // The two vector inputs to the shuffle must be the same length, 1899 // so extend the smaller vector to be the same length as the larger one. 1900 Instruction *NLOp; 1901 if (numElemL > 1) { 1902 1903 std::vector<Constant *> Mask(numElemH); 1904 unsigned v = 0; 1905 for (; v < numElemL; ++v) 1906 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1907 for (; v < numElemH; ++v) 1908 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1909 1910 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 1911 ConstantVector::get(Mask), 1912 getReplacementName(I, true, o, 1)); 1913 } else { 1914 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 1915 getReplacementName(I, true, o, 1)); 1916 } 1917 1918 NLOp->insertBefore(J); 1919 LOp = NLOp; 1920 } 1921 1922 ArgType = ArgTypeH; 1923 } else if (numElemL > numElemH) { 1924 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 1925 ArgTypeH, VArgType)) { 1926 Instruction *S = 1927 InsertElementInst::Create(LOp, HOp, 1928 ConstantInt::get(Type::getInt32Ty(Context), 1929 numElemL), 1930 getReplacementName(I, true, o)); 1931 S->insertBefore(J); 1932 return S; 1933 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 1934 ArgTypeL)) { 1935 Instruction *NHOp; 1936 if (numElemH > 1) { 1937 std::vector<Constant *> Mask(numElemL); 1938 unsigned v = 0; 1939 for (; v < numElemH; ++v) 1940 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1941 for (; v < numElemL; ++v) 1942 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 1943 1944 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 1945 ConstantVector::get(Mask), 1946 getReplacementName(I, true, o, 1)); 1947 } else { 1948 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 1949 getReplacementName(I, true, o, 1)); 1950 } 1951 1952 NHOp->insertBefore(J); 1953 HOp = NHOp; 1954 } 1955 } 1956 1957 if (ArgType->isVectorTy()) { 1958 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1959 std::vector<Constant*> Mask(numElem); 1960 for (unsigned v = 0; v < numElem; ++v) { 1961 unsigned Idx = v; 1962 // If the low vector was expanded, we need to skip the extra 1963 // undefined entries. 1964 if (v >= numElemL && numElemH > numElemL) 1965 Idx += (numElemH - numElemL); 1966 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 1967 } 1968 1969 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 1970 ConstantVector::get(Mask), 1971 getReplacementName(I, true, o)); 1972 BV->insertBefore(J); 1973 return BV; 1974 } 1975 1976 Instruction *BV1 = InsertElementInst::Create( 1977 UndefValue::get(VArgType), LOp, CV0, 1978 getReplacementName(I, true, o, 1)); 1979 BV1->insertBefore(I); 1980 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 1981 getReplacementName(I, true, o, 2)); 1982 BV2->insertBefore(J); 1983 return BV2; 1984 } 1985 1986 // This function creates an array of values that will be used as the inputs 1987 // to the vector instruction that fuses I with J. 1988 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1989 Instruction *I, Instruction *J, 1990 SmallVector<Value *, 3> &ReplacedOperands, 1991 bool FlipMemInputs) { 1992 unsigned NumOperands = I->getNumOperands(); 1993 1994 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1995 // Iterate backward so that we look at the store pointer 1996 // first and know whether or not we need to flip the inputs. 1997 1998 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1999 // This is the pointer for a load/store instruction. 2000 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 2001 FlipMemInputs); 2002 continue; 2003 } else if (isa<CallInst>(I)) { 2004 Function *F = cast<CallInst>(I)->getCalledFunction(); 2005 unsigned IID = F->getIntrinsicID(); 2006 if (o == NumOperands-1) { 2007 BasicBlock &BB = *I->getParent(); 2008 2009 Module *M = BB.getParent()->getParent(); 2010 Type *ArgTypeI = I->getType(); 2011 Type *ArgTypeJ = J->getType(); 2012 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2013 2014 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 2015 (Intrinsic::ID) IID, VArgType); 2016 continue; 2017 } else if (IID == Intrinsic::powi && o == 1) { 2018 // The second argument of powi is a single integer and we've already 2019 // checked that both arguments are equal. As a result, we just keep 2020 // I's second argument. 2021 ReplacedOperands[o] = I->getOperand(o); 2022 continue; 2023 } 2024 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2025 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2026 continue; 2027 } 2028 2029 ReplacedOperands[o] = 2030 getReplacementInput(Context, I, J, o, FlipMemInputs); 2031 } 2032 } 2033 2034 // This function creates two values that represent the outputs of the 2035 // original I and J instructions. These are generally vector shuffles 2036 // or extracts. In many cases, these will end up being unused and, thus, 2037 // eliminated by later passes. 