BBVectorize.cpp revision 00f63b1b84d059a1ffa572e76708e03750a9e523
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 DenseSet<ValuePair> &CandidatePairsSet, 285 std::vector<Value *> &PairableInsts, 286 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 287 DenseMap<VPPair, unsigned> &PairConnectionTypes, 288 ValuePair P); 289 290 bool pairsConflict(ValuePair P, ValuePair Q, 291 DenseSet<ValuePair> &PairableInstUsers, 292 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0, 293 DenseSet<VPPair> *PairableInstUserPairSet = 0); 294 295 bool pairWillFormCycle(ValuePair P, 296 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 297 DenseSet<ValuePair> &CurrentPairs); 298 299 void pruneTreeFor( 300 std::multimap<Value *, Value *> &CandidatePairs, 301 std::vector<Value *> &PairableInsts, 302 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 303 DenseSet<ValuePair> &PairableInstUsers, 304 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 305 DenseSet<VPPair> &PairableInstUserPairSet, 306 DenseMap<Value *, Value *> &ChosenPairs, 307 DenseMap<ValuePair, size_t> &Tree, 308 DenseSet<ValuePair> &PrunedTree, ValuePair J, 309 bool UseCycleCheck); 310 311 void buildInitialTreeFor( 312 std::multimap<Value *, Value *> &CandidatePairs, 313 std::vector<Value *> &PairableInsts, 314 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 315 DenseSet<ValuePair> &PairableInstUsers, 316 DenseMap<Value *, Value *> &ChosenPairs, 317 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 318 319 void findBestTreeFor( 320 std::multimap<Value *, Value *> &CandidatePairs, 321 DenseMap<ValuePair, int> &CandidatePairCostSavings, 322 std::vector<Value *> &PairableInsts, 323 DenseSet<ValuePair> &FixedOrderPairs, 324 DenseMap<VPPair, unsigned> &PairConnectionTypes, 325 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 326 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 327 DenseSet<ValuePair> &PairableInstUsers, 328 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 329 DenseSet<VPPair> &PairableInstUserPairSet, 330 DenseMap<Value *, Value *> &ChosenPairs, 331 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 332 int &BestEffSize, VPIteratorPair ChoiceRange, 333 bool UseCycleCheck); 334 335 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 336 Instruction *J, unsigned o); 337 338 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 339 unsigned MaskOffset, unsigned NumInElem, 340 unsigned NumInElem1, unsigned IdxOffset, 341 std::vector<Constant*> &Mask); 342 343 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 344 Instruction *J); 345 346 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 347 unsigned o, Value *&LOp, unsigned numElemL, 348 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, 349 unsigned IdxOff = 0); 350 351 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 352 Instruction *J, unsigned o, bool IBeforeJ); 353 354 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 355 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 356 bool IBeforeJ); 357 358 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 359 Instruction *J, Instruction *K, 360 Instruction *&InsertionPt, Instruction *&K1, 361 Instruction *&K2); 362 363 void collectPairLoadMoveSet(BasicBlock &BB, 364 DenseMap<Value *, Value *> &ChosenPairs, 365 std::multimap<Value *, Value *> &LoadMoveSet, 366 DenseSet<ValuePair> &LoadMoveSetPairs, 367 Instruction *I); 368 369 void collectLoadMoveSet(BasicBlock &BB, 370 std::vector<Value *> &PairableInsts, 371 DenseMap<Value *, Value *> &ChosenPairs, 372 std::multimap<Value *, Value *> &LoadMoveSet, 373 DenseSet<ValuePair> &LoadMoveSetPairs); 374 375 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 376 DenseSet<ValuePair> &LoadMoveSetPairs, 377 Instruction *I, Instruction *J); 378 379 void moveUsesOfIAfterJ(BasicBlock &BB, 380 DenseSet<ValuePair> &LoadMoveSetPairs, 381 Instruction *&InsertionPt, 382 Instruction *I, Instruction *J); 383 384 void combineMetadata(Instruction *K, const Instruction *J); 385 386 bool vectorizeBB(BasicBlock &BB) { 387 if (!DT->isReachableFromEntry(&BB)) { 388 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 389 " in " << BB.getParent()->getName() << "\n"); 390 return false; 391 } 392 393 DEBUG(if (TTI) dbgs() << "BBV: using target information\n"); 394 395 bool changed = false; 396 // Iterate a sufficient number of times to merge types of size 1 bit, 397 // then 2 bits, then 4, etc. up to half of the target vector width of the 398 // target vector register. 399 unsigned n = 1; 400 for (unsigned v = 2; 401 (TTI || v <= Config.VectorBits) && 402 (!Config.MaxIter || n <= Config.MaxIter); 403 v *= 2, ++n) { 404 DEBUG(dbgs() << "BBV: fusing loop #" << n << 405 " for " << BB.getName() << " in " << 406 BB.getParent()->getName() << "...\n"); 407 if (vectorizePairs(BB)) 408 changed = true; 409 else 410 break; 411 } 412 413 if (changed && !Pow2LenOnly) { 414 ++n; 415 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 416 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 417 n << " for " << BB.getName() << " in " << 418 BB.getParent()->getName() << "...\n"); 419 if (!vectorizePairs(BB, true)) break; 420 } 421 } 422 423 DEBUG(dbgs() << "BBV: done!\n"); 424 return changed; 425 } 426 427 virtual bool runOnBasicBlock(BasicBlock &BB) { 428 AA = &getAnalysis<AliasAnalysis>(); 429 DT = &getAnalysis<DominatorTree>(); 430 SE = &getAnalysis<ScalarEvolution>(); 431 TD = getAnalysisIfAvailable<DataLayout>(); 432 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>(); 433 434 return vectorizeBB(BB); 435 } 436 437 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 438 BasicBlockPass::getAnalysisUsage(AU); 439 AU.addRequired<AliasAnalysis>(); 440 AU.addRequired<DominatorTree>(); 441 AU.addRequired<ScalarEvolution>(); 442 AU.addRequired<TargetTransformInfo>(); 443 AU.addPreserved<AliasAnalysis>(); 444 AU.addPreserved<DominatorTree>(); 445 AU.addPreserved<ScalarEvolution>(); 446 AU.setPreservesCFG(); 447 } 448 449 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 450 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 451 "Cannot form vector from incompatible scalar types"); 452 Type *STy = ElemTy->getScalarType(); 453 454 unsigned numElem; 455 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 456 numElem = VTy->getNumElements(); 457 } else { 458 numElem = 1; 459 } 460 461 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 462 numElem += VTy->getNumElements(); 463 } else { 464 numElem += 1; 465 } 466 467 return VectorType::get(STy, numElem); 468 } 469 470 static inline void getInstructionTypes(Instruction *I, 471 Type *&T1, Type *&T2) { 472 if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 473 // For stores, it is the value type, not the pointer type that matters 474 // because the value is what will come from a vector register. 475 476 Value *IVal = SI->getValueOperand(); 477 T1 = IVal->getType(); 478 } else { 479 T1 = I->getType(); 480 } 481 482 if (CastInst *CI = dyn_cast<CastInst>(I)) 483 T2 = CI->getSrcTy(); 484 else 485 T2 = T1; 486 487 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 488 T2 = SI->getCondition()->getType(); 489 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { 490 T2 = SI->getOperand(0)->getType(); 491 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 492 T2 = CI->getOperand(0)->getType(); 493 } 494 } 495 496 // Returns the weight associated with the provided value. A chain of 497 // candidate pairs has a length given by the sum of the weights of its 498 // members (one weight per pair; the weight of each member of the pair 499 // is assumed to be the same). This length is then compared to the 500 // chain-length threshold to determine if a given chain is significant 501 // enough to be vectorized. The length is also used in comparing 502 // candidate chains where longer chains are considered to be better. 503 // Note: when this function returns 0, the resulting instructions are 504 // not actually fused. 505 inline size_t getDepthFactor(Value *V) { 506 // InsertElement and ExtractElement have a depth factor of zero. This is 507 // for two reasons: First, they cannot be usefully fused. Second, because 508 // the pass generates a lot of these, they can confuse the simple metric 509 // used to compare the trees in the next iteration. Thus, giving them a 510 // weight of zero allows the pass to essentially ignore them in 511 // subsequent iterations when looking for vectorization opportunities 512 // while still tracking dependency chains that flow through those 513 // instructions. 514 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 515 return 0; 516 517 // Give a load or store half of the required depth so that load/store 518 // pairs will vectorize. 519 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 520 return Config.ReqChainDepth/2; 521 522 return 1; 523 } 524 525 // Returns the cost of the provided instruction using TTI. 526 // This does not handle loads and stores. 527 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { 528 switch (Opcode) { 529 default: break; 530 case Instruction::GetElementPtr: 531 // We mark this instruction as zero-cost because scalar GEPs are usually 532 // lowered to the intruction addressing mode. At the moment we don't 533 // generate vector GEPs. 534 return 0; 535 case Instruction::Br: 536 return TTI->getCFInstrCost(Opcode); 537 case Instruction::PHI: 538 return 0; 539 case Instruction::Add: 540 case Instruction::FAdd: 541 case Instruction::Sub: 542 case Instruction::FSub: 543 case Instruction::Mul: 544 case Instruction::FMul: 545 case Instruction::UDiv: 546 case Instruction::SDiv: 547 case Instruction::FDiv: 548 case Instruction::URem: 549 case Instruction::SRem: 550 case Instruction::FRem: 551 case Instruction::Shl: 552 case Instruction::LShr: 553 case Instruction::AShr: 554 case Instruction::And: 555 case Instruction::Or: 556 case Instruction::Xor: 557 return TTI->getArithmeticInstrCost(Opcode, T1); 558 case Instruction::Select: 559 case Instruction::ICmp: 560 case Instruction::FCmp: 561 return TTI->getCmpSelInstrCost(Opcode, T1, T2); 562 case Instruction::ZExt: 563 case Instruction::SExt: 564 case Instruction::FPToUI: 565 case Instruction::FPToSI: 566 case Instruction::FPExt: 567 case Instruction::PtrToInt: 568 case Instruction::IntToPtr: 569 case Instruction::SIToFP: 570 case Instruction::UIToFP: 571 case Instruction::Trunc: 572 case Instruction::FPTrunc: 573 case Instruction::BitCast: 574 case Instruction::ShuffleVector: 575 return TTI->getCastInstrCost(Opcode, T1, T2); 576 } 577 578 return 1; 579 } 580 581 // This determines the relative offset of two loads or stores, returning 582 // true if the offset could be determined to be some constant value. 583 // For example, if OffsetInElmts == 1, then J accesses the memory directly 584 // after I; if OffsetInElmts == -1 then I accesses the memory 585 // directly after J. 586 bool getPairPtrInfo(Instruction *I, Instruction *J, 587 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 588 unsigned &IAddressSpace, unsigned &JAddressSpace, 589 int64_t &OffsetInElmts, bool ComputeOffset = true) { 590 OffsetInElmts = 0; 591 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 592 LoadInst *LJ = cast<LoadInst>(J); 593 IPtr = LI->getPointerOperand(); 594 JPtr = LJ->getPointerOperand(); 595 IAlignment = LI->getAlignment(); 596 JAlignment = LJ->getAlignment(); 597 IAddressSpace = LI->getPointerAddressSpace(); 598 JAddressSpace = LJ->getPointerAddressSpace(); 599 } else { 600 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 601 IPtr = SI->getPointerOperand(); 602 JPtr = SJ->getPointerOperand(); 603 IAlignment = SI->getAlignment(); 604 JAlignment = SJ->getAlignment(); 605 IAddressSpace = SI->getPointerAddressSpace(); 606 JAddressSpace = SJ->getPointerAddressSpace(); 607 } 608 609 if (!ComputeOffset) 610 return true; 611 612 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 613 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 614 615 // If this is a trivial offset, then we'll get something like 616 // 1*sizeof(type). With target data, which we need anyway, this will get 617 // constant folded into a number. 618 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 619 if (const SCEVConstant *ConstOffSCEV = 620 dyn_cast<SCEVConstant>(OffsetSCEV)) { 621 ConstantInt *IntOff = ConstOffSCEV->getValue(); 622 int64_t Offset = IntOff->getSExtValue(); 623 624 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 625 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 626 627 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 628 if (VTy != VTy2 && Offset < 0) { 629 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 630 OffsetInElmts = Offset/VTy2TSS; 631 return (abs64(Offset) % VTy2TSS) == 0; 632 } 633 634 OffsetInElmts = Offset/VTyTSS; 635 return (abs64(Offset) % VTyTSS) == 0; 636 } 637 638 return false; 639 } 640 641 // Returns true if the provided CallInst represents an intrinsic that can 642 // be vectorized. 