BBVectorize.cpp revision 86c88c938aec8006d2ce83325ec1f31e1154620b
1//===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements a basic-block vectorization pass. The algorithm was 11// inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral, 12// et al. It works by looking for chains of pairable operations and then 13// pairing them. 14// 15//===----------------------------------------------------------------------===// 16 17#define BBV_NAME "bb-vectorize" 18#define DEBUG_TYPE BBV_NAME 19#include "llvm/Constants.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Function.h" 22#include "llvm/Instructions.h" 23#include "llvm/IntrinsicInst.h" 24#include "llvm/Intrinsics.h" 25#include "llvm/LLVMContext.h" 26#include "llvm/Metadata.h" 27#include "llvm/Pass.h" 28#include "llvm/Type.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/DenseSet.h" 31#include "llvm/ADT/SmallSet.h" 32#include "llvm/ADT/SmallVector.h" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/ADT/STLExtras.h" 35#include "llvm/ADT/StringExtras.h" 36#include "llvm/Analysis/AliasAnalysis.h" 37#include "llvm/Analysis/AliasSetTracker.h" 38#include "llvm/Analysis/Dominators.h" 39#include "llvm/Analysis/ScalarEvolution.h" 40#include "llvm/Analysis/ScalarEvolutionExpressions.h" 41#include "llvm/Analysis/ValueTracking.h" 42#include "llvm/Support/CommandLine.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/raw_ostream.h" 45#include "llvm/Support/ValueHandle.h" 46#include "llvm/DataLayout.h" 47#include "llvm/TargetTransformInfo.h" 48#include "llvm/Transforms/Utils/Local.h" 49#include "llvm/Transforms/Vectorize.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"); 404assert(n < 10 && "hrmm, really?"); 405 if (vectorizePairs(BB)) 406 changed = true; 407 else 408 break; 409 } 410 411 if (changed && !Pow2LenOnly) { 412 ++n; 413 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 414 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 415 n << " for " << BB.getName() << " in " << 416 BB.getParent()->getName() << "...\n"); 417 if (!vectorizePairs(BB, true)) break; 418 } 419 } 420 421 DEBUG(dbgs() << "BBV: done!\n"); 422 return changed; 423 } 424 425 virtual bool runOnBasicBlock(BasicBlock &BB) { 426 AA = &getAnalysis<AliasAnalysis>(); 427 DT = &getAnalysis<DominatorTree>(); 428 SE = &getAnalysis<ScalarEvolution>(); 429 TD = getAnalysisIfAvailable<DataLayout>(); 430 TTI = IgnoreTargetInfo ? 0 : 431 getAnalysisIfAvailable<TargetTransformInfo>(); 432 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; 433 434 return vectorizeBB(BB); 435 } 436 437 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 438 BasicBlockPass::getAnalysisUsage(AU); 439 AU.addRequired<AliasAnalysis>(); 440 AU.addRequired<DominatorTree>(); 441 AU.addRequired<ScalarEvolution>(); 442 AU.addPreserved<AliasAnalysis>(); 443 AU.addPreserved<DominatorTree>(); 444 AU.addPreserved<ScalarEvolution>(); 445 AU.setPreservesCFG(); 446 } 447 448 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 449 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 450 "Cannot form vector from incompatible scalar types"); 451 Type *STy = ElemTy->getScalarType(); 452 453 unsigned numElem; 454 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 455 numElem = VTy->getNumElements(); 456 } else { 457 numElem = 1; 458 } 459 460 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 461 numElem += VTy->getNumElements(); 462 } else { 463 numElem += 1; 464 } 465 466 return VectorType::get(STy, numElem); 467 } 468 469 static inline void getInstructionTypes(Instruction *I, 470 Type *&T1, Type *&T2) { 471 if (isa<StoreInst>(I)) { 472 // For stores, it is the value type, not the pointer type that matters 473 // because the value is what will come from a vector register. 474 475 Value *IVal = cast<StoreInst>(I)->getValueOperand(); 476 T1 = IVal->getType(); 477 } else { 478 T1 = I->getType(); 479 } 480 481 if (I->isCast()) 482 T2 = cast<CastInst>(I)->getSrcTy(); 483 else 484 T2 = T1; 485 486 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 487 T2 = SI->getCondition()->getType(); 488 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { 489 T2 = SI->getOperand(0)->getType(); 490 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 491 T2 = CI->getOperand(0)->getType(); 492 } 493 } 494 495 // Returns the weight associated with the provided value. A chain of 496 // candidate pairs has a length given by the sum of the weights of its 497 // members (one weight per pair; the weight of each member of the pair 498 // is assumed to be the same). This length is then compared to the 499 // chain-length threshold to determine if a given chain is significant 500 // enough to be vectorized. The length is also used in comparing 501 // candidate chains where longer chains are considered to be better. 502 // Note: when this function returns 0, the resulting instructions are 503 // not actually fused. 504 inline size_t getDepthFactor(Value *V) { 505 // InsertElement and ExtractElement have a depth factor of zero. This is 506 // for two reasons: First, they cannot be usefully fused. Second, because 507 // the pass generates a lot of these, they can confuse the simple metric 508 // used to compare the trees in the next iteration. Thus, giving them a 509 // weight of zero allows the pass to essentially ignore them in 510 // subsequent iterations when looking for vectorization opportunities 511 // while still tracking dependency chains that flow through those 512 // instructions. 513 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 514 return 0; 515 516 // Give a load or store half of the required depth so that load/store 517 // pairs will vectorize. 518 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 519 return Config.ReqChainDepth/2; 520 521 return 1; 522 } 523 524 // Returns the cost of the provided instruction using VTTI. 525 // This does not handle loads and stores. 526 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { 527 switch (Opcode) { 528 default: break; 529 case Instruction::GetElementPtr: 530 // We mark this instruction as zero-cost because scalar GEPs are usually 531 // lowered to the intruction addressing mode. At the moment we don't 532 // generate vector GEPs. 533 return 0; 534 case Instruction::Br: 535 return VTTI->getCFInstrCost(Opcode); 536 case Instruction::PHI: 537 return 0; 538 case Instruction::Add: 539 case Instruction::FAdd: 540 case Instruction::Sub: 541 case Instruction::FSub: 542 case Instruction::Mul: 543 case Instruction::FMul: 544 case Instruction::UDiv: 545 case Instruction::SDiv: 546 case Instruction::FDiv: 547 case Instruction::URem: 548 case Instruction::SRem: 549 case Instruction::FRem: 550 case Instruction::Shl: 551 case Instruction::LShr: 552 case Instruction::AShr: 553 case Instruction::And: 554 case Instruction::Or: 555 case Instruction::Xor: 556 return VTTI->getArithmeticInstrCost(Opcode, T1); 557 case Instruction::Select: 558 case Instruction::ICmp: 559 case Instruction::FCmp: 560 return VTTI->getCmpSelInstrCost(Opcode, T1, T2); 561 case Instruction::ZExt: 562 case Instruction::SExt: 563 case Instruction::FPToUI: 564 case Instruction::FPToSI: 565 case Instruction::FPExt: 566 case Instruction::PtrToInt: 567 case Instruction::IntToPtr: 568 case Instruction::SIToFP: 569 case Instruction::UIToFP: 570 case Instruction::Trunc: 571 case Instruction::FPTrunc: 572 case Instruction::BitCast: 573 case Instruction::ShuffleVector: 574 return VTTI->getCastInstrCost(Opcode, T1, T2); 575 } 576 577 return 1; 578 } 579 580 // This determines the relative offset of two loads or stores, returning 581 // true if the offset could be determined to be some constant value. 582 // For example, if OffsetInElmts == 1, then J accesses the memory directly 583 // after I; if OffsetInElmts == -1 then I accesses the memory 584 // directly after J. 585 bool getPairPtrInfo(Instruction *I, Instruction *J, 586 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 587 unsigned &IAddressSpace, unsigned &JAddressSpace, 588 int64_t &OffsetInElmts, bool ComputeOffset = true) { 589 OffsetInElmts = 0; 590 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 591 LoadInst *LJ = cast<LoadInst>(J); 592 IPtr = LI->getPointerOperand(); 593 JPtr = LJ->getPointerOperand(); 594 IAlignment = LI->getAlignment(); 595 JAlignment = LJ->getAlignment(); 596 IAddressSpace = LI->getPointerAddressSpace(); 597 JAddressSpace = LJ->getPointerAddressSpace(); 598 } else { 599 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 600 IPtr = SI->getPointerOperand(); 601 JPtr = SJ->getPointerOperand(); 602 IAlignment = SI->getAlignment(); 603 JAlignment = SJ->getAlignment(); 604 IAddressSpace = SI->getPointerAddressSpace(); 605 JAddressSpace = SJ->getPointerAddressSpace(); 606 } 607 608 if (!ComputeOffset) 609 return true; 610 611 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 612 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 613 614 // If this is a trivial offset, then we'll get something like 615 // 1*sizeof(type). With target data, which we need anyway, this will get 616 // constant folded into a number. 617 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 618 if (const SCEVConstant *ConstOffSCEV = 619 dyn_cast<SCEVConstant>(OffsetSCEV)) { 620 ConstantInt *IntOff = ConstOffSCEV->getValue(); 621 int64_t Offset = IntOff->getSExtValue(); 622 623 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 624 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 625 626 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 627 if (VTy != VTy2 && Offset < 0) { 628 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 629 OffsetInElmts = Offset/VTy2TSS; 630 return (abs64(Offset) % VTy2TSS) == 0; 631 } 632 633 OffsetInElmts = Offset/VTyTSS; 634 return (abs64(Offset) % VTyTSS) == 0; 635 } 636 637 return false; 638 } 639 640 // Returns true if the provided CallInst represents an intrinsic that can 641 // be vectorized. 642 bool isVectorizableIntrinsic(CallInst* I) { 643 Function *F = I->getCalledFunction(); 644 if (!F) return false; 645 646 unsigned IID = F->getIntrinsicID(); 647 if (!IID) return false; 648 649 switch(IID) { 650 default: 651 return false; 652 case Intrinsic::sqrt: 653 case Intrinsic::powi: 654 case Intrinsic::sin: 655 case Intrinsic::cos: 656 case Intrinsic::log: 657 case Intrinsic::log2: 658 case Intrinsic::log10: 659 case Intrinsic::exp: 660 case Intrinsic::exp2: 661 case Intrinsic::pow: 662 return Config.VectorizeMath; 663 case Intrinsic::fma: 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 }; 681 682 // This function implements one vectorization iteration on the provided 683 // basic block. It returns true if the block is changed. 684 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 685 bool ShouldContinue; 686 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 687 688 std::vector<Value *> AllPairableInsts; 689 DenseMap<Value *, Value *> AllChosenPairs; 690 DenseSet<ValuePair> AllFixedOrderPairs; 691 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 692 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; 693 694 do { 695 std::vector<Value *> PairableInsts; 696 std::multimap<Value *, Value *> CandidatePairs; 697 DenseSet<ValuePair> FixedOrderPairs; 698 DenseMap<ValuePair, int> CandidatePairCostSavings; 699 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 700 FixedOrderPairs, 701 CandidatePairCostSavings, 702 PairableInsts, NonPow2Len); 703 if (PairableInsts.empty()) continue; 704 705 // Now we have a map of all of the pairable instructions and we need to 706 // select the best possible pairing. A good pairing is one such that the 707 // users of the pair are also paired. This defines a (directed) forest 708 // over the pairs such that two pairs are connected iff the second pair 709 // uses the first. 710 711 // Note that it only matters that both members of the second pair use some 712 // element of the first pair (to allow for splatting). 