BBVectorize.cpp revision 6173ed95daf2f209fe3883faee45967e4800ae75
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 must 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 // The powi intrinsic is special because only the first argument is 651 // vectorized, the second arguments must be equal. 652 CallInst *CI = dyn_cast<CallInst>(I); 653 Function *FI; 654 if (CI && (FI = CI->getCalledFunction()) && 655 FI->getIntrinsicID() == Intrinsic::powi) { 656 657 Value *A1I = CI->getArgOperand(1), 658 *A1J = cast<CallInst>(J)->getArgOperand(1); 659 const SCEV *A1ISCEV = SE->getSCEV(A1I), 660 *A1JSCEV = SE->getSCEV(A1J); 661 return (A1ISCEV == A1JSCEV); 662 } 663 664 return true; 665 } 666 667 // Figure out whether or not J uses I and update the users and write-set 668 // structures associated with I. Specifically, Users represents the set of 669 // instructions that depend on I. WriteSet represents the set 670 // of memory locations that are dependent on I. If UpdateUsers is true, 671 // and J uses I, then Users is updated to contain J and WriteSet is updated 672 // to contain any memory locations to which J writes. The function returns 673 // true if J uses I. By default, alias analysis is used to determine 674 // whether J reads from memory that overlaps with a location in WriteSet. 675 // If LoadMoveSet is not null, then it is a previously-computed multimap 676 // where the key is the memory-based user instruction and the value is 677 // the instruction to be compared with I. So, if LoadMoveSet is provided, 678 // then the alias analysis is not used. This is necessary because this 679 // function is called during the process of moving instructions during 680 // vectorization and the results of the alias analysis are not stable during 681 // that process. 682 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 683 AliasSetTracker &WriteSet, Instruction *I, 684 Instruction *J, bool UpdateUsers, 685 std::multimap<Value *, Value *> *LoadMoveSet) { 686 bool UsesI = false; 687 688 // This instruction may already be marked as a user due, for example, to 689 // being a member of a selected pair. 690 if (Users.count(J)) 691 UsesI = true; 692 693 if (!UsesI) 694 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 695 JU != JE; ++JU) { 696 Value *V = *JU; 697 if (I == V || Users.count(V)) { 698 UsesI = true; 699 break; 700 } 701 } 702 if (!UsesI && J->mayReadFromMemory()) { 703 if (LoadMoveSet) { 704 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 705 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 706 } else { 707 for (AliasSetTracker::iterator W = WriteSet.begin(), 708 WE = WriteSet.end(); W != WE; ++W) { 709 if (W->aliasesUnknownInst(J, *AA)) { 710 UsesI = true; 711 break; 712 } 713 } 714 } 715 } 716 717 if (UsesI && UpdateUsers) { 718 if (J->mayWriteToMemory()) WriteSet.add(J); 719 Users.insert(J); 720 } 721 722 return UsesI; 723 } 724 725 // This function iterates over all instruction pairs in the provided 726 // basic block and collects all candidate pairs for vectorization. 727 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 728 BasicBlock::iterator &Start, 729 std::multimap<Value *, Value *> &CandidatePairs, 730 std::vector<Value *> &PairableInsts) { 731 BasicBlock::iterator E = BB.end(); 732 if (Start == E) return false; 733 734 bool ShouldContinue = false, IAfterStart = false; 735 for (BasicBlock::iterator I = Start++; I != E; ++I) { 736 if (I == Start) IAfterStart = true; 737 738 bool IsSimpleLoadStore; 739 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 740 741 // Look for an instruction with which to pair instruction *I... 742 DenseSet<Value *> Users; 743 AliasSetTracker WriteSet(*AA); 744 bool JAfterStart = IAfterStart; 745 BasicBlock::iterator J = llvm::next(I); 746 for (unsigned ss = 0; J != E && ss <= SearchLimit; ++J, ++ss) { 747 if (J == Start) JAfterStart = true; 748 749 // Determine if J uses I, if so, exit the loop. 750 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !FastDep); 751 if (FastDep) { 752 // Note: For this heuristic to be effective, independent operations 753 // must tend to be intermixed. This is likely to be true from some 754 // kinds of grouped loop unrolling (but not the generic LLVM pass), 755 // but otherwise may require some kind of reordering pass. 756 757 // When using fast dependency analysis, 758 // stop searching after first use: 759 if (UsesI) break; 760 } else { 761 if (UsesI) continue; 762 } 763 764 // J does not use I, and comes before the first use of I, so it can be 765 // merged with I if the instructions are compatible. 766 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue; 767 768 // J is a candidate for merging with I. 769 if (!PairableInsts.size() || 770 PairableInsts[PairableInsts.size()-1] != I) { 771 PairableInsts.push_back(I); 772 } 773 774 CandidatePairs.insert(ValuePair(I, J)); 775 776 // The next call to this function must start after the last instruction 777 // selected during this invocation. 778 if (JAfterStart) { 779 Start = llvm::next(J); 780 IAfterStart = JAfterStart = false; 781 } 782 783 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 784 << *I << " <-> " << *J << "\n"); 785 786 // If we have already found too many pairs, break here and this function 787 // will be called again starting after the last instruction selected 788 // during this invocation. 789 if (PairableInsts.size() >= MaxInsts) { 790 ShouldContinue = true; 791 break; 792 } 793 } 794 795 if (ShouldContinue) 796 break; 797 } 798 799 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 800 << " instructions with candidate pairs\n"); 801 802 return ShouldContinue; 803 } 804 805 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 806 // it looks for pairs such that both members have an input which is an 807 // output of PI or PJ. 808 void BBVectorize::computePairsConnectedTo( 809 std::multimap<Value *, Value *> &CandidatePairs, 810 std::vector<Value *> &PairableInsts, 811 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 812 ValuePair P) { 813 // For each possible pairing for this variable, look at the uses of 814 // the first value... 