BBVectorize.cpp revision f3f5a1e6f77a842ccb24cc81766437da5197d712
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/Pass.h" 27#include "llvm/Type.h" 28#include "llvm/ADT/DenseMap.h" 29#include "llvm/ADT/DenseSet.h" 30#include "llvm/ADT/SmallVector.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/ADT/STLExtras.h" 33#include "llvm/ADT/StringExtras.h" 34#include "llvm/Analysis/AliasAnalysis.h" 35#include "llvm/Analysis/AliasSetTracker.h" 36#include "llvm/Analysis/ScalarEvolution.h" 37#include "llvm/Analysis/ScalarEvolutionExpressions.h" 38#include "llvm/Analysis/ValueTracking.h" 39#include "llvm/Support/CommandLine.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/raw_ostream.h" 42#include "llvm/Support/ValueHandle.h" 43#include "llvm/Target/TargetData.h" 44#include "llvm/Transforms/Vectorize.h" 45#include <algorithm> 46#include <map> 47using namespace llvm; 48 49static cl::opt<unsigned> 50ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 51 cl::desc("The required chain depth for vectorization")); 52 53static cl::opt<unsigned> 54SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 55 cl::desc("The maximum search distance for instruction pairs")); 56 57static cl::opt<bool> 58SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 59 cl::desc("Replicating one element to a pair breaks the chain")); 60 61static cl::opt<unsigned> 62VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 63 cl::desc("The size of the native vector registers")); 64 65static cl::opt<unsigned> 66MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 67 cl::desc("The maximum number of pairing iterations")); 68 69static cl::opt<unsigned> 70MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 71 cl::desc("The maximum number of pairable instructions per group")); 72 73static cl::opt<unsigned> 74MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 75 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 76 " a full cycle check")); 77 78static cl::opt<bool> 79NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 80 cl::desc("Don't try to vectorize integer values")); 81 82static cl::opt<bool> 83NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 84 cl::desc("Don't try to vectorize floating-point values")); 85 86static cl::opt<bool> 87NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden, 88 cl::desc("Don't try to vectorize pointer values")); 89 90static cl::opt<bool> 91NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 92 cl::desc("Don't try to vectorize casting (conversion) operations")); 93 94static cl::opt<bool> 95NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 96 cl::desc("Don't try to vectorize floating-point math intrinsics")); 97 98static cl::opt<bool> 99NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 100 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 101 102static cl::opt<bool> 103NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 104 cl::desc("Don't try to vectorize select instructions")); 105 106static cl::opt<bool> 107NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 108 cl::desc("Don't try to vectorize getelementptr instructions")); 109 110static cl::opt<bool> 111NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 112 cl::desc("Don't try to vectorize loads and stores")); 113 114static cl::opt<bool> 115AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 116 cl::desc("Only generate aligned loads and stores")); 117 118static cl::opt<bool> 119NoMemOpBoost("bb-vectorize-no-mem-op-boost", 120 cl::init(false), cl::Hidden, 121 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 122 123static cl::opt<bool> 124FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 125 cl::desc("Use a fast instruction dependency analysis")); 126 127#ifndef NDEBUG 128static cl::opt<bool> 129DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 130 cl::init(false), cl::Hidden, 131 cl::desc("When debugging is enabled, output information on the" 132 " instruction-examination process")); 133static cl::opt<bool> 134DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 135 cl::init(false), cl::Hidden, 136 cl::desc("When debugging is enabled, output information on the" 137 " candidate-selection process")); 138static cl::opt<bool> 139DebugPairSelection("bb-vectorize-debug-pair-selection", 140 cl::init(false), cl::Hidden, 141 cl::desc("When debugging is enabled, output information on the" 142 " pair-selection process")); 143static cl::opt<bool> 144DebugCycleCheck("bb-vectorize-debug-cycle-check", 145 cl::init(false), cl::Hidden, 146 cl::desc("When debugging is enabled, output information on the" 147 " cycle-checking process")); 148#endif 149 150STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 151 152namespace { 153 struct BBVectorize : public BasicBlockPass { 154 static char ID; // Pass identification, replacement for typeid 155 156 const VectorizeConfig Config; 157 158 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 159 : BasicBlockPass(ID), Config(C) { 160 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 161 } 162 163 BBVectorize(Pass *P, const VectorizeConfig &C) 164 : BasicBlockPass(ID), Config(C) { 165 AA = &P->getAnalysis<AliasAnalysis>(); 166 SE = &P->getAnalysis<ScalarEvolution>(); 167 TD = P->getAnalysisIfAvailable<TargetData>(); 168 } 169 170 typedef std::pair<Value *, Value *> ValuePair; 171 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 172 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 173 typedef std::pair<std::multimap<Value *, Value *>::iterator, 174 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 175 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 176 std::multimap<ValuePair, ValuePair>::iterator> 177 VPPIteratorPair; 178 179 AliasAnalysis *AA; 180 ScalarEvolution *SE; 181 TargetData *TD; 182 183 // FIXME: const correct? 184 185 bool vectorizePairs(BasicBlock &BB); 186 187 bool getCandidatePairs(BasicBlock &BB, 188 BasicBlock::iterator &Start, 189 std::multimap<Value *, Value *> &CandidatePairs, 190 std::vector<Value *> &PairableInsts); 191 192 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 193 std::vector<Value *> &PairableInsts, 194 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 195 196 void buildDepMap(BasicBlock &BB, 197 std::multimap<Value *, Value *> &CandidatePairs, 198 std::vector<Value *> &PairableInsts, 199 DenseSet<ValuePair> &PairableInstUsers); 200 201 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 202 std::vector<Value *> &PairableInsts, 203 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 204 DenseSet<ValuePair> &PairableInstUsers, 205 DenseMap<Value *, Value *>& ChosenPairs); 206 207 void fuseChosenPairs(BasicBlock &BB, 208 std::vector<Value *> &PairableInsts, 209 DenseMap<Value *, Value *>& ChosenPairs); 210 211 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 212 213 bool areInstsCompatible(Instruction *I, Instruction *J, 214 bool IsSimpleLoadStore); 215 216 bool trackUsesOfI(DenseSet<Value *> &Users, 217 AliasSetTracker &WriteSet, Instruction *I, 218 Instruction *J, bool UpdateUsers = true, 219 std::multimap<Value *, Value *> *LoadMoveSet = 0); 220 221 void computePairsConnectedTo( 222 std::multimap<Value *, Value *> &CandidatePairs, 223 std::vector<Value *> &PairableInsts, 224 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 225 ValuePair P); 226 227 bool pairsConflict(ValuePair P, ValuePair Q, 228 DenseSet<ValuePair> &PairableInstUsers, 229 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 230 231 bool pairWillFormCycle(ValuePair P, 232 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 233 DenseSet<ValuePair> &CurrentPairs); 234 235 void pruneTreeFor( 236 std::multimap<Value *, Value *> &CandidatePairs, 237 std::vector<Value *> &PairableInsts, 238 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 239 DenseSet<ValuePair> &PairableInstUsers, 240 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 241 DenseMap<Value *, Value *> &ChosenPairs, 242 DenseMap<ValuePair, size_t> &Tree, 243 DenseSet<ValuePair> &PrunedTree, ValuePair J, 244 bool UseCycleCheck); 245 246 void buildInitialTreeFor( 247 std::multimap<Value *, Value *> &CandidatePairs, 248 std::vector<Value *> &PairableInsts, 249 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 250 DenseSet<ValuePair> &PairableInstUsers, 251 DenseMap<Value *, Value *> &ChosenPairs, 252 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 253 254 void findBestTreeFor( 255 std::multimap<Value *, Value *> &CandidatePairs, 256 std::vector<Value *> &PairableInsts, 257 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 258 DenseSet<ValuePair> &PairableInstUsers, 259 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 260 DenseMap<Value *, Value *> &ChosenPairs, 261 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 262 size_t &BestEffSize, VPIteratorPair ChoiceRange, 263 bool UseCycleCheck); 264 265 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 266 Instruction *J, unsigned o, bool &FlipMemInputs); 267 268 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 