1 //===- LoopVectorize.cpp - A Loop 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 is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops 11 // and generates target-independent LLVM-IR. 12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs 13 // of instructions in order to estimate the profitability of vectorization. 14 // 15 // The loop vectorizer combines consecutive loop iterations into a single 16 // 'wide' iteration. After this transformation the index is incremented 17 // by the SIMD vector width, and not by one. 18 // 19 // This pass has three parts: 20 // 1. The main loop pass that drives the different parts. 21 // 2. LoopVectorizationLegality - A unit that checks for the legality 22 // of the vectorization. 23 // 3. InnerLoopVectorizer - A unit that performs the actual 24 // widening of instructions. 25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability 26 // of vectorization. It decides on the optimal vector width, which 27 // can be one, if vectorization is not profitable. 28 // 29 //===----------------------------------------------------------------------===// 30 // 31 // The reduction-variable vectorization is based on the paper: 32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. 33 // 34 // Variable uniformity checks are inspired by: 35 // Karrenberg, R. and Hack, S. Whole Function Vectorization. 36 // 37 // The interleaved access vectorization is based on the paper: 38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved 39 // Data for SIMD 40 // 41 // Other ideas/concepts are from: 42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. 43 // 44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of 45 // Vectorizing Compilers. 46 // 47 //===----------------------------------------------------------------------===// 48 49 #include "llvm/Transforms/Vectorize.h" 50 #include "llvm/ADT/DenseMap.h" 51 #include "llvm/ADT/Hashing.h" 52 #include "llvm/ADT/MapVector.h" 53 #include "llvm/ADT/SetVector.h" 54 #include "llvm/ADT/SmallPtrSet.h" 55 #include "llvm/ADT/SmallSet.h" 56 #include "llvm/ADT/SmallVector.h" 57 #include "llvm/ADT/Statistic.h" 58 #include "llvm/ADT/StringExtras.h" 59 #include "llvm/Analysis/AliasAnalysis.h" 60 #include "llvm/Analysis/BasicAliasAnalysis.h" 61 #include "llvm/Analysis/AliasSetTracker.h" 62 #include "llvm/Analysis/AssumptionCache.h" 63 #include "llvm/Analysis/BlockFrequencyInfo.h" 64 #include "llvm/Analysis/CodeMetrics.h" 65 #include "llvm/Analysis/DemandedBits.h" 66 #include "llvm/Analysis/GlobalsModRef.h" 67 #include "llvm/Analysis/LoopAccessAnalysis.h" 68 #include "llvm/Analysis/LoopInfo.h" 69 #include "llvm/Analysis/LoopIterator.h" 70 #include "llvm/Analysis/LoopPass.h" 71 #include "llvm/Analysis/ScalarEvolution.h" 72 #include "llvm/Analysis/ScalarEvolutionExpander.h" 73 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 74 #include "llvm/Analysis/TargetTransformInfo.h" 75 #include "llvm/Analysis/ValueTracking.h" 76 #include "llvm/IR/Constants.h" 77 #include "llvm/IR/DataLayout.h" 78 #include "llvm/IR/DebugInfo.h" 79 #include "llvm/IR/DerivedTypes.h" 80 #include "llvm/IR/DiagnosticInfo.h" 81 #include "llvm/IR/Dominators.h" 82 #include "llvm/IR/Function.h" 83 #include "llvm/IR/IRBuilder.h" 84 #include "llvm/IR/Instructions.h" 85 #include "llvm/IR/IntrinsicInst.h" 86 #include "llvm/IR/LLVMContext.h" 87 #include "llvm/IR/Module.h" 88 #include "llvm/IR/PatternMatch.h" 89 #include "llvm/IR/Type.h" 90 #include "llvm/IR/Value.h" 91 #include "llvm/IR/ValueHandle.h" 92 #include "llvm/IR/Verifier.h" 93 #include "llvm/Pass.h" 94 #include "llvm/Support/BranchProbability.h" 95 #include "llvm/Support/CommandLine.h" 96 #include "llvm/Support/Debug.h" 97 #include "llvm/Support/raw_ostream.h" 98 #include "llvm/Transforms/Scalar.h" 99 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 100 #include "llvm/Transforms/Utils/Local.h" 101 #include "llvm/Analysis/VectorUtils.h" 102 #include "llvm/Transforms/Utils/LoopUtils.h" 103 #include <algorithm> 104 #include <functional> 105 #include <map> 106 #include <tuple> 107 108 using namespace llvm; 109 using namespace llvm::PatternMatch; 110 111 #define LV_NAME "loop-vectorize" 112 #define DEBUG_TYPE LV_NAME 113 114 STATISTIC(LoopsVectorized, "Number of loops vectorized"); 115 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization"); 116 117 static cl::opt<bool> 118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, 119 cl::desc("Enable if-conversion during vectorization.")); 120 121 /// We don't vectorize loops with a known constant trip count below this number. 122 static cl::opt<unsigned> 123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16), 124 cl::Hidden, 125 cl::desc("Don't vectorize loops with a constant " 126 "trip count that is smaller than this " 127 "value.")); 128 129 static cl::opt<bool> MaximizeBandwidth( 130 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, 131 cl::desc("Maximize bandwidth when selecting vectorization factor which " 132 "will be determined by the smallest type in loop.")); 133 134 /// This enables versioning on the strides of symbolically striding memory 135 /// accesses in code like the following. 136 /// for (i = 0; i < N; ++i) 137 /// A[i * Stride1] += B[i * Stride2] ... 138 /// 139 /// Will be roughly translated to 140 /// if (Stride1 == 1 && Stride2 == 1) { 141 /// for (i = 0; i < N; i+=4) 142 /// A[i:i+3] += ... 143 /// } else 144 /// ... 145 static cl::opt<bool> EnableMemAccessVersioning( 146 "enable-mem-access-versioning", cl::init(true), cl::Hidden, 147 cl::desc("Enable symbolic stride memory access versioning")); 148 149 static cl::opt<bool> EnableInterleavedMemAccesses( 150 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, 151 cl::desc("Enable vectorization on interleaved memory accesses in a loop")); 152 153 /// Maximum factor for an interleaved memory access. 154 static cl::opt<unsigned> MaxInterleaveGroupFactor( 155 "max-interleave-group-factor", cl::Hidden, 156 cl::desc("Maximum factor for an interleaved access group (default = 8)"), 157 cl::init(8)); 158 159 /// We don't interleave loops with a known constant trip count below this 160 /// number. 161 static const unsigned TinyTripCountInterleaveThreshold = 128; 162 163 static cl::opt<unsigned> ForceTargetNumScalarRegs( 164 "force-target-num-scalar-regs", cl::init(0), cl::Hidden, 165 cl::desc("A flag that overrides the target's number of scalar registers.")); 166 167 static cl::opt<unsigned> ForceTargetNumVectorRegs( 168 "force-target-num-vector-regs", cl::init(0), cl::Hidden, 169 cl::desc("A flag that overrides the target's number of vector registers.")); 170 171 /// Maximum vectorization interleave count. 172 static const unsigned MaxInterleaveFactor = 16; 173 174 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( 175 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, 176 cl::desc("A flag that overrides the target's max interleave factor for " 177 "scalar loops.")); 178 179 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( 180 "force-target-max-vector-interleave", cl::init(0), cl::Hidden, 181 cl::desc("A flag that overrides the target's max interleave factor for " 182 "vectorized loops.")); 183 184 static cl::opt<unsigned> ForceTargetInstructionCost( 185 "force-target-instruction-cost", cl::init(0), cl::Hidden, 186 cl::desc("A flag that overrides the target's expected cost for " 187 "an instruction to a single constant value. Mostly " 188 "useful for getting consistent testing.")); 189 190 static cl::opt<unsigned> SmallLoopCost( 191 "small-loop-cost", cl::init(20), cl::Hidden, 192 cl::desc( 193 "The cost of a loop that is considered 'small' by the interleaver.")); 194 195 static cl::opt<bool> LoopVectorizeWithBlockFrequency( 196 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden, 197 cl::desc("Enable the use of the block frequency analysis to access PGO " 198 "heuristics minimizing code growth in cold regions and being more " 199 "aggressive in hot regions.")); 200 201 // Runtime interleave loops for load/store throughput. 202 static cl::opt<bool> EnableLoadStoreRuntimeInterleave( 203 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, 204 cl::desc( 205 "Enable runtime interleaving until load/store ports are saturated")); 206 207 /// The number of stores in a loop that are allowed to need predication. 208 static cl::opt<unsigned> NumberOfStoresToPredicate( 209 "vectorize-num-stores-pred", cl::init(1), cl::Hidden, 210 cl::desc("Max number of stores to be predicated behind an if.")); 211 212 static cl::opt<bool> EnableIndVarRegisterHeur( 213 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, 214 cl::desc("Count the induction variable only once when interleaving")); 215 216 static cl::opt<bool> EnableCondStoresVectorization( 217 "enable-cond-stores-vec", cl::init(false), cl::Hidden, 218 cl::desc("Enable if predication of stores during vectorization.")); 219 220 static cl::opt<unsigned> MaxNestedScalarReductionIC( 221 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, 222 cl::desc("The maximum interleave count to use when interleaving a scalar " 223 "reduction in a nested loop.")); 224 225 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold( 226 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden, 227 cl::desc("The maximum allowed number of runtime memory checks with a " 228 "vectorize(enable) pragma.")); 229 230 static cl::opt<unsigned> VectorizeSCEVCheckThreshold( 231 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden, 232 cl::desc("The maximum number of SCEV checks allowed.")); 233 234 static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold( 235 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden, 236 cl::desc("The maximum number of SCEV checks allowed with a " 237 "vectorize(enable) pragma")); 238 239 namespace { 240 241 // Forward declarations. 242 class LoopVectorizeHints; 243 class LoopVectorizationLegality; 244 class LoopVectorizationCostModel; 245 class LoopVectorizationRequirements; 246 247 /// \brief This modifies LoopAccessReport to initialize message with 248 /// loop-vectorizer-specific part. 249 class VectorizationReport : public LoopAccessReport { 250 public: 251 VectorizationReport(Instruction *I = nullptr) 252 : LoopAccessReport("loop not vectorized: ", I) {} 253 254 /// \brief This allows promotion of the loop-access analysis report into the 255 /// loop-vectorizer report. It modifies the message to add the 256 /// loop-vectorizer-specific part of the message. 257 explicit VectorizationReport(const LoopAccessReport &R) 258 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(), 259 R.getInstr()) {} 260 }; 261 262 /// A helper function for converting Scalar types to vector types. 263 /// If the incoming type is void, we return void. If the VF is 1, we return 264 /// the scalar type. 265 static Type* ToVectorTy(Type *Scalar, unsigned VF) { 266 if (Scalar->isVoidTy() || VF == 1) 267 return Scalar; 268 return VectorType::get(Scalar, VF); 269 } 270 271 /// A helper function that returns GEP instruction and knows to skip a 272 /// 'bitcast'. The 'bitcast' may be skipped if the source and the destination 273 /// pointee types of the 'bitcast' have the same size. 274 /// For example: 275 /// bitcast double** %var to i64* - can be skipped 276 /// bitcast double** %var to i8* - can not 277 static GetElementPtrInst *getGEPInstruction(Value *Ptr) { 278 279 if (isa<GetElementPtrInst>(Ptr)) 280 return cast<GetElementPtrInst>(Ptr); 281 282 if (isa<BitCastInst>(Ptr) && 283 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0))) { 284 Type *BitcastTy = Ptr->getType(); 285 Type *GEPTy = cast<BitCastInst>(Ptr)->getSrcTy(); 286 if (!isa<PointerType>(BitcastTy) || !isa<PointerType>(GEPTy)) 287 return nullptr; 288 Type *Pointee1Ty = cast<PointerType>(BitcastTy)->getPointerElementType(); 289 Type *Pointee2Ty = cast<PointerType>(GEPTy)->getPointerElementType(); 290 const DataLayout &DL = cast<BitCastInst>(Ptr)->getModule()->getDataLayout(); 291 if (DL.getTypeSizeInBits(Pointee1Ty) == DL.getTypeSizeInBits(Pointee2Ty)) 292 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0)); 293 } 294 return nullptr; 295 } 296 297 /// InnerLoopVectorizer vectorizes loops which contain only one basic 298 /// block to a specified vectorization factor (VF). 299 /// This class performs the widening of scalars into vectors, or multiple 300 /// scalars. This class also implements the following features: 301 /// * It inserts an epilogue loop for handling loops that don't have iteration 302 /// counts that are known to be a multiple of the vectorization factor. 303 /// * It handles the code generation for reduction variables. 304 /// * Scalarization (implementation using scalars) of un-vectorizable 305 /// instructions. 306 /// InnerLoopVectorizer does not perform any vectorization-legality 307 /// checks, and relies on the caller to check for the different legality 308 /// aspects. The InnerLoopVectorizer relies on the 309 /// LoopVectorizationLegality class to provide information about the induction 310 /// and reduction variables that were found to a given vectorization factor. 311 class InnerLoopVectorizer { 312 public: 313 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, 314 LoopInfo *LI, DominatorTree *DT, 315 const TargetLibraryInfo *TLI, 316 const TargetTransformInfo *TTI, unsigned VecWidth, 317 unsigned UnrollFactor) 318 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), 319 VF(VecWidth), UF(UnrollFactor), Builder(PSE.getSE()->getContext()), 320 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor), 321 TripCount(nullptr), VectorTripCount(nullptr), Legal(nullptr), 322 AddedSafetyChecks(false) {} 323 324 // Perform the actual loop widening (vectorization). 325 // MinimumBitWidths maps scalar integer values to the smallest bitwidth they 326 // can be validly truncated to. The cost model has assumed this truncation 327 // will happen when vectorizing. 328 void vectorize(LoopVectorizationLegality *L, 329 MapVector<Instruction*,uint64_t> MinimumBitWidths) { 330 MinBWs = MinimumBitWidths; 331 Legal = L; 332 // Create a new empty loop. Unlink the old loop and connect the new one. 333 createEmptyLoop(); 334 // Widen each instruction in the old loop to a new one in the new loop. 335 // Use the Legality module to find the induction and reduction variables. 336 vectorizeLoop(); 337 } 338 339 // Return true if any runtime check is added. 340 bool IsSafetyChecksAdded() { 341 return AddedSafetyChecks; 342 } 343 344 virtual ~InnerLoopVectorizer() {} 345 346 protected: 347 /// A small list of PHINodes. 348 typedef SmallVector<PHINode*, 4> PhiVector; 349 /// When we unroll loops we have multiple vector values for each scalar. 350 /// This data structure holds the unrolled and vectorized values that 351 /// originated from one scalar instruction. 352 typedef SmallVector<Value*, 2> VectorParts; 353 354 // When we if-convert we need to create edge masks. We have to cache values 355 // so that we don't end up with exponential recursion/IR. 356 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>, 357 VectorParts> EdgeMaskCache; 358 359 /// Create an empty loop, based on the loop ranges of the old loop. 360 void createEmptyLoop(); 361 /// Create a new induction variable inside L. 362 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End, 363 Value *Step, Instruction *DL); 364 /// Copy and widen the instructions from the old loop. 365 virtual void vectorizeLoop(); 366 367 /// \brief The Loop exit block may have single value PHI nodes where the 368 /// incoming value is 'Undef'. While vectorizing we only handled real values 369 /// that were defined inside the loop. Here we fix the 'undef case'. 370 /// See PR14725. 371 void fixLCSSAPHIs(); 372 373 /// Shrinks vector element sizes based on information in "MinBWs". 374 void truncateToMinimalBitwidths(); 375 376 /// A helper function that computes the predicate of the block BB, assuming 377 /// that the header block of the loop is set to True. It returns the *entry* 378 /// mask for the block BB. 379 VectorParts createBlockInMask(BasicBlock *BB); 380 /// A helper function that computes the predicate of the edge between SRC 381 /// and DST. 382 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst); 383 384 /// A helper function to vectorize a single BB within the innermost loop. 385 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV); 386 387 /// Vectorize a single PHINode in a block. This method handles the induction 388 /// variable canonicalization. It supports both VF = 1 for unrolled loops and 389 /// arbitrary length vectors. 390 void widenPHIInstruction(Instruction *PN, VectorParts &Entry, 391 unsigned UF, unsigned VF, PhiVector *PV); 392 393 /// Insert the new loop to the loop hierarchy and pass manager 394 /// and update the analysis passes. 395 void updateAnalysis(); 396 397 /// This instruction is un-vectorizable. Implement it as a sequence 398 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each 399 /// scalarized instruction behind an if block predicated on the control 400 /// dependence of the instruction. 401 virtual void scalarizeInstruction(Instruction *Instr, 402 bool IfPredicateStore=false); 403 404 /// Vectorize Load and Store instructions, 405 virtual void vectorizeMemoryInstruction(Instruction *Instr); 406 407 /// Create a broadcast instruction. This method generates a broadcast 408 /// instruction (shuffle) for loop invariant values and for the induction 409 /// value. If this is the induction variable then we extend it to N, N+1, ... 410 /// this is needed because each iteration in the loop corresponds to a SIMD 411 /// element. 412 virtual Value *getBroadcastInstrs(Value *V); 413 414 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...) 415 /// to each vector element of Val. The sequence starts at StartIndex. 416 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step); 417 418 /// When we go over instructions in the basic block we rely on previous 419 /// values within the current basic block or on loop invariant values. 420 /// When we widen (vectorize) values we place them in the map. If the values 421 /// are not within the map, they have to be loop invariant, so we simply 422 /// broadcast them into a vector. 423 VectorParts &getVectorValue(Value *V); 424 425 /// Try to vectorize the interleaved access group that \p Instr belongs to. 426 void vectorizeInterleaveGroup(Instruction *Instr); 427 428 /// Generate a shuffle sequence that will reverse the vector Vec. 429 virtual Value *reverseVector(Value *Vec); 430 431 /// Returns (and creates if needed) the original loop trip count. 432 Value *getOrCreateTripCount(Loop *NewLoop); 433 434 /// Returns (and creates if needed) the trip count of the widened loop. 435 Value *getOrCreateVectorTripCount(Loop *NewLoop); 436 437 /// Emit a bypass check to see if the trip count would overflow, or we 438 /// wouldn't have enough iterations to execute one vector loop. 439 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass); 440 /// Emit a bypass check to see if the vector trip count is nonzero. 441 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass); 442 /// Emit a bypass check to see if all of the SCEV assumptions we've 443 /// had to make are correct. 444 void emitSCEVChecks(Loop *L, BasicBlock *Bypass); 445 /// Emit bypass checks to check any memory assumptions we may have made. 446 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass); 447 448 /// This is a helper class that holds the vectorizer state. It maps scalar 449 /// instructions to vector instructions. When the code is 'unrolled' then 450 /// then a single scalar value is mapped to multiple vector parts. The parts 451 /// are stored in the VectorPart type. 452 struct ValueMap { 453 /// C'tor. UnrollFactor controls the number of vectors ('parts') that 454 /// are mapped. 455 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {} 456 457 /// \return True if 'Key' is saved in the Value Map. 458 bool has(Value *Key) const { return MapStorage.count(Key); } 459 460 /// Initializes a new entry in the map. Sets all of the vector parts to the 461 /// save value in 'Val'. 462 /// \return A reference to a vector with splat values. 463 VectorParts &splat(Value *Key, Value *Val) { 464 VectorParts &Entry = MapStorage[Key]; 465 Entry.assign(UF, Val); 466 return Entry; 467 } 468 469 ///\return A reference to the value that is stored at 'Key'. 470 VectorParts &get(Value *Key) { 471 VectorParts &Entry = MapStorage[Key]; 472 if (Entry.empty()) 473 Entry.resize(UF); 474 assert(Entry.size() == UF); 475 return Entry; 476 } 477 478 private: 479 /// The unroll factor. Each entry in the map stores this number of vector 480 /// elements. 481 unsigned UF; 482 483 /// Map storage. We use std::map and not DenseMap because insertions to a 484 /// dense map invalidates its iterators. 485 std::map<Value *, VectorParts> MapStorage; 486 }; 487 488 /// The original loop. 489 Loop *OrigLoop; 490 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies 491 /// dynamic knowledge to simplify SCEV expressions and converts them to a 492 /// more usable form. 493 PredicatedScalarEvolution &PSE; 494 /// Loop Info. 495 LoopInfo *LI; 496 /// Dominator Tree. 497 DominatorTree *DT; 498 /// Alias Analysis. 499 AliasAnalysis *AA; 500 /// Target Library Info. 501 const TargetLibraryInfo *TLI; 502 /// Target Transform Info. 503 const TargetTransformInfo *TTI; 504 505 /// The vectorization SIMD factor to use. Each vector will have this many 506 /// vector elements. 507 unsigned VF; 508 509 protected: 510 /// The vectorization unroll factor to use. Each scalar is vectorized to this 511 /// many different vector instructions. 512 unsigned UF; 513 514 /// The builder that we use 515 IRBuilder<> Builder; 516 517 // --- Vectorization state --- 518 519 /// The vector-loop preheader. 520 BasicBlock *LoopVectorPreHeader; 521 /// The scalar-loop preheader. 522 BasicBlock *LoopScalarPreHeader; 523 /// Middle Block between the vector and the scalar. 524 BasicBlock *LoopMiddleBlock; 525 ///The ExitBlock of the scalar loop. 526 BasicBlock *LoopExitBlock; 527 ///The vector loop body. 528 SmallVector<BasicBlock *, 4> LoopVectorBody; 529 ///The scalar loop body. 530 BasicBlock *LoopScalarBody; 531 /// A list of all bypass blocks. The first block is the entry of the loop. 532 SmallVector<BasicBlock *, 4> LoopBypassBlocks; 533 534 /// The new Induction variable which was added to the new block. 535 PHINode *Induction; 536 /// The induction variable of the old basic block. 537 PHINode *OldInduction; 538 /// Maps scalars to widened vectors. 539 ValueMap WidenMap; 540 /// Store instructions that should be predicated, as a pair 541 /// <StoreInst, Predicate> 542 SmallVector<std::pair<StoreInst*,Value*>, 4> PredicatedStores; 543 EdgeMaskCache MaskCache; 544 /// Trip count of the original loop. 545 Value *TripCount; 546 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) 547 Value *VectorTripCount; 548 549 /// Map of scalar integer values to the smallest bitwidth they can be legally 550 /// represented as. The vector equivalents of these values should be truncated 551 /// to this type. 552 MapVector<Instruction*,uint64_t> MinBWs; 553 LoopVectorizationLegality *Legal; 554 555 // Record whether runtime check is added. 556 bool AddedSafetyChecks; 557 }; 558 559 class InnerLoopUnroller : public InnerLoopVectorizer { 560 public: 561 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE, 562 LoopInfo *LI, DominatorTree *DT, 563 const TargetLibraryInfo *TLI, 564 const TargetTransformInfo *TTI, unsigned UnrollFactor) 565 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, 1, UnrollFactor) {} 566 567 private: 568 void scalarizeInstruction(Instruction *Instr, 569 bool IfPredicateStore = false) override; 570 void vectorizeMemoryInstruction(Instruction *Instr) override; 571 Value *getBroadcastInstrs(Value *V) override; 572 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override; 573 Value *reverseVector(Value *Vec) override; 574 }; 575 576 /// \brief Look for a meaningful debug location on the instruction or it's 577 /// operands. 578 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { 579 if (!I) 580 return I; 581 582 DebugLoc Empty; 583 if (I->getDebugLoc() != Empty) 584 return I; 585 586 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) { 587 if (Instruction *OpInst = dyn_cast<Instruction>(*OI)) 588 if (OpInst->getDebugLoc() != Empty) 589 return OpInst; 590 } 591 592 return I; 593 } 594 595 /// \brief Set the debug location in the builder using the debug location in the 596 /// instruction. 597 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) { 598 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) 599 B.SetCurrentDebugLocation(Inst->getDebugLoc()); 600 else 601 B.SetCurrentDebugLocation(DebugLoc()); 602 } 603 604 #ifndef NDEBUG 605 /// \return string containing a file name and a line # for the given loop. 606 static std::string getDebugLocString(const Loop *L) { 607 std::string Result; 608 if (L) { 609 raw_string_ostream OS(Result); 610 if (const DebugLoc LoopDbgLoc = L->getStartLoc()) 611 LoopDbgLoc.print(OS); 612 else 613 // Just print the module name. 614 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); 615 OS.flush(); 616 } 617 return Result; 618 } 619 #endif 620 621 /// \brief Propagate known metadata from one instruction to another. 622 static void propagateMetadata(Instruction *To, const Instruction *From) { 623 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata; 624 From->getAllMetadataOtherThanDebugLoc(Metadata); 625 626 for (auto M : Metadata) { 627 unsigned Kind = M.first; 628 629 // These are safe to transfer (this is safe for TBAA, even when we 630 // if-convert, because should that metadata have had a control dependency 631 // on the condition, and thus actually aliased with some other 632 // non-speculated memory access when the condition was false, this would be 633 // caught by the runtime overlap checks). 634 if (Kind != LLVMContext::MD_tbaa && 635 Kind != LLVMContext::MD_alias_scope && 636 Kind != LLVMContext::MD_noalias && 637 Kind != LLVMContext::MD_fpmath && 638 Kind != LLVMContext::MD_nontemporal) 639 continue; 640 641 To->setMetadata(Kind, M.second); 642 } 643 } 644 645 /// \brief Propagate known metadata from one instruction to a vector of others. 646 static void propagateMetadata(SmallVectorImpl<Value *> &To, 647 const Instruction *From) { 648 for (Value *V : To) 649 if (Instruction *I = dyn_cast<Instruction>(V)) 650 propagateMetadata(I, From); 651 } 652 653 /// \brief The group of interleaved loads/stores sharing the same stride and 654 /// close to each other. 655 /// 656 /// Each member in this group has an index starting from 0, and the largest 657 /// index should be less than interleaved factor, which is equal to the absolute 658 /// value of the access's stride. 659 /// 660 /// E.g. An interleaved load group of factor 4: 661 /// for (unsigned i = 0; i < 1024; i+=4) { 662 /// a = A[i]; // Member of index 0 663 /// b = A[i+1]; // Member of index 1 664 /// d = A[i+3]; // Member of index 3 665 /// ... 666 /// } 667 /// 668 /// An interleaved store group of factor 4: 669 /// for (unsigned i = 0; i < 1024; i+=4) { 670 /// ... 671 /// A[i] = a; // Member of index 0 672 /// A[i+1] = b; // Member of index 1 673 /// A[i+2] = c; // Member of index 2 674 /// A[i+3] = d; // Member of index 3 675 /// } 676 /// 677 /// Note: the interleaved load group could have gaps (missing members), but 678 /// the interleaved store group doesn't allow gaps. 679 class InterleaveGroup { 680 public: 681 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align) 682 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) { 683 assert(Align && "The alignment should be non-zero"); 684 685 Factor = std::abs(Stride); 686 assert(Factor > 1 && "Invalid interleave factor"); 687 688 Reverse = Stride < 0; 689 Members[0] = Instr; 690 } 691 692 bool isReverse() const { return Reverse; } 693 unsigned getFactor() const { return Factor; } 694 unsigned getAlignment() const { return Align; } 695 unsigned getNumMembers() const { return Members.size(); } 696 697 /// \brief Try to insert a new member \p Instr with index \p Index and 698 /// alignment \p NewAlign. The index is related to the leader and it could be 699 /// negative if it is the new leader. 