2038 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2039 Instruction *J, Instruction *K, 2040 Instruction *&InsertionPt, 2041 Instruction *&K1, Instruction *&K2, 2042 bool FlipMemInputs) { 2043 if (isa<StoreInst>(I)) { 2044 AA->replaceWithNewValue(I, K); 2045 AA->replaceWithNewValue(J, K); 2046 } else { 2047 Type *IType = I->getType(); 2048 Type *JType = J->getType(); 2049 2050 VectorType *VType = getVecTypeForPair(IType, JType); 2051 unsigned numElem = VType->getNumElements(); 2052 2053 unsigned numElemI, numElemJ; 2054 if (IType->isVectorTy()) 2055 numElemI = cast<VectorType>(IType)->getNumElements(); 2056 else 2057 numElemI = 1; 2058 2059 if (JType->isVectorTy()) 2060 numElemJ = cast<VectorType>(JType)->getNumElements(); 2061 else 2062 numElemJ = 1; 2063 2064 if (IType->isVectorTy()) { 2065 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2066 for (unsigned v = 0; v < numElemI; ++v) { 2067 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2068 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2069 } 2070 2071 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2072 ConstantVector::get( 2073 FlipMemInputs ? Mask2 : Mask1), 2074 getReplacementName(K, false, 1)); 2075 } else { 2076 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2077 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2078 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 2079 getReplacementName(K, false, 1)); 2080 } 2081 2082 if (JType->isVectorTy()) { 2083 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2084 for (unsigned v = 0; v < numElemJ; ++v) { 2085 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2086 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2087 } 2088 2089 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2090 ConstantVector::get( 2091 FlipMemInputs ? Mask1 : Mask2), 2092 getReplacementName(K, false, 2)); 2093 } else { 2094 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2095 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2096 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 2097 getReplacementName(K, false, 2)); 2098 } 2099 2100 K1->insertAfter(K); 2101 K2->insertAfter(K1); 2102 InsertionPt = K2; 2103 } 2104 } 2105 2106 // Move all uses of the function I (including pairing-induced uses) after J. 2107 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2108 std::multimap<Value *, Value *> &LoadMoveSet, 2109 Instruction *I, Instruction *J) { 2110 // Skip to the first instruction past I. 2111 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2112 2113 DenseSet<Value *> Users; 2114 AliasSetTracker WriteSet(*AA); 2115 for (; cast<Instruction>(L) != J; ++L) 2116 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 2117 2118 assert(cast<Instruction>(L) == J && 2119 "Tracking has not proceeded far enough to check for dependencies"); 2120 // If J is now in the use set of I, then trackUsesOfI will return true 2121 // and we have a dependency cycle (and the fusing operation must abort). 2122 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 2123 } 2124 2125 // Move all uses of the function I (including pairing-induced uses) after J. 2126 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2127 std::multimap<Value *, Value *> &LoadMoveSet, 2128 Instruction *&InsertionPt, 2129 Instruction *I, Instruction *J) { 2130 // Skip to the first instruction past I. 2131 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2132 2133 DenseSet<Value *> Users; 2134 AliasSetTracker WriteSet(*AA); 2135 for (; cast<Instruction>(L) != J;) { 2136 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 2137 // Move this instruction 2138 Instruction *InstToMove = L; ++L; 2139 2140 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2141 " to after " << *InsertionPt << "\n"); 2142 InstToMove->removeFromParent(); 2143 InstToMove->insertAfter(InsertionPt); 2144 InsertionPt = InstToMove; 2145 } else { 2146 ++L; 2147 } 2148 } 2149 } 2150 2151 // Collect all load instruction that are in the move set of a given first 2152 // pair member. These loads depend on the first instruction, I, and so need 2153 // to be moved after J (the second instruction) when the pair is fused. 2154 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2155 DenseMap<Value *, Value *> &ChosenPairs, 2156 std::multimap<Value *, Value *> &LoadMoveSet, 2157 Instruction *I) { 2158 // Skip to the first instruction past I. 2159 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2160 2161 DenseSet<Value *> Users; 2162 AliasSetTracker WriteSet(*AA); 2163 2164 // Note: We cannot end the loop when we reach J because J could be moved 2165 // farther down the use chain by another instruction pairing. Also, J 2166 // could be before I if this is an inverted input. 2167 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2168 if (trackUsesOfI(Users, WriteSet, I, L)) { 2169 if (L->mayReadFromMemory()) 2170 LoadMoveSet.insert(ValuePair(L, I)); 2171 } 2172 } 2173 } 2174 2175 // In cases where both load/stores and the computation of their pointers 2176 // are chosen for vectorization, we can end up in a situation where the 2177 // aliasing analysis starts returning different query results as the 2178 // process of fusing instruction pairs continues. Because the algorithm 2179 // relies on finding the same use trees here as were found earlier, we'll 2180 // need to precompute the necessary aliasing information here and then 2181 // manually update it during the fusion process. 