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// Other ideas/concepts are from: 38// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. 39// 40// S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of 41// Vectorizing Compilers. 42// 43//===----------------------------------------------------------------------===// 44 45#include "llvm/Transforms/Vectorize.h" 46#include "llvm/ADT/DenseMap.h" 47#include "llvm/ADT/EquivalenceClasses.h" 48#include "llvm/ADT/Hashing.h" 49#include "llvm/ADT/MapVector.h" 50#include "llvm/ADT/SetVector.h" 51#include "llvm/ADT/SmallPtrSet.h" 52#include "llvm/ADT/SmallSet.h" 53#include "llvm/ADT/SmallVector.h" 54#include "llvm/ADT/Statistic.h" 55#include "llvm/ADT/StringExtras.h" 56#include "llvm/Analysis/AliasAnalysis.h" 57#include "llvm/Analysis/AliasSetTracker.h" 58#include "llvm/Analysis/AssumptionCache.h" 59#include "llvm/Analysis/BlockFrequencyInfo.h" 60#include "llvm/Analysis/CodeMetrics.h" 61#include "llvm/Analysis/LoopAccessAnalysis.h" 62#include "llvm/Analysis/LoopInfo.h" 63#include "llvm/Analysis/LoopIterator.h" 64#include "llvm/Analysis/LoopPass.h" 65#include "llvm/Analysis/ScalarEvolution.h" 66#include "llvm/Analysis/ScalarEvolutionExpander.h" 67#include "llvm/Analysis/ScalarEvolutionExpressions.h" 68#include "llvm/Analysis/TargetTransformInfo.h" 69#include "llvm/Analysis/ValueTracking.h" 70#include "llvm/IR/Constants.h" 71#include "llvm/IR/DataLayout.h" 72#include "llvm/IR/DebugInfo.h" 73#include "llvm/IR/DerivedTypes.h" 74#include "llvm/IR/DiagnosticInfo.h" 75#include "llvm/IR/Dominators.h" 76#include "llvm/IR/Function.h" 77#include "llvm/IR/IRBuilder.h" 78#include "llvm/IR/Instructions.h" 79#include "llvm/IR/IntrinsicInst.h" 80#include "llvm/IR/LLVMContext.h" 81#include "llvm/IR/Module.h" 82#include "llvm/IR/PatternMatch.h" 83#include "llvm/IR/Type.h" 84#include "llvm/IR/Value.h" 85#include "llvm/IR/ValueHandle.h" 86#include "llvm/IR/Verifier.h" 87#include "llvm/Pass.h" 88#include "llvm/Support/BranchProbability.h" 89#include "llvm/Support/CommandLine.h" 90#include "llvm/Support/Debug.h" 91#include "llvm/Support/raw_ostream.h" 92#include "llvm/Transforms/Scalar.h" 93#include "llvm/Transforms/Utils/BasicBlockUtils.h" 94#include "llvm/Transforms/Utils/Local.h" 95#include "llvm/Transforms/Utils/VectorUtils.h" 96#include "llvm/Transforms/Utils/LoopUtils.h" 97#include <algorithm> 98#include <map> 99#include <tuple> 100 101using namespace llvm; 102using namespace llvm::PatternMatch; 103 104#define LV_NAME "loop-vectorize" 105#define DEBUG_TYPE LV_NAME 106 107STATISTIC(LoopsVectorized, "Number of loops vectorized"); 108STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization"); 109 110static cl::opt<bool> 111EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, 112 cl::desc("Enable if-conversion during vectorization.")); 113 114/// We don't vectorize loops with a known constant trip count below this number. 115static cl::opt<unsigned> 116TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16), 117 cl::Hidden, 118 cl::desc("Don't vectorize loops with a constant " 119 "trip count that is smaller than this " 120 "value.")); 121 122/// This enables versioning on the strides of symbolically striding memory 123/// accesses in code like the following. 124/// for (i = 0; i < N; ++i) 125/// A[i * Stride1] += B[i * Stride2] ... 126/// 127/// Will be roughly translated to 128/// if (Stride1 == 1 && Stride2 == 1) { 129/// for (i = 0; i < N; i+=4) 130/// A[i:i+3] += ... 131/// } else 132/// ... 133static cl::opt<bool> EnableMemAccessVersioning( 134 "enable-mem-access-versioning", cl::init(true), cl::Hidden, 135 cl::desc("Enable symblic stride memory access versioning")); 136 137/// We don't unroll loops with a known constant trip count below this number. 138static const unsigned TinyTripCountUnrollThreshold = 128; 139 140static cl::opt<unsigned> ForceTargetNumScalarRegs( 141 "force-target-num-scalar-regs", cl::init(0), cl::Hidden, 142 cl::desc("A flag that overrides the target's number of scalar registers.")); 143 144static cl::opt<unsigned> ForceTargetNumVectorRegs( 145 "force-target-num-vector-regs", cl::init(0), cl::Hidden, 146 cl::desc("A flag that overrides the target's number of vector registers.")); 147 148/// Maximum vectorization interleave count. 149static const unsigned MaxInterleaveFactor = 16; 150 151static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( 152 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, 153 cl::desc("A flag that overrides the target's max interleave factor for " 154 "scalar loops.")); 155 156static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( 157 "force-target-max-vector-interleave", cl::init(0), cl::Hidden, 158 cl::desc("A flag that overrides the target's max interleave factor for " 159 "vectorized loops.")); 160 161static cl::opt<unsigned> ForceTargetInstructionCost( 162 "force-target-instruction-cost", cl::init(0), cl::Hidden, 163 cl::desc("A flag that overrides the target's expected cost for " 164 "an instruction to a single constant value. Mostly " 165 "useful for getting consistent testing.")); 166 167static cl::opt<unsigned> SmallLoopCost( 168 "small-loop-cost", cl::init(20), cl::Hidden, 169 cl::desc("The cost of a loop that is considered 'small' by the unroller.")); 170 171static cl::opt<bool> LoopVectorizeWithBlockFrequency( 172 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden, 173 cl::desc("Enable the use of the block frequency analysis to access PGO " 174 "heuristics minimizing code growth in cold regions and being more " 175 "aggressive in hot regions.")); 176 177// Runtime unroll loops for load/store throughput. 178static cl::opt<bool> EnableLoadStoreRuntimeUnroll( 179 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden, 180 cl::desc("Enable runtime unrolling until load/store ports are saturated")); 181 182/// The number of stores in a loop that are allowed to need predication. 183static cl::opt<unsigned> NumberOfStoresToPredicate( 184 "vectorize-num-stores-pred", cl::init(1), cl::Hidden, 185 cl::desc("Max number of stores to be predicated behind an if.")); 186 187static cl::opt<bool> EnableIndVarRegisterHeur( 188 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, 189 cl::desc("Count the induction variable only once when unrolling")); 190 191static cl::opt<bool> EnableCondStoresVectorization( 192 "enable-cond-stores-vec", cl::init(false), cl::Hidden, 193 cl::desc("Enable if predication of stores during vectorization.")); 194 195static cl::opt<unsigned> MaxNestedScalarReductionUF( 196 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden, 197 cl::desc("The maximum unroll factor to use when unrolling a scalar " 198 "reduction in a nested loop.")); 199 200namespace { 201 202// Forward declarations. 203class LoopVectorizationLegality; 204class LoopVectorizationCostModel; 205class LoopVectorizeHints; 206 207/// \brief This modifies LoopAccessReport to initialize message with 208/// loop-vectorizer-specific part. 209class VectorizationReport : public LoopAccessReport { 210public: 211 VectorizationReport(Instruction *I = nullptr) 212 : LoopAccessReport("loop not vectorized: ", I) {} 213 214 /// \brief This allows promotion of the loop-access analysis report into the 215 /// loop-vectorizer report. It modifies the message to add the 216 /// loop-vectorizer-specific part of the message. 217 explicit VectorizationReport(const LoopAccessReport &R) 218 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(), 219 R.getInstr()) {} 220}; 221 222/// A helper function for converting Scalar types to vector types. 223/// If the incoming type is void, we return void. If the VF is 1, we return 224/// the scalar type. 225static Type* ToVectorTy(Type *Scalar, unsigned VF) { 226 if (Scalar->isVoidTy() || VF == 1) 227 return Scalar; 228 return VectorType::get(Scalar, VF); 229} 230 231/// InnerLoopVectorizer vectorizes loops which contain only one basic 232/// block to a specified vectorization factor (VF). 233/// This class performs the widening of scalars into vectors, or multiple 234/// scalars. This class also implements the following features: 235/// * It inserts an epilogue loop for handling loops that don't have iteration 236/// counts that are known to be a multiple of the vectorization factor. 