700 /// 701 /// \returns false if the instruction doesn't belong to the group. 702 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) { 703 assert(NewAlign && "The new member's alignment should be non-zero"); 704 705 int Key = Index + SmallestKey; 706 707 // Skip if there is already a member with the same index. 708 if (Members.count(Key)) 709 return false; 710 711 if (Key > LargestKey) { 712 // The largest index is always less than the interleave factor. 713 if (Index >= static_cast<int>(Factor)) 714 return false; 715 716 LargestKey = Key; 717 } else if (Key < SmallestKey) { 718 // The largest index is always less than the interleave factor. 719 if (LargestKey - Key >= static_cast<int>(Factor)) 720 return false; 721 722 SmallestKey = Key; 723 } 724 725 // It's always safe to select the minimum alignment. 726 Align = std::min(Align, NewAlign); 727 Members[Key] = Instr; 728 return true; 729 } 730 731 /// \brief Get the member with the given index \p Index 732 /// 733 /// \returns nullptr if contains no such member. 734 Instruction *getMember(unsigned Index) const { 735 int Key = SmallestKey + Index; 736 if (!Members.count(Key)) 737 return nullptr; 738 739 return Members.find(Key)->second; 740 } 741 742 /// \brief Get the index for the given member. Unlike the key in the member 743 /// map, the index starts from 0. 744 unsigned getIndex(Instruction *Instr) const { 745 for (auto I : Members) 746 if (I.second == Instr) 747 return I.first - SmallestKey; 748 749 llvm_unreachable("InterleaveGroup contains no such member"); 750 } 751 752 Instruction *getInsertPos() const { return InsertPos; } 753 void setInsertPos(Instruction *Inst) { InsertPos = Inst; } 754 755 private: 756 unsigned Factor; // Interleave Factor. 757 bool Reverse; 758 unsigned Align; 759 DenseMap<int, Instruction *> Members; 760 int SmallestKey; 761 int LargestKey; 762 763 // To avoid breaking dependences, vectorized instructions of an interleave 764 // group should be inserted at either the first load or the last store in 765 // program order. 766 // 767 // E.g. %even = load i32 // Insert Position 768 // %add = add i32 %even // Use of %even 769 // %odd = load i32 770 // 771 // store i32 %even 772 // %odd = add i32 // Def of %odd 773 // store i32 %odd // Insert Position 774 Instruction *InsertPos; 775 }; 776 777 /// \brief Drive the analysis of interleaved memory accesses in the loop. 778 /// 779 /// Use this class to analyze interleaved accesses only when we can vectorize 780 /// a loop. Otherwise it's meaningless to do analysis as the vectorization 781 /// on interleaved accesses is unsafe. 782 /// 783 /// The analysis collects interleave groups and records the relationships 784 /// between the member and the group in a map. 785 class InterleavedAccessInfo { 786 public: 787 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L, 788 DominatorTree *DT) 789 : PSE(PSE), TheLoop(L), DT(DT) {} 790 791 ~InterleavedAccessInfo() { 792 SmallSet<InterleaveGroup *, 4> DelSet; 793 // Avoid releasing a pointer twice. 794 for (auto &I : InterleaveGroupMap) 795 DelSet.insert(I.second); 796 for (auto *Ptr : DelSet) 797 delete Ptr; 798 } 799 800 /// \brief Analyze the interleaved accesses and collect them in interleave 801 /// groups. Substitute symbolic strides using \p Strides. 802 void analyzeInterleaving(const ValueToValueMap &Strides); 803 804 /// \brief Check if \p Instr belongs to any interleave group. 805 bool isInterleaved(Instruction *Instr) const { 806 return InterleaveGroupMap.count(Instr); 807 } 808 809 /// \brief Get the interleave group that \p Instr belongs to. 810 /// 811 /// \returns nullptr if doesn't have such group. 812 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const { 813 if (InterleaveGroupMap.count(Instr)) 814 return InterleaveGroupMap.find(Instr)->second; 815 return nullptr; 816 } 817 818 private: 819 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks. 820 /// Simplifies SCEV expressions in the context of existing SCEV assumptions. 821 /// The interleaved access analysis can also add new predicates (for example 822 /// by versioning strides of pointers). 823 PredicatedScalarEvolution &PSE; 824 Loop *TheLoop; 825 DominatorTree *DT; 826 827 /// Holds the relationships between the members and the interleave group. 828 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap; 829 830 /// \brief The descriptor for a strided memory access. 831 struct StrideDescriptor { 832 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size, 833 unsigned Align) 834 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {} 835 836 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {} 837 838 int Stride; // The access's stride. It is negative for a reverse access. 839 const SCEV *Scev; // The scalar expression of this access 840 unsigned Size; // The size of the memory object. 841 unsigned Align; // The alignment of this access. 842 }; 843 844 /// \brief Create a new interleave group with the given instruction \p Instr, 845 /// stride \p Stride and alignment \p Align. 846 /// 847 /// \returns the newly created interleave group. 848 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride, 849 unsigned Align) { 850 assert(!InterleaveGroupMap.count(Instr) && 851 "Already in an interleaved access group"); 852 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align); 853 return InterleaveGroupMap[Instr]; 854 } 855 856 /// \brief Release the group and remove all the relationships. 857 void releaseGroup(InterleaveGroup *Group) { 858 for (unsigned i = 0; i < Group->getFactor(); i++) 859 if (Instruction *Member = Group->getMember(i)) 860 InterleaveGroupMap.erase(Member); 861 862 delete Group; 863 } 864 865 /// \brief Collect all the accesses with a constant stride in program order. 866 void collectConstStridedAccesses( 867 MapVector<Instruction *, StrideDescriptor> &StrideAccesses, 868 const ValueToValueMap &Strides); 869 }; 870 871 /// Utility class for getting and setting loop vectorizer hints in the form 872 /// of loop metadata. 873 /// This class keeps a number of loop annotations locally (as member variables) 874 /// and can, upon request, write them back as metadata on the loop. It will 875 /// initially scan the loop for existing metadata, and will update the local 876 /// values based on information in the loop. 877 /// We cannot write all values to metadata, as the mere presence of some info, 878 /// for example 'force', means a decision has been made. So, we need to be 879 /// careful NOT to add them if the user hasn't specifically asked so. 880 class LoopVectorizeHints { 881 enum HintKind { 882 HK_WIDTH, 883 HK_UNROLL, 884 HK_FORCE 885 }; 886 887 /// Hint - associates name and validation with the hint value. 888 struct Hint { 889 const char * Name; 890 unsigned Value; // This may have to change for non-numeric values. 891 HintKind Kind; 892 893 Hint(const char * Name, unsigned Value, HintKind Kind) 894 : Name(Name), Value(Value), Kind(Kind) { } 895 896 bool validate(unsigned Val) { 897 switch (Kind) { 898 case HK_WIDTH: 899 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth; 900 case HK_UNROLL: 901 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor; 902 case HK_FORCE: 903 return (Val <= 1); 904 } 905 return false; 906 } 907 }; 908 909 /// Vectorization width. 910 Hint Width; 911 /// Vectorization interleave factor. 912 Hint Interleave; 913 /// Vectorization forced 914 Hint Force; 915 916 /// Return the loop metadata prefix. 917 static StringRef Prefix() { return "llvm.loop."; } 918 919 public: 920 enum ForceKind { 921 FK_Undefined = -1, ///< Not selected. 922 FK_Disabled = 0, ///< Forcing disabled. 923 FK_Enabled = 1, ///< Forcing enabled. 924 }; 925 926 LoopVectorizeHints(const Loop *L, bool DisableInterleaving) 927 : Width("vectorize.width", VectorizerParams::VectorizationFactor, 928 HK_WIDTH), 929 Interleave("interleave.count", DisableInterleaving, HK_UNROLL), 930 Force("vectorize.enable", FK_Undefined, HK_FORCE), 931 TheLoop(L) { 932 // Populate values with existing loop metadata. 933 getHintsFromMetadata(); 934 935 // force-vector-interleave overrides DisableInterleaving. 936 if (VectorizerParams::isInterleaveForced()) 937 Interleave.Value = VectorizerParams::VectorizationInterleave; 938 939 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs() 940 << "LV: Interleaving disabled by the pass manager\n"); 941 } 942 943 /// Mark the loop L as already vectorized by setting the width to 1. 944 void setAlreadyVectorized() { 945 Width.Value = Interleave.Value = 1; 946 Hint Hints[] = {Width, Interleave}; 947 writeHintsToMetadata(Hints); 948 } 949 950 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const { 951 if (getForce() == LoopVectorizeHints::FK_Disabled) { 952 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n"); 953 emitOptimizationRemarkAnalysis(F->getContext(), 954 vectorizeAnalysisPassName(), *F, 955 L->getStartLoc(), emitRemark()); 956 return false; 957 } 958 959 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) { 960 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n"); 961 emitOptimizationRemarkAnalysis(F->getContext(), 962 vectorizeAnalysisPassName(), *F, 963 L->getStartLoc(), emitRemark()); 964 return false; 965 } 966 967 if (getWidth() == 1 && getInterleave() == 1) { 968 // FIXME: Add a separate metadata to indicate when the loop has already 969 // been vectorized instead of setting width and count to 1. 970 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n"); 971 // FIXME: Add interleave.disable metadata. This will allow 972 // vectorize.disable to be used without disabling the pass and errors 973 // to differentiate between disabled vectorization and a width of 1. 974 emitOptimizationRemarkAnalysis( 975 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(), 976 "loop not vectorized: vectorization and interleaving are explicitly " 977 "disabled, or vectorize width and interleave count are both set to " 978 "1"); 979 return false; 980 } 981 982 return true; 983 } 984 985 /// Dumps all the hint information. 986 std::string emitRemark() const { 987 VectorizationReport R; 988 if (Force.Value == LoopVectorizeHints::FK_Disabled) 989 R << "vectorization is explicitly disabled"; 990 else { 991 R << "use -Rpass-analysis=loop-vectorize for more info"; 992 if (Force.Value == LoopVectorizeHints::FK_Enabled) { 993 R << " (Force=true"; 994 if (Width.Value != 0) 995 R << ", Vector Width=" << Width.Value; 996 if (Interleave.Value != 0) 997 R << ", Interleave Count=" << Interleave.Value; 998 R << ")"; 999 } 1000 } 1001 1002 return R.str(); 1003 } 1004 1005 unsigned getWidth() const { return Width.Value; } 1006 unsigned getInterleave() const { return Interleave.Value; } 1007 enum ForceKind getForce() const { return (ForceKind)Force.Value; } 1008 const char *vectorizeAnalysisPassName() const { 1009 // If hints are provided that don't disable vectorization use the 1010 // AlwaysPrint pass name to force the frontend to print the diagnostic. 1011 if (getWidth() == 1) 1012 return LV_NAME; 1013 if (getForce() == LoopVectorizeHints::FK_Disabled) 1014 return LV_NAME; 1015 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0) 1016 return LV_NAME; 1017 return DiagnosticInfo::AlwaysPrint; 1018 } 1019 1020 bool allowReordering() const { 1021 // When enabling loop hints are provided we allow the vectorizer to change 1022 // the order of operations that is given by the scalar loop. This is not 1023 // enabled by default because can be unsafe or inefficient. For example, 1024 // reordering floating-point operations will change the way round-off 1025 // error accumulates in the loop. 1026 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1; 1027 } 1028 1029 private: 1030 /// Find hints specified in the loop metadata and update local values. 1031 void getHintsFromMetadata() { 1032 MDNode *LoopID = TheLoop->getLoopID(); 1033 if (!LoopID) 1034 return; 1035 1036 // First operand should refer to the loop id itself. 1037 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 1038 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 1039 1040 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 1041 const MDString *S = nullptr; 1042 SmallVector<Metadata *, 4> Args; 1043 1044 // The expected hint is either a MDString or a MDNode with the first 1045 // operand a MDString. 1046 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) { 1047 if (!MD || MD->getNumOperands() == 0) 1048 continue; 1049 S = dyn_cast<MDString>(MD->getOperand(0)); 1050 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i) 1051 Args.push_back(MD->getOperand(i)); 1052 } else { 1053 S = dyn_cast<MDString>(LoopID->getOperand(i)); 1054 assert(Args.size() == 0 && "too many arguments for MDString"); 1055 } 1056 1057 if (!S) 1058 continue; 1059 1060 // Check if the hint starts with the loop metadata prefix. 1061 StringRef Name = S->getString(); 1062 if (Args.size() == 1) 1063 setHint(Name, Args[0]); 1064 } 1065 } 1066 1067 /// Checks string hint with one operand and set value if valid. 1068 void setHint(StringRef Name, Metadata *Arg) { 1069 if (!Name.startswith(Prefix())) 1070 return; 1071 Name = Name.substr(Prefix().size(), StringRef::npos); 1072 1073 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg); 1074 if (!C) return; 1075 unsigned Val = C->getZExtValue(); 1076 1077 Hint *Hints[] = {&Width, &Interleave, &Force}; 1078 for (auto H : Hints) { 1079 if (Name == H->Name) { 1080 if (H->validate(Val)) 1081 H->Value = Val; 1082 else 1083 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n"); 1084 break; 1085 } 1086 } 1087 } 1088 1089 /// Create a new hint from name / value pair. 1090 MDNode *createHintMetadata(StringRef Name, unsigned V) const { 1091 LLVMContext &Context = TheLoop->getHeader()->getContext(); 1092 Metadata *MDs[] = {MDString::get(Context, Name), 1093 ConstantAsMetadata::get( 1094 ConstantInt::get(Type::getInt32Ty(Context), V))}; 1095 return MDNode::get(Context, MDs); 1096 } 1097 1098 /// Matches metadata with hint name. 1099 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) { 1100 MDString* Name = dyn_cast<MDString>(Node->getOperand(0)); 1101 if (!Name) 1102 return false; 1103 1104 for (auto H : HintTypes) 1105 if (Name->getString().endswith(H.Name)) 1106 return true; 1107 return false; 1108 } 1109 1110 /// Sets current hints into loop metadata, keeping other values intact. 1111 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) { 1112 if (HintTypes.size() == 0) 1113 return; 1114 1115 // Reserve the first element to LoopID (see below). 1116 SmallVector<Metadata *, 4> MDs(1); 1117 // If the loop already has metadata, then ignore the existing operands. 1118 MDNode *LoopID = TheLoop->getLoopID(); 1119 if (LoopID) { 1120 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 1121 MDNode *Node = cast<MDNode>(LoopID->getOperand(i)); 1122 // If node in update list, ignore old value. 1123 if (!matchesHintMetadataName(Node, HintTypes)) 1124 MDs.push_back(Node); 1125 } 1126 } 1127 1128 // Now, add the missing hints. 1129 for (auto H : HintTypes) 1130 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value)); 1131 1132 // Replace current metadata node with new one. 1133 LLVMContext &Context = TheLoop->getHeader()->getContext(); 1134 MDNode *NewLoopID = MDNode::get(Context, MDs); 1135 // Set operand 0 to refer to the loop id itself. 1136 NewLoopID->replaceOperandWith(0, NewLoopID); 1137 1138 TheLoop->setLoopID(NewLoopID); 1139 } 1140 1141 /// The loop these hints belong to. 1142 const Loop *TheLoop; 1143 }; 1144 1145 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop, 1146 const LoopVectorizeHints &Hints, 1147 const LoopAccessReport &Message) { 1148 const char *Name = Hints.vectorizeAnalysisPassName(); 1149 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name); 1150 } 1151 1152 static void emitMissedWarning(Function *F, Loop *L, 1153 const LoopVectorizeHints &LH) { 1154 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(), 1155 LH.emitRemark()); 1156 1157 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) { 1158 if (LH.getWidth() != 1) 1159 emitLoopVectorizeWarning( 1160 F->getContext(), *F, L->getStartLoc(), 1161 "failed explicitly specified loop vectorization"); 1162 else if (LH.getInterleave() != 1) 1163 emitLoopInterleaveWarning( 1164 F->getContext(), *F, L->getStartLoc(), 1165 "failed explicitly specified loop interleaving"); 1166 } 1167 } 1168 1169 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and 1170 /// to what vectorization factor. 1171 /// This class does not look at the profitability of vectorization, only the 1172 /// legality. This class has two main kinds of checks: 1173 /// * Memory checks - The code in canVectorizeMemory checks if vectorization 1174 /// will change the order of memory accesses in a way that will change the 1175 /// correctness of the program. 1176 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory 1177 /// checks for a number of different conditions, such as the availability of a 1178 /// single induction variable, that all types are supported and vectorize-able, 1179 /// etc. This code reflects the capabilities of InnerLoopVectorizer. 1180 /// This class is also used by InnerLoopVectorizer for identifying 1181 /// induction variable and the different reduction variables. 1182 class LoopVectorizationLegality { 1183 public: 1184 LoopVectorizationLegality(Loop *L, PredicatedScalarEvolution &PSE, 1185 DominatorTree *DT, TargetLibraryInfo *TLI, 1186 AliasAnalysis *AA, Function *F, 1187 const TargetTransformInfo *TTI, 1188 LoopAccessAnalysis *LAA, 1189 LoopVectorizationRequirements *R, 1190 const LoopVectorizeHints *H) 1191 : NumPredStores(0), TheLoop(L), PSE(PSE), TLI(TLI), TheFunction(F), 1192 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(PSE, L, DT), 1193 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false), 1194 Requirements(R), Hints(H) {} 1195 1196 /// ReductionList contains the reduction descriptors for all 1197 /// of the reductions that were found in the loop. 1198 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList; 1199 1200 /// InductionList saves induction variables and maps them to the 1201 /// induction descriptor. 1202 typedef MapVector<PHINode*, InductionDescriptor> InductionList; 1203 1204 /// Returns true if it is legal to vectorize this loop. 1205 /// This does not mean that it is profitable to vectorize this 1206 /// loop, only that it is legal to do so. 1207 bool canVectorize(); 1208 1209 /// Returns the Induction variable. 1210 PHINode *getInduction() { return Induction; } 1211 1212 /// Returns the reduction variables found in the loop. 1213 ReductionList *getReductionVars() { return &Reductions; } 1214 1215 /// Returns the induction variables found in the loop. 1216 InductionList *getInductionVars() { return &Inductions; } 1217 1218 /// Returns the widest induction type. 1219 Type *getWidestInductionType() { return WidestIndTy; } 1220 1221 /// Returns True if V is an induction variable in this loop. 1222 bool isInductionVariable(const Value *V); 1223 1224 /// Returns True if PN is a reduction variable in this loop. 1225 bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); } 1226 1227 /// Return true if the block BB needs to be predicated in order for the loop 1228 /// to be vectorized. 1229 bool blockNeedsPredication(BasicBlock *BB); 1230 1231 /// Check if this pointer is consecutive when vectorizing. This happens 1232 /// when the last index of the GEP is the induction variable, or that the 1233 /// pointer itself is an induction variable. 1234 /// This check allows us to vectorize A[idx] into a wide load/store. 1235 /// Returns: 1236 /// 0 - Stride is unknown or non-consecutive. 1237 /// 1 - Address is consecutive. 1238 /// -1 - Address is consecutive, and decreasing. 1239 int isConsecutivePtr(Value *Ptr); 1240 1241 /// Returns true if the value V is uniform within the loop. 1242 bool isUniform(Value *V); 1243 1244 /// Returns true if this instruction will remain scalar after vectorization. 1245 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } 1246 1247 /// Returns the information that we collected about runtime memory check. 1248 const RuntimePointerChecking *getRuntimePointerChecking() const { 1249 return LAI->getRuntimePointerChecking(); 1250 } 1251 1252 const LoopAccessInfo *getLAI() const { 1253 return LAI; 1254 } 1255 1256 /// \brief Check if \p Instr belongs to any interleaved access group. 1257 bool isAccessInterleaved(Instruction *Instr) { 1258 return InterleaveInfo.isInterleaved(Instr); 1259 } 1260 1261 /// \brief Get the interleaved access group that \p Instr belongs to. 1262 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) { 1263 return InterleaveInfo.getInterleaveGroup(Instr); 1264 } 1265 1266 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); } 1267 1268 bool hasStride(Value *V) { return StrideSet.count(V); } 1269 bool mustCheckStrides() { return !StrideSet.empty(); } 1270 SmallPtrSet<Value *, 8>::iterator strides_begin() { 1271 return StrideSet.begin(); 1272 } 1273 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); } 1274 1275 /// Returns true if the target machine supports masked store operation 1276 /// for the given \p DataType and kind of access to \p Ptr. 1277 bool isLegalMaskedStore(Type *DataType, Value *Ptr) { 1278 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType); 1279 } 1280 /// Returns true if the target machine supports masked load operation 1281 /// for the given \p DataType and kind of access to \p Ptr. 1282 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) { 1283 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType); 1284 } 1285 /// Returns true if vector representation of the instruction \p I 1286 /// requires mask. 1287 bool isMaskRequired(const Instruction* I) { 1288 return (MaskedOp.count(I) != 0); 1289 } 1290 unsigned getNumStores() const { 1291 return LAI->getNumStores(); 1292 } 1293 unsigned getNumLoads() const { 1294 return LAI->getNumLoads(); 1295 } 1296 unsigned getNumPredStores() const { 1297 return NumPredStores; 1298 } 1299 private: 1300 /// Check if a single basic block loop is vectorizable. 1301 /// At this point we know that this is a loop with a constant trip count 1302 /// and we only need to check individual instructions. 1303 bool canVectorizeInstrs(); 1304 1305 /// When we vectorize loops we may change the order in which 1306 /// we read and write from memory. This method checks if it is 1307 /// legal to vectorize the code, considering only memory constrains. 1308 /// Returns true if the loop is vectorizable 1309 bool canVectorizeMemory(); 1310 1311 /// Return true if we can vectorize this loop using the IF-conversion 1312 /// transformation. 1313 bool canVectorizeWithIfConvert(); 1314 1315 /// Collect the variables that need to stay uniform after vectorization. 1316 void collectLoopUniforms(); 1317 1318 /// Return true if all of the instructions in the block can be speculatively 1319 /// executed. \p SafePtrs is a list of addresses that are known to be legal 1320 /// and we know that we can read from them without segfault. 1321 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs); 1322 1323 /// \brief Collect memory access with loop invariant strides. 1324 /// 1325 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop 1326 /// invariant. 1327 void collectStridedAccess(Value *LoadOrStoreInst); 1328 1329 /// Report an analysis message to assist the user in diagnosing loops that are 1330 /// not vectorized. These are handled as LoopAccessReport rather than 1331 /// VectorizationReport because the << operator of VectorizationReport returns 1332 /// LoopAccessReport. 1333 void emitAnalysis(const LoopAccessReport &Message) const { 1334 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message); 1335 } 1336 1337 unsigned NumPredStores; 1338 1339 /// The loop that we evaluate. 1340 Loop *TheLoop; 1341 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. 1342 /// Applies dynamic knowledge to simplify SCEV expressions in the context 1343 /// of existing SCEV assumptions. The analysis will also add a minimal set 1344 /// of new predicates if this is required to enable vectorization and 1345 /// unrolling. 1346 PredicatedScalarEvolution &PSE; 1347 /// Target Library Info. 1348 TargetLibraryInfo *TLI; 1349 /// Parent function 1350 Function *TheFunction; 1351 /// Target Transform Info 1352 const TargetTransformInfo *TTI; 1353 /// Dominator Tree. 1354 DominatorTree *DT; 1355 // LoopAccess analysis. 1356 LoopAccessAnalysis *LAA; 1357 // And the loop-accesses info corresponding to this loop. This pointer is 1358 // null until canVectorizeMemory sets it up. 1359 const LoopAccessInfo *LAI; 1360 1361 /// The interleave access information contains groups of interleaved accesses 1362 /// with the same stride and close to each other. 1363 InterleavedAccessInfo InterleaveInfo; 1364 1365 // --- vectorization state --- // 1366 1367 /// Holds the integer induction variable. This is the counter of the 1368 /// loop. 1369 PHINode *Induction; 1370 /// Holds the reduction variables. 1371 ReductionList Reductions; 1372 /// Holds all of the induction variables that we found in the loop. 1373 /// Notice that inductions don't need to start at zero and that induction 1374 /// variables can be pointers. 1375 InductionList Inductions; 1376 /// Holds the widest induction type encountered. 1377 Type *WidestIndTy; 1378 1379 /// Allowed outside users. This holds the reduction 1380 /// vars which can be accessed from outside the loop. 1381 SmallPtrSet<Value*, 4> AllowedExit; 1382 /// This set holds the variables which are known to be uniform after 1383 /// vectorization. 1384 SmallPtrSet<Instruction*, 4> Uniforms; 1385 1386 /// Can we assume the absence of NaNs. 1387 bool HasFunNoNaNAttr; 1388 1389 /// Vectorization requirements that will go through late-evaluation. 1390 LoopVectorizationRequirements *Requirements; 1391 1392 /// Used to emit an analysis of any legality issues. 1393 const LoopVectorizeHints *Hints; 1394 1395 ValueToValueMap Strides; 1396 SmallPtrSet<Value *, 8> StrideSet; 1397 1398 /// While vectorizing these instructions we have to generate a 1399 /// call to the appropriate masked intrinsic 1400 SmallPtrSet<const Instruction *, 8> MaskedOp; 1401 }; 1402 1403 /// LoopVectorizationCostModel - estimates the expected speedups due to 1404 /// vectorization. 1405 /// In many cases vectorization is not profitable. This can happen because of 1406 /// a number of reasons. In this class we mainly attempt to predict the 1407 /// expected speedup/slowdowns due to the supported instruction set. We use the 1408 /// TargetTransformInfo to query the different backends for the cost of 1409 /// different operations. 1410 class LoopVectorizationCostModel { 1411 public: 1412 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE, 1413 LoopInfo *LI, LoopVectorizationLegality *Legal, 1414 const TargetTransformInfo &TTI, 1415 const TargetLibraryInfo *TLI, DemandedBits *DB, 1416 AssumptionCache *AC, const Function *F, 1417 const LoopVectorizeHints *Hints) 1418 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB), 1419 AC(AC), TheFunction(F), Hints(Hints) {} 1420 1421 /// Information about vectorization costs 1422 struct VectorizationFactor { 1423 unsigned Width; // Vector width with best cost 1424 unsigned Cost; // Cost of the loop with that width 1425 }; 1426 /// \return The most profitable vectorization factor and the cost of that VF. 1427 /// This method checks every power of two up to VF. If UserVF is not ZERO 1428 /// then this vectorization factor will be selected if vectorization is 1429 /// possible. 