BBVectorize.cpp revision 72465ea23d010507d3746adc126d719005981e05
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/Metadata.h" 27#include "llvm/Pass.h" 28#include "llvm/Type.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/DenseSet.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/ADT/STLExtras.h" 34#include "llvm/ADT/StringExtras.h" 35#include "llvm/Analysis/AliasAnalysis.h" 36#include "llvm/Analysis/AliasSetTracker.h" 37#include "llvm/Analysis/Dominators.h" 38#include "llvm/Analysis/ScalarEvolution.h" 39#include "llvm/Analysis/ScalarEvolutionExpressions.h" 40#include "llvm/Analysis/ValueTracking.h" 41#include "llvm/Support/CommandLine.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/raw_ostream.h" 44#include "llvm/Support/ValueHandle.h" 45#include "llvm/DataLayout.h" 46#include "llvm/TargetTransformInfo.h" 47#include "llvm/Transforms/Utils/Local.h" 48#include "llvm/Transforms/Vectorize.h" 49#include <algorithm> 50#include <map> 51using namespace llvm; 52 53static cl::opt<bool> 54IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false), 55 cl::Hidden, cl::desc("Ignore target information")); 56 57static cl::opt<unsigned> 58ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 59 cl::desc("The required chain depth for vectorization")); 60 61static cl::opt<unsigned> 62SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 63 cl::desc("The maximum search distance for instruction pairs")); 64 65static cl::opt<bool> 66SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 67 cl::desc("Replicating one element to a pair breaks the chain")); 68 69static cl::opt<unsigned> 70VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 71 cl::desc("The size of the native vector registers")); 72 73static cl::opt<unsigned> 74MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 75 cl::desc("The maximum number of pairing iterations")); 76 77static cl::opt<bool> 78Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden, 79 cl::desc("Don't try to form non-2^n-length vectors")); 80 81static cl::opt<unsigned> 82MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 83 cl::desc("The maximum number of pairable instructions per group")); 84 85static cl::opt<unsigned> 86MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 87 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 88 " a full cycle check")); 89 90static cl::opt<bool> 91NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden, 92 cl::desc("Don't try to vectorize boolean (i1) values")); 93 94static cl::opt<bool> 95NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 96 cl::desc("Don't try to vectorize integer values")); 97 98static cl::opt<bool> 99NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 100 cl::desc("Don't try to vectorize floating-point values")); 101 102// FIXME: This should default to false once pointer vector support works. 103static cl::opt<bool> 104NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden, 105 cl::desc("Don't try to vectorize pointer values")); 106 107static cl::opt<bool> 108NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 109 cl::desc("Don't try to vectorize casting (conversion) operations")); 110 111static cl::opt<bool> 112NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 113 cl::desc("Don't try to vectorize floating-point math intrinsics")); 114 115static cl::opt<bool> 116NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 117 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 118 119static cl::opt<bool> 120NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 121 cl::desc("Don't try to vectorize select instructions")); 122 123static cl::opt<bool> 124NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 125 cl::desc("Don't try to vectorize comparison instructions")); 126 127static cl::opt<bool> 128NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 129 cl::desc("Don't try to vectorize getelementptr instructions")); 130 131static cl::opt<bool> 132NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 133 cl::desc("Don't try to vectorize loads and stores")); 134 135static cl::opt<bool> 136AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 137 cl::desc("Only generate aligned loads and stores")); 138 139static cl::opt<bool> 140NoMemOpBoost("bb-vectorize-no-mem-op-boost", 141 cl::init(false), cl::Hidden, 142 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 143 144static cl::opt<bool> 145FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 146 cl::desc("Use a fast instruction dependency analysis")); 147 148#ifndef NDEBUG 149static cl::opt<bool> 150DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 151 cl::init(false), cl::Hidden, 152 cl::desc("When debugging is enabled, output information on the" 153 " instruction-examination process")); 154static cl::opt<bool> 155DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 156 cl::init(false), cl::Hidden, 157 cl::desc("When debugging is enabled, output information on the" 158 " candidate-selection process")); 159static cl::opt<bool> 160DebugPairSelection("bb-vectorize-debug-pair-selection", 161 cl::init(false), cl::Hidden, 162 cl::desc("When debugging is enabled, output information on the" 163 " pair-selection process")); 164static cl::opt<bool> 165DebugCycleCheck("bb-vectorize-debug-cycle-check", 166 cl::init(false), cl::Hidden, 167 cl::desc("When debugging is enabled, output information on the" 168 " cycle-checking process")); 169 170static cl::opt<bool> 171PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair", 172 cl::init(false), cl::Hidden, 173 cl::desc("When debugging is enabled, dump the basic block after" 174 " every pair is fused")); 175#endif 176 177STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 178 179namespace { 180 struct BBVectorize : public BasicBlockPass { 181 static char ID; // Pass identification, replacement for typeid 182 183 const VectorizeConfig Config; 184 185 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 186 : BasicBlockPass(ID), Config(C) { 187 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 188 } 189 190 BBVectorize(Pass *P, const VectorizeConfig &C) 191 : BasicBlockPass(ID), Config(C) { 192 AA = &P->getAnalysis<AliasAnalysis>(); 193 DT = &P->getAnalysis<DominatorTree>(); 194 SE = &P->getAnalysis<ScalarEvolution>(); 195 TD = P->getAnalysisIfAvailable<DataLayout>(); 196 TTI = IgnoreTargetInfo ? 0 : 197 P->getAnalysisIfAvailable<TargetTransformInfo>(); 198 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 199 } 200 201 typedef std::pair<Value *, Value *> ValuePair; 202 typedef std::pair<ValuePair, int> ValuePairWithCost; 203 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 204 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 205 typedef std::pair<VPPair, unsigned> VPPairWithType; 206 typedef std::pair<std::multimap<Value *, Value *>::iterator, 207 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 208 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 209 std::multimap<ValuePair, ValuePair>::iterator> 210 VPPIteratorPair; 211 212 AliasAnalysis *AA; 213 DominatorTree *DT; 214 ScalarEvolution *SE; 215 DataLayout *TD; 216 TargetTransformInfo *TTI; 217 const VectorTargetTransformInfo *VTTI; 218 219 // FIXME: const correct? 220 221 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 222 223 bool getCandidatePairs(BasicBlock &BB, 224 BasicBlock::iterator &Start, 225 std::multimap<Value *, Value *> &CandidatePairs, 226 DenseSet<ValuePair> &FixedOrderPairs, 227 DenseMap<ValuePair, int> &CandidatePairCostSavings, 228 std::vector<Value *> &PairableInsts, bool NonPow2Len); 229 230 enum PairConnectionType { 231 PairConnectionDirect, 232 PairConnectionSwap, 233 PairConnectionSplat 234 }; 235 236 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 237 std::vector<Value *> &PairableInsts, 238 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 239 DenseMap<VPPair, unsigned> &PairConnectionTypes); 240 241 void buildDepMap(BasicBlock &BB, 242 std::multimap<Value *, Value *> &CandidatePairs, 243 std::vector<Value *> &PairableInsts, 244 DenseSet<ValuePair> &PairableInstUsers); 245 246 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 247 DenseMap<ValuePair, int> &CandidatePairCostSavings, 248 std::vector<Value *> &PairableInsts, 249 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 250 DenseSet<ValuePair> &PairableInstUsers, 251 DenseMap<Value *, Value *>& ChosenPairs); 252 253 void fuseChosenPairs(BasicBlock &BB, 254 std::vector<Value *> &PairableInsts, 255 DenseMap<Value *, Value *>& ChosenPairs, 256 DenseSet<ValuePair> &FixedOrderPairs, 257 DenseMap<VPPair, unsigned> &PairConnectionTypes, 258 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 259 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps); 260 261 262 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 263 264 bool areInstsCompatible(Instruction *I, Instruction *J, 265 bool IsSimpleLoadStore, bool NonPow2Len, 266 int &CostSavings, int &FixedOrder); 267 268 bool trackUsesOfI(DenseSet<Value *> &Users, 269 AliasSetTracker &WriteSet, Instruction *I, 270 Instruction *J, bool UpdateUsers = true, 271 std::multimap<Value *, Value *> *LoadMoveSet = 0); 272 273 void computePairsConnectedTo( 274 std::multimap<Value *, Value *> &CandidatePairs, 275 std::vector<Value *> &PairableInsts, 276 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 277 DenseMap<VPPair, unsigned> &PairConnectionTypes, 278 ValuePair P); 279 280 bool pairsConflict(ValuePair P, ValuePair Q, 281 DenseSet<ValuePair> &PairableInstUsers, 282 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 283 284 bool pairWillFormCycle(ValuePair P, 285 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 286 DenseSet<ValuePair> &CurrentPairs); 287 288 void pruneTreeFor( 289 std::multimap<Value *, Value *> &CandidatePairs, 290 std::vector<Value *> &PairableInsts, 291 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 292 DenseSet<ValuePair> &PairableInstUsers, 293 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 294 DenseMap<Value *, Value *> &ChosenPairs, 295 DenseMap<ValuePair, size_t> &Tree, 296 DenseSet<ValuePair> &PrunedTree, ValuePair J, 297 bool UseCycleCheck); 298 299 void buildInitialTreeFor( 300 std::multimap<Value *, Value *> &CandidatePairs, 301 std::vector<Value *> &PairableInsts, 302 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 303 DenseSet<ValuePair> &PairableInstUsers, 304 DenseMap<Value *, Value *> &ChosenPairs, 305 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 306 307 void findBestTreeFor( 308 std::multimap<Value *, Value *> &CandidatePairs, 309 DenseMap<ValuePair, int> &CandidatePairCostSavings, 310 std::vector<Value *> &PairableInsts, 311 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 312 DenseSet<ValuePair> &PairableInstUsers, 313 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 314 DenseMap<Value *, Value *> &ChosenPairs, 315 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 316 int &BestEffSize, VPIteratorPair ChoiceRange, 317 bool UseCycleCheck); 318 319 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 320 Instruction *J, unsigned o); 321 322 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 323 unsigned MaskOffset, unsigned NumInElem, 324 unsigned NumInElem1, unsigned IdxOffset, 325 std::vector<Constant*> &Mask); 326 327 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 328 Instruction *J); 329 330 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 331 unsigned o, Value *&LOp, unsigned numElemL, 332 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, 333 unsigned IdxOff = 0); 334 335 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 336 Instruction *J, unsigned o, bool IBeforeJ); 337 338 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 339 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 340 bool IBeforeJ); 341 342 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 343 Instruction *J, Instruction *K, 344 Instruction *&InsertionPt, Instruction *&K1, 345 Instruction *&K2); 346 347 void collectPairLoadMoveSet(BasicBlock &BB, 348 DenseMap<Value *, Value *> &ChosenPairs, 349 std::multimap<Value *, Value *> &LoadMoveSet, 350 Instruction *I); 351 352 void collectLoadMoveSet(BasicBlock &BB, 353 std::vector<Value *> &PairableInsts, 354 DenseMap<Value *, Value *> &ChosenPairs, 355 std::multimap<Value *, Value *> &LoadMoveSet); 356 357 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 358 std::multimap<Value *, Value *> &LoadMoveSet, 359 Instruction *I, Instruction *J); 360 361 void moveUsesOfIAfterJ(BasicBlock &BB, 362 std::multimap<Value *, Value *> &LoadMoveSet, 363 Instruction *&InsertionPt, 364 Instruction *I, Instruction *J); 365 366 void combineMetadata(Instruction *K, const Instruction *J); 367 368 bool vectorizeBB(BasicBlock &BB) { 369 if (!DT->isReachableFromEntry(&BB)) { 370 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 371 " in " << BB.getParent()->getName() << "\n"); 372 return false; 373 } 374 375 DEBUG(if (VTTI) dbgs() << "BBV: using target information\n"); 376 377 bool changed = false; 378 // Iterate a sufficient number of times to merge types of size 1 bit, 379 // then 2 bits, then 4, etc. up to half of the target vector width of the 380 // target vector register. 