713 714 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; 715 DenseMap<VPPair, unsigned> PairConnectionTypes; 716 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs, 717 PairConnectionTypes); 718 if (ConnectedPairs.empty()) continue; 719 720 for (std::multimap<ValuePair, ValuePair>::iterator 721 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 722 I != IE; ++I) { 723 ConnectedPairDeps.insert(VPPair(I->second, I->first)); 724 } 725 726 // Build the pairable-instruction dependency map 727 DenseSet<ValuePair> PairableInstUsers; 728 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 729 730 // There is now a graph of the connected pairs. For each variable, pick 731 // the pairing with the largest tree meeting the depth requirement on at 732 // least one branch. Then select all pairings that are part of that tree 733 // and remove them from the list of available pairings and pairable 734 // variables. 735 736 DenseMap<Value *, Value *> ChosenPairs; 737 choosePairs(CandidatePairs, CandidatePairCostSavings, 738 PairableInsts, FixedOrderPairs, PairConnectionTypes, 739 ConnectedPairs, ConnectedPairDeps, 740 PairableInstUsers, ChosenPairs); 741 742 if (ChosenPairs.empty()) continue; 743 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 744 PairableInsts.end()); 745 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 746 747 // Only for the chosen pairs, propagate information on fixed-order pairs, 748 // pair connections, and their types to the data structures used by the 749 // pair fusion procedures. 750 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 751 IE = ChosenPairs.end(); I != IE; ++I) { 752 if (FixedOrderPairs.count(*I)) 753 AllFixedOrderPairs.insert(*I); 754 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 755 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 756 757 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 758 J != IE; ++J) { 759 DenseMap<VPPair, unsigned>::iterator K = 760 PairConnectionTypes.find(VPPair(*I, *J)); 761 if (K != PairConnectionTypes.end()) { 762 AllPairConnectionTypes.insert(*K); 763 } else { 764 K = PairConnectionTypes.find(VPPair(*J, *I)); 765 if (K != PairConnectionTypes.end()) 766 AllPairConnectionTypes.insert(*K); 767 } 768 } 769 } 770 771 for (std::multimap<ValuePair, ValuePair>::iterator 772 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 773 I != IE; ++I) { 774 if (AllPairConnectionTypes.count(*I)) { 775 AllConnectedPairs.insert(*I); 776 AllConnectedPairDeps.insert(VPPair(I->second, I->first)); 777 } 778 } 779 } while (ShouldContinue); 780 781 if (AllChosenPairs.empty()) return false; 782 NumFusedOps += AllChosenPairs.size(); 783 784 // A set of pairs has now been selected. It is now necessary to replace the 785 // paired instructions with vector instructions. For this procedure each 786 // operand must be replaced with a vector operand. This vector is formed 787 // by using build_vector on the old operands. The replaced values are then 788 // replaced with a vector_extract on the result. Subsequent optimization 789 // passes should coalesce the build/extract combinations. 790 791 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 792 AllPairConnectionTypes, 793 AllConnectedPairs, AllConnectedPairDeps); 794 795 // It is important to cleanup here so that future iterations of this 796 // function have less work to do. 797 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 798 return true; 799 } 800 801 // This function returns true if the provided instruction is capable of being 802 // fused into a vector instruction. This determination is based only on the 803 // type and other attributes of the instruction. 804 bool BBVectorize::isInstVectorizable(Instruction *I, 805 bool &IsSimpleLoadStore) { 806 IsSimpleLoadStore = false; 807 808 if (CallInst *C = dyn_cast<CallInst>(I)) { 809 if (!isVectorizableIntrinsic(C)) 810 return false; 811 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 812 // Vectorize simple loads if possbile: 813 IsSimpleLoadStore = L->isSimple(); 814 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 815 return false; 816 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 817 // Vectorize simple stores if possbile: 818 IsSimpleLoadStore = S->isSimple(); 819 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 820 return false; 821 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 822 // We can vectorize casts, but not casts of pointer types, etc. 823 if (!Config.VectorizeCasts) 824 return false; 825 826 Type *SrcTy = C->getSrcTy(); 827 if (!SrcTy->isSingleValueType()) 828 return false; 829 830 Type *DestTy = C->getDestTy(); 831 if (!DestTy->isSingleValueType()) 832 return false; 833 } else if (isa<SelectInst>(I)) { 834 if (!Config.VectorizeSelect) 835 return false; 836 } else if (isa<CmpInst>(I)) { 837 if (!Config.VectorizeCmp) 838 return false; 839 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 840 if (!Config.VectorizeGEP) 841 return false; 842 843 // Currently, vector GEPs exist only with one index. 844 if (G->getNumIndices() != 1) 845 return false; 846 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 847 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 848 return false; 849 } 850 851 // We can't vectorize memory operations without target data 852 if (TD == 0 && IsSimpleLoadStore) 853 return false; 854 855 Type *T1, *T2; 856 getInstructionTypes(I, T1, T2); 857 858 // Not every type can be vectorized... 859 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 860 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 861 return false; 862 863 if (T1->getScalarSizeInBits() == 1) { 864 if (!Config.VectorizeBools) 865 return false; 866 } else { 867 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 868 return false; 869 } 870 871 if (T2->getScalarSizeInBits() == 1) { 872 if (!Config.VectorizeBools) 873 return false; 874 } else { 875 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 876 return false; 877 } 878 879 if (!Config.VectorizeFloats 880 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 881 return false; 882 883 // Don't vectorize target-specific types. 884 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 885 return false; 886 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 887 return false; 888 889 if ((!Config.VectorizePointers || TD == 0) && 890 (T1->getScalarType()->isPointerTy() || 891 T2->getScalarType()->isPointerTy())) 892 return false; 893 894 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 895 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 896 return false; 897 898 return true; 899 } 900 901 // This function returns true if the two provided instructions are compatible 902 // (meaning that they can be fused into a vector instruction). This assumes 903 // that I has already been determined to be vectorizable and that J is not 904 // in the use tree of I. 905 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 906 bool IsSimpleLoadStore, bool NonPow2Len, 907 int &CostSavings, int &FixedOrder) { 908 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 909 " <-> " << *J << "\n"); 910 911 CostSavings = 0; 912 FixedOrder = 0; 913 914 // Loads and stores can be merged if they have different alignments, 915 // but are otherwise the same. 916 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 917 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 918 return false; 919 920 Type *IT1, *IT2, *JT1, *JT2; 921 getInstructionTypes(I, IT1, IT2); 922 getInstructionTypes(J, JT1, JT2); 923 unsigned MaxTypeBits = std::max( 924 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 925 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 926 if (!VTTI && MaxTypeBits > Config.VectorBits) 927 return false; 928 929 // FIXME: handle addsub-type operations! 930 931 if (IsSimpleLoadStore) { 932 Value *IPtr, *JPtr; 933 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 934 int64_t OffsetInElmts = 0; 935 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 936 IAddressSpace, JAddressSpace, 937 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 938 FixedOrder = (int) OffsetInElmts; 939 unsigned BottomAlignment = IAlignment; 940 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 941 942 Type *aTypeI = isa<StoreInst>(I) ? 943 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 944 Type *aTypeJ = isa<StoreInst>(J) ? 945 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 946 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 947 948 if (Config.AlignedOnly) { 949 // An aligned load or store is possible only if the instruction 950 // with the lower offset has an alignment suitable for the 951 // vector type. 952 953 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 954 if (BottomAlignment < VecAlignment) 955 return false; 956 } 957 958 if (VTTI) { 959 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(), 960 IAlignment, IAddressSpace); 961 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(), 962 JAlignment, JAddressSpace); 963 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType, 964 BottomAlignment, 965 IAddressSpace); 966 if (VCost > ICost + JCost) 967 return false; 968 969 // We don't want to fuse to a type that will be split, even 970 // if the two input types will also be split and there is no other 971 // associated cost. 972 unsigned VParts = VTTI->getNumberOfParts(VType); 973 if (VParts > 1) 974 return false; 975 else if (!VParts && VCost == ICost + JCost) 976 return false; 977 978 CostSavings = ICost + JCost - VCost; 979 } 980 } else { 981 return false; 982 } 983 } else if (VTTI) { 984 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 985 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 986 Type *VT1 = getVecTypeForPair(IT1, JT1), 987 *VT2 = getVecTypeForPair(IT2, JT2); 988 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 989 990 if (VCost > ICost + JCost) 991 return false; 992 993 // We don't want to fuse to a type that will be split, even 994 // if the two input types will also be split and there is no other 995 // associated cost. 996 unsigned VParts1 = VTTI->getNumberOfParts(VT1), 997 VParts2 = VTTI->getNumberOfParts(VT2); 998 if (VParts1 > 1 || VParts2 > 1) 999 return false; 1000 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) 1001 return false; 1002 1003 CostSavings = ICost + JCost - VCost; 1004 } 1005 1006 // The powi intrinsic is special because only the first argument is 1007 // vectorized, the second arguments must be equal. 1008 CallInst *CI = dyn_cast<CallInst>(I); 1009 Function *FI; 1010 if (CI && (FI = CI->getCalledFunction()) && 1011 FI->getIntrinsicID() == Intrinsic::powi) { 1012 1013 Value *A1I = CI->getArgOperand(1), 1014 *A1J = cast<CallInst>(J)->getArgOperand(1); 1015 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1016 *A1JSCEV = SE->getSCEV(A1J); 1017 return (A1ISCEV == A1JSCEV); 1018 } 1019 1020 return true; 1021 } 1022 1023 // Figure out whether or not J uses I and update the users and write-set 1024 // structures associated with I. Specifically, Users represents the set of 1025 // instructions that depend on I. WriteSet represents the set 1026 // of memory locations that are dependent on I. If UpdateUsers is true, 1027 // and J uses I, then Users is updated to contain J and WriteSet is updated 1028 // to contain any memory locations to which J writes. The function returns 1029 // true if J uses I. By default, alias analysis is used to determine 1030 // whether J reads from memory that overlaps with a location in WriteSet. 1031 // If LoadMoveSet is not null, then it is a previously-computed multimap 1032 // where the key is the memory-based user instruction and the value is 1033 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1034 // then the alias analysis is not used. This is necessary because this 1035 // function is called during the process of moving instructions during 1036 // vectorization and the results of the alias analysis are not stable during 1037 // that process. 