BBVectorize.cpp revision 64a7a24edf719bb6ffacc030c23f4cd99312f3fb
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/ValueTracking.h" 33#include "llvm/Constants.h" 34#include "llvm/DataLayout.h" 35#include "llvm/DerivedTypes.h" 36#include "llvm/Function.h" 37#include "llvm/Instructions.h" 38#include "llvm/IntrinsicInst.h" 39#include "llvm/Intrinsics.h" 40#include "llvm/LLVMContext.h" 41#include "llvm/Metadata.h" 42#include "llvm/Pass.h" 43#include "llvm/Support/CommandLine.h" 44#include "llvm/Support/Debug.h" 45#include "llvm/Support/ValueHandle.h" 46#include "llvm/Support/raw_ostream.h" 47#include "llvm/TargetTransformInfo.h" 48#include "llvm/Transforms/Utils/Local.h" 49#include "llvm/Type.h" 50#include <algorithm> 51#include <map> 52using namespace llvm; 53 54static cl::opt<bool> 55IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false), 56 cl::Hidden, cl::desc("Ignore target information")); 57 58static cl::opt<unsigned> 59ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, 60 cl::desc("The required chain depth for vectorization")); 61 62static cl::opt<bool> 63UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false), 64 cl::Hidden, cl::desc("Use the chain depth requirement with" 65 " target information")); 66 67static cl::opt<unsigned> 68SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, 69 cl::desc("The maximum search distance for instruction pairs")); 70 71static cl::opt<bool> 72SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden, 73 cl::desc("Replicating one element to a pair breaks the chain")); 74 75static cl::opt<unsigned> 76VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden, 77 cl::desc("The size of the native vector registers")); 78 79static cl::opt<unsigned> 80MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden, 81 cl::desc("The maximum number of pairing iterations")); 82 83static cl::opt<bool> 84Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden, 85 cl::desc("Don't try to form non-2^n-length vectors")); 86 87static cl::opt<unsigned> 88MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden, 89 cl::desc("The maximum number of pairable instructions per group")); 90 91static cl::opt<unsigned> 92MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200), 93 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use" 94 " a full cycle check")); 95 96static cl::opt<bool> 97NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden, 98 cl::desc("Don't try to vectorize boolean (i1) values")); 99 100static cl::opt<bool> 101NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden, 102 cl::desc("Don't try to vectorize integer values")); 103 104static cl::opt<bool> 105NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, 106 cl::desc("Don't try to vectorize floating-point values")); 107 108// FIXME: This should default to false once pointer vector support works. 109static cl::opt<bool> 110NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden, 111 cl::desc("Don't try to vectorize pointer values")); 112 113static cl::opt<bool> 114NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, 115 cl::desc("Don't try to vectorize casting (conversion) operations")); 116 117static cl::opt<bool> 118NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden, 119 cl::desc("Don't try to vectorize floating-point math intrinsics")); 120 121static cl::opt<bool> 122NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden, 123 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic")); 124 125static cl::opt<bool> 126NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, 127 cl::desc("Don't try to vectorize select instructions")); 128 129static cl::opt<bool> 130NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden, 131 cl::desc("Don't try to vectorize comparison instructions")); 132 133static cl::opt<bool> 134NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, 135 cl::desc("Don't try to vectorize getelementptr instructions")); 136 137static cl::opt<bool> 138NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, 139 cl::desc("Don't try to vectorize loads and stores")); 140 141static cl::opt<bool> 142AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden, 143 cl::desc("Only generate aligned loads and stores")); 144 145static cl::opt<bool> 146NoMemOpBoost("bb-vectorize-no-mem-op-boost", 147 cl::init(false), cl::Hidden, 148 cl::desc("Don't boost the chain-depth contribution of loads and stores")); 149 150static cl::opt<bool> 151FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden, 152 cl::desc("Use a fast instruction dependency analysis")); 153 154#ifndef NDEBUG 155static cl::opt<bool> 156DebugInstructionExamination("bb-vectorize-debug-instruction-examination", 157 cl::init(false), cl::Hidden, 158 cl::desc("When debugging is enabled, output information on the" 159 " instruction-examination process")); 160static cl::opt<bool> 161DebugCandidateSelection("bb-vectorize-debug-candidate-selection", 162 cl::init(false), cl::Hidden, 163 cl::desc("When debugging is enabled, output information on the" 164 " candidate-selection process")); 165static cl::opt<bool> 166DebugPairSelection("bb-vectorize-debug-pair-selection", 167 cl::init(false), cl::Hidden, 168 cl::desc("When debugging is enabled, output information on the" 169 " pair-selection process")); 170static cl::opt<bool> 171DebugCycleCheck("bb-vectorize-debug-cycle-check", 172 cl::init(false), cl::Hidden, 173 cl::desc("When debugging is enabled, output information on the" 174 " cycle-checking process")); 175 176static cl::opt<bool> 177PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair", 178 cl::init(false), cl::Hidden, 179 cl::desc("When debugging is enabled, dump the basic block after" 180 " every pair is fused")); 181#endif 182 183STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); 184 185namespace { 186 struct BBVectorize : public BasicBlockPass { 187 static char ID; // Pass identification, replacement for typeid 188 189 const VectorizeConfig Config; 190 191 BBVectorize(const VectorizeConfig &C = VectorizeConfig()) 192 : BasicBlockPass(ID), Config(C) { 193 initializeBBVectorizePass(*PassRegistry::getPassRegistry()); 194 } 195 196 BBVectorize(Pass *P, const VectorizeConfig &C) 197 : BasicBlockPass(ID), Config(C) { 198 AA = &P->getAnalysis<AliasAnalysis>(); 199 DT = &P->getAnalysis<DominatorTree>(); 200 SE = &P->getAnalysis<ScalarEvolution>(); 201 TD = P->getAnalysisIfAvailable<DataLayout>(); 202 TTI = IgnoreTargetInfo ? 0 : 203 P->getAnalysisIfAvailable<TargetTransformInfo>(); 204 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 205 } 206 207 typedef std::pair<Value *, Value *> ValuePair; 208 typedef std::pair<ValuePair, int> ValuePairWithCost; 209 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 210 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 211 typedef std::pair<VPPair, unsigned> VPPairWithType; 212 typedef std::pair<std::multimap<Value *, Value *>::iterator, 213 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 214 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 215 std::multimap<ValuePair, ValuePair>::iterator> 216 VPPIteratorPair; 217 218 AliasAnalysis *AA; 219 DominatorTree *DT; 220 ScalarEvolution *SE; 221 DataLayout *TD; 222 TargetTransformInfo *TTI; 223 const VectorTargetTransformInfo *VTTI; 224 225 // FIXME: const correct? 226 227 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 228 229 bool getCandidatePairs(BasicBlock &BB, 230 BasicBlock::iterator &Start, 231 std::multimap<Value *, Value *> &CandidatePairs, 232 DenseSet<ValuePair> &FixedOrderPairs, 233 DenseMap<ValuePair, int> &CandidatePairCostSavings, 234 std::vector<Value *> &PairableInsts, bool NonPow2Len); 235 236 // FIXME: The current implementation does not account for pairs that 237 // are connected in multiple ways. For example: 238 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap) 239 enum PairConnectionType { 240 PairConnectionDirect, 241 PairConnectionSwap, 242 PairConnectionSplat 243 }; 244 245 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 246 std::vector<Value *> &PairableInsts, 247 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 248 DenseMap<VPPair, unsigned> &PairConnectionTypes); 249 250 void buildDepMap(BasicBlock &BB, 251 std::multimap<Value *, Value *> &CandidatePairs, 252 std::vector<Value *> &PairableInsts, 253 DenseSet<ValuePair> &PairableInstUsers); 254 255 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 256 DenseMap<ValuePair, int> &CandidatePairCostSavings, 257 std::vector<Value *> &PairableInsts, 258 DenseSet<ValuePair> &FixedOrderPairs, 259 DenseMap<VPPair, unsigned> &PairConnectionTypes, 260 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 261 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 262 DenseSet<ValuePair> &PairableInstUsers, 263 DenseMap<Value *, Value *>& ChosenPairs); 264 265 void fuseChosenPairs(BasicBlock &BB, 266 std::vector<Value *> &PairableInsts, 267 DenseMap<Value *, Value *>& ChosenPairs, 268 DenseSet<ValuePair> &FixedOrderPairs, 269 DenseMap<VPPair, unsigned> &PairConnectionTypes, 270 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 271 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps); 272 273 274 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 275 276 bool areInstsCompatible(Instruction *I, Instruction *J, 277 bool IsSimpleLoadStore, bool NonPow2Len, 278 int &CostSavings, int &FixedOrder); 279 280 bool trackUsesOfI(DenseSet<Value *> &Users, 281 AliasSetTracker &WriteSet, Instruction *I, 282 Instruction *J, bool UpdateUsers = true, 283 std::multimap<Value *, Value *> *LoadMoveSet = 0); 284 285 void computePairsConnectedTo( 286 std::multimap<Value *, Value *> &CandidatePairs, 287 std::vector<Value *> &PairableInsts, 288 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 289 DenseMap<VPPair, unsigned> &PairConnectionTypes, 290 ValuePair P); 291 292 bool pairsConflict(ValuePair P, ValuePair Q, 293 DenseSet<ValuePair> &PairableInstUsers, 294 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0); 295 296 bool pairWillFormCycle(ValuePair P, 297 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 298 DenseSet<ValuePair> &CurrentPairs); 299 300 void pruneTreeFor( 301 std::multimap<Value *, Value *> &CandidatePairs, 302 std::vector<Value *> &PairableInsts, 303 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 304 DenseSet<ValuePair> &PairableInstUsers, 305 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 306 DenseMap<Value *, Value *> &ChosenPairs, 307 DenseMap<ValuePair, size_t> &Tree, 308 DenseSet<ValuePair> &PrunedTree, ValuePair J, 309 bool UseCycleCheck); 310 311 void buildInitialTreeFor( 312 std::multimap<Value *, Value *> &CandidatePairs, 313 std::vector<Value *> &PairableInsts, 314 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 315 DenseSet<ValuePair> &PairableInstUsers, 316 DenseMap<Value *, Value *> &ChosenPairs, 317 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 318 319 void findBestTreeFor( 320 std::multimap<Value *, Value *> &CandidatePairs, 321 DenseMap<ValuePair, int> &CandidatePairCostSavings, 322 std::vector<Value *> &PairableInsts, 323 DenseSet<ValuePair> &FixedOrderPairs, 324 DenseMap<VPPair, unsigned> &PairConnectionTypes, 325 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 326 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 327 DenseSet<ValuePair> &PairableInstUsers, 328 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 329 DenseMap<Value *, Value *> &ChosenPairs, 330 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 331 int &BestEffSize, VPIteratorPair ChoiceRange, 332 bool UseCycleCheck); 333 334 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 335 Instruction *J, unsigned o); 336 337 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 338 unsigned MaskOffset, unsigned NumInElem, 339 unsigned NumInElem1, unsigned IdxOffset, 340 std::vector<Constant*> &Mask); 341 342 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 343 Instruction *J); 344 345 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 346 unsigned o, Value *&LOp, unsigned numElemL, 347 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, 348 unsigned IdxOff = 0); 349 350 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 351 Instruction *J, unsigned o, bool IBeforeJ); 352 353 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 354 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 355 bool IBeforeJ); 356 357 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 358 Instruction *J, Instruction *K, 359 Instruction *&InsertionPt, Instruction *&K1, 360 Instruction *&K2); 361 362 void collectPairLoadMoveSet(BasicBlock &BB, 363 DenseMap<Value *, Value *> &ChosenPairs, 364 std::multimap<Value *, Value *> &LoadMoveSet, 365 Instruction *I); 366 367 void collectLoadMoveSet(BasicBlock &BB, 368 std::vector<Value *> &PairableInsts, 369 DenseMap<Value *, Value *> &ChosenPairs, 370 std::multimap<Value *, Value *> &LoadMoveSet); 371 372 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 373 std::multimap<Value *, Value *> &LoadMoveSet, 374 Instruction *I, Instruction *J); 375 376 void moveUsesOfIAfterJ(BasicBlock &BB, 377 std::multimap<Value *, Value *> &LoadMoveSet, 378 Instruction *&InsertionPt, 379 Instruction *I, Instruction *J); 380 381 void combineMetadata(Instruction *K, const Instruction *J); 382 383 bool vectorizeBB(BasicBlock &BB) { 384 if (!DT->isReachableFromEntry(&BB)) { 385 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 386 " in " << BB.getParent()->getName() << "\n"); 387 return false; 388 } 389 390 DEBUG(if (VTTI) dbgs() << "BBV: using target information\n"); 391 392 bool changed = false; 393 // Iterate a sufficient number of times to merge types of size 1 bit, 394 // then 2 bits, then 4, etc. up to half of the target vector width of the 395 // target vector register. 