BBVectorize.cpp revision 2f0e63cc16feb39480805bd00f53bbe5e3031d29
1//===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements a basic-block vectorization pass. The algorithm was 11// inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral, 12// et al. It works by looking for chains of pairable operations and then 13// pairing them. 14// 15//===----------------------------------------------------------------------===// 16 17#define BBV_NAME "bb-vectorize" 18#define DEBUG_TYPE BBV_NAME 19#include "llvm/Transforms/Vectorize.h" 20#include "llvm/ADT/DenseMap.h" 21#include "llvm/ADT/DenseSet.h" 22#include "llvm/ADT/STLExtras.h" 23#include "llvm/ADT/SmallSet.h" 24#include "llvm/ADT/SmallVector.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/ADT/StringExtras.h" 27#include "llvm/Analysis/AliasAnalysis.h" 28#include "llvm/Analysis/AliasSetTracker.h" 29#include "llvm/Analysis/Dominators.h" 30#include "llvm/Analysis/ScalarEvolution.h" 31#include "llvm/Analysis/ScalarEvolutionExpressions.h" 32#include "llvm/Analysis/TargetTransformInfo.h" 33#include "llvm/Analysis/ValueTracking.h" 34#include "llvm/IR/Constants.h" 35#include "llvm/IR/DataLayout.h" 36#include "llvm/IR/DerivedTypes.h" 37#include "llvm/IR/Function.h" 38#include "llvm/IR/Instructions.h" 39#include "llvm/IR/IntrinsicInst.h" 40#include "llvm/IR/Intrinsics.h" 41#include "llvm/IR/LLVMContext.h" 42#include "llvm/IR/Metadata.h" 43#include "llvm/IR/Type.h" 44#include "llvm/Pass.h" 45#include "llvm/Support/CommandLine.h" 46#include "llvm/Support/Debug.h" 47#include "llvm/Support/ValueHandle.h" 48#include "llvm/Support/raw_ostream.h" 49#include "llvm/Transforms/Utils/Local.h" 50#include <algorithm> 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 : &P->getAnalysis<TargetTransformInfo>(); 203 } 204 205 typedef std::pair<Value *, Value *> ValuePair; 206 typedef std::pair<ValuePair, int> ValuePairWithCost; 207 typedef std::pair<ValuePair, size_t> ValuePairWithDepth; 208 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair 209 typedef std::pair<VPPair, unsigned> VPPairWithType; 210 typedef std::pair<std::multimap<Value *, Value *>::iterator, 211 std::multimap<Value *, Value *>::iterator> VPIteratorPair; 212 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, 213 std::multimap<ValuePair, ValuePair>::iterator> 214 VPPIteratorPair; 215 216 AliasAnalysis *AA; 217 DominatorTree *DT; 218 ScalarEvolution *SE; 219 DataLayout *TD; 220 const TargetTransformInfo *TTI; 221 222 // FIXME: const correct? 223 224 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false); 225 226 bool getCandidatePairs(BasicBlock &BB, 227 BasicBlock::iterator &Start, 228 std::multimap<Value *, Value *> &CandidatePairs, 229 DenseSet<ValuePair> &FixedOrderPairs, 230 DenseMap<ValuePair, int> &CandidatePairCostSavings, 231 std::vector<Value *> &PairableInsts, bool NonPow2Len); 232 233 // FIXME: The current implementation does not account for pairs that 234 // are connected in multiple ways. For example: 235 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap) 236 enum PairConnectionType { 237 PairConnectionDirect, 238 PairConnectionSwap, 239 PairConnectionSplat 240 }; 241 242 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, 243 std::vector<Value *> &PairableInsts, 244 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 245 DenseMap<VPPair, unsigned> &PairConnectionTypes); 246 247 void buildDepMap(BasicBlock &BB, 248 std::multimap<Value *, Value *> &CandidatePairs, 249 std::vector<Value *> &PairableInsts, 250 DenseSet<ValuePair> &PairableInstUsers); 251 252 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 253 DenseMap<ValuePair, int> &CandidatePairCostSavings, 254 std::vector<Value *> &PairableInsts, 255 DenseSet<ValuePair> &FixedOrderPairs, 256 DenseMap<VPPair, unsigned> &PairConnectionTypes, 257 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 258 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 259 DenseSet<ValuePair> &PairableInstUsers, 260 DenseMap<Value *, Value *>& ChosenPairs); 261 262 void fuseChosenPairs(BasicBlock &BB, 263 std::vector<Value *> &PairableInsts, 264 DenseMap<Value *, Value *>& ChosenPairs, 265 DenseSet<ValuePair> &FixedOrderPairs, 266 DenseMap<VPPair, unsigned> &PairConnectionTypes, 267 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 268 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps); 269 270 271 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 272 273 bool areInstsCompatible(Instruction *I, Instruction *J, 274 bool IsSimpleLoadStore, bool NonPow2Len, 275 int &CostSavings, int &FixedOrder); 276 277 bool trackUsesOfI(DenseSet<Value *> &Users, 278 AliasSetTracker &WriteSet, Instruction *I, 279 Instruction *J, bool UpdateUsers = true, 280 DenseSet<ValuePair> *LoadMoveSetPairs = 0); 281 282 void computePairsConnectedTo( 283 std::multimap<Value *, Value *> &CandidatePairs, 284 std::vector<Value *> &PairableInsts, 285 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 286 DenseMap<VPPair, unsigned> &PairConnectionTypes, 287 ValuePair P); 288 289 bool pairsConflict(ValuePair P, ValuePair Q, 290 DenseSet<ValuePair> &PairableInstUsers, 291 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0, 292 DenseSet<VPPair> *PairableInstUserPairSet = 0); 293 294 bool pairWillFormCycle(ValuePair P, 295 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 296 DenseSet<ValuePair> &CurrentPairs); 297 298 void pruneTreeFor( 299 std::multimap<Value *, Value *> &CandidatePairs, 300 std::vector<Value *> &PairableInsts, 301 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 302 DenseSet<ValuePair> &PairableInstUsers, 303 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 304 DenseSet<VPPair> &PairableInstUserPairSet, 305 DenseMap<Value *, Value *> &ChosenPairs, 306 DenseMap<ValuePair, size_t> &Tree, 307 DenseSet<ValuePair> &PrunedTree, ValuePair J, 308 bool UseCycleCheck); 309 310 void buildInitialTreeFor( 311 std::multimap<Value *, Value *> &CandidatePairs, 312 std::vector<Value *> &PairableInsts, 313 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 314 DenseSet<ValuePair> &PairableInstUsers, 315 DenseMap<Value *, Value *> &ChosenPairs, 316 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 317 318 void findBestTreeFor( 319 std::multimap<Value *, Value *> &CandidatePairs, 320 DenseMap<ValuePair, int> &CandidatePairCostSavings, 321 std::vector<Value *> &PairableInsts, 322 DenseSet<ValuePair> &FixedOrderPairs, 323 DenseMap<VPPair, unsigned> &PairConnectionTypes, 324 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 325 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 326 DenseSet<ValuePair> &PairableInstUsers, 327 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 328 DenseSet<VPPair> &PairableInstUserPairSet, 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 DenseSet<ValuePair> &LoadMoveSetPairs, 366 Instruction *I); 367 368 void collectLoadMoveSet(BasicBlock &BB, 369 std::vector<Value *> &PairableInsts, 370 DenseMap<Value *, Value *> &ChosenPairs, 371 std::multimap<Value *, Value *> &LoadMoveSet, 372 DenseSet<ValuePair> &LoadMoveSetPairs); 373 374 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 375 DenseSet<ValuePair> &LoadMoveSetPairs, 376 Instruction *I, Instruction *J); 377 378 void moveUsesOfIAfterJ(BasicBlock &BB, 379 DenseSet<ValuePair> &LoadMoveSetPairs, 380 Instruction *&InsertionPt, 381 Instruction *I, Instruction *J); 382 383 void combineMetadata(Instruction *K, const Instruction *J); 384 385 bool vectorizeBB(BasicBlock &BB) { 386 if (!DT->isReachableFromEntry(&BB)) { 387 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 388 " in " << BB.getParent()->getName() << "\n"); 389 return false; 390 } 391 392 DEBUG(if (TTI) dbgs() << "BBV: using target information\n"); 393 394 bool changed = false; 395 // Iterate a sufficient number of times to merge types of size 1 bit, 396 // then 2 bits, then 4, etc. up to half of the target vector width of the 397 // target vector register. 398 unsigned n = 1; 399 for (unsigned v = 2; 400 (TTI || v <= Config.VectorBits) && 401 (!Config.MaxIter || n <= Config.MaxIter); 402 v *= 2, ++n) { 403 DEBUG(dbgs() << "BBV: fusing loop #" << n << 404 " for " << BB.getName() << " in " << 405 BB.getParent()->getName() << "...\n"); 406 if (vectorizePairs(BB)) 407 changed = true; 408 else 409 break; 410 } 411 412 if (changed && !Pow2LenOnly) { 413 ++n; 414 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 415 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 416 n << " for " << BB.getName() << " in " << 417 BB.getParent()->getName() << "...\n"); 418 if (!vectorizePairs(BB, true)) break; 419 } 420 } 421 422 DEBUG(dbgs() << "BBV: done!\n"); 423 return changed; 424 } 425 426 virtual bool runOnBasicBlock(BasicBlock &BB) { 427 AA = &getAnalysis<AliasAnalysis>(); 428 DT = &getAnalysis<DominatorTree>(); 429 SE = &getAnalysis<ScalarEvolution>(); 430 TD = getAnalysisIfAvailable<DataLayout>(); 431 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>(); 432 433 return vectorizeBB(BB); 434 } 435 436 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 437 BasicBlockPass::getAnalysisUsage(AU); 438 AU.addRequired<AliasAnalysis>(); 439 AU.addRequired<DominatorTree>(); 440 AU.addRequired<ScalarEvolution>(); 441 AU.addRequired<TargetTransformInfo>(); 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 (StoreInst *SI = dyn_cast<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 = SI->getValueOperand(); 476 T1 = IVal->getType(); 477 } else { 478 T1 = I->getType(); 479 } 480 481 if (CastInst *CI = dyn_cast<CastInst>(I)) 482 T2 = CI->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 TTI. 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 TTI->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 TTI->getArithmeticInstrCost(Opcode, T1); 557 case Instruction::Select: 558 case Instruction::ICmp: 559 case Instruction::FCmp: 560 return TTI->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 TTI->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 Intrinsic::ID IID = (Intrinsic::ID) 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 case Intrinsic::fmuladd: 665 return Config.VectorizeFMA; 666 } 667 } 668 669 // Returns true if J is the second element in some pair referenced by 670 // some multimap pair iterator pair. 671 template <typename V> 672 bool isSecondInIteratorPair(V J, std::pair< 673 typename std::multimap<V, V>::iterator, 674 typename std::multimap<V, V>::iterator> PairRange) { 675 for (typename std::multimap<V, V>::iterator K = PairRange.first; 676 K != PairRange.second; ++K) 677 if (K->second == J) return true; 678 679 return false; 680 } 681 682 bool isPureIEChain(InsertElementInst *IE) { 683 InsertElementInst *IENext = IE; 684 do { 685 if (!isa<UndefValue>(IENext->getOperand(0)) && 686 !isa<InsertElementInst>(IENext->getOperand(0))) { 687 return false; 688 } 689 } while ((IENext = 690 dyn_cast<InsertElementInst>(IENext->getOperand(0)))); 691 692 return true; 693 } 694 }; 695 696 // This function implements one vectorization iteration on the provided 697 // basic block. It returns true if the block is changed. 698 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 699 bool ShouldContinue; 700 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 701 702 std::vector<Value *> AllPairableInsts; 703 DenseMap<Value *, Value *> AllChosenPairs; 704 DenseSet<ValuePair> AllFixedOrderPairs; 705 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 706 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; 707 708 do { 709 std::vector<Value *> PairableInsts; 710 std::multimap<Value *, Value *> CandidatePairs; 711 DenseSet<ValuePair> FixedOrderPairs; 712 DenseMap<ValuePair, int> CandidatePairCostSavings; 713 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 714 FixedOrderPairs, 715 CandidatePairCostSavings, 716 PairableInsts, NonPow2Len); 717 if (PairableInsts.empty()) continue; 718 719 // Now we have a map of all of the pairable instructions and we need to 720 // select the best possible pairing. A good pairing is one such that the 721 // users of the pair are also paired. This defines a (directed) forest 722 // over the pairs such that two pairs are connected iff the second pair 723 // uses the first. 724 725 // Note that it only matters that both members of the second pair use some 726 // element of the first pair (to allow for splatting). 