269 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 270 unsigned IdxOffset, std::vector<Constant*> &Mask); 271 272 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 273 Instruction *J); 274 275 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 276 Instruction *J, unsigned o, bool FlipMemInputs); 277 278 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 279 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 280 bool &FlipMemInputs); 281 282 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 283 Instruction *J, Instruction *K, 284 Instruction *&InsertionPt, Instruction *&K1, 285 Instruction *&K2, bool &FlipMemInputs); 286 287 void collectPairLoadMoveSet(BasicBlock &BB, 288 DenseMap<Value *, Value *> &ChosenPairs, 289 std::multimap<Value *, Value *> &LoadMoveSet, 290 Instruction *I); 291 292 void collectLoadMoveSet(BasicBlock &BB, 293 std::vector<Value *> &PairableInsts, 294 DenseMap<Value *, Value *> &ChosenPairs, 295 std::multimap<Value *, Value *> &LoadMoveSet); 296 297 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 298 std::multimap<Value *, Value *> &LoadMoveSet, 299 Instruction *I, Instruction *J); 300 301 void moveUsesOfIAfterJ(BasicBlock &BB, 302 std::multimap<Value *, Value *> &LoadMoveSet, 303 Instruction *&InsertionPt, 304 Instruction *I, Instruction *J); 305 306 bool vectorizeBB(BasicBlock &BB) { 307 bool changed = false; 308 // Iterate a sufficient number of times to merge types of size 1 bit, 309 // then 2 bits, then 4, etc. up to half of the target vector width of the 310 // target vector register. 311 for (unsigned v = 2, n = 1; 312 v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter); 313 v *= 2, ++n) { 314 DEBUG(dbgs() << "BBV: fusing loop #" << n << 315 " for " << BB.getName() << " in " << 316 BB.getParent()->getName() << "...\n"); 317 if (vectorizePairs(BB)) 318 changed = true; 319 else 320 break; 321 } 322 323 DEBUG(dbgs() << "BBV: done!\n"); 324 return changed; 325 } 326 327 virtual bool runOnBasicBlock(BasicBlock &BB) { 328 AA = &getAnalysis<AliasAnalysis>(); 329 SE = &getAnalysis<ScalarEvolution>(); 330 TD = getAnalysisIfAvailable<TargetData>(); 331 332 return vectorizeBB(BB); 333 } 334 335 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 336 BasicBlockPass::getAnalysisUsage(AU); 337 AU.addRequired<AliasAnalysis>(); 338 AU.addRequired<ScalarEvolution>(); 339 AU.addPreserved<AliasAnalysis>(); 340 AU.addPreserved<ScalarEvolution>(); 341 AU.setPreservesCFG(); 342 } 343 344 // This returns the vector type that holds a pair of the provided type. 345 // If the provided type is already a vector, then its length is doubled. 346 static inline VectorType *getVecTypeForPair(Type *ElemTy) { 347 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 348 unsigned numElem = VTy->getNumElements(); 349 return VectorType::get(ElemTy->getScalarType(), numElem*2); 350 } 351 352 return VectorType::get(ElemTy, 2); 353 } 354 355 // Returns the weight associated with the provided value. A chain of 356 // candidate pairs has a length given by the sum of the weights of its 357 // members (one weight per pair; the weight of each member of the pair 358 // is assumed to be the same). This length is then compared to the 359 // chain-length threshold to determine if a given chain is significant 360 // enough to be vectorized. The length is also used in comparing 361 // candidate chains where longer chains are considered to be better. 362 // Note: when this function returns 0, the resulting instructions are 363 // not actually fused. 364 inline size_t getDepthFactor(Value *V) { 365 // InsertElement and ExtractElement have a depth factor of zero. This is 366 // for two reasons: First, they cannot be usefully fused. Second, because 367 // the pass generates a lot of these, they can confuse the simple metric 368 // used to compare the trees in the next iteration. Thus, giving them a 369 // weight of zero allows the pass to essentially ignore them in 370 // subsequent iterations when looking for vectorization opportunities 371 // while still tracking dependency chains that flow through those 372 // instructions. 373 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 374 return 0; 375 376 // Give a load or store half of the required depth so that load/store 377 // pairs will vectorize. 378 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 379 return Config.ReqChainDepth/2; 380 381 return 1; 382 } 383 384 // This determines the relative offset of two loads or stores, returning 385 // true if the offset could be determined to be some constant value. 386 // For example, if OffsetInElmts == 1, then J accesses the memory directly 387 // after I; if OffsetInElmts == -1 then I accesses the memory 388 // directly after J. This function assumes that both instructions 389 // have the same type. 390 bool getPairPtrInfo(Instruction *I, Instruction *J, 391 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 392 int64_t &OffsetInElmts) { 393 OffsetInElmts = 0; 394 if (isa<LoadInst>(I)) { 395 IPtr = cast<LoadInst>(I)->getPointerOperand(); 396 JPtr = cast<LoadInst>(J)->getPointerOperand(); 397 IAlignment = cast<LoadInst>(I)->getAlignment(); 398 JAlignment = cast<LoadInst>(J)->getAlignment(); 399 } else { 400 IPtr = cast<StoreInst>(I)->getPointerOperand(); 401 JPtr = cast<StoreInst>(J)->getPointerOperand(); 402 IAlignment = cast<StoreInst>(I)->getAlignment(); 403 JAlignment = cast<StoreInst>(J)->getAlignment(); 404 } 405 406 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 407 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 408 409 // If this is a trivial offset, then we'll get something like 410 // 1*sizeof(type). With target data, which we need anyway, this will get 411 // constant folded into a number. 412 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 413 if (const SCEVConstant *ConstOffSCEV = 414 dyn_cast<SCEVConstant>(OffsetSCEV)) { 415 ConstantInt *IntOff = ConstOffSCEV->getValue(); 416 int64_t Offset = IntOff->getSExtValue(); 417 418 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 419 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 420 421 assert(VTy == cast<PointerType>(JPtr->getType())->getElementType()); 422 423 OffsetInElmts = Offset/VTyTSS; 424 return (abs64(Offset) % VTyTSS) == 0; 425 } 426 427 return false; 428 } 429 430 // Returns true if the provided CallInst represents an intrinsic that can 431 // be vectorized. 432 bool isVectorizableIntrinsic(CallInst* I) { 433 Function *F = I->getCalledFunction(); 434 if (!F) return false; 435 436 unsigned IID = F->getIntrinsicID(); 437 if (!IID) return false; 438 439 switch(IID) { 440 default: 441 return false; 442 case Intrinsic::sqrt: 443 case Intrinsic::powi: 444 case Intrinsic::sin: 445 case Intrinsic::cos: 446 case Intrinsic::log: 447 case Intrinsic::log2: 448 case Intrinsic::log10: 449 case Intrinsic::exp: 450 case Intrinsic::exp2: 451 case Intrinsic::pow: 452 return Config.VectorizeMath; 453 case Intrinsic::fma: 454 return Config.VectorizeFMA; 455 } 456 } 457 458 // Returns true if J is the second element in some pair referenced by 459 // some multimap pair iterator pair. 460 template <typename V> 461 bool isSecondInIteratorPair(V J, std::pair< 462 typename std::multimap<V, V>::iterator, 463 typename std::multimap<V, V>::iterator> PairRange) { 464 for (typename std::multimap<V, V>::iterator K = PairRange.first; 465 K != PairRange.second; ++K) 466 if (K->second == J) return true; 467 468 return false; 469 } 470 }; 471 472 // This function implements one vectorization iteration on the provided 473 // basic block. It returns true if the block is changed. 474 bool BBVectorize::vectorizePairs(BasicBlock &BB) { 475 bool ShouldContinue; 476 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 477 478 std::vector<Value *> AllPairableInsts; 479 DenseMap<Value *, Value *> AllChosenPairs; 480 481 do { 482 std::vector<Value *> PairableInsts; 483 std::multimap<Value *, Value *> CandidatePairs; 484 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 485 PairableInsts); 486 if (PairableInsts.empty()) continue; 487 488 // Now we have a map of all of the pairable instructions and we need to 489 // select the best possible pairing. A good pairing is one such that the 490 // users of the pair are also paired. This defines a (directed) forest 491 // over the pairs such that two pairs are connected iff the second pair 492 // uses the first. 493 494 // Note that it only matters that both members of the second pair use some 495 // element of the first pair (to allow for splatting). 496 497 std::multimap<ValuePair, ValuePair> ConnectedPairs; 498 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 499 if (ConnectedPairs.empty()) continue; 500 501 // Build the pairable-instruction dependency map 502 DenseSet<ValuePair> PairableInstUsers; 503 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 504 505 // There is now a graph of the connected pairs. For each variable, pick 506 // the pairing with the largest tree meeting the depth requirement on at 507 // least one branch. Then select all pairings that are part of that tree 508 // and remove them from the list of available pairings and pairable 509 // variables. 