381 unsigned n = 1; 382 for (unsigned v = 2; 383 (VTTI || v <= Config.VectorBits) && 384 (!Config.MaxIter || n <= Config.MaxIter); 385 v *= 2, ++n) { 386 DEBUG(dbgs() << "BBV: fusing loop #" << n << 387 " for " << BB.getName() << " in " << 388 BB.getParent()->getName() << "...\n"); 389 if (vectorizePairs(BB)) 390 changed = true; 391 else 392 break; 393 } 394 395 if (changed && !Pow2LenOnly) { 396 ++n; 397 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 398 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 399 n << " for " << BB.getName() << " in " << 400 BB.getParent()->getName() << "...\n"); 401 if (!vectorizePairs(BB, true)) break; 402 } 403 } 404 405 DEBUG(dbgs() << "BBV: done!\n"); 406 return changed; 407 } 408 409 virtual bool runOnBasicBlock(BasicBlock &BB) { 410 AA = &getAnalysis<AliasAnalysis>(); 411 DT = &getAnalysis<DominatorTree>(); 412 SE = &getAnalysis<ScalarEvolution>(); 413 TD = getAnalysisIfAvailable<DataLayout>(); 414 TTI = IgnoreTargetInfo ? 0 : 415 getAnalysisIfAvailable<TargetTransformInfo>(); 416 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 417 418 return vectorizeBB(BB); 419 } 420 421 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 422 BasicBlockPass::getAnalysisUsage(AU); 423 AU.addRequired<AliasAnalysis>(); 424 AU.addRequired<DominatorTree>(); 425 AU.addRequired<ScalarEvolution>(); 426 AU.addPreserved<AliasAnalysis>(); 427 AU.addPreserved<DominatorTree>(); 428 AU.addPreserved<ScalarEvolution>(); 429 AU.setPreservesCFG(); 430 } 431 432 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 433 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 434 "Cannot form vector from incompatible scalar types"); 435 Type *STy = ElemTy->getScalarType(); 436 437 unsigned numElem; 438 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 439 numElem = VTy->getNumElements(); 440 } else { 441 numElem = 1; 442 } 443 444 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 445 numElem += VTy->getNumElements(); 446 } else { 447 numElem += 1; 448 } 449 450 return VectorType::get(STy, numElem); 451 } 452 453 static inline void getInstructionTypes(Instruction *I, 454 Type *&T1, Type *&T2) { 455 if (isa<StoreInst>(I)) { 456 // For stores, it is the value type, not the pointer type that matters 457 // because the value is what will come from a vector register. 458 459 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 460 T1 = IVal->getType(); 461 } else { 462 T1 = I->getType(); 463 } 464 465 if (I->isCast()) 466 T2 = cast<CastInst>(I)->getSrcTy(); 467 else 468 T2 = T1; 469 470 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 471 T2 = SI->getCondition()->getType(); 472 } 473 } 474 475 // Returns the weight associated with the provided value. A chain of 476 // candidate pairs has a length given by the sum of the weights of its 477 // members (one weight per pair; the weight of each member of the pair 478 // is assumed to be the same). This length is then compared to the 479 // chain-length threshold to determine if a given chain is significant 480 // enough to be vectorized. The length is also used in comparing 481 // candidate chains where longer chains are considered to be better. 482 // Note: when this function returns 0, the resulting instructions are 483 // not actually fused. 484 inline size_t getDepthFactor(Value *V) { 485 // InsertElement and ExtractElement have a depth factor of zero. This is 486 // for two reasons: First, they cannot be usefully fused. Second, because 487 // the pass generates a lot of these, they can confuse the simple metric 488 // used to compare the trees in the next iteration. Thus, giving them a 489 // weight of zero allows the pass to essentially ignore them in 490 // subsequent iterations when looking for vectorization opportunities 491 // while still tracking dependency chains that flow through those 492 // instructions. 493 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 494 return 0; 495 496 // Give a load or store half of the required depth so that load/store 497 // pairs will vectorize. 498 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 499 return Config.ReqChainDepth/2; 500 501 return 1; 502 } 503 504 // Returns the cost of the provided instruction using VTTI. 505 // This does not handle loads and stores. 506 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { 507 switch (Opcode) { 508 default: break; 509 case Instruction::GetElementPtr: 510 // We mark this instruction as zero-cost because scalar GEPs are usually 511 // lowered to the intruction addressing mode. At the moment we don't 512 // generate vector GEPs. 513 return 0; 514 case Instruction::Br: 515 return VTTI->getCFInstrCost(Opcode); 516 case Instruction::PHI: 517 return 0; 518 case Instruction::Add: 519 case Instruction::FAdd: 520 case Instruction::Sub: 521 case Instruction::FSub: 522 case Instruction::Mul: 523 case Instruction::FMul: 524 case Instruction::UDiv: 525 case Instruction::SDiv: 526 case Instruction::FDiv: 527 case Instruction::URem: 528 case Instruction::SRem: 529 case Instruction::FRem: 530 case Instruction::Shl: 531 case Instruction::LShr: 532 case Instruction::AShr: 533 case Instruction::And: 534 case Instruction::Or: 535 case Instruction::Xor: 536 return VTTI->getArithmeticInstrCost(Opcode, T1); 537 case Instruction::Select: 538 case Instruction::ICmp: 539 case Instruction::FCmp: 540 return VTTI->getCmpSelInstrCost(Opcode, T1, T2); 541 case Instruction::ZExt: 542 case Instruction::SExt: 543 case Instruction::FPToUI: 544 case Instruction::FPToSI: 545 case Instruction::FPExt: 546 case Instruction::PtrToInt: 547 case Instruction::IntToPtr: 548 case Instruction::SIToFP: 549 case Instruction::UIToFP: 550 case Instruction::Trunc: 551 case Instruction::FPTrunc: 552 case Instruction::BitCast: 553 return VTTI->getCastInstrCost(Opcode, T1, T2); 554 } 555 556 return 1; 557 } 558 559 // This determines the relative offset of two loads or stores, returning 560 // true if the offset could be determined to be some constant value. 561 // For example, if OffsetInElmts == 1, then J accesses the memory directly 562 // after I; if OffsetInElmts == -1 then I accesses the memory 563 // directly after J. 564 bool getPairPtrInfo(Instruction *I, Instruction *J, 565 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 566 unsigned &IAddressSpace, unsigned &JAddressSpace, 567 int64_t &OffsetInElmts, bool ComputeOffset = true) { 568 OffsetInElmts = 0; 569 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 570 LoadInst *LJ = cast<LoadInst>(J); 571 IPtr = LI->getPointerOperand(); 572 JPtr = LJ->getPointerOperand(); 573 IAlignment = LI->getAlignment(); 574 JAlignment = LJ->getAlignment(); 575 IAddressSpace = LI->getPointerAddressSpace(); 576 JAddressSpace = LJ->getPointerAddressSpace(); 577 } else { 578 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 579 IPtr = SI->getPointerOperand(); 580 JPtr = SJ->getPointerOperand(); 581 IAlignment = SI->getAlignment(); 582 JAlignment = SJ->getAlignment(); 583 IAddressSpace = SI->getPointerAddressSpace(); 584 JAddressSpace = SJ->getPointerAddressSpace(); 585 } 586 587 if (!ComputeOffset) 588 return true; 589 590 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 591 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 592 593 // If this is a trivial offset, then we'll get something like 594 // 1*sizeof(type). With target data, which we need anyway, this will get 595 // constant folded into a number. 596 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 597 if (const SCEVConstant *ConstOffSCEV = 598 dyn_cast<SCEVConstant>(OffsetSCEV)) { 599 ConstantInt *IntOff = ConstOffSCEV->getValue(); 600 int64_t Offset = IntOff->getSExtValue(); 601 602 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 603 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 604 605 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 606 if (VTy != VTy2 && Offset < 0) { 607 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 608 OffsetInElmts = Offset/VTy2TSS; 609 return (abs64(Offset) % VTy2TSS) == 0; 610 } 611 612 OffsetInElmts = Offset/VTyTSS; 613 return (abs64(Offset) % VTyTSS) == 0; 614 } 615 616 return false; 617 } 618 619 // Returns true if the provided CallInst represents an intrinsic that can 620 // be vectorized. 621 bool isVectorizableIntrinsic(CallInst* I) { 622 Function *F = I->getCalledFunction(); 623 if (!F) return false; 624 625 unsigned IID = F->getIntrinsicID(); 626 if (!IID) return false; 627 628 switch(IID) { 629 default: 630 return false; 631 case Intrinsic::sqrt: 632 case Intrinsic::powi: 633 case Intrinsic::sin: 634 case Intrinsic::cos: 635 case Intrinsic::log: 636 case Intrinsic::log2: 637 case Intrinsic::log10: 638 case Intrinsic::exp: 639 case Intrinsic::exp2: 640 case Intrinsic::pow: 641 return Config.VectorizeMath; 642 case Intrinsic::fma: 643 return Config.VectorizeFMA; 644 } 645 } 646 647 // Returns true if J is the second element in some pair referenced by 648 // some multimap pair iterator pair. 649 template <typename V> 650 bool isSecondInIteratorPair(V J, std::pair< 651 typename std::multimap<V, V>::iterator, 652 typename std::multimap<V, V>::iterator> PairRange) { 653 for (typename std::multimap<V, V>::iterator K = PairRange.first; 654 K != PairRange.second; ++K) 655 if (K->second == J) return true; 656 657 return false; 658 } 659 }; 660 661 // This function implements one vectorization iteration on the provided 662 // basic block. It returns true if the block is changed. 663 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 664 bool ShouldContinue; 665 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 666 667 std::vector<Value *> AllPairableInsts; 668 DenseMap<Value *, Value *> AllChosenPairs; 669 DenseSet<ValuePair> AllFixedOrderPairs; 670 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 671 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; 672 673 do { 674 std::vector<Value *> PairableInsts; 675 std::multimap<Value *, Value *> CandidatePairs; 676 DenseSet<ValuePair> FixedOrderPairs; 677 DenseMap<ValuePair, int> CandidatePairCostSavings; 678 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 679 FixedOrderPairs, 680 CandidatePairCostSavings, 681 PairableInsts, NonPow2Len); 682 if (PairableInsts.empty()) continue; 683 684 // Now we have a map of all of the pairable instructions and we need to 685 // select the best possible pairing. A good pairing is one such that the 686 // users of the pair are also paired. This defines a (directed) forest 687 // over the pairs such that two pairs are connected iff the second pair 688 // uses the first. 689 690 // Note that it only matters that both members of the second pair use some 691 // element of the first pair (to allow for splatting). 692 693 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; 694 DenseMap<VPPair, unsigned> PairConnectionTypes; 695 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs, 696 PairConnectionTypes); 697 if (ConnectedPairs.empty()) continue; 698 699 for (std::multimap<ValuePair, ValuePair>::iterator 700 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 701 I != IE; ++I) { 702 ConnectedPairDeps.insert(VPPair(I->second, I->first)); 703 } 704 705 // Build the pairable-instruction dependency map 706 DenseSet<ValuePair> PairableInstUsers; 707 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 708 709 // There is now a graph of the connected pairs. For each variable, pick 710 // the pairing with the largest tree meeting the depth requirement on at 711 // least one branch. Then select all pairings that are part of that tree 712 // and remove them from the list of available pairings and pairable 713 // variables. 