BBVectorize.cpp revision eaa8f5533f9f678fe3c56aec0201a34e46eaaf54
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/Transforms/Vectorize.h" 20#include "llvm/ADT/DenseMap.h" 21#include "llvm/ADT/DenseSet.h" 22#include "llvm/ADT/STLExtras.h" 23#include "llvm/ADT/SmallSet.h" 24#include "llvm/ADT/SmallVector.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/ADT/StringExtras.h" 27#include "llvm/Analysis/AliasAnalysis.h" 28#include "llvm/Analysis/AliasSetTracker.h" 29#include "llvm/Analysis/Dominators.h" 30#include "llvm/Analysis/ScalarEvolution.h" 31#include "llvm/Analysis/ScalarEvolutionExpressions.h" 32#include "llvm/Analysis/TargetTransformInfo.h" 33#include "llvm/Analysis/ValueTracking.h" 34#include "llvm/IR/Constants.h" 35#include "llvm/IR/DataLayout.h" 36#include "llvm/IR/DerivedTypes.h" 37#include "llvm/IR/Function.h" 38#include "llvm/IR/Instructions.h" 39#include "llvm/IR/IntrinsicInst.h" 40#include "llvm/IR/Intrinsics.h" 41#include "llvm/IR/LLVMContext.h" 42#include "llvm/IR/Metadata.h" 43#include "llvm/IR/Type.h" 44#include "llvm/Pass.h" 45#include "llvm/Support/CommandLine.h" 46#include "llvm/Support/Debug.h" 47#include "llvm/Support/ValueHandle.h" 48#include "llvm/Support/raw_ostream.h" 49#include "llvm/Transforms/Utils/Local.h" 50#include <algorithm> 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<bool> 62UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false), 63 cl::Hidden, cl::desc("Use the chain depth requirement with" 64 " target information")); 65 66static cl::opt<unsigned> 67SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 68 cl::desc("The maximum search distance for instruction pairs")); 69 70static cl::opt<bool> 71SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 72 cl::desc("Replicating one element to a pair breaks the chain")); 73 74static cl::opt<unsigned> 75VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 76 cl::desc("The size of the native vector registers")); 77 78static cl::opt<unsigned> 79MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 80 cl::desc("The maximum number of pairing iterations")); 81 82static cl::opt<bool> 83Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden, 84 cl::desc("Don't try to form non-2^n-length vectors")); 85 86static cl::opt<unsigned> 87MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 88 cl::desc("The maximum number of pairable instructions per group")); 89 90static cl::opt<unsigned> 91MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden, 92 cl::desc("The maximum number of candidate instruction pairs per group")); 93 94static cl::opt<unsigned> 95MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 96 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 97 " a full cycle check")); 98 99static cl::opt<bool> 100NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden, 101 cl::desc("Don't try to vectorize boolean (i1) values")); 102 103static cl::opt<bool> 104NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 105 cl::desc("Don't try to vectorize integer values")); 106 107static cl::opt<bool> 108NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 109 cl::desc("Don't try to vectorize floating-point values")); 110 111// FIXME: This should default to false once pointer vector support works. 112static cl::opt<bool> 113NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden, 114 cl::desc("Don't try to vectorize pointer values")); 115 116static cl::opt<bool> 117NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 118 cl::desc("Don't try to vectorize casting (conversion) operations")); 119 120static cl::opt<bool> 121NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 122 cl::desc("Don't try to vectorize floating-point math intrinsics")); 123 124static cl::opt<bool> 125NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 126 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 127 128static cl::opt<bool> 129NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 130 cl::desc("Don't try to vectorize select instructions")); 131 132static cl::opt<bool> 133NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 134 cl::desc("Don't try to vectorize comparison instructions")); 135 136static cl::opt<bool> 137NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 138 cl::desc("Don't try to vectorize getelementptr instructions")); 139 140static cl::opt<bool> 141NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 142 cl::desc("Don't try to vectorize loads and stores")); 143 144static cl::opt<bool> 145AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 146 cl::desc("Only generate aligned loads and stores")); 147 148static cl::opt<bool> 149NoMemOpBoost("bb-vectorize-no-mem-op-boost", 150 cl::init(false), cl::Hidden, 151 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 152 153static cl::opt<bool> 154FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 155 cl::desc("Use a fast instruction dependency analysis")); 156 157#ifndef NDEBUG 158static cl::opt<bool> 159DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 160 cl::init(false), cl::Hidden, 161 cl::desc("When debugging is enabled, output information on the" 162 " instruction-examination process")); 163static cl::opt<bool> 164DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 165 cl::init(false), cl::Hidden, 166 cl::desc("When debugging is enabled, output information on the" 167 " candidate-selection process")); 168static cl::opt<bool> 169DebugPairSelection("bb-vectorize-debug-pair-selection", 170 cl::init(false), cl::Hidden, 171 cl::desc("When debugging is enabled, output information on the" 172 " pair-selection process")); 173static cl::opt<bool> 174DebugCycleCheck("bb-vectorize-debug-cycle-check", 175 cl::init(false), cl::Hidden, 176 cl::desc("When debugging is enabled, output information on the" 177 " cycle-checking process")); 178 179static cl::opt<bool> 180PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair", 181 cl::init(false), cl::Hidden, 182 cl::desc("When debugging is enabled, dump the basic block after" 183 " every pair is fused")); 184#endif 185 186STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 187 188namespace { 189 struct BBVectorize : public BasicBlockPass { 190 static char ID; // Pass identification, replacement for typeid 191 192 const VectorizeConfig Config; 193 194 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 195 : BasicBlockPass(ID), Config(C) { 196 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 197 } 198 199 BBVectorize(Pass *P, const VectorizeConfig &C) 200 : BasicBlockPass(ID), Config(C) { 201 AA = &P->getAnalysis<AliasAnalysis>(); 202 DT = &P->getAnalysis<DominatorTree>(); 203 SE = &P->getAnalysis<ScalarEvolution>(); 204 TD = P->getAnalysisIfAvailable<DataLayout>(); 205 TTI = IgnoreTargetInfo ? 0 : &P->getAnalysis<TargetTransformInfo>(); 206 } 207 208 typedef std::pair<Value *, Value *> ValuePair; 209 typedef std::pair<ValuePair, int> ValuePairWithCost; 210 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 211 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 212 typedef std::pair<VPPair, unsigned> VPPairWithType; 213 214 AliasAnalysis *AA; 215 DominatorTree *DT; 216 ScalarEvolution *SE; 217 DataLayout *TD; 218 const TargetTransformInfo *TTI; 219 220 // FIXME: const correct? 221 222 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 223 224 bool getCandidatePairs(BasicBlock &BB, 225 BasicBlock::iterator &Start, 226 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 227 DenseSet<ValuePair> &FixedOrderPairs, 228 DenseMap<ValuePair, int> &CandidatePairCostSavings, 229 std::vector<Value *> &PairableInsts, bool NonPow2Len); 230 231 // FIXME: The current implementation does not account for pairs that 232 // are connected in multiple ways. For example: 233 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap) 234 enum PairConnectionType { 235 PairConnectionDirect, 236 PairConnectionSwap, 237 PairConnectionSplat 238 }; 239 240 void computeConnectedPairs( 241 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 242 DenseSet<ValuePair> &CandidatePairsSet, 243 std::vector<Value *> &PairableInsts, 244 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 245 DenseMap<VPPair, unsigned> &PairConnectionTypes); 246 247 void buildDepMap(BasicBlock &BB, 248 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 249 std::vector<Value *> &PairableInsts, 250 DenseSet<ValuePair> &PairableInstUsers); 251 252 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 253 DenseSet<ValuePair> &CandidatePairsSet, 254 DenseMap<ValuePair, int> &CandidatePairCostSavings, 255 std::vector<Value *> &PairableInsts, 256 DenseSet<ValuePair> &FixedOrderPairs, 257 DenseMap<VPPair, unsigned> &PairConnectionTypes, 258 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 259 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 260 DenseSet<ValuePair> &PairableInstUsers, 261 DenseMap<Value *, Value *>& ChosenPairs); 262 263 void fuseChosenPairs(BasicBlock &BB, 264 std::vector<Value *> &PairableInsts, 265 DenseMap<Value *, Value *>& ChosenPairs, 266 DenseSet<ValuePair> &FixedOrderPairs, 267 DenseMap<VPPair, unsigned> &PairConnectionTypes, 268 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 269 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps); 270 271 272 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 273 274 bool areInstsCompatible(Instruction *I, Instruction *J, 275 bool IsSimpleLoadStore, bool NonPow2Len, 276 int &CostSavings, int &FixedOrder); 277 278 bool trackUsesOfI(DenseSet<Value *> &Users, 279 AliasSetTracker &WriteSet, Instruction *I, 280 Instruction *J, bool UpdateUsers = true, 281 DenseSet<ValuePair> *LoadMoveSetPairs = 0); 282 283 void computePairsConnectedTo( 284 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 285 DenseSet<ValuePair> &CandidatePairsSet, 286 std::vector<Value *> &PairableInsts, 287 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 288 DenseMap<VPPair, unsigned> &PairConnectionTypes, 289 ValuePair P); 290 291 bool pairsConflict(ValuePair P, ValuePair Q, 292 DenseSet<ValuePair> &PairableInstUsers, 293 DenseMap<ValuePair, std::vector<ValuePair> > 294 *PairableInstUserMap = 0, 295 DenseSet<VPPair> *PairableInstUserPairSet = 0); 296 297 bool pairWillFormCycle(ValuePair P, 298 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers, 299 DenseSet<ValuePair> &CurrentPairs); 300 301 void pruneDAGFor( 302 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 303 std::vector<Value *> &PairableInsts, 304 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 305 DenseSet<ValuePair> &PairableInstUsers, 306 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 307 DenseSet<VPPair> &PairableInstUserPairSet, 308 DenseMap<Value *, Value *> &ChosenPairs, 309 DenseMap<ValuePair, size_t> &DAG, 310 DenseSet<ValuePair> &PrunedDAG, ValuePair J, 311 bool UseCycleCheck); 312 313 void buildInitialDAGFor( 314 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 315 DenseSet<ValuePair> &CandidatePairsSet, 316 std::vector<Value *> &PairableInsts, 317 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 318 DenseSet<ValuePair> &PairableInstUsers, 319 DenseMap<Value *, Value *> &ChosenPairs, 320 DenseMap<ValuePair, size_t> &DAG, ValuePair J); 321 322 void findBestDAGFor( 323 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 324 DenseSet<ValuePair> &CandidatePairsSet, 325 DenseMap<ValuePair, int> &CandidatePairCostSavings, 326 std::vector<Value *> &PairableInsts, 327 DenseSet<ValuePair> &FixedOrderPairs, 328 DenseMap<VPPair, unsigned> &PairConnectionTypes, 329 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 330 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 331 DenseSet<ValuePair> &PairableInstUsers, 332 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 333 DenseSet<VPPair> &PairableInstUserPairSet, 334 DenseMap<Value *, Value *> &ChosenPairs, 335 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth, 336 int &BestEffSize, Value *II, std::vector<Value *>&JJ, 337 bool UseCycleCheck); 338 339 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 340 Instruction *J, unsigned o); 341 342 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 343 unsigned MaskOffset, unsigned NumInElem, 344 unsigned NumInElem1, unsigned IdxOffset, 345 std::vector<Constant*> &Mask); 346 347 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 348 Instruction *J); 349 350 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 351 unsigned o, Value *&LOp, unsigned numElemL, 352 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, 353 unsigned IdxOff = 0); 354 355 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 356 Instruction *J, unsigned o, bool IBeforeJ); 357 358 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 359 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands, 360 bool IBeforeJ); 361 362 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 363 Instruction *J, Instruction *K, 364 Instruction *&InsertionPt, Instruction *&K1, 365 Instruction *&K2); 366 367 void collectPairLoadMoveSet(BasicBlock &BB, 368 DenseMap<Value *, Value *> &ChosenPairs, 369 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 370 DenseSet<ValuePair> &LoadMoveSetPairs, 371 Instruction *I); 372 373 void collectLoadMoveSet(BasicBlock &BB, 374 std::vector<Value *> &PairableInsts, 375 DenseMap<Value *, Value *> &ChosenPairs, 376 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 377 DenseSet<ValuePair> &LoadMoveSetPairs); 378 379 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 380 DenseSet<ValuePair> &LoadMoveSetPairs, 381 Instruction *I, Instruction *J); 382 383 void moveUsesOfIAfterJ(BasicBlock &BB, 384 DenseSet<ValuePair> &LoadMoveSetPairs, 385 Instruction *&InsertionPt, 386 Instruction *I, Instruction *J); 387 388 void combineMetadata(Instruction *K, const Instruction *J); 389 390 bool vectorizeBB(BasicBlock &BB) { 391 if (!DT->isReachableFromEntry(&BB)) { 392 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 393 " in " << BB.getParent()->getName() << "\n"); 394 return false; 395 } 396 397 DEBUG(if (TTI) dbgs() << "BBV: using target information\n"); 398 399 bool changed = false; 400 // Iterate a sufficient number of times to merge types of size 1 bit, 401 // then 2 bits, then 4, etc. up to half of the target vector width of the 402 // target vector register. 403 unsigned n = 1; 404 for (unsigned v = 2; 405 (TTI || v <= Config.VectorBits) && 406 (!Config.MaxIter || n <= Config.MaxIter); 407 v *= 2, ++n) { 408 DEBUG(dbgs() << "BBV: fusing loop #" << n << 409 " for " << BB.getName() << " in " << 410 BB.getParent()->getName() << "...