643 bool isVectorizableIntrinsic(CallInst* I) { 644 Function *F = I->getCalledFunction(); 645 if (!F) return false; 646 647 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 648 if (!IID) return false; 649 650 switch(IID) { 651 default: 652 return false; 653 case Intrinsic::sqrt: 654 case Intrinsic::powi: 655 case Intrinsic::sin: 656 case Intrinsic::cos: 657 case Intrinsic::log: 658 case Intrinsic::log2: 659 case Intrinsic::log10: 660 case Intrinsic::exp: 661 case Intrinsic::exp2: 662 case Intrinsic::pow: 663 return Config.VectorizeMath; 664 case Intrinsic::fma: 665 case Intrinsic::fmuladd: 666 return Config.VectorizeFMA; 667 } 668 } 669 670 bool isPureIEChain(InsertElementInst *IE) { 671 InsertElementInst *IENext = IE; 672 do { 673 if (!isa<UndefValue>(IENext->getOperand(0)) && 674 !isa<InsertElementInst>(IENext->getOperand(0))) { 675 return false; 676 } 677 } while ((IENext = 678 dyn_cast<InsertElementInst>(IENext->getOperand(0)))); 679 680 return true; 681 } 682 }; 683 684 // This function implements one vectorization iteration on the provided 685 // basic block. It returns true if the block is changed. 686 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 687 bool ShouldContinue; 688 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 689 690 std::vector<Value *> AllPairableInsts; 691 DenseMap<Value *, Value *> AllChosenPairs; 692 DenseSet<ValuePair> AllFixedOrderPairs; 693 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 694 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; 695 696 do { 697 std::vector<Value *> PairableInsts; 698 std::multimap<Value *, Value *> CandidatePairs; 699 DenseSet<ValuePair> FixedOrderPairs; 700 DenseMap<ValuePair, int> CandidatePairCostSavings; 701 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 702 FixedOrderPairs, 703 CandidatePairCostSavings, 704 PairableInsts, NonPow2Len); 705 if (PairableInsts.empty()) continue; 706 707 // Now we have a map of all of the pairable instructions and we need to 708 // select the best possible pairing. A good pairing is one such that the 709 // users of the pair are also paired. This defines a (directed) forest 710 // over the pairs such that two pairs are connected iff the second pair 711 // uses the first. 712 713 // Note that it only matters that both members of the second pair use some 714 // element of the first pair (to allow for splatting). 715 716 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; 717 DenseMap<VPPair, unsigned> PairConnectionTypes; 718 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs, 719 PairConnectionTypes); 720 if (ConnectedPairs.empty()) continue; 721 722 for (std::multimap<ValuePair, ValuePair>::iterator 723 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 724 I != IE; ++I) { 725 ConnectedPairDeps.insert(VPPair(I->second, I->first)); 726 } 727 728 // Build the pairable-instruction dependency map 729 DenseSet<ValuePair> PairableInstUsers; 730 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 731 732 // There is now a graph of the connected pairs. For each variable, pick 733 // the pairing with the largest tree meeting the depth requirement on at 734 // least one branch. Then select all pairings that are part of that tree 735 // and remove them from the list of available pairings and pairable 736 // variables. 737 738 DenseMap<Value *, Value *> ChosenPairs; 739 choosePairs(CandidatePairs, CandidatePairCostSavings, 740 PairableInsts, FixedOrderPairs, PairConnectionTypes, 741 ConnectedPairs, ConnectedPairDeps, 742 PairableInstUsers, ChosenPairs); 743 744 if (ChosenPairs.empty()) continue; 745 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 746 PairableInsts.end()); 747 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 748 749 // Only for the chosen pairs, propagate information on fixed-order pairs, 750 // pair connections, and their types to the data structures used by the 751 // pair fusion procedures. 752 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 753 IE = ChosenPairs.end(); I != IE; ++I) { 754 if (FixedOrderPairs.count(*I)) 755 AllFixedOrderPairs.insert(*I); 756 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 757 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 758 759 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 760 J != IE; ++J) { 761 DenseMap<VPPair, unsigned>::iterator K = 762 PairConnectionTypes.find(VPPair(*I, *J)); 763 if (K != PairConnectionTypes.end()) { 764 AllPairConnectionTypes.insert(*K); 765 } else { 766 K = PairConnectionTypes.find(VPPair(*J, *I)); 767 if (K != PairConnectionTypes.end()) 768 AllPairConnectionTypes.insert(*K); 769 } 770 } 771 } 772 773 for (std::multimap<ValuePair, ValuePair>::iterator 774 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 775 I != IE; ++I) { 776 if (AllPairConnectionTypes.count(*I)) { 777 AllConnectedPairs.insert(*I); 778 AllConnectedPairDeps.insert(VPPair(I->second, I->first)); 779 } 780 } 781 } while (ShouldContinue); 782 783 if (AllChosenPairs.empty()) return false; 784 NumFusedOps += AllChosenPairs.size(); 785 786 // A set of pairs has now been selected. It is now necessary to replace the 787 // paired instructions with vector instructions. For this procedure each 788 // operand must be replaced with a vector operand. This vector is formed 789 // by using build_vector on the old operands. The replaced values are then 790 // replaced with a vector_extract on the result. Subsequent optimization 791 // passes should coalesce the build/extract combinations. 792 793 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 794 AllPairConnectionTypes, 795 AllConnectedPairs, AllConnectedPairDeps); 796 797 // It is important to cleanup here so that future iterations of this 798 // function have less work to do. 799 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 800 return true; 801 } 802 803 // This function returns true if the provided instruction is capable of being 804 // fused into a vector instruction. This determination is based only on the 805 // type and other attributes of the instruction. 806 bool BBVectorize::isInstVectorizable(Instruction *I, 807 bool &IsSimpleLoadStore) { 808 IsSimpleLoadStore = false; 809 810 if (CallInst *C = dyn_cast<CallInst>(I)) { 811 if (!isVectorizableIntrinsic(C)) 812 return false; 813 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 814 // Vectorize simple loads if possbile: 815 IsSimpleLoadStore = L->isSimple(); 816 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 817 return false; 818 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 819 // Vectorize simple stores if possbile: 820 IsSimpleLoadStore = S->isSimple(); 821 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 822 return false; 823 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 824 // We can vectorize casts, but not casts of pointer types, etc. 825 if (!Config.VectorizeCasts) 826 return false; 827 828 Type *SrcTy = C->getSrcTy(); 829 if (!SrcTy->isSingleValueType()) 830 return false; 831 832 Type *DestTy = C->getDestTy(); 833 if (!DestTy->isSingleValueType()) 834 return false; 835 } else if (isa<SelectInst>(I)) { 836 if (!Config.VectorizeSelect) 837 return false; 838 } else if (isa<CmpInst>(I)) { 839 if (!Config.VectorizeCmp) 840 return false; 841 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 842 if (!Config.VectorizeGEP) 843 return false; 844 845 // Currently, vector GEPs exist only with one index. 846 if (G->getNumIndices() != 1) 847 return false; 848 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 849 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 850 return false; 851 } 852 853 // We can't vectorize memory operations without target data 854 if (TD == 0 && IsSimpleLoadStore) 855 return false; 856 857 Type *T1, *T2; 858 getInstructionTypes(I, T1, T2); 859 860 // Not every type can be vectorized... 861 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 862 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 863 return false; 864 865 if (T1->getScalarSizeInBits() == 1) { 866 if (!Config.VectorizeBools) 867 return false; 868 } else { 869 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 870 return false; 871 } 872 873 if (T2->getScalarSizeInBits() == 1) { 874 if (!Config.VectorizeBools) 875 return false; 876 } else { 877 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 878 return false; 879 } 880 881 if (!Config.VectorizeFloats 882 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 883 return false; 884 885 // Don't vectorize target-specific types. 886 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 887 return false; 888 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 889 return false; 890 891 if ((!Config.VectorizePointers || TD == 0) && 892 (T1->getScalarType()->isPointerTy() || 893 T2->getScalarType()->isPointerTy())) 894 return false; 895 896 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 897 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 898 return false; 899 900 return true; 901 } 902 903 // This function returns true if the two provided instructions are compatible 904 // (meaning that they can be fused into a vector instruction). This assumes 905 // that I has already been determined to be vectorizable and that J is not 906 // in the use tree of I. 907 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 908 bool IsSimpleLoadStore, bool NonPow2Len, 909 int &CostSavings, int &FixedOrder) { 910 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 911 " <-> " << *J << "\n"); 912 913 CostSavings = 0; 914 FixedOrder = 0; 915 916 // Loads and stores can be merged if they have different alignments, 917 // but are otherwise the same. 918 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 919 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 920 return false; 921 922 Type *IT1, *IT2, *JT1, *JT2; 923 getInstructionTypes(I, IT1, IT2); 924 getInstructionTypes(J, JT1, JT2); 925 unsigned MaxTypeBits = std::max( 926 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 927 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 928 if (!TTI && MaxTypeBits > Config.VectorBits) 929 return false; 930 931 // FIXME: handle addsub-type operations! 932 933 if (IsSimpleLoadStore) { 934 Value *IPtr, *JPtr; 935 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 936 int64_t OffsetInElmts = 0; 937 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 938 IAddressSpace, JAddressSpace, 939 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 940 FixedOrder = (int) OffsetInElmts; 941 unsigned BottomAlignment = IAlignment; 942 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 943 944 Type *aTypeI = isa<StoreInst>(I) ? 945 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 946 Type *aTypeJ = isa<StoreInst>(J) ? 947 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 948 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 949 950 if (Config.AlignedOnly) { 951 // An aligned load or store is possible only if the instruction 952 // with the lower offset has an alignment suitable for the 953 // vector type. 954 955 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 956 if (BottomAlignment < VecAlignment) 957 return false; 958 } 959 960 if (TTI) { 961 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI, 962 IAlignment, IAddressSpace); 963 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ, 964 JAlignment, JAddressSpace); 965 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType, 966 BottomAlignment, 967 IAddressSpace); 968 969 ICost += TTI->getAddressComputationCost(aTypeI); 970 JCost += TTI->getAddressComputationCost(aTypeJ); 971 VCost += TTI->getAddressComputationCost(VType); 972 973 if (VCost > ICost + JCost) 974 return false; 975 976 // We don't want to fuse to a type that will be split, even 977 // if the two input types will also be split and there is no other 978 // associated cost. 979 unsigned VParts = TTI->getNumberOfParts(VType); 980 if (VParts > 1) 981 return false; 982 else if (!VParts && VCost == ICost + JCost) 983 return false; 984 985 CostSavings = ICost + JCost - VCost; 986 } 987 } else { 988 return false; 989 } 990 } else if (TTI) { 991 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 992 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 993 Type *VT1 = getVecTypeForPair(IT1, JT1), 994 *VT2 = getVecTypeForPair(IT2, JT2); 995 996 // Note that this procedure is incorrect for insert and extract element 997 // instructions (because combining these often results in a shuffle), 998 // but this cost is ignored (because insert and extract element 999 // instructions are assigned a zero depth factor and are not really 1000 // fused in general). 