1430 VectorizationFactor selectVectorizationFactor(bool OptForSize); 1431 1432 /// \return The size (in bits) of the smallest and widest types in the code 1433 /// that needs to be vectorized. We ignore values that remain scalar such as 1434 /// 64 bit loop indices. 1435 std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); 1436 1437 /// \return The desired interleave count. 1438 /// If interleave count has been specified by metadata it will be returned. 1439 /// Otherwise, the interleave count is computed and returned. VF and LoopCost 1440 /// are the selected vectorization factor and the cost of the selected VF. 1441 unsigned selectInterleaveCount(bool OptForSize, unsigned VF, 1442 unsigned LoopCost); 1443 1444 /// \return The most profitable unroll factor. 1445 /// This method finds the best unroll-factor based on register pressure and 1446 /// other parameters. VF and LoopCost are the selected vectorization factor 1447 /// and the cost of the selected VF. 1448 unsigned computeInterleaveCount(bool OptForSize, unsigned VF, 1449 unsigned LoopCost); 1450 1451 /// \brief A struct that represents some properties of the register usage 1452 /// of a loop. 1453 struct RegisterUsage { 1454 /// Holds the number of loop invariant values that are used in the loop. 1455 unsigned LoopInvariantRegs; 1456 /// Holds the maximum number of concurrent live intervals in the loop. 1457 unsigned MaxLocalUsers; 1458 /// Holds the number of instructions in the loop. 1459 unsigned NumInstructions; 1460 }; 1461 1462 /// \return Returns information about the register usages of the loop for the 1463 /// given vectorization factors. 1464 SmallVector<RegisterUsage, 8> 1465 calculateRegisterUsage(const SmallVector<unsigned, 8> &VFs); 1466 1467 /// Collect values we want to ignore in the cost model. 1468 void collectValuesToIgnore(); 1469 1470 private: 1471 /// Returns the expected execution cost. The unit of the cost does 1472 /// not matter because we use the 'cost' units to compare different 1473 /// vector widths. The cost that is returned is *not* normalized by 1474 /// the factor width. 1475 unsigned expectedCost(unsigned VF); 1476 1477 /// Returns the execution time cost of an instruction for a given vector 1478 /// width. Vector width of one means scalar. 1479 unsigned getInstructionCost(Instruction *I, unsigned VF); 1480 1481 /// Returns whether the instruction is a load or store and will be a emitted 1482 /// as a vector operation. 1483 bool isConsecutiveLoadOrStore(Instruction *I); 1484 1485 /// Report an analysis message to assist the user in diagnosing loops that are 1486 /// not vectorized. These are handled as LoopAccessReport rather than 1487 /// VectorizationReport because the << operator of VectorizationReport returns 1488 /// LoopAccessReport. 1489 void emitAnalysis(const LoopAccessReport &Message) const { 1490 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message); 1491 } 1492 1493 public: 1494 /// Map of scalar integer values to the smallest bitwidth they can be legally 1495 /// represented as. The vector equivalents of these values should be truncated 1496 /// to this type. 1497 MapVector<Instruction*,uint64_t> MinBWs; 1498 1499 /// The loop that we evaluate. 1500 Loop *TheLoop; 1501 /// Predicated scalar evolution analysis. 1502 PredicatedScalarEvolution &PSE; 1503 /// Loop Info analysis. 1504 LoopInfo *LI; 1505 /// Vectorization legality. 1506 LoopVectorizationLegality *Legal; 1507 /// Vector target information. 1508 const TargetTransformInfo &TTI; 1509 /// Target Library Info. 1510 const TargetLibraryInfo *TLI; 1511 /// Demanded bits analysis. 1512 DemandedBits *DB; 1513 /// Assumption cache. 1514 AssumptionCache *AC; 1515 const Function *TheFunction; 1516 /// Loop Vectorize Hint. 1517 const LoopVectorizeHints *Hints; 1518 /// Values to ignore in the cost model. 1519 SmallPtrSet<const Value *, 16> ValuesToIgnore; 1520 /// Values to ignore in the cost model when VF > 1. 1521 SmallPtrSet<const Value *, 16> VecValuesToIgnore; 1522 }; 1523 1524 /// \brief This holds vectorization requirements that must be verified late in 1525 /// the process. The requirements are set by legalize and costmodel. Once 1526 /// vectorization has been determined to be possible and profitable the 1527 /// requirements can be verified by looking for metadata or compiler options. 1528 /// For example, some loops require FP commutativity which is only allowed if 1529 /// vectorization is explicitly specified or if the fast-math compiler option 1530 /// has been provided. 1531 /// Late evaluation of these requirements allows helpful diagnostics to be 1532 /// composed that tells the user what need to be done to vectorize the loop. For 1533 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late 1534 /// evaluation should be used only when diagnostics can generated that can be 1535 /// followed by a non-expert user. 1536 class LoopVectorizationRequirements { 1537 public: 1538 LoopVectorizationRequirements() 1539 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {} 1540 1541 void addUnsafeAlgebraInst(Instruction *I) { 1542 // First unsafe algebra instruction. 1543 if (!UnsafeAlgebraInst) 1544 UnsafeAlgebraInst = I; 1545 } 1546 1547 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; } 1548 1549 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) { 1550 const char *Name = Hints.vectorizeAnalysisPassName(); 1551 bool Failed = false; 1552 if (UnsafeAlgebraInst && !Hints.allowReordering()) { 1553 emitOptimizationRemarkAnalysisFPCommute( 1554 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(), 1555 VectorizationReport() << "cannot prove it is safe to reorder " 1556 "floating-point operations"); 1557 Failed = true; 1558 } 1559 1560 // Test if runtime memcheck thresholds are exceeded. 1561 bool PragmaThresholdReached = 1562 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold; 1563 bool ThresholdReached = 1564 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold; 1565 if ((ThresholdReached && !Hints.allowReordering()) || 1566 PragmaThresholdReached) { 1567 emitOptimizationRemarkAnalysisAliasing( 1568 F->getContext(), Name, *F, L->getStartLoc(), 1569 VectorizationReport() 1570 << "cannot prove it is safe to reorder memory operations"); 1571 DEBUG(dbgs() << "LV: Too many memory checks needed.\n"); 1572 Failed = true; 1573 } 1574 1575 return Failed; 1576 } 1577 1578 private: 1579 unsigned NumRuntimePointerChecks; 1580 Instruction *UnsafeAlgebraInst; 1581 }; 1582 1583 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) { 1584 if (L.empty()) 1585 return V.push_back(&L); 1586 1587 for (Loop *InnerL : L) 1588 addInnerLoop(*InnerL, V); 1589 } 1590 1591 /// The LoopVectorize Pass. 1592 struct LoopVectorize : public FunctionPass { 1593 /// Pass identification, replacement for typeid 1594 static char ID; 1595 1596 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true) 1597 : FunctionPass(ID), 1598 DisableUnrolling(NoUnrolling), 1599 AlwaysVectorize(AlwaysVectorize) { 1600 initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); 1601 } 1602 1603 ScalarEvolution *SE; 1604 LoopInfo *LI; 1605 TargetTransformInfo *TTI; 1606 DominatorTree *DT; 1607 BlockFrequencyInfo *BFI; 1608 TargetLibraryInfo *TLI; 1609 DemandedBits *DB; 1610 AliasAnalysis *AA; 1611 AssumptionCache *AC; 1612 LoopAccessAnalysis *LAA; 1613 bool DisableUnrolling; 1614 bool AlwaysVectorize; 1615 1616 BlockFrequency ColdEntryFreq; 1617 1618 bool runOnFunction(Function &F) override { 1619 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 1620 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 1621 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 1622 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1623 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); 1624 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 1625 TLI = TLIP ? &TLIP->getTLI() : nullptr; 1626 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 1627 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1628 LAA = &getAnalysis<LoopAccessAnalysis>(); 1629 DB = &getAnalysis<DemandedBits>(); 1630 1631 // Compute some weights outside of the loop over the loops. Compute this 1632 // using a BranchProbability to re-use its scaling math. 1633 const BranchProbability ColdProb(1, 5); // 20% 1634 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb; 1635 1636 // Don't attempt if 1637 // 1. the target claims to have no vector registers, and 1638 // 2. interleaving won't help ILP. 1639 // 1640 // The second condition is necessary because, even if the target has no 1641 // vector registers, loop vectorization may still enable scalar 1642 // interleaving. 1643 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2) 1644 return false; 1645 1646 // Build up a worklist of inner-loops to vectorize. This is necessary as 1647 // the act of vectorizing or partially unrolling a loop creates new loops 1648 // and can invalidate iterators across the loops. 1649 SmallVector<Loop *, 8> Worklist; 1650 1651 for (Loop *L : *LI) 1652 addInnerLoop(*L, Worklist); 1653 1654 LoopsAnalyzed += Worklist.size(); 1655 1656 // Now walk the identified inner loops. 1657 bool Changed = false; 1658 while (!Worklist.empty()) 1659 Changed |= processLoop(Worklist.pop_back_val()); 1660 1661 // Process each loop nest in the function. 1662 return Changed; 1663 } 1664 1665 static void AddRuntimeUnrollDisableMetaData(Loop *L) { 1666 SmallVector<Metadata *, 4> MDs; 1667 // Reserve first location for self reference to the LoopID metadata node. 1668 MDs.push_back(nullptr); 1669 bool IsUnrollMetadata = false; 1670 MDNode *LoopID = L->getLoopID(); 1671 if (LoopID) { 1672 // First find existing loop unrolling disable metadata. 1673 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 1674 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 1675 if (MD) { 1676 const MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 1677 IsUnrollMetadata = 1678 S && S->getString().startswith("llvm.loop.unroll.disable"); 1679 } 1680 MDs.push_back(LoopID->getOperand(i)); 1681 } 1682 } 1683 1684 if (!IsUnrollMetadata) { 1685 // Add runtime unroll disable metadata. 1686 LLVMContext &Context = L->getHeader()->getContext(); 1687 SmallVector<Metadata *, 1> DisableOperands; 1688 DisableOperands.push_back( 1689 MDString::get(Context, "llvm.loop.unroll.runtime.disable")); 1690 MDNode *DisableNode = MDNode::get(Context, DisableOperands); 1691 MDs.push_back(DisableNode); 1692 MDNode *NewLoopID = MDNode::get(Context, MDs); 1693 // Set operand 0 to refer to the loop id itself. 1694 NewLoopID->replaceOperandWith(0, NewLoopID); 1695 L->setLoopID(NewLoopID); 1696 } 1697 } 1698 1699 bool processLoop(Loop *L) { 1700 assert(L->empty() && "Only process inner loops."); 1701 1702 #ifndef NDEBUG 1703 const std::string DebugLocStr = getDebugLocString(L); 1704 #endif /* NDEBUG */ 1705 1706 DEBUG(dbgs() << "\nLV: Checking a loop in \"" 1707 << L->getHeader()->getParent()->getName() << "\" from " 1708 << DebugLocStr << "\n"); 1709 1710 LoopVectorizeHints Hints(L, DisableUnrolling); 1711 1712 DEBUG(dbgs() << "LV: Loop hints:" 1713 << " force=" 1714 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled 1715 ? "disabled" 1716 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled 1717 ? "enabled" 1718 : "?")) << " width=" << Hints.getWidth() 1719 << " unroll=" << Hints.getInterleave() << "\n"); 1720 1721 // Function containing loop 1722 Function *F = L->getHeader()->getParent(); 1723 1724 // Looking at the diagnostic output is the only way to determine if a loop 1725 // was vectorized (other than looking at the IR or machine code), so it 1726 // is important to generate an optimization remark for each loop. Most of 1727 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks 1728 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are 1729 // less verbose reporting vectorized loops and unvectorized loops that may 1730 // benefit from vectorization, respectively. 1731 1732 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) { 1733 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n"); 1734 return false; 1735 } 1736 1737 // Check the loop for a trip count threshold: 1738 // do not vectorize loops with a tiny trip count. 1739 const unsigned TC = SE->getSmallConstantTripCount(L); 1740 if (TC > 0u && TC < TinyTripCountVectorThreshold) { 1741 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " 1742 << "This loop is not worth vectorizing."); 1743 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) 1744 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n"); 1745 else { 1746 DEBUG(dbgs() << "\n"); 1747 emitAnalysisDiag(F, L, Hints, VectorizationReport() 1748 << "vectorization is not beneficial " 1749 "and is not explicitly forced"); 1750 return false; 1751 } 1752 } 1753 1754 PredicatedScalarEvolution PSE(*SE); 1755 1756 // Check if it is legal to vectorize the loop. 1757 LoopVectorizationRequirements Requirements; 1758 LoopVectorizationLegality LVL(L, PSE, DT, TLI, AA, F, TTI, LAA, 1759 &Requirements, &Hints); 1760 if (!LVL.canVectorize()) { 1761 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n"); 1762 emitMissedWarning(F, L, Hints); 1763 return false; 1764 } 1765 1766 // Use the cost model. 1767 LoopVectorizationCostModel CM(L, PSE, LI, &LVL, *TTI, TLI, DB, AC, F, 1768 &Hints); 1769 CM.collectValuesToIgnore(); 1770 1771 // Check the function attributes to find out if this function should be 1772 // optimized for size. 1773 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled && 1774 F->optForSize(); 1775 1776 // Compute the weighted frequency of this loop being executed and see if it 1777 // is less than 20% of the function entry baseline frequency. Note that we 1778 // always have a canonical loop here because we think we *can* vectorize. 1779 // FIXME: This is hidden behind a flag due to pervasive problems with 1780 // exactly what block frequency models. 1781 if (LoopVectorizeWithBlockFrequency) { 1782 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader()); 1783 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled && 1784 LoopEntryFreq < ColdEntryFreq) 1785 OptForSize = true; 1786 } 1787 1788 // Check the function attributes to see if implicit floats are allowed. 1789 // FIXME: This check doesn't seem possibly correct -- what if the loop is 1790 // an integer loop and the vector instructions selected are purely integer 1791 // vector instructions? 1792 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { 1793 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" 1794 "attribute is used.\n"); 1795 emitAnalysisDiag( 1796 F, L, Hints, 1797 VectorizationReport() 1798 << "loop not vectorized due to NoImplicitFloat attribute"); 1799 emitMissedWarning(F, L, Hints); 1800 return false; 1801 } 1802 1803 // Select the optimal vectorization factor. 1804 const LoopVectorizationCostModel::VectorizationFactor VF = 1805 CM.selectVectorizationFactor(OptForSize); 1806 1807 // Select the interleave count. 1808 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost); 1809 1810 // Get user interleave count. 1811 unsigned UserIC = Hints.getInterleave(); 1812 1813 // Identify the diagnostic messages that should be produced. 1814 std::string VecDiagMsg, IntDiagMsg; 1815 bool VectorizeLoop = true, InterleaveLoop = true; 1816 1817 if (Requirements.doesNotMeet(F, L, Hints)) { 1818 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization " 1819 "requirements.\n"); 1820 emitMissedWarning(F, L, Hints); 1821 return false; 1822 } 1823 1824 if (VF.Width == 1) { 1825 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); 1826 VecDiagMsg = 1827 "the cost-model indicates that vectorization is not beneficial"; 1828 VectorizeLoop = false; 1829 } 1830 1831 if (IC == 1 && UserIC <= 1) { 1832 // Tell the user interleaving is not beneficial. 1833 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n"); 1834 IntDiagMsg = 1835 "the cost-model indicates that interleaving is not beneficial"; 1836 InterleaveLoop = false; 1837 if (UserIC == 1) 1838 IntDiagMsg += 1839 " and is explicitly disabled or interleave count is set to 1"; 1840 } else if (IC > 1 && UserIC == 1) { 1841 // Tell the user interleaving is beneficial, but it explicitly disabled. 1842 DEBUG(dbgs() 1843 << "LV: Interleaving is beneficial but is explicitly disabled."); 1844 IntDiagMsg = "the cost-model indicates that interleaving is beneficial " 1845 "but is explicitly disabled or interleave count is set to 1"; 1846 InterleaveLoop = false; 1847 } 1848 1849 // Override IC if user provided an interleave count. 1850 IC = UserIC > 0 ? UserIC : IC; 1851 1852 // Emit diagnostic messages, if any. 1853 const char *VAPassName = Hints.vectorizeAnalysisPassName(); 1854 if (!VectorizeLoop && !InterleaveLoop) { 1855 // Do not vectorize or interleaving the loop. 1856 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F, 1857 L->getStartLoc(), VecDiagMsg); 1858 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F, 1859 L->getStartLoc(), IntDiagMsg); 1860 return false; 1861 } else if (!VectorizeLoop && InterleaveLoop) { 1862 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n'); 1863 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F, 1864 L->getStartLoc(), VecDiagMsg); 1865 } else if (VectorizeLoop && !InterleaveLoop) { 1866 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " 1867 << DebugLocStr << '\n'); 1868 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F, 1869 L->getStartLoc(), IntDiagMsg); 1870 } else if (VectorizeLoop && InterleaveLoop) { 1871 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " 1872 << DebugLocStr << '\n'); 1873 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n'); 1874 } 1875 1876 if (!VectorizeLoop) { 1877 assert(IC > 1 && "interleave count should not be 1 or 0"); 1878 // If we decided that it is not legal to vectorize the loop then 1879 // interleave it. 1880 InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, IC); 1881 Unroller.vectorize(&LVL, CM.MinBWs); 1882 1883 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(), 1884 Twine("interleaved loop (interleaved count: ") + 1885 Twine(IC) + ")"); 1886 } else { 1887 // If we decided that it is *legal* to vectorize the loop then do it. 1888 InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, VF.Width, IC); 1889 LB.vectorize(&LVL, CM.MinBWs); 1890 ++LoopsVectorized; 1891 1892 // Add metadata to disable runtime unrolling scalar loop when there's no 1893 // runtime check about strides and memory. Because at this situation, 1894 // scalar loop is rarely used not worthy to be unrolled. 1895 if (!LB.IsSafetyChecksAdded()) 1896 AddRuntimeUnrollDisableMetaData(L); 1897 1898 // Report the vectorization decision. 1899 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(), 1900 Twine("vectorized loop (vectorization width: ") + 1901 Twine(VF.Width) + ", interleaved count: " + 1902 Twine(IC) + ")"); 1903 } 1904 1905 // Mark the loop as already vectorized to avoid vectorizing again. 1906 Hints.setAlreadyVectorized(); 1907 1908 DEBUG(verifyFunction(*L->getHeader()->getParent())); 1909 return true; 1910 } 1911 1912 void getAnalysisUsage(AnalysisUsage &AU) const override { 1913 AU.addRequired<AssumptionCacheTracker>(); 1914 AU.addRequiredID(LoopSimplifyID); 1915 AU.addRequiredID(LCSSAID); 1916 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 1917 AU.addRequired<DominatorTreeWrapperPass>(); 1918 AU.addRequired<LoopInfoWrapperPass>(); 1919 AU.addRequired<ScalarEvolutionWrapperPass>(); 1920 AU.addRequired<TargetTransformInfoWrapperPass>(); 1921 AU.addRequired<AAResultsWrapperPass>(); 1922 AU.addRequired<LoopAccessAnalysis>(); 1923 AU.addRequired<DemandedBits>(); 1924 AU.addPreserved<LoopInfoWrapperPass>(); 1925 AU.addPreserved<DominatorTreeWrapperPass>(); 1926 AU.addPreserved<BasicAAWrapperPass>(); 1927 AU.addPreserved<AAResultsWrapperPass>(); 1928 AU.addPreserved<GlobalsAAWrapperPass>(); 1929 } 1930 1931 }; 1932 1933 } // end anonymous namespace 1934 1935 //===----------------------------------------------------------------------===// 1936 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and 1937 // LoopVectorizationCostModel. 1938 //===----------------------------------------------------------------------===// 1939 1940 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { 1941 // We need to place the broadcast of invariant variables outside the loop. 1942 Instruction *Instr = dyn_cast<Instruction>(V); 1943 bool NewInstr = 1944 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(), 1945 Instr->getParent()) != LoopVectorBody.end()); 1946 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr; 1947 1948 // Place the code for broadcasting invariant variables in the new preheader. 1949 IRBuilder<>::InsertPointGuard Guard(Builder); 1950 if (Invariant) 1951 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); 1952 1953 // Broadcast the scalar into all locations in the vector. 1954 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); 1955 1956 return Shuf; 1957 } 1958 1959 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, 1960 Value *Step) { 1961 assert(Val->getType()->isVectorTy() && "Must be a vector"); 1962 assert(Val->getType()->getScalarType()->isIntegerTy() && 1963 "Elem must be an integer"); 1964 assert(Step->getType() == Val->getType()->getScalarType() && 1965 "Step has wrong type"); 1966 // Create the types. 1967 Type *ITy = Val->getType()->getScalarType(); 1968 VectorType *Ty = cast<VectorType>(Val->getType()); 1969 int VLen = Ty->getNumElements(); 1970 SmallVector<Constant*, 8> Indices; 1971 1972 // Create a vector of consecutive numbers from zero to VF. 1973 for (int i = 0; i < VLen; ++i) 1974 Indices.push_back(ConstantInt::get(ITy, StartIdx + i)); 1975 1976 // Add the consecutive indices to the vector value. 1977 Constant *Cv = ConstantVector::get(Indices); 1978 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); 1979 Step = Builder.CreateVectorSplat(VLen, Step); 1980 assert(Step->getType() == Val->getType() && "Invalid step vec"); 1981 // FIXME: The newly created binary instructions should contain nsw/nuw flags, 1982 // which can be found from the original scalar operations. 1983 Step = Builder.CreateMul(Cv, Step); 1984 return Builder.CreateAdd(Val, Step, "induction"); 1985 } 1986 1987 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { 1988 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr"); 1989 auto *SE = PSE.getSE(); 1990 // Make sure that the pointer does not point to structs. 1991 if (Ptr->getType()->getPointerElementType()->isAggregateType()) 1992 return 0; 1993 1994 // If this value is a pointer induction variable we know it is consecutive. 1995 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr); 1996 if (Phi && Inductions.count(Phi)) { 1997 InductionDescriptor II = Inductions[Phi]; 1998 return II.getConsecutiveDirection(); 1999 } 2000 2001 GetElementPtrInst *Gep = getGEPInstruction(Ptr); 2002 if (!Gep) 2003 return 0; 2004 2005 unsigned NumOperands = Gep->getNumOperands(); 2006 Value *GpPtr = Gep->getPointerOperand(); 2007 // If this GEP value is a consecutive pointer induction variable and all of 2008 // the indices are constant then we know it is consecutive. We can 2009 Phi = dyn_cast<PHINode>(GpPtr); 2010 if (Phi && Inductions.count(Phi)) { 2011 2012 // Make sure that the pointer does not point to structs. 2013 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType()); 2014 if (GepPtrType->getElementType()->isAggregateType()) 2015 return 0; 2016 2017 // Make sure that all of the index operands are loop invariant. 2018 for (unsigned i = 1; i < NumOperands; ++i) 2019 if (!SE->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)), TheLoop)) 2020 return 0; 2021 2022 InductionDescriptor II = Inductions[Phi]; 2023 return II.getConsecutiveDirection(); 2024 } 2025 2026 unsigned InductionOperand = getGEPInductionOperand(Gep); 2027 2028 // Check that all of the gep indices are uniform except for our induction 2029 // operand. 2030 for (unsigned i = 0; i != NumOperands; ++i) 2031 if (i != InductionOperand && 2032 !SE->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)), TheLoop)) 2033 return 0; 2034 2035 // We can emit wide load/stores only if the last non-zero index is the 2036 // induction variable. 2037 const SCEV *Last = nullptr; 2038 if (!Strides.count(Gep)) 2039 Last = PSE.getSCEV(Gep->getOperand(InductionOperand)); 2040 else { 2041 // Because of the multiplication by a stride we can have a s/zext cast. 2042 // We are going to replace this stride by 1 so the cast is safe to ignore. 2043 // 2044 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 2045 // %0 = trunc i64 %indvars.iv to i32 2046 // %mul = mul i32 %0, %Stride1 2047 // %idxprom = zext i32 %mul to i64 << Safe cast. 2048 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom 2049 // 2050 Last = replaceSymbolicStrideSCEV(PSE, Strides, 2051 Gep->getOperand(InductionOperand), Gep); 2052 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last)) 2053 Last = 2054 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend) 2055 ? C->getOperand() 2056 : Last; 2057 } 2058 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { 2059 const SCEV *Step = AR->getStepRecurrence(*SE); 2060 2061 // The memory is consecutive because the last index is consecutive 2062 // and all other indices are loop invariant. 2063 if (Step->isOne()) 2064 return 1; 2065 if (Step->isAllOnesValue()) 2066 return -1; 2067 } 2068 2069 return 0; 2070 } 2071 2072 bool LoopVectorizationLegality::isUniform(Value *V) { 2073 return LAI->isUniform(V); 2074 } 2075 2076 InnerLoopVectorizer::VectorParts& 2077 InnerLoopVectorizer::getVectorValue(Value *V) { 2078 assert(V != Induction && "The new induction variable should not be used."); 2079 assert(!V->getType()->isVectorTy() && "Can't widen a vector"); 2080 2081 // If we have a stride that is replaced by one, do it here. 2082 if (Legal->hasStride(V)) 2083 V = ConstantInt::get(V->getType(), 1); 2084 2085 // If we have this scalar in the map, return it. 2086 if (WidenMap.has(V)) 2087 return WidenMap.get(V); 2088 2089 // If this scalar is unknown, assume that it is a constant or that it is 2090 // loop invariant. Broadcast V and save the value for future uses. 2091 Value *B = getBroadcastInstrs(V); 2092 return WidenMap.splat(V, B); 2093 } 2094 2095 Value *InnerLoopVectorizer::reverseVector(Value *Vec) { 2096 assert(Vec->getType()->isVectorTy() && "Invalid type"); 2097 SmallVector<Constant*, 8> ShuffleMask; 2098 for (unsigned i = 0; i < VF; ++i) 2099 ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); 2100 2101 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), 2102 ConstantVector::get(ShuffleMask), 2103 "reverse"); 2104 } 2105 2106 // Get a mask to interleave \p NumVec vectors into a wide vector. 2107 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...> 2108 // E.g. For 2 interleaved vectors, if VF is 4, the mask is: 2109 // <0, 4, 1, 5, 2, 6, 3, 7> 2110 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF, 2111 unsigned NumVec) { 2112 SmallVector<Constant *, 16> Mask; 2113 for (unsigned i = 0; i < VF; i++) 2114 for (unsigned j = 0; j < NumVec; j++) 2115 Mask.push_back(Builder.getInt32(j * VF + i)); 2116 2117 return ConstantVector::get(Mask); 2118 } 2119 2120 // Get the strided mask starting from index \p Start. 2121 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)> 2122 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start, 2123 unsigned Stride, unsigned VF) { 2124 SmallVector<Constant *, 16> Mask; 2125 for (unsigned i = 0; i < VF; i++) 2126 Mask.push_back(Builder.getInt32(Start + i * Stride)); 2127 2128 return ConstantVector::get(Mask); 2129 } 2130 2131 // Get a mask of two parts: The first part consists of sequential integers 2132 // starting from 0, The second part consists of UNDEFs. 