\n"); 411 if (vectorizePairs(BB)) 412 changed = true; 413 else 414 break; 415 } 416 417 if (changed && !Pow2LenOnly) { 418 ++n; 419 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 420 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 421 n << " for " << BB.getName() << " in " << 422 BB.getParent()->getName() << "...\n"); 423 if (!vectorizePairs(BB, true)) break; 424 } 425 } 426 427 DEBUG(dbgs() << "BBV: done!\n"); 428 return changed; 429 } 430 431 virtual bool runOnBasicBlock(BasicBlock &BB) { 432 AA = &getAnalysis<AliasAnalysis>(); 433 DT = &getAnalysis<DominatorTree>(); 434 SE = &getAnalysis<ScalarEvolution>(); 435 TD = getAnalysisIfAvailable<DataLayout>(); 436 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>(); 437 438 return vectorizeBB(BB); 439 } 440 441 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 442 BasicBlockPass::getAnalysisUsage(AU); 443 AU.addRequired<AliasAnalysis>(); 444 AU.addRequired<DominatorTree>(); 445 AU.addRequired<ScalarEvolution>(); 446 AU.addRequired<TargetTransformInfo>(); 447 AU.addPreserved<AliasAnalysis>(); 448 AU.addPreserved<DominatorTree>(); 449 AU.addPreserved<ScalarEvolution>(); 450 AU.setPreservesCFG(); 451 } 452 453 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 454 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 455 "Cannot form vector from incompatible scalar types"); 456 Type *STy = ElemTy->getScalarType(); 457 458 unsigned numElem; 459 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 460 numElem = VTy->getNumElements(); 461 } else { 462 numElem = 1; 463 } 464 465 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 466 numElem += VTy->getNumElements(); 467 } else { 468 numElem += 1; 469 } 470 471 return VectorType::get(STy, numElem); 472 } 473 474 static inline void getInstructionTypes(Instruction *I, 475 Type *&T1, Type *&T2) { 476 if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 477 // For stores, it is the value type, not the pointer type that matters 478 // because the value is what will come from a vector register. 479 480 Value *IVal = SI->getValueOperand(); 481 T1 = IVal->getType(); 482 } else { 483 T1 = I->getType(); 484 } 485 486 if (CastInst *CI = dyn_cast<CastInst>(I)) 487 T2 = CI->getSrcTy(); 488 else 489 T2 = T1; 490 491 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 492 T2 = SI->getCondition()->getType(); 493 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { 494 T2 = SI->getOperand(0)->getType(); 495 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 496 T2 = CI->getOperand(0)->getType(); 497 } 498 } 499 500 // Returns the weight associated with the provided value. A chain of 501 // candidate pairs has a length given by the sum of the weights of its 502 // members (one weight per pair; the weight of each member of the pair 503 // is assumed to be the same). This length is then compared to the 504 // chain-length threshold to determine if a given chain is significant 505 // enough to be vectorized. The length is also used in comparing 506 // candidate chains where longer chains are considered to be better. 507 // Note: when this function returns 0, the resulting instructions are 508 // not actually fused. 509 inline size_t getDepthFactor(Value *V) { 510 // InsertElement and ExtractElement have a depth factor of zero. This is 511 // for two reasons: First, they cannot be usefully fused. Second, because 512 // the pass generates a lot of these, they can confuse the simple metric 513 // used to compare the dags in the next iteration. Thus, giving them a 514 // weight of zero allows the pass to essentially ignore them in 515 // subsequent iterations when looking for vectorization opportunities 516 // while still tracking dependency chains that flow through those 517 // instructions. 518 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 519 return 0; 520 521 // Give a load or store half of the required depth so that load/store 522 // pairs will vectorize. 523 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 524 return Config.ReqChainDepth/2; 525 526 return 1; 527 } 528 529 // Returns the cost of the provided instruction using TTI. 530 // This does not handle loads and stores. 531 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { 532 switch (Opcode) { 533 default: break; 534 case Instruction::GetElementPtr: 535 // We mark this instruction as zero-cost because scalar GEPs are usually 536 // lowered to the intruction addressing mode. At the moment we don't 537 // generate vector GEPs. 538 return 0; 539 case Instruction::Br: 540 return TTI->getCFInstrCost(Opcode); 541 case Instruction::PHI: 542 return 0; 543 case Instruction::Add: 544 case Instruction::FAdd: 545 case Instruction::Sub: 546 case Instruction::FSub: 547 case Instruction::Mul: 548 case Instruction::FMul: 549 case Instruction::UDiv: 550 case Instruction::SDiv: 551 case Instruction::FDiv: 552 case Instruction::URem: 553 case Instruction::SRem: 554 case Instruction::FRem: 555 case Instruction::Shl: 556 case Instruction::LShr: 557 case Instruction::AShr: 558 case Instruction::And: 559 case Instruction::Or: 560 case Instruction::Xor: 561 return TTI->getArithmeticInstrCost(Opcode, T1); 562 case Instruction::Select: 563 case Instruction::ICmp: 564 case Instruction::FCmp: 565 return TTI->getCmpSelInstrCost(Opcode, T1, T2); 566 case Instruction::ZExt: 567 case Instruction::SExt: 568 case Instruction::FPToUI: 569 case Instruction::FPToSI: 570 case Instruction::FPExt: 571 case Instruction::PtrToInt: 572 case Instruction::IntToPtr: 573 case Instruction::SIToFP: 574 case Instruction::UIToFP: 575 case Instruction::Trunc: 576 case Instruction::FPTrunc: 577 case Instruction::BitCast: 578 case Instruction::ShuffleVector: 579 return TTI->getCastInstrCost(Opcode, T1, T2); 580 } 581 582 return 1; 583 } 584 585 // This determines the relative offset of two loads or stores, returning 586 // true if the offset could be determined to be some constant value. 587 // For example, if OffsetInElmts == 1, then J accesses the memory directly 588 // after I; if OffsetInElmts == -1 then I accesses the memory 589 // directly after J. 590 bool getPairPtrInfo(Instruction *I, Instruction *J, 591 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 592 unsigned &IAddressSpace, unsigned &JAddressSpace, 593 int64_t &OffsetInElmts, bool ComputeOffset = true) { 594 OffsetInElmts = 0; 595 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 596 LoadInst *LJ = cast<LoadInst>(J); 597 IPtr = LI->getPointerOperand(); 598 JPtr = LJ->getPointerOperand(); 599 IAlignment = LI->getAlignment(); 600 JAlignment = LJ->getAlignment(); 601 IAddressSpace = LI->getPointerAddressSpace(); 602 JAddressSpace = LJ->getPointerAddressSpace(); 603 } else { 604 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 605 IPtr = SI->getPointerOperand(); 606 JPtr = SJ->getPointerOperand(); 607 IAlignment = SI->getAlignment(); 608 JAlignment = SJ->getAlignment(); 609 IAddressSpace = SI->getPointerAddressSpace(); 610 JAddressSpace = SJ->getPointerAddressSpace(); 611 } 612 613 if (!ComputeOffset) 614 return true; 615 616 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 617 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 618 619 // If this is a trivial offset, then we'll get something like 620 // 1*sizeof(type). With target data, which we need anyway, this will get 621 // constant folded into a number. 622 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 623 if (const SCEVConstant *ConstOffSCEV = 624 dyn_cast<SCEVConstant>(OffsetSCEV)) { 625 ConstantInt *IntOff = ConstOffSCEV->getValue(); 626 int64_t Offset = IntOff->getSExtValue(); 627 628 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 629 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 630 631 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 632 if (VTy != VTy2 && Offset < 0) { 633 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 634 OffsetInElmts = Offset/VTy2TSS; 635 return (abs64(Offset) % VTy2TSS) == 0; 636 } 637 638 OffsetInElmts = Offset/VTyTSS; 639 return (abs64(Offset) % VTyTSS) == 0; 640 } 641 642 return false; 643 } 644 645 // Returns true if the provided CallInst represents an intrinsic that can 646 // be vectorized. 647 bool isVectorizableIntrinsic(CallInst* I) { 648 Function *F = I->getCalledFunction(); 649 if (!F) return false; 650 651 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 652 if (!IID) return false; 653 654 switch(IID) { 655 default: 656 return false; 657 case Intrinsic::sqrt: 658 case Intrinsic::powi: 659 case Intrinsic::sin: 660 case Intrinsic::cos: 661 case Intrinsic::log: 662 case Intrinsic::log2: 663 case Intrinsic::log10: 664 case Intrinsic::exp: 665 case Intrinsic::exp2: 666 case Intrinsic::pow: 667 return Config.VectorizeMath; 668 case Intrinsic::fma: 669 case Intrinsic::fmuladd: 670 return Config.VectorizeFMA; 671 } 672 } 673 674 bool isPureIEChain(InsertElementInst *IE) { 675 InsertElementInst *IENext = IE; 676 do { 677 if (!isa<UndefValue>(IENext->getOperand(0)) && 678 !isa<InsertElementInst>(IENext->getOperand(0))) { 679 return false; 680 } 681 } while ((IENext = 682 dyn_cast<InsertElementInst>(IENext->getOperand(0)))); 683 684 return true; 685 } 686 }; 687 688 // This function implements one vectorization iteration on the provided 689 // basic block. It returns true if the block is changed. 690 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 691 bool ShouldContinue; 692 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 693 694 std::vector<Value *> AllPairableInsts; 695 DenseMap<Value *, Value *> AllChosenPairs; 696 DenseSet<ValuePair> AllFixedOrderPairs; 697 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 698 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs, 699 AllConnectedPairDeps; 700 701 do { 702 std::vector<Value *> PairableInsts; 703 DenseMap<Value *, std::vector<Value *> > CandidatePairs; 704 DenseSet<ValuePair> FixedOrderPairs; 705 DenseMap<ValuePair, int> CandidatePairCostSavings; 706 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 707 FixedOrderPairs, 708 CandidatePairCostSavings, 709 PairableInsts, NonPow2Len); 710 if (PairableInsts.empty()) continue; 711 712 // Build the candidate pair set for faster lookups. 713 DenseSet<ValuePair> CandidatePairsSet; 714 for (DenseMap<Value *, std::vector<Value *> >::iterator I = 715 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I) 716 for (std::vector<Value *>::iterator J = I->second.begin(), 717 JE = I->second.end(); J != JE; ++J) 718 CandidatePairsSet.insert(ValuePair(I->first, *J)); 719 720 // Now we have a map of all of the pairable instructions and we need to 721 // select the best possible pairing. A good pairing is one such that the 722 // users of the pair are also paired. This defines a (directed) forest 723 // over the pairs such that two pairs are connected iff the second pair 724 // uses the first. 725 726 // Note that it only matters that both members of the second pair use some 727 // element of the first pair (to allow for splatting). 728 729 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs, 730 ConnectedPairDeps; 731 DenseMap<VPPair, unsigned> PairConnectionTypes; 732 computeConnectedPairs(CandidatePairs, CandidatePairsSet, 733 PairableInsts, ConnectedPairs, PairConnectionTypes); 734 if (ConnectedPairs.empty()) continue; 735 736 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator 737 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 738 I != IE; ++I) 739 for (std::vector<ValuePair>::iterator J = I->second.begin(), 740 JE = I->second.end(); J != JE; ++J) 741 ConnectedPairDeps[*J].push_back(I->first); 742 743 // Build the pairable-instruction dependency map 744 DenseSet<ValuePair> PairableInstUsers; 745 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 746 747 // There is now a graph of the connected pairs. For each variable, pick 748 // the pairing with the largest dag meeting the depth requirement on at 749 // least one branch. Then select all pairings that are part of that dag 750 // and remove them from the list of available pairings and pairable 751 // variables. 752 753 DenseMap<Value *, Value *> ChosenPairs; 754 choosePairs(CandidatePairs, CandidatePairsSet, 755 CandidatePairCostSavings, 756 PairableInsts, FixedOrderPairs, PairConnectionTypes, 757 ConnectedPairs, ConnectedPairDeps, 758 PairableInstUsers, ChosenPairs); 759 760 if (ChosenPairs.empty()) continue; 761 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 762 PairableInsts.end()); 763 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 764 765 // Only for the chosen pairs, propagate information on fixed-order pairs, 766 // pair connections, and their types to the data structures used by the 767 // pair fusion procedures. 768 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 769 IE = ChosenPairs.end(); I != IE; ++I) { 770 if (FixedOrderPairs.count(*I)) 771 AllFixedOrderPairs.insert(*I); 772 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 773 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 774 775 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 776 J != IE; ++J) { 777 DenseMap<VPPair, unsigned>::iterator K = 778 PairConnectionTypes.find(VPPair(*I, *J)); 779 if (K != PairConnectionTypes.end()) { 780 AllPairConnectionTypes.insert(*K); 781 } else { 782 K = PairConnectionTypes.find(VPPair(*J, *I)); 783 if (K != PairConnectionTypes.end()) 784 AllPairConnectionTypes.insert(*K); 785 } 786 } 787 } 788 789 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator 790 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 791 I != IE; ++I) 792 for (std::vector<ValuePair>::iterator J = I->second.begin(), 793 JE = I->second.end(); J != JE; ++J) 794 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) { 795 AllConnectedPairs[I->first].push_back(*J); 796 AllConnectedPairDeps[*J].push_back(I->first); 797 } 798 } while (ShouldContinue); 799 800 if (AllChosenPairs.empty()) return false; 801 NumFusedOps += AllChosenPairs.size(); 802 803 // A set of pairs has now been selected. It is now necessary to replace the 804 // paired instructions with vector instructions. For this procedure each 805 // operand must be replaced with a vector operand. This vector is formed 806 // by using build_vector on the old operands. The replaced values are then 807 // replaced with a vector_extract on the result. Subsequent optimization 808 // passes should coalesce the build/extract combinations. 