BBVectorize.cpp revision ab4684e26fe21857d8c8bc6ba7a5234c35117c83
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/ScalarEvolution.h" 38#include "llvm/Analysis/ScalarEvolutionExpressions.h" 39#include "llvm/Analysis/ValueTracking.h" 40#include "llvm/Support/CommandLine.h" 41#include "llvm/Support/Debug.h" 42#include "llvm/Support/raw_ostream.h" 43#include "llvm/Support/ValueHandle.h" 44#include "llvm/Target/TargetData.h" 45#include "llvm/Transforms/Vectorize.h" 46#include <algorithm> 47#include <map> 48using namespace llvm; 49 50static cl::opt<unsigned> 51ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 52 cl::desc("The required chain depth for vectorization")); 53 54static cl::opt<unsigned> 55SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 56 cl::desc("The maximum search distance for instruction pairs")); 57 58static cl::opt<bool> 59SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 60 cl::desc("Replicating one element to a pair breaks the chain")); 61 62static cl::opt<unsigned> 63VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 64 cl::desc("The size of the native vector registers")); 65 66static cl::opt<unsigned> 67MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 68 cl::desc("The maximum number of pairing iterations")); 69 70static cl::opt<unsigned> 71MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 72 cl::desc("The maximum number of pairable instructions per group")); 73 74static cl::opt<unsigned> 75MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 76 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 77 " a full cycle check")); 78 79static cl::opt<bool> 80NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 81 cl::desc("Don't try to vectorize integer values")); 82 83static cl::opt<bool> 84NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 85 cl::desc("Don't try to vectorize floating-point values")); 86 87static cl::opt<bool> 88NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden, 89 cl::desc("Don't try to vectorize pointer values")); 90 91static cl::opt<bool> 92NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 93 cl::desc("Don't try to vectorize casting (conversion) operations")); 94 95static cl::opt<bool> 96NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 97 cl::desc("Don't try to vectorize floating-point math intrinsics")); 98 99static cl::opt<bool> 100NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 101 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 102 103static cl::opt<bool> 104NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 105 cl::desc("Don't try to vectorize select instructions")); 106 107static cl::opt<bool> 108NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 109 cl::desc("Don't try to vectorize getelementptr instructions")); 110 111static cl::opt<bool> 112NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 113 cl::desc("Don't try to vectorize loads and stores")); 114 115static cl::opt<bool> 116AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 117 cl::desc("Only generate aligned loads and stores")); 118 119static cl::opt<bool> 120NoMemOpBoost("bb-vectorize-no-mem-op-boost", 121 cl::init(false), cl::Hidden, 122 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 123 124static cl::opt<bool> 125FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 126 cl::desc("Use a fast instruction dependency analysis")); 127 128#ifndef NDEBUG 129static cl::opt<bool> 130DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 131 cl::init(false), cl::Hidden, 132 cl::desc("When debugging is enabled, output information on the" 133 " instruction-examination process")); 134static cl::opt<bool> 135DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 136 cl::init(false), cl::Hidden, 137 cl::desc("When debugging is enabled, output information on the" 138 " candidate-selection process")); 139static cl::opt<bool> 140DebugPairSelection("bb-vectorize-debug-pair-selection", 141 cl::init(false), cl::Hidden, 142 cl::desc("When debugging is enabled, output information on the" 143 " pair-selection process")); 144static cl::opt<bool> 145DebugCycleCheck("bb-vectorize-debug-cycle-check", 146 cl::init(false), cl::Hidden, 147 cl::desc("When debugging is enabled, output information on the" 148 " cycle-checking process")); 149#endif 150 151STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 152 153namespace { 154 struct BBVectorize : public BasicBlockPass { 155 static char ID; // Pass identification, replacement for typeid 156 157 const VectorizeConfig Config; 158 159 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 160 : BasicBlockPass(ID), Config(C) { 161 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 162 } 163 164 BBVectorize(Pass *P, const VectorizeConfig &C) 165 : BasicBlockPass(ID), Config(C) { 166 AA = &P->getAnalysis<AliasAnalysis>(); 167 SE = &P->getAnalysis<ScalarEvolution>(); 168 TD = P->getAnalysisIfAvailable<TargetData>(); 169 } 170 171 typedef std::pair<Value *, Value *> ValuePair; 172 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 173 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 174 typedef std::pair<std::multimap<Value *, Value *>::iterator, 175 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 176 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 177 std::multimap<ValuePair, ValuePair>::iterator> 178 VPPIteratorPair; 179 180 AliasAnalysis *AA; 181 ScalarEvolution *SE; 182 TargetData *TD; 183 184 // FIXME: const correct? 185 186 bool vectorizePairs(BasicBlock &BB); 187 188 bool getCandidatePairs(BasicBlock &BB, 189 BasicBlock::iterator &Start, 190 std::multimap<Value *, Value *> &CandidatePairs, 191 std::vector<Value *> &PairableInsts); 192 193 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 194 std::vector<Value *> &PairableInsts, 195 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 196 197 void buildDepMap(BasicBlock &BB, 198 std::multimap<Value *, Value *> &CandidatePairs, 199 std::vector<Value *> &PairableInsts, 200 DenseSet<ValuePair> &PairableInstUsers); 201 202 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 203 std::vector<Value *> &PairableInsts, 204 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 205 DenseSet<ValuePair> &PairableInstUsers, 206 DenseMap<Value *, Value *>& ChosenPairs); 207 208 void fuseChosenPairs(BasicBlock &BB, 209 std::vector<Value *> &PairableInsts, 210 DenseMap<Value *, Value *>& ChosenPairs); 211 212 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 213 214 bool areInstsCompatible(Instruction *I, Instruction *J, 215 bool IsSimpleLoadStore); 216 217 bool trackUsesOfI(DenseSet<Value *> &Users, 218 AliasSetTracker &WriteSet, Instruction *I, 219 Instruction *J, bool UpdateUsers = true, 220 std::multimap<Value *, Value *> *LoadMoveSet = 0); 221 222 void computePairsConnectedTo( 223 std::multimap<Value *, Value *> &CandidatePairs, 224 std::vector<Value *> &PairableInsts, 225 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 226 ValuePair P); 227 228 bool pairsConflict(ValuePair P, ValuePair Q, 229 DenseSet<ValuePair> &PairableInstUsers, 230 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 231 232 bool pairWillFormCycle(ValuePair P, 233 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 234 DenseSet<ValuePair> &CurrentPairs); 235 236 void pruneTreeFor( 237 std::multimap<Value *, Value *> &CandidatePairs, 238 std::vector<Value *> &PairableInsts, 239 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 240 DenseSet<ValuePair> &PairableInstUsers, 241 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 242 DenseMap<Value *, Value *> &ChosenPairs, 243 DenseMap<ValuePair, size_t> &Tree, 244 DenseSet<ValuePair> &PrunedTree, ValuePair J, 245 bool UseCycleCheck); 246 247 void buildInitialTreeFor( 248 std::multimap<Value *, Value *> &CandidatePairs, 249 std::vector<Value *> &PairableInsts, 250 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 251 DenseSet<ValuePair> &PairableInstUsers, 252 DenseMap<Value *, Value *> &ChosenPairs, 253 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 254 255 void findBestTreeFor( 256 std::multimap<Value *, Value *> &CandidatePairs, 257 std::vector<Value *> &PairableInsts, 258 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 259 DenseSet<ValuePair> &PairableInstUsers, 260 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 261 DenseMap<Value *, Value *> &ChosenPairs, 262 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 263 size_t &BestEffSize, VPIteratorPair ChoiceRange, 264 bool UseCycleCheck); 265 266 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 267 Instruction *J, unsigned o, bool &FlipMemInputs); 268 269 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 270 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 271 unsigned IdxOffset, std::vector<Constant*> &Mask); 272 273 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 274 Instruction *J); 275 276 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 277 Instruction *J, unsigned o, bool FlipMemInputs); 278 279 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 280 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 281 bool &FlipMemInputs); 282 283 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 284 Instruction *J, Instruction *K, 285 Instruction *&InsertionPt, Instruction *&K1, 286 Instruction *&K2, bool &FlipMemInputs); 287 288 void collectPairLoadMoveSet(BasicBlock &BB, 289 DenseMap<Value *, Value *> &ChosenPairs, 290 std::multimap<Value *, Value *> &LoadMoveSet, 291 Instruction *I); 292 293 void collectLoadMoveSet(BasicBlock &BB, 294 std::vector<Value *> &PairableInsts, 295 DenseMap<Value *, Value *> &ChosenPairs, 296 std::multimap<Value *, Value *> &LoadMoveSet); 297 298 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 299 std::multimap<Value *, Value *> &LoadMoveSet, 300 Instruction *I, Instruction *J); 301 302 void moveUsesOfIAfterJ(BasicBlock &BB, 303 std::multimap<Value *, Value *> &LoadMoveSet, 304 Instruction *&InsertionPt, 305 Instruction *I, Instruction *J); 306 307 void combineMetadata(Instruction *K, const Instruction *J); 308 309 bool vectorizeBB(BasicBlock &BB) { 310 bool changed = false; 311 // Iterate a sufficient number of times to merge types of size 1 bit, 312 // then 2 bits, then 4, etc. up to half of the target vector width of the 313 // target vector register. 