2133 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef> 2134 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt, 2135 unsigned NumUndef) { 2136 SmallVector<Constant *, 16> Mask; 2137 for (unsigned i = 0; i < NumInt; i++) 2138 Mask.push_back(Builder.getInt32(i)); 2139 2140 Constant *Undef = UndefValue::get(Builder.getInt32Ty()); 2141 for (unsigned i = 0; i < NumUndef; i++) 2142 Mask.push_back(Undef); 2143 2144 return ConstantVector::get(Mask); 2145 } 2146 2147 // Concatenate two vectors with the same element type. The 2nd vector should 2148 // not have more elements than the 1st vector. If the 2nd vector has less 2149 // elements, extend it with UNDEFs. 2150 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1, 2151 Value *V2) { 2152 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType()); 2153 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType()); 2154 assert(VecTy1 && VecTy2 && 2155 VecTy1->getScalarType() == VecTy2->getScalarType() && 2156 "Expect two vectors with the same element type"); 2157 2158 unsigned NumElts1 = VecTy1->getNumElements(); 2159 unsigned NumElts2 = VecTy2->getNumElements(); 2160 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements"); 2161 2162 if (NumElts1 > NumElts2) { 2163 // Extend with UNDEFs. 2164 Constant *ExtMask = 2165 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2); 2166 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask); 2167 } 2168 2169 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0); 2170 return Builder.CreateShuffleVector(V1, V2, Mask); 2171 } 2172 2173 // Concatenate vectors in the given list. All vectors have the same type. 2174 static Value *ConcatenateVectors(IRBuilder<> &Builder, 2175 ArrayRef<Value *> InputList) { 2176 unsigned NumVec = InputList.size(); 2177 assert(NumVec > 1 && "Should be at least two vectors"); 2178 2179 SmallVector<Value *, 8> ResList; 2180 ResList.append(InputList.begin(), InputList.end()); 2181 do { 2182 SmallVector<Value *, 8> TmpList; 2183 for (unsigned i = 0; i < NumVec - 1; i += 2) { 2184 Value *V0 = ResList[i], *V1 = ResList[i + 1]; 2185 assert((V0->getType() == V1->getType() || i == NumVec - 2) && 2186 "Only the last vector may have a different type"); 2187 2188 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1)); 2189 } 2190 2191 // Push the last vector if the total number of vectors is odd. 2192 if (NumVec % 2 != 0) 2193 TmpList.push_back(ResList[NumVec - 1]); 2194 2195 ResList = TmpList; 2196 NumVec = ResList.size(); 2197 } while (NumVec > 1); 2198 2199 return ResList[0]; 2200 } 2201 2202 // Try to vectorize the interleave group that \p Instr belongs to. 2203 // 2204 // E.g. Translate following interleaved load group (factor = 3): 2205 // for (i = 0; i < N; i+=3) { 2206 // R = Pic[i]; // Member of index 0 2207 // G = Pic[i+1]; // Member of index 1 2208 // B = Pic[i+2]; // Member of index 2 2209 // ... // do something to R, G, B 2210 // } 2211 // To: 2212 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B 2213 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements 2214 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements 2215 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements 2216 // 2217 // Or translate following interleaved store group (factor = 3): 2218 // for (i = 0; i < N; i+=3) { 2219 // ... do something to R, G, B 2220 // Pic[i] = R; // Member of index 0 2221 // Pic[i+1] = G; // Member of index 1 2222 // Pic[i+2] = B; // Member of index 2 2223 // } 2224 // To: 2225 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7> 2226 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u> 2227 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec, 2228 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements 2229 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B 2230 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) { 2231 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr); 2232 assert(Group && "Fail to get an interleaved access group."); 2233 2234 // Skip if current instruction is not the insert position. 2235 if (Instr != Group->getInsertPos()) 2236 return; 2237 2238 LoadInst *LI = dyn_cast<LoadInst>(Instr); 2239 StoreInst *SI = dyn_cast<StoreInst>(Instr); 2240 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 2241 2242 // Prepare for the vector type of the interleaved load/store. 2243 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType(); 2244 unsigned InterleaveFactor = Group->getFactor(); 2245 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF); 2246 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace()); 2247 2248 // Prepare for the new pointers. 2249 setDebugLocFromInst(Builder, Ptr); 2250 VectorParts &PtrParts = getVectorValue(Ptr); 2251 SmallVector<Value *, 2> NewPtrs; 2252 unsigned Index = Group->getIndex(Instr); 2253 for (unsigned Part = 0; Part < UF; Part++) { 2254 // Extract the pointer for current instruction from the pointer vector. A 2255 // reverse access uses the pointer in the last lane. 2256 Value *NewPtr = Builder.CreateExtractElement( 2257 PtrParts[Part], 2258 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0)); 2259 2260 // Notice current instruction could be any index. Need to adjust the address 2261 // to the member of index 0. 2262 // 2263 // E.g. a = A[i+1]; // Member of index 1 (Current instruction) 2264 // b = A[i]; // Member of index 0 2265 // Current pointer is pointed to A[i+1], adjust it to A[i]. 2266 // 2267 // E.g. A[i+1] = a; // Member of index 1 2268 // A[i] = b; // Member of index 0 2269 // A[i+2] = c; // Member of index 2 (Current instruction) 2270 // Current pointer is pointed to A[i+2], adjust it to A[i]. 2271 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index)); 2272 2273 // Cast to the vector pointer type. 2274 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy)); 2275 } 2276 2277 setDebugLocFromInst(Builder, Instr); 2278 Value *UndefVec = UndefValue::get(VecTy); 2279 2280 // Vectorize the interleaved load group. 2281 if (LI) { 2282 for (unsigned Part = 0; Part < UF; Part++) { 2283 Instruction *NewLoadInstr = Builder.CreateAlignedLoad( 2284 NewPtrs[Part], Group->getAlignment(), "wide.vec"); 2285 2286 for (unsigned i = 0; i < InterleaveFactor; i++) { 2287 Instruction *Member = Group->getMember(i); 2288 2289 // Skip the gaps in the group. 2290 if (!Member) 2291 continue; 2292 2293 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF); 2294 Value *StridedVec = Builder.CreateShuffleVector( 2295 NewLoadInstr, UndefVec, StrideMask, "strided.vec"); 2296 2297 // If this member has different type, cast the result type. 2298 if (Member->getType() != ScalarTy) { 2299 VectorType *OtherVTy = VectorType::get(Member->getType(), VF); 2300 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy); 2301 } 2302 2303 VectorParts &Entry = WidenMap.get(Member); 2304 Entry[Part] = 2305 Group->isReverse() ? reverseVector(StridedVec) : StridedVec; 2306 } 2307 2308 propagateMetadata(NewLoadInstr, Instr); 2309 } 2310 return; 2311 } 2312 2313 // The sub vector type for current instruction. 2314 VectorType *SubVT = VectorType::get(ScalarTy, VF); 2315 2316 // Vectorize the interleaved store group. 2317 for (unsigned Part = 0; Part < UF; Part++) { 2318 // Collect the stored vector from each member. 2319 SmallVector<Value *, 4> StoredVecs; 2320 for (unsigned i = 0; i < InterleaveFactor; i++) { 2321 // Interleaved store group doesn't allow a gap, so each index has a member 2322 Instruction *Member = Group->getMember(i); 2323 assert(Member && "Fail to get a member from an interleaved store group"); 2324 2325 Value *StoredVec = 2326 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part]; 2327 if (Group->isReverse()) 2328 StoredVec = reverseVector(StoredVec); 2329 2330 // If this member has different type, cast it to an unified type. 2331 if (StoredVec->getType() != SubVT) 2332 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT); 2333 2334 StoredVecs.push_back(StoredVec); 2335 } 2336 2337 // Concatenate all vectors into a wide vector. 2338 Value *WideVec = ConcatenateVectors(Builder, StoredVecs); 2339 2340 // Interleave the elements in the wide vector. 2341 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor); 2342 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask, 2343 "interleaved.vec"); 2344 2345 Instruction *NewStoreInstr = 2346 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment()); 2347 propagateMetadata(NewStoreInstr, Instr); 2348 } 2349 } 2350 2351 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) { 2352 // Attempt to issue a wide load. 2353 LoadInst *LI = dyn_cast<LoadInst>(Instr); 2354 StoreInst *SI = dyn_cast<StoreInst>(Instr); 2355 2356 assert((LI || SI) && "Invalid Load/Store instruction"); 2357 2358 // Try to vectorize the interleave group if this access is interleaved. 2359 if (Legal->isAccessInterleaved(Instr)) 2360 return vectorizeInterleaveGroup(Instr); 2361 2362 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType(); 2363 Type *DataTy = VectorType::get(ScalarDataTy, VF); 2364 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 2365 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment(); 2366 // An alignment of 0 means target abi alignment. We need to use the scalar's 2367 // target abi alignment in such a case. 2368 const DataLayout &DL = Instr->getModule()->getDataLayout(); 2369 if (!Alignment) 2370 Alignment = DL.getABITypeAlignment(ScalarDataTy); 2371 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); 2372 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy); 2373 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF; 2374 2375 if (SI && Legal->blockNeedsPredication(SI->getParent()) && 2376 !Legal->isMaskRequired(SI)) 2377 return scalarizeInstruction(Instr, true); 2378 2379 if (ScalarAllocatedSize != VectorElementSize) 2380 return scalarizeInstruction(Instr); 2381 2382 // If the pointer is loop invariant or if it is non-consecutive, 2383 // scalarize the load. 2384 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 2385 bool Reverse = ConsecutiveStride < 0; 2386 bool UniformLoad = LI && Legal->isUniform(Ptr); 2387 if (!ConsecutiveStride || UniformLoad) 2388 return scalarizeInstruction(Instr); 2389 2390 Constant *Zero = Builder.getInt32(0); 2391 VectorParts &Entry = WidenMap.get(Instr); 2392 2393 // Handle consecutive loads/stores. 2394 GetElementPtrInst *Gep = getGEPInstruction(Ptr); 2395 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) { 2396 setDebugLocFromInst(Builder, Gep); 2397 Value *PtrOperand = Gep->getPointerOperand(); 2398 Value *FirstBasePtr = getVectorValue(PtrOperand)[0]; 2399 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero); 2400 2401 // Create the new GEP with the new induction variable. 2402 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 2403 Gep2->setOperand(0, FirstBasePtr); 2404 Gep2->setName("gep.indvar.base"); 2405 Ptr = Builder.Insert(Gep2); 2406 } else if (Gep) { 2407 setDebugLocFromInst(Builder, Gep); 2408 assert(PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getPointerOperand()), 2409 OrigLoop) && 2410 "Base ptr must be invariant"); 2411 2412 // The last index does not have to be the induction. It can be 2413 // consecutive and be a function of the index. For example A[I+1]; 2414 unsigned NumOperands = Gep->getNumOperands(); 2415 unsigned InductionOperand = getGEPInductionOperand(Gep); 2416 // Create the new GEP with the new induction variable. 2417 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 2418 2419 for (unsigned i = 0; i < NumOperands; ++i) { 2420 Value *GepOperand = Gep->getOperand(i); 2421 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand); 2422 2423 // Update last index or loop invariant instruction anchored in loop. 2424 if (i == InductionOperand || 2425 (GepOperandInst && OrigLoop->contains(GepOperandInst))) { 2426 assert((i == InductionOperand || 2427 PSE.getSE()->isLoopInvariant(PSE.getSCEV(GepOperandInst), 2428 OrigLoop)) && 2429 "Must be last index or loop invariant"); 2430 2431 VectorParts &GEPParts = getVectorValue(GepOperand); 2432 Value *Index = GEPParts[0]; 2433 Index = Builder.CreateExtractElement(Index, Zero); 2434 Gep2->setOperand(i, Index); 2435 Gep2->setName("gep.indvar.idx"); 2436 } 2437 } 2438 Ptr = Builder.Insert(Gep2); 2439 } else { 2440 // Use the induction element ptr. 2441 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 2442 setDebugLocFromInst(Builder, Ptr); 2443 VectorParts &PtrVal = getVectorValue(Ptr); 2444 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 2445 } 2446 2447 VectorParts Mask = createBlockInMask(Instr->getParent()); 2448 // Handle Stores: 2449 if (SI) { 2450 assert(!Legal->isUniform(SI->getPointerOperand()) && 2451 "We do not allow storing to uniform addresses"); 2452 setDebugLocFromInst(Builder, SI); 2453 // We don't want to update the value in the map as it might be used in 2454 // another expression. So don't use a reference type for "StoredVal". 2455 VectorParts StoredVal = getVectorValue(SI->getValueOperand()); 2456 2457 for (unsigned Part = 0; Part < UF; ++Part) { 2458 // Calculate the pointer for the specific unroll-part. 2459 Value *PartPtr = 2460 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF)); 2461 2462 if (Reverse) { 2463 // If we store to reverse consecutive memory locations, then we need 2464 // to reverse the order of elements in the stored value. 2465 StoredVal[Part] = reverseVector(StoredVal[Part]); 2466 // If the address is consecutive but reversed, then the 2467 // wide store needs to start at the last vector element. 2468 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF)); 2469 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF)); 2470 Mask[Part] = reverseVector(Mask[Part]); 2471 } 2472 2473 Value *VecPtr = Builder.CreateBitCast(PartPtr, 2474 DataTy->getPointerTo(AddressSpace)); 2475 2476 Instruction *NewSI; 2477 if (Legal->isMaskRequired(SI)) 2478 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment, 2479 Mask[Part]); 2480 else 2481 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment); 2482 propagateMetadata(NewSI, SI); 2483 } 2484 return; 2485 } 2486 2487 // Handle loads. 2488 assert(LI && "Must have a load instruction"); 2489 setDebugLocFromInst(Builder, LI); 2490 for (unsigned Part = 0; Part < UF; ++Part) { 2491 // Calculate the pointer for the specific unroll-part. 2492 Value *PartPtr = 2493 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF)); 2494 2495 if (Reverse) { 2496 // If the address is consecutive but reversed, then the 2497 // wide load needs to start at the last vector element. 2498 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF)); 2499 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF)); 2500 Mask[Part] = reverseVector(Mask[Part]); 2501 } 2502 2503 Instruction* NewLI; 2504 Value *VecPtr = Builder.CreateBitCast(PartPtr, 2505 DataTy->getPointerTo(AddressSpace)); 2506 if (Legal->isMaskRequired(LI)) 2507 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part], 2508 UndefValue::get(DataTy), 2509 "wide.masked.load"); 2510 else 2511 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load"); 2512 propagateMetadata(NewLI, LI); 2513 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI; 2514 } 2515 } 2516 2517 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, 2518 bool IfPredicateStore) { 2519 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 2520 // Holds vector parameters or scalars, in case of uniform vals. 2521 SmallVector<VectorParts, 4> Params; 2522 2523 setDebugLocFromInst(Builder, Instr); 2524 2525 // Find all of the vectorized parameters. 2526 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 2527 Value *SrcOp = Instr->getOperand(op); 2528 2529 // If we are accessing the old induction variable, use the new one. 2530 if (SrcOp == OldInduction) { 2531 Params.push_back(getVectorValue(SrcOp)); 2532 continue; 2533 } 2534 2535 // Try using previously calculated values. 2536 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 2537 2538 // If the src is an instruction that appeared earlier in the basic block, 2539 // then it should already be vectorized. 2540 if (SrcInst && OrigLoop->contains(SrcInst)) { 2541 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 2542 // The parameter is a vector value from earlier. 2543 Params.push_back(WidenMap.get(SrcInst)); 2544 } else { 2545 // The parameter is a scalar from outside the loop. Maybe even a constant. 2546 VectorParts Scalars; 2547 Scalars.append(UF, SrcOp); 2548 Params.push_back(Scalars); 2549 } 2550 } 2551 2552 assert(Params.size() == Instr->getNumOperands() && 2553 "Invalid number of operands"); 2554 2555 // Does this instruction return a value ? 2556 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 2557 2558 Value *UndefVec = IsVoidRetTy ? nullptr : 2559 UndefValue::get(VectorType::get(Instr->getType(), VF)); 2560 // Create a new entry in the WidenMap and initialize it to Undef or Null. 2561 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 2562 2563 VectorParts Cond; 2564 if (IfPredicateStore) { 2565 assert(Instr->getParent()->getSinglePredecessor() && 2566 "Only support single predecessor blocks"); 2567 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(), 2568 Instr->getParent()); 2569 } 2570 2571 // For each vector unroll 'part': 2572 for (unsigned Part = 0; Part < UF; ++Part) { 2573 // For each scalar that we create: 2574 for (unsigned Width = 0; Width < VF; ++Width) { 2575 2576 // Start if-block. 2577 Value *Cmp = nullptr; 2578 if (IfPredicateStore) { 2579 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width)); 2580 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, 2581 ConstantInt::get(Cmp->getType(), 1)); 2582 } 2583 2584 Instruction *Cloned = Instr->clone(); 2585 if (!IsVoidRetTy) 2586 Cloned->setName(Instr->getName() + ".cloned"); 2587 // Replace the operands of the cloned instructions with extracted scalars. 2588 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 2589 Value *Op = Params[op][Part]; 2590 // Param is a vector. Need to extract the right lane. 2591 if (Op->getType()->isVectorTy()) 2592 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); 2593 Cloned->setOperand(op, Op); 2594 } 2595 2596 // Place the cloned scalar in the new loop. 2597 Builder.Insert(Cloned); 2598 2599 // If the original scalar returns a value we need to place it in a vector 2600 // so that future users will be able to use it. 2601 if (!IsVoidRetTy) 2602 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, 2603 Builder.getInt32(Width)); 2604 // End if-block. 2605 if (IfPredicateStore) 2606 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned), 2607 Cmp)); 2608 } 2609 } 2610 } 2611 2612 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start, 2613 Value *End, Value *Step, 2614 Instruction *DL) { 2615 BasicBlock *Header = L->getHeader(); 2616 BasicBlock *Latch = L->getLoopLatch(); 2617 // As we're just creating this loop, it's possible no latch exists 2618 // yet. If so, use the header as this will be a single block loop. 2619 if (!Latch) 2620 Latch = Header; 2621 2622 IRBuilder<> Builder(&*Header->getFirstInsertionPt()); 2623 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction)); 2624 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index"); 2625 2626 Builder.SetInsertPoint(Latch->getTerminator()); 2627 2628 // Create i+1 and fill the PHINode. 2629 Value *Next = Builder.CreateAdd(Induction, Step, "index.next"); 2630 Induction->addIncoming(Start, L->getLoopPreheader()); 2631 Induction->addIncoming(Next, Latch); 2632 // Create the compare. 2633 Value *ICmp = Builder.CreateICmpEQ(Next, End); 2634 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header); 2635 2636 // Now we have two terminators. Remove the old one from the block. 2637 Latch->getTerminator()->eraseFromParent(); 2638 2639 return Induction; 2640 } 2641 2642 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) { 2643 if (TripCount) 2644 return TripCount; 2645 2646 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 2647 // Find the loop boundaries. 2648 ScalarEvolution *SE = PSE.getSE(); 2649 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(OrigLoop); 2650 assert(BackedgeTakenCount != SE->getCouldNotCompute() && 2651 "Invalid loop count"); 2652 2653 Type *IdxTy = Legal->getWidestInductionType(); 2654 2655 // The exit count might have the type of i64 while the phi is i32. This can 2656 // happen if we have an induction variable that is sign extended before the 2657 // compare. The only way that we get a backedge taken count is that the 2658 // induction variable was signed and as such will not overflow. In such a case 2659 // truncation is legal. 2660 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() > 2661 IdxTy->getPrimitiveSizeInBits()) 2662 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy); 2663 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); 2664 2665 // Get the total trip count from the count by adding 1. 2666 const SCEV *ExitCount = SE->getAddExpr( 2667 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); 2668 2669 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 2670 2671 // Expand the trip count and place the new instructions in the preheader. 2672 // Notice that the pre-header does not change, only the loop body. 2673 SCEVExpander Exp(*SE, DL, "induction"); 2674 2675 // Count holds the overall loop count (N). 2676 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 2677 L->getLoopPreheader()->getTerminator()); 2678 2679 if (TripCount->getType()->isPointerTy()) 2680 TripCount = 2681 CastInst::CreatePointerCast(TripCount, IdxTy, 2682 "exitcount.ptrcnt.to.int", 2683 L->getLoopPreheader()->getTerminator()); 2684 2685 return TripCount; 2686 } 2687 2688 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { 2689 if (VectorTripCount) 2690 return VectorTripCount; 2691 2692 Value *TC = getOrCreateTripCount(L); 2693 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 2694 2695 // Now we need to generate the expression for N - (N % VF), which is 2696 // the part that the vectorized body will execute. 2697 // The loop step is equal to the vectorization factor (num of SIMD elements) 2698 // times the unroll factor (num of SIMD instructions). 2699 Constant *Step = ConstantInt::get(TC->getType(), VF * UF); 2700 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf"); 2701 VectorTripCount = Builder.CreateSub(TC, R, "n.vec"); 2702 2703 return VectorTripCount; 2704 } 2705 2706 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, 2707 BasicBlock *Bypass) { 2708 Value *Count = getOrCreateTripCount(L); 2709 BasicBlock *BB = L->getLoopPreheader(); 2710 IRBuilder<> Builder(BB->getTerminator()); 2711 2712 // Generate code to check that the loop's trip count that we computed by 2713 // adding one to the backedge-taken count will not overflow. 2714 Value *CheckMinIters = 2715 Builder.CreateICmpULT(Count, 2716 ConstantInt::get(Count->getType(), VF * UF), 2717 "min.iters.check"); 2718 2719 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), 2720 "min.iters.checked"); 2721 if (L->getParentLoop()) 2722 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI); 2723 ReplaceInstWithInst(BB->getTerminator(), 2724 BranchInst::Create(Bypass, NewBB, CheckMinIters)); 2725 LoopBypassBlocks.push_back(BB); 2726 } 2727 2728 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L, 2729 BasicBlock *Bypass) { 2730 Value *TC = getOrCreateVectorTripCount(L); 2731 BasicBlock *BB = L->getLoopPreheader(); 2732 IRBuilder<> Builder(BB->getTerminator()); 2733 2734 // Now, compare the new count to zero. If it is zero skip the vector loop and 2735 // jump to the scalar loop. 2736 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()), 2737 "cmp.zero"); 2738 2739 // Generate code to check that the loop's trip count that we computed by 2740 // adding one to the backedge-taken count will not overflow. 2741 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), 2742 "vector.ph"); 2743 if (L->getParentLoop()) 2744 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI); 2745 ReplaceInstWithInst(BB->getTerminator(), 2746 BranchInst::Create(Bypass, NewBB, Cmp)); 2747 LoopBypassBlocks.push_back(BB); 2748 } 2749 2750 void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) { 2751 BasicBlock *BB = L->getLoopPreheader(); 2752 2753 // Generate the code to check that the SCEV assumptions that we made. 2754 // We want the new basic block to start at the first instruction in a 2755 // sequence of instructions that form a check. 2756 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(), 2757 "scev.check"); 2758 Value *SCEVCheck = 2759 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator()); 2760 2761 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck)) 2762 if (C->isZero()) 2763 return; 2764 2765 // Create a new block containing the stride check. 2766 BB->setName("vector.scevcheck"); 2767 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph"); 2768 if (L->getParentLoop()) 2769 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI); 2770 ReplaceInstWithInst(BB->getTerminator(), 2771 BranchInst::Create(Bypass, NewBB, SCEVCheck)); 2772 LoopBypassBlocks.push_back(BB); 2773 AddedSafetyChecks = true; 2774 } 2775 2776 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, 2777 BasicBlock *Bypass) { 2778 BasicBlock *BB = L->getLoopPreheader(); 2779 2780 // Generate the code that checks in runtime if arrays overlap. We put the 2781 // checks into a separate block to make the more common case of few elements 2782 // faster. 2783 Instruction *FirstCheckInst; 2784 Instruction *MemRuntimeCheck; 2785 std::tie(FirstCheckInst, MemRuntimeCheck) = 2786 Legal->getLAI()->addRuntimeChecks(BB->getTerminator()); 2787 if (!MemRuntimeCheck) 2788 return; 2789 2790 // Create a new block containing the memory check. 2791 BB->setName("vector.memcheck"); 2792 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph"); 2793 if (L->getParentLoop()) 2794 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI); 2795 ReplaceInstWithInst(BB->getTerminator(), 2796 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck)); 2797 LoopBypassBlocks.push_back(BB); 2798 AddedSafetyChecks = true; 2799 } 2800 2801 2802 void InnerLoopVectorizer::createEmptyLoop() { 2803 /* 2804 In this function we generate a new loop. The new loop will contain 2805 the vectorized instructions while the old loop will continue to run the 2806 scalar remainder. 2807 2808 [ ] <-- loop iteration number check. 2809 / | 2810 / v 2811 | [ ] <-- vector loop bypass (may consist of multiple blocks). 2812 | / | 2813 | / v 2814 || [ ] <-- vector pre header. 2815 |/ | 2816 | v 2817 | [ ] \ 2818 | [ ]_| <-- vector loop. 2819 | | 2820 | v 2821 | -[ ] <--- middle-block. 2822 | / | 2823 | / v 2824 -|- >[ ] <--- new preheader. 2825 | | 2826 | v 2827 | [ ] \ 2828 | [ ]_| <-- old scalar loop to handle remainder. 2829 \ | 2830 \ v 2831 >[ ] <-- exit block. 2832 ... 2833 */ 2834 2835 BasicBlock *OldBasicBlock = OrigLoop->getHeader(); 2836 BasicBlock *VectorPH = OrigLoop->getLoopPreheader(); 2837 BasicBlock *ExitBlock = OrigLoop->getExitBlock(); 2838 assert(VectorPH && "Invalid loop structure"); 2839 assert(ExitBlock && "Must have an exit block"); 2840 2841 // Some loops have a single integer induction variable, while other loops 2842 // don't. One example is c++ iterators that often have multiple pointer 2843 // induction variables. In the code below we also support a case where we 2844 // don't have a single induction variable. 2845 // 2846 // We try to obtain an induction variable from the original loop as hard 2847 // as possible. However if we don't find one that: 2848 // - is an integer 2849 // - counts from zero, stepping by one 2850 // - is the size of the widest induction variable type 2851 // then we create a new one. 2852 OldInduction = Legal->getInduction(); 2853 Type *IdxTy = Legal->getWidestInductionType(); 2854 2855 // Split the single block loop into the two loop structure described above. 