809 810 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 811 AllPairConnectionTypes, 812 AllConnectedPairs, AllConnectedPairDeps); 813 814 // It is important to cleanup here so that future iterations of this 815 // function have less work to do. 816 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 817 return true; 818 } 819 820 // This function returns true if the provided instruction is capable of being 821 // fused into a vector instruction. This determination is based only on the 822 // type and other attributes of the instruction. 823 bool BBVectorize::isInstVectorizable(Instruction *I, 824 bool &IsSimpleLoadStore) { 825 IsSimpleLoadStore = false; 826 827 if (CallInst *C = dyn_cast<CallInst>(I)) { 828 if (!isVectorizableIntrinsic(C)) 829 return false; 830 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 831 // Vectorize simple loads if possbile: 832 IsSimpleLoadStore = L->isSimple(); 833 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 834 return false; 835 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 836 // Vectorize simple stores if possbile: 837 IsSimpleLoadStore = S->isSimple(); 838 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 839 return false; 840 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 841 // We can vectorize casts, but not casts of pointer types, etc. 842 if (!Config.VectorizeCasts) 843 return false; 844 845 Type *SrcTy = C->getSrcTy(); 846 if (!SrcTy->isSingleValueType()) 847 return false; 848 849 Type *DestTy = C->getDestTy(); 850 if (!DestTy->isSingleValueType()) 851 return false; 852 } else if (isa<SelectInst>(I)) { 853 if (!Config.VectorizeSelect) 854 return false; 855 } else if (isa<CmpInst>(I)) { 856 if (!Config.VectorizeCmp) 857 return false; 858 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 859 if (!Config.VectorizeGEP) 860 return false; 861 862 // Currently, vector GEPs exist only with one index. 863 if (G->getNumIndices() != 1) 864 return false; 865 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 866 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 867 return false; 868 } 869 870 // We can't vectorize memory operations without target data 871 if (TD == 0 && IsSimpleLoadStore) 872 return false; 873 874 Type *T1, *T2; 875 getInstructionTypes(I, T1, T2); 876 877 // Not every type can be vectorized... 878 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 879 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 880 return false; 881 882 if (T1->getScalarSizeInBits() == 1) { 883 if (!Config.VectorizeBools) 884 return false; 885 } else { 886 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 887 return false; 888 } 889 890 if (T2->getScalarSizeInBits() == 1) { 891 if (!Config.VectorizeBools) 892 return false; 893 } else { 894 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 895 return false; 896 } 897 898 if (!Config.VectorizeFloats 899 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 900 return false; 901 902 // Don't vectorize target-specific types. 903 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 904 return false; 905 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 906 return false; 907 908 if ((!Config.VectorizePointers || TD == 0) && 909 (T1->getScalarType()->isPointerTy() || 910 T2->getScalarType()->isPointerTy())) 911 return false; 912 913 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 914 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 915 return false; 916 917 return true; 918 } 919 920 // This function returns true if the two provided instructions are compatible 921 // (meaning that they can be fused into a vector instruction). This assumes 922 // that I has already been determined to be vectorizable and that J is not 923 // in the use dag of I. 924 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 925 bool IsSimpleLoadStore, bool NonPow2Len, 926 int &CostSavings, int &FixedOrder) { 927 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 928 " <-> " << *J << "\n"); 929 930 CostSavings = 0; 931 FixedOrder = 0; 932 933 // Loads and stores can be merged if they have different alignments, 934 // but are otherwise the same. 935 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 936 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 937 return false; 938 939 Type *IT1, *IT2, *JT1, *JT2; 940 getInstructionTypes(I, IT1, IT2); 941 getInstructionTypes(J, JT1, JT2); 942 unsigned MaxTypeBits = std::max( 943 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 944 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 945 if (!TTI && MaxTypeBits > Config.VectorBits) 946 return false; 947 948 // FIXME: handle addsub-type operations! 949 950 if (IsSimpleLoadStore) { 951 Value *IPtr, *JPtr; 952 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 953 int64_t OffsetInElmts = 0; 954 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 955 IAddressSpace, JAddressSpace, 956 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 957 FixedOrder = (int) OffsetInElmts; 958 unsigned BottomAlignment = IAlignment; 959 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 960 961 Type *aTypeI = isa<StoreInst>(I) ? 962 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 963 Type *aTypeJ = isa<StoreInst>(J) ? 964 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 965 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 966 967 if (Config.AlignedOnly) { 968 // An aligned load or store is possible only if the instruction 969 // with the lower offset has an alignment suitable for the 970 // vector type. 971 972 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 973 if (BottomAlignment < VecAlignment) 974 return false; 975 } 976 977 if (TTI) { 978 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI, 979 IAlignment, IAddressSpace); 980 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ, 981 JAlignment, JAddressSpace); 982 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType, 983 BottomAlignment, 984 IAddressSpace); 985 986 ICost += TTI->getAddressComputationCost(aTypeI); 987 JCost += TTI->getAddressComputationCost(aTypeJ); 988 VCost += TTI->getAddressComputationCost(VType); 989 990 if (VCost > ICost + JCost) 991 return false; 992 993 // We don't want to fuse to a type that will be split, even 994 // if the two input types will also be split and there is no other 995 // associated cost. 996 unsigned VParts = TTI->getNumberOfParts(VType); 997 if (VParts > 1) 998 return false; 999 else if (!VParts && VCost == ICost + JCost) 1000 return false; 1001 1002 CostSavings = ICost + JCost - VCost; 1003 } 1004 } else { 1005 return false; 1006 } 1007 } else if (TTI) { 1008 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 1009 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 1010 Type *VT1 = getVecTypeForPair(IT1, JT1), 1011 *VT2 = getVecTypeForPair(IT2, JT2); 1012 1013 // Note that this procedure is incorrect for insert and extract element 1014 // instructions (because combining these often results in a shuffle), 1015 // but this cost is ignored (because insert and extract element 1016 // instructions are assigned a zero depth factor and are not really 1017 // fused in general). 1018 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 1019 1020 if (VCost > ICost + JCost) 1021 return false; 1022 1023 // We don't want to fuse to a type that will be split, even 1024 // if the two input types will also be split and there is no other 1025 // associated cost. 1026 unsigned VParts1 = TTI->getNumberOfParts(VT1), 1027 VParts2 = TTI->getNumberOfParts(VT2); 1028 if (VParts1 > 1 || VParts2 > 1) 1029 return false; 1030 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) 1031 return false; 1032 1033 CostSavings = ICost + JCost - VCost; 1034 } 1035 1036 // The powi intrinsic is special because only the first argument is 1037 // vectorized, the second arguments must be equal. 1038 CallInst *CI = dyn_cast<CallInst>(I); 1039 Function *FI; 1040 if (CI && (FI = CI->getCalledFunction())) { 1041 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID(); 1042 if (IID == Intrinsic::powi) { 1043 Value *A1I = CI->getArgOperand(1), 1044 *A1J = cast<CallInst>(J)->getArgOperand(1); 1045 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1046 *A1JSCEV = SE->getSCEV(A1J); 1047 return (A1ISCEV == A1JSCEV); 1048 } 1049 1050 if (IID && TTI) { 1051 SmallVector<Type*, 4> Tys; 1052 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 1053 Tys.push_back(CI->getArgOperand(i)->getType()); 1054 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys); 1055 1056 Tys.clear(); 1057 CallInst *CJ = cast<CallInst>(J); 1058 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i) 1059 Tys.push_back(CJ->getArgOperand(i)->getType()); 1060 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys); 1061 1062 Tys.clear(); 1063 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() && 1064 "Intrinsic argument counts differ"); 1065 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1066 if (IID == Intrinsic::powi && i == 1) 1067 Tys.push_back(CI->getArgOperand(i)->getType()); 1068 else 1069 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(), 1070 CJ->getArgOperand(i)->getType())); 1071 } 1072 1073 Type *RetTy = getVecTypeForPair(IT1, JT1); 1074 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys); 1075 1076 if (VCost > ICost + JCost) 1077 return false; 1078 1079 // We don't want to fuse to a type that will be split, even 1080 // if the two input types will also be split and there is no other 1081 // associated cost. 1082 unsigned RetParts = TTI->getNumberOfParts(RetTy); 1083 if (RetParts > 1) 1084 return false; 1085 else if (!RetParts && VCost == ICost + JCost) 1086 return false; 1087 1088 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1089 if (!Tys[i]->isVectorTy()) 1090 continue; 1091 1092 unsigned NumParts = TTI->getNumberOfParts(Tys[i]); 1093 if (NumParts > 1) 1094 return false; 1095 else if (!NumParts && VCost == ICost + JCost) 1096 return false; 1097 } 1098 1099 CostSavings = ICost + JCost - VCost; 1100 } 1101 } 1102 1103 return true; 1104 } 1105 1106 // Figure out whether or not J uses I and update the users and write-set 1107 // structures associated with I. Specifically, Users represents the set of 1108 // instructions that depend on I. WriteSet represents the set 1109 // of memory locations that are dependent on I. If UpdateUsers is true, 1110 // and J uses I, then Users is updated to contain J and WriteSet is updated 1111 // to contain any memory locations to which J writes. The function returns 1112 // true if J uses I. By default, alias analysis is used to determine 1113 // whether J reads from memory that overlaps with a location in WriteSet. 1114 // If LoadMoveSet is not null, then it is a previously-computed map 1115 // where the key is the memory-based user instruction and the value is 1116 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1117 // then the alias analysis is not used. This is necessary because this 1118 // function is called during the process of moving instructions during 1119 // vectorization and the results of the alias analysis are not stable during 1120 // that process. 1121 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1122 AliasSetTracker &WriteSet, Instruction *I, 1123 Instruction *J, bool UpdateUsers, 1124 DenseSet<ValuePair> *LoadMoveSetPairs) { 1125 bool UsesI = false; 1126 1127 // This instruction may already be marked as a user due, for example, to 1128 // being a member of a selected pair. 1129 if (Users.count(J)) 1130 UsesI = true; 1131 1132 if (!UsesI) 1133 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1134 JU != JE; ++JU) { 1135 Value *V = *JU; 1136 if (I == V || Users.count(V)) { 1137 UsesI = true; 1138 break; 1139 } 1140 } 1141 if (!UsesI && J->mayReadFromMemory()) { 1142 if (LoadMoveSetPairs) { 1143 UsesI = LoadMoveSetPairs->count(ValuePair(J, I)); 1144 } else { 1145 for (AliasSetTracker::iterator W = WriteSet.begin(), 1146 WE = WriteSet.end(); W != WE; ++W) { 1147 if (W->aliasesUnknownInst(J, *AA)) { 1148 UsesI = true; 1149 break; 1150 } 1151 } 1152 } 1153 } 1154 1155 if (UsesI && UpdateUsers) { 1156 if (J->mayWriteToMemory()) WriteSet.add(J); 1157 Users.insert(J); 1158 } 1159 1160 return UsesI; 1161 } 1162 1163 // This function iterates over all instruction pairs in the provided 1164 // basic block and collects all candidate pairs for vectorization. 1165 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1166 BasicBlock::iterator &Start, 1167 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1168 DenseSet<ValuePair> &FixedOrderPairs, 1169 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1170 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1171 size_t TotalPairs = 0; 1172 BasicBlock::iterator E = BB.end(); 1173 if (Start == E) return false; 1174 1175 bool ShouldContinue = false, IAfterStart = false; 1176 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1177 if (I == Start) IAfterStart = true; 1178 1179 bool IsSimpleLoadStore; 1180 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1181 1182 // Look for an instruction with which to pair instruction *I... 1183 DenseSet<Value *> Users; 1184 AliasSetTracker WriteSet(*AA); 1185 if (I->mayWriteToMemory()) WriteSet.add(I); 1186 1187 bool JAfterStart = IAfterStart; 1188 BasicBlock::iterator J = llvm::next(I); 1189 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1190 if (J == Start) JAfterStart = true; 1191 1192 // Determine if J uses I, if so, exit the loop. 1193 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1194 if (Config.FastDep) { 1195 // Note: For this heuristic to be effective, independent operations 1196 // must tend to be intermixed. This is likely to be true from some 1197 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1198 // but otherwise may require some kind of reordering pass. 1199 1200 // When using fast dependency analysis, 1201 // stop searching after first use: 1202 if (UsesI) break; 1203 } else { 1204 if (UsesI) continue; 1205 } 1206 1207 // J does not use I, and comes before the first use of I, so it can be 1208 // merged with I if the instructions are compatible. 