BBVectorize.cpp revision 8f3359a4b396d3f1a7b2726e02f199be74c62e4c
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 if (IsInPair.find(*U) == IsInPair.end()) continue; 1395 PairableInstUsers.insert(ValuePair(I, *U)); 1396 } 1397 } 1398 } 1399 1400 // Returns true if an input to pair P is an output of pair Q and also an 1401 // input of pair Q is an output of pair P. If this is the case, then these 1402 // two pairs cannot be simultaneously fused. 1403 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q, 1404 DenseSet<ValuePair> &PairableInstUsers, 1405 std::multimap<ValuePair, ValuePair> *PairableInstUserMap, 1406 DenseSet<VPPair> *PairableInstUserPairSet) { 1407 // Two pairs are in conflict if they are mutual Users of eachother. 1408 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) || 1409 PairableInstUsers.count(ValuePair(P.first, Q.second)) || 1410 PairableInstUsers.count(ValuePair(P.second, Q.first)) || 1411 PairableInstUsers.count(ValuePair(P.second, Q.second)); 1412 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) || 1413 PairableInstUsers.count(ValuePair(Q.first, P.second)) || 1414 PairableInstUsers.count(ValuePair(Q.second, P.first)) || 1415 PairableInstUsers.count(ValuePair(Q.second, P.second)); 1416 if (PairableInstUserMap) { 1417 // FIXME: The expensive part of the cycle check is not so much the cycle 1418 // check itself but this edge insertion procedure. This needs some 1419 // profiling and probably a different data structure (same is true of 1420 // most uses of std::multimap). 1421 if (PUsesQ) { 1422 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second) 1423 PairableInstUserMap->insert(VPPair(Q, P)); 1424 } 1425 if (QUsesP) { 1426 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second) 1427 PairableInstUserMap->insert(VPPair(P, Q)); 1428 } 1429 } 1430 1431 return (QUsesP && PUsesQ); 1432 } 1433 1434 // This function walks the use graph of current pairs to see if, starting 1435 // from P, the walk returns to P. 1436 bool BBVectorize::pairWillFormCycle(ValuePair P, 1437 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1438 DenseSet<ValuePair> &CurrentPairs) { 1439 DEBUG(if (DebugCycleCheck) 1440 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> " 1441 << *P.second << "\n"); 1442 // A lookup table of visisted pairs is kept because the PairableInstUserMap 1443 // contains non-direct associations. 1444 DenseSet<ValuePair> Visited; 1445 SmallVector<ValuePair, 32> Q; 1446 // General depth-first post-order traversal: 1447 Q.push_back(P); 1448 do { 1449 ValuePair QTop = Q.pop_back_val(); 1450 Visited.insert(QTop); 1451 1452 DEBUG(if (DebugCycleCheck) 1453 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> " 1454 << *QTop.second << "\n"); 1455 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop); 1456 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first; 1457 C != QPairRange.second; ++C) { 1458 if (C->second == P) { 1459 DEBUG(dbgs() 1460 << "BBV: rejected to prevent non-trivial cycle formation: " 1461 << *C->first.first << " <-> " << *C->first.second << "\n"); 1462 return true; 1463 } 1464 1465 if (CurrentPairs.count(C->second) && !Visited.count(C->second)) 1466 Q.push_back(C->second); 1467 } 1468 } while (!Q.empty()); 1469 1470 return false; 1471 } 1472 1473 // This function builds the initial tree of connected pairs with the 1474 // pair J at the root. 1475 void BBVectorize::buildInitialTreeFor( 1476 std::multimap<Value *, Value *> &CandidatePairs, 1477 DenseSet<ValuePair> &CandidatePairsSet, 1478 std::vector<Value *> &PairableInsts, 1479 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1480 DenseSet<ValuePair> &PairableInstUsers, 1481 DenseMap<Value *, Value *> &ChosenPairs, 1482 DenseMap<ValuePair, size_t> &Tree, ValuePair J) { 1483 // Each of these pairs is viewed as the root node of a Tree. The Tree 1484 // is then walked (depth-first). As this happens, we keep track of 1485 // the pairs that compose the Tree and the maximum depth of the Tree. 1486 SmallVector<ValuePairWithDepth, 32> Q; 1487 // General depth-first post-order traversal: 1488 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1489 do { 1490 ValuePairWithDepth QTop = Q.back(); 1491 1492 // Push each child onto the queue: 1493 bool MoreChildren = false; 1494 size_t MaxChildDepth = QTop.second; 1495 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first); 1496 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first; 1497 k != qtRange.second; ++k) { 1498 // Make sure that this child pair is still a candidate: 1499 if (CandidatePairsSet.count(ValuePair(k->second))) { 1500 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second); 1501 if (C == Tree.end()) { 1502 size_t d = getDepthFactor(k->second.first); 1503 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d)); 1504 MoreChildren = true; 1505 } else { 1506 MaxChildDepth = std::max(MaxChildDepth, C->second); 1507 } 1508 } 1509 } 1510 1511 if (!MoreChildren) { 1512 // Record the current pair as part of the Tree: 1513 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth)); 1514 Q.pop_back(); 1515 } 1516 } while (!Q.empty()); 1517 } 1518 1519 // Given some initial tree, prune it by removing conflicting pairs (pairs 1520 // that cannot be simultaneously chosen for vectorization). 1521 void BBVectorize::pruneTreeFor( 1522 std::multimap<Value *, Value *> &CandidatePairs, 1523 std::vector<Value *> &PairableInsts, 1524 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1525 DenseSet<ValuePair> &PairableInstUsers, 1526 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1527 DenseSet<VPPair> &PairableInstUserPairSet, 1528 DenseMap<Value *, Value *> &ChosenPairs, 1529 DenseMap<ValuePair, size_t> &Tree, 1530 DenseSet<ValuePair> &PrunedTree, ValuePair J, 1531 bool UseCycleCheck) { 1532 SmallVector<ValuePairWithDepth, 32> Q; 1533 // General depth-first post-order traversal: 1534 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first))); 1535 do { 1536 ValuePairWithDepth QTop = Q.pop_back_val(); 1537 PrunedTree.insert(QTop.first); 1538 1539 // Visit each child, pruning as necessary... 1540 SmallVector<ValuePairWithDepth, 8> BestChildren; 1541 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); 1542 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; 1543 K != QTopRange.second; ++K) { 1544 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second); 1545 if (C == Tree.end()) continue; 1546 1547 // This child is in the Tree, now we need to make sure it is the 1548 // best of any conflicting children. There could be multiple 1549 // conflicting children, so first, determine if we're keeping 1550 // this child, then delete conflicting children as necessary. 1551 1552 // It is also necessary to guard against pairing-induced 1553 // dependencies. Consider instructions a .. x .. y .. b 1554 // such that (a,b) are to be fused and (x,y) are to be fused 1555 // but a is an input to x and b is an output from y. This 1556 // means that y cannot be moved after b but x must be moved 1557 // after b for (a,b) to be fused. In other words, after 1558 // fusing (a,b) we have y .. a/b .. x where y is an input 1559 // to a/b and x is an output to a/b: x and y can no longer 1560 // be legally fused. To prevent this condition, we must 1561 // make sure that a child pair added to the Tree is not 1562 // both an input and output of an already-selected pair. 1563 1564 // Pairing-induced dependencies can also form from more complicated 1565 // cycles. The pair vs. pair conflicts are easy to check, and so 1566 // that is done explicitly for "fast rejection", and because for 1567 // child vs. child conflicts, we may prefer to keep the current 1568 // pair in preference to the already-selected child. 