237/// * It handles the code generation for reduction variables. 238/// * Scalarization (implementation using scalars) of un-vectorizable 239/// instructions. 240/// InnerLoopVectorizer does not perform any vectorization-legality 241/// checks, and relies on the caller to check for the different legality 242/// aspects. The InnerLoopVectorizer relies on the 243/// LoopVectorizationLegality class to provide information about the induction 244/// and reduction variables that were found to a given vectorization factor. 245class InnerLoopVectorizer { 246public: 247 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, 248 DominatorTree *DT, const TargetLibraryInfo *TLI, 249 const TargetTransformInfo *TTI, unsigned VecWidth, 250 unsigned UnrollFactor) 251 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), 252 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), 253 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor), 254 Legal(nullptr), AddedSafetyChecks(false) {} 255 256 // Perform the actual loop widening (vectorization). 257 void vectorize(LoopVectorizationLegality *L) { 258 Legal = L; 259 // Create a new empty loop. Unlink the old loop and connect the new one. 260 createEmptyLoop(); 261 // Widen each instruction in the old loop to a new one in the new loop. 262 // Use the Legality module to find the induction and reduction variables. 263 vectorizeLoop(); 264 // Register the new loop and update the analysis passes. 265 updateAnalysis(); 266 } 267 268 // Return true if any runtime check is added. 269 bool IsSafetyChecksAdded() { 270 return AddedSafetyChecks; 271 } 272 273 virtual ~InnerLoopVectorizer() {} 274 275protected: 276 /// A small list of PHINodes. 277 typedef SmallVector<PHINode*, 4> PhiVector; 278 /// When we unroll loops we have multiple vector values for each scalar. 279 /// This data structure holds the unrolled and vectorized values that 280 /// originated from one scalar instruction. 281 typedef SmallVector<Value*, 2> VectorParts; 282 283 // When we if-convert we need create edge masks. We have to cache values so 284 // that we don't end up with exponential recursion/IR. 285 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>, 286 VectorParts> EdgeMaskCache; 287 288 /// \brief Add checks for strides that where assumed to be 1. 289 /// 290 /// Returns the last check instruction and the first check instruction in the 291 /// pair as (first, last). 292 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc); 293 294 /// Create an empty loop, based on the loop ranges of the old loop. 295 void createEmptyLoop(); 296 /// Copy and widen the instructions from the old loop. 297 virtual void vectorizeLoop(); 298 299 /// \brief The Loop exit block may have single value PHI nodes where the 300 /// incoming value is 'Undef'. While vectorizing we only handled real values 301 /// that were defined inside the loop. Here we fix the 'undef case'. 302 /// See PR14725. 303 void fixLCSSAPHIs(); 304 305 /// A helper function that computes the predicate of the block BB, assuming 306 /// that the header block of the loop is set to True. It returns the *entry* 307 /// mask for the block BB. 308 VectorParts createBlockInMask(BasicBlock *BB); 309 /// A helper function that computes the predicate of the edge between SRC 310 /// and DST. 311 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst); 312 313 /// A helper function to vectorize a single BB within the innermost loop. 314 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV); 315 316 /// Vectorize a single PHINode in a block. This method handles the induction 317 /// variable canonicalization. It supports both VF = 1 for unrolled loops and 318 /// arbitrary length vectors. 319 void widenPHIInstruction(Instruction *PN, VectorParts &Entry, 320 unsigned UF, unsigned VF, PhiVector *PV); 321 322 /// Insert the new loop to the loop hierarchy and pass manager 323 /// and update the analysis passes. 324 void updateAnalysis(); 325 326 /// This instruction is un-vectorizable. Implement it as a sequence 327 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each 328 /// scalarized instruction behind an if block predicated on the control 329 /// dependence of the instruction. 330 virtual void scalarizeInstruction(Instruction *Instr, 331 bool IfPredicateStore=false); 332 333 /// Vectorize Load and Store instructions, 334 virtual void vectorizeMemoryInstruction(Instruction *Instr); 335 336 /// Create a broadcast instruction. This method generates a broadcast 337 /// instruction (shuffle) for loop invariant values and for the induction 338 /// value. If this is the induction variable then we extend it to N, N+1, ... 339 /// this is needed because each iteration in the loop corresponds to a SIMD 340 /// element. 341 virtual Value *getBroadcastInstrs(Value *V); 342 343 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...) 344 /// to each vector element of Val. The sequence starts at StartIndex. 345 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step); 346 347 /// When we go over instructions in the basic block we rely on previous 348 /// values within the current basic block or on loop invariant values. 349 /// When we widen (vectorize) values we place them in the map. If the values 350 /// are not within the map, they have to be loop invariant, so we simply 351 /// broadcast them into a vector. 352 VectorParts &getVectorValue(Value *V); 353 354 /// Generate a shuffle sequence that will reverse the vector Vec. 355 virtual Value *reverseVector(Value *Vec); 356 357 /// This is a helper class that holds the vectorizer state. It maps scalar 358 /// instructions to vector instructions. When the code is 'unrolled' then 359 /// then a single scalar value is mapped to multiple vector parts. The parts 360 /// are stored in the VectorPart type. 361 struct ValueMap { 362 /// C'tor. UnrollFactor controls the number of vectors ('parts') that 363 /// are mapped. 364 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {} 365 366 /// \return True if 'Key' is saved in the Value Map. 367 bool has(Value *Key) const { return MapStorage.count(Key); } 368 369 /// Initializes a new entry in the map. Sets all of the vector parts to the 370 /// save value in 'Val'. 371 /// \return A reference to a vector with splat values. 372 VectorParts &splat(Value *Key, Value *Val) { 373 VectorParts &Entry = MapStorage[Key]; 374 Entry.assign(UF, Val); 375 return Entry; 376 } 377 378 ///\return A reference to the value that is stored at 'Key'. 379 VectorParts &get(Value *Key) { 380 VectorParts &Entry = MapStorage[Key]; 381 if (Entry.empty()) 382 Entry.resize(UF); 383 assert(Entry.size() == UF); 384 return Entry; 385 } 386 387 private: 388 /// The unroll factor. Each entry in the map stores this number of vector 389 /// elements. 390 unsigned UF; 391 392 /// Map storage. We use std::map and not DenseMap because insertions to a 393 /// dense map invalidates its iterators. 394 std::map<Value *, VectorParts> MapStorage; 395 }; 396 397 /// The original loop. 398 Loop *OrigLoop; 399 /// Scev analysis to use. 400 ScalarEvolution *SE; 401 /// Loop Info. 402 LoopInfo *LI; 403 /// Dominator Tree. 404 DominatorTree *DT; 405 /// Alias Analysis. 406 AliasAnalysis *AA; 407 /// Target Library Info. 408 const TargetLibraryInfo *TLI; 409 /// Target Transform Info. 410 const TargetTransformInfo *TTI; 411 412 /// The vectorization SIMD factor to use. Each vector will have this many 413 /// vector elements. 414 unsigned VF; 415 416protected: 417 /// The vectorization unroll factor to use. Each scalar is vectorized to this 418 /// many different vector instructions. 419 unsigned UF; 420 421 /// The builder that we use 422 IRBuilder<> Builder; 423 424 // --- Vectorization state --- 425 426 /// The vector-loop preheader. 427 BasicBlock *LoopVectorPreHeader; 428 /// The scalar-loop preheader. 429 BasicBlock *LoopScalarPreHeader; 430 /// Middle Block between the vector and the scalar. 