BBVectorize.cpp revision 5d4e18bc39fea892f523d960213906d296d3cb38
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> 107FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 108 cl::desc("Use a fast instruction dependency analysis")); 109 110#ifndef NDEBUG 111static cl::opt<bool> 112DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 113 cl::init(false), cl::Hidden, 114 cl::desc("When debugging is enabled, output information on the" 115 " instruction-examination process")); 116static cl::opt<bool> 117DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 118 cl::init(false), cl::Hidden, 119 cl::desc("When debugging is enabled, output information on the" 120 " candidate-selection process")); 121static cl::opt<bool> 122DebugPairSelection("bb-vectorize-debug-pair-selection", 123 cl::init(false), cl::Hidden, 124 cl::desc("When debugging is enabled, output information on the" 125 " pair-selection process")); 126static cl::opt<bool> 127DebugCycleCheck("bb-vectorize-debug-cycle-check", 128 cl::init(false), cl::Hidden, 129 cl::desc("When debugging is enabled, output information on the" 130 " cycle-checking process")); 131#endif 132 133STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 134 135namespace { 136 struct BBVectorize : public BasicBlockPass { 137 static char ID; // Pass identification, replacement for typeid 138 BBVectorize() : BasicBlockPass(ID) { 139 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 140 } 141 142 typedef std::pair<Value *, Value *> ValuePair; 143 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 144 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 145 typedef std::pair<std::multimap<Value *, Value *>::iterator, 146 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 147 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 148 std::multimap<ValuePair, ValuePair>::iterator> 149 VPPIteratorPair; 150 151 AliasAnalysis *AA; 152 ScalarEvolution *SE; 153 TargetData *TD; 154 155 // FIXME: const correct? 156 157 bool vectorizePairs(BasicBlock &BB); 158 159 bool getCandidatePairs(BasicBlock &BB, 160 BasicBlock::iterator &Start, 161 std::multimap<Value *, Value *> &CandidatePairs, 162 std::vector<Value *> &PairableInsts); 163 164 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 165 std::vector<Value *> &PairableInsts, 166 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 167 168 void buildDepMap(BasicBlock &BB, 169 std::multimap<Value *, Value *> &CandidatePairs, 170 std::vector<Value *> &PairableInsts, 171 DenseSet<ValuePair> &PairableInstUsers); 172 173 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 174 std::vector<Value *> &PairableInsts, 175 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 176 DenseSet<ValuePair> &PairableInstUsers, 177 DenseMap<Value *, Value *>& ChosenPairs); 178 179 void fuseChosenPairs(BasicBlock &BB, 180 std::vector<Value *> &PairableInsts, 181 DenseMap<Value *, Value *>& ChosenPairs); 182 183 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 184 185 bool areInstsCompatible(Instruction *I, Instruction *J, 186 bool IsSimpleLoadStore); 187 188 bool trackUsesOfI(DenseSet<Value *> &Users, 189 AliasSetTracker &WriteSet, Instruction *I, 190 Instruction *J, bool UpdateUsers = true, 191 std::multimap<Value *, Value *> *LoadMoveSet = 0); 192 193 void computePairsConnectedTo( 194 std::multimap<Value *, Value *> &CandidatePairs, 195 std::vector<Value *> &PairableInsts, 196 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 197 ValuePair P); 198 199 bool pairsConflict(ValuePair P, ValuePair Q, 200 DenseSet<ValuePair> &PairableInstUsers, 201 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 202 203 bool pairWillFormCycle(ValuePair P, 204 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 205 DenseSet<ValuePair> &CurrentPairs); 206 207 void pruneTreeFor( 208 std::multimap<Value *, Value *> &CandidatePairs, 209 std::vector<Value *> &PairableInsts, 210 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 211 DenseSet<ValuePair> &PairableInstUsers, 212 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 213 DenseMap<Value *, Value *> &ChosenPairs, 214 DenseMap<ValuePair, size_t> &Tree, 215 DenseSet<ValuePair> &PrunedTree, ValuePair J, 216 bool UseCycleCheck); 217 218 void buildInitialTreeFor( 219 std::multimap<Value *, Value *> &CandidatePairs, 220 std::vector<Value *> &PairableInsts, 221 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 222 DenseSet<ValuePair> &PairableInstUsers, 223 DenseMap<Value *, Value *> &ChosenPairs, 224 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 225 226 void findBestTreeFor( 227 std::multimap<Value *, Value *> &CandidatePairs, 228 std::vector<Value *> &PairableInsts, 229 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 230 DenseSet<ValuePair> &PairableInstUsers, 231 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 232 DenseMap<Value *, Value *> &ChosenPairs, 233 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 234 size_t &BestEffSize, VPIteratorPair ChoiceRange, 235 bool UseCycleCheck); 236 237 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 238 Instruction *J, unsigned o, bool &FlipMemInputs); 239 240 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 241 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 242 unsigned IdxOffset, std::vector<Constant*> &Mask); 243 244 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 245 Instruction *J); 246 247 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 248 Instruction *J, unsigned o, bool FlipMemInputs); 249 250 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 251 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 252 bool &FlipMemInputs); 253 254 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 255 Instruction *J, Instruction *K, 256 Instruction *&InsertionPt, Instruction *&K1, 257 Instruction *&K2, bool &FlipMemInputs); 258 259 void collectPairLoadMoveSet(BasicBlock &BB, 260 DenseMap<Value *, Value *> &ChosenPairs, 261 std::multimap<Value *, Value *> &LoadMoveSet, 262 Instruction *I); 263 264 void collectLoadMoveSet(BasicBlock &BB, 265 std::vector<Value *> &PairableInsts, 266 DenseMap<Value *, Value *> &ChosenPairs, 267 std::multimap<Value *, Value *> &LoadMoveSet); 268 269 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 270 std::multimap<Value *, Value *> &LoadMoveSet, 271 Instruction *I, Instruction *J); 272 273 void moveUsesOfIAfterJ(BasicBlock &BB, 274 std::multimap<Value *, Value *> &LoadMoveSet, 275 Instruction *&InsertionPt, 276 Instruction *I, Instruction *J); 277 278 virtual bool runOnBasicBlock(BasicBlock &BB) { 279 AA = &getAnalysis<AliasAnalysis>(); 280 SE = &getAnalysis<ScalarEvolution>(); 281 TD = getAnalysisIfAvailable<TargetData>(); 282 283 bool changed = false; 284 // Iterate a sufficient number of times to merge types of size 1 bit, 285 // then 2 bits, then 4, etc. up to half of the target vector width of the 286 // target vector register. 287 for (unsigned v = 2, n = 1; v <= VectorBits && (!MaxIter || n <= MaxIter); 288 v *= 2, ++n) { 289 DEBUG(dbgs() << "BBV: fusing loop #" << n << 290 " for " << BB.getName() << " in " << 291 BB.getParent()->getName() << "...\n"); 292 if (vectorizePairs(BB)) 293 changed = true; 294 else 295 break; 296 } 297 298 DEBUG(dbgs() << "BBV: done!\n"); 299 return changed; 300 } 301 302 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 303 BasicBlockPass::getAnalysisUsage(AU); 304 AU.addRequired<AliasAnalysis>(); 305 AU.addRequired<ScalarEvolution>(); 306 AU.addPreserved<AliasAnalysis>(); 307 AU.addPreserved<ScalarEvolution>(); 308 AU.setPreservesCFG(); 309 } 310 311 // This returns the vector type that holds a pair of the provided type. 312 // If the provided type is already a vector, then its length is doubled. 313 static inline VectorType *getVecTypeForPair(Type *ElemTy) { 314 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 315 unsigned numElem = VTy->getNumElements(); 316 return VectorType::get(ElemTy->getScalarType(), numElem*2); 317 } 318 319 return VectorType::get(ElemTy, 2); 320 } 321 322 // Returns the weight associated with the provided value. A chain of 323 // candidate pairs has a length given by the sum of the weights of its 324 // members (one weight per pair; the weight of each member of the pair 325 // is assumed to be the same). This length is then compared to the 326 // chain-length threshold to determine if a given chain is significant 327 // enough to be vectorized. The length is also used in comparing 328 // candidate chains where longer chains are considered to be better. 