1001 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 1002 1003 if (VCost > ICost + JCost) 1004 return false; 1005 1006 // We don't want to fuse to a type that will be split, even 1007 // if the two input types will also be split and there is no other 1008 // associated cost. 1009 unsigned VParts1 = TTI->getNumberOfParts(VT1), 1010 VParts2 = TTI->getNumberOfParts(VT2); 1011 if (VParts1 > 1 || VParts2 > 1) 1012 return false; 1013 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) 1014 return false; 1015 1016 CostSavings = ICost + JCost - VCost; 1017 } 1018 1019 // The powi intrinsic is special because only the first argument is 1020 // vectorized, the second arguments must be equal. 1021 CallInst *CI = dyn_cast<CallInst>(I); 1022 Function *FI; 1023 if (CI && (FI = CI->getCalledFunction())) { 1024 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID(); 1025 if (IID == Intrinsic::powi) { 1026 Value *A1I = CI->getArgOperand(1), 1027 *A1J = cast<CallInst>(J)->getArgOperand(1); 1028 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1029 *A1JSCEV = SE->getSCEV(A1J); 1030 return (A1ISCEV == A1JSCEV); 1031 } 1032 1033 if (IID && TTI) { 1034 SmallVector<Type*, 4> Tys; 1035 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 1036 Tys.push_back(CI->getArgOperand(i)->getType()); 1037 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys); 1038 1039 Tys.clear(); 1040 CallInst *CJ = cast<CallInst>(J); 1041 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i) 1042 Tys.push_back(CJ->getArgOperand(i)->getType()); 1043 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys); 1044 1045 Tys.clear(); 1046 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() && 1047 "Intrinsic argument counts differ"); 1048 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1049 if (IID == Intrinsic::powi && i == 1) 1050 Tys.push_back(CI->getArgOperand(i)->getType()); 1051 else 1052 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(), 1053 CJ->getArgOperand(i)->getType())); 1054 } 1055 1056 Type *RetTy = getVecTypeForPair(IT1, JT1); 1057 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys); 1058 1059 if (VCost > ICost + JCost) 1060 return false; 1061 1062 // We don't want to fuse to a type that will be split, even 1063 // if the two input types will also be split and there is no other 1064 // associated cost. 1065 unsigned RetParts = TTI->getNumberOfParts(RetTy); 1066 if (RetParts > 1) 1067 return false; 1068 else if (!RetParts && VCost == ICost + JCost) 1069 return false; 1070 1071 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1072 if (!Tys[i]->isVectorTy()) 1073 continue; 1074 1075 unsigned NumParts = TTI->getNumberOfParts(Tys[i]); 1076 if (NumParts > 1) 1077 return false; 1078 else if (!NumParts && VCost == ICost + JCost) 1079 return false; 1080 } 1081 1082 CostSavings = ICost + JCost - VCost; 1083 } 1084 } 1085 1086 return true; 1087 } 1088 1089 // Figure out whether or not J uses I and update the users and write-set 1090 // structures associated with I. Specifically, Users represents the set of 1091 // instructions that depend on I. WriteSet represents the set 1092 // of memory locations that are dependent on I. If UpdateUsers is true, 1093 // and J uses I, then Users is updated to contain J and WriteSet is updated 1094 // to contain any memory locations to which J writes. The function returns 1095 // true if J uses I. By default, alias analysis is used to determine 1096 // whether J reads from memory that overlaps with a location in WriteSet. 1097 // If LoadMoveSet is not null, then it is a previously-computed multimap 1098 // where the key is the memory-based user instruction and the value is 1099 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1100 // then the alias analysis is not used. This is necessary because this 1101 // function is called during the process of moving instructions during 1102 // vectorization and the results of the alias analysis are not stable during 1103 // that process. 1104 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1105 AliasSetTracker &WriteSet, Instruction *I, 1106 Instruction *J, bool UpdateUsers, 1107 DenseSet<ValuePair> *LoadMoveSetPairs) { 1108 bool UsesI = false; 1109 1110 // This instruction may already be marked as a user due, for example, to 1111 // being a member of a selected pair. 1112 if (Users.count(J)) 1113 UsesI = true; 1114 1115 if (!UsesI) 1116 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1117 JU != JE; ++JU) { 1118 Value *V = *JU; 1119 if (I == V || Users.count(V)) { 1120 UsesI = true; 1121 break; 1122 } 1123 } 1124 if (!UsesI && J->mayReadFromMemory()) { 1125 if (LoadMoveSetPairs) { 1126 UsesI = LoadMoveSetPairs->count(ValuePair(J, I)); 1127 } else { 1128 for (AliasSetTracker::iterator W = WriteSet.begin(), 1129 WE = WriteSet.end(); W != WE; ++W) { 1130 if (W->aliasesUnknownInst(J, *AA)) { 1131 UsesI = true; 1132 break; 1133 } 1134 } 1135 } 1136 } 1137 1138 if (UsesI && UpdateUsers) { 1139 if (J->mayWriteToMemory()) WriteSet.add(J); 1140 Users.insert(J); 1141 } 1142 1143 return UsesI; 1144 } 1145 1146 // This function iterates over all instruction pairs in the provided 1147 // basic block and collects all candidate pairs for vectorization. 1148 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1149 BasicBlock::iterator &Start, 1150 std::multimap<Value *, Value *> &CandidatePairs, 1151 DenseSet<ValuePair> &FixedOrderPairs, 1152 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1153 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1154 BasicBlock::iterator E = BB.end(); 1155 if (Start == E) return false; 1156 1157 bool ShouldContinue = false, IAfterStart = false; 1158 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1159 if (I == Start) IAfterStart = true; 1160 1161 bool IsSimpleLoadStore; 1162 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1163 1164 // Look for an instruction with which to pair instruction *I... 1165 DenseSet<Value *> Users; 1166 AliasSetTracker WriteSet(*AA); 1167 bool JAfterStart = IAfterStart; 1168 BasicBlock::iterator J = llvm::next(I); 1169 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1170 if (J == Start) JAfterStart = true; 1171 1172 // Determine if J uses I, if so, exit the loop. 1173 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1174 if (Config.FastDep) { 1175 // Note: For this heuristic to be effective, independent operations 1176 // must tend to be intermixed. This is likely to be true from some 1177 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1178 // but otherwise may require some kind of reordering pass. 1179 1180 // When using fast dependency analysis, 1181 // stop searching after first use: 1182 if (UsesI) break; 1183 } else { 1184 if (UsesI) continue; 1185 } 1186 1187 // J does not use I, and comes before the first use of I, so it can be 1188 // merged with I if the instructions are compatible. 1189 int CostSavings, FixedOrder; 1190 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1191 CostSavings, FixedOrder)) continue; 1192 1193 // J is a candidate for merging with I. 1194 if (!PairableInsts.size() || 1195 PairableInsts[PairableInsts.size()-1] != I) { 1196 PairableInsts.push_back(I); 1197 } 1198 1199 CandidatePairs.insert(ValuePair(I, J)); 1200 if (TTI) 1201 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1202 CostSavings)); 1203 1204 if (FixedOrder == 1) 1205 FixedOrderPairs.insert(ValuePair(I, J)); 1206 else if (FixedOrder == -1) 1207 FixedOrderPairs.insert(ValuePair(J, I)); 1208 1209 // The next call to this function must start after the last instruction 1210 // selected during this invocation. 1211 if (JAfterStart) { 1212 Start = llvm::next(J); 1213 IAfterStart = JAfterStart = false; 1214 } 1215 1216 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1217 << *I << " <-> " << *J << " (cost savings: " << 1218 CostSavings << ")\n"); 1219 1220 // If we have already found too many pairs, break here and this function 1221 // will be called again starting after the last instruction selected 1222 // during this invocation. 1223 if (PairableInsts.size() >= Config.MaxInsts) { 1224 ShouldContinue = true; 1225 break; 1226 } 1227 } 1228 1229 if (ShouldContinue) 1230 break; 1231 } 1232 1233 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1234 << " instructions with candidate pairs\n"); 1235 1236 return ShouldContinue; 1237 } 1238 1239 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1240 // it looks for pairs such that both members have an input which is an 1241 // output of PI or PJ. 1242 void BBVectorize::computePairsConnectedTo( 1243 std::multimap<Value *, Value *> &CandidatePairs, 1244 DenseSet<ValuePair> &CandidatePairsSet, 1245 std::vector<Value *> &PairableInsts, 1246 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1247 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1248 ValuePair P) { 1249 StoreInst *SI, *SJ; 1250 1251 // For each possible pairing for this variable, look at the uses of 1252 // the first value... 1253 for (Value::use_iterator I = P.first->use_begin(), 1254 E = P.first->use_end(); I != E; ++I) { 1255 if (isa<LoadInst>(*I)) { 1256 // A pair cannot be connected to a load because the load only takes one 1257 // operand (the address) and it is a scalar even after vectorization. 1258 continue; 1259 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1260 P.first == SI->getPointerOperand()) { 1261 // Similarly, a pair cannot be connected to a store through its 1262 // pointer operand. 1263 continue; 1264 } 1265 1266 // For each use of the first variable, look for uses of the second 1267 // variable... 1268 for (Value::use_iterator J = P.second->use_begin(), 1269 E2 = P.second->use_end(); J != E2; ++J) { 1270 if ((SJ = dyn_cast<StoreInst>(*J)) && 1271 P.second == SJ->getPointerOperand()) 1272 continue; 1273 1274 // Look for <I, J>: 1275 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1276 VPPair VP(P, ValuePair(*I, *J)); 1277 ConnectedPairs.insert(VP); 1278 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1279 } 1280 1281 // Look for <J, I>: 1282 if (CandidatePairsSet.count(ValuePair(*J, *I))) { 1283 VPPair VP(P, ValuePair(*J, *I)); 1284 ConnectedPairs.insert(VP); 1285 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1286 } 1287 } 1288 1289 if (Config.SplatBreaksChain) continue; 1290 // Look for cases where just the first value in the pair is used by 1291 // both members of another pair (splatting). 1292 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1293 if ((SJ = dyn_cast<StoreInst>(*J)) && 1294 P.first == SJ->getPointerOperand()) 1295 continue; 1296 1297 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1298 VPPair VP(P, ValuePair(*I, *J)); 1299 ConnectedPairs.insert(VP); 1300 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1301 } 1302 } 1303 } 1304 1305 if (Config.SplatBreaksChain) return; 1306 // Look for cases where just the second value in the pair is used by 1307 // both members of another pair (splatting). 1308 for (Value::use_iterator I = P.second->use_begin(), 1309 E = P.second->use_end(); I != E; ++I) { 1310 if (isa<LoadInst>(*I)) 1311 continue; 1312 else if ((SI = dyn_cast<StoreInst>(*I)) && 1313 P.second == SI->getPointerOperand()) 1314 continue; 1315 1316 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1317 if ((SJ = dyn_cast<StoreInst>(*J)) && 1318 P.second == SJ->getPointerOperand()) 1319 continue; 1320 1321 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1322 VPPair VP(P, ValuePair(*I, *J)); 1323 ConnectedPairs.insert(VP); 1324 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1325 } 1326 } 1327 } 1328 } 1329 1330 // This function figures out which pairs are connected. Two pairs are 1331 // connected if some output of the first pair forms an input to both members 1332 // of the second pair. 1333 void BBVectorize::computeConnectedPairs( 1334 std::multimap<Value *, Value *> &CandidatePairs, 1335 std::vector<Value *> &PairableInsts, 1336 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1337 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1338 DenseSet<ValuePair> CandidatePairsSet; 1339 for (std::multimap<Value *, Value *>::iterator I = CandidatePairs.begin(), 1340 E = CandidatePairs.end(); I != E; ++I) 1341 CandidatePairsSet.insert(*I); 1342 1343 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1344 PE = PairableInsts.end(); PI != PE; ++PI) { 1345 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1346 1347 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1348 P != choiceRange.second; ++P) 1349 computePairsConnectedTo(CandidatePairs, CandidatePairsSet, PairableInsts, 1350 ConnectedPairs, PairConnectionTypes, *P); 1351 } 1352 1353 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1354 << " pair connections.