2856 BasicBlock *VecBody = 2857 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 2858 BasicBlock *MiddleBlock = 2859 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 2860 BasicBlock *ScalarPH = 2861 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 2862 2863 // Create and register the new vector loop. 2864 Loop* Lp = new Loop(); 2865 Loop *ParentLoop = OrigLoop->getParentLoop(); 2866 2867 // Insert the new loop into the loop nest and register the new basic blocks 2868 // before calling any utilities such as SCEV that require valid LoopInfo. 2869 if (ParentLoop) { 2870 ParentLoop->addChildLoop(Lp); 2871 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI); 2872 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI); 2873 } else { 2874 LI->addTopLevelLoop(Lp); 2875 } 2876 Lp->addBasicBlockToLoop(VecBody, *LI); 2877 2878 // Find the loop boundaries. 2879 Value *Count = getOrCreateTripCount(Lp); 2880 2881 Value *StartIdx = ConstantInt::get(IdxTy, 0); 2882 2883 // We need to test whether the backedge-taken count is uint##_max. Adding one 2884 // to it will cause overflow and an incorrect loop trip count in the vector 2885 // body. In case of overflow we want to directly jump to the scalar remainder 2886 // loop. 2887 emitMinimumIterationCountCheck(Lp, ScalarPH); 2888 // Now, compare the new count to zero. If it is zero skip the vector loop and 2889 // jump to the scalar loop. 2890 emitVectorLoopEnteredCheck(Lp, ScalarPH); 2891 // Generate the code to check any assumptions that we've made for SCEV 2892 // expressions. 2893 emitSCEVChecks(Lp, ScalarPH); 2894 2895 // Generate the code that checks in runtime if arrays overlap. We put the 2896 // checks into a separate block to make the more common case of few elements 2897 // faster. 2898 emitMemRuntimeChecks(Lp, ScalarPH); 2899 2900 // Generate the induction variable. 2901 // The loop step is equal to the vectorization factor (num of SIMD elements) 2902 // times the unroll factor (num of SIMD instructions). 2903 Value *CountRoundDown = getOrCreateVectorTripCount(Lp); 2904 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 2905 Induction = 2906 createInductionVariable(Lp, StartIdx, CountRoundDown, Step, 2907 getDebugLocFromInstOrOperands(OldInduction)); 2908 2909 // We are going to resume the execution of the scalar loop. 2910 // Go over all of the induction variables that we found and fix the 2911 // PHIs that are left in the scalar version of the loop. 2912 // The starting values of PHI nodes depend on the counter of the last 2913 // iteration in the vectorized loop. 2914 // If we come from a bypass edge then we need to start from the original 2915 // start value. 2916 2917 // This variable saves the new starting index for the scalar loop. It is used 2918 // to test if there are any tail iterations left once the vector loop has 2919 // completed. 2920 LoopVectorizationLegality::InductionList::iterator I, E; 2921 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 2922 for (I = List->begin(), E = List->end(); I != E; ++I) { 2923 PHINode *OrigPhi = I->first; 2924 InductionDescriptor II = I->second; 2925 2926 // Create phi nodes to merge from the backedge-taken check block. 2927 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3, 2928 "bc.resume.val", 2929 ScalarPH->getTerminator()); 2930 Value *EndValue; 2931 if (OrigPhi == OldInduction) { 2932 // We know what the end value is. 2933 EndValue = CountRoundDown; 2934 } else { 2935 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator()); 2936 Value *CRD = B.CreateSExtOrTrunc(CountRoundDown, 2937 II.getStepValue()->getType(), 2938 "cast.crd"); 2939 EndValue = II.transform(B, CRD); 2940 EndValue->setName("ind.end"); 2941 } 2942 2943 // The new PHI merges the original incoming value, in case of a bypass, 2944 // or the value at the end of the vectorized loop. 2945 BCResumeVal->addIncoming(EndValue, MiddleBlock); 2946 2947 // Fix the scalar body counter (PHI node). 2948 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 2949 2950 // The old induction's phi node in the scalar body needs the truncated 2951 // value. 2952 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 2953 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]); 2954 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal); 2955 } 2956 2957 // Add a check in the middle block to see if we have completed 2958 // all of the iterations in the first vector loop. 2959 // If (N - N%VF) == N, then we *don't* need to run the remainder. 2960 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count, 2961 CountRoundDown, "cmp.n", 2962 MiddleBlock->getTerminator()); 2963 ReplaceInstWithInst(MiddleBlock->getTerminator(), 2964 BranchInst::Create(ExitBlock, ScalarPH, CmpN)); 2965 2966 // Get ready to start creating new instructions into the vectorized body. 2967 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt()); 2968 2969 // Save the state. 2970 LoopVectorPreHeader = Lp->getLoopPreheader(); 2971 LoopScalarPreHeader = ScalarPH; 2972 LoopMiddleBlock = MiddleBlock; 2973 LoopExitBlock = ExitBlock; 2974 LoopVectorBody.push_back(VecBody); 2975 LoopScalarBody = OldBasicBlock; 2976 2977 LoopVectorizeHints Hints(Lp, true); 2978 Hints.setAlreadyVectorized(); 2979 } 2980 2981 namespace { 2982 struct CSEDenseMapInfo { 2983 static bool canHandle(Instruction *I) { 2984 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 2985 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); 2986 } 2987 static inline Instruction *getEmptyKey() { 2988 return DenseMapInfo<Instruction *>::getEmptyKey(); 2989 } 2990 static inline Instruction *getTombstoneKey() { 2991 return DenseMapInfo<Instruction *>::getTombstoneKey(); 2992 } 2993 static unsigned getHashValue(Instruction *I) { 2994 assert(canHandle(I) && "Unknown instruction!"); 2995 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), 2996 I->value_op_end())); 2997 } 2998 static bool isEqual(Instruction *LHS, Instruction *RHS) { 2999 if (LHS == getEmptyKey() || RHS == getEmptyKey() || 3000 LHS == getTombstoneKey() || RHS == getTombstoneKey()) 3001 return LHS == RHS; 3002 return LHS->isIdenticalTo(RHS); 3003 } 3004 }; 3005 } 3006 3007 /// \brief Check whether this block is a predicated block. 3008 /// Due to if predication of stores we might create a sequence of "if(pred) a[i] 3009 /// = ...; " blocks. We start with one vectorized basic block. For every 3010 /// conditional block we split this vectorized block. Therefore, every second 3011 /// block will be a predicated one. 3012 static bool isPredicatedBlock(unsigned BlockNum) { 3013 return BlockNum % 2; 3014 } 3015 3016 ///\brief Perform cse of induction variable instructions. 3017 static void cse(SmallVector<BasicBlock *, 4> &BBs) { 3018 // Perform simple cse. 3019 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; 3020 for (unsigned i = 0, e = BBs.size(); i != e; ++i) { 3021 BasicBlock *BB = BBs[i]; 3022 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { 3023 Instruction *In = &*I++; 3024 3025 if (!CSEDenseMapInfo::canHandle(In)) 3026 continue; 3027 3028 // Check if we can replace this instruction with any of the 3029 // visited instructions. 3030 if (Instruction *V = CSEMap.lookup(In)) { 3031 In->replaceAllUsesWith(V); 3032 In->eraseFromParent(); 3033 continue; 3034 } 3035 // Ignore instructions in conditional blocks. We create "if (pred) a[i] = 3036 // ...;" blocks for predicated stores. Every second block is a predicated 3037 // block. 3038 if (isPredicatedBlock(i)) 3039 continue; 3040 3041 CSEMap[In] = In; 3042 } 3043 } 3044 } 3045 3046 /// \brief Adds a 'fast' flag to floating point operations. 3047 static Value *addFastMathFlag(Value *V) { 3048 if (isa<FPMathOperator>(V)){ 3049 FastMathFlags Flags; 3050 Flags.setUnsafeAlgebra(); 3051 cast<Instruction>(V)->setFastMathFlags(Flags); 3052 } 3053 return V; 3054 } 3055 3056 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if 3057 /// the result needs to be inserted and/or extracted from vectors. 3058 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract, 3059 const TargetTransformInfo &TTI) { 3060 if (Ty->isVoidTy()) 3061 return 0; 3062 3063 assert(Ty->isVectorTy() && "Can only scalarize vectors"); 3064 unsigned Cost = 0; 3065 3066 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) { 3067 if (Insert) 3068 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i); 3069 if (Extract) 3070 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i); 3071 } 3072 3073 return Cost; 3074 } 3075 3076 // Estimate cost of a call instruction CI if it were vectorized with factor VF. 3077 // Return the cost of the instruction, including scalarization overhead if it's 3078 // needed. The flag NeedToScalarize shows if the call needs to be scalarized - 3079 // i.e. either vector version isn't available, or is too expensive. 3080 static unsigned getVectorCallCost(CallInst *CI, unsigned VF, 3081 const TargetTransformInfo &TTI, 3082 const TargetLibraryInfo *TLI, 3083 bool &NeedToScalarize) { 3084 Function *F = CI->getCalledFunction(); 3085 StringRef FnName = CI->getCalledFunction()->getName(); 3086 Type *ScalarRetTy = CI->getType(); 3087 SmallVector<Type *, 4> Tys, ScalarTys; 3088 for (auto &ArgOp : CI->arg_operands()) 3089 ScalarTys.push_back(ArgOp->getType()); 3090 3091 // Estimate cost of scalarized vector call. The source operands are assumed 3092 // to be vectors, so we need to extract individual elements from there, 3093 // execute VF scalar calls, and then gather the result into the vector return 3094 // value. 3095 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys); 3096 if (VF == 1) 3097 return ScalarCallCost; 3098 3099 // Compute corresponding vector type for return value and arguments. 3100 Type *RetTy = ToVectorTy(ScalarRetTy, VF); 3101 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i) 3102 Tys.push_back(ToVectorTy(ScalarTys[i], VF)); 3103 3104 // Compute costs of unpacking argument values for the scalar calls and 3105 // packing the return values to a vector. 3106 unsigned ScalarizationCost = 3107 getScalarizationOverhead(RetTy, true, false, TTI); 3108 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) 3109 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI); 3110 3111 unsigned Cost = ScalarCallCost * VF + ScalarizationCost; 3112 3113 // If we can't emit a vector call for this function, then the currently found 3114 // cost is the cost we need to return. 3115 NeedToScalarize = true; 3116 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin()) 3117 return Cost; 3118 3119 // If the corresponding vector cost is cheaper, return its cost. 3120 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys); 3121 if (VectorCallCost < Cost) { 3122 NeedToScalarize = false; 3123 return VectorCallCost; 3124 } 3125 return Cost; 3126 } 3127 3128 // Estimate cost of an intrinsic call instruction CI if it were vectorized with 3129 // factor VF. Return the cost of the instruction, including scalarization 3130 // overhead if it's needed. 3131 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF, 3132 const TargetTransformInfo &TTI, 3133 const TargetLibraryInfo *TLI) { 3134 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 3135 assert(ID && "Expected intrinsic call!"); 3136 3137 Type *RetTy = ToVectorTy(CI->getType(), VF); 3138 SmallVector<Type *, 4> Tys; 3139 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 3140 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 3141 3142 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); 3143 } 3144 3145 static Type *smallestIntegerVectorType(Type *T1, Type *T2) { 3146 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType()); 3147 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType()); 3148 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2; 3149 } 3150 static Type *largestIntegerVectorType(Type *T1, Type *T2) { 3151 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType()); 3152 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType()); 3153 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; 3154 } 3155 3156 void InnerLoopVectorizer::truncateToMinimalBitwidths() { 3157 // For every instruction `I` in MinBWs, truncate the operands, create a 3158 // truncated version of `I` and reextend its result. InstCombine runs 3159 // later and will remove any ext/trunc pairs. 3160 // 3161 for (auto &KV : MinBWs) { 3162 VectorParts &Parts = WidenMap.get(KV.first); 3163 for (Value *&I : Parts) { 3164 if (I->use_empty()) 3165 continue; 3166 Type *OriginalTy = I->getType(); 3167 Type *ScalarTruncatedTy = IntegerType::get(OriginalTy->getContext(), 3168 KV.second); 3169 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy, 3170 OriginalTy->getVectorNumElements()); 3171 if (TruncatedTy == OriginalTy) 3172 continue; 3173 3174 IRBuilder<> B(cast<Instruction>(I)); 3175 auto ShrinkOperand = [&](Value *V) -> Value* { 3176 if (auto *ZI = dyn_cast<ZExtInst>(V)) 3177 if (ZI->getSrcTy() == TruncatedTy) 3178 return ZI->getOperand(0); 3179 return B.CreateZExtOrTrunc(V, TruncatedTy); 3180 }; 3181 3182 // The actual instruction modification depends on the instruction type, 3183 // unfortunately. 3184 Value *NewI = nullptr; 3185 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 3186 NewI = B.CreateBinOp(BO->getOpcode(), 3187 ShrinkOperand(BO->getOperand(0)), 3188 ShrinkOperand(BO->getOperand(1))); 3189 cast<BinaryOperator>(NewI)->copyIRFlags(I); 3190 } else if (ICmpInst *CI = dyn_cast<ICmpInst>(I)) { 3191 NewI = B.CreateICmp(CI->getPredicate(), 3192 ShrinkOperand(CI->getOperand(0)), 3193 ShrinkOperand(CI->getOperand(1))); 3194 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 3195 NewI = B.CreateSelect(SI->getCondition(), 3196 ShrinkOperand(SI->getTrueValue()), 3197 ShrinkOperand(SI->getFalseValue())); 3198 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 3199 switch (CI->getOpcode()) { 3200 default: llvm_unreachable("Unhandled cast!"); 3201 case Instruction::Trunc: 3202 NewI = ShrinkOperand(CI->getOperand(0)); 3203 break; 3204 case Instruction::SExt: 3205 NewI = B.CreateSExtOrTrunc(CI->getOperand(0), 3206 smallestIntegerVectorType(OriginalTy, 3207 TruncatedTy)); 3208 break; 3209 case Instruction::ZExt: 3210 NewI = B.CreateZExtOrTrunc(CI->getOperand(0), 3211 smallestIntegerVectorType(OriginalTy, 3212 TruncatedTy)); 3213 break; 3214 } 3215 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { 3216 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements(); 3217 auto *O0 = 3218 B.CreateZExtOrTrunc(SI->getOperand(0), 3219 VectorType::get(ScalarTruncatedTy, Elements0)); 3220 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements(); 3221 auto *O1 = 3222 B.CreateZExtOrTrunc(SI->getOperand(1), 3223 VectorType::get(ScalarTruncatedTy, Elements1)); 3224 3225 NewI = B.CreateShuffleVector(O0, O1, SI->getMask()); 3226 } else if (isa<LoadInst>(I)) { 3227 // Don't do anything with the operands, just extend the result. 3228 continue; 3229 } else { 3230 llvm_unreachable("Unhandled instruction type!"); 3231 } 3232 3233 // Lastly, extend the result. 3234 NewI->takeName(cast<Instruction>(I)); 3235 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); 3236 I->replaceAllUsesWith(Res); 3237 cast<Instruction>(I)->eraseFromParent(); 3238 I = Res; 3239 } 3240 } 3241 3242 // We'll have created a bunch of ZExts that are now parentless. Clean up. 3243 for (auto &KV : MinBWs) { 3244 VectorParts &Parts = WidenMap.get(KV.first); 3245 for (Value *&I : Parts) { 3246 ZExtInst *Inst = dyn_cast<ZExtInst>(I); 3247 if (Inst && Inst->use_empty()) { 3248 Value *NewI = Inst->getOperand(0); 3249 Inst->eraseFromParent(); 3250 I = NewI; 3251 } 3252 } 3253 } 3254 } 3255 3256 void InnerLoopVectorizer::vectorizeLoop() { 3257 //===------------------------------------------------===// 3258 // 3259 // Notice: any optimization or new instruction that go 3260 // into the code below should be also be implemented in 3261 // the cost-model. 3262 // 3263 //===------------------------------------------------===// 3264 Constant *Zero = Builder.getInt32(0); 3265 3266 // In order to support reduction variables we need to be able to vectorize 3267 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two 3268 // stages. First, we create a new vector PHI node with no incoming edges. 3269 // We use this value when we vectorize all of the instructions that use the 3270 // PHI. Next, after all of the instructions in the block are complete we 3271 // add the new incoming edges to the PHI. At this point all of the 3272 // instructions in the basic block are vectorized, so we can use them to 3273 // construct the PHI. 3274 PhiVector RdxPHIsToFix; 3275 3276 // Scan the loop in a topological order to ensure that defs are vectorized 3277 // before users. 3278 LoopBlocksDFS DFS(OrigLoop); 3279 DFS.perform(LI); 3280 3281 // Vectorize all of the blocks in the original loop. 3282 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 3283 be = DFS.endRPO(); bb != be; ++bb) 3284 vectorizeBlockInLoop(*bb, &RdxPHIsToFix); 3285 3286 // Insert truncates and extends for any truncated instructions as hints to 3287 // InstCombine. 3288 if (VF > 1) 3289 truncateToMinimalBitwidths(); 3290 3291 // At this point every instruction in the original loop is widened to 3292 // a vector form. We are almost done. Now, we need to fix the PHI nodes 3293 // that we vectorized. The PHI nodes are currently empty because we did 3294 // not want to introduce cycles. Notice that the remaining PHI nodes 3295 // that we need to fix are reduction variables. 3296 3297 // Create the 'reduced' values for each of the induction vars. 3298 // The reduced values are the vector values that we scalarize and combine 3299 // after the loop is finished. 3300 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); 3301 it != e; ++it) { 3302 PHINode *RdxPhi = *it; 3303 assert(RdxPhi && "Unable to recover vectorized PHI"); 3304 3305 // Find the reduction variable descriptor. 3306 assert(Legal->isReductionVariable(RdxPhi) && 3307 "Unable to find the reduction variable"); 3308 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi]; 3309 3310 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind(); 3311 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); 3312 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); 3313 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind = 3314 RdxDesc.getMinMaxRecurrenceKind(); 3315 setDebugLocFromInst(Builder, ReductionStartValue); 3316 3317 // We need to generate a reduction vector from the incoming scalar. 3318 // To do so, we need to generate the 'identity' vector and override 3319 // one of the elements with the incoming scalar reduction. We need 3320 // to do it in the vector-loop preheader. 3321 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator()); 3322 3323 // This is the vector-clone of the value that leaves the loop. 3324 VectorParts &VectorExit = getVectorValue(LoopExitInst); 3325 Type *VecTy = VectorExit[0]->getType(); 3326 3327 // Find the reduction identity variable. Zero for addition, or, xor, 3328 // one for multiplication, -1 for And. 3329 Value *Identity; 3330 Value *VectorStart; 3331 if (RK == RecurrenceDescriptor::RK_IntegerMinMax || 3332 RK == RecurrenceDescriptor::RK_FloatMinMax) { 3333 // MinMax reduction have the start value as their identify. 3334 if (VF == 1) { 3335 VectorStart = Identity = ReductionStartValue; 3336 } else { 3337 VectorStart = Identity = 3338 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident"); 3339 } 3340 } else { 3341 // Handle other reduction kinds: 3342 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity( 3343 RK, VecTy->getScalarType()); 3344 if (VF == 1) { 3345 Identity = Iden; 3346 // This vector is the Identity vector where the first element is the 3347 // incoming scalar reduction. 3348 VectorStart = ReductionStartValue; 3349 } else { 3350 Identity = ConstantVector::getSplat(VF, Iden); 3351 3352 // This vector is the Identity vector where the first element is the 3353 // incoming scalar reduction. 3354 VectorStart = 3355 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero); 3356 } 3357 } 3358 3359 // Fix the vector-loop phi. 3360 3361 // Reductions do not have to start at zero. They can start with 3362 // any loop invariant values. 3363 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 3364 BasicBlock *Latch = OrigLoop->getLoopLatch(); 3365 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 3366 VectorParts &Val = getVectorValue(LoopVal); 3367 for (unsigned part = 0; part < UF; ++part) { 3368 // Make sure to add the reduction stat value only to the 3369 // first unroll part. 3370 Value *StartVal = (part == 0) ? VectorStart : Identity; 3371 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, 3372 LoopVectorPreHeader); 3373 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], 3374 LoopVectorBody.back()); 3375 } 3376 3377 // Before each round, move the insertion point right between 3378 // the PHIs and the values we are going to write. 3379 // This allows us to write both PHINodes and the extractelement 3380 // instructions. 3381 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); 3382 3383 VectorParts RdxParts = getVectorValue(LoopExitInst); 3384 setDebugLocFromInst(Builder, LoopExitInst); 3385 3386 // If the vector reduction can be performed in a smaller type, we truncate 3387 // then extend the loop exit value to enable InstCombine to evaluate the 3388 // entire expression in the smaller type. 3389 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) { 3390 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); 3391 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator()); 3392 for (unsigned part = 0; part < UF; ++part) { 3393 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy); 3394 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) 3395 : Builder.CreateZExt(Trunc, VecTy); 3396 for (Value::user_iterator UI = RdxParts[part]->user_begin(); 3397 UI != RdxParts[part]->user_end();) 3398 if (*UI != Trunc) { 3399 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd); 3400 RdxParts[part] = Extnd; 3401 } else { 3402 ++UI; 3403 } 3404 } 3405 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); 3406 for (unsigned part = 0; part < UF; ++part) 3407 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy); 3408 } 3409 3410 // Reduce all of the unrolled parts into a single vector. 3411 Value *ReducedPartRdx = RdxParts[0]; 3412 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK); 3413 setDebugLocFromInst(Builder, ReducedPartRdx); 3414 for (unsigned part = 1; part < UF; ++part) { 3415 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 3416 // Floating point operations had to be 'fast' to enable the reduction. 3417 ReducedPartRdx = addFastMathFlag( 3418 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part], 3419 ReducedPartRdx, "bin.rdx")); 3420 else 3421 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp( 3422 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]); 3423 } 3424 3425 if (VF > 1) { 3426 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 3427 // and vector ops, reducing the set of values being computed by half each 3428 // round. 3429 assert(isPowerOf2_32(VF) && 3430 "Reduction emission only supported for pow2 vectors!"); 3431 Value *TmpVec = ReducedPartRdx; 3432 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr); 3433 for (unsigned i = VF; i != 1; i >>= 1) { 3434 // Move the upper half of the vector to the lower half. 3435 for (unsigned j = 0; j != i/2; ++j) 3436 ShuffleMask[j] = Builder.getInt32(i/2 + j); 3437 3438 // Fill the rest of the mask with undef. 3439 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 3440 UndefValue::get(Builder.getInt32Ty())); 3441 3442 Value *Shuf = 3443 Builder.CreateShuffleVector(TmpVec, 3444 UndefValue::get(TmpVec->getType()), 3445 ConstantVector::get(ShuffleMask), 3446 "rdx.shuf"); 3447 3448 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 3449 // Floating point operations had to be 'fast' to enable the reduction. 3450 TmpVec = addFastMathFlag(Builder.CreateBinOp( 3451 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx")); 3452 else 3453 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind, 3454 TmpVec, Shuf); 3455 } 3456 3457 // The result is in the first element of the vector. 3458 ReducedPartRdx = Builder.CreateExtractElement(TmpVec, 3459 Builder.getInt32(0)); 3460 3461 // If the reduction can be performed in a smaller type, we need to extend 3462 // the reduction to the wider type before we branch to the original loop. 3463 if (RdxPhi->getType() != RdxDesc.getRecurrenceType()) 3464 ReducedPartRdx = 3465 RdxDesc.isSigned() 3466 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType()) 3467 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType()); 3468 } 3469 3470 // Create a phi node that merges control-flow from the backedge-taken check 3471 // block and the middle block. 3472 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx", 3473 LoopScalarPreHeader->getTerminator()); 3474 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 3475 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]); 3476 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); 3477 3478 // Now, we need to fix the users of the reduction variable 3479 // inside and outside of the scalar remainder loop. 3480 // We know that the loop is in LCSSA form. We need to update the 3481 // PHI nodes in the exit blocks. 3482 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 3483 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 3484 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 3485 if (!LCSSAPhi) break; 3486 3487 // All PHINodes need to have a single entry edge, or two if 3488 // we already fixed them. 3489 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 3490 3491 // We found our reduction value exit-PHI. Update it with the 3492 // incoming bypass edge. 3493 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) { 3494 // Add an edge coming from the bypass. 3495 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); 3496 break; 3497 } 3498 }// end of the LCSSA phi scan. 3499 3500 // Fix the scalar loop reduction variable with the incoming reduction sum 3501 // from the vector body and from the backedge value. 3502 int IncomingEdgeBlockIdx = 3503 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 3504 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 3505 // Pick the other block. 3506 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 3507 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); 3508 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); 3509 }// end of for each redux variable. 3510 3511 fixLCSSAPHIs(); 3512 3513 // Make sure DomTree is updated. 3514 updateAnalysis(); 3515 3516 // Predicate any stores. 3517 for (auto KV : PredicatedStores) { 3518 BasicBlock::iterator I(KV.first); 3519 auto *BB = SplitBlock(I->getParent(), &*std::next(I), DT, LI); 3520 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false, 3521 /*BranchWeights=*/nullptr, DT); 3522 I->moveBefore(T); 3523 I->getParent()->setName("pred.store.if"); 3524 BB->setName("pred.store.continue"); 3525 } 3526 DEBUG(DT->verifyDomTree()); 3527 // Remove redundant induction instructions. 3528 cse(LoopVectorBody); 3529 } 3530 3531 void InnerLoopVectorizer::fixLCSSAPHIs() { 3532 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 3533 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 3534 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 3535 if (!LCSSAPhi) break; 3536 if (LCSSAPhi->getNumIncomingValues() == 1) 3537 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 3538 LoopMiddleBlock); 3539 } 3540 } 3541 3542 InnerLoopVectorizer::VectorParts 3543 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 3544 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 3545 "Invalid edge"); 3546 3547 // Look for cached value. 