1209 int CostSavings, FixedOrder; 1210 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1211 CostSavings, FixedOrder)) continue; 1212 1213 // J is a candidate for merging with I. 1214 if (!PairableInsts.size() || 1215 PairableInsts[PairableInsts.size()-1] != I) { 1216 PairableInsts.push_back(I); 1217 } 1218 1219 CandidatePairs[I].push_back(J); 1220 ++TotalPairs; 1221 if (TTI) 1222 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1223 CostSavings)); 1224 1225 if (FixedOrder == 1) 1226 FixedOrderPairs.insert(ValuePair(I, J)); 1227 else if (FixedOrder == -1) 1228 FixedOrderPairs.insert(ValuePair(J, I)); 1229 1230 // The next call to this function must start after the last instruction 1231 // selected during this invocation. 1232 if (JAfterStart) { 1233 Start = llvm::next(J); 1234 IAfterStart = JAfterStart = false; 1235 } 1236 1237 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1238 << *I << " <-> " << *J << " (cost savings: " << 1239 CostSavings << ")\n"); 1240 1241 // If we have already found too many pairs, break here and this function 1242 // will be called again starting after the last instruction selected 1243 // during this invocation. 1244 if (PairableInsts.size() >= Config.MaxInsts || 1245 TotalPairs >= Config.MaxPairs) { 1246 ShouldContinue = true; 1247 break; 1248 } 1249 } 1250 1251 if (ShouldContinue) 1252 break; 1253 } 1254 1255 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1256 << " instructions with candidate pairs\n"); 1257 1258 return ShouldContinue; 1259 } 1260 1261 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1262 // it looks for pairs such that both members have an input which is an 1263 // output of PI or PJ. 1264 void BBVectorize::computePairsConnectedTo( 1265 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1266 DenseSet<ValuePair> &CandidatePairsSet, 1267 std::vector<Value *> &PairableInsts, 1268 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1269 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1270 ValuePair P) { 1271 StoreInst *SI, *SJ; 1272 1273 // For each possible pairing for this variable, look at the uses of 1274 // the first value... 1275 for (Value::use_iterator I = P.first->use_begin(), 1276 E = P.first->use_end(); I != E; ++I) { 1277 if (isa<LoadInst>(*I)) { 1278 // A pair cannot be connected to a load because the load only takes one 1279 // operand (the address) and it is a scalar even after vectorization. 1280 continue; 1281 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1282 P.first == SI->getPointerOperand()) { 1283 // Similarly, a pair cannot be connected to a store through its 1284 // pointer operand. 1285 continue; 1286 } 1287 1288 // For each use of the first variable, look for uses of the second 1289 // variable... 1290 for (Value::use_iterator J = P.second->use_begin(), 1291 E2 = P.second->use_end(); J != E2; ++J) { 1292 if ((SJ = dyn_cast<StoreInst>(*J)) && 1293 P.second == SJ->getPointerOperand()) 1294 continue; 1295 1296 // Look for <I, J>: 1297 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1298 VPPair VP(P, ValuePair(*I, *J)); 1299 ConnectedPairs[VP.first].push_back(VP.second); 1300 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1301 } 1302 1303 // Look for <J, I>: 1304 if (CandidatePairsSet.count(ValuePair(*J, *I))) { 1305 VPPair VP(P, ValuePair(*J, *I)); 1306 ConnectedPairs[VP.first].push_back(VP.second); 1307 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1308 } 1309 } 1310 1311 if (Config.SplatBreaksChain) continue; 1312 // Look for cases where just the first value in the pair is used by 1313 // both members of another pair (splatting). 1314 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1315 if ((SJ = dyn_cast<StoreInst>(*J)) && 1316 P.first == SJ->getPointerOperand()) 1317 continue; 1318 1319 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1320 VPPair VP(P, ValuePair(*I, *J)); 1321 ConnectedPairs[VP.first].push_back(VP.second); 1322 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1323 } 1324 } 1325 } 1326 1327 if (Config.SplatBreaksChain) return; 1328 // Look for cases where just the second value in the pair is used by 1329 // both members of another pair (splatting). 1330 for (Value::use_iterator I = P.second->use_begin(), 1331 E = P.second->use_end(); I != E; ++I) { 1332 if (isa<LoadInst>(*I)) 1333 continue; 1334 else if ((SI = dyn_cast<StoreInst>(*I)) && 1335 P.second == SI->getPointerOperand()) 1336 continue; 1337 1338 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1339 if ((SJ = dyn_cast<StoreInst>(*J)) && 1340 P.second == SJ->getPointerOperand()) 1341 continue; 1342 1343 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1344 VPPair VP(P, ValuePair(*I, *J)); 1345 ConnectedPairs[VP.first].push_back(VP.second); 1346 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1347 } 1348 } 1349 } 1350 } 1351 1352 // This function figures out which pairs are connected. Two pairs are 1353 // connected if some output of the first pair forms an input to both members 1354 // of the second pair. 1355 void BBVectorize::computeConnectedPairs( 1356 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1357 DenseSet<ValuePair> &CandidatePairsSet, 1358 std::vector<Value *> &PairableInsts, 1359 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1360 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1361 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1362 PE = PairableInsts.end(); PI != PE; ++PI) { 1363 DenseMap<Value *, std::vector<Value *> >::iterator PP = 1364 CandidatePairs.find(*PI); 1365 if (PP == CandidatePairs.end()) 1366 continue; 1367 1368 for (std::vector<Value *>::iterator P = PP->second.begin(), 1369 E = PP->second.end(); P != E; ++P) 1370 computePairsConnectedTo(CandidatePairs, CandidatePairsSet, 1371 PairableInsts, ConnectedPairs, 1372 PairConnectionTypes, ValuePair(*PI, *P)); 1373 } 1374 1375 DEBUG(size_t TotalPairs = 0; 1376 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I = 1377 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I) 1378 TotalPairs += I->second.size(); 1379 dbgs() << "BBV: found " << TotalPairs 1380 << " pair connections.\n"); 1381 } 1382 1383 // This function builds a set of use tuples such that <A, B> is in the set 1384 // if B is in the use dag of A. If B is in the use dag of A, then B 1385 // depends on the output of A. 1386 void BBVectorize::buildDepMap( 1387 BasicBlock &BB, 1388 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1389 std::vector<Value *> &PairableInsts, 1390 DenseSet<ValuePair> &PairableInstUsers) { 1391 DenseSet<Value *> IsInPair; 1392 for (DenseMap<Value *, std::vector<Value *> >::iterator C = 1393 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) { 1394 IsInPair.insert(C->first); 1395 IsInPair.insert(C->second.begin(), C->second.end()); 1396 } 1397 1398 // Iterate through the basic block, recording all users of each 1399 // pairable instruction. 1400 1401 BasicBlock::iterator E = BB.end(), EL = 1402 BasicBlock::iterator(cast<Instruction>(PairableInsts.back())); 1403 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1404 if (IsInPair.find(I) == IsInPair.end()) continue; 1405 1406 DenseSet<Value *> Users; 1407 AliasSetTracker WriteSet(*AA); 1408 if (I->mayWriteToMemory()) WriteSet.add(I); 1409 1410 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) { 1411 (void) trackUsesOfI(Users, WriteSet, I, J); 1412 1413 if (J == EL) 1414 break; 1415 } 1416 1417 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1418 U != E; ++U) { 1419 if (IsInPair.find(*U) == IsInPair.end()) continue; 1420 PairableInstUsers.insert(ValuePair(I, *U)); 1421 } 1422 1423 if (I == EL) 1424 break; 1425 } 1426 } 1427 1428 // Returns true if an input to pair P is an output of pair Q and also an 1429 // input of pair Q is an output of pair P. If this is the case, then these 1430 // two pairs cannot be simultaneously fused. 1431 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1432 DenseSet<ValuePair> &PairableInstUsers, 1433 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap, 1434 DenseSet<VPPair> *PairableInstUserPairSet) { 1435 // Two pairs are in conflict if they are mutual Users of eachother. 1436 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1437 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1438 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1439 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1440 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1441 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1442 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1443 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1444 if (PairableInstUserMap) { 1445 // FIXME: The expensive part of the cycle check is not so much the cycle 1446 // check itself but this edge insertion procedure. This needs some 1447 // profiling and probably a different data structure. 1448 if (PUsesQ) { 1449 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second) 1450 (*PairableInstUserMap)[Q].push_back(P); 1451 } 1452 if (QUsesP) { 1453 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second) 1454 (*PairableInstUserMap)[P].push_back(Q); 1455 } 1456 } 1457 1458 return (QUsesP && PUsesQ); 1459 } 1460 1461 // This function walks the use graph of current pairs to see if, starting 1462 // from P, the walk returns to P. 1463 bool BBVectorize::pairWillFormCycle(ValuePair P, 1464 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 1465 DenseSet<ValuePair> &CurrentPairs) { 1466 DEBUG(if (DebugCycleCheck) 1467 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1468 << *P.second << "\n"); 1469 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1470 // contains non-direct associations. 1471 DenseSet<ValuePair> Visited; 1472 SmallVector<ValuePair, 32> Q; 1473 // General depth-first post-order traversal: 1474 Q.push_back(P); 1475 do { 1476 ValuePair QTop = Q.pop_back_val(); 1477 Visited.insert(QTop); 1478 1479 DEBUG(if (DebugCycleCheck) 1480 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1481 << *QTop.second << "\n"); 1482 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ = 1483 PairableInstUserMap.find(QTop); 1484 if (QQ == PairableInstUserMap.end()) 1485 continue; 1486 1487 for (std::vector<ValuePair>::iterator C = QQ->second.begin(), 1488 CE = QQ->second.end(); C != CE; ++C) { 1489 if (*C == P) { 1490 DEBUG(dbgs() 1491 << "BBV: rejected to prevent non-trivial cycle formation: " 1492 << QTop.first << " <-> " << C->second << "\n"); 1493 return true; 1494 } 1495 1496 if (CurrentPairs.count(*C) && !Visited.count(*C)) 1497 Q.push_back(*C); 1498 } 1499 } while (!Q.empty()); 1500 1501 return false; 1502 } 1503 1504 // This function builds the initial dag of connected pairs with the 1505 // pair J at the root. 1506 void BBVectorize::buildInitialDAGFor( 1507 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1508 DenseSet<ValuePair> &CandidatePairsSet, 1509 std::vector<Value *> &PairableInsts, 1510 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1511 DenseSet<ValuePair> &PairableInstUsers, 1512 DenseMap<Value *, Value *> &ChosenPairs, 1513 DenseMap<ValuePair, size_t> &DAG, ValuePair J) { 1514 // Each of these pairs is viewed as the root node of a DAG. The DAG 1515 // is then walked (depth-first). As this happens, we keep track of 1516 // the pairs that compose the DAG and the maximum depth of the DAG. 1517 SmallVector<ValuePairWithDepth, 32> Q; 1518 // General depth-first post-order traversal: 1519 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1520 do { 1521 ValuePairWithDepth QTop = Q.back(); 1522 1523 // Push each child onto the queue: 1524 bool MoreChildren = false; 1525 size_t MaxChildDepth = QTop.second; 1526 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ = 1527 ConnectedPairs.find(QTop.first); 1528 if (QQ != ConnectedPairs.end()) 1529 for (std::vector<ValuePair>::iterator k = QQ->second.begin(), 1530 ke = QQ->second.end(); k != ke; ++k) { 1531 // Make sure that this child pair is still a candidate: 1532 if (CandidatePairsSet.count(*k)) { 1533 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k); 1534 if (C == DAG.end()) { 1535 size_t d = getDepthFactor(k->first); 1536 Q.push_back(ValuePairWithDepth(*k, QTop.second+d)); 1537 MoreChildren = true; 1538 } else { 1539 MaxChildDepth = std::max(MaxChildDepth, C->second); 1540 } 1541 } 1542 } 1543 1544 if (!MoreChildren) { 1545 // Record the current pair as part of the DAG: 1546 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1547 Q.pop_back(); 1548 } 1549 } while (!Q.empty()); 1550 } 1551 1552 // Given some initial dag, prune it by removing conflicting pairs (pairs 1553 // that cannot be simultaneously chosen for vectorization). 1554 void BBVectorize::pruneDAGFor( 1555 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1556 std::vector<Value *> &PairableInsts, 1557 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1558 DenseSet<ValuePair> &PairableInstUsers, 1559 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 1560 DenseSet<VPPair> &PairableInstUserPairSet, 1561 DenseMap<Value *, Value *> &ChosenPairs, 1562 DenseMap<ValuePair, size_t> &DAG, 1563 DenseSet<ValuePair> &PrunedDAG, ValuePair J, 1564 bool UseCycleCheck) { 1565 SmallVector<ValuePairWithDepth, 32> Q; 1566 // General depth-first post-order traversal: 1567 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1568 do { 1569 ValuePairWithDepth QTop = Q.pop_back_val(); 1570 PrunedDAG.insert(QTop.first); 1571 1572 // Visit each child, pruning as necessary... 1573 SmallVector<ValuePairWithDepth, 8> BestChildren; 1574 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ = 1575 ConnectedPairs.find(QTop.first); 1576 if (QQ == ConnectedPairs.end()) 1577 continue; 1578 1579 for (std::vector<ValuePair>::iterator K = QQ->second.begin(), 1580 KE = QQ->second.end(); K != KE; ++K) { 1581 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K); 1582 if (C == DAG.end()) continue; 1583 1584 // This child is in the DAG, now we need to make sure it is the 1585 // best of any conflicting children. There could be multiple 1586 // conflicting children, so first, determine if we're keeping 1587 // this child, then delete conflicting children as necessary. 1588 1589 // It is also necessary to guard against pairing-induced 1590 // dependencies. Consider instructions a .. x .. y .. b 1591 // such that (a,b) are to be fused and (x,y) are to be fused 1592 // but a is an input to x and b is an output from y. This 1593 // means that y cannot be moved after b but x must be moved 1594 // after b for (a,b) to be fused. In other words, after 1595 // fusing (a,b) we have y .. a/b .. x where y is an input 1596 // to a/b and x is an output to a/b: x and y can no longer 1597 // be legally fused. To prevent this condition, we must 1598 // make sure that a child pair added to the DAG is not 1599 // both an input and output of an already-selected pair. 1600 1601 // Pairing-induced dependencies can also form from more complicated 1602 // cycles. The pair vs. pair conflicts are easy to check, and so 1603 // that is done explicitly for "fast rejection", and because for 1604 // child vs. child conflicts, we may prefer to keep the current 1605 // pair in preference to the already-selected child. 