\n"); 1355 } 1356 1357 // This function builds a set of use tuples such that <A, B> is in the set 1358 // if B is in the use tree of A. If B is in the use tree of A, then B 1359 // depends on the output of A. 1360 void BBVectorize::buildDepMap( 1361 BasicBlock &BB, 1362 std::multimap<Value *, Value *> &CandidatePairs, 1363 std::vector<Value *> &PairableInsts, 1364 DenseSet<ValuePair> &PairableInstUsers) { 1365 DenseSet<Value *> IsInPair; 1366 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1367 E = CandidatePairs.end(); C != E; ++C) { 1368 IsInPair.insert(C->first); 1369 IsInPair.insert(C->second); 1370 } 1371 1372 // Iterate through the basic block, recording all users of each 1373 // pairable instruction. 1374 1375 BasicBlock::iterator E = BB.end(); 1376 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1377 if (IsInPair.find(I) == IsInPair.end()) continue; 1378 1379 DenseSet<Value *> Users; 1380 AliasSetTracker WriteSet(*AA); 1381 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1382 (void) trackUsesOfI(Users, WriteSet, I, J); 1383 1384 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1385 U != E; ++U) 1386 PairableInstUsers.insert(ValuePair(I, *U)); 1387 } 1388 } 1389 1390 // Returns true if an input to pair P is an output of pair Q and also an 1391 // input of pair Q is an output of pair P. If this is the case, then these 1392 // two pairs cannot be simultaneously fused. 1393 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1394 DenseSet<ValuePair> &PairableInstUsers, 1395 std::multimap<ValuePair, ValuePair> *PairableInstUserMap, 1396 DenseSet<VPPair> *PairableInstUserPairSet) { 1397 // Two pairs are in conflict if they are mutual Users of eachother. 1398 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1399 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1400 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1401 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1402 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1403 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1404 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1405 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1406 if (PairableInstUserMap) { 1407 // FIXME: The expensive part of the cycle check is not so much the cycle 1408 // check itself but this edge insertion procedure. This needs some 1409 // profiling and probably a different data structure (same is true of 1410 // most uses of std::multimap). 1411 if (PUsesQ) { 1412 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second) 1413 PairableInstUserMap->insert(VPPair(Q, P)); 1414 } 1415 if (QUsesP) { 1416 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second) 1417 PairableInstUserMap->insert(VPPair(P, Q)); 1418 } 1419 } 1420 1421 return (QUsesP && PUsesQ); 1422 } 1423 1424 // This function walks the use graph of current pairs to see if, starting 1425 // from P, the walk returns to P. 1426 bool BBVectorize::pairWillFormCycle(ValuePair P, 1427 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1428 DenseSet<ValuePair> &CurrentPairs) { 1429 DEBUG(if (DebugCycleCheck) 1430 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1431 << *P.second << "\n"); 1432 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1433 // contains non-direct associations. 1434 DenseSet<ValuePair> Visited; 1435 SmallVector<ValuePair, 32> Q; 1436 // General depth-first post-order traversal: 1437 Q.push_back(P); 1438 do { 1439 ValuePair QTop = Q.pop_back_val(); 1440 Visited.insert(QTop); 1441 1442 DEBUG(if (DebugCycleCheck) 1443 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1444 << *QTop.second << "\n"); 1445 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1446 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1447 C != QPairRange.second; ++C) { 1448 if (C->second == P) { 1449 DEBUG(dbgs() 1450 << "BBV: rejected to prevent non-trivial cycle formation: " 1451 << *C->first.first << " <-> " << *C->first.second << "\n"); 1452 return true; 1453 } 1454 1455 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1456 Q.push_back(C->second); 1457 } 1458 } while (!Q.empty()); 1459 1460 return false; 1461 } 1462 1463 // This function builds the initial tree of connected pairs with the 1464 // pair J at the root. 1465 void BBVectorize::buildInitialTreeFor( 1466 std::multimap<Value *, Value *> &CandidatePairs, 1467 std::vector<Value *> &PairableInsts, 1468 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1469 DenseSet<ValuePair> &PairableInstUsers, 1470 DenseMap<Value *, Value *> &ChosenPairs, 1471 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1472 // Each of these pairs is viewed as the root node of a Tree. The Tree 1473 // is then walked (depth-first). As this happens, we keep track of 1474 // the pairs that compose the Tree and the maximum depth of the Tree. 1475 SmallVector<ValuePairWithDepth, 32> Q; 1476 // General depth-first post-order traversal: 1477 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1478 do { 1479 ValuePairWithDepth QTop = Q.back(); 1480 1481 // Push each child onto the queue: 1482 bool MoreChildren = false; 1483 size_t MaxChildDepth = QTop.second; 1484 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1485 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1486 k != qtRange.second; ++k) { 1487 // Make sure that this child pair is still a candidate: 1488 bool IsStillCand = false; 1489 VPIteratorPair checkRange = 1490 CandidatePairs.equal_range(k->second.first); 1491 for (std::multimap<Value *, Value *>::iterator m = checkRange.first; 1492 m != checkRange.second; ++m) { 1493 if (m->second == k->second.second) { 1494 IsStillCand = true; 1495 break; 1496 } 1497 } 1498 1499 if (IsStillCand) { 1500 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1501 if (C == Tree.end()) { 1502 size_t d = getDepthFactor(k->second.first); 1503 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1504 MoreChildren = true; 1505 } else { 1506 MaxChildDepth = std::max(MaxChildDepth, C->second); 1507 } 1508 } 1509 } 1510 1511 if (!MoreChildren) { 1512 // Record the current pair as part of the Tree: 1513 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1514 Q.pop_back(); 1515 } 1516 } while (!Q.empty()); 1517 } 1518 1519 // Given some initial tree, prune it by removing conflicting pairs (pairs 1520 // that cannot be simultaneously chosen for vectorization). 1521 void BBVectorize::pruneTreeFor( 1522 std::multimap<Value *, Value *> &CandidatePairs, 1523 std::vector<Value *> &PairableInsts, 1524 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1525 DenseSet<ValuePair> &PairableInstUsers, 1526 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1527 DenseSet<VPPair> &PairableInstUserPairSet, 1528 DenseMap<Value *, Value *> &ChosenPairs, 1529 DenseMap<ValuePair, size_t> &Tree, 1530 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1531 bool UseCycleCheck) { 1532 SmallVector<ValuePairWithDepth, 32> Q; 1533 // General depth-first post-order traversal: 1534 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1535 do { 1536 ValuePairWithDepth QTop = Q.pop_back_val(); 1537 PrunedTree.insert(QTop.first); 1538 1539 // Visit each child, pruning as necessary... 1540 SmallVector<ValuePairWithDepth, 8> BestChildren; 1541 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1542 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1543 K != QTopRange.second; ++K) { 1544 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1545 if (C == Tree.end()) continue; 1546 1547 // This child is in the Tree, now we need to make sure it is the 1548 // best of any conflicting children. There could be multiple 1549 // conflicting children, so first, determine if we're keeping 1550 // this child, then delete conflicting children as necessary. 1551 1552 // It is also necessary to guard against pairing-induced 1553 // dependencies. Consider instructions a .. x .. y .. b 1554 // such that (a,b) are to be fused and (x,y) are to be fused 1555 // but a is an input to x and b is an output from y. This 1556 // means that y cannot be moved after b but x must be moved 1557 // after b for (a,b) to be fused. In other words, after 1558 // fusing (a,b) we have y .. a/b .. x where y is an input 1559 // to a/b and x is an output to a/b: x and y can no longer 1560 // be legally fused. To prevent this condition, we must 1561 // make sure that a child pair added to the Tree is not 1562 // both an input and output of an already-selected pair. 1563 1564 // Pairing-induced dependencies can also form from more complicated 1565 // cycles. The pair vs. pair conflicts are easy to check, and so 1566 // that is done explicitly for "fast rejection", and because for 1567 // child vs. child conflicts, we may prefer to keep the current 1568 // pair in preference to the already-selected child. 1569 DenseSet<ValuePair> CurrentPairs; 1570 1571 bool CanAdd = true; 1572 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1573 = BestChildren.begin(), E2 = BestChildren.end(); 1574 C2 != E2; ++C2) { 1575 if (C2->first.first == C->first.first || 1576 C2->first.first == C->first.second || 1577 C2->first.second == C->first.first || 1578 C2->first.second == C->first.second || 1579 pairsConflict(C2->first, C->first, PairableInstUsers, 1580 UseCycleCheck ? &PairableInstUserMap : 0, 1581 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1582 if (C2->second >= C->second) { 1583 CanAdd = false; 1584 break; 1585 } 1586 1587 CurrentPairs.insert(C2->first); 1588 } 1589 } 1590 if (!CanAdd) continue; 1591 1592 // Even worse, this child could conflict with another node already 1593 // selected for the Tree. If that is the case, ignore this child. 1594 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1595 E2 = PrunedTree.end(); T != E2; ++T) { 1596 if (T->first == C->first.first || 1597 T->first == C->first.second || 1598 T->second == C->first.first || 1599 T->second == C->first.second || 1600 pairsConflict(*T, C->first, PairableInstUsers, 1601 UseCycleCheck ? &PairableInstUserMap : 0, 1602 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1603 CanAdd = false; 1604 break; 1605 } 1606 1607 CurrentPairs.insert(*T); 1608 } 1609 if (!CanAdd) continue; 1610 1611 // And check the queue too... 1612 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1613 E2 = Q.end(); C2 != E2; ++C2) { 1614 if (C2->first.first == C->first.first || 1615 C2->first.first == C->first.second || 1616 C2->first.second == C->first.first || 1617 C2->first.second == C->first.second || 1618 pairsConflict(C2->first, C->first, PairableInstUsers, 1619 UseCycleCheck ? &PairableInstUserMap : 0, 1620 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1621 CanAdd = false; 1622 break; 1623 } 1624 1625 CurrentPairs.insert(C2->first); 1626 } 1627 if (!CanAdd) continue; 1628 1629 // Last but not least, check for a conflict with any of the 1630 // already-chosen pairs. 1631 for (DenseMap<Value *, Value *>::iterator C2 = 1632 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1633 C2 != E2; ++C2) { 1634 if (pairsConflict(*C2, C->first, PairableInstUsers, 1635 UseCycleCheck ? &PairableInstUserMap : 0, 1636 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1637 CanAdd = false; 1638 break; 1639 } 1640 1641 CurrentPairs.insert(*C2); 1642 } 1643 if (!CanAdd) continue; 1644 1645 // To check for non-trivial cycles formed by the addition of the 1646 // current pair we've formed a list of all relevant pairs, now use a 1647 // graph walk to check for a cycle. We start from the current pair and 1648 // walk the use tree to see if we again reach the current pair. If we 1649 // do, then the current pair is rejected. 1650 1651 // FIXME: It may be more efficient to use a topological-ordering 1652 // algorithm to improve the cycle check. This should be investigated. 1653 if (UseCycleCheck && 1654 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1655 continue; 1656 1657 // This child can be added, but we may have chosen it in preference 1658 // to an already-selected child. Check for this here, and if a 1659 // conflict is found, then remove the previously-selected child 1660 // before adding this one in its place. 1661 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1662 = BestChildren.begin(); C2 != BestChildren.end();) { 1663 if (C2->first.first == C->first.first || 1664 C2->first.first == C->first.second || 1665 C2->first.second == C->first.first || 1666 C2->first.second == C->first.second || 1667 pairsConflict(C2->first, C->first, PairableInstUsers)) 1668 C2 = BestChildren.erase(C2); 1669 else 1670 ++C2; 1671 } 1672 1673 BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); 1674 } 1675 1676 for (SmallVector<ValuePairWithDepth, 8>::iterator C 1677 = BestChildren.begin(), E2 = BestChildren.end(); 1678 C != E2; ++C) { 1679 size_t DepthF = getDepthFactor(C->first.first); 1680 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1681 } 1682 } while (!Q.empty()); 1683 } 1684 1685 // This function finds the best tree of mututally-compatible connected 1686 // pairs, given the choice of root pairs as an iterator range. 