396 unsigned n = 1; 397 for (unsigned v = 2; 398 (VTTI || v <= Config.VectorBits) && 399 (!Config.MaxIter || n <= Config.MaxIter); 400 v *= 2, ++n) { 401 DEBUG(dbgs() << "BBV: fusing loop #" << n << 402 " for " << BB.getName() << " in " << 403 BB.getParent()->getName() << "...\n"); 404 if (vectorizePairs(BB)) 405 changed = true; 406 else 407 break; 408 } 409 410 if (changed && !Pow2LenOnly) { 411 ++n; 412 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 413 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 414 n << " for " << BB.getName() << " in " << 415 BB.getParent()->getName() << "...\n"); 416 if (!vectorizePairs(BB, true)) break; 417 } 418 } 419 420 DEBUG(dbgs() << "BBV: done!\n"); 421 return changed; 422 } 423 424 virtual bool runOnBasicBlock(BasicBlock &BB) { 425 AA = &getAnalysis<AliasAnalysis>(); 426 DT = &getAnalysis<DominatorTree>(); 427 SE = &getAnalysis<ScalarEvolution>(); 428 TD = getAnalysisIfAvailable<DataLayout>(); 429 TTI = IgnoreTargetInfo ? 0 : 430 getAnalysisIfAvailable<TargetTransformInfo>(); 431 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 432 433 return vectorizeBB(BB); 434 } 435 436 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 437 BasicBlockPass::getAnalysisUsage(AU); 438 AU.addRequired<AliasAnalysis>(); 439 AU.addRequired<DominatorTree>(); 440 AU.addRequired<ScalarEvolution>(); 441 AU.addPreserved<AliasAnalysis>(); 442 AU.addPreserved<DominatorTree>(); 443 AU.addPreserved<ScalarEvolution>(); 444 AU.setPreservesCFG(); 445 } 446 447 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 448 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 449 "Cannot form vector from incompatible scalar types"); 450 Type *STy = ElemTy->getScalarType(); 451 452 unsigned numElem; 453 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 454 numElem = VTy->getNumElements(); 455 } else { 456 numElem = 1; 457 } 458 459 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 460 numElem += VTy->getNumElements(); 461 } else { 462 numElem += 1; 463 } 464 465 return VectorType::get(STy, numElem); 466 } 467 468 static inline void getInstructionTypes(Instruction *I, 469 Type *&T1, Type *&T2) { 470 if (isa<StoreInst>(I)) { 471 // For stores, it is the value type, not the pointer type that matters 472 // because the value is what will come from a vector register. 473 474 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 475 T1 = IVal->getType(); 476 } else { 477 T1 = I->getType(); 478 } 479 480 if (I->isCast()) 481 T2 = cast<CastInst>(I)->getSrcTy(); 482 else 483 T2 = T1; 484 485 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 486 T2 = SI->getCondition()->getType(); 487 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { 488 T2 = SI->getOperand(0)->getType(); 489 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 490 T2 = CI->getOperand(0)->getType(); 491 } 492 } 493 494 // Returns the weight associated with the provided value. A chain of 495 // candidate pairs has a length given by the sum of the weights of its 496 // members (one weight per pair; the weight of each member of the pair 497 // is assumed to be the same). This length is then compared to the 498 // chain-length threshold to determine if a given chain is significant 499 // enough to be vectorized. The length is also used in comparing 500 // candidate chains where longer chains are considered to be better. 501 // Note: when this function returns 0, the resulting instructions are 502 // not actually fused. 503 inline size_t getDepthFactor(Value *V) { 504 // InsertElement and ExtractElement have a depth factor of zero. This is 505 // for two reasons: First, they cannot be usefully fused. Second, because 506 // the pass generates a lot of these, they can confuse the simple metric 507 // used to compare the trees in the next iteration. Thus, giving them a 508 // weight of zero allows the pass to essentially ignore them in 509 // subsequent iterations when looking for vectorization opportunities 510 // while still tracking dependency chains that flow through those 511 // instructions. 512 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 513 return 0; 514 515 // Give a load or store half of the required depth so that load/store 516 // pairs will vectorize. 517 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 518 return Config.ReqChainDepth/2; 519 520 return 1; 521 } 522 523 // Returns the cost of the provided instruction using VTTI. 524 // This does not handle loads and stores. 525 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { 526 switch (Opcode) { 527 default: break; 528 case Instruction::GetElementPtr: 529 // We mark this instruction as zero-cost because scalar GEPs are usually 530 // lowered to the intruction addressing mode. At the moment we don't 531 // generate vector GEPs. 532 return 0; 533 case Instruction::Br: 534 return VTTI->getCFInstrCost(Opcode); 535 case Instruction::PHI: 536 return 0; 537 case Instruction::Add: 538 case Instruction::FAdd: 539 case Instruction::Sub: 540 case Instruction::FSub: 541 case Instruction::Mul: 542 case Instruction::FMul: 543 case Instruction::UDiv: 544 case Instruction::SDiv: 545 case Instruction::FDiv: 546 case Instruction::URem: 547 case Instruction::SRem: 548 case Instruction::FRem: 549 case Instruction::Shl: 550 case Instruction::LShr: 551 case Instruction::AShr: 552 case Instruction::And: 553 case Instruction::Or: 554 case Instruction::Xor: 555 return VTTI->getArithmeticInstrCost(Opcode, T1); 556 case Instruction::Select: 557 case Instruction::ICmp: 558 case Instruction::FCmp: 559 return VTTI->getCmpSelInstrCost(Opcode, T1, T2); 560 case Instruction::ZExt: 561 case Instruction::SExt: 562 case Instruction::FPToUI: 563 case Instruction::FPToSI: 564 case Instruction::FPExt: 565 case Instruction::PtrToInt: 566 case Instruction::IntToPtr: 567 case Instruction::SIToFP: 568 case Instruction::UIToFP: 569 case Instruction::Trunc: 570 case Instruction::FPTrunc: 571 case Instruction::BitCast: 572 case Instruction::ShuffleVector: 573 return VTTI->getCastInstrCost(Opcode, T1, T2); 574 } 575 576 return 1; 577 } 578 579 // This determines the relative offset of two loads or stores, returning 580 // true if the offset could be determined to be some constant value. 581 // For example, if OffsetInElmts == 1, then J accesses the memory directly 582 // after I; if OffsetInElmts == -1 then I accesses the memory 583 // directly after J. 584 bool getPairPtrInfo(Instruction *I, Instruction *J, 585 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 586 unsigned &IAddressSpace, unsigned &JAddressSpace, 587 int64_t &OffsetInElmts, bool ComputeOffset = true) { 588 OffsetInElmts = 0; 589 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 590 LoadInst *LJ = cast<LoadInst>(J); 591 IPtr = LI->getPointerOperand(); 592 JPtr = LJ->getPointerOperand(); 593 IAlignment = LI->getAlignment(); 594 JAlignment = LJ->getAlignment(); 595 IAddressSpace = LI->getPointerAddressSpace(); 596 JAddressSpace = LJ->getPointerAddressSpace(); 597 } else { 598 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 599 IPtr = SI->getPointerOperand(); 600 JPtr = SJ->getPointerOperand(); 601 IAlignment = SI->getAlignment(); 602 JAlignment = SJ->getAlignment(); 603 IAddressSpace = SI->getPointerAddressSpace(); 604 JAddressSpace = SJ->getPointerAddressSpace(); 605 } 606 607 if (!ComputeOffset) 608 return true; 609 610 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 611 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 612 613 // If this is a trivial offset, then we'll get something like 614 // 1*sizeof(type). With target data, which we need anyway, this will get 615 // constant folded into a number. 616 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 617 if (const SCEVConstant *ConstOffSCEV = 618 dyn_cast<SCEVConstant>(OffsetSCEV)) { 619 ConstantInt *IntOff = ConstOffSCEV->getValue(); 620 int64_t Offset = IntOff->getSExtValue(); 621 622 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 623 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 624 625 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 626 if (VTy != VTy2 && Offset < 0) { 627 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 628 OffsetInElmts = Offset/VTy2TSS; 629 return (abs64(Offset) % VTy2TSS) == 0; 630 } 631 632 OffsetInElmts = Offset/VTyTSS; 633 return (abs64(Offset) % VTyTSS) == 0; 634 } 635 636 return false; 637 } 638 639 // Returns true if the provided CallInst represents an intrinsic that can 640 // be vectorized. 641 bool isVectorizableIntrinsic(CallInst* I) { 642 Function *F = I->getCalledFunction(); 643 if (!F) return false; 644 645 unsigned IID = F->getIntrinsicID(); 646 if (!IID) return false; 647 648 switch(IID) { 649 default: 650 return false; 651 case Intrinsic::sqrt: 652 case Intrinsic::powi: 653 case Intrinsic::sin: 654 case Intrinsic::cos: 655 case Intrinsic::log: 656 case Intrinsic::log2: 657 case Intrinsic::log10: 658 case Intrinsic::exp: 659 case Intrinsic::exp2: 660 case Intrinsic::pow: 661 return Config.VectorizeMath; 662 case Intrinsic::fma: 663 case Intrinsic::fmuladd: 664 return Config.VectorizeFMA; 665 } 666 } 667 668 // Returns true if J is the second element in some pair referenced by 669 // some multimap pair iterator pair. 670 template <typename V> 671 bool isSecondInIteratorPair(V J, std::pair< 672 typename std::multimap<V, V>::iterator, 673 typename std::multimap<V, V>::iterator> PairRange) { 674 for (typename std::multimap<V, V>::iterator K = PairRange.first; 675 K != PairRange.second; ++K) 676 if (K->second == J) return true; 677 678 return false; 679 } 680 681 bool isPureIEChain(InsertElementInst *IE) { 682 InsertElementInst *IENext = IE; 683 do { 684 if (!isa<UndefValue>(IENext->getOperand(0)) && 685 !isa<InsertElementInst>(IENext->getOperand(0))) { 686 return false; 687 } 688 } while ((IENext = 689 dyn_cast<InsertElementInst>(IENext->getOperand(0)))); 690 691 return true; 692 } 693 }; 694 695 // This function implements one vectorization iteration on the provided 696 // basic block. It returns true if the block is changed. 697 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 698 bool ShouldContinue; 699 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 700 701 std::vector<Value *> AllPairableInsts; 702 DenseMap<Value *, Value *> AllChosenPairs; 703 DenseSet<ValuePair> AllFixedOrderPairs; 704 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 705 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; 706 707 do { 708 std::vector<Value *> PairableInsts; 709 std::multimap<Value *, Value *> CandidatePairs; 710 DenseSet<ValuePair> FixedOrderPairs; 711 DenseMap<ValuePair, int> CandidatePairCostSavings; 712 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 713 FixedOrderPairs, 714 CandidatePairCostSavings, 715 PairableInsts, NonPow2Len); 716 if (PairableInsts.empty()) continue; 717 718 // Now we have a map of all of the pairable instructions and we need to 719 // select the best possible pairing. A good pairing is one such that the 720 // users of the pair are also paired. This defines a (directed) forest 721 // over the pairs such that two pairs are connected iff the second pair 722 // uses the first. 723 724 // Note that it only matters that both members of the second pair use some 725 // element of the first pair (to allow for splatting). 726 727 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; 728 DenseMap<VPPair, unsigned> PairConnectionTypes; 729 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs, 730 PairConnectionTypes); 731 if (ConnectedPairs.empty()) continue; 732 733 for (std::multimap<ValuePair, ValuePair>::iterator 734 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 735 I != IE; ++I) { 736 ConnectedPairDeps.insert(VPPair(I->second, I->first)); 737 } 738 739 // Build the pairable-instruction dependency map 740 DenseSet<ValuePair> PairableInstUsers; 741 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 742 743 // There is now a graph of the connected pairs. For each variable, pick 744 // the pairing with the largest tree meeting the depth requirement on at 745 // least one branch. Then select all pairings that are part of that tree 746 // and remove them from the list of available pairings and pairable 747 // variables. 748 749 DenseMap<Value *, Value *> ChosenPairs; 750 choosePairs(CandidatePairs, CandidatePairCostSavings, 751 PairableInsts, FixedOrderPairs, PairConnectionTypes, 752 ConnectedPairs, ConnectedPairDeps, 753 PairableInstUsers, ChosenPairs); 754 755 if (ChosenPairs.empty()) continue; 756 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 757 PairableInsts.end()); 758 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 759 760 // Only for the chosen pairs, propagate information on fixed-order pairs, 761 // pair connections, and their types to the data structures used by the 762 // pair fusion procedures. 