815 for (Value::use_iterator I = P.first->use_begin(), 816 E = P.first->use_end(); I != E; ++I) { 817 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 818 819 // For each use of the first variable, look for uses of the second 820 // variable... 821 for (Value::use_iterator J = P.second->use_begin(), 822 E2 = P.second->use_end(); J != E2; ++J) { 823 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 824 825 // Look for <I, J>: 826 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 827 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 828 829 // Look for <J, I>: 830 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 831 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 832 } 833 834 if (SplatBreaksChain) continue; 835 // Look for cases where just the first value in the pair is used by 836 // both members of another pair (splatting). 837 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 838 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 839 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 840 } 841 } 842 843 if (SplatBreaksChain) return; 844 // Look for cases where just the second value in the pair is used by 845 // both members of another pair (splatting). 846 for (Value::use_iterator I = P.second->use_begin(), 847 E = P.second->use_end(); I != E; ++I) { 848 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 849 850 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 851 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 852 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 853 } 854 } 855 } 856 857 // This function figures out which pairs are connected. Two pairs are 858 // connected if some output of the first pair forms an input to both members 859 // of the second pair. 860 void BBVectorize::computeConnectedPairs( 861 std::multimap<Value *, Value *> &CandidatePairs, 862 std::vector<Value *> &PairableInsts, 863 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 864 865 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 866 PE = PairableInsts.end(); PI != PE; ++PI) { 867 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 868 869 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 870 P != choiceRange.second; ++P) 871 computePairsConnectedTo(CandidatePairs, PairableInsts, 872 ConnectedPairs, *P); 873 } 874 875 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 876 << " pair connections.\n"); 877 } 878 879 // This function builds a set of use tuples such that <A, B> is in the set 880 // if B is in the use tree of A. If B is in the use tree of A, then B 881 // depends on the output of A. 882 void BBVectorize::buildDepMap( 883 BasicBlock &BB, 884 std::multimap<Value *, Value *> &CandidatePairs, 885 std::vector<Value *> &PairableInsts, 886 DenseSet<ValuePair> &PairableInstUsers) { 887 DenseSet<Value *> IsInPair; 888 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 889 E = CandidatePairs.end(); C != E; ++C) { 890 IsInPair.insert(C->first); 891 IsInPair.insert(C->second); 892 } 893 894 // Iterate through the basic block, recording all Users of each 895 // pairable instruction. 896 897 BasicBlock::iterator E = BB.end(); 898 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 899 if (IsInPair.find(I) == IsInPair.end()) continue; 900 901 DenseSet<Value *> Users; 902 AliasSetTracker WriteSet(*AA); 903 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 904 (void) trackUsesOfI(Users, WriteSet, I, J); 905 906 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 907 U != E; ++U) 908 PairableInstUsers.insert(ValuePair(I, *U)); 909 } 910 } 911 912 // Returns true if an input to pair P is an output of pair Q and also an 913 // input of pair Q is an output of pair P. If this is the case, then these 914 // two pairs cannot be simultaneously fused. 915 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 916 DenseSet<ValuePair> &PairableInstUsers, 917 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 918 // Two pairs are in conflict if they are mutual Users of eachother. 919 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 920 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 921 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 922 PairableInstUsers.count(ValuePair(P.second, Q.second)); 923 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 924 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 925 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 926 PairableInstUsers.count(ValuePair(Q.second, P.second)); 927 if (PairableInstUserMap) { 928 // FIXME: The expensive part of the cycle check is not so much the cycle 929 // check itself but this edge insertion procedure. This needs some 930 // profiling and probably a different data structure (same is true of 931 // most uses of std::multimap). 932 if (PUsesQ) { 933 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 934 if (!isSecondInIteratorPair(P, QPairRange)) 935 PairableInstUserMap->insert(VPPair(Q, P)); 936 } 937 if (QUsesP) { 938 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 939 if (!isSecondInIteratorPair(Q, PPairRange)) 940 PairableInstUserMap->insert(VPPair(P, Q)); 941 } 942 } 943 944 return (QUsesP && PUsesQ); 945 } 946 947 // This function walks the use graph of current pairs to see if, starting 948 // from P, the walk returns to P. 949 bool BBVectorize::pairWillFormCycle(ValuePair P, 950 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 951 DenseSet<ValuePair> &CurrentPairs) { 952 DEBUG(if (DebugCycleCheck) 953 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 954 << *P.second << "\n"); 955 // A lookup table of visisted pairs is kept because the PairableInstUserMap 956 // contains non-direct associations. 