1038 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1039 AliasSetTracker &WriteSet, Instruction *I, 1040 Instruction *J, bool UpdateUsers, 1041 std::multimap<Value *, Value *> *LoadMoveSet) { 1042 bool UsesI = false; 1043 1044 // This instruction may already be marked as a user due, for example, to 1045 // being a member of a selected pair. 1046 if (Users.count(J)) 1047 UsesI = true; 1048 1049 if (!UsesI) 1050 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1051 JU != JE; ++JU) { 1052 Value *V = *JU; 1053 if (I == V || Users.count(V)) { 1054 UsesI = true; 1055 break; 1056 } 1057 } 1058 if (!UsesI && J->mayReadFromMemory()) { 1059 if (LoadMoveSet) { 1060 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J); 1061 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange); 1062 } else { 1063 for (AliasSetTracker::iterator W = WriteSet.begin(), 1064 WE = WriteSet.end(); W != WE; ++W) { 1065 if (W->aliasesUnknownInst(J, *AA)) { 1066 UsesI = true; 1067 break; 1068 } 1069 } 1070 } 1071 } 1072 1073 if (UsesI && UpdateUsers) { 1074 if (J->mayWriteToMemory()) WriteSet.add(J); 1075 Users.insert(J); 1076 } 1077 1078 return UsesI; 1079 } 1080 1081 // This function iterates over all instruction pairs in the provided 1082 // basic block and collects all candidate pairs for vectorization. 1083 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1084 BasicBlock::iterator &Start, 1085 std::multimap<Value *, Value *> &CandidatePairs, 1086 DenseSet<ValuePair> &FixedOrderPairs, 1087 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1088 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1089 BasicBlock::iterator E = BB.end(); 1090 if (Start == E) return false; 1091 1092 bool ShouldContinue = false, IAfterStart = false; 1093 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1094 if (I == Start) IAfterStart = true; 1095 1096 bool IsSimpleLoadStore; 1097 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1098 1099 // Look for an instruction with which to pair instruction *I... 1100 DenseSet<Value *> Users; 1101 AliasSetTracker WriteSet(*AA); 1102 bool JAfterStart = IAfterStart; 1103 BasicBlock::iterator J = llvm::next(I); 1104 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1105 if (J == Start) JAfterStart = true; 1106 1107 // Determine if J uses I, if so, exit the loop. 1108 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1109 if (Config.FastDep) { 1110 // Note: For this heuristic to be effective, independent operations 1111 // must tend to be intermixed. This is likely to be true from some 1112 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1113 // but otherwise may require some kind of reordering pass. 1114 1115 // When using fast dependency analysis, 1116 // stop searching after first use: 1117 if (UsesI) break; 1118 } else { 1119 if (UsesI) continue; 1120 } 1121 1122 // J does not use I, and comes before the first use of I, so it can be 1123 // merged with I if the instructions are compatible. 1124 int CostSavings, FixedOrder; 1125 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1126 CostSavings, FixedOrder)) continue; 1127 1128 // J is a candidate for merging with I. 1129 if (!PairableInsts.size() || 1130 PairableInsts[PairableInsts.size()-1] != I) { 1131 PairableInsts.push_back(I); 1132 } 1133 1134 CandidatePairs.insert(ValuePair(I, J)); 1135 if (VTTI) 1136 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1137 CostSavings)); 1138 1139 if (FixedOrder == 1) 1140 FixedOrderPairs.insert(ValuePair(I, J)); 1141 else if (FixedOrder == -1) 1142 FixedOrderPairs.insert(ValuePair(J, I)); 1143 1144 // The next call to this function must start after the last instruction 1145 // selected during this invocation. 1146 if (JAfterStart) { 1147 Start = llvm::next(J); 1148 IAfterStart = JAfterStart = false; 1149 } 1150 1151 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1152 << *I << " <-> " << *J << " (cost savings: " << 1153 CostSavings << ")\n"); 1154 1155 // If we have already found too many pairs, break here and this function 1156 // will be called again starting after the last instruction selected 1157 // during this invocation. 1158 if (PairableInsts.size() >= Config.MaxInsts) { 1159 ShouldContinue = true; 1160 break; 1161 } 1162 } 1163 1164 if (ShouldContinue) 1165 break; 1166 } 1167 1168 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1169 << " instructions with candidate pairs\n"); 1170 1171 return ShouldContinue; 1172 } 1173 1174 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1175 // it looks for pairs such that both members have an input which is an 1176 // output of PI or PJ. 1177 void BBVectorize::computePairsConnectedTo( 1178 std::multimap<Value *, Value *> &CandidatePairs, 1179 std::vector<Value *> &PairableInsts, 1180 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1181 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1182 ValuePair P) { 1183 StoreInst *SI, *SJ; 1184 1185 // For each possible pairing for this variable, look at the uses of 1186 // the first value... 1187 for (Value::use_iterator I = P.first->use_begin(), 1188 E = P.first->use_end(); I != E; ++I) { 1189 if (isa<LoadInst>(*I)) { 1190 // A pair cannot be connected to a load because the load only takes one 1191 // operand (the address) and it is a scalar even after vectorization. 1192 continue; 1193 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1194 P.first == SI->getPointerOperand()) { 1195 // Similarly, a pair cannot be connected to a store through its 1196 // pointer operand. 1197 continue; 1198 } 1199 1200 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1201 1202 // For each use of the first variable, look for uses of the second 1203 // variable... 1204 for (Value::use_iterator J = P.second->use_begin(), 1205 E2 = P.second->use_end(); J != E2; ++J) { 1206 if ((SJ = dyn_cast<StoreInst>(*J)) && 1207 P.second == SJ->getPointerOperand()) 1208 continue; 1209 1210 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 1211 1212 // Look for <I, J>: 1213 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1214 VPPair VP(P, ValuePair(*I, *J)); 1215 ConnectedPairs.insert(VP); 1216 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1217 } 1218 1219 // Look for <J, I>: 1220 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) { 1221 VPPair VP(P, ValuePair(*J, *I)); 1222 ConnectedPairs.insert(VP); 1223 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1224 } 1225 } 1226 1227 if (Config.SplatBreaksChain) continue; 1228 // Look for cases where just the first value in the pair is used by 1229 // both members of another pair (splatting). 1230 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1231 if ((SJ = dyn_cast<StoreInst>(*J)) && 1232 P.first == SJ->getPointerOperand()) 1233 continue; 1234 1235 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1236 VPPair VP(P, ValuePair(*I, *J)); 1237 ConnectedPairs.insert(VP); 1238 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1239 } 1240 } 1241 } 1242 1243 if (Config.SplatBreaksChain) return; 1244 // Look for cases where just the second value in the pair is used by 1245 // both members of another pair (splatting). 1246 for (Value::use_iterator I = P.second->use_begin(), 1247 E = P.second->use_end(); I != E; ++I) { 1248 if (isa<LoadInst>(*I)) 1249 continue; 1250 else if ((SI = dyn_cast<StoreInst>(*I)) && 1251 P.second == SI->getPointerOperand()) 1252 continue; 1253 1254 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1255 1256 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1257 if ((SJ = dyn_cast<StoreInst>(*J)) && 1258 P.second == SJ->getPointerOperand()) 1259 continue; 1260 1261 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1262 VPPair VP(P, ValuePair(*I, *J)); 1263 ConnectedPairs.insert(VP); 1264 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1265 } 1266 } 1267 } 1268 } 1269 1270 // This function figures out which pairs are connected. Two pairs are 1271 // connected if some output of the first pair forms an input to both members 1272 // of the second pair. 1273 void BBVectorize::computeConnectedPairs( 1274 std::multimap<Value *, Value *> &CandidatePairs, 1275 std::vector<Value *> &PairableInsts, 1276 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1277 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1278 1279 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1280 PE = PairableInsts.end(); PI != PE; ++PI) { 1281 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1282 1283 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1284 P != choiceRange.second; ++P) 1285 computePairsConnectedTo(CandidatePairs, PairableInsts, 1286 ConnectedPairs, PairConnectionTypes, *P); 1287 } 1288 1289 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1290 << " pair connections.\n"); 1291 } 1292 1293 // This function builds a set of use tuples such that <A, B> is in the set 1294 // if B is in the use tree of A. If B is in the use tree of A, then B 1295 // depends on the output of A. 1296 void BBVectorize::buildDepMap( 1297 BasicBlock &BB, 1298 std::multimap<Value *, Value *> &CandidatePairs, 1299 std::vector<Value *> &PairableInsts, 1300 DenseSet<ValuePair> &PairableInstUsers) { 1301 DenseSet<Value *> IsInPair; 1302 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1303 E = CandidatePairs.end(); C != E; ++C) { 1304 IsInPair.insert(C->first); 1305 IsInPair.insert(C->second); 1306 } 1307 1308 // Iterate through the basic block, recording all Users of each 1309 // pairable instruction. 1310 1311 BasicBlock::iterator E = BB.end(); 1312 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1313 if (IsInPair.find(I) == IsInPair.end()) continue; 1314 1315 DenseSet<Value *> Users; 1316 AliasSetTracker WriteSet(*AA); 1317 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1318 (void) trackUsesOfI(Users, WriteSet, I, J); 1319 1320 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1321 U != E; ++U) 1322 PairableInstUsers.insert(ValuePair(I, *U)); 1323 } 1324 } 1325 1326 // Returns true if an input to pair P is an output of pair Q and also an 1327 // input of pair Q is an output of pair P. If this is the case, then these 1328 // two pairs cannot be simultaneously fused. 1329 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1330 DenseSet<ValuePair> &PairableInstUsers, 1331 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) { 1332 // Two pairs are in conflict if they are mutual Users of eachother. 1333 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1334 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1335 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1336 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1337 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1338 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1339 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1340 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1341 if (PairableInstUserMap) { 1342 // FIXME: The expensive part of the cycle check is not so much the cycle 1343 // check itself but this edge insertion procedure. This needs some 1344 // profiling and probably a different data structure (same is true of 1345 // most uses of std::multimap). 1346 if (PUsesQ) { 1347 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q); 1348 if (!isSecondInIteratorPair(P, QPairRange)) 1349 PairableInstUserMap->insert(VPPair(Q, P)); 1350 } 1351 if (QUsesP) { 1352 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P); 1353 if (!isSecondInIteratorPair(Q, PPairRange)) 1354 PairableInstUserMap->insert(VPPair(P, Q)); 1355 } 1356 } 1357 1358 return (QUsesP && PUsesQ); 1359 } 1360 1361 // This function walks the use graph of current pairs to see if, starting 1362 // from P, the walk returns to P. 1363 bool BBVectorize::pairWillFormCycle(ValuePair P, 1364 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1365 DenseSet<ValuePair> &CurrentPairs) { 1366 DEBUG(if (DebugCycleCheck) 1367 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1368 << *P.second << "\n"); 1369 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1370 // contains non-direct associations. 