727 728 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; 729 DenseMap<VPPair, unsigned> PairConnectionTypes; 730 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs, 731 PairConnectionTypes); 732 if (ConnectedPairs.empty()) continue; 733 734 for (std::multimap<ValuePair, ValuePair>::iterator 735 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 736 I != IE; ++I) { 737 ConnectedPairDeps.insert(VPPair(I->second, I->first)); 738 } 739 740 // Build the pairable-instruction dependency map 741 DenseSet<ValuePair> PairableInstUsers; 742 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 743 744 // There is now a graph of the connected pairs. For each variable, pick 745 // the pairing with the largest tree meeting the depth requirement on at 746 // least one branch. Then select all pairings that are part of that tree 747 // and remove them from the list of available pairings and pairable 748 // variables. 749 750 DenseMap<Value *, Value *> ChosenPairs; 751 choosePairs(CandidatePairs, CandidatePairCostSavings, 752 PairableInsts, FixedOrderPairs, PairConnectionTypes, 753 ConnectedPairs, ConnectedPairDeps, 754 PairableInstUsers, ChosenPairs); 755 756 if (ChosenPairs.empty()) continue; 757 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 758 PairableInsts.end()); 759 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 760 761 // Only for the chosen pairs, propagate information on fixed-order pairs, 762 // pair connections, and their types to the data structures used by the 763 // pair fusion procedures. 764 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 765 IE = ChosenPairs.end(); I != IE; ++I) { 766 if (FixedOrderPairs.count(*I)) 767 AllFixedOrderPairs.insert(*I); 768 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 769 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 770 771 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 772 J != IE; ++J) { 773 DenseMap<VPPair, unsigned>::iterator K = 774 PairConnectionTypes.find(VPPair(*I, *J)); 775 if (K != PairConnectionTypes.end()) { 776 AllPairConnectionTypes.insert(*K); 777 } else { 778 K = PairConnectionTypes.find(VPPair(*J, *I)); 779 if (K != PairConnectionTypes.end()) 780 AllPairConnectionTypes.insert(*K); 781 } 782 } 783 } 784 785 for (std::multimap<ValuePair, ValuePair>::iterator 786 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 787 I != IE; ++I) { 788 if (AllPairConnectionTypes.count(*I)) { 789 AllConnectedPairs.insert(*I); 790 AllConnectedPairDeps.insert(VPPair(I->second, I->first)); 791 } 792 } 793 } while (ShouldContinue); 794 795 if (AllChosenPairs.empty()) return false; 796 NumFusedOps += AllChosenPairs.size(); 797 798 // A set of pairs has now been selected. It is now necessary to replace the 799 // paired instructions with vector instructions. For this procedure each 800 // operand must be replaced with a vector operand. This vector is formed 801 // by using build_vector on the old operands. The replaced values are then 802 // replaced with a vector_extract on the result. Subsequent optimization 803 // passes should coalesce the build/extract combinations. 804 805 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 806 AllPairConnectionTypes, 807 AllConnectedPairs, AllConnectedPairDeps); 808 809 // It is important to cleanup here so that future iterations of this 810 // function have less work to do. 811 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 812 return true; 813 } 814 815 // This function returns true if the provided instruction is capable of being 816 // fused into a vector instruction. This determination is based only on the 817 // type and other attributes of the instruction. 818 bool BBVectorize::isInstVectorizable(Instruction *I, 819 bool &IsSimpleLoadStore) { 820 IsSimpleLoadStore = false; 821 822 if (CallInst *C = dyn_cast<CallInst>(I)) { 823 if (!isVectorizableIntrinsic(C)) 824 return false; 825 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 826 // Vectorize simple loads if possbile: 827 IsSimpleLoadStore = L->isSimple(); 828 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 829 return false; 830 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 831 // Vectorize simple stores if possbile: 832 IsSimpleLoadStore = S->isSimple(); 833 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 834 return false; 835 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 836 // We can vectorize casts, but not casts of pointer types, etc. 837 if (!Config.VectorizeCasts) 838 return false; 839 840 Type *SrcTy = C->getSrcTy(); 841 if (!SrcTy->isSingleValueType()) 842 return false; 843 844 Type *DestTy = C->getDestTy(); 845 if (!DestTy->isSingleValueType()) 846 return false; 847 } else if (isa<SelectInst>(I)) { 848 if (!Config.VectorizeSelect) 849 return false; 850 } else if (isa<CmpInst>(I)) { 851 if (!Config.VectorizeCmp) 852 return false; 853 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 854 if (!Config.VectorizeGEP) 855 return false; 856 857 // Currently, vector GEPs exist only with one index. 858 if (G->getNumIndices() != 1) 859 return false; 860 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 861 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 862 return false; 863 } 864 865 // We can't vectorize memory operations without target data 866 if (TD == 0 && IsSimpleLoadStore) 867 return false; 868 869 Type *T1, *T2; 870 getInstructionTypes(I, T1, T2); 871 872 // Not every type can be vectorized... 873 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 874 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 875 return false; 876 877 if (T1->getScalarSizeInBits() == 1) { 878 if (!Config.VectorizeBools) 879 return false; 880 } else { 881 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 882 return false; 883 } 884 885 if (T2->getScalarSizeInBits() == 1) { 886 if (!Config.VectorizeBools) 887 return false; 888 } else { 889 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 890 return false; 891 } 892 893 if (!Config.VectorizeFloats 894 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 895 return false; 896 897 // Don't vectorize target-specific types. 898 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 899 return false; 900 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 901 return false; 902 903 if ((!Config.VectorizePointers || TD == 0) && 904 (T1->getScalarType()->isPointerTy() || 905 T2->getScalarType()->isPointerTy())) 906 return false; 907 908 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 909 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 910 return false; 911 912 return true; 913 } 914 915 // This function returns true if the two provided instructions are compatible 916 // (meaning that they can be fused into a vector instruction). This assumes 917 // that I has already been determined to be vectorizable and that J is not 918 // in the use tree of I. 919 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 920 bool IsSimpleLoadStore, bool NonPow2Len, 921 int &CostSavings, int &FixedOrder) { 922 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 923 " <-> " << *J << "\n"); 924 925 CostSavings = 0; 926 FixedOrder = 0; 927 928 // Loads and stores can be merged if they have different alignments, 929 // but are otherwise the same. 930 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 931 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 932 return false; 933 934 Type *IT1, *IT2, *JT1, *JT2; 935 getInstructionTypes(I, IT1, IT2); 936 getInstructionTypes(J, JT1, JT2); 937 unsigned MaxTypeBits = std::max( 938 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 939 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 940 if (!TTI && MaxTypeBits > Config.VectorBits) 941 return false; 942 943 // FIXME: handle addsub-type operations! 944 945 if (IsSimpleLoadStore) { 946 Value *IPtr, *JPtr; 947 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 948 int64_t OffsetInElmts = 0; 949 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 950 IAddressSpace, JAddressSpace, 951 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 952 FixedOrder = (int) OffsetInElmts; 953 unsigned BottomAlignment = IAlignment; 954 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 955 956 Type *aTypeI = isa<StoreInst>(I) ? 957 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 958 Type *aTypeJ = isa<StoreInst>(J) ? 959 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 960 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 961 962 if (Config.AlignedOnly) { 963 // An aligned load or store is possible only if the instruction 964 // with the lower offset has an alignment suitable for the 965 // vector type. 966 967 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 968 if (BottomAlignment < VecAlignment) 969 return false; 970 } 971 972 if (TTI) { 973 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI, 974 IAlignment, IAddressSpace); 975 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ, 976 JAlignment, JAddressSpace); 977 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType, 978 BottomAlignment, 979 IAddressSpace); 980 981 ICost += TTI->getAddressComputationCost(aTypeI); 982 JCost += TTI->getAddressComputationCost(aTypeJ); 983 VCost += TTI->getAddressComputationCost(VType); 984 985 if (VCost > ICost + JCost) 986 return false; 987 988 // We don't want to fuse to a type that will be split, even 989 // if the two input types will also be split and there is no other 990 // associated cost. 991 unsigned VParts = TTI->getNumberOfParts(VType); 992 if (VParts > 1) 993 return false; 994 else if (!VParts && VCost == ICost + JCost) 995 return false; 996 997 CostSavings = ICost + JCost - VCost; 998 } 999 } else { 1000 return false; 1001 } 1002 } else if (TTI) { 1003 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 1004 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 1005 Type *VT1 = getVecTypeForPair(IT1, JT1), 1006 *VT2 = getVecTypeForPair(IT2, JT2); 1007 1008 // Note that this procedure is incorrect for insert and extract element 1009 // instructions (because combining these often results in a shuffle), 1010 // but this cost is ignored (because insert and extract element 1011 // instructions are assigned a zero depth factor and are not really 1012 // fused in general). 1013 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 1014 1015 if (VCost > ICost + JCost) 1016 return false; 1017 1018 // We don't want to fuse to a type that will be split, even 1019 // if the two input types will also be split and there is no other 1020 // associated cost. 1021 unsigned VParts1 = TTI->getNumberOfParts(VT1), 1022 VParts2 = TTI->getNumberOfParts(VT2); 1023 if (VParts1 > 1 || VParts2 > 1) 1024 return false; 1025 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) 1026 return false; 1027 1028 CostSavings = ICost + JCost - VCost; 1029 } 1030 1031 // The powi intrinsic is special because only the first argument is 1032 // vectorized, the second arguments must be equal. 1033 CallInst *CI = dyn_cast<CallInst>(I); 1034 Function *FI; 1035 if (CI && (FI = CI->getCalledFunction())) { 1036 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID(); 1037 if (IID == Intrinsic::powi) { 1038 Value *A1I = CI->getArgOperand(1), 1039 *A1J = cast<CallInst>(J)->getArgOperand(1); 1040 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1041 *A1JSCEV = SE->getSCEV(A1J); 1042 return (A1ISCEV == A1JSCEV); 1043 } 1044 1045 if (IID && TTI) { 1046 SmallVector<Type*, 4> Tys; 1047 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 1048 Tys.push_back(CI->getArgOperand(i)->getType()); 1049 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys); 1050 1051 Tys.clear(); 1052 CallInst *CJ = cast<CallInst>(J); 1053 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i) 1054 Tys.push_back(CJ->getArgOperand(i)->getType()); 1055 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys); 1056 1057 Tys.clear(); 1058 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() && 1059 "Intrinsic argument counts differ"); 1060 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1061 if (IID == Intrinsic::powi && i == 1) 1062 Tys.push_back(CI->getArgOperand(i)->getType()); 1063 else 1064 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(), 1065 CJ->getArgOperand(i)->getType())); 1066 } 1067 1068 Type *RetTy = getVecTypeForPair(IT1, JT1); 1069 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys); 1070 1071 if (VCost > ICost + JCost) 1072 return false; 1073 1074 // We don't want to fuse to a type that will be split, even 1075 // if the two input types will also be split and there is no other 1076 // associated cost. 1077 unsigned RetParts = TTI->getNumberOfParts(RetTy); 1078 if (RetParts > 1) 1079 return false; 1080 else if (!RetParts && VCost == ICost + JCost) 1081 return false; 1082 1083 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1084 if (!Tys[i]->isVectorTy()) 1085 continue; 1086 1087 unsigned NumParts = TTI->getNumberOfParts(Tys[i]); 1088 if (NumParts > 1) 1089 return false; 1090 else if (!NumParts && VCost == ICost + JCost) 1091 return false; 1092 } 1093 1094 CostSavings = ICost + JCost - VCost; 1095 } 1096 } 1097 1098 return true; 1099 } 1100 1101 // Figure out whether or not J uses I and update the users and write-set 1102 // structures associated with I. Specifically, Users represents the set of 1103 // instructions that depend on I. WriteSet represents the set 1104 // of memory locations that are dependent on I. If UpdateUsers is true, 1105 // and J uses I, then Users is updated to contain J and WriteSet is updated 1106 // to contain any memory locations to which J writes. The function returns 1107 // true if J uses I. By default, alias analysis is used to determine 1108 // whether J reads from memory that overlaps with a location in WriteSet. 