314 for (unsigned v = 2, n = 1; 315 v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter); 316 v *= 2, ++n) { 317 DEBUG(dbgs() << "BBV: fusing loop #" << n << 318 " for " << BB.getName() << " in " << 319 BB.getParent()->getName() << "...\n"); 320 if (vectorizePairs(BB)) 321 changed = true; 322 else 323 break; 324 } 325 326 DEBUG(dbgs() << "BBV: done!\n"); 327 return changed; 328 } 329 330 virtual bool runOnBasicBlock(BasicBlock &BB) { 331 AA = &getAnalysis<AliasAnalysis>(); 332 SE = &getAnalysis<ScalarEvolution>(); 333 TD = getAnalysisIfAvailable<TargetData>(); 334 335 return vectorizeBB(BB); 336 } 337 338 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 339 BasicBlockPass::getAnalysisUsage(AU); 340 AU.addRequired<AliasAnalysis>(); 341 AU.addRequired<ScalarEvolution>(); 342 AU.addPreserved<AliasAnalysis>(); 343 AU.addPreserved<ScalarEvolution>(); 344 AU.setPreservesCFG(); 345 } 346 347 // This returns the vector type that holds a pair of the provided type. 348 // If the provided type is already a vector, then its length is doubled. 349 static inline VectorType *getVecTypeForPair(Type *ElemTy) { 350 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 351 unsigned numElem = VTy->getNumElements(); 352 return VectorType::get(ElemTy->getScalarType(), numElem*2); 353 } 354 355 return VectorType::get(ElemTy, 2); 356 } 357 358 // Returns the weight associated with the provided value. A chain of 359 // candidate pairs has a length given by the sum of the weights of its 360 // members (one weight per pair; the weight of each member of the pair 361 // is assumed to be the same). This length is then compared to the 362 // chain-length threshold to determine if a given chain is significant 363 // enough to be vectorized. The length is also used in comparing 364 // candidate chains where longer chains are considered to be better. 365 // Note: when this function returns 0, the resulting instructions are 366 // not actually fused. 367 inline size_t getDepthFactor(Value *V) { 368 // InsertElement and ExtractElement have a depth factor of zero. This is 369 // for two reasons: First, they cannot be usefully fused. Second, because 370 // the pass generates a lot of these, they can confuse the simple metric 371 // used to compare the trees in the next iteration. Thus, giving them a 372 // weight of zero allows the pass to essentially ignore them in 373 // subsequent iterations when looking for vectorization opportunities 374 // while still tracking dependency chains that flow through those 375 // instructions. 376 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 377 return 0; 378 379 // Give a load or store half of the required depth so that load/store 380 // pairs will vectorize. 381 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 382 return Config.ReqChainDepth/2; 383 384 return 1; 385 } 386 387 // This determines the relative offset of two loads or stores, returning 388 // true if the offset could be determined to be some constant value. 389 // For example, if OffsetInElmts == 1, then J accesses the memory directly 390 // after I; if OffsetInElmts == -1 then I accesses the memory 391 // directly after J. This function assumes that both instructions 392 // have the same type. 393 bool getPairPtrInfo(Instruction *I, Instruction *J, 394 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 395 int64_t &OffsetInElmts) { 396 OffsetInElmts = 0; 397 if (isa<LoadInst>(I)) { 398 IPtr = cast<LoadInst>(I)->getPointerOperand(); 399 JPtr = cast<LoadInst>(J)->getPointerOperand(); 400 IAlignment = cast<LoadInst>(I)->getAlignment(); 401 JAlignment = cast<LoadInst>(J)->getAlignment(); 402 } else { 403 IPtr = cast<StoreInst>(I)->getPointerOperand(); 404 JPtr = cast<StoreInst>(J)->getPointerOperand(); 405 IAlignment = cast<StoreInst>(I)->getAlignment(); 406 JAlignment = cast<StoreInst>(J)->getAlignment(); 407 } 408 409 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 410 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 411 412 // If this is a trivial offset, then we'll get something like 413 // 1*sizeof(type). With target data, which we need anyway, this will get 414 // constant folded into a number. 415 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 416 if (const SCEVConstant *ConstOffSCEV = 417 dyn_cast<SCEVConstant>(OffsetSCEV)) { 418 ConstantInt *IntOff = ConstOffSCEV->getValue(); 419 int64_t Offset = IntOff->getSExtValue(); 420 421 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 422 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 423 424 assert(VTy == cast<PointerType>(JPtr->getType())->getElementType()); 425 426 OffsetInElmts = Offset/VTyTSS; 427 return (abs64(Offset) % VTyTSS) == 0; 428 } 429 430 return false; 431 } 432 433 // Returns true if the provided CallInst represents an intrinsic that can 434 // be vectorized. 435 bool isVectorizableIntrinsic(CallInst* I) { 436 Function *F = I->getCalledFunction(); 437 if (!F) return false; 438 439 unsigned IID = F->getIntrinsicID(); 440 if (!IID) return false; 441 442 switch(IID) { 443 default: 444 return false; 445 case Intrinsic::sqrt: 446 case Intrinsic::powi: 447 case Intrinsic::sin: 448 case Intrinsic::cos: 449 case Intrinsic::log: 450 case Intrinsic::log2: 451 case Intrinsic::log10: 452 case Intrinsic::exp: 453 case Intrinsic::exp2: 454 case Intrinsic::pow: 455 return Config.VectorizeMath; 456 case Intrinsic::fma: 457 return Config.VectorizeFMA; 458 } 459 } 460 461 // Returns true if J is the second element in some pair referenced by 462 // some multimap pair iterator pair. 463 template <typename V> 464 bool isSecondInIteratorPair(V J, std::pair< 465 typename std::multimap<V, V>::iterator, 466 typename std::multimap<V, V>::iterator> PairRange) { 467 for (typename std::multimap<V, V>::iterator K = PairRange.first; 468 K != PairRange.second; ++K) 469 if (K->second == J) return true; 470 471 return false; 472 } 473 }; 474 475 // This function implements one vectorization iteration on the provided 476 // basic block. It returns true if the block is changed. 477 bool BBVectorize::vectorizePairs(BasicBlock &BB) { 478 bool ShouldContinue; 479 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 480 481 std::vector<Value *> AllPairableInsts; 482 DenseMap<Value *, Value *> AllChosenPairs; 483 484 do { 485 std::vector<Value *> PairableInsts; 486 std::multimap<Value *, Value *> CandidatePairs; 487 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 488 PairableInsts); 489 if (PairableInsts.empty()) continue; 490 491 // Now we have a map of all of the pairable instructions and we need to 492 // select the best possible pairing. A good pairing is one such that the 493 // users of the pair are also paired. This defines a (directed) forest 494 // over the pairs such that two pairs are connected iff the second pair 495 // uses the first. 496 497 // Note that it only matters that both members of the second pair use some 498 // element of the first pair (to allow for splatting). 499 500 std::multimap<ValuePair, ValuePair> ConnectedPairs; 501 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 502 if (ConnectedPairs.empty()) continue; 503 504 // Build the pairable-instruction dependency map 505 DenseSet<ValuePair> PairableInstUsers; 506 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 507 508 // There is now a graph of the connected pairs. For each variable, pick 509 // the pairing with the largest tree meeting the depth requirement on at 510 // least one branch. Then select all pairings that are part of that tree 511 // and remove them from the list of available pairings and pairable 512 // variables. 