3548 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst); 3549 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge); 3550 if (ECEntryIt != MaskCache.end()) 3551 return ECEntryIt->second; 3552 3553 VectorParts SrcMask = createBlockInMask(Src); 3554 3555 // The terminator has to be a branch inst! 3556 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 3557 assert(BI && "Unexpected terminator found"); 3558 3559 if (BI->isConditional()) { 3560 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 3561 3562 if (BI->getSuccessor(0) != Dst) 3563 for (unsigned part = 0; part < UF; ++part) 3564 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 3565 3566 for (unsigned part = 0; part < UF; ++part) 3567 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 3568 3569 MaskCache[Edge] = EdgeMask; 3570 return EdgeMask; 3571 } 3572 3573 MaskCache[Edge] = SrcMask; 3574 return SrcMask; 3575 } 3576 3577 InnerLoopVectorizer::VectorParts 3578 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 3579 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 3580 3581 // Loop incoming mask is all-one. 3582 if (OrigLoop->getHeader() == BB) { 3583 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 3584 return getVectorValue(C); 3585 } 3586 3587 // This is the block mask. We OR all incoming edges, and with zero. 3588 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 3589 VectorParts BlockMask = getVectorValue(Zero); 3590 3591 // For each pred: 3592 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 3593 VectorParts EM = createEdgeMask(*it, BB); 3594 for (unsigned part = 0; part < UF; ++part) 3595 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 3596 } 3597 3598 return BlockMask; 3599 } 3600 3601 void InnerLoopVectorizer::widenPHIInstruction( 3602 Instruction *PN, InnerLoopVectorizer::VectorParts &Entry, unsigned UF, 3603 unsigned VF, PhiVector *PV) { 3604 PHINode* P = cast<PHINode>(PN); 3605 // Handle reduction variables: 3606 if (Legal->isReductionVariable(P)) { 3607 for (unsigned part = 0; part < UF; ++part) { 3608 // This is phase one of vectorizing PHIs. 3609 Type *VecTy = (VF == 1) ? PN->getType() : 3610 VectorType::get(PN->getType(), VF); 3611 Entry[part] = PHINode::Create( 3612 VecTy, 2, "vec.phi", &*LoopVectorBody.back()->getFirstInsertionPt()); 3613 } 3614 PV->push_back(P); 3615 return; 3616 } 3617 3618 setDebugLocFromInst(Builder, P); 3619 // Check for PHI nodes that are lowered to vector selects. 3620 if (P->getParent() != OrigLoop->getHeader()) { 3621 // We know that all PHIs in non-header blocks are converted into 3622 // selects, so we don't have to worry about the insertion order and we 3623 // can just use the builder. 3624 // At this point we generate the predication tree. There may be 3625 // duplications since this is a simple recursive scan, but future 3626 // optimizations will clean it up. 3627 3628 unsigned NumIncoming = P->getNumIncomingValues(); 3629 3630 // Generate a sequence of selects of the form: 3631 // SELECT(Mask3, In3, 3632 // SELECT(Mask2, In2, 3633 // ( ...))) 3634 for (unsigned In = 0; In < NumIncoming; In++) { 3635 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), 3636 P->getParent()); 3637 VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); 3638 3639 for (unsigned part = 0; part < UF; ++part) { 3640 // We might have single edge PHIs (blocks) - use an identity 3641 // 'select' for the first PHI operand. 3642 if (In == 0) 3643 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 3644 In0[part]); 3645 else 3646 // Select between the current value and the previous incoming edge 3647 // based on the incoming mask. 3648 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 3649 Entry[part], "predphi"); 3650 } 3651 } 3652 return; 3653 } 3654 3655 // This PHINode must be an induction variable. 3656 // Make sure that we know about it. 3657 assert(Legal->getInductionVars()->count(P) && 3658 "Not an induction variable"); 3659 3660 InductionDescriptor II = Legal->getInductionVars()->lookup(P); 3661 3662 // FIXME: The newly created binary instructions should contain nsw/nuw flags, 3663 // which can be found from the original scalar operations. 3664 switch (II.getKind()) { 3665 case InductionDescriptor::IK_NoInduction: 3666 llvm_unreachable("Unknown induction"); 3667 case InductionDescriptor::IK_IntInduction: { 3668 assert(P->getType() == II.getStartValue()->getType() && 3669 "Types must match"); 3670 // Handle other induction variables that are now based on the 3671 // canonical one. 3672 Value *V = Induction; 3673 if (P != OldInduction) { 3674 V = Builder.CreateSExtOrTrunc(Induction, P->getType()); 3675 V = II.transform(Builder, V); 3676 V->setName("offset.idx"); 3677 } 3678 Value *Broadcasted = getBroadcastInstrs(V); 3679 // After broadcasting the induction variable we need to make the vector 3680 // consecutive by adding 0, 1, 2, etc. 3681 for (unsigned part = 0; part < UF; ++part) 3682 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue()); 3683 return; 3684 } 3685 case InductionDescriptor::IK_PtrInduction: 3686 // Handle the pointer induction variable case. 3687 assert(P->getType()->isPointerTy() && "Unexpected type."); 3688 // This is the normalized GEP that starts counting at zero. 3689 Value *PtrInd = Induction; 3690 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType()); 3691 // This is the vector of results. Notice that we don't generate 3692 // vector geps because scalar geps result in better code. 3693 for (unsigned part = 0; part < UF; ++part) { 3694 if (VF == 1) { 3695 int EltIndex = part; 3696 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex); 3697 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx); 3698 Value *SclrGep = II.transform(Builder, GlobalIdx); 3699 SclrGep->setName("next.gep"); 3700 Entry[part] = SclrGep; 3701 continue; 3702 } 3703 3704 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 3705 for (unsigned int i = 0; i < VF; ++i) { 3706 int EltIndex = i + part * VF; 3707 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex); 3708 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx); 3709 Value *SclrGep = II.transform(Builder, GlobalIdx); 3710 SclrGep->setName("next.gep"); 3711 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 3712 Builder.getInt32(i), 3713 "insert.gep"); 3714 } 3715 Entry[part] = VecVal; 3716 } 3717 return; 3718 } 3719 } 3720 3721 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) { 3722 // For each instruction in the old loop. 3723 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3724 VectorParts &Entry = WidenMap.get(&*it); 3725 3726 switch (it->getOpcode()) { 3727 case Instruction::Br: 3728 // Nothing to do for PHIs and BR, since we already took care of the 3729 // loop control flow instructions. 3730 continue; 3731 case Instruction::PHI: { 3732 // Vectorize PHINodes. 3733 widenPHIInstruction(&*it, Entry, UF, VF, PV); 3734 continue; 3735 }// End of PHI. 3736 3737 case Instruction::Add: 3738 case Instruction::FAdd: 3739 case Instruction::Sub: 3740 case Instruction::FSub: 3741 case Instruction::Mul: 3742 case Instruction::FMul: 3743 case Instruction::UDiv: 3744 case Instruction::SDiv: 3745 case Instruction::FDiv: 3746 case Instruction::URem: 3747 case Instruction::SRem: 3748 case Instruction::FRem: 3749 case Instruction::Shl: 3750 case Instruction::LShr: 3751 case Instruction::AShr: 3752 case Instruction::And: 3753 case Instruction::Or: 3754 case Instruction::Xor: { 3755 // Just widen binops. 3756 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 3757 setDebugLocFromInst(Builder, BinOp); 3758 VectorParts &A = getVectorValue(it->getOperand(0)); 3759 VectorParts &B = getVectorValue(it->getOperand(1)); 3760 3761 // Use this vector value for all users of the original instruction. 3762 for (unsigned Part = 0; Part < UF; ++Part) { 3763 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 3764 3765 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V)) 3766 VecOp->copyIRFlags(BinOp); 3767 3768 Entry[Part] = V; 3769 } 3770 3771 propagateMetadata(Entry, &*it); 3772 break; 3773 } 3774 case Instruction::Select: { 3775 // Widen selects. 3776 // If the selector is loop invariant we can create a select 3777 // instruction with a scalar condition. Otherwise, use vector-select. 3778 auto *SE = PSE.getSE(); 3779 bool InvariantCond = 3780 SE->isLoopInvariant(PSE.getSCEV(it->getOperand(0)), OrigLoop); 3781 setDebugLocFromInst(Builder, &*it); 3782 3783 // The condition can be loop invariant but still defined inside the 3784 // loop. This means that we can't just use the original 'cond' value. 3785 // We have to take the 'vectorized' value and pick the first lane. 3786 // Instcombine will make this a no-op. 3787 VectorParts &Cond = getVectorValue(it->getOperand(0)); 3788 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 3789 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 3790 3791 Value *ScalarCond = (VF == 1) ? Cond[0] : 3792 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0)); 3793 3794 for (unsigned Part = 0; Part < UF; ++Part) { 3795 Entry[Part] = Builder.CreateSelect( 3796 InvariantCond ? ScalarCond : Cond[Part], 3797 Op0[Part], 3798 Op1[Part]); 3799 } 3800 3801 propagateMetadata(Entry, &*it); 3802 break; 3803 } 3804 3805 case Instruction::ICmp: 3806 case Instruction::FCmp: { 3807 // Widen compares. Generate vector compares. 3808 bool FCmp = (it->getOpcode() == Instruction::FCmp); 3809 CmpInst *Cmp = dyn_cast<CmpInst>(it); 3810 setDebugLocFromInst(Builder, &*it); 3811 VectorParts &A = getVectorValue(it->getOperand(0)); 3812 VectorParts &B = getVectorValue(it->getOperand(1)); 3813 for (unsigned Part = 0; Part < UF; ++Part) { 3814 Value *C = nullptr; 3815 if (FCmp) { 3816 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 3817 cast<FCmpInst>(C)->copyFastMathFlags(&*it); 3818 } else { 3819 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 3820 } 3821 Entry[Part] = C; 3822 } 3823 3824 propagateMetadata(Entry, &*it); 3825 break; 3826 } 3827 3828 case Instruction::Store: 3829 case Instruction::Load: 3830 vectorizeMemoryInstruction(&*it); 3831 break; 3832 case Instruction::ZExt: 3833 case Instruction::SExt: 3834 case Instruction::FPToUI: 3835 case Instruction::FPToSI: 3836 case Instruction::FPExt: 3837 case Instruction::PtrToInt: 3838 case Instruction::IntToPtr: 3839 case Instruction::SIToFP: 3840 case Instruction::UIToFP: 3841 case Instruction::Trunc: 3842 case Instruction::FPTrunc: 3843 case Instruction::BitCast: { 3844 CastInst *CI = dyn_cast<CastInst>(it); 3845 setDebugLocFromInst(Builder, &*it); 3846 /// Optimize the special case where the source is the induction 3847 /// variable. Notice that we can only optimize the 'trunc' case 3848 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 3849 /// c. other casts depend on pointer size. 3850 if (CI->getOperand(0) == OldInduction && 3851 it->getOpcode() == Instruction::Trunc) { 3852 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 3853 CI->getType()); 3854 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 3855 InductionDescriptor II = 3856 Legal->getInductionVars()->lookup(OldInduction); 3857 Constant *Step = ConstantInt::getSigned( 3858 CI->getType(), II.getStepValue()->getSExtValue()); 3859 for (unsigned Part = 0; Part < UF; ++Part) 3860 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step); 3861 propagateMetadata(Entry, &*it); 3862 break; 3863 } 3864 /// Vectorize casts. 3865 Type *DestTy = (VF == 1) ? CI->getType() : 3866 VectorType::get(CI->getType(), VF); 3867 3868 VectorParts &A = getVectorValue(it->getOperand(0)); 3869 for (unsigned Part = 0; Part < UF; ++Part) 3870 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 3871 propagateMetadata(Entry, &*it); 3872 break; 3873 } 3874 3875 case Instruction::Call: { 3876 // Ignore dbg intrinsics. 3877 if (isa<DbgInfoIntrinsic>(it)) 3878 break; 3879 setDebugLocFromInst(Builder, &*it); 3880 3881 Module *M = BB->getParent()->getParent(); 3882 CallInst *CI = cast<CallInst>(it); 3883 3884 StringRef FnName = CI->getCalledFunction()->getName(); 3885 Function *F = CI->getCalledFunction(); 3886 Type *RetTy = ToVectorTy(CI->getType(), VF); 3887 SmallVector<Type *, 4> Tys; 3888 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 3889 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 3890 3891 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 3892 if (ID && 3893 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || 3894 ID == Intrinsic::lifetime_start)) { 3895 scalarizeInstruction(&*it); 3896 break; 3897 } 3898 // The flag shows whether we use Intrinsic or a usual Call for vectorized 3899 // version of the instruction. 3900 // Is it beneficial to perform intrinsic call compared to lib call? 3901 bool NeedToScalarize; 3902 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize); 3903 bool UseVectorIntrinsic = 3904 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost; 3905 if (!UseVectorIntrinsic && NeedToScalarize) { 3906 scalarizeInstruction(&*it); 3907 break; 3908 } 3909 3910 for (unsigned Part = 0; Part < UF; ++Part) { 3911 SmallVector<Value *, 4> Args; 3912 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 3913 Value *Arg = CI->getArgOperand(i); 3914 // Some intrinsics have a scalar argument - don't replace it with a 3915 // vector. 3916 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) { 3917 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i)); 3918 Arg = VectorArg[Part]; 3919 } 3920 Args.push_back(Arg); 3921 } 3922 3923 Function *VectorF; 3924 if (UseVectorIntrinsic) { 3925 // Use vector version of the intrinsic. 3926 Type *TysForDecl[] = {CI->getType()}; 3927 if (VF > 1) 3928 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF); 3929 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl); 3930 } else { 3931 // Use vector version of the library call. 3932 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF); 3933 assert(!VFnName.empty() && "Vector function name is empty."); 3934 VectorF = M->getFunction(VFnName); 3935 if (!VectorF) { 3936 // Generate a declaration 3937 FunctionType *FTy = FunctionType::get(RetTy, Tys, false); 3938 VectorF = 3939 Function::Create(FTy, Function::ExternalLinkage, VFnName, M); 3940 VectorF->copyAttributesFrom(F); 3941 } 3942 } 3943 assert(VectorF && "Can't create vector function."); 3944 Entry[Part] = Builder.CreateCall(VectorF, Args); 3945 } 3946 3947 propagateMetadata(Entry, &*it); 3948 break; 3949 } 3950 3951 default: 3952 // All other instructions are unsupported. Scalarize them. 3953 scalarizeInstruction(&*it); 3954 break; 3955 }// end of switch. 3956 }// end of for_each instr. 3957 } 3958 3959 void InnerLoopVectorizer::updateAnalysis() { 3960 // Forget the original basic block. 3961 PSE.getSE()->forgetLoop(OrigLoop); 3962 3963 // Update the dominator tree information. 3964 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && 3965 "Entry does not dominate exit."); 3966 3967 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 3968 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); 3969 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); 3970 3971 // We don't predicate stores by this point, so the vector body should be a 3972 // single loop. 3973 assert(LoopVectorBody.size() == 1 && "Expected single block loop!"); 3974 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader); 3975 3976 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody.back()); 3977 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]); 3978 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 3979 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]); 3980 3981 DEBUG(DT->verifyDomTree()); 3982 } 3983 3984 /// \brief Check whether it is safe to if-convert this phi node. 3985 /// 3986 /// Phi nodes with constant expressions that can trap are not safe to if 3987 /// convert. 3988 static bool canIfConvertPHINodes(BasicBlock *BB) { 3989 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 3990 PHINode *Phi = dyn_cast<PHINode>(I); 3991 if (!Phi) 3992 return true; 3993 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p) 3994 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p))) 3995 if (C->canTrap()) 3996 return false; 3997 } 3998 return true; 3999 } 4000 4001 bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 4002 if (!EnableIfConversion) { 4003 emitAnalysis(VectorizationReport() << "if-conversion is disabled"); 4004 return false; 4005 } 4006 4007 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 4008 4009 // A list of pointers that we can safely read and write to. 4010 SmallPtrSet<Value *, 8> SafePointes; 4011 4012 // Collect safe addresses. 4013 for (Loop::block_iterator BI = TheLoop->block_begin(), 4014 BE = TheLoop->block_end(); BI != BE; ++BI) { 4015 BasicBlock *BB = *BI; 4016 4017 if (blockNeedsPredication(BB)) 4018 continue; 4019 4020 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 4021 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 4022 SafePointes.insert(LI->getPointerOperand()); 4023 else if (StoreInst *SI = dyn_cast<StoreInst>(I)) 4024 SafePointes.insert(SI->getPointerOperand()); 4025 } 4026 } 4027 4028 // Collect the blocks that need predication. 4029 BasicBlock *Header = TheLoop->getHeader(); 4030 for (Loop::block_iterator BI = TheLoop->block_begin(), 4031 BE = TheLoop->block_end(); BI != BE; ++BI) { 4032 BasicBlock *BB = *BI; 4033 4034 // We don't support switch statements inside loops. 4035 if (!isa<BranchInst>(BB->getTerminator())) { 4036 emitAnalysis(VectorizationReport(BB->getTerminator()) 4037 << "loop contains a switch statement"); 4038 return false; 4039 } 4040 4041 // We must be able to predicate all blocks that need to be predicated. 4042 if (blockNeedsPredication(BB)) { 4043 if (!blockCanBePredicated(BB, SafePointes)) { 4044 emitAnalysis(VectorizationReport(BB->getTerminator()) 4045 << "control flow cannot be substituted for a select"); 4046 return false; 4047 } 4048 } else if (BB != Header && !canIfConvertPHINodes(BB)) { 4049 emitAnalysis(VectorizationReport(BB->getTerminator()) 4050 << "control flow cannot be substituted for a select"); 4051 return false; 4052 } 4053 } 4054 4055 // We can if-convert this loop. 4056 return true; 4057 } 4058 4059 bool LoopVectorizationLegality::canVectorize() { 4060 // We must have a loop in canonical form. Loops with indirectbr in them cannot 4061 // be canonicalized. 4062 if (!TheLoop->getLoopPreheader()) { 4063 emitAnalysis( 4064 VectorizationReport() << 4065 "loop control flow is not understood by vectorizer"); 4066 return false; 4067 } 4068 4069 // We can only vectorize innermost loops. 4070 if (!TheLoop->empty()) { 4071 emitAnalysis(VectorizationReport() << "loop is not the innermost loop"); 4072 return false; 4073 } 4074 4075 // We must have a single backedge. 4076 if (TheLoop->getNumBackEdges() != 1) { 4077 emitAnalysis( 4078 VectorizationReport() << 4079 "loop control flow is not understood by vectorizer"); 4080 return false; 4081 } 4082 4083 // We must have a single exiting block. 4084 if (!TheLoop->getExitingBlock()) { 4085 emitAnalysis( 4086 VectorizationReport() << 4087 "loop control flow is not understood by vectorizer"); 4088 return false; 4089 } 4090 4091 // We only handle bottom-tested loops, i.e. loop in which the condition is 4092 // checked at the end of each iteration. With that we can assume that all 4093 // instructions in the loop are executed the same number of times. 4094 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { 4095 emitAnalysis( 4096 VectorizationReport() << 4097 "loop control flow is not understood by vectorizer"); 4098 return false; 4099 } 4100 4101 // We need to have a loop header. 4102 DEBUG(dbgs() << "LV: Found a loop: " << 4103 TheLoop->getHeader()->getName() << '\n'); 4104 4105 // Check if we can if-convert non-single-bb loops. 4106 unsigned NumBlocks = TheLoop->getNumBlocks(); 4107 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 4108 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 4109 return false; 4110 } 4111 4112 // ScalarEvolution needs to be able to find the exit count. 4113 const SCEV *ExitCount = PSE.getSE()->getBackedgeTakenCount(TheLoop); 4114 if (ExitCount == PSE.getSE()->getCouldNotCompute()) { 4115 emitAnalysis(VectorizationReport() 4116 << "could not determine number of loop iterations"); 4117 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 4118 return false; 4119 } 4120 4121 // Check if we can vectorize the instructions and CFG in this loop. 4122 if (!canVectorizeInstrs()) { 4123 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 4124 return false; 4125 } 4126 4127 // Go over each instruction and look at memory deps. 4128 if (!canVectorizeMemory()) { 4129 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 4130 return false; 4131 } 4132 4133 // Collect all of the variables that remain uniform after vectorization. 4134 collectLoopUniforms(); 4135 4136 DEBUG(dbgs() << "LV: We can vectorize this loop" 4137 << (LAI->getRuntimePointerChecking()->Need 4138 ? " (with a runtime bound check)" 4139 : "") 4140 << "!\n"); 4141 4142 bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); 4143 4144 // If an override option has been passed in for interleaved accesses, use it. 4145 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) 4146 UseInterleaved = EnableInterleavedMemAccesses; 4147 4148 // Analyze interleaved memory accesses. 4149 if (UseInterleaved) 4150 InterleaveInfo.analyzeInterleaving(Strides); 4151 4152 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold; 4153 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled) 4154 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold; 4155 4156 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) { 4157 emitAnalysis(VectorizationReport() 4158 << "Too many SCEV assumptions need to be made and checked " 4159 << "at runtime"); 4160 DEBUG(dbgs() << "LV: Too many SCEV checks needed.\n"); 4161 return false; 4162 } 4163 4164 // Okay! We can vectorize. At this point we don't have any other mem analysis 4165 // which may limit our maximum vectorization factor, so just return true with 4166 // no restrictions. 4167 return true; 4168 } 4169 4170 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) { 4171 if (Ty->isPointerTy()) 4172 return DL.getIntPtrType(Ty); 4173 4174 // It is possible that char's or short's overflow when we ask for the loop's 4175 // trip count, work around this by changing the type size. 4176 if (Ty->getScalarSizeInBits() < 32) 4177 return Type::getInt32Ty(Ty->getContext()); 4178 4179 return Ty; 4180 } 4181 4182 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) { 4183 Ty0 = convertPointerToIntegerType(DL, Ty0); 4184 Ty1 = convertPointerToIntegerType(DL, Ty1); 4185 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits()) 4186 return Ty0; 4187 return Ty1; 4188 } 4189 4190 /// \brief Check that the instruction has outside loop users and is not an 4191 /// identified reduction variable. 4192 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst, 4193 SmallPtrSetImpl<Value *> &Reductions) { 4194 // Reduction instructions are allowed to have exit users. All other 4195 // instructions must not have external users. 4196 if (!Reductions.count(Inst)) 4197 //Check that all of the users of the loop are inside the BB. 4198 for (User *U : Inst->users()) { 4199 Instruction *UI = cast<Instruction>(U); 4200 // This user may be a reduction exit value. 4201 if (!TheLoop->contains(UI)) { 4202 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n'); 4203 return true; 4204 } 4205 } 4206 return false; 4207 } 4208 4209 bool LoopVectorizationLegality::canVectorizeInstrs() { 4210 BasicBlock *Header = TheLoop->getHeader(); 4211 4212 // Look for the attribute signaling the absence of NaNs. 4213 Function &F = *Header->getParent(); 4214 const DataLayout &DL = F.getParent()->getDataLayout(); 4215 if (F.hasFnAttribute("no-nans-fp-math")) 4216 HasFunNoNaNAttr = 4217 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true"; 4218 4219 // For each block in the loop. 4220 for (Loop::block_iterator bb = TheLoop->block_begin(), 4221 be = TheLoop->block_end(); bb != be; ++bb) { 4222 4223 // Scan the instructions in the block and look for hazards. 4224 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 4225 ++it) { 4226 4227 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 4228 Type *PhiTy = Phi->getType(); 4229 // Check that this PHI type is allowed. 4230 if (!PhiTy->isIntegerTy() && 4231 !PhiTy->isFloatingPointTy() && 4232 !PhiTy->isPointerTy()) { 4233 emitAnalysis(VectorizationReport(&*it) 4234 << "loop control flow is not understood by vectorizer"); 4235 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 4236 return false; 4237 } 4238 4239 // If this PHINode is not in the header block, then we know that we 4240 // can convert it to select during if-conversion. No need to check if 4241 // the PHIs in this block are induction or reduction variables. 4242 if (*bb != Header) { 4243 // Check that this instruction has no outside users or is an 4244 // identified reduction value with an outside user. 4245 if (!hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) 4246 continue; 4247 emitAnalysis(VectorizationReport(&*it) << 4248 "value could not be identified as " 4249 "an induction or reduction variable"); 4250 return false; 4251 } 4252 4253 // We only allow if-converted PHIs with exactly two incoming values. 4254 if (Phi->getNumIncomingValues() != 2) { 4255 emitAnalysis(VectorizationReport(&*it) 4256 << "control flow not understood by vectorizer"); 4257 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 4258 return false; 4259 } 4260 4261 InductionDescriptor ID; 4262 if (InductionDescriptor::isInductionPHI(Phi, PSE.getSE(), ID)) { 4263 Inductions[Phi] = ID; 4264 // Get the widest type. 4265 if (!WidestIndTy) 4266 WidestIndTy = convertPointerToIntegerType(DL, PhiTy); 4267 else 4268 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy); 4269 4270 // Int inductions are special because we only allow one IV. 4271 if (ID.getKind() == InductionDescriptor::IK_IntInduction && 4272 ID.getStepValue()->isOne() && 4273 isa<Constant>(ID.getStartValue()) && 4274 cast<Constant>(ID.getStartValue())->isNullValue()) { 4275 // Use the phi node with the widest type as induction. Use the last 4276 // one if there are multiple (no good reason for doing this other 4277 // than it is expedient). We've checked that it begins at zero and 4278 // steps by one, so this is a canonical induction variable. 4279 if (!Induction || PhiTy == WidestIndTy) 4280 Induction = Phi; 4281 } 4282 4283 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 4284 4285 // Until we explicitly handle the case of an induction variable with 4286 // an outside loop user we have to give up vectorizing this loop. 4287 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) { 4288 emitAnalysis(VectorizationReport(&*it) << 4289 "use of induction value outside of the " 4290 "loop is not handled by vectorizer"); 4291 return false; 4292 } 4293 4294 continue; 4295 } 4296 4297 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, 4298 Reductions[Phi])) { 4299 if (Reductions[Phi].hasUnsafeAlgebra()) 4300 Requirements->addUnsafeAlgebraInst( 4301 Reductions[Phi].getUnsafeAlgebraInst()); 4302 AllowedExit.insert(Reductions[Phi].getLoopExitInstr()); 4303 continue; 4304 } 4305 4306 emitAnalysis(VectorizationReport(&*it) << 4307 "value that could not be identified as " 4308 "reduction is used outside the loop"); 4309 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 4310 return false; 4311 }// end of PHI handling 4312 4313 // We handle calls that: 4314 // * Are debug info intrinsics. 4315 // * Have a mapping to an IR intrinsic. 4316 // * Have a vector version available. 4317 CallInst *CI = dyn_cast<CallInst>(it); 4318 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) && 4319 !