BBVectorize.cpp revision b1a82589339fed148c12b052d30861a539552f1a
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 DenseSet<ValuePair> &CandidatePairsSet, 244 std::vector<Value *> &PairableInsts, 245 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 246 DenseMap<VPPair, unsigned> &PairConnectionTypes); 247 248 void buildDepMap(BasicBlock &BB, 249 std::multimap<Value *, Value *> &CandidatePairs, 250 std::vector<Value *> &PairableInsts, 251 DenseSet<ValuePair> &PairableInstUsers); 252 253 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, 254 DenseSet<ValuePair> &CandidatePairsSet, 255 DenseMap<ValuePair, int> &CandidatePairCostSavings, 256 std::vector<Value *> &PairableInsts, 257 DenseSet<ValuePair> &FixedOrderPairs, 258 DenseMap<VPPair, unsigned> &PairConnectionTypes, 259 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 260 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 261 DenseSet<ValuePair> &PairableInstUsers, 262 DenseMap<Value *, Value *>& ChosenPairs); 263 264 void fuseChosenPairs(BasicBlock &BB, 265 std::vector<Value *> &PairableInsts, 266 DenseMap<Value *, Value *>& ChosenPairs, 267 DenseSet<ValuePair> &FixedOrderPairs, 268 DenseMap<VPPair, unsigned> &PairConnectionTypes, 269 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 270 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps); 271 272 273 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); 274 275 bool areInstsCompatible(Instruction *I, Instruction *J, 276 bool IsSimpleLoadStore, bool NonPow2Len, 277 int &CostSavings, int &FixedOrder); 278 279 bool trackUsesOfI(DenseSet<Value *> &Users, 280 AliasSetTracker &WriteSet, Instruction *I, 281 Instruction *J, bool UpdateUsers = true, 282 DenseSet<ValuePair> *LoadMoveSetPairs = 0); 283 284 void computePairsConnectedTo( 285 std::multimap<Value *, Value *> &CandidatePairs, 286 DenseSet<ValuePair> &CandidatePairsSet, 287 std::vector<Value *> &PairableInsts, 288 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 289 DenseMap<VPPair, unsigned> &PairConnectionTypes, 290 ValuePair P); 291 292 bool pairsConflict(ValuePair P, ValuePair Q, 293 DenseSet<ValuePair> &PairableInstUsers, 294 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0, 295 DenseSet<VPPair> *PairableInstUserPairSet = 0); 296 297 bool pairWillFormCycle(ValuePair P, 298 std::multimap<ValuePair, ValuePair> &PairableInstUsers, 299 DenseSet<ValuePair> &CurrentPairs); 300 301 void pruneTreeFor( 302 std::multimap<Value *, Value *> &CandidatePairs, 303 std::vector<Value *> &PairableInsts, 304 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 305 DenseSet<ValuePair> &PairableInstUsers, 306 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 307 DenseSet<VPPair> &PairableInstUserPairSet, 308 DenseMap<Value *, Value *> &ChosenPairs, 309 DenseMap<ValuePair, size_t> &Tree, 310 DenseSet<ValuePair> &PrunedTree, ValuePair J, 311 bool UseCycleCheck); 312 313 void buildInitialTreeFor( 314 std::multimap<Value *, Value *> &CandidatePairs, 315 DenseSet<ValuePair> &CandidatePairsSet, 316 std::vector<Value *> &PairableInsts, 317 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 318 DenseSet<ValuePair> &PairableInstUsers, 319 DenseMap<Value *, Value *> &ChosenPairs, 320 DenseMap<ValuePair, size_t> &Tree, ValuePair J); 321 322 void findBestTreeFor( 323 std::multimap<Value *, Value *> &CandidatePairs, 324 DenseSet<ValuePair> &CandidatePairsSet, 325 DenseMap<ValuePair, int> &CandidatePairCostSavings, 326 std::vector<Value *> &PairableInsts, 327 DenseSet<ValuePair> &FixedOrderPairs, 328 DenseMap<VPPair, unsigned> &PairConnectionTypes, 329 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 330 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 331 DenseSet<ValuePair> &PairableInstUsers, 332 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 333 DenseSet<VPPair> &PairableInstUserPairSet, 334 DenseMap<Value *, Value *> &ChosenPairs, 335 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 336 int &BestEffSize, VPIteratorPair ChoiceRange, 337 bool UseCycleCheck); 338 339 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, 340 Instruction *J, unsigned o); 341 342 void fillNewShuffleMask(LLVMContext& Context, Instruction *J, 343 unsigned MaskOffset, unsigned NumInElem, 344 unsigned NumInElem1, unsigned IdxOffset, 345 std::vector<Constant*> &Mask); 346 347 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I, 348 Instruction *J); 349 350 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, 351 unsigned o, Value *&LOp, unsigned numElemL, 352 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, 353 unsigned IdxOff = 0); 354 355 Value *getReplacementInput(LLVMContext& Context, Instruction *I, 356 Instruction *J, unsigned o, bool IBeforeJ); 357 358 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, 359 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, 360 bool IBeforeJ); 361 362 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 363 Instruction *J, Instruction *K, 364 Instruction *&InsertionPt, Instruction *&K1, 365 Instruction *&K2); 366 367 void collectPairLoadMoveSet(BasicBlock &BB, 368 DenseMap<Value *, Value *> &ChosenPairs, 369 std::multimap<Value *, Value *> &LoadMoveSet, 370 DenseSet<ValuePair> &LoadMoveSetPairs, 371 Instruction *I); 372 373 void collectLoadMoveSet(BasicBlock &BB, 374 std::vector<Value *> &PairableInsts, 375 DenseMap<Value *, Value *> &ChosenPairs, 376 std::multimap<Value *, Value *> &LoadMoveSet, 377 DenseSet<ValuePair> &LoadMoveSetPairs); 378 379 bool canMoveUsesOfIAfterJ(BasicBlock &BB, 380 DenseSet<ValuePair> &LoadMoveSetPairs, 381 Instruction *I, Instruction *J); 382 383 void moveUsesOfIAfterJ(BasicBlock &BB, 384 DenseSet<ValuePair> &LoadMoveSetPairs, 385 Instruction *&InsertionPt, 386 Instruction *I, Instruction *J); 387 388 void combineMetadata(Instruction *K, const Instruction *J); 389 390 bool vectorizeBB(BasicBlock &BB) { 391 if (!DT->isReachableFromEntry(&BB)) { 392 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << 393 " in " << BB.getParent()->getName() << "\n"); 394 return false; 395 } 396 397 DEBUG(if (TTI) dbgs() << "BBV: using target information\n"); 398 399 bool changed = false; 400 // Iterate a sufficient number of times to merge types of size 1 bit, 401 // then 2 bits, then 4, etc. up to half of the target vector width of the 402 // target vector register. 403 unsigned n = 1; 404 for (unsigned v = 2; 405 (TTI || v <= Config.VectorBits) && 406 (!Config.MaxIter || n <= Config.MaxIter); 407 v *= 2, ++n) { 408 DEBUG(dbgs() << "BBV: fusing loop #" << n << 409 " for " << BB.getName() << " in " << 410 BB.getParent()->getName() << "...\n"); 411 if (vectorizePairs(BB)) 412 changed = true; 413 else 414 break; 415 } 416 417 if (changed && !Pow2LenOnly) { 418 ++n; 419 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) { 420 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " << 421 n << " for " << BB.getName() << " in " << 422 BB.getParent()->getName() << "...\n"); 423 if (!vectorizePairs(BB, true)) break; 424 } 425 } 426 427 DEBUG(dbgs() << "BBV: done!\n"); 428 return changed; 429 } 430 431 virtual bool runOnBasicBlock(BasicBlock &BB) { 432 AA = &getAnalysis<AliasAnalysis>(); 433 DT = &getAnalysis<DominatorTree>(); 434 SE = &getAnalysis<ScalarEvolution>(); 435 TD = getAnalysisIfAvailable<DataLayout>(); 436 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>(); 437 438 return vectorizeBB(BB); 439 } 440 441 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 442 BasicBlockPass::getAnalysisUsage(AU); 443 AU.addRequired<AliasAnalysis>(); 444 AU.addRequired<DominatorTree>(); 445 AU.addRequired<ScalarEvolution>(); 446 AU.addRequired<TargetTransformInfo>(); 447 AU.addPreserved<AliasAnalysis>(); 448 AU.addPreserved<DominatorTree>(); 449 AU.addPreserved<ScalarEvolution>(); 450 AU.setPreservesCFG(); 451 } 452 453 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) { 454 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() && 455 "Cannot form vector from incompatible scalar types"); 456 Type *STy = ElemTy->getScalarType(); 457 458 unsigned numElem; 459 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) { 460 numElem = VTy->getNumElements(); 461 } else { 462 numElem = 1; 463 } 464 465 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) { 466 numElem += VTy->getNumElements(); 467 } else { 468 numElem += 1; 469 } 470 471 return VectorType::get(STy, numElem); 472 } 473 474 static inline void getInstructionTypes(Instruction *I, 475 Type *&T1, Type *&T2) { 476 if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 477 // For stores, it is the value type, not the pointer type that matters 478 // because the value is what will come from a vector register. 479 480 Value *IVal = SI->getValueOperand(); 481 T1 = IVal->getType(); 482 } else { 483 T1 = I->getType(); 484 } 485 486 if (CastInst *CI = dyn_cast<CastInst>(I)) 487 T2 = CI->getSrcTy(); 488 else 489 T2 = T1; 490 491 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 492 T2 = SI->getCondition()->getType(); 493 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { 494 T2 = SI->getOperand(0)->getType(); 495 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 496 T2 = CI->getOperand(0)->getType(); 497 } 498 } 499 500 // Returns the weight associated with the provided value. A chain of 501 // candidate pairs has a length given by the sum of the weights of its 502 // members (one weight per pair; the weight of each member of the pair 503 // is assumed to be the same). This length is then compared to the 504 // chain-length threshold to determine if a given chain is significant 505 // enough to be vectorized. The length is also used in comparing 506 // candidate chains where longer chains are considered to be better. 507 // Note: when this function returns 0, the resulting instructions are 508 // not actually fused. 509 inline size_t getDepthFactor(Value *V) { 510 // InsertElement and ExtractElement have a depth factor of zero. This is 511 // for two reasons: First, they cannot be usefully fused. Second, because 512 // the pass generates a lot of these, they can confuse the simple metric 513 // used to compare the trees in the next iteration. Thus, giving them a 514 // weight of zero allows the pass to essentially ignore them in 515 // subsequent iterations when looking for vectorization opportunities 516 // while still tracking dependency chains that flow through those 517 // instructions. 518 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V)) 519 return 0; 520 521 // Give a load or store half of the required depth so that load/store 522 // pairs will vectorize. 523 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V))) 524 return Config.ReqChainDepth/2; 525 526 return 1; 527 } 528 529 // Returns the cost of the provided instruction using TTI. 530 // This does not handle loads and stores. 531 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { 532 switch (Opcode) { 533 default: break; 534 case Instruction::GetElementPtr: 535 // We mark this instruction as zero-cost because scalar GEPs are usually 536 // lowered to the intruction addressing mode. At the moment we don't 537 // generate vector GEPs. 538 return 0; 539 case Instruction::Br: 540 return TTI->getCFInstrCost(Opcode); 541 case Instruction::PHI: 542 return 0; 543 case Instruction::Add: 544 case Instruction::FAdd: 545 case Instruction::Sub: 546 case Instruction::FSub: 547 case Instruction::Mul: 548 case Instruction::FMul: 549 case Instruction::UDiv: 550 case Instruction::SDiv: 551 case Instruction::FDiv: 552 case Instruction::URem: 553 case Instruction::SRem: 554 case Instruction::FRem: 555 case Instruction::Shl: 556 case Instruction::LShr: 557 case Instruction::AShr: 558 case Instruction::And: 559 case Instruction::Or: 560 case Instruction::Xor: 561 return TTI->getArithmeticInstrCost(Opcode, T1); 562 case Instruction::Select: 563 case Instruction::ICmp: 564 case Instruction::FCmp: 565 return TTI->getCmpSelInstrCost(Opcode, T1, T2); 566 case Instruction::ZExt: 567 case Instruction::SExt: 568 case Instruction::FPToUI: 569 case Instruction::FPToSI: 570 case Instruction::FPExt: 571 case Instruction::PtrToInt: 572 case Instruction::IntToPtr: 573 case Instruction::SIToFP: 574 case Instruction::UIToFP: 575 case Instruction::Trunc: 576 case Instruction::FPTrunc: 577 case Instruction::BitCast: 578 case Instruction::ShuffleVector: 579 return TTI->getCastInstrCost(Opcode, T1, T2); 580 } 581 582 return 1; 583 } 584 585 // This determines the relative offset of two loads or stores, returning 586 // true if the offset could be determined to be some constant value. 