513 514 DenseMap<Value *, Value *> ChosenPairs; 515 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, 516 PairableInstUsers, ChosenPairs); 517 518 if (ChosenPairs.empty()) continue; 519 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 520 PairableInsts.end()); 521 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 522 } while (ShouldContinue); 523 524 if (AllChosenPairs.empty()) return false; 525 NumFusedOps += AllChosenPairs.size(); 526 527 // A set of pairs has now been selected. It is now necessary to replace the 528 // paired instructions with vector instructions. For this procedure each 529 // operand must be replaced with a vector operand. This vector is formed 530 // by using build_vector on the old operands. The replaced values are then 531 // replaced with a vector_extract on the result. Subsequent optimization 532 // passes should coalesce the build/extract combinations. 533 534 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 535 return true; 536 } 537 538 // This function returns true if the provided instruction is capable of being 539 // fused into a vector instruction. This determination is based only on the 540 // type and other attributes of the instruction. 541 bool BBVectorize::isInstVectorizable(Instruction *I, 542 bool &IsSimpleLoadStore) { 543 IsSimpleLoadStore = false; 544 545 if (CallInst *C = dyn_cast<CallInst>(I)) { 546 if (!isVectorizableIntrinsic(C)) 547 return false; 548 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 549 // Vectorize simple loads if possbile: 550 IsSimpleLoadStore = L->isSimple(); 551 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 552 return false; 553 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 554 // Vectorize simple stores if possbile: 555 IsSimpleLoadStore = S->isSimple(); 556 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 557 return false; 558 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 559 // We can vectorize casts, but not casts of pointer types, etc. 560 if (!Config.VectorizeCasts) 561 return false; 562 563 Type *SrcTy = C->getSrcTy(); 564 if (!SrcTy->isSingleValueType()) 565 return false; 566 567 Type *DestTy = C->getDestTy(); 568 if (!DestTy->isSingleValueType()) 569 return false; 570 } else if (isa<SelectInst>(I)) { 571 if (!Config.VectorizeSelect) 572 return false; 573 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 574 if (!Config.VectorizeGEP) 575 return false; 576 577 // Currently, vector GEPs exist only with one index. 578 if (G->getNumIndices() != 1) 579 return false; 580 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 581 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 582 return false; 583 } 584 585 // We can't vectorize memory operations without target data 586 if (TD == 0 && IsSimpleLoadStore) 587 return false; 588 589 Type *T1, *T2; 590 if (isa<StoreInst>(I)) { 591 // For stores, it is the value type, not the pointer type that matters 592 // because the value is what will come from a vector register. 593 594 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 595 T1 = IVal->getType(); 596 } else { 597 T1 = I->getType(); 598 } 599 600 if (I->isCast()) 601 T2 = cast<CastInst>(I)->getSrcTy(); 602 else 603 T2 = T1; 604 605 // Not every type can be vectorized... 606 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 607 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 608 return false; 609 610 if (!Config.VectorizeInts 611 && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) 612 return false; 613 614 if (!Config.VectorizeFloats 615 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 616 return false; 617 618 // Don't vectorize target-specific types. 619 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 620 return false; 621 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 622 return false; 623 624 if ((!Config.VectorizePointers || TD == 0) && 625 (T1->getScalarType()->isPointerTy() || 626 T2->getScalarType()->isPointerTy())) 627 return false; 628 629 if (T1->getPrimitiveSizeInBits() > Config.VectorBits/2 || 630 T2->getPrimitiveSizeInBits() > Config.VectorBits/2) 631 return false; 632 633 return true; 634 } 635 636 // This function returns true if the two provided instructions are compatible 637 // (meaning that they can be fused into a vector instruction). This assumes 638 // that I has already been determined to be vectorizable and that J is not 639 // in the use tree of I. 640 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 641 bool IsSimpleLoadStore) { 642 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 643 " <-> " << *J << "\n"); 644 645 // Loads and stores can be merged if they have different alignments, 646 // but are otherwise the same. 647 LoadInst *LI, *LJ; 648 StoreInst *SI, *SJ; 649 if ((LI = dyn_cast<LoadInst>(I)) && (LJ = dyn_cast<LoadInst>(J))) { 650 if (I->getType() != J->getType()) 651 return false; 652 653 if (LI->getPointerOperand()->getType() != 654 LJ->getPointerOperand()->getType() || 655 LI->isVolatile() != LJ->isVolatile() || 656 LI->getOrdering() != LJ->getOrdering() || 657 LI->getSynchScope() != LJ->getSynchScope()) 658 return false; 659 } else if ((SI = dyn_cast<StoreInst>(I)) && (SJ = dyn_cast<StoreInst>(J))) { 660 if (SI->getValueOperand()->getType() != 661 SJ->getValueOperand()->getType() || 662 SI->getPointerOperand()->getType() != 663 SJ->getPointerOperand()->getType() || 664 SI->isVolatile() != SJ->isVolatile() || 665 SI->getOrdering() != SJ->getOrdering() || 666 SI->getSynchScope() != SJ->getSynchScope()) 667 return false; 668 } else if (!J->isSameOperationAs(I)) { 669 return false; 670 } 671 // FIXME: handle addsub-type operations! 672 673 if (IsSimpleLoadStore) { 674 Value *IPtr, *JPtr; 675 unsigned IAlignment, JAlignment; 676 int64_t OffsetInElmts = 0; 677 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 678 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 679 if (Config.AlignedOnly) { 680 Type *aType = isa<StoreInst>(I) ? 681 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 682 // An aligned load or store is possible only if the instruction 683 // with the lower offset has an alignment suitable for the 684 // vector type. 685 686 unsigned BottomAlignment = IAlignment; 687 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 688 689 Type *VType = getVecTypeForPair(aType); 690 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 691 if (BottomAlignment < VecAlignment) 692 return false; 693 } 694 } else { 695 return false; 696 } 697 } else if (isa<ShuffleVectorInst>(I)) { 698 // Only merge two shuffles if they're both constant 699 return isa<Constant>(I->getOperand(2)) && 700 isa<Constant>(J->getOperand(2)); 701 // FIXME: We may want to vectorize non-constant shuffles also. 702 } 703 704 // The powi intrinsic is special because only the first argument is 705 // vectorized, the second arguments must be equal. 706 CallInst *CI = dyn_cast<CallInst>(I); 707 Function *FI; 708 if (CI && (FI = CI->getCalledFunction()) && 709 FI->getIntrinsicID() == Intrinsic::powi) { 710 711 Value *A1I = CI->getArgOperand(1), 712 *A1J = cast<CallInst>(J)->getArgOperand(1); 713 const SCEV *A1ISCEV = SE->getSCEV(A1I), 714 *A1JSCEV = SE->getSCEV(A1J); 715 return (A1ISCEV == A1JSCEV); 716 } 717 718 return true; 719 } 720 721 // Figure out whether or not J uses I and update the users and write-set 722 // structures associated with I. Specifically, Users represents the set of 723 // instructions that depend on I. WriteSet represents the set 724 // of memory locations that are dependent on I. If UpdateUsers is true, 725 // and J uses I, then Users is updated to contain J and WriteSet is updated 726 // to contain any memory locations to which J writes. The function returns 727 // true if J uses I. By default, alias analysis is used to determine 728 // whether J reads from memory that overlaps with a location in WriteSet. 729 // If LoadMoveSet is not null, then it is a previously-computed multimap 730 // where the key is the memory-based user instruction and the value is 731 // the instruction to be compared with I. So, if LoadMoveSet is provided, 732 // then the alias analysis is not used. This is necessary because this 733 // function is called during the process of moving instructions during 734 // vectorization and the results of the alias analysis are not stable during 735 // that process. 736 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 737 AliasSetTracker &WriteSet, Instruction *I, 738 Instruction *J, bool UpdateUsers, 739 std::multimap<Value *, Value *> *LoadMoveSet) { 740 bool UsesI = false; 741 742 // This instruction may already be marked as a user due, for example, to 743 // being a member of a selected pair. 744 if (Users.count(J)) 745 UsesI = true; 746 747 if (!UsesI) 748 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 749 JU != JE; ++JU) { 750 Value *V = *JU; 751 if (I == V || Users.count(V)) { 752 UsesI = true; 753 break; 754 } 755 } 756 if (!UsesI && J->mayReadFromMemory()) { 757 if (LoadMoveSet) { 758 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 759 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 760 } else { 761 for (AliasSetTracker::iterator W = WriteSet.begin(), 762 WE = WriteSet.end(); W != WE; ++W) { 763 if (W->aliasesUnknownInst(J, *AA)) { 764 UsesI = true; 765 break; 766 } 767 } 768 } 769 } 770 771 if (UsesI && UpdateUsers) { 772 if (J->mayWriteToMemory()) WriteSet.add(J); 773 Users.insert(J); 774 } 775 776 return UsesI; 777 } 778 779 // This function iterates over all instruction pairs in the provided 780 // basic block and collects all candidate pairs for vectorization. 