1569 DenseSet<ValuePair> CurrentPairs; 1570 1571 bool CanAdd = true; 1572 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1573 = BestChildren.begin(), E2 = BestChildren.end(); 1574 C2 != E2; ++C2) { 1575 if (C2->first.first == C->first.first || 1576 C2->first.first == C->first.second || 1577 C2->first.second == C->first.first || 1578 C2->first.second == C->first.second || 1579 pairsConflict(C2->first, C->first, PairableInstUsers, 1580 UseCycleCheck ? &PairableInstUserMap : 0, 1581 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1582 if (C2->second >= C->second) { 1583 CanAdd = false; 1584 break; 1585 } 1586 1587 CurrentPairs.insert(C2->first); 1588 } 1589 } 1590 if (!CanAdd) continue; 1591 1592 // Even worse, this child could conflict with another node already 1593 // selected for the Tree. If that is the case, ignore this child. 1594 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(), 1595 E2 = PrunedTree.end(); T != E2; ++T) { 1596 if (T->first == C->first.first || 1597 T->first == C->first.second || 1598 T->second == C->first.first || 1599 T->second == C->first.second || 1600 pairsConflict(*T, C->first, PairableInstUsers, 1601 UseCycleCheck ? &PairableInstUserMap : 0, 1602 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1603 CanAdd = false; 1604 break; 1605 } 1606 1607 CurrentPairs.insert(*T); 1608 } 1609 if (!CanAdd) continue; 1610 1611 // And check the queue too... 1612 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(), 1613 E2 = Q.end(); C2 != E2; ++C2) { 1614 if (C2->first.first == C->first.first || 1615 C2->first.first == C->first.second || 1616 C2->first.second == C->first.first || 1617 C2->first.second == C->first.second || 1618 pairsConflict(C2->first, C->first, PairableInstUsers, 1619 UseCycleCheck ? &PairableInstUserMap : 0, 1620 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1621 CanAdd = false; 1622 break; 1623 } 1624 1625 CurrentPairs.insert(C2->first); 1626 } 1627 if (!CanAdd) continue; 1628 1629 // Last but not least, check for a conflict with any of the 1630 // already-chosen pairs. 1631 for (DenseMap<Value *, Value *>::iterator C2 = 1632 ChosenPairs.begin(), E2 = ChosenPairs.end(); 1633 C2 != E2; ++C2) { 1634 if (pairsConflict(*C2, C->first, PairableInstUsers, 1635 UseCycleCheck ? &PairableInstUserMap : 0, 1636 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1637 CanAdd = false; 1638 break; 1639 } 1640 1641 CurrentPairs.insert(*C2); 1642 } 1643 if (!CanAdd) continue; 1644 1645 // To check for non-trivial cycles formed by the addition of the 1646 // current pair we've formed a list of all relevant pairs, now use a 1647 // graph walk to check for a cycle. We start from the current pair and 1648 // walk the use tree to see if we again reach the current pair. If we 1649 // do, then the current pair is rejected. 1650 1651 // FIXME: It may be more efficient to use a topological-ordering 1652 // algorithm to improve the cycle check. This should be investigated. 1653 if (UseCycleCheck && 1654 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs)) 1655 continue; 1656 1657 // This child can be added, but we may have chosen it in preference 1658 // to an already-selected child. Check for this here, and if a 1659 // conflict is found, then remove the previously-selected child 1660 // before adding this one in its place. 1661 for (SmallVector<ValuePairWithDepth, 8>::iterator C2 1662 = BestChildren.begin(); C2 != BestChildren.end();) { 1663 if (C2->first.first == C->first.first || 1664 C2->first.first == C->first.second || 1665 C2->first.second == C->first.first || 1666 C2->first.second == C->first.second || 1667 pairsConflict(C2->first, C->first, PairableInstUsers)) 1668 C2 = BestChildren.erase(C2); 1669 else 1670 ++C2; 1671 } 1672 1673 BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); 1674 } 1675 1676 for (SmallVector<ValuePairWithDepth, 8>::iterator C 1677 = BestChildren.begin(), E2 = BestChildren.end(); 1678 C != E2; ++C) { 1679 size_t DepthF = getDepthFactor(C->first.first); 1680 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF)); 1681 } 1682 } while (!Q.empty()); 1683 } 1684 1685 // This function finds the best tree of mututally-compatible connected 1686 // pairs, given the choice of root pairs as an iterator range. 1687 void BBVectorize::findBestTreeFor( 1688 std::multimap<Value *, Value *> &CandidatePairs, 1689 DenseSet<ValuePair> &CandidatePairsSet, 1690 DenseMap<ValuePair, int> &CandidatePairCostSavings, 1691 std::vector<Value *> &PairableInsts, 1692 DenseSet<ValuePair> &FixedOrderPairs, 1693 DenseMap<VPPair, unsigned> &PairConnectionTypes, 1694 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 1695 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 1696 DenseSet<ValuePair> &PairableInstUsers, 1697 std::multimap<ValuePair, ValuePair> &PairableInstUserMap, 1698 DenseSet<VPPair> &PairableInstUserPairSet, 1699 DenseMap<Value *, Value *> &ChosenPairs, 1700 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, 1701 int &BestEffSize, VPIteratorPair ChoiceRange, 1702 bool UseCycleCheck) { 1703 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; 1704 J != ChoiceRange.second; ++J) { 1705 1706 // Before going any further, make sure that this pair does not 1707 // conflict with any already-selected pairs (see comment below 1708 // near the Tree pruning for more details). 1709 DenseSet<ValuePair> ChosenPairSet; 1710 bool DoesConflict = false; 1711 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(), 1712 E = ChosenPairs.end(); C != E; ++C) { 1713 if (pairsConflict(*C, *J, PairableInstUsers, 1714 UseCycleCheck ? &PairableInstUserMap : 0, 1715 UseCycleCheck ? &PairableInstUserPairSet : 0)) { 1716 DoesConflict = true; 1717 break; 1718 } 1719 1720 ChosenPairSet.insert(*C); 1721 } 1722 if (DoesConflict) continue; 1723 1724 if (UseCycleCheck && 1725 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet)) 1726 continue; 1727 1728 DenseMap<ValuePair, size_t> Tree; 1729 buildInitialTreeFor(CandidatePairs, CandidatePairsSet, 1730 PairableInsts, ConnectedPairs, 1731 PairableInstUsers, ChosenPairs, Tree, *J); 1732 1733 // Because we'll keep the child with the largest depth, the largest 1734 // depth is still the same in the unpruned Tree. 1735 size_t MaxDepth = Tree.lookup(*J); 1736 1737 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {" 1738 << *J->first << " <-> " << *J->second << "} of depth " << 1739 MaxDepth << " and size " << Tree.size() << "\n"); 1740 1741 // At this point the Tree has been constructed, but, may contain 1742 // contradictory children (meaning that different children of 1743 // some tree node may be attempting to fuse the same instruction). 1744 // So now we walk the tree again, in the case of a conflict, 1745 // keep only the child with the largest depth. To break a tie, 1746 // favor the first child. 1747 1748 DenseSet<ValuePair> PrunedTree; 1749 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, 1750 PairableInstUsers, PairableInstUserMap, 1751 PairableInstUserPairSet, 1752 ChosenPairs, Tree, PrunedTree, *J, UseCycleCheck); 1753 1754 int EffSize = 0; 1755 if (TTI) { 1756 DenseSet<Value *> PrunedTreeInstrs; 1757 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1758 E = PrunedTree.end(); S != E; ++S) { 1759 PrunedTreeInstrs.insert(S->first); 1760 PrunedTreeInstrs.insert(S->second); 1761 } 1762 1763 // The set of pairs that have already contributed to the total cost. 1764 DenseSet<ValuePair> IncomingPairs; 1765 1766 // If the cost model were perfect, this might not be necessary; but we 1767 // need to make sure that we don't get stuck vectorizing our own 1768 // shuffle chains. 1769 bool HasNontrivialInsts = false; 1770 1771 // The node weights represent the cost savings associated with 1772 // fusing the pair of instructions. 1773 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 1774 E = PrunedTree.end(); S != E; ++S) { 1775 if (!isa<ShuffleVectorInst>(S->first) && 1776 !isa<InsertElementInst>(S->first) && 1777 !isa<ExtractElementInst>(S->first)) 1778 HasNontrivialInsts = true; 1779 1780 bool FlipOrder = false; 1781 1782 if (getDepthFactor(S->first)) { 1783 int ESContrib = CandidatePairCostSavings.find(*S)->second; 1784 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" 1785 << *S->first << " <-> " << *S->second << "} = " << 1786 ESContrib << "\n"); 1787 EffSize += ESContrib; 1788 } 1789 1790 // The edge weights contribute in a negative sense: they represent 1791 // the cost of shuffles. 