329 // Note: when this function returns 0, the resulting instructions are 330 // not actually fused. 331 static inline size_t getDepthFactor(Value *V) { 332 // InsertElement and ExtractElement have a depth factor of zero. This is 333 // for two reasons: First, they cannot be usefully fused. Second, because 334 // the pass generates a lot of these, they can confuse the simple metric 335 // used to compare the trees in the next iteration. Thus, giving them a 336 // weight of zero allows the pass to essentially ignore them in 337 // subsequent iterations when looking for vectorization opportunities 338 // while still tracking dependency chains that flow through those 339 // instructions. 340 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 341 return 0; 342 343 return 1; 344 } 345 346 // This determines the relative offset of two loads or stores, returning 347 // true if the offset could be determined to be some constant value. 348 // For example, if OffsetInElmts == 1, then J accesses the memory directly 349 // after I; if OffsetInElmts == -1 then I accesses the memory 350 // directly after J. This function assumes that both instructions 351 // have the same type. 352 bool getPairPtrInfo(Instruction *I, Instruction *J, 353 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 354 int64_t &OffsetInElmts) { 355 OffsetInElmts = 0; 356 if (isa<LoadInst>(I)) { 357 IPtr = cast<LoadInst>(I)->getPointerOperand(); 358 JPtr = cast<LoadInst>(J)->getPointerOperand(); 359 IAlignment = cast<LoadInst>(I)->getAlignment(); 360 JAlignment = cast<LoadInst>(J)->getAlignment(); 361 } else { 362 IPtr = cast<StoreInst>(I)->getPointerOperand(); 363 JPtr = cast<StoreInst>(J)->getPointerOperand(); 364 IAlignment = cast<StoreInst>(I)->getAlignment(); 365 JAlignment = cast<StoreInst>(J)->getAlignment(); 366 } 367 368 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 369 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 370 371 // If this is a trivial offset, then we'll get something like 372 // 1*sizeof(type). With target data, which we need anyway, this will get 373 // constant folded into a number. 374 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 375 if (const SCEVConstant *ConstOffSCEV = 376 dyn_cast<SCEVConstant>(OffsetSCEV)) { 377 ConstantInt *IntOff = ConstOffSCEV->getValue(); 378 int64_t Offset = IntOff->getSExtValue(); 379 380 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 381 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 382 383 assert(VTy == cast<PointerType>(JPtr->getType())->getElementType()); 384 385 OffsetInElmts = Offset/VTyTSS; 386 return (abs64(Offset) % VTyTSS) == 0; 387 } 388 389 return false; 390 } 391 392 // Returns true if the provided CallInst represents an intrinsic that can 393 // be vectorized. 394 bool isVectorizableIntrinsic(CallInst* I) { 395 Function *F = I->getCalledFunction(); 396 if (!F) return false; 397 398 unsigned IID = F->getIntrinsicID(); 399 if (!IID) return false; 400 401 switch(IID) { 402 default: 403 return false; 404 case Intrinsic::sqrt: 405 case Intrinsic::powi: 406 case Intrinsic::sin: 407 case Intrinsic::cos: 408 case Intrinsic::log: 409 case Intrinsic::log2: 410 case Intrinsic::log10: 411 case Intrinsic::exp: 412 case Intrinsic::exp2: 413 case Intrinsic::pow: 414 return !NoMath; 415 case Intrinsic::fma: 416 return !NoFMA; 417 } 418 } 419 420 // Returns true if J is the second element in some pair referenced by 421 // some multimap pair iterator pair. 422 template <typename V> 423 bool isSecondInIteratorPair(V J, std::pair< 424 typename std::multimap<V, V>::iterator, 425 typename std::multimap<V, V>::iterator> PairRange) { 426 for (typename std::multimap<V, V>::iterator K = PairRange.first; 427 K != PairRange.second; ++K) 428 if (K->second == J) return true; 429 430 return false; 431 } 432 }; 433 434 // This function implements one vectorization iteration on the provided 435 // basic block. It returns true if the block is changed. 436 bool BBVectorize::vectorizePairs(BasicBlock &BB) { 437 bool ShouldContinue; 438 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 439 440 std::vector<Value *> AllPairableInsts; 441 DenseMap<Value *, Value *> AllChosenPairs; 442 443 do { 444 std::vector<Value *> PairableInsts; 445 std::multimap<Value *, Value *> CandidatePairs; 446 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 447 PairableInsts); 448 if (PairableInsts.empty()) continue; 449 450 // Now we have a map of all of the pairable instructions and we need to 451 // select the best possible pairing. A good pairing is one such that the 452 // users of the pair are also paired. This defines a (directed) forest 453 // over the pairs such that two pairs are connected iff the second pair 454 // uses the first. 455 456 // Note that it only matters that both members of the second pair use some 457 // element of the first pair (to allow for splatting). 458 459 std::multimap<ValuePair, ValuePair> ConnectedPairs; 460 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 461 if (ConnectedPairs.empty()) continue; 462 463 // Build the pairable-instruction dependency map 464 DenseSet<ValuePair> PairableInstUsers; 465 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 466 467 // There is now a graph of the connected pairs. For each variable, pick the 468 // pairing with the largest tree meeting the depth requirement on at least 469 // one branch. Then select all pairings that are part of that tree and 470 // remove them from the list of available pairings and pairable variables. 471 472 DenseMap<Value *, Value *> ChosenPairs; 473 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, 474 PairableInstUsers, ChosenPairs); 475 476 if (ChosenPairs.empty()) continue; 477 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 478 PairableInsts.end()); 479 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 480 } while (ShouldContinue); 481 482 if (AllChosenPairs.empty()) return false; 483 NumFusedOps += AllChosenPairs.size(); 484 485 // A set of pairs has now been selected. It is now necessary to replace the 486 // paired instructions with vector instructions. For this procedure each 487 // operand much be replaced with a vector operand. This vector is formed 488 // by using build_vector on the old operands. The replaced values are then 489 // replaced with a vector_extract on the result. Subsequent optimization 490 // passes should coalesce the build/extract combinations. 491 492 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); 493 return true; 494 } 495 496 // This function returns true if the provided instruction is capable of being 497 // fused into a vector instruction. This determination is based only on the 498 // type and other attributes of the instruction. 499 bool BBVectorize::isInstVectorizable(Instruction *I, 500 bool &IsSimpleLoadStore) { 501 IsSimpleLoadStore = false; 502 503 if (CallInst *C = dyn_cast<CallInst>(I)) { 504 if (!isVectorizableIntrinsic(C)) 505 return false; 506 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 507 // Vectorize simple loads if possbile: 508 IsSimpleLoadStore = L->isSimple(); 509 if (!IsSimpleLoadStore || NoMemOps) 510 return false; 511 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 512 // Vectorize simple stores if possbile: 513 IsSimpleLoadStore = S->isSimple(); 514 if (!IsSimpleLoadStore || NoMemOps) 515 return false; 516 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 517 // We can vectorize casts, but not casts of pointer types, etc. 518 if (NoCasts) 519 return false; 520 521 Type *SrcTy = C->getSrcTy(); 522 if (!SrcTy->isSingleValueType() || SrcTy->isPointerTy()) 523 return false; 524 525 Type *DestTy = C->getDestTy(); 526 if (!DestTy->isSingleValueType() || DestTy->isPointerTy()) 527 return false; 528 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 529 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 530 return false; 531 } 532 533 // We can't vectorize memory operations without target data 534 if (TD == 0 && IsSimpleLoadStore) 535 return false; 536 537 Type *T1, *T2; 538 if (isa<StoreInst>(I)) { 539 // For stores, it is the value type, not the pointer type that matters 540 // because the value is what will come from a vector register. 