1606 DenseSet<ValuePair> CurrentPairs; 1607 1608 bool CanAdd = true; 1609 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 1610 = BestChildren.begin(), E2 = BestChildren.end(); 1611 C2 != E2; ++C2) { 1612 if (C2->first.first == C->first.first || 1613 C2->first.first == C->first.second || 1614 C2->first.second == C->first.first || 1615 C2->first.second == C->first.second || 1616 pairsConflict(C2->first, C->first, PairableInstUsers, 1617 UseCycleCheck ? &PairableInstUserMap : 0, 1618 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1619 if (C2->second >= C->second) { 1620 CanAdd = false; 1621 break; 1622 } 1623 1624 CurrentPairs.insert(C2->first); 1625 } 1626 } 1627 if (!CanAdd) continue; 1628 1629 // Even worse, this child could conflict with another node already 1630 // selected for the DAG. If that is the case, ignore this child. 1631 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(), 1632 E2 = PrunedDAG.end(); T != E2; ++T) { 1633 if (T->first == C->first.first || 1634 T->first == C->first.second || 1635 T->second == C->first.first || 1636 T->second == C->first.second || 1637 pairsConflict(*T, C->first, PairableInstUsers, 1638 UseCycleCheck ? &PairableInstUserMap : 0, 1639 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1640 CanAdd = false; 1641 break; 1642 } 1643 1644 CurrentPairs.insert(*T); 1645 } 1646 if (!CanAdd) continue; 1647 1648 // And check the queue too... 1649 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(), 1650 E2 = Q.end(); C2 != E2; ++C2) { 1651 if (C2->first.first == C->first.first || 1652 C2->first.first == C->first.second || 1653 C2->first.second == C->first.first || 1654 C2->first.second == C->first.second || 1655 pairsConflict(C2->first, C->first, PairableInstUsers, 1656 UseCycleCheck ? &PairableInstUserMap : 0, 1657 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1658 CanAdd = false; 1659 break; 1660 } 1661 1662 CurrentPairs.insert(C2->first); 1663 } 1664 if (!CanAdd) continue; 1665 1666 // Last but not least, check for a conflict with any of the 1667 // already-chosen pairs. 1668 for (DenseMap<Value *, Value *>::iterator C2 = 1669 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1670 C2 != E2; ++C2) { 1671 if (pairsConflict(*C2, C->first, PairableInstUsers, 1672 UseCycleCheck ? &PairableInstUserMap : 0, 1673 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1674 CanAdd = false; 1675 break; 1676 } 1677 1678 CurrentPairs.insert(*C2); 1679 } 1680 if (!CanAdd) continue; 1681 1682 // To check for non-trivial cycles formed by the addition of the 1683 // current pair we've formed a list of all relevant pairs, now use a 1684 // graph walk to check for a cycle. We start from the current pair and 1685 // walk the use dag to see if we again reach the current pair. If we 1686 // do, then the current pair is rejected. 1687 1688 // FIXME: It may be more efficient to use a topological-ordering 1689 // algorithm to improve the cycle check. This should be investigated. 1690 if (UseCycleCheck && 1691 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1692 continue; 1693 1694 // This child can be added, but we may have chosen it in preference 1695 // to an already-selected child. Check for this here, and if a 1696 // conflict is found, then remove the previously-selected child 1697 // before adding this one in its place. 1698 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 1699 = BestChildren.begin(); C2 != BestChildren.end();) { 1700 if (C2->first.first == C->first.first || 1701 C2->first.first == C->first.second || 1702 C2->first.second == C->first.first || 1703 C2->first.second == C->first.second || 1704 pairsConflict(C2->first, C->first, PairableInstUsers)) 1705 C2 = BestChildren.erase(C2); 1706 else 1707 ++C2; 1708 } 1709 1710 BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); 1711 } 1712 1713 for (SmallVectorImpl<ValuePairWithDepth>::iterator C 1714 = BestChildren.begin(), E2 = BestChildren.end(); 1715 C != E2; ++C) { 1716 size_t DepthF = getDepthFactor(C->first.first); 1717 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1718 } 1719 } while (!Q.empty()); 1720 } 1721 1722 // This function finds the best dag of mututally-compatible connected 1723 // pairs, given the choice of root pairs as an iterator range. 1724 void BBVectorize::findBestDAGFor( 1725 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 1726 DenseSet<ValuePair> &CandidatePairsSet, 1727 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1728 std::vector<Value *> &PairableInsts, 1729 DenseSet<ValuePair> &FixedOrderPairs, 1730 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1731 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 1732 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 1733 DenseSet<ValuePair> &PairableInstUsers, 1734 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap, 1735 DenseSet<VPPair> &PairableInstUserPairSet, 1736 DenseMap<Value *, Value *> &ChosenPairs, 1737 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth, 1738 int &BestEffSize, Value *II, std::vector<Value *>&JJ, 1739 bool UseCycleCheck) { 1740 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end(); 1741 J != JE; ++J) { 1742 ValuePair IJ(II, *J); 1743 if (!CandidatePairsSet.count(IJ)) 1744 continue; 1745 1746 // Before going any further, make sure that this pair does not 1747 // conflict with any already-selected pairs (see comment below 1748 // near the DAG pruning for more details). 1749 DenseSet<ValuePair> ChosenPairSet; 1750 bool DoesConflict = false; 1751 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1752 E = ChosenPairs.end(); C != E; ++C) { 1753 if (pairsConflict(*C, IJ, PairableInstUsers, 1754 UseCycleCheck ? &PairableInstUserMap : 0, 1755 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1756 DoesConflict = true; 1757 break; 1758 } 1759 1760 ChosenPairSet.insert(*C); 1761 } 1762 if (DoesConflict) continue; 1763 1764 if (UseCycleCheck && 1765 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet)) 1766 continue; 1767 1768 DenseMap<ValuePair, size_t> DAG; 1769 buildInitialDAGFor(CandidatePairs, CandidatePairsSet, 1770 PairableInsts, ConnectedPairs, 1771 PairableInstUsers, ChosenPairs, DAG, IJ); 1772 1773 // Because we'll keep the child with the largest depth, the largest 1774 // depth is still the same in the unpruned DAG. 1775 size_t MaxDepth = DAG.lookup(IJ); 1776 1777 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {" 1778 << *IJ.first << " <-> " << *IJ.second << "} of depth " << 1779 MaxDepth << " and size " << DAG.size() << "\n"); 1780 1781 // At this point the DAG has been constructed, but, may contain 1782 // contradictory children (meaning that different children of 1783 // some dag node may be attempting to fuse the same instruction). 1784 // So now we walk the dag again, in the case of a conflict, 1785 // keep only the child with the largest depth. To break a tie, 1786 // favor the first child. 1787 1788 DenseSet<ValuePair> PrunedDAG; 1789 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs, 1790 PairableInstUsers, PairableInstUserMap, 1791 PairableInstUserPairSet, 1792 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck); 1793 1794 int EffSize = 0; 1795 if (TTI) { 1796 DenseSet<Value *> PrunedDAGInstrs; 1797 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(), 1798 E = PrunedDAG.end(); S != E; ++S) { 1799 PrunedDAGInstrs.insert(S->first); 1800 PrunedDAGInstrs.insert(S->second); 1801 } 1802 1803 // The set of pairs that have already contributed to the total cost. 1804 DenseSet<ValuePair> IncomingPairs; 1805 1806 // If the cost model were perfect, this might not be necessary; but we 1807 // need to make sure that we don't get stuck vectorizing our own 1808 // shuffle chains. 1809 bool HasNontrivialInsts = false; 1810 1811 // The node weights represent the cost savings associated with 1812 // fusing the pair of instructions. 1813 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(), 1814 E = PrunedDAG.end(); S != E; ++S) { 1815 if (!isa<ShuffleVectorInst>(S->first) && 1816 !isa<InsertElementInst>(S->first) && 1817 !isa<ExtractElementInst>(S->first)) 1818 HasNontrivialInsts = true; 1819 1820 bool FlipOrder = false; 1821 1822 if (getDepthFactor(S->first)) { 1823 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1824 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1825 << *S->first << " <-> " << *S->second << "} = " << 1826 ESContrib << "\n"); 1827 EffSize += ESContrib; 1828 } 1829 1830 // The edge weights contribute in a negative sense: they represent 1831 // the cost of shuffles. 1832 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS = 1833 ConnectedPairDeps.find(*S); 1834 if (SS != ConnectedPairDeps.end()) { 1835 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1836 for (std::vector<ValuePair>::iterator T = SS->second.begin(), 1837 TE = SS->second.end(); T != TE; ++T) { 1838 VPPair Q(*S, *T); 1839 if (!PrunedDAG.count(Q.second)) 1840 continue; 1841 DenseMap<VPPair, unsigned>::iterator R = 1842 PairConnectionTypes.find(VPPair(Q.second, Q.first)); 1843 assert(R != PairConnectionTypes.end() && 1844 "Cannot find pair connection type"); 1845 if (R->second == PairConnectionDirect) 1846 ++NumDepsDirect; 1847 else if (R->second == PairConnectionSwap) 1848 ++NumDepsSwap; 1849 } 1850 1851 // If there are more swaps than direct connections, then 1852 // the pair order will be flipped during fusion. So the real 1853 // number of swaps is the minimum number. 1854 FlipOrder = !FixedOrderPairs.count(*S) && 1855 ((NumDepsSwap > NumDepsDirect) || 1856 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1857 1858 for (std::vector<ValuePair>::iterator T = SS->second.begin(), 1859 TE = SS->second.end(); T != TE; ++T) { 1860 VPPair Q(*S, *T); 1861 if (!PrunedDAG.count(Q.second)) 1862 continue; 1863 DenseMap<VPPair, unsigned>::iterator R = 1864 PairConnectionTypes.find(VPPair(Q.second, Q.first)); 1865 assert(R != PairConnectionTypes.end() && 1866 "Cannot find pair connection type"); 1867 Type *Ty1 = Q.second.first->getType(), 1868 *Ty2 = Q.second.second->getType(); 1869 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1870 if ((R->second == PairConnectionDirect && FlipOrder) || 1871 (R->second == PairConnectionSwap && !FlipOrder) || 1872 R->second == PairConnectionSplat) { 1873 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1874 VTy, VTy); 1875 1876 if (VTy->getVectorNumElements() == 2) { 1877 if (R->second == PairConnectionSplat) 1878 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1879 TargetTransformInfo::SK_Broadcast, VTy)); 1880 else 1881 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1882 TargetTransformInfo::SK_Reverse, VTy)); 1883 } 1884 1885 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1886 *Q.second.first << " <-> " << *Q.second.second << 1887 "} -> {" << 1888 *S->first << " <-> " << *S->second << "} = " << 1889 ESContrib << "\n"); 1890 EffSize -= ESContrib; 1891 } 1892 } 1893 } 1894 1895 // Compute the cost of outgoing edges. We assume that edges outgoing 1896 // to shuffles, inserts or extracts can be merged, and so contribute 1897 // no additional cost. 1898 if (!S->first->getType()->isVoidTy()) { 1899 Type *Ty1 = S->first->getType(), 1900 *Ty2 = S->second->getType(); 1901 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1902 1903 bool NeedsExtraction = false; 1904 for (Value::use_iterator I = S->first->use_begin(), 1905 IE = S->first->use_end(); I != IE; ++I) { 1906 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1907 // Shuffle can be folded if it has no other input 1908 if (isa<UndefValue>(SI->getOperand(1))) 1909 continue; 1910 } 1911 if (isa<ExtractElementInst>(*I)) 1912 continue; 1913 if (PrunedDAGInstrs.count(*I)) 1914 continue; 1915 NeedsExtraction = true; 1916 break; 1917 } 1918 1919 if (NeedsExtraction) { 1920 int ESContrib; 1921 if (Ty1->isVectorTy()) { 1922 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1923 Ty1, VTy); 1924 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1925 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1)); 1926 } else 1927 ESContrib = (int) TTI->getVectorInstrCost( 1928 Instruction::ExtractElement, VTy, 0); 1929 1930 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1931 *S->first << "} = " << ESContrib << "\n"); 1932 EffSize -= ESContrib; 1933 } 1934 1935 NeedsExtraction = false; 1936 for (Value::use_iterator I = S->second->use_begin(), 1937 IE = S->second->use_end(); I != IE; ++I) { 1938 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1939 // Shuffle can be folded if it has no other input 1940 if (isa<UndefValue>(SI->getOperand(1))) 1941 continue; 1942 } 1943 if (isa<ExtractElementInst>(*I)) 1944 continue; 1945 if (PrunedDAGInstrs.count(*I)) 1946 continue; 1947 NeedsExtraction = true; 1948 break; 1949 } 1950 1951 if (NeedsExtraction) { 1952 int ESContrib; 1953 if (Ty2->isVectorTy()) { 1954 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1955 Ty2, VTy); 1956 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1957 TargetTransformInfo::SK_ExtractSubvector, VTy, 1958 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2)); 1959 } else 1960 ESContrib = (int) TTI->getVectorInstrCost( 1961 Instruction::ExtractElement, VTy, 1); 1962 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1963 *S->second << "} = " << ESContrib << "\n"); 1964 EffSize -= ESContrib; 1965 } 1966 } 1967 1968 // Compute the cost of incoming edges. 1969 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 1970 Instruction *S1 = cast<Instruction>(S->first), 1971 *S2 = cast<Instruction>(S->second); 1972 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 1973 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 1974 1975 // Combining constants into vector constants (or small vector 1976 // constants into larger ones are assumed free). 1977 if (isa<Constant>(O1) && isa<Constant>(O2)) 1978 continue; 1979 1980 if (FlipOrder) 1981 std::swap(O1, O2); 1982 1983 ValuePair VP = ValuePair(O1, O2); 1984 ValuePair VPR = ValuePair(O2, O1); 1985 1986 // Internal edges are not handled here. 1987 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR)) 1988 continue; 1989 1990 Type *Ty1 = O1->getType(), 1991 *Ty2 = O2->getType(); 1992 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1993 1994 // Combining vector operations of the same type is also assumed 1995 // folded with other operations. 