587 // For example, if OffsetInElmts == 1, then J accesses the memory directly 588 // after I; if OffsetInElmts == -1 then I accesses the memory 589 // directly after J. 590 bool getPairPtrInfo(Instruction *I, Instruction *J, 591 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, 592 unsigned &IAddressSpace, unsigned &JAddressSpace, 593 int64_t &OffsetInElmts, bool ComputeOffset = true) { 594 OffsetInElmts = 0; 595 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 596 LoadInst *LJ = cast<LoadInst>(J); 597 IPtr = LI->getPointerOperand(); 598 JPtr = LJ->getPointerOperand(); 599 IAlignment = LI->getAlignment(); 600 JAlignment = LJ->getAlignment(); 601 IAddressSpace = LI->getPointerAddressSpace(); 602 JAddressSpace = LJ->getPointerAddressSpace(); 603 } else { 604 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); 605 IPtr = SI->getPointerOperand(); 606 JPtr = SJ->getPointerOperand(); 607 IAlignment = SI->getAlignment(); 608 JAlignment = SJ->getAlignment(); 609 IAddressSpace = SI->getPointerAddressSpace(); 610 JAddressSpace = SJ->getPointerAddressSpace(); 611 } 612 613 if (!ComputeOffset) 614 return true; 615 616 const SCEV *IPtrSCEV = SE->getSCEV(IPtr); 617 const SCEV *JPtrSCEV = SE->getSCEV(JPtr); 618 619 // If this is a trivial offset, then we'll get something like 620 // 1*sizeof(type). With target data, which we need anyway, this will get 621 // constant folded into a number. 622 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV); 623 if (const SCEVConstant *ConstOffSCEV = 624 dyn_cast<SCEVConstant>(OffsetSCEV)) { 625 ConstantInt *IntOff = ConstOffSCEV->getValue(); 626 int64_t Offset = IntOff->getSExtValue(); 627 628 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType(); 629 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy); 630 631 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType(); 632 if (VTy != VTy2 && Offset < 0) { 633 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2); 634 OffsetInElmts = Offset/VTy2TSS; 635 return (abs64(Offset) % VTy2TSS) == 0; 636 } 637 638 OffsetInElmts = Offset/VTyTSS; 639 return (abs64(Offset) % VTyTSS) == 0; 640 } 641 642 return false; 643 } 644 645 // Returns true if the provided CallInst represents an intrinsic that can 646 // be vectorized. 647 bool isVectorizableIntrinsic(CallInst* I) { 648 Function *F = I->getCalledFunction(); 649 if (!F) return false; 650 651 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 652 if (!IID) return false; 653 654 switch(IID) { 655 default: 656 return false; 657 case Intrinsic::sqrt: 658 case Intrinsic::powi: 659 case Intrinsic::sin: 660 case Intrinsic::cos: 661 case Intrinsic::log: 662 case Intrinsic::log2: 663 case Intrinsic::log10: 664 case Intrinsic::exp: 665 case Intrinsic::exp2: 666 case Intrinsic::pow: 667 return Config.VectorizeMath; 668 case Intrinsic::fma: 669 case Intrinsic::fmuladd: 670 return Config.VectorizeFMA; 671 } 672 } 673 674 bool isPureIEChain(InsertElementInst *IE) { 675 InsertElementInst *IENext = IE; 676 do { 677 if (!isa<UndefValue>(IENext->getOperand(0)) && 678 !isa<InsertElementInst>(IENext->getOperand(0))) { 679 return false; 680 } 681 } while ((IENext = 682 dyn_cast<InsertElementInst>(IENext->getOperand(0)))); 683 684 return true; 685 } 686 }; 687 688 // This function implements one vectorization iteration on the provided 689 // basic block. It returns true if the block is changed. 690 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) { 691 bool ShouldContinue; 692 BasicBlock::iterator Start = BB.getFirstInsertionPt(); 693 694 std::vector<Value *> AllPairableInsts; 695 DenseMap<Value *, Value *> AllChosenPairs; 696 DenseSet<ValuePair> AllFixedOrderPairs; 697 DenseMap<VPPair, unsigned> AllPairConnectionTypes; 698 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; 699 700 do { 701 std::vector<Value *> PairableInsts; 702 std::multimap<Value *, Value *> CandidatePairs; 703 DenseSet<ValuePair> FixedOrderPairs; 704 DenseMap<ValuePair, int> CandidatePairCostSavings; 705 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, 706 FixedOrderPairs, 707 CandidatePairCostSavings, 708 PairableInsts, NonPow2Len); 709 if (PairableInsts.empty()) continue; 710 711 // Build the candidate pair set for faster lookups. 712 DenseSet<ValuePair> CandidatePairsSet; 713 for (std::multimap<Value *, Value *>::iterator I = CandidatePairs.begin(), 714 E = CandidatePairs.end(); I != E; ++I) 715 CandidatePairsSet.insert(*I); 716 717 // Now we have a map of all of the pairable instructions and we need to 718 // select the best possible pairing. A good pairing is one such that the 719 // users of the pair are also paired. This defines a (directed) forest 720 // over the pairs such that two pairs are connected iff the second pair 721 // uses the first. 722 723 // Note that it only matters that both members of the second pair use some 724 // element of the first pair (to allow for splatting). 725 726 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; 727 DenseMap<VPPair, unsigned> PairConnectionTypes; 728 computeConnectedPairs(CandidatePairs, CandidatePairsSet, 729 PairableInsts, ConnectedPairs, PairConnectionTypes); 730 if (ConnectedPairs.empty()) continue; 731 732 for (std::multimap<ValuePair, ValuePair>::iterator 733 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 734 I != IE; ++I) { 735 ConnectedPairDeps.insert(VPPair(I->second, I->first)); 736 } 737 738 // Build the pairable-instruction dependency map 739 DenseSet<ValuePair> PairableInstUsers; 740 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); 741 742 // There is now a graph of the connected pairs. For each variable, pick 743 // the pairing with the largest tree meeting the depth requirement on at 744 // least one branch. Then select all pairings that are part of that tree 745 // and remove them from the list of available pairings and pairable 746 // variables. 747 748 DenseMap<Value *, Value *> ChosenPairs; 749 choosePairs(CandidatePairs, CandidatePairsSet, 750 CandidatePairCostSavings, 751 PairableInsts, FixedOrderPairs, PairConnectionTypes, 752 ConnectedPairs, ConnectedPairDeps, 753 PairableInstUsers, ChosenPairs); 754 755 if (ChosenPairs.empty()) continue; 756 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), 757 PairableInsts.end()); 758 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); 759 760 // Only for the chosen pairs, propagate information on fixed-order pairs, 761 // pair connections, and their types to the data structures used by the 762 // pair fusion procedures. 763 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), 764 IE = ChosenPairs.end(); I != IE; ++I) { 765 if (FixedOrderPairs.count(*I)) 766 AllFixedOrderPairs.insert(*I); 767 else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) 768 AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); 769 770 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); 771 J != IE; ++J) { 772 DenseMap<VPPair, unsigned>::iterator K = 773 PairConnectionTypes.find(VPPair(*I, *J)); 774 if (K != PairConnectionTypes.end()) { 775 AllPairConnectionTypes.insert(*K); 776 } else { 777 K = PairConnectionTypes.find(VPPair(*J, *I)); 778 if (K != PairConnectionTypes.end()) 779 AllPairConnectionTypes.insert(*K); 780 } 781 } 782 } 783 784 for (std::multimap<ValuePair, ValuePair>::iterator 785 I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); 786 I != IE; ++I) { 787 if (AllPairConnectionTypes.count(*I)) { 788 AllConnectedPairs.insert(*I); 789 AllConnectedPairDeps.insert(VPPair(I->second, I->first)); 790 } 791 } 792 } while (ShouldContinue); 793 794 if (AllChosenPairs.empty()) return false; 795 NumFusedOps += AllChosenPairs.size(); 796 797 // A set of pairs has now been selected. It is now necessary to replace the 798 // paired instructions with vector instructions. For this procedure each 799 // operand must be replaced with a vector operand. This vector is formed 800 // by using build_vector on the old operands. The replaced values are then 801 // replaced with a vector_extract on the result. Subsequent optimization 802 // passes should coalesce the build/extract combinations. 803 804 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, 805 AllPairConnectionTypes, 806 AllConnectedPairs, AllConnectedPairDeps); 807 808 // It is important to cleanup here so that future iterations of this 809 // function have less work to do. 810 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); 811 return true; 812 } 813 814 // This function returns true if the provided instruction is capable of being 815 // fused into a vector instruction. This determination is based only on the 816 // type and other attributes of the instruction. 817 bool BBVectorize::isInstVectorizable(Instruction *I, 818 bool &IsSimpleLoadStore) { 819 IsSimpleLoadStore = false; 820 821 if (CallInst *C = dyn_cast<CallInst>(I)) { 822 if (!isVectorizableIntrinsic(C)) 823 return false; 824 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) { 825 // Vectorize simple loads if possbile: 826 IsSimpleLoadStore = L->isSimple(); 827 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 828 return false; 829 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { 830 // Vectorize simple stores if possbile: 831 IsSimpleLoadStore = S->isSimple(); 832 if (!IsSimpleLoadStore || !Config.VectorizeMemOps) 833 return false; 834 } else if (CastInst *C = dyn_cast<CastInst>(I)) { 835 // We can vectorize casts, but not casts of pointer types, etc. 836 if (!Config.VectorizeCasts) 837 return false; 838 839 Type *SrcTy = C->getSrcTy(); 840 if (!SrcTy->isSingleValueType()) 841 return false; 842 843 Type *DestTy = C->getDestTy(); 844 if (!DestTy->isSingleValueType()) 845 return false; 846 } else if (isa<SelectInst>(I)) { 847 if (!Config.VectorizeSelect) 848 return false; 849 } else if (isa<CmpInst>(I)) { 850 if (!Config.VectorizeCmp) 851 return false; 852 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) { 853 if (!Config.VectorizeGEP) 854 return false; 855 856 // Currently, vector GEPs exist only with one index. 857 if (G->getNumIndices() != 1) 858 return false; 859 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) || 860 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) { 861 return false; 862 } 863 864 // We can't vectorize memory operations without target data 865 if (TD == 0 && IsSimpleLoadStore) 866 return false; 867 868 Type *T1, *T2; 869 getInstructionTypes(I, T1, T2); 870 871 // Not every type can be vectorized... 872 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) || 873 !(VectorType::isValidElementType(T2) || T2->isVectorTy())) 874 return false; 875 876 if (T1->getScalarSizeInBits() == 1) { 877 if (!Config.VectorizeBools) 878 return false; 879 } else { 880 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) 881 return false; 882 } 883 884 if (T2->getScalarSizeInBits() == 1) { 885 if (!Config.VectorizeBools) 886 return false; 887 } else { 888 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) 889 return false; 890 } 891 892 if (!Config.VectorizeFloats 893 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) 894 return false; 895 896 // Don't vectorize target-specific types. 897 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy()) 898 return false; 899 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy()) 900 return false; 901 902 if ((!Config.VectorizePointers || TD == 0) && 903 (T1->getScalarType()->isPointerTy() || 904 T2->getScalarType()->isPointerTy())) 905 return false; 906 907 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || 908 T2->getPrimitiveSizeInBits() >= Config.VectorBits)) 909 return false; 910 911 return true; 912 } 913 914 // This function returns true if the two provided instructions are compatible 915 // (meaning that they can be fused into a vector instruction). This assumes 916 // that I has already been determined to be vectorizable and that J is not 917 // in the use tree of I. 918 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, 919 bool IsSimpleLoadStore, bool NonPow2Len, 920 int &CostSavings, int &FixedOrder) { 921 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << 922 " <-> " << *J << "\n"); 923 924 CostSavings = 0; 925 FixedOrder = 0; 926 927 // Loads and stores can be merged if they have different alignments, 928 // but are otherwise the same. 929 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | 930 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0))) 931 return false; 932 933 Type *IT1, *IT2, *JT1, *JT2; 934 getInstructionTypes(I, IT1, IT2); 935 getInstructionTypes(J, JT1, JT2); 936 unsigned MaxTypeBits = std::max( 937 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), 938 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); 939 if (!TTI && MaxTypeBits > Config.VectorBits) 940 return false; 941 942 // FIXME: handle addsub-type operations! 943 944 if (IsSimpleLoadStore) { 945 Value *IPtr, *JPtr; 946 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 947 int64_t OffsetInElmts = 0; 948 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 949 IAddressSpace, JAddressSpace, 950 OffsetInElmts) && abs64(OffsetInElmts) == 1) { 951 FixedOrder = (int) OffsetInElmts; 952 unsigned BottomAlignment = IAlignment; 953 if (OffsetInElmts < 0) BottomAlignment = JAlignment; 954 955 Type *aTypeI = isa<StoreInst>(I) ? 956 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); 957 Type *aTypeJ = isa<StoreInst>(J) ? 