714 715 DenseMap<Value *, Value *> ChosenPairs; 716 choosePairs(CandidatePairs, CandidatePairCostSavings, 717 PairableInsts, ConnectedPairs, 718 PairableInstUsers, ChosenPairs); 719 720 if (ChosenPairs.empty()) continue; 721 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 722 PairableInsts.end()); 723 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 724 725 // Only for the chosen pairs, propagate information on fixed-order pairs, 726 // pair connections, and their types to the data structures used by the 727 // pair fusion procedures. 728 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 729 IE = ChosenPairs.end(); I != IE; ++I) { 730 if (FixedOrderPairs.count(*I)) 731 AllFixedOrderPairs.insert(*I); 732 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 733 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 734 735 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 736 J != IE; ++J) { 737 DenseMap<VPPair, unsigned>::iterator K = 738 PairConnectionTypes.find(VPPair(*I, *J)); 739 if (K != PairConnectionTypes.end()) { 740 AllPairConnectionTypes.insert(*K); 741 } else { 742 K = PairConnectionTypes.find(VPPair(*J, *I)); 743 if (K != PairConnectionTypes.end()) 744 AllPairConnectionTypes.insert(*K); 745 } 746 } 747 } 748 749 for (std::multimap<ValuePair, ValuePair>::iterator 750 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 751 I != IE; ++I) { 752 if (AllPairConnectionTypes.count(*I)) { 753 AllConnectedPairs.insert(*I); 754 AllConnectedPairDeps.insert(VPPair(I->second, I->first)); 755 } 756 } 757 } while (ShouldContinue); 758 759 if (AllChosenPairs.empty()) return false; 760 NumFusedOps += AllChosenPairs.size(); 761 762 // A set of pairs has now been selected. It is now necessary to replace the 763 // paired instructions with vector instructions. For this procedure each 764 // operand must be replaced with a vector operand. This vector is formed 765 // by using build_vector on the old operands. The replaced values are then 766 // replaced with a vector_extract on the result. Subsequent optimization 767 // passes should coalesce the build/extract combinations. 768 769 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 770 AllPairConnectionTypes, 771 AllConnectedPairs, AllConnectedPairDeps); 772 773 // It is important to cleanup here so that future iterations of this 774 // function have less work to do. 775 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 776 return true; 777 } 778 779 // This function returns true if the provided instruction is capable of being 780 // fused into a vector instruction. This determination is based only on the 781 // type and other attributes of the instruction. 782 bool BBVectorize::isInstVectorizable(Instruction *I, 783 bool &IsSimpleLoadStore) { 784 IsSimpleLoadStore = false; 785 786 if (CallInst *C = dyn_cast<CallInst>(I)) { 787 if (!isVectorizableIntrinsic(C)) 788 return false; 789 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 790 // Vectorize simple loads if possbile: 791 IsSimpleLoadStore = L->isSimple(); 792 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 793 return false; 794 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 795 // Vectorize simple stores if possbile: 796 IsSimpleLoadStore = S->isSimple(); 797 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 798 return false; 799 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 800 // We can vectorize casts, but not casts of pointer types, etc. 801 if (!Config.VectorizeCasts) 802 return false; 803 804 Type *SrcTy = C->getSrcTy(); 805 if (!SrcTy->isSingleValueType()) 806 return false; 807 808 Type *DestTy = C->getDestTy(); 809 if (!DestTy->isSingleValueType()) 810 return false; 811 } else if (isa<SelectInst>(I)) { 812 if (!Config.VectorizeSelect) 813 return false; 814 } else if (isa<CmpInst>(I)) { 815 if (!Config.VectorizeCmp) 816 return false; 817 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 818 if (!Config.VectorizeGEP) 819 return false; 820 821 // Currently, vector GEPs exist only with one index. 822 if (G->getNumIndices() != 1) 823 return false; 824 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 825 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 826 return false; 827 } 828 829 // We can't vectorize memory operations without target data 830 if (TD == 0 && IsSimpleLoadStore) 831 return false; 832 833 Type *T1, *T2; 834 getInstructionTypes(I, T1, T2); 835 836 // Not every type can be vectorized... 837 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 838 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 839 return false; 840 841 if (T1->getScalarSizeInBits() == 1) { 842 if (!Config.VectorizeBools) 843 return false; 844 } else { 845 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 846 return false; 847 } 848 849 if (T2->getScalarSizeInBits() == 1) { 850 if (!Config.VectorizeBools) 851 return false; 852 } else { 853 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 854 return false; 855 } 856 857 if (!Config.VectorizeFloats 858 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 859 return false; 860 861 // Don't vectorize target-specific types. 862 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 863 return false; 864 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 865 return false; 866 867 if ((!Config.VectorizePointers || TD == 0) && 868 (T1->getScalarType()->isPointerTy() || 869 T2->getScalarType()->isPointerTy())) 870 return false; 871 872 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 873 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 874 return false; 875 876 return true; 877 } 878 879 // This function returns true if the two provided instructions are compatible 880 // (meaning that they can be fused into a vector instruction). This assumes 881 // that I has already been determined to be vectorizable and that J is not 882 // in the use tree of I. 883 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 884 bool IsSimpleLoadStore, bool NonPow2Len, 885 int &CostSavings, int &FixedOrder) { 886 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 887 " <-> " << *J << "\n"); 888 889 CostSavings = 0; 890 FixedOrder = 0; 891 892 // Loads and stores can be merged if they have different alignments, 893 // but are otherwise the same. 894 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 895 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 896 return false; 897 898 Type *IT1, *IT2, *JT1, *JT2; 899 getInstructionTypes(I, IT1, IT2); 900 getInstructionTypes(J, JT1, JT2); 901 unsigned MaxTypeBits = std::max( 902 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 903 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 904 if (!VTTI && MaxTypeBits > Config.VectorBits) 905 return false; 906 907 // FIXME: handle addsub-type operations! 908 909 if (IsSimpleLoadStore) { 910 Value *IPtr, *JPtr; 911 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 912 int64_t OffsetInElmts = 0; 913 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 914 IAddressSpace, JAddressSpace, 915 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 916 FixedOrder = (int) OffsetInElmts; 917 unsigned BottomAlignment = IAlignment; 918 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 919 920 Type *aTypeI = isa<StoreInst>(I) ? 921 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 922 Type *aTypeJ = isa<StoreInst>(J) ? 923 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 924 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 925 926 if (Config.AlignedOnly) { 927 // An aligned load or store is possible only if the instruction 928 // with the lower offset has an alignment suitable for the 929 // vector type. 930 931 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 932 if (BottomAlignment < VecAlignment) 933 return false; 934 } 935 936 if (VTTI) { 937 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(), 938 IAlignment, IAddressSpace); 939 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(), 940 JAlignment, JAddressSpace); 941 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType, 942 BottomAlignment, 943 IAddressSpace); 944 if (VCost > ICost + JCost) 945 return false; 946 947 // We don't want to fuse to a type that will be split, even 948 // if the two input types will also be split and there is no other 949 // associated cost. 950 unsigned VParts = VTTI->getNumberOfParts(VType); 951 if (VParts > 1) 952 return false; 953 else if (!VParts && VCost == ICost + JCost) 954 return false; 955 956 CostSavings = ICost + JCost - VCost; 957 } 958 } else { 959 return false; 960 } 961 } else if (VTTI) { 962 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 963 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 964 Type *VT1 = getVecTypeForPair(IT1, JT1), 965 *VT2 = getVecTypeForPair(IT2, JT2); 966 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 967 968 if (VCost > ICost + JCost) 969 return false; 970 971 // We don't want to fuse to a type that will be split, even 972 // if the two input types will also be split and there is no other 973 // associated cost. 974 unsigned VParts = VTTI->getNumberOfParts(VT1); 975 if (VParts > 1) 976 return false; 977 else if (!VParts && VCost == ICost + JCost) 978 return false; 979 980 CostSavings = ICost + JCost - VCost; 981 } 982 983 // The powi intrinsic is special because only the first argument is 984 // vectorized, the second arguments must be equal. 985 CallInst *CI = dyn_cast<CallInst>(I); 986 Function *FI; 987 if (CI && (FI = CI->getCalledFunction()) && 988 FI->getIntrinsicID() == Intrinsic::powi) { 989 990 Value *A1I = CI->getArgOperand(1), 991 *A1J = cast<CallInst>(J)->getArgOperand(1); 992 const SCEV *A1ISCEV = SE->getSCEV(A1I), 993 *A1JSCEV = SE->getSCEV(A1J); 994 return (A1ISCEV == A1JSCEV); 995 } 996 997 return true; 998 } 999 1000 // Figure out whether or not J uses I and update the users and write-set 1001 // structures associated with I. Specifically, Users represents the set of 1002 // instructions that depend on I. WriteSet represents the set 1003 // of memory locations that are dependent on I. If UpdateUsers is true, 1004 // and J uses I, then Users is updated to contain J and WriteSet is updated 1005 // to contain any memory locations to which J writes. The function returns 1006 // true if J uses I. By default, alias analysis is used to determine 1007 // whether J reads from memory that overlaps with a location in WriteSet. 1008 // If LoadMoveSet is not null, then it is a previously-computed multimap 1009 // where the key is the memory-based user instruction and the value is 1010 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1011 // then the alias analysis is not used. This is necessary because this 1012 // function is called during the process of moving instructions during 1013 // vectorization and the results of the alias analysis are not stable during 1014 // that process. 1015 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1016 AliasSetTracker &WriteSet, Instruction *I, 1017 Instruction *J, bool UpdateUsers, 1018 std::multimap<Value *, Value *> *LoadMoveSet) { 1019 bool UsesI = false; 1020 1021 // This instruction may already be marked as a user due, for example, to 1022 // being a member of a selected pair. 1023 if (Users.count(J)) 1024 UsesI = true; 1025 1026 if (!UsesI) 1027 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1028 JU != JE; ++JU) { 1029 Value *V = *JU; 1030 if (I == V || Users.count(V)) { 1031 UsesI = true; 1032 break; 1033 } 1034 } 1035 if (!UsesI && J->mayReadFromMemory()) { 1036 if (LoadMoveSet) { 1037 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 1038 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 1039 } else { 1040 for (AliasSetTracker::iterator W = WriteSet.begin(), 1041 WE = WriteSet.end(); W != WE; ++W) { 1042 if (W->aliasesUnknownInst(J, *AA)) { 1043 UsesI = true; 1044 break; 1045 } 1046 } 1047 } 1048 } 1049 1050 if (UsesI && UpdateUsers) { 1051 if (J->mayWriteToMemory()) WriteSet.add(J); 1052 Users.insert(J); 1053 } 1054 1055 return UsesI; 1056 } 1057 1058 // This function iterates over all instruction pairs in the provided 1059 // basic block and collects all candidate pairs for vectorization. 