(CI->getCalledFunction() && TLI && 4320 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) { 4321 emitAnalysis(VectorizationReport(&*it) 4322 << "call instruction cannot be vectorized"); 4323 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n"); 4324 return false; 4325 } 4326 4327 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the 4328 // second argument is the same (i.e. loop invariant) 4329 if (CI && 4330 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) { 4331 auto *SE = PSE.getSE(); 4332 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) { 4333 emitAnalysis(VectorizationReport(&*it) 4334 << "intrinsic instruction cannot be vectorized"); 4335 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n"); 4336 return false; 4337 } 4338 } 4339 4340 // Check that the instruction return type is vectorizable. 4341 // Also, we can't vectorize extractelement instructions. 4342 if ((!VectorType::isValidElementType(it->getType()) && 4343 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) { 4344 emitAnalysis(VectorizationReport(&*it) 4345 << "instruction return type cannot be vectorized"); 4346 DEBUG(dbgs() << "LV: Found unvectorizable type.\n"); 4347 return false; 4348 } 4349 4350 // Check that the stored type is vectorizable. 4351 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 4352 Type *T = ST->getValueOperand()->getType(); 4353 if (!VectorType::isValidElementType(T)) { 4354 emitAnalysis(VectorizationReport(ST) << 4355 "store instruction cannot be vectorized"); 4356 return false; 4357 } 4358 if (EnableMemAccessVersioning) 4359 collectStridedAccess(ST); 4360 } 4361 4362 if (EnableMemAccessVersioning) 4363 if (LoadInst *LI = dyn_cast<LoadInst>(it)) 4364 collectStridedAccess(LI); 4365 4366 // Reduction instructions are allowed to have exit users. 4367 // All other instructions must not have external users. 4368 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) { 4369 emitAnalysis(VectorizationReport(&*it) << 4370 "value cannot be used outside the loop"); 4371 return false; 4372 } 4373 4374 } // next instr. 4375 4376 } 4377 4378 if (!Induction) { 4379 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 4380 if (Inductions.empty()) { 4381 emitAnalysis(VectorizationReport() 4382 << "loop induction variable could not be identified"); 4383 return false; 4384 } 4385 } 4386 4387 // Now we know the widest induction type, check if our found induction 4388 // is the same size. If it's not, unset it here and InnerLoopVectorizer 4389 // will create another. 4390 if (Induction && WidestIndTy != Induction->getType()) 4391 Induction = nullptr; 4392 4393 return true; 4394 } 4395 4396 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) { 4397 Value *Ptr = nullptr; 4398 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess)) 4399 Ptr = LI->getPointerOperand(); 4400 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess)) 4401 Ptr = SI->getPointerOperand(); 4402 else 4403 return; 4404 4405 Value *Stride = getStrideFromPointer(Ptr, PSE.getSE(), TheLoop); 4406 if (!Stride) 4407 return; 4408 4409 DEBUG(dbgs() << "LV: Found a strided access that we can version"); 4410 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n"); 4411 Strides[Ptr] = Stride; 4412 StrideSet.insert(Stride); 4413 } 4414 4415 void LoopVectorizationLegality::collectLoopUniforms() { 4416 // We now know that the loop is vectorizable! 4417 // Collect variables that will remain uniform after vectorization. 4418 std::vector<Value*> Worklist; 4419 BasicBlock *Latch = TheLoop->getLoopLatch(); 4420 4421 // Start with the conditional branch and walk up the block. 4422 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 4423 4424 // Also add all consecutive pointer values; these values will be uniform 4425 // after vectorization (and subsequent cleanup) and, until revectorization is 4426 // supported, all dependencies must also be uniform. 4427 for (Loop::block_iterator B = TheLoop->block_begin(), 4428 BE = TheLoop->block_end(); B != BE; ++B) 4429 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end(); 4430 I != IE; ++I) 4431 if (I->getType()->isPointerTy() && isConsecutivePtr(&*I)) 4432 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end()); 4433 4434 while (!Worklist.empty()) { 4435 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 4436 Worklist.pop_back(); 4437 4438 // Look at instructions inside this loop. 4439 // Stop when reaching PHI nodes. 4440 // TODO: we need to follow values all over the loop, not only in this block. 4441 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 4442 continue; 4443 4444 // This is a known uniform. 4445 Uniforms.insert(I); 4446 4447 // Insert all operands. 4448 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end()); 4449 } 4450 } 4451 4452 bool LoopVectorizationLegality::canVectorizeMemory() { 4453 LAI = &LAA->getInfo(TheLoop, Strides); 4454 auto &OptionalReport = LAI->getReport(); 4455 if (OptionalReport) 4456 emitAnalysis(VectorizationReport(*OptionalReport)); 4457 if (!LAI->canVectorizeMemory()) 4458 return false; 4459 4460 if (LAI->hasStoreToLoopInvariantAddress()) { 4461 emitAnalysis( 4462 VectorizationReport() 4463 << "write to a loop invariant address could not be vectorized"); 4464 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 4465 return false; 4466 } 4467 4468 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks()); 4469 PSE.addPredicate(LAI->PSE.getUnionPredicate()); 4470 4471 return true; 4472 } 4473 4474 bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 4475 Value *In0 = const_cast<Value*>(V); 4476 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 4477 if (!PN) 4478 return false; 4479 4480 return Inductions.count(PN); 4481 } 4482 4483 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 4484 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT); 4485 } 4486 4487 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB, 4488 SmallPtrSetImpl<Value *> &SafePtrs) { 4489 4490 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4491 // Check that we don't have a constant expression that can trap as operand. 4492 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end(); 4493 OI != OE; ++OI) { 4494 if (Constant *C = dyn_cast<Constant>(*OI)) 4495 if (C->canTrap()) 4496 return false; 4497 } 4498 // We might be able to hoist the load. 4499 if (it->mayReadFromMemory()) { 4500 LoadInst *LI = dyn_cast<LoadInst>(it); 4501 if (!LI) 4502 return false; 4503 if (!SafePtrs.count(LI->getPointerOperand())) { 4504 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) { 4505 MaskedOp.insert(LI); 4506 continue; 4507 } 4508 return false; 4509 } 4510 } 4511 4512 // We don't predicate stores at the moment. 4513 if (it->mayWriteToMemory()) { 4514 StoreInst *SI = dyn_cast<StoreInst>(it); 4515 // We only support predication of stores in basic blocks with one 4516 // predecessor. 4517 if (!SI) 4518 return false; 4519 4520 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0); 4521 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor(); 4522 4523 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr || 4524 !isSinglePredecessor) { 4525 // Build a masked store if it is legal for the target, otherwise 4526 // scalarize the block. 4527 bool isLegalMaskedOp = 4528 isLegalMaskedStore(SI->getValueOperand()->getType(), 4529 SI->getPointerOperand()); 4530 if (isLegalMaskedOp) { 4531 --NumPredStores; 4532 MaskedOp.insert(SI); 4533 continue; 4534 } 4535 return false; 4536 } 4537 } 4538 if (it->mayThrow()) 4539 return false; 4540 4541 // The instructions below can trap. 4542 switch (it->getOpcode()) { 4543 default: continue; 4544 case Instruction::UDiv: 4545 case Instruction::SDiv: 4546 case Instruction::URem: 4547 case Instruction::SRem: 4548 return false; 4549 } 4550 } 4551 4552 return true; 4553 } 4554 4555 void InterleavedAccessInfo::collectConstStridedAccesses( 4556 MapVector<Instruction *, StrideDescriptor> &StrideAccesses, 4557 const ValueToValueMap &Strides) { 4558 // Holds load/store instructions in program order. 4559 SmallVector<Instruction *, 16> AccessList; 4560 4561 for (auto *BB : TheLoop->getBlocks()) { 4562 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT); 4563 4564 for (auto &I : *BB) { 4565 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I)) 4566 continue; 4567 // FIXME: Currently we can't handle mixed accesses and predicated accesses 4568 if (IsPred) 4569 return; 4570 4571 AccessList.push_back(&I); 4572 } 4573 } 4574 4575 if (AccessList.empty()) 4576 return; 4577 4578 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout(); 4579 for (auto I : AccessList) { 4580 LoadInst *LI = dyn_cast<LoadInst>(I); 4581 StoreInst *SI = dyn_cast<StoreInst>(I); 4582 4583 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 4584 int Stride = isStridedPtr(PSE, Ptr, TheLoop, Strides); 4585 4586 // The factor of the corresponding interleave group. 4587 unsigned Factor = std::abs(Stride); 4588 4589 // Ignore the access if the factor is too small or too large. 4590 if (Factor < 2 || Factor > MaxInterleaveGroupFactor) 4591 continue; 4592 4593 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); 4594 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 4595 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType()); 4596 4597 // An alignment of 0 means target ABI alignment. 4598 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment(); 4599 if (!Align) 4600 Align = DL.getABITypeAlignment(PtrTy->getElementType()); 4601 4602 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align); 4603 } 4604 } 4605 4606 // Analyze interleaved accesses and collect them into interleave groups. 4607 // 4608 // Notice that the vectorization on interleaved groups will change instruction 4609 // orders and may break dependences. But the memory dependence check guarantees 4610 // that there is no overlap between two pointers of different strides, element 4611 // sizes or underlying bases. 4612 // 4613 // For pointers sharing the same stride, element size and underlying base, no 4614 // need to worry about Read-After-Write dependences and Write-After-Read 4615 // dependences. 4616 // 4617 // E.g. The RAW dependence: A[i] = a; 4618 // b = A[i]; 4619 // This won't exist as it is a store-load forwarding conflict, which has 4620 // already been checked and forbidden in the dependence check. 4621 // 4622 // E.g. The WAR dependence: a = A[i]; // (1) 4623 // A[i] = b; // (2) 4624 // The store group of (2) is always inserted at or below (2), and the load group 4625 // of (1) is always inserted at or above (1). The dependence is safe. 4626 void InterleavedAccessInfo::analyzeInterleaving( 4627 const ValueToValueMap &Strides) { 4628 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n"); 4629 4630 // Holds all the stride accesses. 4631 MapVector<Instruction *, StrideDescriptor> StrideAccesses; 4632 collectConstStridedAccesses(StrideAccesses, Strides); 4633 4634 if (StrideAccesses.empty()) 4635 return; 4636 4637 // Holds all interleaved store groups temporarily. 4638 SmallSetVector<InterleaveGroup *, 4> StoreGroups; 4639 4640 // Search the load-load/write-write pair B-A in bottom-up order and try to 4641 // insert B into the interleave group of A according to 3 rules: 4642 // 1. A and B have the same stride. 4643 // 2. A and B have the same memory object size. 4644 // 3. B belongs to the group according to the distance. 4645 // 4646 // The bottom-up order can avoid breaking the Write-After-Write dependences 4647 // between two pointers of the same base. 4648 // E.g. A[i] = a; (1) 4649 // A[i] = b; (2) 4650 // A[i+1] = c (3) 4651 // We form the group (2)+(3) in front, so (1) has to form groups with accesses 4652 // above (1), which guarantees that (1) is always above (2). 4653 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E; 4654 ++I) { 4655 Instruction *A = I->first; 4656 StrideDescriptor DesA = I->second; 4657 4658 InterleaveGroup *Group = getInterleaveGroup(A); 4659 if (!Group) { 4660 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n'); 4661 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align); 4662 } 4663 4664 if (A->mayWriteToMemory()) 4665 StoreGroups.insert(Group); 4666 4667 for (auto II = std::next(I); II != E; ++II) { 4668 Instruction *B = II->first; 4669 StrideDescriptor DesB = II->second; 4670 4671 // Ignore if B is already in a group or B is a different memory operation. 4672 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory()) 4673 continue; 4674 4675 // Check the rule 1 and 2. 4676 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size) 4677 continue; 4678 4679 // Calculate the distance and prepare for the rule 3. 4680 const SCEVConstant *DistToA = dyn_cast<SCEVConstant>( 4681 PSE.getSE()->getMinusSCEV(DesB.Scev, DesA.Scev)); 4682 if (!DistToA) 4683 continue; 4684 4685 int DistanceToA = DistToA->getAPInt().getSExtValue(); 4686 4687 // Skip if the distance is not multiple of size as they are not in the 4688 // same group. 4689 if (DistanceToA % static_cast<int>(DesA.Size)) 4690 continue; 4691 4692 // The index of B is the index of A plus the related index to A. 4693 int IndexB = 4694 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size); 4695 4696 // Try to insert B into the group. 4697 if (Group->insertMember(B, IndexB, DesB.Align)) { 4698 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n' 4699 << " into the interleave group with" << *A << '\n'); 4700 InterleaveGroupMap[B] = Group; 4701 4702 // Set the first load in program order as the insert position. 4703 if (B->mayReadFromMemory()) 4704 Group->setInsertPos(B); 4705 } 4706 } // Iteration on instruction B 4707 } // Iteration on instruction A 4708 4709 // Remove interleaved store groups with gaps. 4710 for (InterleaveGroup *Group : StoreGroups) 4711 if (Group->getNumMembers() != Group->getFactor()) 4712 releaseGroup(Group); 4713 } 4714 4715 LoopVectorizationCostModel::VectorizationFactor 4716 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) { 4717 // Width 1 means no vectorize 4718 VectorizationFactor Factor = { 1U, 0U }; 4719 if (OptForSize && Legal->getRuntimePointerChecking()->Need) { 4720 emitAnalysis(VectorizationReport() << 4721 "runtime pointer checks needed. Enable vectorization of this " 4722 "loop with '#pragma clang loop vectorize(enable)' when " 4723 "compiling with -Os/-Oz"); 4724 DEBUG(dbgs() << 4725 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n"); 4726 return Factor; 4727 } 4728 4729 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) { 4730 emitAnalysis(VectorizationReport() << 4731 "store that is conditionally executed prevents vectorization"); 4732 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n"); 4733 return Factor; 4734 } 4735 4736 // Find the trip count. 4737 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); 4738 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n'); 4739 4740 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); 4741 unsigned SmallestType, WidestType; 4742 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); 4743 unsigned WidestRegister = TTI.getRegisterBitWidth(true); 4744 unsigned MaxSafeDepDist = -1U; 4745 if (Legal->getMaxSafeDepDistBytes() != -1U) 4746 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8; 4747 WidestRegister = ((WidestRegister < MaxSafeDepDist) ? 4748 WidestRegister : MaxSafeDepDist); 4749 unsigned MaxVectorSize = WidestRegister / WidestType; 4750 4751 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " 4752 << WidestType << " bits.\n"); 4753 DEBUG(dbgs() << "LV: The Widest register is: " 4754 << WidestRegister << " bits.\n"); 4755 4756 if (MaxVectorSize == 0) { 4757 DEBUG(dbgs() << "LV: The target has no vector registers.\n"); 4758 MaxVectorSize = 1; 4759 } 4760 4761 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements" 4762 " into one vector!"); 4763 4764 unsigned VF = MaxVectorSize; 4765 if (MaximizeBandwidth && !OptForSize) { 4766 // Collect all viable vectorization factors. 4767 SmallVector<unsigned, 8> VFs; 4768 unsigned NewMaxVectorSize = WidestRegister / SmallestType; 4769 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2) 4770 VFs.push_back(VS); 4771 4772 // For each VF calculate its register usage. 4773 auto RUs = calculateRegisterUsage(VFs); 4774 4775 // Select the largest VF which doesn't require more registers than existing 4776 // ones. 4777 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true); 4778 for (int i = RUs.size() - 1; i >= 0; --i) { 4779 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) { 4780 VF = VFs[i]; 4781 break; 4782 } 4783 } 4784 } 4785 4786 // If we optimize the program for size, avoid creating the tail loop. 4787 if (OptForSize) { 4788 // If we are unable to calculate the trip count then don't try to vectorize. 4789 if (TC < 2) { 4790 emitAnalysis 4791 (VectorizationReport() << 4792 "unable to calculate the loop count due to complex control flow"); 4793 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n"); 4794 return Factor; 4795 } 4796 4797 // Find the maximum SIMD width that can fit within the trip count. 4798 VF = TC % MaxVectorSize; 4799 4800 if (VF == 0) 4801 VF = MaxVectorSize; 4802 else { 4803 // If the trip count that we found modulo the vectorization factor is not 4804 // zero then we require a tail. 4805 emitAnalysis(VectorizationReport() << 4806 "cannot optimize for size and vectorize at the " 4807 "same time. Enable vectorization of this loop " 4808 "with '#pragma clang loop vectorize(enable)' " 4809 "when compiling with -Os/-Oz"); 4810 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n"); 4811 return Factor; 4812 } 4813 } 4814 4815 int UserVF = Hints->getWidth(); 4816 if (UserVF != 0) { 4817 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 4818 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n"); 4819 4820 Factor.Width = UserVF; 4821 return Factor; 4822 } 4823 4824 float Cost = expectedCost(1); 4825 #ifndef NDEBUG 4826 const float ScalarCost = Cost; 4827 #endif /* NDEBUG */ 4828 unsigned Width = 1; 4829 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n"); 4830 4831 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; 4832 // Ignore scalar width, because the user explicitly wants vectorization. 4833 if (ForceVectorization && VF > 1) { 4834 Width = 2; 4835 Cost = expectedCost(Width) / (float)Width; 4836 } 4837 4838 for (unsigned i=2; i <= VF; i*=2) { 4839 // Notice that the vector loop needs to be executed less times, so 4840 // we need to divide the cost of the vector loops by the width of 4841 // the vector elements. 4842 float VectorCost = expectedCost(i) / (float)i; 4843 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " << 4844 (int)VectorCost << ".\n"); 4845 if (VectorCost < Cost) { 4846 Cost = VectorCost; 4847 Width = i; 4848 } 4849 } 4850 4851 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs() 4852 << "LV: Vectorization seems to be not beneficial, " 4853 << "but was forced by a user.\n"); 4854 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n"); 4855 Factor.Width = Width; 4856 Factor.Cost = Width * Cost; 4857 return Factor; 4858 } 4859 4860 std::pair<unsigned, unsigned> 4861 LoopVectorizationCostModel::getSmallestAndWidestTypes() { 4862 unsigned MinWidth = -1U; 4863 unsigned MaxWidth = 8; 4864 const DataLayout &DL = TheFunction->getParent()->getDataLayout(); 4865 4866 // For each block. 4867 for (Loop::block_iterator bb = TheLoop->block_begin(), 4868 be = TheLoop->block_end(); bb != be; ++bb) { 4869 BasicBlock *BB = *bb; 4870 4871 // For each instruction in the loop. 4872 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4873 Type *T = it->getType(); 4874 4875 // Skip ignored values. 4876 if (ValuesToIgnore.count(&*it)) 4877 continue; 4878 4879 // Only examine Loads, Stores and PHINodes. 4880 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it)) 4881 continue; 4882 4883 // Examine PHI nodes that are reduction variables. Update the type to 4884 // account for the recurrence type. 4885 if (PHINode *PN = dyn_cast<PHINode>(it)) { 4886 if (!Legal->isReductionVariable(PN)) 4887 continue; 4888 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN]; 4889 T = RdxDesc.getRecurrenceType(); 4890 } 4891 4892 // Examine the stored values. 4893 if (StoreInst *ST = dyn_cast<StoreInst>(it)) 4894 T = ST->getValueOperand()->getType(); 4895 4896 // Ignore loaded pointer types and stored pointer types that are not 4897 // consecutive. However, we do want to take consecutive stores/loads of 4898 // pointer vectors into account. 4899 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&*it)) 4900 continue; 4901 4902 MinWidth = std::min(MinWidth, 4903 (unsigned)DL.getTypeSizeInBits(T->getScalarType())); 4904 MaxWidth = std::max(MaxWidth, 4905 (unsigned)DL.getTypeSizeInBits(T->getScalarType())); 4906 } 4907 } 4908 4909 return {MinWidth, MaxWidth}; 4910 } 4911 4912 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize, 4913 unsigned VF, 4914 unsigned LoopCost) { 4915 4916 // -- The interleave heuristics -- 4917 // We interleave the loop in order to expose ILP and reduce the loop overhead. 4918 // There are many micro-architectural considerations that we can't predict 4919 // at this level. For example, frontend pressure (on decode or fetch) due to 4920 // code size, or the number and capabilities of the execution ports. 4921 // 4922 // We use the following heuristics to select the interleave count: 4923 // 1. If the code has reductions, then we interleave to break the cross 4924 // iteration dependency. 4925 // 2. If the loop is really small, then we interleave to reduce the loop 4926 // overhead. 4927 // 3. We don't interleave if we think that we will spill registers to memory 4928 // due to the increased register pressure. 4929 4930 // When we optimize for size, we don't interleave. 4931 if (OptForSize) 4932 return 1; 4933 4934 // We used the distance for the interleave count. 4935 if (Legal->getMaxSafeDepDistBytes() != -1U) 4936 return 1; 4937 4938 // Do not interleave loops with a relatively small trip count. 4939 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); 4940 if (TC > 1 && TC < TinyTripCountInterleaveThreshold) 4941 return 1; 4942 4943 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1); 4944 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters << 4945 " registers\n"); 4946 4947 if (VF == 1) { 4948 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) 4949 TargetNumRegisters = ForceTargetNumScalarRegs; 4950 } else { 4951 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) 4952 TargetNumRegisters = ForceTargetNumVectorRegs; 4953 } 4954 4955 RegisterUsage R = calculateRegisterUsage({VF})[0]; 4956 // We divide by these constants so assume that we have at least one 4957 // instruction that uses at least one register. 4958 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 4959 R.NumInstructions = std::max(R.NumInstructions, 1U); 4960 4961 // We calculate the interleave count using the following formula. 4962 // Subtract the number of loop invariants from the number of available 4963 // registers. These registers are used by all of the interleaved instances. 4964 // Next, divide the remaining registers by the number of registers that is 4965 // required by the loop, in order to estimate how many parallel instances 4966 // fit without causing spills. All of this is rounded down if necessary to be 4967 // a power of two. We want power of two interleave count to simplify any 4968 // addressing operations or alignment considerations. 4969 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) / 4970 R.MaxLocalUsers); 4971 4972 // Don't count the induction variable as interleaved. 4973 if (EnableIndVarRegisterHeur) 4974 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) / 4975 std::max(1U, (R.MaxLocalUsers - 1))); 4976 4977 // Clamp the interleave ranges to reasonable counts. 4978 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF); 4979 4980 // Check if the user has overridden the max. 4981 if (VF == 1) { 4982 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) 4983 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; 4984 } else { 4985 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) 4986 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; 4987 } 4988 4989 // If we did not calculate the cost for VF (because the user selected the VF) 4990 // then we calculate the cost of VF here. 4991 if (LoopCost == 0) 4992 LoopCost = expectedCost(VF); 4993 4994 // Clamp the calculated IC to be between the 1 and the max interleave count 4995 // that the target allows. 4996 if (IC > MaxInterleaveCount) 4997 IC = MaxInterleaveCount; 4998 else if (IC < 1) 4999 IC = 1; 5000 5001 // Interleave if we vectorized this loop and there is a reduction that could 5002 // benefit from interleaving. 5003 if (VF > 1 && Legal->getReductionVars()->size()) { 5004 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n"); 5005 return IC; 5006 } 5007 5008 // Note that if we've already vectorized the loop we will have done the 5009 // runtime check and so interleaving won't require further checks. 5010 bool InterleavingRequiresRuntimePointerCheck = 5011 (VF == 1 && Legal->getRuntimePointerChecking()->Need); 5012 5013 // We want to interleave small loops in order to reduce the loop overhead and 5014 // potentially expose ILP opportunities. 5015 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n'); 5016 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) { 5017 // We assume that the cost overhead is 1 and we use the cost model 5018 // to estimate the cost of the loop and interleave until the cost of the 5019 // loop overhead is about 5% of the cost of the loop. 5020 unsigned SmallIC = 5021 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost)); 5022 5023 // Interleave until store/load ports (estimated by max interleave count) are 5024 // saturated. 5025 unsigned NumStores = Legal->getNumStores(); 5026 unsigned NumLoads = Legal->getNumLoads(); 5027 unsigned StoresIC = IC / (NumStores ? NumStores : 1); 5028 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); 5029 5030 // If we have a scalar reduction (vector reductions are already dealt with 5031 // by this point), we can increase the critical path length if the loop 5032 // we're interleaving is inside another loop. Limit, by default to 2, so the 5033 // critical path only gets increased by one reduction operation. 5034 if (Legal->getReductionVars()->size() && 5035 TheLoop->getLoopDepth() > 1) { 5036 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC); 5037 SmallIC = std::min(SmallIC, F); 5038 StoresIC = std::min(StoresIC, F); 5039 LoadsIC = std::min(LoadsIC, F); 5040 } 5041 5042 if (EnableLoadStoreRuntimeInterleave && 5043 std::max(StoresIC, LoadsIC) > SmallIC) { 5044 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n"); 5045 return std::max(StoresIC, LoadsIC); 5046 } 5047 5048 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n"); 5049 return SmallIC; 5050 } 5051 5052 // Interleave if this is a large loop (small loops are already dealt with by 5053 // this point) that could benefit from interleaving. 5054 bool HasReductions = (Legal->getReductionVars()->size() > 0); 5055 if (TTI.enableAggressiveInterleaving(HasReductions)) { 5056 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n"); 5057 return IC; 5058 } 5059 5060 DEBUG(dbgs() << "LV: Not Interleaving.\n"); 5061 return 1; 5062 } 5063 5064 SmallVector<LoopVectorizationCostModel::RegisterUsage, 8> 5065 LoopVectorizationCostModel::calculateRegisterUsage( 5066 const SmallVector<unsigned, 8> &VFs) { 5067 // This function calculates the register usage by measuring the highest number 5068 // of values that are alive at a single location. Obviously, this is a very 5069 // rough estimation. We scan the loop in a topological order in order and 5070 // assign a number to each instruction. We use RPO to ensure that defs are 5071 // met before their users. We assume that each instruction that has in-loop 5072 // users starts an interval. We record every time that an in-loop value is 5073 // used, so we have a list of the first and last occurrences of each 5074 // instruction. Next, we transpose this data structure into a multi map that 5075 // holds the list of intervals that *end* at a specific location. This multi 5076 // map allows us to perform a linear search. We scan the instructions linearly 5077 // and record each time that a new interval starts, by placing it in a set. 5078 // If we find this value in the multi-map then we remove it from the set. 5079 // The max register usage is the maximum size of the set. 5080 // We also search for instructions that are defined outside the loop, but are 5081 // used inside the loop. We need this number separately from the max-interval 5082 // usage number because when we unroll, loop-invariant values do not take 5083 // more register. 