2182 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2183 std::vector<Value *> &PairableInsts, 2184 DenseMap<Value *, Value *> &ChosenPairs, 2185 std::multimap<Value *, Value *> &LoadMoveSet) { 2186 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2187 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2188 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2189 if (P == ChosenPairs.end()) continue; 2190 2191 Instruction *I = cast<Instruction>(P->first); 2192 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 2193 } 2194 } 2195 2196 // As with the aliasing information, SCEV can also change because of 2197 // vectorization. This information is used to compute relative pointer 2198 // offsets; the necessary information will be cached here prior to 2199 // fusion. 2200 void BBVectorize::collectPtrInfo(std::vector<Value *> &PairableInsts, 2201 DenseMap<Value *, Value *> &ChosenPairs, 2202 DenseSet<Value *> &LowPtrInsts) { 2203 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2204 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2205 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2206 if (P == ChosenPairs.end()) continue; 2207 2208 Instruction *I = cast<Instruction>(P->first); 2209 Instruction *J = cast<Instruction>(P->second); 2210 2211 if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) 2212 continue; 2213 2214 Value *IPtr, *JPtr; 2215 unsigned IAlignment, JAlignment; 2216 int64_t OffsetInElmts; 2217 if (!getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2218 OffsetInElmts) || abs64(OffsetInElmts) != 1) 2219 llvm_unreachable("Pre-fusion pointer analysis failed"); 2220 2221 Value *LowPI = (OffsetInElmts > 0) ? I : J; 2222 LowPtrInsts.insert(LowPI); 2223 } 2224 } 2225 2226 // When the first instruction in each pair is cloned, it will inherit its 2227 // parent's metadata. This metadata must be combined with that of the other 2228 // instruction in a safe way. 2229 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2230 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2231 K->getAllMetadataOtherThanDebugLoc(Metadata); 2232 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2233 unsigned Kind = Metadata[i].first; 2234 MDNode *JMD = J->getMetadata(Kind); 2235 MDNode *KMD = Metadata[i].second; 2236 2237 switch (Kind) { 2238 default: 2239 K->setMetadata(Kind, 0); // Remove unknown metadata 2240 break; 2241 case LLVMContext::MD_tbaa: 2242 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2243 break; 2244 case LLVMContext::MD_fpmath: 2245 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2246 break; 2247 } 2248 } 2249 } 2250 2251 // This function fuses the chosen instruction pairs into vector instructions, 2252 // taking care preserve any needed scalar outputs and, then, it reorders the 2253 // remaining instructions as needed (users of the first member of the pair 2254 // need to be moved to after the location of the second member of the pair 2255 // because the vector instruction is inserted in the location of the pair's 2256 // second member). 2257 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2258 std::vector<Value *> &PairableInsts, 2259 DenseMap<Value *, Value *> &ChosenPairs) { 2260 LLVMContext& Context = BB.getContext(); 2261 2262 // During the vectorization process, the order of the pairs to be fused 2263 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2264 // list. After a pair is fused, the flipped pair is removed from the list. 2265 std::vector<ValuePair> FlippedPairs; 2266 FlippedPairs.reserve(ChosenPairs.size()); 2267 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2268 E = ChosenPairs.end(); P != E; ++P) 2269 FlippedPairs.push_back(ValuePair(P->second, P->first)); 2270 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 2271 E = FlippedPairs.end(); P != E; ++P) 2272 ChosenPairs.insert(*P); 2273 2274 std::multimap<Value *, Value *> LoadMoveSet; 2275 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 2276 2277 DenseSet<Value *> LowPtrInsts; 2278 collectPtrInfo(PairableInsts, ChosenPairs, LowPtrInsts); 2279 2280 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2281 2282 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2283 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2284 if (P == ChosenPairs.end()) { 2285 ++PI; 2286 continue; 2287 } 2288 2289 if (getDepthFactor(P->first) == 0) { 2290 // These instructions are not really fused, but are tracked as though 2291 // they are. Any case in which it would be interesting to fuse them 2292 // will be taken care of by InstCombine. 