BBVectorize.cpp revision 86312cc15f29ce2bbd9647b94862e068045280c3
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> 87NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 88 cl::desc("Don't try to vectorize casting (conversion) operations")); 89 90static cl::opt<bool> 91NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 92 cl::desc("Don't try to vectorize floating-point math intrinsics")); 93 94static cl::opt<bool> 95NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 96 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 97 98static cl::opt<bool> 99NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 100 cl::desc("Don't try to vectorize loads and stores")); 101 102static cl::opt<bool> 103AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 104 cl::desc("Only generate aligned loads and stores")); 105 106static cl::opt<bool> 107NoMemOpBoost("bb-vectorize-no-mem-op-boost", 108 cl::init(false), cl::Hidden, 109 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 110 111static cl::opt<bool> 112FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 113 cl::desc("Use a fast instruction dependency analysis")); 114 115#ifndef NDEBUG 116static cl::opt<bool> 117DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 118 cl::init(false), cl::Hidden, 119 cl::desc("When debugging is enabled, output information on the" 120 " instruction-examination process")); 121static cl::opt<bool> 122DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 123 cl::init(false), cl::Hidden, 124 cl::desc("When debugging is enabled, output information on the" 125 " candidate-selection process")); 126static cl::opt<bool> 127DebugPairSelection("bb-vectorize-debug-pair-selection", 128 cl::init(false), cl::Hidden, 129 cl::desc("When debugging is enabled, output information on the" 130 " pair-selection process")); 131static cl::opt<bool> 132DebugCycleCheck("bb-vectorize-debug-cycle-check", 133 cl::init(false), cl::Hidden, 134 cl::desc("When debugging is enabled, output information on the" 135 " cycle-checking process")); 136#endif 137 138STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 139 140namespace { 141 struct BBVectorize : public BasicBlockPass { 142 static char ID; // Pass identification, replacement for typeid 143 144 const VectorizeConfig Config; 145 146 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 147 : BasicBlockPass(ID), Config(C) { 148 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 149 } 150 151 BBVectorize(Pass *P, const VectorizeConfig &C) 152 : BasicBlockPass(ID), Config(C) { 153 AA = &P->getAnalysis<AliasAnalysis>(); 154 SE = &P->getAnalysis<ScalarEvolution>(); 155 TD = P->getAnalysisIfAvailable<TargetData>(); 156 } 157 158 typedef std::pair<Value *, Value *> ValuePair; 159 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 160 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 161 typedef std::pair<std::multimap<Value *, Value *>::iterator, 162 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 163 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 164 std::multimap<ValuePair, ValuePair>::iterator> 165 VPPIteratorPair; 166 167 AliasAnalysis *AA; 168 ScalarEvolution *SE; 169 TargetData *TD; 170 171 // FIXME: const correct? 172 173 bool vectorizePairs(BasicBlock &BB); 174 175 bool getCandidatePairs(BasicBlock &BB, 176 BasicBlock::iterator &Start, 177 std::multimap<Value *, Value *> &CandidatePairs, 178 std::vector<Value *> &PairableInsts); 179 180 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 181 std::vector<Value *> &PairableInsts, 182 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 183 184 void buildDepMap(BasicBlock &BB, 185 std::multimap<Value *, Value *> &CandidatePairs, 186 std::vector<Value *> &PairableInsts, 187 DenseSet<ValuePair> &PairableInstUsers); 188 189 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 190 std::vector<Value *> &PairableInsts, 191 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 192 DenseSet<ValuePair> &PairableInstUsers, 193 DenseMap<Value *, Value *>& ChosenPairs); 194 195 void fuseChosenPairs(BasicBlock &BB, 196 std::vector<Value *> &PairableInsts, 197 DenseMap<Value *, Value *>& ChosenPairs); 198 199 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 200 201 bool areInstsCompatible(Instruction *I, Instruction *J, 202 bool IsSimpleLoadStore); 203 204 bool trackUsesOfI(DenseSet<Value *> &Users, 205 AliasSetTracker &WriteSet, Instruction *I, 206 Instruction *J, bool UpdateUsers = true, 207 std::multimap<Value *, Value *> *LoadMoveSet = 0); 208 209 void computePairsConnectedTo( 210 std::multimap<Value *, Value *> &CandidatePairs, 211 std::vector<Value *> &PairableInsts, 212 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 213 ValuePair P); 214 215 bool pairsConflict(ValuePair P, ValuePair Q, 216 DenseSet<ValuePair> &PairableInstUsers, 217 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 218 219 bool pairWillFormCycle(ValuePair P, 220 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 221 DenseSet<ValuePair> &CurrentPairs); 222 223 void pruneTreeFor( 224 std::multimap<Value *, Value *> &CandidatePairs, 225 std::vector<Value *> &PairableInsts, 226 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 227 DenseSet<ValuePair> &PairableInstUsers, 228 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 229 DenseMap<Value *, Value *> &ChosenPairs, 230 DenseMap<ValuePair, size_t> &Tree, 231 DenseSet<ValuePair> &PrunedTree, ValuePair J, 232 bool UseCycleCheck); 233 234 void buildInitialTreeFor( 235 std::multimap<Value *, Value *> &CandidatePairs, 236 std::vector<Value *> &PairableInsts, 237 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 238 DenseSet<ValuePair> &PairableInstUsers, 239 DenseMap<Value *, Value *> &ChosenPairs, 240 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 241 242 void findBestTreeFor( 243 std::multimap<Value *, Value *> &CandidatePairs, 244 std::vector<Value *> &PairableInsts, 245 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 246 DenseSet<ValuePair> &PairableInstUsers, 247 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 248 DenseMap<Value *, Value *> &ChosenPairs, 249 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 250 size_t &BestEffSize, VPIteratorPair ChoiceRange, 251 bool UseCycleCheck); 252 253 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 254 Instruction *J, unsigned o, bool &FlipMemInputs); 255 256 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 257 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 258 unsigned IdxOffset, std::vector<Constant*> &Mask); 259 260 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 261 Instruction *J); 262 263 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 264 Instruction *J, unsigned o, bool FlipMemInputs); 265 266 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 267 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 268 bool &FlipMemInputs); 269 270 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 271 Instruction *J, Instruction *K, 272 Instruction *&InsertionPt, Instruction *&K1, 273 Instruction *&K2, bool &FlipMemInputs); 274 275 void collectPairLoadMoveSet(BasicBlock &BB, 276 DenseMap<Value *, Value *> &ChosenPairs, 277 std::multimap<Value *, Value *> &LoadMoveSet, 278 Instruction *I); 279 280 void collectLoadMoveSet(BasicBlock &BB, 281 std::vector<Value *> &PairableInsts, 282 DenseMap<Value *, Value *> &ChosenPairs, 283 std::multimap<Value *, Value *> &LoadMoveSet); 284 285 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 286 std::multimap<Value *, Value *> &LoadMoveSet, 287 Instruction *I, Instruction *J); 288 289 void moveUsesOfIAfterJ(BasicBlock &BB, 290 std::multimap<Value *, Value *> &LoadMoveSet, 291 Instruction *&InsertionPt, 292 Instruction *I, Instruction *J); 293 294 bool vectorizeBB(BasicBlock &BB) { 295 bool changed = false; 296 // Iterate a sufficient number of times to merge types of size 1 bit, 297 // then 2 bits, then 4, etc. up to half of the target vector width of the 298 // target vector register. 299 for (unsigned v = 2, n = 1; 300 v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter); 301 v *= 2, ++n) { 302 DEBUG(dbgs() << "BBV: fusing loop #" << n << 303 " for " << BB.getName() << " in " << 304 BB.getParent()->getName() << "...\n"); 305 if (vectorizePairs(BB)) 306 changed = true; 307 else 308 break; 309 } 310 311 DEBUG(dbgs() << "BBV: done!