1060 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1061 BasicBlock::iterator &Start, 1062 std::multimap<Value *, Value *> &CandidatePairs, 1063 DenseSet<ValuePair> &FixedOrderPairs, 1064 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1065 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1066 BasicBlock::iterator E = BB.end(); 1067 if (Start == E) return false; 1068 1069 bool ShouldContinue = false, IAfterStart = false; 1070 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1071 if (I == Start) IAfterStart = true; 1072 1073 bool IsSimpleLoadStore; 1074 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1075 1076 // Look for an instruction with which to pair instruction *I... 1077 DenseSet<Value *> Users; 1078 AliasSetTracker WriteSet(*AA); 1079 bool JAfterStart = IAfterStart; 1080 BasicBlock::iterator J = llvm::next(I); 1081 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1082 if (J == Start) JAfterStart = true; 1083 1084 // Determine if J uses I, if so, exit the loop. 1085 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1086 if (Config.FastDep) { 1087 // Note: For this heuristic to be effective, independent operations 1088 // must tend to be intermixed. This is likely to be true from some 1089 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1090 // but otherwise may require some kind of reordering pass. 1091 1092 // When using fast dependency analysis, 1093 // stop searching after first use: 1094 if (UsesI) break; 1095 } else { 1096 if (UsesI) continue; 1097 } 1098 1099 // J does not use I, and comes before the first use of I, so it can be 1100 // merged with I if the instructions are compatible. 1101 int CostSavings, FixedOrder; 1102 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1103 CostSavings, FixedOrder)) continue; 1104 1105 // J is a candidate for merging with I. 1106 if (!PairableInsts.size() || 1107 PairableInsts[PairableInsts.size()-1] != I) { 1108 PairableInsts.push_back(I); 1109 } 1110 1111 CandidatePairs.insert(ValuePair(I, J)); 1112 if (VTTI) 1113 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1114 CostSavings)); 1115 1116 if (FixedOrder == 1) 1117 FixedOrderPairs.insert(ValuePair(I, J)); 1118 else if (FixedOrder == -1) 1119 FixedOrderPairs.insert(ValuePair(J, I)); 1120 1121 // The next call to this function must start after the last instruction 1122 // selected during this invocation. 1123 if (JAfterStart) { 1124 Start = llvm::next(J); 1125 IAfterStart = JAfterStart = false; 1126 } 1127 1128 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1129 << *I << " <-> " << *J << " (cost savings: " << 1130 CostSavings << ")\n"); 1131 1132 // If we have already found too many pairs, break here and this function 1133 // will be called again starting after the last instruction selected 1134 // during this invocation. 1135 if (PairableInsts.size() >= Config.MaxInsts) { 1136 ShouldContinue = true; 1137 break; 1138 } 1139 } 1140 1141 if (ShouldContinue) 1142 break; 1143 } 1144 1145 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1146 << " instructions with candidate pairs\n"); 1147 1148 return ShouldContinue; 1149 } 1150 1151 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1152 // it looks for pairs such that both members have an input which is an 1153 // output of PI or PJ. 1154 void BBVectorize::computePairsConnectedTo( 1155 std::multimap<Value *, Value *> &CandidatePairs, 1156 std::vector<Value *> &PairableInsts, 1157 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1158 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1159 ValuePair P) { 1160 StoreInst *SI, *SJ; 1161 1162 // For each possible pairing for this variable, look at the uses of 1163 // the first value... 1164 for (Value::use_iterator I = P.first->use_begin(), 1165 E = P.first->use_end(); I != E; ++I) { 1166 if (isa<LoadInst>(*I)) { 1167 // A pair cannot be connected to a load because the load only takes one 1168 // operand (the address) and it is a scalar even after vectorization. 1169 continue; 1170 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1171 P.first == SI->getPointerOperand()) { 1172 // Similarly, a pair cannot be connected to a store through its 1173 // pointer operand. 1174 continue; 1175 } 1176 1177 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1178 1179 // For each use of the first variable, look for uses of the second 1180 // variable... 1181 for (Value::use_iterator J = P.second->use_begin(), 1182 E2 = P.second->use_end(); J != E2; ++J) { 1183 if ((SJ = dyn_cast<StoreInst>(*J)) && 1184 P.second == SJ->getPointerOperand()) 1185 continue; 1186 1187 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 1188 1189 // Look for <I, J>: 1190 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1191 VPPair VP(P, ValuePair(*I, *J)); 1192 ConnectedPairs.insert(VP); 1193 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1194 } 1195 1196 // Look for <J, I>: 1197 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) { 1198 VPPair VP(P, ValuePair(*J, *I)); 1199 ConnectedPairs.insert(VP); 1200 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1201 } 1202 } 1203 1204 if (Config.SplatBreaksChain) continue; 1205 // Look for cases where just the first value in the pair is used by 1206 // both members of another pair (splatting). 1207 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1208 if ((SJ = dyn_cast<StoreInst>(*J)) && 1209 P.first == SJ->getPointerOperand()) 1210 continue; 1211 1212 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1213 VPPair VP(P, ValuePair(*I, *J)); 1214 ConnectedPairs.insert(VP); 1215 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1216 } 1217 } 1218 } 1219 1220 if (Config.SplatBreaksChain) return; 1221 // Look for cases where just the second value in the pair is used by 1222 // both members of another pair (splatting). 1223 for (Value::use_iterator I = P.second->use_begin(), 1224 E = P.second->use_end(); I != E; ++I) { 1225 if (isa<LoadInst>(*I)) 1226 continue; 1227 else if ((SI = dyn_cast<StoreInst>(*I)) && 1228 P.second == SI->getPointerOperand()) 1229 continue; 1230 1231 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1232 1233 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1234 if ((SJ = dyn_cast<StoreInst>(*J)) && 1235 P.second == SJ->getPointerOperand()) 1236 continue; 1237 1238 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1239 VPPair VP(P, ValuePair(*I, *J)); 1240 ConnectedPairs.insert(VP); 1241 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1242 } 1243 } 1244 } 1245 } 1246 1247 // This function figures out which pairs are connected. Two pairs are 1248 // connected if some output of the first pair forms an input to both members 1249 // of the second pair. 1250 void BBVectorize::computeConnectedPairs( 1251 std::multimap<Value *, Value *> &CandidatePairs, 1252 std::vector<Value *> &PairableInsts, 1253 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1254 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1255 1256 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1257 PE = PairableInsts.end(); PI != PE; ++PI) { 1258 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1259 1260 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1261 P != choiceRange.second; ++P) 1262 computePairsConnectedTo(CandidatePairs, PairableInsts, 1263 ConnectedPairs, PairConnectionTypes, *P); 1264 } 1265 1266 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1267 << " pair connections.\n"); 1268 } 1269 1270 // This function builds a set of use tuples such that <A, B> is in the set 1271 // if B is in the use tree of A. If B is in the use tree of A, then B 1272 // depends on the output of A. 1273 void BBVectorize::buildDepMap( 1274 BasicBlock &BB, 1275 std::multimap<Value *, Value *> &CandidatePairs, 1276 std::vector<Value *> &PairableInsts, 1277 DenseSet<ValuePair> &PairableInstUsers) { 1278 DenseSet<Value *> IsInPair; 1279 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1280 E = CandidatePairs.end(); C != E; ++C) { 1281 IsInPair.insert(C->first); 1282 IsInPair.insert(C->second); 1283 } 1284 1285 // Iterate through the basic block, recording all Users of each 1286 // pairable instruction. 1287 1288 BasicBlock::iterator E = BB.end(); 1289 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1290 if (IsInPair.find(I) == IsInPair.end()) continue; 1291 1292 DenseSet<Value *> Users; 1293 AliasSetTracker WriteSet(*AA); 1294 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1295 (void) trackUsesOfI(Users, WriteSet, I, J); 1296 1297 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1298 U != E; ++U) 1299 PairableInstUsers.insert(ValuePair(I, *U)); 1300 } 1301 } 1302 1303 // Returns true if an input to pair P is an output of pair Q and also an 1304 // input of pair Q is an output of pair P. If this is the case, then these 1305 // two pairs cannot be simultaneously fused. 1306 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1307 DenseSet<ValuePair> &PairableInstUsers, 1308 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1309 // Two pairs are in conflict if they are mutual Users of eachother. 1310 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1311 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1312 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1313 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1314 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1315 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1316 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1317 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1318 if (PairableInstUserMap) { 1319 // FIXME: The expensive part of the cycle check is not so much the cycle 1320 // check itself but this edge insertion procedure. This needs some 1321 // profiling and probably a different data structure (same is true of 1322 // most uses of std::multimap). 1323 if (PUsesQ) { 1324 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1325 if (!isSecondInIteratorPair(P, QPairRange)) 1326 PairableInstUserMap->insert(VPPair(Q, P)); 1327 } 1328 if (QUsesP) { 1329 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1330 if (!isSecondInIteratorPair(Q, PPairRange)) 1331 PairableInstUserMap->insert(VPPair(P, Q)); 1332 } 1333 } 1334 1335 return (QUsesP && PUsesQ); 1336 } 1337 1338 // This function walks the use graph of current pairs to see if, starting 1339 // from P, the walk returns to P. 1340 bool BBVectorize::pairWillFormCycle(ValuePair P, 1341 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1342 DenseSet<ValuePair> &CurrentPairs) { 1343 DEBUG(if (DebugCycleCheck) 1344 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1345 << *P.second << "\n"); 1346 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1347 // contains non-direct associations. 