1792 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); 1793 if (IP.first != ConnectedPairDeps.end()) { 1794 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 1795 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1796 Q != IP.second; ++Q) { 1797 if (!PrunedTree.count(Q->second)) 1798 continue; 1799 DenseMap<VPPair, unsigned>::iterator R = 1800 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1801 assert(R != PairConnectionTypes.end() && 1802 "Cannot find pair connection type"); 1803 if (R->second == PairConnectionDirect) 1804 ++NumDepsDirect; 1805 else if (R->second == PairConnectionSwap) 1806 ++NumDepsSwap; 1807 } 1808 1809 // If there are more swaps than direct connections, then 1810 // the pair order will be flipped during fusion. So the real 1811 // number of swaps is the minimum number. 1812 FlipOrder = !FixedOrderPairs.count(*S) && 1813 ((NumDepsSwap > NumDepsDirect) || 1814 FixedOrderPairs.count(ValuePair(S->second, S->first))); 1815 1816 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 1817 Q != IP.second; ++Q) { 1818 if (!PrunedTree.count(Q->second)) 1819 continue; 1820 DenseMap<VPPair, unsigned>::iterator R = 1821 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 1822 assert(R != PairConnectionTypes.end() && 1823 "Cannot find pair connection type"); 1824 Type *Ty1 = Q->second.first->getType(), 1825 *Ty2 = Q->second.second->getType(); 1826 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1827 if ((R->second == PairConnectionDirect && FlipOrder) || 1828 (R->second == PairConnectionSwap && !FlipOrder) || 1829 R->second == PairConnectionSplat) { 1830 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1831 VTy, VTy); 1832 1833 if (VTy->getVectorNumElements() == 2) { 1834 if (R->second == PairConnectionSplat) 1835 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1836 TargetTransformInfo::SK_Broadcast, VTy)); 1837 else 1838 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1839 TargetTransformInfo::SK_Reverse, VTy)); 1840 } 1841 1842 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1843 *Q->second.first << " <-> " << *Q->second.second << 1844 "} -> {" << 1845 *S->first << " <-> " << *S->second << "} = " << 1846 ESContrib << "\n"); 1847 EffSize -= ESContrib; 1848 } 1849 } 1850 } 1851 1852 // Compute the cost of outgoing edges. We assume that edges outgoing 1853 // to shuffles, inserts or extracts can be merged, and so contribute 1854 // no additional cost. 1855 if (!S->first->getType()->isVoidTy()) { 1856 Type *Ty1 = S->first->getType(), 1857 *Ty2 = S->second->getType(); 1858 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1859 1860 bool NeedsExtraction = false; 1861 for (Value::use_iterator I = S->first->use_begin(), 1862 IE = S->first->use_end(); I != IE; ++I) { 1863 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1864 // Shuffle can be folded if it has no other input 1865 if (isa<UndefValue>(SI->getOperand(1))) 1866 continue; 1867 } 1868 if (isa<ExtractElementInst>(*I)) 1869 continue; 1870 if (PrunedTreeInstrs.count(*I)) 1871 continue; 1872 NeedsExtraction = true; 1873 break; 1874 } 1875 1876 if (NeedsExtraction) { 1877 int ESContrib; 1878 if (Ty1->isVectorTy()) { 1879 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1880 Ty1, VTy); 1881 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1882 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1)); 1883 } else 1884 ESContrib = (int) TTI->getVectorInstrCost( 1885 Instruction::ExtractElement, VTy, 0); 1886 1887 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1888 *S->first << "} = " << ESContrib << "\n"); 1889 EffSize -= ESContrib; 1890 } 1891 1892 NeedsExtraction = false; 1893 for (Value::use_iterator I = S->second->use_begin(), 1894 IE = S->second->use_end(); I != IE; ++I) { 1895 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { 1896 // Shuffle can be folded if it has no other input 1897 if (isa<UndefValue>(SI->getOperand(1))) 1898 continue; 1899 } 1900 if (isa<ExtractElementInst>(*I)) 1901 continue; 1902 if (PrunedTreeInstrs.count(*I)) 1903 continue; 1904 NeedsExtraction = true; 1905 break; 1906 } 1907 1908 if (NeedsExtraction) { 1909 int ESContrib; 1910 if (Ty2->isVectorTy()) { 1911 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1912 Ty2, VTy); 1913 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1914 TargetTransformInfo::SK_ExtractSubvector, VTy, 1915 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2)); 1916 } else 1917 ESContrib = (int) TTI->getVectorInstrCost( 1918 Instruction::ExtractElement, VTy, 1); 1919 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << 1920 *S->second << "} = " << ESContrib << "\n"); 1921 EffSize -= ESContrib; 1922 } 1923 } 1924 1925 // Compute the cost of incoming edges. 1926 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { 1927 Instruction *S1 = cast<Instruction>(S->first), 1928 *S2 = cast<Instruction>(S->second); 1929 for (unsigned o = 0; o < S1->getNumOperands(); ++o) { 1930 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); 1931 1932 // Combining constants into vector constants (or small vector 1933 // constants into larger ones are assumed free). 1934 if (isa<Constant>(O1) && isa<Constant>(O2)) 1935 continue; 1936 1937 if (FlipOrder) 1938 std::swap(O1, O2); 1939 1940 ValuePair VP = ValuePair(O1, O2); 1941 ValuePair VPR = ValuePair(O2, O1); 1942 1943 // Internal edges are not handled here. 1944 if (PrunedTree.count(VP) || PrunedTree.count(VPR)) 1945 continue; 1946 1947 Type *Ty1 = O1->getType(), 1948 *Ty2 = O2->getType(); 1949 Type *VTy = getVecTypeForPair(Ty1, Ty2); 1950 1951 // Combining vector operations of the same type is also assumed 1952 // folded with other operations. 1953 if (Ty1 == Ty2) { 1954 // If both are insert elements, then both can be widened. 1955 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), 1956 *IEO2 = dyn_cast<InsertElementInst>(O2); 1957 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) 1958 continue; 1959 // If both are extract elements, and both have the same input 1960 // type, then they can be replaced with a shuffle 1961 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), 1962 *EIO2 = dyn_cast<ExtractElementInst>(O2); 1963 if (EIO1 && EIO2 && 1964 EIO1->getOperand(0)->getType() == 1965 EIO2->getOperand(0)->getType()) 1966 continue; 1967 // If both are a shuffle with equal operand types and only two 1968 // unqiue operands, then they can be replaced with a single 1969 // shuffle 1970 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), 1971 *SIO2 = dyn_cast<ShuffleVectorInst>(O2); 1972 if (SIO1 && SIO2 && 1973 SIO1->getOperand(0)->getType() == 1974 SIO2->getOperand(0)->getType()) { 1975 SmallSet<Value *, 4> SIOps; 1976 SIOps.insert(SIO1->getOperand(0)); 1977 SIOps.insert(SIO1->getOperand(1)); 1978 SIOps.insert(SIO2->getOperand(0)); 1979 SIOps.insert(SIO2->getOperand(1)); 1980 if (SIOps.size() <= 2) 1981 continue; 1982 } 1983 } 1984 1985 int ESContrib; 1986 // This pair has already been formed. 1987 if (IncomingPairs.count(VP)) { 1988 continue; 1989 } else if (IncomingPairs.count(VPR)) { 1990 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 1991 VTy, VTy); 1992 1993 if (VTy->getVectorNumElements() == 2) 1994 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost( 1995 TargetTransformInfo::SK_Reverse, VTy)); 1996 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { 1997 ESContrib = (int) TTI->getVectorInstrCost( 1998 Instruction::InsertElement, VTy, 0); 1999 ESContrib += (int) TTI->getVectorInstrCost( 2000 Instruction::InsertElement, VTy, 1); 2001 } else if (!Ty1->isVectorTy()) { 2002 // O1 needs to be inserted into a vector of size O2, and then 2003 // both need to be shuffled together. 