1109 // If LoadMoveSet is not null, then it is a previously-computed multimap 1110 // where the key is the memory-based user instruction and the value is 1111 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1112 // then the alias analysis is not used. This is necessary because this 1113 // function is called during the process of moving instructions during 1114 // vectorization and the results of the alias analysis are not stable during 1115 // that process. 1116 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1117 AliasSetTracker &WriteSet, Instruction *I, 1118 Instruction *J, bool UpdateUsers, 1119 DenseSet<ValuePair> *LoadMoveSetPairs) { 1120 bool UsesI = false; 1121 1122 // This instruction may already be marked as a user due, for example, to 1123 // being a member of a selected pair. 1124 if (Users.count(J)) 1125 UsesI = true; 1126 1127 if (!UsesI) 1128 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1129 JU != JE; ++JU) { 1130 Value *V = *JU; 1131 if (I == V || Users.count(V)) { 1132 UsesI = true; 1133 break; 1134 } 1135 } 1136 if (!UsesI && J->mayReadFromMemory()) { 1137 if (LoadMoveSetPairs) { 1138 UsesI = LoadMoveSetPairs->count(ValuePair(J, I)); 1139 } else { 1140 for (AliasSetTracker::iterator W = WriteSet.begin(), 1141 WE = WriteSet.end(); W != WE; ++W) { 1142 if (W->aliasesUnknownInst(J, *AA)) { 1143 UsesI = true; 1144 break; 1145 } 1146 } 1147 } 1148 } 1149 1150 if (UsesI && UpdateUsers) { 1151 if (J->mayWriteToMemory()) WriteSet.add(J); 1152 Users.insert(J); 1153 } 1154 1155 return UsesI; 1156 } 1157 1158 // This function iterates over all instruction pairs in the provided 1159 // basic block and collects all candidate pairs for vectorization. 1160 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1161 BasicBlock::iterator &Start, 1162 std::multimap<Value *, Value *> &CandidatePairs, 1163 DenseSet<ValuePair> &FixedOrderPairs, 1164 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1165 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1166 BasicBlock::iterator E = BB.end(); 1167 if (Start == E) return false; 1168 1169 bool ShouldContinue = false, IAfterStart = false; 1170 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1171 if (I == Start) IAfterStart = true; 1172 1173 bool IsSimpleLoadStore; 1174 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1175 1176 // Look for an instruction with which to pair instruction *I... 1177 DenseSet<Value *> Users; 1178 AliasSetTracker WriteSet(*AA); 1179 bool JAfterStart = IAfterStart; 1180 BasicBlock::iterator J = llvm::next(I); 1181 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1182 if (J == Start) JAfterStart = true; 1183 1184 // Determine if J uses I, if so, exit the loop. 1185 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1186 if (Config.FastDep) { 1187 // Note: For this heuristic to be effective, independent operations 1188 // must tend to be intermixed. This is likely to be true from some 1189 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1190 // but otherwise may require some kind of reordering pass. 1191 1192 // When using fast dependency analysis, 1193 // stop searching after first use: 1194 if (UsesI) break; 1195 } else { 1196 if (UsesI) continue; 1197 } 1198 1199 // J does not use I, and comes before the first use of I, so it can be 1200 // merged with I if the instructions are compatible. 1201 int CostSavings, FixedOrder; 1202 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1203 CostSavings, FixedOrder)) continue; 1204 1205 // J is a candidate for merging with I. 1206 if (!PairableInsts.size() || 1207 PairableInsts[PairableInsts.size()-1] != I) { 1208 PairableInsts.push_back(I); 1209 } 1210 1211 CandidatePairs.insert(ValuePair(I, J)); 1212 if (TTI) 1213 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1214 CostSavings)); 1215 1216 if (FixedOrder == 1) 1217 FixedOrderPairs.insert(ValuePair(I, J)); 1218 else if (FixedOrder == -1) 1219 FixedOrderPairs.insert(ValuePair(J, I)); 1220 1221 // The next call to this function must start after the last instruction 1222 // selected during this invocation. 1223 if (JAfterStart) { 1224 Start = llvm::next(J); 1225 IAfterStart = JAfterStart = false; 1226 } 1227 1228 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1229 << *I << " <-> " << *J << " (cost savings: " << 1230 CostSavings << ")\n"); 1231 1232 // If we have already found too many pairs, break here and this function 1233 // will be called again starting after the last instruction selected 1234 // during this invocation. 1235 if (PairableInsts.size() >= Config.MaxInsts) { 1236 ShouldContinue = true; 1237 break; 1238 } 1239 } 1240 1241 if (ShouldContinue) 1242 break; 1243 } 1244 1245 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1246 << " instructions with candidate pairs\n"); 1247 1248 return ShouldContinue; 1249 } 1250 1251 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1252 // it looks for pairs such that both members have an input which is an 1253 // output of PI or PJ. 1254 void BBVectorize::computePairsConnectedTo( 1255 std::multimap<Value *, Value *> &CandidatePairs, 1256 std::vector<Value *> &PairableInsts, 1257 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1258 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1259 ValuePair P) { 1260 StoreInst *SI, *SJ; 1261 1262 // For each possible pairing for this variable, look at the uses of 1263 // the first value... 1264 for (Value::use_iterator I = P.first->use_begin(), 1265 E = P.first->use_end(); I != E; ++I) { 1266 if (isa<LoadInst>(*I)) { 1267 // A pair cannot be connected to a load because the load only takes one 1268 // operand (the address) and it is a scalar even after vectorization. 1269 continue; 1270 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1271 P.first == SI->getPointerOperand()) { 1272 // Similarly, a pair cannot be connected to a store through its 1273 // pointer operand. 1274 continue; 1275 } 1276 1277 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1278 1279 // For each use of the first variable, look for uses of the second 1280 // variable... 1281 for (Value::use_iterator J = P.second->use_begin(), 1282 E2 = P.second->use_end(); J != E2; ++J) { 1283 if ((SJ = dyn_cast<StoreInst>(*J)) && 1284 P.second == SJ->getPointerOperand()) 1285 continue; 1286 1287 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); 1288 1289 // Look for <I, J>: 1290 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1291 VPPair VP(P, ValuePair(*I, *J)); 1292 ConnectedPairs.insert(VP); 1293 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1294 } 1295 1296 // Look for <J, I>: 1297 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) { 1298 VPPair VP(P, ValuePair(*J, *I)); 1299 ConnectedPairs.insert(VP); 1300 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1301 } 1302 } 1303 1304 if (Config.SplatBreaksChain) continue; 1305 // Look for cases where just the first value in the pair is used by 1306 // both members of another pair (splatting). 1307 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1308 if ((SJ = dyn_cast<StoreInst>(*J)) && 1309 P.first == SJ->getPointerOperand()) 1310 continue; 1311 1312 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1313 VPPair VP(P, ValuePair(*I, *J)); 1314 ConnectedPairs.insert(VP); 1315 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1316 } 1317 } 1318 } 1319 1320 if (Config.SplatBreaksChain) return; 1321 // Look for cases where just the second value in the pair is used by 1322 // both members of another pair (splatting). 1323 for (Value::use_iterator I = P.second->use_begin(), 1324 E = P.second->use_end(); I != E; ++I) { 1325 if (isa<LoadInst>(*I)) 1326 continue; 1327 else if ((SI = dyn_cast<StoreInst>(*I)) && 1328 P.second == SI->getPointerOperand()) 1329 continue; 1330 1331 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I); 1332 1333 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1334 if ((SJ = dyn_cast<StoreInst>(*J)) && 1335 P.second == SJ->getPointerOperand()) 1336 continue; 1337 1338 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { 1339 VPPair VP(P, ValuePair(*I, *J)); 1340 ConnectedPairs.insert(VP); 1341 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1342 } 1343 } 1344 } 1345 } 1346 1347 // This function figures out which pairs are connected. Two pairs are 1348 // connected if some output of the first pair forms an input to both members 1349 // of the second pair. 1350 void BBVectorize::computeConnectedPairs( 1351 std::multimap<Value *, Value *> &CandidatePairs, 1352 std::vector<Value *> &PairableInsts, 1353 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1354 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1355 1356 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1357 PE = PairableInsts.end(); PI != PE; ++PI) { 1358 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1359 1360 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1361 P != choiceRange.second; ++P) 1362 computePairsConnectedTo(CandidatePairs, PairableInsts, 1363 ConnectedPairs, PairConnectionTypes, *P); 1364 } 1365 1366 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1367 << " pair connections.\n"); 1368 } 1369 1370 // This function builds a set of use tuples such that <A, B> is in the set 1371 // if B is in the use tree of A. If B is in the use tree of A, then B 1372 // depends on the output of A. 1373 void BBVectorize::buildDepMap( 1374 BasicBlock &BB, 1375 std::multimap<Value *, Value *> &CandidatePairs, 1376 std::vector<Value *> &PairableInsts, 1377 DenseSet<ValuePair> &PairableInstUsers) { 1378 DenseSet<Value *> IsInPair; 1379 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1380 E = CandidatePairs.end(); C != E; ++C) { 1381 IsInPair.insert(C->first); 1382 IsInPair.insert(C->second); 1383 } 1384 1385 // Iterate through the basic block, recording all users of each 1386 // pairable instruction. 1387 1388 BasicBlock::iterator E = BB.end(); 1389 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1390 if (IsInPair.find(I) == IsInPair.end()) continue; 1391 1392 DenseSet<Value *> Users; 1393 AliasSetTracker WriteSet(*AA); 1394 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1395 (void) trackUsesOfI(Users, WriteSet, I, J); 1396 1397 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1398 U != E; ++U) 1399 PairableInstUsers.insert(ValuePair(I, *U)); 1400 } 1401 } 1402 1403 // Returns true if an input to pair P is an output of pair Q and also an 1404 // input of pair Q is an output of pair P. If this is the case, then these 1405 // two pairs cannot be simultaneously fused. 1406 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1407 DenseSet<ValuePair> &PairableInstUsers, 1408 std::multimap<ValuePair, ValuePair> *PairableInstUserMap, 1409 DenseSet<VPPair> *PairableInstUserPairSet) { 1410 // Two pairs are in conflict if they are mutual Users of eachother. 1411 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1412 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1413 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1414 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1415 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1416 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1417 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1418 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1419 if (PairableInstUserMap) { 1420 // FIXME: The expensive part of the cycle check is not so much the cycle 1421 // check itself but this edge insertion procedure. This needs some 1422 // profiling and probably a different data structure (same is true of 1423 // most uses of std::multimap). 