958 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); 959 Type *VType = getVecTypeForPair(aTypeI, aTypeJ); 960 961 if (Config.AlignedOnly) { 962 // An aligned load or store is possible only if the instruction 963 // with the lower offset has an alignment suitable for the 964 // vector type. 965 966 unsigned VecAlignment = TD->getPrefTypeAlignment(VType); 967 if (BottomAlignment < VecAlignment) 968 return false; 969 } 970 971 if (TTI) { 972 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI, 973 IAlignment, IAddressSpace); 974 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ, 975 JAlignment, JAddressSpace); 976 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType, 977 BottomAlignment, 978 IAddressSpace); 979 980 ICost += TTI->getAddressComputationCost(aTypeI); 981 JCost += TTI->getAddressComputationCost(aTypeJ); 982 VCost += TTI->getAddressComputationCost(VType); 983 984 if (VCost > ICost + JCost) 985 return false; 986 987 // We don't want to fuse to a type that will be split, even 988 // if the two input types will also be split and there is no other 989 // associated cost. 990 unsigned VParts = TTI->getNumberOfParts(VType); 991 if (VParts > 1) 992 return false; 993 else if (!VParts && VCost == ICost + JCost) 994 return false; 995 996 CostSavings = ICost + JCost - VCost; 997 } 998 } else { 999 return false; 1000 } 1001 } else if (TTI) { 1002 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); 1003 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); 1004 Type *VT1 = getVecTypeForPair(IT1, JT1), 1005 *VT2 = getVecTypeForPair(IT2, JT2); 1006 1007 // Note that this procedure is incorrect for insert and extract element 1008 // instructions (because combining these often results in a shuffle), 1009 // but this cost is ignored (because insert and extract element 1010 // instructions are assigned a zero depth factor and are not really 1011 // fused in general). 1012 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); 1013 1014 if (VCost > ICost + JCost) 1015 return false; 1016 1017 // We don't want to fuse to a type that will be split, even 1018 // if the two input types will also be split and there is no other 1019 // associated cost. 1020 unsigned VParts1 = TTI->getNumberOfParts(VT1), 1021 VParts2 = TTI->getNumberOfParts(VT2); 1022 if (VParts1 > 1 || VParts2 > 1) 1023 return false; 1024 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) 1025 return false; 1026 1027 CostSavings = ICost + JCost - VCost; 1028 } 1029 1030 // The powi intrinsic is special because only the first argument is 1031 // vectorized, the second arguments must be equal. 1032 CallInst *CI = dyn_cast<CallInst>(I); 1033 Function *FI; 1034 if (CI && (FI = CI->getCalledFunction())) { 1035 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID(); 1036 if (IID == Intrinsic::powi) { 1037 Value *A1I = CI->getArgOperand(1), 1038 *A1J = cast<CallInst>(J)->getArgOperand(1); 1039 const SCEV *A1ISCEV = SE->getSCEV(A1I), 1040 *A1JSCEV = SE->getSCEV(A1J); 1041 return (A1ISCEV == A1JSCEV); 1042 } 1043 1044 if (IID && TTI) { 1045 SmallVector<Type*, 4> Tys; 1046 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 1047 Tys.push_back(CI->getArgOperand(i)->getType()); 1048 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys); 1049 1050 Tys.clear(); 1051 CallInst *CJ = cast<CallInst>(J); 1052 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i) 1053 Tys.push_back(CJ->getArgOperand(i)->getType()); 1054 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys); 1055 1056 Tys.clear(); 1057 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() && 1058 "Intrinsic argument counts differ"); 1059 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1060 if (IID == Intrinsic::powi && i == 1) 1061 Tys.push_back(CI->getArgOperand(i)->getType()); 1062 else 1063 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(), 1064 CJ->getArgOperand(i)->getType())); 1065 } 1066 1067 Type *RetTy = getVecTypeForPair(IT1, JT1); 1068 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys); 1069 1070 if (VCost > ICost + JCost) 1071 return false; 1072 1073 // We don't want to fuse to a type that will be split, even 1074 // if the two input types will also be split and there is no other 1075 // associated cost. 1076 unsigned RetParts = TTI->getNumberOfParts(RetTy); 1077 if (RetParts > 1) 1078 return false; 1079 else if (!RetParts && VCost == ICost + JCost) 1080 return false; 1081 1082 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 1083 if (!Tys[i]->isVectorTy()) 1084 continue; 1085 1086 unsigned NumParts = TTI->getNumberOfParts(Tys[i]); 1087 if (NumParts > 1) 1088 return false; 1089 else if (!NumParts && VCost == ICost + JCost) 1090 return false; 1091 } 1092 1093 CostSavings = ICost + JCost - VCost; 1094 } 1095 } 1096 1097 return true; 1098 } 1099 1100 // Figure out whether or not J uses I and update the users and write-set 1101 // structures associated with I. Specifically, Users represents the set of 1102 // instructions that depend on I. WriteSet represents the set 1103 // of memory locations that are dependent on I. If UpdateUsers is true, 1104 // and J uses I, then Users is updated to contain J and WriteSet is updated 1105 // to contain any memory locations to which J writes. The function returns 1106 // true if J uses I. By default, alias analysis is used to determine 1107 // whether J reads from memory that overlaps with a location in WriteSet. 1108 // If LoadMoveSet is not null, then it is a previously-computed multimap 1109 // where the key is the memory-based user instruction and the value is 1110 // the instruction to be compared with I. So, if LoadMoveSet is provided, 1111 // then the alias analysis is not used. This is necessary because this 1112 // function is called during the process of moving instructions during 1113 // vectorization and the results of the alias analysis are not stable during 1114 // that process. 1115 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users, 1116 AliasSetTracker &WriteSet, Instruction *I, 1117 Instruction *J, bool UpdateUsers, 1118 DenseSet<ValuePair> *LoadMoveSetPairs) { 1119 bool UsesI = false; 1120 1121 // This instruction may already be marked as a user due, for example, to 1122 // being a member of a selected pair. 1123 if (Users.count(J)) 1124 UsesI = true; 1125 1126 if (!UsesI) 1127 for (User::op_iterator JU = J->op_begin(), JE = J->op_end(); 1128 JU != JE; ++JU) { 1129 Value *V = *JU; 1130 if (I == V || Users.count(V)) { 1131 UsesI = true; 1132 break; 1133 } 1134 } 1135 if (!UsesI && J->mayReadFromMemory()) { 1136 if (LoadMoveSetPairs) { 1137 UsesI = LoadMoveSetPairs->count(ValuePair(J, I)); 1138 } else { 1139 for (AliasSetTracker::iterator W = WriteSet.begin(), 1140 WE = WriteSet.end(); W != WE; ++W) { 1141 if (W->aliasesUnknownInst(J, *AA)) { 1142 UsesI = true; 1143 break; 1144 } 1145 } 1146 } 1147 } 1148 1149 if (UsesI && UpdateUsers) { 1150 if (J->mayWriteToMemory()) WriteSet.add(J); 1151 Users.insert(J); 1152 } 1153 1154 return UsesI; 1155 } 1156 1157 // This function iterates over all instruction pairs in the provided 1158 // basic block and collects all candidate pairs for vectorization. 1159 bool BBVectorize::getCandidatePairs(BasicBlock &BB, 1160 BasicBlock::iterator &Start, 1161 std::multimap<Value *, Value *> &CandidatePairs, 1162 DenseSet<ValuePair> &FixedOrderPairs, 1163 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1164 std::vector<Value *> &PairableInsts, bool NonPow2Len) { 1165 BasicBlock::iterator E = BB.end(); 1166 if (Start == E) return false; 1167 1168 bool ShouldContinue = false, IAfterStart = false; 1169 for (BasicBlock::iterator I = Start++; I != E; ++I) { 1170 if (I == Start) IAfterStart = true; 1171 1172 bool IsSimpleLoadStore; 1173 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue; 1174 1175 // Look for an instruction with which to pair instruction *I... 1176 DenseSet<Value *> Users; 1177 AliasSetTracker WriteSet(*AA); 1178 bool JAfterStart = IAfterStart; 1179 BasicBlock::iterator J = llvm::next(I); 1180 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) { 1181 if (J == Start) JAfterStart = true; 1182 1183 // Determine if J uses I, if so, exit the loop. 1184 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep); 1185 if (Config.FastDep) { 1186 // Note: For this heuristic to be effective, independent operations 1187 // must tend to be intermixed. This is likely to be true from some 1188 // kinds of grouped loop unrolling (but not the generic LLVM pass), 1189 // but otherwise may require some kind of reordering pass. 1190 1191 // When using fast dependency analysis, 1192 // stop searching after first use: 1193 if (UsesI) break; 1194 } else { 1195 if (UsesI) continue; 1196 } 1197 1198 // J does not use I, and comes before the first use of I, so it can be 1199 // merged with I if the instructions are compatible. 1200 int CostSavings, FixedOrder; 1201 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, 1202 CostSavings, FixedOrder)) continue; 1203 1204 // J is a candidate for merging with I. 1205 if (!PairableInsts.size() || 1206 PairableInsts[PairableInsts.size()-1] != I) { 1207 PairableInsts.push_back(I); 1208 } 1209 1210 CandidatePairs.insert(ValuePair(I, J)); 1211 if (TTI) 1212 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), 1213 CostSavings)); 1214 1215 if (FixedOrder == 1) 1216 FixedOrderPairs.insert(ValuePair(I, J)); 1217 else if (FixedOrder == -1) 1218 FixedOrderPairs.insert(ValuePair(J, I)); 1219 1220 // The next call to this function must start after the last instruction 1221 // selected during this invocation. 1222 if (JAfterStart) { 1223 Start = llvm::next(J); 1224 IAfterStart = JAfterStart = false; 1225 } 1226 1227 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " 1228 << *I << " <-> " << *J << " (cost savings: " << 1229 CostSavings << ")\n"); 1230 1231 // If we have already found too many pairs, break here and this function 1232 // will be called again starting after the last instruction selected 1233 // during this invocation. 1234 if (PairableInsts.size() >= Config.MaxInsts) { 1235 ShouldContinue = true; 1236 break; 1237 } 1238 } 1239 1240 if (ShouldContinue) 1241 break; 1242 } 1243 1244 DEBUG(dbgs() << "BBV: found " << PairableInsts.size() 1245 << " instructions with candidate pairs\n"); 1246 1247 return ShouldContinue; 1248 } 1249 1250 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that 1251 // it looks for pairs such that both members have an input which is an 1252 // output of PI or PJ. 1253 void BBVectorize::computePairsConnectedTo( 1254 std::multimap<Value *, Value *> &CandidatePairs, 1255 DenseSet<ValuePair> &CandidatePairsSet, 1256 std::vector<Value *> &PairableInsts, 1257 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1258 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1259 ValuePair P) { 1260 StoreInst *SI, *SJ; 1261 1262 // For each possible pairing for this variable, look at the uses of 1263 // the first value... 1264 for (Value::use_iterator I = P.first->use_begin(), 1265 E = P.first->use_end(); I != E; ++I) { 1266 if (isa<LoadInst>(*I)) { 1267 // A pair cannot be connected to a load because the load only takes one 1268 // operand (the address) and it is a scalar even after vectorization. 1269 continue; 1270 } else if ((SI = dyn_cast<StoreInst>(*I)) && 1271 P.first == SI->getPointerOperand()) { 1272 // Similarly, a pair cannot be connected to a store through its 1273 // pointer operand. 1274 continue; 1275 } 1276 1277 // For each use of the first variable, look for uses of the second 1278 // variable... 1279 for (Value::use_iterator J = P.second->use_begin(), 1280 E2 = P.second->use_end(); J != E2; ++J) { 1281 if ((SJ = dyn_cast<StoreInst>(*J)) && 1282 P.second == SJ->getPointerOperand()) 1283 continue; 1284 1285 // Look for <I, J>: 1286 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1287 VPPair VP(P, ValuePair(*I, *J)); 1288 ConnectedPairs.insert(VP); 1289 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); 1290 } 1291 1292 // Look for <J, I>: 1293 if (CandidatePairsSet.count(ValuePair(*J, *I))) { 1294 VPPair VP(P, ValuePair(*J, *I)); 1295 ConnectedPairs.insert(VP); 1296 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); 1297 } 1298 } 1299 1300 if (Config.SplatBreaksChain) continue; 1301 // Look for cases where just the first value in the pair is used by 1302 // both members of another pair (splatting). 1303 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) { 1304 if ((SJ = dyn_cast<StoreInst>(*J)) && 1305 P.first == SJ->getPointerOperand()) 1306 continue; 1307 1308 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1309 VPPair VP(P, ValuePair(*I, *J)); 1310 ConnectedPairs.insert(VP); 1311 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1312 } 1313 } 1314 } 1315 1316 if (Config.SplatBreaksChain) return; 1317 // Look for cases where just the second value in the pair is used by 1318 // both members of another pair (splatting). 