\n"); 312 return changed; 313 } 314 315 virtual bool runOnBasicBlock(BasicBlock &BB) { 316 AA = &getAnalysis<AliasAnalysis>(); 317 SE = &getAnalysis<ScalarEvolution>(); 318 TD = getAnalysisIfAvailable<TargetData>(); 319 320 return vectorizeBB(BB); 321 } 322 323 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 324 BasicBlockPass::getAnalysisUsage(AU); 325 AU.addRequired<AliasAnalysis>(); 326 AU.addRequired<ScalarEvolution>(); 327 AU.addPreserved<AliasAnalysis>(); 328 AU.addPreserved<ScalarEvolution>(); 329 AU.setPreservesCFG(); 330 } 331 332 // This returns the vector type that holds a pair of the provided type. 333 // If the provided type is already a vector, then its length is doubled. 334 static inline VectorType *getVecTypeForPair(Type *ElemTy) { 335 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 336 unsigned numElem = VTy->getNumElements(); 337 return VectorType::get(ElemTy->getScalarType(), numElem*2); 338 } 339 340 return VectorType::get(ElemTy, 2); 341 } 342 343 // Returns the weight associated with the provided value. A chain of 344 // candidate pairs has a length given by the sum of the weights of its 345 // members (one weight per pair; the weight of each member of the pair 346 // is assumed to be the same). This length is then compared to the 347 // chain-length threshold to determine if a given chain is significant 348 // enough to be vectorized. The length is also used in comparing 349 // candidate chains where longer chains are considered to be better. 350 // Note: when this function returns 0, the resulting instructions are 351 // not actually fused. 352 inline size_t getDepthFactor(Value *V) { 353 // InsertElement and ExtractElement have a depth factor of zero. This is 354 // for two reasons: First, they cannot be usefully fused. Second, because 355 // the pass generates a lot of these, they can confuse the simple metric 356 // used to compare the trees in the next iteration. Thus, giving them a 357 // weight of zero allows the pass to essentially ignore them in 358 // subsequent iterations when looking for vectorization opportunities 359 // while still tracking dependency chains that flow through those 360 // instructions. 361 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 362 return 0; 363 364 // Give a load or store half of the required depth so that load/store 365 // pairs will vectorize. 366 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 367 return Config.ReqChainDepth/2; 368 369 return 1; 370 } 371 372 // This determines the relative offset of two loads or stores, returning 373 // true if the offset could be determined to be some constant value. 374 // For example, if OffsetInElmts == 1, then J accesses the memory directly 375 // after I; if OffsetInElmts == -1 then I accesses the memory 376 // directly after J. This function assumes that both instructions 377 // have the same type. 378 bool getPairPtrInfo(Instruction *I, Instruction *J, 379 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 380 int64_t &OffsetInElmts) { 381 OffsetInElmts = 0; 382 if (isa<LoadInst>(I)) { 383 IPtr = cast<LoadInst>(I)->getPointerOperand(); 384 JPtr = cast<LoadInst>(J)->getPointerOperand(); 385 IAlignment = cast<LoadInst>(I)->getAlignment(); 386 JAlignment = cast<LoadInst>(J)->getAlignment(); 387 } else { 388 IPtr = cast<StoreInst>(I)->getPointerOperand(); 389 JPtr = cast<StoreInst>(J)->getPointerOperand(); 390 IAlignment = cast<StoreInst>(I)->getAlignment(); 391 JAlignment = cast<StoreInst>(J)->getAlignment(); 392 } 393 394 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 395 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 396 397 // If this is a trivial offset, then we'll get something like 398 // 1*sizeof(type). With target data, which we need anyway, this will get 399 // constant folded into a number. 400 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 401 if (const SCEVConstant *ConstOffSCEV = 402 dyn_cast<SCEVConstant>(OffsetSCEV)) { 403 ConstantInt *IntOff = ConstOffSCEV->getValue(); 404 int64_t Offset = IntOff->getSExtValue(); 405 406 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 407 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 408 409 assert(VTy == cast<PointerType>(JPtr->getType())->getElementType()); 410 411 OffsetInElmts = Offset/VTyTSS; 412 return (abs64(Offset) % VTyTSS) == 0; 413 } 414 415 return false; 416 } 417 418 // Returns true if the provided CallInst represents an intrinsic that can 419 // be vectorized. 420 bool isVectorizableIntrinsic(CallInst* I) { 421 Function *F = I->getCalledFunction(); 422 if (!F) return false; 423 424 unsigned IID = F->getIntrinsicID(); 425 if (!IID) return false; 426 427 switch(IID) { 428 default: 429 return false; 430 case Intrinsic::sqrt: 431 case Intrinsic::powi: 432 case Intrinsic::sin: 433 case Intrinsic::cos: 434 case Intrinsic::log: 435 case Intrinsic::log2: 436 case Intrinsic::log10: 437 case Intrinsic::exp: 438 case Intrinsic::exp2: 439 case Intrinsic::pow: 440 return Config.VectorizeMath; 441 case Intrinsic::fma: 442 return Config.VectorizeFMA; 443 } 444 } 445 446 // Returns true if J is the second element in some pair referenced by 447 // some multimap pair iterator pair. 448 template <typename V> 449 bool isSecondInIteratorPair(V J, std::pair< 450 typename std::multimap<V, V>::iterator, 451 typename std::multimap<V, V>::iterator> PairRange) { 452 for (typename std::multimap<V, V>::iterator K = PairRange.first; 453 K != PairRange.second; ++K) 454 if (K->second == J) return true; 455 456 return false; 457 } 458 }; 459 460 // This function implements one vectorization iteration on the provided 461 // basic block. It returns true if the block is changed. 462 bool BBVectorize::vectorizePairs(BasicBlock &BB) { 463 bool ShouldContinue; 464 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 465 466 std::vector<Value *> AllPairableInsts; 467 DenseMap<Value *, Value *> AllChosenPairs; 468 469 do { 470 std::vector<Value *> PairableInsts; 471 std::multimap<Value *, Value *> CandidatePairs; 472 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 473 PairableInsts); 474 if (PairableInsts.empty()) continue; 475 476 // Now we have a map of all of the pairable instructions and we need to 477 // select the best possible pairing. A good pairing is one such that the 478 // users of the pair are also paired. This defines a (directed) forest 479 // over the pairs such that two pairs are connected iff the second pair 480 // uses the first. 481 482 // Note that it only matters that both members of the second pair use some 483 // element of the first pair (to allow for splatting). 484 485 std::multimap<ValuePair, ValuePair> ConnectedPairs; 486 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 487 if (ConnectedPairs.empty()) continue; 488 489 // Build the pairable-instruction dependency map 490 DenseSet<ValuePair> PairableInstUsers; 491 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 492 493 // There is now a graph of the connected pairs. For each variable, pick 494 // the pairing with the largest tree meeting the depth requirement on at 495 // least one branch. Then select all pairings that are part of that tree 496 // and remove them from the list of available pairings and pairable 497 // variables. 498 499 DenseMap<Value *, Value *> ChosenPairs; 500 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, 501 PairableInstUsers, ChosenPairs); 502 503 if (ChosenPairs.empty()) continue; 504 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 505 PairableInsts.end()); 506 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 507 } while (ShouldContinue); 508 509 if (AllChosenPairs.empty()) return false; 510 NumFusedOps += AllChosenPairs.size(); 511 512 // A set of pairs has now been selected. It is now necessary to replace the 513 // paired instructions with vector instructions. For this procedure each 514 // operand must be replaced with a vector operand. This vector is formed 515 // by using build_vector on the old operands. The replaced values are then 516 // replaced with a vector_extract on the result. Subsequent optimization 517 // passes should coalesce the build/extract combinations. 518 519 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 520 return true; 521 } 522 523 // This function returns true if the provided instruction is capable of being 524 // fused into a vector instruction. This determination is based only on the 525 // type and other attributes of the instruction. 