2004 ESContrib = (int) TTI->getVectorInstrCost( 2005 Instruction::InsertElement, Ty2, 0); 2006 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2007 VTy, Ty2); 2008 } else if (!Ty2->isVectorTy()) { 2009 // O2 needs to be inserted into a vector of size O1, and then 2010 // both need to be shuffled together. 2011 ESContrib = (int) TTI->getVectorInstrCost( 2012 Instruction::InsertElement, Ty1, 0); 2013 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2014 VTy, Ty1); 2015 } else { 2016 Type *TyBig = Ty1, *TySmall = Ty2; 2017 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) 2018 std::swap(TyBig, TySmall); 2019 2020 ESContrib = (int) getInstrCost(Instruction::ShuffleVector, 2021 VTy, TyBig); 2022 if (TyBig != TySmall) 2023 ESContrib += (int) getInstrCost(Instruction::ShuffleVector, 2024 TyBig, TySmall); 2025 } 2026 2027 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" 2028 << *O1 << " <-> " << *O2 << "} = " << 2029 ESContrib << "\n"); 2030 EffSize -= ESContrib; 2031 IncomingPairs.insert(VP); 2032 } 2033 } 2034 } 2035 2036 if (!HasNontrivialInsts) { 2037 DEBUG(if (DebugPairSelection) dbgs() << 2038 "\tNo non-trivial instructions in tree;" 2039 " override to zero effective size\n"); 2040 EffSize = 0; 2041 } 2042 } else { 2043 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), 2044 E = PrunedTree.end(); S != E; ++S) 2045 EffSize += (int) getDepthFactor(S->first); 2046 } 2047 2048 DEBUG(if (DebugPairSelection) 2049 dbgs() << "BBV: found pruned Tree for pair {" 2050 << *J->first << " <-> " << *J->second << "} of depth " << 2051 MaxDepth << " and size " << PrunedTree.size() << 2052 " (effective size: " << EffSize << ")\n"); 2053 if (((TTI && !UseChainDepthWithTI) || 2054 MaxDepth >= Config.ReqChainDepth) && 2055 EffSize > 0 && EffSize > BestEffSize) { 2056 BestMaxDepth = MaxDepth; 2057 BestEffSize = EffSize; 2058 BestTree = PrunedTree; 2059 } 2060 } 2061 } 2062 2063 // Given the list of candidate pairs, this function selects those 2064 // that will be fused into vector instructions. 2065 void BBVectorize::choosePairs( 2066 std::multimap<Value *, Value *> &CandidatePairs, 2067 DenseSet<ValuePair> &CandidatePairsSet, 2068 DenseMap<ValuePair, int> &CandidatePairCostSavings, 2069 std::vector<Value *> &PairableInsts, 2070 DenseSet<ValuePair> &FixedOrderPairs, 2071 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2072 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2073 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, 2074 DenseSet<ValuePair> &PairableInstUsers, 2075 DenseMap<Value *, Value *>& ChosenPairs) { 2076 bool UseCycleCheck = 2077 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck; 2078 std::multimap<ValuePair, ValuePair> PairableInstUserMap; 2079 DenseSet<VPPair> PairableInstUserPairSet; 2080 for (std::vector<Value *>::iterator I = PairableInsts.begin(), 2081 E = PairableInsts.end(); I != E; ++I) { 2082 // The number of possible pairings for this variable: 2083 size_t NumChoices = CandidatePairs.count(*I); 2084 if (!NumChoices) continue; 2085 2086 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); 2087 2088 // The best pair to choose and its tree: 2089 size_t BestMaxDepth = 0; 2090 int BestEffSize = 0; 2091 DenseSet<ValuePair> BestTree; 2092 findBestTreeFor(CandidatePairs, CandidatePairsSet, 2093 CandidatePairCostSavings, 2094 PairableInsts, FixedOrderPairs, PairConnectionTypes, 2095 ConnectedPairs, ConnectedPairDeps, 2096 PairableInstUsers, PairableInstUserMap, 2097 PairableInstUserPairSet, ChosenPairs, 2098 BestTree, BestMaxDepth, BestEffSize, ChoiceRange, 2099 UseCycleCheck); 2100 2101 // A tree has been chosen (or not) at this point. If no tree was 2102 // chosen, then this instruction, I, cannot be paired (and is no longer 2103 // considered). 2104 2105 DEBUG(if (BestTree.size() > 0) 2106 dbgs() << "BBV: selected pairs in the best tree for: " 2107 << *cast<Instruction>(*I) << "\n"); 2108 2109 for (DenseSet<ValuePair>::iterator S = BestTree.begin(), 2110 SE2 = BestTree.end(); S != SE2; ++S) { 2111 // Insert the members of this tree into the list of chosen pairs. 2112 ChosenPairs.insert(ValuePair(S->first, S->second)); 2113 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " << 2114 *S->second << "\n"); 2115 2116 // Remove all candidate pairs that have values in the chosen tree. 2117 for (std::multimap<Value *, Value *>::iterator K = 2118 CandidatePairs.begin(); K != CandidatePairs.end();) { 2119 if (K->first == S->first || K->second == S->first || 2120 K->second == S->second || K->first == S->second) { 2121 // Don't remove the actual pair chosen so that it can be used 2122 // in subsequent tree selections. 2123 if (!(K->first == S->first && K->second == S->second)) { 2124 CandidatePairsSet.erase(*K); 2125 CandidatePairs.erase(K++); 2126 } else 2127 ++K; 2128 } else { 2129 ++K; 2130 } 2131 } 2132 } 2133 } 2134 2135 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n"); 2136 } 2137 2138 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o, 2139 unsigned n = 0) { 2140 if (!I->hasName()) 2141 return ""; 2142 2143 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) + 2144 (n > 0 ? "." + utostr(n) : "")).str(); 2145 } 2146 2147 // Returns the value that is to be used as the pointer input to the vector 2148 // instruction that fuses I with J. 2149 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, 2150 Instruction *I, Instruction *J, unsigned o) { 2151 Value *IPtr, *JPtr; 2152 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; 2153 int64_t OffsetInElmts; 2154 2155 // Note: the analysis might fail here, that is why the pair order has 2156 // been precomputed (OffsetInElmts must be unused here). 2157 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, 2158 IAddressSpace, JAddressSpace, 2159 OffsetInElmts, false); 2160 2161 // The pointer value is taken to be the one with the lowest offset. 2162 Value *VPtr = IPtr; 2163 2164 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); 2165 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); 2166 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2167 Type *VArgPtrType = PointerType::get(VArgType, 2168 cast<PointerType>(IPtr->getType())->getAddressSpace()); 2169 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), 2170 /* insert before */ I); 2171 } 2172 2173 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, 2174 unsigned MaskOffset, unsigned NumInElem, 2175 unsigned NumInElem1, unsigned IdxOffset, 2176 std::vector<Constant*> &Mask) { 2177 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements(); 2178 for (unsigned v = 0; v < NumElem1; ++v) { 2179 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v); 2180 if (m < 0) { 2181 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context)); 2182 } else { 2183 unsigned mm = m + (int) IdxOffset; 2184 if (m >= (int) NumInElem1) 2185 mm += (int) NumInElem; 2186 2187 Mask[v+MaskOffset] = 2188 ConstantInt::get(Type::getInt32Ty(Context), mm); 2189 } 2190 } 2191 } 2192 2193 // Returns the value that is to be used as the vector-shuffle mask to the 2194 // vector instruction that fuses I with J. 2195 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context, 2196 Instruction *I, Instruction *J) { 2197 // This is the shuffle mask. We need to append the second 2198 // mask to the first, and the numbers need to be adjusted. 