5084 LoopBlocksDFS DFS(TheLoop); 5085 DFS.perform(LI); 5086 5087 RegisterUsage RU; 5088 RU.NumInstructions = 0; 5089 5090 // Each 'key' in the map opens a new interval. The values 5091 // of the map are the index of the 'last seen' usage of the 5092 // instruction that is the key. 5093 typedef DenseMap<Instruction*, unsigned> IntervalMap; 5094 // Maps instruction to its index. 5095 DenseMap<unsigned, Instruction*> IdxToInstr; 5096 // Marks the end of each interval. 5097 IntervalMap EndPoint; 5098 // Saves the list of instruction indices that are used in the loop. 5099 SmallSet<Instruction*, 8> Ends; 5100 // Saves the list of values that are used in the loop but are 5101 // defined outside the loop, such as arguments and constants. 5102 SmallPtrSet<Value*, 8> LoopInvariants; 5103 5104 unsigned Index = 0; 5105 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 5106 be = DFS.endRPO(); bb != be; ++bb) { 5107 RU.NumInstructions += (*bb)->size(); 5108 for (Instruction &I : **bb) { 5109 IdxToInstr[Index++] = &I; 5110 5111 // Save the end location of each USE. 5112 for (unsigned i = 0; i < I.getNumOperands(); ++i) { 5113 Value *U = I.getOperand(i); 5114 Instruction *Instr = dyn_cast<Instruction>(U); 5115 5116 // Ignore non-instruction values such as arguments, constants, etc. 5117 if (!Instr) continue; 5118 5119 // If this instruction is outside the loop then record it and continue. 5120 if (!TheLoop->contains(Instr)) { 5121 LoopInvariants.insert(Instr); 5122 continue; 5123 } 5124 5125 // Overwrite previous end points. 5126 EndPoint[Instr] = Index; 5127 Ends.insert(Instr); 5128 } 5129 } 5130 } 5131 5132 // Saves the list of intervals that end with the index in 'key'. 5133 typedef SmallVector<Instruction*, 2> InstrList; 5134 DenseMap<unsigned, InstrList> TransposeEnds; 5135 5136 // Transpose the EndPoints to a list of values that end at each index. 5137 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 5138 it != e; ++it) 5139 TransposeEnds[it->second].push_back(it->first); 5140 5141 SmallSet<Instruction*, 8> OpenIntervals; 5142 5143 // Get the size of the widest register. 5144 unsigned MaxSafeDepDist = -1U; 5145 if (Legal->getMaxSafeDepDistBytes() != -1U) 5146 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8; 5147 unsigned WidestRegister = 5148 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist); 5149 const DataLayout &DL = TheFunction->getParent()->getDataLayout(); 5150 5151 SmallVector<RegisterUsage, 8> RUs(VFs.size()); 5152 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0); 5153 5154 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 5155 5156 // A lambda that gets the register usage for the given type and VF. 5157 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) { 5158 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType()); 5159 return std::max<unsigned>(1, VF * TypeSize / WidestRegister); 5160 }; 5161 5162 for (unsigned int i = 0; i < Index; ++i) { 5163 Instruction *I = IdxToInstr[i]; 5164 // Ignore instructions that are never used within the loop. 5165 if (!Ends.count(I)) continue; 5166 5167 // Remove all of the instructions that end at this location. 5168 InstrList &List = TransposeEnds[i]; 5169 for (unsigned int j = 0, e = List.size(); j < e; ++j) 5170 OpenIntervals.erase(List[j]); 5171 5172 // Skip ignored values. 5173 if (ValuesToIgnore.count(I)) 5174 continue; 5175 5176 // For each VF find the maximum usage of registers. 5177 for (unsigned j = 0, e = VFs.size(); j < e; ++j) { 5178 if (VFs[j] == 1) { 5179 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size()); 5180 continue; 5181 } 5182 5183 // Count the number of live intervals. 5184 unsigned RegUsage = 0; 5185 for (auto Inst : OpenIntervals) { 5186 // Skip ignored values for VF > 1. 5187 if (VecValuesToIgnore.count(Inst)) 5188 continue; 5189 RegUsage += GetRegUsage(Inst->getType(), VFs[j]); 5190 } 5191 MaxUsages[j] = std::max(MaxUsages[j], RegUsage); 5192 } 5193 5194 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " 5195 << OpenIntervals.size() << '\n'); 5196 5197 // Add the current instruction to the list of open intervals. 5198 OpenIntervals.insert(I); 5199 } 5200 5201 for (unsigned i = 0, e = VFs.size(); i < e; ++i) { 5202 unsigned Invariant = 0; 5203 if (VFs[i] == 1) 5204 Invariant = LoopInvariants.size(); 5205 else { 5206 for (auto Inst : LoopInvariants) 5207 Invariant += GetRegUsage(Inst->getType(), VFs[i]); 5208 } 5209 5210 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n'); 5211 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n'); 5212 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n'); 5213 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n'); 5214 5215 RU.LoopInvariantRegs = Invariant; 5216 RU.MaxLocalUsers = MaxUsages[i]; 5217 RUs[i] = RU; 5218 } 5219 5220 return RUs; 5221 } 5222 5223 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 5224 unsigned Cost = 0; 5225 5226 // For each block. 5227 for (Loop::block_iterator bb = TheLoop->block_begin(), 5228 be = TheLoop->block_end(); bb != be; ++bb) { 5229 unsigned BlockCost = 0; 5230 BasicBlock *BB = *bb; 5231 5232 // For each instruction in the old loop. 5233 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 5234 // Skip dbg intrinsics. 5235 if (isa<DbgInfoIntrinsic>(it)) 5236 continue; 5237 5238 // Skip ignored values. 5239 if (ValuesToIgnore.count(&*it)) 5240 continue; 5241 5242 unsigned C = getInstructionCost(&*it, VF); 5243 5244 // Check if we should override the cost. 5245 if (ForceTargetInstructionCost.getNumOccurrences() > 0) 5246 C = ForceTargetInstructionCost; 5247 5248 BlockCost += C; 5249 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " << 5250 VF << " For instruction: " << *it << '\n'); 5251 } 5252 5253 // We assume that if-converted blocks have a 50% chance of being executed. 5254 // When the code is scalar then some of the blocks are avoided due to CF. 5255 // When the code is vectorized we execute all code paths. 5256 if (VF == 1 && Legal->blockNeedsPredication(*bb)) 5257 BlockCost /= 2; 5258 5259 Cost += BlockCost; 5260 } 5261 5262 return Cost; 5263 } 5264 5265 /// \brief Check whether the address computation for a non-consecutive memory 5266 /// access looks like an unlikely candidate for being merged into the indexing 5267 /// mode. 5268 /// 5269 /// We look for a GEP which has one index that is an induction variable and all 5270 /// other indices are loop invariant. If the stride of this access is also 5271 /// within a small bound we decide that this address computation can likely be 5272 /// merged into the addressing mode. 5273 /// In all other cases, we identify the address computation as complex. 5274 static bool isLikelyComplexAddressComputation(Value *Ptr, 5275 LoopVectorizationLegality *Legal, 5276 ScalarEvolution *SE, 5277 const Loop *TheLoop) { 5278 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 5279 if (!Gep) 5280 return true; 5281 5282 // We are looking for a gep with all loop invariant indices except for one 5283 // which should be an induction variable. 5284 unsigned NumOperands = Gep->getNumOperands(); 5285 for (unsigned i = 1; i < NumOperands; ++i) { 5286 Value *Opd = Gep->getOperand(i); 5287 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && 5288 !Legal->isInductionVariable(Opd)) 5289 return true; 5290 } 5291 5292 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step 5293 // can likely be merged into the address computation. 5294 unsigned MaxMergeDistance = 64; 5295 5296 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr)); 5297 if (!AddRec) 5298 return true; 5299 5300 // Check the step is constant. 5301 const SCEV *Step = AddRec->getStepRecurrence(*SE); 5302 // Calculate the pointer stride and check if it is consecutive. 5303 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 5304 if (!C) 5305 return true; 5306 5307 const APInt &APStepVal = C->getAPInt(); 5308 5309 // Huge step value - give up. 5310 if (APStepVal.getBitWidth() > 64) 5311 return true; 5312 5313 int64_t StepVal = APStepVal.getSExtValue(); 5314 5315 return StepVal > MaxMergeDistance; 5316 } 5317 5318 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { 5319 return Legal->hasStride(I->getOperand(0)) || 5320 Legal->hasStride(I->getOperand(1)); 5321 } 5322 5323 unsigned 5324 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 5325 // If we know that this instruction will remain uniform, check the cost of 5326 // the scalar version. 5327 if (Legal->isUniformAfterVectorization(I)) 5328 VF = 1; 5329 5330 Type *RetTy = I->getType(); 5331 if (VF > 1 && MinBWs.count(I)) 5332 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]); 5333 Type *VectorTy = ToVectorTy(RetTy, VF); 5334 auto SE = PSE.getSE(); 5335 5336 // TODO: We need to estimate the cost of intrinsic calls. 5337 switch (I->getOpcode()) { 5338 case Instruction::GetElementPtr: 5339 // We mark this instruction as zero-cost because the cost of GEPs in 5340 // vectorized code depends on whether the corresponding memory instruction 5341 // is scalarized or not. Therefore, we handle GEPs with the memory 5342 // instruction cost. 5343 return 0; 5344 case Instruction::Br: { 5345 return TTI.getCFInstrCost(I->getOpcode()); 5346 } 5347 case Instruction::PHI: 5348 //TODO: IF-converted IFs become selects. 5349 return 0; 5350 case Instruction::Add: 5351 case Instruction::FAdd: 5352 case Instruction::Sub: 5353 case Instruction::FSub: 5354 case Instruction::Mul: 5355 case Instruction::FMul: 5356 case Instruction::UDiv: 5357 case Instruction::SDiv: 5358 case Instruction::FDiv: 5359 case Instruction::URem: 5360 case Instruction::SRem: 5361 case Instruction::FRem: 5362 case Instruction::Shl: 5363 case Instruction::LShr: 5364 case Instruction::AShr: 5365 case Instruction::And: 5366 case Instruction::Or: 5367 case Instruction::Xor: { 5368 // Since we will replace the stride by 1 the multiplication should go away. 5369 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) 5370 return 0; 5371 // Certain instructions can be cheaper to vectorize if they have a constant 5372 // second vector operand. One example of this are shifts on x86. 5373 TargetTransformInfo::OperandValueKind Op1VK = 5374 TargetTransformInfo::OK_AnyValue; 5375 TargetTransformInfo::OperandValueKind Op2VK = 5376 TargetTransformInfo::OK_AnyValue; 5377 TargetTransformInfo::OperandValueProperties Op1VP = 5378 TargetTransformInfo::OP_None; 5379 TargetTransformInfo::OperandValueProperties Op2VP = 5380 TargetTransformInfo::OP_None; 5381 Value *Op2 = I->getOperand(1); 5382 5383 // Check for a splat of a constant or for a non uniform vector of constants. 5384 if (isa<ConstantInt>(Op2)) { 5385 ConstantInt *CInt = cast<ConstantInt>(Op2); 5386 if (CInt && CInt->getValue().isPowerOf2()) 5387 Op2VP = TargetTransformInfo::OP_PowerOf2; 5388 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 5389 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) { 5390 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; 5391 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue(); 5392 if (SplatValue) { 5393 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue); 5394 if (CInt && CInt->getValue().isPowerOf2()) 5395 Op2VP = TargetTransformInfo::OP_PowerOf2; 5396 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 5397 } 5398 } 5399 5400 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK, 5401 Op1VP, Op2VP); 5402 } 5403 case Instruction::Select: { 5404 SelectInst *SI = cast<SelectInst>(I); 5405 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 5406 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 5407 Type *CondTy = SI->getCondition()->getType(); 5408 if (!ScalarCond) 5409 CondTy = VectorType::get(CondTy, VF); 5410 5411 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 5412 } 5413 case Instruction::ICmp: 5414 case Instruction::FCmp: { 5415 Type *ValTy = I->getOperand(0)->getType(); 5416 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0)); 5417 auto It = MinBWs.find(Op0AsInstruction); 5418 if (VF > 1 && It != MinBWs.end()) 5419 ValTy = IntegerType::get(ValTy->getContext(), It->second); 5420 VectorTy = ToVectorTy(ValTy, VF); 5421 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); 5422 } 5423 case Instruction::Store: 5424 case Instruction::Load: { 5425 StoreInst *SI = dyn_cast<StoreInst>(I); 5426 LoadInst *LI = dyn_cast<LoadInst>(I); 5427 Type *ValTy = (SI ? SI->getValueOperand()->getType() : 5428 LI->getType()); 5429 VectorTy = ToVectorTy(ValTy, VF); 5430 5431 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); 5432 unsigned AS = SI ? SI->getPointerAddressSpace() : 5433 LI->getPointerAddressSpace(); 5434 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); 5435 // We add the cost of address computation here instead of with the gep 5436 // instruction because only here we know whether the operation is 5437 // scalarized. 5438 if (VF == 1) 5439 return TTI.getAddressComputationCost(VectorTy) + 5440 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 5441 5442 // For an interleaved access, calculate the total cost of the whole 5443 // interleave group. 5444 if (Legal->isAccessInterleaved(I)) { 5445 auto Group = Legal->getInterleavedAccessGroup(I); 5446 assert(Group && "Fail to get an interleaved access group."); 5447 5448 // Only calculate the cost once at the insert position. 5449 if (Group->getInsertPos() != I) 5450 return 0; 5451 5452 unsigned InterleaveFactor = Group->getFactor(); 5453 Type *WideVecTy = 5454 VectorType::get(VectorTy->getVectorElementType(), 5455 VectorTy->getVectorNumElements() * InterleaveFactor); 5456 5457 // Holds the indices of existing members in an interleaved load group. 5458 // An interleaved store group doesn't need this as it dones't allow gaps. 5459 SmallVector<unsigned, 4> Indices; 5460 if (LI) { 5461 for (unsigned i = 0; i < InterleaveFactor; i++) 5462 if (Group->getMember(i)) 5463 Indices.push_back(i); 5464 } 5465 5466 // Calculate the cost of the whole interleaved group. 5467 unsigned Cost = TTI.getInterleavedMemoryOpCost( 5468 I->getOpcode(), WideVecTy, Group->getFactor(), Indices, 5469 Group->getAlignment(), AS); 5470 5471 if (Group->isReverse()) 5472 Cost += 5473 Group->getNumMembers() * 5474 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0); 5475 5476 // FIXME: The interleaved load group with a huge gap could be even more 5477 // expensive than scalar operations. Then we could ignore such group and 5478 // use scalar operations instead. 5479 return Cost; 5480 } 5481 5482 // Scalarized loads/stores. 5483 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 5484 bool Reverse = ConsecutiveStride < 0; 5485 const DataLayout &DL = I->getModule()->getDataLayout(); 5486 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy); 5487 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF; 5488 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { 5489 bool IsComplexComputation = 5490 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop); 5491 unsigned Cost = 0; 5492 // The cost of extracting from the value vector and pointer vector. 5493 Type *PtrTy = ToVectorTy(Ptr->getType(), VF); 5494 for (unsigned i = 0; i < VF; ++i) { 5495 // The cost of extracting the pointer operand. 5496 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); 5497 // In case of STORE, the cost of ExtractElement from the vector. 5498 // In case of LOAD, the cost of InsertElement into the returned 5499 // vector. 5500 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : 5501 Instruction::InsertElement, 5502 VectorTy, i); 5503 } 5504 5505 // The cost of the scalar loads/stores. 5506 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation); 5507 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), 5508 Alignment, AS); 5509 return Cost; 5510 } 5511 5512 // Wide load/stores. 5513 unsigned Cost = TTI.getAddressComputationCost(VectorTy); 5514 if (Legal->isMaskRequired(I)) 5515 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, 5516 AS); 5517 else 5518 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 5519 5520 if (Reverse) 5521 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, 5522 VectorTy, 0); 5523 return Cost; 5524 } 5525 case Instruction::ZExt: 5526 case Instruction::SExt: 5527 case Instruction::FPToUI: 5528 case Instruction::FPToSI: 5529 case Instruction::FPExt: 5530 case Instruction::PtrToInt: 5531 case Instruction::IntToPtr: 5532 case Instruction::SIToFP: 5533 case Instruction::UIToFP: 5534 case Instruction::Trunc: 5535 case Instruction::FPTrunc: 5536 case Instruction::BitCast: { 5537 // We optimize the truncation of induction variable. 5538 // The cost of these is the same as the scalar operation. 5539 if (I->getOpcode() == Instruction::Trunc && 5540 Legal->isInductionVariable(I->getOperand(0))) 5541 return TTI.getCastInstrCost(I->getOpcode(), I->getType(), 5542 I->getOperand(0)->getType()); 5543 5544 Type *SrcScalarTy = I->getOperand(0)->getType(); 5545 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF); 5546 if (VF > 1 && MinBWs.count(I)) { 5547 // This cast is going to be shrunk. This may remove the cast or it might 5548 // turn it into slightly different cast. For example, if MinBW == 16, 5549 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16". 5550 // 5551 // Calculate the modified src and dest types. 5552 Type *MinVecTy = VectorTy; 5553 if (I->getOpcode() == Instruction::Trunc) { 5554 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy); 5555 VectorTy = largestIntegerVectorType(ToVectorTy(I->getType(), VF), 5556 MinVecTy); 5557 } else if (I->getOpcode() == Instruction::ZExt || 5558 I->getOpcode() == Instruction::SExt) { 5559 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy); 5560 VectorTy = smallestIntegerVectorType(ToVectorTy(I->getType(), VF), 5561 MinVecTy); 5562 } 5563 } 5564 5565 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 5566 } 5567 case Instruction::Call: { 5568 bool NeedToScalarize; 5569 CallInst *CI = cast<CallInst>(I); 5570 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize); 5571 if (getIntrinsicIDForCall(CI, TLI)) 5572 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI)); 5573 return CallCost; 5574 } 5575 default: { 5576 // We are scalarizing the instruction. Return the cost of the scalar 5577 // instruction, plus the cost of insert and extract into vector 5578 // elements, times the vector width. 5579 unsigned Cost = 0; 5580 5581 if (!RetTy->isVoidTy() && VF != 1) { 5582 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, 5583 VectorTy); 5584 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, 5585 VectorTy); 5586 5587 // The cost of inserting the results plus extracting each one of the 5588 // operands. 5589 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 5590 } 5591 5592 // The cost of executing VF copies of the scalar instruction. This opcode 5593 // is unknown. Assume that it is the same as 'mul'. 5594 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); 5595 return Cost; 5596 } 5597 }// end of switch. 5598 } 5599 5600 char LoopVectorize::ID = 0; 5601 static const char lv_name[] = "Loop Vectorization"; 5602 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 5603 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 5604 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass) 5605 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 5606 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) 5607 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 5608 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 5609 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 5610 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 5611 INITIALIZE_PASS_DEPENDENCY(LCSSA) 5612 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 5613 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 5614 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis) 5615 INITIALIZE_PASS_DEPENDENCY(DemandedBits) 5616 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 5617 5618 namespace llvm { 5619 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) { 5620 return new LoopVectorize(NoUnrolling, AlwaysVectorize); 5621 } 5622 } 5623 5624 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { 5625 // Check for a store. 5626 if (StoreInst *ST = dyn_cast<StoreInst>(Inst)) 5627 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; 5628 5629 // Check for a load. 5630 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 5631 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; 5632 5633 return false; 5634 } 5635 5636 void LoopVectorizationCostModel::collectValuesToIgnore() { 5637 // Ignore ephemeral values. 5638 CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); 5639 5640 // Ignore type-promoting instructions we identified during reduction 5641 // detection. 5642 for (auto &Reduction : *Legal->getReductionVars()) { 5643 RecurrenceDescriptor &RedDes = Reduction.second; 5644 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); 5645 VecValuesToIgnore.insert(Casts.begin(), Casts.end()); 5646 } 5647 5648 // Ignore induction phis that are only used in either GetElementPtr or ICmp 5649 // instruction to exit loop. Induction variables usually have large types and 5650 // can have big impact when estimating register usage. 5651 // This is for when VF > 1. 5652 for (auto &Induction : *Legal->getInductionVars()) { 5653 auto *PN = Induction.first; 5654 auto *UpdateV = PN->getIncomingValueForBlock(TheLoop->getLoopLatch()); 5655 5656 // Check that the PHI is only used by the induction increment (UpdateV) or 5657 // by GEPs. Then check that UpdateV is only used by a compare instruction or 5658 // the loop header PHI. 5659 // FIXME: Need precise def-use analysis to determine if this instruction 5660 // variable will be vectorized. 5661 if (std::all_of(PN->user_begin(), PN->user_end(), 5662 [&](const User *U) -> bool { 5663 return U == UpdateV || isa<GetElementPtrInst>(U); 5664 }) && 5665 std::all_of(UpdateV->user_begin(), UpdateV->user_end(), 5666 [&](const User *U) -> bool { 5667 return U == PN || isa<ICmpInst>(U); 5668 })) { 5669 VecValuesToIgnore.insert(PN); 5670 VecValuesToIgnore.insert(UpdateV); 5671 } 5672 } 5673 5674 // Ignore instructions that will not be vectorized. 5675 // This is for when VF > 1. 5676 for (auto bb = TheLoop->block_begin(), be = TheLoop->block_end(); bb != be; 5677 ++bb) { 5678 for (auto &Inst : **bb) { 5679 switch (Inst.getOpcode()) { 5680 case Instruction::GetElementPtr: { 5681 // Ignore GEP if its last operand is an induction variable so that it is 5682 // a consecutive load/store and won't be vectorized as scatter/gather 5683 // pattern. 5684 5685 GetElementPtrInst *Gep = cast<GetElementPtrInst>(&Inst); 5686 unsigned NumOperands = Gep->getNumOperands(); 5687 unsigned InductionOperand = getGEPInductionOperand(Gep); 5688 bool GepToIgnore = true; 5689 5690 // Check that all of the gep indices are uniform except for the 5691 // induction operand. 5692 for (unsigned i = 0; i != NumOperands; ++i) { 5693 if (i != InductionOperand && 5694 !PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)), 5695 TheLoop)) { 5696 GepToIgnore = false; 5697 break; 5698 } 5699 } 5700 5701 if (GepToIgnore) 5702 VecValuesToIgnore.insert(&Inst); 5703 break; 5704 } 5705 } 5706 } 5707 } 5708 } 5709 5710 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr, 5711 bool IfPredicateStore) { 5712 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 5713 // Holds vector parameters or scalars, in case of uniform vals. 5714 SmallVector<VectorParts, 4> Params; 5715 5716 setDebugLocFromInst(Builder, Instr); 5717 5718 // Find all of the vectorized parameters. 5719 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5720 Value *SrcOp = Instr->getOperand(op); 5721 5722 // If we are accessing the old induction variable, use the new one. 5723 if (SrcOp == OldInduction) { 5724 Params.push_back(getVectorValue(SrcOp)); 5725 continue; 5726 } 5727 5728 // Try using previously calculated values. 5729 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 5730 5731 // If the src is an instruction that appeared earlier in the basic block 5732 // then it should already be vectorized. 5733 if (SrcInst && OrigLoop->contains(SrcInst)) { 5734 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 5735 // The parameter is a vector value from earlier. 5736 Params.push_back(WidenMap.get(SrcInst)); 5737 } else { 5738 // The parameter is a scalar from outside the loop. Maybe even a constant. 5739 VectorParts Scalars; 5740 Scalars.append(UF, SrcOp); 5741 Params.push_back(Scalars); 5742 } 5743 } 5744 5745 assert(Params.size() == Instr->getNumOperands() && 5746 "Invalid number of operands"); 5747 5748 // Does this instruction return a value ? 5749 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 5750 5751 Value *UndefVec = IsVoidRetTy ? nullptr : 5752 UndefValue::get(Instr->getType()); 5753 // Create a new entry in the WidenMap and initialize it to Undef or Null. 5754 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 5755 5756 VectorParts Cond; 5757 if (IfPredicateStore) { 5758 assert(Instr->getParent()->getSinglePredecessor() && 5759 "Only support single predecessor blocks"); 5760 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(), 5761 Instr->getParent()); 5762 } 5763 5764 // For each vector unroll 'part': 5765 for (unsigned Part = 0; Part < UF; ++Part) { 5766 // For each scalar that we create: 5767 5768 // Start an "if (pred) a[i] = ..." block. 5769 Value *Cmp = nullptr; 5770 if (IfPredicateStore) { 5771 if (Cond[Part]->getType()->isVectorTy()) 5772 Cond[Part] = 5773 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0)); 5774 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part], 5775 ConstantInt::get(Cond[Part]->getType(), 1)); 5776 } 5777 5778 Instruction *Cloned = Instr->clone(); 5779 if (!IsVoidRetTy) 5780 Cloned->setName(Instr->getName() + ".cloned"); 5781 // Replace the operands of the cloned instructions with extracted scalars. 5782 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5783 Value *Op = Params[op][Part]; 5784 Cloned->setOperand(op, Op); 5785 } 5786 5787 // Place the cloned scalar in the new loop. 5788 Builder.Insert(Cloned); 5789 5790 // If the original scalar returns a value we need to place it in a vector 5791 // so that future users will be able to use it. 5792 if (!IsVoidRetTy) 5793 VecResults[Part] = Cloned; 5794 5795 // End if-block. 5796 if (IfPredicateStore) 5797 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned), 5798 Cmp)); 5799 } 5800 } 5801 5802 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) { 5803 StoreInst *SI = dyn_cast<StoreInst>(Instr); 5804 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent())); 5805 5806 return scalarizeInstruction(Instr, IfPredicateStore); 5807 } 5808 5809 Value *InnerLoopUnroller::reverseVector(Value *Vec) { 5810 return Vec; 5811 } 5812 5813 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { 5814 return V; 5815 } 5816 5817 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) { 5818 // When unrolling and the VF is 1, we only need to add a simple scalar. 5819 Type *ITy = Val->getType(); 5820 assert(!ITy->isVectorTy() && "Val must be a scalar"); 5821 Constant *C = ConstantInt::get(ITy, StartIdx); 5822 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction"); 5823 } 5824