1996 if (Ty1 == Ty2) { 1997 // If both are insert elements, then both can be widened. 1998 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), 1999 *IEO2 = dyn_cast<InsertElementInst>(O2); 2000 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) 2001 continue; 2002 // If both are extract elements, and both have the same input 2003 // type, then they can be replaced with a shuffle 2004 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 2005 *EIO2 = dyn_cast<ExtractElementInst>(O2); 2006 if (EIO1 && EIO2 && 2007 EIO1->getOperand(0)->getType() == 2008 EIO2->getOperand(0)->getType()) 2009 continue; 2010 // If both are a shuffle with equal operand types and only two 2011 // unqiue operands, then they can be replaced with a single 2012 // shuffle 2013 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 2014 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 2015 if (SIO1 && SIO2 && 2016 SIO1->getOperand(0)->getType() == 2017 SIO2->getOperand(0)->getType()) { 2018 SmallSet<Value *, 4> SIOps; 2019 SIOps.insert(SIO1->getOperand(0)); 2020 SIOps.insert(SIO1->getOperand(1)); 2021 SIOps.insert(SIO2->getOperand(0)); 2022 SIOps.insert(SIO2->getOperand(1)); 2023 if (SIOps.size() <= 2) 2024 continue; 2025 } 2026 } 2027 2028 int ESContrib; 2029 // This pair has already been formed. 2030 if (IncomingPairs.count(VP)) { 2031 continue; 2032 } else if (IncomingPairs.count(VPR)) { 2033 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2034 VTy, VTy); 2035 2036 if (VTy->getVectorNumElements() == 2) 2037 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 2038 TargetTransformInfo::SK_Reverse, VTy)); 2039 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 2040 ESContrib = (int) TTI->getVectorInstrCost( 2041 Instruction::InsertElement, VTy, 0); 2042 ESContrib += (int) TTI->getVectorInstrCost( 2043 Instruction::InsertElement, VTy, 1); 2044 } else if (!Ty1->isVectorTy()) { 2045 // O1 needs to be inserted into a vector of size O2, and then 2046 // both need to be shuffled together. 2047 ESContrib = (int) TTI->getVectorInstrCost( 2048 Instruction::InsertElement, Ty2, 0); 2049 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2050 VTy, Ty2); 2051 } else if (!Ty2->isVectorTy()) { 2052 // O2 needs to be inserted into a vector of size O1, and then 2053 // both need to be shuffled together. 2054 ESContrib = (int) TTI->getVectorInstrCost( 2055 Instruction::InsertElement, Ty1, 0); 2056 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2057 VTy, Ty1); 2058 } else { 2059 Type *TyBig = Ty1, *TySmall = Ty2; 2060 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 2061 std::swap(TyBig, TySmall); 2062 2063 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2064 VTy, TyBig); 2065 if (TyBig != TySmall) 2066 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2067 TyBig, TySmall); 2068 } 2069 2070 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 2071 << *O1 << " <-> " << *O2 << "} = " << 2072 ESContrib << "\n"); 2073 EffSize -= ESContrib; 2074 IncomingPairs.insert(VP); 2075 } 2076 } 2077 } 2078 2079 if (!HasNontrivialInsts) { 2080 DEBUG(if (DebugPairSelection) dbgs() << 2081 "\tNo non-trivial instructions in DAG;" 2082 " override to zero effective size\n"); 2083 EffSize = 0; 2084 } 2085 } else { 2086 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(), 2087 E = PrunedDAG.end(); S != E; ++S) 2088 EffSize += (int) getDepthFactor(S->first); 2089 } 2090 2091 DEBUG(if (DebugPairSelection) 2092 dbgs() << "BBV: found pruned DAG for pair {" 2093 << *IJ.first << " <-> " << *IJ.second << "} of depth " << 2094 MaxDepth << " and size " << PrunedDAG.size() << 2095 " (effective size: " << EffSize << ")\n"); 2096 if (((TTI && !UseChainDepthWithTI) || 2097 MaxDepth >= Config.ReqChainDepth) && 2098 EffSize > 0 && EffSize > BestEffSize) { 2099 BestMaxDepth = MaxDepth; 2100 BestEffSize = EffSize; 2101 BestDAG = PrunedDAG; 2102 } 2103 } 2104 } 2105 2106 // Given the list of candidate pairs, this function selects those 2107 // that will be fused into vector instructions. 2108 void BBVectorize::choosePairs( 2109 DenseMap<Value *, std::vector<Value *> > &CandidatePairs, 2110 DenseSet<ValuePair> &CandidatePairsSet, 2111 DenseMap<ValuePair, int> &CandidatePairCostSavings, 2112 std::vector<Value *> &PairableInsts, 2113 DenseSet<ValuePair> &FixedOrderPairs, 2114 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2115 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 2116 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps, 2117 DenseSet<ValuePair> &PairableInstUsers, 2118 DenseMap<Value *, Value *>& ChosenPairs) { 2119 bool UseCycleCheck = 2120 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck; 2121 2122 DenseMap<Value *, std::vector<Value *> > CandidatePairs2; 2123 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(), 2124 E = CandidatePairsSet.end(); I != E; ++I) { 2125 std::vector<Value *> &JJ = CandidatePairs2[I->second]; 2126 if (JJ.empty()) JJ.reserve(32); 2127 JJ.push_back(I->first); 2128 } 2129 2130 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap; 2131 DenseSet<VPPair> PairableInstUserPairSet; 2132 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 2133 E = PairableInsts.end(); I != E; ++I) { 2134 // The number of possible pairings for this variable: 2135 size_t NumChoices = CandidatePairs.lookup(*I).size(); 2136 if (!NumChoices) continue; 2137 2138 std::vector<Value *> &JJ = CandidatePairs[*I]; 2139 2140 // The best pair to choose and its dag: 2141 size_t BestMaxDepth = 0; 2142 int BestEffSize = 0; 2143 DenseSet<ValuePair> BestDAG; 2144 findBestDAGFor(CandidatePairs, CandidatePairsSet, 2145 CandidatePairCostSavings, 2146 PairableInsts, FixedOrderPairs, PairConnectionTypes, 2147 ConnectedPairs, ConnectedPairDeps, 2148 PairableInstUsers, PairableInstUserMap, 2149 PairableInstUserPairSet, ChosenPairs, 2150 BestDAG, BestMaxDepth, BestEffSize, *I, JJ, 2151 UseCycleCheck); 2152 2153 if (BestDAG.empty()) 2154 continue; 2155 2156 // A dag has been chosen (or not) at this point. If no dag was 2157 // chosen, then this instruction, I, cannot be paired (and is no longer 2158 // considered). 2159 2160 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: " 2161 << *cast<Instruction>(*I) << "\n"); 2162 2163 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(), 2164 SE2 = BestDAG.end(); S != SE2; ++S) { 2165 // Insert the members of this dag into the list of chosen pairs. 2166 ChosenPairs.insert(ValuePair(S->first, S->second)); 2167 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 2168 *S->second << "\n"); 2169 2170 // Remove all candidate pairs that have values in the chosen dag. 2171 std::vector<Value *> &KK = CandidatePairs[S->first]; 2172 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end(); 2173 K != KE; ++K) { 2174 if (*K == S->second) 2175 continue; 2176 2177 CandidatePairsSet.erase(ValuePair(S->first, *K)); 2178 } 2179 2180 std::vector<Value *> &LL = CandidatePairs2[S->second]; 2181 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end(); 2182 L != LE; ++L) { 2183 if (*L == S->first) 2184 continue; 2185 2186 CandidatePairsSet.erase(ValuePair(*L, S->second)); 2187 } 2188 2189 std::vector<Value *> &MM = CandidatePairs[S->second]; 2190 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end(); 2191 M != ME; ++M) { 2192 assert(*M != S->first && "Flipped pair in candidate list?"); 2193 CandidatePairsSet.erase(ValuePair(S->second, *M)); 2194 } 2195 2196 std::vector<Value *> &NN = CandidatePairs2[S->first]; 2197 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end(); 2198 N != NE; ++N) { 2199 assert(*N != S->second && "Flipped pair in candidate list?"); 2200 CandidatePairsSet.erase(ValuePair(*N, S->first)); 2201 } 2202 } 2203 } 2204 2205 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2206 } 2207 2208 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2209 unsigned n = 0) { 2210 if (!I->hasName()) 2211 return ""; 2212 2213 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2214 (n > 0 ? "." + utostr(n) : "")).str(); 2215 } 2216 2217 // Returns the value that is to be used as the pointer input to the vector 2218 // instruction that fuses I with J. 2219 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2220 Instruction *I, Instruction *J, unsigned o) { 2221 Value *IPtr, *JPtr; 2222 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2223 int64_t OffsetInElmts; 2224 2225 // Note: the analysis might fail here, that is why the pair order has 2226 // been precomputed (OffsetInElmts must be unused here). 2227 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2228 IAddressSpace, JAddressSpace, 2229 OffsetInElmts, false); 2230 2231 // The pointer value is taken to be the one with the lowest offset. 2232 Value *VPtr = IPtr; 2233 2234 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 2235 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 2236 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2237 Type *VArgPtrType = PointerType::get(VArgType, 2238 cast<PointerType>(IPtr->getType())->getAddressSpace()); 2239 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2240 /* insert before */ I); 2241 } 2242 2243 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2244 unsigned MaskOffset, unsigned NumInElem, 2245 unsigned NumInElem1, unsigned IdxOffset, 2246 std::vector<Constant*> &Mask) { 2247 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 2248 for (unsigned v = 0; v < NumElem1; ++v) { 2249 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2250 if (m < 0) { 2251 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2252 } else { 2253 unsigned mm = m + (int) IdxOffset; 2254 if (m >= (int) NumInElem1) 2255 mm += (int) NumInElem; 2256 2257 Mask[v+MaskOffset] = 2258 ConstantInt::get(Type::getInt32Ty(Context), mm); 2259 } 2260 } 2261 } 2262 2263 // Returns the value that is to be used as the vector-shuffle mask to the 2264 // vector instruction that fuses I with J. 2265 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2266 Instruction *I, Instruction *J) { 2267 // This is the shuffle mask. We need to append the second 2268 // mask to the first, and the numbers need to be adjusted. 2269 2270 Type *ArgTypeI = I->getType(); 2271 Type *ArgTypeJ = J->getType(); 2272 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2273 2274 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 2275 2276 // Get the total number of elements in the fused vector type. 2277 // By definition, this must equal the number of elements in 2278 // the final mask. 2279 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 2280 std::vector<Constant*> Mask(NumElem); 2281 2282 Type *OpTypeI = I->getOperand(0)->getType(); 2283 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 2284 Type *OpTypeJ = J->getOperand(0)->getType(); 2285 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 2286 2287 // The fused vector will be: 2288 // ----------------------------------------------------- 2289 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2290 // ----------------------------------------------------- 2291 // from which we'll extract NumElem total elements (where the first NumElemI 2292 // of them come from the mask in I and the remainder come from the mask 2293 // in J. 2294 2295 // For the mask from the first pair... 2296 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2297 0, Mask); 2298 2299 // For the mask from the second pair... 2300 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2301 NumInElemI, Mask); 2302 2303 return ConstantVector::get(Mask); 2304 } 2305 2306 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2307 Instruction *J, unsigned o, Value *&LOp, 2308 unsigned numElemL, 2309 Type *ArgTypeL, Type *ArgTypeH, 2310 bool IBeforeJ, unsigned IdxOff) { 2311 bool ExpandedIEChain = false; 2312 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2313 // If we have a pure insertelement chain, then this can be rewritten 2314 // into a chain that directly builds the larger type. 2315 if (isPureIEChain(LIE)) { 2316 SmallVector<Value *, 8> VectElemts(numElemL, 2317 UndefValue::get(ArgTypeL->getScalarType())); 2318 InsertElementInst *LIENext = LIE; 2319 do { 2320 unsigned Idx = 2321 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2322 VectElemts[Idx] = LIENext->getOperand(1); 2323 } while ((LIENext = 2324 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2325 2326 LIENext = 0; 2327 Value *LIEPrev = UndefValue::get(ArgTypeH); 2328 for (unsigned i = 0; i < numElemL; ++i) { 2329 if (isa<UndefValue>(VectElemts[i])) continue; 2330 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2331 ConstantInt::get(Type::getInt32Ty(Context), 2332 i + IdxOff), 2333 getReplacementName(IBeforeJ ? I : J, 2334 true, o, i+1)); 2335 LIENext->insertBefore(IBeforeJ ? J : I); 2336 LIEPrev = LIENext; 2337 } 2338 2339 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2340 ExpandedIEChain = true; 2341 } 2342 } 2343 2344 return ExpandedIEChain; 2345 } 2346 2347 // Returns the value to be used as the specified operand of the vector 2348 // instruction that fuses I with J. 2349 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2350 Instruction *J, unsigned o, bool IBeforeJ) { 2351 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2352 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2353 2354 // Compute the fused vector type for this operand 2355 Type *ArgTypeI = I->getOperand(o)->getType(); 2356 Type *ArgTypeJ = J->getOperand(o)->getType(); 2357 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2358 2359 Instruction *L = I, *H = J; 2360 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2361 2362 unsigned numElemL; 2363 if (ArgTypeL->isVectorTy()) 2364 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 2365 else 2366 numElemL = 1; 2367 2368 unsigned numElemH; 2369 if (ArgTypeH->isVectorTy()) 2370 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 2371 else 2372 numElemH = 1; 2373 2374 Value *LOp = L->getOperand(o); 2375 Value *HOp = H->getOperand(o); 2376 unsigned numElem = VArgType->getNumElements(); 2377 2378 // First, we check if we can reuse the "original" vector outputs (if these 2379 // exist). We might need a shuffle. 