957 DenseSet<ValuePair> Visited; 958 SmallVector<ValuePair, 32> Q; 959 // General depth-first post-order traversal: 960 Q.push_back(P); 961 do { 962 ValuePair QTop = Q.pop_back_val(); 963 Visited.insert(QTop); 964 965 DEBUG(if (DebugCycleCheck) 966 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 967 << *QTop.second << "\n"); 968 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 969 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 970 C != QPairRange.second; ++C) { 971 if (C->second == P) { 972 DEBUG(dbgs() 973 << "BBV: rejected to prevent non-trivial cycle formation: " 974 << *C->first.first << " <-> " << *C->first.second << "\n"); 975 return true; 976 } 977 978 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 979 Q.push_back(C->second); 980 } 981 } while (!Q.empty()); 982 983 return false; 984 } 985 986 // This function builds the initial tree of connected pairs with the 987 // pair J at the root. 988 void BBVectorize::buildInitialTreeFor( 989 std::multimap<Value *, Value *> &CandidatePairs, 990 std::vector<Value *> &PairableInsts, 991 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 992 DenseSet<ValuePair> &PairableInstUsers, 993 DenseMap<Value *, Value *> &ChosenPairs, 994 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 995 // Each of these pairs is viewed as the root node of a Tree. The Tree 996 // is then walked (depth-first). As this happens, we keep track of 997 // the pairs that compose the Tree and the maximum depth of the Tree. 998 SmallVector<ValuePairWithDepth, 32> Q; 999 // General depth-first post-order traversal: 1000 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1001 do { 1002 ValuePairWithDepth QTop = Q.back(); 1003 1004 // Push each child onto the queue: 1005 bool MoreChildren = false; 1006 size_t MaxChildDepth = QTop.second; 1007 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1008 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1009 k != qtRange.second; ++k) { 1010 // Make sure that this child pair is still a candidate: 1011 bool IsStillCand = false; 1012 VPIteratorPair checkRange = 1013 CandidatePairs.equal_range(k->second.first); 1014 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1015 m != checkRange.second; ++m) { 1016 if (m->second == k->second.second) { 1017 IsStillCand = true; 1018 break; 1019 } 1020 } 1021 1022 if (IsStillCand) { 1023 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1024 if (C == Tree.end()) { 1025 size_t d = getDepthFactor(k->second.first); 1026 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1027 MoreChildren = true; 1028 } else { 1029 MaxChildDepth = std::max(MaxChildDepth, C->second); 1030 } 1031 } 1032 } 1033 1034 if (!MoreChildren) { 1035 // Record the current pair as part of the Tree: 1036 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1037 Q.pop_back(); 1038 } 1039 } while (!Q.empty()); 1040 } 1041 1042 // Given some initial tree, prune it by removing conflicting pairs (pairs 1043 // that cannot be simultaneously chosen for vectorization). 1044 void BBVectorize::pruneTreeFor( 1045 std::multimap<Value *, Value *> &CandidatePairs, 1046 std::vector<Value *> &PairableInsts, 1047 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1048 DenseSet<ValuePair> &PairableInstUsers, 1049 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1050 DenseMap<Value *, Value *> &ChosenPairs, 1051 DenseMap<ValuePair, size_t> &Tree, 1052 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1053 bool UseCycleCheck) { 1054 SmallVector<ValuePairWithDepth, 32> Q; 1055 // General depth-first post-order traversal: 1056 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1057 do { 1058 ValuePairWithDepth QTop = Q.pop_back_val(); 1059 PrunedTree.insert(QTop.first); 1060 1061 // Visit each child, pruning as necessary... 1062 DenseMap<ValuePair, size_t> BestChildren; 1063 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1064 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1065 K != QTopRange.second; ++K) { 1066 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1067 if (C == Tree.end()) continue; 1068 1069 // This child is in the Tree, now we need to make sure it is the 1070 // best of any conflicting children. There could be multiple 1071 // conflicting children, so first, determine if we're keeping 1072 // this child, then delete conflicting children as necessary. 1073 1074 // It is also necessary to guard against pairing-induced 1075 // dependencies. Consider instructions a .. x .. y .. b 1076 // such that (a,b) are to be fused and (x,y) are to be fused 1077 // but a is an input to x and b is an output from y. This 1078 // means that y cannot be moved after b but x must be moved 1079 // after b for (a,b) to be fused. In other words, after 1080 // fusing (a,b) we have y .. a/b .. x where y is an input 1081 // to a/b and x is an output to a/b: x and y can no longer 1082 // be legally fused. To prevent this condition, we must 1083 // make sure that a child pair added to the Tree is not 1084 // both an input and output of an already-selected pair. 1085 1086 // Pairing-induced dependencies can also form from more complicated 1087 // cycles. The pair vs. pair conflicts are easy to check, and so 1088 // that is done explicitly for "fast rejection", and because for 1089 // child vs. child conflicts, we may prefer to keep the current 1090 // pair in preference to the already-selected child. 