BBVectorize.cpp revision 5c05f2b99e6d377786b4384925e61577f385f49c
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 BBVectorize() : BasicBlockPass(ID) { 144 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 145 } 146 147 typedef std::pair<Value *, Value *> ValuePair; 148 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 149 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 150 typedef std::pair<std::multimap<Value *, Value *>::iterator, 151 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 152 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 153 std::multimap<ValuePair, ValuePair>::iterator> 154 VPPIteratorPair; 155 156 AliasAnalysis *AA; 157 ScalarEvolution *SE; 158 TargetData *TD; 159 160 // FIXME: const correct? 161 162 bool vectorizePairs(BasicBlock &BB); 163 164 bool getCandidatePairs(BasicBlock &BB, 165 BasicBlock::iterator &Start, 166 std::multimap<Value *, Value *> &CandidatePairs, 167 std::vector<Value *> &PairableInsts); 168 169 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 170 std::vector<Value *> &PairableInsts, 171 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 172 173 void buildDepMap(BasicBlock &BB, 174 std::multimap<Value *, Value *> &CandidatePairs, 175 std::vector<Value *> &PairableInsts, 176 DenseSet<ValuePair> &PairableInstUsers); 177 178 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 179 std::vector<Value *> &PairableInsts, 180 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 181 DenseSet<ValuePair> &PairableInstUsers, 182 DenseMap<Value *, Value *>& ChosenPairs); 183 184 void fuseChosenPairs(BasicBlock &BB, 185 std::vector<Value *> &PairableInsts, 186 DenseMap<Value *, Value *>& ChosenPairs); 187 188 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 189 190 bool areInstsCompatible(Instruction *I, Instruction *J, 191 bool IsSimpleLoadStore); 192 193 bool trackUsesOfI(DenseSet<Value *> &Users, 194 AliasSetTracker &WriteSet, Instruction *I, 195 Instruction *J, bool UpdateUsers = true, 196 std::multimap<Value *, Value *> *LoadMoveSet = 0); 197 198 void computePairsConnectedTo( 199 std::multimap<Value *, Value *> &CandidatePairs, 200 std::vector<Value *> &PairableInsts, 201 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 202 ValuePair P); 203 204 bool pairsConflict(ValuePair P, ValuePair Q, 205 DenseSet<ValuePair> &PairableInstUsers, 206 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 207 208 bool pairWillFormCycle(ValuePair P, 209 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 210 DenseSet<ValuePair> &CurrentPairs); 211 212 void pruneTreeFor( 213 std::multimap<Value *, Value *> &CandidatePairs, 214 std::vector<Value *> &PairableInsts, 215 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 216 DenseSet<ValuePair> &PairableInstUsers, 217 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 218 DenseMap<Value *, Value *> &ChosenPairs, 219 DenseMap<ValuePair, size_t> &Tree, 220 DenseSet<ValuePair> &PrunedTree, ValuePair J, 221 bool UseCycleCheck); 222 223 void buildInitialTreeFor( 224 std::multimap<Value *, Value *> &CandidatePairs, 225 std::vector<Value *> &PairableInsts, 226 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 227 DenseSet<ValuePair> &PairableInstUsers, 228 DenseMap<Value *, Value *> &ChosenPairs, 229 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 230 231 void findBestTreeFor( 232 std::multimap<Value *, Value *> &CandidatePairs, 233 std::vector<Value *> &PairableInsts, 234 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 235 DenseSet<ValuePair> &PairableInstUsers, 236 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 237 DenseMap<Value *, Value *> &ChosenPairs, 238 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 239 size_t &BestEffSize, VPIteratorPair ChoiceRange, 240 bool UseCycleCheck); 241 242 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 243 Instruction *J, unsigned o, bool &FlipMemInputs); 244 245 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 246 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 247 unsigned IdxOffset, std::vector<Constant*> &Mask); 248 249 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 250 Instruction *J); 251 252 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 253 Instruction *J, unsigned o, bool FlipMemInputs); 254 255 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 256 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 257 bool &FlipMemInputs); 258 259 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 260 Instruction *J, Instruction *K, 261 Instruction *&InsertionPt, Instruction *&K1, 262 Instruction *&K2, bool &FlipMemInputs); 263 264 void collectPairLoadMoveSet(BasicBlock &BB, 265 DenseMap<Value *, Value *> &ChosenPairs, 266 std::multimap<Value *, Value *> &LoadMoveSet, 267 Instruction *I); 268 269 void collectLoadMoveSet(BasicBlock &BB, 270 std::vector<Value *> &PairableInsts, 271 DenseMap<Value *, Value *> &ChosenPairs, 272 std::multimap<Value *, Value *> &LoadMoveSet); 273 274 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 275 std::multimap<Value *, Value *> &LoadMoveSet, 276 Instruction *I, Instruction *J); 277 278 void moveUsesOfIAfterJ(BasicBlock &BB, 279 std::multimap<Value *, Value *> &LoadMoveSet, 280 Instruction *&InsertionPt, 281 Instruction *I, Instruction *J); 282 283 virtual bool runOnBasicBlock(BasicBlock &BB) { 284 AA = &getAnalysis<AliasAnalysis>(); 285 SE = &getAnalysis<ScalarEvolution>(); 286 TD = getAnalysisIfAvailable<TargetData>(); 287 288 bool changed = false; 289 // Iterate a sufficient number of times to merge types of size 1 bit, 290 // then 2 bits, then 4, etc. up to half of the target vector width of the 291 // target vector register. 292 for (unsigned v = 2, n = 1; v <= VectorBits && (!MaxIter || n <= MaxIter); 293 v *= 2, ++n) { 294 DEBUG(dbgs() << "BBV: fusing loop #" << n << 295 " for " << BB.getName() << " in " << 296 BB.getParent()->getName() << "...\n"); 297 if (vectorizePairs(BB)) 298 changed = true; 299 else 300 break; 301 } 302 303 DEBUG(dbgs() << "BBV: done!\n"); 304 return changed; 305 } 306 307 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 308 BasicBlockPass::getAnalysisUsage(AU); 309 AU.addRequired<AliasAnalysis>(); 310 AU.addRequired<ScalarEvolution>(); 311 AU.addPreserved<AliasAnalysis>(); 312 AU.addPreserved<ScalarEvolution>(); 313 AU.setPreservesCFG(); 314 } 315 316 // This returns the vector type that holds a pair of the provided type. 317 // If the provided type is already a vector, then its length is doubled. 318 static inline VectorType *getVecTypeForPair(Type *ElemTy) { 319 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 320 unsigned numElem = VTy->getNumElements(); 321 return VectorType::get(ElemTy->getScalarType(), numElem*2); 322 } 323 324 return VectorType::get(ElemTy, 2); 325 } 326 327 // Returns the weight associated with the provided value. A chain of 328 // candidate pairs has a length given by the sum of the weights of its 329 // members (one weight per pair; the weight of each member of the pair 330 // is assumed to be the same). This length is then compared to the 331 // chain-length threshold to determine if a given chain is significant 332 // enough to be vectorized. The length is also used in comparing 333 // candidate chains where longer chains are considered to be better. 334 // Note: when this function returns 0, the resulting instructions are 335 // not actually fused. 336 static inline size_t getDepthFactor(Value *V) { 337 // InsertElement and ExtractElement have a depth factor of zero. This is 338 // for two reasons: First, they cannot be usefully fused. Second, because 339 // the pass generates a lot of these, they can confuse the simple metric 340 // used to compare the trees in the next iteration. Thus, giving them a 341 // weight of zero allows the pass to essentially ignore them in 342 // subsequent iterations when looking for vectorization opportunities 343 // while still tracking dependency chains that flow through those 344 // instructions. 