781 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 782 BasicBlock::iterator &Start, 783 std::multimap<Value *, Value *> &CandidatePairs, 784 std::vector<Value *> &PairableInsts) { 785 BasicBlock::iterator E = BB.end(); 786 if (Start == E) return false; 787 788 bool ShouldContinue = false, IAfterStart = false; 789 for (BasicBlock::iterator I = Start++; I != E; ++I) { 790 if (I == Start) IAfterStart = true; 791 792 bool IsSimpleLoadStore; 793 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 794 795 // Look for an instruction with which to pair instruction *I... 796 DenseSet<Value *> Users; 797 AliasSetTracker WriteSet(*AA); 798 bool JAfterStart = IAfterStart; 799 BasicBlock::iterator J = llvm::next(I); 800 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 801 if (J == Start) JAfterStart = true; 802 803 // Determine if J uses I, if so, exit the loop. 804 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 805 if (Config.FastDep) { 806 // Note: For this heuristic to be effective, independent operations 807 // must tend to be intermixed. This is likely to be true from some 808 // kinds of grouped loop unrolling (but not the generic LLVM pass), 809 // but otherwise may require some kind of reordering pass. 810 811 // When using fast dependency analysis, 812 // stop searching after first use: 813 if (UsesI) break; 814 } else { 815 if (UsesI) continue; 816 } 817 818 // J does not use I, and comes before the first use of I, so it can be 819 // merged with I if the instructions are compatible. 820 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue; 821 822 // J is a candidate for merging with I. 823 if (!PairableInsts.size() || 824 PairableInsts[PairableInsts.size()-1] != I) { 825 PairableInsts.push_back(I); 826 } 827 828 CandidatePairs.insert(ValuePair(I, J)); 829 830 // The next call to this function must start after the last instruction 831 // selected during this invocation. 832 if (JAfterStart) { 833 Start = llvm::next(J); 834 IAfterStart = JAfterStart = false; 835 } 836 837 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 838 << *I << " <-> " << *J << "\n"); 839 840 // If we have already found too many pairs, break here and this function 841 // will be called again starting after the last instruction selected 842 // during this invocation. 843 if (PairableInsts.size() >= Config.MaxInsts) { 844 ShouldContinue = true; 845 break; 846 } 847 } 848 849 if (ShouldContinue) 850 break; 851 } 852 853 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 854 << " instructions with candidate pairs\n"); 855 856 return ShouldContinue; 857 } 858 859 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 860 // it looks for pairs such that both members have an input which is an 861 // output of PI or PJ. 862 void BBVectorize::computePairsConnectedTo( 863 std::multimap<Value *, Value *> &CandidatePairs, 864 std::vector<Value *> &PairableInsts, 865 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 866 ValuePair P) { 867 StoreInst *SI, *SJ; 868 869 // For each possible pairing for this variable, look at the uses of 870 // the first value... 871 for (Value::use_iterator I = P.first->use_begin(), 872 E = P.first->use_end(); I != E; ++I) { 873 if (isa<LoadInst>(*I)) { 874 // A pair cannot be connected to a load because the load only takes one 875 // operand (the address) and it is a scalar even after vectorization. 876 continue; 877 } else if ((SI = dyn_cast<StoreInst>(*I)) && 878 P.first == SI->getPointerOperand()) { 879 // Similarly, a pair cannot be connected to a store through its 880 // pointer operand. 881 continue; 882 } 883 884 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 885 886 // For each use of the first variable, look for uses of the second 887 // variable... 888 for (Value::use_iterator J = P.second->use_begin(), 889 E2 = P.second->use_end(); J != E2; ++J) { 890 if ((SJ = dyn_cast<StoreInst>(*J)) && 891 P.second == SJ->getPointerOperand()) 892 continue; 893 894 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 895 896 // Look for <I, J>: 897 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 898 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 899 900 // Look for <J, I>: 901 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 902 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 903 } 904 905 if (Config.SplatBreaksChain) continue; 906 // Look for cases where just the first value in the pair is used by 907 // both members of another pair (splatting). 908 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 909 if ((SJ = dyn_cast<StoreInst>(*J)) && 910 P.first == SJ->getPointerOperand()) 911 continue; 912 913 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 914 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 915 } 916 } 917 918 if (Config.SplatBreaksChain) return; 919 // Look for cases where just the second value in the pair is used by 920 // both members of another pair (splatting). 921 for (Value::use_iterator I = P.second->use_begin(), 922 E = P.second->use_end(); I != E; ++I) { 923 if (isa<LoadInst>(*I)) 924 continue; 925 else if ((SI = dyn_cast<StoreInst>(*I)) && 926 P.second == SI->getPointerOperand()) 927 continue; 928 929 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 930 931 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 932 if ((SJ = dyn_cast<StoreInst>(*J)) && 933 P.second == SJ->getPointerOperand()) 934 continue; 935 936 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 937 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 938 } 939 } 940 } 941 942 // This function figures out which pairs are connected. Two pairs are 943 // connected if some output of the first pair forms an input to both members 944 // of the second pair. 945 void BBVectorize::computeConnectedPairs( 946 std::multimap<Value *, Value *> &CandidatePairs, 947 std::vector<Value *> &PairableInsts, 948 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 949 950 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 951 PE = PairableInsts.end(); PI != PE; ++PI) { 952 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 953 954 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 955 P != choiceRange.second; ++P) 956 computePairsConnectedTo(CandidatePairs, PairableInsts, 957 ConnectedPairs, *P); 958 } 959 960 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 961 << " pair connections.\n"); 962 } 963 964 // This function builds a set of use tuples such that <A, B> is in the set 965 // if B is in the use tree of A. If B is in the use tree of A, then B 966 // depends on the output of A. 967 void BBVectorize::buildDepMap( 968 BasicBlock &BB, 969 std::multimap<Value *, Value *> &CandidatePairs, 970 std::vector<Value *> &PairableInsts, 971 DenseSet<ValuePair> &PairableInstUsers) { 972 DenseSet<Value *> IsInPair; 973 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 974 E = CandidatePairs.end(); C != E; ++C) { 975 IsInPair.insert(C->first); 976 IsInPair.insert(C->second); 977 } 978 979 // Iterate through the basic block, recording all Users of each 980 // pairable instruction. 981 982 BasicBlock::iterator E = BB.end(); 983 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 984 if (IsInPair.find(I) == IsInPair.end()) continue; 985 986 DenseSet<Value *> Users; 987 AliasSetTracker WriteSet(*AA); 988 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 989 (void) trackUsesOfI(Users, WriteSet, I, J); 990 991 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 992 U != E; ++U) 993 PairableInstUsers.insert(ValuePair(I, *U)); 994 } 995 } 996 997 // Returns true if an input to pair P is an output of pair Q and also an 998 // input of pair Q is an output of pair P. If this is the case, then these 999 // two pairs cannot be simultaneously fused. 1000 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1001 DenseSet<ValuePair> &PairableInstUsers, 1002 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1003 // Two pairs are in conflict if they are mutual Users of eachother. 1004 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1005 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1006 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1007 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1008 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1009 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1010 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1011 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1012 if (PairableInstUserMap) { 1013 // FIXME: The expensive part of the cycle check is not so much the cycle 1014 // check itself but this edge insertion procedure. This needs some 1015 // profiling and probably a different data structure (same is true of 1016 // most uses of std::multimap). 