541 542 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 543 T1 = IVal->getType(); 544 } else { 545 T1 = I->getType(); 546 } 547 548 if (I->isCast()) 549 T2 = cast<CastInst>(I)->getSrcTy(); 550 else 551 T2 = T1; 552 553 // Not every type can be vectorized... 554 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 555 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 556 return false; 557 558 if (NoInts && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) 559 return false; 560 561 if (NoFloats && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 562 return false; 563 564 if (T1->getPrimitiveSizeInBits() > VectorBits/2 || 565 T2->getPrimitiveSizeInBits() > VectorBits/2) 566 return false; 567 568 return true; 569 } 570 571 // This function returns true if the two provided instructions are compatible 572 // (meaning that they can be fused into a vector instruction). This assumes 573 // that I has already been determined to be vectorizable and that J is not 574 // in the use tree of I. 575 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 576 bool IsSimpleLoadStore) { 577 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 578 " <-> " << *J << "\n"); 579 580 // Loads and stores can be merged if they have different alignments, 581 // but are otherwise the same. 582 LoadInst *LI, *LJ; 583 StoreInst *SI, *SJ; 584 if ((LI = dyn_cast<LoadInst>(I)) && (LJ = dyn_cast<LoadInst>(J))) { 585 if (I->getType() != J->getType()) 586 return false; 587 588 if (LI->getPointerOperand()->getType() != 589 LJ->getPointerOperand()->getType() || 590 LI->isVolatile() != LJ->isVolatile() || 591 LI->getOrdering() != LJ->getOrdering() || 592 LI->getSynchScope() != LJ->getSynchScope()) 593 return false; 594 } else if ((SI = dyn_cast<StoreInst>(I)) && (SJ = dyn_cast<StoreInst>(J))) { 595 if (SI->getValueOperand()->getType() != 596 SJ->getValueOperand()->getType() || 597 SI->getPointerOperand()->getType() != 598 SJ->getPointerOperand()->getType() || 599 SI->isVolatile() != SJ->isVolatile() || 600 SI->getOrdering() != SJ->getOrdering() || 601 SI->getSynchScope() != SJ->getSynchScope()) 602 return false; 603 } else if (!J->isSameOperationAs(I)) { 604 return false; 605 } 606 // FIXME: handle addsub-type operations! 607 608 if (IsSimpleLoadStore) { 609 Value *IPtr, *JPtr; 610 unsigned IAlignment, JAlignment; 611 int64_t OffsetInElmts = 0; 612 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 613 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 614 if (AlignedOnly) { 615 Type *aType = isa<StoreInst>(I) ? 616 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 617 // An aligned load or store is possible only if the instruction 618 // with the lower offset has an alignment suitable for the 619 // vector type. 620 621 unsigned BottomAlignment = IAlignment; 622 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 623 624 Type *VType = getVecTypeForPair(aType); 625 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 626 if (BottomAlignment < VecAlignment) 627 return false; 628 } 629 } else { 630 return false; 631 } 632 } else if (isa<ShuffleVectorInst>(I)) { 633 // Only merge two shuffles if they're both constant 634 return isa<Constant>(I->getOperand(2)) && 635 isa<Constant>(J->getOperand(2)); 636 // FIXME: We may want to vectorize non-constant shuffles also. 637 } 638 639 return true; 640 } 641 642 // Figure out whether or not J uses I and update the users and write-set 643 // structures associated with I. Specifically, Users represents the set of 644 // instructions that depend on I. WriteSet represents the set 645 // of memory locations that are dependent on I. If UpdateUsers is true, 646 // and J uses I, then Users is updated to contain J and WriteSet is updated 647 // to contain any memory locations to which J writes. The function returns 648 // true if J uses I. By default, alias analysis is used to determine 649 // whether J reads from memory that overlaps with a location in WriteSet. 650 // If LoadMoveSet is not null, then it is a previously-computed multimap 651 // where the key is the memory-based user instruction and the value is 652 // the instruction to be compared with I. So, if LoadMoveSet is provided, 653 // then the alias analysis is not used. This is necessary because this 654 // function is called during the process of moving instructions during 655 // vectorization and the results of the alias analysis are not stable during 656 // that process. 657 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 658 AliasSetTracker &WriteSet, Instruction *I, 659 Instruction *J, bool UpdateUsers, 660 std::multimap<Value *, Value *> *LoadMoveSet) { 661 bool UsesI = false; 662 663 // This instruction may already be marked as a user due, for example, to 664 // being a member of a selected pair. 665 if (Users.count(J)) 666 UsesI = true; 667 668 if (!UsesI) 669 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 670 JU != JE; ++JU) { 671 Value *V = *JU; 672 if (I == V || Users.count(V)) { 673 UsesI = true; 674 break; 675 } 676 } 677 if (!UsesI && J->mayReadFromMemory()) { 678 if (LoadMoveSet) { 679 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 680 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 681 } else { 682 for (AliasSetTracker::iterator W = WriteSet.begin(), 683 WE = WriteSet.end(); W != WE; ++W) { 684 for (AliasSet::iterator A = W->begin(), AE = W->end(); 685 A != AE; ++A) { 686 AliasAnalysis::Location ptrLoc(A->getValue(), A->getSize(), 687 A->getTBAAInfo()); 688 if (AA->getModRefInfo(J, ptrLoc) != AliasAnalysis::NoModRef) { 689 UsesI = true; 690 break; 691 } 692 } 693 if (UsesI) break; 694 } 695 } 696 } 697 698 if (UsesI && UpdateUsers) { 699 if (J->mayWriteToMemory()) WriteSet.add(J); 700 Users.insert(J); 701 } 702 703 return UsesI; 704 } 705 706 // This function iterates over all instruction pairs in the provided 707 // basic block and collects all candidate pairs for vectorization. 708 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 709 BasicBlock::iterator &Start, 710 std::multimap<Value *, Value *> &CandidatePairs, 711 std::vector<Value *> &PairableInsts) { 712 BasicBlock::iterator E = BB.end(); 713 if (Start == E) return false; 714 715 bool ShouldContinue = false, IAfterStart = false; 716 for (BasicBlock::iterator I = Start++; I != E; ++I) { 717 if (I == Start) IAfterStart = true; 718 719 bool IsSimpleLoadStore; 720 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 721 722 // Look for an instruction with which to pair instruction *I... 723 DenseSet<Value *> Users; 724 AliasSetTracker WriteSet(*AA); 725 bool JAfterStart = IAfterStart; 726 BasicBlock::iterator J = llvm::next(I); 727 for (unsigned ss = 0; J != E && ss <= SearchLimit; ++J, ++ss) { 728 if (J == Start) JAfterStart = true; 729 730 // Determine if J uses I, if so, exit the loop. 731 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !FastDep); 732 if (FastDep) { 733 // Note: For this heuristic to be effective, independent operations 734 // must tend to be intermixed. This is likely to be true from some 735 // kinds of grouped loop unrolling (but not the generic LLVM pass), 736 // but otherwise may require some kind of reordering pass. 737 738 // When using fast dependency analysis, 739 // stop searching after first use: 740 if (UsesI) break; 741 } else { 742 if (UsesI) continue; 743 } 744 745 // J does not use I, and comes before the first use of I, so it can be 746 // merged with I if the instructions are compatible. 747 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue; 748 749 // J is a candidate for merging with I. 750 if (!PairableInsts.size() || 751 PairableInsts[PairableInsts.size()-1] != I) { 752 PairableInsts.push_back(I); 753 } 754 755 CandidatePairs.insert(ValuePair(I, J)); 756 757 // The next call to this function must start after the last instruction 758 // selected during this invocation. 