526 bool BBVectorize::isInstVectorizable(Instruction *I, 527 bool &IsSimpleLoadStore) { 528 IsSimpleLoadStore = false; 529 530 if (CallInst *C = dyn_cast<CallInst>(I)) { 531 if (!isVectorizableIntrinsic(C)) 532 return false; 533 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 534 // Vectorize simple loads if possbile: 535 IsSimpleLoadStore = L->isSimple(); 536 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 537 return false; 538 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 539 // Vectorize simple stores if possbile: 540 IsSimpleLoadStore = S->isSimple(); 541 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 542 return false; 543 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 544 // We can vectorize casts, but not casts of pointer types, etc. 545 if (!Config.VectorizeCasts) 546 return false; 547 548 Type *SrcTy = C->getSrcTy(); 549 if (!SrcTy->isSingleValueType() || SrcTy->isPointerTy()) 550 return false; 551 552 Type *DestTy = C->getDestTy(); 553 if (!DestTy->isSingleValueType() || DestTy->isPointerTy()) 554 return false; 555 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 556 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 557 return false; 558 } 559 560 // We can't vectorize memory operations without target data 561 if (TD == 0 && IsSimpleLoadStore) 562 return false; 563 564 Type *T1, *T2; 565 if (isa<StoreInst>(I)) { 566 // For stores, it is the value type, not the pointer type that matters 567 // because the value is what will come from a vector register. 568 569 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 570 T1 = IVal->getType(); 571 } else { 572 T1 = I->getType(); 573 } 574 575 if (I->isCast()) 576 T2 = cast<CastInst>(I)->getSrcTy(); 577 else 578 T2 = T1; 579 580 // Not every type can be vectorized... 581 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 582 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 583 return false; 584 585 if (!Config.VectorizeInts 586 && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) 587 return false; 588 589 if (!Config.VectorizeFloats 590 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 591 return false; 592 593 if (T1->getPrimitiveSizeInBits() > Config.VectorBits/2 || 594 T2->getPrimitiveSizeInBits() > Config.VectorBits/2) 595 return false; 596 597 return true; 598 } 599 600 // This function returns true if the two provided instructions are compatible 601 // (meaning that they can be fused into a vector instruction). This assumes 602 // that I has already been determined to be vectorizable and that J is not 603 // in the use tree of I. 604 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 605 bool IsSimpleLoadStore) { 606 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 607 " <-> " << *J << "\n"); 608 609 // Loads and stores can be merged if they have different alignments, 610 // but are otherwise the same. 611 LoadInst *LI, *LJ; 612 StoreInst *SI, *SJ; 613 if ((LI = dyn_cast<LoadInst>(I)) && (LJ = dyn_cast<LoadInst>(J))) { 614 if (I->getType() != J->getType()) 615 return false; 616 617 if (LI->getPointerOperand()->getType() != 618 LJ->getPointerOperand()->getType() || 619 LI->isVolatile() != LJ->isVolatile() || 620 LI->getOrdering() != LJ->getOrdering() || 621 LI->getSynchScope() != LJ->getSynchScope()) 622 return false; 623 } else if ((SI = dyn_cast<StoreInst>(I)) && (SJ = dyn_cast<StoreInst>(J))) { 624 if (SI->getValueOperand()->getType() != 625 SJ->getValueOperand()->getType() || 626 SI->getPointerOperand()->getType() != 627 SJ->getPointerOperand()->getType() || 628 SI->isVolatile() != SJ->isVolatile() || 629 SI->getOrdering() != SJ->getOrdering() || 630 SI->getSynchScope() != SJ->getSynchScope()) 631 return false; 632 } else if (!J->isSameOperationAs(I)) { 633 return false; 634 } 635 // FIXME: handle addsub-type operations! 636 637 if (IsSimpleLoadStore) { 638 Value *IPtr, *JPtr; 639 unsigned IAlignment, JAlignment; 640 int64_t OffsetInElmts = 0; 641 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 642 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 643 if (Config.AlignedOnly) { 644 Type *aType = isa<StoreInst>(I) ? 645 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 646 // An aligned load or store is possible only if the instruction 647 // with the lower offset has an alignment suitable for the 648 // vector type. 649 650 unsigned BottomAlignment = IAlignment; 651 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 652 653 Type *VType = getVecTypeForPair(aType); 654 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 655 if (BottomAlignment < VecAlignment) 656 return false; 657 } 658 } else { 659 return false; 660 } 661 } else if (isa<ShuffleVectorInst>(I)) { 662 // Only merge two shuffles if they're both constant 663 return isa<Constant>(I->getOperand(2)) && 664 isa<Constant>(J->getOperand(2)); 665 // FIXME: We may want to vectorize non-constant shuffles also. 666 } 667 668 // The powi intrinsic is special because only the first argument is 669 // vectorized, the second arguments must be equal. 670 CallInst *CI = dyn_cast<CallInst>(I); 671 Function *FI; 672 if (CI && (FI = CI->getCalledFunction()) && 673 FI->getIntrinsicID() == Intrinsic::powi) { 674 675 Value *A1I = CI->getArgOperand(1), 676 *A1J = cast<CallInst>(J)->getArgOperand(1); 677 const SCEV *A1ISCEV = SE->getSCEV(A1I), 678 *A1JSCEV = SE->getSCEV(A1J); 679 return (A1ISCEV == A1JSCEV); 680 } 681 682 return true; 683 } 684 685 // Figure out whether or not J uses I and update the users and write-set 686 // structures associated with I. Specifically, Users represents the set of 687 // instructions that depend on I. WriteSet represents the set 688 // of memory locations that are dependent on I. If UpdateUsers is true, 689 // and J uses I, then Users is updated to contain J and WriteSet is updated 690 // to contain any memory locations to which J writes. The function returns 691 // true if J uses I. By default, alias analysis is used to determine 692 // whether J reads from memory that overlaps with a location in WriteSet. 693 // If LoadMoveSet is not null, then it is a previously-computed multimap 694 // where the key is the memory-based user instruction and the value is 695 // the instruction to be compared with I. So, if LoadMoveSet is provided, 696 // then the alias analysis is not used. This is necessary because this 697 // function is called during the process of moving instructions during 698 // vectorization and the results of the alias analysis are not stable during 699 // that process. 700 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 701 AliasSetTracker &WriteSet, Instruction *I, 702 Instruction *J, bool UpdateUsers, 703 std::multimap<Value *, Value *> *LoadMoveSet) { 704 bool UsesI = false; 705 706 // This instruction may already be marked as a user due, for example, to 707 // being a member of a selected pair. 708 if (Users.count(J)) 709 UsesI = true; 710 711 if (!UsesI) 712 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 713 JU != JE; ++JU) { 714 Value *V = *JU; 715 if (I == V || Users.count(V)) { 716 UsesI = true; 717 break; 718 } 719 } 720 if (!UsesI && J->mayReadFromMemory()) { 721 if (LoadMoveSet) { 722 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 723 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 724 } else { 725 for (AliasSetTracker::iterator W = WriteSet.begin(), 726 WE = WriteSet.end(); W != WE; ++W) { 727 if (W->aliasesUnknownInst(J, *AA)) { 728 UsesI = true; 729 break; 730 } 731 } 732 } 733 } 734 735 if (UsesI && UpdateUsers) { 736 if (J->mayWriteToMemory()) WriteSet.add(J); 737 Users.insert(J); 738 } 739 740 return UsesI; 741 } 742 743 // This function iterates over all instruction pairs in the provided 744 // basic block and collects all candidate pairs for vectorization. 745 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 746 BasicBlock::iterator &Start, 747 std::multimap<Value *, Value *> &CandidatePairs, 748 std::vector<Value *> &PairableInsts) { 749 BasicBlock::iterator E = BB.end(); 750 if (Start == E) return false; 751 752 bool ShouldContinue = false, IAfterStart = false; 753 for (BasicBlock::iterator I = Start++; I != E; ++I) { 754 if (I == Start) IAfterStart = true; 755 756 bool IsSimpleLoadStore; 757 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 758 759 // Look for an instruction with which to pair instruction *I... 760 DenseSet<Value *> Users; 761 AliasSetTracker WriteSet(*AA); 762 bool JAfterStart = IAfterStart; 763 BasicBlock::iterator J = llvm::next(I); 764 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 765 if (J == Start) JAfterStart = true; 766 767 // Determine if J uses I, if so, exit the loop. 768 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 769 if (Config.FastDep) { 770 // Note: For this heuristic to be effective, independent operations 771 // must tend to be intermixed. This is likely to be true from some 772 // kinds of grouped loop unrolling (but not the generic LLVM pass), 773 // but otherwise may require some kind of reordering pass. 774 775 // When using fast dependency analysis, 776 // stop searching after first use: 777 if (UsesI) break; 778 } else { 779 if (UsesI) continue; 780 } 781 782 // J does not use I, and comes before the first use of I, so it can be 783 // merged with I if the instructions are compatible. 