1091 DenseSet<ValuePair> CurrentPairs; 1092 1093 bool CanAdd = true; 1094 for (DenseMap<ValuePair, size_t>::iterator C2 1095 = BestChildren.begin(), E2 = BestChildren.end(); 1096 C2 != E2; ++C2) { 1097 if (C2->first.first == C->first.first || 1098 C2->first.first == C->first.second || 1099 C2->first.second == C->first.first || 1100 C2->first.second == C->first.second || 1101 pairsConflict(C2->first, C->first, PairableInstUsers, 1102 UseCycleCheck ? &PairableInstUserMap : 0)) { 1103 if (C2->second >= C->second) { 1104 CanAdd = false; 1105 break; 1106 } 1107 1108 CurrentPairs.insert(C2->first); 1109 } 1110 } 1111 if (!CanAdd) continue; 1112 1113 // Even worse, this child could conflict with another node already 1114 // selected for the Tree. If that is the case, ignore this child. 1115 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1116 E2 = PrunedTree.end(); T != E2; ++T) { 1117 if (T->first == C->first.first || 1118 T->first == C->first.second || 1119 T->second == C->first.first || 1120 T->second == C->first.second || 1121 pairsConflict(*T, C->first, PairableInstUsers, 1122 UseCycleCheck ? &PairableInstUserMap : 0)) { 1123 CanAdd = false; 1124 break; 1125 } 1126 1127 CurrentPairs.insert(*T); 1128 } 1129 if (!CanAdd) continue; 1130 1131 // And check the queue too... 1132 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1133 E2 = Q.end(); C2 != E2; ++C2) { 1134 if (C2->first.first == C->first.first || 1135 C2->first.first == C->first.second || 1136 C2->first.second == C->first.first || 1137 C2->first.second == C->first.second || 1138 pairsConflict(C2->first, C->first, PairableInstUsers, 1139 UseCycleCheck ? &PairableInstUserMap : 0)) { 1140 CanAdd = false; 1141 break; 1142 } 1143 1144 CurrentPairs.insert(C2->first); 1145 } 1146 if (!CanAdd) continue; 1147 1148 // Last but not least, check for a conflict with any of the 1149 // already-chosen pairs. 1150 for (DenseMap<Value *, Value *>::iterator C2 = 1151 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1152 C2 != E2; ++C2) { 1153 if (pairsConflict(*C2, C->first, PairableInstUsers, 1154 UseCycleCheck ? &PairableInstUserMap : 0)) { 1155 CanAdd = false; 1156 break; 1157 } 1158 1159 CurrentPairs.insert(*C2); 1160 } 1161 if (!CanAdd) continue; 1162 1163 // To check for non-trivial cycles formed by the addition of the 1164 // current pair we've formed a list of all relevant pairs, now use a 1165 // graph walk to check for a cycle. We start from the current pair and 1166 // walk the use tree to see if we again reach the current pair. If we 1167 // do, then the current pair is rejected. 1168 1169 // FIXME: It may be more efficient to use a topological-ordering 1170 // algorithm to improve the cycle check. This should be investigated. 1171 if (UseCycleCheck && 1172 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1173 continue; 1174 1175 // This child can be added, but we may have chosen it in preference 1176 // to an already-selected child. Check for this here, and if a 1177 // conflict is found, then remove the previously-selected child 1178 // before adding this one in its place. 1179 for (DenseMap<ValuePair, size_t>::iterator C2 1180 = BestChildren.begin(); C2 != BestChildren.end();) { 1181 if (C2->first.first == C->first.first || 1182 C2->first.first == C->first.second || 1183 C2->first.second == C->first.first || 1184 C2->first.second == C->first.second || 1185 pairsConflict(C2->first, C->first, PairableInstUsers)) 1186 BestChildren.erase(C2++); 1187 else 1188 ++C2; 1189 } 1190 1191 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1192 } 1193 1194 for (DenseMap<ValuePair, size_t>::iterator C 1195 = BestChildren.begin(), E2 = BestChildren.end(); 1196 C != E2; ++C) { 1197 size_t DepthF = getDepthFactor(C->first.first); 1198 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1199 } 1200 } while (!Q.empty()); 1201 } 1202 1203 // This function finds the best tree of mututally-compatible connected 1204 // pairs, given the choice of root pairs as an iterator range. 1205 void BBVectorize::findBestTreeFor( 1206 std::multimap<Value *, Value *> &CandidatePairs, 1207 std::vector<Value *> &PairableInsts, 1208 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1209 DenseSet<ValuePair> &PairableInstUsers, 1210 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1211 DenseMap<Value *, Value *> &ChosenPairs, 1212 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1213 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1214 bool UseCycleCheck) { 1215 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1216 J != ChoiceRange.second; ++J) { 1217 1218 // Before going any further, make sure that this pair does not 1219 // conflict with any already-selected pairs (see comment below 1220 // near the Tree pruning for more details). 1221 DenseSet<ValuePair> ChosenPairSet; 1222 bool DoesConflict = false; 1223 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1224 E = ChosenPairs.end(); C != E; ++C) { 1225 if (pairsConflict(*C, *J, PairableInstUsers, 1226 UseCycleCheck ? &PairableInstUserMap : 0)) { 1227 DoesConflict = true; 1228 break; 1229 } 1230 1231 ChosenPairSet.insert(*C); 1232 } 1233 if (DoesConflict) continue; 1234 1235 if (UseCycleCheck && 1236 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1237 continue; 1238 1239 DenseMap<ValuePair, size_t> Tree; 1240 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1241 PairableInstUsers, ChosenPairs, Tree, *J); 1242 1243 // Because we'll keep the child with the largest depth, the largest 1244 // depth is still the same in the unpruned Tree. 1245 size_t MaxDepth = Tree.lookup(*J); 1246 1247 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1248 << *J->first << " <-> " << *J->second << "} of depth " << 1249 MaxDepth << " and size " << Tree.size() << "\n"); 1250 1251 // At this point the Tree has been constructed, but, may contain 1252 // contradictory children (meaning that different children of 1253 // some tree node may be attempting to fuse the same instruction). 