759 if (JAfterStart) { 760 Start = llvm::next(J); 761 IAfterStart = JAfterStart = false; 762 } 763 764 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 765 << *I << " <-> " << *J << "\n"); 766 767 // If we have already found too many pairs, break here and this function 768 // will be called again starting after the last instruction selected 769 // during this invocation. 770 if (PairableInsts.size() >= MaxInsts) { 771 ShouldContinue = true; 772 break; 773 } 774 } 775 776 if (ShouldContinue) 777 break; 778 } 779 780 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 781 << " instructions with candidate pairs\n"); 782 783 return ShouldContinue; 784 } 785 786 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 787 // it looks for pairs such that both members have an input which is an 788 // output of PI or PJ. 789 void BBVectorize::computePairsConnectedTo( 790 std::multimap<Value *, Value *> &CandidatePairs, 791 std::vector<Value *> &PairableInsts, 792 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 793 ValuePair P) { 794 // For each possible pairing for this variable, look at the uses of 795 // the first value... 796 for (Value::use_iterator I = P.first->use_begin(), 797 E = P.first->use_end(); I != E; ++I) { 798 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 799 800 // For each use of the first variable, look for uses of the second 801 // variable... 802 for (Value::use_iterator J = P.second->use_begin(), 803 E2 = P.second->use_end(); J != E2; ++J) { 804 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 805 806 // Look for <I, J>: 807 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 808 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 809 810 // Look for <J, I>: 811 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 812 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 813 } 814 815 if (SplatBreaksChain) continue; 816 // Look for cases where just the first value in the pair is used by 817 // both members of another pair (splatting). 818 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 819 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 820 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 821 } 822 } 823 824 if (SplatBreaksChain) return; 825 // Look for cases where just the second value in the pair is used by 826 // both members of another pair (splatting). 827 for (Value::use_iterator I = P.second->use_begin(), 828 E = P.second->use_end(); I != E; ++I) { 829 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 830 831 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 832 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 833 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 834 } 835 } 836 } 837 838 // This function figures out which pairs are connected. Two pairs are 839 // connected if some output of the first pair forms an input to both members 840 // of the second pair. 841 void BBVectorize::computeConnectedPairs( 842 std::multimap<Value *, Value *> &CandidatePairs, 843 std::vector<Value *> &PairableInsts, 844 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 845 846 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 847 PE = PairableInsts.end(); PI != PE; ++PI) { 848 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 849 850 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 851 P != choiceRange.second; ++P) 852 computePairsConnectedTo(CandidatePairs, PairableInsts, 853 ConnectedPairs, *P); 854 } 855 856 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 857 << " pair connections.\n"); 858 } 859 860 // This function builds a set of use tuples such that <A, B> is in the set 861 // if B is in the use tree of A. If B is in the use tree of A, then B 862 // depends on the output of A. 863 void BBVectorize::buildDepMap( 864 BasicBlock &BB, 865 std::multimap<Value *, Value *> &CandidatePairs, 866 std::vector<Value *> &PairableInsts, 867 DenseSet<ValuePair> &PairableInstUsers) { 868 DenseSet<Value *> IsInPair; 869 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 870 E = CandidatePairs.end(); C != E; ++C) { 871 IsInPair.insert(C->first); 872 IsInPair.insert(C->second); 873 } 874 875 // Iterate through the basic block, recording all Users of each 876 // pairable instruction. 877 878 BasicBlock::iterator E = BB.end(); 879 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 880 if (IsInPair.find(I) == IsInPair.end()) continue; 881 882 DenseSet<Value *> Users; 883 AliasSetTracker WriteSet(*AA); 884 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 885 (void) trackUsesOfI(Users, WriteSet, I, J); 886 887 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 888 U != E; ++U) 889 PairableInstUsers.insert(ValuePair(I, *U)); 890 } 891 } 892 893 // Returns true if an input to pair P is an output of pair Q and also an 894 // input of pair Q is an output of pair P. If this is the case, then these 895 // two pairs cannot be simultaneously fused. 896 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 897 DenseSet<ValuePair> &PairableInstUsers, 898 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 899 // Two pairs are in conflict if they are mutual Users of eachother. 900 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 901 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 902 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 903 PairableInstUsers.count(ValuePair(P.second, Q.second)); 904 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 905 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 906 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 907 PairableInstUsers.count(ValuePair(Q.second, P.second)); 908 if (PairableInstUserMap) { 909 // FIXME: The expensive part of the cycle check is not so much the cycle 910 // check itself but this edge insertion procedure. This needs some 911 // profiling and probably a different data structure (same is true of 912 // most uses of std::multimap). 913 if (PUsesQ) { 914 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 915 if (!isSecondInIteratorPair(P, QPairRange)) 916 PairableInstUserMap->insert(VPPair(Q, P)); 917 } 918 if (QUsesP) { 919 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 920 if (!isSecondInIteratorPair(Q, PPairRange)) 921 PairableInstUserMap->insert(VPPair(P, Q)); 922 } 923 } 924 925 return (QUsesP && PUsesQ); 926 } 927 928 // This function walks the use graph of current pairs to see if, starting 929 // from P, the walk returns to P. 930 bool BBVectorize::pairWillFormCycle(ValuePair P, 931 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 932 DenseSet<ValuePair> &CurrentPairs) { 933 DEBUG(if (DebugCycleCheck) 934 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 935 << *P.second << "\n"); 936 // A lookup table of visisted pairs is kept because the PairableInstUserMap 937 // contains non-direct associations. 938 DenseSet<ValuePair> Visited; 939 std::vector<ValuePair> Q; 940 // General depth-first post-order traversal: 941 Q.push_back(P); 942 while (!Q.empty()) { 943 ValuePair QTop = Q.back(); 944 945 Visited.insert(QTop); 946 Q.pop_back(); 947 948 DEBUG(if (DebugCycleCheck) 949 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 950 << *QTop.second << "\n"); 951 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 952 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 953 C != QPairRange.second; ++C) { 954 if (C->second == P) { 955 DEBUG(dbgs() 956 << "BBV: rejected to prevent non-trivial cycle formation: " 957 << *C->first.first << " <-> " << *C->first.second << "\n"); 958 return true; 959 } 960 961 if (CurrentPairs.count(C->second) > 0 && 962 Visited.count(C->second) == 0) 963 Q.push_back(C->second); 964 } 965 } 966 967 return false; 968 } 969 970 // This function builds the initial tree of connected pairs with the 971 // pair J at the root. 972 void BBVectorize::buildInitialTreeFor( 973 std::multimap<Value *, Value *> &CandidatePairs, 974 std::vector<Value *> &PairableInsts, 975 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 976 DenseSet<ValuePair> &PairableInstUsers, 977 DenseMap<Value *, Value *> &ChosenPairs, 978 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 979 // Each of these pairs is viewed as the root node of a Tree. The Tree 980 // is then walked (depth-first). As this happens, we keep track of 981 // the pairs that compose the Tree and the