431 BasicBlock *LoopMiddleBlock; 432 ///The ExitBlock of the scalar loop. 433 BasicBlock *LoopExitBlock; 434 ///The vector loop body. 435 SmallVector<BasicBlock *, 4> LoopVectorBody; 436 ///The scalar loop body. 437 BasicBlock *LoopScalarBody; 438 /// A list of all bypass blocks. The first block is the entry of the loop. 439 SmallVector<BasicBlock *, 4> LoopBypassBlocks; 440 441 /// The new Induction variable which was added to the new block. 442 PHINode *Induction; 443 /// The induction variable of the old basic block. 444 PHINode *OldInduction; 445 /// Holds the extended (to the widest induction type) start index. 446 Value *ExtendedIdx; 447 /// Maps scalars to widened vectors. 448 ValueMap WidenMap; 449 EdgeMaskCache MaskCache; 450 451 LoopVectorizationLegality *Legal; 452 453 // Record whether runtime check is added. 454 bool AddedSafetyChecks; 455}; 456 457class InnerLoopUnroller : public InnerLoopVectorizer { 458public: 459 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, 460 DominatorTree *DT, const TargetLibraryInfo *TLI, 461 const TargetTransformInfo *TTI, unsigned UnrollFactor) 462 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {} 463 464private: 465 void scalarizeInstruction(Instruction *Instr, 466 bool IfPredicateStore = false) override; 467 void vectorizeMemoryInstruction(Instruction *Instr) override; 468 Value *getBroadcastInstrs(Value *V) override; 469 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override; 470 Value *reverseVector(Value *Vec) override; 471}; 472 473/// \brief Look for a meaningful debug location on the instruction or it's 474/// operands. 475static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { 476 if (!I) 477 return I; 478 479 DebugLoc Empty; 480 if (I->getDebugLoc() != Empty) 481 return I; 482 483 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) { 484 if (Instruction *OpInst = dyn_cast<Instruction>(*OI)) 485 if (OpInst->getDebugLoc() != Empty) 486 return OpInst; 487 } 488 489 return I; 490} 491 492/// \brief Set the debug location in the builder using the debug location in the 493/// instruction. 494static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) { 495 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) 496 B.SetCurrentDebugLocation(Inst->getDebugLoc()); 497 else 498 B.SetCurrentDebugLocation(DebugLoc()); 499} 500 501#ifndef NDEBUG 502/// \return string containing a file name and a line # for the given loop. 503static std::string getDebugLocString(const Loop *L) { 504 std::string Result; 505 if (L) { 506 raw_string_ostream OS(Result); 507 if (const DebugLoc LoopDbgLoc = L->getStartLoc()) 508 LoopDbgLoc.print(OS); 509 else 510 // Just print the module name. 511 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); 512 OS.flush(); 513 } 514 return Result; 515} 516#endif 517 518/// \brief Propagate known metadata from one instruction to another. 519static void propagateMetadata(Instruction *To, const Instruction *From) { 520 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata; 521 From->getAllMetadataOtherThanDebugLoc(Metadata); 522 523 for (auto M : Metadata) { 524 unsigned Kind = M.first; 525 526 // These are safe to transfer (this is safe for TBAA, even when we 527 // if-convert, because should that metadata have had a control dependency 528 // on the condition, and thus actually aliased with some other 529 // non-speculated memory access when the condition was false, this would be 530 // caught by the runtime overlap checks). 531 if (Kind != LLVMContext::MD_tbaa && 532 Kind != LLVMContext::MD_alias_scope && 533 Kind != LLVMContext::MD_noalias && 534 Kind != LLVMContext::MD_fpmath) 535 continue; 536 537 To->setMetadata(Kind, M.second); 538 } 539} 540 541/// \brief Propagate known metadata from one instruction to a vector of others. 542static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) { 543 for (Value *V : To) 544 if (Instruction *I = dyn_cast<Instruction>(V)) 545 propagateMetadata(I, From); 546} 547 548/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and 549/// to what vectorization factor. 550/// This class does not look at the profitability of vectorization, only the 551/// legality. This class has two main kinds of checks: 552/// * Memory checks - The code in canVectorizeMemory checks if vectorization 553/// will change the order of memory accesses in a way that will change the 554/// correctness of the program. 555/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory 556/// checks for a number of different conditions, such as the availability of a 557/// single induction variable, that all types are supported and vectorize-able, 558/// etc. This code reflects the capabilities of InnerLoopVectorizer. 559/// This class is also used by InnerLoopVectorizer for identifying 560/// induction variable and the different reduction variables. 561class LoopVectorizationLegality { 562public: 563 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT, 564 TargetLibraryInfo *TLI, AliasAnalysis *AA, 565 Function *F, const TargetTransformInfo *TTI, 566 LoopAccessAnalysis *LAA) 567 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F), 568 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), Induction(nullptr), 569 WidestIndTy(nullptr), HasFunNoNaNAttr(false) {} 570 571 /// This enum represents the kinds of reductions that we support. 572 enum ReductionKind { 573 RK_NoReduction, ///< Not a reduction. 574 RK_IntegerAdd, ///< Sum of integers. 575 RK_IntegerMult, ///< Product of integers. 576 RK_IntegerOr, ///< Bitwise or logical OR of numbers. 577 RK_IntegerAnd, ///< Bitwise or logical AND of numbers. 578 RK_IntegerXor, ///< Bitwise or logical XOR of numbers. 579 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()). 580 RK_FloatAdd, ///< Sum of floats. 581 RK_FloatMult, ///< Product of floats. 582 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()). 583 }; 584 585 /// This enum represents the kinds of inductions that we support. 586 enum InductionKind { 587 IK_NoInduction, ///< Not an induction variable. 588 IK_IntInduction, ///< Integer induction variable. Step = C. 589 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem). 590 }; 591 592 // This enum represents the kind of minmax reduction. 593 enum MinMaxReductionKind { 594 MRK_Invalid, 595 MRK_UIntMin, 596 MRK_UIntMax, 597 MRK_SIntMin, 598 MRK_SIntMax, 599 MRK_FloatMin, 600 MRK_FloatMax 601 }; 602 603 /// This struct holds information about reduction variables. 604 struct ReductionDescriptor { 605 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr), 606 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {} 607 608 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K, 609 MinMaxReductionKind MK) 610 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {} 611 612 // The starting value of the reduction. 613 // It does not have to be zero! 614 TrackingVH<Value> StartValue; 615 // The instruction who's value is used outside the loop. 616 Instruction *LoopExitInstr; 617 // The kind of the reduction. 618 ReductionKind Kind; 619 // If this a min/max reduction the kind of reduction. 620 MinMaxReductionKind MinMaxKind; 621 }; 622 623 /// This POD struct holds information about a potential reduction operation. 624 struct ReductionInstDesc { 625 ReductionInstDesc(bool IsRedux, Instruction *I) : 626 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {} 627 628 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) : 629 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {} 630 631 // Is this instruction a reduction candidate. 632 bool IsReduction; 633 // The last instruction in a min/max pattern (select of the select(icmp()) 634 // pattern), or the current reduction instruction otherwise. 635 Instruction *PatternLastInst; 636 // If this is a min/max pattern the comparison predicate. 637 MinMaxReductionKind MinMaxKind; 638 }; 639 640 /// A struct for saving information about induction variables. 