2199 2200 Type *ArgTypeI = I->getType(); 2201 Type *ArgTypeJ = J->getType(); 2202 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2203 2204 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements(); 2205 2206 // Get the total number of elements in the fused vector type. 2207 // By definition, this must equal the number of elements in 2208 // the final mask. 2209 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements(); 2210 std::vector<Constant*> Mask(NumElem); 2211 2212 Type *OpTypeI = I->getOperand(0)->getType(); 2213 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements(); 2214 Type *OpTypeJ = J->getOperand(0)->getType(); 2215 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements(); 2216 2217 // The fused vector will be: 2218 // ----------------------------------------------------- 2219 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ | 2220 // ----------------------------------------------------- 2221 // from which we'll extract NumElem total elements (where the first NumElemI 2222 // of them come from the mask in I and the remainder come from the mask 2223 // in J. 2224 2225 // For the mask from the first pair... 2226 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI, 2227 0, Mask); 2228 2229 // For the mask from the second pair... 2230 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ, 2231 NumInElemI, Mask); 2232 2233 return ConstantVector::get(Mask); 2234 } 2235 2236 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I, 2237 Instruction *J, unsigned o, Value *&LOp, 2238 unsigned numElemL, 2239 Type *ArgTypeL, Type *ArgTypeH, 2240 bool IBeforeJ, unsigned IdxOff) { 2241 bool ExpandedIEChain = false; 2242 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { 2243 // If we have a pure insertelement chain, then this can be rewritten 2244 // into a chain that directly builds the larger type. 2245 if (isPureIEChain(LIE)) { 2246 SmallVector<Value *, 8> VectElemts(numElemL, 2247 UndefValue::get(ArgTypeL->getScalarType())); 2248 InsertElementInst *LIENext = LIE; 2249 do { 2250 unsigned Idx = 2251 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue(); 2252 VectElemts[Idx] = LIENext->getOperand(1); 2253 } while ((LIENext = 2254 dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); 2255 2256 LIENext = 0; 2257 Value *LIEPrev = UndefValue::get(ArgTypeH); 2258 for (unsigned i = 0; i < numElemL; ++i) { 2259 if (isa<UndefValue>(VectElemts[i])) continue; 2260 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], 2261 ConstantInt::get(Type::getInt32Ty(Context), 2262 i + IdxOff), 2263 getReplacementName(IBeforeJ ? I : J, 2264 true, o, i+1)); 2265 LIENext->insertBefore(IBeforeJ ? J : I); 2266 LIEPrev = LIENext; 2267 } 2268 2269 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH); 2270 ExpandedIEChain = true; 2271 } 2272 } 2273 2274 return ExpandedIEChain; 2275 } 2276 2277 // Returns the value to be used as the specified operand of the vector 2278 // instruction that fuses I with J. 2279 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, 2280 Instruction *J, unsigned o, bool IBeforeJ) { 2281 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2282 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); 2283 2284 // Compute the fused vector type for this operand 2285 Type *ArgTypeI = I->getOperand(o)->getType(); 2286 Type *ArgTypeJ = J->getOperand(o)->getType(); 2287 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2288 2289 Instruction *L = I, *H = J; 2290 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; 2291 2292 unsigned numElemL; 2293 if (ArgTypeL->isVectorTy()) 2294 numElemL = cast<VectorType>(ArgTypeL)->getNumElements(); 2295 else 2296 numElemL = 1; 2297 2298 unsigned numElemH; 2299 if (ArgTypeH->isVectorTy()) 2300 numElemH = cast<VectorType>(ArgTypeH)->getNumElements(); 2301 else 2302 numElemH = 1; 2303 2304 Value *LOp = L->getOperand(o); 2305 Value *HOp = H->getOperand(o); 2306 unsigned numElem = VArgType->getNumElements(); 2307 2308 // First, we check if we can reuse the "original" vector outputs (if these 2309 // exist). We might need a shuffle. 2310 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp); 2311 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp); 2312 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp); 2313 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp); 2314 2315 // FIXME: If we're fusing shuffle instructions, then we can't apply this 2316 // optimization. The input vectors to the shuffle might be a different 2317 // length from the shuffle outputs. Unfortunately, the replacement 2318 // shuffle mask has already been formed, and the mask entries are sensitive 2319 // to the sizes of the inputs. 2320 bool IsSizeChangeShuffle = 2321 isa<ShuffleVectorInst>(L) && 2322 (LOp->getType() != L->getType() || HOp->getType() != H->getType()); 2323 2324 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) { 2325 // We can have at most two unique vector inputs. 2326 bool CanUseInputs = true; 2327 Value *I1, *I2 = 0; 2328 if (LEE) { 2329 I1 = LEE->getOperand(0); 2330 } else { 2331 I1 = LSV->getOperand(0); 2332 I2 = LSV->getOperand(1); 2333 if (I2 == I1 || isa<UndefValue>(I2)) 2334 I2 = 0; 2335 } 2336 2337 if (HEE) { 2338 Value *I3 = HEE->getOperand(0); 2339 if (!I2 && I3 != I1) 2340 I2 = I3; 2341 else if (I3 != I1 && I3 != I2) 2342 CanUseInputs = false; 2343 } else { 2344 Value *I3 = HSV->getOperand(0); 2345 if (!I2 && I3 != I1) 2346 I2 = I3; 2347 else if (I3 != I1 && I3 != I2) 2348 CanUseInputs = false; 2349 2350 if (CanUseInputs) { 2351 Value *I4 = HSV->getOperand(1); 2352 if (!isa<UndefValue>(I4)) { 2353 if (!I2 && I4 != I1) 2354 I2 = I4; 2355 else if (I4 != I1 && I4 != I2) 2356 CanUseInputs = false; 2357 } 2358 } 2359 } 2360 2361 if (CanUseInputs) { 2362 unsigned LOpElem = 2363 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType()) 2364 ->getNumElements(); 2365 unsigned HOpElem = 2366 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType()) 2367 ->getNumElements(); 2368 2369 // We have one or two input vectors. We need to map each index of the 2370 // operands to the index of the original vector. 2371 SmallVector<std::pair<int, int>, 8> II(numElem); 2372 for (unsigned i = 0; i < numElemL; ++i) { 2373 int Idx, INum; 2374 if (LEE) { 2375 Idx = 2376 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue(); 2377 INum = LEE->getOperand(0) == I1 ? 0 : 1; 2378 } else { 2379 Idx = LSV->getMaskValue(i); 2380 if (Idx < (int) LOpElem) { 2381 INum = LSV->getOperand(0) == I1 ? 0 : 1; 2382 } else { 2383 Idx -= LOpElem; 2384 INum = LSV->getOperand(1) == I1 ? 0 : 1; 2385 } 2386 } 2387 2388 II[i] = std::pair<int, int>(Idx, INum); 2389 } 2390 for (unsigned i = 0; i < numElemH; ++i) { 2391 int Idx, INum; 2392 if (HEE) { 2393 Idx = 2394 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue(); 2395 INum = HEE->getOperand(0) == I1 ? 0 : 1; 2396 } else { 2397 Idx = HSV->getMaskValue(i); 2398 if (Idx < (int) HOpElem) { 2399 INum = HSV->getOperand(0) == I1 ? 0 : 1; 2400 } else { 2401 Idx -= HOpElem; 2402 INum = HSV->getOperand(1) == I1 ? 0 : 1; 2403 } 2404 } 2405 2406 II[i + numElemL] = std::pair<int, int>(Idx, INum); 2407 } 2408 2409 // We now have an array which tells us from which index of which 2410 // input vector each element of the operand comes. 2411 VectorType *I1T = cast<VectorType>(I1->getType()); 2412 unsigned I1Elem = I1T->getNumElements(); 2413 2414 if (!I2) { 2415 // In this case there is only one underlying vector input. Check for 2416 // the trivial case where we can use the input directly. 2417 if (I1Elem == numElem) { 2418 bool ElemInOrder = true; 2419 for (unsigned i = 0; i < numElem; ++i) { 2420 if (II[i].first != (int) i && II[i].first != -1) { 2421 ElemInOrder = false; 2422 break; 2423 } 2424 } 2425 2426 if (ElemInOrder) 2427 return I1; 2428 } 2429 2430 // A shuffle is needed. 2431 std::vector<Constant *> Mask(numElem); 2432 for (unsigned i = 0; i < numElem; ++i) { 2433 int Idx = II[i].first; 2434 if (Idx == -1) 2435 Mask[i] = UndefValue::get(Type::getInt32Ty(Context)); 2436 else 2437 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2438 } 2439 2440 Instruction *S = 2441 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2442 ConstantVector::get(Mask), 2443 getReplacementName(IBeforeJ ? I : J, 2444 true, o)); 2445 S->insertBefore(IBeforeJ ? J : I); 2446 return S; 2447 } 2448 2449 VectorType *I2T = cast<VectorType>(I2->getType()); 2450 unsigned I2Elem = I2T->getNumElements(); 2451 2452 // This input comes from two distinct vectors. The first step is to 2453 // make sure that both vectors are the same length. If not, the 2454 // smaller one will need to grow before they can be shuffled together. 