345 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 346 return 0; 347 348 // Give a load or store half of the required depth so that load/store 349 // pairs will vectorize. 350 if (!NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 351 return ReqChainDepth/2; 352 353 return 1; 354 } 355 356 // This determines the relative offset of two loads or stores, returning 357 // true if the offset could be determined to be some constant value. 358 // For example, if OffsetInElmts == 1, then J accesses the memory directly 359 // after I; if OffsetInElmts == -1 then I accesses the memory 360 // directly after J. This function assumes that both instructions 361 // have the same type. 362 bool getPairPtrInfo(Instruction *I, Instruction *J, 363 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 364 int64_t &OffsetInElmts) { 365 OffsetInElmts = 0; 366 if (isa<LoadInst>(I)) { 367 IPtr = cast<LoadInst>(I)->getPointerOperand(); 368 JPtr = cast<LoadInst>(J)->getPointerOperand(); 369 IAlignment = cast<LoadInst>(I)->getAlignment(); 370 JAlignment = cast<LoadInst>(J)->getAlignment(); 371 } else { 372 IPtr = cast<StoreInst>(I)->getPointerOperand(); 373 JPtr = cast<StoreInst>(J)->getPointerOperand(); 374 IAlignment = cast<StoreInst>(I)->getAlignment(); 375 JAlignment = cast<StoreInst>(J)->getAlignment(); 376 } 377 378 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 379 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 380 381 // If this is a trivial offset, then we'll get something like 382 // 1*sizeof(type). With target data, which we need anyway, this will get 383 // constant folded into a number. 384 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 385 if (const SCEVConstant *ConstOffSCEV = 386 dyn_cast<SCEVConstant>(OffsetSCEV)) { 387 ConstantInt *IntOff = ConstOffSCEV->getValue(); 388 int64_t Offset = IntOff->getSExtValue(); 389 390 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 391 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 392 393 assert(VTy == cast<PointerType>(JPtr->getType())->getElementType()); 394 395 OffsetInElmts = Offset/VTyTSS; 396 return (abs64(Offset) % VTyTSS) == 0; 397 } 398 399 return false; 400 } 401 402 // Returns true if the provided CallInst represents an intrinsic that can 403 // be vectorized. 404 bool isVectorizableIntrinsic(CallInst* I) { 405 Function *F = I->getCalledFunction(); 406 if (!F) return false; 407 408 unsigned IID = F->getIntrinsicID(); 409 if (!IID) return false; 410 411 switch(IID) { 412 default: 413 return false; 414 case Intrinsic::sqrt: 415 case Intrinsic::powi: 416 case Intrinsic::sin: 417 case Intrinsic::cos: 418 case Intrinsic::log: 419 case Intrinsic::log2: 420 case Intrinsic::log10: 421 case Intrinsic::exp: 422 case Intrinsic::exp2: 423 case Intrinsic::pow: 424 return !NoMath; 425 case Intrinsic::fma: 426 return !NoFMA; 427 } 428 } 429 430 // Returns true if J is the second element in some pair referenced by 431 // some multimap pair iterator pair. 432 template <typename V> 433 bool isSecondInIteratorPair(V J, std::pair< 434 typename std::multimap<V, V>::iterator, 435 typename std::multimap<V, V>::iterator> PairRange) { 436 for (typename std::multimap<V, V>::iterator K = PairRange.first; 437 K != PairRange.second; ++K) 438 if (K->second == J) return true; 439 440 return false; 441 } 442 }; 443 444 // This function implements one vectorization iteration on the provided 445 // basic block. It returns true if the block is changed. 446 bool BBVectorize::vectorizePairs(BasicBlock &BB) { 447 bool ShouldContinue; 448 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 449 450 std::vector<Value *> AllPairableInsts; 451 DenseMap<Value *, Value *> AllChosenPairs; 452 453 do { 454 std::vector<Value *> PairableInsts; 455 std::multimap<Value *, Value *> CandidatePairs; 456 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 457 PairableInsts); 458 if (PairableInsts.empty()) continue; 459 460 // Now we have a map of all of the pairable instructions and we need to 461 // select the best possible pairing. A good pairing is one such that the 462 // users of the pair are also paired. This defines a (directed) forest 463 // over the pairs such that two pairs are connected iff the second pair 464 // uses the first. 465 466 // Note that it only matters that both members of the second pair use some 467 // element of the first pair (to allow for splatting). 468 469 std::multimap<ValuePair, ValuePair> ConnectedPairs; 470 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 471 if (ConnectedPairs.empty()) continue; 472 473 // Build the pairable-instruction dependency map 474 DenseSet<ValuePair> PairableInstUsers; 475 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 476 477 // There is now a graph of the connected pairs. For each variable, pick 478 // the pairing with the largest tree meeting the depth requirement on at 479 // least one branch. Then select all pairings that are part of that tree 480 // and remove them from the list of available pairings and pairable 481 // variables. 482 483 DenseMap<Value *, Value *> ChosenPairs; 484 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, 485 PairableInstUsers, ChosenPairs); 486 487 if (ChosenPairs.empty()) continue; 488 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 489 PairableInsts.end()); 490 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 491 } while (ShouldContinue); 492 493 if (AllChosenPairs.empty()) return false; 494 NumFusedOps += AllChosenPairs.size(); 495 496 // A set of pairs has now been selected. It is now necessary to replace the 497 // paired instructions with vector instructions. For this procedure each 498 // operand much be replaced with a vector operand. This vector is formed 499 // by using build_vector on the old operands. The replaced values are then 500 // replaced with a vector_extract on the result. Subsequent optimization 501 // passes should coalesce the build/extract combinations. 502 503 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 504 return true; 505 } 506 507 // This function returns true if the provided instruction is capable of being 508 // fused into a vector instruction. This determination is based only on the 509 // type and other attributes of the instruction. 510 bool BBVectorize::isInstVectorizable(Instruction *I, 511 bool &IsSimpleLoadStore) { 512 IsSimpleLoadStore = false; 513 514 if (CallInst *C = dyn_cast<CallInst>(I)) { 515 if (!isVectorizableIntrinsic(C)) 516 return false; 517 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 518 // Vectorize simple loads if possbile: 519 IsSimpleLoadStore = L->isSimple(); 520 if (!IsSimpleLoadStore || NoMemOps) 521 return false; 522 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 523 // Vectorize simple stores if possbile: 524 IsSimpleLoadStore = S->isSimple(); 525 if (!IsSimpleLoadStore || NoMemOps) 526 return false; 527 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 528 // We can vectorize casts, but not casts of pointer types, etc. 529 if (NoCasts) 530 return false; 531 532 Type *SrcTy = C->getSrcTy(); 533 if (!SrcTy->isSingleValueType() || SrcTy->isPointerTy()) 534 return false; 535 536 Type *DestTy = C->getDestTy(); 537 if (!DestTy->isSingleValueType() || DestTy->isPointerTy()) 538 return false; 539 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 540 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 541 return false; 542 } 543 544 // We can't vectorize memory operations without target data 545 if (TD == 0 && IsSimpleLoadStore) 546 return false; 547 548 Type *T1, *T2; 549 if (isa<StoreInst>(I)) { 550 // For stores, it is the value type, not the pointer type that matters 551 // because the value is what will come from a vector register. 552 553 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 554 T1 = IVal->getType(); 555 } else { 556 T1 = I->getType(); 557 } 558 559 if (I->isCast()) 560 T2 = cast<CastInst>(I)->getSrcTy(); 561 else 562 T2 = T1; 563 564 // Not every type can be vectorized... 