641 struct InductionInfo { 642 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step) 643 : StartValue(Start), IK(K), StepValue(Step) { 644 assert(IK != IK_NoInduction && "Not an induction"); 645 assert(StartValue && "StartValue is null"); 646 assert(StepValue && !StepValue->isZero() && "StepValue is zero"); 647 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) && 648 "StartValue is not a pointer for pointer induction"); 649 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) && 650 "StartValue is not an integer for integer induction"); 651 assert(StepValue->getType()->isIntegerTy() && 652 "StepValue is not an integer"); 653 } 654 InductionInfo() 655 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {} 656 657 /// Get the consecutive direction. Returns: 658 /// 0 - unknown or non-consecutive. 659 /// 1 - consecutive and increasing. 660 /// -1 - consecutive and decreasing. 661 int getConsecutiveDirection() const { 662 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne())) 663 return StepValue->getSExtValue(); 664 return 0; 665 } 666 667 /// Compute the transformed value of Index at offset StartValue using step 668 /// StepValue. 669 /// For integer induction, returns StartValue + Index * StepValue. 670 /// For pointer induction, returns StartValue[Index * StepValue]. 671 /// FIXME: The newly created binary instructions should contain nsw/nuw 672 /// flags, which can be found from the original scalar operations. 673 Value *transform(IRBuilder<> &B, Value *Index) const { 674 switch (IK) { 675 case IK_IntInduction: 676 assert(Index->getType() == StartValue->getType() && 677 "Index type does not match StartValue type"); 678 if (StepValue->isMinusOne()) 679 return B.CreateSub(StartValue, Index); 680 if (!StepValue->isOne()) 681 Index = B.CreateMul(Index, StepValue); 682 return B.CreateAdd(StartValue, Index); 683 684 case IK_PtrInduction: 685 if (StepValue->isMinusOne()) 686 Index = B.CreateNeg(Index); 687 else if (!StepValue->isOne()) 688 Index = B.CreateMul(Index, StepValue); 689 return B.CreateGEP(nullptr, StartValue, Index); 690 691 case IK_NoInduction: 692 return nullptr; 693 } 694 llvm_unreachable("invalid enum"); 695 } 696 697 /// Start value. 698 TrackingVH<Value> StartValue; 699 /// Induction kind. 700 InductionKind IK; 701 /// Step value. 702 ConstantInt *StepValue; 703 }; 704 705 /// ReductionList contains the reduction descriptors for all 706 /// of the reductions that were found in the loop. 707 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList; 708 709 /// InductionList saves induction variables and maps them to the 710 /// induction descriptor. 711 typedef MapVector<PHINode*, InductionInfo> InductionList; 712 713 /// Returns true if it is legal to vectorize this loop. 714 /// This does not mean that it is profitable to vectorize this 715 /// loop, only that it is legal to do so. 716 bool canVectorize(); 717 718 /// Returns the Induction variable. 719 PHINode *getInduction() { return Induction; } 720 721 /// Returns the reduction variables found in the loop. 722 ReductionList *getReductionVars() { return &Reductions; } 723 724 /// Returns the induction variables found in the loop. 725 InductionList *getInductionVars() { return &Inductions; } 726 727 /// Returns the widest induction type. 728 Type *getWidestInductionType() { return WidestIndTy; } 729 730 /// Returns True if V is an induction variable in this loop. 731 bool isInductionVariable(const Value *V); 732 733 /// Return true if the block BB needs to be predicated in order for the loop 734 /// to be vectorized. 735 bool blockNeedsPredication(BasicBlock *BB); 736 737 /// Check if this pointer is consecutive when vectorizing. This happens 738 /// when the last index of the GEP is the induction variable, or that the 739 /// pointer itself is an induction variable. 740 /// This check allows us to vectorize A[idx] into a wide load/store. 741 /// Returns: 742 /// 0 - Stride is unknown or non-consecutive. 743 /// 1 - Address is consecutive. 744 /// -1 - Address is consecutive, and decreasing. 745 int isConsecutivePtr(Value *Ptr); 746 747 /// Returns true if the value V is uniform within the loop. 748 bool isUniform(Value *V); 749 750 /// Returns true if this instruction will remain scalar after vectorization. 751 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } 752 753 /// Returns the information that we collected about runtime memory check. 754 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const { 755 return LAI->getRuntimePointerCheck(); 756 } 757 758 const LoopAccessInfo *getLAI() const { 759 return LAI; 760 } 761 762 /// This function returns the identity element (or neutral element) for 763 /// the operation K. 764 static Constant *getReductionIdentity(ReductionKind K, Type *Tp); 765 766 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); } 767 768 bool hasStride(Value *V) { return StrideSet.count(V); } 769 bool mustCheckStrides() { return !StrideSet.empty(); } 770 SmallPtrSet<Value *, 8>::iterator strides_begin() { 771 return StrideSet.begin(); 772 } 773 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); } 774 775 /// Returns true if the target machine supports masked store operation 776 /// for the given \p DataType and kind of access to \p Ptr. 777 bool isLegalMaskedStore(Type *DataType, Value *Ptr) { 778 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr)); 779 } 780 /// Returns true if the target machine supports masked load operation 781 /// for the given \p DataType and kind of access to \p Ptr. 782 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) { 783 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr)); 784 } 785 /// Returns true if vector representation of the instruction \p I 786 /// requires mask. 787 bool isMaskRequired(const Instruction* I) { 788 return (MaskedOp.count(I) != 0); 789 } 790 unsigned getNumStores() const { 791 return LAI->getNumStores(); 792 } 793 unsigned getNumLoads() const { 794 return LAI->getNumLoads(); 795 } 796 unsigned getNumPredStores() const { 797 return NumPredStores; 798 } 799private: 800 /// Check if a single basic block loop is vectorizable. 801 /// At this point we know that this is a loop with a constant trip count 802 /// and we only need to check individual instructions. 803 bool canVectorizeInstrs(); 804 805 /// When we vectorize loops we may change the order in which 806 /// we read and write from memory. This method checks if it is 807 /// legal to vectorize the code, considering only memory constrains. 808 /// Returns true if the loop is vectorizable 809 bool canVectorizeMemory(); 810 811 /// Return true if we can vectorize this loop using the IF-conversion 812 /// transformation. 813 bool canVectorizeWithIfConvert(); 814 815 /// Collect the variables that need to stay uniform after vectorization. 816 void collectLoopUniforms(); 817 818 /// Return true if all of the instructions in the block can be speculatively 819 /// executed. \p SafePtrs is a list of addresses that are known to be legal 820 /// and we know that we can read from them without segfault. 821 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs); 822 823 /// Returns True, if 'Phi' is the kind of reduction variable for type 824 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. 825 bool AddReductionVar(PHINode *Phi, ReductionKind Kind); 826 /// Returns a struct describing if the instruction 'I' can be a reduction 827 /// variable of type 'Kind'. If the reduction is a min/max pattern of 828 /// select(icmp()) this function advances the instruction pointer 'I' from the 829 /// compare instruction to the select instruction and stores this pointer in 830 /// 'PatternLastInst' member of the returned struct. 831 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind, 832 ReductionInstDesc &Desc); 833 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 834 /// pattern corresponding to a min(X, Y) or max(X, Y). 835 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I, 836 ReductionInstDesc &Prev); 837 /// Returns the induction kind of Phi and record the step. This function may 838 /// return NoInduction if the PHI is not an induction variable. 839 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue); 840 841 /// \brief Collect memory access with loop invariant strides. 842 /// 843 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop 844 /// invariant. 