1254 // So now we walk the tree again, in the case of a conflict, 1255 // keep only the child with the largest depth. To break a tie, 1256 // favor the first child. 1257 1258 DenseSet<ValuePair> PrunedTree; 1259 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1260 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1261 PrunedTree, *J, UseCycleCheck); 1262 1263 size_t EffSize = 0; 1264 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1265 E = PrunedTree.end(); S != E; ++S) 1266 EffSize += getDepthFactor(S->first); 1267 1268 DEBUG(if (DebugPairSelection) 1269 dbgs() << "BBV: found pruned Tree for pair {" 1270 << *J->first << " <-> " << *J->second << "} of depth " << 1271 MaxDepth << " and size " << PrunedTree.size() << 1272 " (effective size: " << EffSize << ")\n"); 1273 if (MaxDepth >= ReqChainDepth && EffSize > BestEffSize) { 1274 BestMaxDepth = MaxDepth; 1275 BestEffSize = EffSize; 1276 BestTree = PrunedTree; 1277 } 1278 } 1279 } 1280 1281 // Given the list of candidate pairs, this function selects those 1282 // that will be fused into vector instructions. 1283 void BBVectorize::choosePairs( 1284 std::multimap<Value *, Value *> &CandidatePairs, 1285 std::vector<Value *> &PairableInsts, 1286 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1287 DenseSet<ValuePair> &PairableInstUsers, 1288 DenseMap<Value *, Value *>& ChosenPairs) { 1289 bool UseCycleCheck = CandidatePairs.size() <= MaxCandPairsForCycleCheck; 1290 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1291 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1292 E = PairableInsts.end(); I != E; ++I) { 1293 // The number of possible pairings for this variable: 1294 size_t NumChoices = CandidatePairs.count(*I); 1295 if (!NumChoices) continue; 1296 1297 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1298 1299 // The best pair to choose and its tree: 1300 size_t BestMaxDepth = 0, BestEffSize = 0; 1301 DenseSet<ValuePair> BestTree; 1302 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1303 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1304 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1305 UseCycleCheck); 1306 1307 // A tree has been chosen (or not) at this point. If no tree was 1308 // chosen, then this instruction, I, cannot be paired (and is no longer 1309 // considered). 1310 1311 DEBUG(if (BestTree.size() > 0) 1312 dbgs() << "BBV: selected pairs in the best tree for: " 1313 << *cast<Instruction>(*I) << "\n"); 1314 1315 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1316 SE2 = BestTree.end(); S != SE2; ++S) { 1317 // Insert the members of this tree into the list of chosen pairs. 1318 ChosenPairs.insert(ValuePair(S->first, S->second)); 1319 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1320 *S->second << "\n"); 1321 1322 // Remove all candidate pairs that have values in the chosen tree. 1323 for (std::multimap<Value *, Value *>::iterator K = 1324 CandidatePairs.begin(); K != CandidatePairs.end();) { 1325 if (K->first == S->first || K->second == S->first || 1326 K->second == S->second || K->first == S->second) { 1327 // Don't remove the actual pair chosen so that it can be used 1328 // in subsequent tree selections. 1329 if (!(K->first == S->first && K->second == S->second)) 1330 CandidatePairs.erase(K++); 1331 else 1332 ++K; 1333 } else { 1334 ++K; 1335 } 1336 } 1337 } 1338 } 1339 1340 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1341 } 1342 1343 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1344 unsigned n = 0) { 1345 if (!I->hasName()) 1346 return ""; 1347 1348 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1349 (n > 0 ? "." + utostr(n) : "")).str(); 1350 } 1351 1352 // Returns the value that is to be used as the pointer input to the vector 1353 // instruction that fuses I with J. 1354 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1355 Instruction *I, Instruction *J, unsigned o, 1356 bool &FlipMemInputs) { 1357 Value *IPtr, *JPtr; 1358 unsigned IAlignment, JAlignment; 1359 int64_t OffsetInElmts; 1360 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1361 OffsetInElmts); 1362 1363 // The pointer value is taken to be the one with the lowest offset. 1364 Value *VPtr; 1365 if (OffsetInElmts > 0) { 1366 VPtr = IPtr; 1367 } else { 1368 FlipMemInputs = true; 1369 VPtr = JPtr; 1370 } 1371 1372 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType(); 1373 Type *VArgType = getVecTypeForPair(ArgType); 1374 Type *VArgPtrType = PointerType::get(VArgType, 1375 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1376 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1377 /* insert before */ FlipMemInputs ? J : I); 1378 } 1379 1380 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1381 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 1382 unsigned IdxOffset, std::vector<Constant*> &Mask) { 1383 for (unsigned v = 0; v < NumElem/2; ++v) { 1384 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1385 if (m < 0) { 1386 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1387 } else { 1388 unsigned mm = m + (int) IdxOffset; 1389 if (m >= (int) NumInElem) 1390 mm += (int) NumInElem; 1391 1392 Mask[v+MaskOffset] = 1393 ConstantInt::get(Type::getInt32Ty(Context), mm); 1394 } 1395 } 1396 } 1397 1398 // Returns the value that is to be used as the vector-shuffle mask to the 1399 // vector instruction that fuses I with J. 1400 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1401 Instruction *I, Instruction *J) { 1402 // This is the shuffle mask. We need to append the second 1403 // mask to the first, and the numbers need to be adjusted. 