784 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue; 785 786 // J is a candidate for merging with I. 787 if (!PairableInsts.size() || 788 PairableInsts[PairableInsts.size()-1] != I) { 789 PairableInsts.push_back(I); 790 } 791 792 CandidatePairs.insert(ValuePair(I, J)); 793 794 // The next call to this function must start after the last instruction 795 // selected during this invocation. 796 if (JAfterStart) { 797 Start = llvm::next(J); 798 IAfterStart = JAfterStart = false; 799 } 800 801 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 802 << *I << " <-> " << *J << "\n"); 803 804 // If we have already found too many pairs, break here and this function 805 // will be called again starting after the last instruction selected 806 // during this invocation. 807 if (PairableInsts.size() >= Config.MaxInsts) { 808 ShouldContinue = true; 809 break; 810 } 811 } 812 813 if (ShouldContinue) 814 break; 815 } 816 817 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 818 << " instructions with candidate pairs\n"); 819 820 return ShouldContinue; 821 } 822 823 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 824 // it looks for pairs such that both members have an input which is an 825 // output of PI or PJ. 826 void BBVectorize::computePairsConnectedTo( 827 std::multimap<Value *, Value *> &CandidatePairs, 828 std::vector<Value *> &PairableInsts, 829 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 830 ValuePair P) { 831 // For each possible pairing for this variable, look at the uses of 832 // the first value... 833 for (Value::use_iterator I = P.first->use_begin(), 834 E = P.first->use_end(); I != E; ++I) { 835 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 836 837 // For each use of the first variable, look for uses of the second 838 // variable... 839 for (Value::use_iterator J = P.second->use_begin(), 840 E2 = P.second->use_end(); J != E2; ++J) { 841 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 842 843 // Look for <I, J>: 844 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 845 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 846 847 // Look for <J, I>: 848 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 849 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 850 } 851 852 if (Config.SplatBreaksChain) continue; 853 // Look for cases where just the first value in the pair is used by 854 // both members of another pair (splatting). 855 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 856 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 857 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 858 } 859 } 860 861 if (Config.SplatBreaksChain) return; 862 // Look for cases where just the second value in the pair is used by 863 // both members of another pair (splatting). 864 for (Value::use_iterator I = P.second->use_begin(), 865 E = P.second->use_end(); I != E; ++I) { 866 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 867 868 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 869 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 870 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 871 } 872 } 873 } 874 875 // This function figures out which pairs are connected. Two pairs are 876 // connected if some output of the first pair forms an input to both members 877 // of the second pair. 878 void BBVectorize::computeConnectedPairs( 879 std::multimap<Value *, Value *> &CandidatePairs, 880 std::vector<Value *> &PairableInsts, 881 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 882 883 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 884 PE = PairableInsts.end(); PI != PE; ++PI) { 885 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 886 887 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 888 P != choiceRange.second; ++P) 889 computePairsConnectedTo(CandidatePairs, PairableInsts, 890 ConnectedPairs, *P); 891 } 892 893 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 894 << " pair connections.\n"); 895 } 896 897 // This function builds a set of use tuples such that <A, B> is in the set 898 // if B is in the use tree of A. If B is in the use tree of A, then B 899 // depends on the output of A. 900 void BBVectorize::buildDepMap( 901 BasicBlock &BB, 902 std::multimap<Value *, Value *> &CandidatePairs, 903 std::vector<Value *> &PairableInsts, 904 DenseSet<ValuePair> &PairableInstUsers) { 905 DenseSet<Value *> IsInPair; 906 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 907 E = CandidatePairs.end(); C != E; ++C) { 908 IsInPair.insert(C->first); 909 IsInPair.insert(C->second); 910 } 911 912 // Iterate through the basic block, recording all Users of each 913 // pairable instruction. 914 915 BasicBlock::iterator E = BB.end(); 916 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 917 if (IsInPair.find(I) == IsInPair.end()) continue; 918 919 DenseSet<Value *> Users; 920 AliasSetTracker WriteSet(*AA); 921 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 922 (void) trackUsesOfI(Users, WriteSet, I, J); 923 924 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 925 U != E; ++U) 926 PairableInstUsers.insert(ValuePair(I, *U)); 927 } 928 } 929 930 // Returns true if an input to pair P is an output of pair Q and also an 931 // input of pair Q is an output of pair P. If this is the case, then these 932 // two pairs cannot be simultaneously fused. 933 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 934 DenseSet<ValuePair> &PairableInstUsers, 935 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 936 // Two pairs are in conflict if they are mutual Users of eachother. 937 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 938 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 939 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 940 PairableInstUsers.count(ValuePair(P.second, Q.second)); 941 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 942 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 943 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 944 PairableInstUsers.count(ValuePair(Q.second, P.second)); 945 if (PairableInstUserMap) { 946 // FIXME: The expensive part of the cycle check is not so much the cycle 947 // check itself but this edge insertion procedure. This needs some 948 // profiling and probably a different data structure (same is true of 949 // most uses of std::multimap). 950 if (PUsesQ) { 951 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 952 if (!isSecondInIteratorPair(P, QPairRange)) 953 PairableInstUserMap->insert(VPPair(Q, P)); 954 } 955 if (QUsesP) { 956 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 957 if (!isSecondInIteratorPair(Q, PPairRange)) 958 PairableInstUserMap->insert(VPPair(P, Q)); 959 } 960 } 961 962 return (QUsesP && PUsesQ); 963 } 964 965 // This function walks the use graph of current pairs to see if, starting 966 // from P, the walk returns to P. 967 bool BBVectorize::pairWillFormCycle(ValuePair P, 968 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 969 DenseSet<ValuePair> &CurrentPairs) { 970 DEBUG(if (DebugCycleCheck) 971 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 972 << *P.second << "\n"); 973 // A lookup table of visisted pairs is kept because the PairableInstUserMap 974 // contains non-direct associations. 