2380 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2381 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2382 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2383 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2384 2385 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2386 // optimization. The input vectors to the shuffle might be a different 2387 // length from the shuffle outputs. Unfortunately, the replacement 2388 // shuffle mask has already been formed, and the mask entries are sensitive 2389 // to the sizes of the inputs. 2390 bool IsSizeChangeShuffle = 2391 isa<ShuffleVectorInst>(L) && 2392 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2393 2394 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2395 // We can have at most two unique vector inputs. 2396 bool CanUseInputs = true; 2397 Value *I1, *I2 = 0; 2398 if (LEE) { 2399 I1 = LEE->getOperand(0); 2400 } else { 2401 I1 = LSV->getOperand(0); 2402 I2 = LSV->getOperand(1); 2403 if (I2 == I1 || isa<UndefValue>(I2)) 2404 I2 = 0; 2405 } 2406 2407 if (HEE) { 2408 Value *I3 = HEE->getOperand(0); 2409 if (!I2 && I3 != I1) 2410 I2 = I3; 2411 else if (I3 != I1 && I3 != I2) 2412 CanUseInputs = false; 2413 } else { 2414 Value *I3 = HSV->getOperand(0); 2415 if (!I2 && I3 != I1) 2416 I2 = I3; 2417 else if (I3 != I1 && I3 != I2) 2418 CanUseInputs = false; 2419 2420 if (CanUseInputs) { 2421 Value *I4 = HSV->getOperand(1); 2422 if (!isa<UndefValue>(I4)) { 2423 if (!I2 && I4 != I1) 2424 I2 = I4; 2425 else if (I4 != I1 && I4 != I2) 2426 CanUseInputs = false; 2427 } 2428 } 2429 } 2430 2431 if (CanUseInputs) { 2432 unsigned LOpElem = 2433 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2434 ->getNumElements(); 2435 unsigned HOpElem = 2436 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2437 ->getNumElements(); 2438 2439 // We have one or two input vectors. We need to map each index of the 2440 // operands to the index of the original vector. 2441 SmallVector<std::pair<int, int>, 8> II(numElem); 2442 for (unsigned i = 0; i < numElemL; ++i) { 2443 int Idx, INum; 2444 if (LEE) { 2445 Idx = 2446 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2447 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2448 } else { 2449 Idx = LSV->getMaskValue(i); 2450 if (Idx < (int) LOpElem) { 2451 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2452 } else { 2453 Idx -= LOpElem; 2454 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2455 } 2456 } 2457 2458 II[i] = std::pair<int, int>(Idx, INum); 2459 } 2460 for (unsigned i = 0; i < numElemH; ++i) { 2461 int Idx, INum; 2462 if (HEE) { 2463 Idx = 2464 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2465 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2466 } else { 2467 Idx = HSV->getMaskValue(i); 2468 if (Idx < (int) HOpElem) { 2469 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2470 } else { 2471 Idx -= HOpElem; 2472 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2473 } 2474 } 2475 2476 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2477 } 2478 2479 // We now have an array which tells us from which index of which 2480 // input vector each element of the operand comes. 2481 VectorType *I1T = cast<VectorType>(I1->getType()); 2482 unsigned I1Elem = I1T->getNumElements(); 2483 2484 if (!I2) { 2485 // In this case there is only one underlying vector input. Check for 2486 // the trivial case where we can use the input directly. 2487 if (I1Elem == numElem) { 2488 bool ElemInOrder = true; 2489 for (unsigned i = 0; i < numElem; ++i) { 2490 if (II[i].first != (int) i && II[i].first != -1) { 2491 ElemInOrder = false; 2492 break; 2493 } 2494 } 2495 2496 if (ElemInOrder) 2497 return I1; 2498 } 2499 2500 // A shuffle is needed. 2501 std::vector<Constant *> Mask(numElem); 2502 for (unsigned i = 0; i < numElem; ++i) { 2503 int Idx = II[i].first; 2504 if (Idx == -1) 2505 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2506 else 2507 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2508 } 2509 2510 Instruction *S = 2511 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2512 ConstantVector::get(Mask), 2513 getReplacementName(IBeforeJ ? I : J, 2514 true, o)); 2515 S->insertBefore(IBeforeJ ? J : I); 2516 return S; 2517 } 2518 2519 VectorType *I2T = cast<VectorType>(I2->getType()); 2520 unsigned I2Elem = I2T->getNumElements(); 2521 2522 // This input comes from two distinct vectors. The first step is to 2523 // make sure that both vectors are the same length. If not, the 2524 // smaller one will need to grow before they can be shuffled together. 2525 if (I1Elem < I2Elem) { 2526 std::vector<Constant *> Mask(I2Elem); 2527 unsigned v = 0; 2528 for (; v < I1Elem; ++v) 2529 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2530 for (; v < I2Elem; ++v) 2531 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2532 2533 Instruction *NewI1 = 2534 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2535 ConstantVector::get(Mask), 2536 getReplacementName(IBeforeJ ? I : J, 2537 true, o, 1)); 2538 NewI1->insertBefore(IBeforeJ ? J : I); 2539 I1 = NewI1; 2540 I1T = I2T; 2541 I1Elem = I2Elem; 2542 } else if (I1Elem > I2Elem) { 2543 std::vector<Constant *> Mask(I1Elem); 2544 unsigned v = 0; 2545 for (; v < I2Elem; ++v) 2546 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2547 for (; v < I1Elem; ++v) 2548 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2549 2550 Instruction *NewI2 = 2551 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2552 ConstantVector::get(Mask), 2553 getReplacementName(IBeforeJ ? I : J, 2554 true, o, 1)); 2555 NewI2->insertBefore(IBeforeJ ? J : I); 2556 I2 = NewI2; 2557 I2T = I1T; 2558 I2Elem = I1Elem; 2559 } 2560 2561 // Now that both I1 and I2 are the same length we can shuffle them 2562 // together (and use the result). 2563 std::vector<Constant *> Mask(numElem); 2564 for (unsigned v = 0; v < numElem; ++v) { 2565 if (II[v].first == -1) { 2566 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2567 } else { 2568 int Idx = II[v].first + II[v].second * I1Elem; 2569 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2570 } 2571 } 2572 2573 Instruction *NewOp = 2574 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2575 getReplacementName(IBeforeJ ? I : J, true, o)); 2576 NewOp->insertBefore(IBeforeJ ? J : I); 2577 return NewOp; 2578 } 2579 } 2580 2581 Type *ArgType = ArgTypeL; 2582 if (numElemL < numElemH) { 2583 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2584 ArgTypeL, VArgType, IBeforeJ, 1)) { 2585 // This is another short-circuit case: we're combining a scalar into 2586 // a vector that is formed by an IE chain. We've just expanded the IE 2587 // chain, now insert the scalar and we're done. 2588 2589 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2590 getReplacementName(IBeforeJ ? I : J, true, o)); 2591 S->insertBefore(IBeforeJ ? J : I); 2592 return S; 2593 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2594 ArgTypeH, IBeforeJ)) { 2595 // The two vector inputs to the shuffle must be the same length, 2596 // so extend the smaller vector to be the same length as the larger one. 2597 Instruction *NLOp; 2598 if (numElemL > 1) { 2599 2600 std::vector<Constant *> Mask(numElemH); 2601 unsigned v = 0; 2602 for (; v < numElemL; ++v) 2603 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2604 for (; v < numElemH; ++v) 2605 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2606 2607 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2608 ConstantVector::get(Mask), 2609 getReplacementName(IBeforeJ ? I : J, 2610 true, o, 1)); 2611 } else { 2612 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2613 getReplacementName(IBeforeJ ? I : J, 2614 true, o, 1)); 2615 } 2616 2617 NLOp->insertBefore(IBeforeJ ? J : I); 2618 LOp = NLOp; 2619 } 2620 2621 ArgType = ArgTypeH; 2622 } else if (numElemL > numElemH) { 2623 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2624 ArgTypeH, VArgType, IBeforeJ)) { 2625 Instruction *S = 2626 InsertElementInst::Create(LOp, HOp, 2627 ConstantInt::get(Type::getInt32Ty(Context), 2628 numElemL), 2629 getReplacementName(IBeforeJ ? I : J, 2630 true, o)); 2631 S->insertBefore(IBeforeJ ? J : I); 2632 return S; 2633 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2634 ArgTypeL, IBeforeJ)) { 2635 Instruction *NHOp; 2636 if (numElemH > 1) { 2637 std::vector<Constant *> Mask(numElemL); 2638 unsigned v = 0; 2639 for (; v < numElemH; ++v) 2640 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2641 for (; v < numElemL; ++v) 2642 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2643 2644 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2645 ConstantVector::get(Mask), 2646 getReplacementName(IBeforeJ ? I : J, 2647 true, o, 1)); 2648 } else { 2649 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2650 getReplacementName(IBeforeJ ? I : J, 2651 true, o, 1)); 2652 } 2653 2654 NHOp->insertBefore(IBeforeJ ? J : I); 2655 HOp = NHOp; 2656 } 2657 } 2658 2659 if (ArgType->isVectorTy()) { 2660 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2661 std::vector<Constant*> Mask(numElem); 2662 for (unsigned v = 0; v < numElem; ++v) { 2663 unsigned Idx = v; 2664 // If the low vector was expanded, we need to skip the extra 2665 // undefined entries. 2666 if (v >= numElemL && numElemH > numElemL) 2667 Idx += (numElemH - numElemL); 2668 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2669 } 2670 2671 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2672 ConstantVector::get(Mask), 2673 getReplacementName(IBeforeJ ? I : J, true, o)); 2674 BV->insertBefore(IBeforeJ ? J : I); 2675 return BV; 2676 } 2677 2678 Instruction *BV1 = InsertElementInst::Create( 2679 UndefValue::get(VArgType), LOp, CV0, 2680 getReplacementName(IBeforeJ ? I : J, 2681 true, o, 1)); 2682 BV1->insertBefore(IBeforeJ ? J : I); 2683 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2684 getReplacementName(IBeforeJ ? I : J, 2685 true, o, 2)); 2686 BV2->insertBefore(IBeforeJ ? J : I); 2687 return BV2; 2688 } 2689 2690 // This function creates an array of values that will be used as the inputs 2691 // to the vector instruction that fuses I with J. 2692 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2693 Instruction *I, Instruction *J, 2694 SmallVectorImpl<Value *> &ReplacedOperands, 2695 bool IBeforeJ) { 2696 unsigned NumOperands = I->getNumOperands(); 2697 2698 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2699 // Iterate backward so that we look at the store pointer 2700 // first and know whether or not we need to flip the inputs. 2701 2702 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2703 // This is the pointer for a load/store instruction. 2704 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2705 continue; 2706 } else if (isa<CallInst>(I)) { 2707 Function *F = cast<CallInst>(I)->getCalledFunction(); 2708 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 2709 if (o == NumOperands-1) { 2710 BasicBlock &BB = *I->getParent(); 2711 2712 Module *M = BB.getParent()->getParent(); 2713 Type *ArgTypeI = I->getType(); 2714 Type *ArgTypeJ = J->getType(); 2715 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2716 2717 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType); 2718 continue; 2719 } else if (IID == Intrinsic::powi && o == 1) { 2720 // The second argument of powi is a single integer and we've already 2721 // checked that both arguments are equal. As a result, we just keep 2722 // I's second argument. 2723 ReplacedOperands[o] = I->getOperand(o); 2724 continue; 2725 } 2726 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2727 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2728 continue; 2729 } 2730 2731 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2732 } 2733 } 2734 2735 // This function creates two values that represent the outputs of the 2736 // original I and J instructions. These are generally vector shuffles 2737 // or extracts. In many cases, these will end up being unused and, thus, 2738 // eliminated by later passes. 2739 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2740 Instruction *J, Instruction *K, 2741 Instruction *&InsertionPt, 2742 Instruction *&K1, Instruction *&K2) { 2743 if (isa<StoreInst>(I)) { 2744 AA->replaceWithNewValue(I, K); 2745 AA->replaceWithNewValue(J, K); 2746 } else { 2747 Type *IType = I->getType(); 2748 Type *JType = J->getType(); 2749 2750 VectorType *VType = getVecTypeForPair(IType, JType); 2751 unsigned numElem = VType->getNumElements(); 2752 2753 unsigned numElemI, numElemJ; 2754 if (IType->isVectorTy()) 2755 numElemI = cast<VectorType>(IType)->getNumElements(); 2756 else 2757 numElemI = 1; 2758 2759 if (JType->isVectorTy()) 2760 numElemJ = cast<VectorType>(JType)->getNumElements(); 2761 else 2762 numElemJ = 1; 2763 2764 if (IType->isVectorTy()) { 2765 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2766 for (unsigned v = 0; v < numElemI; ++v) { 2767 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2768 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2769 } 2770 2771 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2772 ConstantVector::get( Mask1), 2773 getReplacementName(K, false, 1)); 2774 } else { 2775 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2776 K1 = ExtractElementInst::Create(K, CV0, 2777 getReplacementName(K, false, 1)); 2778 } 2779 2780 if (JType->isVectorTy()) { 2781 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2782 for (unsigned v = 0; v < numElemJ; ++v) { 2783 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2784 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2785 } 2786 2787 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2788 ConstantVector::get( Mask2), 2789 getReplacementName(K, false, 2)); 2790 } else { 2791 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2792 K2 = ExtractElementInst::Create(K, CV1, 2793 getReplacementName(K, false, 2)); 2794 } 2795 2796 K1->insertAfter(K); 2797 K2->insertAfter(K1); 2798 InsertionPt = K2; 2799 } 2800 } 2801 2802 // Move all uses of the function I (including pairing-induced uses) after J. 2803 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2804 DenseSet<ValuePair> &LoadMoveSetPairs, 2805 Instruction *I, Instruction *J) { 2806 // Skip to the first instruction past I. 2807 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2808 2809 DenseSet<Value *> Users; 2810 AliasSetTracker WriteSet(*AA); 2811 if (I->mayWriteToMemory()) WriteSet.add(I); 2812 2813 for (; cast<Instruction>(L) != J; ++L) 2814 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs); 2815 2816 assert(cast<Instruction>(L) == J && 2817 "Tracking has not proceeded far enough to check for dependencies"); 2818 // If J is now in the use set of I, then trackUsesOfI will return true 2819 // and we have a dependency cycle (and the fusing operation must abort). 