BBVectorize.cpp revision 7e004d177fe76145f75a9417ed2e281f1b9abaf7
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> 70MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 71 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 72 " a full cycle check")); 73 74static cl::opt<bool> 75NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 76 cl::desc("Don't try to vectorize integer values")); 77 78static cl::opt<bool> 79NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 80 cl::desc("Don't try to vectorize floating-point values")); 81 82static cl::opt<bool> 83NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 84 cl::desc("Don't try to vectorize casting (conversion) operations")); 85 86static cl::opt<bool> 87NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 88 cl::desc("Don't try to vectorize floating-point math intrinsics")); 89 90static cl::opt<bool> 91NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 92 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 93 94static cl::opt<bool> 95NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 96 cl::desc("Don't try to vectorize loads and stores")); 97 98static cl::opt<bool> 99AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 100 cl::desc("Only generate aligned loads and stores")); 101 102static cl::opt<bool> 103FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 104 cl::desc("Use a fast instruction dependency analysis")); 105 106#ifndef NDEBUG 107static cl::opt<bool> 108DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 109 cl::init(false), cl::Hidden, 110 cl::desc("When debugging is enabled, output information on the" 111 " instruction-examination process")); 112static cl::opt<bool> 113DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 114 cl::init(false), cl::Hidden, 115 cl::desc("When debugging is enabled, output information on the" 116 " candidate-selection process")); 117static cl::opt<bool> 118DebugPairSelection("bb-vectorize-debug-pair-selection", 119 cl::init(false), cl::Hidden, 120 cl::desc("When debugging is enabled, output information on the" 121 " pair-selection process")); 122static cl::opt<bool> 123DebugCycleCheck("bb-vectorize-debug-cycle-check", 124 cl::init(false), cl::Hidden, 125 cl::desc("When debugging is enabled, output information on the" 126 " cycle-checking process")); 127#endif 128 129STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 130 131namespace { 132 struct BBVectorize : public BasicBlockPass { 133 static char ID; // Pass identification, replacement for typeid 134 BBVectorize() : BasicBlockPass(ID) { 135 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 136 } 137 138 typedef std::pair<Value *, Value *> ValuePair; 139 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 140 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 141 typedef std::pair<std::multimap<Value *, Value *>::iterator, 142 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 143 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 144 std::multimap<ValuePair, ValuePair>::iterator> 145 VPPIteratorPair; 146 147 AliasAnalysis *AA; 148 ScalarEvolution *SE; 149 TargetData *TD; 150 151 // FIXME: const correct? 152 153 bool vectorizePairs(BasicBlock &BB); 154 155 void getCandidatePairs(BasicBlock &BB, 156 std::multimap<Value *, Value *> &CandidatePairs, 157 std::vector<Value *> &PairableInsts); 158 159 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 160 std::vector<Value *> &PairableInsts, 161 std::multimap<ValuePair, ValuePair> &ConnectedPairs); 162 163 void buildDepMap(BasicBlock &BB, 164 std::multimap<Value *, Value *> &CandidatePairs, 165 std::vector<Value *> &PairableInsts, 166 DenseSet<ValuePair> &PairableInstUsers); 167 168 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 169 std::vector<Value *> &PairableInsts, 170 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 171 DenseSet<ValuePair> &PairableInstUsers, 172 DenseMap<Value *, Value *>& ChosenPairs); 173 174 void fuseChosenPairs(BasicBlock &BB, 175 std::vector<Value *> &PairableInsts, 176 DenseMap<Value *, Value *>& ChosenPairs); 177 178 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 179 180 bool areInstsCompatible(Instruction *I, Instruction *J, 181 bool IsSimpleLoadStore); 182 183 bool trackUsesOfI(DenseSet<Value *> &Users, 184 AliasSetTracker &WriteSet, Instruction *I, 185 Instruction *J, bool UpdateUsers = true, 186 std::multimap<Value *, Value *> *LoadMoveSet = 0); 187 188 void computePairsConnectedTo( 189 std::multimap<Value *, Value *> &CandidatePairs, 190 std::vector<Value *> &PairableInsts, 191 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 192 ValuePair P); 193 194 bool pairsConflict(ValuePair P, ValuePair Q, 195 DenseSet<ValuePair> &PairableInstUsers, 196 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 197 198 bool pairWillFormCycle(ValuePair P, 199 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 200 DenseSet<ValuePair> &CurrentPairs); 201 202 void pruneTreeFor( 203 std::multimap<Value *, Value *> &CandidatePairs, 204 std::vector<Value *> &PairableInsts, 205 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 206 DenseSet<ValuePair> &PairableInstUsers, 207 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 208 DenseMap<Value *, Value *> &ChosenPairs, 209 DenseMap<ValuePair, size_t> &Tree, 210 DenseSet<ValuePair> &PrunedTree, ValuePair J, 211 bool UseCycleCheck); 212 213 void buildInitialTreeFor( 214 std::multimap<Value *, Value *> &CandidatePairs, 215 std::vector<Value *> &PairableInsts, 216 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 217 DenseSet<ValuePair> &PairableInstUsers, 218 DenseMap<Value *, Value *> &ChosenPairs, 219 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 220 221 void findBestTreeFor( 222 std::multimap<Value *, Value *> &CandidatePairs, 223 std::vector<Value *> &PairableInsts, 224 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 225 DenseSet<ValuePair> &PairableInstUsers, 226 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 227 DenseMap<Value *, Value *> &ChosenPairs, 228 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 229 size_t &BestEffSize, VPIteratorPair ChoiceRange, 230 bool UseCycleCheck); 231 232 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 233 Instruction *J, unsigned o, bool &FlipMemInputs); 234 235 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 236 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 237 unsigned IdxOffset, std::vector<Constant*> &Mask); 238 239 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 240 Instruction *J); 241 242 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 243 Instruction *J, unsigned o, bool FlipMemInputs); 244 245 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 246 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 247 bool &FlipMemInputs); 248 249 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 250 Instruction *J, Instruction *K, 251 Instruction *&InsertionPt, Instruction *&K1, 252 Instruction *&K2, bool &FlipMemInputs); 253 254 void collectPairLoadMoveSet(BasicBlock &BB, 255 DenseMap<Value *, Value *> &ChosenPairs, 256 std::multimap<Value *, Value *> &LoadMoveSet, 257 Instruction *I); 258 259 void collectLoadMoveSet(BasicBlock &BB, 260 std::vector<Value *> &PairableInsts, 261 DenseMap<Value *, Value *> &ChosenPairs, 262 std::multimap<Value *, Value *> &LoadMoveSet); 263 264 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 265 std::multimap<Value *, Value *> &LoadMoveSet, 266 Instruction *I, Instruction *J); 267 268 void moveUsesOfIAfterJ(BasicBlock &BB, 269 std::multimap<Value *, Value *> &LoadMoveSet, 270 Instruction *&InsertionPt, 271 Instruction *I, Instruction *J); 272 273 virtual bool runOnBasicBlock(BasicBlock &BB) { 274 AA = &getAnalysis<AliasAnalysis>(); 275 SE = &getAnalysis<ScalarEvolution>(); 276 TD = getAnalysisIfAvailable<TargetData>(); 277 278 bool changed = false; 279 // Iterate a sufficient number of times to merge types of size 1 bit, 280 // then 2 bits, then 4, etc. up to half of the target vector width of the 281 // target vector register. 282 for (unsigned v = 2, n = 1; v <= VectorBits && (!MaxIter || n <= MaxIter); 283 v *= 2, ++n) { 284 DEBUG(dbgs() << "BBV: fusing loop #" << n << 285 " for " << BB.getName() << " in " << 286 BB.getParent()->getName() << "...