975 DenseSet<ValuePair> Visited; 976 SmallVector<ValuePair, 32> Q; 977 // General depth-first post-order traversal: 978 Q.push_back(P); 979 do { 980 ValuePair QTop = Q.pop_back_val(); 981 Visited.insert(QTop); 982 983 DEBUG(if (DebugCycleCheck) 984 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 985 << *QTop.second << "\n"); 986 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 987 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 988 C != QPairRange.second; ++C) { 989 if (C->second == P) { 990 DEBUG(dbgs() 991 << "BBV: rejected to prevent non-trivial cycle formation: " 992 << *C->first.first << " <-> " << *C->first.second << "\n"); 993 return true; 994 } 995 996 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 997 Q.push_back(C->second); 998 } 999 } while (!Q.empty()); 1000 1001 return false; 1002 } 1003 1004 // This function builds the initial tree of connected pairs with the 1005 // pair J at the root. 1006 void BBVectorize::buildInitialTreeFor( 1007 std::multimap<Value *, Value *> &CandidatePairs, 1008 std::vector<Value *> &PairableInsts, 1009 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1010 DenseSet<ValuePair> &PairableInstUsers, 1011 DenseMap<Value *, Value *> &ChosenPairs, 1012 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1013 // Each of these pairs is viewed as the root node of a Tree. The Tree 1014 // is then walked (depth-first). As this happens, we keep track of 1015 // the pairs that compose the Tree and the maximum depth of the Tree. 1016 SmallVector<ValuePairWithDepth, 32> Q; 1017 // General depth-first post-order traversal: 1018 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1019 do { 1020 ValuePairWithDepth QTop = Q.back(); 1021 1022 // Push each child onto the queue: 1023 bool MoreChildren = false; 1024 size_t MaxChildDepth = QTop.second; 1025 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1026 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1027 k != qtRange.second; ++k) { 1028 // Make sure that this child pair is still a candidate: 1029 bool IsStillCand = false; 1030 VPIteratorPair checkRange = 1031 CandidatePairs.equal_range(k->second.first); 1032 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1033 m != checkRange.second; ++m) { 1034 if (m->second == k->second.second) { 1035 IsStillCand = true; 1036 break; 1037 } 1038 } 1039 1040 if (IsStillCand) { 1041 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1042 if (C == Tree.end()) { 1043 size_t d = getDepthFactor(k->second.first); 1044 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1045 MoreChildren = true; 1046 } else { 1047 MaxChildDepth = std::max(MaxChildDepth, C->second); 1048 } 1049 } 1050 } 1051 1052 if (!MoreChildren) { 1053 // Record the current pair as part of the Tree: 1054 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1055 Q.pop_back(); 1056 } 1057 } while (!Q.empty()); 1058 } 1059 1060 // Given some initial tree, prune it by removing conflicting pairs (pairs 1061 // that cannot be simultaneously chosen for vectorization). 1062 void BBVectorize::pruneTreeFor( 1063 std::multimap<Value *, Value *> &CandidatePairs, 1064 std::vector<Value *> &PairableInsts, 1065 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1066 DenseSet<ValuePair> &PairableInstUsers, 1067 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1068 DenseMap<Value *, Value *> &ChosenPairs, 1069 DenseMap<ValuePair, size_t> &Tree, 1070 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1071 bool UseCycleCheck) { 1072 SmallVector<ValuePairWithDepth, 32> Q; 1073 // General depth-first post-order traversal: 1074 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1075 do { 1076 ValuePairWithDepth QTop = Q.pop_back_val(); 1077 PrunedTree.insert(QTop.first); 1078 1079 // Visit each child, pruning as necessary... 1080 DenseMap<ValuePair, size_t> BestChildren; 1081 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1082 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1083 K != QTopRange.second; ++K) { 1084 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1085 if (C == Tree.end()) continue; 1086 1087 // This child is in the Tree, now we need to make sure it is the 1088 // best of any conflicting children. There could be multiple 1089 // conflicting children, so first, determine if we're keeping 1090 // this child, then delete conflicting children as necessary. 1091 1092 // It is also necessary to guard against pairing-induced 1093 // dependencies. Consider instructions a .. x .. y .. b 1094 // such that (a,b) are to be fused and (x,y) are to be fused 1095 // but a is an input to x and b is an output from y. This 1096 // means that y cannot be moved after b but x must be moved 1097 // after b for (a,b) to be fused. In other words, after 1098 // fusing (a,b) we have y .. a/b .. x where y is an input 1099 // to a/b and x is an output to a/b: x and y can no longer 1100 // be legally fused. To prevent this condition, we must 1101 // make sure that a child pair added to the Tree is not 1102 // both an input and output of an already-selected pair. 1103 1104 // Pairing-induced dependencies can also form from more complicated 1105 // cycles. The pair vs. pair conflicts are easy to check, and so 1106 // that is done explicitly for "fast rejection", and because for 1107 // child vs. child conflicts, we may prefer to keep the current 1108 // pair in preference to the already-selected child. 1109 DenseSet<ValuePair> CurrentPairs; 1110 1111 bool CanAdd = true; 1112 for (DenseMap<ValuePair, size_t>::iterator C2 1113 = BestChildren.begin(), E2 = BestChildren.end(); 1114 C2 != E2; ++C2) { 1115 if (C2->first.first == C->first.first || 1116 C2->first.first == C->first.second || 1117 C2->first.second == C->first.first || 1118 C2->first.second == C->first.second || 1119 pairsConflict(C2->first, C->first, PairableInstUsers, 1120 UseCycleCheck ? &PairableInstUserMap : 0)) { 1121 if (C2->second >= C->second) { 1122 CanAdd = false; 1123 break; 1124 } 1125 1126 CurrentPairs.insert(C2->first); 1127 } 1128 } 1129 if (!CanAdd) continue; 1130 1131 // Even worse, this child could conflict with another node already 1132 // selected for the Tree. If that is the case, ignore this child. 1133 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1134 E2 = PrunedTree.end(); T != E2; ++T) { 1135 if (T->first == C->first.first || 1136 T->first == C->first.second || 1137 T->second == C->first.first || 1138 T->second == C->first.second || 1139 pairsConflict(*T, C->first, PairableInstUsers, 1140 UseCycleCheck ? &PairableInstUserMap : 0)) { 1141 CanAdd = false; 1142 break; 1143 } 1144 1145 CurrentPairs.insert(*T); 1146 } 1147 if (!CanAdd) continue; 1148 1149 // And check the queue too... 1150 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1151 E2 = Q.end(); C2 != E2; ++C2) { 1152 if (C2->first.first == C->first.first || 1153 C2->first.first == C->first.second || 1154 C2->first.second == C->first.first || 1155 C2->first.second == C->first.second || 1156 pairsConflict(C2->first, C->first, PairableInstUsers, 1157 UseCycleCheck ? &PairableInstUserMap : 0)) { 1158 CanAdd = false; 1159 break; 1160 } 1161 1162 CurrentPairs.insert(C2->first); 1163 } 1164 if (!CanAdd) continue; 1165 1166 // Last but not least, check for a conflict with any of the 1167 // already-chosen pairs. 1168 for (DenseMap<Value *, Value *>::iterator C2 = 1169 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1170 C2 != E2; ++C2) { 1171 if (pairsConflict(*C2, C->first, PairableInstUsers, 1172 UseCycleCheck ? &PairableInstUserMap : 0)) { 1173 CanAdd = false; 1174 break; 1175 } 1176 1177 CurrentPairs.insert(*C2); 1178 } 1179 if (!CanAdd) continue; 1180 1181 // To check for non-trivial cycles formed by the addition of the 1182 // current pair we've formed a list of all relevant pairs, now use a 1183 // graph walk to check for a cycle. We start from the current pair and 1184 // walk the use tree to see if we again reach the current pair. If we 1185 // do, then the current pair is rejected. 1186 1187 // FIXME: It may be more efficient to use a topological-ordering 1188 // algorithm to improve the cycle check. This should be investigated. 1189 if (UseCycleCheck && 1190 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1191 continue; 1192 1193 // This child can be added, but we may have chosen it in preference 1194 // to an already-selected child. Check for this here, and if a 1195 // conflict is found, then remove the previously-selected child 1196 // before adding this one in its place. 1197 for (DenseMap<ValuePair, size_t>::iterator C2 1198 = BestChildren.begin(); C2 != BestChildren.end();) { 1199 if (C2->first.first == C->first.first || 1200 C2->first.first == C->first.second || 1201 C2->first.second == C->first.first || 1202 C2->first.second == C->first.second || 1203 pairsConflict(C2->first, C->first, PairableInstUsers)) 1204 BestChildren.erase(C2++); 1205 else 1206 ++C2; 1207 } 1208 1209 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1210 } 1211 1212 for (DenseMap<ValuePair, size_t>::iterator C 1213 = BestChildren.begin(), E2 = BestChildren.end(); 1214 C != E2; ++C) { 1215 size_t DepthF = getDepthFactor(C->first.first); 1216 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1217 } 1218 } while (!Q.empty()); 1219 } 1220 1221 // This function finds the best tree of mututally-compatible connected 1222 // pairs, given the choice of root pairs as an iterator range. 