763 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 764 IE = ChosenPairs.end(); I != IE; ++I) { 765 if (FixedOrderPairs.count(*I)) 766 AllFixedOrderPairs.insert(*I); 767 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 768 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 769 770 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 771 J != IE; ++J) { 772 DenseMap<VPPair, unsigned>::iterator K = 773 PairConnectionTypes.find(VPPair(*I, *J)); 774 if (K != PairConnectionTypes.end()) { 775 AllPairConnectionTypes.insert(*K); 776 } else { 777 K = PairConnectionTypes.find(VPPair(*J, *I)); 778 if (K != PairConnectionTypes.end()) 779 AllPairConnectionTypes.insert(*K); 780 } 781 } 782 } 783 784 for (std::multimap<ValuePair, ValuePair>::iterator 785 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 786 I != IE; ++I) { 787 if (AllPairConnectionTypes.count(*I)) { 788 AllConnectedPairs.insert(*I); 789 AllConnectedPairDeps.insert(VPPair(I->second, I->first)); 790 } 791 } 792 } while (ShouldContinue); 793 794 if (AllChosenPairs.empty()) return false; 795 NumFusedOps += AllChosenPairs.size(); 796 797 // A set of pairs has now been selected. It is now necessary to replace the 798 // paired instructions with vector instructions. For this procedure each 799 // operand must be replaced with a vector operand. This vector is formed 800 // by using build_vector on the old operands. The replaced values are then 801 // replaced with a vector_extract on the result. Subsequent optimization 802 // passes should coalesce the build/extract combinations. 803 804 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 805 AllPairConnectionTypes, 806 AllConnectedPairs, AllConnectedPairDeps); 807 808 // It is important to cleanup here so that future iterations of this 809 // function have less work to do. 810 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 811 return true; 812 } 813 814 // This function returns true if the provided instruction is capable of being 815 // fused into a vector instruction. This determination is based only on the 816 // type and other attributes of the instruction. 817 bool BBVectorize::isInstVectorizable(Instruction *I, 818 bool &IsSimpleLoadStore) { 819 IsSimpleLoadStore = false; 820 821 if (CallInst *C = dyn_cast<CallInst>(I)) { 822 if (!isVectorizableIntrinsic(C)) 823 return false; 824 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 825 // Vectorize simple loads if possbile: 826 IsSimpleLoadStore = L->isSimple(); 827 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 828 return false; 829 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 830 // Vectorize simple stores if possbile: 831 IsSimpleLoadStore = S->isSimple(); 832 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 833 return false; 834 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 835 // We can vectorize casts, but not casts of pointer types, etc. 836 if (!Config.VectorizeCasts) 837 return false; 838 839 Type *SrcTy = C->getSrcTy(); 840 if (!SrcTy->isSingleValueType()) 841 return false; 842 843 Type *DestTy = C->getDestTy(); 844 if (!DestTy->isSingleValueType()) 845 return false; 846 } else if (isa<SelectInst>(I)) { 847 if (!Config.VectorizeSelect) 848 return false; 849 } else if (isa<CmpInst>(I)) { 850 if (!Config.VectorizeCmp) 851 return false; 852 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 853 if (!Config.VectorizeGEP) 854 return false; 855 856 // Currently, vector GEPs exist only with one index. 857 if (G->getNumIndices() != 1) 858 return false; 859 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 860 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 861 return false; 862 } 863 864 // We can't vectorize memory operations without target data 865 if (TD == 0 && IsSimpleLoadStore) 866 return false; 867 868 Type *T1, *T2; 869 getInstructionTypes(I, T1, T2); 870 871 // Not every type can be vectorized... 872 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 873 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 874 return false; 875 876 if (T1->getScalarSizeInBits() == 1) { 877 if (!Config.VectorizeBools) 878 return false; 879 } else { 880 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 881 return false; 882 } 883 884 if (T2->getScalarSizeInBits() == 1) { 885 if (!Config.VectorizeBools) 886 return false; 887 } else { 888 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 889 return false; 890 } 891 892 if (!Config.VectorizeFloats 893 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 894 return false; 895 896 // Don't vectorize target-specific types. 897 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 898 return false; 899 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 900 return false; 901 902 if ((!Config.VectorizePointers || TD == 0) && 903 (T1->getScalarType()->isPointerTy() || 904 T2->getScalarType()->isPointerTy())) 905 return false; 906 907 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 908 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 909 return false; 910 911 return true; 912 } 913 914 // This function returns true if the two provided instructions are compatible 915 // (meaning that they can be fused into a vector instruction). This assumes 916 // that I has already been determined to be vectorizable and that J is not 917 // in the use tree of I. 918 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 919 bool IsSimpleLoadStore, bool NonPow2Len, 920 int &CostSavings, int &FixedOrder) { 921 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 922 " <-> " << *J << "\n"); 923 924 CostSavings = 0; 925 FixedOrder = 0; 926 927 // Loads and stores can be merged if they have different alignments, 928 // but are otherwise the same. 929 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 930 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 931 return false; 932 933 Type *IT1, *IT2, *JT1, *JT2; 934 getInstructionTypes(I, IT1, IT2); 935 getInstructionTypes(J, JT1, JT2); 936 unsigned MaxTypeBits = std::max( 937 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 938 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 939 if (!VTTI && MaxTypeBits > Config.VectorBits) 940 return false; 941 942 // FIXME: handle addsub-type operations! 943 944 if (IsSimpleLoadStore) { 945 Value *IPtr, *JPtr; 946 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 947 int64_t OffsetInElmts = 0; 948 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 949 IAddressSpace, JAddressSpace, 950 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 951 FixedOrder = (int) OffsetInElmts; 952 unsigned BottomAlignment = IAlignment; 953 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 954 955 Type *aTypeI = isa<StoreInst>(I) ? 956 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 957 Type *aTypeJ = isa<StoreInst>(J) ? 958 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 959 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 960 961 if (Config.AlignedOnly) { 962 // An aligned load or store is possible only if the instruction 963 // with the lower offset has an alignment suitable for the 964 // vector type. 965 966 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 967 if (BottomAlignment < VecAlignment) 968 return false; 969 } 970 971 if (VTTI) { 972 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), aTypeI, 973 IAlignment, IAddressSpace); 974 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), aTypeJ, 975 JAlignment, JAddressSpace); 976 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType, 977 BottomAlignment, 978 IAddressSpace); 979 if (VCost > ICost + JCost) 980 return false; 981 982 // We don't want to fuse to a type that will be split, even 983 // if the two input types will also be split and there is no other 984 // associated cost. 985 unsigned VParts = VTTI->getNumberOfParts(VType); 986 if (VParts > 1) 987 return false; 988 else if (!VParts && VCost == ICost + JCost) 989 return false; 990 991 CostSavings = ICost + JCost - VCost; 992 } 993 } else { 994 return false; 995 } 996 } else if (VTTI) { 997 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 998 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 999 Type *VT1 = getVecTypeForPair(IT1, JT1), 1000 *VT2 = getVecTypeForPair(IT2, JT2); 1001 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 1002 1003 if (VCost > ICost + JCost) 1004 return false; 1005 1006 // We don't want to fuse to a type that will be split, even 1007 // if the two input types will also be split and there is no other 1008 // associated cost. 1009 unsigned VParts1 = VTTI->getNumberOfParts(VT1), 1010 VParts2 = VTTI->getNumberOfParts(VT2); 1011 if (VParts1 > 1 || VParts2 > 1) 1012 return false; 1013 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) 1014 return false; 1015 1016 CostSavings = ICost + JCost - VCost; 1017 } 1018 1019 // The powi intrinsic is special because only the first argument is 1020 // vectorized, the second arguments must be equal. 1021 CallInst *CI = dyn_cast<CallInst>(I); 1022 Function *FI; 1023 if (CI && (FI = CI->getCalledFunction()) && 1024 FI->getIntrinsicID() == Intrinsic::powi) { 1025 1026 Value *A1I = CI->getArgOperand(1), 1027 *A1J = cast<CallInst>(J)->getArgOperand(1); 1028 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1029 *A1JSCEV = SE->getSCEV(A1J); 1030 return (A1ISCEV == A1JSCEV); 1031 } 1032 1033 return true; 1034 } 1035 1036 // Figure out whether or not J uses I and update the users and write-set 1037 // structures associated with I. Specifically, Users represents the set of 1038 // instructions that depend on I. WriteSet represents the set 1039 // of memory locations that are dependent on I. If UpdateUsers is true, 1040 // and J uses I, then Users is updated to contain J and WriteSet is updated 1041 // to contain any memory locations to which J writes. The function returns 1042 // true if J uses I. By default, alias analysis is used to determine 1043 // whether J reads from memory that overlaps with a location in WriteSet. 1044 // If LoadMoveSet is not null, then it is a previously-computed multimap 1045 // where the key is the memory-based user instruction and the value is 1046 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1047 // then the alias analysis is not used. This is necessary because this 1048 // function is called during the process of moving instructions during 1049 // vectorization and the results of the alias analysis are not stable during 1050 // that process. 1051 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1052 AliasSetTracker &WriteSet, Instruction *I, 1053 Instruction *J, bool UpdateUsers, 1054 std::multimap<Value *, Value *> *LoadMoveSet) { 1055 bool UsesI = false; 1056 1057 // This instruction may already be marked as a user due, for example, to 1058 // being a member of a selected pair. 1059 if (Users.count(J)) 1060 UsesI = true; 1061 1062 if (!UsesI) 1063 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1064 JU != JE; ++JU) { 1065 Value *V = *JU; 1066 if (I == V || Users.count(V)) { 1067 UsesI = true; 1068 break; 1069 } 1070 } 1071 if (!UsesI && J->mayReadFromMemory()) { 1072 if (LoadMoveSet) { 1073 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 1074 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 1075 } else { 1076 for (AliasSetTracker::iterator W = WriteSet.begin(), 1077 WE = WriteSet.end(); W != WE; ++W) { 1078 if (W->aliasesUnknownInst(J, *AA)) { 1079 UsesI = true; 1080 break; 1081 } 1082 } 1083 } 1084 } 1085 1086 if (UsesI && UpdateUsers) { 1087 if (J->mayWriteToMemory()) WriteSet.add(J); 1088 Users.insert(J); 1089 } 1090 1091 return UsesI; 1092 } 1093 1094 // This function iterates over all instruction pairs in the provided 1095 // basic block and collects all candidate pairs for vectorization. 1096 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1097 BasicBlock::iterator &Start, 1098 std::multimap<Value *, Value *> &CandidatePairs, 1099 DenseSet<ValuePair> &FixedOrderPairs, 1100 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1101 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1102 BasicBlock::iterator E = BB.end(); 1103 if (Start == E) return false; 1104 1105 bool ShouldContinue = false, IAfterStart = false; 1106 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1107 if (I == Start) IAfterStart = true; 1108 1109 bool IsSimpleLoadStore; 1110 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1111 1112 // Look for an instruction with which to pair instruction *I... 1113 DenseSet<Value *> Users; 1114 AliasSetTracker WriteSet(*AA); 1115 bool JAfterStart = IAfterStart; 1116 BasicBlock::iterator J = llvm::next(I); 1117 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1118 if (J == Start) JAfterStart = true; 1119 1120 // Determine if J uses I, if so, exit the loop. 1121 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1122 if (Config.FastDep) { 1123 // Note: For this heuristic to be effective, independent operations 1124 // must tend to be intermixed. This is likely to be true from some 1125 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1126 // but otherwise may require some kind of reordering pass. 1127 1128 // When using fast dependency analysis, 1129 // stop searching after first use: 1130 if (UsesI) break; 1131 } else { 1132 if (UsesI) continue; 1133 } 1134 1135 // J does not use I, and comes before the first use of I, so it can be 1136 // merged with I if the instructions are compatible. 1137 int CostSavings, FixedOrder; 1138 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1139 CostSavings, FixedOrder)) continue; 1140 1141 // J is a candidate for merging with I. 1142 if (!PairableInsts.size() || 1143 PairableInsts[PairableInsts.size()-1] != I) { 1144 PairableInsts.push_back(I); 1145 } 1146 1147 CandidatePairs.insert(ValuePair(I, J)); 1148 if (VTTI) 1149 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1150 CostSavings)); 1151 1152 if (FixedOrder == 1) 1153 FixedOrderPairs.insert(ValuePair(I, J)); 1154 else if (FixedOrder == -1) 1155 FixedOrderPairs.insert(ValuePair(J, I)); 1156 1157 // The next call to this function must start after the last instruction 1158 // selected during this invocation. 