\n"); 287 if (vectorizePairs(BB)) 288 changed = true; 289 else 290 break; 291 } 292 293 DEBUG(dbgs() << "BBV: done!\n"); 294 return changed; 295 } 296 297 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 298 BasicBlockPass::getAnalysisUsage(AU); 299 AU.addRequired<AliasAnalysis>(); 300 AU.addRequired<ScalarEvolution>(); 301 AU.addPreserved<AliasAnalysis>(); 302 AU.addPreserved<ScalarEvolution>(); 303 AU.setPreservesCFG(); 304 } 305 306 // This returns the vector type that holds a pair of the provided type. 307 // If the provided type is already a vector, then its length is doubled. 308 static inline VectorType *getVecTypeForPair(Type *ElemTy) { 309 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 310 unsigned numElem = VTy->getNumElements(); 311 return VectorType::get(ElemTy->getScalarType(), numElem*2); 312 } 313 314 return VectorType::get(ElemTy, 2); 315 } 316 317 // Returns the weight associated with the provided value. A chain of 318 // candidate pairs has a length given by the sum of the weights of its 319 // members (one weight per pair; the weight of each member of the pair 320 // is assumed to be the same). This length is then compared to the 321 // chain-length threshold to determine if a given chain is significant 322 // enough to be vectorized. The length is also used in comparing 323 // candidate chains where longer chains are considered to be better. 324 // Note: when this function returns 0, the resulting instructions are 325 // not actually fused. 326 static inline size_t getDepthFactor(Value *V) { 327 // InsertElement and ExtractElement have a depth factor of zero. This is 328 // for two reasons: First, they cannot be usefully fused. Second, because 329 // the pass generates a lot of these, they can confuse the simple metric 330 // used to compare the trees in the next iteration. Thus, giving them a 331 // weight of zero allows the pass to essentially ignore them in 332 // subsequent iterations when looking for vectorization opportunities 333 // while still tracking dependency chains that flow through those 334 // instructions. 335 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 336 return 0; 337 338 return 1; 339 } 340 341 // This determines the relative offset of two loads or stores, returning 342 // true if the offset could be determined to be some constant value. 343 // For example, if OffsetInElmts == 1, then J accesses the memory directly 344 // after I; if OffsetInElmts == -1 then I accesses the memory 345 // directly after J. This function assumes that both instructions 346 // have the same type. 347 bool getPairPtrInfo(Instruction *I, Instruction *J, 348 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 349 int64_t &OffsetInElmts) { 350 OffsetInElmts = 0; 351 if (isa<LoadInst>(I)) { 352 IPtr = cast<LoadInst>(I)->getPointerOperand(); 353 JPtr = cast<LoadInst>(J)->getPointerOperand(); 354 IAlignment = cast<LoadInst>(I)->getAlignment(); 355 JAlignment = cast<LoadInst>(J)->getAlignment(); 356 } else { 357 IPtr = cast<StoreInst>(I)->getPointerOperand(); 358 JPtr = cast<StoreInst>(J)->getPointerOperand(); 359 IAlignment = cast<StoreInst>(I)->getAlignment(); 360 JAlignment = cast<StoreInst>(J)->getAlignment(); 361 } 362 363 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 364 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 365 366 // If this is a trivial offset, then we'll get something like 367 // 1*sizeof(type). With target data, which we need anyway, this will get 368 // constant folded into a number. 369 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 370 if (const SCEVConstant *ConstOffSCEV = 371 dyn_cast<SCEVConstant>(OffsetSCEV)) { 372 ConstantInt *IntOff = ConstOffSCEV->getValue(); 373 int64_t Offset = IntOff->getSExtValue(); 374 375 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 376 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 377 378 assert(VTy == cast<PointerType>(JPtr->getType())->getElementType()); 379 380 OffsetInElmts = Offset/VTyTSS; 381 return (abs64(Offset) % VTyTSS) == 0; 382 } 383 384 return false; 385 } 386 387 // Returns true if the provided CallInst represents an intrinsic that can 388 // be vectorized. 389 bool isVectorizableIntrinsic(CallInst* I) { 390 Function *F = I->getCalledFunction(); 391 if (!F) return false; 392 393 unsigned IID = F->getIntrinsicID(); 394 if (!IID) return false; 395 396 switch(IID) { 397 default: 398 return false; 399 case Intrinsic::sqrt: 400 case Intrinsic::powi: 401 case Intrinsic::sin: 402 case Intrinsic::cos: 403 case Intrinsic::log: 404 case Intrinsic::log2: 405 case Intrinsic::log10: 406 case Intrinsic::exp: 407 case Intrinsic::exp2: 408 case Intrinsic::pow: 409 return !NoMath; 410 case Intrinsic::fma: 411 return !NoFMA; 412 } 413 } 414 415 // Returns true if J is the second element in some pair referenced by 416 // some multimap pair iterator pair. 417 template <typename V> 418 bool isSecondInIteratorPair(V J, std::pair< 419 typename std::multimap<V, V>::iterator, 420 typename std::multimap<V, V>::iterator> PairRange) { 421 for (typename std::multimap<V, V>::iterator K = PairRange.first; 422 K != PairRange.second; ++K) 423 if (K->second == J) return true; 424 425 return false; 426 } 427 }; 428 429 // This function implements one vectorization iteration on the provided 430 // basic block. It returns true if the block is changed. 431 bool BBVectorize::vectorizePairs(BasicBlock &BB) { 432 std::vector<Value *> PairableInsts; 433 std::multimap<Value *, Value *> CandidatePairs; 434 getCandidatePairs(BB, CandidatePairs, PairableInsts); 435 if (PairableInsts.empty()) return false; 436 437 // Now we have a map of all of the pairable instructions and we need to 438 // select the best possible pairing. A good pairing is one such that the 439 // users of the pair are also paired. This defines a (directed) forest 440 // over the pairs such that two pairs are connected iff the second pair 441 // uses the first. 442 443 // Note that it only matters that both members of the second pair use some 444 // element of the first pair (to allow for splatting). 445 446 std::multimap<ValuePair, ValuePair> ConnectedPairs; 447 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); 448 if (ConnectedPairs.empty()) return false; 449 450 // Build the pairable-instruction dependency map 451 DenseSet<ValuePair> PairableInstUsers; 452 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 453 454 // There is now a graph of the connected pairs. For each variable, pick the 455 // pairing with the largest tree meeting the depth requirement on at least 456 // one branch. Then select all pairings that are part of that tree and 457 // remove them from the list of available pairings and pairable variables. 458 459 DenseMap<Value *, Value *> ChosenPairs; 460 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, 461 PairableInstUsers, ChosenPairs); 462 463 if (ChosenPairs.empty()) return false; 464 NumFusedOps += ChosenPairs.size(); 465 466 // A set of pairs has now been selected. It is now necessary to replace the 467 // paired instructions with vector instructions. For this procedure each 468 // operand much be replaced with a vector operand. This vector is formed 469 // by using build_vector on the old operands. The replaced values are then 470 // replaced with a vector_extract on the result. Subsequent optimization 471 // passes should coalesce the build/extract combinations. 472 473 fuseChosenPairs(BB, PairableInsts, ChosenPairs); 474 475 return true; 476 } 477 478 // This function returns true if the provided instruction is capable of being 479 // fused into a vector instruction. This determination is based only on the 480 // type and other attributes of the instruction. 