1371 DenseSet<ValuePair> Visited; 1372 SmallVector<ValuePair, 32> Q; 1373 // General depth-first post-order traversal: 1374 Q.push_back(P); 1375 do { 1376 ValuePair QTop = Q.pop_back_val(); 1377 Visited.insert(QTop); 1378 1379 DEBUG(if (DebugCycleCheck) 1380 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1381 << *QTop.second << "\n"); 1382 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1383 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1384 C != QPairRange.second; ++C) { 1385 if (C->second == P) { 1386 DEBUG(dbgs() 1387 << "BBV: rejected to prevent non-trivial cycle formation: " 1388 << *C->first.first << " <-> " << *C->first.second << "\n"); 1389 return true; 1390 } 1391 1392 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1393 Q.push_back(C->second); 1394 } 1395 } while (!Q.empty()); 1396 1397 return false; 1398 } 1399 1400 // This function builds the initial tree of connected pairs with the 1401 // pair J at the root. 1402 void BBVectorize::buildInitialTreeFor( 1403 std::multimap<Value *, Value *> &CandidatePairs, 1404 std::vector<Value *> &PairableInsts, 1405 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1406 DenseSet<ValuePair> &PairableInstUsers, 1407 DenseMap<Value *, Value *> &ChosenPairs, 1408 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1409 // Each of these pairs is viewed as the root node of a Tree. The Tree 1410 // is then walked (depth-first). As this happens, we keep track of 1411 // the pairs that compose the Tree and the maximum depth of the Tree. 1412 SmallVector<ValuePairWithDepth, 32> Q; 1413 // General depth-first post-order traversal: 1414 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1415 do { 1416 ValuePairWithDepth QTop = Q.back(); 1417 1418 // Push each child onto the queue: 1419 bool MoreChildren = false; 1420 size_t MaxChildDepth = QTop.second; 1421 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1422 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1423 k != qtRange.second; ++k) { 1424 // Make sure that this child pair is still a candidate: 1425 bool IsStillCand = false; 1426 VPIteratorPair checkRange = 1427 CandidatePairs.equal_range(k->second.first); 1428 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1429 m != checkRange.second; ++m) { 1430 if (m->second == k->second.second) { 1431 IsStillCand = true; 1432 break; 1433 } 1434 } 1435 1436 if (IsStillCand) { 1437 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1438 if (C == Tree.end()) { 1439 size_t d = getDepthFactor(k->second.first); 1440 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1441 MoreChildren = true; 1442 } else { 1443 MaxChildDepth = std::max(MaxChildDepth, C->second); 1444 } 1445 } 1446 } 1447 1448 if (!MoreChildren) { 1449 // Record the current pair as part of the Tree: 1450 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1451 Q.pop_back(); 1452 } 1453 } while (!Q.empty()); 1454 } 1455 1456 // Given some initial tree, prune it by removing conflicting pairs (pairs 1457 // that cannot be simultaneously chosen for vectorization). 1458 void BBVectorize::pruneTreeFor( 1459 std::multimap<Value *, Value *> &CandidatePairs, 1460 std::vector<Value *> &PairableInsts, 1461 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1462 DenseSet<ValuePair> &PairableInstUsers, 1463 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1464 DenseMap<Value *, Value *> &ChosenPairs, 1465 DenseMap<ValuePair, size_t> &Tree, 1466 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1467 bool UseCycleCheck) { 1468 SmallVector<ValuePairWithDepth, 32> Q; 1469 // General depth-first post-order traversal: 1470 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1471 do { 1472 ValuePairWithDepth QTop = Q.pop_back_val(); 1473 PrunedTree.insert(QTop.first); 1474 1475 // Visit each child, pruning as necessary... 1476 DenseMap<ValuePair, size_t> BestChildren; 1477 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1478 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1479 K != QTopRange.second; ++K) { 1480 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1481 if (C == Tree.end()) continue; 1482 1483 // This child is in the Tree, now we need to make sure it is the 1484 // best of any conflicting children. There could be multiple 1485 // conflicting children, so first, determine if we're keeping 1486 // this child, then delete conflicting children as necessary. 1487 1488 // It is also necessary to guard against pairing-induced 1489 // dependencies. Consider instructions a .. x .. y .. b 1490 // such that (a,b) are to be fused and (x,y) are to be fused 1491 // but a is an input to x and b is an output from y. This 1492 // means that y cannot be moved after b but x must be moved 1493 // after b for (a,b) to be fused. In other words, after 1494 // fusing (a,b) we have y .. a/b .. x where y is an input 1495 // to a/b and x is an output to a/b: x and y can no longer 1496 // be legally fused. To prevent this condition, we must 1497 // make sure that a child pair added to the Tree is not 1498 // both an input and output of an already-selected pair. 1499 1500 // Pairing-induced dependencies can also form from more complicated 1501 // cycles. The pair vs. pair conflicts are easy to check, and so 1502 // that is done explicitly for "fast rejection", and because for 1503 // child vs. child conflicts, we may prefer to keep the current 1504 // pair in preference to the already-selected child. 1505 DenseSet<ValuePair> CurrentPairs; 1506 1507 bool CanAdd = true; 1508 for (DenseMap<ValuePair, size_t>::iterator C2 1509 = BestChildren.begin(), E2 = BestChildren.end(); 1510 C2 != E2; ++C2) { 1511 if (C2->first.first == C->first.first || 1512 C2->first.first == C->first.second || 1513 C2->first.second == C->first.first || 1514 C2->first.second == C->first.second || 1515 pairsConflict(C2->first, C->first, PairableInstUsers, 1516 UseCycleCheck ? &PairableInstUserMap : 0)) { 1517 if (C2->second >= C->second) { 1518 CanAdd = false; 1519 break; 1520 } 1521 1522 CurrentPairs.insert(C2->first); 1523 } 1524 } 1525 if (!CanAdd) continue; 1526 1527 // Even worse, this child could conflict with another node already 1528 // selected for the Tree. If that is the case, ignore this child. 1529 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1530 E2 = PrunedTree.end(); T != E2; ++T) { 1531 if (T->first == C->first.first || 1532 T->first == C->first.second || 1533 T->second == C->first.first || 1534 T->second == C->first.second || 1535 pairsConflict(*T, C->first, PairableInstUsers, 1536 UseCycleCheck ? &PairableInstUserMap : 0)) { 1537 CanAdd = false; 1538 break; 1539 } 1540 1541 CurrentPairs.insert(*T); 1542 } 1543 if (!CanAdd) continue; 1544 1545 // And check the queue too... 1546 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1547 E2 = Q.end(); C2 != E2; ++C2) { 1548 if (C2->first.first == C->first.first || 1549 C2->first.first == C->first.second || 1550 C2->first.second == C->first.first || 1551 C2->first.second == C->first.second || 1552 pairsConflict(C2->first, C->first, PairableInstUsers, 1553 UseCycleCheck ? &PairableInstUserMap : 0)) { 1554 CanAdd = false; 1555 break; 1556 } 1557 1558 CurrentPairs.insert(C2->first); 1559 } 1560 if (!CanAdd) continue; 1561 1562 // Last but not least, check for a conflict with any of the 1563 // already-chosen pairs. 1564 for (DenseMap<Value *, Value *>::iterator C2 = 1565 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1566 C2 != E2; ++C2) { 1567 if (pairsConflict(*C2, C->first, PairableInstUsers, 1568 UseCycleCheck ? &PairableInstUserMap : 0)) { 1569 CanAdd = false; 1570 break; 1571 } 1572 1573 CurrentPairs.insert(*C2); 1574 } 1575 if (!CanAdd) continue; 1576 1577 // To check for non-trivial cycles formed by the addition of the 1578 // current pair we've formed a list of all relevant pairs, now use a 1579 // graph walk to check for a cycle. We start from the current pair and 1580 // walk the use tree to see if we again reach the current pair. If we 1581 // do, then the current pair is rejected. 1582 1583 // FIXME: It may be more efficient to use a topological-ordering 1584 // algorithm to improve the cycle check. This should be investigated. 1585 if (UseCycleCheck && 1586 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1587 continue; 1588 1589 // This child can be added, but we may have chosen it in preference 1590 // to an already-selected child. Check for this here, and if a 1591 // conflict is found, then remove the previously-selected child 1592 // before adding this one in its place. 1593 for (DenseMap<ValuePair, size_t>::iterator C2 1594 = BestChildren.begin(); C2 != BestChildren.end();) { 1595 if (C2->first.first == C->first.first || 1596 C2->first.first == C->first.second || 1597 C2->first.second == C->first.first || 1598 C2->first.second == C->first.second || 1599 pairsConflict(C2->first, C->first, PairableInstUsers)) 1600 BestChildren.erase(C2++); 1601 else 1602 ++C2; 1603 } 1604 1605 BestChildren.insert(ValuePairWithDepth(C->first, C->second)); 1606 } 1607 1608 for (DenseMap<ValuePair, size_t>::iterator C 1609 = BestChildren.begin(), E2 = BestChildren.end(); 1610 C != E2; ++C) { 1611 size_t DepthF = getDepthFactor(C->first.first); 1612 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1613 } 1614 } while (!Q.empty()); 1615 } 1616 1617 // This function finds the best tree of mututally-compatible connected 1618 // pairs, given the choice of root pairs as an iterator range. 1619 void BBVectorize::findBestTreeFor( 1620 std::multimap<Value *, Value *> &CandidatePairs, 1621 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1622 std::vector<Value *> &PairableInsts, 1623 DenseSet<ValuePair> &FixedOrderPairs, 1624 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1625 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1626 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1627 DenseSet<ValuePair> &PairableInstUsers, 1628 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1629 DenseMap<Value *, Value *> &ChosenPairs, 1630 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1631 int &BestEffSize, VPIteratorPair ChoiceRange, 1632 bool UseCycleCheck) { 1633 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1634 J != ChoiceRange.second; ++J) { 1635 1636 // Before going any further, make sure that this pair does not 1637 // conflict with any already-selected pairs (see comment below 1638 // near the Tree pruning for more details). 1639 DenseSet<ValuePair> ChosenPairSet; 1640 bool DoesConflict = false; 1641 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1642 E = ChosenPairs.end(); C != E; ++C) { 1643 if (pairsConflict(*C, *J, PairableInstUsers, 1644 UseCycleCheck ? &PairableInstUserMap : 0)) { 1645 DoesConflict = true; 1646 break; 1647 } 1648 1649 ChosenPairSet.insert(*C); 1650 } 1651 if (DoesConflict) continue; 1652 1653 if (UseCycleCheck && 1654 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1655 continue; 1656 1657 DenseMap<ValuePair, size_t> Tree; 1658 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1659 PairableInstUsers, ChosenPairs, Tree, *J); 1660 1661 // Because we'll keep the child with the largest depth, the largest 1662 // depth is still the same in the unpruned Tree. 1663 size_t MaxDepth = Tree.lookup(*J); 1664 1665 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1666 << *J->first << " <-> " << *J->second << "} of depth " << 1667 MaxDepth << " and size " << Tree.size() << "\n"); 1668 1669 // At this point the Tree has been constructed, but, may contain 1670 // contradictory children (meaning that different children of 1671 // some tree node may be attempting to fuse the same instruction). 