1017 if (PUsesQ) { 1018 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1019 if (!isSecondInIteratorPair(P, QPairRange)) 1020 PairableInstUserMap->insert(VPPair(Q, P)); 1021 } 1022 if (QUsesP) { 1023 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1024 if (!isSecondInIteratorPair(Q, PPairRange)) 1025 PairableInstUserMap->insert(VPPair(P, Q)); 1026 } 1027 } 1028 1029 return (QUsesP && PUsesQ); 1030 } 1031 1032 // This function walks the use graph of current pairs to see if, starting 1033 // from P, the walk returns to P. 1034 bool BBVectorize::pairWillFormCycle(ValuePair P, 1035 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1036 DenseSet<ValuePair> &CurrentPairs) { 1037 DEBUG(if (DebugCycleCheck) 1038 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1039 << *P.second << "\n"); 1040 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1041 // contains non-direct associations. 1042 DenseSet<ValuePair> Visited; 1043 SmallVector<ValuePair, 32> Q; 1044 // General depth-first post-order traversal: 1045 Q.push_back(P); 1046 do { 1047 ValuePair QTop = Q.pop_back_val(); 1048 Visited.insert(QTop); 1049 1050 DEBUG(if (DebugCycleCheck) 1051 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1052 << *QTop.second << "\n"); 1053 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1054 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1055 C != QPairRange.second; ++C) { 1056 if (C->second == P) { 1057 DEBUG(dbgs() 1058 << "BBV: rejected to prevent non-trivial cycle formation: " 1059 << *C->first.first << " <-> " << *C->first.second << "\n"); 1060 return true; 1061 } 1062 1063 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1064 Q.push_back(C->second); 1065 } 1066 } while (!Q.empty()); 1067 1068 return false; 1069 } 1070 1071 // This function builds the initial tree of connected pairs with the 1072 // pair J at the root. 1073 void BBVectorize::buildInitialTreeFor( 1074 std::multimap<Value *, Value *> &CandidatePairs, 1075 std::vector<Value *> &PairableInsts, 1076 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1077 DenseSet<ValuePair> &PairableInstUsers, 1078 DenseMap<Value *, Value *> &ChosenPairs, 1079 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1080 // Each of these pairs is viewed as the root node of a Tree. The Tree 1081 // is then walked (depth-first). As this happens, we keep track of 1082 // the pairs that compose the Tree and the maximum depth of the Tree. 1083 SmallVector<ValuePairWithDepth, 32> Q; 1084 // General depth-first post-order traversal: 1085 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1086 do { 1087 ValuePairWithDepth QTop = Q.back(); 1088 1089 // Push each child onto the queue: 1090 bool MoreChildren = false; 1091 size_t MaxChildDepth = QTop.second; 1092 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1093 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1094 k != qtRange.second; ++k) { 1095 // Make sure that this child pair is still a candidate: 1096 bool IsStillCand = false; 1097 VPIteratorPair checkRange = 1098 CandidatePairs.equal_range(k->second.first); 1099 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1100 m != checkRange.second; ++m) { 1101 if (m->second == k->second.second) { 1102 IsStillCand = true; 1103 break; 1104 } 1105 } 1106 1107 if (IsStillCand) { 1108 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1109 if (C == Tree.end()) { 1110 size_t d = getDepthFactor(k->second.first); 1111 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1112 MoreChildren = true; 1113 } else { 1114 MaxChildDepth = std::max(MaxChildDepth, C->second); 1115 } 1116 } 1117 } 1118 1119 if (!MoreChildren) { 1120 // Record the current pair as part of the Tree: 1121 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1122 Q.pop_back(); 1123 } 1124 } while (!Q.empty()); 1125 } 1126 1127 // Given some initial tree, prune it by removing conflicting pairs (pairs 1128 // that cannot be simultaneously chosen for vectorization). 1129 void BBVectorize::pruneTreeFor( 1130 std::multimap<Value *, Value *> &CandidatePairs, 1131 std::vector<Value *> &PairableInsts, 1132 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1133 DenseSet<ValuePair> &PairableInstUsers, 1134 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1135 DenseMap<Value *, Value *> &ChosenPairs, 1136 DenseMap<ValuePair, size_t> &Tree, 1137 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1138 bool UseCycleCheck) { 1139 SmallVector<ValuePairWithDepth, 32> Q; 1140 // General depth-first post-order traversal: 1141 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1142 do { 1143 ValuePairWithDepth QTop = Q.pop_back_val(); 1144 PrunedTree.insert(QTop.first); 1145 1146 // Visit each child, pruning as necessary... 1147 DenseMap<ValuePair, size_t> BestChildren; 1148 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1149 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1150 K != QTopRange.second; ++K) { 1151 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1152 if (C == Tree.end()) continue; 1153 1154 // This child is in the Tree, now we need to make sure it is the 1155 // best of any conflicting children. There could be multiple 1156 // conflicting children, so first, determine if we're keeping 1157 // this child, then delete conflicting children as necessary. 1158 1159 // It is also necessary to guard against pairing-induced 1160 // dependencies. Consider instructions a .. x .. y .. b 1161 // such that (a,b) are to be fused and (x,y) are to be fused 1162 // but a is an input to x and b is an output from y. This 1163 // means that y cannot be moved after b but x must be moved 1164 // after b for (a,b) to be fused. In other words, after 1165 // fusing (a,b) we have y .. a/b .. x where y is an input 1166 // to a/b and x is an output to a/b: x and y can no longer 1167 // be legally fused. To prevent this condition, we must 1168 // make sure that a child pair added to the Tree is not 1169 // both an input and output of an already-selected pair. 1170 1171 // Pairing-induced dependencies can also form from more complicated 1172 // cycles. The pair vs. pair conflicts are easy to check, and so 1173 // that is done explicitly for "fast rejection", and because for 1174 // child vs. child conflicts, we may prefer to keep the current 1175 // pair in preference to the already-selected child. 1176 DenseSet<ValuePair> CurrentPairs; 1177 1178 bool CanAdd = true; 1179 for (DenseMap<ValuePair, size_t>::iterator C2 1180 = BestChildren.begin(), E2 = BestChildren.end(); 1181 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 if (C2->second >= C->second) { 1189 CanAdd = false; 1190 break; 1191 } 1192 1193 CurrentPairs.insert(C2->first); 1194 } 1195 } 1196 if (!CanAdd) continue; 1197 1198 // Even worse, this child could conflict with another node already 1199 // selected for the Tree. If that is the case, ignore this child. 1200 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1201 E2 = PrunedTree.end(); T != E2; ++T) { 1202 if (T->first == C->first.first || 1203 T->first == C->first.second || 1204 T->second == C->first.first || 1205 T->second == C->first.second || 1206 pairsConflict(*T, C->first, PairableInstUsers, 1207 UseCycleCheck ? &PairableInstUserMap : 0)) { 1208 CanAdd = false; 1209 break; 1210 } 1211 1212 CurrentPairs.insert(*T); 1213 } 1214 if (!CanAdd) continue; 1215 1216 // And check the queue too... 1217 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1218 E2 = Q.end(); C2 != E2; ++C2) { 1219 if (C2->first.first == C->first.first || 1220 C2->first.first == C->first.second || 1221 C2->first.second == C->first.first || 1222 C2->first.second == C->first.second || 1223 pairsConflict(C2->first, C->first, PairableInstUsers, 1224 UseCycleCheck ? &PairableInstUserMap : 0)) { 1225 CanAdd = false; 1226 break; 1227 } 1228 1229 CurrentPairs.insert(C2->first); 1230 } 1231 if (!CanAdd) continue; 1232 1233 // Last but not least, check for a conflict with any of the 1234 // already-chosen pairs. 1235 for (DenseMap<Value *, Value *>::iterator C2 = 1236 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1237 C2 != E2; ++C2) { 1238 if (pairsConflict(*C2, C->first, PairableInstUsers, 1239 UseCycleCheck ? &PairableInstUserMap : 0)) { 1240 CanAdd = false; 1241 break; 1242 } 1243 1244 CurrentPairs.insert(*C2); 1245 } 1246 if (!CanAdd) continue; 1247 1248 // To check for non-trivial cycles formed by the addition of the 1249 // current pair we've formed a list of all relevant pairs, now use a 1250 // graph walk to check for a cycle. We start from the current pair and 1251 // walk the use tree to see if we again reach the current pair. If we 1252 // do, then the current pair is rejected. 1253 1254 // FIXME: It may be more efficient to use a topological-ordering 1255 // algorithm to improve the cycle check. This should be investigated. 1256 if (UseCycleCheck && 1257 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1258 continue; 1259 1260 // This child can be added, but we may have chosen it in preference 1261 // to an already-selected child. Check for this here, and if a 1262 // conflict is found, then remove the previously-selected child 1263 // before adding this one in its place. 