1687 void BBVectorize::findBestTreeFor( 1688 std::multimap<Value *, Value *> &CandidatePairs, 1689 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1690 std::vector<Value *> &PairableInsts, 1691 DenseSet<ValuePair> &FixedOrderPairs, 1692 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1693 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1694 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1695 DenseSet<ValuePair> &PairableInstUsers, 1696 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1697 DenseSet<VPPair> &PairableInstUserPairSet, 1698 DenseMap<Value *, Value *> &ChosenPairs, 1699 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1700 int &BestEffSize, VPIteratorPair ChoiceRange, 1701 bool UseCycleCheck) { 1702 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1703 J != ChoiceRange.second; ++J) { 1704 1705 // Before going any further, make sure that this pair does not 1706 // conflict with any already-selected pairs (see comment below 1707 // near the Tree pruning for more details). 1708 DenseSet<ValuePair> ChosenPairSet; 1709 bool DoesConflict = false; 1710 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1711 E = ChosenPairs.end(); C != E; ++C) { 1712 if (pairsConflict(*C, *J, PairableInstUsers, 1713 UseCycleCheck ? &PairableInstUserMap : 0, 1714 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1715 DoesConflict = true; 1716 break; 1717 } 1718 1719 ChosenPairSet.insert(*C); 1720 } 1721 if (DoesConflict) continue; 1722 1723 if (UseCycleCheck && 1724 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1725 continue; 1726 1727 DenseMap<ValuePair, size_t> Tree; 1728 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1729 PairableInstUsers, ChosenPairs, Tree, *J); 1730 1731 // Because we'll keep the child with the largest depth, the largest 1732 // depth is still the same in the unpruned Tree. 1733 size_t MaxDepth = Tree.lookup(*J); 1734 1735 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1736 << *J->first << " <-> " << *J->second << "} of depth " << 1737 MaxDepth << " and size " << Tree.size() << "\n"); 1738 1739 // At this point the Tree has been constructed, but, may contain 1740 // contradictory children (meaning that different children of 1741 // some tree node may be attempting to fuse the same instruction). 1742 // So now we walk the tree again, in the case of a conflict, 1743 // keep only the child with the largest depth. To break a tie, 1744 // favor the first child. 1745 1746 DenseSet<ValuePair> PrunedTree; 1747 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1748 PairableInstUsers, PairableInstUserMap, PairableInstUserPairSet, 1749 ChosenPairs, Tree, PrunedTree, *J, UseCycleCheck); 1750 1751 int EffSize = 0; 1752 if (TTI) { 1753 DenseSet<Value *> PrunedTreeInstrs; 1754 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1755 E = PrunedTree.end(); S != E; ++S) { 1756 PrunedTreeInstrs.insert(S->first); 1757 PrunedTreeInstrs.insert(S->second); 1758 } 1759 1760 // The set of pairs that have already contributed to the total cost. 1761 DenseSet<ValuePair> IncomingPairs; 1762 1763 // If the cost model were perfect, this might not be necessary; but we 1764 // need to make sure that we don't get stuck vectorizing our own 1765 // shuffle chains. 1766 bool HasNontrivialInsts = false; 1767 1768 // The node weights represent the cost savings associated with 1769 // fusing the pair of instructions. 1770 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1771 E = PrunedTree.end(); S != E; ++S) { 1772 if (!isa<ShuffleVectorInst>(S->first) && 1773 !isa<InsertElementInst>(S->first) && 1774 !isa<ExtractElementInst>(S->first)) 1775 HasNontrivialInsts = true; 1776 1777 bool FlipOrder = false; 1778 1779 if (getDepthFactor(S->first)) { 1780 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1781 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1782 << *S->first << " <-> " << *S->second << "} = " << 1783 ESContrib << "\n"); 1784 EffSize += ESContrib; 1785 } 1786 1787 // The edge weights contribute in a negative sense: they represent 1788 // the cost of shuffles. 1789 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); 1790 if (IP.first != ConnectedPairDeps.end()) { 1791 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1792 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1793 Q != IP.second; ++Q) { 1794 if (!PrunedTree.count(Q->second)) 1795 continue; 1796 DenseMap<VPPair, unsigned>::iterator R = 1797 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1798 assert(R != PairConnectionTypes.end() && 1799 "Cannot find pair connection type"); 1800 if (R->second == PairConnectionDirect) 1801 ++NumDepsDirect; 1802 else if (R->second == PairConnectionSwap) 1803 ++NumDepsSwap; 1804 } 1805 1806 // If there are more swaps than direct connections, then 1807 // the pair order will be flipped during fusion. So the real 1808 // number of swaps is the minimum number. 1809 FlipOrder = !FixedOrderPairs.count(*S) && 1810 ((NumDepsSwap > NumDepsDirect) || 1811 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1812 1813 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1814 Q != IP.second; ++Q) { 1815 if (!PrunedTree.count(Q->second)) 1816 continue; 1817 DenseMap<VPPair, unsigned>::iterator R = 1818 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1819 assert(R != PairConnectionTypes.end() && 1820 "Cannot find pair connection type"); 1821 Type *Ty1 = Q->second.first->getType(), 1822 *Ty2 = Q->second.second->getType(); 1823 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1824 if ((R->second == PairConnectionDirect && FlipOrder) || 1825 (R->second == PairConnectionSwap && !FlipOrder) || 1826 R->second == PairConnectionSplat) { 1827 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1828 VTy, VTy); 1829 1830 if (VTy->getVectorNumElements() == 2) { 1831 if (R->second == PairConnectionSplat) 1832 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1833 TargetTransformInfo::SK_Broadcast, VTy)); 1834 else 1835 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1836 TargetTransformInfo::SK_Reverse, VTy)); 1837 } 1838 1839 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1840 *Q->second.first << " <-> " << *Q->second.second << 1841 "} -> {" << 1842 *S->first << " <-> " << *S->second << "} = " << 1843 ESContrib << "\n"); 1844 EffSize -= ESContrib; 1845 } 1846 } 1847 } 1848 1849 // Compute the cost of outgoing edges. We assume that edges outgoing 1850 // to shuffles, inserts or extracts can be merged, and so contribute 1851 // no additional cost. 1852 if (!S->first->getType()->isVoidTy()) { 1853 Type *Ty1 = S->first->getType(), 1854 *Ty2 = S->second->getType(); 1855 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1856 1857 bool NeedsExtraction = false; 1858 for (Value::use_iterator I = S->first->use_begin(), 1859 IE = S->first->use_end(); I != IE; ++I) { 1860 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1861 // Shuffle can be folded if it has no other input 1862 if (isa<UndefValue>(SI->getOperand(1))) 1863 continue; 1864 } 1865 if (isa<ExtractElementInst>(*I)) 1866 continue; 1867 if (PrunedTreeInstrs.count(*I)) 1868 continue; 1869 NeedsExtraction = true; 1870 break; 1871 } 1872 1873 if (NeedsExtraction) { 1874 int ESContrib; 1875 if (Ty1->isVectorTy()) { 1876 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1877 Ty1, VTy); 1878 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1879 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1)); 1880 } else 1881 ESContrib = (int) TTI->getVectorInstrCost( 1882 Instruction::ExtractElement, VTy, 0); 1883 1884 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1885 *S->first << "} = " << ESContrib << "\n"); 1886 EffSize -= ESContrib; 1887 } 1888 1889 NeedsExtraction = false; 1890 for (Value::use_iterator I = S->second->use_begin(), 1891 IE = S->second->use_end(); I != IE; ++I) { 1892 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1893 // Shuffle can be folded if it has no other input 1894 if (isa<UndefValue>(SI->getOperand(1))) 1895 continue; 1896 } 1897 if (isa<ExtractElementInst>(*I)) 1898 continue; 1899 if (PrunedTreeInstrs.count(*I)) 1900 continue; 1901 NeedsExtraction = true; 1902 break; 1903 } 1904 1905 if (NeedsExtraction) { 1906 int ESContrib; 1907 if (Ty2->isVectorTy()) { 1908 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1909 Ty2, VTy); 1910 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1911 TargetTransformInfo::SK_ExtractSubvector, VTy, 1912 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2)); 1913 } else 1914 ESContrib = (int) TTI->getVectorInstrCost( 1915 Instruction::ExtractElement, VTy, 1); 1916 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1917 *S->second << "} = " << ESContrib << "\n"); 1918 EffSize -= ESContrib; 1919 } 1920 } 1921 1922 // Compute the cost of incoming edges. 1923 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 1924 Instruction *S1 = cast<Instruction>(S->first), 1925 *S2 = cast<Instruction>(S->second); 1926 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 1927 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 1928 1929 // Combining constants into vector constants (or small vector 1930 // constants into larger ones are assumed free). 1931 if (isa<Constant>(O1) && isa<Constant>(O2)) 1932 continue; 1933 1934 if (FlipOrder) 1935 std::swap(O1, O2); 1936 1937 ValuePair VP = ValuePair(O1, O2); 1938 ValuePair VPR = ValuePair(O2, O1); 1939 1940 // Internal edges are not handled here. 1941 if (PrunedTree.count(VP) || PrunedTree.count(VPR)) 1942 continue; 1943 1944 Type *Ty1 = O1->getType(), 1945 *Ty2 = O2->getType(); 1946 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1947 1948 // Combining vector operations of the same type is also assumed 1949 // folded with other operations. 1950 if (Ty1 == Ty2) { 1951 // If both are insert elements, then both can be widened. 1952 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), 1953 *IEO2 = dyn_cast<InsertElementInst>(O2); 1954 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) 1955 continue; 1956 // If both are extract elements, and both have the same input 1957 // type, then they can be replaced with a shuffle 1958 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 1959 *EIO2 = dyn_cast<ExtractElementInst>(O2); 1960 if (EIO1 && EIO2 && 1961 EIO1->getOperand(0)->getType() == 1962 EIO2->getOperand(0)->getType()) 1963 continue; 1964 // If both are a shuffle with equal operand types and only two 1965 // unqiue operands, then they can be replaced with a single 1966 // shuffle 1967 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 1968 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 1969 if (SIO1 && SIO2 && 1970 SIO1->getOperand(0)->getType() == 1971 SIO2->getOperand(0)->getType()) { 1972 SmallSet<Value *, 4> SIOps; 1973 SIOps.insert(SIO1->getOperand(0)); 1974 SIOps.insert(SIO1->getOperand(1)); 1975 SIOps.insert(SIO2->getOperand(0)); 1976 SIOps.insert(SIO2->getOperand(1)); 1977 if (SIOps.size() <= 2) 1978 continue; 1979 } 1980 } 1981 1982 int ESContrib; 1983 // This pair has already been formed. 1984 if (IncomingPairs.count(VP)) { 1985 continue; 1986 } else if (IncomingPairs.count(VPR)) { 1987 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1988 VTy, VTy); 1989 1990 if (VTy->getVectorNumElements() == 2) 1991 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1992 TargetTransformInfo::SK_Reverse, VTy)); 1993 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 1994 ESContrib = (int) TTI->getVectorInstrCost( 1995 Instruction::InsertElement, VTy, 0); 1996 ESContrib += (int) TTI->getVectorInstrCost( 1997 Instruction::InsertElement, VTy, 1); 1998 } else if (!Ty1->isVectorTy()) { 1999 // O1 needs to be inserted into a vector of size O2, and then 2000 // both need to be shuffled together. 2001 ESContrib = (int) TTI->getVectorInstrCost( 2002 Instruction::InsertElement, Ty2, 0); 2003 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2004 VTy, Ty2); 2005 } else if (!Ty2->isVectorTy()) { 2006 // O2 needs to be inserted into a vector of size O1, and then 2007 // both need to be shuffled together. 