1672 // So now we walk the tree again, in the case of a conflict, 1673 // keep only the child with the largest depth. To break a tie, 1674 // favor the first child. 1675 1676 DenseSet<ValuePair> PrunedTree; 1677 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1678 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, 1679 PrunedTree, *J, UseCycleCheck); 1680 1681 int EffSize = 0; 1682 if (VTTI) { 1683 DenseSet<Value *> PrunedTreeInstrs; 1684 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1685 E = PrunedTree.end(); S != E; ++S) { 1686 PrunedTreeInstrs.insert(S->first); 1687 PrunedTreeInstrs.insert(S->second); 1688 } 1689 1690 // The set of pairs that have already contributed to the total cost. 1691 DenseSet<ValuePair> IncomingPairs; 1692 1693 // The node weights represent the cost savings associated with 1694 // fusing the pair of instructions. 1695 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1696 E = PrunedTree.end(); S != E; ++S) { 1697 bool FlipOrder = false; 1698 1699 if (getDepthFactor(S->first)) { 1700 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1701 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1702 << *S->first << " <-> " << *S->second << "} = " << 1703 ESContrib << "\n"); 1704 EffSize += ESContrib; 1705 } 1706 1707 // The edge weights contribute in a negative sense: they represent 1708 // the cost of shuffles. 1709 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); 1710 if (IP.first != ConnectedPairDeps.end()) { 1711 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1712 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1713 Q != IP.second; ++Q) { 1714 if (!PrunedTree.count(Q->second)) 1715 continue; 1716 DenseMap<VPPair, unsigned>::iterator R = 1717 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1718 assert(R != PairConnectionTypes.end() && 1719 "Cannot find pair connection type"); 1720 if (R->second == PairConnectionDirect) 1721 ++NumDepsDirect; 1722 else if (R->second == PairConnectionSwap) 1723 ++NumDepsSwap; 1724 } 1725 1726 // If there are more swaps than direct connections, then 1727 // the pair order will be flipped during fusion. So the real 1728 // number of swaps is the minimum number. 1729 FlipOrder = !FixedOrderPairs.count(*S) && 1730 ((NumDepsSwap > NumDepsDirect) || 1731 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1732 1733 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1734 Q != IP.second; ++Q) { 1735 if (!PrunedTree.count(Q->second)) 1736 continue; 1737 DenseMap<VPPair, unsigned>::iterator R = 1738 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1739 assert(R != PairConnectionTypes.end() && 1740 "Cannot find pair connection type"); 1741 Type *Ty1 = Q->second.first->getType(), 1742 *Ty2 = Q->second.second->getType(); 1743 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1744 if ((R->second == PairConnectionDirect && FlipOrder) || 1745 (R->second == PairConnectionSwap && !FlipOrder) || 1746 R->second == PairConnectionSplat) { 1747 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1748 VTy, VTy); 1749 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1750 *Q->second.first << " <-> " << *Q->second.second << 1751 "} -> {" << 1752 *S->first << " <-> " << *S->second << "} = " << 1753 ESContrib << "\n"); 1754 EffSize -= ESContrib; 1755 } 1756 } 1757 } 1758 1759 // Compute the cost of outgoing edges. We assume that edges outgoing 1760 // to shuffles, inserts or extracts can be merged, and so contribute 1761 // no additional cost. 1762 if (!S->first->getType()->isVoidTy()) { 1763 Type *Ty1 = S->first->getType(), 1764 *Ty2 = S->second->getType(); 1765 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1766 1767 bool NeedsExtraction = false; 1768 for (Value::use_iterator I = S->first->use_begin(), 1769 IE = S->first->use_end(); I != IE; ++I) { 1770 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1771 // Shuffle can be folded if it has no other input 1772 if (isa<UndefValue>(SI->getOperand(1))) 1773 continue; 1774 } 1775 if (isa<ExtractElementInst>(*I)) 1776 continue; 1777 if (PrunedTreeInstrs.count(*I)) 1778 continue; 1779 NeedsExtraction = true; 1780 break; 1781 } 1782 1783 if (NeedsExtraction) { 1784 int ESContrib; 1785 if (Ty1->isVectorTy()) 1786 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1787 Ty1, VTy); 1788 else 1789 ESContrib = (int) VTTI->getVectorInstrCost( 1790 Instruction::ExtractElement, VTy, 0); 1791 1792 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1793 *S->first << "} = " << ESContrib << "\n"); 1794 EffSize -= ESContrib; 1795 } 1796 1797 NeedsExtraction = false; 1798 for (Value::use_iterator I = S->second->use_begin(), 1799 IE = S->second->use_end(); I != IE; ++I) { 1800 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1801 // Shuffle can be folded if it has no other input 1802 if (isa<UndefValue>(SI->getOperand(1))) 1803 continue; 1804 } 1805 if (isa<ExtractElementInst>(*I)) 1806 continue; 1807 if (PrunedTreeInstrs.count(*I)) 1808 continue; 1809 NeedsExtraction = true; 1810 break; 1811 } 1812 1813 if (NeedsExtraction) { 1814 int ESContrib; 1815 if (Ty2->isVectorTy()) 1816 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1817 Ty2, VTy); 1818 else 1819 ESContrib = (int) VTTI->getVectorInstrCost( 1820 Instruction::ExtractElement, VTy, 1); 1821 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1822 *S->second << "} = " << ESContrib << "\n"); 1823 EffSize -= ESContrib; 1824 } 1825 } 1826 1827 // Compute the cost of incoming edges. 1828 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 1829 Instruction *S1 = cast<Instruction>(S->first), 1830 *S2 = cast<Instruction>(S->second); 1831 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 1832 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 1833 1834 // Combining constants into vector constants (or small vector 1835 // constants into larger ones are assumed free). 1836 if (isa<Constant>(O1) && isa<Constant>(O2)) 1837 continue; 1838 1839 if (FlipOrder) 1840 std::swap(O1, O2); 1841 1842 ValuePair VP = ValuePair(O1, O2); 1843 ValuePair VPR = ValuePair(O2, O1); 1844 1845 // Internal edges are not handled here. 1846 if (PrunedTree.count(VP) || PrunedTree.count(VPR)) 1847 continue; 1848 1849 Type *Ty1 = O1->getType(), 1850 *Ty2 = O2->getType(); 1851 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1852 1853 // Combining vector operations of the same type is also assumed 1854 // folded with other operations. 1855 if (Ty1 == Ty2) { 1856 // If both are insert elements, then both can be widened. 1857 if (isa<InsertElementInst>(O1) && isa<InsertElementInst>(O2)) 1858 continue; 1859 // If both are extract elements, and both have the same input 1860 // type, then they can be replaced with a shuffle 1861 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 1862 *EIO2 = dyn_cast<ExtractElementInst>(O2); 1863 if (EIO1 && EIO2 && 1864 EIO1->getOperand(0)->getType() == 1865 EIO2->getOperand(0)->getType()) 1866 continue; 1867 // If both are a shuffle with equal operand types and only two 1868 // unqiue operands, then they can be replaced with a single 1869 // shuffle 1870 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 1871 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 1872 if (SIO1 && SIO2 && 1873 SIO1->getOperand(0)->getType() == 1874 SIO2->getOperand(0)->getType()) { 1875 SmallSet<Value *, 4> SIOps; 1876 SIOps.insert(SIO1->getOperand(0)); 1877 SIOps.insert(SIO1->getOperand(1)); 1878 SIOps.insert(SIO2->getOperand(0)); 1879 SIOps.insert(SIO2->getOperand(1)); 1880 if (SIOps.size() <= 2) 1881 continue; 1882 } 1883 } 1884 1885 int ESContrib; 1886 // This pair has already been formed. 1887 if (IncomingPairs.count(VP)) { 1888 continue; 1889 } else if (IncomingPairs.count(VPR)) { 1890 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1891 VTy, VTy); 1892 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 1893 ESContrib = (int) VTTI->getVectorInstrCost( 1894 Instruction::InsertElement, VTy, 0); 1895 ESContrib += (int) VTTI->getVectorInstrCost( 1896 Instruction::InsertElement, VTy, 1); 1897 } else if (!Ty1->isVectorTy()) { 1898 // O1 needs to be inserted into a vector of size O2, and then 1899 // both need to be shuffled together. 1900 ESContrib = (int) VTTI->getVectorInstrCost( 1901 Instruction::InsertElement, Ty2, 0); 1902 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 1903 VTy, Ty2); 1904 } else if (!Ty2->isVectorTy()) { 1905 // O2 needs to be inserted into a vector of size O1, and then 1906 // both need to be shuffled together. 1907 ESContrib = (int) VTTI->getVectorInstrCost( 1908 Instruction::InsertElement, Ty1, 0); 1909 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 1910 VTy, Ty1); 1911 } else { 1912 Type *TyBig = Ty1, *TySmall = Ty2; 1913 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 1914 std::swap(TyBig, TySmall); 1915 1916 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1917 VTy, TyBig); 1918 if (TyBig != TySmall) 1919 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 1920 TyBig, TySmall); 1921 } 1922 1923 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 1924 << *O1 << " <-> " << *O2 << "} = " << 1925 ESContrib << "\n"); 1926 EffSize -= ESContrib; 1927 IncomingPairs.insert(VP); 1928 } 1929 } 1930 } 1931 } else { 1932 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1933 E = PrunedTree.end(); S != E; ++S) 1934 EffSize += (int) getDepthFactor(S->first); 1935 } 1936 1937 DEBUG(if (DebugPairSelection) 1938 dbgs() << "BBV: found pruned Tree for pair {" 1939 << *J->first << " <-> " << *J->second << "} of depth " << 1940 MaxDepth << " and size " << PrunedTree.size() << 1941 " (effective size: " << EffSize << ")\n"); 1942 if (((VTTI && !UseChainDepthWithTI) || 1943 MaxDepth >= Config.ReqChainDepth) && 1944 EffSize > 0 && EffSize > BestEffSize) { 1945 BestMaxDepth = MaxDepth; 1946 BestEffSize = EffSize; 1947 BestTree = PrunedTree; 1948 } 1949 } 1950 } 1951 1952 // Given the list of candidate pairs, this function selects those 1953 // that will be fused into vector instructions. 1954 void BBVectorize::choosePairs( 1955 std::multimap<Value *, Value *> &CandidatePairs, 1956 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1957 std::vector<Value *> &PairableInsts, 1958 DenseSet<ValuePair> &FixedOrderPairs, 1959 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1960 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1961 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1962 DenseSet<ValuePair> &PairableInstUsers, 1963 DenseMap<Value *, Value *>& ChosenPairs) { 1964 bool UseCycleCheck = 1965 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 1966 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 1967 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 1968 E = PairableInsts.end(); I != E; ++I) { 1969 // The number of possible pairings for this variable: 1970 size_t NumChoices = CandidatePairs.count(*I); 1971 if (!NumChoices) continue; 1972 1973 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 1974 1975 // The best pair to choose and its tree: 1976 size_t BestMaxDepth = 0; 1977 int BestEffSize = 0; 1978 DenseSet<ValuePair> BestTree; 1979 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 1980 PairableInsts, FixedOrderPairs, PairConnectionTypes, 1981 ConnectedPairs, ConnectedPairDeps, 1982 PairableInstUsers, PairableInstUserMap, ChosenPairs, 1983 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 1984 UseCycleCheck); 1985 1986 // A tree has been chosen (or not) at this point. If no tree was 1987 // chosen, then this instruction, I, cannot be paired (and is no longer 1988 // considered). 1989 1990 DEBUG(if (BestTree.size() > 0) 1991 dbgs() << "BBV: selected pairs in the best tree for: " 1992 << *cast<Instruction>(*I) << "\n"); 1993 1994 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 1995 SE2 = BestTree.end(); S != SE2; ++S) { 1996 // Insert the members of this tree into the list of chosen pairs. 