1319 for (Value::use_iterator I = P.second->use_begin(), 1320 E = P.second->use_end(); I != E; ++I) { 1321 if (isa<LoadInst>(*I)) 1322 continue; 1323 else if ((SI = dyn_cast<StoreInst>(*I)) && 1324 P.second == SI->getPointerOperand()) 1325 continue; 1326 1327 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) { 1328 if ((SJ = dyn_cast<StoreInst>(*J)) && 1329 P.second == SJ->getPointerOperand()) 1330 continue; 1331 1332 if (CandidatePairsSet.count(ValuePair(*I, *J))) { 1333 VPPair VP(P, ValuePair(*I, *J)); 1334 ConnectedPairs.insert(VP); 1335 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); 1336 } 1337 } 1338 } 1339 } 1340 1341 // This function figures out which pairs are connected. Two pairs are 1342 // connected if some output of the first pair forms an input to both members 1343 // of the second pair. 1344 void BBVectorize::computeConnectedPairs( 1345 std::multimap<Value *, Value *> &CandidatePairs, 1346 DenseSet<ValuePair> &CandidatePairsSet, 1347 std::vector<Value *> &PairableInsts, 1348 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1349 DenseMap<VPPair, unsigned> &PairConnectionTypes) { 1350 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 1351 PE = PairableInsts.end(); PI != PE; ++PI) { 1352 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI); 1353 1354 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; 1355 P != choiceRange.second; ++P) 1356 computePairsConnectedTo(CandidatePairs, CandidatePairsSet, 1357 PairableInsts, ConnectedPairs, 1358 PairConnectionTypes, *P); 1359 } 1360 1361 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() 1362 << " pair connections.\n"); 1363 } 1364 1365 // This function builds a set of use tuples such that <A, B> is in the set 1366 // if B is in the use tree of A. If B is in the use tree of A, then B 1367 // depends on the output of A. 1368 void BBVectorize::buildDepMap( 1369 BasicBlock &BB, 1370 std::multimap<Value *, Value *> &CandidatePairs, 1371 std::vector<Value *> &PairableInsts, 1372 DenseSet<ValuePair> &PairableInstUsers) { 1373 DenseSet<Value *> IsInPair; 1374 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(), 1375 E = CandidatePairs.end(); C != E; ++C) { 1376 IsInPair.insert(C->first); 1377 IsInPair.insert(C->second); 1378 } 1379 1380 // Iterate through the basic block, recording all users of each 1381 // pairable instruction. 1382 1383 BasicBlock::iterator E = BB.end(); 1384 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) { 1385 if (IsInPair.find(I) == IsInPair.end()) continue; 1386 1387 DenseSet<Value *> Users; 1388 AliasSetTracker WriteSet(*AA); 1389 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) 1390 (void) trackUsesOfI(Users, WriteSet, I, J); 1391 1392 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end(); 1393 U != E; ++U) 1394 PairableInstUsers.insert(ValuePair(I, *U)); 1395 } 1396 } 1397 1398 // Returns true if an input to pair P is an output of pair Q and also an 1399 // input of pair Q is an output of pair P. If this is the case, then these 1400 // two pairs cannot be simultaneously fused. 1401 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1402 DenseSet<ValuePair> &PairableInstUsers, 1403 std::multimap<ValuePair, ValuePair> *PairableInstUserMap, 1404 DenseSet<VPPair> *PairableInstUserPairSet) { 1405 // Two pairs are in conflict if they are mutual Users of eachother. 1406 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1407 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1408 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1409 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1410 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1411 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1412 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1413 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1414 if (PairableInstUserMap) { 1415 // FIXME: The expensive part of the cycle check is not so much the cycle 1416 // check itself but this edge insertion procedure. This needs some 1417 // profiling and probably a different data structure (same is true of 1418 // most uses of std::multimap). 1419 if (PUsesQ) { 1420 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second) 1421 PairableInstUserMap->insert(VPPair(Q, P)); 1422 } 1423 if (QUsesP) { 1424 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second) 1425 PairableInstUserMap->insert(VPPair(P, Q)); 1426 } 1427 } 1428 1429 return (QUsesP && PUsesQ); 1430 } 1431 1432 // This function walks the use graph of current pairs to see if, starting 1433 // from P, the walk returns to P. 1434 bool BBVectorize::pairWillFormCycle(ValuePair P, 1435 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1436 DenseSet<ValuePair> &CurrentPairs) { 1437 DEBUG(if (DebugCycleCheck) 1438 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1439 << *P.second << "\n"); 1440 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1441 // contains non-direct associations. 1442 DenseSet<ValuePair> Visited; 1443 SmallVector<ValuePair, 32> Q; 1444 // General depth-first post-order traversal: 1445 Q.push_back(P); 1446 do { 1447 ValuePair QTop = Q.pop_back_val(); 1448 Visited.insert(QTop); 1449 1450 DEBUG(if (DebugCycleCheck) 1451 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1452 << *QTop.second << "\n"); 1453 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1454 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1455 C != QPairRange.second; ++C) { 1456 if (C->second == P) { 1457 DEBUG(dbgs() 1458 << "BBV: rejected to prevent non-trivial cycle formation: " 1459 << *C->first.first << " <-> " << *C->first.second << "\n"); 1460 return true; 1461 } 1462 1463 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1464 Q.push_back(C->second); 1465 } 1466 } while (!Q.empty()); 1467 1468 return false; 1469 } 1470 1471 // This function builds the initial tree of connected pairs with the 1472 // pair J at the root. 1473 void BBVectorize::buildInitialTreeFor( 1474 std::multimap<Value *, Value *> &CandidatePairs, 1475 DenseSet<ValuePair> &CandidatePairsSet, 1476 std::vector<Value *> &PairableInsts, 1477 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1478 DenseSet<ValuePair> &PairableInstUsers, 1479 DenseMap<Value *, Value *> &ChosenPairs, 1480 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1481 // Each of these pairs is viewed as the root node of a Tree. The Tree 1482 // is then walked (depth-first). As this happens, we keep track of 1483 // the pairs that compose the Tree and the maximum depth of the Tree. 1484 SmallVector<ValuePairWithDepth, 32> Q; 1485 // General depth-first post-order traversal: 1486 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1487 do { 1488 ValuePairWithDepth QTop = Q.back(); 1489 1490 // Push each child onto the queue: 1491 bool MoreChildren = false; 1492 size_t MaxChildDepth = QTop.second; 1493 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1494 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1495 k != qtRange.second; ++k) { 1496 // Make sure that this child pair is still a candidate: 1497 if (CandidatePairsSet.count(ValuePair(k->second))) { 1498 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1499 if (C == Tree.end()) { 1500 size_t d = getDepthFactor(k->second.first); 1501 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1502 MoreChildren = true; 1503 } else { 1504 MaxChildDepth = std::max(MaxChildDepth, C->second); 1505 } 1506 } 1507 } 1508 1509 if (!MoreChildren) { 1510 // Record the current pair as part of the Tree: 1511 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1512 Q.pop_back(); 1513 } 1514 } while (!Q.empty()); 1515 } 1516 1517 // Given some initial tree, prune it by removing conflicting pairs (pairs 1518 // that cannot be simultaneously chosen for vectorization). 1519 void BBVectorize::pruneTreeFor( 1520 std::multimap<Value *, Value *> &CandidatePairs, 1521 std::vector<Value *> &PairableInsts, 1522 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1523 DenseSet<ValuePair> &PairableInstUsers, 1524 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1525 DenseSet<VPPair> &PairableInstUserPairSet, 1526 DenseMap<Value *, Value *> &ChosenPairs, 1527 DenseMap<ValuePair, size_t> &Tree, 1528 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1529 bool UseCycleCheck) { 1530 SmallVector<ValuePairWithDepth, 32> Q; 1531 // General depth-first post-order traversal: 1532 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1533 do { 1534 ValuePairWithDepth QTop = Q.pop_back_val(); 1535 PrunedTree.insert(QTop.first); 1536 1537 // Visit each child, pruning as necessary... 1538 SmallVector<ValuePairWithDepth, 8> BestChildren; 1539 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1540 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1541 K != QTopRange.second; ++K) { 1542 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1543 if (C == Tree.end()) continue; 1544 1545 // This child is in the Tree, now we need to make sure it is the 1546 // best of any conflicting children. There could be multiple 1547 // conflicting children, so first, determine if we're keeping 1548 // this child, then delete conflicting children as necessary. 1549 1550 // It is also necessary to guard against pairing-induced 1551 // dependencies. Consider instructions a .. x .. y .. b 1552 // such that (a,b) are to be fused and (x,y) are to be fused 1553 // but a is an input to x and b is an output from y. This 1554 // means that y cannot be moved after b but x must be moved 1555 // after b for (a,b) to be fused. In other words, after 1556 // fusing (a,b) we have y .. a/b .. x where y is an input 1557 // to a/b and x is an output to a/b: x and y can no longer 1558 // be legally fused. To prevent this condition, we must 1559 // make sure that a child pair added to the Tree is not 1560 // both an input and output of an already-selected pair. 1561 1562 // Pairing-induced dependencies can also form from more complicated 1563 // cycles. The pair vs. pair conflicts are easy to check, and so 1564 // that is done explicitly for "fast rejection", and because for 1565 // child vs. child conflicts, we may prefer to keep the current 1566 // pair in preference to the already-selected child. 1567 DenseSet<ValuePair> CurrentPairs; 1568 1569 bool CanAdd = true; 1570 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1571 = BestChildren.begin(), E2 = BestChildren.end(); 1572 C2 != E2; ++C2) { 1573 if (C2->first.first == C->first.first || 1574 C2->first.first == C->first.second || 1575 C2->first.second == C->first.first || 1576 C2->first.second == C->first.second || 1577 pairsConflict(C2->first, C->first, PairableInstUsers, 1578 UseCycleCheck ? &PairableInstUserMap : 0, 1579 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1580 if (C2->second >= C->second) { 1581 CanAdd = false; 1582 break; 1583 } 1584 1585 CurrentPairs.insert(C2->first); 1586 } 1587 } 1588 if (!CanAdd) continue; 1589 1590 // Even worse, this child could conflict with another node already 1591 // selected for the Tree. If that is the case, ignore this child. 1592 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1593 E2 = PrunedTree.end(); T != E2; ++T) { 1594 if (T->first == C->first.first || 1595 T->first == C->first.second || 1596 T->second == C->first.first || 1597 T->second == C->first.second || 1598 pairsConflict(*T, C->first, PairableInstUsers, 1599 UseCycleCheck ? &PairableInstUserMap : 0, 1600 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1601 CanAdd = false; 1602 break; 1603 } 1604 1605 CurrentPairs.insert(*T); 1606 } 1607 if (!CanAdd) continue; 1608 1609 // And check the queue too... 1610 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1611 E2 = Q.end(); C2 != E2; ++C2) { 1612 if (C2->first.first == C->first.first || 1613 C2->first.first == C->first.second || 1614 C2->first.second == C->first.first || 1615 C2->first.second == C->first.second || 1616 pairsConflict(C2->first, C->first, PairableInstUsers, 1617 UseCycleCheck ? &PairableInstUserMap : 0, 1618 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1619 CanAdd = false; 1620 break; 1621 } 1622 1623 CurrentPairs.insert(C2->first); 1624 } 1625 if (!CanAdd) continue; 1626 1627 // Last but not least, check for a conflict with any of the 1628 // already-chosen pairs. 1629 for (DenseMap<Value *, Value *>::iterator C2 = 1630 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1631 C2 != E2; ++C2) { 1632 if (pairsConflict(*C2, C->first, PairableInstUsers, 1633 UseCycleCheck ? &PairableInstUserMap : 0, 1634 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1635 CanAdd = false; 1636 break; 1637 } 1638 1639 CurrentPairs.insert(*C2); 1640 } 1641 if (!CanAdd) continue; 1642 1643 // To check for non-trivial cycles formed by the addition of the 1644 // current pair we've formed a list of all relevant pairs, now use a 1645 // graph walk to check for a cycle. We start from the current pair and 1646 // walk the use tree to see if we again reach the current pair. If we 1647 // do, then the current pair is rejected. 1648 1649 // FIXME: It may be more efficient to use a topological-ordering 1650 // algorithm to improve the cycle check. This should be investigated. 