481 bool BBVectorize::isInstVectorizable(Instruction *I, 482 bool &IsSimpleLoadStore) { 483 IsSimpleLoadStore = false; 484 485 if (CallInst *C = dyn_cast<CallInst>(I)) { 486 if (!isVectorizableIntrinsic(C)) 487 return false; 488 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 489 // Vectorize simple loads if possbile: 490 IsSimpleLoadStore = L->isSimple(); 491 if (!IsSimpleLoadStore || NoMemOps) 492 return false; 493 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 494 // Vectorize simple stores if possbile: 495 IsSimpleLoadStore = S->isSimple(); 496 if (!IsSimpleLoadStore || NoMemOps) 497 return false; 498 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 499 // We can vectorize casts, but not casts of pointer types, etc. 500 if (NoCasts) 501 return false; 502 503 Type *SrcTy = C->getSrcTy(); 504 if (!SrcTy->isSingleValueType() || SrcTy->isPointerTy()) 505 return false; 506 507 Type *DestTy = C->getDestTy(); 508 if (!DestTy->isSingleValueType() || DestTy->isPointerTy()) 509 return false; 510 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 511 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 512 return false; 513 } 514 515 // We can't vectorize memory operations without target data 516 if (TD == 0 && IsSimpleLoadStore) 517 return false; 518 519 Type *T1, *T2; 520 if (isa<StoreInst>(I)) { 521 // For stores, it is the value type, not the pointer type that matters 522 // because the value is what will come from a vector register. 523 524 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 525 T1 = IVal->getType(); 526 } else { 527 T1 = I->getType(); 528 } 529 530 if (I->isCast()) 531 T2 = cast<CastInst>(I)->getSrcTy(); 532 else 533 T2 = T1; 534 535 // Not every type can be vectorized... 536 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 537 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 538 return false; 539 540 if (NoInts && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) 541 return false; 542 543 if (NoFloats && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 544 return false; 545 546 if (T1->getPrimitiveSizeInBits() > VectorBits/2 || 547 T2->getPrimitiveSizeInBits() > VectorBits/2) 548 return false; 549 550 return true; 551 } 552 553 // This function returns true if the two provided instructions are compatible 554 // (meaning that they can be fused into a vector instruction). This assumes 555 // that I has already been determined to be vectorizable and that J is not 556 // in the use tree of I. 557 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 558 bool IsSimpleLoadStore) { 559 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 560 " <-> " << *J << "\n"); 561 562 // Loads and stores can be merged if they have different alignments, 563 // but are otherwise the same. 564 LoadInst *LI, *LJ; 565 StoreInst *SI, *SJ; 566 if ((LI = dyn_cast<LoadInst>(I)) && (LJ = dyn_cast<LoadInst>(J))) { 567 if (I->getType() != J->getType()) 568 return false; 569 570 if (LI->getPointerOperand()->getType() != 571 LJ->getPointerOperand()->getType() || 572 LI->isVolatile() != LJ->isVolatile() || 573 LI->getOrdering() != LJ->getOrdering() || 574 LI->getSynchScope() != LJ->getSynchScope()) 575 return false; 576 } else if ((SI = dyn_cast<StoreInst>(I)) && (SJ = dyn_cast<StoreInst>(J))) { 577 if (SI->getValueOperand()->getType() != 578 SJ->getValueOperand()->getType() || 579 SI->getPointerOperand()->getType() != 580 SJ->getPointerOperand()->getType() || 581 SI->isVolatile() != SJ->isVolatile() || 582 SI->getOrdering() != SJ->getOrdering() || 583 SI->getSynchScope() != SJ->getSynchScope()) 584 return false; 585 } else if (!J->isSameOperationAs(I)) { 586 return false; 587 } 588 // FIXME: handle addsub-type operations! 589 590 if (IsSimpleLoadStore) { 591 Value *IPtr, *JPtr; 592 unsigned IAlignment, JAlignment; 593 int64_t OffsetInElmts = 0; 594 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 595 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 596 if (AlignedOnly) { 597 Type *aType = isa<StoreInst>(I) ? 598 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 599 // An aligned load or store is possible only if the instruction 600 // with the lower offset has an alignment suitable for the 601 // vector type. 602 603 unsigned BottomAlignment = IAlignment; 604 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 605 606 Type *VType = getVecTypeForPair(aType); 607 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 608 if (BottomAlignment < VecAlignment) 609 return false; 610 } 611 } else { 612 return false; 613 } 614 } else if (isa<ShuffleVectorInst>(I)) { 615 // Only merge two shuffles if they're both constant 616 return isa<Constant>(I->getOperand(2)) && 617 isa<Constant>(J->getOperand(2)); 618 // FIXME: We may want to vectorize non-constant shuffles also. 619 } 620 621 return true; 622 } 623 624 // Figure out whether or not J uses I and update the users and write-set 625 // structures associated with I. Specifically, Users represents the set of 626 // instructions that depend on I. WriteSet represents the set 627 // of memory locations that are dependent on I. If UpdateUsers is true, 628 // and J uses I, then Users is updated to contain J and WriteSet is updated 629 // to contain any memory locations to which J writes. The function returns 630 // true if J uses I. By default, alias analysis is used to determine 631 // whether J reads from memory that overlaps with a location in WriteSet. 632 // If LoadMoveSet is not null, then it is a previously-computed multimap 633 // where the key is the memory-based user instruction and the value is 634 // the instruction to be compared with I. So, if LoadMoveSet is provided, 635 // then the alias analysis is not used. This is necessary because this 636 // function is called during the process of moving instructions during 637 // vectorization and the results of the alias analysis are not stable during 638 // that process. 639 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 640 AliasSetTracker &WriteSet, Instruction *I, 641 Instruction *J, bool UpdateUsers, 642 std::multimap<Value *, Value *> *LoadMoveSet) { 643 bool UsesI = false; 644 645 // This instruction may already be marked as a user due, for example, to 646 // being a member of a selected pair. 647 if (Users.count(J)) 648 UsesI = true; 649 650 if (!UsesI) 651 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 652 JU != JE; ++JU) { 653 Value *V = *JU; 654 if (I == V || Users.count(V)) { 655 UsesI = true; 656 break; 657 } 658 } 659 if (!UsesI && J->mayReadFromMemory()) { 660 if (LoadMoveSet) { 661 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 662 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 663 } else { 664 for (AliasSetTracker::iterator W = WriteSet.begin(), 665 WE = WriteSet.end(); W != WE; ++W) { 666 for (AliasSet::iterator A = W->begin(), AE = W->end(); 667 A != AE; ++A) { 668 AliasAnalysis::Location ptrLoc(A->getValue(), A->getSize(), 669 A->getTBAAInfo()); 670 if (AA->getModRefInfo(J, ptrLoc) != AliasAnalysis::NoModRef) { 671 UsesI = true; 672 break; 673 } 674 } 675 if (UsesI) break; 676 } 677 } 678 } 679 680 if (UsesI && UpdateUsers) { 681 if (J->mayWriteToMemory()) WriteSet.add(J); 682 Users.insert(J); 683 } 684 685 return UsesI; 686 } 687 688 // This function iterates over all instruction pairs in the provided 689 // basic block and collects all candidate pairs for vectorization. 690 void BBVectorize::getCandidatePairs(BasicBlock &BB, 691 std::multimap<Value *, Value *> &CandidatePairs, 692 std::vector<Value *> &PairableInsts) { 693 BasicBlock::iterator E = BB.end(); 694 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 695 bool IsSimpleLoadStore; 696 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 697 698 // Look for an instruction with which to pair instruction *I... 699 DenseSet<Value *> Users; 700 AliasSetTracker WriteSet(*AA); 701 BasicBlock::iterator J = I; ++J; 702 for (unsigned ss = 0; J != E && ss <= SearchLimit; ++J, ++ss) { 703 // Determine if J uses I, if so, exit the loop. 704 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !FastDep); 705 if (FastDep) { 706 // Note: For this heuristic to be effective, independent operations 707 // must tend to be intermixed. This is likely to be true from some 708 // kinds of grouped loop unrolling (but not the generic LLVM pass), 709 // but otherwise may require some kind of reordering pass. 710 711 // When using fast dependency analysis, 712 // stop searching after first use: 713 if (UsesI) break; 714 } else { 715 if (UsesI) continue; 716 } 717 718 // J does not use I, and comes before the first use of I, so it can be 719 // merged with I if the instructions are compatible. 