maximum depth of the Tree. 982 std::vector<ValuePairWithDepth> Q; 983 // General depth-first post-order traversal: 984 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 985 while (!Q.empty()) { 986 ValuePairWithDepth QTop = Q.back(); 987 988 // Push each child onto the queue: 989 bool MoreChildren = false; 990 size_t MaxChildDepth = QTop.second; 991 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 992 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 993 k != qtRange.second; ++k) { 994 // Make sure that this child pair is still a candidate: 995 bool IsStillCand = false; 996 VPIteratorPair checkRange = 997 CandidatePairs.equal_range(k->second.first); 998 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 999 m != checkRange.second; ++m) { 1000 if (m->second == k->second.second) { 1001 IsStillCand = true; 1002 break; 1003 } 1004 } 1005 1006 if (IsStillCand) { 1007 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1008 if (C == Tree.end()) { 1009 size_t d = getDepthFactor(k->second.first); 1010 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1011 MoreChildren = true; 1012 } else { 1013 MaxChildDepth = std::max(MaxChildDepth, C->second); 1014 } 1015 } 1016 } 1017 1018 if (!MoreChildren) { 1019 // Record the current pair as part of the Tree: 1020 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1021 Q.pop_back(); 1022 } 1023 } 1024 } 1025 1026 // Given some initial tree, prune it by removing conflicting pairs (pairs 1027 // that cannot be simultaneously chosen for vectorization). 1028 void BBVectorize::pruneTreeFor( 1029 std::multimap<Value *, Value *> &CandidatePairs, 1030 std::vector<Value *> &PairableInsts, 1031 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1032 DenseSet<ValuePair> &PairableInstUsers, 1033 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1034 DenseMap<Value *, Value *> &ChosenPairs, 1035 DenseMap<ValuePair, size_t> &Tree, 1036 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1037 bool UseCycleCheck) { 1038 std::vector<ValuePairWithDepth> Q; 1039 // General depth-first post-order traversal: 1040 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1041 while (!Q.empty()) { 1042 ValuePairWithDepth QTop = Q.back(); 1043 PrunedTree.insert(QTop.first); 1044 Q.pop_back(); 1045 1046 // Visit each child, pruning as necessary... 1047 DenseMap<ValuePair, size_t> BestChilden; 1048 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1049 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1050 K != QTopRange.second; ++K) { 1051 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1052 if (C == Tree.end()) continue; 1053 1054 // This child is in the Tree, now we need to make sure it is the 1055 // best of any conflicting children. There could be multiple 1056 // conflicting children, so first, determine if we're keeping 1057 // this child, then delete conflicting children as necessary. 1058 1059 // It is also necessary to guard against pairing-induced 1060 // dependencies. Consider instructions a .. x .. y .. b 1061 // such that (a,b) are to be fused and (x,y) are to be fused 1062 // but a is an input to x and b is an output from y. This 1063 // means that y cannot be moved after b but x must be moved 1064 // after b for (a,b) to be fused. In other words, after 1065 // fusing (a,b) we have y .. a/b .. x where y is an input 1066 // to a/b and x is an output to a/b: x and y can no longer 1067 // be legally fused. To prevent this condition, we must 1068 // make sure that a child pair added to the Tree is not 1069 // both an input and output of an already-selected pair. 1070 1071 // Pairing-induced dependencies can also form from more complicated 1072 // cycles. The pair vs. pair conflicts are easy to check, and so 1073 // that is done explicitly for "fast rejection", and because for 1074 // child vs. child conflicts, we may prefer to keep the current 1075 // pair in preference to the already-selected child. 1076 DenseSet<ValuePair> CurrentPairs; 1077 1078 bool CanAdd = true; 1079 for (DenseMap<ValuePair, size_t>::iterator C2 1080 = BestChilden.begin(), E2 = BestChilden.end(); 1081 C2 != E2; ++C2) { 1082 if (C2->first.first == C->first.first || 1083 C2->first.first == C->first.second || 1084 C2->first.second == C->first.first || 1085 C2->first.second == C->first.second || 1086 pairsConflict(C2->first, C->first, PairableInstUsers, 1087 UseCycleCheck ? &PairableInstUserMap : 0)) { 1088 if (C2->second >= C->second) { 1089 CanAdd = false; 1090 break; 1091 } 1092 1093 CurrentPairs.insert(C2->first); 1094 } 1095 } 1096 if (!CanAdd) continue; 1097 1098 // Even worse, this child could conflict with another node already 1099 // selected for the Tree. If that is the case, ignore this child. 1100 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1101 E2 = PrunedTree.end(); T != E2; ++T) { 1102 if (T->first == C->first.first || 1103 T->first == C->first.second || 1104 T->second == C->first.first || 1105 T->second == C->first.second || 1106 pairsConflict(*T, C->first, PairableInstUsers, 1107 UseCycleCheck ? &PairableInstUserMap : 0)) { 1108 CanAdd = false; 1109 break; 1110 } 1111 1112 CurrentPairs.insert(*T); 1113 } 1114 if (!CanAdd) continue; 1115 1116 // And check the queue too... 1117 for (std::vector<ValuePairWithDepth>::iterator C2 = Q.begin(), 1118 E2 = Q.end(); C2 != E2; ++C2) { 1119 if (C2->first.first == C->first.first || 1120 C2->first.first == C->first.second || 1121 C2->first.second == C->first.first || 1122 C2->first.second == C->first.second || 1123 pairsConflict(C2->first, C->first, PairableInstUsers, 1124 UseCycleCheck ? &PairableInstUserMap : 0)) { 1125 CanAdd = false; 1126 break; 1127 } 1128 1129 CurrentPairs.insert(C2->first); 1130 } 1131 if (!CanAdd) continue; 1132 1133 // Last but not least, check for a conflict with any of the 1134 // already-chosen pairs. 1135 for (DenseMap<Value *, Value *>::iterator C2 = 1136 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1137 C2 != E2; ++C2) { 1138 if (pairsConflict(*C2, C->first, PairableInstUsers, 1139 UseCycleCheck ? &PairableInstUserMap : 0)) { 1140 CanAdd = false; 1141 break; 1142 } 1143 1144 CurrentPairs.insert(*C2); 1145 } 1146 if (!CanAdd) continue; 1147 1148 // To check for non-trivial cycles formed by the addition of the 1149 // current pair we've formed a list of all relevant pairs, now use a 1150 // graph walk to check for a cycle. We start from the current pair and 1151 // walk the use tree to see if we again reach the current pair. If we 1152 // do, then the current pair is rejected. 1153 1154 // FIXME: It may be more efficient to use a topological-ordering 1155 // algorithm to improve the cycle check. This should be investigated. 1156 if (UseCycleCheck && 1157 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1158 continue; 1159 1160 // This child can be added, but we may have chosen it in preference 1161 // to an already-selected child. Check for this here, and if a 1162 // conflict is found, then remove the previously-selected child 1163 // before adding this one in its place. 1164 for (DenseMap<ValuePair, size_t>::iterator C2 1165 = BestChilden.begin(); C2 != BestChilden.end();) { 1166 if (C2->first.first == C->first.first || 1167 C2->first.first == C->first.second || 1168 C2->first.second == C->first.first || 1169 C2->first.second == C->first.second || 1170 pairsConflict(C2->first, C->first, PairableInstUsers)) 1171 BestChilden.erase(C2++); 1172 else 1173 ++C2; 1174 } 1175 1176 BestChilden.insert(ValuePairWithDepth(C->first, C->second)); 1177 } 1178 1179 for (DenseMap<ValuePair, size_t>::iterator C 1180 = BestChilden.begin(), E2 = BestChilden.end(); 1181 C != E2; ++C) { 1182 size_t DepthF = getDepthFactor(C->first.first); 1183 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1184 } 1185 } 1186 } 1187 1188 // This function finds the best tree of mututally-compatible connected 1189 // pairs, given the choice of root pairs as an iterator range. 