1651 if (UseCycleCheck && 1652 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1653 continue; 1654 1655 // This child can be added, but we may have chosen it in preference 1656 // to an already-selected child. Check for this here, and if a 1657 // conflict is found, then remove the previously-selected child 1658 // before adding this one in its place. 1659 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1660 = BestChildren.begin(); C2 != BestChildren.end();) { 1661 if (C2->first.first == C->first.first || 1662 C2->first.first == C->first.second || 1663 C2->first.second == C->first.first || 1664 C2->first.second == C->first.second || 1665 pairsConflict(C2->first, C->first, PairableInstUsers)) 1666 C2 = BestChildren.erase(C2); 1667 else 1668 ++C2; 1669 } 1670 1671 BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); 1672 } 1673 1674 for (SmallVector<ValuePairWithDepth, 8>::iterator C 1675 = BestChildren.begin(), E2 = BestChildren.end(); 1676 C != E2; ++C) { 1677 size_t DepthF = getDepthFactor(C->first.first); 1678 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1679 } 1680 } while (!Q.empty()); 1681 } 1682 1683 // This function finds the best tree of mututally-compatible connected 1684 // pairs, given the choice of root pairs as an iterator range. 1685 void BBVectorize::findBestTreeFor( 1686 std::multimap<Value *, Value *> &CandidatePairs, 1687 DenseSet<ValuePair> &CandidatePairsSet, 1688 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1689 std::vector<Value *> &PairableInsts, 1690 DenseSet<ValuePair> &FixedOrderPairs, 1691 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1692 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1693 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1694 DenseSet<ValuePair> &PairableInstUsers, 1695 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1696 DenseSet<VPPair> &PairableInstUserPairSet, 1697 DenseMap<Value *, Value *> &ChosenPairs, 1698 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1699 int &BestEffSize, VPIteratorPair ChoiceRange, 1700 bool UseCycleCheck) { 1701 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1702 J != ChoiceRange.second; ++J) { 1703 1704 // Before going any further, make sure that this pair does not 1705 // conflict with any already-selected pairs (see comment below 1706 // near the Tree pruning for more details). 1707 DenseSet<ValuePair> ChosenPairSet; 1708 bool DoesConflict = false; 1709 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1710 E = ChosenPairs.end(); C != E; ++C) { 1711 if (pairsConflict(*C, *J, PairableInstUsers, 1712 UseCycleCheck ? &PairableInstUserMap : 0, 1713 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1714 DoesConflict = true; 1715 break; 1716 } 1717 1718 ChosenPairSet.insert(*C); 1719 } 1720 if (DoesConflict) continue; 1721 1722 if (UseCycleCheck && 1723 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1724 continue; 1725 1726 DenseMap<ValuePair, size_t> Tree; 1727 buildInitialTreeFor(CandidatePairs, CandidatePairsSet, 1728 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, 1749 PairableInstUserPairSet, 1750 ChosenPairs, Tree, PrunedTree, *J, UseCycleCheck); 1751 1752 int EffSize = 0; 1753 if (TTI) { 1754 DenseSet<Value *> PrunedTreeInstrs; 1755 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1756 E = PrunedTree.end(); S != E; ++S) { 1757 PrunedTreeInstrs.insert(S->first); 1758 PrunedTreeInstrs.insert(S->second); 1759 } 1760 1761 // The set of pairs that have already contributed to the total cost. 1762 DenseSet<ValuePair> IncomingPairs; 1763 1764 // If the cost model were perfect, this might not be necessary; but we 1765 // need to make sure that we don't get stuck vectorizing our own 1766 // shuffle chains. 1767 bool HasNontrivialInsts = false; 1768 1769 // The node weights represent the cost savings associated with 1770 // fusing the pair of instructions. 1771 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1772 E = PrunedTree.end(); S != E; ++S) { 1773 if (!isa<ShuffleVectorInst>(S->first) && 1774 !isa<InsertElementInst>(S->first) && 1775 !isa<ExtractElementInst>(S->first)) 1776 HasNontrivialInsts = true; 1777 1778 bool FlipOrder = false; 1779 1780 if (getDepthFactor(S->first)) { 1781 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1782 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1783 << *S->first << " <-> " << *S->second << "} = " << 1784 ESContrib << "\n"); 1785 EffSize += ESContrib; 1786 } 1787 1788 // The edge weights contribute in a negative sense: they represent 1789 // the cost of shuffles. 1790 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); 1791 if (IP.first != ConnectedPairDeps.end()) { 1792 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1793 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1794 Q != IP.second; ++Q) { 1795 if (!PrunedTree.count(Q->second)) 1796 continue; 1797 DenseMap<VPPair, unsigned>::iterator R = 1798 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1799 assert(R != PairConnectionTypes.end() && 1800 "Cannot find pair connection type"); 1801 if (R->second == PairConnectionDirect) 1802 ++NumDepsDirect; 1803 else if (R->second == PairConnectionSwap) 1804 ++NumDepsSwap; 1805 } 1806 1807 // If there are more swaps than direct connections, then 1808 // the pair order will be flipped during fusion. So the real 1809 // number of swaps is the minimum number. 1810 FlipOrder = !FixedOrderPairs.count(*S) && 1811 ((NumDepsSwap > NumDepsDirect) || 1812 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1813 1814 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1815 Q != IP.second; ++Q) { 1816 if (!PrunedTree.count(Q->second)) 1817 continue; 1818 DenseMap<VPPair, unsigned>::iterator R = 1819 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1820 assert(R != PairConnectionTypes.end() && 1821 "Cannot find pair connection type"); 1822 Type *Ty1 = Q->second.first->getType(), 1823 *Ty2 = Q->second.second->getType(); 1824 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1825 if ((R->second == PairConnectionDirect && FlipOrder) || 1826 (R->second == PairConnectionSwap && !FlipOrder) || 1827 R->second == PairConnectionSplat) { 1828 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1829 VTy, VTy); 1830 1831 if (VTy->getVectorNumElements() == 2) { 1832 if (R->second == PairConnectionSplat) 1833 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1834 TargetTransformInfo::SK_Broadcast, VTy)); 1835 else 1836 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1837 TargetTransformInfo::SK_Reverse, VTy)); 1838 } 1839 1840 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1841 *Q->second.first << " <-> " << *Q->second.second << 1842 "} -> {" << 1843 *S->first << " <-> " << *S->second << "} = " << 1844 ESContrib << "\n"); 1845 EffSize -= ESContrib; 1846 } 1847 } 1848 } 1849 1850 // Compute the cost of outgoing edges. We assume that edges outgoing 1851 // to shuffles, inserts or extracts can be merged, and so contribute 1852 // no additional cost. 1853 if (!S->first->getType()->isVoidTy()) { 1854 Type *Ty1 = S->first->getType(), 1855 *Ty2 = S->second->getType(); 1856 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1857 1858 bool NeedsExtraction = false; 1859 for (Value::use_iterator I = S->first->use_begin(), 1860 IE = S->first->use_end(); I != IE; ++I) { 1861 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1862 // Shuffle can be folded if it has no other input 1863 if (isa<UndefValue>(SI->getOperand(1))) 1864 continue; 1865 } 1866 if (isa<ExtractElementInst>(*I)) 1867 continue; 1868 if (PrunedTreeInstrs.count(*I)) 1869 continue; 1870 NeedsExtraction = true; 1871 break; 1872 } 1873 1874 if (NeedsExtraction) { 1875 int ESContrib; 1876 if (Ty1->isVectorTy()) { 1877 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1878 Ty1, VTy); 1879 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1880 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1)); 1881 } else 1882 ESContrib = (int) TTI->getVectorInstrCost( 1883 Instruction::ExtractElement, VTy, 0); 1884 1885 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1886 *S->first << "} = " << ESContrib << "\n"); 1887 EffSize -= ESContrib; 1888 } 1889 1890 NeedsExtraction = false; 1891 for (Value::use_iterator I = S->second->use_begin(), 1892 IE = S->second->use_end(); I != IE; ++I) { 1893 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1894 // Shuffle can be folded if it has no other input 1895 if (isa<UndefValue>(SI->getOperand(1))) 1896 continue; 1897 } 1898 if (isa<ExtractElementInst>(*I)) 1899 continue; 1900 if (PrunedTreeInstrs.count(*I)) 1901 continue; 1902 NeedsExtraction = true; 1903 break; 1904 } 1905 1906 if (NeedsExtraction) { 1907 int ESContrib; 1908 if (Ty2->isVectorTy()) { 1909 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1910 Ty2, VTy); 1911 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1912 TargetTransformInfo::SK_ExtractSubvector, VTy, 1913 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2)); 1914 } else 1915 ESContrib = (int) TTI->getVectorInstrCost( 1916 Instruction::ExtractElement, VTy, 1); 1917 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1918 *S->second << "} = " << ESContrib << "\n"); 1919 EffSize -= ESContrib; 1920 } 1921 } 1922 1923 // Compute the cost of incoming edges. 1924 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 1925 Instruction *S1 = cast<Instruction>(S->first), 1926 *S2 = cast<Instruction>(S->second); 1927 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 1928 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 1929 1930 // Combining constants into vector constants (or small vector 1931 // constants into larger ones are assumed free). 1932 if (isa<Constant>(O1) && isa<Constant>(O2)) 1933 continue; 1934 1935 if (FlipOrder) 1936 std::swap(O1, O2); 1937 1938 ValuePair VP = ValuePair(O1, O2); 1939 ValuePair VPR = ValuePair(O2, O1); 1940 1941 // Internal edges are not handled here. 1942 if (PrunedTree.count(VP) || PrunedTree.count(VPR)) 1943 continue; 1944 1945 Type *Ty1 = O1->getType(), 1946 *Ty2 = O2->getType(); 1947 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1948 1949 // Combining vector operations of the same type is also assumed 1950 // folded with other operations. 1951 if (Ty1 == Ty2) { 1952 // If both are insert elements, then both can be widened. 1953 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), 1954 *IEO2 = dyn_cast<InsertElementInst>(O2); 1955 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) 1956 continue; 1957 // If both are extract elements, and both have the same input 1958 // type, then they can be replaced with a shuffle 1959 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 1960 *EIO2 = dyn_cast<ExtractElementInst>(O2); 1961 if (EIO1 && EIO2 && 1962 EIO1->getOperand(0)->getType() == 1963 EIO2->getOperand(0)->getType()) 1964 continue; 1965 // If both are a shuffle with equal operand types and only two 1966 // unqiue operands, then they can be replaced with a single 1967 // shuffle 1968 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 1969 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 1970 if (SIO1 && SIO2 && 1971 SIO1->getOperand(0)->getType() == 1972 SIO2->getOperand(0)->getType()) { 1973 SmallSet<Value *, 4> SIOps; 1974 SIOps.insert(SIO1->getOperand(0)); 1975 SIOps.insert(SIO1->getOperand(1)); 1976 SIOps.insert(SIO2->getOperand(0)); 1977 SIOps.insert(SIO2->getOperand(1)); 1978 if (SIOps.size() <= 2) 1979 continue; 1980 } 1981 } 1982 1983 int ESContrib; 1984 // This pair has already been formed. 1985 if (IncomingPairs.count(VP)) { 1986 continue; 1987 } else if (IncomingPairs.count(VPR)) { 1988 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1989 VTy, VTy); 1990 1991 if (VTy->getVectorNumElements() == 2) 1992 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1993 TargetTransformInfo::SK_Reverse, VTy)); 1994 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 1995 ESContrib = (int) TTI->getVectorInstrCost( 1996 Instruction::InsertElement, VTy, 0); 1997 ESContrib += (int) TTI->getVectorInstrCost( 1998 Instruction::InsertElement, VTy, 1); 1999 } else if (!Ty1->isVectorTy()) { 2000 // O1 needs to be inserted into a vector of size O2, and then 2001 // both need to be shuffled together. 2002 ESContrib = (int) TTI->getVectorInstrCost( 2003 Instruction::InsertElement, Ty2, 0); 2004 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2005 VTy, Ty2); 2006 } else if (!Ty2->isVectorTy()) { 2007 // O2 needs to be inserted into a vector of size O1, and then 2008 // both need to be shuffled together. 