845 void collectStridedAccess(Value *LoadOrStoreInst); 846 847 /// Report an analysis message to assist the user in diagnosing loops that are 848 /// not vectorized. These are handled as LoopAccessReport rather than 849 /// VectorizationReport because the << operator of VectorizationReport returns 850 /// LoopAccessReport. 851 void emitAnalysis(const LoopAccessReport &Message) { 852 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME); 853 } 854 855 unsigned NumPredStores; 856 857 /// The loop that we evaluate. 858 Loop *TheLoop; 859 /// Scev analysis. 860 ScalarEvolution *SE; 861 /// Target Library Info. 862 TargetLibraryInfo *TLI; 863 /// Parent function 864 Function *TheFunction; 865 /// Target Transform Info 866 const TargetTransformInfo *TTI; 867 /// Dominator Tree. 868 DominatorTree *DT; 869 // LoopAccess analysis. 870 LoopAccessAnalysis *LAA; 871 // And the loop-accesses info corresponding to this loop. This pointer is 872 // null until canVectorizeMemory sets it up. 873 const LoopAccessInfo *LAI; 874 875 // --- vectorization state --- // 876 877 /// Holds the integer induction variable. This is the counter of the 878 /// loop. 879 PHINode *Induction; 880 /// Holds the reduction variables. 881 ReductionList Reductions; 882 /// Holds all of the induction variables that we found in the loop. 883 /// Notice that inductions don't need to start at zero and that induction 884 /// variables can be pointers. 885 InductionList Inductions; 886 /// Holds the widest induction type encountered. 887 Type *WidestIndTy; 888 889 /// Allowed outside users. This holds the reduction 890 /// vars which can be accessed from outside the loop. 891 SmallPtrSet<Value*, 4> AllowedExit; 892 /// This set holds the variables which are known to be uniform after 893 /// vectorization. 894 SmallPtrSet<Instruction*, 4> Uniforms; 895 896 /// Can we assume the absence of NaNs. 897 bool HasFunNoNaNAttr; 898 899 ValueToValueMap Strides; 900 SmallPtrSet<Value *, 8> StrideSet; 901 902 /// While vectorizing these instructions we have to generate a 903 /// call to the appropriate masked intrinsic 904 SmallPtrSet<const Instruction*, 8> MaskedOp; 905}; 906 907/// LoopVectorizationCostModel - estimates the expected speedups due to 908/// vectorization. 909/// In many cases vectorization is not profitable. This can happen because of 910/// a number of reasons. In this class we mainly attempt to predict the 911/// expected speedup/slowdowns due to the supported instruction set. We use the 912/// TargetTransformInfo to query the different backends for the cost of 913/// different operations. 914class LoopVectorizationCostModel { 915public: 916 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI, 917 LoopVectorizationLegality *Legal, 918 const TargetTransformInfo &TTI, 919 const TargetLibraryInfo *TLI, AssumptionCache *AC, 920 const Function *F, const LoopVectorizeHints *Hints) 921 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), 922 TheFunction(F), Hints(Hints) { 923 CodeMetrics::collectEphemeralValues(L, AC, EphValues); 924 } 925 926 /// Information about vectorization costs 927 struct VectorizationFactor { 928 unsigned Width; // Vector width with best cost 929 unsigned Cost; // Cost of the loop with that width 930 }; 931 /// \return The most profitable vectorization factor and the cost of that VF. 932 /// This method checks every power of two up to VF. If UserVF is not ZERO 933 /// then this vectorization factor will be selected if vectorization is 934 /// possible. 935 VectorizationFactor selectVectorizationFactor(bool OptForSize); 936 937 /// \return The size (in bits) of the widest type in the code that 938 /// needs to be vectorized. We ignore values that remain scalar such as 939 /// 64 bit loop indices. 940 unsigned getWidestType(); 941 942 /// \return The most profitable unroll factor. 943 /// If UserUF is non-zero then this method finds the best unroll-factor 944 /// based on register pressure and other parameters. 945 /// VF and LoopCost are the selected vectorization factor and the cost of the 946 /// selected VF. 947 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost); 948 949 /// \brief A struct that represents some properties of the register usage 950 /// of a loop. 951 struct RegisterUsage { 952 /// Holds the number of loop invariant values that are used in the loop. 953 unsigned LoopInvariantRegs; 954 /// Holds the maximum number of concurrent live intervals in the loop. 955 unsigned MaxLocalUsers; 956 /// Holds the number of instructions in the loop. 957 unsigned NumInstructions; 958 }; 959 960 /// \return information about the register usage of the loop. 961 RegisterUsage calculateRegisterUsage(); 962 963private: 964 /// Returns the expected execution cost. The unit of the cost does 965 /// not matter because we use the 'cost' units to compare different 966 /// vector widths. The cost that is returned is *not* normalized by 967 /// the factor width. 968 unsigned expectedCost(unsigned VF); 969 970 /// Returns the execution time cost of an instruction for a given vector 971 /// width. Vector width of one means scalar. 972 unsigned getInstructionCost(Instruction *I, unsigned VF); 973 974 /// Returns whether the instruction is a load or store and will be a emitted 975 /// as a vector operation. 976 bool isConsecutiveLoadOrStore(Instruction *I); 977 978 /// Report an analysis message to assist the user in diagnosing loops that are 979 /// not vectorized. These are handled as LoopAccessReport rather than 980 /// VectorizationReport because the << operator of VectorizationReport returns 981 /// LoopAccessReport. 982 void emitAnalysis(const LoopAccessReport &Message) { 983 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME); 984 } 985 986 /// Values used only by @llvm.assume calls. 987 SmallPtrSet<const Value *, 32> EphValues; 988 989 /// The loop that we evaluate. 990 Loop *TheLoop; 991 /// Scev analysis. 992 ScalarEvolution *SE; 993 /// Loop Info analysis. 994 LoopInfo *LI; 995 /// Vectorization legality. 996 LoopVectorizationLegality *Legal; 997 /// Vector target information. 998 const TargetTransformInfo &TTI; 999 /// Target Library Info. 1000 const TargetLibraryInfo *TLI; 1001 const Function *TheFunction; 1002 // Loop Vectorize Hint. 1003 const LoopVectorizeHints *Hints; 1004}; 1005 1006/// Utility class for getting and setting loop vectorizer hints in the form 1007/// of loop metadata. 1008/// This class keeps a number of loop annotations locally (as member variables) 1009/// and can, upon request, write them back as metadata on the loop. It will 1010/// initially scan the loop for existing metadata, and will update the local 1011/// values based on information in the loop. 1012/// We cannot write all values to metadata, as the mere presence of some info, 1013/// for example 'force', means a decision has been made. So, we need to be 1014/// careful NOT to add them if the user hasn't specifically asked so. 1015class LoopVectorizeHints { 1016 enum HintKind { 1017 HK_WIDTH, 1018 HK_UNROLL, 1019 HK_FORCE 1020 }; 1021 1022 /// Hint - associates name and validation with the hint value. 1023 struct Hint { 1024 const char * Name; 1025 unsigned Value; // This may have to change for non-numeric values. 1026 HintKind Kind; 1027 1028 Hint(const char * Name, unsigned Value, HintKind Kind) 1029 : Name(Name), Value(Value), Kind(Kind) { } 1030 1031 bool validate(unsigned Val) { 1032 switch (Kind) { 1033 case HK_WIDTH: 1034 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth; 1035 case HK_UNROLL: 1036 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor; 1037 case HK_FORCE: 1038 return (Val <= 1); 1039 } 1040 return false; 1041 } 1042 }; 1043 1044 /// Vectorization width. 1045 Hint Width; 1046 /// Vectorization interleave factor. 1047 Hint Interleave; 1048 /// Vectorization forced 1049 Hint Force; 1050 1051 /// Return the loop metadata prefix. 1052 static StringRef Prefix() { return "llvm.loop."; } 1053 1054public: 1055 enum ForceKind { 1056 FK_Undefined = -1, ///< Not selected. 1057 FK_Disabled = 0, ///< Forcing disabled. 1058 FK_Enabled = 1, ///< Forcing enabled. 1059 }; 1060 1061 LoopVectorizeHints(const Loop *L, bool DisableInterleaving) 1062 : Width("vectorize.width", VectorizerParams::VectorizationFactor, 1063 HK_WIDTH), 1064 Interleave("interleave.count", DisableInterleaving, HK_UNROLL), 1065 Force("vectorize.enable", FK_Undefined, HK_FORCE), 1066 TheLoop(L) { 1067 // Populate values with existing loop metadata. 