2455 if (I1Elem < I2Elem) { 2456 std::vector<Constant *> Mask(I2Elem); 2457 unsigned v = 0; 2458 for (; v < I1Elem; ++v) 2459 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2460 for (; v < I2Elem; ++v) 2461 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2462 2463 Instruction *NewI1 = 2464 new ShuffleVectorInst(I1, UndefValue::get(I1T), 2465 ConstantVector::get(Mask), 2466 getReplacementName(IBeforeJ ? I : J, 2467 true, o, 1)); 2468 NewI1->insertBefore(IBeforeJ ? J : I); 2469 I1 = NewI1; 2470 I1T = I2T; 2471 I1Elem = I2Elem; 2472 } else if (I1Elem > I2Elem) { 2473 std::vector<Constant *> Mask(I1Elem); 2474 unsigned v = 0; 2475 for (; v < I2Elem; ++v) 2476 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2477 for (; v < I1Elem; ++v) 2478 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2479 2480 Instruction *NewI2 = 2481 new ShuffleVectorInst(I2, UndefValue::get(I2T), 2482 ConstantVector::get(Mask), 2483 getReplacementName(IBeforeJ ? I : J, 2484 true, o, 1)); 2485 NewI2->insertBefore(IBeforeJ ? J : I); 2486 I2 = NewI2; 2487 I2T = I1T; 2488 I2Elem = I1Elem; 2489 } 2490 2491 // Now that both I1 and I2 are the same length we can shuffle them 2492 // together (and use the result). 2493 std::vector<Constant *> Mask(numElem); 2494 for (unsigned v = 0; v < numElem; ++v) { 2495 if (II[v].first == -1) { 2496 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2497 } else { 2498 int Idx = II[v].first + II[v].second * I1Elem; 2499 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2500 } 2501 } 2502 2503 Instruction *NewOp = 2504 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), 2505 getReplacementName(IBeforeJ ? I : J, true, o)); 2506 NewOp->insertBefore(IBeforeJ ? J : I); 2507 return NewOp; 2508 } 2509 } 2510 2511 Type *ArgType = ArgTypeL; 2512 if (numElemL < numElemH) { 2513 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, 2514 ArgTypeL, VArgType, IBeforeJ, 1)) { 2515 // This is another short-circuit case: we're combining a scalar into 2516 // a vector that is formed by an IE chain. We've just expanded the IE 2517 // chain, now insert the scalar and we're done. 2518 2519 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, 2520 getReplacementName(IBeforeJ ? I : J, true, o)); 2521 S->insertBefore(IBeforeJ ? J : I); 2522 return S; 2523 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, 2524 ArgTypeH, IBeforeJ)) { 2525 // The two vector inputs to the shuffle must be the same length, 2526 // so extend the smaller vector to be the same length as the larger one. 2527 Instruction *NLOp; 2528 if (numElemL > 1) { 2529 2530 std::vector<Constant *> Mask(numElemH); 2531 unsigned v = 0; 2532 for (; v < numElemL; ++v) 2533 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2534 for (; v < numElemH; ++v) 2535 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2536 2537 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), 2538 ConstantVector::get(Mask), 2539 getReplacementName(IBeforeJ ? I : J, 2540 true, o, 1)); 2541 } else { 2542 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, 2543 getReplacementName(IBeforeJ ? I : J, 2544 true, o, 1)); 2545 } 2546 2547 NLOp->insertBefore(IBeforeJ ? J : I); 2548 LOp = NLOp; 2549 } 2550 2551 ArgType = ArgTypeH; 2552 } else if (numElemL > numElemH) { 2553 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, 2554 ArgTypeH, VArgType, IBeforeJ)) { 2555 Instruction *S = 2556 InsertElementInst::Create(LOp, HOp, 2557 ConstantInt::get(Type::getInt32Ty(Context), 2558 numElemL), 2559 getReplacementName(IBeforeJ ? I : J, 2560 true, o)); 2561 S->insertBefore(IBeforeJ ? J : I); 2562 return S; 2563 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, 2564 ArgTypeL, IBeforeJ)) { 2565 Instruction *NHOp; 2566 if (numElemH > 1) { 2567 std::vector<Constant *> Mask(numElemL); 2568 unsigned v = 0; 2569 for (; v < numElemH; ++v) 2570 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2571 for (; v < numElemL; ++v) 2572 Mask[v] = UndefValue::get(Type::getInt32Ty(Context)); 2573 2574 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), 2575 ConstantVector::get(Mask), 2576 getReplacementName(IBeforeJ ? I : J, 2577 true, o, 1)); 2578 } else { 2579 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, 2580 getReplacementName(IBeforeJ ? I : J, 2581 true, o, 1)); 2582 } 2583 2584 NHOp->insertBefore(IBeforeJ ? J : I); 2585 HOp = NHOp; 2586 } 2587 } 2588 2589 if (ArgType->isVectorTy()) { 2590 unsigned numElem = cast<VectorType>(VArgType)->getNumElements(); 2591 std::vector<Constant*> Mask(numElem); 2592 for (unsigned v = 0; v < numElem; ++v) { 2593 unsigned Idx = v; 2594 // If the low vector was expanded, we need to skip the extra 2595 // undefined entries. 2596 if (v >= numElemL && numElemH > numElemL) 2597 Idx += (numElemH - numElemL); 2598 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx); 2599 } 2600 2601 Instruction *BV = new ShuffleVectorInst(LOp, HOp, 2602 ConstantVector::get(Mask), 2603 getReplacementName(IBeforeJ ? I : J, true, o)); 2604 BV->insertBefore(IBeforeJ ? J : I); 2605 return BV; 2606 } 2607 2608 Instruction *BV1 = InsertElementInst::Create( 2609 UndefValue::get(VArgType), LOp, CV0, 2610 getReplacementName(IBeforeJ ? I : J, 2611 true, o, 1)); 2612 BV1->insertBefore(IBeforeJ ? J : I); 2613 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, 2614 getReplacementName(IBeforeJ ? I : J, 2615 true, o, 2)); 2616 BV2->insertBefore(IBeforeJ ? J : I); 2617 return BV2; 2618 } 2619 2620 // This function creates an array of values that will be used as the inputs 2621 // to the vector instruction that fuses I with J. 2622 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, 2623 Instruction *I, Instruction *J, 2624 SmallVector<Value *, 3> &ReplacedOperands, 2625 bool IBeforeJ) { 2626 unsigned NumOperands = I->getNumOperands(); 2627 2628 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { 2629 // Iterate backward so that we look at the store pointer 2630 // first and know whether or not we need to flip the inputs. 2631 2632 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { 2633 // This is the pointer for a load/store instruction. 2634 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); 2635 continue; 2636 } else if (isa<CallInst>(I)) { 2637 Function *F = cast<CallInst>(I)->getCalledFunction(); 2638 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID(); 2639 if (o == NumOperands-1) { 2640 BasicBlock &BB = *I->getParent(); 2641 2642 Module *M = BB.getParent()->getParent(); 2643 Type *ArgTypeI = I->getType(); 2644 Type *ArgTypeJ = J->getType(); 2645 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ); 2646 2647 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType); 2648 continue; 2649 } else if (IID == Intrinsic::powi && o == 1) { 2650 // The second argument of powi is a single integer and we've already 2651 // checked that both arguments are equal. As a result, we just keep 2652 // I's second argument. 2653 ReplacedOperands[o] = I->getOperand(o); 2654 continue; 2655 } 2656 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) { 2657 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J); 2658 continue; 2659 } 2660 2661 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); 2662 } 2663 } 2664 2665 // This function creates two values that represent the outputs of the 2666 // original I and J instructions. These are generally vector shuffles 2667 // or extracts. In many cases, these will end up being unused and, thus, 2668 // eliminated by later passes. 