1159 if (JAfterStart) { 1160 Start = llvm::next(J); 1161 IAfterStart = JAfterStart = false; 1162 } 1163 1164 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1165 << *I << " <-> " << *J << " (cost savings: " << 1166 CostSavings << ")\n"); 1167 1168 // If we have already found too many pairs, break here and this function 1169 // will be called again starting after the last instruction selected 1170 // during this invocation. 1171 if (PairableInsts.size() >= Config.MaxInsts) { 1172 ShouldContinue = true; 1173 break; 1174 } 1175 } 1176 1177 if (ShouldContinue) 1178 break; 1179 } 1180 1181 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1182 << " instructions with candidate pairs\n"); 1183 1184 return ShouldContinue; 1185 } 1186 1187 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1188 // it looks for pairs such that both members have an input which is an 1189 // output of PI or PJ. 1190 void BBVectorize::computePairsConnectedTo( 1191 std::multimap<Value *, Value *> &CandidatePairs, 1192 std::vector<Value *> &PairableInsts, 1193 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1194 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1195 ValuePair P) { 1196 StoreInst *SI, *SJ; 1197 1198 // For each possible pairing for this variable, look at the uses of 1199 // the first value... 1200 for (Value::use_iterator I = P.first->use_begin(), 1201 E = P.first->use_end(); I != E; ++I) { 1202 if (isa<LoadInst>(*I)) { 1203 // A pair cannot be connected to a load because the load only takes one 1204 // operand (the address) and it is a scalar even after vectorization. 1205 continue; 1206 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1207 P.first == SI->getPointerOperand()) { 1208 // Similarly, a pair cannot be connected to a store through its 1209 // pointer operand. 1210 continue; 1211 } 1212 1213 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1214 1215 // For each use of the first variable, look for uses of the second 1216 // variable... 1217 for (Value::use_iterator J = P.second->use_begin(), 1218 E2 = P.second->use_end(); J != E2; ++J) { 1219 if ((SJ = dyn_cast<StoreInst>(*J)) && 1220 P.second == SJ->getPointerOperand()) 1221 continue; 1222 1223 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 1224 1225 // Look for <I, J>: 1226 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1227 VPPair VP(P, ValuePair(*I, *J)); 1228 ConnectedPairs.insert(VP); 1229 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1230 } 1231 1232 // Look for <J, I>: 1233 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) { 1234 VPPair VP(P, ValuePair(*J, *I)); 1235 ConnectedPairs.insert(VP); 1236 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1237 } 1238 } 1239 1240 if (Config.SplatBreaksChain) continue; 1241 // Look for cases where just the first value in the pair is used by 1242 // both members of another pair (splatting). 1243 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1244 if ((SJ = dyn_cast<StoreInst>(*J)) && 1245 P.first == SJ->getPointerOperand()) 1246 continue; 1247 1248 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1249 VPPair VP(P, ValuePair(*I, *J)); 1250 ConnectedPairs.insert(VP); 1251 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1252 } 1253 } 1254 } 1255 1256 if (Config.SplatBreaksChain) return; 1257 // Look for cases where just the second value in the pair is used by 1258 // both members of another pair (splatting). 1259 for (Value::use_iterator I = P.second->use_begin(), 1260 E = P.second->use_end(); I != E; ++I) { 1261 if (isa<LoadInst>(*I)) 1262 continue; 1263 else if ((SI = dyn_cast<StoreInst>(*I)) && 1264 P.second == SI->getPointerOperand()) 1265 continue; 1266 1267 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1268 1269 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1270 if ((SJ = dyn_cast<StoreInst>(*J)) && 1271 P.second == SJ->getPointerOperand()) 1272 continue; 1273 1274 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1275 VPPair VP(P, ValuePair(*I, *J)); 1276 ConnectedPairs.insert(VP); 1277 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1278 } 1279 } 1280 } 1281 } 1282 1283 // This function figures out which pairs are connected. Two pairs are 1284 // connected if some output of the first pair forms an input to both members 1285 // of the second pair. 1286 void BBVectorize::computeConnectedPairs( 1287 std::multimap<Value *, Value *> &CandidatePairs, 1288 std::vector<Value *> &PairableInsts, 1289 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1290 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1291 1292 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1293 PE = PairableInsts.end(); PI != PE; ++PI) { 1294 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1295 1296 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1297 P != choiceRange.second; ++P) 1298 computePairsConnectedTo(CandidatePairs, PairableInsts, 1299 ConnectedPairs, PairConnectionTypes, *P); 1300 } 1301 1302 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1303 << " pair connections.\n"); 1304 } 1305 1306 // This function builds a set of use tuples such that <A, B> is in the set 1307 // if B is in the use tree of A. If B is in the use tree of A, then B 1308 // depends on the output of A. 1309 void BBVectorize::buildDepMap( 1310 BasicBlock &BB, 1311 std::multimap<Value *, Value *> &CandidatePairs, 1312 std::vector<Value *> &PairableInsts, 1313 DenseSet<ValuePair> &PairableInstUsers) { 1314 DenseSet<Value *> IsInPair; 1315 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1316 E = CandidatePairs.end(); C != E; ++C) { 1317 IsInPair.insert(C->first); 1318 IsInPair.insert(C->second); 1319 } 1320 1321 // Iterate through the basic block, recording all Users of each 1322 // pairable instruction. 1323 1324 BasicBlock::iterator E = BB.end(); 1325 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1326 if (IsInPair.find(I) == IsInPair.end()) continue; 1327 1328 DenseSet<Value *> Users; 1329 AliasSetTracker WriteSet(*AA); 1330 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1331 (void) trackUsesOfI(Users, WriteSet, I, J); 1332 1333 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1334 U != E; ++U) 1335 PairableInstUsers.insert(ValuePair(I, *U)); 1336 } 1337 } 1338 1339 // Returns true if an input to pair P is an output of pair Q and also an 1340 // input of pair Q is an output of pair P. If this is the case, then these 1341 // two pairs cannot be simultaneously fused. 1342 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1343 DenseSet<ValuePair> &PairableInstUsers, 1344 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1345 // Two pairs are in conflict if they are mutual Users of eachother. 1346 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1347 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1348 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1349 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1350 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1351 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1352 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1353 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1354 if (PairableInstUserMap) { 1355 // FIXME: The expensive part of the cycle check is not so much the cycle 1356 // check itself but this edge insertion procedure. This needs some 1357 // profiling and probably a different data structure (same is true of 1358 // most uses of std::multimap). 1359 if (PUsesQ) { 1360 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1361 if (!isSecondInIteratorPair(P, QPairRange)) 1362 PairableInstUserMap->insert(VPPair(Q, P)); 1363 } 1364 if (QUsesP) { 1365 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1366 if (!isSecondInIteratorPair(Q, PPairRange)) 1367 PairableInstUserMap->insert(VPPair(P, Q)); 1368 } 1369 } 1370 1371 return (QUsesP && PUsesQ); 1372 } 1373 1374 // This function walks the use graph of current pairs to see if, starting 1375 // from P, the walk returns to P. 1376 bool BBVectorize::pairWillFormCycle(ValuePair P, 1377 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1378 DenseSet<ValuePair> &CurrentPairs) { 1379 DEBUG(if (DebugCycleCheck) 1380 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1381 << *P.second << "\n"); 1382 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1383 // contains non-direct associations. 1384 DenseSet<ValuePair> Visited; 1385 SmallVector<ValuePair, 32> Q; 1386 // General depth-first post-order traversal: 1387 Q.push_back(P); 1388 do { 1389 ValuePair QTop = Q.pop_back_val(); 1390 Visited.insert(QTop); 1391 1392 DEBUG(if (DebugCycleCheck) 1393 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1394 << *QTop.second << "\n"); 1395 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1396 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1397 C != QPairRange.second; ++C) { 1398 if (C->second == P) { 1399 DEBUG(dbgs() 1400 << "BBV: rejected to prevent non-trivial cycle formation: " 1401 << *C->first.first << " <-> " << *C->first.second << "\n"); 1402 return true; 1403 } 1404 1405 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1406 Q.push_back(C->second); 1407 } 1408 } while (!Q.empty()); 1409 1410 return false; 1411 } 1412 1413 // This function builds the initial tree of connected pairs with the 1414 // pair J at the root. 1415 void BBVectorize::buildInitialTreeFor( 1416 std::multimap<Value *, Value *> &CandidatePairs, 1417 std::vector<Value *> &PairableInsts, 1418 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1419 DenseSet<ValuePair> &PairableInstUsers, 1420 DenseMap<Value *, Value *> &ChosenPairs, 1421 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1422 // Each of these pairs is viewed as the root node of a Tree. The Tree 1423 // is then walked (depth-first). As this happens, we keep track of 1424 // the pairs that compose the Tree and the maximum depth of the Tree. 1425 SmallVector<ValuePairWithDepth, 32> Q; 1426 // General depth-first post-order traversal: 1427 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1428 do { 1429 ValuePairWithDepth QTop = Q.back(); 1430 1431 // Push each child onto the queue: 1432 bool MoreChildren = false; 1433 size_t MaxChildDepth = QTop.second; 1434 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1435 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1436 k != qtRange.second; ++k) { 1437 // Make sure that this child pair is still a candidate: 1438 bool IsStillCand = false; 1439 VPIteratorPair checkRange = 1440 CandidatePairs.equal_range(k->second.first); 1441 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1442 m != checkRange.second; ++m) { 1443 if (m->second == k->second.second) { 1444 IsStillCand = true; 1445 break; 1446 } 1447 } 1448 1449 if (IsStillCand) { 1450 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1451 if (C == Tree.end()) { 1452 size_t d = getDepthFactor(k->second.first); 1453 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1454 MoreChildren = true; 1455 } else { 1456 MaxChildDepth = std::max(MaxChildDepth, C->second); 1457 } 1458 } 1459 } 1460 1461 if (!MoreChildren) { 1462 // Record the current pair as part of the Tree: 1463 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1464 Q.pop_back(); 1465 } 1466 } while (!Q.empty()); 1467 } 1468 1469 // Given some initial tree, prune it by removing conflicting pairs (pairs 1470 // that cannot be simultaneously chosen for vectorization). 1471 void BBVectorize::pruneTreeFor( 1472 std::multimap<Value *, Value *> &CandidatePairs, 1473 std::vector<Value *> &PairableInsts, 1474 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1475 DenseSet<ValuePair> &PairableInstUsers, 1476 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1477 DenseMap<Value *, Value *> &ChosenPairs, 1478 DenseMap<ValuePair, size_t> &Tree, 1479 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1480 bool UseCycleCheck) { 1481 SmallVector<ValuePairWithDepth, 32> Q; 1482 // General depth-first post-order traversal: 1483 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1484 do { 1485 ValuePairWithDepth QTop = Q.pop_back_val(); 1486 PrunedTree.insert(QTop.first); 1487 1488 // Visit each child, pruning as necessary... 1489 SmallVector<ValuePairWithDepth, 8> BestChildren; 1490 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1491 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1492 K != QTopRange.second; ++K) { 1493 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1494 if (C == Tree.end()) continue; 1495 1496 // This child is in the Tree, now we need to make sure it is the 1497 // best of any conflicting children. There could be multiple 1498 // conflicting children, so first, determine if we're keeping 1499 // this child, then delete conflicting children as necessary. 1500 1501 // It is also necessary to guard against pairing-induced 1502 // dependencies. Consider instructions a .. x .. y .. b 1503 // such that (a,b) are to be fused and (x,y) are to be fused 1504 // but a is an input to x and b is an output from y. This 1505 // means that y cannot be moved after b but x must be moved 1506 // after b for (a,b) to be fused. In other words, after 1507 // fusing (a,b) we have y .. a/b .. x where y is an input 1508 // to a/b and x is an output to a/b: x and y can no longer 1509 // be legally fused. To prevent this condition, we must 1510 // make sure that a child pair added to the Tree is not 1511 // both an input and output of an already-selected pair. 1512 1513 // Pairing-induced dependencies can also form from more complicated 1514 // cycles. The pair vs. pair conflicts are easy to check, and so 1515 // that is done explicitly for "fast rejection", and because for 1516 // child vs. child conflicts, we may prefer to keep the current 1517 // pair in preference to the already-selected child. 