720 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue; 721 722 // J is a candidate for merging with I. 723 if (!PairableInsts.size() || 724 PairableInsts[PairableInsts.size()-1] != I) { 725 PairableInsts.push_back(I); 726 } 727 CandidatePairs.insert(ValuePair(I, J)); 728 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 729 << *I << " <-> " << *J << "\n"); 730 } 731 } 732 733 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 734 << " instructions with candidate pairs\n"); 735 } 736 737 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 738 // it looks for pairs such that both members have an input which is an 739 // output of PI or PJ. 740 void BBVectorize::computePairsConnectedTo( 741 std::multimap<Value *, Value *> &CandidatePairs, 742 std::vector<Value *> &PairableInsts, 743 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 744 ValuePair P) { 745 // For each possible pairing for this variable, look at the uses of 746 // the first value... 747 for (Value::use_iterator I = P.first->use_begin(), 748 E = P.first->use_end(); I != E; ++I) { 749 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 750 751 // For each use of the first variable, look for uses of the second 752 // variable... 753 for (Value::use_iterator J = P.second->use_begin(), 754 E2 = P.second->use_end(); J != E2; ++J) { 755 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 756 757 // Look for <I, J>: 758 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 759 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 760 761 // Look for <J, I>: 762 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) 763 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); 764 } 765 766 if (SplatBreaksChain) continue; 767 // Look for cases where just the first value in the pair is used by 768 // both members of another pair (splatting). 769 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 770 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 771 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 772 } 773 } 774 775 if (SplatBreaksChain) return; 776 // Look for cases where just the second value in the pair is used by 777 // both members of another pair (splatting). 778 for (Value::use_iterator I = P.second->use_begin(), 779 E = P.second->use_end(); I != E; ++I) { 780 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 781 782 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 783 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) 784 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); 785 } 786 } 787 } 788 789 // This function figures out which pairs are connected. Two pairs are 790 // connected if some output of the first pair forms an input to both members 791 // of the second pair. 792 void BBVectorize::computeConnectedPairs( 793 std::multimap<Value *, Value *> &CandidatePairs, 794 std::vector<Value *> &PairableInsts, 795 std::multimap<ValuePair, ValuePair> &ConnectedPairs) { 796 797 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 798 PE = PairableInsts.end(); PI != PE; ++PI) { 799 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 800 801 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 802 P != choiceRange.second; ++P) 803 computePairsConnectedTo(CandidatePairs, PairableInsts, 804 ConnectedPairs, *P); 805 } 806 807 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 808 << " pair connections.\n"); 809 } 810 811 // This function builds a set of use tuples such that <A, B> is in the set 812 // if B is in the use tree of A. If B is in the use tree of A, then B 813 // depends on the output of A. 814 void BBVectorize::buildDepMap( 815 BasicBlock &BB, 816 std::multimap<Value *, Value *> &CandidatePairs, 817 std::vector<Value *> &PairableInsts, 818 DenseSet<ValuePair> &PairableInstUsers) { 819 DenseSet<Value *> IsInPair; 820 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 821 E = CandidatePairs.end(); C != E; ++C) { 822 IsInPair.insert(C->first); 823 IsInPair.insert(C->second); 824 } 825 826 // Iterate through the basic block, recording all Users of each 827 // pairable instruction. 828 829 BasicBlock::iterator E = BB.end(); 830 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 831 if (IsInPair.find(I) == IsInPair.end()) continue; 832 833 DenseSet<Value *> Users; 834 AliasSetTracker WriteSet(*AA); 835 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 836 (void) trackUsesOfI(Users, WriteSet, I, J); 837 838 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 839 U != E; ++U) 840 PairableInstUsers.insert(ValuePair(I, *U)); 841 } 842 } 843 844 // Returns true if an input to pair P is an output of pair Q and also an 845 // input of pair Q is an output of pair P. If this is the case, then these 846 // two pairs cannot be simultaneously fused. 847 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 848 DenseSet<ValuePair> &PairableInstUsers, 849 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 850 // Two pairs are in conflict if they are mutual Users of eachother. 851 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 852 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 853 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 854 PairableInstUsers.count(ValuePair(P.second, Q.second)); 855 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 856 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 857 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 858 PairableInstUsers.count(ValuePair(Q.second, P.second)); 859 if (PairableInstUserMap) { 860 // FIXME: The expensive part of the cycle check is not so much the cycle 861 // check itself but this edge insertion procedure. This needs some 862 // profiling and probably a different data structure (same is true of 863 // most uses of std::multimap). 864 if (PUsesQ) { 865 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 866 if (!isSecondInIteratorPair(P, QPairRange)) 867 PairableInstUserMap->insert(VPPair(Q, P)); 868 } 869 if (QUsesP) { 870 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 871 if (!isSecondInIteratorPair(Q, PPairRange)) 872 PairableInstUserMap->insert(VPPair(P, Q)); 873 } 874 } 875 876 return (QUsesP && PUsesQ); 877 } 878 879 // This function walks the use graph of current pairs to see if, starting 880 // from P, the walk returns to P. 881 bool BBVectorize::pairWillFormCycle(ValuePair P, 882 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 883 DenseSet<ValuePair> &CurrentPairs) { 884 DEBUG(if (DebugCycleCheck) 885 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 886 << *P.second << "\n"); 887 // A lookup table of visisted pairs is kept because the PairableInstUserMap 888 // contains non-direct associations. 889 DenseSet<ValuePair> Visited; 890 std::vector<ValuePair> Q; 891 // General depth-first post-order traversal: 892 Q.push_back(P); 893 while (!Q.empty()) { 894 ValuePair QTop = Q.back(); 895 896 Visited.insert(QTop); 897 Q.pop_back(); 898 899 DEBUG(if (DebugCycleCheck) 900 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 901 << *QTop.second << "\n"); 902 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 903 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 904 C != QPairRange.second; ++C) { 905 if (C->second == P) { 906 DEBUG(dbgs() 907 << "BBV: rejected to prevent non-trivial cycle formation: " 908 << *C->first.first << " <-> " << *C->first.second << "\n"); 909 return true; 910 } 911 912 if (CurrentPairs.count(C->second) > 0 && 913 Visited.count(C->second) == 0) 914 Q.push_back(C->second); 915 } 916 } 917 918 return false; 919 } 920 921 // This function builds the initial tree of connected pairs with the 922 // pair J at the root. 923 void BBVectorize::buildInitialTreeFor( 924 std::multimap<Value *, Value *> &CandidatePairs, 925 std::vector<Value *> &PairableInsts, 926 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 927 DenseSet<ValuePair> &PairableInstUsers, 928 DenseMap<Value *, Value *> &ChosenPairs, 929 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 930 // Each of these pairs is viewed as the root node of a Tree. The Tree 931 // is then walked (depth-first). As this happens, we keep track of 932 // the pairs that compose the Tree and the maximum depth of the Tree. 933 std::vector<ValuePairWithDepth> Q; 