1424 if (PUsesQ) { 1425 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second) 1426 PairableInstUserMap->insert(VPPair(Q, P)); 1427 } 1428 if (QUsesP) { 1429 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second) 1430 PairableInstUserMap->insert(VPPair(P, Q)); 1431 } 1432 } 1433 1434 return (QUsesP && PUsesQ); 1435 } 1436 1437 // This function walks the use graph of current pairs to see if, starting 1438 // from P, the walk returns to P. 1439 bool BBVectorize::pairWillFormCycle(ValuePair P, 1440 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1441 DenseSet<ValuePair> &CurrentPairs) { 1442 DEBUG(if (DebugCycleCheck) 1443 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1444 << *P.second << "\n"); 1445 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1446 // contains non-direct associations. 1447 DenseSet<ValuePair> Visited; 1448 SmallVector<ValuePair, 32> Q; 1449 // General depth-first post-order traversal: 1450 Q.push_back(P); 1451 do { 1452 ValuePair QTop = Q.pop_back_val(); 1453 Visited.insert(QTop); 1454 1455 DEBUG(if (DebugCycleCheck) 1456 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1457 << *QTop.second << "\n"); 1458 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1459 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1460 C != QPairRange.second; ++C) { 1461 if (C->second == P) { 1462 DEBUG(dbgs() 1463 << "BBV: rejected to prevent non-trivial cycle formation: " 1464 << *C->first.first << " <-> " << *C->first.second << "\n"); 1465 return true; 1466 } 1467 1468 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1469 Q.push_back(C->second); 1470 } 1471 } while (!Q.empty()); 1472 1473 return false; 1474 } 1475 1476 // This function builds the initial tree of connected pairs with the 1477 // pair J at the root. 1478 void BBVectorize::buildInitialTreeFor( 1479 std::multimap<Value *, Value *> &CandidatePairs, 1480 std::vector<Value *> &PairableInsts, 1481 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1482 DenseSet<ValuePair> &PairableInstUsers, 1483 DenseMap<Value *, Value *> &ChosenPairs, 1484 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1485 // Each of these pairs is viewed as the root node of a Tree. The Tree 1486 // is then walked (depth-first). As this happens, we keep track of 1487 // the pairs that compose the Tree and the maximum depth of the Tree. 1488 SmallVector<ValuePairWithDepth, 32> Q; 1489 // General depth-first post-order traversal: 1490 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1491 do { 1492 ValuePairWithDepth QTop = Q.back(); 1493 1494 // Push each child onto the queue: 1495 bool MoreChildren = false; 1496 size_t MaxChildDepth = QTop.second; 1497 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1498 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1499 k != qtRange.second; ++k) { 1500 // Make sure that this child pair is still a candidate: 1501 bool IsStillCand = false; 1502 VPIteratorPair checkRange = 1503 CandidatePairs.equal_range(k->second.first); 1504 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1505 m != checkRange.second; ++m) { 1506 if (m->second == k->second.second) { 1507 IsStillCand = true; 1508 break; 1509 } 1510 } 1511 1512 if (IsStillCand) { 1513 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1514 if (C == Tree.end()) { 1515 size_t d = getDepthFactor(k->second.first); 1516 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1517 MoreChildren = true; 1518 } else { 1519 MaxChildDepth = std::max(MaxChildDepth, C->second); 1520 } 1521 } 1522 } 1523 1524 if (!MoreChildren) { 1525 // Record the current pair as part of the Tree: 1526 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1527 Q.pop_back(); 1528 } 1529 } while (!Q.empty()); 1530 } 1531 1532 // Given some initial tree, prune it by removing conflicting pairs (pairs 1533 // that cannot be simultaneously chosen for vectorization). 1534 void BBVectorize::pruneTreeFor( 1535 std::multimap<Value *, Value *> &CandidatePairs, 1536 std::vector<Value *> &PairableInsts, 1537 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1538 DenseSet<ValuePair> &PairableInstUsers, 1539 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1540 DenseSet<VPPair> &PairableInstUserPairSet, 1541 DenseMap<Value *, Value *> &ChosenPairs, 1542 DenseMap<ValuePair, size_t> &Tree, 1543 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1544 bool UseCycleCheck) { 1545 SmallVector<ValuePairWithDepth, 32> Q; 1546 // General depth-first post-order traversal: 1547 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1548 do { 1549 ValuePairWithDepth QTop = Q.pop_back_val(); 1550 PrunedTree.insert(QTop.first); 1551 1552 // Visit each child, pruning as necessary... 1553 SmallVector<ValuePairWithDepth, 8> BestChildren; 1554 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1555 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1556 K != QTopRange.second; ++K) { 1557 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1558 if (C == Tree.end()) continue; 1559 1560 // This child is in the Tree, now we need to make sure it is the 1561 // best of any conflicting children. There could be multiple 1562 // conflicting children, so first, determine if we're keeping 1563 // this child, then delete conflicting children as necessary. 1564 1565 // It is also necessary to guard against pairing-induced 1566 // dependencies. Consider instructions a .. x .. y .. b 1567 // such that (a,b) are to be fused and (x,y) are to be fused 1568 // but a is an input to x and b is an output from y. This 1569 // means that y cannot be moved after b but x must be moved 1570 // after b for (a,b) to be fused. In other words, after 1571 // fusing (a,b) we have y .. a/b .. x where y is an input 1572 // to a/b and x is an output to a/b: x and y can no longer 1573 // be legally fused. To prevent this condition, we must 1574 // make sure that a child pair added to the Tree is not 1575 // both an input and output of an already-selected pair. 1576 1577 // Pairing-induced dependencies can also form from more complicated 1578 // cycles. The pair vs. pair conflicts are easy to check, and so 1579 // that is done explicitly for "fast rejection", and because for 1580 // child vs. child conflicts, we may prefer to keep the current 1581 // pair in preference to the already-selected child. 1582 DenseSet<ValuePair> CurrentPairs; 1583 1584 bool CanAdd = true; 1585 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1586 = BestChildren.begin(), E2 = BestChildren.end(); 1587 C2 != E2; ++C2) { 1588 if (C2->first.first == C->first.first || 1589 C2->first.first == C->first.second || 1590 C2->first.second == C->first.first || 1591 C2->first.second == C->first.second || 1592 pairsConflict(C2->first, C->first, PairableInstUsers, 1593 UseCycleCheck ? &PairableInstUserMap : 0, 1594 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1595 if (C2->second >= C->second) { 1596 CanAdd = false; 1597 break; 1598 } 1599 1600 CurrentPairs.insert(C2->first); 1601 } 1602 } 1603 if (!CanAdd) continue; 1604 1605 // Even worse, this child could conflict with another node already 1606 // selected for the Tree. If that is the case, ignore this child. 1607 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1608 E2 = PrunedTree.end(); T != E2; ++T) { 1609 if (T->first == C->first.first || 1610 T->first == C->first.second || 1611 T->second == C->first.first || 1612 T->second == C->first.second || 1613 pairsConflict(*T, C->first, PairableInstUsers, 1614 UseCycleCheck ? &PairableInstUserMap : 0, 1615 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1616 CanAdd = false; 1617 break; 1618 } 1619 1620 CurrentPairs.insert(*T); 1621 } 1622 if (!CanAdd) continue; 1623 1624 // And check the queue too... 1625 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1626 E2 = Q.end(); C2 != E2; ++C2) { 1627 if (C2->first.first == C->first.first || 1628 C2->first.first == C->first.second || 1629 C2->first.second == C->first.first || 1630 C2->first.second == C->first.second || 1631 pairsConflict(C2->first, C->first, PairableInstUsers, 1632 UseCycleCheck ? &PairableInstUserMap : 0, 1633 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1634 CanAdd = false; 1635 break; 1636 } 1637 1638 CurrentPairs.insert(C2->first); 1639 } 1640 if (!CanAdd) continue; 1641 1642 // Last but not least, check for a conflict with any of the 1643 // already-chosen pairs. 1644 for (DenseMap<Value *, Value *>::iterator C2 = 1645 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1646 C2 != E2; ++C2) { 1647 if (pairsConflict(*C2, C->first, PairableInstUsers, 1648 UseCycleCheck ? &PairableInstUserMap : 0, 1649 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1650 CanAdd = false; 1651 break; 1652 } 1653 1654 CurrentPairs.insert(*C2); 1655 } 1656 if (!CanAdd) continue; 1657 1658 // To check for non-trivial cycles formed by the addition of the 1659 // current pair we've formed a list of all relevant pairs, now use a 1660 // graph walk to check for a cycle. We start from the current pair and 1661 // walk the use tree to see if we again reach the current pair. If we 1662 // do, then the current pair is rejected. 1663 1664 // FIXME: It may be more efficient to use a topological-ordering 1665 // algorithm to improve the cycle check. This should be investigated. 1666 if (UseCycleCheck && 1667 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1668 continue; 1669 1670 // This child can be added, but we may have chosen it in preference 1671 // to an already-selected child. Check for this here, and if a 1672 // conflict is found, then remove the previously-selected child 1673 // before adding this one in its place. 1674 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1675 = BestChildren.begin(); C2 != BestChildren.end();) { 1676 if (C2->first.first == C->first.first || 1677 C2->first.first == C->first.second || 1678 C2->first.second == C->first.first || 1679 C2->first.second == C->first.second || 1680 pairsConflict(C2->first, C->first, PairableInstUsers)) 1681 C2 = BestChildren.erase(C2); 1682 else 1683 ++C2; 1684 } 1685 1686 BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); 1687 } 1688 1689 for (SmallVector<ValuePairWithDepth, 8>::iterator C 1690 = BestChildren.begin(), E2 = BestChildren.end(); 1691 C != E2; ++C) { 1692 size_t DepthF = getDepthFactor(C->first.first); 1693 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1694 } 1695 } while (!Q.empty()); 1696 } 1697 1698 // This function finds the best tree of mututally-compatible connected 1699 // pairs, given the choice of root pairs as an iterator range. 1700 void BBVectorize::findBestTreeFor( 1701 std::multimap<Value *, Value *> &CandidatePairs, 1702 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1703 std::vector<Value *> &PairableInsts, 1704 DenseSet<ValuePair> &FixedOrderPairs, 1705 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1706 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1707 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1708 DenseSet<ValuePair> &PairableInstUsers, 1709 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1710 DenseSet<VPPair> &PairableInstUserPairSet, 1711 DenseMap<Value *, Value *> &ChosenPairs, 1712 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1713 int &BestEffSize, VPIteratorPair ChoiceRange, 1714 bool UseCycleCheck) { 1715 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1716 J != ChoiceRange.second; ++J) { 1717 1718 // Before going any further, make sure that this pair does not 1719 // conflict with any already-selected pairs (see comment below 1720 // near the Tree pruning for more details). 