1997 ChosenPairs.insert(ValuePair(S->first, S->second)); 1998 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 1999 *S->second << "\n"); 2000 2001 // Remove all candidate pairs that have values in the chosen tree. 2002 for (std::multimap<Value *, Value *>::iterator K = 2003 CandidatePairs.begin(); K != CandidatePairs.end();) { 2004 if (K->first == S->first || K->second == S->first || 2005 K->second == S->second || K->first == S->second) { 2006 // Don't remove the actual pair chosen so that it can be used 2007 // in subsequent tree selections. 2008 if (!(K->first == S->first && K->second == S->second)) 2009 CandidatePairs.erase(K++); 2010 else 2011 ++K; 2012 } else { 2013 ++K; 2014 } 2015 } 2016 } 2017 } 2018 2019 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2020 } 2021 2022 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2023 unsigned n = 0) { 2024 if (!I->hasName()) 2025 return ""; 2026 2027 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2028 (n > 0 ? "." + utostr(n) : "")).str(); 2029 } 2030 2031 // Returns the value that is to be used as the pointer input to the vector 2032 // instruction that fuses I with J. 2033 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2034 Instruction *I, Instruction *J, unsigned o) { 2035 Value *IPtr, *JPtr; 2036 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2037 int64_t OffsetInElmts; 2038 2039 // Note: the analysis might fail here, that is why the pair order has 2040 // been precomputed (OffsetInElmts must be unused here). 2041 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2042 IAddressSpace, JAddressSpace, 2043 OffsetInElmts, false); 2044 2045 // The pointer value is taken to be the one with the lowest offset. 2046 Value *VPtr = IPtr; 2047 2048 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 2049 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 2050 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2051 Type *VArgPtrType = PointerType::get(VArgType, 2052 cast<PointerType>(IPtr->getType())->getAddressSpace()); 2053 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2054 /* insert before */ I); 2055 } 2056 2057 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2058 unsigned MaskOffset, unsigned NumInElem, 2059 unsigned NumInElem1, unsigned IdxOffset, 2060 std::vector<Constant*> &Mask) { 2061 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 2062 for (unsigned v = 0; v < NumElem1; ++v) { 2063 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2064 if (m < 0) { 2065 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2066 } else { 2067 unsigned mm = m + (int) IdxOffset; 2068 if (m >= (int) NumInElem1) 2069 mm += (int) NumInElem; 2070 2071 Mask[v+MaskOffset] = 2072 ConstantInt::get(Type::getInt32Ty(Context), mm); 2073 } 2074 } 2075 } 2076 2077 // Returns the value that is to be used as the vector-shuffle mask to the 2078 // vector instruction that fuses I with J. 2079 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2080 Instruction *I, Instruction *J) { 2081 // This is the shuffle mask. We need to append the second 2082 // mask to the first, and the numbers need to be adjusted. 2083 2084 Type *ArgTypeI = I->getType(); 2085 Type *ArgTypeJ = J->getType(); 2086 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2087 2088 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 2089 2090 // Get the total number of elements in the fused vector type. 2091 // By definition, this must equal the number of elements in 2092 // the final mask. 2093 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 2094 std::vector<Constant*> Mask(NumElem); 2095 2096 Type *OpTypeI = I->getOperand(0)->getType(); 2097 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 2098 Type *OpTypeJ = J->getOperand(0)->getType(); 2099 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 2100 2101 // The fused vector will be: 2102 // ----------------------------------------------------- 2103 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2104 // ----------------------------------------------------- 2105 // from which we'll extract NumElem total elements (where the first NumElemI 2106 // of them come from the mask in I and the remainder come from the mask 2107 // in J. 2108 2109 // For the mask from the first pair... 2110 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2111 0, Mask); 2112 2113 // For the mask from the second pair... 2114 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2115 NumInElemI, Mask); 2116 2117 return ConstantVector::get(Mask); 2118 } 2119 2120 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2121 Instruction *J, unsigned o, Value *&LOp, 2122 unsigned numElemL, 2123 Type *ArgTypeL, Type *ArgTypeH, 2124 bool IBeforeJ, unsigned IdxOff) { 2125 bool ExpandedIEChain = false; 2126 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2127 // If we have a pure insertelement chain, then this can be rewritten 2128 // into a chain that directly builds the larger type. 2129 bool PureChain = true; 2130 InsertElementInst *LIENext = LIE; 2131 do { 2132 if (!isa<UndefValue>(LIENext->getOperand(0)) && 2133 !isa<InsertElementInst>(LIENext->getOperand(0))) { 2134 PureChain = false; 2135 break; 2136 } 2137 } while ((LIENext = 2138 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2139 2140 if (PureChain) { 2141 SmallVector<Value *, 8> VectElemts(numElemL, 2142 UndefValue::get(ArgTypeL->getScalarType())); 2143 InsertElementInst *LIENext = LIE; 2144 do { 2145 unsigned Idx = 2146 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2147 VectElemts[Idx] = LIENext->getOperand(1); 2148 } while ((LIENext = 2149 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2150 2151 LIENext = 0; 2152 Value *LIEPrev = UndefValue::get(ArgTypeH); 2153 for (unsigned i = 0; i < numElemL; ++i) { 2154 if (isa<UndefValue>(VectElemts[i])) continue; 2155 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2156 ConstantInt::get(Type::getInt32Ty(Context), 2157 i + IdxOff), 2158 getReplacementName(IBeforeJ ? I : J, 2159 true, o, i+1)); 2160 LIENext->insertBefore(IBeforeJ ? J : I); 2161 LIEPrev = LIENext; 2162 } 2163 2164 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2165 ExpandedIEChain = true; 2166 } 2167 } 2168 2169 return ExpandedIEChain; 2170 } 2171 2172 // Returns the value to be used as the specified operand of the vector 2173 // instruction that fuses I with J. 2174 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2175 Instruction *J, unsigned o, bool IBeforeJ) { 2176 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2177 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2178 2179 // Compute the fused vector type for this operand 2180 Type *ArgTypeI = I->getOperand(o)->getType(); 2181 Type *ArgTypeJ = J->getOperand(o)->getType(); 2182 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2183 2184 Instruction *L = I, *H = J; 2185 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2186 2187 unsigned numElemL; 2188 if (ArgTypeL->isVectorTy()) 2189 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 2190 else 2191 numElemL = 1; 2192 2193 unsigned numElemH; 2194 if (ArgTypeH->isVectorTy()) 2195 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 2196 else 2197 numElemH = 1; 2198 2199 Value *LOp = L->getOperand(o); 2200 Value *HOp = H->getOperand(o); 2201 unsigned numElem = VArgType->getNumElements(); 2202 2203 // First, we check if we can reuse the "original" vector outputs (if these 2204 // exist). We might need a shuffle. 2205 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2206 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2207 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2208 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2209 2210 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2211 // optimization. The input vectors to the shuffle might be a different 2212 // length from the shuffle outputs. Unfortunately, the replacement 2213 // shuffle mask has already been formed, and the mask entries are sensitive 2214 // to the sizes of the inputs. 2215 bool IsSizeChangeShuffle = 2216 isa<ShuffleVectorInst>(L) && 2217 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2218 2219 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2220 // We can have at most two unique vector inputs. 2221 bool CanUseInputs = true; 2222 Value *I1, *I2 = 0; 2223 if (LEE) { 2224 I1 = LEE->getOperand(0); 2225 } else { 2226 I1 = LSV->getOperand(0); 2227 I2 = LSV->getOperand(1); 2228 if (I2 == I1 || isa<UndefValue>(I2)) 2229 I2 = 0; 2230 } 2231 2232 if (HEE) { 2233 Value *I3 = HEE->getOperand(0); 2234 if (!I2 && I3 != I1) 2235 I2 = I3; 2236 else if (I3 != I1 && I3 != I2) 2237 CanUseInputs = false; 2238 } else { 2239 Value *I3 = HSV->getOperand(0); 2240 if (!I2 && I3 != I1) 2241 I2 = I3; 2242 else if (I3 != I1 && I3 != I2) 2243 CanUseInputs = false; 2244 2245 if (CanUseInputs) { 2246 Value *I4 = HSV->getOperand(1); 2247 if (!isa<UndefValue>(I4)) { 2248 if (!I2 && I4 != I1) 2249 I2 = I4; 2250 else if (I4 != I1 && I4 != I2) 2251 CanUseInputs = false; 2252 } 2253 } 2254 } 2255 2256 if (CanUseInputs) { 2257 unsigned LOpElem = 2258 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2259 ->getNumElements(); 2260 unsigned HOpElem = 2261 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2262 ->getNumElements(); 2263 2264 // We have one or two input vectors. We need to map each index of the 2265 // operands to the index of the original vector. 2266 SmallVector<std::pair<int, int>, 8> II(numElem); 2267 for (unsigned i = 0; i < numElemL; ++i) { 2268 int Idx, INum; 2269 if (LEE) { 2270 Idx = 2271 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2272 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2273 } else { 2274 Idx = LSV->getMaskValue(i); 2275 if (Idx < (int) LOpElem) { 2276 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2277 } else { 2278 Idx -= LOpElem; 2279 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2280 } 2281 } 2282 2283 II[i] = std::pair<int, int>(Idx, INum); 2284 } 2285 for (unsigned i = 0; i < numElemH; ++i) { 2286 int Idx, INum; 2287 if (HEE) { 2288 Idx = 2289 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2290 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2291 } else { 2292 Idx = HSV->getMaskValue(i); 2293 if (Idx < (int) HOpElem) { 2294 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2295 } else { 2296 Idx -= HOpElem; 2297 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2298 } 2299 } 2300 2301 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2302 } 2303 2304 // We now have an array which tells us from which index of which 2305 // input vector each element of the operand comes. 2306 VectorType *I1T = cast<VectorType>(I1->getType()); 2307 unsigned I1Elem = I1T->getNumElements(); 2308 2309 if (!I2) { 2310 // In this case there is only one underlying vector input. Check for 2311 // the trivial case where we can use the input directly. 2312 if (I1Elem == numElem) { 2313 bool ElemInOrder = true; 2314 for (unsigned i = 0; i < numElem; ++i) { 2315 if (II[i].first != (int) i && II[i].first != -1) { 2316 ElemInOrder = false; 2317 break; 2318 } 2319 } 2320 2321 if (ElemInOrder) 2322 return I1; 2323 } 2324 2325 // A shuffle is needed. 2326 std::vector<Constant *> Mask(numElem); 2327 for (unsigned i = 0; i < numElem; ++i) { 2328 int Idx = II[i].first; 2329 if (Idx == -1) 2330 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2331 else 2332 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2333 } 2334 2335 Instruction *S = 2336 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2337 ConstantVector::get(Mask), 2338 getReplacementName(IBeforeJ ? I : J, 2339 true, o)); 2340 S->insertBefore(IBeforeJ ? J : I); 2341 return S; 2342 } 2343 2344 VectorType *I2T = cast<VectorType>(I2->getType()); 2345 unsigned I2Elem = I2T->getNumElements(); 2346 2347 // This input comes from two distinct vectors. The first step is to 2348 // make sure that both vectors are the same length. If not, the 2349 // smaller one will need to grow before they can be shuffled together. 