2008 ESContrib = (int) TTI->getVectorInstrCost( 2009 Instruction::InsertElement, Ty1, 0); 2010 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2011 VTy, Ty1); 2012 } else { 2013 Type *TyBig = Ty1, *TySmall = Ty2; 2014 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 2015 std::swap(TyBig, TySmall); 2016 2017 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2018 VTy, TyBig); 2019 if (TyBig != TySmall) 2020 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2021 TyBig, TySmall); 2022 } 2023 2024 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 2025 << *O1 << " <-> " << *O2 << "} = " << 2026 ESContrib << "\n"); 2027 EffSize -= ESContrib; 2028 IncomingPairs.insert(VP); 2029 } 2030 } 2031 } 2032 2033 if (!HasNontrivialInsts) { 2034 DEBUG(if (DebugPairSelection) dbgs() << 2035 "\tNo non-trivial instructions in tree;" 2036 " override to zero effective size\n"); 2037 EffSize = 0; 2038 } 2039 } else { 2040 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 2041 E = PrunedTree.end(); S != E; ++S) 2042 EffSize += (int) getDepthFactor(S->first); 2043 } 2044 2045 DEBUG(if (DebugPairSelection) 2046 dbgs() << "BBV: found pruned Tree for pair {" 2047 << *J->first << " <-> " << *J->second << "} of depth " << 2048 MaxDepth << " and size " << PrunedTree.size() << 2049 " (effective size: " << EffSize << ")\n"); 2050 if (((TTI && !UseChainDepthWithTI) || 2051 MaxDepth >= Config.ReqChainDepth) && 2052 EffSize > 0 && EffSize > BestEffSize) { 2053 BestMaxDepth = MaxDepth; 2054 BestEffSize = EffSize; 2055 BestTree = PrunedTree; 2056 } 2057 } 2058 } 2059 2060 // Given the list of candidate pairs, this function selects those 2061 // that will be fused into vector instructions. 2062 void BBVectorize::choosePairs( 2063 std::multimap<Value *, Value *> &CandidatePairs, 2064 DenseMap<ValuePair, int> &CandidatePairCostSavings, 2065 std::vector<Value *> &PairableInsts, 2066 DenseSet<ValuePair> &FixedOrderPairs, 2067 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2068 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2069 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 2070 DenseSet<ValuePair> &PairableInstUsers, 2071 DenseMap<Value *, Value *>& ChosenPairs) { 2072 bool UseCycleCheck = 2073 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 2074 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 2075 DenseSet<VPPair> PairableInstUserPairSet; 2076 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 2077 E = PairableInsts.end(); I != E; ++I) { 2078 // The number of possible pairings for this variable: 2079 size_t NumChoices = CandidatePairs.count(*I); 2080 if (!NumChoices) continue; 2081 2082 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 2083 2084 // The best pair to choose and its tree: 2085 size_t BestMaxDepth = 0; 2086 int BestEffSize = 0; 2087 DenseSet<ValuePair> BestTree; 2088 findBestTreeFor(CandidatePairs, CandidatePairCostSavings, 2089 PairableInsts, FixedOrderPairs, PairConnectionTypes, 2090 ConnectedPairs, ConnectedPairDeps, 2091 PairableInstUsers, PairableInstUserMap, 2092 PairableInstUserPairSet, ChosenPairs, 2093 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 2094 UseCycleCheck); 2095 2096 // A tree has been chosen (or not) at this point. If no tree was 2097 // chosen, then this instruction, I, cannot be paired (and is no longer 2098 // considered). 2099 2100 DEBUG(if (BestTree.size() > 0) 2101 dbgs() << "BBV: selected pairs in the best tree for: " 2102 << *cast<Instruction>(*I) << "\n"); 2103 2104 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 2105 SE2 = BestTree.end(); S != SE2; ++S) { 2106 // Insert the members of this tree into the list of chosen pairs. 2107 ChosenPairs.insert(ValuePair(S->first, S->second)); 2108 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 2109 *S->second << "\n"); 2110 2111 // Remove all candidate pairs that have values in the chosen tree. 2112 for (std::multimap<Value *, Value *>::iterator K = 2113 CandidatePairs.begin(); K != CandidatePairs.end();) { 2114 if (K->first == S->first || K->second == S->first || 2115 K->second == S->second || K->first == S->second) { 2116 // Don't remove the actual pair chosen so that it can be used 2117 // in subsequent tree selections. 2118 if (!(K->first == S->first && K->second == S->second)) 2119 CandidatePairs.erase(K++); 2120 else 2121 ++K; 2122 } else { 2123 ++K; 2124 } 2125 } 2126 } 2127 } 2128 2129 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2130 } 2131 2132 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2133 unsigned n = 0) { 2134 if (!I->hasName()) 2135 return ""; 2136 2137 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2138 (n > 0 ? "." + utostr(n) : "")).str(); 2139 } 2140 2141 // Returns the value that is to be used as the pointer input to the vector 2142 // instruction that fuses I with J. 2143 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2144 Instruction *I, Instruction *J, unsigned o) { 2145 Value *IPtr, *JPtr; 2146 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2147 int64_t OffsetInElmts; 2148 2149 // Note: the analysis might fail here, that is why the pair order has 2150 // been precomputed (OffsetInElmts must be unused here). 2151 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2152 IAddressSpace, JAddressSpace, 2153 OffsetInElmts, false); 2154 2155 // The pointer value is taken to be the one with the lowest offset. 2156 Value *VPtr = IPtr; 2157 2158 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 2159 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 2160 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2161 Type *VArgPtrType = PointerType::get(VArgType, 2162 cast<PointerType>(IPtr->getType())->getAddressSpace()); 2163 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2164 /* insert before */ I); 2165 } 2166 2167 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2168 unsigned MaskOffset, unsigned NumInElem, 2169 unsigned NumInElem1, unsigned IdxOffset, 2170 std::vector<Constant*> &Mask) { 2171 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 2172 for (unsigned v = 0; v < NumElem1; ++v) { 2173 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2174 if (m < 0) { 2175 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2176 } else { 2177 unsigned mm = m + (int) IdxOffset; 2178 if (m >= (int) NumInElem1) 2179 mm += (int) NumInElem; 2180 2181 Mask[v+MaskOffset] = 2182 ConstantInt::get(Type::getInt32Ty(Context), mm); 2183 } 2184 } 2185 } 2186 2187 // Returns the value that is to be used as the vector-shuffle mask to the 2188 // vector instruction that fuses I with J. 2189 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2190 Instruction *I, Instruction *J) { 2191 // This is the shuffle mask. We need to append the second 2192 // mask to the first, and the numbers need to be adjusted. 2193 2194 Type *ArgTypeI = I->getType(); 2195 Type *ArgTypeJ = J->getType(); 2196 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2197 2198 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 2199 2200 // Get the total number of elements in the fused vector type. 2201 // By definition, this must equal the number of elements in 2202 // the final mask. 2203 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 2204 std::vector<Constant*> Mask(NumElem); 2205 2206 Type *OpTypeI = I->getOperand(0)->getType(); 2207 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 2208 Type *OpTypeJ = J->getOperand(0)->getType(); 2209 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 2210 2211 // The fused vector will be: 2212 // ----------------------------------------------------- 2213 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2214 // ----------------------------------------------------- 2215 // from which we'll extract NumElem total elements (where the first NumElemI 2216 // of them come from the mask in I and the remainder come from the mask 2217 // in J. 2218 2219 // For the mask from the first pair... 2220 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2221 0, Mask); 2222 2223 // For the mask from the second pair... 2224 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2225 NumInElemI, Mask); 2226 2227 return ConstantVector::get(Mask); 2228 } 2229 2230 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2231 Instruction *J, unsigned o, Value *&LOp, 2232 unsigned numElemL, 2233 Type *ArgTypeL, Type *ArgTypeH, 2234 bool IBeforeJ, unsigned IdxOff) { 2235 bool ExpandedIEChain = false; 2236 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2237 // If we have a pure insertelement chain, then this can be rewritten 2238 // into a chain that directly builds the larger type. 2239 if (isPureIEChain(LIE)) { 2240 SmallVector<Value *, 8> VectElemts(numElemL, 2241 UndefValue::get(ArgTypeL->getScalarType())); 2242 InsertElementInst *LIENext = LIE; 2243 do { 2244 unsigned Idx = 2245 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2246 VectElemts[Idx] = LIENext->getOperand(1); 2247 } while ((LIENext = 2248 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2249 2250 LIENext = 0; 2251 Value *LIEPrev = UndefValue::get(ArgTypeH); 2252 for (unsigned i = 0; i < numElemL; ++i) { 2253 if (isa<UndefValue>(VectElemts[i])) continue; 2254 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2255 ConstantInt::get(Type::getInt32Ty(Context), 2256 i + IdxOff), 2257 getReplacementName(IBeforeJ ? I : J, 2258 true, o, i+1)); 2259 LIENext->insertBefore(IBeforeJ ? J : I); 2260 LIEPrev = LIENext; 2261 } 2262 2263 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2264 ExpandedIEChain = true; 2265 } 2266 } 2267 2268 return ExpandedIEChain; 2269 } 2270 2271 // Returns the value to be used as the specified operand of the vector 2272 // instruction that fuses I with J. 2273 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2274 Instruction *J, unsigned o, bool IBeforeJ) { 2275 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2276 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2277 2278 // Compute the fused vector type for this operand 2279 Type *ArgTypeI = I->getOperand(o)->getType(); 2280 Type *ArgTypeJ = J->getOperand(o)->getType(); 2281 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2282 2283 Instruction *L = I, *H = J; 2284 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2285 2286 unsigned numElemL; 2287 if (ArgTypeL->isVectorTy()) 2288 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 2289 else 2290 numElemL = 1; 2291 2292 unsigned numElemH; 2293 if (ArgTypeH->isVectorTy()) 2294 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 2295 else 2296 numElemH = 1; 2297 2298 Value *LOp = L->getOperand(o); 2299 Value *HOp = H->getOperand(o); 2300 unsigned numElem = VArgType->getNumElements(); 2301 2302 // First, we check if we can reuse the "original" vector outputs (if these 2303 // exist). We might need a shuffle. 2304 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2305 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2306 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2307 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2308 2309 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2310 // optimization. The input vectors to the shuffle might be a different 2311 // length from the shuffle outputs. Unfortunately, the replacement 2312 // shuffle mask has already been formed, and the mask entries are sensitive 2313 // to the sizes of the inputs. 2314 bool IsSizeChangeShuffle = 2315 isa<ShuffleVectorInst>(L) && 2316 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2317 2318 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2319 // We can have at most two unique vector inputs. 2320 bool CanUseInputs = true; 2321 Value *I1, *I2 = 0; 2322 if (LEE) { 2323 I1 = LEE->getOperand(0); 2324 } else { 2325 I1 = LSV->getOperand(0); 2326 I2 = LSV->getOperand(1); 2327 if (I2 == I1 || isa<UndefValue>(I2)) 2328 I2 = 0; 2329 } 2330 2331 if (HEE) { 2332 Value *I3 = HEE->getOperand(0); 2333 if (!I2 && I3 != I1) 2334 I2 = I3; 2335 else if (I3 != I1 && I3 != I2) 2336 CanUseInputs = false; 2337 } else { 2338 Value *I3 = HSV->getOperand(0); 2339 if (!I2 && I3 != I1) 2340 I2 = I3; 2341 else if (I3 != I1 && I3 != I2) 2342 CanUseInputs = false; 2343 2344 if (CanUseInputs) { 2345 Value *I4 = HSV->getOperand(1); 2346 if (!isa<UndefValue>(I4)) { 2347 if (!I2 && I4 != I1) 2348 I2 = I4; 2349 else if (I4 != I1 && I4 != I2) 2350 CanUseInputs = false; 2351 } 2352 } 2353 } 2354 2355 if (CanUseInputs) { 2356 unsigned LOpElem = 2357 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2358 ->getNumElements(); 2359 unsigned HOpElem = 2360 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2361 ->getNumElements(); 2362 2363 // We have one or two input vectors. We need to map each index of the 2364 // operands to the index of the original vector. 2365 SmallVector<std::pair<int, int>, 8> II(numElem); 2366 for (unsigned i = 0; i < numElemL; ++i) { 2367 int Idx, INum; 2368 if (LEE) { 2369 Idx = 2370 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2371 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2372 } else { 2373 Idx = LSV->getMaskValue(i); 2374 if (Idx < (int) LOpElem) { 2375 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2376 } else { 2377 Idx -= LOpElem; 2378 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2379 } 2380 } 2381 2382 II[i] = std::pair<int, int>(Idx, INum); 2383 } 2384 for (unsigned i = 0; i < numElemH; ++i) { 2385 int Idx, INum; 2386 if (HEE) { 2387 Idx = 2388 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2389 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2390 } else { 2391 Idx = HSV->getMaskValue(i); 2392 if (Idx < (int) HOpElem) { 2393 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2394 } else { 2395 Idx -= HOpElem; 2396 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2397 } 2398 } 2399 2400 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2401 } 2402 2403 // We now have an array which tells us from which index of which 2404 // input vector each element of the operand comes. 