934 // General depth-first post-order traversal: 935 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 936 while (!Q.empty()) { 937 ValuePairWithDepth QTop = Q.back(); 938 939 // Push each child onto the queue: 940 bool MoreChildren = false; 941 size_t MaxChildDepth = QTop.second; 942 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 943 for (std::map<ValuePair, ValuePair>::iterator k = qtRange.first; 944 k != qtRange.second; ++k) { 945 // Make sure that this child pair is still a candidate: 946 bool IsStillCand = false; 947 VPIteratorPair checkRange = 948 CandidatePairs.equal_range(k->second.first); 949 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 950 m != checkRange.second; ++m) { 951 if (m->second == k->second.second) { 952 IsStillCand = true; 953 break; 954 } 955 } 956 957 if (IsStillCand) { 958 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 959 if (C == Tree.end()) { 960 size_t d = getDepthFactor(k->second.first); 961 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 962 MoreChildren = true; 963 } else { 964 MaxChildDepth = std::max(MaxChildDepth, C->second); 965 } 966 } 967 } 968 969 if (!MoreChildren) { 970 // Record the current pair as part of the Tree: 971 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 972 Q.pop_back(); 973 } 974 } 975 } 976 977 // Given some initial tree, prune it by removing conflicting pairs (pairs 978 // that cannot be simultaneously chosen for vectorization). 979 void BBVectorize::pruneTreeFor( 980 std::multimap<Value *, Value *> &CandidatePairs, 981 std::vector<Value *> &PairableInsts, 982 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 983 DenseSet<ValuePair> &PairableInstUsers, 984 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 985 DenseMap<Value *, Value *> &ChosenPairs, 986 DenseMap<ValuePair, size_t> &Tree, 987 DenseSet<ValuePair> &PrunedTree, ValuePair J, 988 bool UseCycleCheck) { 989 std::vector<ValuePairWithDepth> Q; 990 // General depth-first post-order traversal: 991 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 992 while (!Q.empty()) { 993 ValuePairWithDepth QTop = Q.back(); 994 PrunedTree.insert(QTop.first); 995 Q.pop_back(); 996 997 // Visit each child, pruning as necessary... 998 DenseMap<ValuePair, size_t> BestChilden; 999 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1000 for (std::map<ValuePair, ValuePair>::iterator K = QTopRange.first; 1001 K != QTopRange.second; ++K) { 1002 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1003 if (C == Tree.end()) continue; 1004 1005 // This child is in the Tree, now we need to make sure it is the 1006 // best of any conflicting children. There could be multiple 1007 // conflicting children, so first, determine if we're keeping 1008 // this child, then delete conflicting children as necessary. 1009 1010 // It is also necessary to guard against pairing-induced 1011 // dependencies. Consider instructions a .. x .. y .. b 1012 // such that (a,b) are to be fused and (x,y) are to be fused 1013 // but a is an input to x and b is an output from y. This 1014 // means that y cannot be moved after b but x must be moved 1015 // after b for (a,b) to be fused. In other words, after 1016 // fusing (a,b) we have y .. a/b .. x where y is an input 1017 // to a/b and x is an output to a/b: x and y can no longer 1018 // be legally fused. To prevent this condition, we must 1019 // make sure that a child pair added to the Tree is not 1020 // both an input and output of an already-selected pair. 1021 1022 // Pairing-induced dependencies can also form from more complicated 1023 // cycles. The pair vs. pair conflicts are easy to check, and so 1024 // that is done explicitly for "fast rejection", and because for 1025 // child vs. child conflicts, we may prefer to keep the current 1026 // pair in preference to the already-selected child. 1027 DenseSet<ValuePair> CurrentPairs; 1028 1029 bool CanAdd = true; 1030 for (DenseMap<ValuePair, size_t>::iterator C2 1031 = BestChilden.begin(), E2 = BestChilden.end(); 1032 C2 != E2; ++C2) { 1033 if (C2->first.first == C->first.first || 1034 C2->first.first == C->first.second || 1035 C2->first.second == C->first.first || 1036 C2->first.second == C->first.second || 1037 pairsConflict(C2->first, C->first, PairableInstUsers, 1038 UseCycleCheck ? &PairableInstUserMap : 0)) { 1039 if (C2->second >= C->second) { 1040 CanAdd = false; 1041 break; 1042 } 1043 1044 CurrentPairs.insert(C2->first); 1045 } 1046 } 1047 if (!CanAdd) continue; 1048 1049 // Even worse, this child could conflict with another node already 1050 // selected for the Tree. If that is the case, ignore this child. 1051 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1052 E2 = PrunedTree.end(); T != E2; ++T) { 1053 if (T->first == C->first.first || 1054 T->first == C->first.second || 1055 T->second == C->first.first || 1056 T->second == C->first.second || 1057 pairsConflict(*T, C->first, PairableInstUsers, 1058 UseCycleCheck ? &PairableInstUserMap : 0)) { 1059 CanAdd = false; 1060 break; 1061 } 1062 1063 CurrentPairs.insert(*T); 1064 } 1065 if (!CanAdd) continue; 1066 1067 // And check the queue too... 1068 for (std::vector<ValuePairWithDepth>::iterator C2 = Q.begin(), 1069 E2 = Q.end(); C2 != E2; ++C2) { 1070 if (C2->first.first == C->first.first || 1071 C2->first.first == C->first.second || 1072 C2->first.second == C->first.first || 1073 C2->first.second == C->first.second || 1074 pairsConflict(C2->first, C->first, PairableInstUsers, 1075 UseCycleCheck ? &PairableInstUserMap : 0)) { 1076 CanAdd = false; 1077 break; 1078 } 1079 1080 CurrentPairs.insert(C2->first); 1081 } 1082 if (!CanAdd) continue; 1083 1084 // Last but not least, check for a conflict with any of the 1085 // already-chosen pairs. 1086 for (DenseMap<Value *, Value *>::iterator C2 = 1087 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1088 C2 != E2; ++C2) { 1089 if (pairsConflict(*C2, C->first, PairableInstUsers, 1090 UseCycleCheck ? &PairableInstUserMap : 0)) { 1091 CanAdd = false; 1092 break; 1093 } 1094 1095 CurrentPairs.insert(*C2); 1096 } 1097 if (!CanAdd) continue; 1098 1099 // To check for non-trivial cycles formed by the addition of the 1100 // current pair we've formed a list of all relevant pairs, now use a 1101 // graph walk to check for a cycle. We start from the current pair and 1102 // walk the use tree to see if we again reach the current pair. If we 1103 // do, then the current pair is rejected. 1104 1105 // FIXME: It may be more efficient to use a topological-ordering 1106 // algorithm to improve the cycle check. This should be investigated. 1107 if (UseCycleCheck && 1108 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1109 continue; 1110 1111 // This child can be added, but we may have chosen it in preference 1112 // to an already-selected child. Check for this here, and if a 1113 // conflict is found, then remove the previously-selected child 1114 // before adding this one in its place. 1115 for (DenseMap<ValuePair, size_t>::iterator C2 1116 = BestChilden.begin(); C2 != BestChilden.end();) { 1117 if (C2->first.first == C->first.first || 1118 C2->first.first == C->first.second || 1119 C2->first.second == C->first.first || 1120 C2->first.second == C->first.second || 1121 pairsConflict(C2->first, C->first, PairableInstUsers)) 1122 BestChilden.erase(C2++); 1123 else 1124 ++C2; 1125 } 1126 1127 BestChilden.insert(ValuePairWithDepth(C->first, C->second)); 1128 } 1129 1130 for (DenseMap<ValuePair, size_t>::iterator C 1131 = BestChilden.begin(), E2 = BestChilden.end(); 1132 C != E2; ++C) { 1133 size_t DepthF = getDepthFactor(C->first.first); 1134 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1135 } 1136 } 1137 } 1138 1139 // This function finds the best tree of mututally-compatible connected 1140 // pairs, given the choice of root pairs as an iterator range. 