1404 1405 Type *ArgType = I->getType(); 1406 Type *VArgType = getVecTypeForPair(ArgType); 1407 1408 // Get the total number of elements in the fused vector type. 1409 // By definition, this must equal the number of elements in 1410 // the final mask. 1411 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1412 std::vector<Constant*> Mask(NumElem); 1413 1414 Type *OpType = I->getOperand(0)->getType(); 1415 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements(); 1416 1417 // For the mask from the first pair... 1418 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask); 1419 1420 // For the mask from the second pair... 1421 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem, 1422 Mask); 1423 1424 return ConstantVector::get(Mask); 1425 } 1426 1427 // Returns the value to be used as the specified operand of the vector 1428 // instruction that fuses I with J. 1429 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1430 Instruction *J, unsigned o, bool FlipMemInputs) { 1431 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1432 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1433 1434 // Compute the fused vector type for this operand 1435 Type *ArgType = I->getOperand(o)->getType(); 1436 VectorType *VArgType = getVecTypeForPair(ArgType); 1437 1438 Instruction *L = I, *H = J; 1439 if (FlipMemInputs) { 1440 L = J; 1441 H = I; 1442 } 1443 1444 if (ArgType->isVectorTy()) { 1445 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1446 std::vector<Constant*> Mask(numElem); 1447 for (unsigned v = 0; v < numElem; ++v) 1448 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1449 1450 Instruction *BV = new ShuffleVectorInst(L->getOperand(o), 1451 H->getOperand(o), 1452 ConstantVector::get(Mask), 1453 getReplacementName(I, true, o)); 1454 BV->insertBefore(J); 1455 return BV; 1456 } 1457 1458 // If these two inputs are the output of another vector instruction, 1459 // then we should use that output directly. It might be necessary to 1460 // permute it first. [When pairings are fused recursively, you can 1461 // end up with cases where a large vector is decomposed into scalars 1462 // using extractelement instructions, then built into size-2 1463 // vectors using insertelement and the into larger vectors using 1464 // shuffles. InstCombine does not simplify all of these cases well, 1465 // and so we make sure that shuffles are generated here when possible. 1466 ExtractElementInst *LEE 1467 = dyn_cast<ExtractElementInst>(L->getOperand(o)); 1468 ExtractElementInst *HEE 1469 = dyn_cast<ExtractElementInst>(H->getOperand(o)); 1470 1471 if (LEE && HEE && 1472 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) { 1473 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType()); 1474 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue(); 1475 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue(); 1476 if (LEE->getOperand(0) == HEE->getOperand(0)) { 1477 if (LowIndx == 0 && HighIndx == 1) 1478 return LEE->getOperand(0); 1479 1480 std::vector<Constant*> Mask(2); 1481 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1482 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1483 1484 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1485 UndefValue::get(EEType), 1486 ConstantVector::get(Mask), 1487 getReplacementName(I, true, o)); 1488 BV->insertBefore(J); 1489 return BV; 1490 } 1491 1492 std::vector<Constant*> Mask(2); 1493 HighIndx += EEType->getNumElements(); 1494 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1495 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1496 1497 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1498 HEE->getOperand(0), 1499 ConstantVector::get(Mask), 1500 getReplacementName(I, true, o)); 1501 BV->insertBefore(J); 1502 return BV; 1503 } 1504 1505 Instruction *BV1 = InsertElementInst::Create( 1506 UndefValue::get(VArgType), 1507 L->getOperand(o), CV0, 1508 getReplacementName(I, true, o, 1)); 1509 BV1->insertBefore(I); 1510 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o), 1511 CV1, 1512 getReplacementName(I, true, o, 2)); 1513 BV2->insertBefore(J); 1514 return BV2; 1515 } 1516 1517 // This function creates an array of values that will be used as the inputs 1518 // to the vector instruction that fuses I with J. 1519 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1520 Instruction *I, Instruction *J, 1521 SmallVector<Value *, 3> &ReplacedOperands, 1522 bool &FlipMemInputs) { 1523 FlipMemInputs = false; 1524 unsigned NumOperands = I->getNumOperands(); 1525 1526 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1527 // Iterate backward so that we look at the store pointer 1528 // first and know whether or not we need to flip the inputs. 1529 1530 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1531 // This is the pointer for a load/store instruction. 1532 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 1533 FlipMemInputs); 1534 continue; 1535 } else if (isa<CallInst>(I)) { 1536 Function *F = cast<CallInst>(I)->getCalledFunction(); 1537 unsigned IID = F->getIntrinsicID(); 1538 if (o == NumOperands-1) { 1539 BasicBlock &BB = *I->getParent(); 1540 1541 Module *M = BB.getParent()->getParent(); 1542 Type *ArgType = I->getType(); 1543 Type *VArgType = getVecTypeForPair(ArgType); 1544 1545 // FIXME: is it safe to do this here? 1546 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 1547 (Intrinsic::ID) IID, VArgType); 1548 continue; 1549 } else if (IID == Intrinsic::powi && o == 1) { 1550 // The second argument of powi is a single integer and we've already 1551 // checked that both arguments are equal. As a result, we just keep 1552 // I's second argument. 1553 ReplacedOperands[o] = I->getOperand(o); 1554 continue; 1555 } 1556 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 1557 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 1558 continue; 1559 } 1560 1561 ReplacedOperands[o] = 1562 getReplacementInput(Context, I, J, o, FlipMemInputs); 1563 } 1564 } 1565 1566 // This function creates two values that represent the outputs of the 1567 // original I and J instructions. These are generally vector shuffles 1568 // or extracts. In many cases, these will end up being unused and, thus, 1569 // eliminated by later passes. 