2405 VectorType *I1T = cast<VectorType>(I1->getType()); 2406 unsigned I1Elem = I1T->getNumElements(); 2407 2408 if (!I2) { 2409 // In this case there is only one underlying vector input. Check for 2410 // the trivial case where we can use the input directly. 2411 if (I1Elem == numElem) { 2412 bool ElemInOrder = true; 2413 for (unsigned i = 0; i < numElem; ++i) { 2414 if (II[i].first != (int) i && II[i].first != -1) { 2415 ElemInOrder = false; 2416 break; 2417 } 2418 } 2419 2420 if (ElemInOrder) 2421 return I1; 2422 } 2423 2424 // A shuffle is needed. 2425 std::vector<Constant *> Mask(numElem); 2426 for (unsigned i = 0; i < numElem; ++i) { 2427 int Idx = II[i].first; 2428 if (Idx == -1) 2429 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2430 else 2431 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2432 } 2433 2434 Instruction *S = 2435 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2436 ConstantVector::get(Mask), 2437 getReplacementName(IBeforeJ ? I : J, 2438 true, o)); 2439 S->insertBefore(IBeforeJ ? J : I); 2440 return S; 2441 } 2442 2443 VectorType *I2T = cast<VectorType>(I2->getType()); 2444 unsigned I2Elem = I2T->getNumElements(); 2445 2446 // This input comes from two distinct vectors. The first step is to 2447 // make sure that both vectors are the same length. If not, the 2448 // smaller one will need to grow before they can be shuffled together. 2449 if (I1Elem < I2Elem) { 2450 std::vector<Constant *> Mask(I2Elem); 2451 unsigned v = 0; 2452 for (; v < I1Elem; ++v) 2453 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2454 for (; v < I2Elem; ++v) 2455 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2456 2457 Instruction *NewI1 = 2458 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2459 ConstantVector::get(Mask), 2460 getReplacementName(IBeforeJ ? I : J, 2461 true, o, 1)); 2462 NewI1->insertBefore(IBeforeJ ? J : I); 2463 I1 = NewI1; 2464 I1T = I2T; 2465 I1Elem = I2Elem; 2466 } else if (I1Elem > I2Elem) { 2467 std::vector<Constant *> Mask(I1Elem); 2468 unsigned v = 0; 2469 for (; v < I2Elem; ++v) 2470 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2471 for (; v < I1Elem; ++v) 2472 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2473 2474 Instruction *NewI2 = 2475 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2476 ConstantVector::get(Mask), 2477 getReplacementName(IBeforeJ ? I : J, 2478 true, o, 1)); 2479 NewI2->insertBefore(IBeforeJ ? J : I); 2480 I2 = NewI2; 2481 I2T = I1T; 2482 I2Elem = I1Elem; 2483 } 2484 2485 // Now that both I1 and I2 are the same length we can shuffle them 2486 // together (and use the result). 2487 std::vector<Constant *> Mask(numElem); 2488 for (unsigned v = 0; v < numElem; ++v) { 2489 if (II[v].first == -1) { 2490 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2491 } else { 2492 int Idx = II[v].first + II[v].second * I1Elem; 2493 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2494 } 2495 } 2496 2497 Instruction *NewOp = 2498 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2499 getReplacementName(IBeforeJ ? I : J, true, o)); 2500 NewOp->insertBefore(IBeforeJ ? J : I); 2501 return NewOp; 2502 } 2503 } 2504 2505 Type *ArgType = ArgTypeL; 2506 if (numElemL < numElemH) { 2507 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2508 ArgTypeL, VArgType, IBeforeJ, 1)) { 2509 // This is another short-circuit case: we're combining a scalar into 2510 // a vector that is formed by an IE chain. We've just expanded the IE 2511 // chain, now insert the scalar and we're done. 2512 2513 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2514 getReplacementName(IBeforeJ ? I : J, true, o)); 2515 S->insertBefore(IBeforeJ ? J : I); 2516 return S; 2517 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2518 ArgTypeH, IBeforeJ)) { 2519 // The two vector inputs to the shuffle must be the same length, 2520 // so extend the smaller vector to be the same length as the larger one. 2521 Instruction *NLOp; 2522 if (numElemL > 1) { 2523 2524 std::vector<Constant *> Mask(numElemH); 2525 unsigned v = 0; 2526 for (; v < numElemL; ++v) 2527 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2528 for (; v < numElemH; ++v) 2529 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2530 2531 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2532 ConstantVector::get(Mask), 2533 getReplacementName(IBeforeJ ? I : J, 2534 true, o, 1)); 2535 } else { 2536 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2537 getReplacementName(IBeforeJ ? I : J, 2538 true, o, 1)); 2539 } 2540 2541 NLOp->insertBefore(IBeforeJ ? J : I); 2542 LOp = NLOp; 2543 } 2544 2545 ArgType = ArgTypeH; 2546 } else if (numElemL > numElemH) { 2547 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2548 ArgTypeH, VArgType, IBeforeJ)) { 2549 Instruction *S = 2550 InsertElementInst::Create(LOp, HOp, 2551 ConstantInt::get(Type::getInt32Ty(Context), 2552 numElemL), 2553 getReplacementName(IBeforeJ ? I : J, 2554 true, o)); 2555 S->insertBefore(IBeforeJ ? J : I); 2556 return S; 2557 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2558 ArgTypeL, IBeforeJ)) { 2559 Instruction *NHOp; 2560 if (numElemH > 1) { 2561 std::vector<Constant *> Mask(numElemL); 2562 unsigned v = 0; 2563 for (; v < numElemH; ++v) 2564 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2565 for (; v < numElemL; ++v) 2566 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2567 2568 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2569 ConstantVector::get(Mask), 2570 getReplacementName(IBeforeJ ? I : J, 2571 true, o, 1)); 2572 } else { 2573 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2574 getReplacementName(IBeforeJ ? I : J, 2575 true, o, 1)); 2576 } 2577 2578 NHOp->insertBefore(IBeforeJ ? J : I); 2579 HOp = NHOp; 2580 } 2581 } 2582 2583 if (ArgType->isVectorTy()) { 2584 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2585 std::vector<Constant*> Mask(numElem); 2586 for (unsigned v = 0; v < numElem; ++v) { 2587 unsigned Idx = v; 2588 // If the low vector was expanded, we need to skip the extra 2589 // undefined entries. 2590 if (v >= numElemL && numElemH > numElemL) 2591 Idx += (numElemH - numElemL); 2592 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2593 } 2594 2595 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2596 ConstantVector::get(Mask), 2597 getReplacementName(IBeforeJ ? I : J, true, o)); 2598 BV->insertBefore(IBeforeJ ? J : I); 2599 return BV; 2600 } 2601 2602 Instruction *BV1 = InsertElementInst::Create( 2603 UndefValue::get(VArgType), LOp, CV0, 2604 getReplacementName(IBeforeJ ? I : J, 2605 true, o, 1)); 2606 BV1->insertBefore(IBeforeJ ? J : I); 2607 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2608 getReplacementName(IBeforeJ ? I : J, 2609 true, o, 2)); 2610 BV2->insertBefore(IBeforeJ ? J : I); 2611 return BV2; 2612 } 2613 2614 // This function creates an array of values that will be used as the inputs 2615 // to the vector instruction that fuses I with J. 2616 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2617 Instruction *I, Instruction *J, 2618 SmallVector<Value *, 3> &ReplacedOperands, 2619 bool IBeforeJ) { 2620 unsigned NumOperands = I->getNumOperands(); 2621 2622 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2623 // Iterate backward so that we look at the store pointer 2624 // first and know whether or not we need to flip the inputs. 2625 2626 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2627 // This is the pointer for a load/store instruction. 2628 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2629 continue; 2630 } else if (isa<CallInst>(I)) { 2631 Function *F = cast<CallInst>(I)->getCalledFunction(); 2632 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 2633 if (o == NumOperands-1) { 2634 BasicBlock &BB = *I->getParent(); 2635 2636 Module *M = BB.getParent()->getParent(); 2637 Type *ArgTypeI = I->getType(); 2638 Type *ArgTypeJ = J->getType(); 2639 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2640 2641 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType); 2642 continue; 2643 } else if (IID == Intrinsic::powi && o == 1) { 2644 // The second argument of powi is a single integer and we've already 2645 // checked that both arguments are equal. As a result, we just keep 2646 // I's second argument. 2647 ReplacedOperands[o] = I->getOperand(o); 2648 continue; 2649 } 2650 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2651 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2652 continue; 2653 } 2654 2655 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2656 } 2657 } 2658 2659 // This function creates two values that represent the outputs of the 2660 // original I and J instructions. These are generally vector shuffles 2661 // or extracts. In many cases, these will end up being unused and, thus, 2662 // eliminated by later passes. 2663 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2664 Instruction *J, Instruction *K, 2665 Instruction *&InsertionPt, 2666 Instruction *&K1, Instruction *&K2) { 2667 if (isa<StoreInst>(I)) { 2668 AA->replaceWithNewValue(I, K); 2669 AA->replaceWithNewValue(J, K); 2670 } else { 2671 Type *IType = I->getType(); 2672 Type *JType = J->getType(); 2673 2674 VectorType *VType = getVecTypeForPair(IType, JType); 2675 unsigned numElem = VType->getNumElements(); 2676 2677 unsigned numElemI, numElemJ; 2678 if (IType->isVectorTy()) 2679 numElemI = cast<VectorType>(IType)->getNumElements(); 2680 else 2681 numElemI = 1; 2682 2683 if (JType->isVectorTy()) 2684 numElemJ = cast<VectorType>(JType)->getNumElements(); 2685 else 2686 numElemJ = 1; 2687 2688 if (IType->isVectorTy()) { 2689 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2690 for (unsigned v = 0; v < numElemI; ++v) { 2691 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2692 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2693 } 2694 2695 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2696 ConstantVector::get( Mask1), 2697 getReplacementName(K, false, 1)); 2698 } else { 2699 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2700 K1 = ExtractElementInst::Create(K, CV0, 2701 getReplacementName(K, false, 1)); 2702 } 2703 2704 if (JType->isVectorTy()) { 2705 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2706 for (unsigned v = 0; v < numElemJ; ++v) { 2707 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2708 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2709 } 2710 2711 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2712 ConstantVector::get( Mask2), 2713 getReplacementName(K, false, 2)); 2714 } else { 2715 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2716 K2 = ExtractElementInst::Create(K, CV1, 2717 getReplacementName(K, false, 2)); 2718 } 2719 2720 K1->insertAfter(K); 2721 K2->insertAfter(K1); 2722 InsertionPt = K2; 2723 } 2724 } 2725 2726 // Move all uses of the function I (including pairing-induced uses) after J. 2727 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2728 DenseSet<ValuePair> &LoadMoveSetPairs, 2729 Instruction *I, Instruction *J) { 2730 // Skip to the first instruction past I. 2731 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2732 2733 DenseSet<Value *> Users; 2734 AliasSetTracker WriteSet(*AA); 2735 for (; cast<Instruction>(L) != J; ++L) 2736 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs); 2737 2738 assert(cast<Instruction>(L) == J && 2739 "Tracking has not proceeded far enough to check for dependencies"); 2740 // If J is now in the use set of I, then trackUsesOfI will return true 2741 // and we have a dependency cycle (and the fusing operation must abort). 2742 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs); 2743 } 2744 2745 // Move all uses of the function I (including pairing-induced uses) after J. 2746 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2747 DenseSet<ValuePair> &LoadMoveSetPairs, 2748 Instruction *&InsertionPt, 2749 Instruction *I, Instruction *J) { 2750 // Skip to the first instruction past I. 2751 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2752 2753 DenseSet<Value *> Users; 2754 AliasSetTracker WriteSet(*AA); 2755 for (; cast<Instruction>(L) != J;) { 2756 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) { 2757 // Move this instruction 2758 Instruction *InstToMove = L; ++L; 2759 2760 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2761 " to after " << *InsertionPt << "\n"); 2762 InstToMove->removeFromParent(); 2763 InstToMove->insertAfter(InsertionPt); 2764 InsertionPt = InstToMove; 2765 } else { 2766 ++L; 2767 } 2768 } 2769 } 2770 2771 // Collect all load instruction that are in the move set of a given first 2772 // pair member. These loads depend on the first instruction, I, and so need 2773 // to be moved after J (the second instruction) when the pair is fused. 