565 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 566 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 567 return false; 568 569 if (NoInts && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) 570 return false; 571 572 if (NoFloats && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 573 return false; 574 575 if (T1->getPrimitiveSizeInBits() > VectorBits/2 || 576 T2->getPrimitiveSizeInBits() > VectorBits/2) 577 return false; 578 579 return true; 580 } 581 582 // This function returns true if the two provided instructions are compatible 583 // (meaning that they can be fused into a vector instruction). This assumes 584 // that I has already been determined to be vectorizable and that J is not 585 // in the use tree of I. 586 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 587 bool IsSimpleLoadStore) { 588 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 589 " <-> " << *J << "\n"); 590 591 // Loads and stores can be merged if they have different alignments, 592 // but are otherwise the same. 593 LoadInst *LI, *LJ; 594 StoreInst *SI, *SJ; 595 if ((LI = dyn_cast<LoadInst>(I)) && (LJ = dyn_cast<LoadInst>(J))) { 596 if (I->getType() != J->getType()) 597 return false; 598 599 if (LI->getPointerOperand()->getType() != 600 LJ->getPointerOperand()->getType() || 601 LI->isVolatile() != LJ->isVolatile() || 602 LI->getOrdering() != LJ->getOrdering() || 603 LI->getSynchScope() != LJ->getSynchScope()) 604 return false; 605 } else if ((SI = dyn_cast<StoreInst>(I)) && (SJ = dyn_cast<StoreInst>(J))) { 606 if (SI->getValueOperand()->getType() != 607 SJ->getValueOperand()->getType() || 608 SI->getPointerOperand()->getType() != 609 SJ->getPointerOperand()->getType() || 610 SI->isVolatile() != SJ->isVolatile() || 611 SI->getOrdering() != SJ->getOrdering() || 612 SI->getSynchScope() != SJ->getSynchScope()) 613 return false; 614 } else if (!J->isSameOperationAs(I)) { 615 return false; 616 } 617 // FIXME: handle addsub-type operations! 618 619 if (IsSimpleLoadStore) { 620 Value *IPtr, *JPtr; 621 unsigned IAlignment, JAlignment; 622 int64_t OffsetInElmts = 0; 623 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 624 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 625 if (AlignedOnly) { 626 Type *aType = isa<StoreInst>(I) ? 627 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 628 // An aligned load or store is possible only if the instruction 629 // with the lower offset has an alignment suitable for the 630 // vector type. 631 632 unsigned BottomAlignment = IAlignment; 633 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 634 635 Type *VType = getVecTypeForPair(aType); 636 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 637 if (BottomAlignment < VecAlignment) 638 return false; 639 } 640 } else { 641 return false; 642 } 643 } else if (isa<ShuffleVectorInst>(I)) { 644 // Only merge two shuffles if they're both constant 645 return isa<Constant>(I->getOperand(2)) && 646 isa<Constant>(J->getOperand(2)); 647 // FIXME: We may want to vectorize non-constant shuffles also. 648 } 649 650 return true; 651 } 652 653 // Figure out whether or not J uses I and update the users and write-set 654 // structures associated with I. Specifically, Users represents the set of 655 // instructions that depend on I. WriteSet represents the set 656 // of memory locations that are dependent on I. If UpdateUsers is true, 657 // and J uses I, then Users is updated to contain J and WriteSet is updated 658 // to contain any memory locations to which J writes. The function returns 659 // true if J uses I. By default, alias analysis is used to determine 660 // whether J reads from memory that overlaps with a location in WriteSet. 661 // If LoadMoveSet is not null, then it is a previously-computed multimap 662 // where the key is the memory-based user instruction and the value is 663 // the instruction to be compared with I. So, if LoadMoveSet is provided, 664 // then the alias analysis is not used. This is necessary because this 665 // function is called during the process of moving instructions during 666 // vectorization and the results of the alias analysis are not stable during 667 // that process. 668 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 669 AliasSetTracker &WriteSet, Instruction *I, 670 Instruction *J, bool UpdateUsers, 671 std::multimap<Value *, Value *> *LoadMoveSet) { 672 bool UsesI = false; 673 674 // This instruction may already be marked as a user due, for example, to 675 // being a member of a selected pair. 676 if (Users.count(J)) 677 UsesI = true; 678 679 if (!UsesI) 680 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 681 JU != JE; ++JU) { 682 Value *V = *JU; 683 if (I == V || Users.count(V)) { 684 UsesI = true; 685 break; 686 } 687 } 688 if (!UsesI && J->mayReadFromMemory()) { 689 if (LoadMoveSet) { 690 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 691 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 692 } else { 693 for (AliasSetTracker::iterator W = WriteSet.begin(), 694 WE = WriteSet.end(); W != WE; ++W) { 695 for (AliasSet::iterator A = W->begin(), AE = W->end(); 696 A != AE; ++A) { 697 AliasAnalysis::Location ptrLoc(A->getValue(), A->getSize(), 698 A->getTBAAInfo()); 699 if (AA->getModRefInfo(J, ptrLoc) != AliasAnalysis::NoModRef) { 700 UsesI = true; 701 break; 702 } 703 } 704 if (UsesI) break; 705 } 706 } 707 } 708 709 if (UsesI && UpdateUsers) { 710 if (J->mayWriteToMemory()) WriteSet.add(J); 711 Users.insert(J); 712 } 713 714 return UsesI; 715 } 716 717 // This function iterates over all instruction pairs in the provided 718 // basic block and collects all candidate pairs for vectorization. 719 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 720 BasicBlock::iterator &Start, 721 std::multimap<Value *, Value *> &CandidatePairs, 722 std::vector<Value *> &PairableInsts) { 723 BasicBlock::iterator E = BB.end(); 724 if (Start == E) return false; 725 726 bool ShouldContinue = false, IAfterStart = false; 727 for (BasicBlock::iterator I = Start++; I != E; ++I) { 728 if (I == Start) IAfterStart = true; 729 730 bool IsSimpleLoadStore; 731 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 732 733 // Look for an instruction with which to pair instruction *I... 734 DenseSet<Value *> Users; 735 AliasSetTracker WriteSet(*AA); 736 bool JAfterStart = IAfterStart; 737 BasicBlock::iterator J = llvm::next(I); 738 for (unsigned ss = 0; J != E && ss <= SearchLimit; ++J, ++ss) { 739 if (J == Start) JAfterStart = true; 740 741 // Determine if J uses I, if so, exit the loop. 742 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !FastDep); 743 if (FastDep) { 744 // Note: For this heuristic to be effective, independent operations 745 // must tend to be intermixed. This is likely to be true from some 746 // kinds of grouped loop unrolling (but not the generic LLVM pass), 747 // but otherwise may require some kind of reordering pass. 748 749 // When using fast dependency analysis, 750 // stop searching after first use: 751 if (UsesI) break; 752 } else { 753 if (UsesI) continue; 754 } 755 756 // J does not use I, and comes before the first use of I, so it can be 757 // merged with I if the instructions are compatible. 758 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue; 759 760 // J is a candidate for merging with I. 761 if (!PairableInsts.size() || 762 PairableInsts[PairableInsts.size()-1] != I) { 763 PairableInsts.push_back(I); 764 } 765 766 CandidatePairs.insert(ValuePair(I, J)); 767 768 // The next call to this function must start after the last instruction 769 // selected during this invocation. 770 if (JAfterStart) { 771 Start = llvm::next(J); 772 IAfterStart = JAfterStart = false; 773 } 774 775 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 776 << *I << " <-> " << *J << "\n"); 777 778 // If we have already found too many pairs, break here and this function 779 // will be called again starting after the last instruction selected 780 // during this invocation. 