1264 for (DenseMap<ValuePair, size_t>::iterator C2 1265 = BestChildren.begin(); C2 != BestChildren.end();) { 1266 if (C2->first.first == C->first.first || 1267 C2->first.first == C->first.second || 1268 C2->first.second == C->first.first || 1269 C2->first.second == C->first.second || 1270 pairsConflict(C2->first, C->first, PairableInstUsers)) 1271 BestChildren.erase(C2++); 1272 else 1273 ++C2; 1274 } 1275 1276 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1277 } 1278 1279 for (DenseMap<ValuePair, size_t>::iterator C 1280 = BestChildren.begin(), E2 = BestChildren.end(); 1281 C != E2; ++C) { 1282 size_t DepthF = getDepthFactor(C->first.first); 1283 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1284 } 1285 } while (!Q.empty()); 1286 } 1287 1288 // This function finds the best tree of mututally-compatible connected 1289 // pairs, given the choice of root pairs as an iterator range. 1290 void BBVectorize::findBestTreeFor( 1291 std::multimap<Value *, Value *> &CandidatePairs, 1292 std::vector<Value *> &PairableInsts, 1293 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1294 DenseSet<ValuePair> &PairableInstUsers, 1295 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1296 DenseMap<Value *, Value *> &ChosenPairs, 1297 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1298 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1299 bool UseCycleCheck) { 1300 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1301 J != ChoiceRange.second; ++J) { 1302 1303 // Before going any further, make sure that this pair does not 1304 // conflict with any already-selected pairs (see comment below 1305 // near the Tree pruning for more details). 1306 DenseSet<ValuePair> ChosenPairSet; 1307 bool DoesConflict = false; 1308 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1309 E = ChosenPairs.end(); C != E; ++C) { 1310 if (pairsConflict(*C, *J, PairableInstUsers, 1311 UseCycleCheck ? &PairableInstUserMap : 0)) { 1312 DoesConflict = true; 1313 break; 1314 } 1315 1316 ChosenPairSet.insert(*C); 1317 } 1318 if (DoesConflict) continue; 1319 1320 if (UseCycleCheck && 1321 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1322 continue; 1323 1324 DenseMap<ValuePair, size_t> Tree; 1325 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1326 PairableInstUsers, ChosenPairs, Tree, *J); 1327 1328 // Because we'll keep the child with the largest depth, the largest 1329 // depth is still the same in the unpruned Tree. 1330 size_t MaxDepth = Tree.lookup(*J); 1331 1332 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1333 << *J->first << " <-> " << *J->second << "} of depth " << 1334 MaxDepth << " and size " << Tree.size() << "\n"); 1335 1336 // At this point the Tree has been constructed, but, may contain 1337 // contradictory children (meaning that different children of 1338 // some tree node may be attempting to fuse the same instruction). 1339 // So now we walk the tree again, in the case of a conflict, 1340 // keep only the child with the largest depth. To break a tie, 1341 // favor the first child. 1342 1343 DenseSet<ValuePair> PrunedTree; 1344 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1345 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1346 PrunedTree, *J, UseCycleCheck); 1347 1348 size_t EffSize = 0; 1349 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1350 E = PrunedTree.end(); S != E; ++S) 1351 EffSize += getDepthFactor(S->first); 1352 1353 DEBUG(if (DebugPairSelection) 1354 dbgs() << "BBV: found pruned Tree for pair {" 1355 << *J->first << " <-> " << *J->second << "} of depth " << 1356 MaxDepth << " and size " << PrunedTree.size() << 1357 " (effective size: " << EffSize << ")\n"); 1358 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) { 1359 BestMaxDepth = MaxDepth; 1360 BestEffSize = EffSize; 1361 BestTree = PrunedTree; 1362 } 1363 } 1364 } 1365 1366 // Given the list of candidate pairs, this function selects those 1367 // that will be fused into vector instructions. 1368 void BBVectorize::choosePairs( 1369 std::multimap<Value *, Value *> &CandidatePairs, 1370 std::vector<Value *> &PairableInsts, 1371 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1372 DenseSet<ValuePair> &PairableInstUsers, 1373 DenseMap<Value *, Value *>& ChosenPairs) { 1374 bool UseCycleCheck = 1375 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1376 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1377 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1378 E = PairableInsts.end(); I != E; ++I) { 1379 // The number of possible pairings for this variable: 1380 size_t NumChoices = CandidatePairs.count(*I); 1381 if (!NumChoices) continue; 1382 1383 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1384 1385 // The best pair to choose and its tree: 1386 size_t BestMaxDepth = 0, BestEffSize = 0; 1387 DenseSet<ValuePair> BestTree; 1388 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1389 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1390 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1391 UseCycleCheck); 1392 1393 // A tree has been chosen (or not) at this point. If no tree was 1394 // chosen, then this instruction, I, cannot be paired (and is no longer 1395 // considered). 1396 1397 DEBUG(if (BestTree.size() > 0) 1398 dbgs() << "BBV: selected pairs in the best tree for: " 1399 << *cast<Instruction>(*I) << "\n"); 1400 1401 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1402 SE2 = BestTree.end(); S != SE2; ++S) { 1403 // Insert the members of this tree into the list of chosen pairs. 1404 ChosenPairs.insert(ValuePair(S->first, S->second)); 1405 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1406 *S->second << "\n"); 1407 1408 // Remove all candidate pairs that have values in the chosen tree. 1409 for (std::multimap<Value *, Value *>::iterator K = 1410 CandidatePairs.begin(); K != CandidatePairs.end();) { 1411 if (K->first == S->first || K->second == S->first || 1412 K->second == S->second || K->first == S->second) { 1413 // Don't remove the actual pair chosen so that it can be used 1414 // in subsequent tree selections. 1415 if (!(K->first == S->first && K->second == S->second)) 1416 CandidatePairs.erase(K++); 1417 else 1418 ++K; 1419 } else { 1420 ++K; 1421 } 1422 } 1423 } 1424 } 1425 1426 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1427 } 1428 1429 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1430 unsigned n = 0) { 1431 if (!I->hasName()) 1432 return ""; 1433 1434 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1435 (n > 0 ? "." + utostr(n) : "")).str(); 1436 } 1437 1438 // Returns the value that is to be used as the pointer input to the vector 1439 // instruction that fuses I with J. 1440 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1441 Instruction *I, Instruction *J, unsigned o, 1442 bool &FlipMemInputs) { 1443 Value *IPtr, *JPtr; 1444 unsigned IAlignment, JAlignment; 1445 int64_t OffsetInElmts; 1446 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1447 OffsetInElmts); 1448 1449 // The pointer value is taken to be the one with the lowest offset. 1450 Value *VPtr; 1451 if (OffsetInElmts > 0) { 1452 VPtr = IPtr; 1453 } else { 1454 FlipMemInputs = true; 1455 VPtr = JPtr; 1456 } 1457 1458 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType(); 1459 Type *VArgType = getVecTypeForPair(ArgType); 1460 Type *VArgPtrType = PointerType::get(VArgType, 1461 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1462 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1463 /* insert before */ FlipMemInputs ? J : I); 1464 } 1465 1466 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1467 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 1468 unsigned IdxOffset, std::vector<Constant*> &Mask) { 1469 for (unsigned v = 0; v < NumElem/2; ++v) { 1470 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1471 if (m < 0) { 1472 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1473 } else { 1474 unsigned mm = m + (int) IdxOffset; 1475 if (m >= (int) NumInElem) 1476 mm += (int) NumInElem; 1477 1478 Mask[v+MaskOffset] = 1479 ConstantInt::get(Type::getInt32Ty(Context), mm); 1480 } 1481 } 1482 } 1483 1484 // Returns the value that is to be used as the vector-shuffle mask to the 1485 // vector instruction that fuses I with J. 1486 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1487 Instruction *I, Instruction *J) { 1488 // This is the shuffle mask. We need to append the second 1489 // mask to the first, and the numbers need to be adjusted. 