1518 DenseSet<ValuePair> CurrentPairs; 1519 1520 bool CanAdd = true; 1521 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1522 = BestChildren.begin(), E2 = BestChildren.end(); 1523 C2 != E2; ++C2) { 1524 if (C2->first.first == C->first.first || 1525 C2->first.first == C->first.second || 1526 C2->first.second == C->first.first || 1527 C2->first.second == C->first.second || 1528 pairsConflict(C2->first, C->first, PairableInstUsers, 1529 UseCycleCheck ? &PairableInstUserMap : 0)) { 1530 if (C2->second >= C->second) { 1531 CanAdd = false; 1532 break; 1533 } 1534 1535 CurrentPairs.insert(C2->first); 1536 } 1537 } 1538 if (!CanAdd) continue; 1539 1540 // Even worse, this child could conflict with another node already 1541 // selected for the Tree. If that is the case, ignore this child. 1542 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1543 E2 = PrunedTree.end(); T != E2; ++T) { 1544 if (T->first == C->first.first || 1545 T->first == C->first.second || 1546 T->second == C->first.first || 1547 T->second == C->first.second || 1548 pairsConflict(*T, C->first, PairableInstUsers, 1549 UseCycleCheck ? &PairableInstUserMap : 0)) { 1550 CanAdd = false; 1551 break; 1552 } 1553 1554 CurrentPairs.insert(*T); 1555 } 1556 if (!CanAdd) continue; 1557 1558 // And check the queue too... 1559 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1560 E2 = Q.end(); C2 != E2; ++C2) { 1561 if (C2->first.first == C->first.first || 1562 C2->first.first == C->first.second || 1563 C2->first.second == C->first.first || 1564 C2->first.second == C->first.second || 1565 pairsConflict(C2->first, C->first, PairableInstUsers, 1566 UseCycleCheck ? &PairableInstUserMap : 0)) { 1567 CanAdd = false; 1568 break; 1569 } 1570 1571 CurrentPairs.insert(C2->first); 1572 } 1573 if (!CanAdd) continue; 1574 1575 // Last but not least, check for a conflict with any of the 1576 // already-chosen pairs. 1577 for (DenseMap<Value *, Value *>::iterator C2 = 1578 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1579 C2 != E2; ++C2) { 1580 if (pairsConflict(*C2, C->first, PairableInstUsers, 1581 UseCycleCheck ? &PairableInstUserMap : 0)) { 1582 CanAdd = false; 1583 break; 1584 } 1585 1586 CurrentPairs.insert(*C2); 1587 } 1588 if (!CanAdd) continue; 1589 1590 // To check for non-trivial cycles formed by the addition of the 1591 // current pair we've formed a list of all relevant pairs, now use a 1592 // graph walk to check for a cycle. We start from the current pair and 1593 // walk the use tree to see if we again reach the current pair. If we 1594 // do, then the current pair is rejected. 1595 1596 // FIXME: It may be more efficient to use a topological-ordering 1597 // algorithm to improve the cycle check. This should be investigated. 1598 if (UseCycleCheck && 1599 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1600 continue; 1601 1602 // This child can be added, but we may have chosen it in preference 1603 // to an already-selected child. Check for this here, and if a 1604 // conflict is found, then remove the previously-selected child 1605 // before adding this one in its place. 1606 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1607 = BestChildren.begin(); C2 != BestChildren.end();) { 1608 if (C2->first.first == C->first.first || 1609 C2->first.first == C->first.second || 1610 C2->first.second == C->first.first || 1611 C2->first.second == C->first.second || 1612 pairsConflict(C2->first, C->first, PairableInstUsers)) 1613 C2 = BestChildren.erase(C2); 1614 else 1615 ++C2; 1616 } 1617 1618 BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); 1619 } 1620 1621 for (SmallVector<ValuePairWithDepth, 8>::iterator C 1622 = BestChildren.begin(), E2 = BestChildren.end(); 1623 C != E2; ++C) { 1624 size_t DepthF = getDepthFactor(C->first.first); 1625 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1626 } 1627 } while (!Q.empty()); 1628 } 1629 1630 // This function finds the best tree of mututally-compatible connected 1631 // pairs, given the choice of root pairs as an iterator range. 1632 void BBVectorize::findBestTreeFor( 1633 std::multimap<Value *, Value *> &CandidatePairs, 1634 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1635 std::vector<Value *> &PairableInsts, 1636 DenseSet<ValuePair> &FixedOrderPairs, 1637 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1638 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1639 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1640 DenseSet<ValuePair> &PairableInstUsers, 1641 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1642 DenseMap<Value *, Value *> &ChosenPairs, 1643 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1644 int &BestEffSize, VPIteratorPair ChoiceRange, 1645 bool UseCycleCheck) { 1646 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1647 J != ChoiceRange.second; ++J) { 1648 1649 // Before going any further, make sure that this pair does not 1650 // conflict with any already-selected pairs (see comment below 1651 // near the Tree pruning for more details). 1652 DenseSet<ValuePair> ChosenPairSet; 1653 bool DoesConflict = false; 1654 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1655 E = ChosenPairs.end(); C != E; ++C) { 1656 if (pairsConflict(*C, *J, PairableInstUsers, 1657 UseCycleCheck ? &PairableInstUserMap : 0)) { 1658 DoesConflict = true; 1659 break; 1660 } 1661 1662 ChosenPairSet.insert(*C); 1663 } 1664 if (DoesConflict) continue; 1665 1666 if (UseCycleCheck && 1667 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1668 continue; 1669 1670 DenseMap<ValuePair, size_t> Tree; 1671 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1672 PairableInstUsers, ChosenPairs, Tree, *J); 1673 1674 // Because we'll keep the child with the largest depth, the largest 1675 // depth is still the same in the unpruned Tree. 1676 size_t MaxDepth = Tree.lookup(*J); 1677 1678 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1679 << *J->first << " <-> " << *J->second << "} of depth " << 1680 MaxDepth << " and size " << Tree.size() << "\n"); 1681 1682 // At this point the Tree has been constructed, but, may contain 1683 // contradictory children (meaning that different children of 1684 // some tree node may be attempting to fuse the same instruction). 1685 // So now we walk the tree again, in the case of a conflict, 1686 // keep only the child with the largest depth. To break a tie, 1687 // favor the first child. 1688 1689 DenseSet<ValuePair> PrunedTree; 1690 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1691 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1692 PrunedTree, *J, UseCycleCheck); 1693 1694 int EffSize = 0; 1695 if (VTTI) { 1696 DenseSet<Value *> PrunedTreeInstrs; 1697 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1698 E = PrunedTree.end(); S != E; ++S) { 1699 PrunedTreeInstrs.insert(S->first); 1700 PrunedTreeInstrs.insert(S->second); 1701 } 1702 1703 // The set of pairs that have already contributed to the total cost. 1704 DenseSet<ValuePair> IncomingPairs; 1705 1706 // If the cost model were perfect, this might not be necessary; but we 1707 // need to make sure that we don't get stuck vectorizing our own 1708 // shuffle chains. 1709 bool HasNontrivialInsts = false; 1710 1711 // The node weights represent the cost savings associated with 1712 // fusing the pair of instructions. 1713 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1714 E = PrunedTree.end(); S != E; ++S) { 1715 if (!isa<ShuffleVectorInst>(S->first) && 1716 !isa<InsertElementInst>(S->first) && 1717 !isa<ExtractElementInst>(S->first)) 1718 HasNontrivialInsts = true; 1719 1720 bool FlipOrder = false; 1721 1722 if (getDepthFactor(S->first)) { 1723 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1724 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1725 << *S->first << " <-> " << *S->second << "} = " << 1726 ESContrib << "\n"); 1727 EffSize += ESContrib; 1728 } 1729 1730 // The edge weights contribute in a negative sense: they represent 1731 // the cost of shuffles. 1732 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); 1733 if (IP.first != ConnectedPairDeps.end()) { 1734 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1735 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1736 Q != IP.second; ++Q) { 1737 if (!PrunedTree.count(Q->second)) 1738 continue; 1739 DenseMap<VPPair, unsigned>::iterator R = 1740 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1741 assert(R != PairConnectionTypes.end() && 1742 "Cannot find pair connection type"); 1743 if (R->second == PairConnectionDirect) 1744 ++NumDepsDirect; 1745 else if (R->second == PairConnectionSwap) 1746 ++NumDepsSwap; 1747 } 1748 1749 // If there are more swaps than direct connections, then 1750 // the pair order will be flipped during fusion. So the real 1751 // number of swaps is the minimum number. 1752 FlipOrder = !FixedOrderPairs.count(*S) && 1753 ((NumDepsSwap > NumDepsDirect) || 1754 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1755 1756 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1757 Q != IP.second; ++Q) { 1758 if (!PrunedTree.count(Q->second)) 1759 continue; 1760 DenseMap<VPPair, unsigned>::iterator R = 1761 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1762 assert(R != PairConnectionTypes.end() && 1763 "Cannot find pair connection type"); 1764 Type *Ty1 = Q->second.first->getType(), 1765 *Ty2 = Q->second.second->getType(); 1766 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1767 if ((R->second == PairConnectionDirect && FlipOrder) || 1768 (R->second == PairConnectionSwap && !FlipOrder) || 1769 R->second == PairConnectionSplat) { 1770 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1771 VTy, VTy); 1772 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1773 *Q->second.first << " <-> " << *Q->second.second << 1774 "} -> {" << 1775 *S->first << " <-> " << *S->second << "} = " << 1776 ESContrib << "\n"); 1777 EffSize -= ESContrib; 1778 } 1779 } 1780 } 1781 1782 // Compute the cost of outgoing edges. We assume that edges outgoing 1783 // to shuffles, inserts or extracts can be merged, and so contribute 1784 // no additional cost. 1785 if (!S->first->getType()->isVoidTy()) { 1786 Type *Ty1 = S->first->getType(), 1787 *Ty2 = S->second->getType(); 1788 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1789 1790 bool NeedsExtraction = false; 1791 for (Value::use_iterator I = S->first->use_begin(), 1792 IE = S->first->use_end(); I != IE; ++I) { 1793 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1794 // Shuffle can be folded if it has no other input 1795 if (isa<UndefValue>(SI->getOperand(1))) 1796 continue; 1797 } 1798 if (isa<ExtractElementInst>(*I)) 1799 continue; 1800 if (PrunedTreeInstrs.count(*I)) 1801 continue; 1802 NeedsExtraction = true; 1803 break; 1804 } 1805 1806 if (NeedsExtraction) { 1807 int ESContrib; 1808 if (Ty1->isVectorTy()) 1809 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1810 Ty1, VTy); 1811 else 1812 ESContrib = (int) VTTI->getVectorInstrCost( 1813 Instruction::ExtractElement, VTy, 0); 1814 1815 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1816 *S->first << "} = " << ESContrib << "\n"); 1817 EffSize -= ESContrib; 1818 } 1819 1820 NeedsExtraction = false; 1821 for (Value::use_iterator I = S->second->use_begin(), 1822 IE = S->second->use_end(); I != IE; ++I) { 1823 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1824 // Shuffle can be folded if it has no other input 1825 if (isa<UndefValue>(SI->getOperand(1))) 1826 continue; 1827 } 1828 if (isa<ExtractElementInst>(*I)) 1829 continue; 1830 if (PrunedTreeInstrs.count(*I)) 1831 continue; 1832 NeedsExtraction = true; 1833 break; 1834 } 1835 1836 if (NeedsExtraction) { 1837 int ESContrib; 1838 if (Ty2->isVectorTy()) 1839 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1840 Ty2, VTy); 1841 else 1842 ESContrib = (int) VTTI->getVectorInstrCost( 1843 Instruction::ExtractElement, VTy, 1); 1844 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1845 *S->second << "} = " << ESContrib << "\n"); 1846 EffSize -= ESContrib; 1847 } 1848 } 1849 1850 // Compute the cost of incoming edges. 1851 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 1852 Instruction *S1 = cast<Instruction>(S->first), 1853 *S2 = cast<Instruction>(S->second); 1854 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 1855 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 1856 1857 // Combining constants into vector constants (or small vector 1858 // constants into larger ones are assumed free). 1859 if (isa<Constant>(O1) && isa<Constant>(O2)) 1860 continue; 1861 1862 if (FlipOrder) 1863 std::swap(O1, O2); 1864 1865 ValuePair VP = ValuePair(O1, O2); 1866 ValuePair VPR = ValuePair(O2, O1); 1867 1868 // Internal edges are not handled here. 1869 if (PrunedTree.count(VP) || PrunedTree.count(VPR)) 1870 continue; 1871 1872 Type *Ty1 = O1->getType(), 1873 *Ty2 = O2->getType(); 1874 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1875 1876 // Combining vector operations of the same type is also assumed 1877 // folded with other operations. 1878 if (Ty1 == Ty2) { 1879 // If both are insert elements, then both can be widened. 