781 if (PairableInsts.size() >= MaxInsts) { 782 ShouldContinue = true; 783 break; 784 } 785 } 786 787 if (ShouldContinue) 788 break; 789 } 790 791 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 792 << " instructions with candidate pairs\n"); 793 794 return ShouldContinue; 795 } 796 797 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 798 // it looks for pairs such that both members have an input which is an 799 // output of PI or PJ. 800 void BBVectorize::computePairsConnectedTo( 801 std::multimap<Value *, Value *> &CandidatePairs, 802 std::vector<Value *> &PairableInsts, 803 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 804 ValuePair P) { 805 // For each possible pairing for this variable, look at the uses of 806 // the first value... 807 for (Value::use_iterator I = P.first->use_begin(), 808 E = P.first->use_end(); I != E; ++I) { 809 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 810 811 // For each use of the first variable, look for uses of the second 812 // variable... 813 for (Value::use_iterator J = P.second->use_begin(), 814 E2 = P.second->use_end(); J != E2; ++J) { 815 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 816 817 // Look for <I, J>: 818 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 819 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 820 821 // Look for <J, I>: 822 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 823 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 824 } 825 826 if (SplatBreaksChain) continue; 827 // Look for cases where just the first value in the pair is used by 828 // both members of another pair (splatting). 829 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 830 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 831 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 832 } 833 } 834 835 if (SplatBreaksChain) return; 836 // Look for cases where just the second value in the pair is used by 837 // both members of another pair (splatting). 838 for (Value::use_iterator I = P.second->use_begin(), 839 E = P.second->use_end(); I != E; ++I) { 840 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 841 842 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 843 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 844 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 845 } 846 } 847 } 848 849 // This function figures out which pairs are connected. Two pairs are 850 // connected if some output of the first pair forms an input to both members 851 // of the second pair. 852 void BBVectorize::computeConnectedPairs( 853 std::multimap<Value *, Value *> &CandidatePairs, 854 std::vector<Value *> &PairableInsts, 855 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 856 857 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 858 PE = PairableInsts.end(); PI != PE; ++PI) { 859 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 860 861 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 862 P != choiceRange.second; ++P) 863 computePairsConnectedTo(CandidatePairs, PairableInsts, 864 ConnectedPairs, *P); 865 } 866 867 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 868 << " pair connections.\n"); 869 } 870 871 // This function builds a set of use tuples such that <A, B> is in the set 872 // if B is in the use tree of A. If B is in the use tree of A, then B 873 // depends on the output of A. 874 void BBVectorize::buildDepMap( 875 BasicBlock &BB, 876 std::multimap<Value *, Value *> &CandidatePairs, 877 std::vector<Value *> &PairableInsts, 878 DenseSet<ValuePair> &PairableInstUsers) { 879 DenseSet<Value *> IsInPair; 880 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 881 E = CandidatePairs.end(); C != E; ++C) { 882 IsInPair.insert(C->first); 883 IsInPair.insert(C->second); 884 } 885 886 // Iterate through the basic block, recording all Users of each 887 // pairable instruction. 888 889 BasicBlock::iterator E = BB.end(); 890 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 891 if (IsInPair.find(I) == IsInPair.end()) continue; 892 893 DenseSet<Value *> Users; 894 AliasSetTracker WriteSet(*AA); 895 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 896 (void) trackUsesOfI(Users, WriteSet, I, J); 897 898 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 899 U != E; ++U) 900 PairableInstUsers.insert(ValuePair(I, *U)); 901 } 902 } 903 904 // Returns true if an input to pair P is an output of pair Q and also an 905 // input of pair Q is an output of pair P. If this is the case, then these 906 // two pairs cannot be simultaneously fused. 907 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 908 DenseSet<ValuePair> &PairableInstUsers, 909 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 910 // Two pairs are in conflict if they are mutual Users of eachother. 911 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 912 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 913 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 914 PairableInstUsers.count(ValuePair(P.second, Q.second)); 915 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 916 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 917 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 918 PairableInstUsers.count(ValuePair(Q.second, P.second)); 919 if (PairableInstUserMap) { 920 // FIXME: The expensive part of the cycle check is not so much the cycle 921 // check itself but this edge insertion procedure. This needs some 922 // profiling and probably a different data structure (same is true of 923 // most uses of std::multimap). 924 if (PUsesQ) { 925 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 926 if (!isSecondInIteratorPair(P, QPairRange)) 927 PairableInstUserMap->insert(VPPair(Q, P)); 928 } 929 if (QUsesP) { 930 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 931 if (!isSecondInIteratorPair(Q, PPairRange)) 932 PairableInstUserMap->insert(VPPair(P, Q)); 933 } 934 } 935 936 return (QUsesP && PUsesQ); 937 } 938 939 // This function walks the use graph of current pairs to see if, starting 940 // from P, the walk returns to P. 941 bool BBVectorize::pairWillFormCycle(ValuePair P, 942 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 943 DenseSet<ValuePair> &CurrentPairs) { 944 DEBUG(if (DebugCycleCheck) 945 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 946 << *P.second << "\n"); 947 // A lookup table of visisted pairs is kept because the PairableInstUserMap 948 // contains non-direct associations. 949 DenseSet<ValuePair> Visited; 950 SmallVector<ValuePair, 32> Q; 951 // General depth-first post-order traversal: 952 Q.push_back(P); 953 do { 954 ValuePair QTop = Q.pop_back_val(); 955 Visited.insert(QTop); 956 957 DEBUG(if (DebugCycleCheck) 958 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 959 << *QTop.second << "\n"); 960 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 961 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 962 C != QPairRange.second; ++C) { 963 if (C->second == P) { 964 DEBUG(dbgs() 965 << "BBV: rejected to prevent non-trivial cycle formation: " 966 << *C->first.first << " <-> " << *C->first.second << "\n"); 967 return true; 968 } 969 970 if (CurrentPairs.count(C->second) != 0 && 971 Visited.count(C->second) == 0) 972 Q.push_back(C->second); 973 } 974 } while (!Q.empty()); 975 976 return false; 977 } 978 979 // This function builds the initial tree of connected pairs with the 980 // pair J at the root. 981 void BBVectorize::buildInitialTreeFor( 982 std::multimap<Value *, Value *> &CandidatePairs, 983 std::vector<Value *> &PairableInsts, 984 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 985 DenseSet<ValuePair> &PairableInstUsers, 986 DenseMap<Value *, Value *> &ChosenPairs, 987 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 988 // Each of these pairs is viewed as the root node of a Tree. The Tree 989 // is then walked (depth-first). As this happens, we keep track of 990 // the pairs that compose the Tree and the maximum depth of the Tree. 991 SmallVector<ValuePairWithDepth, 32> Q; 992 // General depth-first post-order traversal: 993 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 994 do { 995 ValuePairWithDepth QTop = Q.back(); 996 997 // Push each child onto the queue: 998 bool MoreChildren = false; 999 size_t MaxChildDepth = QTop.second; 1000 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1001 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1002 k != qtRange.second; ++k) { 1003 // Make sure that this child pair is still a candidate: 1004 bool IsStillCand = false; 1005 VPIteratorPair checkRange = 1006 CandidatePairs.equal_range(k->second.first); 1007 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1008 m != checkRange.second; ++m) { 1009 if (m->second == k->second.second) { 1010 IsStillCand = true; 1011 break; 1012 } 1013 } 1014 1015 if (IsStillCand) { 1016 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1017 if (C == Tree.end()) { 1018 size_t d = getDepthFactor(k->second.first); 1019 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1020 MoreChildren = true; 1021 } else { 1022 MaxChildDepth = std::max(MaxChildDepth, C->second); 1023 } 1024 } 1025 } 1026 1027 if (!MoreChildren) { 1028 // Record the current pair as part of the Tree: 1029 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1030 Q.pop_back(); 1031 } 1032 } while (!Q.empty()); 1033 } 1034 1035 // Given some initial tree, prune it by removing conflicting pairs (pairs 1036 // that cannot be simultaneously chosen for vectorization). 