1190 void BBVectorize::findBestTreeFor( 1191 std::multimap<Value *, Value *> &CandidatePairs, 1192 std::vector<Value *> &PairableInsts, 1193 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1194 DenseSet<ValuePair> &PairableInstUsers, 1195 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1196 DenseMap<Value *, Value *> &ChosenPairs, 1197 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1198 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1199 bool UseCycleCheck) { 1200 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1201 J != ChoiceRange.second; ++J) { 1202 1203 // Before going any further, make sure that this pair does not 1204 // conflict with any already-selected pairs (see comment below 1205 // near the Tree pruning for more details). 1206 DenseSet<ValuePair> ChosenPairSet; 1207 bool DoesConflict = false; 1208 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1209 E = ChosenPairs.end(); C != E; ++C) { 1210 if (pairsConflict(*C, *J, PairableInstUsers, 1211 UseCycleCheck ? &PairableInstUserMap : 0)) { 1212 DoesConflict = true; 1213 break; 1214 } 1215 1216 ChosenPairSet.insert(*C); 1217 } 1218 if (DoesConflict) continue; 1219 1220 if (UseCycleCheck && 1221 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1222 continue; 1223 1224 DenseMap<ValuePair, size_t> Tree; 1225 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1226 PairableInstUsers, ChosenPairs, Tree, *J); 1227 1228 // Because we'll keep the child with the largest depth, the largest 1229 // depth is still the same in the unpruned Tree. 1230 size_t MaxDepth = Tree.lookup(*J); 1231 1232 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1233 << *J->first << " <-> " << *J->second << "} of depth " << 1234 MaxDepth << " and size " << Tree.size() << "\n"); 1235 1236 // At this point the Tree has been constructed, but, may contain 1237 // contradictory children (meaning that different children of 1238 // some tree node may be attempting to fuse the same instruction). 1239 // So now we walk the tree again, in the case of a conflict, 1240 // keep only the child with the largest depth. To break a tie, 1241 // favor the first child. 1242 1243 DenseSet<ValuePair> PrunedTree; 1244 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1245 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1246 PrunedTree, *J, UseCycleCheck); 1247 1248 size_t EffSize = 0; 1249 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1250 E = PrunedTree.end(); S != E; ++S) 1251 EffSize += getDepthFactor(S->first); 1252 1253 DEBUG(if (DebugPairSelection) 1254 dbgs() << "BBV: found pruned Tree for pair {" 1255 << *J->first << " <-> " << *J->second << "} of depth " << 1256 MaxDepth << " and size " << PrunedTree.size() << 1257 " (effective size: " << EffSize << ")\n"); 1258 if (MaxDepth >= ReqChainDepth && EffSize > BestEffSize) { 1259 BestMaxDepth = MaxDepth; 1260 BestEffSize = EffSize; 1261 BestTree = PrunedTree; 1262 } 1263 } 1264 } 1265 1266 // Given the list of candidate pairs, this function selects those 1267 // that will be fused into vector instructions. 1268 void BBVectorize::choosePairs( 1269 std::multimap<Value *, Value *> &CandidatePairs, 1270 std::vector<Value *> &PairableInsts, 1271 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1272 DenseSet<ValuePair> &PairableInstUsers, 1273 DenseMap<Value *, Value *>& ChosenPairs) { 1274 bool UseCycleCheck = CandidatePairs.size() <= MaxCandPairsForCycleCheck; 1275 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1276 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1277 E = PairableInsts.end(); I != E; ++I) { 1278 // The number of possible pairings for this variable: 1279 size_t NumChoices = CandidatePairs.count(*I); 1280 if (!NumChoices) continue; 1281 1282 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1283 1284 // The best pair to choose and its tree: 1285 size_t BestMaxDepth = 0, BestEffSize = 0; 1286 DenseSet<ValuePair> BestTree; 1287 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1288 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1289 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1290 UseCycleCheck); 1291 1292 // A tree has been chosen (or not) at this point. If no tree was 1293 // chosen, then this instruction, I, cannot be paired (and is no longer 1294 // considered). 1295 1296 DEBUG(if (BestTree.size() > 0) 1297 dbgs() << "BBV: selected pairs in the best tree for: " 1298 << *cast<Instruction>(*I) << "\n"); 1299 1300 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1301 SE2 = BestTree.end(); S != SE2; ++S) { 1302 // Insert the members of this tree into the list of chosen pairs. 1303 ChosenPairs.insert(ValuePair(S->first, S->second)); 1304 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1305 *S->second << "\n"); 1306 1307 // Remove all candidate pairs that have values in the chosen tree. 1308 for (std::multimap<Value *, Value *>::iterator K = 1309 CandidatePairs.begin(); K != CandidatePairs.end();) { 1310 if (K->first == S->first || K->second == S->first || 1311 K->second == S->second || K->first == S->second) { 1312 // Don't remove the actual pair chosen so that it can be used 1313 // in subsequent tree selections. 1314 if (!(K->first == S->first && K->second == S->second)) 1315 CandidatePairs.erase(K++); 1316 else 1317 ++K; 1318 } else { 1319 ++K; 1320 } 1321 } 1322 } 1323 } 1324 1325 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1326 } 1327 1328 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1329 unsigned n = 0) { 1330 if (!I->hasName()) 1331 return ""; 1332 1333 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1334 (n > 0 ? "." + utostr(n) : "")).str(); 1335 } 1336 1337 // Returns the value that is to be used as the pointer input to the vector 1338 // instruction that fuses I with J. 1339 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1340 Instruction *I, Instruction *J, unsigned o, 1341 bool &FlipMemInputs) { 1342 Value *IPtr, *JPtr; 1343 unsigned IAlignment, JAlignment; 1344 int64_t OffsetInElmts; 1345 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1346 OffsetInElmts); 1347 1348 // The pointer value is taken to be the one with the lowest offset. 1349 Value *VPtr; 1350 if (OffsetInElmts > 0) { 1351 VPtr = IPtr; 1352 } else { 1353 FlipMemInputs = true; 1354 VPtr = JPtr; 1355 } 1356 1357 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType(); 1358 Type *VArgType = getVecTypeForPair(ArgType); 1359 Type *VArgPtrType = PointerType::get(VArgType, 1360 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1361 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1362 /* insert before */ FlipMemInputs ? J : I); 1363 } 1364 1365 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1366 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 1367 unsigned IdxOffset, std::vector<Constant*> &Mask) { 1368 for (unsigned v = 0; v < NumElem/2; ++v) { 1369 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1370 if (m < 0) { 1371 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1372 } else { 1373 unsigned mm = m + (int) IdxOffset; 1374 if (m >= (int) NumInElem) 1375 mm += (int) NumInElem; 1376 1377 Mask[v+MaskOffset] = 1378 ConstantInt::get(Type::getInt32Ty(Context), mm); 1379 } 1380 } 1381 } 1382 1383 // Returns the value that is to be used as the vector-shuffle mask to the 1384 // vector instruction that fuses I with J. 1385 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1386 Instruction *I, Instruction *J) { 1387 // This is the shuffle mask. We need to append the second 1388 // mask to the first, and the numbers need to be adjusted. 