1348 DenseSet<ValuePair> Visited; 1349 SmallVector<ValuePair, 32> Q; 1350 // General depth-first post-order traversal: 1351 Q.push_back(P); 1352 do { 1353 ValuePair QTop = Q.pop_back_val(); 1354 Visited.insert(QTop); 1355 1356 DEBUG(if (DebugCycleCheck) 1357 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1358 << *QTop.second << "\n"); 1359 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1360 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1361 C != QPairRange.second; ++C) { 1362 if (C->second == P) { 1363 DEBUG(dbgs() 1364 << "BBV: rejected to prevent non-trivial cycle formation: " 1365 << *C->first.first << " <-> " << *C->first.second << "\n"); 1366 return true; 1367 } 1368 1369 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1370 Q.push_back(C->second); 1371 } 1372 } while (!Q.empty()); 1373 1374 return false; 1375 } 1376 1377 // This function builds the initial tree of connected pairs with the 1378 // pair J at the root. 1379 void BBVectorize::buildInitialTreeFor( 1380 std::multimap<Value *, Value *> &CandidatePairs, 1381 std::vector<Value *> &PairableInsts, 1382 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1383 DenseSet<ValuePair> &PairableInstUsers, 1384 DenseMap<Value *, Value *> &ChosenPairs, 1385 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1386 // Each of these pairs is viewed as the root node of a Tree. The Tree 1387 // is then walked (depth-first). As this happens, we keep track of 1388 // the pairs that compose the Tree and the maximum depth of the Tree. 1389 SmallVector<ValuePairWithDepth, 32> Q; 1390 // General depth-first post-order traversal: 1391 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1392 do { 1393 ValuePairWithDepth QTop = Q.back(); 1394 1395 // Push each child onto the queue: 1396 bool MoreChildren = false; 1397 size_t MaxChildDepth = QTop.second; 1398 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1399 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1400 k != qtRange.second; ++k) { 1401 // Make sure that this child pair is still a candidate: 1402 bool IsStillCand = false; 1403 VPIteratorPair checkRange = 1404 CandidatePairs.equal_range(k->second.first); 1405 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1406 m != checkRange.second; ++m) { 1407 if (m->second == k->second.second) { 1408 IsStillCand = true; 1409 break; 1410 } 1411 } 1412 1413 if (IsStillCand) { 1414 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1415 if (C == Tree.end()) { 1416 size_t d = getDepthFactor(k->second.first); 1417 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1418 MoreChildren = true; 1419 } else { 1420 MaxChildDepth = std::max(MaxChildDepth, C->second); 1421 } 1422 } 1423 } 1424 1425 if (!MoreChildren) { 1426 // Record the current pair as part of the Tree: 1427 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1428 Q.pop_back(); 1429 } 1430 } while (!Q.empty()); 1431 } 1432 1433 // Given some initial tree, prune it by removing conflicting pairs (pairs 1434 // that cannot be simultaneously chosen for vectorization). 1435 void BBVectorize::pruneTreeFor( 1436 std::multimap<Value *, Value *> &CandidatePairs, 1437 std::vector<Value *> &PairableInsts, 1438 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1439 DenseSet<ValuePair> &PairableInstUsers, 1440 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1441 DenseMap<Value *, Value *> &ChosenPairs, 1442 DenseMap<ValuePair, size_t> &Tree, 1443 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1444 bool UseCycleCheck) { 1445 SmallVector<ValuePairWithDepth, 32> Q; 1446 // General depth-first post-order traversal: 1447 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1448 do { 1449 ValuePairWithDepth QTop = Q.pop_back_val(); 1450 PrunedTree.insert(QTop.first); 1451 1452 // Visit each child, pruning as necessary... 1453 DenseMap<ValuePair, size_t> BestChildren; 1454 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1455 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1456 K != QTopRange.second; ++K) { 1457 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1458 if (C == Tree.end()) continue; 1459 1460 // This child is in the Tree, now we need to make sure it is the 1461 // best of any conflicting children. There could be multiple 1462 // conflicting children, so first, determine if we're keeping 1463 // this child, then delete conflicting children as necessary. 1464 1465 // It is also necessary to guard against pairing-induced 1466 // dependencies. Consider instructions a .. x .. y .. b 1467 // such that (a,b) are to be fused and (x,y) are to be fused 1468 // but a is an input to x and b is an output from y. This 1469 // means that y cannot be moved after b but x must be moved 1470 // after b for (a,b) to be fused. In other words, after 1471 // fusing (a,b) we have y .. a/b .. x where y is an input 1472 // to a/b and x is an output to a/b: x and y can no longer 1473 // be legally fused. To prevent this condition, we must 1474 // make sure that a child pair added to the Tree is not 1475 // both an input and output of an already-selected pair. 1476 1477 // Pairing-induced dependencies can also form from more complicated 1478 // cycles. The pair vs. pair conflicts are easy to check, and so 1479 // that is done explicitly for "fast rejection", and because for 1480 // child vs. child conflicts, we may prefer to keep the current 1481 // pair in preference to the already-selected child. 1482 DenseSet<ValuePair> CurrentPairs; 1483 1484 bool CanAdd = true; 1485 for (DenseMap<ValuePair, size_t>::iterator C2 1486 = BestChildren.begin(), E2 = BestChildren.end(); 1487 C2 != E2; ++C2) { 1488 if (C2->first.first == C->first.first || 1489 C2->first.first == C->first.second || 1490 C2->first.second == C->first.first || 1491 C2->first.second == C->first.second || 1492 pairsConflict(C2->first, C->first, PairableInstUsers, 1493 UseCycleCheck ? &PairableInstUserMap : 0)) { 1494 if (C2->second >= C->second) { 1495 CanAdd = false; 1496 break; 1497 } 1498 1499 CurrentPairs.insert(C2->first); 1500 } 1501 } 1502 if (!CanAdd) continue; 1503 1504 // Even worse, this child could conflict with another node already 1505 // selected for the Tree. If that is the case, ignore this child. 1506 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1507 E2 = PrunedTree.end(); T != E2; ++T) { 1508 if (T->first == C->first.first || 1509 T->first == C->first.second || 1510 T->second == C->first.first || 1511 T->second == C->first.second || 1512 pairsConflict(*T, C->first, PairableInstUsers, 1513 UseCycleCheck ? &PairableInstUserMap : 0)) { 1514 CanAdd = false; 1515 break; 1516 } 1517 1518 CurrentPairs.insert(*T); 1519 } 1520 if (!CanAdd) continue; 1521 1522 // And check the queue too... 1523 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1524 E2 = Q.end(); C2 != E2; ++C2) { 1525 if (C2->first.first == C->first.first || 1526 C2->first.first == C->first.second || 1527 C2->first.second == C->first.first || 1528 C2->first.second == C->first.second || 1529 pairsConflict(C2->first, C->first, PairableInstUsers, 1530 UseCycleCheck ? &PairableInstUserMap : 0)) { 1531 CanAdd = false; 1532 break; 1533 } 1534 1535 CurrentPairs.insert(C2->first); 1536 } 1537 if (!CanAdd) continue; 1538 1539 // Last but not least, check for a conflict with any of the 1540 // already-chosen pairs. 1541 for (DenseMap<Value *, Value *>::iterator C2 = 1542 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1543 C2 != E2; ++C2) { 1544 if (pairsConflict(*C2, C->first, PairableInstUsers, 1545 UseCycleCheck ? &PairableInstUserMap : 0)) { 1546 CanAdd = false; 1547 break; 1548 } 1549 1550 CurrentPairs.insert(*C2); 1551 } 1552 if (!CanAdd) continue; 1553 1554 // To check for non-trivial cycles formed by the addition of the 1555 // current pair we've formed a list of all relevant pairs, now use a 1556 // graph walk to check for a cycle. We start from the current pair and 1557 // walk the use tree to see if we again reach the current pair. If we 1558 // do, then the current pair is rejected. 1559 1560 // FIXME: It may be more efficient to use a topological-ordering 1561 // algorithm to improve the cycle check. This should be investigated. 1562 if (UseCycleCheck && 1563 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1564 continue; 1565 1566 // This child can be added, but we may have chosen it in preference 1567 // to an already-selected child. Check for this here, and if a 1568 // conflict is found, then remove the previously-selected child 1569 // before adding this one in its place. 1570 for (DenseMap<ValuePair, size_t>::iterator C2 1571 = BestChildren.begin(); C2 != BestChildren.end();) { 1572 if (C2->first.first == C->first.first || 1573 C2->first.first == C->first.second || 1574 C2->first.second == C->first.first || 1575 C2->first.second == C->first.second || 1576 pairsConflict(C2->first, C->first, PairableInstUsers)) 1577 BestChildren.erase(C2++); 1578 else 1579 ++C2; 1580 } 1581 1582 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1583 } 1584 1585 for (DenseMap<ValuePair, size_t>::iterator C 1586 = BestChildren.begin(), E2 = BestChildren.end(); 1587 C != E2; ++C) { 1588 size_t DepthF = getDepthFactor(C->first.first); 1589 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1590 } 1591 } while (!Q.empty()); 1592 } 1593 1594 // This function finds the best tree of mututally-compatible connected 1595 // pairs, given the choice of root pairs as an iterator range. 1596 void BBVectorize::findBestTreeFor( 1597 std::multimap<Value *, Value *> &CandidatePairs, 1598 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1599 std::vector<Value *> &PairableInsts, 1600 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1601 DenseSet<ValuePair> &PairableInstUsers, 1602 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1603 DenseMap<Value *, Value *> &ChosenPairs, 1604 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1605 int &BestEffSize, VPIteratorPair ChoiceRange, 1606 bool UseCycleCheck) { 1607 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1608 J != ChoiceRange.second; ++J) { 1609 1610 // Before going any further, make sure that this pair does not 1611 // conflict with any already-selected pairs (see comment below 1612 // near the Tree pruning for more details). 1613 DenseSet<ValuePair> ChosenPairSet; 1614 bool DoesConflict = false; 1615 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1616 E = ChosenPairs.end(); C != E; ++C) { 1617 if (pairsConflict(*C, *J, PairableInstUsers, 1618 UseCycleCheck ? &PairableInstUserMap : 0)) { 1619 DoesConflict = true; 1620 break; 1621 } 1622 1623 ChosenPairSet.insert(*C); 1624 } 1625 if (DoesConflict) continue; 1626 1627 if (UseCycleCheck && 1628 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1629 continue; 1630 1631 DenseMap<ValuePair, size_t> Tree; 1632 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1633 PairableInstUsers, ChosenPairs, Tree, *J); 1634 1635 // Because we'll keep the child with the largest depth, the largest 1636 // depth is still the same in the unpruned Tree. 1637 size_t MaxDepth = Tree.lookup(*J); 1638 1639 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1640 << *J->first << " <-> " << *J->second << "} of depth " << 1641 MaxDepth << " and size " << Tree.size() << "\n"); 1642 1643 // At this point the Tree has been constructed, but, may contain 1644 // contradictory children (meaning that different children of 1645 // some tree node may be attempting to fuse the same instruction). 1646 // So now we walk the tree again, in the case of a conflict, 1647 // keep only the child with the largest depth. To break a tie, 1648 // favor the first child. 1649 1650 DenseSet<ValuePair> PrunedTree; 1651 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1652 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1653 PrunedTree, *J, UseCycleCheck); 1654 1655 int EffSize = 0; 1656 if (VTTI) { 1657 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1658 E = PrunedTree.end(); S != E; ++S) { 1659 if (getDepthFactor(S->first)) 1660 EffSize += CandidatePairCostSavings.find(*S)->second; 1661 } 1662 } else { 1663 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1664 E = PrunedTree.end(); S != E; ++S) 1665 EffSize += (int) getDepthFactor(S->first); 1666 } 1667 1668 DEBUG(if (DebugPairSelection) 1669 dbgs() << "BBV: found pruned Tree for pair {" 1670 << *J->first << " <-> " << *J->second << "} of depth " << 1671 MaxDepth << " and size " << PrunedTree.size() << 1672 " (effective size: " << EffSize << ")\n"); 1673 if (MaxDepth >= Config.ReqChainDepth && 1674 EffSize > 0 && EffSize > BestEffSize) { 1675 BestMaxDepth = MaxDepth; 1676 BestEffSize = EffSize; 1677 BestTree = PrunedTree; 1678 } 1679 } 1680 } 1681 1682 // Given the list of candidate pairs, this function selects those 1683 // that will be fused into vector instructions. 