510 511 DenseMap<Value *, Value *> ChosenPairs; 512 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, 513 PairableInstUsers, ChosenPairs); 514 515 if (ChosenPairs.empty()) continue; 516 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 517 PairableInsts.end()); 518 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 519 } while (ShouldContinue); 520 521 if (AllChosenPairs.empty()) return false; 522 NumFusedOps += AllChosenPairs.size(); 523 524 // A set of pairs has now been selected. It is now necessary to replace the 525 // paired instructions with vector instructions. For this procedure each 526 // operand must be replaced with a vector operand. This vector is formed 527 // by using build_vector on the old operands. The replaced values are then 528 // replaced with a vector_extract on the result. Subsequent optimization 529 // passes should coalesce the build/extract combinations. 530 531 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 532 return true; 533 } 534 535 // This function returns true if the provided instruction is capable of being 536 // fused into a vector instruction. This determination is based only on the 537 // type and other attributes of the instruction. 538 bool BBVectorize::isInstVectorizable(Instruction *I, 539 bool &IsSimpleLoadStore) { 540 IsSimpleLoadStore = false; 541 542 if (CallInst *C = dyn_cast<CallInst>(I)) { 543 if (!isVectorizableIntrinsic(C)) 544 return false; 545 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 546 // Vectorize simple loads if possbile: 547 IsSimpleLoadStore = L->isSimple(); 548 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 549 return false; 550 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 551 // Vectorize simple stores if possbile: 552 IsSimpleLoadStore = S->isSimple(); 553 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 554 return false; 555 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 556 // We can vectorize casts, but not casts of pointer types, etc. 557 if (!Config.VectorizeCasts) 558 return false; 559 560 Type *SrcTy = C->getSrcTy(); 561 if (!SrcTy->isSingleValueType()) 562 return false; 563 564 Type *DestTy = C->getDestTy(); 565 if (!DestTy->isSingleValueType()) 566 return false; 567 } else if (isa<SelectInst>(I)) { 568 if (!Config.VectorizeSelect) 569 return false; 570 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 571 if (!Config.VectorizeGEP) 572 return false; 573 574 // Currently, vector GEPs exist only with one index. 575 if (G->getNumIndices() != 1) 576 return false; 577 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 578 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 579 return false; 580 } 581 582 // We can't vectorize memory operations without target data 583 if (TD == 0 && IsSimpleLoadStore) 584 return false; 585 586 Type *T1, *T2; 587 if (isa<StoreInst>(I)) { 588 // For stores, it is the value type, not the pointer type that matters 589 // because the value is what will come from a vector register. 590 591 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 592 T1 = IVal->getType(); 593 } else { 594 T1 = I->getType(); 595 } 596 597 if (I->isCast()) 598 T2 = cast<CastInst>(I)->getSrcTy(); 599 else 600 T2 = T1; 601 602 // Not every type can be vectorized... 603 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 604 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 605 return false; 606 607 if (!Config.VectorizeInts 608 && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) 609 return false; 610 611 if (!Config.VectorizeFloats 612 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 613 return false; 614 615 if ((!Config.VectorizePointers || TD == 0) 616 && ((T1->isPointerTy() || 617 (T1->isVectorTy() && T1->getScalarType()->isPointerTy())) || 618 (T2->isPointerTy() || 619 (T2->isVectorTy() && T2->getScalarType()->isPointerTy())) 620 )) 621 return false; 622 623 if (T1->getPrimitiveSizeInBits() > Config.VectorBits/2 || 624 T2->getPrimitiveSizeInBits() > Config.VectorBits/2) 625 return false; 626 627 return true; 628 } 629 630 // This function returns true if the two provided instructions are compatible 631 // (meaning that they can be fused into a vector instruction). This assumes 632 // that I has already been determined to be vectorizable and that J is not 633 // in the use tree of I. 634 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 635 bool IsSimpleLoadStore) { 636 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 637 " <-> " << *J << "\n"); 638 639 // Loads and stores can be merged if they have different alignments, 640 // but are otherwise the same. 641 LoadInst *LI, *LJ; 642 StoreInst *SI, *SJ; 643 if ((LI = dyn_cast<LoadInst>(I)) && (LJ = dyn_cast<LoadInst>(J))) { 644 if (I->getType() != J->getType()) 645 return false; 646 647 if (LI->getPointerOperand()->getType() != 648 LJ->getPointerOperand()->getType() || 649 LI->isVolatile() != LJ->isVolatile() || 650 LI->getOrdering() != LJ->getOrdering() || 651 LI->getSynchScope() != LJ->getSynchScope()) 652 return false; 653 } else if ((SI = dyn_cast<StoreInst>(I)) && (SJ = dyn_cast<StoreInst>(J))) { 654 if (SI->getValueOperand()->getType() != 655 SJ->getValueOperand()->getType() || 656 SI->getPointerOperand()->getType() != 657 SJ->getPointerOperand()->getType() || 658 SI->isVolatile() != SJ->isVolatile() || 659 SI->getOrdering() != SJ->getOrdering() || 660 SI->getSynchScope() != SJ->getSynchScope()) 661 return false; 662 } else if (!J->isSameOperationAs(I)) { 663 return false; 664 } 665 // FIXME: handle addsub-type operations! 666 667 if (IsSimpleLoadStore) { 668 Value *IPtr, *JPtr; 669 unsigned IAlignment, JAlignment; 670 int64_t OffsetInElmts = 0; 671 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 672 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 673 if (Config.AlignedOnly) { 674 Type *aType = isa<StoreInst>(I) ? 675 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 676 // An aligned load or store is possible only if the instruction 677 // with the lower offset has an alignment suitable for the 678 // vector type. 679 680 unsigned BottomAlignment = IAlignment; 681 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 682 683 Type *VType = getVecTypeForPair(aType); 684 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 685 if (BottomAlignment < VecAlignment) 686 return false; 687 } 688 } else { 689 return false; 690 } 691 } else if (isa<ShuffleVectorInst>(I)) { 692 // Only merge two shuffles if they're both constant 693 return isa<Constant>(I->getOperand(2)) && 694 isa<Constant>(J->getOperand(2)); 695 // FIXME: We may want to vectorize non-constant shuffles also. 696 } 697 698 // The powi intrinsic is special because only the first argument is 699 // vectorized, the second arguments must be equal. 700 CallInst *CI = dyn_cast<CallInst>(I); 701 Function *FI; 702 if (CI && (FI = CI->getCalledFunction()) && 703 FI->getIntrinsicID() == Intrinsic::powi) { 704 705 Value *A1I = CI->getArgOperand(1), 706 *A1J = cast<CallInst>(J)->getArgOperand(1); 707 const SCEV *A1ISCEV = SE->getSCEV(A1I), 708 *A1JSCEV = SE->getSCEV(A1J); 709 return (A1ISCEV == A1JSCEV); 710 } 711 712 return true; 713 } 714 715 // Figure out whether or not J uses I and update the users and write-set 716 // structures associated with I. Specifically, Users represents the set of 717 // instructions that depend on I. WriteSet represents the set 718 // of memory locations that are dependent on I. If UpdateUsers is true, 719 // and J uses I, then Users is updated to contain J and WriteSet is updated 720 // to contain any memory locations to which J writes. The function returns 721 // true if J uses I. By default, alias analysis is used to determine 722 // whether J reads from memory that overlaps with a location in WriteSet. 723 // If LoadMoveSet is not null, then it is a previously-computed multimap 724 // where the key is the memory-based user instruction and the value is 725 // the instruction to be compared with I. So, if LoadMoveSet is provided, 726 // then the alias analysis is not used. This is necessary because this 727 // function is called during the process of moving instructions during 728 // vectorization and the results of the alias analysis are not stable during 729 // that process. 