1068 getHintsFromMetadata(); 1069 1070 // force-vector-interleave overrides DisableInterleaving. 1071 if (VectorizerParams::isInterleaveForced()) 1072 Interleave.Value = VectorizerParams::VectorizationInterleave; 1073 1074 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs() 1075 << "LV: Interleaving disabled by the pass manager\n"); 1076 } 1077 1078 /// Mark the loop L as already vectorized by setting the width to 1. 1079 void setAlreadyVectorized() { 1080 Width.Value = Interleave.Value = 1; 1081 Hint Hints[] = {Width, Interleave}; 1082 writeHintsToMetadata(Hints); 1083 } 1084 1085 /// Dumps all the hint information. 1086 std::string emitRemark() const { 1087 VectorizationReport R; 1088 if (Force.Value == LoopVectorizeHints::FK_Disabled) 1089 R << "vectorization is explicitly disabled"; 1090 else { 1091 R << "use -Rpass-analysis=loop-vectorize for more info"; 1092 if (Force.Value == LoopVectorizeHints::FK_Enabled) { 1093 R << " (Force=true"; 1094 if (Width.Value != 0) 1095 R << ", Vector Width=" << Width.Value; 1096 if (Interleave.Value != 0) 1097 R << ", Interleave Count=" << Interleave.Value; 1098 R << ")"; 1099 } 1100 } 1101 1102 return R.str(); 1103 } 1104 1105 unsigned getWidth() const { return Width.Value; } 1106 unsigned getInterleave() const { return Interleave.Value; } 1107 enum ForceKind getForce() const { return (ForceKind)Force.Value; } 1108 1109private: 1110 /// Find hints specified in the loop metadata and update local values. 1111 void getHintsFromMetadata() { 1112 MDNode *LoopID = TheLoop->getLoopID(); 1113 if (!LoopID) 1114 return; 1115 1116 // First operand should refer to the loop id itself. 1117 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 1118 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 1119 1120 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 1121 const MDString *S = nullptr; 1122 SmallVector<Metadata *, 4> Args; 1123 1124 // The expected hint is either a MDString or a MDNode with the first 1125 // operand a MDString. 1126 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) { 1127 if (!MD || MD->getNumOperands() == 0) 1128 continue; 1129 S = dyn_cast<MDString>(MD->getOperand(0)); 1130 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i) 1131 Args.push_back(MD->getOperand(i)); 1132 } else { 1133 S = dyn_cast<MDString>(LoopID->getOperand(i)); 1134 assert(Args.size() == 0 && "too many arguments for MDString"); 1135 } 1136 1137 if (!S) 1138 continue; 1139 1140 // Check if the hint starts with the loop metadata prefix. 1141 StringRef Name = S->getString(); 1142 if (Args.size() == 1) 1143 setHint(Name, Args[0]); 1144 } 1145 } 1146 1147 /// Checks string hint with one operand and set value if valid. 1148 void setHint(StringRef Name, Metadata *Arg) { 1149 if (!Name.startswith(Prefix())) 1150 return; 1151 Name = Name.substr(Prefix().size(), StringRef::npos); 1152 1153 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg); 1154 if (!C) return; 1155 unsigned Val = C->getZExtValue(); 1156 1157 Hint *Hints[] = {&Width, &Interleave, &Force}; 1158 for (auto H : Hints) { 1159 if (Name == H->Name) { 1160 if (H->validate(Val)) 1161 H->Value = Val; 1162 else 1163 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n"); 1164 break; 1165 } 1166 } 1167 } 1168 1169 /// Create a new hint from name / value pair. 1170 MDNode *createHintMetadata(StringRef Name, unsigned V) const { 1171 LLVMContext &Context = TheLoop->getHeader()->getContext(); 1172 Metadata *MDs[] = {MDString::get(Context, Name), 1173 ConstantAsMetadata::get( 1174 ConstantInt::get(Type::getInt32Ty(Context), V))}; 1175 return MDNode::get(Context, MDs); 1176 } 1177 1178 /// Matches metadata with hint name. 1179 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) { 1180 MDString* Name = dyn_cast<MDString>(Node->getOperand(0)); 1181 if (!Name) 1182 return false; 1183 1184 for (auto H : HintTypes) 1185 if (Name->getString().endswith(H.Name)) 1186 return true; 1187 return false; 1188 } 1189 1190 /// Sets current hints into loop metadata, keeping other values intact. 1191 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) { 1192 if (HintTypes.size() == 0) 1193 return; 1194 1195 // Reserve the first element to LoopID (see below). 1196 SmallVector<Metadata *, 4> MDs(1); 1197 // If the loop already has metadata, then ignore the existing operands. 1198 MDNode *LoopID = TheLoop->getLoopID(); 1199 if (LoopID) { 1200 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 1201 MDNode *Node = cast<MDNode>(LoopID->getOperand(i)); 1202 // If node in update list, ignore old value. 1203 if (!matchesHintMetadataName(Node, HintTypes)) 1204 MDs.push_back(Node); 1205 } 1206 } 1207 1208 // Now, add the missing hints. 1209 for (auto H : HintTypes) 1210 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value)); 1211 1212 // Replace current metadata node with new one. 1213 LLVMContext &Context = TheLoop->getHeader()->getContext(); 1214 MDNode *NewLoopID = MDNode::get(Context, MDs); 1215 // Set operand 0 to refer to the loop id itself. 1216 NewLoopID->replaceOperandWith(0, NewLoopID); 1217 1218 TheLoop->setLoopID(NewLoopID); 1219 } 1220 1221 /// The loop these hints belong to. 1222 const Loop *TheLoop; 1223}; 1224 1225static void emitMissedWarning(Function *F, Loop *L, 1226 const LoopVectorizeHints &LH) { 1227 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F, 1228 L->getStartLoc(), LH.emitRemark()); 1229 1230 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) { 1231 if (LH.getWidth() != 1) 1232 emitLoopVectorizeWarning( 1233 F->getContext(), *F, L->getStartLoc(), 1234 "failed explicitly specified loop vectorization"); 1235 else if (LH.getInterleave() != 1) 1236 emitLoopInterleaveWarning( 1237 F->getContext(), *F, L->getStartLoc(), 1238 "failed explicitly specified loop interleaving"); 1239 } 1240} 1241 1242static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) { 1243 if (L.empty()) 1244 return V.push_back(&L); 1245 1246 for (Loop *InnerL : L) 1247 addInnerLoop(*InnerL, V); 1248} 1249 1250/// The LoopVectorize Pass. 1251struct LoopVectorize : public FunctionPass { 1252 /// Pass identification, replacement for typeid 1253 static char ID; 1254 1255 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true) 1256 : FunctionPass(ID), 1257 DisableUnrolling(NoUnrolling), 1258 AlwaysVectorize(AlwaysVectorize) { 1259 initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); 1260 } 1261 1262 ScalarEvolution *SE; 1263 LoopInfo *LI; 1264 TargetTransformInfo *TTI; 1265 DominatorTree *DT; 1266 BlockFrequencyInfo *BFI; 1267 TargetLibraryInfo *TLI; 1268 AliasAnalysis *AA; 1269 AssumptionCache *AC; 1270 LoopAccessAnalysis *LAA; 1271 bool DisableUnrolling; 1272 bool AlwaysVectorize; 1273 1274 BlockFrequency ColdEntryFreq; 1275 1276 bool runOnFunction(Function &F) override { 1277 SE = &getAnalysis<ScalarEvolution>(); 1278 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 1279 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 1280 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1281 BFI = &getAnalysis<BlockFrequencyInfo>(); 1282 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 1283 TLI = TLIP ? &TLIP->getTLI() : nullptr; 1284 AA = &getAnalysis<AliasAnalysis>(); 1285 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1286 LAA = &getAnalysis<LoopAccessAnalysis>(); 1287 1288 // Compute some weights outside of the loop over the loops. Compute this 1289 // using a BranchProbability to re-use its scaling math. 1290 const BranchProbability ColdProb(1, 5); // 20% 1291 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb; 1292 1293 // If the target claims to have no vector registers don't attempt 1294 // vectorization. 1295 if (!TTI->getNumberOfRegisters(true)) 1296 return false; 1297 1298 // Build up a worklist of inner-loops to vectorize. This is necessary as 1299 // the act of vectorizing or partially unrolling a loop creates new loops 1300 // and can invalidate iterators across the loops. 