1389 1390 Type *ArgType = I->getType(); 1391 Type *VArgType = getVecTypeForPair(ArgType); 1392 1393 // Get the total number of elements in the fused vector type. 1394 // By definition, this must equal the number of elements in 1395 // the final mask. 1396 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1397 std::vector<Constant*> Mask(NumElem); 1398 1399 Type *OpType = I->getOperand(0)->getType(); 1400 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements(); 1401 1402 // For the mask from the first pair... 1403 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask); 1404 1405 // For the mask from the second pair... 1406 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem, 1407 Mask); 1408 1409 return ConstantVector::get(Mask); 1410 } 1411 1412 // Returns the value to be used as the specified operand of the vector 1413 // instruction that fuses I with J. 1414 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1415 Instruction *J, unsigned o, bool FlipMemInputs) { 1416 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1417 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1418 1419 // Compute the fused vector type for this operand 1420 Type *ArgType = I->getOperand(o)->getType(); 1421 VectorType *VArgType = getVecTypeForPair(ArgType); 1422 1423 Instruction *L = I, *H = J; 1424 if (FlipMemInputs) { 1425 L = J; 1426 H = I; 1427 } 1428 1429 if (ArgType->isVectorTy()) { 1430 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1431 std::vector<Constant*> Mask(numElem); 1432 for (unsigned v = 0; v < numElem; ++v) 1433 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1434 1435 Instruction *BV = new ShuffleVectorInst(L->getOperand(o), 1436 H->getOperand(o), 1437 ConstantVector::get(Mask), 1438 getReplacementName(I, true, o)); 1439 BV->insertBefore(J); 1440 return BV; 1441 } 1442 1443 // If these two inputs are the output of another vector instruction, 1444 // then we should use that output directly. It might be necessary to 1445 // permute it first. [When pairings are fused recursively, you can 1446 // end up with cases where a large vector is decomposed into scalars 1447 // using extractelement instructions, then built into size-2 1448 // vectors using insertelement and the into larger vectors using 1449 // shuffles. InstCombine does not simplify all of these cases well, 1450 // and so we make sure that shuffles are generated here when possible. 1451 ExtractElementInst *LEE 1452 = dyn_cast<ExtractElementInst>(L->getOperand(o)); 1453 ExtractElementInst *HEE 1454 = dyn_cast<ExtractElementInst>(H->getOperand(o)); 1455 1456 if (LEE && HEE && 1457 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) { 1458 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType()); 1459 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue(); 1460 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue(); 1461 if (LEE->getOperand(0) == HEE->getOperand(0)) { 1462 if (LowIndx == 0 && HighIndx == 1) 1463 return LEE->getOperand(0); 1464 1465 std::vector<Constant*> Mask(2); 1466 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1467 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1468 1469 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1470 UndefValue::get(EEType), 1471 ConstantVector::get(Mask), 1472 getReplacementName(I, true, o)); 1473 BV->insertBefore(J); 1474 return BV; 1475 } 1476 1477 std::vector<Constant*> Mask(2); 1478 HighIndx += EEType->getNumElements(); 1479 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1480 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1481 1482 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1483 HEE->getOperand(0), 1484 ConstantVector::get(Mask), 1485 getReplacementName(I, true, o)); 1486 BV->insertBefore(J); 1487 return BV; 1488 } 1489 1490 Instruction *BV1 = InsertElementInst::Create( 1491 UndefValue::get(VArgType), 1492 L->getOperand(o), CV0, 1493 getReplacementName(I, true, o, 1)); 1494 BV1->insertBefore(I); 1495 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o), 1496 CV1, 1497 getReplacementName(I, true, o, 2)); 1498 BV2->insertBefore(J); 1499 return BV2; 1500 } 1501 1502 // This function creates an array of values that will be used as the inputs 1503 // to the vector instruction that fuses I with J. 1504 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1505 Instruction *I, Instruction *J, 1506 SmallVector<Value *, 3> &ReplacedOperands, 1507 bool &FlipMemInputs) { 1508 FlipMemInputs = false; 1509 unsigned NumOperands = I->getNumOperands(); 1510 1511 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1512 // Iterate backward so that we look at the store pointer 1513 // first and know whether or not we need to flip the inputs. 1514 1515 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1516 // This is the pointer for a load/store instruction. 1517 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 1518 FlipMemInputs); 1519 continue; 1520 } else if (isa<CallInst>(I) && o == NumOperands-1) { 1521 Function *F = cast<CallInst>(I)->getCalledFunction(); 1522 unsigned IID = F->getIntrinsicID(); 1523 BasicBlock &BB = *I->getParent(); 1524 1525 Module *M = BB.getParent()->getParent(); 1526 Type *ArgType = I->getType(); 1527 Type *VArgType = getVecTypeForPair(ArgType); 1528 1529 // FIXME: is it safe to do this here? 1530 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 1531 (Intrinsic::ID) IID, VArgType); 1532 continue; 1533 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 1534 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 1535 continue; 1536 } 1537 1538 ReplacedOperands[o] = 1539 getReplacementInput(Context, I, J, o, FlipMemInputs); 1540 } 1541 } 1542 1543 // This function creates two values that represent the outputs of the 1544 // original I and J instructions. These are generally vector shuffles 1545 // or extracts. In many cases, these will end up being unused and, thus, 1546 // eliminated by later passes. 1547 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 1548 Instruction *J, Instruction *K, 1549 Instruction *&InsertionPt, 1550 Instruction *&K1, Instruction *&K2, 1551 bool &FlipMemInputs) { 1552 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1553 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1554 1555 if (isa<StoreInst>(I)) { 1556 AA->replaceWithNewValue(I, K); 1557 AA->replaceWithNewValue(J, K); 1558 } else { 1559 Type *IType = I->getType(); 1560 Type *VType = getVecTypeForPair(IType); 1561 1562 if (IType->isVectorTy()) { 1563 unsigned numElem = cast<VectorType>(IType)->getNumElements(); 1564 std::vector<Constant*> Mask1(numElem), Mask2(numElem); 1565 for (unsigned v = 0; v < numElem; ++v) { 1566 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1567 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v); 1568 } 1569 1570 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 1571 ConstantVector::get( 1572 FlipMemInputs ? Mask2 : Mask1), 1573 getReplacementName(K, false, 1)); 1574 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 1575 ConstantVector::get( 1576 FlipMemInputs ? Mask1 : Mask2), 1577 getReplacementName(K, false, 2)); 1578 } else { 1579 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 1580 getReplacementName(K, false, 1)); 1581 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 1582 getReplacementName(K, false, 2)); 1583 } 1584 1585 K1->insertAfter(K); 1586 K2->insertAfter(K1); 1587 InsertionPt = K2; 1588 } 1589 } 1590 1591 // Move all uses of the function I (including pairing-induced uses) after J. 1592 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 1593 std::multimap<Value *, Value *> &LoadMoveSet, 1594 Instruction *I, Instruction *J) { 1595 // Skip to the first instruction past I. 1596 BasicBlock::iterator L = BB.begin(); 1597 for (; cast<Instruction>(L) != I; ++L); 1598 ++L; 1599 1600 DenseSet<Value *> Users; 1601 AliasSetTracker WriteSet(*AA); 1602 for (; cast<Instruction>(L) != J; ++L) 1603 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 1604 1605 assert(cast<Instruction>(L) == J && 1606 "Tracking has not proceeded far enough to check for dependencies"); 1607 // If J is now in the use set of I, then trackUsesOfI will return true 1608 // and we have a dependency cycle (and the fusing operation must abort). 1609 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 1610 } 1611 1612 // Move all uses of the function I (including pairing-induced uses) after J. 1613 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 1614 std::multimap<Value *, Value *> &LoadMoveSet, 1615 Instruction *&InsertionPt, 1616 Instruction *I, Instruction *J) { 1617 // Skip to the first instruction past I. 1618 BasicBlock::iterator L = BB.begin(); 1619 for (; cast<Instruction>(L) != I; ++L); 1620 ++L; 1621 1622 DenseSet<Value *> Users; 1623 AliasSetTracker WriteSet(*AA); 1624 for (; cast<Instruction>(L) != J;) { 1625 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 1626 // Move this instruction 1627 Instruction *InstToMove = L; ++L; 1628 1629 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 1630 " to after " << *InsertionPt << "\n"); 1631 InstToMove->removeFromParent(); 1632 InstToMove->insertAfter(InsertionPt); 1633 InsertionPt = InstToMove; 1634 } else { 1635 ++L; 1636 } 1637 } 1638 } 1639 1640 // Collect all load instruction that are in the move set of a given first 1641 // pair member. These loads depend on the first instruction, I, and so need 1642 // to be moved after J (the second instruction) when the pair is fused. 