1721 DenseSet<ValuePair> ChosenPairSet; 1722 bool DoesConflict = false; 1723 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1724 E = ChosenPairs.end(); C != E; ++C) { 1725 if (pairsConflict(*C, *J, PairableInstUsers, 1726 UseCycleCheck ? &PairableInstUserMap : 0, 1727 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1728 DoesConflict = true; 1729 break; 1730 } 1731 1732 ChosenPairSet.insert(*C); 1733 } 1734 if (DoesConflict) continue; 1735 1736 if (UseCycleCheck && 1737 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1738 continue; 1739 1740 DenseMap<ValuePair, size_t> Tree; 1741 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1742 PairableInstUsers, ChosenPairs, Tree, *J); 1743 1744 // Because we'll keep the child with the largest depth, the largest 1745 // depth is still the same in the unpruned Tree. 1746 size_t MaxDepth = Tree.lookup(*J); 1747 1748 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1749 << *J->first << " <-> " << *J->second << "} of depth " << 1750 MaxDepth << " and size " << Tree.size() << "\n"); 1751 1752 // At this point the Tree has been constructed, but, may contain 1753 // contradictory children (meaning that different children of 1754 // some tree node may be attempting to fuse the same instruction). 1755 // So now we walk the tree again, in the case of a conflict, 1756 // keep only the child with the largest depth. To break a tie, 1757 // favor the first child. 1758 1759 DenseSet<ValuePair> PrunedTree; 1760 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1761 PairableInstUsers, PairableInstUserMap, PairableInstUserPairSet, 1762 ChosenPairs, Tree, PrunedTree, *J, UseCycleCheck); 1763 1764 int EffSize = 0; 1765 if (TTI) { 1766 DenseSet<Value *> PrunedTreeInstrs; 1767 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1768 E = PrunedTree.end(); S != E; ++S) { 1769 PrunedTreeInstrs.insert(S->first); 1770 PrunedTreeInstrs.insert(S->second); 1771 } 1772 1773 // The set of pairs that have already contributed to the total cost. 1774 DenseSet<ValuePair> IncomingPairs; 1775 1776 // If the cost model were perfect, this might not be necessary; but we 1777 // need to make sure that we don't get stuck vectorizing our own 1778 // shuffle chains. 1779 bool HasNontrivialInsts = false; 1780 1781 // The node weights represent the cost savings associated with 1782 // fusing the pair of instructions. 1783 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1784 E = PrunedTree.end(); S != E; ++S) { 1785 if (!isa<ShuffleVectorInst>(S->first) && 1786 !isa<InsertElementInst>(S->first) && 1787 !isa<ExtractElementInst>(S->first)) 1788 HasNontrivialInsts = true; 1789 1790 bool FlipOrder = false; 1791 1792 if (getDepthFactor(S->first)) { 1793 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1794 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1795 << *S->first << " <-> " << *S->second << "} = " << 1796 ESContrib << "\n"); 1797 EffSize += ESContrib; 1798 } 1799 1800 // The edge weights contribute in a negative sense: they represent 1801 // the cost of shuffles. 1802 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); 1803 if (IP.first != ConnectedPairDeps.end()) { 1804 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1805 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1806 Q != IP.second; ++Q) { 1807 if (!PrunedTree.count(Q->second)) 1808 continue; 1809 DenseMap<VPPair, unsigned>::iterator R = 1810 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1811 assert(R != PairConnectionTypes.end() && 1812 "Cannot find pair connection type"); 1813 if (R->second == PairConnectionDirect) 1814 ++NumDepsDirect; 1815 else if (R->second == PairConnectionSwap) 1816 ++NumDepsSwap; 1817 } 1818 1819 // If there are more swaps than direct connections, then 1820 // the pair order will be flipped during fusion. So the real 1821 // number of swaps is the minimum number. 1822 FlipOrder = !FixedOrderPairs.count(*S) && 1823 ((NumDepsSwap > NumDepsDirect) || 1824 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1825 1826 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1827 Q != IP.second; ++Q) { 1828 if (!PrunedTree.count(Q->second)) 1829 continue; 1830 DenseMap<VPPair, unsigned>::iterator R = 1831 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1832 assert(R != PairConnectionTypes.end() && 1833 "Cannot find pair connection type"); 1834 Type *Ty1 = Q->second.first->getType(), 1835 *Ty2 = Q->second.second->getType(); 1836 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1837 if ((R->second == PairConnectionDirect && FlipOrder) || 1838 (R->second == PairConnectionSwap && !FlipOrder) || 1839 R->second == PairConnectionSplat) { 1840 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1841 VTy, VTy); 1842 1843 if (VTy->getVectorNumElements() == 2) { 1844 if (R->second == PairConnectionSplat) 1845 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1846 TargetTransformInfo::SK_Broadcast, VTy)); 1847 else 1848 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1849 TargetTransformInfo::SK_Reverse, VTy)); 1850 } 1851 1852 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1853 *Q->second.first << " <-> " << *Q->second.second << 1854 "} -> {" << 1855 *S->first << " <-> " << *S->second << "} = " << 1856 ESContrib << "\n"); 1857 EffSize -= ESContrib; 1858 } 1859 } 1860 } 1861 1862 // Compute the cost of outgoing edges. We assume that edges outgoing 1863 // to shuffles, inserts or extracts can be merged, and so contribute 1864 // no additional cost. 1865 if (!S->first->getType()->isVoidTy()) { 1866 Type *Ty1 = S->first->getType(), 1867 *Ty2 = S->second->getType(); 1868 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1869 1870 bool NeedsExtraction = false; 1871 for (Value::use_iterator I = S->first->use_begin(), 1872 IE = S->first->use_end(); I != IE; ++I) { 1873 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1874 // Shuffle can be folded if it has no other input 1875 if (isa<UndefValue>(SI->getOperand(1))) 1876 continue; 1877 } 1878 if (isa<ExtractElementInst>(*I)) 1879 continue; 1880 if (PrunedTreeInstrs.count(*I)) 1881 continue; 1882 NeedsExtraction = true; 1883 break; 1884 } 1885 1886 if (NeedsExtraction) { 1887 int ESContrib; 1888 if (Ty1->isVectorTy()) { 1889 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1890 Ty1, VTy); 1891 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1892 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1)); 1893 } else 1894 ESContrib = (int) TTI->getVectorInstrCost( 1895 Instruction::ExtractElement, VTy, 0); 1896 1897 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1898 *S->first << "} = " << ESContrib << "\n"); 1899 EffSize -= ESContrib; 1900 } 1901 1902 NeedsExtraction = false; 1903 for (Value::use_iterator I = S->second->use_begin(), 1904 IE = S->second->use_end(); I != IE; ++I) { 1905 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1906 // Shuffle can be folded if it has no other input 1907 if (isa<UndefValue>(SI->getOperand(1))) 1908 continue; 1909 } 1910 if (isa<ExtractElementInst>(*I)) 1911 continue; 1912 if (PrunedTreeInstrs.count(*I)) 1913 continue; 1914 NeedsExtraction = true; 1915 break; 1916 } 1917 1918 if (NeedsExtraction) { 1919 int ESContrib; 1920 if (Ty2->isVectorTy()) { 1921 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1922 Ty2, VTy); 1923 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1924 TargetTransformInfo::SK_ExtractSubvector, VTy, 1925 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2)); 1926 } else 1927 ESContrib = (int) TTI->getVectorInstrCost( 1928 Instruction::ExtractElement, VTy, 1); 1929 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1930 *S->second << "} = " << ESContrib << "\n"); 1931 EffSize -= ESContrib; 1932 } 1933 } 1934 1935 // Compute the cost of incoming edges. 1936 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 1937 Instruction *S1 = cast<Instruction>(S->first), 1938 *S2 = cast<Instruction>(S->second); 1939 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 1940 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 1941 1942 // Combining constants into vector constants (or small vector 1943 // constants into larger ones are assumed free). 1944 if (isa<Constant>(O1) && isa<Constant>(O2)) 1945 continue; 1946 1947 if (FlipOrder) 1948 std::swap(O1, O2); 1949 1950 ValuePair VP = ValuePair(O1, O2); 1951 ValuePair VPR = ValuePair(O2, O1); 1952 1953 // Internal edges are not handled here. 1954 if (PrunedTree.count(VP) || PrunedTree.count(VPR)) 1955 continue; 1956 1957 Type *Ty1 = O1->getType(), 1958 *Ty2 = O2->getType(); 1959 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1960 1961 // Combining vector operations of the same type is also assumed 1962 // folded with other operations. 1963 if (Ty1 == Ty2) { 1964 // If both are insert elements, then both can be widened. 1965 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), 1966 *IEO2 = dyn_cast<InsertElementInst>(O2); 1967 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) 1968 continue; 1969 // If both are extract elements, and both have the same input 1970 // type, then they can be replaced with a shuffle 1971 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 1972 *EIO2 = dyn_cast<ExtractElementInst>(O2); 1973 if (EIO1 && EIO2 && 1974 EIO1->getOperand(0)->getType() == 1975 EIO2->getOperand(0)->getType()) 1976 continue; 1977 // If both are a shuffle with equal operand types and only two 1978 // unqiue operands, then they can be replaced with a single 1979 // shuffle 1980 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 1981 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 1982 if (SIO1 && SIO2 && 1983 SIO1->getOperand(0)->getType() == 1984 SIO2->getOperand(0)->getType()) { 1985 SmallSet<Value *, 4> SIOps; 1986 SIOps.insert(SIO1->getOperand(0)); 1987 SIOps.insert(SIO1->getOperand(1)); 1988 SIOps.insert(SIO2->getOperand(0)); 1989 SIOps.insert(SIO2->getOperand(1)); 1990 if (SIOps.size() <= 2) 1991 continue; 1992 } 1993 } 1994 1995 int ESContrib; 1996 // This pair has already been formed. 1997 if (IncomingPairs.count(VP)) { 1998 continue; 1999 } else if (IncomingPairs.count(VPR)) { 2000 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2001 VTy, VTy); 2002 2003 if (VTy->getVectorNumElements() == 2) 2004 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 2005 TargetTransformInfo::SK_Reverse, VTy)); 2006 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 2007 ESContrib = (int) TTI->getVectorInstrCost( 2008 Instruction::InsertElement, VTy, 0); 2009 ESContrib += (int) TTI->getVectorInstrCost( 2010 Instruction::InsertElement, VTy, 1); 2011 } else if (!Ty1->isVectorTy()) { 2012 // O1 needs to be inserted into a vector of size O2, and then 2013 // both need to be shuffled together. 2014 ESContrib = (int) TTI->getVectorInstrCost( 2015 Instruction::InsertElement, Ty2, 0); 2016 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2017 VTy, Ty2); 2018 } else if (!Ty2->isVectorTy()) { 2019 // O2 needs to be inserted into a vector of size O1, and then 2020 // both need to be shuffled together. 2021 ESContrib = (int) TTI->getVectorInstrCost( 2022 Instruction::InsertElement, Ty1, 0); 2023 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2024 VTy, Ty1); 2025 } else { 2026 Type *TyBig = Ty1, *TySmall = Ty2; 2027 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 2028 std::swap(TyBig, TySmall); 2029 2030 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2031 VTy, TyBig); 2032 if (TyBig != TySmall) 2033 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2034 TyBig, TySmall); 2035 } 2036 2037 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 2038 << *O1 << " <-> " << *O2 << "} = " << 2039 ESContrib << "\n"); 2040 EffSize -= ESContrib; 2041 IncomingPairs.insert(VP); 2042 } 2043 } 2044 } 2045 2046 if (!HasNontrivialInsts) { 2047 DEBUG(if (DebugPairSelection) dbgs() << 2048 "\tNo non-trivial instructions in tree;" 2049 " override to zero effective size\n"); 2050 EffSize = 0; 2051 } 2052 } else { 2053 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 2054 E = PrunedTree.end(); S != E; ++S) 2055 EffSize += (int) getDepthFactor(S->first); 2056 } 2057 2058 DEBUG(if (DebugPairSelection) 2059 dbgs() << "BBV: found pruned Tree for pair {" 2060 << *J->first << " <-> " << *J->second << "} of depth " << 2061 MaxDepth << " and size " << PrunedTree.size() << 2062 " (effective size: " << EffSize << ")\n"); 2063 if (((TTI && !UseChainDepthWithTI) || 2064 MaxDepth >= Config.ReqChainDepth) && 2065 EffSize > 0 && EffSize > BestEffSize) { 2066 BestMaxDepth = MaxDepth; 2067 BestEffSize = EffSize; 2068 BestTree = PrunedTree; 2069 } 2070 } 2071 } 2072 2073 // Given the list of candidate pairs, this function selects those 2074 // that will be fused into vector instructions. 