1301 SmallVector<Loop *, 8> Worklist; 1302 1303 for (Loop *L : *LI) 1304 addInnerLoop(*L, Worklist); 1305 1306 LoopsAnalyzed += Worklist.size(); 1307 1308 // Now walk the identified inner loops. 1309 bool Changed = false; 1310 while (!Worklist.empty()) 1311 Changed |= processLoop(Worklist.pop_back_val()); 1312 1313 // Process each loop nest in the function. 1314 return Changed; 1315 } 1316 1317 static void AddRuntimeUnrollDisableMetaData(Loop *L) { 1318 SmallVector<Metadata *, 4> MDs; 1319 // Reserve first location for self reference to the LoopID metadata node. 1320 MDs.push_back(nullptr); 1321 bool IsUnrollMetadata = false; 1322 MDNode *LoopID = L->getLoopID(); 1323 if (LoopID) { 1324 // First find existing loop unrolling disable metadata. 1325 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 1326 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 1327 if (MD) { 1328 const MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 1329 IsUnrollMetadata = 1330 S && S->getString().startswith("llvm.loop.unroll.disable"); 1331 } 1332 MDs.push_back(LoopID->getOperand(i)); 1333 } 1334 } 1335 1336 if (!IsUnrollMetadata) { 1337 // Add runtime unroll disable metadata. 1338 LLVMContext &Context = L->getHeader()->getContext(); 1339 SmallVector<Metadata *, 1> DisableOperands; 1340 DisableOperands.push_back( 1341 MDString::get(Context, "llvm.loop.unroll.runtime.disable")); 1342 MDNode *DisableNode = MDNode::get(Context, DisableOperands); 1343 MDs.push_back(DisableNode); 1344 MDNode *NewLoopID = MDNode::get(Context, MDs); 1345 // Set operand 0 to refer to the loop id itself. 1346 NewLoopID->replaceOperandWith(0, NewLoopID); 1347 L->setLoopID(NewLoopID); 1348 } 1349 } 1350 1351 bool processLoop(Loop *L) { 1352 assert(L->empty() && "Only process inner loops."); 1353 1354#ifndef NDEBUG 1355 const std::string DebugLocStr = getDebugLocString(L); 1356#endif /* NDEBUG */ 1357 1358 DEBUG(dbgs() << "\nLV: Checking a loop in \"" 1359 << L->getHeader()->getParent()->getName() << "\" from " 1360 << DebugLocStr << "\n"); 1361 1362 LoopVectorizeHints Hints(L, DisableUnrolling); 1363 1364 DEBUG(dbgs() << "LV: Loop hints:" 1365 << " force=" 1366 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled 1367 ? "disabled" 1368 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled 1369 ? "enabled" 1370 : "?")) << " width=" << Hints.getWidth() 1371 << " unroll=" << Hints.getInterleave() << "\n"); 1372 1373 // Function containing loop 1374 Function *F = L->getHeader()->getParent(); 1375 1376 // Looking at the diagnostic output is the only way to determine if a loop 1377 // was vectorized (other than looking at the IR or machine code), so it 1378 // is important to generate an optimization remark for each loop. Most of 1379 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks 1380 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are 1381 // less verbose reporting vectorized loops and unvectorized loops that may 1382 // benefit from vectorization, respectively. 1383 1384 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) { 1385 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n"); 1386 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F, 1387 L->getStartLoc(), Hints.emitRemark()); 1388 return false; 1389 } 1390 1391 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) { 1392 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n"); 1393 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F, 1394 L->getStartLoc(), Hints.emitRemark()); 1395 return false; 1396 } 1397 1398 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) { 1399 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n"); 1400 emitOptimizationRemarkAnalysis( 1401 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(), 1402 "loop not vectorized: vector width and interleave count are " 1403 "explicitly set to 1"); 1404 return false; 1405 } 1406 1407 // Check the loop for a trip count threshold: 1408 // do not vectorize loops with a tiny trip count. 1409 const unsigned TC = SE->getSmallConstantTripCount(L); 1410 if (TC > 0u && TC < TinyTripCountVectorThreshold) { 1411 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " 1412 << "This loop is not worth vectorizing."); 1413 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) 1414 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n"); 1415 else { 1416 DEBUG(dbgs() << "\n"); 1417 emitOptimizationRemarkAnalysis( 1418 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(), 1419 "vectorization is not beneficial and is not explicitly forced"); 1420 return false; 1421 } 1422 } 1423 1424 // Check if it is legal to vectorize the loop. 1425 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA); 1426 if (!LVL.canVectorize()) { 1427 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n"); 1428 emitMissedWarning(F, L, Hints); 1429 return false; 1430 } 1431 1432 // Use the cost model. 1433 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints); 1434 1435 // Check the function attributes to find out if this function should be 1436 // optimized for size. 1437 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled && 1438 F->hasFnAttribute(Attribute::OptimizeForSize); 1439 1440 // Compute the weighted frequency of this loop being executed and see if it 1441 // is less than 20% of the function entry baseline frequency. Note that we 1442 // always have a canonical loop here because we think we *can* vectoriez. 1443 // FIXME: This is hidden behind a flag due to pervasive problems with 1444 // exactly what block frequency models. 1445 if (LoopVectorizeWithBlockFrequency) { 1446 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader()); 1447 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled && 1448 LoopEntryFreq < ColdEntryFreq) 1449 OptForSize = true; 1450 } 1451 1452 // Check the function attributes to see if implicit floats are allowed.a 1453 // FIXME: This check doesn't seem possibly correct -- what if the loop is 1454 // an integer loop and the vector instructions selected are purely integer 1455 // vector instructions? 1456 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { 1457 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" 1458 "attribute is used.\n"); 1459 emitOptimizationRemarkAnalysis( 1460 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(), 1461 "loop not vectorized due to NoImplicitFloat attribute"); 1462 emitMissedWarning(F, L, Hints); 1463 return false; 1464 } 1465 1466 // Select the optimal vectorization factor. 1467 const LoopVectorizationCostModel::VectorizationFactor VF = 1468 CM.selectVectorizationFactor(OptForSize); 1469 1470 // Select the unroll factor. 1471 const unsigned UF = 1472 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost); 1473 1474 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " 1475 << DebugLocStr << '\n'); 1476 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n'); 1477 1478 if (VF.Width == 1) { 1479 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n"); 1480 1481 if (UF == 1) { 1482 emitOptimizationRemarkAnalysis( 1483 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(), 1484 "not beneficial to vectorize and user disabled interleaving"); 1485 return false; 1486 } 1487 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n"); 1488 1489 // Report the unrolling decision. 1490 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(), 1491 Twine("unrolled with interleaving factor " + 1492 Twine(UF) + 1493 " (vectorization not beneficial)")); 1494 1495 // We decided not to vectorize, but we may want to unroll. 1496 1497 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, UF); 1498 Unroller.vectorize(&LVL); 1499 } else { 1500 // If we decided that it is *legal* to vectorize the loop then do it. 1501 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, UF); 1502 LB.vectorize(&LVL); 1503 ++LoopsVectorized; 1504 1505 // Add metadata to disable runtime unrolling scalar loop when there's no 1506 // runtime check about strides and memory. Because at this situation, 1507 // scalar loop is rarely used not worthy to be unrolled. 1508 if (!LB.IsSafetyChecksAdded()) 1509 AddRuntimeUnrollDisableMetaData(L); 1510 1511 // Report the vectorization decision. 