1684 void BBVectorize::choosePairs( 1685 std::multimap<Value *, Value *> &CandidatePairs, 1686 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1687 std::vector<Value *> &PairableInsts, 1688 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1689 DenseSet<ValuePair> &PairableInstUsers, 1690 DenseMap<Value *, Value *>& ChosenPairs) { 1691 bool UseCycleCheck = 1692 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1693 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1694 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1695 E = PairableInsts.end(); I != E; ++I) { 1696 // The number of possible pairings for this variable: 1697 size_t NumChoices = CandidatePairs.count(*I); 1698 if (!NumChoices) continue; 1699 1700 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1701 1702 // The best pair to choose and its tree: 1703 size_t BestMaxDepth = 0; 1704 int BestEffSize = 0; 1705 DenseSet<ValuePair> BestTree; 1706 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 1707 PairableInsts, ConnectedPairs, 1708 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1709 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1710 UseCycleCheck); 1711 1712 // A tree has been chosen (or not) at this point. If no tree was 1713 // chosen, then this instruction, I, cannot be paired (and is no longer 1714 // considered). 1715 1716 DEBUG(if (BestTree.size() > 0) 1717 dbgs() << "BBV: selected pairs in the best tree for: " 1718 << *cast<Instruction>(*I) << "\n"); 1719 1720 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1721 SE2 = BestTree.end(); S != SE2; ++S) { 1722 // Insert the members of this tree into the list of chosen pairs. 1723 ChosenPairs.insert(ValuePair(S->first, S->second)); 1724 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1725 *S->second << "\n"); 1726 1727 // Remove all candidate pairs that have values in the chosen tree. 1728 for (std::multimap<Value *, Value *>::iterator K = 1729 CandidatePairs.begin(); K != CandidatePairs.end();) { 1730 if (K->first == S->first || K->second == S->first || 1731 K->second == S->second || K->first == S->second) { 1732 // Don't remove the actual pair chosen so that it can be used 1733 // in subsequent tree selections. 1734 if (!(K->first == S->first && K->second == S->second)) 1735 CandidatePairs.erase(K++); 1736 else 1737 ++K; 1738 } else { 1739 ++K; 1740 } 1741 } 1742 } 1743 } 1744 1745 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 1746 } 1747 1748 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 1749 unsigned n = 0) { 1750 if (!I->hasName()) 1751 return ""; 1752 1753 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 1754 (n > 0 ? "." + utostr(n) : "")).str(); 1755 } 1756 1757 // Returns the value that is to be used as the pointer input to the vector 1758 // instruction that fuses I with J. 1759 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 1760 Instruction *I, Instruction *J, unsigned o) { 1761 Value *IPtr, *JPtr; 1762 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 1763 int64_t OffsetInElmts; 1764 1765 // Note: the analysis might fail here, that is why the pair order has 1766 // been precomputed (OffsetInElmts must be unused here). 1767 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 1768 IAddressSpace, JAddressSpace, 1769 OffsetInElmts, false); 1770 1771 // The pointer value is taken to be the one with the lowest offset. 1772 Value *VPtr = IPtr; 1773 1774 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 1775 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 1776 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1777 Type *VArgPtrType = PointerType::get(VArgType, 1778 cast<PointerType>(IPtr->getType())->getAddressSpace()); 1779 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 1780 /* insert before */ I); 1781 } 1782 1783 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 1784 unsigned MaskOffset, unsigned NumInElem, 1785 unsigned NumInElem1, unsigned IdxOffset, 1786 std::vector<Constant*> &Mask) { 1787 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 1788 for (unsigned v = 0; v < NumElem1; ++v) { 1789 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 1790 if (m < 0) { 1791 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 1792 } else { 1793 unsigned mm = m + (int) IdxOffset; 1794 if (m >= (int) NumInElem1) 1795 mm += (int) NumInElem; 1796 1797 Mask[v+MaskOffset] = 1798 ConstantInt::get(Type::getInt32Ty(Context), mm); 1799 } 1800 } 1801 } 1802 1803 // Returns the value that is to be used as the vector-shuffle mask to the 1804 // vector instruction that fuses I with J. 1805 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 1806 Instruction *I, Instruction *J) { 1807 // This is the shuffle mask. We need to append the second 1808 // mask to the first, and the numbers need to be adjusted. 1809 1810 Type *ArgTypeI = I->getType(); 1811 Type *ArgTypeJ = J->getType(); 1812 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1813 1814 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 1815 1816 // Get the total number of elements in the fused vector type. 1817 // By definition, this must equal the number of elements in 1818 // the final mask. 1819 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 1820 std::vector<Constant*> Mask(NumElem); 1821 1822 Type *OpTypeI = I->getOperand(0)->getType(); 1823 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 1824 Type *OpTypeJ = J->getOperand(0)->getType(); 1825 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 1826 1827 // The fused vector will be: 1828 // ----------------------------------------------------- 1829 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 1830 // ----------------------------------------------------- 1831 // from which we'll extract NumElem total elements (where the first NumElemI 1832 // of them come from the mask in I and the remainder come from the mask 1833 // in J. 1834 1835 // For the mask from the first pair... 1836 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 1837 0, Mask); 1838 1839 // For the mask from the second pair... 1840 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 1841 NumInElemI, Mask); 1842 1843 return ConstantVector::get(Mask); 1844 } 1845 1846 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 1847 Instruction *J, unsigned o, Value *&LOp, 1848 unsigned numElemL, 1849 Type *ArgTypeL, Type *ArgTypeH, 1850 bool IBeforeJ, unsigned IdxOff) { 1851 bool ExpandedIEChain = false; 1852 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 1853 // If we have a pure insertelement chain, then this can be rewritten 1854 // into a chain that directly builds the larger type. 1855 bool PureChain = true; 1856 InsertElementInst *LIENext = LIE; 1857 do { 1858 if (!isa<UndefValue>(LIENext->getOperand(0)) && 1859 !isa<InsertElementInst>(LIENext->getOperand(0))) { 1860 PureChain = false; 1861 break; 1862 } 1863 } while ((LIENext = 1864 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1865 1866 if (PureChain) { 1867 SmallVector<Value *, 8> VectElemts(numElemL, 1868 UndefValue::get(ArgTypeL->getScalarType())); 1869 InsertElementInst *LIENext = LIE; 1870 do { 1871 unsigned Idx = 1872 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 1873 VectElemts[Idx] = LIENext->getOperand(1); 1874 } while ((LIENext = 1875 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 1876 1877 LIENext = 0; 1878 Value *LIEPrev = UndefValue::get(ArgTypeH); 1879 for (unsigned i = 0; i < numElemL; ++i) { 1880 if (isa<UndefValue>(VectElemts[i])) continue; 1881 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 1882 ConstantInt::get(Type::getInt32Ty(Context), 1883 i + IdxOff), 1884 getReplacementName(IBeforeJ ? I : J, 1885 true, o, i+1)); 1886 LIENext->insertBefore(IBeforeJ ? J : I); 1887 LIEPrev = LIENext; 1888 } 1889 1890 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 1891 ExpandedIEChain = true; 1892 } 1893 } 1894 1895 return ExpandedIEChain; 1896 } 1897 1898 // Returns the value to be used as the specified operand of the vector 1899 // instruction that fuses I with J. 1900 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 1901 Instruction *J, unsigned o, bool IBeforeJ) { 1902 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 1903 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 1904 1905 // Compute the fused vector type for this operand 1906 Type *ArgTypeI = I->getOperand(o)->getType(); 1907 Type *ArgTypeJ = J->getOperand(o)->getType(); 1908 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 1909 1910 Instruction *L = I, *H = J; 1911 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 1912 1913 unsigned numElemL; 1914 if (ArgTypeL->isVectorTy()) 1915 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 1916 else 1917 numElemL = 1; 1918 1919 unsigned numElemH; 1920 if (ArgTypeH->isVectorTy()) 1921 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 1922 else 1923 numElemH = 1; 1924 1925 Value *LOp = L->getOperand(o); 1926 Value *HOp = H->getOperand(o); 1927 unsigned numElem = VArgType->getNumElements(); 1928 1929 // First, we check if we can reuse the "original" vector outputs (if these 1930 // exist). We might need a shuffle. 1931 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 1932 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 1933 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 1934 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 1935 1936 // FIXME: If we're fusing shuffle instructions, then we can't apply this 1937 // optimization. The input vectors to the shuffle might be a different 1938 // length from the shuffle outputs. Unfortunately, the replacement 1939 // shuffle mask has already been formed, and the mask entries are sensitive 1940 // to the sizes of the inputs. 1941 bool IsSizeChangeShuffle = 1942 isa<ShuffleVectorInst>(L) && 1943 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 1944 1945 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 1946 // We can have at most two unique vector inputs. 1947 bool CanUseInputs = true; 1948 Value *I1, *I2 = 0; 1949 if (LEE) { 1950 I1 = LEE->getOperand(0); 1951 } else { 1952 I1 = LSV->getOperand(0); 1953 I2 = LSV->getOperand(1); 1954 if (I2 == I1 || isa<UndefValue>(I2)) 1955 I2 = 0; 1956 } 1957 1958 if (HEE) { 1959 Value *I3 = HEE->getOperand(0); 1960 if (!I2 && I3 != I1) 1961 I2 = I3; 1962 else if (I3 != I1 && I3 != I2) 1963 CanUseInputs = false; 1964 } else { 1965 Value *I3 = HSV->getOperand(0); 1966 if (!I2 && I3 != I1) 1967 I2 = I3; 1968 else if (I3 != I1 && I3 != I2) 1969 CanUseInputs = false; 1970 1971 if (CanUseInputs) { 1972 Value *I4 = HSV->getOperand(1); 1973 if (!isa<UndefValue>(I4)) { 1974 if (!I2 && I4 != I1) 1975 I2 = I4; 1976 else if (I4 != I1 && I4 != I2) 1977 CanUseInputs = false; 1978 } 1979 } 1980 } 1981 1982 if (CanUseInputs) { 1983 unsigned LOpElem = 1984 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 1985 ->getNumElements(); 1986 unsigned HOpElem = 1987 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 1988 ->getNumElements(); 1989 1990 // We have one or two input vectors. We need to map each index of the 1991 // operands to the index of the original vector. 1992 SmallVector<std::pair<int, int>, 8> II(numElem); 1993 for (unsigned i = 0; i < numElemL; ++i) { 1994 int Idx, INum; 1995 if (LEE) { 1996 Idx = 1997 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 1998 INum = LEE->getOperand(0) == I1 ? 0 : 1; 1999 } else { 2000 Idx = LSV->getMaskValue(i); 2001 if (Idx < (int) LOpElem) { 2002 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2003 } else { 2004 Idx -= LOpElem; 2005 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2006 } 2007 } 2008 2009 II[i] = std::pair<int, int>(Idx, INum); 2010 } 2011 for (unsigned i = 0; i < numElemH; ++i) { 2012 int Idx, INum; 2013 if (HEE) { 2014 Idx = 2015 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2016 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2017 } else { 2018 Idx = HSV->getMaskValue(i); 2019 if (Idx < (int) HOpElem) { 2020 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2021 } else { 2022 Idx -= HOpElem; 2023 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2024 } 2025 } 2026 2027 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2028 } 2029 2030 // We now have an array which tells us from which index of which 2031 // input vector each element of the operand comes. 2032 VectorType *I1T = cast<VectorType>(I1->getType()); 2033 unsigned I1Elem = I1T->getNumElements(); 2034 2035 if (!I2) { 2036 // In this case there is only one underlying vector input. Check for 2037 // the trivial case where we can use the input directly. 2038 if (I1Elem == numElem) { 2039 bool ElemInOrder = true; 2040 for (unsigned i = 0; i < numElem; ++i) { 2041 if (II[i].first != (int) i && II[i].first != -1) { 2042 ElemInOrder = false; 2043 break; 2044 } 2045 } 2046 2047 if (ElemInOrder) 2048 return I1; 2049 } 2050 2051 // A shuffle is needed. 