2293 --NumFusedOps; 2294 ++PI; 2295 continue; 2296 } 2297 2298 Instruction *I = cast<Instruction>(P->first), 2299 *J = cast<Instruction>(P->second); 2300 2301 DEBUG(dbgs() << "BBV: fusing: " << *I << 2302 " <-> " << *J << "\n"); 2303 2304 // Remove the pair and flipped pair from the list. 2305 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2306 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2307 ChosenPairs.erase(FP); 2308 ChosenPairs.erase(P); 2309 2310 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 2311 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2312 " <-> " << *J << 2313 " aborted because of non-trivial dependency cycle\n"); 2314 --NumFusedOps; 2315 ++PI; 2316 continue; 2317 } 2318 2319 bool FlipMemInputs = false; 2320 if (isa<LoadInst>(I) || isa<StoreInst>(I)) 2321 FlipMemInputs = (LowPtrInsts.find(I) == LowPtrInsts.end()); 2322 2323 unsigned NumOperands = I->getNumOperands(); 2324 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2325 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 2326 FlipMemInputs); 2327 2328 // Make a copy of the original operation, change its type to the vector 2329 // type and replace its operands with the vector operands. 2330 Instruction *K = I->clone(); 2331 if (I->hasName()) K->takeName(I); 2332 2333 if (!isa<StoreInst>(K)) 2334 K->mutateType(getVecTypeForPair(I->getType(), J->getType())); 2335 2336 combineMetadata(K, J); 2337 2338 for (unsigned o = 0; o < NumOperands; ++o) 2339 K->setOperand(o, ReplacedOperands[o]); 2340 2341 // If we've flipped the memory inputs, make sure that we take the correct 2342 // alignment. 2343 if (FlipMemInputs) { 2344 if (isa<StoreInst>(K)) 2345 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 2346 else 2347 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 2348 } 2349 2350 K->insertAfter(J); 2351 2352 // Instruction insertion point: 2353 Instruction *InsertionPt = K; 2354 Instruction *K1 = 0, *K2 = 0; 2355 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 2356 FlipMemInputs); 2357 2358 // The use tree of the first original instruction must be moved to after 2359 // the location of the second instruction. The entire use tree of the 2360 // first instruction is disjoint from the input tree of the second 2361 // (by definition), and so commutes with it. 2362 2363 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 2364 2365 if (!isa<StoreInst>(I)) { 2366 I->replaceAllUsesWith(K1); 2367 J->replaceAllUsesWith(K2); 2368 AA->replaceWithNewValue(I, K1); 2369 AA->replaceWithNewValue(J, K2); 2370 } 2371 2372 // Instructions that may read from memory may be in the load move set. 2373 // Once an instruction is fused, we no longer need its move set, and so 2374 // the values of the map never need to be updated. However, when a load 2375 // is fused, we need to merge the entries from both instructions in the 2376 // pair in case those instructions were in the move set of some other 2377 // yet-to-be-fused pair. The loads in question are the keys of the map. 2378 if (I->mayReadFromMemory()) { 2379 std::vector<ValuePair> NewSetMembers; 2380 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 2381 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 2382 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 2383 N != IPairRange.second; ++N) 2384 NewSetMembers.push_back(ValuePair(K, N->second)); 2385 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 2386 N != JPairRange.second; ++N) 2387 NewSetMembers.push_back(ValuePair(K, N->second)); 2388 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 2389 AE = NewSetMembers.end(); A != AE; ++A) 2390 LoadMoveSet.insert(*A); 2391 } 2392 2393 // Before removing I, set the iterator to the next instruction. 2394 PI = llvm::next(BasicBlock::iterator(I)); 2395 if (cast<Instruction>(PI) == J) 2396 ++PI; 2397 2398 SE->forgetValue(I); 2399 SE->forgetValue(J); 2400 I->eraseFromParent(); 2401 J->eraseFromParent(); 2402 } 2403 2404 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 2405 } 2406} 2407 2408char BBVectorize::ID = 0; 2409static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 2410INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2411INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2412INITIALIZE_PASS_DEPENDENCY(DominatorTree) 2413INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2414INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2415 2416BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 2417 return new BBVectorize(C); 2418} 2419 2420bool 2421llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 2422 BBVectorize BBVectorizer(P, C); 2423 return BBVectorizer.vectorizeBB(BB); 2424} 2425 2426//===----------------------------------------------------------------------===// 2427VectorizeConfig::VectorizeConfig() { 2428 VectorBits = ::VectorBits; 2429 VectorizeBools = !::NoBools; 2430 VectorizeInts = !::NoInts; 2431 VectorizeFloats = !::NoFloats; 2432 VectorizePointers = !::NoPointers; 2433 VectorizeCasts = !::NoCasts; 2434 VectorizeMath = !::NoMath; 2435 VectorizeFMA = !::NoFMA; 2436 VectorizeSelect = !::NoSelect; 2437 VectorizeCmp = !::NoCmp; 2438 VectorizeGEP = !::NoGEP; 2439 VectorizeMemOps = !::NoMemOps; 2440 AlignedOnly = ::AlignedOnly; 2441 ReqChainDepth= ::ReqChainDepth; 2442 SearchLimit = ::SearchLimit; 2443 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 2444 SplatBreaksChain = ::SplatBreaksChain; 2445 MaxInsts = ::MaxInsts; 2446 MaxIter = ::MaxIter; 2447 Pow2LenOnly = ::Pow2LenOnly; 2448 NoMemOpBoost = ::NoMemOpBoost; 2449 FastDep = ::FastDep; 2450} 2451