1037 void BBVectorize::pruneTreeFor( 1038 std::multimap<Value *, Value *> &CandidatePairs, 1039 std::vector<Value *> &PairableInsts, 1040 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1041 DenseSet<ValuePair> &PairableInstUsers, 1042 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1043 DenseMap<Value *, Value *> &ChosenPairs, 1044 DenseMap<ValuePair, size_t> &Tree, 1045 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1046 bool UseCycleCheck) { 1047 SmallVector<ValuePairWithDepth, 32> Q; 1048 // General depth-first post-order traversal: 1049 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1050 do { 1051 ValuePairWithDepth QTop = Q.pop_back_val(); 1052 PrunedTree.insert(QTop.first); 1053 1054 // Visit each child, pruning as necessary... 1055 DenseMap<ValuePair, size_t> BestChilden; 1056 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1057 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1058 K != QTopRange.second; ++K) { 1059 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1060 if (C == Tree.end()) continue; 1061 1062 // This child is in the Tree, now we need to make sure it is the 1063 // best of any conflicting children. There could be multiple 1064 // conflicting children, so first, determine if we're keeping 1065 // this child, then delete conflicting children as necessary. 1066 1067 // It is also necessary to guard against pairing-induced 1068 // dependencies. Consider instructions a .. x .. y .. b 1069 // such that (a,b) are to be fused and (x,y) are to be fused 1070 // but a is an input to x and b is an output from y. This 1071 // means that y cannot be moved after b but x must be moved 1072 // after b for (a,b) to be fused. In other words, after 1073 // fusing (a,b) we have y .. a/b .. x where y is an input 1074 // to a/b and x is an output to a/b: x and y can no longer 1075 // be legally fused. To prevent this condition, we must 1076 // make sure that a child pair added to the Tree is not 1077 // both an input and output of an already-selected pair. 1078 1079 // Pairing-induced dependencies can also form from more complicated 1080 // cycles. The pair vs. pair conflicts are easy to check, and so 1081 // that is done explicitly for "fast rejection", and because for 1082 // child vs. child conflicts, we may prefer to keep the current 1083 // pair in preference to the already-selected child. 1084 DenseSet<ValuePair> CurrentPairs; 1085 1086 bool CanAdd = true; 1087 for (DenseMap<ValuePair, size_t>::iterator C2 1088 = BestChilden.begin(), E2 = BestChilden.end(); 1089 C2 != E2; ++C2) { 1090 if (C2->first.first == C->first.first || 1091 C2->first.first == C->first.second || 1092 C2->first.second == C->first.first || 1093 C2->first.second == C->first.second || 1094 pairsConflict(C2->first, C->first, PairableInstUsers, 1095 UseCycleCheck ? &PairableInstUserMap : 0)) { 1096 if (C2->second >= C->second) { 1097 CanAdd = false; 1098 break; 1099 } 1100 1101 CurrentPairs.insert(C2->first); 1102 } 1103 } 1104 if (!CanAdd) continue; 1105 1106 // Even worse, this child could conflict with another node already 1107 // selected for the Tree. If that is the case, ignore this child. 1108 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1109 E2 = PrunedTree.end(); T != E2; ++T) { 1110 if (T->first == C->first.first || 1111 T->first == C->first.second || 1112 T->second == C->first.first || 1113 T->second == C->first.second || 1114 pairsConflict(*T, C->first, PairableInstUsers, 1115 UseCycleCheck ? &PairableInstUserMap : 0)) { 1116 CanAdd = false; 1117 break; 1118 } 1119 1120 CurrentPairs.insert(*T); 1121 } 1122 if (!CanAdd) continue; 1123 1124 // And check the queue too... 1125 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1126 E2 = Q.end(); C2 != E2; ++C2) { 1127 if (C2->first.first == C->first.first || 1128 C2->first.first == C->first.second || 1129 C2->first.second == C->first.first || 1130 C2->first.second == C->first.second || 1131 pairsConflict(C2->first, C->first, PairableInstUsers, 1132 UseCycleCheck ? &PairableInstUserMap : 0)) { 1133 CanAdd = false; 1134 break; 1135 } 1136 1137 CurrentPairs.insert(C2->first); 1138 } 1139 if (!CanAdd) continue; 1140 1141 // Last but not least, check for a conflict with any of the 1142 // already-chosen pairs. 1143 for (DenseMap<Value *, Value *>::iterator C2 = 1144 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1145 C2 != E2; ++C2) { 1146 if (pairsConflict(*C2, C->first, PairableInstUsers, 1147 UseCycleCheck ? &PairableInstUserMap : 0)) { 1148 CanAdd = false; 1149 break; 1150 } 1151 1152 CurrentPairs.insert(*C2); 1153 } 1154 if (!CanAdd) continue; 1155 1156 // To check for non-trivial cycles formed by the addition of the 1157 // current pair we've formed a list of all relevant pairs, now use a 1158 // graph walk to check for a cycle. We start from the current pair and 1159 // walk the use tree to see if we again reach the current pair. If we 1160 // do, then the current pair is rejected. 1161 1162 // FIXME: It may be more efficient to use a topological-ordering 1163 // algorithm to improve the cycle check. This should be investigated. 1164 if (UseCycleCheck && 1165 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1166 continue; 1167 1168 // This child can be added, but we may have chosen it in preference 1169 // to an already-selected child. Check for this here, and if a 1170 // conflict is found, then remove the previously-selected child 1171 // before adding this one in its place. 1172 for (DenseMap<ValuePair, size_t>::iterator C2 1173 = BestChilden.begin(); C2 != BestChilden.end();) { 1174 if (C2->first.first == C->first.first || 1175 C2->first.first == C->first.second || 1176 C2->first.second == C->first.first || 1177 C2->first.second == C->first.second || 1178 pairsConflict(C2->first, C->first, PairableInstUsers)) 1179 BestChilden.erase(C2++); 1180 else 1181 ++C2; 1182 } 1183 1184 BestChilden.insert(ValuePairWithDepth(C->first, C->second)); 1185 } 1186 1187 for (DenseMap<ValuePair, size_t>::iterator C 1188 = BestChilden.begin(), E2 = BestChilden.end(); 1189 C != E2; ++C) { 1190 size_t DepthF = getDepthFactor(C->first.first); 1191 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1192 } 1193 } while (!Q.empty()); 1194 } 1195 1196 // This function finds the best tree of mututally-compatible connected 1197 // pairs, given the choice of root pairs as an iterator range. 1198 void BBVectorize::findBestTreeFor( 1199 std::multimap<Value *, Value *> &CandidatePairs, 1200 std::vector<Value *> &PairableInsts, 1201 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1202 DenseSet<ValuePair> &PairableInstUsers, 1203 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1204 DenseMap<Value *, Value *> &ChosenPairs, 1205 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1206 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1207 bool UseCycleCheck) { 1208 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1209 J != ChoiceRange.second; ++J) { 1210 1211 // Before going any further, make sure that this pair does not 1212 // conflict with any already-selected pairs (see comment below 1213 // near the Tree pruning for more details). 