1141 void BBVectorize::findBestTreeFor( 1142 std::multimap<Value *, Value *> &CandidatePairs, 1143 std::vector<Value *> &PairableInsts, 1144 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1145 DenseSet<ValuePair> &PairableInstUsers, 1146 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1147 DenseMap<Value *, Value *> &ChosenPairs, 1148 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1149 size_t &BestEffSize, VPIteratorPair ChoiceRange, 1150 bool UseCycleCheck) { 1151 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1152 J != ChoiceRange.second; ++J) { 1153 1154 // Before going any further, make sure that this pair does not 1155 // conflict with any already-selected pairs (see comment below 1156 // near the Tree pruning for more details). 1157 DenseSet<ValuePair> ChosenPairSet; 1158 bool DoesConflict = false; 1159 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1160 E = ChosenPairs.end(); C != E; ++C) { 1161 if (pairsConflict(*C, *J, PairableInstUsers, 1162 UseCycleCheck ? &PairableInstUserMap : 0)) { 1163 DoesConflict = true; 1164 break; 1165 } 1166 1167 ChosenPairSet.insert(*C); 1168 } 1169 if (DoesConflict) continue; 1170 1171 if (UseCycleCheck && 1172 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1173 continue; 1174 1175 DenseMap<ValuePair, size_t> Tree; 1176 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1177 PairableInstUsers, ChosenPairs, Tree, *J); 1178 1179 // Because we'll keep the child with the largest depth, the largest 1180 // depth is still the same in the unpruned Tree. 1181 size_t MaxDepth = Tree.lookup(*J); 1182 1183 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1184 << *J->first << " <-> " << *J->second << "} of depth " << 1185 MaxDepth << " and size " << Tree.size() << "\n"); 1186 1187 // At this point the Tree has been constructed, but, may contain 1188 // contradictory children (meaning that different children of 1189 // some tree node may be attempting to fuse the same instruction). 1190 // So now we walk the tree again, in the case of a conflict, 1191 // keep only the child with the largest depth. To break a tie, 1192 // favor the first child. 1193 1194 DenseSet<ValuePair> PrunedTree; 1195 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1196 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1197 PrunedTree, *J, UseCycleCheck); 1198 1199 size_t EffSize = 0; 1200 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1201 E = PrunedTree.end(); S != E; ++S) 1202 EffSize += getDepthFactor(S->first); 1203 1204 DEBUG(if (DebugPairSelection) 1205 dbgs() << "BBV: found pruned Tree for pair {" 1206 << *J->first << " <-> " << *J->second << "} of depth " << 1207 MaxDepth << " and size " << PrunedTree.size() << 1208 " (effective size: " << EffSize << ")\n"); 1209 if (MaxDepth >= ReqChainDepth && EffSize > BestEffSize) { 1210 BestMaxDepth = MaxDepth; 1211 BestEffSize = EffSize; 1212 BestTree = PrunedTree; 1213 } 1214 } 1215 } 1216 1217 // Given the list of candidate pairs, this function selects those 1218 // that will be fused into vector instructions. 1219 void BBVectorize::choosePairs( 1220 std::multimap<Value *, Value *> &CandidatePairs, 1221 std::vector<Value *> &PairableInsts, 1222 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1223 DenseSet<ValuePair> &PairableInstUsers, 1224 DenseMap<Value *, Value *>& ChosenPairs) { 1225 bool UseCycleCheck = CandidatePairs.size() <= MaxCandPairsForCycleCheck; 1226 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1227 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1228 E = PairableInsts.end(); I != E; ++I) { 1229 // The number of possible pairings for this variable: 1230 size_t NumChoices = CandidatePairs.count(*I); 1231 if (!NumChoices) continue; 1232 1233 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1234 1235 // The best pair to choose and its tree: 1236 size_t BestMaxDepth = 0, BestEffSize = 0; 1237 DenseSet<ValuePair> BestTree; 1238 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1239 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1240 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1241 UseCycleCheck); 1242 1243 // A tree has been chosen (or not) at this point. If no tree was 1244 // chosen, then this instruction, I, cannot be paired (and is no longer 1245 // considered). 1246 1247 DEBUG(if (BestTree.size() > 0) 1248 dbgs() << "BBV: selected pairs in the best tree for: " 1249 << *cast<Instruction>(*I) << "\n"); 1250 1251 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1252 SE2 = BestTree.end(); S != SE2; ++S) { 1253 // Insert the members of this tree into the list of chosen pairs. 1254 ChosenPairs.insert(ValuePair(S->first, S->second)); 1255 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1256 *S->second << "\n"); 1257 1258 // Remove all candidate pairs that have values in the chosen tree. 1259 for (std::multimap<Value *, Value *>::iterator K = 1260 CandidatePairs.begin(); K != CandidatePairs.end();) { 1261 if (K->first == S->first || K->second == S->first || 1262 K->second == S->second || K->first == S->second) { 1263 // Don't remove the actual pair chosen so that it can be used 1264 // in subsequent tree selections. 1265 if (!(K->first == S->first && K->second == S->second)) 1266 CandidatePairs.erase(K++); 1267 else 1268 ++K; 1269 } else { 1270 ++K; 1271 } 1272 } 1273 } 1274 } 1275 1276 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1277 } 1278 1279 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1280 unsigned n = 0) { 1281 if (!I->hasName()) 1282 return ""; 1283 1284 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1285 (n > 0 ? "." + utostr(n) : "")).str(); 1286 } 1287 1288 // Returns the value that is to be used as the pointer input to the vector 1289 // instruction that fuses I with J. 1290 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1291 Instruction *I, Instruction *J, unsigned o, 1292 bool &FlipMemInputs) { 1293 Value *IPtr, *JPtr; 1294 unsigned IAlignment, JAlignment; 1295 int64_t OffsetInElmts; 1296 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1297 OffsetInElmts); 1298 1299 // The pointer value is taken to be the one with the lowest offset. 1300 Value *VPtr; 1301 if (OffsetInElmts > 0) { 1302 VPtr = IPtr; 1303 } else { 1304 FlipMemInputs = true; 1305 VPtr = JPtr; 1306 } 1307 1308 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType(); 1309 Type *VArgType = getVecTypeForPair(ArgType); 1310 Type *VArgPtrType = PointerType::get(VArgType, 1311 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1312 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1313 /* insert before */ FlipMemInputs ? J : I); 1314 } 1315 1316 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1317 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem, 1318 unsigned IdxOffset, std::vector<Constant*> &Mask) { 1319 for (unsigned v = 0; v < NumElem/2; ++v) { 1320 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1321 if (m < 0) { 1322 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1323 } else { 1324 unsigned mm = m + (int) IdxOffset; 1325 if (m >= (int) NumInElem) 1326 mm += (int) NumInElem; 1327 1328 Mask[v+MaskOffset] = 1329 ConstantInt::get(Type::getInt32Ty(Context), mm); 1330 } 1331 } 1332 } 1333 1334 // Returns the value that is to be used as the vector-shuffle mask to the 1335 // vector instruction that fuses I with J. 1336 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1337 Instruction *I, Instruction *J) { 1338 // This is the shuffle mask. We need to append the second 1339 // mask to the first, and the numbers need to be adjusted. 