2052 std::vector<Constant *> Mask(numElem); 2053 for (unsigned i = 0; i < numElem; ++i) { 2054 int Idx = II[i].first; 2055 if (Idx == -1) 2056 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2057 else 2058 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2059 } 2060 2061 Instruction *S = 2062 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2063 ConstantVector::get(Mask), 2064 getReplacementName(IBeforeJ ? I : J, 2065 true, o)); 2066 S->insertBefore(IBeforeJ ? J : I); 2067 return S; 2068 } 2069 2070 VectorType *I2T = cast<VectorType>(I2->getType()); 2071 unsigned I2Elem = I2T->getNumElements(); 2072 2073 // This input comes from two distinct vectors. The first step is to 2074 // make sure that both vectors are the same length. If not, the 2075 // smaller one will need to grow before they can be shuffled together. 2076 if (I1Elem < I2Elem) { 2077 std::vector<Constant *> Mask(I2Elem); 2078 unsigned v = 0; 2079 for (; v < I1Elem; ++v) 2080 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2081 for (; v < I2Elem; ++v) 2082 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2083 2084 Instruction *NewI1 = 2085 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2086 ConstantVector::get(Mask), 2087 getReplacementName(IBeforeJ ? I : J, 2088 true, o, 1)); 2089 NewI1->insertBefore(IBeforeJ ? J : I); 2090 I1 = NewI1; 2091 I1T = I2T; 2092 I1Elem = I2Elem; 2093 } else if (I1Elem > I2Elem) { 2094 std::vector<Constant *> Mask(I1Elem); 2095 unsigned v = 0; 2096 for (; v < I2Elem; ++v) 2097 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2098 for (; v < I1Elem; ++v) 2099 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2100 2101 Instruction *NewI2 = 2102 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2103 ConstantVector::get(Mask), 2104 getReplacementName(IBeforeJ ? I : J, 2105 true, o, 1)); 2106 NewI2->insertBefore(IBeforeJ ? J : I); 2107 I2 = NewI2; 2108 I2T = I1T; 2109 I2Elem = I1Elem; 2110 } 2111 2112 // Now that both I1 and I2 are the same length we can shuffle them 2113 // together (and use the result). 2114 std::vector<Constant *> Mask(numElem); 2115 for (unsigned v = 0; v < numElem; ++v) { 2116 if (II[v].first == -1) { 2117 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2118 } else { 2119 int Idx = II[v].first + II[v].second * I1Elem; 2120 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2121 } 2122 } 2123 2124 Instruction *NewOp = 2125 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2126 getReplacementName(IBeforeJ ? I : J, true, o)); 2127 NewOp->insertBefore(IBeforeJ ? J : I); 2128 return NewOp; 2129 } 2130 } 2131 2132 Type *ArgType = ArgTypeL; 2133 if (numElemL < numElemH) { 2134 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2135 ArgTypeL, VArgType, IBeforeJ, 1)) { 2136 // This is another short-circuit case: we're combining a scalar into 2137 // a vector that is formed by an IE chain. We've just expanded the IE 2138 // chain, now insert the scalar and we're done. 2139 2140 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2141 getReplacementName(IBeforeJ ? I : J, true, o)); 2142 S->insertBefore(IBeforeJ ? J : I); 2143 return S; 2144 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2145 ArgTypeH, IBeforeJ)) { 2146 // The two vector inputs to the shuffle must be the same length, 2147 // so extend the smaller vector to be the same length as the larger one. 2148 Instruction *NLOp; 2149 if (numElemL > 1) { 2150 2151 std::vector<Constant *> Mask(numElemH); 2152 unsigned v = 0; 2153 for (; v < numElemL; ++v) 2154 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2155 for (; v < numElemH; ++v) 2156 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2157 2158 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2159 ConstantVector::get(Mask), 2160 getReplacementName(IBeforeJ ? I : J, 2161 true, o, 1)); 2162 } else { 2163 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2164 getReplacementName(IBeforeJ ? I : J, 2165 true, o, 1)); 2166 } 2167 2168 NLOp->insertBefore(IBeforeJ ? J : I); 2169 LOp = NLOp; 2170 } 2171 2172 ArgType = ArgTypeH; 2173 } else if (numElemL > numElemH) { 2174 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2175 ArgTypeH, VArgType, IBeforeJ)) { 2176 Instruction *S = 2177 InsertElementInst::Create(LOp, HOp, 2178 ConstantInt::get(Type::getInt32Ty(Context), 2179 numElemL), 2180 getReplacementName(IBeforeJ ? I : J, 2181 true, o)); 2182 S->insertBefore(IBeforeJ ? J : I); 2183 return S; 2184 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2185 ArgTypeL, IBeforeJ)) { 2186 Instruction *NHOp; 2187 if (numElemH > 1) { 2188 std::vector<Constant *> Mask(numElemL); 2189 unsigned v = 0; 2190 for (; v < numElemH; ++v) 2191 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2192 for (; v < numElemL; ++v) 2193 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2194 2195 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2196 ConstantVector::get(Mask), 2197 getReplacementName(IBeforeJ ? I : J, 2198 true, o, 1)); 2199 } else { 2200 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2201 getReplacementName(IBeforeJ ? I : J, 2202 true, o, 1)); 2203 } 2204 2205 NHOp->insertBefore(IBeforeJ ? J : I); 2206 HOp = NHOp; 2207 } 2208 } 2209 2210 if (ArgType->isVectorTy()) { 2211 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2212 std::vector<Constant*> Mask(numElem); 2213 for (unsigned v = 0; v < numElem; ++v) { 2214 unsigned Idx = v; 2215 // If the low vector was expanded, we need to skip the extra 2216 // undefined entries. 2217 if (v >= numElemL && numElemH > numElemL) 2218 Idx += (numElemH - numElemL); 2219 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2220 } 2221 2222 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2223 ConstantVector::get(Mask), 2224 getReplacementName(IBeforeJ ? I : J, true, o)); 2225 BV->insertBefore(IBeforeJ ? J : I); 2226 return BV; 2227 } 2228 2229 Instruction *BV1 = InsertElementInst::Create( 2230 UndefValue::get(VArgType), LOp, CV0, 2231 getReplacementName(IBeforeJ ? I : J, 2232 true, o, 1)); 2233 BV1->insertBefore(IBeforeJ ? J : I); 2234 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2235 getReplacementName(IBeforeJ ? I : J, 2236 true, o, 2)); 2237 BV2->insertBefore(IBeforeJ ? J : I); 2238 return BV2; 2239 } 2240 2241 // This function creates an array of values that will be used as the inputs 2242 // to the vector instruction that fuses I with J. 2243 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2244 Instruction *I, Instruction *J, 2245 SmallVector<Value *, 3> &ReplacedOperands, 2246 bool IBeforeJ) { 2247 unsigned NumOperands = I->getNumOperands(); 2248 2249 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2250 // Iterate backward so that we look at the store pointer 2251 // first and know whether or not we need to flip the inputs. 2252 2253 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2254 // This is the pointer for a load/store instruction. 2255 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2256 continue; 2257 } else if (isa<CallInst>(I)) { 2258 Function *F = cast<CallInst>(I)->getCalledFunction(); 2259 unsigned IID = F->getIntrinsicID(); 2260 if (o == NumOperands-1) { 2261 BasicBlock &BB = *I->getParent(); 2262 2263 Module *M = BB.getParent()->getParent(); 2264 Type *ArgTypeI = I->getType(); 2265 Type *ArgTypeJ = J->getType(); 2266 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2267 2268 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 2269 (Intrinsic::ID) IID, VArgType); 2270 continue; 2271 } else if (IID == Intrinsic::powi && o == 1) { 2272 // The second argument of powi is a single integer and we've already 2273 // checked that both arguments are equal. As a result, we just keep 2274 // I's second argument. 2275 ReplacedOperands[o] = I->getOperand(o); 2276 continue; 2277 } 2278 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2279 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2280 continue; 2281 } 2282 2283 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2284 } 2285 } 2286 2287 // This function creates two values that represent the outputs of the 2288 // original I and J instructions. These are generally vector shuffles 2289 // or extracts. In many cases, these will end up being unused and, thus, 2290 // eliminated by later passes. 2291 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2292 Instruction *J, Instruction *K, 2293 Instruction *&InsertionPt, 2294 Instruction *&K1, Instruction *&K2) { 2295 if (isa<StoreInst>(I)) { 2296 AA->replaceWithNewValue(I, K); 2297 AA->replaceWithNewValue(J, K); 2298 } else { 2299 Type *IType = I->getType(); 2300 Type *JType = J->getType(); 2301 2302 VectorType *VType = getVecTypeForPair(IType, JType); 2303 unsigned numElem = VType->getNumElements(); 2304 2305 unsigned numElemI, numElemJ; 2306 if (IType->isVectorTy()) 2307 numElemI = cast<VectorType>(IType)->getNumElements(); 2308 else 2309 numElemI = 1; 2310 2311 if (JType->isVectorTy()) 2312 numElemJ = cast<VectorType>(JType)->getNumElements(); 2313 else 2314 numElemJ = 1; 2315 2316 if (IType->isVectorTy()) { 2317 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2318 for (unsigned v = 0; v < numElemI; ++v) { 2319 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2320 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2321 } 2322 2323 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2324 ConstantVector::get( Mask1), 2325 getReplacementName(K, false, 1)); 2326 } else { 2327 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2328 K1 = ExtractElementInst::Create(K, CV0, 2329 getReplacementName(K, false, 1)); 2330 } 2331 2332 if (JType->isVectorTy()) { 2333 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2334 for (unsigned v = 0; v < numElemJ; ++v) { 2335 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2336 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2337 } 2338 2339 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2340 ConstantVector::get( Mask2), 2341 getReplacementName(K, false, 2)); 2342 } else { 2343 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2344 K2 = ExtractElementInst::Create(K, CV1, 2345 getReplacementName(K, false, 2)); 2346 } 2347 2348 K1->insertAfter(K); 2349 K2->insertAfter(K1); 2350 InsertionPt = K2; 2351 } 2352 } 2353 2354 // Move all uses of the function I (including pairing-induced uses) after J. 2355 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2356 std::multimap<Value *, Value *> &LoadMoveSet, 2357 Instruction *I, Instruction *J) { 2358 // Skip to the first instruction past I. 2359 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2360 2361 DenseSet<Value *> Users; 2362 AliasSetTracker WriteSet(*AA); 2363 for (; cast<Instruction>(L) != J; ++L) 2364 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 2365 2366 assert(cast<Instruction>(L) == J && 2367 "Tracking has not proceeded far enough to check for dependencies"); 2368 // If J is now in the use set of I, then trackUsesOfI will return true 2369 // and we have a dependency cycle (and the fusing operation must abort). 2370 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 2371 } 2372 2373 // Move all uses of the function I (including pairing-induced uses) after J. 2374 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2375 std::multimap<Value *, Value *> &LoadMoveSet, 2376 Instruction *&InsertionPt, 2377 Instruction *I, Instruction *J) { 2378 // Skip to the first instruction past I. 2379 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2380 2381 DenseSet<Value *> Users; 2382 AliasSetTracker WriteSet(*AA); 2383 for (; cast<Instruction>(L) != J;) { 2384 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 2385 // Move this instruction 2386 Instruction *InstToMove = L; ++L; 2387 2388 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2389 " to after " << *InsertionPt << "\n"); 2390 InstToMove->removeFromParent(); 2391 InstToMove->insertAfter(InsertionPt); 2392 InsertionPt = InstToMove; 2393 } else { 2394 ++L; 2395 } 2396 } 2397 } 2398 2399 // Collect all load instruction that are in the move set of a given first 2400 // pair member. These loads depend on the first instruction, I, and so need 2401 // to be moved after J (the second instruction) when the pair is fused. 