1880 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), 1881 *IEO2 = dyn_cast<InsertElementInst>(O2); 1882 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) 1883 continue; 1884 // If both are extract elements, and both have the same input 1885 // type, then they can be replaced with a shuffle 1886 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 1887 *EIO2 = dyn_cast<ExtractElementInst>(O2); 1888 if (EIO1 && EIO2 && 1889 EIO1->getOperand(0)->getType() == 1890 EIO2->getOperand(0)->getType()) 1891 continue; 1892 // If both are a shuffle with equal operand types and only two 1893 // unqiue operands, then they can be replaced with a single 1894 // shuffle 1895 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 1896 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 1897 if (SIO1 && SIO2 && 1898 SIO1->getOperand(0)->getType() == 1899 SIO2->getOperand(0)->getType()) { 1900 SmallSet<Value *, 4> SIOps; 1901 SIOps.insert(SIO1->getOperand(0)); 1902 SIOps.insert(SIO1->getOperand(1)); 1903 SIOps.insert(SIO2->getOperand(0)); 1904 SIOps.insert(SIO2->getOperand(1)); 1905 if (SIOps.size() <= 2) 1906 continue; 1907 } 1908 } 1909 1910 int ESContrib; 1911 // This pair has already been formed. 1912 if (IncomingPairs.count(VP)) { 1913 continue; 1914 } else if (IncomingPairs.count(VPR)) { 1915 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1916 VTy, VTy); 1917 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 1918 ESContrib = (int) VTTI->getVectorInstrCost( 1919 Instruction::InsertElement, VTy, 0); 1920 ESContrib += (int) VTTI->getVectorInstrCost( 1921 Instruction::InsertElement, VTy, 1); 1922 } else if (!Ty1->isVectorTy()) { 1923 // O1 needs to be inserted into a vector of size O2, and then 1924 // both need to be shuffled together. 1925 ESContrib = (int) VTTI->getVectorInstrCost( 1926 Instruction::InsertElement, Ty2, 0); 1927 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 1928 VTy, Ty2); 1929 } else if (!Ty2->isVectorTy()) { 1930 // O2 needs to be inserted into a vector of size O1, and then 1931 // both need to be shuffled together. 1932 ESContrib = (int) VTTI->getVectorInstrCost( 1933 Instruction::InsertElement, Ty1, 0); 1934 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 1935 VTy, Ty1); 1936 } else { 1937 Type *TyBig = Ty1, *TySmall = Ty2; 1938 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 1939 std::swap(TyBig, TySmall); 1940 1941 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1942 VTy, TyBig); 1943 if (TyBig != TySmall) 1944 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 1945 TyBig, TySmall); 1946 } 1947 1948 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 1949 << *O1 << " <-> " << *O2 << "} = " << 1950 ESContrib << "\n"); 1951 EffSize -= ESContrib; 1952 IncomingPairs.insert(VP); 1953 } 1954 } 1955 } 1956 1957 if (!HasNontrivialInsts) { 1958 DEBUG(if (DebugPairSelection) dbgs() << 1959 "\tNo non-trivial instructions in tree;" 1960 " override to zero effective size\n"); 1961 EffSize = 0; 1962 } 1963 } else { 1964 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1965 E = PrunedTree.end(); S != E; ++S) 1966 EffSize += (int) getDepthFactor(S->first); 1967 } 1968 1969 DEBUG(if (DebugPairSelection) 1970 dbgs() << "BBV: found pruned Tree for pair {" 1971 << *J->first << " <-> " << *J->second << "} of depth " << 1972 MaxDepth << " and size " << PrunedTree.size() << 1973 " (effective size: " << EffSize << ")\n"); 1974 if (((VTTI && !UseChainDepthWithTI) || 1975 MaxDepth >= Config.ReqChainDepth) && 1976 EffSize > 0 && EffSize > BestEffSize) { 1977 BestMaxDepth = MaxDepth; 1978 BestEffSize = EffSize; 1979 BestTree = PrunedTree; 1980 } 1981 } 1982 } 1983 1984 // Given the list of candidate pairs, this function selects those 1985 // that will be fused into vector instructions. 1986 void BBVectorize::choosePairs( 1987 std::multimap<Value *, Value *> &CandidatePairs, 1988 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1989 std::vector<Value *> &PairableInsts, 1990 DenseSet<ValuePair> &FixedOrderPairs, 1991 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1992 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1993 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1994 DenseSet<ValuePair> &PairableInstUsers, 1995 DenseMap<Value *, Value *>& ChosenPairs) { 1996 bool UseCycleCheck = 1997 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1998 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1999 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 2000 E = PairableInsts.end(); I != E; ++I) { 2001 // The number of possible pairings for this variable: 2002 size_t NumChoices = CandidatePairs.count(*I); 2003 if (!NumChoices) continue; 2004 2005 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 2006 2007 // The best pair to choose and its tree: 2008 size_t BestMaxDepth = 0; 2009 int BestEffSize = 0; 2010 DenseSet<ValuePair> BestTree; 2011 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 2012 PairableInsts, FixedOrderPairs, PairConnectionTypes, 2013 ConnectedPairs, ConnectedPairDeps, 2014 PairableInstUsers, PairableInstUserMap, ChosenPairs, 2015 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 2016 UseCycleCheck); 2017 2018 // A tree has been chosen (or not) at this point. If no tree was 2019 // chosen, then this instruction, I, cannot be paired (and is no longer 2020 // considered). 2021 2022 DEBUG(if (BestTree.size() > 0) 2023 dbgs() << "BBV: selected pairs in the best tree for: " 2024 << *cast<Instruction>(*I) << "\n"); 2025 2026 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 2027 SE2 = BestTree.end(); S != SE2; ++S) { 2028 // Insert the members of this tree into the list of chosen pairs. 2029 ChosenPairs.insert(ValuePair(S->first, S->second)); 2030 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 2031 *S->second << "\n"); 2032 2033 // Remove all candidate pairs that have values in the chosen tree. 2034 for (std::multimap<Value *, Value *>::iterator K = 2035 CandidatePairs.begin(); K != CandidatePairs.end();) { 2036 if (K->first == S->first || K->second == S->first || 2037 K->second == S->second || K->first == S->second) { 2038 // Don't remove the actual pair chosen so that it can be used 2039 // in subsequent tree selections. 2040 if (!(K->first == S->first && K->second == S->second)) 2041 CandidatePairs.erase(K++); 2042 else 2043 ++K; 2044 } else { 2045 ++K; 2046 } 2047 } 2048 } 2049 } 2050 2051 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2052 } 2053 2054 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2055 unsigned n = 0) { 2056 if (!I->hasName()) 2057 return ""; 2058 2059 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2060 (n > 0 ? "." + utostr(n) : "")).str(); 2061 } 2062 2063 // Returns the value that is to be used as the pointer input to the vector 2064 // instruction that fuses I with J. 2065 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2066 Instruction *I, Instruction *J, unsigned o) { 2067 Value *IPtr, *JPtr; 2068 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2069 int64_t OffsetInElmts; 2070 2071 // Note: the analysis might fail here, that is why the pair order has 2072 // been precomputed (OffsetInElmts must be unused here). 2073 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2074 IAddressSpace, JAddressSpace, 2075 OffsetInElmts, false); 2076 2077 // The pointer value is taken to be the one with the lowest offset. 2078 Value *VPtr = IPtr; 2079 2080 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 2081 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 2082 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2083 Type *VArgPtrType = PointerType::get(VArgType, 2084 cast<PointerType>(IPtr->getType())->getAddressSpace()); 2085 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2086 /* insert before */ I); 2087 } 2088 2089 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2090 unsigned MaskOffset, unsigned NumInElem, 2091 unsigned NumInElem1, unsigned IdxOffset, 2092 std::vector<Constant*> &Mask) { 2093 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 2094 for (unsigned v = 0; v < NumElem1; ++v) { 2095 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2096 if (m < 0) { 2097 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2098 } else { 2099 unsigned mm = m + (int) IdxOffset; 2100 if (m >= (int) NumInElem1) 2101 mm += (int) NumInElem; 2102 2103 Mask[v+MaskOffset] = 2104 ConstantInt::get(Type::getInt32Ty(Context), mm); 2105 } 2106 } 2107 } 2108 2109 // Returns the value that is to be used as the vector-shuffle mask to the 2110 // vector instruction that fuses I with J. 2111 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2112 Instruction *I, Instruction *J) { 2113 // This is the shuffle mask. We need to append the second 2114 // mask to the first, and the numbers need to be adjusted. 2115 2116 Type *ArgTypeI = I->getType(); 2117 Type *ArgTypeJ = J->getType(); 2118 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2119 2120 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 2121 2122 // Get the total number of elements in the fused vector type. 2123 // By definition, this must equal the number of elements in 2124 // the final mask. 2125 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 2126 std::vector<Constant*> Mask(NumElem); 2127 2128 Type *OpTypeI = I->getOperand(0)->getType(); 2129 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 2130 Type *OpTypeJ = J->getOperand(0)->getType(); 2131 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 2132 2133 // The fused vector will be: 2134 // ----------------------------------------------------- 2135 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2136 // ----------------------------------------------------- 2137 // from which we'll extract NumElem total elements (where the first NumElemI 2138 // of them come from the mask in I and the remainder come from the mask 2139 // in J. 2140 2141 // For the mask from the first pair... 2142 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2143 0, Mask); 2144 2145 // For the mask from the second pair... 2146 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2147 NumInElemI, Mask); 2148 2149 return ConstantVector::get(Mask); 2150 } 2151 2152 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2153 Instruction *J, unsigned o, Value *&LOp, 2154 unsigned numElemL, 2155 Type *ArgTypeL, Type *ArgTypeH, 2156 bool IBeforeJ, unsigned IdxOff) { 2157 bool ExpandedIEChain = false; 2158 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2159 // If we have a pure insertelement chain, then this can be rewritten 2160 // into a chain that directly builds the larger type. 2161 if (isPureIEChain(LIE)) { 2162 SmallVector<Value *, 8> VectElemts(numElemL, 2163 UndefValue::get(ArgTypeL->getScalarType())); 2164 InsertElementInst *LIENext = LIE; 2165 do { 2166 unsigned Idx = 2167 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2168 VectElemts[Idx] = LIENext->getOperand(1); 2169 } while ((LIENext = 2170 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2171 2172 LIENext = 0; 2173 Value *LIEPrev = UndefValue::get(ArgTypeH); 2174 for (unsigned i = 0; i < numElemL; ++i) { 2175 if (isa<UndefValue>(VectElemts[i])) continue; 2176 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2177 ConstantInt::get(Type::getInt32Ty(Context), 2178 i + IdxOff), 2179 getReplacementName(IBeforeJ ? I : J, 2180 true, o, i+1)); 2181 LIENext->insertBefore(IBeforeJ ? J : I); 2182 LIEPrev = LIENext; 2183 } 2184 2185 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2186 ExpandedIEChain = true; 2187 } 2188 } 2189 2190 return ExpandedIEChain; 2191 } 2192 2193 // Returns the value to be used as the specified operand of the vector 2194 // instruction that fuses I with J. 2195 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2196 Instruction *J, unsigned o, bool IBeforeJ) { 2197 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2198 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2199 2200 // Compute the fused vector type for this operand 2201 Type *ArgTypeI = I->getOperand(o)->getType(); 2202 Type *ArgTypeJ = J->getOperand(o)->getType(); 2203 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2204 2205 Instruction *L = I, *H = J; 2206 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2207 2208 unsigned numElemL; 2209 if (ArgTypeL->isVectorTy()) 2210 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 2211 else 2212 numElemL = 1; 2213 2214 unsigned numElemH; 2215 if (ArgTypeH->isVectorTy()) 2216 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 2217 else 2218 numElemH = 1; 2219 2220 Value *LOp = L->getOperand(o); 2221 Value *HOp = H->getOperand(o); 2222 unsigned numElem = VArgType->getNumElements(); 2223 2224 // First, we check if we can reuse the "original" vector outputs (if these 2225 // exist). We might need a shuffle. 2226 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2227 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2228 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2229 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2230 2231 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2232 // optimization. The input vectors to the shuffle might be a different 2233 // length from the shuffle outputs. Unfortunately, the replacement 2234 // shuffle mask has already been formed, and the mask entries are sensitive 2235 // to the sizes of the inputs. 2236 bool IsSizeChangeShuffle = 2237 isa<ShuffleVectorInst>(L) && 2238 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2239 2240 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2241 // We can have at most two unique vector inputs. 