1570 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 1571 Instruction *J, Instruction *K, 1572 Instruction *&InsertionPt, 1573 Instruction *&K1, Instruction *&K2, 1574 bool &FlipMemInputs) { 1575 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1576 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1577 1578 if (isa<StoreInst>(I)) { 1579 AA->replaceWithNewValue(I, K); 1580 AA->replaceWithNewValue(J, K); 1581 } else { 1582 Type *IType = I->getType(); 1583 Type *VType = getVecTypeForPair(IType); 1584 1585 if (IType->isVectorTy()) { 1586 unsigned numElem = cast<VectorType>(IType)->getNumElements(); 1587 std::vector<Constant*> Mask1(numElem), Mask2(numElem); 1588 for (unsigned v = 0; v < numElem; ++v) { 1589 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1590 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v); 1591 } 1592 1593 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 1594 ConstantVector::get( 1595 FlipMemInputs ? Mask2 : Mask1), 1596 getReplacementName(K, false, 1)); 1597 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 1598 ConstantVector::get( 1599 FlipMemInputs ? Mask1 : Mask2), 1600 getReplacementName(K, false, 2)); 1601 } else { 1602 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 1603 getReplacementName(K, false, 1)); 1604 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 1605 getReplacementName(K, false, 2)); 1606 } 1607 1608 K1->insertAfter(K); 1609 K2->insertAfter(K1); 1610 InsertionPt = K2; 1611 } 1612 } 1613 1614 // Move all uses of the function I (including pairing-induced uses) after J. 1615 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 1616 std::multimap<Value *, Value *> &LoadMoveSet, 1617 Instruction *I, Instruction *J) { 1618 // Skip to the first instruction past I. 1619 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1620 1621 DenseSet<Value *> Users; 1622 AliasSetTracker WriteSet(*AA); 1623 for (; cast<Instruction>(L) != J; ++L) 1624 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 1625 1626 assert(cast<Instruction>(L) == J && 1627 "Tracking has not proceeded far enough to check for dependencies"); 1628 // If J is now in the use set of I, then trackUsesOfI will return true 1629 // and we have a dependency cycle (and the fusing operation must abort). 1630 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 1631 } 1632 1633 // Move all uses of the function I (including pairing-induced uses) after J. 1634 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 1635 std::multimap<Value *, Value *> &LoadMoveSet, 1636 Instruction *&InsertionPt, 1637 Instruction *I, Instruction *J) { 1638 // Skip to the first instruction past I. 1639 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1640 1641 DenseSet<Value *> Users; 1642 AliasSetTracker WriteSet(*AA); 1643 for (; cast<Instruction>(L) != J;) { 1644 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 1645 // Move this instruction 1646 Instruction *InstToMove = L; ++L; 1647 1648 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 1649 " to after " << *InsertionPt << "\n"); 1650 InstToMove->removeFromParent(); 1651 InstToMove->insertAfter(InsertionPt); 1652 InsertionPt = InstToMove; 1653 } else { 1654 ++L; 1655 } 1656 } 1657 } 1658 1659 // Collect all load instruction that are in the move set of a given first 1660 // pair member. These loads depend on the first instruction, I, and so need 1661 // to be moved after J (the second instruction) when the pair is fused. 1662 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 1663 DenseMap<Value *, Value *> &ChosenPairs, 1664 std::multimap<Value *, Value *> &LoadMoveSet, 1665 Instruction *I) { 1666 // Skip to the first instruction past I. 1667 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 1668 1669 DenseSet<Value *> Users; 1670 AliasSetTracker WriteSet(*AA); 1671 1672 // Note: We cannot end the loop when we reach J because J could be moved 1673 // farther down the use chain by another instruction pairing. Also, J 1674 // could be before I if this is an inverted input. 1675 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 1676 if (trackUsesOfI(Users, WriteSet, I, L)) { 1677 if (L->mayReadFromMemory()) 1678 LoadMoveSet.insert(ValuePair(L, I)); 1679 } 1680 } 1681 } 1682 1683 // In cases where both load/stores and the computation of their pointers 1684 // are chosen for vectorization, we can end up in a situation where the 1685 // aliasing analysis starts returning different query results as the 1686 // process of fusing instruction pairs continues. Because the algorithm 1687 // relies on finding the same use trees here as were found earlier, we'll 1688 // need to precompute the necessary aliasing information here and then 1689 // manually update it during the fusion process. 1690 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 1691 std::vector<Value *> &PairableInsts, 1692 DenseMap<Value *, Value *> &ChosenPairs, 1693 std::multimap<Value *, Value *> &LoadMoveSet) { 1694 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1695 PIE = PairableInsts.end(); PI != PIE; ++PI) { 1696 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 1697 if (P == ChosenPairs.end()) continue; 1698 1699 Instruction *I = cast<Instruction>(P->first); 1700 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 1701 } 1702 } 1703 1704 // This function fuses the chosen instruction pairs into vector instructions, 1705 // taking care preserve any needed scalar outputs and, then, it reorders the 1706 // remaining instructions as needed (users of the first member of the pair 1707 // need to be moved to after the location of the second member of the pair 1708 // because the vector instruction is inserted in the location of the pair's 1709 // second member). 