2669 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, 2670 Instruction *J, Instruction *K, 2671 Instruction *&InsertionPt, 2672 Instruction *&K1, Instruction *&K2) { 2673 if (isa<StoreInst>(I)) { 2674 AA->replaceWithNewValue(I, K); 2675 AA->replaceWithNewValue(J, K); 2676 } else { 2677 Type *IType = I->getType(); 2678 Type *JType = J->getType(); 2679 2680 VectorType *VType = getVecTypeForPair(IType, JType); 2681 unsigned numElem = VType->getNumElements(); 2682 2683 unsigned numElemI, numElemJ; 2684 if (IType->isVectorTy()) 2685 numElemI = cast<VectorType>(IType)->getNumElements(); 2686 else 2687 numElemI = 1; 2688 2689 if (JType->isVectorTy()) 2690 numElemJ = cast<VectorType>(JType)->getNumElements(); 2691 else 2692 numElemJ = 1; 2693 2694 if (IType->isVectorTy()) { 2695 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI); 2696 for (unsigned v = 0; v < numElemI; ++v) { 2697 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2698 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v); 2699 } 2700 2701 K1 = new ShuffleVectorInst(K, UndefValue::get(VType), 2702 ConstantVector::get( Mask1), 2703 getReplacementName(K, false, 1)); 2704 } else { 2705 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); 2706 K1 = ExtractElementInst::Create(K, CV0, 2707 getReplacementName(K, false, 1)); 2708 } 2709 2710 if (JType->isVectorTy()) { 2711 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ); 2712 for (unsigned v = 0; v < numElemJ; ++v) { 2713 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v); 2714 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v); 2715 } 2716 2717 K2 = new ShuffleVectorInst(K, UndefValue::get(VType), 2718 ConstantVector::get( Mask2), 2719 getReplacementName(K, false, 2)); 2720 } else { 2721 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); 2722 K2 = ExtractElementInst::Create(K, CV1, 2723 getReplacementName(K, false, 2)); 2724 } 2725 2726 K1->insertAfter(K); 2727 K2->insertAfter(K1); 2728 InsertionPt = K2; 2729 } 2730 } 2731 2732 // Move all uses of the function I (including pairing-induced uses) after J. 2733 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB, 2734 DenseSet<ValuePair> &LoadMoveSetPairs, 2735 Instruction *I, Instruction *J) { 2736 // Skip to the first instruction past I. 2737 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2738 2739 DenseSet<Value *> Users; 2740 AliasSetTracker WriteSet(*AA); 2741 for (; cast<Instruction>(L) != J; ++L) 2742 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs); 2743 2744 assert(cast<Instruction>(L) == J && 2745 "Tracking has not proceeded far enough to check for dependencies"); 2746 // If J is now in the use set of I, then trackUsesOfI will return true 2747 // and we have a dependency cycle (and the fusing operation must abort). 2748 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs); 2749 } 2750 2751 // Move all uses of the function I (including pairing-induced uses) after J. 2752 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB, 2753 DenseSet<ValuePair> &LoadMoveSetPairs, 2754 Instruction *&InsertionPt, 2755 Instruction *I, Instruction *J) { 2756 // Skip to the first instruction past I. 2757 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2758 2759 DenseSet<Value *> Users; 2760 AliasSetTracker WriteSet(*AA); 2761 for (; cast<Instruction>(L) != J;) { 2762 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) { 2763 // Move this instruction 2764 Instruction *InstToMove = L; ++L; 2765 2766 DEBUG(dbgs() << "BBV: moving: " << *InstToMove << 2767 " to after " << *InsertionPt << "\n"); 2768 InstToMove->removeFromParent(); 2769 InstToMove->insertAfter(InsertionPt); 2770 InsertionPt = InstToMove; 2771 } else { 2772 ++L; 2773 } 2774 } 2775 } 2776 2777 // Collect all load instruction that are in the move set of a given first 2778 // pair member. These loads depend on the first instruction, I, and so need 2779 // to be moved after J (the second instruction) when the pair is fused. 2780 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB, 2781 DenseMap<Value *, Value *> &ChosenPairs, 2782 std::multimap<Value *, Value *> &LoadMoveSet, 2783 DenseSet<ValuePair> &LoadMoveSetPairs, 2784 Instruction *I) { 2785 // Skip to the first instruction past I. 2786 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I)); 2787 2788 DenseSet<Value *> Users; 2789 AliasSetTracker WriteSet(*AA); 2790 2791 // Note: We cannot end the loop when we reach J because J could be moved 2792 // farther down the use chain by another instruction pairing. Also, J 2793 // could be before I if this is an inverted input. 2794 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) { 2795 if (trackUsesOfI(Users, WriteSet, I, L)) { 2796 if (L->mayReadFromMemory()) { 2797 LoadMoveSet.insert(ValuePair(L, I)); 2798 LoadMoveSetPairs.insert(ValuePair(L, I)); 2799 } 2800 } 2801 } 2802 } 2803 2804 // In cases where both load/stores and the computation of their pointers 2805 // are chosen for vectorization, we can end up in a situation where the 2806 // aliasing analysis starts returning different query results as the 2807 // process of fusing instruction pairs continues. Because the algorithm 2808 // relies on finding the same use trees here as were found earlier, we'll 2809 // need to precompute the necessary aliasing information here and then 2810 // manually update it during the fusion process. 2811 void BBVectorize::collectLoadMoveSet(BasicBlock &BB, 2812 std::vector<Value *> &PairableInsts, 2813 DenseMap<Value *, Value *> &ChosenPairs, 2814 std::multimap<Value *, Value *> &LoadMoveSet, 2815 DenseSet<ValuePair> &LoadMoveSetPairs) { 2816 for (std::vector<Value *>::iterator PI = PairableInsts.begin(), 2817 PIE = PairableInsts.end(); PI != PIE; ++PI) { 2818 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); 2819 if (P == ChosenPairs.end()) continue; 2820 2821 Instruction *I = cast<Instruction>(P->first); 2822 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, 2823 LoadMoveSetPairs, I); 2824 } 2825 } 2826 2827 // When the first instruction in each pair is cloned, it will inherit its 2828 // parent's metadata. This metadata must be combined with that of the other 2829 // instruction in a safe way. 2830 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) { 2831 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata; 2832 K->getAllMetadataOtherThanDebugLoc(Metadata); 2833 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 2834 unsigned Kind = Metadata[i].first; 2835 MDNode *JMD = J->getMetadata(Kind); 2836 MDNode *KMD = Metadata[i].second; 2837 2838 switch (Kind) { 2839 default: 2840 K->setMetadata(Kind, 0); // Remove unknown metadata 2841 break; 2842 case LLVMContext::MD_tbaa: 2843 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2844 break; 2845 case LLVMContext::MD_fpmath: 2846 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2847 break; 2848 } 2849 } 2850 } 2851 2852 // This function fuses the chosen instruction pairs into vector instructions, 2853 // taking care preserve any needed scalar outputs and, then, it reorders the 2854 // remaining instructions as needed (users of the first member of the pair 2855 // need to be moved to after the location of the second member of the pair 2856 // because the vector instruction is inserted in the location of the pair's 2857 // second member). 2858 void BBVectorize::fuseChosenPairs(BasicBlock &BB, 2859 std::vector<Value *> &PairableInsts, 2860 DenseMap<Value *, Value *> &ChosenPairs, 2861 DenseSet<ValuePair> &FixedOrderPairs, 2862 DenseMap<VPPair, unsigned> &PairConnectionTypes, 2863 std::multimap<ValuePair, ValuePair> &ConnectedPairs, 2864 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { 2865 LLVMContext& Context = BB.getContext(); 2866 2867 // During the vectorization process, the order of the pairs to be fused 2868 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs 2869 // list. After a pair is fused, the flipped pair is removed from the list. 2870 DenseSet<ValuePair> FlippedPairs; 2871 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), 2872 E = ChosenPairs.end(); P != E; ++P) 2873 FlippedPairs.insert(ValuePair(P->second, P->first)); 2874 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), 2875 E = FlippedPairs.end(); P != E; ++P) 2876 ChosenPairs.insert(*P); 2877 2878 std::multimap<Value *, Value *> LoadMoveSet; 2879 DenseSet<ValuePair> LoadMoveSetPairs; 2880 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, 2881 LoadMoveSet, LoadMoveSetPairs); 2882 2883 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); 2884 2885 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { 2886 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI); 2887 if (P == ChosenPairs.end()) { 2888 ++PI; 2889 continue; 2890 } 2891 2892 if (getDepthFactor(P->first) == 0) { 2893 // These instructions are not really fused, but are tracked as though 2894 // they are. Any case in which it would be interesting to fuse them 2895 // will be taken care of by InstCombine. 