1643 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 1644 DenseMap<Value *, Value *> &ChosenPairs, 1645 std::multimap<Value *, Value *> &LoadMoveSet, 1646 Instruction *I) { 1647 // Skip to the first instruction past I. 1648 BasicBlock::iterator L = BB.begin(); 1649 for (; cast<Instruction>(L) != I; ++L); 1650 ++L; 1651 1652 DenseSet<Value *> Users; 1653 AliasSetTracker WriteSet(*AA); 1654 1655 // Note: We cannot end the loop when we reach J because J could be moved 1656 // farther down the use chain by another instruction pairing. Also, J 1657 // could be before I if this is an inverted input. 1658 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 1659 if (trackUsesOfI(Users, WriteSet, I, L)) { 1660 if (L->mayReadFromMemory()) 1661 LoadMoveSet.insert(ValuePair(L, I)); 1662 } 1663 } 1664 } 1665 1666 // In cases where both load/stores and the computation of their pointers 1667 // are chosen for vectorization, we can end up in a situation where the 1668 // aliasing analysis starts returning different query results as the 1669 // process of fusing instruction pairs continues. Because the algorithm 1670 // relies on finding the same use trees here as were found earlier, we'll 1671 // need to precompute the necessary aliasing information here and then 1672 // manually update it during the fusion process. 1673 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 1674 std::vector<Value *> &PairableInsts, 1675 DenseMap<Value *, Value *> &ChosenPairs, 1676 std::multimap<Value *, Value *> &LoadMoveSet) { 1677 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1678 PIE = PairableInsts.end(); PI != PIE; ++PI) { 1679 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 1680 if (P == ChosenPairs.end()) continue; 1681 1682 Instruction *I = cast<Instruction>(P->first); 1683 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 1684 } 1685 } 1686 1687 // This function fuses the chosen instruction pairs into vector instructions, 1688 // taking care preserve any needed scalar outputs and, then, it reorders the 1689 // remaining instructions as needed (users of the first member of the pair 1690 // need to be moved to after the location of the second member of the pair 1691 // because the vector instruction is inserted in the location of the pair's 1692 // second member). 1693 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 1694 std::vector<Value *> &PairableInsts, 1695 DenseMap<Value *, Value *> &ChosenPairs) { 1696 LLVMContext& Context = BB.getContext(); 1697 1698 // During the vectorization process, the order of the pairs to be fused 1699 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 1700 // list. After a pair is fused, the flipped pair is removed from the list. 1701 std::vector<ValuePair> FlippedPairs; 1702 FlippedPairs.reserve(ChosenPairs.size()); 1703 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 1704 E = ChosenPairs.end(); P != E; ++P) 1705 FlippedPairs.push_back(ValuePair(P->second, P->first)); 1706 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 1707 E = FlippedPairs.end(); P != E; ++P) 1708 ChosenPairs.insert(*P); 1709 1710 std::multimap<Value *, Value *> LoadMoveSet; 1711 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 1712 1713 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 1714 1715 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 1716 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 1717 if (P == ChosenPairs.end()) { 1718 ++PI; 1719 continue; 1720 } 1721 1722 if (getDepthFactor(P->first) == 0) { 1723 // These instructions are not really fused, but are tracked as though 1724 // they are. Any case in which it would be interesting to fuse them 1725 // will be taken care of by InstCombine. 1726 --NumFusedOps; 1727 ++PI; 1728 continue; 1729 } 1730 1731 Instruction *I = cast<Instruction>(P->first), 1732 *J = cast<Instruction>(P->second); 1733 1734 DEBUG(dbgs() << "BBV: fusing: " << *I << 1735 " <-> " << *J << "\n"); 1736 1737 // Remove the pair and flipped pair from the list. 1738 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 1739 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 1740 ChosenPairs.erase(FP); 1741 ChosenPairs.erase(P); 1742 1743 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 1744 DEBUG(dbgs() << "BBV: fusion of: " << *I << 1745 " <-> " << *J << 1746 " aborted because of non-trivial dependency cycle\n"); 1747 --NumFusedOps; 1748 ++PI; 1749 continue; 1750 } 1751 1752 bool FlipMemInputs; 1753 unsigned NumOperands = I->getNumOperands(); 1754 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 1755 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 1756 FlipMemInputs); 1757 1758 // Make a copy of the original operation, change its type to the vector 1759 // type and replace its operands with the vector operands. 1760 Instruction *K = I->clone(); 1761 if (I->hasName()) K->takeName(I); 1762 1763 if (!isa<StoreInst>(K)) 1764 K->mutateType(getVecTypeForPair(I->getType())); 1765 1766 for (unsigned o = 0; o < NumOperands; ++o) 1767 K->setOperand(o, ReplacedOperands[o]); 1768 1769 // If we've flipped the memory inputs, make sure that we take the correct 1770 // alignment. 1771 if (FlipMemInputs) { 1772 if (isa<StoreInst>(K)) 1773 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 1774 else 1775 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 1776 } 1777 1778 K->insertAfter(J); 1779 1780 // Instruction insertion point: 1781 Instruction *InsertionPt = K; 1782 Instruction *K1 = 0, *K2 = 0; 1783 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 1784 FlipMemInputs); 1785 1786 // The use tree of the first original instruction must be moved to after 1787 // the location of the second instruction. The entire use tree of the 1788 // first instruction is disjoint from the input tree of the second 1789 // (by definition), and so commutes with it. 1790 1791 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 1792 1793 if (!isa<StoreInst>(I)) { 1794 I->replaceAllUsesWith(K1); 1795 J->replaceAllUsesWith(K2); 1796 AA->replaceWithNewValue(I, K1); 1797 AA->replaceWithNewValue(J, K2); 1798 } 1799 1800 // Instructions that may read from memory may be in the load move set. 1801 // Once an instruction is fused, we no longer need its move set, and so 1802 // the values of the map never need to be updated. However, when a load 1803 // is fused, we need to merge the entries from both instructions in the 1804 // pair in case those instructions were in the move set of some other 1805 // yet-to-be-fused pair. The loads in question are the keys of the map. 1806 if (I->mayReadFromMemory()) { 1807 std::vector<ValuePair> NewSetMembers; 1808 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 1809 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 1810 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 1811 N != IPairRange.second; ++N) 1812 NewSetMembers.push_back(ValuePair(K, N->second)); 1813 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 1814 N != JPairRange.second; ++N) 1815 NewSetMembers.push_back(ValuePair(K, N->second)); 1816 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 1817 AE = NewSetMembers.end(); A != AE; ++A) 1818 LoadMoveSet.insert(*A); 1819 } 1820 1821 // Before removing I, set the iterator to the next instruction. 1822 PI = llvm::next(BasicBlock::iterator(I)); 1823 if (cast<Instruction>(PI) == J) 1824 ++PI; 1825 1826 SE->forgetValue(I); 1827 SE->forgetValue(J); 1828 I->eraseFromParent(); 1829 J->eraseFromParent(); 1830 } 1831 1832 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 1833 } 1834} 1835 1836char BBVectorize::ID = 0; 1837static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 1838INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1839INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 1840INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 1841INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1842 1843BasicBlockPass *llvm::createBBVectorizePass() { 1844 return new BBVectorize(); 1845} 1846 1847