1490 1491 Type *ArgType = I->getType(); 1492 Type *VArgType = getVecTypeForPair(ArgType); 1493 1494 // Get the total number of elements in the fused vector type. 1495 // By definition, this must equal the number of elements in 1496 // the final mask. 1497 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1498 std::vector<Constant*> Mask(NumElem); 1499 1500 Type *OpType = I->getOperand(0)->getType(); 1501 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements(); 1502 1503 // For the mask from the first pair... 1504 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask); 1505 1506 // For the mask from the second pair... 1507 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem, 1508 Mask); 1509 1510 return ConstantVector::get(Mask); 1511 } 1512 1513 // Returns the value to be used as the specified operand of the vector 1514 // instruction that fuses I with J. 1515 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1516 Instruction *J, unsigned o, bool FlipMemInputs) { 1517 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1518 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1519 1520 // Compute the fused vector type for this operand 1521 Type *ArgType = I->getOperand(o)->getType(); 1522 VectorType *VArgType = getVecTypeForPair(ArgType); 1523 1524 Instruction *L = I, *H = J; 1525 if (FlipMemInputs) { 1526 L = J; 1527 H = I; 1528 } 1529 1530 if (ArgType->isVectorTy()) { 1531 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1532 std::vector<Constant*> Mask(numElem); 1533 for (unsigned v = 0; v < numElem; ++v) 1534 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1535 1536 Instruction *BV = new ShuffleVectorInst(L->getOperand(o), 1537 H->getOperand(o), 1538 ConstantVector::get(Mask), 1539 getReplacementName(I, true, o)); 1540 BV->insertBefore(J); 1541 return BV; 1542 } 1543 1544 // If these two inputs are the output of another vector instruction, 1545 // then we should use that output directly. It might be necessary to 1546 // permute it first. [When pairings are fused recursively, you can 1547 // end up with cases where a large vector is decomposed into scalars 1548 // using extractelement instructions, then built into size-2 1549 // vectors using insertelement and the into larger vectors using 1550 // shuffles. InstCombine does not simplify all of these cases well, 1551 // and so we make sure that shuffles are generated here when possible. 1552 ExtractElementInst *LEE 1553 = dyn_cast<ExtractElementInst>(L->getOperand(o)); 1554 ExtractElementInst *HEE 1555 = dyn_cast<ExtractElementInst>(H->getOperand(o)); 1556 1557 if (LEE && HEE && 1558 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) { 1559 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType()); 1560 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue(); 1561 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue(); 1562 if (LEE->getOperand(0) == HEE->getOperand(0)) { 1563 if (LowIndx == 0 && HighIndx == 1) 1564 return LEE->getOperand(0); 1565 1566 std::vector<Constant*> Mask(2); 1567 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1568 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1569 1570 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1571 UndefValue::get(EEType), 1572 ConstantVector::get(Mask), 1573 getReplacementName(I, true, o)); 1574 BV->insertBefore(J); 1575 return BV; 1576 } 1577 1578 std::vector<Constant*> Mask(2); 1579 HighIndx += EEType->getNumElements(); 1580 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1581 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1582 1583 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1584 HEE->getOperand(0), 1585 ConstantVector::get(Mask), 1586 getReplacementName(I, true, o)); 1587 BV->insertBefore(J); 1588 return BV; 1589 } 1590 1591 Instruction *BV1 = InsertElementInst::Create( 1592 UndefValue::get(VArgType), 1593 L->getOperand(o), CV0, 1594 getReplacementName(I, true, o, 1)); 1595 BV1->insertBefore(I); 1596 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o), 1597 CV1, 1598 getReplacementName(I, true, o, 2)); 1599 BV2->insertBefore(J); 1600 return BV2; 1601 } 1602 1603 // This function creates an array of values that will be used as the inputs 1604 // to the vector instruction that fuses I with J. 1605 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1606 Instruction *I, Instruction *J, 1607 SmallVector<Value *, 3> &ReplacedOperands, 1608 bool &FlipMemInputs) { 1609 FlipMemInputs = false; 1610 unsigned NumOperands = I->getNumOperands(); 1611 1612 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1613 // Iterate backward so that we look at the store pointer 1614 // first and know whether or not we need to flip the inputs. 1615 1616 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1617 // This is the pointer for a load/store instruction. 1618 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 1619 FlipMemInputs); 1620 continue; 1621 } else if (isa<CallInst>(I)) { 1622 Function *F = cast<CallInst>(I)->getCalledFunction(); 1623 unsigned IID = F->getIntrinsicID(); 1624 if (o == NumOperands-1) { 1625 BasicBlock &BB = *I->getParent(); 1626 1627 Module *M = BB.getParent()->getParent(); 1628 Type *ArgType = I->getType(); 1629 Type *VArgType = getVecTypeForPair(ArgType); 1630 1631 // FIXME: is it safe to do this here? 1632 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 1633 (Intrinsic::ID) IID, VArgType); 1634 continue; 1635 } else if (IID == Intrinsic::powi && o == 1) { 1636 // The second argument of powi is a single integer and we've already 1637 // checked that both arguments are equal. As a result, we just keep 1638 // I's second argument. 1639 ReplacedOperands[o] = I->getOperand(o); 1640 continue; 1641 } 1642 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 1643 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 1644 continue; 1645 } 1646 1647 ReplacedOperands[o] = 1648 getReplacementInput(Context, I, J, o, FlipMemInputs); 1649 } 1650 } 1651 1652 // This function creates two values that represent the outputs of the 1653 // original I and J instructions. These are generally vector shuffles 1654 // or extracts. In many cases, these will end up being unused and, thus, 1655 // eliminated by later passes. 1656 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 1657 Instruction *J, Instruction *K, 1658 Instruction *&InsertionPt, 1659 Instruction *&K1, Instruction *&K2, 1660 bool &FlipMemInputs) { 1661 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1662 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1663 1664 if (isa<StoreInst>(I)) { 1665 AA->replaceWithNewValue(I, K); 1666 AA->replaceWithNewValue(J, K); 1667 } else { 1668 Type *IType = I->getType(); 1669 Type *VType = getVecTypeForPair(IType); 1670 1671 if (IType->isVectorTy()) { 1672 unsigned numElem = cast<VectorType>(IType)->getNumElements(); 1673 std::vector<Constant*> Mask1(numElem), Mask2(numElem); 1674 for (unsigned v = 0; v < numElem; ++v) { 1675 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1676 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v); 1677 } 1678 1679 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 1680 ConstantVector::get( 1681 FlipMemInputs ? Mask2 : Mask1), 1682 getReplacementName(K, false, 1)); 1683 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 1684 ConstantVector::get( 1685 FlipMemInputs ? Mask1 : Mask2), 1686 getReplacementName(K, false, 2)); 1687 } else { 1688 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 1689 getReplacementName(K, false, 1)); 1690 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 1691 getReplacementName(K, false, 2)); 1692 } 1693 1694 K1->insertAfter(K); 1695 K2->insertAfter(K1); 1696 InsertionPt = K2; 1697 } 1698 } 1699 1700 // Move all uses of the function I (including pairing-induced uses) after J. 1701 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 1702 std::multimap<Value *, Value *> &LoadMoveSet, 1703 Instruction *I, Instruction *J) { 1704 // Skip to the first instruction past I. 1705 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1706 1707 DenseSet<Value *> Users; 1708 AliasSetTracker WriteSet(*AA); 1709 for (; cast<Instruction>(L) != J; ++L) 1710 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 1711 1712 assert(cast<Instruction>(L) == J && 1713 "Tracking has not proceeded far enough to check for dependencies"); 1714 // If J is now in the use set of I, then trackUsesOfI will return true 1715 // and we have a dependency cycle (and the fusing operation must abort). 1716 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 1717 } 1718 1719 // Move all uses of the function I (including pairing-induced uses) after J. 1720 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 1721 std::multimap<Value *, Value *> &LoadMoveSet, 1722 Instruction *&InsertionPt, 1723 Instruction *I, Instruction *J) { 1724 // Skip to the first instruction past I. 1725 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1726 1727 DenseSet<Value *> Users; 1728 AliasSetTracker WriteSet(*AA); 1729 for (; cast<Instruction>(L) != J;) { 1730 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 1731 // Move this instruction 1732 Instruction *InstToMove = L; ++L; 1733 1734 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 1735 " to after " << *InsertionPt << "\n"); 1736 InstToMove->removeFromParent(); 1737 InstToMove->insertAfter(InsertionPt); 1738 InsertionPt = InstToMove; 1739 } else { 1740 ++L; 1741 } 1742 } 1743 } 1744 1745 // Collect all load instruction that are in the move set of a given first 1746 // pair member. These loads depend on the first instruction, I, and so need 1747 // to be moved after J (the second instruction) when the pair is fused. 1748 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 1749 DenseMap<Value *, Value *> &ChosenPairs, 1750 std::multimap<Value *, Value *> &LoadMoveSet, 1751 Instruction *I) { 1752 // Skip to the first instruction past I. 1753 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1754 1755 DenseSet<Value *> Users; 1756 AliasSetTracker WriteSet(*AA); 1757 1758 // Note: We cannot end the loop when we reach J because J could be moved 1759 // farther down the use chain by another instruction pairing. Also, J 1760 // could be before I if this is an inverted input. 