2075 void BBVectorize::choosePairs( 2076 std::multimap<Value *, Value *> &CandidatePairs, 2077 DenseMap<ValuePair, int> &CandidatePairCostSavings, 2078 std::vector<Value *> &PairableInsts, 2079 DenseSet<ValuePair> &FixedOrderPairs, 2080 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2081 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2082 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 2083 DenseSet<ValuePair> &PairableInstUsers, 2084 DenseMap<Value *, Value *>& ChosenPairs) { 2085 bool UseCycleCheck = 2086 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 2087 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 2088 DenseSet<VPPair> PairableInstUserPairSet; 2089 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 2090 E = PairableInsts.end(); I != E; ++I) { 2091 // The number of possible pairings for this variable: 2092 size_t NumChoices = CandidatePairs.count(*I); 2093 if (!NumChoices) continue; 2094 2095 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 2096 2097 // The best pair to choose and its tree: 2098 size_t BestMaxDepth = 0; 2099 int BestEffSize = 0; 2100 DenseSet<ValuePair> BestTree; 2101 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 2102 PairableInsts, FixedOrderPairs, PairConnectionTypes, 2103 ConnectedPairs, ConnectedPairDeps, 2104 PairableInstUsers, PairableInstUserMap, 2105 PairableInstUserPairSet, ChosenPairs, 2106 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 2107 UseCycleCheck); 2108 2109 // A tree has been chosen (or not) at this point. If no tree was 2110 // chosen, then this instruction, I, cannot be paired (and is no longer 2111 // considered). 2112 2113 DEBUG(if (BestTree.size() > 0) 2114 dbgs() << "BBV: selected pairs in the best tree for: " 2115 << *cast<Instruction>(*I) << "\n"); 2116 2117 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 2118 SE2 = BestTree.end(); S != SE2; ++S) { 2119 // Insert the members of this tree into the list of chosen pairs. 2120 ChosenPairs.insert(ValuePair(S->first, S->second)); 2121 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 2122 *S->second << "\n"); 2123 2124 // Remove all candidate pairs that have values in the chosen tree. 2125 for (std::multimap<Value *, Value *>::iterator K = 2126 CandidatePairs.begin(); K != CandidatePairs.end();) { 2127 if (K->first == S->first || K->second == S->first || 2128 K->second == S->second || K->first == S->second) { 2129 // Don't remove the actual pair chosen so that it can be used 2130 // in subsequent tree selections. 2131 if (!(K->first == S->first && K->second == S->second)) 2132 CandidatePairs.erase(K++); 2133 else 2134 ++K; 2135 } else { 2136 ++K; 2137 } 2138 } 2139 } 2140 } 2141 2142 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2143 } 2144 2145 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2146 unsigned n = 0) { 2147 if (!I->hasName()) 2148 return ""; 2149 2150 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2151 (n > 0 ? "." + utostr(n) : "")).str(); 2152 } 2153 2154 // Returns the value that is to be used as the pointer input to the vector 2155 // instruction that fuses I with J. 2156 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2157 Instruction *I, Instruction *J, unsigned o) { 2158 Value *IPtr, *JPtr; 2159 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2160 int64_t OffsetInElmts; 2161 2162 // Note: the analysis might fail here, that is why the pair order has 2163 // been precomputed (OffsetInElmts must be unused here). 2164 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2165 IAddressSpace, JAddressSpace, 2166 OffsetInElmts, false); 2167 2168 // The pointer value is taken to be the one with the lowest offset. 2169 Value *VPtr = IPtr; 2170 2171 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 2172 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 2173 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2174 Type *VArgPtrType = PointerType::get(VArgType, 2175 cast<PointerType>(IPtr->getType())->getAddressSpace()); 2176 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2177 /* insert before */ I); 2178 } 2179 2180 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2181 unsigned MaskOffset, unsigned NumInElem, 2182 unsigned NumInElem1, unsigned IdxOffset, 2183 std::vector<Constant*> &Mask) { 2184 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 2185 for (unsigned v = 0; v < NumElem1; ++v) { 2186 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2187 if (m < 0) { 2188 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2189 } else { 2190 unsigned mm = m + (int) IdxOffset; 2191 if (m >= (int) NumInElem1) 2192 mm += (int) NumInElem; 2193 2194 Mask[v+MaskOffset] = 2195 ConstantInt::get(Type::getInt32Ty(Context), mm); 2196 } 2197 } 2198 } 2199 2200 // Returns the value that is to be used as the vector-shuffle mask to the 2201 // vector instruction that fuses I with J. 2202 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2203 Instruction *I, Instruction *J) { 2204 // This is the shuffle mask. We need to append the second 2205 // mask to the first, and the numbers need to be adjusted. 2206 2207 Type *ArgTypeI = I->getType(); 2208 Type *ArgTypeJ = J->getType(); 2209 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2210 2211 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 2212 2213 // Get the total number of elements in the fused vector type. 2214 // By definition, this must equal the number of elements in 2215 // the final mask. 2216 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 2217 std::vector<Constant*> Mask(NumElem); 2218 2219 Type *OpTypeI = I->getOperand(0)->getType(); 2220 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 2221 Type *OpTypeJ = J->getOperand(0)->getType(); 2222 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 2223 2224 // The fused vector will be: 2225 // ----------------------------------------------------- 2226 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2227 // ----------------------------------------------------- 2228 // from which we'll extract NumElem total elements (where the first NumElemI 2229 // of them come from the mask in I and the remainder come from the mask 2230 // in J. 2231 2232 // For the mask from the first pair... 2233 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2234 0, Mask); 2235 2236 // For the mask from the second pair... 2237 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2238 NumInElemI, Mask); 2239 2240 return ConstantVector::get(Mask); 2241 } 2242 2243 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2244 Instruction *J, unsigned o, Value *&LOp, 2245 unsigned numElemL, 2246 Type *ArgTypeL, Type *ArgTypeH, 2247 bool IBeforeJ, unsigned IdxOff) { 2248 bool ExpandedIEChain = false; 2249 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2250 // If we have a pure insertelement chain, then this can be rewritten 2251 // into a chain that directly builds the larger type. 2252 if (isPureIEChain(LIE)) { 2253 SmallVector<Value *, 8> VectElemts(numElemL, 2254 UndefValue::get(ArgTypeL->getScalarType())); 2255 InsertElementInst *LIENext = LIE; 2256 do { 2257 unsigned Idx = 2258 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2259 VectElemts[Idx] = LIENext->getOperand(1); 2260 } while ((LIENext = 2261 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2262 2263 LIENext = 0; 2264 Value *LIEPrev = UndefValue::get(ArgTypeH); 2265 for (unsigned i = 0; i < numElemL; ++i) { 2266 if (isa<UndefValue>(VectElemts[i])) continue; 2267 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2268 ConstantInt::get(Type::getInt32Ty(Context), 2269 i + IdxOff), 2270 getReplacementName(IBeforeJ ? I : J, 2271 true, o, i+1)); 2272 LIENext->insertBefore(IBeforeJ ? J : I); 2273 LIEPrev = LIENext; 2274 } 2275 2276 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2277 ExpandedIEChain = true; 2278 } 2279 } 2280 2281 return ExpandedIEChain; 2282 } 2283 2284 // Returns the value to be used as the specified operand of the vector 2285 // instruction that fuses I with J. 2286 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2287 Instruction *J, unsigned o, bool IBeforeJ) { 2288 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2289 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2290 2291 // Compute the fused vector type for this operand 2292 Type *ArgTypeI = I->getOperand(o)->getType(); 2293 Type *ArgTypeJ = J->getOperand(o)->getType(); 2294 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2295 2296 Instruction *L = I, *H = J; 2297 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2298 2299 unsigned numElemL; 2300 if (ArgTypeL->isVectorTy()) 2301 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 2302 else 2303 numElemL = 1; 2304 2305 unsigned numElemH; 2306 if (ArgTypeH->isVectorTy()) 2307 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 2308 else 2309 numElemH = 1; 2310 2311 Value *LOp = L->getOperand(o); 2312 Value *HOp = H->getOperand(o); 2313 unsigned numElem = VArgType->getNumElements(); 2314 2315 // First, we check if we can reuse the "original" vector outputs (if these 2316 // exist). We might need a shuffle. 2317 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2318 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2319 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2320 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2321 2322 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2323 // optimization. The input vectors to the shuffle might be a different 2324 // length from the shuffle outputs. Unfortunately, the replacement 2325 // shuffle mask has already been formed, and the mask entries are sensitive 2326 // to the sizes of the inputs. 2327 bool IsSizeChangeShuffle = 2328 isa<ShuffleVectorInst>(L) && 2329 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2330 2331 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2332 // We can have at most two unique vector inputs. 2333 bool CanUseInputs = true; 2334 Value *I1, *I2 = 0; 2335 if (LEE) { 2336 I1 = LEE->getOperand(0); 2337 } else { 2338 I1 = LSV->getOperand(0); 2339 I2 = LSV->getOperand(1); 2340 if (I2 == I1 || isa<UndefValue>(I2)) 2341 I2 = 0; 2342 } 2343 2344 if (HEE) { 2345 Value *I3 = HEE->getOperand(0); 2346 if (!I2 && I3 != I1) 2347 I2 = I3; 2348 else if (I3 != I1 && I3 != I2) 2349 CanUseInputs = false; 2350 } else { 2351 Value *I3 = HSV->getOperand(0); 2352 if (!I2 && I3 != I1) 2353 I2 = I3; 2354 else if (I3 != I1 && I3 != I2) 2355 CanUseInputs = false; 2356 2357 if (CanUseInputs) { 2358 Value *I4 = HSV->getOperand(1); 2359 if (!isa<UndefValue>(I4)) { 2360 if (!I2 && I4 != I1) 2361 I2 = I4; 2362 else if (I4 != I1 && I4 != I2) 2363 CanUseInputs = false; 2364 } 2365 } 2366 } 2367 2368 if (CanUseInputs) { 2369 unsigned LOpElem = 2370 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2371 ->getNumElements(); 2372 unsigned HOpElem = 2373 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2374 ->getNumElements(); 2375 2376 // We have one or two input vectors. We need to map each index of the 2377 // operands to the index of the original vector. 2378 SmallVector<std::pair<int, int>, 8> II(numElem); 2379 for (unsigned i = 0; i < numElemL; ++i) { 2380 int Idx, INum; 2381 if (LEE) { 2382 Idx = 2383 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2384 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2385 } else { 2386 Idx = LSV->getMaskValue(i); 2387 if (Idx < (int) LOpElem) { 2388 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2389 } else { 2390 Idx -= LOpElem; 2391 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2392 } 2393 } 2394 2395 II[i] = std::pair<int, int>(Idx, INum); 2396 } 2397 for (unsigned i = 0; i < numElemH; ++i) { 2398 int Idx, INum; 2399 if (HEE) { 2400 Idx = 2401 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2402 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2403 } else { 2404 Idx = HSV->getMaskValue(i); 2405 if (Idx < (int) HOpElem) { 2406 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2407 } else { 2408 Idx -= HOpElem; 2409 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2410 } 2411 } 2412 2413 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2414 } 2415 2416 // We now have an array which tells us from which index of which 2417 // input vector each element of the operand comes. 2418 VectorType *I1T = cast<VectorType>(I1->getType()); 2419 unsigned I1Elem = I1T->getNumElements(); 2420 2421 if (!I2) { 2422 // In this case there is only one underlying vector input. Check for 2423 // the trivial case where we can use the input directly. 