730 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 731 AliasSetTracker &WriteSet, Instruction *I, 732 Instruction *J, bool UpdateUsers, 733 std::multimap<Value *, Value *> *LoadMoveSet) { 734 bool UsesI = false; 735 736 // This instruction may already be marked as a user due, for example, to 737 // being a member of a selected pair. 738 if (Users.count(J)) 739 UsesI = true; 740 741 if (!UsesI) 742 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 743 JU != JE; ++JU) { 744 Value *V = *JU; 745 if (I == V || Users.count(V)) { 746 UsesI = true; 747 break; 748 } 749 } 750 if (!UsesI && J->mayReadFromMemory()) { 751 if (LoadMoveSet) { 752 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 753 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 754 } else { 755 for (AliasSetTracker::iterator W = WriteSet.begin(), 756 WE = WriteSet.end(); W != WE; ++W) { 757 if (W->aliasesUnknownInst(J, *AA)) { 758 UsesI = true; 759 break; 760 } 761 } 762 } 763 } 764 765 if (UsesI && UpdateUsers) { 766 if (J->mayWriteToMemory()) WriteSet.add(J); 767 Users.insert(J); 768 } 769 770 return UsesI; 771 } 772 773 // This function iterates over all instruction pairs in the provided 774 // basic block and collects all candidate pairs for vectorization. 775 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 776 BasicBlock::iterator &Start, 777 std::multimap<Value *, Value *> &CandidatePairs, 778 std::vector<Value *> &PairableInsts) { 779 BasicBlock::iterator E = BB.end(); 780 if (Start == E) return false; 781 782 bool ShouldContinue = false, IAfterStart = false; 783 for (BasicBlock::iterator I = Start++; I != E; ++I) { 784 if (I == Start) IAfterStart = true; 785 786 bool IsSimpleLoadStore; 787 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 788 789 // Look for an instruction with which to pair instruction *I... 790 DenseSet<Value *> Users; 791 AliasSetTracker WriteSet(*AA); 792 bool JAfterStart = IAfterStart; 793 BasicBlock::iterator J = llvm::next(I); 794 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 795 if (J == Start) JAfterStart = true; 796 797 // Determine if J uses I, if so, exit the loop. 798 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 799 if (Config.FastDep) { 800 // Note: For this heuristic to be effective, independent operations 801 // must tend to be intermixed. This is likely to be true from some 802 // kinds of grouped loop unrolling (but not the generic LLVM pass), 803 // but otherwise may require some kind of reordering pass. 804 805 // When using fast dependency analysis, 806 // stop searching after first use: 807 if (UsesI) break; 808 } else { 809 if (UsesI) continue; 810 } 811 812 // J does not use I, and comes before the first use of I, so it can be 813 // merged with I if the instructions are compatible. 814 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue; 815 816 // J is a candidate for merging with I. 817 if (!PairableInsts.size() || 818 PairableInsts[PairableInsts.size()-1] != I) { 819 PairableInsts.push_back(I); 820 } 821 822 CandidatePairs.insert(ValuePair(I, J)); 823 824 // The next call to this function must start after the last instruction 825 // selected during this invocation. 826 if (JAfterStart) { 827 Start = llvm::next(J); 828 IAfterStart = JAfterStart = false; 829 } 830 831 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 832 << *I << " <-> " << *J << "\n"); 833 834 // If we have already found too many pairs, break here and this function 835 // will be called again starting after the last instruction selected 836 // during this invocation. 837 if (PairableInsts.size() >= Config.MaxInsts) { 838 ShouldContinue = true; 839 break; 840 } 841 } 842 843 if (ShouldContinue) 844 break; 845 } 846 847 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 848 << " instructions with candidate pairs\n"); 849 850 return ShouldContinue; 851 } 852 853 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 854 // it looks for pairs such that both members have an input which is an 855 // output of PI or PJ. 856 void BBVectorize::computePairsConnectedTo( 857 std::multimap<Value *, Value *> &CandidatePairs, 858 std::vector<Value *> &PairableInsts, 859 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 860 ValuePair P) { 861 // For each possible pairing for this variable, look at the uses of 862 // the first value... 863 for (Value::use_iterator I = P.first->use_begin(), 864 E = P.first->use_end(); I != E; ++I) { 865 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 866 867 // For each use of the first variable, look for uses of the second 868 // variable... 869 for (Value::use_iterator J = P.second->use_begin(), 870 E2 = P.second->use_end(); J != E2; ++J) { 871 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 872 873 // Look for <I, J>: 874 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 875 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 876 877 // Look for <J, I>: 878 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 879 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 880 } 881 882 if (Config.SplatBreaksChain) continue; 883 // Look for cases where just the first value in the pair is used by 884 // both members of another pair (splatting). 885 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 886 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 887 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 888 } 889 } 890 891 if (Config.SplatBreaksChain) return; 892 // Look for cases where just the second value in the pair is used by 893 // both members of another pair (splatting). 894 for (Value::use_iterator I = P.second->use_begin(), 895 E = P.second->use_end(); I != E; ++I) { 896 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 897 898 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 899 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 900 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 901 } 902 } 903 } 904 905 // This function figures out which pairs are connected. Two pairs are 906 // connected if some output of the first pair forms an input to both members 907 // of the second pair. 908 void BBVectorize::computeConnectedPairs( 909 std::multimap<Value *, Value *> &CandidatePairs, 910 std::vector<Value *> &PairableInsts, 911 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 912 913 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 914 PE = PairableInsts.end(); PI != PE; ++PI) { 915 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 916 917 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 918 P != choiceRange.second; ++P) 919 computePairsConnectedTo(CandidatePairs, PairableInsts, 920 ConnectedPairs, *P); 921 } 922 923 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 924 << " pair connections.\n"); 925 } 926 927 // This function builds a set of use tuples such that <A, B> is in the set 928 // if B is in the use tree of A. If B is in the use tree of A, then B 929 // depends on the output of A. 930 void BBVectorize::buildDepMap( 931 BasicBlock &BB, 932 std::multimap<Value *, Value *> &CandidatePairs, 933 std::vector<Value *> &PairableInsts, 934 DenseSet<ValuePair> &PairableInstUsers) { 935 DenseSet<Value *> IsInPair; 936 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 937 E = CandidatePairs.end(); C != E; ++C) { 938 IsInPair.insert(C->first); 939 IsInPair.insert(C->second); 940 } 941 942 // Iterate through the basic block, recording all Users of each 943 // pairable instruction. 944 945 BasicBlock::iterator E = BB.end(); 946 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 947 if (IsInPair.find(I) == IsInPair.end()) continue; 948 949 DenseSet<Value *> Users; 950 AliasSetTracker WriteSet(*AA); 951 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 952 (void) trackUsesOfI(Users, WriteSet, I, J); 953 954 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 955 U != E; ++U) 956 PairableInstUsers.insert(ValuePair(I, *U)); 957 } 958 } 959 960 // Returns true if an input to pair P is an output of pair Q and also an 961 // input of pair Q is an output of pair P. If this is the case, then these 962 // two pairs cannot be simultaneously fused. 963 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 964 DenseSet<ValuePair> &PairableInstUsers, 965 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 966 // Two pairs are in conflict if they are mutual Users of eachother. 967 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 968 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 969 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 970 PairableInstUsers.count(ValuePair(P.second, Q.second)); 971 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 972 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 973 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 974 PairableInstUsers.count(ValuePair(Q.second, P.second)); 975 if (PairableInstUserMap) { 976 // FIXME: The expensive part of the cycle check is not so much the cycle 977 // check itself but this edge insertion procedure. This needs some 978 // profiling and probably a different data structure (same is true of 979 // most uses of std::multimap). 980 if (PUsesQ) { 981 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 982 if (!isSecondInIteratorPair(P, QPairRange)) 983 PairableInstUserMap->insert(VPPair(Q, P)); 984 } 985 if (QUsesP) { 986 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 987 if (!isSecondInIteratorPair(Q, PPairRange)) 988 PairableInstUserMap->insert(VPPair(P, Q)); 989 } 990 } 991 992 return (QUsesP && PUsesQ); 993 } 994 995 // This function walks the use graph of current pairs to see if, starting 996 // from P, the walk returns to P. 997 bool BBVectorize::pairWillFormCycle(ValuePair P, 998 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 999 DenseSet<ValuePair> &CurrentPairs) { 1000 DEBUG(if (DebugCycleCheck) 1001 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1002 << *P.second << "\n"); 1003 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1004 // contains non-direct associations. 