2774 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2775 DenseMap<Value *, Value *> &ChosenPairs, 2776 std::multimap<Value *, Value *> &LoadMoveSet, 2777 DenseSet<ValuePair> &LoadMoveSetPairs, 2778 Instruction *I) { 2779 // Skip to the first instruction past I. 2780 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2781 2782 DenseSet<Value *> Users; 2783 AliasSetTracker WriteSet(*AA); 2784 2785 // Note: We cannot end the loop when we reach J because J could be moved 2786 // farther down the use chain by another instruction pairing. Also, J 2787 // could be before I if this is an inverted input. 2788 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2789 if (trackUsesOfI(Users, WriteSet, I, L)) { 2790 if (L->mayReadFromMemory()) { 2791 LoadMoveSet.insert(ValuePair(L, I)); 2792 LoadMoveSetPairs.insert(ValuePair(L, I)); 2793 } 2794 } 2795 } 2796 } 2797 2798 // In cases where both load/stores and the computation of their pointers 2799 // are chosen for vectorization, we can end up in a situation where the 2800 // aliasing analysis starts returning different query results as the 2801 // process of fusing instruction pairs continues. Because the algorithm 2802 // relies on finding the same use trees here as were found earlier, we'll 2803 // need to precompute the necessary aliasing information here and then 2804 // manually update it during the fusion process. 2805 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2806 std::vector<Value *> &PairableInsts, 2807 DenseMap<Value *, Value *> &ChosenPairs, 2808 std::multimap<Value *, Value *> &LoadMoveSet, 2809 DenseSet<ValuePair> &LoadMoveSetPairs) { 2810 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2811 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2812 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2813 if (P == ChosenPairs.end()) continue; 2814 2815 Instruction *I = cast<Instruction>(P->first); 2816 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, 2817 LoadMoveSetPairs, I); 2818 } 2819 } 2820 2821 // When the first instruction in each pair is cloned, it will inherit its 2822 // parent's metadata. This metadata must be combined with that of the other 2823 // instruction in a safe way. 2824 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2825 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2826 K->getAllMetadataOtherThanDebugLoc(Metadata); 2827 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2828 unsigned Kind = Metadata[i].first; 2829 MDNode *JMD = J->getMetadata(Kind); 2830 MDNode *KMD = Metadata[i].second; 2831 2832 switch (Kind) { 2833 default: 2834 K->setMetadata(Kind, 0); // Remove unknown metadata 2835 break; 2836 case LLVMContext::MD_tbaa: 2837 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2838 break; 2839 case LLVMContext::MD_fpmath: 2840 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2841 break; 2842 } 2843 } 2844 } 2845 2846 // This function fuses the chosen instruction pairs into vector instructions, 2847 // taking care preserve any needed scalar outputs and, then, it reorders the 2848 // remaining instructions as needed (users of the first member of the pair 2849 // need to be moved to after the location of the second member of the pair 2850 // because the vector instruction is inserted in the location of the pair's 2851 // second member). 2852 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2853 std::vector<Value *> &PairableInsts, 2854 DenseMap<Value *, Value *> &ChosenPairs, 2855 DenseSet<ValuePair> &FixedOrderPairs, 2856 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2857 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2858 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2859 LLVMContext& Context = BB.getContext(); 2860 2861 // During the vectorization process, the order of the pairs to be fused 2862 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2863 // list. After a pair is fused, the flipped pair is removed from the list. 2864 DenseSet<ValuePair> FlippedPairs; 2865 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2866 E = ChosenPairs.end(); P != E; ++P) 2867 FlippedPairs.insert(ValuePair(P->second, P->first)); 2868 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2869 E = FlippedPairs.end(); P != E; ++P) 2870 ChosenPairs.insert(*P); 2871 2872 std::multimap<Value *, Value *> LoadMoveSet; 2873 DenseSet<ValuePair> LoadMoveSetPairs; 2874 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, 2875 LoadMoveSet, LoadMoveSetPairs); 2876 2877 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2878 2879 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2880 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2881 if (P == ChosenPairs.end()) { 2882 ++PI; 2883 continue; 2884 } 2885 2886 if (getDepthFactor(P->first) == 0) { 2887 // These instructions are not really fused, but are tracked as though 2888 // they are. Any case in which it would be interesting to fuse them 2889 // will be taken care of by InstCombine. 2890 --NumFusedOps; 2891 ++PI; 2892 continue; 2893 } 2894 2895 Instruction *I = cast<Instruction>(P->first), 2896 *J = cast<Instruction>(P->second); 2897 2898 DEBUG(dbgs() << "BBV: fusing: " << *I << 2899 " <-> " << *J << "\n"); 2900 2901 // Remove the pair and flipped pair from the list. 2902 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2903 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2904 ChosenPairs.erase(FP); 2905 ChosenPairs.erase(P); 2906 2907 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) { 2908 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2909 " <-> " << *J << 2910 " aborted because of non-trivial dependency cycle\n"); 2911 --NumFusedOps; 2912 ++PI; 2913 continue; 2914 } 2915 2916 // If the pair must have the other order, then flip it. 2917 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2918 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2919 // This pair does not have a fixed order, and so we might want to 2920 // flip it if that will yield fewer shuffles. We count the number 2921 // of dependencies connected via swaps, and those directly connected, 2922 // and flip the order if the number of swaps is greater. 2923 bool OrigOrder = true; 2924 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2925 if (IP.first == ConnectedPairDeps.end()) { 2926 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2927 OrigOrder = false; 2928 } 2929 2930 if (IP.first != ConnectedPairDeps.end()) { 2931 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2932 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2933 Q != IP.second; ++Q) { 2934 DenseMap<VPPair, unsigned>::iterator R = 2935 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2936 assert(R != PairConnectionTypes.end() && 2937 "Cannot find pair connection type"); 2938 if (R->second == PairConnectionDirect) 2939 ++NumDepsDirect; 2940 else if (R->second == PairConnectionSwap) 2941 ++NumDepsSwap; 2942 } 2943 2944 if (!OrigOrder) 2945 std::swap(NumDepsDirect, NumDepsSwap); 2946 2947 if (NumDepsSwap > NumDepsDirect) { 2948 FlipPairOrder = true; 2949 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2950 " <-> " << *J << "\n"); 2951 } 2952 } 2953 } 2954 2955 Instruction *L = I, *H = J; 2956 if (FlipPairOrder) 2957 std::swap(H, L); 2958 2959 // If the pair being fused uses the opposite order from that in the pair 2960 // connection map, then we need to flip the types. 2961 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2962 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2963 Q != IP.second; ++Q) { 2964 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2965 assert(R != PairConnectionTypes.end() && 2966 "Cannot find pair connection type"); 2967 if (R->second == PairConnectionDirect) 2968 R->second = PairConnectionSwap; 2969 else if (R->second == PairConnectionSwap) 2970 R->second = PairConnectionDirect; 2971 } 2972 2973 bool LBeforeH = !FlipPairOrder; 2974 unsigned NumOperands = I->getNumOperands(); 2975 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2976 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2977 LBeforeH); 2978 2979 // Make a copy of the original operation, change its type to the vector 2980 // type and replace its operands with the vector operands. 2981 Instruction *K = L->clone(); 2982 if (L->hasName()) 2983 K->takeName(L); 2984 else if (H->hasName()) 2985 K->takeName(H); 2986 2987 if (!isa<StoreInst>(K)) 2988 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 2989 2990 combineMetadata(K, H); 2991 K->intersectOptionalDataWith(H); 2992 2993 for (unsigned o = 0; o < NumOperands; ++o) 2994 K->setOperand(o, ReplacedOperands[o]); 2995 2996 K->insertAfter(J); 2997 2998 // Instruction insertion point: 2999 Instruction *InsertionPt = K; 3000 Instruction *K1 = 0, *K2 = 0; 3001 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 3002 3003 // The use tree of the first original instruction must be moved to after 3004 // the location of the second instruction. The entire use tree of the 3005 // first instruction is disjoint from the input tree of the second 3006 // (by definition), and so commutes with it. 3007 3008 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J); 3009 3010 if (!isa<StoreInst>(I)) { 3011 L->replaceAllUsesWith(K1); 3012 H->replaceAllUsesWith(K2); 3013 AA->replaceWithNewValue(L, K1); 3014 AA->replaceWithNewValue(H, K2); 3015 } 3016 3017 // Instructions that may read from memory may be in the load move set. 3018 // Once an instruction is fused, we no longer need its move set, and so 3019 // the values of the map never need to be updated. However, when a load 3020 // is fused, we need to merge the entries from both instructions in the 3021 // pair in case those instructions were in the move set of some other 3022 // yet-to-be-fused pair. The loads in question are the keys of the map. 3023 if (I->mayReadFromMemory()) { 3024 std::vector<ValuePair> NewSetMembers; 3025 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 3026 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 3027 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 3028 N != IPairRange.second; ++N) 3029 NewSetMembers.push_back(ValuePair(K, N->second)); 3030 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 3031 N != JPairRange.second; ++N) 3032 NewSetMembers.push_back(ValuePair(K, N->second)); 3033 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 3034 AE = NewSetMembers.end(); A != AE; ++A) { 3035 LoadMoveSet.insert(*A); 3036 LoadMoveSetPairs.insert(*A); 3037 } 3038 } 3039 3040 // Before removing I, set the iterator to the next instruction. 3041 PI = llvm::next(BasicBlock::iterator(I)); 3042 if (cast<Instruction>(PI) == J) 3043 ++PI; 3044 3045 SE->forgetValue(I); 3046 SE->forgetValue(J); 3047 I->eraseFromParent(); 3048 J->eraseFromParent(); 3049 3050 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 3051 BB << "\n"); 3052 } 3053 3054 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 3055 } 3056} 3057 3058char BBVectorize::ID = 0; 3059static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 3060INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3061INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3062INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3063INITIALIZE_PASS_DEPENDENCY(DominatorTree) 3064INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3065INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3066 3067BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 3068 return new BBVectorize(C); 3069} 3070 3071bool 3072llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 3073 BBVectorize BBVectorizer(P, C); 3074 return BBVectorizer.vectorizeBB(BB); 3075} 3076 3077//===----------------------------------------------------------------------===// 3078VectorizeConfig::VectorizeConfig() { 3079 VectorBits = ::VectorBits; 3080 VectorizeBools = !::NoBools; 3081 VectorizeInts = !::NoInts; 3082 VectorizeFloats = !::NoFloats; 3083 VectorizePointers = !::NoPointers; 3084 VectorizeCasts = !::NoCasts; 3085 VectorizeMath = !::NoMath; 3086 VectorizeFMA = !::NoFMA; 3087 VectorizeSelect = !::NoSelect; 3088 VectorizeCmp = !::NoCmp; 3089 VectorizeGEP = !::NoGEP; 3090 VectorizeMemOps = !::NoMemOps; 3091 AlignedOnly = ::AlignedOnly; 3092 ReqChainDepth= ::ReqChainDepth; 3093 SearchLimit = ::SearchLimit; 3094 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 3095 SplatBreaksChain = ::SplatBreaksChain; 3096 MaxInsts = ::MaxInsts; 3097 MaxIter = ::MaxIter; 3098 Pow2LenOnly = ::Pow2LenOnly; 3099 NoMemOpBoost = ::NoMemOpBoost; 3100 FastDep = ::FastDep; 3101} 3102