2350 if (I1Elem < I2Elem) { 2351 std::vector<Constant *> Mask(I2Elem); 2352 unsigned v = 0; 2353 for (; v < I1Elem; ++v) 2354 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2355 for (; v < I2Elem; ++v) 2356 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2357 2358 Instruction *NewI1 = 2359 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2360 ConstantVector::get(Mask), 2361 getReplacementName(IBeforeJ ? I : J, 2362 true, o, 1)); 2363 NewI1->insertBefore(IBeforeJ ? J : I); 2364 I1 = NewI1; 2365 I1T = I2T; 2366 I1Elem = I2Elem; 2367 } else if (I1Elem > I2Elem) { 2368 std::vector<Constant *> Mask(I1Elem); 2369 unsigned v = 0; 2370 for (; v < I2Elem; ++v) 2371 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2372 for (; v < I1Elem; ++v) 2373 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2374 2375 Instruction *NewI2 = 2376 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2377 ConstantVector::get(Mask), 2378 getReplacementName(IBeforeJ ? I : J, 2379 true, o, 1)); 2380 NewI2->insertBefore(IBeforeJ ? J : I); 2381 I2 = NewI2; 2382 I2T = I1T; 2383 I2Elem = I1Elem; 2384 } 2385 2386 // Now that both I1 and I2 are the same length we can shuffle them 2387 // together (and use the result). 2388 std::vector<Constant *> Mask(numElem); 2389 for (unsigned v = 0; v < numElem; ++v) { 2390 if (II[v].first == -1) { 2391 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2392 } else { 2393 int Idx = II[v].first + II[v].second * I1Elem; 2394 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2395 } 2396 } 2397 2398 Instruction *NewOp = 2399 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2400 getReplacementName(IBeforeJ ? I : J, true, o)); 2401 NewOp->insertBefore(IBeforeJ ? J : I); 2402 return NewOp; 2403 } 2404 } 2405 2406 Type *ArgType = ArgTypeL; 2407 if (numElemL < numElemH) { 2408 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2409 ArgTypeL, VArgType, IBeforeJ, 1)) { 2410 // This is another short-circuit case: we're combining a scalar into 2411 // a vector that is formed by an IE chain. We've just expanded the IE 2412 // chain, now insert the scalar and we're done. 2413 2414 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2415 getReplacementName(IBeforeJ ? I : J, true, o)); 2416 S->insertBefore(IBeforeJ ? J : I); 2417 return S; 2418 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2419 ArgTypeH, IBeforeJ)) { 2420 // The two vector inputs to the shuffle must be the same length, 2421 // so extend the smaller vector to be the same length as the larger one. 2422 Instruction *NLOp; 2423 if (numElemL > 1) { 2424 2425 std::vector<Constant *> Mask(numElemH); 2426 unsigned v = 0; 2427 for (; v < numElemL; ++v) 2428 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2429 for (; v < numElemH; ++v) 2430 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2431 2432 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2433 ConstantVector::get(Mask), 2434 getReplacementName(IBeforeJ ? I : J, 2435 true, o, 1)); 2436 } else { 2437 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2438 getReplacementName(IBeforeJ ? I : J, 2439 true, o, 1)); 2440 } 2441 2442 NLOp->insertBefore(IBeforeJ ? J : I); 2443 LOp = NLOp; 2444 } 2445 2446 ArgType = ArgTypeH; 2447 } else if (numElemL > numElemH) { 2448 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2449 ArgTypeH, VArgType, IBeforeJ)) { 2450 Instruction *S = 2451 InsertElementInst::Create(LOp, HOp, 2452 ConstantInt::get(Type::getInt32Ty(Context), 2453 numElemL), 2454 getReplacementName(IBeforeJ ? I : J, 2455 true, o)); 2456 S->insertBefore(IBeforeJ ? J : I); 2457 return S; 2458 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2459 ArgTypeL, IBeforeJ)) { 2460 Instruction *NHOp; 2461 if (numElemH > 1) { 2462 std::vector<Constant *> Mask(numElemL); 2463 unsigned v = 0; 2464 for (; v < numElemH; ++v) 2465 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2466 for (; v < numElemL; ++v) 2467 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2468 2469 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2470 ConstantVector::get(Mask), 2471 getReplacementName(IBeforeJ ? I : J, 2472 true, o, 1)); 2473 } else { 2474 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2475 getReplacementName(IBeforeJ ? I : J, 2476 true, o, 1)); 2477 } 2478 2479 NHOp->insertBefore(IBeforeJ ? J : I); 2480 HOp = NHOp; 2481 } 2482 } 2483 2484 if (ArgType->isVectorTy()) { 2485 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2486 std::vector<Constant*> Mask(numElem); 2487 for (unsigned v = 0; v < numElem; ++v) { 2488 unsigned Idx = v; 2489 // If the low vector was expanded, we need to skip the extra 2490 // undefined entries. 2491 if (v >= numElemL && numElemH > numElemL) 2492 Idx += (numElemH - numElemL); 2493 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2494 } 2495 2496 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2497 ConstantVector::get(Mask), 2498 getReplacementName(IBeforeJ ? I : J, true, o)); 2499 BV->insertBefore(IBeforeJ ? J : I); 2500 return BV; 2501 } 2502 2503 Instruction *BV1 = InsertElementInst::Create( 2504 UndefValue::get(VArgType), LOp, CV0, 2505 getReplacementName(IBeforeJ ? I : J, 2506 true, o, 1)); 2507 BV1->insertBefore(IBeforeJ ? J : I); 2508 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2509 getReplacementName(IBeforeJ ? I : J, 2510 true, o, 2)); 2511 BV2->insertBefore(IBeforeJ ? J : I); 2512 return BV2; 2513 } 2514 2515 // This function creates an array of values that will be used as the inputs 2516 // to the vector instruction that fuses I with J. 2517 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2518 Instruction *I, Instruction *J, 2519 SmallVector<Value *, 3> &ReplacedOperands, 2520 bool IBeforeJ) { 2521 unsigned NumOperands = I->getNumOperands(); 2522 2523 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2524 // Iterate backward so that we look at the store pointer 2525 // first and know whether or not we need to flip the inputs. 2526 2527 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2528 // This is the pointer for a load/store instruction. 2529 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2530 continue; 2531 } else if (isa<CallInst>(I)) { 2532 Function *F = cast<CallInst>(I)->getCalledFunction(); 2533 unsigned IID = F->getIntrinsicID(); 2534 if (o == NumOperands-1) { 2535 BasicBlock &BB = *I->getParent(); 2536 2537 Module *M = BB.getParent()->getParent(); 2538 Type *ArgTypeI = I->getType(); 2539 Type *ArgTypeJ = J->getType(); 2540 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2541 2542 ReplacedOperands[o] = Intrinsic::getDeclaration(M, 2543 (Intrinsic::ID) IID, VArgType); 2544 continue; 2545 } else if (IID == Intrinsic::powi && o == 1) { 2546 // The second argument of powi is a single integer and we've already 2547 // checked that both arguments are equal. As a result, we just keep 2548 // I's second argument. 2549 ReplacedOperands[o] = I->getOperand(o); 2550 continue; 2551 } 2552 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2553 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2554 continue; 2555 } 2556 2557 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2558 } 2559 } 2560 2561 // This function creates two values that represent the outputs of the 2562 // original I and J instructions. These are generally vector shuffles 2563 // or extracts. In many cases, these will end up being unused and, thus, 2564 // eliminated by later passes. 2565 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2566 Instruction *J, Instruction *K, 2567 Instruction *&InsertionPt, 2568 Instruction *&K1, Instruction *&K2) { 2569 if (isa<StoreInst>(I)) { 2570 AA->replaceWithNewValue(I, K); 2571 AA->replaceWithNewValue(J, K); 2572 } else { 2573 Type *IType = I->getType(); 2574 Type *JType = J->getType(); 2575 2576 VectorType *VType = getVecTypeForPair(IType, JType); 2577 unsigned numElem = VType->getNumElements(); 2578 2579 unsigned numElemI, numElemJ; 2580 if (IType->isVectorTy()) 2581 numElemI = cast<VectorType>(IType)->getNumElements(); 2582 else 2583 numElemI = 1; 2584 2585 if (JType->isVectorTy()) 2586 numElemJ = cast<VectorType>(JType)->getNumElements(); 2587 else 2588 numElemJ = 1; 2589 2590 if (IType->isVectorTy()) { 2591 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2592 for (unsigned v = 0; v < numElemI; ++v) { 2593 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2594 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2595 } 2596 2597 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2598 ConstantVector::get( Mask1), 2599 getReplacementName(K, false, 1)); 2600 } else { 2601 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2602 K1 = ExtractElementInst::Create(K, CV0, 2603 getReplacementName(K, false, 1)); 2604 } 2605 2606 if (JType->isVectorTy()) { 2607 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2608 for (unsigned v = 0; v < numElemJ; ++v) { 2609 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2610 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2611 } 2612 2613 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2614 ConstantVector::get( Mask2), 2615 getReplacementName(K, false, 2)); 2616 } else { 2617 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2618 K2 = ExtractElementInst::Create(K, CV1, 2619 getReplacementName(K, false, 2)); 2620 } 2621 2622 K1->insertAfter(K); 2623 K2->insertAfter(K1); 2624 InsertionPt = K2; 2625 } 2626 } 2627 2628 // Move all uses of the function I (including pairing-induced uses) after J. 2629 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2630 std::multimap<Value *, Value *> &LoadMoveSet, 2631 Instruction *I, Instruction *J) { 2632 // Skip to the first instruction past I. 2633 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2634 2635 DenseSet<Value *> Users; 2636 AliasSetTracker WriteSet(*AA); 2637 for (; cast<Instruction>(L) != J; ++L) 2638 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet); 2639 2640 assert(cast<Instruction>(L) == J && 2641 "Tracking has not proceeded far enough to check for dependencies"); 2642 // If J is now in the use set of I, then trackUsesOfI will return true 2643 // and we have a dependency cycle (and the fusing operation must abort). 2644 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet); 2645 } 2646 2647 // Move all uses of the function I (including pairing-induced uses) after J. 2648 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2649 std::multimap<Value *, Value *> &LoadMoveSet, 2650 Instruction *&InsertionPt, 2651 Instruction *I, Instruction *J) { 2652 // Skip to the first instruction past I. 2653 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2654 2655 DenseSet<Value *> Users; 2656 AliasSetTracker WriteSet(*AA); 2657 for (; cast<Instruction>(L) != J;) { 2658 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) { 2659 // Move this instruction 2660 Instruction *InstToMove = L; ++L; 2661 2662 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2663 " to after " << *InsertionPt << "\n"); 2664 InstToMove->removeFromParent(); 2665 InstToMove->insertAfter(InsertionPt); 2666 InsertionPt = InstToMove; 2667 } else { 2668 ++L; 2669 } 2670 } 2671 } 2672 2673 // Collect all load instruction that are in the move set of a given first 2674 // pair member. These loads depend on the first instruction, I, and so need 2675 // to be moved after J (the second instruction) when the pair is fused. 