1214 DenseSet<ValuePair> ChosenPairSet; 1215 bool DoesConflict = false; 1216 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1217 E = ChosenPairs.end(); C != E; ++C) { 1218 if (pairsConflict(*C, *J, PairableInstUsers, 1219 UseCycleCheck ? &PairableInstUserMap : 0)) { 1220 DoesConflict = true; 1221 break; 1222 } 1223 1224 ChosenPairSet.insert(*C); 1225 } 1226 if (DoesConflict) continue; 1227 1228 if (UseCycleCheck && 1229 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1230 continue; 1231 1232 DenseMap<ValuePair, size_t> Tree; 1233 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1234 PairableInstUsers, ChosenPairs, Tree, *J); 1235 1236 // Because we'll keep the child with the largest depth, the largest 1237 // depth is still the same in the unpruned Tree. 1238 size_t MaxDepth = Tree.lookup(*J); 1239 1240 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1241 << *J->first << " <-> " << *J->second << "} of depth " << 1242 MaxDepth << " and size " << Tree.size() << "\n"); 1243 1244 // At this point the Tree has been constructed, but, may contain 1245 // contradictory children (meaning that different children of 1246 // some tree node may be attempting to fuse the same instruction). 1247 // So now we walk the tree again, in the case of a conflict, 1248 // keep only the child with the largest depth. To break a tie, 1249 // favor the first child. 1250 1251 DenseSet<ValuePair> PrunedTree; 1252 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1253 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1254 PrunedTree, *J, UseCycleCheck); 1255 1256 size_t EffSize = 0; 1257 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1258 E = PrunedTree.end(); S != E; ++S) 1259 EffSize += getDepthFactor(S->first); 1260 1261 DEBUG(if (DebugPairSelection) 1262 dbgs() << "BBV: found pruned Tree for pair {" 1263 << *J->first << " <-> " << *J->second << "} of depth " << 1264 MaxDepth << " and size " << PrunedTree.size() << 1265 " (effective size: " << EffSize << ")\n"); 1266 if (MaxDepth >= ReqChainDepth && EffSize > BestEffSize) { 1267 BestMaxDepth = MaxDepth; 1268 BestEffSize = EffSize; 1269 BestTree = PrunedTree; 1270 } 1271 } 1272 } 1273 1274 // Given the list of candidate pairs, this function selects those 1275 // that will be fused into vector instructions. 1276 void BBVectorize::choosePairs( 1277 std::multimap<Value *, Value *> &CandidatePairs, 1278 std::vector<Value *> &PairableInsts, 1279 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1280 DenseSet<ValuePair> &PairableInstUsers, 1281 DenseMap<Value *, Value *>& ChosenPairs) { 1282 bool UseCycleCheck = CandidatePairs.size() <= MaxCandPairsForCycleCheck; 1283 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1284 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1285 E = PairableInsts.end(); I != E; ++I) { 1286 // The number of possible pairings for this variable: 1287 size_t NumChoices = CandidatePairs.count(*I); 1288 if (!NumChoices) continue; 1289 1290 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1291 1292 // The best pair to choose and its tree: 1293 size_t BestMaxDepth = 0, BestEffSize = 0; 1294 DenseSet<ValuePair> BestTree; 1295 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1296 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1297 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1298 UseCycleCheck); 1299 1300 // A tree has been chosen (or not) at this point. If no tree was 1301 // chosen, then this instruction, I, cannot be paired (and is no longer 1302 // considered). 1303 1304 DEBUG(if (BestTree.size() > 0) 1305 dbgs() << "BBV: selected pairs in the best tree for: " 1306 << *cast<Instruction>(*I) << "\n"); 1307 1308 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1309 SE2 = BestTree.end(); S != SE2; ++S) { 1310 // Insert the members of this tree into the list of chosen pairs. 1311 ChosenPairs.insert(ValuePair(S->first, S->second)); 1312 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1313 *S->second << "\n"); 1314 1315 // Remove all candidate pairs that have values in the chosen tree. 1316 for (std::multimap<Value *, Value *>::iterator K = 1317 CandidatePairs.begin(); K != CandidatePairs.end();) { 1318 if (K->first == S->first || K->second == S->first || 1319 K->second == S->second || K->first == S->second) { 1320 // Don't remove the actual pair chosen so that it can be used 1321 // in subsequent tree selections. 1322 if (!(K->first == S->first && K->second == S->second)) 1323 CandidatePairs.erase(K++); 1324 else 1325 ++K; 1326 } else { 1327 ++K; 1328 } 1329 } 1330 } 1331 } 1332 1333 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1334 } 1335 1336 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1337 unsigned n = 0) { 1338 if (!I->hasName()) 1339 return ""; 1340 1341 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1342 (n > 0 ? "." + utostr(n) : "")).str(); 1343 } 1344 1345 // Returns the value that is to be used as the pointer input to the vector 1346 // instruction that fuses I with J. 1347 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1348 Instruction *I, Instruction *J, unsigned o, 1349 bool &FlipMemInputs) { 1350 Value *IPtr, *JPtr; 1351 unsigned IAlignment, JAlignment; 1352 int64_t OffsetInElmts; 1353 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1354 OffsetInElmts); 1355 1356 // The pointer value is taken to be the one with the lowest offset. 1357 Value *VPtr; 1358 if (OffsetInElmts > 0) { 1359 VPtr = IPtr; 1360 } else { 1361 FlipMemInputs = true; 1362 VPtr = JPtr; 1363 } 1364 1365 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType(); 1366 Type *VArgType = getVecTypeForPair(ArgType); 1367 Type *VArgPtrType = PointerType::get(VArgType, 1368 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1369 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1370 /* insert before */ FlipMemInputs ? J : I); 1371 } 1372 1373 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1374 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 1375 unsigned IdxOffset, std::vector<Constant*> &Mask) { 1376 for (unsigned v = 0; v < NumElem/2; ++v) { 1377 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1378 if (m < 0) { 1379 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1380 } else { 1381 unsigned mm = m + (int) IdxOffset; 1382 if (m >= (int) NumInElem) 1383 mm += (int) NumInElem; 1384 1385 Mask[v+MaskOffset] = 1386 ConstantInt::get(Type::getInt32Ty(Context), mm); 1387 } 1388 } 1389 } 1390 1391 // Returns the value that is to be used as the vector-shuffle mask to the 1392 // vector instruction that fuses I with J. 1393 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1394 Instruction *I, Instruction *J) { 1395 // This is the shuffle mask. We need to append the second 1396 // mask to the first, and the numbers need to be adjusted. 