1005 DenseSet<ValuePair> Visited; 1006 SmallVector<ValuePair, 32> Q; 1007 // General depth-first post-order traversal: 1008 Q.push_back(P); 1009 do { 1010 ValuePair QTop = Q.pop_back_val(); 1011 Visited.insert(QTop); 1012 1013 DEBUG(if (DebugCycleCheck) 1014 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1015 << *QTop.second << "\n"); 1016 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1017 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1018 C != QPairRange.second; ++C) { 1019 if (C->second == P) { 1020 DEBUG(dbgs() 1021 << "BBV: rejected to prevent non-trivial cycle formation: " 1022 << *C->first.first << " <-> " << *C->first.second << "\n"); 1023 return true; 1024 } 1025 1026 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1027 Q.push_back(C->second); 1028 } 1029 } while (!Q.empty()); 1030 1031 return false; 1032 } 1033 1034 // This function builds the initial tree of connected pairs with the 1035 // pair J at the root. 1036 void BBVectorize::buildInitialTreeFor( 1037 std::multimap<Value *, Value *> &CandidatePairs, 1038 std::vector<Value *> &PairableInsts, 1039 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1040 DenseSet<ValuePair> &PairableInstUsers, 1041 DenseMap<Value *, Value *> &ChosenPairs, 1042 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1043 // Each of these pairs is viewed as the root node of a Tree. The Tree 1044 // is then walked (depth-first). As this happens, we keep track of 1045 // the pairs that compose the Tree and the maximum depth of the Tree. 1046 SmallVector<ValuePairWithDepth, 32> Q; 1047 // General depth-first post-order traversal: 1048 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1049 do { 1050 ValuePairWithDepth QTop = Q.back(); 1051 1052 // Push each child onto the queue: 1053 bool MoreChildren = false; 1054 size_t MaxChildDepth = QTop.second; 1055 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1056 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1057 k != qtRange.second; ++k) { 1058 // Make sure that this child pair is still a candidate: 1059 bool IsStillCand = false; 1060 VPIteratorPair checkRange = 1061 CandidatePairs.equal_range(k->second.first); 1062 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1063 m != checkRange.second; ++m) { 1064 if (m->second == k->second.second) { 1065 IsStillCand = true; 1066 break; 1067 } 1068 } 1069 1070 if (IsStillCand) { 1071 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1072 if (C == Tree.end()) { 1073 size_t d = getDepthFactor(k->second.first); 1074 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1075 MoreChildren = true; 1076 } else { 1077 MaxChildDepth = std::max(MaxChildDepth, C->second); 1078 } 1079 } 1080 } 1081 1082 if (!MoreChildren) { 1083 // Record the current pair as part of the Tree: 1084 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1085 Q.pop_back(); 1086 } 1087 } while (!Q.empty()); 1088 } 1089 1090 // Given some initial tree, prune it by removing conflicting pairs (pairs 1091 // that cannot be simultaneously chosen for vectorization). 1092 void BBVectorize::pruneTreeFor( 1093 std::multimap<Value *, Value *> &CandidatePairs, 1094 std::vector<Value *> &PairableInsts, 1095 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1096 DenseSet<ValuePair> &PairableInstUsers, 1097 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1098 DenseMap<Value *, Value *> &ChosenPairs, 1099 DenseMap<ValuePair, size_t> &Tree, 1100 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1101 bool UseCycleCheck) { 1102 SmallVector<ValuePairWithDepth, 32> Q; 1103 // General depth-first post-order traversal: 1104 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1105 do { 1106 ValuePairWithDepth QTop = Q.pop_back_val(); 1107 PrunedTree.insert(QTop.first); 1108 1109 // Visit each child, pruning as necessary... 1110 DenseMap<ValuePair, size_t> BestChildren; 1111 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1112 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1113 K != QTopRange.second; ++K) { 1114 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1115 if (C == Tree.end()) continue; 1116 1117 // This child is in the Tree, now we need to make sure it is the 1118 // best of any conflicting children. There could be multiple 1119 // conflicting children, so first, determine if we're keeping 1120 // this child, then delete conflicting children as necessary. 1121 1122 // It is also necessary to guard against pairing-induced 1123 // dependencies. Consider instructions a .. x .. y .. b 1124 // such that (a,b) are to be fused and (x,y) are to be fused 1125 // but a is an input to x and b is an output from y. This 1126 // means that y cannot be moved after b but x must be moved 1127 // after b for (a,b) to be fused. In other words, after 1128 // fusing (a,b) we have y .. a/b .. x where y is an input 1129 // to a/b and x is an output to a/b: x and y can no longer 1130 // be legally fused. To prevent this condition, we must 1131 // make sure that a child pair added to the Tree is not 1132 // both an input and output of an already-selected pair. 1133 1134 // Pairing-induced dependencies can also form from more complicated 1135 // cycles. The pair vs. pair conflicts are easy to check, and so 1136 // that is done explicitly for "fast rejection", and because for 1137 // child vs. child conflicts, we may prefer to keep the current 1138 // pair in preference to the already-selected child. 1139 DenseSet<ValuePair> CurrentPairs; 1140 1141 bool CanAdd = true; 1142 for (DenseMap<ValuePair, size_t>::iterator C2 1143 = BestChildren.begin(), E2 = BestChildren.end(); 1144 C2 != E2; ++C2) { 1145 if (C2->first.first == C->first.first || 1146 C2->first.first == C->first.second || 1147 C2->first.second == C->first.first || 1148 C2->first.second == C->first.second || 1149 pairsConflict(C2->first, C->first, PairableInstUsers, 1150 UseCycleCheck ? &PairableInstUserMap : 0)) { 1151 if (C2->second >= C->second) { 1152 CanAdd = false; 1153 break; 1154 } 1155 1156 CurrentPairs.insert(C2->first); 1157 } 1158 } 1159 if (!CanAdd) continue; 1160 1161 // Even worse, this child could conflict with another node already 1162 // selected for the Tree. If that is the case, ignore this child. 1163 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1164 E2 = PrunedTree.end(); T != E2; ++T) { 1165 if (T->first == C->first.first || 1166 T->first == C->first.second || 1167 T->second == C->first.first || 1168 T->second == C->first.second || 1169 pairsConflict(*T, C->first, PairableInstUsers, 1170 UseCycleCheck ? &PairableInstUserMap : 0)) { 1171 CanAdd = false; 1172 break; 1173 } 1174 1175 CurrentPairs.insert(*T); 1176 } 1177 if (!CanAdd) continue; 1178 1179 // And check the queue too... 1180 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1181 E2 = Q.end(); C2 != E2; ++C2) { 1182 if (C2->first.first == C->first.first || 1183 C2->first.first == C->first.second || 1184 C2->first.second == C->first.first || 1185 C2->first.second == C->first.second || 1186 pairsConflict(C2->first, C->first, PairableInstUsers, 1187 UseCycleCheck ? &PairableInstUserMap : 0)) { 1188 CanAdd = false; 1189 break; 1190 } 1191 1192 CurrentPairs.insert(C2->first); 1193 } 1194 if (!CanAdd) continue; 1195 1196 // Last but not least, check for a conflict with any of the 1197 // already-chosen pairs. 1198 for (DenseMap<Value *, Value *>::iterator C2 = 1199 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1200 C2 != E2; ++C2) { 1201 if (pairsConflict(*C2, C->first, PairableInstUsers, 1202 UseCycleCheck ? &PairableInstUserMap : 0)) { 1203 CanAdd = false; 1204 break; 1205 } 1206 1207 CurrentPairs.insert(*C2); 1208 } 1209 if (!CanAdd) continue; 1210 1211 // To check for non-trivial cycles formed by the addition of the 1212 // current pair we've formed a list of all relevant pairs, now use a 1213 // graph walk to check for a cycle. We start from the current pair and 1214 // walk the use tree to see if we again reach the current pair. If we 1215 // do, then the current pair is rejected. 1216 1217 // FIXME: It may be more efficient to use a topological-ordering 1218 // algorithm to improve the cycle check. This should be investigated. 1219 if (UseCycleCheck && 1220 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1221 continue; 1222 1223 // This child can be added, but we may have chosen it in preference 1224 // to an already-selected child. Check for this here, and if a 1225 // conflict is found, then remove the previously-selected child 1226 // before adding this one in its place. 