1761 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 1762 if (trackUsesOfI(Users, WriteSet, I, L)) { 1763 if (L->mayReadFromMemory()) 1764 LoadMoveSet.insert(ValuePair(L, I)); 1765 } 1766 } 1767 } 1768 1769 // In cases where both load/stores and the computation of their pointers 1770 // are chosen for vectorization, we can end up in a situation where the 1771 // aliasing analysis starts returning different query results as the 1772 // process of fusing instruction pairs continues. Because the algorithm 1773 // relies on finding the same use trees here as were found earlier, we'll 1774 // need to precompute the necessary aliasing information here and then 1775 // manually update it during the fusion process. 1776 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 1777 std::vector<Value *> &PairableInsts, 1778 DenseMap<Value *, Value *> &ChosenPairs, 1779 std::multimap<Value *, Value *> &LoadMoveSet) { 1780 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1781 PIE = PairableInsts.end(); PI != PIE; ++PI) { 1782 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 1783 if (P == ChosenPairs.end()) continue; 1784 1785 Instruction *I = cast<Instruction>(P->first); 1786 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 1787 } 1788 } 1789 1790 // When the first instruction in each pair is cloned, it will inherit its 1791 // parent's metadata. This metadata must be combined with that of the other 1792 // instruction in a safe way. 1793 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 1794 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 1795 K->getAllMetadataOtherThanDebugLoc(Metadata); 1796 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 1797 unsigned Kind = Metadata[i].first; 1798 MDNode *JMD = J->getMetadata(Kind); 1799 MDNode *KMD = Metadata[i].second; 1800 1801 switch (Kind) { 1802 default: 1803 K->setMetadata(Kind, 0); // Remove unknown metadata 1804 break; 1805 case LLVMContext::MD_tbaa: 1806 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 1807 break; 1808 case LLVMContext::MD_fpmath: 1809 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 1810 break; 1811 } 1812 } 1813 } 1814 1815 // This function fuses the chosen instruction pairs into vector instructions, 1816 // taking care preserve any needed scalar outputs and, then, it reorders the 1817 // remaining instructions as needed (users of the first member of the pair 1818 // need to be moved to after the location of the second member of the pair 1819 // because the vector instruction is inserted in the location of the pair's 1820 // second member). 1821 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 1822 std::vector<Value *> &PairableInsts, 1823 DenseMap<Value *, Value *> &ChosenPairs) { 1824 LLVMContext& Context = BB.getContext(); 1825 1826 // During the vectorization process, the order of the pairs to be fused 1827 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 1828 // list. After a pair is fused, the flipped pair is removed from the list. 1829 std::vector<ValuePair> FlippedPairs; 1830 FlippedPairs.reserve(ChosenPairs.size()); 1831 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 1832 E = ChosenPairs.end(); P != E; ++P) 1833 FlippedPairs.push_back(ValuePair(P->second, P->first)); 1834 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 1835 E = FlippedPairs.end(); P != E; ++P) 1836 ChosenPairs.insert(*P); 1837 1838 std::multimap<Value *, Value *> LoadMoveSet; 1839 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 1840 1841 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 1842 1843 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 1844 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 1845 if (P == ChosenPairs.end()) { 1846 ++PI; 1847 continue; 1848 } 1849 1850 if (getDepthFactor(P->first) == 0) { 1851 // These instructions are not really fused, but are tracked as though 1852 // they are. Any case in which it would be interesting to fuse them 1853 // will be taken care of by InstCombine. 1854 --NumFusedOps; 1855 ++PI; 1856 continue; 1857 } 1858 1859 Instruction *I = cast<Instruction>(P->first), 1860 *J = cast<Instruction>(P->second); 1861 1862 DEBUG(dbgs() << "BBV: fusing: " << *I << 1863 " <-> " << *J << "\n"); 1864 1865 // Remove the pair and flipped pair from the list. 1866 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 1867 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 1868 ChosenPairs.erase(FP); 1869 ChosenPairs.erase(P); 1870 1871 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 1872 DEBUG(dbgs() << "BBV: fusion of: " << *I << 1873 " <-> " << *J << 1874 " aborted because of non-trivial dependency cycle\n"); 1875 --NumFusedOps; 1876 ++PI; 1877 continue; 1878 } 1879 1880 bool FlipMemInputs; 1881 unsigned NumOperands = I->getNumOperands(); 1882 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 1883 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 1884 FlipMemInputs); 1885 1886 // Make a copy of the original operation, change its type to the vector 1887 // type and replace its operands with the vector operands. 1888 Instruction *K = I->clone(); 1889 if (I->hasName()) K->takeName(I); 1890 1891 if (!isa<StoreInst>(K)) 1892 K->mutateType(getVecTypeForPair(I->getType())); 1893 1894 combineMetadata(K, J); 1895 1896 for (unsigned o = 0; o < NumOperands; ++o) 1897 K->setOperand(o, ReplacedOperands[o]); 1898 1899 // If we've flipped the memory inputs, make sure that we take the correct 1900 // alignment. 1901 if (FlipMemInputs) { 1902 if (isa<StoreInst>(K)) 1903 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 1904 else 1905 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 1906 } 1907 1908 K->insertAfter(J); 1909 1910 // Instruction insertion point: 1911 Instruction *InsertionPt = K; 1912 Instruction *K1 = 0, *K2 = 0; 1913 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 1914 FlipMemInputs); 1915 1916 // The use tree of the first original instruction must be moved to after 1917 // the location of the second instruction. The entire use tree of the 1918 // first instruction is disjoint from the input tree of the second 1919 // (by definition), and so commutes with it. 1920 1921 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 1922 1923 if (!isa<StoreInst>(I)) { 1924 I->replaceAllUsesWith(K1); 1925 J->replaceAllUsesWith(K2); 1926 AA->replaceWithNewValue(I, K1); 1927 AA->replaceWithNewValue(J, K2); 1928 } 1929 1930 // Instructions that may read from memory may be in the load move set. 1931 // Once an instruction is fused, we no longer need its move set, and so 1932 // the values of the map never need to be updated. However, when a load 1933 // is fused, we need to merge the entries from both instructions in the 1934 // pair in case those instructions were in the move set of some other 1935 // yet-to-be-fused pair. The loads in question are the keys of the map. 1936 if (I->mayReadFromMemory()) { 1937 std::vector<ValuePair> NewSetMembers; 1938 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 1939 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 1940 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 1941 N != IPairRange.second; ++N) 1942 NewSetMembers.push_back(ValuePair(K, N->second)); 1943 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 1944 N != JPairRange.second; ++N) 1945 NewSetMembers.push_back(ValuePair(K, N->second)); 1946 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 1947 AE = NewSetMembers.end(); A != AE; ++A) 1948 LoadMoveSet.insert(*A); 1949 } 1950 1951 // Before removing I, set the iterator to the next instruction. 1952 PI = llvm::next(BasicBlock::iterator(I)); 1953 if (cast<Instruction>(PI) == J) 1954 ++PI; 1955 1956 SE->forgetValue(I); 1957 SE->forgetValue(J); 1958 I->eraseFromParent(); 1959 J->eraseFromParent(); 1960 } 1961 1962 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 1963 } 1964} 1965 1966char BBVectorize::ID = 0; 1967static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 1968INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1969INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 1970INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 1971INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1972 1973BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 1974 return new BBVectorize(C); 1975} 1976 1977bool 1978llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 1979 BBVectorize BBVectorizer(P, C); 1980 return BBVectorizer.vectorizeBB(BB); 1981} 1982 1983//===----------------------------------------------------------------------===// 1984VectorizeConfig::VectorizeConfig() { 1985 VectorBits = ::VectorBits; 1986 VectorizeInts = !::NoInts; 1987 VectorizeFloats = !::NoFloats; 1988 VectorizePointers = !::NoPointers; 1989 VectorizeCasts = !::NoCasts; 1990 VectorizeMath = !::NoMath; 1991 VectorizeFMA = !::NoFMA; 1992 VectorizeSelect = !::NoSelect; 1993 VectorizeGEP = !::NoGEP; 1994 VectorizeMemOps = !::NoMemOps; 1995 AlignedOnly = ::AlignedOnly; 1996 ReqChainDepth= ::ReqChainDepth; 1997 SearchLimit = ::SearchLimit; 1998 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 1999 SplatBreaksChain = ::SplatBreaksChain; 2000 MaxInsts = ::MaxInsts; 2001 MaxIter = ::MaxIter; 2002 NoMemOpBoost = ::NoMemOpBoost; 2003 FastDep = ::FastDep; 2004} 2005