2009 ESContrib = (int) TTI->getVectorInstrCost( 2010 Instruction::InsertElement, Ty1, 0); 2011 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2012 VTy, Ty1); 2013 } else { 2014 Type *TyBig = Ty1, *TySmall = Ty2; 2015 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 2016 std::swap(TyBig, TySmall); 2017 2018 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2019 VTy, TyBig); 2020 if (TyBig != TySmall) 2021 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2022 TyBig, TySmall); 2023 } 2024 2025 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 2026 << *O1 << " <-> " << *O2 << "} = " << 2027 ESContrib << "\n"); 2028 EffSize -= ESContrib; 2029 IncomingPairs.insert(VP); 2030 } 2031 } 2032 } 2033 2034 if (!HasNontrivialInsts) { 2035 DEBUG(if (DebugPairSelection) dbgs() << 2036 "\tNo non-trivial instructions in tree;" 2037 " override to zero effective size\n"); 2038 EffSize = 0; 2039 } 2040 } else { 2041 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 2042 E = PrunedTree.end(); S != E; ++S) 2043 EffSize += (int) getDepthFactor(S->first); 2044 } 2045 2046 DEBUG(if (DebugPairSelection) 2047 dbgs() << "BBV: found pruned Tree for pair {" 2048 << *J->first << " <-> " << *J->second << "} of depth " << 2049 MaxDepth << " and size " << PrunedTree.size() << 2050 " (effective size: " << EffSize << ")\n"); 2051 if (((TTI && !UseChainDepthWithTI) || 2052 MaxDepth >= Config.ReqChainDepth) && 2053 EffSize > 0 && EffSize > BestEffSize) { 2054 BestMaxDepth = MaxDepth; 2055 BestEffSize = EffSize; 2056 BestTree = PrunedTree; 2057 } 2058 } 2059 } 2060 2061 // Given the list of candidate pairs, this function selects those 2062 // that will be fused into vector instructions. 2063 void BBVectorize::choosePairs( 2064 std::multimap<Value *, Value *> &CandidatePairs, 2065 DenseSet<ValuePair> &CandidatePairsSet, 2066 DenseMap<ValuePair, int> &CandidatePairCostSavings, 2067 std::vector<Value *> &PairableInsts, 2068 DenseSet<ValuePair> &FixedOrderPairs, 2069 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2070 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2071 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 2072 DenseSet<ValuePair> &PairableInstUsers, 2073 DenseMap<Value *, Value *>& ChosenPairs) { 2074 bool UseCycleCheck = 2075 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 2076 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 2077 DenseSet<VPPair> PairableInstUserPairSet; 2078 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 2079 E = PairableInsts.end(); I != E; ++I) { 2080 // The number of possible pairings for this variable: 2081 size_t NumChoices = CandidatePairs.count(*I); 2082 if (!NumChoices) continue; 2083 2084 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 2085 2086 // The best pair to choose and its tree: 2087 size_t BestMaxDepth = 0; 2088 int BestEffSize = 0; 2089 DenseSet<ValuePair> BestTree; 2090 findBestTreeFor(CandidatePairs, CandidatePairsSet, 2091 CandidatePairCostSavings, 2092 PairableInsts, FixedOrderPairs, PairConnectionTypes, 2093 ConnectedPairs, ConnectedPairDeps, 2094 PairableInstUsers, PairableInstUserMap, 2095 PairableInstUserPairSet, ChosenPairs, 2096 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 2097 UseCycleCheck); 2098 2099 // A tree has been chosen (or not) at this point. If no tree was 2100 // chosen, then this instruction, I, cannot be paired (and is no longer 2101 // considered). 2102 2103 DEBUG(if (BestTree.size() > 0) 2104 dbgs() << "BBV: selected pairs in the best tree for: " 2105 << *cast<Instruction>(*I) << "\n"); 2106 2107 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 2108 SE2 = BestTree.end(); S != SE2; ++S) { 2109 // Insert the members of this tree into the list of chosen pairs. 2110 ChosenPairs.insert(ValuePair(S->first, S->second)); 2111 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 2112 *S->second << "\n"); 2113 2114 // Remove all candidate pairs that have values in the chosen tree. 2115 for (std::multimap<Value *, Value *>::iterator K = 2116 CandidatePairs.begin(); K != CandidatePairs.end();) { 2117 if (K->first == S->first || K->second == S->first || 2118 K->second == S->second || K->first == S->second) { 2119 // Don't remove the actual pair chosen so that it can be used 2120 // in subsequent tree selections. 2121 if (!(K->first == S->first && K->second == S->second)) { 2122 CandidatePairsSet.erase(*K); 2123 CandidatePairs.erase(K++); 2124 } else 2125 ++K; 2126 } else { 2127 ++K; 2128 } 2129 } 2130 } 2131 } 2132 2133 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2134 } 2135 2136 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2137 unsigned n = 0) { 2138 if (!I->hasName()) 2139 return ""; 2140 2141 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2142 (n > 0 ? "." + utostr(n) : "")).str(); 2143 } 2144 2145 // Returns the value that is to be used as the pointer input to the vector 2146 // instruction that fuses I with J. 2147 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2148 Instruction *I, Instruction *J, unsigned o) { 2149 Value *IPtr, *JPtr; 2150 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2151 int64_t OffsetInElmts; 2152 2153 // Note: the analysis might fail here, that is why the pair order has 2154 // been precomputed (OffsetInElmts must be unused here). 2155 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2156 IAddressSpace, JAddressSpace, 2157 OffsetInElmts, false); 2158 2159 // The pointer value is taken to be the one with the lowest offset. 2160 Value *VPtr = IPtr; 2161 2162 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 2163 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 2164 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2165 Type *VArgPtrType = PointerType::get(VArgType, 2166 cast<PointerType>(IPtr->getType())->getAddressSpace()); 2167 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2168 /* insert before */ I); 2169 } 2170 2171 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2172 unsigned MaskOffset, unsigned NumInElem, 2173 unsigned NumInElem1, unsigned IdxOffset, 2174 std::vector<Constant*> &Mask) { 2175 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 2176 for (unsigned v = 0; v < NumElem1; ++v) { 2177 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2178 if (m < 0) { 2179 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2180 } else { 2181 unsigned mm = m + (int) IdxOffset; 2182 if (m >= (int) NumInElem1) 2183 mm += (int) NumInElem; 2184 2185 Mask[v+MaskOffset] = 2186 ConstantInt::get(Type::getInt32Ty(Context), mm); 2187 } 2188 } 2189 } 2190 2191 // Returns the value that is to be used as the vector-shuffle mask to the 2192 // vector instruction that fuses I with J. 2193 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2194 Instruction *I, Instruction *J) { 2195 // This is the shuffle mask. We need to append the second 2196 // mask to the first, and the numbers need to be adjusted. 2197 2198 Type *ArgTypeI = I->getType(); 2199 Type *ArgTypeJ = J->getType(); 2200 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2201 2202 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 2203 2204 // Get the total number of elements in the fused vector type. 2205 // By definition, this must equal the number of elements in 2206 // the final mask. 2207 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 2208 std::vector<Constant*> Mask(NumElem); 2209 2210 Type *OpTypeI = I->getOperand(0)->getType(); 2211 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 2212 Type *OpTypeJ = J->getOperand(0)->getType(); 2213 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 2214 2215 // The fused vector will be: 2216 // ----------------------------------------------------- 2217 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2218 // ----------------------------------------------------- 2219 // from which we'll extract NumElem total elements (where the first NumElemI 2220 // of them come from the mask in I and the remainder come from the mask 2221 // in J. 2222 2223 // For the mask from the first pair... 2224 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2225 0, Mask); 2226 2227 // For the mask from the second pair... 2228 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2229 NumInElemI, Mask); 2230 2231 return ConstantVector::get(Mask); 2232 } 2233 2234 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2235 Instruction *J, unsigned o, Value *&LOp, 2236 unsigned numElemL, 2237 Type *ArgTypeL, Type *ArgTypeH, 2238 bool IBeforeJ, unsigned IdxOff) { 2239 bool ExpandedIEChain = false; 2240 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2241 // If we have a pure insertelement chain, then this can be rewritten 2242 // into a chain that directly builds the larger type. 2243 if (isPureIEChain(LIE)) { 2244 SmallVector<Value *, 8> VectElemts(numElemL, 2245 UndefValue::get(ArgTypeL->getScalarType())); 2246 InsertElementInst *LIENext = LIE; 2247 do { 2248 unsigned Idx = 2249 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2250 VectElemts[Idx] = LIENext->getOperand(1); 2251 } while ((LIENext = 2252 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2253 2254 LIENext = 0; 2255 Value *LIEPrev = UndefValue::get(ArgTypeH); 2256 for (unsigned i = 0; i < numElemL; ++i) { 2257 if (isa<UndefValue>(VectElemts[i])) continue; 2258 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2259 ConstantInt::get(Type::getInt32Ty(Context), 2260 i + IdxOff), 2261 getReplacementName(IBeforeJ ? I : J, 2262 true, o, i+1)); 2263 LIENext->insertBefore(IBeforeJ ? J : I); 2264 LIEPrev = LIENext; 2265 } 2266 2267 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2268 ExpandedIEChain = true; 2269 } 2270 } 2271 2272 return ExpandedIEChain; 2273 } 2274 2275 // Returns the value to be used as the specified operand of the vector 2276 // instruction that fuses I with J. 2277 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2278 Instruction *J, unsigned o, bool IBeforeJ) { 2279 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2280 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2281 2282 // Compute the fused vector type for this operand 2283 Type *ArgTypeI = I->getOperand(o)->getType(); 2284 Type *ArgTypeJ = J->getOperand(o)->getType(); 2285 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2286 2287 Instruction *L = I, *H = J; 2288 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2289 2290 unsigned numElemL; 2291 if (ArgTypeL->isVectorTy()) 2292 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 2293 else 2294 numElemL = 1; 2295 2296 unsigned numElemH; 2297 if (ArgTypeH->isVectorTy()) 2298 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 2299 else 2300 numElemH = 1; 2301 2302 Value *LOp = L->getOperand(o); 2303 Value *HOp = H->getOperand(o); 2304 unsigned numElem = VArgType->getNumElements(); 2305 2306 // First, we check if we can reuse the "original" vector outputs (if these 2307 // exist). We might need a shuffle. 2308 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2309 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2310 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2311 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2312 2313 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2314 // optimization. The input vectors to the shuffle might be a different 2315 // length from the shuffle outputs. Unfortunately, the replacement 2316 // shuffle mask has already been formed, and the mask entries are sensitive 2317 // to the sizes of the inputs. 2318 bool IsSizeChangeShuffle = 2319 isa<ShuffleVectorInst>(L) && 2320 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2321 2322 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2323 // We can have at most two unique vector inputs. 2324 bool CanUseInputs = true; 2325 Value *I1, *I2 = 0; 2326 if (LEE) { 2327 I1 = LEE->getOperand(0); 2328 } else { 2329 I1 = LSV->getOperand(0); 2330 I2 = LSV->getOperand(1); 2331 if (I2 == I1 || isa<UndefValue>(I2)) 2332 I2 = 0; 2333 } 2334 2335 if (HEE) { 2336 Value *I3 = HEE->getOperand(0); 2337 if (!I2 && I3 != I1) 2338 I2 = I3; 2339 else if (I3 != I1 && I3 != I2) 2340 CanUseInputs = false; 2341 } else { 2342 Value *I3 = HSV->getOperand(0); 2343 if (!I2 && I3 != I1) 2344 I2 = I3; 2345 else if (I3 != I1 && I3 != I2) 2346 CanUseInputs = false; 2347 2348 if (CanUseInputs) { 2349 Value *I4 = HSV->getOperand(1); 2350 if (!isa<UndefValue>(I4)) { 2351 if (!I2 && I4 != I1) 2352 I2 = I4; 2353 else if (I4 != I1 && I4 != I2) 2354 CanUseInputs = false; 2355 } 2356 } 2357 } 2358 2359 if (CanUseInputs) { 2360 unsigned LOpElem = 2361 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2362 ->getNumElements(); 2363 unsigned HOpElem = 2364 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2365 ->getNumElements(); 2366 2367 // We have one or two input vectors. We need to map each index of the 2368 // operands to the index of the original vector. 