1397 1398 Type *ArgType = I->getType(); 1399 Type *VArgType = getVecTypeForPair(ArgType); 1400 1401 // Get the total number of elements in the fused vector type. 1402 // By definition, this must equal the number of elements in 1403 // the final mask. 1404 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1405 std::vector<Constant*> Mask(NumElem); 1406 1407 Type *OpType = I->getOperand(0)->getType(); 1408 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements(); 1409 1410 // For the mask from the first pair... 1411 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask); 1412 1413 // For the mask from the second pair... 1414 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem, 1415 Mask); 1416 1417 return ConstantVector::get(Mask); 1418 } 1419 1420 // Returns the value to be used as the specified operand of the vector 1421 // instruction that fuses I with J. 1422 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1423 Instruction *J, unsigned o, bool FlipMemInputs) { 1424 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1425 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1426 1427 // Compute the fused vector type for this operand 1428 Type *ArgType = I->getOperand(o)->getType(); 1429 VectorType *VArgType = getVecTypeForPair(ArgType); 1430 1431 Instruction *L = I, *H = J; 1432 if (FlipMemInputs) { 1433 L = J; 1434 H = I; 1435 } 1436 1437 if (ArgType->isVectorTy()) { 1438 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1439 std::vector<Constant*> Mask(numElem); 1440 for (unsigned v = 0; v < numElem; ++v) 1441 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1442 1443 Instruction *BV = new ShuffleVectorInst(L->getOperand(o), 1444 H->getOperand(o), 1445 ConstantVector::get(Mask), 1446 getReplacementName(I, true, o)); 1447 BV->insertBefore(J); 1448 return BV; 1449 } 1450 1451 // If these two inputs are the output of another vector instruction, 1452 // then we should use that output directly. It might be necessary to 1453 // permute it first. [When pairings are fused recursively, you can 1454 // end up with cases where a large vector is decomposed into scalars 1455 // using extractelement instructions, then built into size-2 1456 // vectors using insertelement and the into larger vectors using 1457 // shuffles. InstCombine does not simplify all of these cases well, 1458 // and so we make sure that shuffles are generated here when possible. 1459 ExtractElementInst *LEE 1460 = dyn_cast<ExtractElementInst>(L->getOperand(o)); 1461 ExtractElementInst *HEE 1462 = dyn_cast<ExtractElementInst>(H->getOperand(o)); 1463 1464 if (LEE && HEE && 1465 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) { 1466 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType()); 1467 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue(); 1468 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue(); 1469 if (LEE->getOperand(0) == HEE->getOperand(0)) { 1470 if (LowIndx == 0 && HighIndx == 1) 1471 return LEE->getOperand(0); 1472 1473 std::vector<Constant*> Mask(2); 1474 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1475 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1476 1477 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1478 UndefValue::get(EEType), 1479 ConstantVector::get(Mask), 1480 getReplacementName(I, true, o)); 1481 BV->insertBefore(J); 1482 return BV; 1483 } 1484 1485 std::vector<Constant*> Mask(2); 1486 HighIndx += EEType->getNumElements(); 1487 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1488 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1489 1490 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1491 HEE->getOperand(0), 1492 ConstantVector::get(Mask), 1493 getReplacementName(I, true, o)); 1494 BV->insertBefore(J); 1495 return BV; 1496 } 1497 1498 Instruction *BV1 = InsertElementInst::Create( 1499 UndefValue::get(VArgType), 1500 L->getOperand(o), CV0, 1501 getReplacementName(I, true, o, 1)); 1502 BV1->insertBefore(I); 1503 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o), 1504 CV1, 1505 getReplacementName(I, true, o, 2)); 1506 BV2->insertBefore(J); 1507 return BV2; 1508 } 1509 1510 // This function creates an array of values that will be used as the inputs 1511 // to the vector instruction that fuses I with J. 1512 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1513 Instruction *I, Instruction *J, 1514 SmallVector<Value *, 3> &ReplacedOperands, 1515 bool &FlipMemInputs) { 1516 FlipMemInputs = false; 1517 unsigned NumOperands = I->getNumOperands(); 1518 1519 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1520 // Iterate backward so that we look at the store pointer 1521 // first and know whether or not we need to flip the inputs. 1522 1523 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1524 // This is the pointer for a load/store instruction. 1525 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 1526 FlipMemInputs); 1527 continue; 1528 } else if (isa<CallInst>(I) && o == NumOperands-1) { 1529 Function *F = cast<CallInst>(I)->getCalledFunction(); 1530 unsigned IID = F->getIntrinsicID(); 1531 BasicBlock &BB = *I->getParent(); 1532 1533 Module *M = BB.getParent()->getParent(); 1534 Type *ArgType = I->getType(); 1535 Type *VArgType = getVecTypeForPair(ArgType); 1536 1537 // FIXME: is it safe to do this here? 1538 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 1539 (Intrinsic::ID) IID, VArgType); 1540 continue; 1541 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 1542 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 1543 continue; 1544 } 1545 1546 ReplacedOperands[o] = 1547 getReplacementInput(Context, I, J, o, FlipMemInputs); 1548 } 1549 } 1550 1551 // This function creates two values that represent the outputs of the 1552 // original I and J instructions. These are generally vector shuffles 1553 // or extracts. In many cases, these will end up being unused and, thus, 1554 // eliminated by later passes. 1555 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 1556 Instruction *J, Instruction *K, 1557 Instruction *&InsertionPt, 1558 Instruction *&K1, Instruction *&K2, 1559 bool &FlipMemInputs) { 1560 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1561 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1562 1563 if (isa<StoreInst>(I)) { 1564 AA->replaceWithNewValue(I, K); 1565 AA->replaceWithNewValue(J, K); 1566 } else { 1567 Type *IType = I->getType(); 1568 Type *VType = getVecTypeForPair(IType); 1569 1570 if (IType->isVectorTy()) { 1571 unsigned numElem = cast<VectorType>(IType)->getNumElements(); 1572 std::vector<Constant*> Mask1(numElem), Mask2(numElem); 1573 for (unsigned v = 0; v < numElem; ++v) { 1574 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1575 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v); 1576 } 1577 1578 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 1579 ConstantVector::get( 1580 FlipMemInputs ? Mask2 : Mask1), 1581 getReplacementName(K, false, 1)); 1582 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 1583 ConstantVector::get( 1584 FlipMemInputs ? Mask1 : Mask2), 1585 getReplacementName(K, false, 2)); 1586 } else { 1587 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 1588 getReplacementName(K, false, 1)); 1589 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 1590 getReplacementName(K, false, 2)); 1591 } 1592 1593 K1->insertAfter(K); 1594 K2->insertAfter(K1); 1595 InsertionPt = K2; 1596 } 1597 } 1598 1599 // Move all uses of the function I (including pairing-induced uses) after J. 1600 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 1601 std::multimap<Value *, Value *> &LoadMoveSet, 1602 Instruction *I, Instruction *J) { 1603 // Skip to the first instruction past I. 1604 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1605 1606 DenseSet<Value *> Users; 1607 AliasSetTracker WriteSet(*AA); 1608 for (; cast<Instruction>(L) != J; ++L) 1609 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 1610 1611 assert(cast<Instruction>(L) == J && 1612 "Tracking has not proceeded far enough to check for dependencies"); 1613 // If J is now in the use set of I, then trackUsesOfI will return true 1614 // and we have a dependency cycle (and the fusing operation must abort). 1615 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 1616 } 1617 1618 // Move all uses of the function I (including pairing-induced uses) after J. 1619 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 1620 std::multimap<Value *, Value *> &LoadMoveSet, 1621 Instruction *&InsertionPt, 1622 Instruction *I, Instruction *J) { 1623 // Skip to the first instruction past I. 1624 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1625 1626 DenseSet<Value *> Users; 1627 AliasSetTracker WriteSet(*AA); 1628 for (; cast<Instruction>(L) != J;) { 1629 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 1630 // Move this instruction 1631 Instruction *InstToMove = L; ++L; 1632 1633 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 1634 " to after " << *InsertionPt << "\n"); 1635 InstToMove->removeFromParent(); 1636 InstToMove->insertAfter(InsertionPt); 1637 InsertionPt = InstToMove; 1638 } else { 1639 ++L; 1640 } 1641 } 1642 } 1643 1644 // Collect all load instruction that are in the move set of a given first 1645 // pair member. These loads depend on the first instruction, I, and so need 1646 // to be moved after J (the second instruction) when the pair is fused. 