1223 void BBVectorize::findBestTreeFor( 1224 std::multimap<Value *, Value *> &CandidatePairs, 1225 std::vector<Value *> &PairableInsts, 1226 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1227 DenseSet<ValuePair> &PairableInstUsers, 1228 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1229 DenseMap<Value *, Value *> &ChosenPairs, 1230 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1231 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1232 bool UseCycleCheck) { 1233 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1234 J != ChoiceRange.second; ++J) { 1235 1236 // Before going any further, make sure that this pair does not 1237 // conflict with any already-selected pairs (see comment below 1238 // near the Tree pruning for more details). 1239 DenseSet<ValuePair> ChosenPairSet; 1240 bool DoesConflict = false; 1241 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1242 E = ChosenPairs.end(); C != E; ++C) { 1243 if (pairsConflict(*C, *J, PairableInstUsers, 1244 UseCycleCheck ? &PairableInstUserMap : 0)) { 1245 DoesConflict = true; 1246 break; 1247 } 1248 1249 ChosenPairSet.insert(*C); 1250 } 1251 if (DoesConflict) continue; 1252 1253 if (UseCycleCheck && 1254 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1255 continue; 1256 1257 DenseMap<ValuePair, size_t> Tree; 1258 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1259 PairableInstUsers, ChosenPairs, Tree, *J); 1260 1261 // Because we'll keep the child with the largest depth, the largest 1262 // depth is still the same in the unpruned Tree. 1263 size_t MaxDepth = Tree.lookup(*J); 1264 1265 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1266 << *J->first << " <-> " << *J->second << "} of depth " << 1267 MaxDepth << " and size " << Tree.size() << "\n"); 1268 1269 // At this point the Tree has been constructed, but, may contain 1270 // contradictory children (meaning that different children of 1271 // some tree node may be attempting to fuse the same instruction). 1272 // So now we walk the tree again, in the case of a conflict, 1273 // keep only the child with the largest depth. To break a tie, 1274 // favor the first child. 1275 1276 DenseSet<ValuePair> PrunedTree; 1277 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1278 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1279 PrunedTree, *J, UseCycleCheck); 1280 1281 size_t EffSize = 0; 1282 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1283 E = PrunedTree.end(); S != E; ++S) 1284 EffSize += getDepthFactor(S->first); 1285 1286 DEBUG(if (DebugPairSelection) 1287 dbgs() << "BBV: found pruned Tree for pair {" 1288 << *J->first << " <-> " << *J->second << "} of depth " << 1289 MaxDepth << " and size " << PrunedTree.size() << 1290 " (effective size: " << EffSize << ")\n"); 1291 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) { 1292 BestMaxDepth = MaxDepth; 1293 BestEffSize = EffSize; 1294 BestTree = PrunedTree; 1295 } 1296 } 1297 } 1298 1299 // Given the list of candidate pairs, this function selects those 1300 // that will be fused into vector instructions. 1301 void BBVectorize::choosePairs( 1302 std::multimap<Value *, Value *> &CandidatePairs, 1303 std::vector<Value *> &PairableInsts, 1304 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1305 DenseSet<ValuePair> &PairableInstUsers, 1306 DenseMap<Value *, Value *>& ChosenPairs) { 1307 bool UseCycleCheck = 1308 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1309 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1310 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1311 E = PairableInsts.end(); I != E; ++I) { 1312 // The number of possible pairings for this variable: 1313 size_t NumChoices = CandidatePairs.count(*I); 1314 if (!NumChoices) continue; 1315 1316 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1317 1318 // The best pair to choose and its tree: 1319 size_t BestMaxDepth = 0, BestEffSize = 0; 1320 DenseSet<ValuePair> BestTree; 1321 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1322 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1323 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1324 UseCycleCheck); 1325 1326 // A tree has been chosen (or not) at this point. If no tree was 1327 // chosen, then this instruction, I, cannot be paired (and is no longer 1328 // considered). 1329 1330 DEBUG(if (BestTree.size() > 0) 1331 dbgs() << "BBV: selected pairs in the best tree for: " 1332 << *cast<Instruction>(*I) << "\n"); 1333 1334 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1335 SE2 = BestTree.end(); S != SE2; ++S) { 1336 // Insert the members of this tree into the list of chosen pairs. 1337 ChosenPairs.insert(ValuePair(S->first, S->second)); 1338 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1339 *S->second << "\n"); 1340 1341 // Remove all candidate pairs that have values in the chosen tree. 1342 for (std::multimap<Value *, Value *>::iterator K = 1343 CandidatePairs.begin(); K != CandidatePairs.end();) { 1344 if (K->first == S->first || K->second == S->first || 1345 K->second == S->second || K->first == S->second) { 1346 // Don't remove the actual pair chosen so that it can be used 1347 // in subsequent tree selections. 1348 if (!(K->first == S->first && K->second == S->second)) 1349 CandidatePairs.erase(K++); 1350 else 1351 ++K; 1352 } else { 1353 ++K; 1354 } 1355 } 1356 } 1357 } 1358 1359 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1360 } 1361 1362 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1363 unsigned n = 0) { 1364 if (!I->hasName()) 1365 return ""; 1366 1367 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1368 (n > 0 ? "." + utostr(n) : "")).str(); 1369 } 1370 1371 // Returns the value that is to be used as the pointer input to the vector 1372 // instruction that fuses I with J. 1373 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1374 Instruction *I, Instruction *J, unsigned o, 1375 bool &FlipMemInputs) { 1376 Value *IPtr, *JPtr; 1377 unsigned IAlignment, JAlignment; 1378 int64_t OffsetInElmts; 1379 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1380 OffsetInElmts); 1381 1382 // The pointer value is taken to be the one with the lowest offset. 1383 Value *VPtr; 1384 if (OffsetInElmts > 0) { 1385 VPtr = IPtr; 1386 } else { 1387 FlipMemInputs = true; 1388 VPtr = JPtr; 1389 } 1390 1391 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType(); 1392 Type *VArgType = getVecTypeForPair(ArgType); 1393 Type *VArgPtrType = PointerType::get(VArgType, 1394 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1395 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1396 /* insert before */ FlipMemInputs ? J : I); 1397 } 1398 1399 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1400 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 1401 unsigned IdxOffset, std::vector<Constant*> &Mask) { 1402 for (unsigned v = 0; v < NumElem/2; ++v) { 1403 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1404 if (m < 0) { 1405 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1406 } else { 1407 unsigned mm = m + (int) IdxOffset; 1408 if (m >= (int) NumInElem) 1409 mm += (int) NumInElem; 1410 1411 Mask[v+MaskOffset] = 1412 ConstantInt::get(Type::getInt32Ty(Context), mm); 1413 } 1414 } 1415 } 1416 1417 // Returns the value that is to be used as the vector-shuffle mask to the 1418 // vector instruction that fuses I with J. 1419 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1420 Instruction *I, Instruction *J) { 1421 // This is the shuffle mask. We need to append the second 1422 // mask to the first, and the numbers need to be adjusted. 1423 1424 Type *ArgType = I->getType(); 1425 Type *VArgType = getVecTypeForPair(ArgType); 1426 1427 // Get the total number of elements in the fused vector type. 1428 // By definition, this must equal the number of elements in 1429 // the final mask. 1430 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1431 std::vector<Constant*> Mask(NumElem); 1432 1433 Type *OpType = I->getOperand(0)->getType(); 1434 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements(); 1435 1436 // For the mask from the first pair... 1437 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask); 1438 1439 // For the mask from the second pair... 1440 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem, 1441 Mask); 1442 1443 return ConstantVector::get(Mask); 1444 } 1445 1446 // Returns the value to be used as the specified operand of the vector 1447 // instruction that fuses I with J. 1448 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1449 Instruction *J, unsigned o, bool FlipMemInputs) { 1450 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1451 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1452 1453 // Compute the fused vector type for this operand 1454 Type *ArgType = I->getOperand(o)->getType(); 1455 VectorType *VArgType = getVecTypeForPair(ArgType); 1456 1457 Instruction *L = I, *H = J; 1458 if (FlipMemInputs) { 1459 L = J; 1460 H = I; 1461 } 1462 1463 if (ArgType->isVectorTy()) { 1464 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1465 std::vector<Constant*> Mask(numElem); 1466 for (unsigned v = 0; v < numElem; ++v) 1467 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1468 1469 Instruction *BV = new ShuffleVectorInst(L->getOperand(o), 1470 H->getOperand(o), 1471 ConstantVector::get(Mask), 1472 getReplacementName(I, true, o)); 1473 BV->insertBefore(J); 1474 return BV; 1475 } 1476 1477 // If these two inputs are the output of another vector instruction, 1478 // then we should use that output directly. It might be necessary to 1479 // permute it first. [When pairings are fused recursively, you can 1480 // end up with cases where a large vector is decomposed into scalars 1481 // using extractelement instructions, then built into size-2 1482 // vectors using insertelement and the into larger vectors using 1483 // shuffles. InstCombine does not simplify all of these cases well, 1484 // and so we make sure that shuffles are generated here when possible. 