2369 SmallVector<std::pair<int, int>, 8> II(numElem); 2370 for (unsigned i = 0; i < numElemL; ++i) { 2371 int Idx, INum; 2372 if (LEE) { 2373 Idx = 2374 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2375 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2376 } else { 2377 Idx = LSV->getMaskValue(i); 2378 if (Idx < (int) LOpElem) { 2379 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2380 } else { 2381 Idx -= LOpElem; 2382 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2383 } 2384 } 2385 2386 II[i] = std::pair<int, int>(Idx, INum); 2387 } 2388 for (unsigned i = 0; i < numElemH; ++i) { 2389 int Idx, INum; 2390 if (HEE) { 2391 Idx = 2392 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2393 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2394 } else { 2395 Idx = HSV->getMaskValue(i); 2396 if (Idx < (int) HOpElem) { 2397 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2398 } else { 2399 Idx -= HOpElem; 2400 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2401 } 2402 } 2403 2404 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2405 } 2406 2407 // We now have an array which tells us from which index of which 2408 // input vector each element of the operand comes. 2409 VectorType *I1T = cast<VectorType>(I1->getType()); 2410 unsigned I1Elem = I1T->getNumElements(); 2411 2412 if (!I2) { 2413 // In this case there is only one underlying vector input. Check for 2414 // the trivial case where we can use the input directly. 2415 if (I1Elem == numElem) { 2416 bool ElemInOrder = true; 2417 for (unsigned i = 0; i < numElem; ++i) { 2418 if (II[i].first != (int) i && II[i].first != -1) { 2419 ElemInOrder = false; 2420 break; 2421 } 2422 } 2423 2424 if (ElemInOrder) 2425 return I1; 2426 } 2427 2428 // A shuffle is needed. 2429 std::vector<Constant *> Mask(numElem); 2430 for (unsigned i = 0; i < numElem; ++i) { 2431 int Idx = II[i].first; 2432 if (Idx == -1) 2433 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2434 else 2435 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2436 } 2437 2438 Instruction *S = 2439 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2440 ConstantVector::get(Mask), 2441 getReplacementName(IBeforeJ ? I : J, 2442 true, o)); 2443 S->insertBefore(IBeforeJ ? J : I); 2444 return S; 2445 } 2446 2447 VectorType *I2T = cast<VectorType>(I2->getType()); 2448 unsigned I2Elem = I2T->getNumElements(); 2449 2450 // This input comes from two distinct vectors. The first step is to 2451 // make sure that both vectors are the same length. If not, the 2452 // smaller one will need to grow before they can be shuffled together. 2453 if (I1Elem < I2Elem) { 2454 std::vector<Constant *> Mask(I2Elem); 2455 unsigned v = 0; 2456 for (; v < I1Elem; ++v) 2457 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2458 for (; v < I2Elem; ++v) 2459 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2460 2461 Instruction *NewI1 = 2462 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2463 ConstantVector::get(Mask), 2464 getReplacementName(IBeforeJ ? I : J, 2465 true, o, 1)); 2466 NewI1->insertBefore(IBeforeJ ? J : I); 2467 I1 = NewI1; 2468 I1T = I2T; 2469 I1Elem = I2Elem; 2470 } else if (I1Elem > I2Elem) { 2471 std::vector<Constant *> Mask(I1Elem); 2472 unsigned v = 0; 2473 for (; v < I2Elem; ++v) 2474 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2475 for (; v < I1Elem; ++v) 2476 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2477 2478 Instruction *NewI2 = 2479 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2480 ConstantVector::get(Mask), 2481 getReplacementName(IBeforeJ ? I : J, 2482 true, o, 1)); 2483 NewI2->insertBefore(IBeforeJ ? J : I); 2484 I2 = NewI2; 2485 I2T = I1T; 2486 I2Elem = I1Elem; 2487 } 2488 2489 // Now that both I1 and I2 are the same length we can shuffle them 2490 // together (and use the result). 2491 std::vector<Constant *> Mask(numElem); 2492 for (unsigned v = 0; v < numElem; ++v) { 2493 if (II[v].first == -1) { 2494 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2495 } else { 2496 int Idx = II[v].first + II[v].second * I1Elem; 2497 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2498 } 2499 } 2500 2501 Instruction *NewOp = 2502 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2503 getReplacementName(IBeforeJ ? I : J, true, o)); 2504 NewOp->insertBefore(IBeforeJ ? J : I); 2505 return NewOp; 2506 } 2507 } 2508 2509 Type *ArgType = ArgTypeL; 2510 if (numElemL < numElemH) { 2511 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2512 ArgTypeL, VArgType, IBeforeJ, 1)) { 2513 // This is another short-circuit case: we're combining a scalar into 2514 // a vector that is formed by an IE chain. We've just expanded the IE 2515 // chain, now insert the scalar and we're done. 2516 2517 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2518 getReplacementName(IBeforeJ ? I : J, true, o)); 2519 S->insertBefore(IBeforeJ ? J : I); 2520 return S; 2521 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2522 ArgTypeH, IBeforeJ)) { 2523 // The two vector inputs to the shuffle must be the same length, 2524 // so extend the smaller vector to be the same length as the larger one. 2525 Instruction *NLOp; 2526 if (numElemL > 1) { 2527 2528 std::vector<Constant *> Mask(numElemH); 2529 unsigned v = 0; 2530 for (; v < numElemL; ++v) 2531 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2532 for (; v < numElemH; ++v) 2533 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2534 2535 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2536 ConstantVector::get(Mask), 2537 getReplacementName(IBeforeJ ? I : J, 2538 true, o, 1)); 2539 } else { 2540 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2541 getReplacementName(IBeforeJ ? I : J, 2542 true, o, 1)); 2543 } 2544 2545 NLOp->insertBefore(IBeforeJ ? J : I); 2546 LOp = NLOp; 2547 } 2548 2549 ArgType = ArgTypeH; 2550 } else if (numElemL > numElemH) { 2551 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2552 ArgTypeH, VArgType, IBeforeJ)) { 2553 Instruction *S = 2554 InsertElementInst::Create(LOp, HOp, 2555 ConstantInt::get(Type::getInt32Ty(Context), 2556 numElemL), 2557 getReplacementName(IBeforeJ ? I : J, 2558 true, o)); 2559 S->insertBefore(IBeforeJ ? J : I); 2560 return S; 2561 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2562 ArgTypeL, IBeforeJ)) { 2563 Instruction *NHOp; 2564 if (numElemH > 1) { 2565 std::vector<Constant *> Mask(numElemL); 2566 unsigned v = 0; 2567 for (; v < numElemH; ++v) 2568 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2569 for (; v < numElemL; ++v) 2570 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2571 2572 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2573 ConstantVector::get(Mask), 2574 getReplacementName(IBeforeJ ? I : J, 2575 true, o, 1)); 2576 } else { 2577 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2578 getReplacementName(IBeforeJ ? I : J, 2579 true, o, 1)); 2580 } 2581 2582 NHOp->insertBefore(IBeforeJ ? J : I); 2583 HOp = NHOp; 2584 } 2585 } 2586 2587 if (ArgType->isVectorTy()) { 2588 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2589 std::vector<Constant*> Mask(numElem); 2590 for (unsigned v = 0; v < numElem; ++v) { 2591 unsigned Idx = v; 2592 // If the low vector was expanded, we need to skip the extra 2593 // undefined entries. 2594 if (v >= numElemL && numElemH > numElemL) 2595 Idx += (numElemH - numElemL); 2596 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2597 } 2598 2599 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2600 ConstantVector::get(Mask), 2601 getReplacementName(IBeforeJ ? I : J, true, o)); 2602 BV->insertBefore(IBeforeJ ? J : I); 2603 return BV; 2604 } 2605 2606 Instruction *BV1 = InsertElementInst::Create( 2607 UndefValue::get(VArgType), LOp, CV0, 2608 getReplacementName(IBeforeJ ? I : J, 2609 true, o, 1)); 2610 BV1->insertBefore(IBeforeJ ? J : I); 2611 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2612 getReplacementName(IBeforeJ ? I : J, 2613 true, o, 2)); 2614 BV2->insertBefore(IBeforeJ ? J : I); 2615 return BV2; 2616 } 2617 2618 // This function creates an array of values that will be used as the inputs 2619 // to the vector instruction that fuses I with J. 2620 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2621 Instruction *I, Instruction *J, 2622 SmallVector<Value *, 3> &ReplacedOperands, 2623 bool IBeforeJ) { 2624 unsigned NumOperands = I->getNumOperands(); 2625 2626 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2627 // Iterate backward so that we look at the store pointer 2628 // first and know whether or not we need to flip the inputs. 2629 2630 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2631 // This is the pointer for a load/store instruction. 2632 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2633 continue; 2634 } else if (isa<CallInst>(I)) { 2635 Function *F = cast<CallInst>(I)->getCalledFunction(); 2636 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 2637 if (o == NumOperands-1) { 2638 BasicBlock &BB = *I->getParent(); 2639 2640 Module *M = BB.getParent()->getParent(); 2641 Type *ArgTypeI = I->getType(); 2642 Type *ArgTypeJ = J->getType(); 2643 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2644 2645 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType); 2646 continue; 2647 } else if (IID == Intrinsic::powi && o == 1) { 2648 // The second argument of powi is a single integer and we've already 2649 // checked that both arguments are equal. As a result, we just keep 2650 // I's second argument. 2651 ReplacedOperands[o] = I->getOperand(o); 2652 continue; 2653 } 2654 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2655 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2656 continue; 2657 } 2658 2659 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2660 } 2661 } 2662 2663 // This function creates two values that represent the outputs of the 2664 // original I and J instructions. These are generally vector shuffles 2665 // or extracts. In many cases, these will end up being unused and, thus, 2666 // eliminated by later passes. 2667 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2668 Instruction *J, Instruction *K, 2669 Instruction *&InsertionPt, 2670 Instruction *&K1, Instruction *&K2) { 2671 if (isa<StoreInst>(I)) { 2672 AA->replaceWithNewValue(I, K); 2673 AA->replaceWithNewValue(J, K); 2674 } else { 2675 Type *IType = I->getType(); 2676 Type *JType = J->getType(); 2677 2678 VectorType *VType = getVecTypeForPair(IType, JType); 2679 unsigned numElem = VType->getNumElements(); 2680 2681 unsigned numElemI, numElemJ; 2682 if (IType->isVectorTy()) 2683 numElemI = cast<VectorType>(IType)->getNumElements(); 2684 else 2685 numElemI = 1; 2686 2687 if (JType->isVectorTy()) 2688 numElemJ = cast<VectorType>(JType)->getNumElements(); 2689 else 2690 numElemJ = 1; 2691 2692 if (IType->isVectorTy()) { 2693 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2694 for (unsigned v = 0; v < numElemI; ++v) { 2695 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2696 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2697 } 2698 2699 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2700 ConstantVector::get( Mask1), 2701 getReplacementName(K, false, 1)); 2702 } else { 2703 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2704 K1 = ExtractElementInst::Create(K, CV0, 2705 getReplacementName(K, false, 1)); 2706 } 2707 2708 if (JType->isVectorTy()) { 2709 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2710 for (unsigned v = 0; v < numElemJ; ++v) { 2711 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2712 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2713 } 2714 2715 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2716 ConstantVector::get( Mask2), 2717 getReplacementName(K, false, 2)); 2718 } else { 2719 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2720 K2 = ExtractElementInst::Create(K, CV1, 2721 getReplacementName(K, false, 2)); 2722 } 2723 2724 K1->insertAfter(K); 2725 K2->insertAfter(K1); 2726 InsertionPt = K2; 2727 } 2728 } 2729 2730 // Move all uses of the function I (including pairing-induced uses) after J. 2731 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2732 DenseSet<ValuePair> &LoadMoveSetPairs, 2733 Instruction *I, Instruction *J) { 2734 // Skip to the first instruction past I. 2735 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2736 2737 DenseSet<Value *> Users; 2738 AliasSetTracker WriteSet(*AA); 2739 for (; cast<Instruction>(L) != J; ++L) 2740 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs); 2741 2742 assert(cast<Instruction>(L) == J && 2743 "Tracking has not proceeded far enough to check for dependencies"); 2744 // If J is now in the use set of I, then trackUsesOfI will return true 2745 // and we have a dependency cycle (and the fusing operation must abort). 2746 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs); 2747 } 2748 2749 // Move all uses of the function I (including pairing-induced uses) after J. 2750 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2751 DenseSet<ValuePair> &LoadMoveSetPairs, 2752 Instruction *&InsertionPt, 2753 Instruction *I, Instruction *J) { 2754 // Skip to the first instruction past I. 2755 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2756 2757 DenseSet<Value *> Users; 2758 AliasSetTracker WriteSet(*AA); 2759 for (; cast<Instruction>(L) != J;) { 2760 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) { 2761 // Move this instruction 2762 Instruction *InstToMove = L; ++L; 2763 2764 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2765 " to after " << *InsertionPt << "\n"); 2766 InstToMove->removeFromParent(); 2767 InstToMove->insertAfter(InsertionPt); 2768 InsertionPt = InstToMove; 2769 } else { 2770 ++L; 2771 } 2772 } 2773 } 2774 2775 // Collect all load instruction that are in the move set of a given first 2776 // pair member. These loads depend on the first instruction, I, and so need 2777 // to be moved after J (the second instruction) when the pair is fused. 