1227 for (DenseMap<ValuePair, size_t>::iterator C2 1228 = BestChildren.begin(); C2 != BestChildren.end();) { 1229 if (C2->first.first == C->first.first || 1230 C2->first.first == C->first.second || 1231 C2->first.second == C->first.first || 1232 C2->first.second == C->first.second || 1233 pairsConflict(C2->first, C->first, PairableInstUsers)) 1234 BestChildren.erase(C2++); 1235 else 1236 ++C2; 1237 } 1238 1239 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1240 } 1241 1242 for (DenseMap<ValuePair, size_t>::iterator C 1243 = BestChildren.begin(), E2 = BestChildren.end(); 1244 C != E2; ++C) { 1245 size_t DepthF = getDepthFactor(C->first.first); 1246 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1247 } 1248 } while (!Q.empty()); 1249 } 1250 1251 // This function finds the best tree of mututally-compatible connected 1252 // pairs, given the choice of root pairs as an iterator range. 1253 void BBVectorize::findBestTreeFor( 1254 std::multimap<Value *, Value *> &CandidatePairs, 1255 std::vector<Value *> &PairableInsts, 1256 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1257 DenseSet<ValuePair> &PairableInstUsers, 1258 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1259 DenseMap<Value *, Value *> &ChosenPairs, 1260 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1261 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1262 bool UseCycleCheck) { 1263 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1264 J != ChoiceRange.second; ++J) { 1265 1266 // Before going any further, make sure that this pair does not 1267 // conflict with any already-selected pairs (see comment below 1268 // near the Tree pruning for more details). 1269 DenseSet<ValuePair> ChosenPairSet; 1270 bool DoesConflict = false; 1271 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1272 E = ChosenPairs.end(); C != E; ++C) { 1273 if (pairsConflict(*C, *J, PairableInstUsers, 1274 UseCycleCheck ? &PairableInstUserMap : 0)) { 1275 DoesConflict = true; 1276 break; 1277 } 1278 1279 ChosenPairSet.insert(*C); 1280 } 1281 if (DoesConflict) continue; 1282 1283 if (UseCycleCheck && 1284 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1285 continue; 1286 1287 DenseMap<ValuePair, size_t> Tree; 1288 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1289 PairableInstUsers, ChosenPairs, Tree, *J); 1290 1291 // Because we'll keep the child with the largest depth, the largest 1292 // depth is still the same in the unpruned Tree. 1293 size_t MaxDepth = Tree.lookup(*J); 1294 1295 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1296 << *J->first << " <-> " << *J->second << "} of depth " << 1297 MaxDepth << " and size " << Tree.size() << "\n"); 1298 1299 // At this point the Tree has been constructed, but, may contain 1300 // contradictory children (meaning that different children of 1301 // some tree node may be attempting to fuse the same instruction). 1302 // So now we walk the tree again, in the case of a conflict, 1303 // keep only the child with the largest depth. To break a tie, 1304 // favor the first child. 1305 1306 DenseSet<ValuePair> PrunedTree; 1307 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1308 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1309 PrunedTree, *J, UseCycleCheck); 1310 1311 size_t EffSize = 0; 1312 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1313 E = PrunedTree.end(); S != E; ++S) 1314 EffSize += getDepthFactor(S->first); 1315 1316 DEBUG(if (DebugPairSelection) 1317 dbgs() << "BBV: found pruned Tree for pair {" 1318 << *J->first << " <-> " << *J->second << "} of depth " << 1319 MaxDepth << " and size " << PrunedTree.size() << 1320 " (effective size: " << EffSize << ")\n"); 1321 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) { 1322 BestMaxDepth = MaxDepth; 1323 BestEffSize = EffSize; 1324 BestTree = PrunedTree; 1325 } 1326 } 1327 } 1328 1329 // Given the list of candidate pairs, this function selects those 1330 // that will be fused into vector instructions. 1331 void BBVectorize::choosePairs( 1332 std::multimap<Value *, Value *> &CandidatePairs, 1333 std::vector<Value *> &PairableInsts, 1334 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1335 DenseSet<ValuePair> &PairableInstUsers, 1336 DenseMap<Value *, Value *>& ChosenPairs) { 1337 bool UseCycleCheck = 1338 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1339 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1340 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1341 E = PairableInsts.end(); I != E; ++I) { 1342 // The number of possible pairings for this variable: 1343 size_t NumChoices = CandidatePairs.count(*I); 1344 if (!NumChoices) continue; 1345 1346 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1347 1348 // The best pair to choose and its tree: 1349 size_t BestMaxDepth = 0, BestEffSize = 0; 1350 DenseSet<ValuePair> BestTree; 1351 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1352 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1353 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1354 UseCycleCheck); 1355 1356 // A tree has been chosen (or not) at this point. If no tree was 1357 // chosen, then this instruction, I, cannot be paired (and is no longer 1358 // considered). 1359 1360 DEBUG(if (BestTree.size() > 0) 1361 dbgs() << "BBV: selected pairs in the best tree for: " 1362 << *cast<Instruction>(*I) << "\n"); 1363 1364 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1365 SE2 = BestTree.end(); S != SE2; ++S) { 1366 // Insert the members of this tree into the list of chosen pairs. 1367 ChosenPairs.insert(ValuePair(S->first, S->second)); 1368 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1369 *S->second << "\n"); 1370 1371 // Remove all candidate pairs that have values in the chosen tree. 1372 for (std::multimap<Value *, Value *>::iterator K = 1373 CandidatePairs.begin(); K != CandidatePairs.end();) { 1374 if (K->first == S->first || K->second == S->first || 1375 K->second == S->second || K->first == S->second) { 1376 // Don't remove the actual pair chosen so that it can be used 1377 // in subsequent tree selections. 1378 if (!(K->first == S->first && K->second == S->second)) 1379 CandidatePairs.erase(K++); 1380 else 1381 ++K; 1382 } else { 1383 ++K; 1384 } 1385 } 1386 } 1387 } 1388 1389 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1390 } 1391 1392 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1393 unsigned n = 0) { 1394 if (!I->hasName()) 1395 return ""; 1396 1397 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1398 (n > 0 ? "." + utostr(n) : "")).str(); 1399 } 1400 1401 // Returns the value that is to be used as the pointer input to the vector 1402 // instruction that fuses I with J. 1403 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1404 Instruction *I, Instruction *J, unsigned o, 1405 bool &FlipMemInputs) { 1406 Value *IPtr, *JPtr; 1407 unsigned IAlignment, JAlignment; 1408 int64_t OffsetInElmts; 1409 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1410 OffsetInElmts); 1411 1412 // The pointer value is taken to be the one with the lowest offset. 1413 Value *VPtr; 1414 if (OffsetInElmts > 0) { 1415 VPtr = IPtr; 1416 } else { 1417 FlipMemInputs = true; 1418 VPtr = JPtr; 1419 } 1420 1421 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType(); 1422 Type *VArgType = getVecTypeForPair(ArgType); 1423 Type *VArgPtrType = PointerType::get(VArgType, 1424 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1425 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1426 /* insert before */ FlipMemInputs ? J : I); 1427 } 1428 1429 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1430 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 1431 unsigned IdxOffset, std::vector<Constant*> &Mask) { 1432 for (unsigned v = 0; v < NumElem/2; ++v) { 1433 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1434 if (m < 0) { 1435 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1436 } else { 1437 unsigned mm = m + (int) IdxOffset; 1438 if (m >= (int) NumInElem) 1439 mm += (int) NumInElem; 1440 1441 Mask[v+MaskOffset] = 1442 ConstantInt::get(Type::getInt32Ty(Context), mm); 1443 } 1444 } 1445 } 1446 1447 // Returns the value that is to be used as the vector-shuffle mask to the 1448 // vector instruction that fuses I with J. 1449 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1450 Instruction *I, Instruction *J) { 1451 // This is the shuffle mask. We need to append the second 1452 // mask to the first, and the numbers need to be adjusted. 