2424 if (I1Elem == numElem) { 2425 bool ElemInOrder = true; 2426 for (unsigned i = 0; i < numElem; ++i) { 2427 if (II[i].first != (int) i && II[i].first != -1) { 2428 ElemInOrder = false; 2429 break; 2430 } 2431 } 2432 2433 if (ElemInOrder) 2434 return I1; 2435 } 2436 2437 // A shuffle is needed. 2438 std::vector<Constant *> Mask(numElem); 2439 for (unsigned i = 0; i < numElem; ++i) { 2440 int Idx = II[i].first; 2441 if (Idx == -1) 2442 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2443 else 2444 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2445 } 2446 2447 Instruction *S = 2448 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2449 ConstantVector::get(Mask), 2450 getReplacementName(IBeforeJ ? I : J, 2451 true, o)); 2452 S->insertBefore(IBeforeJ ? J : I); 2453 return S; 2454 } 2455 2456 VectorType *I2T = cast<VectorType>(I2->getType()); 2457 unsigned I2Elem = I2T->getNumElements(); 2458 2459 // This input comes from two distinct vectors. The first step is to 2460 // make sure that both vectors are the same length. If not, the 2461 // smaller one will need to grow before they can be shuffled together. 2462 if (I1Elem < I2Elem) { 2463 std::vector<Constant *> Mask(I2Elem); 2464 unsigned v = 0; 2465 for (; v < I1Elem; ++v) 2466 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2467 for (; v < I2Elem; ++v) 2468 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2469 2470 Instruction *NewI1 = 2471 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2472 ConstantVector::get(Mask), 2473 getReplacementName(IBeforeJ ? I : J, 2474 true, o, 1)); 2475 NewI1->insertBefore(IBeforeJ ? J : I); 2476 I1 = NewI1; 2477 I1T = I2T; 2478 I1Elem = I2Elem; 2479 } else if (I1Elem > I2Elem) { 2480 std::vector<Constant *> Mask(I1Elem); 2481 unsigned v = 0; 2482 for (; v < I2Elem; ++v) 2483 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2484 for (; v < I1Elem; ++v) 2485 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2486 2487 Instruction *NewI2 = 2488 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2489 ConstantVector::get(Mask), 2490 getReplacementName(IBeforeJ ? I : J, 2491 true, o, 1)); 2492 NewI2->insertBefore(IBeforeJ ? J : I); 2493 I2 = NewI2; 2494 I2T = I1T; 2495 I2Elem = I1Elem; 2496 } 2497 2498 // Now that both I1 and I2 are the same length we can shuffle them 2499 // together (and use the result). 2500 std::vector<Constant *> Mask(numElem); 2501 for (unsigned v = 0; v < numElem; ++v) { 2502 if (II[v].first == -1) { 2503 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2504 } else { 2505 int Idx = II[v].first + II[v].second * I1Elem; 2506 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2507 } 2508 } 2509 2510 Instruction *NewOp = 2511 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2512 getReplacementName(IBeforeJ ? I : J, true, o)); 2513 NewOp->insertBefore(IBeforeJ ? J : I); 2514 return NewOp; 2515 } 2516 } 2517 2518 Type *ArgType = ArgTypeL; 2519 if (numElemL < numElemH) { 2520 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2521 ArgTypeL, VArgType, IBeforeJ, 1)) { 2522 // This is another short-circuit case: we're combining a scalar into 2523 // a vector that is formed by an IE chain. We've just expanded the IE 2524 // chain, now insert the scalar and we're done. 2525 2526 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2527 getReplacementName(IBeforeJ ? I : J, true, o)); 2528 S->insertBefore(IBeforeJ ? J : I); 2529 return S; 2530 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2531 ArgTypeH, IBeforeJ)) { 2532 // The two vector inputs to the shuffle must be the same length, 2533 // so extend the smaller vector to be the same length as the larger one. 2534 Instruction *NLOp; 2535 if (numElemL > 1) { 2536 2537 std::vector<Constant *> Mask(numElemH); 2538 unsigned v = 0; 2539 for (; v < numElemL; ++v) 2540 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2541 for (; v < numElemH; ++v) 2542 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2543 2544 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2545 ConstantVector::get(Mask), 2546 getReplacementName(IBeforeJ ? I : J, 2547 true, o, 1)); 2548 } else { 2549 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2550 getReplacementName(IBeforeJ ? I : J, 2551 true, o, 1)); 2552 } 2553 2554 NLOp->insertBefore(IBeforeJ ? J : I); 2555 LOp = NLOp; 2556 } 2557 2558 ArgType = ArgTypeH; 2559 } else if (numElemL > numElemH) { 2560 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2561 ArgTypeH, VArgType, IBeforeJ)) { 2562 Instruction *S = 2563 InsertElementInst::Create(LOp, HOp, 2564 ConstantInt::get(Type::getInt32Ty(Context), 2565 numElemL), 2566 getReplacementName(IBeforeJ ? I : J, 2567 true, o)); 2568 S->insertBefore(IBeforeJ ? J : I); 2569 return S; 2570 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2571 ArgTypeL, IBeforeJ)) { 2572 Instruction *NHOp; 2573 if (numElemH > 1) { 2574 std::vector<Constant *> Mask(numElemL); 2575 unsigned v = 0; 2576 for (; v < numElemH; ++v) 2577 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2578 for (; v < numElemL; ++v) 2579 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2580 2581 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2582 ConstantVector::get(Mask), 2583 getReplacementName(IBeforeJ ? I : J, 2584 true, o, 1)); 2585 } else { 2586 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2587 getReplacementName(IBeforeJ ? I : J, 2588 true, o, 1)); 2589 } 2590 2591 NHOp->insertBefore(IBeforeJ ? J : I); 2592 HOp = NHOp; 2593 } 2594 } 2595 2596 if (ArgType->isVectorTy()) { 2597 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2598 std::vector<Constant*> Mask(numElem); 2599 for (unsigned v = 0; v < numElem; ++v) { 2600 unsigned Idx = v; 2601 // If the low vector was expanded, we need to skip the extra 2602 // undefined entries. 2603 if (v >= numElemL && numElemH > numElemL) 2604 Idx += (numElemH - numElemL); 2605 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2606 } 2607 2608 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2609 ConstantVector::get(Mask), 2610 getReplacementName(IBeforeJ ? I : J, true, o)); 2611 BV->insertBefore(IBeforeJ ? J : I); 2612 return BV; 2613 } 2614 2615 Instruction *BV1 = InsertElementInst::Create( 2616 UndefValue::get(VArgType), LOp, CV0, 2617 getReplacementName(IBeforeJ ? I : J, 2618 true, o, 1)); 2619 BV1->insertBefore(IBeforeJ ? J : I); 2620 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2621 getReplacementName(IBeforeJ ? I : J, 2622 true, o, 2)); 2623 BV2->insertBefore(IBeforeJ ? J : I); 2624 return BV2; 2625 } 2626 2627 // This function creates an array of values that will be used as the inputs 2628 // to the vector instruction that fuses I with J. 2629 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2630 Instruction *I, Instruction *J, 2631 SmallVector<Value *, 3> &ReplacedOperands, 2632 bool IBeforeJ) { 2633 unsigned NumOperands = I->getNumOperands(); 2634 2635 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2636 // Iterate backward so that we look at the store pointer 2637 // first and know whether or not we need to flip the inputs. 2638 2639 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2640 // This is the pointer for a load/store instruction. 2641 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2642 continue; 2643 } else if (isa<CallInst>(I)) { 2644 Function *F = cast<CallInst>(I)->getCalledFunction(); 2645 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 2646 if (o == NumOperands-1) { 2647 BasicBlock &BB = *I->getParent(); 2648 2649 Module *M = BB.getParent()->getParent(); 2650 Type *ArgTypeI = I->getType(); 2651 Type *ArgTypeJ = J->getType(); 2652 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2653 2654 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType); 2655 continue; 2656 } else if (IID == Intrinsic::powi && o == 1) { 2657 // The second argument of powi is a single integer and we've already 2658 // checked that both arguments are equal. As a result, we just keep 2659 // I's second argument. 2660 ReplacedOperands[o] = I->getOperand(o); 2661 continue; 2662 } 2663 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2664 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2665 continue; 2666 } 2667 2668 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2669 } 2670 } 2671 2672 // This function creates two values that represent the outputs of the 2673 // original I and J instructions. These are generally vector shuffles 2674 // or extracts. In many cases, these will end up being unused and, thus, 2675 // eliminated by later passes. 2676 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2677 Instruction *J, Instruction *K, 2678 Instruction *&InsertionPt, 2679 Instruction *&K1, Instruction *&K2) { 2680 if (isa<StoreInst>(I)) { 2681 AA->replaceWithNewValue(I, K); 2682 AA->replaceWithNewValue(J, K); 2683 } else { 2684 Type *IType = I->getType(); 2685 Type *JType = J->getType(); 2686 2687 VectorType *VType = getVecTypeForPair(IType, JType); 2688 unsigned numElem = VType->getNumElements(); 2689 2690 unsigned numElemI, numElemJ; 2691 if (IType->isVectorTy()) 2692 numElemI = cast<VectorType>(IType)->getNumElements(); 2693 else 2694 numElemI = 1; 2695 2696 if (JType->isVectorTy()) 2697 numElemJ = cast<VectorType>(JType)->getNumElements(); 2698 else 2699 numElemJ = 1; 2700 2701 if (IType->isVectorTy()) { 2702 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2703 for (unsigned v = 0; v < numElemI; ++v) { 2704 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2705 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2706 } 2707 2708 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2709 ConstantVector::get( Mask1), 2710 getReplacementName(K, false, 1)); 2711 } else { 2712 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2713 K1 = ExtractElementInst::Create(K, CV0, 2714 getReplacementName(K, false, 1)); 2715 } 2716 2717 if (JType->isVectorTy()) { 2718 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2719 for (unsigned v = 0; v < numElemJ; ++v) { 2720 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2721 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2722 } 2723 2724 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2725 ConstantVector::get( Mask2), 2726 getReplacementName(K, false, 2)); 2727 } else { 2728 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2729 K2 = ExtractElementInst::Create(K, CV1, 2730 getReplacementName(K, false, 2)); 2731 } 2732 2733 K1->insertAfter(K); 2734 K2->insertAfter(K1); 2735 InsertionPt = K2; 2736 } 2737 } 2738 2739 // Move all uses of the function I (including pairing-induced uses) after J. 2740 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2741 DenseSet<ValuePair> &LoadMoveSetPairs, 2742 Instruction *I, Instruction *J) { 2743 // Skip to the first instruction past I. 2744 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2745 2746 DenseSet<Value *> Users; 2747 AliasSetTracker WriteSet(*AA); 2748 for (; cast<Instruction>(L) != J; ++L) 2749 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs); 2750 2751 assert(cast<Instruction>(L) == J && 2752 "Tracking has not proceeded far enough to check for dependencies"); 2753 // If J is now in the use set of I, then trackUsesOfI will return true 2754 // and we have a dependency cycle (and the fusing operation must abort). 2755 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs); 2756 } 2757 2758 // Move all uses of the function I (including pairing-induced uses) after J. 2759 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2760 DenseSet<ValuePair> &LoadMoveSetPairs, 2761 Instruction *&InsertionPt, 2762 Instruction *I, Instruction *J) { 2763 // Skip to the first instruction past I. 2764 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2765 2766 DenseSet<Value *> Users; 2767 AliasSetTracker WriteSet(*AA); 2768 for (; cast<Instruction>(L) != J;) { 2769 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) { 2770 // Move this instruction 2771 Instruction *InstToMove = L; ++L; 2772 2773 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2774 " to after " << *InsertionPt << "\n"); 2775 InstToMove->removeFromParent(); 2776 InstToMove->insertAfter(InsertionPt); 2777 InsertionPt = InstToMove; 2778 } else { 2779 ++L; 2780 } 2781 } 2782 } 2783 2784 // Collect all load instruction that are in the move set of a given first 2785 // pair member. These loads depend on the first instruction, I, and so need 2786 // to be moved after J (the second instruction) when the pair is fused. 