1647 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 1648 DenseMap<Value *, Value *> &ChosenPairs, 1649 std::multimap<Value *, Value *> &LoadMoveSet, 1650 Instruction *I) { 1651 // Skip to the first instruction past I. 1652 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1653 1654 DenseSet<Value *> Users; 1655 AliasSetTracker WriteSet(*AA); 1656 1657 // Note: We cannot end the loop when we reach J because J could be moved 1658 // farther down the use chain by another instruction pairing. Also, J 1659 // could be before I if this is an inverted input. 1660 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 1661 if (trackUsesOfI(Users, WriteSet, I, L)) { 1662 if (L->mayReadFromMemory()) 1663 LoadMoveSet.insert(ValuePair(L, I)); 1664 } 1665 } 1666 } 1667 1668 // In cases where both load/stores and the computation of their pointers 1669 // are chosen for vectorization, we can end up in a situation where the 1670 // aliasing analysis starts returning different query results as the 1671 // process of fusing instruction pairs continues. Because the algorithm 1672 // relies on finding the same use trees here as were found earlier, we'll 1673 // need to precompute the necessary aliasing information here and then 1674 // manually update it during the fusion process. 1675 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 1676 std::vector<Value *> &PairableInsts, 1677 DenseMap<Value *, Value *> &ChosenPairs, 1678 std::multimap<Value *, Value *> &LoadMoveSet) { 1679 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1680 PIE = PairableInsts.end(); PI != PIE; ++PI) { 1681 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 1682 if (P == ChosenPairs.end()) continue; 1683 1684 Instruction *I = cast<Instruction>(P->first); 1685 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 1686 } 1687 } 1688 1689 // This function fuses the chosen instruction pairs into vector instructions, 1690 // taking care preserve any needed scalar outputs and, then, it reorders the 1691 // remaining instructions as needed (users of the first member of the pair 1692 // need to be moved to after the location of the second member of the pair 1693 // because the vector instruction is inserted in the location of the pair's 1694 // second member). 1695 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 1696 std::vector<Value *> &PairableInsts, 1697 DenseMap<Value *, Value *> &ChosenPairs) { 1698 LLVMContext& Context = BB.getContext(); 1699 1700 // During the vectorization process, the order of the pairs to be fused 1701 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 1702 // list. After a pair is fused, the flipped pair is removed from the list. 1703 std::vector<ValuePair> FlippedPairs; 1704 FlippedPairs.reserve(ChosenPairs.size()); 1705 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 1706 E = ChosenPairs.end(); P != E; ++P) 1707 FlippedPairs.push_back(ValuePair(P->second, P->first)); 1708 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 1709 E = FlippedPairs.end(); P != E; ++P) 1710 ChosenPairs.insert(*P); 1711 1712 std::multimap<Value *, Value *> LoadMoveSet; 1713 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 1714 1715 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 1716 1717 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 1718 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 1719 if (P == ChosenPairs.end()) { 1720 ++PI; 1721 continue; 1722 } 1723 1724 if (getDepthFactor(P->first) == 0) { 1725 // These instructions are not really fused, but are tracked as though 1726 // they are. Any case in which it would be interesting to fuse them 1727 // will be taken care of by InstCombine. 1728 --NumFusedOps; 1729 ++PI; 1730 continue; 1731 } 1732 1733 Instruction *I = cast<Instruction>(P->first), 1734 *J = cast<Instruction>(P->second); 1735 1736 DEBUG(dbgs() << "BBV: fusing: " << *I << 1737 " <-> " << *J << "\n"); 1738 1739 // Remove the pair and flipped pair from the list. 1740 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 1741 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 1742 ChosenPairs.erase(FP); 1743 ChosenPairs.erase(P); 1744 1745 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 1746 DEBUG(dbgs() << "BBV: fusion of: " << *I << 1747 " <-> " << *J << 1748 " aborted because of non-trivial dependency cycle\n"); 1749 --NumFusedOps; 1750 ++PI; 1751 continue; 1752 } 1753 1754 bool FlipMemInputs; 1755 unsigned NumOperands = I->getNumOperands(); 1756 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 1757 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 1758 FlipMemInputs); 1759 1760 // Make a copy of the original operation, change its type to the vector 1761 // type and replace its operands with the vector operands. 1762 Instruction *K = I->clone(); 1763 if (I->hasName()) K->takeName(I); 1764 1765 if (!isa<StoreInst>(K)) 1766 K->mutateType(getVecTypeForPair(I->getType())); 1767 1768 for (unsigned o = 0; o < NumOperands; ++o) 1769 K->setOperand(o, ReplacedOperands[o]); 1770 1771 // If we've flipped the memory inputs, make sure that we take the correct 1772 // alignment. 1773 if (FlipMemInputs) { 1774 if (isa<StoreInst>(K)) 1775 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 1776 else 1777 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 1778 } 1779 1780 K->insertAfter(J); 1781 1782 // Instruction insertion point: 1783 Instruction *InsertionPt = K; 1784 Instruction *K1 = 0, *K2 = 0; 1785 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 1786 FlipMemInputs); 1787 1788 // The use tree of the first original instruction must be moved to after 1789 // the location of the second instruction. The entire use tree of the 1790 // first instruction is disjoint from the input tree of the second 1791 // (by definition), and so commutes with it. 1792 1793 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 1794 1795 if (!isa<StoreInst>(I)) { 1796 I->replaceAllUsesWith(K1); 1797 J->replaceAllUsesWith(K2); 1798 AA->replaceWithNewValue(I, K1); 1799 AA->replaceWithNewValue(J, K2); 1800 } 1801 1802 // Instructions that may read from memory may be in the load move set. 1803 // Once an instruction is fused, we no longer need its move set, and so 1804 // the values of the map never need to be updated. However, when a load 1805 // is fused, we need to merge the entries from both instructions in the 1806 // pair in case those instructions were in the move set of some other 1807 // yet-to-be-fused pair. The loads in question are the keys of the map. 1808 if (I->mayReadFromMemory()) { 1809 std::vector<ValuePair> NewSetMembers; 1810 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 1811 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 1812 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 1813 N != IPairRange.second; ++N) 1814 NewSetMembers.push_back(ValuePair(K, N->second)); 1815 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 1816 N != JPairRange.second; ++N) 1817 NewSetMembers.push_back(ValuePair(K, N->second)); 1818 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 1819 AE = NewSetMembers.end(); A != AE; ++A) 1820 LoadMoveSet.insert(*A); 1821 } 1822 1823 // Before removing I, set the iterator to the next instruction. 1824 PI = llvm::next(BasicBlock::iterator(I)); 1825 if (cast<Instruction>(PI) == J) 1826 ++PI; 1827 1828 SE->forgetValue(I); 1829 SE->forgetValue(J); 1830 I->eraseFromParent(); 1831 J->eraseFromParent(); 1832 } 1833 1834 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 1835 } 1836} 1837 1838char BBVectorize::ID = 0; 1839static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 1840INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1841INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 1842INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 1843INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1844 1845BasicBlockPass *llvm::createBBVectorizePass() { 1846 return new BBVectorize(); 1847} 1848 1849