1453 1454 Type *ArgType = I->getType(); 1455 Type *VArgType = getVecTypeForPair(ArgType); 1456 1457 // Get the total number of elements in the fused vector type. 1458 // By definition, this must equal the number of elements in 1459 // the final mask. 1460 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1461 std::vector<Constant*> Mask(NumElem); 1462 1463 Type *OpType = I->getOperand(0)->getType(); 1464 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements(); 1465 1466 // For the mask from the first pair... 1467 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask); 1468 1469 // For the mask from the second pair... 1470 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem, 1471 Mask); 1472 1473 return ConstantVector::get(Mask); 1474 } 1475 1476 // Returns the value to be used as the specified operand of the vector 1477 // instruction that fuses I with J. 1478 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1479 Instruction *J, unsigned o, bool FlipMemInputs) { 1480 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1481 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1482 1483 // Compute the fused vector type for this operand 1484 Type *ArgType = I->getOperand(o)->getType(); 1485 VectorType *VArgType = getVecTypeForPair(ArgType); 1486 1487 Instruction *L = I, *H = J; 1488 if (FlipMemInputs) { 1489 L = J; 1490 H = I; 1491 } 1492 1493 if (ArgType->isVectorTy()) { 1494 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1495 std::vector<Constant*> Mask(numElem); 1496 for (unsigned v = 0; v < numElem; ++v) 1497 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1498 1499 Instruction *BV = new ShuffleVectorInst(L->getOperand(o), 1500 H->getOperand(o), 1501 ConstantVector::get(Mask), 1502 getReplacementName(I, true, o)); 1503 BV->insertBefore(J); 1504 return BV; 1505 } 1506 1507 // If these two inputs are the output of another vector instruction, 1508 // then we should use that output directly. It might be necessary to 1509 // permute it first. [When pairings are fused recursively, you can 1510 // end up with cases where a large vector is decomposed into scalars 1511 // using extractelement instructions, then built into size-2 1512 // vectors using insertelement and the into larger vectors using 1513 // shuffles. InstCombine does not simplify all of these cases well, 1514 // and so we make sure that shuffles are generated here when possible. 1515 ExtractElementInst *LEE 1516 = dyn_cast<ExtractElementInst>(L->getOperand(o)); 1517 ExtractElementInst *HEE 1518 = dyn_cast<ExtractElementInst>(H->getOperand(o)); 1519 1520 if (LEE && HEE && 1521 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) { 1522 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType()); 1523 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue(); 1524 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue(); 1525 if (LEE->getOperand(0) == HEE->getOperand(0)) { 1526 if (LowIndx == 0 && HighIndx == 1) 1527 return LEE->getOperand(0); 1528 1529 std::vector<Constant*> Mask(2); 1530 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1531 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1532 1533 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1534 UndefValue::get(EEType), 1535 ConstantVector::get(Mask), 1536 getReplacementName(I, true, o)); 1537 BV->insertBefore(J); 1538 return BV; 1539 } 1540 1541 std::vector<Constant*> Mask(2); 1542 HighIndx += EEType->getNumElements(); 1543 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1544 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1545 1546 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1547 HEE->getOperand(0), 1548 ConstantVector::get(Mask), 1549 getReplacementName(I, true, o)); 1550 BV->insertBefore(J); 1551 return BV; 1552 } 1553 1554 Instruction *BV1 = InsertElementInst::Create( 1555 UndefValue::get(VArgType), 1556 L->getOperand(o), CV0, 1557 getReplacementName(I, true, o, 1)); 1558 BV1->insertBefore(I); 1559 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o), 1560 CV1, 1561 getReplacementName(I, true, o, 2)); 1562 BV2->insertBefore(J); 1563 return BV2; 1564 } 1565 1566 // This function creates an array of values that will be used as the inputs 1567 // to the vector instruction that fuses I with J. 1568 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1569 Instruction *I, Instruction *J, 1570 SmallVector<Value *, 3> &ReplacedOperands, 1571 bool &FlipMemInputs) { 1572 FlipMemInputs = false; 1573 unsigned NumOperands = I->getNumOperands(); 1574 1575 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1576 // Iterate backward so that we look at the store pointer 1577 // first and know whether or not we need to flip the inputs. 1578 1579 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1580 // This is the pointer for a load/store instruction. 1581 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 1582 FlipMemInputs); 1583 continue; 1584 } else if (isa<CallInst>(I)) { 1585 Function *F = cast<CallInst>(I)->getCalledFunction(); 1586 unsigned IID = F->getIntrinsicID(); 1587 if (o == NumOperands-1) { 1588 BasicBlock &BB = *I->getParent(); 1589 1590 Module *M = BB.getParent()->getParent(); 1591 Type *ArgType = I->getType(); 1592 Type *VArgType = getVecTypeForPair(ArgType); 1593 1594 // FIXME: is it safe to do this here? 1595 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 1596 (Intrinsic::ID) IID, VArgType); 1597 continue; 1598 } else if (IID == Intrinsic::powi && o == 1) { 1599 // The second argument of powi is a single integer and we've already 1600 // checked that both arguments are equal. As a result, we just keep 1601 // I's second argument. 1602 ReplacedOperands[o] = I->getOperand(o); 1603 continue; 1604 } 1605 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 1606 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 1607 continue; 1608 } 1609 1610 ReplacedOperands[o] = 1611 getReplacementInput(Context, I, J, o, FlipMemInputs); 1612 } 1613 } 1614 1615 // This function creates two values that represent the outputs of the 1616 // original I and J instructions. These are generally vector shuffles 1617 // or extracts. In many cases, these will end up being unused and, thus, 1618 // eliminated by later passes. 1619 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 1620 Instruction *J, Instruction *K, 1621 Instruction *&InsertionPt, 1622 Instruction *&K1, Instruction *&K2, 1623 bool &FlipMemInputs) { 1624 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1625 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1626 1627 if (isa<StoreInst>(I)) { 1628 AA->replaceWithNewValue(I, K); 1629 AA->replaceWithNewValue(J, K); 1630 } else { 1631 Type *IType = I->getType(); 1632 Type *VType = getVecTypeForPair(IType); 1633 1634 if (IType->isVectorTy()) { 1635 unsigned numElem = cast<VectorType>(IType)->getNumElements(); 1636 std::vector<Constant*> Mask1(numElem), Mask2(numElem); 1637 for (unsigned v = 0; v < numElem; ++v) { 1638 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1639 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v); 1640 } 1641 1642 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 1643 ConstantVector::get( 1644 FlipMemInputs ? Mask2 : Mask1), 1645 getReplacementName(K, false, 1)); 1646 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 1647 ConstantVector::get( 1648 FlipMemInputs ? Mask1 : Mask2), 1649 getReplacementName(K, false, 2)); 1650 } else { 1651 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 1652 getReplacementName(K, false, 1)); 1653 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 1654 getReplacementName(K, false, 2)); 1655 } 1656 1657 K1->insertAfter(K); 1658 K2->insertAfter(K1); 1659 InsertionPt = K2; 1660 } 1661 } 1662 1663 // Move all uses of the function I (including pairing-induced uses) after J. 1664 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 1665 std::multimap<Value *, Value *> &LoadMoveSet, 1666 Instruction *I, Instruction *J) { 1667 // Skip to the first instruction past I. 1668 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1669 1670 DenseSet<Value *> Users; 1671 AliasSetTracker WriteSet(*AA); 1672 for (; cast<Instruction>(L) != J; ++L) 1673 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 1674 1675 assert(cast<Instruction>(L) == J && 1676 "Tracking has not proceeded far enough to check for dependencies"); 1677 // If J is now in the use set of I, then trackUsesOfI will return true 1678 // and we have a dependency cycle (and the fusing operation must abort). 1679 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 1680 } 1681 1682 // Move all uses of the function I (including pairing-induced uses) after J. 1683 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 1684 std::multimap<Value *, Value *> &LoadMoveSet, 1685 Instruction *&InsertionPt, 1686 Instruction *I, Instruction *J) { 1687 // Skip to the first instruction past I. 1688 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1689 1690 DenseSet<Value *> Users; 1691 AliasSetTracker WriteSet(*AA); 1692 for (; cast<Instruction>(L) != J;) { 1693 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 1694 // Move this instruction 1695 Instruction *InstToMove = L; ++L; 1696 1697 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 1698 " to after " << *InsertionPt << "\n"); 1699 InstToMove->removeFromParent(); 1700 InstToMove->insertAfter(InsertionPt); 1701 InsertionPt = InstToMove; 1702 } else { 1703 ++L; 1704 } 1705 } 1706 } 1707 1708 // Collect all load instruction that are in the move set of a given first 1709 // pair member. These loads depend on the first instruction, I, and so need 1710 // to be moved after J (the second instruction) when the pair is fused. 