2402 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2403 DenseMap<Value *, Value *> &ChosenPairs, 2404 std::multimap<Value *, Value *> &LoadMoveSet, 2405 Instruction *I) { 2406 // Skip to the first instruction past I. 2407 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2408 2409 DenseSet<Value *> Users; 2410 AliasSetTracker WriteSet(*AA); 2411 2412 // Note: We cannot end the loop when we reach J because J could be moved 2413 // farther down the use chain by another instruction pairing. Also, J 2414 // could be before I if this is an inverted input. 2415 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2416 if (trackUsesOfI(Users, WriteSet, I, L)) { 2417 if (L->mayReadFromMemory()) 2418 LoadMoveSet.insert(ValuePair(L, I)); 2419 } 2420 } 2421 } 2422 2423 // In cases where both load/stores and the computation of their pointers 2424 // are chosen for vectorization, we can end up in a situation where the 2425 // aliasing analysis starts returning different query results as the 2426 // process of fusing instruction pairs continues. Because the algorithm 2427 // relies on finding the same use trees here as were found earlier, we'll 2428 // need to precompute the necessary aliasing information here and then 2429 // manually update it during the fusion process. 2430 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2431 std::vector<Value *> &PairableInsts, 2432 DenseMap<Value *, Value *> &ChosenPairs, 2433 std::multimap<Value *, Value *> &LoadMoveSet) { 2434 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2435 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2436 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2437 if (P == ChosenPairs.end()) continue; 2438 2439 Instruction *I = cast<Instruction>(P->first); 2440 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 2441 } 2442 } 2443 2444 // When the first instruction in each pair is cloned, it will inherit its 2445 // parent's metadata. This metadata must be combined with that of the other 2446 // instruction in a safe way. 2447 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2448 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2449 K->getAllMetadataOtherThanDebugLoc(Metadata); 2450 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2451 unsigned Kind = Metadata[i].first; 2452 MDNode *JMD = J->getMetadata(Kind); 2453 MDNode *KMD = Metadata[i].second; 2454 2455 switch (Kind) { 2456 default: 2457 K->setMetadata(Kind, 0); // Remove unknown metadata 2458 break; 2459 case LLVMContext::MD_tbaa: 2460 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2461 break; 2462 case LLVMContext::MD_fpmath: 2463 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2464 break; 2465 } 2466 } 2467 } 2468 2469 // This function fuses the chosen instruction pairs into vector instructions, 2470 // taking care preserve any needed scalar outputs and, then, it reorders the 2471 // remaining instructions as needed (users of the first member of the pair 2472 // need to be moved to after the location of the second member of the pair 2473 // because the vector instruction is inserted in the location of the pair's 2474 // second member). 2475 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2476 std::vector<Value *> &PairableInsts, 2477 DenseMap<Value *, Value *> &ChosenPairs, 2478 DenseSet<ValuePair> &FixedOrderPairs, 2479 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2480 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2481 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2482 LLVMContext& Context = BB.getContext(); 2483 2484 // During the vectorization process, the order of the pairs to be fused 2485 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2486 // list. After a pair is fused, the flipped pair is removed from the list. 2487 DenseSet<ValuePair> FlippedPairs; 2488 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2489 E = ChosenPairs.end(); P != E; ++P) 2490 FlippedPairs.insert(ValuePair(P->second, P->first)); 2491 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2492 E = FlippedPairs.end(); P != E; ++P) 2493 ChosenPairs.insert(*P); 2494 2495 std::multimap<Value *, Value *> LoadMoveSet; 2496 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 2497 2498 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2499 2500 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2501 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2502 if (P == ChosenPairs.end()) { 2503 ++PI; 2504 continue; 2505 } 2506 2507 if (getDepthFactor(P->first) == 0) { 2508 // These instructions are not really fused, but are tracked as though 2509 // they are. Any case in which it would be interesting to fuse them 2510 // will be taken care of by InstCombine. 2511 --NumFusedOps; 2512 ++PI; 2513 continue; 2514 } 2515 2516 Instruction *I = cast<Instruction>(P->first), 2517 *J = cast<Instruction>(P->second); 2518 2519 DEBUG(dbgs() << "BBV: fusing: " << *I << 2520 " <-> " << *J << "\n"); 2521 2522 // Remove the pair and flipped pair from the list. 2523 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2524 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2525 ChosenPairs.erase(FP); 2526 ChosenPairs.erase(P); 2527 2528 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 2529 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2530 " <-> " << *J << 2531 " aborted because of non-trivial dependency cycle\n"); 2532 --NumFusedOps; 2533 ++PI; 2534 continue; 2535 } 2536 2537 // If the pair must have the other order, then flip it. 2538 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2539 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2540 // This pair does not have a fixed order, and so we might want to 2541 // flip it if that will yield fewer shuffles. We count the number 2542 // of dependencies connected via swaps, and those directly connected, 2543 // and flip the order if the number of swaps is greater. 2544 bool OrigOrder = true; 2545 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2546 if (IP.first == ConnectedPairDeps.end()) { 2547 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2548 OrigOrder = false; 2549 } 2550 2551 if (IP.first != ConnectedPairDeps.end()) { 2552 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2553 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2554 Q != IP.second; ++Q) { 2555 DenseMap<VPPair, unsigned>::iterator R = 2556 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2557 assert(R != PairConnectionTypes.end() && 2558 "Cannot find pair connection type"); 2559 if (R->second == PairConnectionDirect) 2560 ++NumDepsDirect; 2561 else if (R->second == PairConnectionSwap) 2562 ++NumDepsSwap; 2563 } 2564 2565 if (!OrigOrder) 2566 std::swap(NumDepsDirect, NumDepsSwap); 2567 2568 if (NumDepsSwap > NumDepsDirect) { 2569 FlipPairOrder = true; 2570 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2571 " <-> " << *J << "\n"); 2572 } 2573 } 2574 } 2575 2576 Instruction *L = I, *H = J; 2577 if (FlipPairOrder) 2578 std::swap(H, L); 2579 2580 // If the pair being fused uses the opposite order from that in the pair 2581 // connection map, then we need to flip the types. 2582 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2583 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2584 Q != IP.second; ++Q) { 2585 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2586 assert(R != PairConnectionTypes.end() && 2587 "Cannot find pair connection type"); 2588 if (R->second == PairConnectionDirect) 2589 R->second = PairConnectionSwap; 2590 else if (R->second == PairConnectionSwap) 2591 R->second = PairConnectionDirect; 2592 } 2593 2594 bool LBeforeH = !FlipPairOrder; 2595 unsigned NumOperands = I->getNumOperands(); 2596 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2597 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2598 LBeforeH); 2599 2600 // Make a copy of the original operation, change its type to the vector 2601 // type and replace its operands with the vector operands. 2602 Instruction *K = L->clone(); 2603 if (L->hasName()) 2604 K->takeName(L); 2605 else if (H->hasName()) 2606 K->takeName(H); 2607 2608 if (!isa<StoreInst>(K)) 2609 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 2610 2611 combineMetadata(K, H); 2612 2613 for (unsigned o = 0; o < NumOperands; ++o) 2614 K->setOperand(o, ReplacedOperands[o]); 2615 2616 K->insertAfter(J); 2617 2618 // Instruction insertion point: 2619 Instruction *InsertionPt = K; 2620 Instruction *K1 = 0, *K2 = 0; 2621 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 2622 2623 // The use tree of the first original instruction must be moved to after 2624 // the location of the second instruction. The entire use tree of the 2625 // first instruction is disjoint from the input tree of the second 2626 // (by definition), and so commutes with it. 2627 2628 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 2629 2630 if (!isa<StoreInst>(I)) { 2631 L->replaceAllUsesWith(K1); 2632 H->replaceAllUsesWith(K2); 2633 AA->replaceWithNewValue(L, K1); 2634 AA->replaceWithNewValue(H, K2); 2635 } 2636 2637 // Instructions that may read from memory may be in the load move set. 2638 // Once an instruction is fused, we no longer need its move set, and so 2639 // the values of the map never need to be updated. However, when a load 2640 // is fused, we need to merge the entries from both instructions in the 2641 // pair in case those instructions were in the move set of some other 2642 // yet-to-be-fused pair. The loads in question are the keys of the map. 2643 if (I->mayReadFromMemory()) { 2644 std::vector<ValuePair> NewSetMembers; 2645 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 2646 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 2647 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 2648 N != IPairRange.second; ++N) 2649 NewSetMembers.push_back(ValuePair(K, N->second)); 2650 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 2651 N != JPairRange.second; ++N) 2652 NewSetMembers.push_back(ValuePair(K, N->second)); 2653 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 2654 AE = NewSetMembers.end(); A != AE; ++A) 2655 LoadMoveSet.insert(*A); 2656 } 2657 2658 // Before removing I, set the iterator to the next instruction. 2659 PI = llvm::next(BasicBlock::iterator(I)); 2660 if (cast<Instruction>(PI) == J) 2661 ++PI; 2662 2663 SE->forgetValue(I); 2664 SE->forgetValue(J); 2665 I->eraseFromParent(); 2666 J->eraseFromParent(); 2667 2668 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 2669 BB << "\n"); 2670 } 2671 2672 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 2673 } 2674} 2675 2676char BBVectorize::ID = 0; 2677static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 2678INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2679INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2680INITIALIZE_PASS_DEPENDENCY(DominatorTree) 2681INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2682INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2683 2684BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 2685 return new BBVectorize(C); 2686} 2687 2688bool 2689llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 2690 BBVectorize BBVectorizer(P, C); 2691 return BBVectorizer.vectorizeBB(BB); 2692} 2693 2694//===----------------------------------------------------------------------===// 2695VectorizeConfig::VectorizeConfig() { 2696 VectorBits = ::VectorBits; 2697 VectorizeBools = !::NoBools; 2698 VectorizeInts = !::NoInts; 2699 VectorizeFloats = !::NoFloats; 2700 VectorizePointers = !::NoPointers; 2701 VectorizeCasts = !::NoCasts; 2702 VectorizeMath = !::NoMath; 2703 VectorizeFMA = !::NoFMA; 2704 VectorizeSelect = !::NoSelect; 2705 VectorizeCmp = !::NoCmp; 2706 VectorizeGEP = !::NoGEP; 2707 VectorizeMemOps = !::NoMemOps; 2708 AlignedOnly = ::AlignedOnly; 2709 ReqChainDepth= ::ReqChainDepth; 2710 SearchLimit = ::SearchLimit; 2711 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 2712 SplatBreaksChain = ::SplatBreaksChain; 2713 MaxInsts = ::MaxInsts; 2714 MaxIter = ::MaxIter; 2715 Pow2LenOnly = ::Pow2LenOnly; 2716 NoMemOpBoost = ::NoMemOpBoost; 2717 FastDep = ::FastDep; 2718} 2719