2820 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs); 2821 } 2822 2823 // Move all uses of the function I (including pairing-induced uses) after J. 2824 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2825 DenseSet<ValuePair> &LoadMoveSetPairs, 2826 Instruction *&InsertionPt, 2827 Instruction *I, Instruction *J) { 2828 // Skip to the first instruction past I. 2829 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2830 2831 DenseSet<Value *> Users; 2832 AliasSetTracker WriteSet(*AA); 2833 if (I->mayWriteToMemory()) WriteSet.add(I); 2834 2835 for (; cast<Instruction>(L) != J;) { 2836 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) { 2837 // Move this instruction 2838 Instruction *InstToMove = L; ++L; 2839 2840 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2841 " to after " << *InsertionPt << "\n"); 2842 InstToMove->removeFromParent(); 2843 InstToMove->insertAfter(InsertionPt); 2844 InsertionPt = InstToMove; 2845 } else { 2846 ++L; 2847 } 2848 } 2849 } 2850 2851 // Collect all load instruction that are in the move set of a given first 2852 // pair member. These loads depend on the first instruction, I, and so need 2853 // to be moved after J (the second instruction) when the pair is fused. 2854 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2855 DenseMap<Value *, Value *> &ChosenPairs, 2856 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 2857 DenseSet<ValuePair> &LoadMoveSetPairs, 2858 Instruction *I) { 2859 // Skip to the first instruction past I. 2860 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2861 2862 DenseSet<Value *> Users; 2863 AliasSetTracker WriteSet(*AA); 2864 if (I->mayWriteToMemory()) WriteSet.add(I); 2865 2866 // Note: We cannot end the loop when we reach J because J could be moved 2867 // farther down the use chain by another instruction pairing. Also, J 2868 // could be before I if this is an inverted input. 2869 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2870 if (trackUsesOfI(Users, WriteSet, I, L)) { 2871 if (L->mayReadFromMemory()) { 2872 LoadMoveSet[L].push_back(I); 2873 LoadMoveSetPairs.insert(ValuePair(L, I)); 2874 } 2875 } 2876 } 2877 } 2878 2879 // In cases where both load/stores and the computation of their pointers 2880 // are chosen for vectorization, we can end up in a situation where the 2881 // aliasing analysis starts returning different query results as the 2882 // process of fusing instruction pairs continues. Because the algorithm 2883 // relies on finding the same use dags here as were found earlier, we'll 2884 // need to precompute the necessary aliasing information here and then 2885 // manually update it during the fusion process. 2886 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2887 std::vector<Value *> &PairableInsts, 2888 DenseMap<Value *, Value *> &ChosenPairs, 2889 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet, 2890 DenseSet<ValuePair> &LoadMoveSetPairs) { 2891 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2892 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2893 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2894 if (P == ChosenPairs.end()) continue; 2895 2896 Instruction *I = cast<Instruction>(P->first); 2897 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, 2898 LoadMoveSetPairs, I); 2899 } 2900 } 2901 2902 // When the first instruction in each pair is cloned, it will inherit its 2903 // parent's metadata. This metadata must be combined with that of the other 2904 // instruction in a safe way. 2905 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2906 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2907 K->getAllMetadataOtherThanDebugLoc(Metadata); 2908 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2909 unsigned Kind = Metadata[i].first; 2910 MDNode *JMD = J->getMetadata(Kind); 2911 MDNode *KMD = Metadata[i].second; 2912 2913 switch (Kind) { 2914 default: 2915 K->setMetadata(Kind, 0); // Remove unknown metadata 2916 break; 2917 case LLVMContext::MD_tbaa: 2918 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2919 break; 2920 case LLVMContext::MD_fpmath: 2921 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2922 break; 2923 } 2924 } 2925 } 2926 2927 // This function fuses the chosen instruction pairs into vector instructions, 2928 // taking care preserve any needed scalar outputs and, then, it reorders the 2929 // remaining instructions as needed (users of the first member of the pair 2930 // need to be moved to after the location of the second member of the pair 2931 // because the vector instruction is inserted in the location of the pair's 2932 // second member). 2933 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2934 std::vector<Value *> &PairableInsts, 2935 DenseMap<Value *, Value *> &ChosenPairs, 2936 DenseSet<ValuePair> &FixedOrderPairs, 2937 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2938 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs, 2939 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) { 2940 LLVMContext& Context = BB.getContext(); 2941 2942 // During the vectorization process, the order of the pairs to be fused 2943 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2944 // list. After a pair is fused, the flipped pair is removed from the list. 2945 DenseSet<ValuePair> FlippedPairs; 2946 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2947 E = ChosenPairs.end(); P != E; ++P) 2948 FlippedPairs.insert(ValuePair(P->second, P->first)); 2949 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2950 E = FlippedPairs.end(); P != E; ++P) 2951 ChosenPairs.insert(*P); 2952 2953 DenseMap<Value *, std::vector<Value *> > LoadMoveSet; 2954 DenseSet<ValuePair> LoadMoveSetPairs; 2955 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, 2956 LoadMoveSet, LoadMoveSetPairs); 2957 2958 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2959 2960 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2961 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2962 if (P == ChosenPairs.end()) { 2963 ++PI; 2964 continue; 2965 } 2966 2967 if (getDepthFactor(P->first) == 0) { 2968 // These instructions are not really fused, but are tracked as though 2969 // they are. Any case in which it would be interesting to fuse them 2970 // will be taken care of by InstCombine. 2971 --NumFusedOps; 2972 ++PI; 2973 continue; 2974 } 2975 2976 Instruction *I = cast<Instruction>(P->first), 2977 *J = cast<Instruction>(P->second); 2978 2979 DEBUG(dbgs() << "BBV: fusing: " << *I << 2980 " <-> " << *J << "\n"); 2981 2982 // Remove the pair and flipped pair from the list. 2983 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2984 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2985 ChosenPairs.erase(FP); 2986 ChosenPairs.erase(P); 2987 2988 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) { 2989 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2990 " <-> " << *J << 2991 " aborted because of non-trivial dependency cycle\n"); 2992 --NumFusedOps; 2993 ++PI; 2994 continue; 2995 } 2996 2997 // If the pair must have the other order, then flip it. 2998 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2999 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 3000 // This pair does not have a fixed order, and so we might want to 3001 // flip it if that will yield fewer shuffles. We count the number 3002 // of dependencies connected via swaps, and those directly connected, 3003 // and flip the order if the number of swaps is greater. 3004 bool OrigOrder = true; 3005 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ = 3006 ConnectedPairDeps.find(ValuePair(I, J)); 3007 if (IJ == ConnectedPairDeps.end()) { 3008 IJ = ConnectedPairDeps.find(ValuePair(J, I)); 3009 OrigOrder = false; 3010 } 3011 3012 if (IJ != ConnectedPairDeps.end()) { 3013 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 3014 for (std::vector<ValuePair>::iterator T = IJ->second.begin(), 3015 TE = IJ->second.end(); T != TE; ++T) { 3016 VPPair Q(IJ->first, *T); 3017 DenseMap<VPPair, unsigned>::iterator R = 3018 PairConnectionTypes.find(VPPair(Q.second, Q.first)); 3019 assert(R != PairConnectionTypes.end() && 3020 "Cannot find pair connection type"); 3021 if (R->second == PairConnectionDirect) 3022 ++NumDepsDirect; 3023 else if (R->second == PairConnectionSwap) 3024 ++NumDepsSwap; 3025 } 3026 3027 if (!OrigOrder) 3028 std::swap(NumDepsDirect, NumDepsSwap); 3029 3030 if (NumDepsSwap > NumDepsDirect) { 3031 FlipPairOrder = true; 3032 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 3033 " <-> " << *J << "\n"); 3034 } 3035 } 3036 } 3037 3038 Instruction *L = I, *H = J; 3039 if (FlipPairOrder) 3040 std::swap(H, L); 3041 3042 // If the pair being fused uses the opposite order from that in the pair 3043 // connection map, then we need to flip the types. 3044 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL = 3045 ConnectedPairs.find(ValuePair(H, L)); 3046 if (HL != ConnectedPairs.end()) 3047 for (std::vector<ValuePair>::iterator T = HL->second.begin(), 3048 TE = HL->second.end(); T != TE; ++T) { 3049 VPPair Q(HL->first, *T); 3050 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q); 3051 assert(R != PairConnectionTypes.end() && 3052 "Cannot find pair connection type"); 3053 if (R->second == PairConnectionDirect) 3054 R->second = PairConnectionSwap; 3055 else if (R->second == PairConnectionSwap) 3056 R->second = PairConnectionDirect; 3057 } 3058 3059 bool LBeforeH = !FlipPairOrder; 3060 unsigned NumOperands = I->getNumOperands(); 3061 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 3062 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 3063 LBeforeH); 3064 3065 // Make a copy of the original operation, change its type to the vector 3066 // type and replace its operands with the vector operands. 3067 Instruction *K = L->clone(); 3068 if (L->hasName()) 3069 K->takeName(L); 3070 else if (H->hasName()) 3071 K->takeName(H); 3072 3073 if (!isa<StoreInst>(K)) 3074 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 3075 3076 combineMetadata(K, H); 3077 K->intersectOptionalDataWith(H); 3078 3079 for (unsigned o = 0; o < NumOperands; ++o) 3080 K->setOperand(o, ReplacedOperands[o]); 3081 3082 K->insertAfter(J); 3083 3084 // Instruction insertion point: 3085 Instruction *InsertionPt = K; 3086 Instruction *K1 = 0, *K2 = 0; 3087 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 3088 3089 // The use dag of the first original instruction must be moved to after 3090 // the location of the second instruction. The entire use dag of the 3091 // first instruction is disjoint from the input dag of the second 3092 // (by definition), and so commutes with it. 3093 3094 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J); 3095 3096 if (!isa<StoreInst>(I)) { 3097 L->replaceAllUsesWith(K1); 3098 H->replaceAllUsesWith(K2); 3099 AA->replaceWithNewValue(L, K1); 3100 AA->replaceWithNewValue(H, K2); 3101 } 3102 3103 // Instructions that may read from memory may be in the load move set. 3104 // Once an instruction is fused, we no longer need its move set, and so 3105 // the values of the map never need to be updated. However, when a load 3106 // is fused, we need to merge the entries from both instructions in the 3107 // pair in case those instructions were in the move set of some other 3108 // yet-to-be-fused pair. The loads in question are the keys of the map. 3109 if (I->mayReadFromMemory()) { 3110 std::vector<ValuePair> NewSetMembers; 3111 DenseMap<Value *, std::vector<Value *> >::iterator II = 3112 LoadMoveSet.find(I); 3113 if (II != LoadMoveSet.end()) 3114 for (std::vector<Value *>::iterator N = II->second.begin(), 3115 NE = II->second.end(); N != NE; ++N) 3116 NewSetMembers.push_back(ValuePair(K, *N)); 3117 DenseMap<Value *, std::vector<Value *> >::iterator JJ = 3118 LoadMoveSet.find(J); 3119 if (JJ != LoadMoveSet.end()) 3120 for (std::vector<Value *>::iterator N = JJ->second.begin(), 3121 NE = JJ->second.end(); N != NE; ++N) 3122 NewSetMembers.push_back(ValuePair(K, *N)); 3123 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 3124 AE = NewSetMembers.end(); A != AE; ++A) { 3125 LoadMoveSet[A->first].push_back(A->second); 3126 LoadMoveSetPairs.insert(*A); 3127 } 3128 } 3129 3130 // Before removing I, set the iterator to the next instruction. 3131 PI = llvm::next(BasicBlock::iterator(I)); 3132 if (cast<Instruction>(PI) == J) 3133 ++PI; 3134 3135 SE->forgetValue(I); 3136 SE->forgetValue(J); 3137 I->eraseFromParent(); 3138 J->eraseFromParent(); 3139 3140 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 3141 BB << "\n"); 3142 } 3143 3144 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 3145 } 3146} 3147 3148char BBVectorize::ID = 0; 3149static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 3150INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3151INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3152INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3153INITIALIZE_PASS_DEPENDENCY(DominatorTree) 3154INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3155INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3156 3157BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 3158 return new BBVectorize(C); 3159} 3160 3161bool 3162llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 3163 BBVectorize BBVectorizer(P, C); 3164 return BBVectorizer.vectorizeBB(BB); 3165} 3166 3167//===----------------------------------------------------------------------===// 3168VectorizeConfig::VectorizeConfig() { 3169 VectorBits = ::VectorBits; 3170 VectorizeBools = !::NoBools; 3171 VectorizeInts = !::NoInts; 3172 VectorizeFloats = !::NoFloats; 3173 VectorizePointers = !::NoPointers; 3174 VectorizeCasts = !::NoCasts; 3175 VectorizeMath = !::NoMath; 3176 VectorizeFMA = !::NoFMA; 3177 VectorizeSelect = !::NoSelect; 3178 VectorizeCmp = !::NoCmp; 3179 VectorizeGEP = !::NoGEP; 3180 VectorizeMemOps = !::NoMemOps; 3181 AlignedOnly = ::AlignedOnly; 3182 ReqChainDepth= ::ReqChainDepth; 3183 SearchLimit = ::SearchLimit; 3184 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 3185 SplatBreaksChain = ::SplatBreaksChain; 3186 MaxInsts = ::MaxInsts; 3187 MaxPairs = ::MaxPairs; 3188 MaxIter = ::MaxIter; 3189 Pow2LenOnly = ::Pow2LenOnly; 3190 NoMemOpBoost = ::NoMemOpBoost; 3191 FastDep = ::FastDep; 3192} 3193