1485 ExtractElementInst *LEE 1486 = dyn_cast<ExtractElementInst>(L->getOperand(o)); 1487 ExtractElementInst *HEE 1488 = dyn_cast<ExtractElementInst>(H->getOperand(o)); 1489 1490 if (LEE && HEE && 1491 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) { 1492 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType()); 1493 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue(); 1494 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue(); 1495 if (LEE->getOperand(0) == HEE->getOperand(0)) { 1496 if (LowIndx == 0 && HighIndx == 1) 1497 return LEE->getOperand(0); 1498 1499 std::vector<Constant*> Mask(2); 1500 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1501 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1502 1503 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1504 UndefValue::get(EEType), 1505 ConstantVector::get(Mask), 1506 getReplacementName(I, true, o)); 1507 BV->insertBefore(J); 1508 return BV; 1509 } 1510 1511 std::vector<Constant*> Mask(2); 1512 HighIndx += EEType->getNumElements(); 1513 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1514 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1515 1516 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1517 HEE->getOperand(0), 1518 ConstantVector::get(Mask), 1519 getReplacementName(I, true, o)); 1520 BV->insertBefore(J); 1521 return BV; 1522 } 1523 1524 Instruction *BV1 = InsertElementInst::Create( 1525 UndefValue::get(VArgType), 1526 L->getOperand(o), CV0, 1527 getReplacementName(I, true, o, 1)); 1528 BV1->insertBefore(I); 1529 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o), 1530 CV1, 1531 getReplacementName(I, true, o, 2)); 1532 BV2->insertBefore(J); 1533 return BV2; 1534 } 1535 1536 // This function creates an array of values that will be used as the inputs 1537 // to the vector instruction that fuses I with J. 1538 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1539 Instruction *I, Instruction *J, 1540 SmallVector<Value *, 3> &ReplacedOperands, 1541 bool &FlipMemInputs) { 1542 FlipMemInputs = false; 1543 unsigned NumOperands = I->getNumOperands(); 1544 1545 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1546 // Iterate backward so that we look at the store pointer 1547 // first and know whether or not we need to flip the inputs. 1548 1549 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1550 // This is the pointer for a load/store instruction. 1551 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 1552 FlipMemInputs); 1553 continue; 1554 } else if (isa<CallInst>(I)) { 1555 Function *F = cast<CallInst>(I)->getCalledFunction(); 1556 unsigned IID = F->getIntrinsicID(); 1557 if (o == NumOperands-1) { 1558 BasicBlock &BB = *I->getParent(); 1559 1560 Module *M = BB.getParent()->getParent(); 1561 Type *ArgType = I->getType(); 1562 Type *VArgType = getVecTypeForPair(ArgType); 1563 1564 // FIXME: is it safe to do this here? 1565 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 1566 (Intrinsic::ID) IID, VArgType); 1567 continue; 1568 } else if (IID == Intrinsic::powi && o == 1) { 1569 // The second argument of powi is a single integer and we've already 1570 // checked that both arguments are equal. As a result, we just keep 1571 // I's second argument. 1572 ReplacedOperands[o] = I->getOperand(o); 1573 continue; 1574 } 1575 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 1576 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 1577 continue; 1578 } 1579 1580 ReplacedOperands[o] = 1581 getReplacementInput(Context, I, J, o, FlipMemInputs); 1582 } 1583 } 1584 1585 // This function creates two values that represent the outputs of the 1586 // original I and J instructions. These are generally vector shuffles 1587 // or extracts. In many cases, these will end up being unused and, thus, 1588 // eliminated by later passes. 1589 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 1590 Instruction *J, Instruction *K, 1591 Instruction *&InsertionPt, 1592 Instruction *&K1, Instruction *&K2, 1593 bool &FlipMemInputs) { 1594 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1595 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1596 1597 if (isa<StoreInst>(I)) { 1598 AA->replaceWithNewValue(I, K); 1599 AA->replaceWithNewValue(J, K); 1600 } else { 1601 Type *IType = I->getType(); 1602 Type *VType = getVecTypeForPair(IType); 1603 1604 if (IType->isVectorTy()) { 1605 unsigned numElem = cast<VectorType>(IType)->getNumElements(); 1606 std::vector<Constant*> Mask1(numElem), Mask2(numElem); 1607 for (unsigned v = 0; v < numElem; ++v) { 1608 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1609 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v); 1610 } 1611 1612 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 1613 ConstantVector::get( 1614 FlipMemInputs ? Mask2 : Mask1), 1615 getReplacementName(K, false, 1)); 1616 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 1617 ConstantVector::get( 1618 FlipMemInputs ? Mask1 : Mask2), 1619 getReplacementName(K, false, 2)); 1620 } else { 1621 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 1622 getReplacementName(K, false, 1)); 1623 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 1624 getReplacementName(K, false, 2)); 1625 } 1626 1627 K1->insertAfter(K); 1628 K2->insertAfter(K1); 1629 InsertionPt = K2; 1630 } 1631 } 1632 1633 // Move all uses of the function I (including pairing-induced uses) after J. 1634 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 1635 std::multimap<Value *, Value *> &LoadMoveSet, 1636 Instruction *I, Instruction *J) { 1637 // Skip to the first instruction past I. 1638 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1639 1640 DenseSet<Value *> Users; 1641 AliasSetTracker WriteSet(*AA); 1642 for (; cast<Instruction>(L) != J; ++L) 1643 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 1644 1645 assert(cast<Instruction>(L) == J && 1646 "Tracking has not proceeded far enough to check for dependencies"); 1647 // If J is now in the use set of I, then trackUsesOfI will return true 1648 // and we have a dependency cycle (and the fusing operation must abort). 1649 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 1650 } 1651 1652 // Move all uses of the function I (including pairing-induced uses) after J. 1653 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 1654 std::multimap<Value *, Value *> &LoadMoveSet, 1655 Instruction *&InsertionPt, 1656 Instruction *I, Instruction *J) { 1657 // Skip to the first instruction past I. 1658 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1659 1660 DenseSet<Value *> Users; 1661 AliasSetTracker WriteSet(*AA); 1662 for (; cast<Instruction>(L) != J;) { 1663 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 1664 // Move this instruction 1665 Instruction *InstToMove = L; ++L; 1666 1667 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 1668 " to after " << *InsertionPt << "\n"); 1669 InstToMove->removeFromParent(); 1670 InstToMove->insertAfter(InsertionPt); 1671 InsertionPt = InstToMove; 1672 } else { 1673 ++L; 1674 } 1675 } 1676 } 1677 1678 // Collect all load instruction that are in the move set of a given first 1679 // pair member. These loads depend on the first instruction, I, and so need 1680 // to be moved after J (the second instruction) when the pair is fused. 