1340 1341 Type *ArgType = I->getType(); 1342 Type *VArgType = getVecTypeForPair(ArgType); 1343 1344 // Get the total number of elements in the fused vector type. 1345 // By definition, this must equal the number of elements in 1346 // the final mask. 1347 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1348 std::vector<Constant*> Mask(NumElem); 1349 1350 Type *OpType = I->getOperand(0)->getType(); 1351 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements(); 1352 1353 // For the mask from the first pair... 1354 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask); 1355 1356 // For the mask from the second pair... 1357 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem, 1358 Mask); 1359 1360 return ConstantVector::get(Mask); 1361 } 1362 1363 // Returns the value to be used as the specified operand of the vector 1364 // instruction that fuses I with J. 1365 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1366 Instruction *J, unsigned o, bool FlipMemInputs) { 1367 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1368 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1369 1370 // Compute the fused vector type for this operand 1371 Type *ArgType = I->getOperand(o)->getType(); 1372 VectorType *VArgType = getVecTypeForPair(ArgType); 1373 1374 Instruction *L = I, *H = J; 1375 if (FlipMemInputs) { 1376 L = J; 1377 H = I; 1378 } 1379 1380 if (ArgType->isVectorTy()) { 1381 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 1382 std::vector<Constant*> Mask(numElem); 1383 for (unsigned v = 0; v < numElem; ++v) 1384 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1385 1386 Instruction *BV = new ShuffleVectorInst(L->getOperand(o), 1387 H->getOperand(o), 1388 ConstantVector::get(Mask), 1389 getReplacementName(I, true, o)); 1390 BV->insertBefore(J); 1391 return BV; 1392 } 1393 1394 // If these two inputs are the output of another vector instruction, 1395 // then we should use that output directly. It might be necessary to 1396 // permute it first. [When pairings are fused recursively, you can 1397 // end up with cases where a large vector is decomposed into scalars 1398 // using extractelement instructions, then built into size-2 1399 // vectors using insertelement and the into larger vectors using 1400 // shuffles. InstCombine does not simplify all of these cases well, 1401 // and so we make sure that shuffles are generated here when possible. 1402 ExtractElementInst *LEE 1403 = dyn_cast<ExtractElementInst>(L->getOperand(o)); 1404 ExtractElementInst *HEE 1405 = dyn_cast<ExtractElementInst>(H->getOperand(o)); 1406 1407 if (LEE && HEE && 1408 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) { 1409 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType()); 1410 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue(); 1411 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue(); 1412 if (LEE->getOperand(0) == HEE->getOperand(0)) { 1413 if (LowIndx == 0 && HighIndx == 1) 1414 return LEE->getOperand(0); 1415 1416 std::vector<Constant*> Mask(2); 1417 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1418 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1419 1420 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1421 UndefValue::get(EEType), 1422 ConstantVector::get(Mask), 1423 getReplacementName(I, true, o)); 1424 BV->insertBefore(J); 1425 return BV; 1426 } 1427 1428 std::vector<Constant*> Mask(2); 1429 HighIndx += EEType->getNumElements(); 1430 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx); 1431 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx); 1432 1433 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0), 1434 HEE->getOperand(0), 1435 ConstantVector::get(Mask), 1436 getReplacementName(I, true, o)); 1437 BV->insertBefore(J); 1438 return BV; 1439 } 1440 1441 Instruction *BV1 = InsertElementInst::Create( 1442 UndefValue::get(VArgType), 1443 L->getOperand(o), CV0, 1444 getReplacementName(I, true, o, 1)); 1445 BV1->insertBefore(I); 1446 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o), 1447 CV1, 1448 getReplacementName(I, true, o, 2)); 1449 BV2->insertBefore(J); 1450 return BV2; 1451 } 1452 1453 // This function creates an array of values that will be used as the inputs 1454 // to the vector instruction that fuses I with J. 1455 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 1456 Instruction *I, Instruction *J, 1457 SmallVector<Value *, 3> &ReplacedOperands, 1458 bool &FlipMemInputs) { 1459 FlipMemInputs = false; 1460 unsigned NumOperands = I->getNumOperands(); 1461 1462 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 1463 // Iterate backward so that we look at the store pointer 1464 // first and know whether or not we need to flip the inputs. 1465 1466 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 1467 // This is the pointer for a load/store instruction. 1468 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, 1469 FlipMemInputs); 1470 continue; 1471 } else if (isa<CallInst>(I) && o == NumOperands-1) { 1472 Function *F = cast<CallInst>(I)->getCalledFunction(); 1473 unsigned IID = F->getIntrinsicID(); 1474 BasicBlock &BB = *I->getParent(); 1475 1476 Module *M = BB.getParent()->getParent(); 1477 Type *ArgType = I->getType(); 1478 Type *VArgType = getVecTypeForPair(ArgType); 1479 1480 // FIXME: is it safe to do this here? 1481 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 1482 (Intrinsic::ID) IID, VArgType); 1483 continue; 1484 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 1485 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 1486 continue; 1487 } 1488 1489 ReplacedOperands[o] = 1490 getReplacementInput(Context, I, J, o, FlipMemInputs); 1491 } 1492 } 1493 1494 // This function creates two values that represent the outputs of the 1495 // original I and J instructions. These are generally vector shuffles 1496 // or extracts. In many cases, these will end up being unused and, thus, 1497 // eliminated by later passes. 1498 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 1499 Instruction *J, Instruction *K, 1500 Instruction *&InsertionPt, 1501 Instruction *&K1, Instruction *&K2, 1502 bool &FlipMemInputs) { 1503 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1504 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1505 1506 if (isa<StoreInst>(I)) { 1507 AA->replaceWithNewValue(I, K); 1508 AA->replaceWithNewValue(J, K); 1509 } else { 1510 Type *IType = I->getType(); 1511 Type *VType = getVecTypeForPair(IType); 1512 1513 if (IType->isVectorTy()) { 1514 unsigned numElem = cast<VectorType>(IType)->getNumElements(); 1515 std::vector<Constant*> Mask1(numElem), Mask2(numElem); 1516 for (unsigned v = 0; v < numElem; ++v) { 1517 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 1518 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v); 1519 } 1520 1521 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 1522 ConstantVector::get( 1523 FlipMemInputs ? Mask2 : Mask1), 1524 getReplacementName(K, false, 1)); 1525 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 1526 ConstantVector::get( 1527 FlipMemInputs ? Mask1 : Mask2), 1528 getReplacementName(K, false, 2)); 1529 } else { 1530 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, 1531 getReplacementName(K, false, 1)); 1532 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, 1533 getReplacementName(K, false, 2)); 1534 } 1535 1536 K1->insertAfter(K); 1537 K2->insertAfter(K1); 1538 InsertionPt = K2; 1539 } 1540 } 1541 1542 // Move all uses of the function I (including pairing-induced uses) after J. 1543 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 1544 std::multimap<Value *, Value *> &LoadMoveSet, 1545 Instruction *I, Instruction *J) { 1546 // Skip to the first instruction past I. 1547 BasicBlock::iterator L = BB.begin(); 1548 for (; cast<Instruction>(L) != I; ++L); 1549 ++L; 1550 1551 DenseSet<Value *> Users; 1552 AliasSetTracker WriteSet(*AA); 1553 for (; cast<Instruction>(L) != J; ++L) 1554 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 1555 1556 assert(cast<Instruction>(L) == J && 1557 "Tracking has not proceeded far enough to check for dependencies"); 1558 // If J is now in the use set of I, then trackUsesOfI will return true 1559 // and we have a dependency cycle (and the fusing operation must abort). 1560 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 1561 } 1562 1563 // Move all uses of the function I (including pairing-induced uses) after J. 1564 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 1565 std::multimap<Value *, Value *> &LoadMoveSet, 1566 Instruction *&InsertionPt, 1567 Instruction *I, Instruction *J) { 1568 // Skip to the first instruction past I. 1569 BasicBlock::iterator L = BB.begin(); 1570 for (; cast<Instruction>(L) != I; ++L); 1571 ++L; 1572 1573 DenseSet<Value *> Users; 1574 AliasSetTracker WriteSet(*AA); 1575 for (; cast<Instruction>(L) != J;) { 1576 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 1577 // Move this instruction 1578 Instruction *InstToMove = L; ++L; 1579 1580 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 1581 " to after " << *InsertionPt << "\n"); 1582 InstToMove->removeFromParent(); 1583 InstToMove->insertAfter(InsertionPt); 1584 InsertionPt = InstToMove; 1585 } else { 1586 ++L; 1587 } 1588 } 1589 } 1590 1591 // Collect all load instruction that are in the move set of a given first 1592 // pair member. These loads depend on the first instruction, I, and so need 1593 // to be moved after J (the second instruction) when the pair is fused. 