2778 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2779 DenseMap<Value *, Value *> &ChosenPairs, 2780 std::multimap<Value *, Value *> &LoadMoveSet, 2781 DenseSet<ValuePair> &LoadMoveSetPairs, 2782 Instruction *I) { 2783 // Skip to the first instruction past I. 2784 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2785 2786 DenseSet<Value *> Users; 2787 AliasSetTracker WriteSet(*AA); 2788 2789 // Note: We cannot end the loop when we reach J because J could be moved 2790 // farther down the use chain by another instruction pairing. Also, J 2791 // could be before I if this is an inverted input. 2792 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2793 if (trackUsesOfI(Users, WriteSet, I, L)) { 2794 if (L->mayReadFromMemory()) { 2795 LoadMoveSet.insert(ValuePair(L, I)); 2796 LoadMoveSetPairs.insert(ValuePair(L, I)); 2797 } 2798 } 2799 } 2800 } 2801 2802 // In cases where both load/stores and the computation of their pointers 2803 // are chosen for vectorization, we can end up in a situation where the 2804 // aliasing analysis starts returning different query results as the 2805 // process of fusing instruction pairs continues. Because the algorithm 2806 // relies on finding the same use trees here as were found earlier, we'll 2807 // need to precompute the necessary aliasing information here and then 2808 // manually update it during the fusion process. 2809 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2810 std::vector<Value *> &PairableInsts, 2811 DenseMap<Value *, Value *> &ChosenPairs, 2812 std::multimap<Value *, Value *> &LoadMoveSet, 2813 DenseSet<ValuePair> &LoadMoveSetPairs) { 2814 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2815 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2816 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2817 if (P == ChosenPairs.end()) continue; 2818 2819 Instruction *I = cast<Instruction>(P->first); 2820 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, 2821 LoadMoveSetPairs, I); 2822 } 2823 } 2824 2825 // When the first instruction in each pair is cloned, it will inherit its 2826 // parent's metadata. This metadata must be combined with that of the other 2827 // instruction in a safe way. 2828 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2829 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2830 K->getAllMetadataOtherThanDebugLoc(Metadata); 2831 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2832 unsigned Kind = Metadata[i].first; 2833 MDNode *JMD = J->getMetadata(Kind); 2834 MDNode *KMD = Metadata[i].second; 2835 2836 switch (Kind) { 2837 default: 2838 K->setMetadata(Kind, 0); // Remove unknown metadata 2839 break; 2840 case LLVMContext::MD_tbaa: 2841 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2842 break; 2843 case LLVMContext::MD_fpmath: 2844 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2845 break; 2846 } 2847 } 2848 } 2849 2850 // This function fuses the chosen instruction pairs into vector instructions, 2851 // taking care preserve any needed scalar outputs and, then, it reorders the 2852 // remaining instructions as needed (users of the first member of the pair 2853 // need to be moved to after the location of the second member of the pair 2854 // because the vector instruction is inserted in the location of the pair's 2855 // second member). 2856 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2857 std::vector<Value *> &PairableInsts, 2858 DenseMap<Value *, Value *> &ChosenPairs, 2859 DenseSet<ValuePair> &FixedOrderPairs, 2860 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2861 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2862 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2863 LLVMContext& Context = BB.getContext(); 2864 2865 // During the vectorization process, the order of the pairs to be fused 2866 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2867 // list. After a pair is fused, the flipped pair is removed from the list. 2868 DenseSet<ValuePair> FlippedPairs; 2869 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2870 E = ChosenPairs.end(); P != E; ++P) 2871 FlippedPairs.insert(ValuePair(P->second, P->first)); 2872 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2873 E = FlippedPairs.end(); P != E; ++P) 2874 ChosenPairs.insert(*P); 2875 2876 std::multimap<Value *, Value *> LoadMoveSet; 2877 DenseSet<ValuePair> LoadMoveSetPairs; 2878 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, 2879 LoadMoveSet, LoadMoveSetPairs); 2880 2881 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2882 2883 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2884 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2885 if (P == ChosenPairs.end()) { 2886 ++PI; 2887 continue; 2888 } 2889 2890 if (getDepthFactor(P->first) == 0) { 2891 // These instructions are not really fused, but are tracked as though 2892 // they are. Any case in which it would be interesting to fuse them 2893 // will be taken care of by InstCombine. 2894 --NumFusedOps; 2895 ++PI; 2896 continue; 2897 } 2898 2899 Instruction *I = cast<Instruction>(P->first), 2900 *J = cast<Instruction>(P->second); 2901 2902 DEBUG(dbgs() << "BBV: fusing: " << *I << 2903 " <-> " << *J << "\n"); 2904 2905 // Remove the pair and flipped pair from the list. 2906 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2907 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2908 ChosenPairs.erase(FP); 2909 ChosenPairs.erase(P); 2910 2911 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) { 2912 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2913 " <-> " << *J << 2914 " aborted because of non-trivial dependency cycle\n"); 2915 --NumFusedOps; 2916 ++PI; 2917 continue; 2918 } 2919 2920 // If the pair must have the other order, then flip it. 2921 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2922 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2923 // This pair does not have a fixed order, and so we might want to 2924 // flip it if that will yield fewer shuffles. We count the number 2925 // of dependencies connected via swaps, and those directly connected, 2926 // and flip the order if the number of swaps is greater. 2927 bool OrigOrder = true; 2928 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2929 if (IP.first == ConnectedPairDeps.end()) { 2930 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2931 OrigOrder = false; 2932 } 2933 2934 if (IP.first != ConnectedPairDeps.end()) { 2935 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2936 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2937 Q != IP.second; ++Q) { 2938 DenseMap<VPPair, unsigned>::iterator R = 2939 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2940 assert(R != PairConnectionTypes.end() && 2941 "Cannot find pair connection type"); 2942 if (R->second == PairConnectionDirect) 2943 ++NumDepsDirect; 2944 else if (R->second == PairConnectionSwap) 2945 ++NumDepsSwap; 2946 } 2947 2948 if (!OrigOrder) 2949 std::swap(NumDepsDirect, NumDepsSwap); 2950 2951 if (NumDepsSwap > NumDepsDirect) { 2952 FlipPairOrder = true; 2953 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2954 " <-> " << *J << "\n"); 2955 } 2956 } 2957 } 2958 2959 Instruction *L = I, *H = J; 2960 if (FlipPairOrder) 2961 std::swap(H, L); 2962 2963 // If the pair being fused uses the opposite order from that in the pair 2964 // connection map, then we need to flip the types. 2965 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2966 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2967 Q != IP.second; ++Q) { 2968 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2969 assert(R != PairConnectionTypes.end() && 2970 "Cannot find pair connection type"); 2971 if (R->second == PairConnectionDirect) 2972 R->second = PairConnectionSwap; 2973 else if (R->second == PairConnectionSwap) 2974 R->second = PairConnectionDirect; 2975 } 2976 2977 bool LBeforeH = !FlipPairOrder; 2978 unsigned NumOperands = I->getNumOperands(); 2979 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2980 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2981 LBeforeH); 2982 2983 // Make a copy of the original operation, change its type to the vector 2984 // type and replace its operands with the vector operands. 2985 Instruction *K = L->clone(); 2986 if (L->hasName()) 2987 K->takeName(L); 2988 else if (H->hasName()) 2989 K->takeName(H); 2990 2991 if (!isa<StoreInst>(K)) 2992 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 2993 2994 combineMetadata(K, H); 2995 K->intersectOptionalDataWith(H); 2996 2997 for (unsigned o = 0; o < NumOperands; ++o) 2998 K->setOperand(o, ReplacedOperands[o]); 2999 3000 K->insertAfter(J); 3001 3002 // Instruction insertion point: 3003 Instruction *InsertionPt = K; 3004 Instruction *K1 = 0, *K2 = 0; 3005 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 3006 3007 // The use tree of the first original instruction must be moved to after 3008 // the location of the second instruction. The entire use tree of the 3009 // first instruction is disjoint from the input tree of the second 3010 // (by definition), and so commutes with it. 3011 3012 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J); 3013 3014 if (!isa<StoreInst>(I)) { 3015 L->replaceAllUsesWith(K1); 3016 H->replaceAllUsesWith(K2); 3017 AA->replaceWithNewValue(L, K1); 3018 AA->replaceWithNewValue(H, K2); 3019 } 3020 3021 // Instructions that may read from memory may be in the load move set. 3022 // Once an instruction is fused, we no longer need its move set, and so 3023 // the values of the map never need to be updated. However, when a load 3024 // is fused, we need to merge the entries from both instructions in the 3025 // pair in case those instructions were in the move set of some other 3026 // yet-to-be-fused pair. The loads in question are the keys of the map. 3027 if (I->mayReadFromMemory()) { 3028 std::vector<ValuePair> NewSetMembers; 3029 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 3030 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 3031 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 3032 N != IPairRange.second; ++N) 3033 NewSetMembers.push_back(ValuePair(K, N->second)); 3034 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 3035 N != JPairRange.second; ++N) 3036 NewSetMembers.push_back(ValuePair(K, N->second)); 3037 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 3038 AE = NewSetMembers.end(); A != AE; ++A) { 3039 LoadMoveSet.insert(*A); 3040 LoadMoveSetPairs.insert(*A); 3041 } 3042 } 3043 3044 // Before removing I, set the iterator to the next instruction. 3045 PI = llvm::next(BasicBlock::iterator(I)); 3046 if (cast<Instruction>(PI) == J) 3047 ++PI; 3048 3049 SE->forgetValue(I); 3050 SE->forgetValue(J); 3051 I->eraseFromParent(); 3052 J->eraseFromParent(); 3053 3054 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 3055 BB << "\n"); 3056 } 3057 3058 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 3059 } 3060} 3061 3062char BBVectorize::ID = 0; 3063static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 3064INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3065INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3066INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3067INITIALIZE_PASS_DEPENDENCY(DominatorTree) 3068INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3069INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3070 3071BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 3072 return new BBVectorize(C); 3073} 3074 3075bool 3076llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 3077 BBVectorize BBVectorizer(P, C); 3078 return BBVectorizer.vectorizeBB(BB); 3079} 3080 3081//===----------------------------------------------------------------------===// 3082VectorizeConfig::VectorizeConfig() { 3083 VectorBits = ::VectorBits; 3084 VectorizeBools = !::NoBools; 3085 VectorizeInts = !::NoInts; 3086 VectorizeFloats = !::NoFloats; 3087 VectorizePointers = !::NoPointers; 3088 VectorizeCasts = !::NoCasts; 3089 VectorizeMath = !::NoMath; 3090 VectorizeFMA = !::NoFMA; 3091 VectorizeSelect = !::NoSelect; 3092 VectorizeCmp = !::NoCmp; 3093 VectorizeGEP = !::NoGEP; 3094 VectorizeMemOps = !::NoMemOps; 3095 AlignedOnly = ::AlignedOnly; 3096 ReqChainDepth= ::ReqChainDepth; 3097 SearchLimit = ::SearchLimit; 3098 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 3099 SplatBreaksChain = ::SplatBreaksChain; 3100 MaxInsts = ::MaxInsts; 3101 MaxIter = ::MaxIter; 3102 Pow2LenOnly = ::Pow2LenOnly; 3103 NoMemOpBoost = ::NoMemOpBoost; 3104 FastDep = ::FastDep; 3105} 3106