2896 --NumFusedOps; 2897 ++PI; 2898 continue; 2899 } 2900 2901 Instruction *I = cast<Instruction>(P->first), 2902 *J = cast<Instruction>(P->second); 2903 2904 DEBUG(dbgs() << "BBV: fusing: " << *I << 2905 " <-> " << *J << "\n"); 2906 2907 // Remove the pair and flipped pair from the list. 2908 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second); 2909 assert(FP != ChosenPairs.end() && "Flipped pair not found in list"); 2910 ChosenPairs.erase(FP); 2911 ChosenPairs.erase(P); 2912 2913 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) { 2914 DEBUG(dbgs() << "BBV: fusion of: " << *I << 2915 " <-> " << *J << 2916 " aborted because of non-trivial dependency cycle\n"); 2917 --NumFusedOps; 2918 ++PI; 2919 continue; 2920 } 2921 2922 // If the pair must have the other order, then flip it. 2923 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); 2924 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { 2925 // This pair does not have a fixed order, and so we might want to 2926 // flip it if that will yield fewer shuffles. We count the number 2927 // of dependencies connected via swaps, and those directly connected, 2928 // and flip the order if the number of swaps is greater. 2929 bool OrigOrder = true; 2930 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); 2931 if (IP.first == ConnectedPairDeps.end()) { 2932 IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); 2933 OrigOrder = false; 2934 } 2935 2936 if (IP.first != ConnectedPairDeps.end()) { 2937 unsigned NumDepsDirect = 0, NumDepsSwap = 0; 2938 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2939 Q != IP.second; ++Q) { 2940 DenseMap<VPPair, unsigned>::iterator R = 2941 PairConnectionTypes.find(VPPair(Q->second, Q->first)); 2942 assert(R != PairConnectionTypes.end() && 2943 "Cannot find pair connection type"); 2944 if (R->second == PairConnectionDirect) 2945 ++NumDepsDirect; 2946 else if (R->second == PairConnectionSwap) 2947 ++NumDepsSwap; 2948 } 2949 2950 if (!OrigOrder) 2951 std::swap(NumDepsDirect, NumDepsSwap); 2952 2953 if (NumDepsSwap > NumDepsDirect) { 2954 FlipPairOrder = true; 2955 DEBUG(dbgs() << "BBV: reordering pair: " << *I << 2956 " <-> " << *J << "\n"); 2957 } 2958 } 2959 } 2960 2961 Instruction *L = I, *H = J; 2962 if (FlipPairOrder) 2963 std::swap(H, L); 2964 2965 // If the pair being fused uses the opposite order from that in the pair 2966 // connection map, then we need to flip the types. 2967 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); 2968 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; 2969 Q != IP.second; ++Q) { 2970 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); 2971 assert(R != PairConnectionTypes.end() && 2972 "Cannot find pair connection type"); 2973 if (R->second == PairConnectionDirect) 2974 R->second = PairConnectionSwap; 2975 else if (R->second == PairConnectionSwap) 2976 R->second = PairConnectionDirect; 2977 } 2978 2979 bool LBeforeH = !FlipPairOrder; 2980 unsigned NumOperands = I->getNumOperands(); 2981 SmallVector<Value *, 3> ReplacedOperands(NumOperands); 2982 getReplacementInputsForPair(Context, L, H, ReplacedOperands, 2983 LBeforeH); 2984 2985 // Make a copy of the original operation, change its type to the vector 2986 // type and replace its operands with the vector operands. 2987 Instruction *K = L->clone(); 2988 if (L->hasName()) 2989 K->takeName(L); 2990 else if (H->hasName()) 2991 K->takeName(H); 2992 2993 if (!isa<StoreInst>(K)) 2994 K->mutateType(getVecTypeForPair(L->getType(), H->getType())); 2995 2996 combineMetadata(K, H); 2997 K->intersectOptionalDataWith(H); 2998 2999 for (unsigned o = 0; o < NumOperands; ++o) 3000 K->setOperand(o, ReplacedOperands[o]); 3001 3002 K->insertAfter(J); 3003 3004 // Instruction insertion point: 3005 Instruction *InsertionPt = K; 3006 Instruction *K1 = 0, *K2 = 0; 3007 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); 3008 3009 // The use tree of the first original instruction must be moved to after 3010 // the location of the second instruction. The entire use tree of the 3011 // first instruction is disjoint from the input tree of the second 3012 // (by definition), and so commutes with it. 3013 3014 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J); 3015 3016 if (!isa<StoreInst>(I)) { 3017 L->replaceAllUsesWith(K1); 3018 H->replaceAllUsesWith(K2); 3019 AA->replaceWithNewValue(L, K1); 3020 AA->replaceWithNewValue(H, K2); 3021 } 3022 3023 // Instructions that may read from memory may be in the load move set. 3024 // Once an instruction is fused, we no longer need its move set, and so 3025 // the values of the map never need to be updated. However, when a load 3026 // is fused, we need to merge the entries from both instructions in the 3027 // pair in case those instructions were in the move set of some other 3028 // yet-to-be-fused pair. The loads in question are the keys of the map. 3029 if (I->mayReadFromMemory()) { 3030 std::vector<ValuePair> NewSetMembers; 3031 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I); 3032 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J); 3033 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first; 3034 N != IPairRange.second; ++N) 3035 NewSetMembers.push_back(ValuePair(K, N->second)); 3036 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first; 3037 N != JPairRange.second; ++N) 3038 NewSetMembers.push_back(ValuePair(K, N->second)); 3039 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(), 3040 AE = NewSetMembers.end(); A != AE; ++A) { 3041 LoadMoveSet.insert(*A); 3042 LoadMoveSetPairs.insert(*A); 3043 } 3044 } 3045 3046 // Before removing I, set the iterator to the next instruction. 3047 PI = llvm::next(BasicBlock::iterator(I)); 3048 if (cast<Instruction>(PI) == J) 3049 ++PI; 3050 3051 SE->forgetValue(I); 3052 SE->forgetValue(J); 3053 I->eraseFromParent(); 3054 J->eraseFromParent(); 3055 3056 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << 3057 BB << "\n"); 3058 } 3059 3060 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); 3061 } 3062} 3063 3064char BBVectorize::ID = 0; 3065static const char bb_vectorize_name[] = "Basic-Block Vectorization"; 3066INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3067INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3068INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3069INITIALIZE_PASS_DEPENDENCY(DominatorTree) 3070INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3071INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) 3072 3073BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) { 3074 return new BBVectorize(C); 3075} 3076 3077bool 3078llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) { 3079 BBVectorize BBVectorizer(P, C); 3080 return BBVectorizer.vectorizeBB(BB); 3081} 3082 3083//===----------------------------------------------------------------------===// 3084VectorizeConfig::VectorizeConfig() { 3085 VectorBits = ::VectorBits; 3086 VectorizeBools = !::NoBools; 3087 VectorizeInts = !::NoInts; 3088 VectorizeFloats = !::NoFloats; 3089 VectorizePointers = !::NoPointers; 3090 VectorizeCasts = !::NoCasts; 3091 VectorizeMath = !::NoMath; 3092 VectorizeFMA = !::NoFMA; 3093 VectorizeSelect = !::NoSelect; 3094 VectorizeCmp = !::NoCmp; 3095 VectorizeGEP = !::NoGEP; 3096 VectorizeMemOps = !::NoMemOps; 3097 AlignedOnly = ::AlignedOnly; 3098 ReqChainDepth= ::ReqChainDepth; 3099 SearchLimit = ::SearchLimit; 3100 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck; 3101 SplatBreaksChain = ::SplatBreaksChain; 3102 MaxInsts = ::MaxInsts; 3103 MaxIter = ::MaxIter; 3104 Pow2LenOnly = ::Pow2LenOnly; 3105 NoMemOpBoost = ::NoMemOpBoost; 3106 FastDep = ::FastDep; 3107} 3108