2242 bool CanUseInputs = true; 2243 Value *I1, *I2 = 0; 2244 if (LEE) { 2245 I1 = LEE->getOperand(0); 2246 } else { 2247 I1 = LSV->getOperand(0); 2248 I2 = LSV->getOperand(1); 2249 if (I2 == I1 || isa<UndefValue>(I2)) 2250 I2 = 0; 2251 } 2252 2253 if (HEE) { 2254 Value *I3 = HEE->getOperand(0); 2255 if (!I2 && I3 != I1) 2256 I2 = I3; 2257 else if (I3 != I1 && I3 != I2) 2258 CanUseInputs = false; 2259 } else { 2260 Value *I3 = HSV->getOperand(0); 2261 if (!I2 && I3 != I1) 2262 I2 = I3; 2263 else if (I3 != I1 && I3 != I2) 2264 CanUseInputs = false; 2265 2266 if (CanUseInputs) { 2267 Value *I4 = HSV->getOperand(1); 2268 if (!isa<UndefValue>(I4)) { 2269 if (!I2 && I4 != I1) 2270 I2 = I4; 2271 else if (I4 != I1 && I4 != I2) 2272 CanUseInputs = false; 2273 } 2274 } 2275 } 2276 2277 if (CanUseInputs) { 2278 unsigned LOpElem = 2279 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2280 ->getNumElements(); 2281 unsigned HOpElem = 2282 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2283 ->getNumElements(); 2284 2285 // We have one or two input vectors. We need to map each index of the 2286 // operands to the index of the original vector. 2287 SmallVector<std::pair<int, int>, 8> II(numElem); 2288 for (unsigned i = 0; i < numElemL; ++i) { 2289 int Idx, INum; 2290 if (LEE) { 2291 Idx = 2292 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2293 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2294 } else { 2295 Idx = LSV->getMaskValue(i); 2296 if (Idx < (int) LOpElem) { 2297 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2298 } else { 2299 Idx -= LOpElem; 2300 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2301 } 2302 } 2303 2304 II[i] = std::pair<int, int>(Idx, INum); 2305 } 2306 for (unsigned i = 0; i < numElemH; ++i) { 2307 int Idx, INum; 2308 if (HEE) { 2309 Idx = 2310 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2311 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2312 } else { 2313 Idx = HSV->getMaskValue(i); 2314 if (Idx < (int) HOpElem) { 2315 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2316 } else { 2317 Idx -= HOpElem; 2318 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2319 } 2320 } 2321 2322 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2323 } 2324 2325 // We now have an array which tells us from which index of which 2326 // input vector each element of the operand comes. 2327 VectorType *I1T = cast<VectorType>(I1->getType()); 2328 unsigned I1Elem = I1T->getNumElements(); 2329 2330 if (!I2) { 2331 // In this case there is only one underlying vector input. Check for 2332 // the trivial case where we can use the input directly. 2333 if (I1Elem == numElem) { 2334 bool ElemInOrder = true; 2335 for (unsigned i = 0; i < numElem; ++i) { 2336 if (II[i].first != (int) i && II[i].first != -1) { 2337 ElemInOrder = false; 2338 break; 2339 } 2340 } 2341 2342 if (ElemInOrder) 2343 return I1; 2344 } 2345 2346 // A shuffle is needed. 2347 std::vector<Constant *> Mask(numElem); 2348 for (unsigned i = 0; i < numElem; ++i) { 2349 int Idx = II[i].first; 2350 if (Idx == -1) 2351 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2352 else 2353 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2354 } 2355 2356 Instruction *S = 2357 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2358 ConstantVector::get(Mask), 2359 getReplacementName(IBeforeJ ? I : J, 2360 true, o)); 2361 S->insertBefore(IBeforeJ ? J : I); 2362 return S; 2363 } 2364 2365 VectorType *I2T = cast<VectorType>(I2->getType()); 2366 unsigned I2Elem = I2T->getNumElements(); 2367 2368 // This input comes from two distinct vectors. The first step is to 2369 // make sure that both vectors are the same length. If not, the 2370 // smaller one will need to grow before they can be shuffled together. 2371 if (I1Elem < I2Elem) { 2372 std::vector<Constant *> Mask(I2Elem); 2373 unsigned v = 0; 2374 for (; v < I1Elem; ++v) 2375 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2376 for (; v < I2Elem; ++v) 2377 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2378 2379 Instruction *NewI1 = 2380 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2381 ConstantVector::get(Mask), 2382 getReplacementName(IBeforeJ ? I : J, 2383 true, o, 1)); 2384 NewI1->insertBefore(IBeforeJ ? J : I); 2385 I1 = NewI1; 2386 I1T = I2T; 2387 I1Elem = I2Elem; 2388 } else if (I1Elem > I2Elem) { 2389 std::vector<Constant *> Mask(I1Elem); 2390 unsigned v = 0; 2391 for (; v < I2Elem; ++v) 2392 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2393 for (; v < I1Elem; ++v) 2394 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2395 2396 Instruction *NewI2 = 2397 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2398 ConstantVector::get(Mask), 2399 getReplacementName(IBeforeJ ? I : J, 2400 true, o, 1)); 2401 NewI2->insertBefore(IBeforeJ ? J : I); 2402 I2 = NewI2; 2403 I2T = I1T; 2404 I2Elem = I1Elem; 2405 } 2406 2407 // Now that both I1 and I2 are the same length we can shuffle them 2408 // together (and use the result). 2409 std::vector<Constant *> Mask(numElem); 2410 for (unsigned v = 0; v < numElem; ++v) { 2411 if (II[v].first == -1) { 2412 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2413 } else { 2414 int Idx = II[v].first + II[v].second * I1Elem; 2415 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2416 } 2417 } 2418 2419 Instruction *NewOp = 2420 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2421 getReplacementName(IBeforeJ ? I : J, true, o)); 2422 NewOp->insertBefore(IBeforeJ ? J : I); 2423 return NewOp; 2424 } 2425 } 2426 2427 Type *ArgType = ArgTypeL; 2428 if (numElemL < numElemH) { 2429 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2430 ArgTypeL, VArgType, IBeforeJ, 1)) { 2431 // This is another short-circuit case: we're combining a scalar into 2432 // a vector that is formed by an IE chain. We've just expanded the IE 2433 // chain, now insert the scalar and we're done. 2434 2435 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2436 getReplacementName(IBeforeJ ? I : J, true, o)); 2437 S->insertBefore(IBeforeJ ? J : I); 2438 return S; 2439 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2440 ArgTypeH, IBeforeJ)) { 2441 // The two vector inputs to the shuffle must be the same length, 2442 // so extend the smaller vector to be the same length as the larger one. 2443 Instruction *NLOp; 2444 if (numElemL > 1) { 2445 2446 std::vector<Constant *> Mask(numElemH); 2447 unsigned v = 0; 2448 for (; v < numElemL; ++v) 2449 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2450 for (; v < numElemH; ++v) 2451 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2452 2453 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2454 ConstantVector::get(Mask), 2455 getReplacementName(IBeforeJ ? I : J, 2456 true, o, 1)); 2457 } else { 2458 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2459 getReplacementName(IBeforeJ ? I : J, 2460 true, o, 1)); 2461 } 2462 2463 NLOp->insertBefore(IBeforeJ ? J : I); 2464 LOp = NLOp; 2465 } 2466 2467 ArgType = ArgTypeH; 2468 } else if (numElemL > numElemH) { 2469 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2470 ArgTypeH, VArgType, IBeforeJ)) { 2471 Instruction *S = 2472 InsertElementInst::Create(LOp, HOp, 2473 ConstantInt::get(Type::getInt32Ty(Context), 2474 numElemL), 2475 getReplacementName(IBeforeJ ? I : J, 2476 true, o)); 2477 S->insertBefore(IBeforeJ ? J : I); 2478 return S; 2479 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2480 ArgTypeL, IBeforeJ)) { 2481 Instruction *NHOp; 2482 if (numElemH > 1) { 2483 std::vector<Constant *> Mask(numElemL); 2484 unsigned v = 0; 2485 for (; v < numElemH; ++v) 2486 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2487 for (; v < numElemL; ++v) 2488 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2489 2490 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2491 ConstantVector::get(Mask), 2492 getReplacementName(IBeforeJ ? I : J, 2493 true, o, 1)); 2494 } else { 2495 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2496 getReplacementName(IBeforeJ ? I : J, 2497 true, o, 1)); 2498 } 2499 2500 NHOp->insertBefore(IBeforeJ ? J : I); 2501 HOp = NHOp; 2502 } 2503 } 2504 2505 if (ArgType->isVectorTy()) { 2506 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2507 std::vector<Constant*> Mask(numElem); 2508 for (unsigned v = 0; v < numElem; ++v) { 2509 unsigned Idx = v; 2510 // If the low vector was expanded, we need to skip the extra 2511 // undefined entries. 2512 if (v >= numElemL && numElemH > numElemL) 2513 Idx += (numElemH - numElemL); 2514 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2515 } 2516 2517 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2518 ConstantVector::get(Mask), 2519 getReplacementName(IBeforeJ ? I : J, true, o)); 2520 BV->insertBefore(IBeforeJ ? J : I); 2521 return BV; 2522 } 2523 2524 Instruction *BV1 = InsertElementInst::Create( 2525 UndefValue::get(VArgType), LOp, CV0, 2526 getReplacementName(IBeforeJ ? I : J, 2527 true, o, 1)); 2528 BV1->insertBefore(IBeforeJ ? J : I); 2529 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2530 getReplacementName(IBeforeJ ? I : J, 2531 true, o, 2)); 2532 BV2->insertBefore(IBeforeJ ? J : I); 2533 return BV2; 2534 } 2535 2536 // This function creates an array of values that will be used as the inputs 2537 // to the vector instruction that fuses I with J. 2538 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2539 Instruction *I, Instruction *J, 2540 SmallVector<Value *, 3> &ReplacedOperands, 2541 bool IBeforeJ) { 2542 unsigned NumOperands = I->getNumOperands(); 2543 2544 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2545 // Iterate backward so that we look at the store pointer 2546 // first and know whether or not we need to flip the inputs. 2547 2548 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2549 // This is the pointer for a load/store instruction. 2550 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2551 continue; 2552 } else if (isa<CallInst>(I)) { 2553 Function *F = cast<CallInst>(I)->getCalledFunction(); 2554 unsigned IID = F->getIntrinsicID(); 2555 if (o == NumOperands-1) { 2556 BasicBlock &BB = *I->getParent(); 2557 2558 Module *M = BB.getParent()->getParent(); 2559 Type *ArgTypeI = I->getType(); 2560 Type *ArgTypeJ = J->getType(); 2561 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2562 2563 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 2564 (Intrinsic::ID) IID, VArgType); 2565 continue; 2566 } else if (IID == Intrinsic::powi && o == 1) { 2567 // The second argument of powi is a single integer and we've already 2568 // checked that both arguments are equal. As a result, we just keep 2569 // I's second argument. 2570 ReplacedOperands[o] = I->getOperand(o); 2571 continue; 2572 } 2573 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2574 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2575 continue; 2576 } 2577 2578 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2579 } 2580 } 2581 2582 // This function creates two values that represent the outputs of the 2583 // original I and J instructions. These are generally vector shuffles 2584 // or extracts. In many cases, these will end up being unused and, thus, 2585 // eliminated by later passes. 2586 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2587 Instruction *J, Instruction *K, 2588 Instruction *&InsertionPt, 2589 Instruction *&K1, Instruction *&K2) { 2590 if (isa<StoreInst>(I)) { 2591 AA->replaceWithNewValue(I, K); 2592 AA->replaceWithNewValue(J, K); 2593 } else { 2594 Type *IType = I->getType(); 2595 Type *JType = J->getType(); 2596 2597 VectorType *VType = getVecTypeForPair(IType, JType); 2598 unsigned numElem = VType->getNumElements(); 2599 2600 unsigned numElemI, numElemJ; 2601 if (IType->isVectorTy()) 2602 numElemI = cast<VectorType>(IType)->getNumElements(); 2603 else 2604 numElemI = 1; 2605 2606 if (JType->isVectorTy()) 2607 numElemJ = cast<VectorType>(JType)->getNumElements(); 2608 else 2609 numElemJ = 1; 2610 2611 if (IType->isVectorTy()) { 2612 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2613 for (unsigned v = 0; v < numElemI; ++v) { 2614 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2615 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2616 } 2617 2618 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2619 ConstantVector::get( Mask1), 2620 getReplacementName(K, false, 1)); 2621 } else { 2622 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2623 K1 = ExtractElementInst::Create(K, CV0, 2624 getReplacementName(K, false, 1)); 2625 } 2626 2627 if (JType->isVectorTy()) { 2628 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2629 for (unsigned v = 0; v < numElemJ; ++v) { 2630 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2631 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2632 } 2633 2634 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2635 ConstantVector::get( Mask2), 2636 getReplacementName(K, false, 2)); 2637 } else { 2638 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2639 K2 = ExtractElementInst::Create(K, CV1, 2640 getReplacementName(K, false, 2)); 2641 } 2642 2643 K1->insertAfter(K); 2644 K2->insertAfter(K1); 2645 InsertionPt = K2; 2646 } 2647 } 2648 2649 // Move all uses of the function I (including pairing-induced uses) after J. 2650 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2651 std::multimap<Value *, Value *> &LoadMoveSet, 2652 Instruction *I, Instruction *J) { 2653 // Skip to the first instruction past I. 2654 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2655 2656 DenseSet<Value *> Users; 2657 AliasSetTracker WriteSet(*AA); 2658 for (; cast<Instruction>(L) != J; ++L) 2659 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 2660 2661 assert(cast<Instruction>(L) == J && 2662 "Tracking has not proceeded far enough to check for dependencies"); 2663 // If J is now in the use set of I, then trackUsesOfI will return true 2664 // and we have a dependency cycle (and the fusing operation must abort). 