2676 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2677 DenseMap<Value *, Value *> &ChosenPairs, 2678 std::multimap<Value *, Value *> &LoadMoveSet, 2679 Instruction *I) { 2680 // Skip to the first instruction past I. 2681 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2682 2683 DenseSet<Value *> Users; 2684 AliasSetTracker WriteSet(*AA); 2685 2686 // Note: We cannot end the loop when we reach J because J could be moved 2687 // farther down the use chain by another instruction pairing. Also, J 2688 // could be before I if this is an inverted input. 2689 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2690 if (trackUsesOfI(Users, WriteSet, I, L)) { 2691 if (L->mayReadFromMemory()) 2692 LoadMoveSet.insert(ValuePair(L, I)); 2693 } 2694 } 2695 } 2696 2697 // In cases where both load/stores and the computation of their pointers 2698 // are chosen for vectorization, we can end up in a situation where the 2699 // aliasing analysis starts returning different query results as the 2700 // process of fusing instruction pairs continues. Because the algorithm 2701 // relies on finding the same use trees here as were found earlier, we'll 2702 // need to precompute the necessary aliasing information here and then 2703 // manually update it during the fusion process. 2704 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2705 std::vector<Value *> &PairableInsts, 2706 DenseMap<Value *, Value *> &ChosenPairs, 2707 std::multimap<Value *, Value *> &LoadMoveSet) { 2708 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2709 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2710 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2711 if (P == ChosenPairs.end()) continue; 2712 2713 Instruction *I = cast<Instruction>(P->first); 2714 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I); 2715 } 2716 } 2717 2718 // When the first instruction in each pair is cloned, it will inherit its 2719 // parent's metadata. This metadata must be combined with that of the other 2720 // instruction in a safe way. 2721 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2722 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2723 K->getAllMetadataOtherThanDebugLoc(Metadata); 2724 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2725 unsigned Kind = Metadata[i].first; 2726 MDNode *JMD = J->getMetadata(Kind); 2727 MDNode *KMD = Metadata[i].second; 2728 2729 switch (Kind) { 2730 default: 2731 K->setMetadata(Kind, 0); // Remove unknown metadata 2732 break; 2733 case LLVMContext::MD_tbaa: 2734 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2735 break; 2736 case LLVMContext::MD_fpmath: 2737 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2738 break; 2739 } 2740 } 2741 } 2742 2743 // This function fuses the chosen instruction pairs into vector instructions, 2744 // taking care preserve any needed scalar outputs and, then, it reorders the 2745 // remaining instructions as needed (users of the first member of the pair 2746 // need to be moved to after the location of the second member of the pair 2747 // because the vector instruction is inserted in the location of the pair's 2748 // second member). 2749 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2750 std::vector<Value *> &PairableInsts, 2751 DenseMap<Value *, Value *> &ChosenPairs, 2752 DenseSet<ValuePair> &FixedOrderPairs, 2753 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2754 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2755 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2756 LLVMContext& Context = BB.getContext(); 2757 2758 // During the vectorization process, the order of the pairs to be fused 2759 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2760 // list. After a pair is fused, the flipped pair is removed from the list. 2761 DenseSet<ValuePair> FlippedPairs; 2762 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2763 E = ChosenPairs.end(); P != E; ++P) 2764 FlippedPairs.insert(ValuePair(P->second, P->first)); 2765 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2766 E = FlippedPairs.end(); P != E; ++P) 2767 ChosenPairs.insert(*P); 2768 2769 std::multimap<Value *, Value *> LoadMoveSet; 2770 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); 2771 2772 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2773 2774 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2775 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2776 if (P == ChosenPairs.end()) { 2777 ++PI; 2778 continue; 2779 } 2780 2781 if (getDepthFactor(P->first) == 0) { 2782 // These instructions are not really fused, but are tracked as though 2783 // they are. Any case in which it would be interesting to fuse them 2784 // will be taken care of by InstCombine. 2785 --NumFusedOps; 2786 ++PI; 2787 continue; 2788 } 2789 2790 Instruction *I = cast<Instruction>(P->first), 2791 *J = cast<Instruction>(P->second); 2792 2793 DEBUG(dbgs() << "BBV: fusing: " << *I << 2794 " <-> " << *J << "\n"); 2795 2796 // Remove the pair and flipped pair from the list. 2797 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2798 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2799 ChosenPairs.erase(FP); 2800 ChosenPairs.erase(P); 2801 2802 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) { 2803 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2804 " <-> " << *J << 2805 " aborted because of non-trivial dependency cycle\n"); 2806 --NumFusedOps; 2807 ++PI; 2808 continue; 2809 } 2810 2811 // If the pair must have the other order, then flip it. 2812 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2813 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2814 // This pair does not have a fixed order, and so we might want to 2815 // flip it if that will yield fewer shuffles. We count the number 2816 // of dependencies connected via swaps, and those directly connected, 2817 // and flip the order if the number of swaps is greater. 2818 bool OrigOrder = true; 2819 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2820 if (IP.first == ConnectedPairDeps.end()) { 2821 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2822 OrigOrder = false; 2823 } 2824 2825 if (IP.first != ConnectedPairDeps.end()) { 2826 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2827 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2828 Q != IP.second; ++Q) { 2829 DenseMap<VPPair, unsigned>::iterator R = 2830 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2831 assert(R != PairConnectionTypes.end() && 2832 "Cannot find pair connection type"); 2833 if (R->second == PairConnectionDirect) 2834 ++NumDepsDirect; 2835 else if (R->second == PairConnectionSwap) 2836 ++NumDepsSwap; 2837 } 2838 2839 if (!OrigOrder) 2840 std::swap(NumDepsDirect, NumDepsSwap); 2841 2842 if (NumDepsSwap > NumDepsDirect) { 2843 FlipPairOrder = true; 2844 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2845 " <-> " << *J << "\n"); 2846 } 2847 } 2848 } 2849 2850 Instruction *L = I, *H = J; 2851 if (FlipPairOrder) 2852 std::swap(H, L); 2853 2854 // If the pair being fused uses the opposite order from that in the pair 2855 // connection map, then we need to flip the types. 2856 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2857 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2858 Q != IP.second; ++Q) { 2859 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2860 assert(R != PairConnectionTypes.end() && 2861 "Cannot find pair connection type"); 2862 if (R->second == PairConnectionDirect) 2863 R->second = PairConnectionSwap; 2864 else if (R->second == PairConnectionSwap) 2865 R->second = PairConnectionDirect; 2866 } 2867 2868 bool LBeforeH = !FlipPairOrder; 2869 unsigned NumOperands = I->getNumOperands(); 2870 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2871 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2872 LBeforeH); 2873 2874 // Make a copy of the original operation, change its type to the vector 2875 // type and replace its operands with the vector operands. 2876 Instruction *K = L->clone(); 2877 if (L->hasName()) 2878 K->takeName(L); 2879 else if (H->hasName()) 2880 K->takeName(H); 2881 2882 if (!isa<StoreInst>(K)) 2883 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 2884 2885 combineMetadata(K, H); 2886 2887 for (unsigned o = 0; o < NumOperands; ++o) 2888 K->setOperand(o, ReplacedOperands[o]); 2889 2890 K->insertAfter(J); 2891 2892 // Instruction insertion point: 2893 Instruction *InsertionPt = K; 2894 Instruction *K1 = 0, *K2 = 0; 2895 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 2896 2897 // The use tree of the first original instruction must be moved to after 2898 // the location of the second instruction. The entire use tree of the 2899 // first instruction is disjoint from the input tree of the second 2900 // (by definition), and so commutes with it. 2901 2902 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); 2903 2904 if (!isa<StoreInst>(I)) { 2905 L->replaceAllUsesWith(K1); 2906 H->replaceAllUsesWith(K2); 2907 AA->replaceWithNewValue(L, K1); 2908 AA->replaceWithNewValue(H, K2); 2909 } 2910 2911 // Instructions that may read from memory may be in the load move set. 2912 // Once an instruction is fused, we no longer need its move set, and so 2913 // the values of the map never need to be updated. However, when a load 2914 // is fused, we need to merge the entries from both instructions in the 2915 // pair in case those instructions were in the move set of some other 2916 // yet-to-be-fused pair. The loads in question are the keys of the map. 2917 if (I->mayReadFromMemory()) { 2918 std::vector<ValuePair> NewSetMembers; 2919 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 2920 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 2921 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 2922 N != IPairRange.second; ++N) 2923 NewSetMembers.push_back(ValuePair(K, N->second)); 2924 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 2925 N != JPairRange.second; ++N) 2926 NewSetMembers.push_back(ValuePair(K, N->second)); 2927 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 2928 AE = NewSetMembers.end(); A != AE; ++A) 2929 LoadMoveSet.insert(*A); 2930 } 2931 2932 // Before removing I, set the iterator to the next instruction. 2933 PI = llvm::next(BasicBlock::iterator(I)); 2934 if (cast<Instruction>(PI) == J) 2935 ++PI; 2936 2937 SE->forgetValue(I); 2938 SE->forgetValue(J); 2939 I->eraseFromParent(); 2940 J->eraseFromParent(); 2941 2942 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 2943 BB << "\n"); 2944 } 2945 2946 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 2947 } 2948} 2949 2950char BBVectorize::ID = 0; 2951static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 2952INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2953INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2954INITIALIZE_PASS_DEPENDENCY(DominatorTree) 2955INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2956INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 2957 2958BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 2959 return new BBVectorize(C); 2960} 2961 2962bool 2963llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 2964 BBVectorize BBVectorizer(P, C); 2965 return BBVectorizer.vectorizeBB(BB); 2966} 2967 2968//===----------------------------------------------------------------------===// 2969VectorizeConfig::VectorizeConfig() { 2970 VectorBits = ::VectorBits; 2971 VectorizeBools = !::NoBools; 2972 VectorizeInts = !::NoInts; 2973 VectorizeFloats = !::NoFloats; 2974 VectorizePointers = !::NoPointers; 2975 VectorizeCasts = !::NoCasts; 2976 VectorizeMath = !::NoMath; 2977 VectorizeFMA = !::NoFMA; 2978 VectorizeSelect = !::NoSelect; 2979 VectorizeCmp = !::NoCmp; 2980 VectorizeGEP = !::NoGEP; 2981 VectorizeMemOps = !::NoMemOps; 2982 AlignedOnly = ::AlignedOnly; 2983 ReqChainDepth= ::ReqChainDepth; 2984 SearchLimit = ::SearchLimit; 2985 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 2986 SplatBreaksChain = ::SplatBreaksChain; 2987 MaxInsts = ::MaxInsts; 2988 MaxIter = ::MaxIter; 2989 Pow2LenOnly = ::Pow2LenOnly; 2990 NoMemOpBoost = ::NoMemOpBoost; 2991 FastDep = ::FastDep; 2992} 2993