2787 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2788 DenseMap<Value *, Value *> &ChosenPairs, 2789 std::multimap<Value *, Value *> &LoadMoveSet, 2790 DenseSet<ValuePair> &LoadMoveSetPairs, 2791 Instruction *I) { 2792 // Skip to the first instruction past I. 2793 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2794 2795 DenseSet<Value *> Users; 2796 AliasSetTracker WriteSet(*AA); 2797 2798 // Note: We cannot end the loop when we reach J because J could be moved 2799 // farther down the use chain by another instruction pairing. Also, J 2800 // could be before I if this is an inverted input. 2801 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2802 if (trackUsesOfI(Users, WriteSet, I, L)) { 2803 if (L->mayReadFromMemory()) { 2804 LoadMoveSet.insert(ValuePair(L, I)); 2805 LoadMoveSetPairs.insert(ValuePair(L, I)); 2806 } 2807 } 2808 } 2809 } 2810 2811 // In cases where both load/stores and the computation of their pointers 2812 // are chosen for vectorization, we can end up in a situation where the 2813 // aliasing analysis starts returning different query results as the 2814 // process of fusing instruction pairs continues. Because the algorithm 2815 // relies on finding the same use trees here as were found earlier, we'll 2816 // need to precompute the necessary aliasing information here and then 2817 // manually update it during the fusion process. 2818 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2819 std::vector<Value *> &PairableInsts, 2820 DenseMap<Value *, Value *> &ChosenPairs, 2821 std::multimap<Value *, Value *> &LoadMoveSet, 2822 DenseSet<ValuePair> &LoadMoveSetPairs) { 2823 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2824 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2825 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2826 if (P == ChosenPairs.end()) continue; 2827 2828 Instruction *I = cast<Instruction>(P->first); 2829 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, 2830 LoadMoveSetPairs, I); 2831 } 2832 } 2833 2834 // When the first instruction in each pair is cloned, it will inherit its 2835 // parent's metadata. This metadata must be combined with that of the other 2836 // instruction in a safe way. 2837 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2838 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2839 K->getAllMetadataOtherThanDebugLoc(Metadata); 2840 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2841 unsigned Kind = Metadata[i].first; 2842 MDNode *JMD = J->getMetadata(Kind); 2843 MDNode *KMD = Metadata[i].second; 2844 2845 switch (Kind) { 2846 default: 2847 K->setMetadata(Kind, 0); // Remove unknown metadata 2848 break; 2849 case LLVMContext::MD_tbaa: 2850 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2851 break; 2852 case LLVMContext::MD_fpmath: 2853 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2854 break; 2855 } 2856 } 2857 } 2858 2859 // This function fuses the chosen instruction pairs into vector instructions, 2860 // taking care preserve any needed scalar outputs and, then, it reorders the 2861 // remaining instructions as needed (users of the first member of the pair 2862 // need to be moved to after the location of the second member of the pair 2863 // because the vector instruction is inserted in the location of the pair's 2864 // second member). 2865 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2866 std::vector<Value *> &PairableInsts, 2867 DenseMap<Value *, Value *> &ChosenPairs, 2868 DenseSet<ValuePair> &FixedOrderPairs, 2869 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2870 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2871 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2872 LLVMContext& Context = BB.getContext(); 2873 2874 // During the vectorization process, the order of the pairs to be fused 2875 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2876 // list. After a pair is fused, the flipped pair is removed from the list. 2877 DenseSet<ValuePair> FlippedPairs; 2878 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2879 E = ChosenPairs.end(); P != E; ++P) 2880 FlippedPairs.insert(ValuePair(P->second, P->first)); 2881 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2882 E = FlippedPairs.end(); P != E; ++P) 2883 ChosenPairs.insert(*P); 2884 2885 std::multimap<Value *, Value *> LoadMoveSet; 2886 DenseSet<ValuePair> LoadMoveSetPairs; 2887 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, 2888 LoadMoveSet, LoadMoveSetPairs); 2889 2890 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2891 2892 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2893 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2894 if (P == ChosenPairs.end()) { 2895 ++PI; 2896 continue; 2897 } 2898 2899 if (getDepthFactor(P->first) == 0) { 2900 // These instructions are not really fused, but are tracked as though 2901 // they are. Any case in which it would be interesting to fuse them 2902 // will be taken care of by InstCombine. 2903 --NumFusedOps; 2904 ++PI; 2905 continue; 2906 } 2907 2908 Instruction *I = cast<Instruction>(P->first), 2909 *J = cast<Instruction>(P->second); 2910 2911 DEBUG(dbgs() << "BBV: fusing: " << *I << 2912 " <-> " << *J << "\n"); 2913 2914 // Remove the pair and flipped pair from the list. 2915 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2916 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2917 ChosenPairs.erase(FP); 2918 ChosenPairs.erase(P); 2919 2920 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) { 2921 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2922 " <-> " << *J << 2923 " aborted because of non-trivial dependency cycle\n"); 2924 --NumFusedOps; 2925 ++PI; 2926 continue; 2927 } 2928 2929 // If the pair must have the other order, then flip it. 2930 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2931 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2932 // This pair does not have a fixed order, and so we might want to 2933 // flip it if that will yield fewer shuffles. We count the number 2934 // of dependencies connected via swaps, and those directly connected, 2935 // and flip the order if the number of swaps is greater. 2936 bool OrigOrder = true; 2937 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2938 if (IP.first == ConnectedPairDeps.end()) { 2939 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2940 OrigOrder = false; 2941 } 2942 2943 if (IP.first != ConnectedPairDeps.end()) { 2944 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2945 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2946 Q != IP.second; ++Q) { 2947 DenseMap<VPPair, unsigned>::iterator R = 2948 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2949 assert(R != PairConnectionTypes.end() && 2950 "Cannot find pair connection type"); 2951 if (R->second == PairConnectionDirect) 2952 ++NumDepsDirect; 2953 else if (R->second == PairConnectionSwap) 2954 ++NumDepsSwap; 2955 } 2956 2957 if (!OrigOrder) 2958 std::swap(NumDepsDirect, NumDepsSwap); 2959 2960 if (NumDepsSwap > NumDepsDirect) { 2961 FlipPairOrder = true; 2962 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2963 " <-> " << *J << "\n"); 2964 } 2965 } 2966 } 2967 2968 Instruction *L = I, *H = J; 2969 if (FlipPairOrder) 2970 std::swap(H, L); 2971 2972 // If the pair being fused uses the opposite order from that in the pair 2973 // connection map, then we need to flip the types. 2974 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2975 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2976 Q != IP.second; ++Q) { 2977 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2978 assert(R != PairConnectionTypes.end() && 2979 "Cannot find pair connection type"); 2980 if (R->second == PairConnectionDirect) 2981 R->second = PairConnectionSwap; 2982 else if (R->second == PairConnectionSwap) 2983 R->second = PairConnectionDirect; 2984 } 2985 2986 bool LBeforeH = !FlipPairOrder; 2987 unsigned NumOperands = I->getNumOperands(); 2988 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2989 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2990 LBeforeH); 2991 2992 // Make a copy of the original operation, change its type to the vector 2993 // type and replace its operands with the vector operands. 2994 Instruction *K = L->clone(); 2995 if (L->hasName()) 2996 K->takeName(L); 2997 else if (H->hasName()) 2998 K->takeName(H); 2999 3000 if (!isa<StoreInst>(K)) 3001 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 3002 3003 combineMetadata(K, H); 3004 K->intersectOptionalDataWith(H); 3005 3006 for (unsigned o = 0; o < NumOperands; ++o) 3007 K->setOperand(o, ReplacedOperands[o]); 3008 3009 K->insertAfter(J); 3010 3011 // Instruction insertion point: 3012 Instruction *InsertionPt = K; 3013 Instruction *K1 = 0, *K2 = 0; 3014 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 3015 3016 // The use tree of the first original instruction must be moved to after 3017 // the location of the second instruction. The entire use tree of the 3018 // first instruction is disjoint from the input tree of the second 3019 // (by definition), and so commutes with it. 3020 3021 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J); 3022 3023 if (!isa<StoreInst>(I)) { 3024 L->replaceAllUsesWith(K1); 3025 H->replaceAllUsesWith(K2); 3026 AA->replaceWithNewValue(L, K1); 3027 AA->replaceWithNewValue(H, K2); 3028 } 3029 3030 // Instructions that may read from memory may be in the load move set. 3031 // Once an instruction is fused, we no longer need its move set, and so 3032 // the values of the map never need to be updated. However, when a load 3033 // is fused, we need to merge the entries from both instructions in the 3034 // pair in case those instructions were in the move set of some other 3035 // yet-to-be-fused pair. The loads in question are the keys of the map. 3036 if (I->mayReadFromMemory()) { 3037 std::vector<ValuePair> NewSetMembers; 3038 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 3039 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 3040 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 3041 N != IPairRange.second; ++N) 3042 NewSetMembers.push_back(ValuePair(K, N->second)); 3043 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 3044 N != JPairRange.second; ++N) 3045 NewSetMembers.push_back(ValuePair(K, N->second)); 3046 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 3047 AE = NewSetMembers.end(); A != AE; ++A) { 3048 LoadMoveSet.insert(*A); 3049 LoadMoveSetPairs.insert(*A); 3050 } 3051 } 3052 3053 // Before removing I, set the iterator to the next instruction. 3054 PI = llvm::next(BasicBlock::iterator(I)); 3055 if (cast<Instruction>(PI) == J) 3056 ++PI; 3057 3058 SE->forgetValue(I); 3059 SE->forgetValue(J); 3060 I->eraseFromParent(); 3061 J->eraseFromParent(); 3062 3063 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 3064 BB << "\n"); 3065 } 3066 3067 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 3068 } 3069} 3070 3071char BBVectorize::ID = 0; 3072static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 3073INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3074INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3075INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3076INITIALIZE_PASS_DEPENDENCY(DominatorTree) 3077INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3078INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3079 3080BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 3081 return new BBVectorize(C); 3082} 3083 3084bool 3085llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 3086 BBVectorize BBVectorizer(P, C); 3087 return BBVectorizer.vectorizeBB(BB); 3088} 3089 3090//===----------------------------------------------------------------------===// 3091VectorizeConfig::VectorizeConfig() { 3092 VectorBits = ::VectorBits; 3093 VectorizeBools = !::NoBools; 3094 VectorizeInts = !::NoInts; 3095 VectorizeFloats = !::NoFloats; 3096 VectorizePointers = !::NoPointers; 3097 VectorizeCasts = !::NoCasts; 3098 VectorizeMath = !::NoMath; 3099 VectorizeFMA = !::NoFMA; 3100 VectorizeSelect = !::NoSelect; 3101 VectorizeCmp = !::NoCmp; 3102 VectorizeGEP = !::NoGEP; 3103 VectorizeMemOps = !::NoMemOps; 3104 AlignedOnly = ::AlignedOnly; 3105 ReqChainDepth= ::ReqChainDepth; 3106 SearchLimit = ::SearchLimit; 3107 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 3108 SplatBreaksChain = ::SplatBreaksChain; 3109 MaxInsts = ::MaxInsts; 3110 MaxIter = ::MaxIter; 3111 Pow2LenOnly = ::Pow2LenOnly; 3112 NoMemOpBoost = ::NoMemOpBoost; 3113 FastDep = ::FastDep; 3114} 3115