1594 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 1595 DenseMap<Value *, Value *> &ChosenPairs, 1596 std::multimap<Value *, Value *> &LoadMoveSet, 1597 Instruction *I) { 1598 // Skip to the first instruction past I. 1599 BasicBlock::iterator L = BB.begin(); 1600 for (; cast<Instruction>(L) != I; ++L); 1601 ++L; 1602 1603 DenseSet<Value *> Users; 1604 AliasSetTracker WriteSet(*AA); 1605 1606 // Note: We cannot end the loop when we reach J because J could be moved 1607 // farther down the use chain by another instruction pairing. Also, J 1608 // could be before I if this is an inverted input. 1609 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 1610 if (trackUsesOfI(Users, WriteSet, I, L)) { 1611 if (L->mayReadFromMemory()) 1612 LoadMoveSet.insert(ValuePair(L, I)); 1613 } 1614 } 1615 } 1616 1617 // In cases where both load/stores and the computation of their pointers 1618 // are chosen for vectorization, we can end up in a situation where the 1619 // aliasing analysis starts returning different query results as the 1620 // process of fusing instruction pairs continues. Because the algorithm 1621 // relies on finding the same use trees here as were found earlier, we'll 1622 // need to precompute the necessary aliasing information here and then 1623 // manually update it during the fusion process. 1624 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 1625 std::vector<Value *> &PairableInsts, 1626 DenseMap<Value *, Value *> &ChosenPairs, 1627 std::multimap<Value *, Value *> &LoadMoveSet) { 1628 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1629 PIE = PairableInsts.end(); PI != PIE; ++PI) { 1630 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 1631 if (P == ChosenPairs.end()) continue; 1632 1633 Instruction *I = cast<Instruction>(P->first); 1634 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 1635 } 1636 } 1637 1638 // This function fuses the chosen instruction pairs into vector instructions, 1639 // taking care preserve any needed scalar outputs and, then, it reorders the 1640 // remaining instructions as needed (users of the first member of the pair 1641 // need to be moved to after the location of the second member of the pair 1642 // because the vector instruction is inserted in the location of the pair's 1643 // second member). 1644 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 1645 std::vector<Value *> &PairableInsts, 1646 DenseMap<Value *, Value *> &ChosenPairs) { 1647 LLVMContext& Context = BB.getContext(); 1648 1649 // During the vectorization process, the order of the pairs to be fused 1650 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 1651 // list. After a pair is fused, the flipped pair is removed from the list. 1652 std::vector<ValuePair> FlippedPairs; 1653 FlippedPairs.reserve(ChosenPairs.size()); 1654 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 1655 E = ChosenPairs.end(); P != E; ++P) 1656 FlippedPairs.push_back(ValuePair(P->second, P->first)); 1657 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), 1658 E = FlippedPairs.end(); P != E; ++P) 1659 ChosenPairs.insert(*P); 1660 1661 std::multimap<Value *, Value *> LoadMoveSet; 1662 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 1663 1664 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 1665 1666 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 1667 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 1668 if (P == ChosenPairs.end()) { 1669 ++PI; 1670 continue; 1671 } 1672 1673 if (getDepthFactor(P->first) == 0) { 1674 // These instructions are not really fused, but are tracked as though 1675 // they are. Any case in which it would be interesting to fuse them 1676 // will be taken care of by InstCombine. 1677 --NumFusedOps; 1678 ++PI; 1679 continue; 1680 } 1681 1682 Instruction *I = cast<Instruction>(P->first), 1683 *J = cast<Instruction>(P->second); 1684 1685 DEBUG(dbgs() << "BBV: fusing: " << *I << 1686 " <-> " << *J << "\n"); 1687 1688 // Remove the pair and flipped pair from the list. 1689 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 1690 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 1691 ChosenPairs.erase(FP); 1692 ChosenPairs.erase(P); 1693 1694 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 1695 DEBUG(dbgs() << "BBV: fusion of: " << *I << 1696 " <-> " << *J << 1697 " aborted because of non-trivial dependency cycle\n"); 1698 --NumFusedOps; 1699 ++PI; 1700 continue; 1701 } 1702 1703 bool FlipMemInputs; 1704 unsigned NumOperands = I->getNumOperands(); 1705 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 1706 getReplacementInputsForPair(Context, I, J, ReplacedOperands, 1707 FlipMemInputs); 1708 1709 // Make a copy of the original operation, change its type to the vector 1710 // type and replace its operands with the vector operands. 1711 Instruction *K = I->clone(); 1712 if (I->hasName()) K->takeName(I); 1713 1714 if (!isa<StoreInst>(K)) 1715 K->mutateType(getVecTypeForPair(I->getType())); 1716 1717 for (unsigned o = 0; o < NumOperands; ++o) 1718 K->setOperand(o, ReplacedOperands[o]); 1719 1720 // If we've flipped the memory inputs, make sure that we take the correct 1721 // alignment. 1722 if (FlipMemInputs) { 1723 if (isa<StoreInst>(K)) 1724 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); 1725 else 1726 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); 1727 } 1728 1729 K->insertAfter(J); 1730 1731 // Instruction insertion point: 1732 Instruction *InsertionPt = K; 1733 Instruction *K1 = 0, *K2 = 0; 1734 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, 1735 FlipMemInputs); 1736 1737 // The use tree of the first original instruction must be moved to after 1738 // the location of the second instruction. The entire use tree of the 1739 // first instruction is disjoint from the input tree of the second 1740 // (by definition), and so commutes with it. 1741 1742 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 1743 1744 if (!isa<StoreInst>(I)) { 1745 I->replaceAllUsesWith(K1); 1746 J->replaceAllUsesWith(K2); 1747 AA->replaceWithNewValue(I, K1); 1748 AA->replaceWithNewValue(J, K2); 1749 } 1750 1751 // Instructions that may read from memory may be in the load move set. 1752 // Once an instruction is fused, we no longer need its move set, and so 1753 // the values of the map never need to be updated. However, when a load 1754 // is fused, we need to merge the entries from both instructions in the 1755 // pair in case those instructions were in the move set of some other 1756 // yet-to-be-fused pair. The loads in question are the keys of the map. 1757 if (I->mayReadFromMemory()) { 1758 std::vector<ValuePair> NewSetMembers; 1759 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 1760 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 1761 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 1762 N != IPairRange.second; ++N) 1763 NewSetMembers.push_back(ValuePair(K, N->second)); 1764 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 1765 N != JPairRange.second; ++N) 1766 NewSetMembers.push_back(ValuePair(K, N->second)); 1767 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 1768 AE = NewSetMembers.end(); A != AE; ++A) 1769 LoadMoveSet.insert(*A); 1770 } 1771 1772 // Before removing I, set the iterator to the next instruction. 1773 PI = llvm::next(BasicBlock::iterator(I)); 1774 if (cast<Instruction>(PI) == J) 1775 ++PI; 1776 1777 SE->forgetValue(I); 1778 SE->forgetValue(J); 1779 I->eraseFromParent(); 1780 J->eraseFromParent(); 1781 } 1782 1783 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 1784 } 1785} 1786 1787char BBVectorize::ID = 0; 1788static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 1789INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1790INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 1791INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 1792INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 1793 1794BasicBlockPass *llvm::createBBVectorizePass() { 1795 return new BBVectorize(); 1796} 1797 1798