1710 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 1711 std::vector<Value *> &PairableInsts, 1712 DenseMap<Value *, Value *> &ChosenPairs) { 1713 LLVMContext& Context = BB.getContext(); 1714 1715 // During the vectorization process, the order of the pairs to be fused 1716 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 1717 // list. After a pair is fused, the flipped pair is removed from the list. 1718 std::vector<ValuePair> FlippedPairs; 1719 FlippedPairs.reserve(ChosenPairs.size()); 1720 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 1721 E = ChosenPairs.end(); P != E; ++P) 1722 FlippedPairs.push_back(ValuePair(P->second, P->first)); 1723 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 1724 E = FlippedPairs.end(); P != E; ++P) 1725 ChosenPairs.insert(*P); 1726 1727 std::multimap<Value *, Value *> LoadMoveSet; 1728 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 1729 1730 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 1731 1732 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 1733 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 1734 if (P == ChosenPairs.end()) { 1735 ++PI; 1736 continue; 1737 } 1738 1739 if (getDepthFactor(P->first) == 0) { 1740 // These instructions are not really fused, but are tracked as though 1741 // they are. Any case in which it would be interesting to fuse them 1742 // will be taken care of by InstCombine. 1743 --NumFusedOps; 1744 ++PI; 1745 continue; 1746 } 1747 1748 Instruction *I = cast<Instruction>(P->first), 1749 *J = cast<Instruction>(P->second); 1750 1751 DEBUG(dbgs() << "BBV: fusing: " << *I << 1752 " <-> " << *J << "\n"); 1753 1754 // Remove the pair and flipped pair from the list. 1755 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 1756 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 1757 ChosenPairs.erase(FP); 1758 ChosenPairs.erase(P); 1759 1760 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 1761 DEBUG(dbgs() << "BBV: fusion of: " << *I << 1762 " <-> " << *J << 1763 " aborted because of non-trivial dependency cycle\n"); 1764 --NumFusedOps; 1765 ++PI; 1766 continue; 1767 } 1768 1769 bool FlipMemInputs; 1770 unsigned NumOperands = I->getNumOperands(); 1771 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 1772 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 1773 FlipMemInputs); 1774 1775 // Make a copy of the original operation, change its type to the vector 1776 // type and replace its operands with the vector operands. 1777 Instruction *K = I->clone(); 1778 if (I->hasName()) K->takeName(I); 1779 1780 if (!isa<StoreInst>(K)) 1781 K->mutateType(getVecTypeForPair(I->getType())); 1782 1783 for (unsigned o = 0; o < NumOperands; ++o) 1784 K->setOperand(o, ReplacedOperands[o]); 1785 1786 // If we've flipped the memory inputs, make sure that we take the correct 1787 // alignment. 1788 if (FlipMemInputs) { 1789 if (isa<StoreInst>(K)) 1790 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 1791 else 1792 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 1793 } 1794 1795 K->insertAfter(J); 1796 1797 // Instruction insertion point: 1798 Instruction *InsertionPt = K; 1799 Instruction *K1 = 0, *K2 = 0; 1800 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 1801 FlipMemInputs); 1802 1803 // The use tree of the first original instruction must be moved to after 1804 // the location of the second instruction. The entire use tree of the 1805 // first instruction is disjoint from the input tree of the second 1806 // (by definition), and so commutes with it. 1807 1808 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 1809 1810 if (!isa<StoreInst>(I)) { 1811 I->replaceAllUsesWith(K1); 1812 J->replaceAllUsesWith(K2); 1813 AA->replaceWithNewValue(I, K1); 1814 AA->replaceWithNewValue(J, K2); 1815 } 1816 1817 // Instructions that may read from memory may be in the load move set. 1818 // Once an instruction is fused, we no longer need its move set, and so 1819 // the values of the map never need to be updated. However, when a load 1820 // is fused, we need to merge the entries from both instructions in the 1821 // pair in case those instructions were in the move set of some other 1822 // yet-to-be-fused pair. The loads in question are the keys of the map. 1823 if (I->mayReadFromMemory()) { 1824 std::vector<ValuePair> NewSetMembers; 1825 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 1826 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 1827 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 1828 N != IPairRange.second; ++N) 1829 NewSetMembers.push_back(ValuePair(K, N->second)); 1830 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 1831 N != JPairRange.second; ++N) 1832 NewSetMembers.push_back(ValuePair(K, N->second)); 1833 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 1834 AE = NewSetMembers.end(); A != AE; ++A) 1835 LoadMoveSet.insert(*A); 1836 } 1837 1838 // Before removing I, set the iterator to the next instruction. 1839 PI = llvm::next(BasicBlock::iterator(I)); 1840 if (cast<Instruction>(PI) == J) 1841 ++PI; 1842 1843 SE->forgetValue(I); 1844 SE->forgetValue(J); 1845 I->eraseFromParent(); 1846 J->eraseFromParent(); 1847 } 1848 1849 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 1850 } 1851} 1852 1853char BBVectorize::ID = 0; 1854static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 1855INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1856INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 1857INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 1858INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1859 1860BasicBlockPass *llvm::createBBVectorizePass() { 1861 return new BBVectorize(); 1862} 1863 1864