1711 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 1712 DenseMap<Value *, Value *> &ChosenPairs, 1713 std::multimap<Value *, Value *> &LoadMoveSet, 1714 Instruction *I) { 1715 // Skip to the first instruction past I. 1716 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1717 1718 DenseSet<Value *> Users; 1719 AliasSetTracker WriteSet(*AA); 1720 1721 // Note: We cannot end the loop when we reach J because J could be moved 1722 // farther down the use chain by another instruction pairing. Also, J 1723 // could be before I if this is an inverted input. 1724 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 1725 if (trackUsesOfI(Users, WriteSet, I, L)) { 1726 if (L->mayReadFromMemory()) 1727 LoadMoveSet.insert(ValuePair(L, I)); 1728 } 1729 } 1730 } 1731 1732 // In cases where both load/stores and the computation of their pointers 1733 // are chosen for vectorization, we can end up in a situation where the 1734 // aliasing analysis starts returning different query results as the 1735 // process of fusing instruction pairs continues. Because the algorithm 1736 // relies on finding the same use trees here as were found earlier, we'll 1737 // need to precompute the necessary aliasing information here and then 1738 // manually update it during the fusion process. 1739 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 1740 std::vector<Value *> &PairableInsts, 1741 DenseMap<Value *, Value *> &ChosenPairs, 1742 std::multimap<Value *, Value *> &LoadMoveSet) { 1743 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1744 PIE = PairableInsts.end(); PI != PIE; ++PI) { 1745 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 1746 if (P == ChosenPairs.end()) continue; 1747 1748 Instruction *I = cast<Instruction>(P->first); 1749 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 1750 } 1751 } 1752 1753 // This function fuses the chosen instruction pairs into vector instructions, 1754 // taking care preserve any needed scalar outputs and, then, it reorders the 1755 // remaining instructions as needed (users of the first member of the pair 1756 // need to be moved to after the location of the second member of the pair 1757 // because the vector instruction is inserted in the location of the pair's 1758 // second member). 1759 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 1760 std::vector<Value *> &PairableInsts, 1761 DenseMap<Value *, Value *> &ChosenPairs) { 1762 LLVMContext& Context = BB.getContext(); 1763 1764 // During the vectorization process, the order of the pairs to be fused 1765 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 1766 // list. After a pair is fused, the flipped pair is removed from the list. 1767 std::vector<ValuePair> FlippedPairs; 1768 FlippedPairs.reserve(ChosenPairs.size()); 1769 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 1770 E = ChosenPairs.end(); P != E; ++P) 1771 FlippedPairs.push_back(ValuePair(P->second, P->first)); 1772 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 1773 E = FlippedPairs.end(); P != E; ++P) 1774 ChosenPairs.insert(*P); 1775 1776 std::multimap<Value *, Value *> LoadMoveSet; 1777 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 1778 1779 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 1780 1781 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 1782 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 1783 if (P == ChosenPairs.end()) { 1784 ++PI; 1785 continue; 1786 } 1787 1788 if (getDepthFactor(P->first) == 0) { 1789 // These instructions are not really fused, but are tracked as though 1790 // they are. Any case in which it would be interesting to fuse them 1791 // will be taken care of by InstCombine. 1792 --NumFusedOps; 1793 ++PI; 1794 continue; 1795 } 1796 1797 Instruction *I = cast<Instruction>(P->first), 1798 *J = cast<Instruction>(P->second); 1799 1800 DEBUG(dbgs() << "BBV: fusing: " << *I << 1801 " <-> " << *J << "\n"); 1802 1803 // Remove the pair and flipped pair from the list. 1804 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 1805 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 1806 ChosenPairs.erase(FP); 1807 ChosenPairs.erase(P); 1808 1809 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 1810 DEBUG(dbgs() << "BBV: fusion of: " << *I << 1811 " <-> " << *J << 1812 " aborted because of non-trivial dependency cycle\n"); 1813 --NumFusedOps; 1814 ++PI; 1815 continue; 1816 } 1817 1818 bool FlipMemInputs; 1819 unsigned NumOperands = I->getNumOperands(); 1820 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 1821 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 1822 FlipMemInputs); 1823 1824 // Make a copy of the original operation, change its type to the vector 1825 // type and replace its operands with the vector operands. 1826 Instruction *K = I->clone(); 1827 if (I->hasName()) K->takeName(I); 1828 1829 if (!isa<StoreInst>(K)) 1830 K->mutateType(getVecTypeForPair(I->getType())); 1831 1832 for (unsigned o = 0; o < NumOperands; ++o) 1833 K->setOperand(o, ReplacedOperands[o]); 1834 1835 // If we've flipped the memory inputs, make sure that we take the correct 1836 // alignment. 1837 if (FlipMemInputs) { 1838 if (isa<StoreInst>(K)) 1839 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 1840 else 1841 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 1842 } 1843 1844 K->insertAfter(J); 1845 1846 // Instruction insertion point: 1847 Instruction *InsertionPt = K; 1848 Instruction *K1 = 0, *K2 = 0; 1849 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 1850 FlipMemInputs); 1851 1852 // The use tree of the first original instruction must be moved to after 1853 // the location of the second instruction. The entire use tree of the 1854 // first instruction is disjoint from the input tree of the second 1855 // (by definition), and so commutes with it. 1856 1857 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 1858 1859 if (!isa<StoreInst>(I)) { 1860 I->replaceAllUsesWith(K1); 1861 J->replaceAllUsesWith(K2); 1862 AA->replaceWithNewValue(I, K1); 1863 AA->replaceWithNewValue(J, K2); 1864 } 1865 1866 // Instructions that may read from memory may be in the load move set. 1867 // Once an instruction is fused, we no longer need its move set, and so 1868 // the values of the map never need to be updated. However, when a load 1869 // is fused, we need to merge the entries from both instructions in the 1870 // pair in case those instructions were in the move set of some other 1871 // yet-to-be-fused pair. The loads in question are the keys of the map. 1872 if (I->mayReadFromMemory()) { 1873 std::vector<ValuePair> NewSetMembers; 1874 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 1875 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 1876 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 1877 N != IPairRange.second; ++N) 1878 NewSetMembers.push_back(ValuePair(K, N->second)); 1879 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 1880 N != JPairRange.second; ++N) 1881 NewSetMembers.push_back(ValuePair(K, N->second)); 1882 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 1883 AE = NewSetMembers.end(); A != AE; ++A) 1884 LoadMoveSet.insert(*A); 1885 } 1886 1887 // Before removing I, set the iterator to the next instruction. 1888 PI = llvm::next(BasicBlock::iterator(I)); 1889 if (cast<Instruction>(PI) == J) 1890 ++PI; 1891 1892 SE->forgetValue(I); 1893 SE->forgetValue(J); 1894 I->eraseFromParent(); 1895 J->eraseFromParent(); 1896 } 1897 1898 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 1899 } 1900} 1901 1902char BBVectorize::ID = 0; 1903static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 1904INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1905INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 1906INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 1907INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1908 1909BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 1910 return new BBVectorize(C); 1911} 1912 1913bool 1914llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 1915 BBVectorize BBVectorizer(P, C); 1916 return BBVectorizer.vectorizeBB(BB); 1917} 1918 1919//===----------------------------------------------------------------------===// 1920VectorizeConfig::VectorizeConfig() { 1921 VectorBits = ::VectorBits; 1922 VectorizeInts = !::NoInts; 1923 VectorizeFloats = !::NoFloats; 1924 VectorizePointers = !::NoPointers; 1925 VectorizeCasts = !::NoCasts; 1926 VectorizeMath = !::NoMath; 1927 VectorizeFMA = !::NoFMA; 1928 VectorizeSelect = !::NoSelect; 1929 VectorizeGEP = !::NoGEP; 1930 VectorizeMemOps = !::NoMemOps; 1931 AlignedOnly = ::AlignedOnly; 1932 ReqChainDepth= ::ReqChainDepth; 1933 SearchLimit = ::SearchLimit; 1934 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 1935 SplatBreaksChain = ::SplatBreaksChain; 1936 MaxInsts = ::MaxInsts; 1937 MaxIter = ::MaxIter; 1938 NoMemOpBoost = ::NoMemOpBoost; 1939 FastDep = ::FastDep; 1940} 1941