1681 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 1682 DenseMap<Value *, Value *> &ChosenPairs, 1683 std::multimap<Value *, Value *> &LoadMoveSet, 1684 Instruction *I) { 1685 // Skip to the first instruction past I. 1686 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1687 1688 DenseSet<Value *> Users; 1689 AliasSetTracker WriteSet(*AA); 1690 1691 // Note: We cannot end the loop when we reach J because J could be moved 1692 // farther down the use chain by another instruction pairing. Also, J 1693 // could be before I if this is an inverted input. 1694 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 1695 if (trackUsesOfI(Users, WriteSet, I, L)) { 1696 if (L->mayReadFromMemory()) 1697 LoadMoveSet.insert(ValuePair(L, I)); 1698 } 1699 } 1700 } 1701 1702 // In cases where both load/stores and the computation of their pointers 1703 // are chosen for vectorization, we can end up in a situation where the 1704 // aliasing analysis starts returning different query results as the 1705 // process of fusing instruction pairs continues. Because the algorithm 1706 // relies on finding the same use trees here as were found earlier, we'll 1707 // need to precompute the necessary aliasing information here and then 1708 // manually update it during the fusion process. 1709 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 1710 std::vector<Value *> &PairableInsts, 1711 DenseMap<Value *, Value *> &ChosenPairs, 1712 std::multimap<Value *, Value *> &LoadMoveSet) { 1713 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1714 PIE = PairableInsts.end(); PI != PIE; ++PI) { 1715 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 1716 if (P == ChosenPairs.end()) continue; 1717 1718 Instruction *I = cast<Instruction>(P->first); 1719 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 1720 } 1721 } 1722 1723 // This function fuses the chosen instruction pairs into vector instructions, 1724 // taking care preserve any needed scalar outputs and, then, it reorders the 1725 // remaining instructions as needed (users of the first member of the pair 1726 // need to be moved to after the location of the second member of the pair 1727 // because the vector instruction is inserted in the location of the pair's 1728 // second member). 1729 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 1730 std::vector<Value *> &PairableInsts, 1731 DenseMap<Value *, Value *> &ChosenPairs) { 1732 LLVMContext& Context = BB.getContext(); 1733 1734 // During the vectorization process, the order of the pairs to be fused 1735 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 1736 // list. After a pair is fused, the flipped pair is removed from the list. 1737 std::vector<ValuePair> FlippedPairs; 1738 FlippedPairs.reserve(ChosenPairs.size()); 1739 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 1740 E = ChosenPairs.end(); P != E; ++P) 1741 FlippedPairs.push_back(ValuePair(P->second, P->first)); 1742 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 1743 E = FlippedPairs.end(); P != E; ++P) 1744 ChosenPairs.insert(*P); 1745 1746 std::multimap<Value *, Value *> LoadMoveSet; 1747 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 1748 1749 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 1750 1751 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 1752 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 1753 if (P == ChosenPairs.end()) { 1754 ++PI; 1755 continue; 1756 } 1757 1758 if (getDepthFactor(P->first) == 0) { 1759 // These instructions are not really fused, but are tracked as though 1760 // they are. Any case in which it would be interesting to fuse them 1761 // will be taken care of by InstCombine. 1762 --NumFusedOps; 1763 ++PI; 1764 continue; 1765 } 1766 1767 Instruction *I = cast<Instruction>(P->first), 1768 *J = cast<Instruction>(P->second); 1769 1770 DEBUG(dbgs() << "BBV: fusing: " << *I << 1771 " <-> " << *J << "\n"); 1772 1773 // Remove the pair and flipped pair from the list. 1774 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 1775 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 1776 ChosenPairs.erase(FP); 1777 ChosenPairs.erase(P); 1778 1779 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 1780 DEBUG(dbgs() << "BBV: fusion of: " << *I << 1781 " <-> " << *J << 1782 " aborted because of non-trivial dependency cycle\n"); 1783 --NumFusedOps; 1784 ++PI; 1785 continue; 1786 } 1787 1788 bool FlipMemInputs; 1789 unsigned NumOperands = I->getNumOperands(); 1790 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 1791 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 1792 FlipMemInputs); 1793 1794 // Make a copy of the original operation, change its type to the vector 1795 // type and replace its operands with the vector operands. 1796 Instruction *K = I->clone(); 1797 if (I->hasName()) K->takeName(I); 1798 1799 if (!isa<StoreInst>(K)) 1800 K->mutateType(getVecTypeForPair(I->getType())); 1801 1802 for (unsigned o = 0; o < NumOperands; ++o) 1803 K->setOperand(o, ReplacedOperands[o]); 1804 1805 // If we've flipped the memory inputs, make sure that we take the correct 1806 // alignment. 1807 if (FlipMemInputs) { 1808 if (isa<StoreInst>(K)) 1809 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 1810 else 1811 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 1812 } 1813 1814 K->insertAfter(J); 1815 1816 // Instruction insertion point: 1817 Instruction *InsertionPt = K; 1818 Instruction *K1 = 0, *K2 = 0; 1819 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 1820 FlipMemInputs); 1821 1822 // The use tree of the first original instruction must be moved to after 1823 // the location of the second instruction. The entire use tree of the 1824 // first instruction is disjoint from the input tree of the second 1825 // (by definition), and so commutes with it. 1826 1827 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 1828 1829 if (!isa<StoreInst>(I)) { 1830 I->replaceAllUsesWith(K1); 1831 J->replaceAllUsesWith(K2); 1832 AA->replaceWithNewValue(I, K1); 1833 AA->replaceWithNewValue(J, K2); 1834 } 1835 1836 // Instructions that may read from memory may be in the load move set. 1837 // Once an instruction is fused, we no longer need its move set, and so 1838 // the values of the map never need to be updated. However, when a load 1839 // is fused, we need to merge the entries from both instructions in the 1840 // pair in case those instructions were in the move set of some other 1841 // yet-to-be-fused pair. The loads in question are the keys of the map. 1842 if (I->mayReadFromMemory()) { 1843 std::vector<ValuePair> NewSetMembers; 1844 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 1845 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 1846 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 1847 N != IPairRange.second; ++N) 1848 NewSetMembers.push_back(ValuePair(K, N->second)); 1849 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 1850 N != JPairRange.second; ++N) 1851 NewSetMembers.push_back(ValuePair(K, N->second)); 1852 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 1853 AE = NewSetMembers.end(); A != AE; ++A) 1854 LoadMoveSet.insert(*A); 1855 } 1856 1857 // Before removing I, set the iterator to the next instruction. 1858 PI = llvm::next(BasicBlock::iterator(I)); 1859 if (cast<Instruction>(PI) == J) 1860 ++PI; 1861 1862 SE->forgetValue(I); 1863 SE->forgetValue(J); 1864 I->eraseFromParent(); 1865 J->eraseFromParent(); 1866 } 1867 1868 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 1869 } 1870} 1871 1872char BBVectorize::ID = 0; 1873static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 1874INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1875INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 1876INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 1877INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1878 1879BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 1880 return new BBVectorize(C); 1881} 1882 1883bool 1884llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 1885 BBVectorize BBVectorizer(P, C); 1886 return BBVectorizer.vectorizeBB(BB); 1887} 1888 1889//===----------------------------------------------------------------------===// 1890VectorizeConfig::VectorizeConfig() { 1891 VectorBits = ::VectorBits; 1892 VectorizeInts = !::NoInts; 1893 VectorizeFloats = !::NoFloats; 1894 VectorizeCasts = !::NoCasts; 1895 VectorizeMath = !::NoMath; 1896 VectorizeFMA = !::NoFMA; 1897 VectorizeMemOps = !::NoMemOps; 1898 AlignedOnly = ::AlignedOnly; 1899 ReqChainDepth= ::ReqChainDepth; 1900 SearchLimit = ::SearchLimit; 1901 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 1902 SplatBreaksChain = ::SplatBreaksChain; 1903 MaxInsts = ::MaxInsts; 1904 MaxIter = ::MaxIter; 1905 NoMemOpBoost = ::NoMemOpBoost; 1906 FastDep = ::FastDep; 1907} 1908