2665 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 2666 } 2667 2668 // Move all uses of the function I (including pairing-induced uses) after J. 2669 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2670 std::multimap<Value *, Value *> &LoadMoveSet, 2671 Instruction *&InsertionPt, 2672 Instruction *I, Instruction *J) { 2673 // Skip to the first instruction past I. 2674 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2675 2676 DenseSet<Value *> Users; 2677 AliasSetTracker WriteSet(*AA); 2678 for (; cast<Instruction>(L) != J;) { 2679 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 2680 // Move this instruction 2681 Instruction *InstToMove = L; ++L; 2682 2683 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2684 " to after " << *InsertionPt << "\n"); 2685 InstToMove->removeFromParent(); 2686 InstToMove->insertAfter(InsertionPt); 2687 InsertionPt = InstToMove; 2688 } else { 2689 ++L; 2690 } 2691 } 2692 } 2693 2694 // Collect all load instruction that are in the move set of a given first 2695 // pair member. These loads depend on the first instruction, I, and so need 2696 // to be moved after J (the second instruction) when the pair is fused. 2697 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2698 DenseMap<Value *, Value *> &ChosenPairs, 2699 std::multimap<Value *, Value *> &LoadMoveSet, 2700 Instruction *I) { 2701 // Skip to the first instruction past I. 2702 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2703 2704 DenseSet<Value *> Users; 2705 AliasSetTracker WriteSet(*AA); 2706 2707 // Note: We cannot end the loop when we reach J because J could be moved 2708 // farther down the use chain by another instruction pairing. Also, J 2709 // could be before I if this is an inverted input. 2710 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2711 if (trackUsesOfI(Users, WriteSet, I, L)) { 2712 if (L->mayReadFromMemory()) 2713 LoadMoveSet.insert(ValuePair(L, I)); 2714 } 2715 } 2716 } 2717 2718 // In cases where both load/stores and the computation of their pointers 2719 // are chosen for vectorization, we can end up in a situation where the 2720 // aliasing analysis starts returning different query results as the 2721 // process of fusing instruction pairs continues. Because the algorithm 2722 // relies on finding the same use trees here as were found earlier, we'll 2723 // need to precompute the necessary aliasing information here and then 2724 // manually update it during the fusion process. 2725 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2726 std::vector<Value *> &PairableInsts, 2727 DenseMap<Value *, Value *> &ChosenPairs, 2728 std::multimap<Value *, Value *> &LoadMoveSet) { 2729 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2730 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2731 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2732 if (P == ChosenPairs.end()) continue; 2733 2734 Instruction *I = cast<Instruction>(P->first); 2735 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 2736 } 2737 } 2738 2739 // When the first instruction in each pair is cloned, it will inherit its 2740 // parent's metadata. This metadata must be combined with that of the other 2741 // instruction in a safe way. 2742 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2743 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2744 K->getAllMetadataOtherThanDebugLoc(Metadata); 2745 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2746 unsigned Kind = Metadata[i].first; 2747 MDNode *JMD = J->getMetadata(Kind); 2748 MDNode *KMD = Metadata[i].second; 2749 2750 switch (Kind) { 2751 default: 2752 K->setMetadata(Kind, 0); // Remove unknown metadata 2753 break; 2754 case LLVMContext::MD_tbaa: 2755 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2756 break; 2757 case LLVMContext::MD_fpmath: 2758 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2759 break; 2760 } 2761 } 2762 } 2763 2764 // This function fuses the chosen instruction pairs into vector instructions, 2765 // taking care preserve any needed scalar outputs and, then, it reorders the 2766 // remaining instructions as needed (users of the first member of the pair 2767 // need to be moved to after the location of the second member of the pair 2768 // because the vector instruction is inserted in the location of the pair's 2769 // second member). 2770 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2771 std::vector<Value *> &PairableInsts, 2772 DenseMap<Value *, Value *> &ChosenPairs, 2773 DenseSet<ValuePair> &FixedOrderPairs, 2774 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2775 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2776 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2777 LLVMContext& Context = BB.getContext(); 2778 2779 // During the vectorization process, the order of the pairs to be fused 2780 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2781 // list. After a pair is fused, the flipped pair is removed from the list. 2782 DenseSet<ValuePair> FlippedPairs; 2783 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2784 E = ChosenPairs.end(); P != E; ++P) 2785 FlippedPairs.insert(ValuePair(P->second, P->first)); 2786 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2787 E = FlippedPairs.end(); P != E; ++P) 2788 ChosenPairs.insert(*P); 2789 2790 std::multimap<Value *, Value *> LoadMoveSet; 2791 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 2792 2793 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2794 2795 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2796 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2797 if (P == ChosenPairs.end()) { 2798 ++PI; 2799 continue; 2800 } 2801 2802 if (getDepthFactor(P->first) == 0) { 2803 // These instructions are not really fused, but are tracked as though 2804 // they are. Any case in which it would be interesting to fuse them 2805 // will be taken care of by InstCombine. 2806 --NumFusedOps; 2807 ++PI; 2808 continue; 2809 } 2810 2811 Instruction *I = cast<Instruction>(P->first), 2812 *J = cast<Instruction>(P->second); 2813 2814 DEBUG(dbgs() << "BBV: fusing: " << *I << 2815 " <-> " << *J << "\n"); 2816 2817 // Remove the pair and flipped pair from the list. 2818 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2819 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2820 ChosenPairs.erase(FP); 2821 ChosenPairs.erase(P); 2822 2823 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 2824 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2825 " <-> " << *J << 2826 " aborted because of non-trivial dependency cycle\n"); 2827 --NumFusedOps; 2828 ++PI; 2829 continue; 2830 } 2831 2832 // If the pair must have the other order, then flip it. 2833 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2834 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2835 // This pair does not have a fixed order, and so we might want to 2836 // flip it if that will yield fewer shuffles. We count the number 2837 // of dependencies connected via swaps, and those directly connected, 2838 // and flip the order if the number of swaps is greater. 2839 bool OrigOrder = true; 2840 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2841 if (IP.first == ConnectedPairDeps.end()) { 2842 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2843 OrigOrder = false; 2844 } 2845 2846 if (IP.first != ConnectedPairDeps.end()) { 2847 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2848 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2849 Q != IP.second; ++Q) { 2850 DenseMap<VPPair, unsigned>::iterator R = 2851 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2852 assert(R != PairConnectionTypes.end() && 2853 "Cannot find pair connection type"); 2854 if (R->second == PairConnectionDirect) 2855 ++NumDepsDirect; 2856 else if (R->second == PairConnectionSwap) 2857 ++NumDepsSwap; 2858 } 2859 2860 if (!OrigOrder) 2861 std::swap(NumDepsDirect, NumDepsSwap); 2862 2863 if (NumDepsSwap > NumDepsDirect) { 2864 FlipPairOrder = true; 2865 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2866 " <-> " << *J << "\n"); 2867 } 2868 } 2869 } 2870 2871 Instruction *L = I, *H = J; 2872 if (FlipPairOrder) 2873 std::swap(H, L); 2874 2875 // If the pair being fused uses the opposite order from that in the pair 2876 // connection map, then we need to flip the types. 2877 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2878 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2879 Q != IP.second; ++Q) { 2880 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2881 assert(R != PairConnectionTypes.end() && 2882 "Cannot find pair connection type"); 2883 if (R->second == PairConnectionDirect) 2884 R->second = PairConnectionSwap; 2885 else if (R->second == PairConnectionSwap) 2886 R->second = PairConnectionDirect; 2887 } 2888 2889 bool LBeforeH = !FlipPairOrder; 2890 unsigned NumOperands = I->getNumOperands(); 2891 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2892 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2893 LBeforeH); 2894 2895 // Make a copy of the original operation, change its type to the vector 2896 // type and replace its operands with the vector operands. 2897 Instruction *K = L->clone(); 2898 if (L->hasName()) 2899 K->takeName(L); 2900 else if (H->hasName()) 2901 K->takeName(H); 2902 2903 if (!isa<StoreInst>(K)) 2904 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 2905 2906 combineMetadata(K, H); 2907 K->intersectOptionalDataWith(H); 2908 2909 for (unsigned o = 0; o < NumOperands; ++o) 2910 K->setOperand(o, ReplacedOperands[o]); 2911 2912 K->insertAfter(J); 2913 2914 // Instruction insertion point: 2915 Instruction *InsertionPt = K; 2916 Instruction *K1 = 0, *K2 = 0; 2917 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 2918 2919 // The use tree of the first original instruction must be moved to after 2920 // the location of the second instruction. The entire use tree of the 2921 // first instruction is disjoint from the input tree of the second 2922 // (by definition), and so commutes with it. 2923 2924 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 2925 2926 if (!isa<StoreInst>(I)) { 2927 L->replaceAllUsesWith(K1); 2928 H->replaceAllUsesWith(K2); 2929 AA->replaceWithNewValue(L, K1); 2930 AA->replaceWithNewValue(H, K2); 2931 } 2932 2933 // Instructions that may read from memory may be in the load move set. 2934 // Once an instruction is fused, we no longer need its move set, and so 2935 // the values of the map never need to be updated. However, when a load 2936 // is fused, we need to merge the entries from both instructions in the 2937 // pair in case those instructions were in the move set of some other 2938 // yet-to-be-fused pair. The loads in question are the keys of the map. 2939 if (I->mayReadFromMemory()) { 2940 std::vector<ValuePair> NewSetMembers; 2941 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 2942 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 2943 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 2944 N != IPairRange.second; ++N) 2945 NewSetMembers.push_back(ValuePair(K, N->second)); 2946 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 2947 N != JPairRange.second; ++N) 2948 NewSetMembers.push_back(ValuePair(K, N->second)); 2949 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 2950 AE = NewSetMembers.end(); A != AE; ++A) 2951 LoadMoveSet.insert(*A); 2952 } 2953 2954 // Before removing I, set the iterator to the next instruction. 2955 PI = llvm::next(BasicBlock::iterator(I)); 2956 if (cast<Instruction>(PI) == J) 2957 ++PI; 2958 2959 SE->forgetValue(I); 2960 SE->forgetValue(J); 2961 I->eraseFromParent(); 2962 J->eraseFromParent(); 2963 2964 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 2965 BB << "\n"); 2966 } 2967 2968 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 2969 } 2970} 2971 2972char BBVectorize::ID = 0; 2973static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 2974INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2975INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2976INITIALIZE_PASS_DEPENDENCY(DominatorTree) 2977INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2978INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2979 2980BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 2981 return new BBVectorize(C); 2982} 2983 2984bool 2985llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 2986 BBVectorize BBVectorizer(P, C); 2987 return BBVectorizer.vectorizeBB(BB); 2988} 2989 2990//===----------------------------------------------------------------------===// 2991VectorizeConfig::VectorizeConfig() { 2992 VectorBits = ::VectorBits; 2993 VectorizeBools = !::NoBools; 2994 VectorizeInts = !::NoInts; 2995 VectorizeFloats = !::NoFloats; 2996 VectorizePointers = !::NoPointers; 2997 VectorizeCasts = !::NoCasts; 2998 VectorizeMath = !::NoMath; 2999 VectorizeFMA = !::NoFMA; 3000 VectorizeSelect = !::NoSelect; 3001 VectorizeCmp = !::NoCmp; 3002 VectorizeGEP = !::NoGEP; 3003 VectorizeMemOps = !::NoMemOps; 3004 AlignedOnly = ::AlignedOnly; 3005 ReqChainDepth= ::ReqChainDepth; 3006 SearchLimit = ::SearchLimit; 3007 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 3008 SplatBreaksChain = ::SplatBreaksChain; 3009 MaxInsts = ::MaxInsts; 3010 MaxIter = ::MaxIter; 3011 Pow2LenOnly = ::Pow2LenOnly; 3012 NoMemOpBoost = ::NoMemOpBoost; 3013 FastDep = ::FastDep; 3014} 3015