1512 emitOptimizationRemark( 1513 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(), 1514 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) + 1515 ", unrolling interleave factor: " + Twine(UF) + ")"); 1516 } 1517 1518 // Mark the loop as already vectorized to avoid vectorizing again. 1519 Hints.setAlreadyVectorized(); 1520 1521 DEBUG(verifyFunction(*L->getHeader()->getParent())); 1522 return true; 1523 } 1524 1525 void getAnalysisUsage(AnalysisUsage &AU) const override { 1526 AU.addRequired<AssumptionCacheTracker>(); 1527 AU.addRequiredID(LoopSimplifyID); 1528 AU.addRequiredID(LCSSAID); 1529 AU.addRequired<BlockFrequencyInfo>(); 1530 AU.addRequired<DominatorTreeWrapperPass>(); 1531 AU.addRequired<LoopInfoWrapperPass>(); 1532 AU.addRequired<ScalarEvolution>(); 1533 AU.addRequired<TargetTransformInfoWrapperPass>(); 1534 AU.addRequired<AliasAnalysis>(); 1535 AU.addRequired<LoopAccessAnalysis>(); 1536 AU.addPreserved<LoopInfoWrapperPass>(); 1537 AU.addPreserved<DominatorTreeWrapperPass>(); 1538 AU.addPreserved<AliasAnalysis>(); 1539 } 1540 1541}; 1542 1543} // end anonymous namespace 1544 1545//===----------------------------------------------------------------------===// 1546// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and 1547// LoopVectorizationCostModel. 1548//===----------------------------------------------------------------------===// 1549 1550Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { 1551 // We need to place the broadcast of invariant variables outside the loop. 1552 Instruction *Instr = dyn_cast<Instruction>(V); 1553 bool NewInstr = 1554 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(), 1555 Instr->getParent()) != LoopVectorBody.end()); 1556 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr; 1557 1558 // Place the code for broadcasting invariant variables in the new preheader. 1559 IRBuilder<>::InsertPointGuard Guard(Builder); 1560 if (Invariant) 1561 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); 1562 1563 // Broadcast the scalar into all locations in the vector. 1564 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); 1565 1566 return Shuf; 1567} 1568 1569Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, 1570 Value *Step) { 1571 assert(Val->getType()->isVectorTy() && "Must be a vector"); 1572 assert(Val->getType()->getScalarType()->isIntegerTy() && 1573 "Elem must be an integer"); 1574 assert(Step->getType() == Val->getType()->getScalarType() && 1575 "Step has wrong type"); 1576 // Create the types. 1577 Type *ITy = Val->getType()->getScalarType(); 1578 VectorType *Ty = cast<VectorType>(Val->getType()); 1579 int VLen = Ty->getNumElements(); 1580 SmallVector<Constant*, 8> Indices; 1581 1582 // Create a vector of consecutive numbers from zero to VF. 1583 for (int i = 0; i < VLen; ++i) 1584 Indices.push_back(ConstantInt::get(ITy, StartIdx + i)); 1585 1586 // Add the consecutive indices to the vector value. 1587 Constant *Cv = ConstantVector::get(Indices); 1588 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); 1589 Step = Builder.CreateVectorSplat(VLen, Step); 1590 assert(Step->getType() == Val->getType() && "Invalid step vec"); 1591 // FIXME: The newly created binary instructions should contain nsw/nuw flags, 1592 // which can be found from the original scalar operations. 1593 Step = Builder.CreateMul(Cv, Step); 1594 return Builder.CreateAdd(Val, Step, "induction"); 1595} 1596 1597/// \brief Find the operand of the GEP that should be checked for consecutive 1598/// stores. This ignores trailing indices that have no effect on the final 1599/// pointer. 1600static unsigned getGEPInductionOperand(const GetElementPtrInst *Gep) { 1601 const DataLayout &DL = Gep->getModule()->getDataLayout(); 1602 unsigned LastOperand = Gep->getNumOperands() - 1; 1603 unsigned GEPAllocSize = DL.getTypeAllocSize( 1604 cast<PointerType>(Gep->getType()->getScalarType())->getElementType()); 1605 1606 // Walk backwards and try to peel off zeros. 1607 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) { 1608 // Find the type we're currently indexing into. 1609 gep_type_iterator GEPTI = gep_type_begin(Gep); 1610 std::advance(GEPTI, LastOperand - 1); 1611 1612 // If it's a type with the same allocation size as the result of the GEP we 1613 // can peel off the zero index. 1614 if (DL.getTypeAllocSize(*GEPTI) != GEPAllocSize) 1615 break; 1616 --LastOperand; 1617 } 1618 1619 return LastOperand; 1620} 1621 1622int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { 1623 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr"); 1624 // Make sure that the pointer does not point to structs. 1625 if (Ptr->getType()->getPointerElementType()->isAggregateType()) 1626 return 0; 1627 1628 // If this value is a pointer induction variable we know it is consecutive. 1629 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr); 1630 if (Phi && Inductions.count(Phi)) { 1631 InductionInfo II = Inductions[Phi]; 1632 return II.getConsecutiveDirection(); 1633 } 1634 1635 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr); 1636 if (!Gep) 1637 return 0; 1638 1639 unsigned NumOperands = Gep->getNumOperands(); 1640 Value *GpPtr = Gep->getPointerOperand(); 1641 // If this GEP value is a consecutive pointer induction variable and all of 1642 // the indices are constant then we know it is consecutive. We can 1643 Phi = dyn_cast<PHINode>(GpPtr); 1644 if (Phi && Inductions.count(Phi)) { 1645 1646 // Make sure that the pointer does not point to structs. 1647 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType()); 1648 if (GepPtrType->getElementType()->isAggregateType()) 1649 return 0; 1650 1651 // Make sure that all of the index operands are loop invariant. 1652 for (unsigned i = 1; i < NumOperands; ++i) 1653 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 1654 return 0; 1655 1656 InductionInfo II = Inductions[Phi]; 1657 return II.getConsecutiveDirection(); 1658 } 1659 1660 unsigned InductionOperand = getGEPInductionOperand(Gep); 1661 1662 // Check that all of the gep indices are uniform except for our induction 1663 // operand. 1664 for (unsigned i = 0; i != NumOperands; ++i) 1665 if (i != InductionOperand && 1666 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 1667 return 0; 1668 1669 // We can emit wide load/stores only if the last non-zero index is the 1670 // induction variable. 1671 const SCEV *Last = nullptr; 1672 if (!Strides.count(Gep)) 1673 Last = SE->getSCEV(Gep->getOperand(InductionOperand)); 1674 else { 1675 // Because of the multiplication by a stride we can have a s/zext cast. 1676 // We are going to replace this stride by 1 so the cast is safe to ignore. 1677 // 1678 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 1679 // %0 = trunc i64 %indvars.iv to i32 1680 // %mul = mul i32 %0, %Stride1 1681 // %idxprom = zext i32 %mul to i64 << Safe cast. 1682 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom 1683 // 1684 Last = replaceSymbolicStrideSCEV(SE, Strides, 1685 Gep->getOperand(InductionOperand), Gep); 1686 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last)) 1687 Last = 1688 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend) 1689 ? C->getOperand() 1690 : Last; 1691 } 1692 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { 1693 const SCEV *Step = AR->getStepRecurrence(*SE); 1694 1695 // The memory is consecutive because the last index is consecutive 1696 // and all other indices are loop invariant. 1697 if (Step->isOne()) 1698 return 1; 1699 if (Step->isAllOnesValue()) 1700 return -1; 1701 } 1702 1703 return 0; 1704} 1705 1706bool LoopVectorizationLegality::isUniform(Value *V) { 1707 return LAI->isUniform(V); 1708} 1709 1710InnerLoopVectorizer::VectorParts& 1711InnerLoopVectorizer::getVectorValue(Value *V) { 1712 assert(V != Induction && "The new induction variable should not be used."); 1713 assert(!V->getType()->isVectorTy() && "Can't widen a vector"); 1714 1715 // If we have a stride that is replaced by one, do it here. 1716 if (Legal->hasStride(V)) 1717 V = ConstantInt::get(V->getType(), 1); 1718 1719 // If we have this scalar in the map, return it. 1720 if (WidenMap.has(V)) 1721 return WidenMap.get(V); 1722 1723 // If this scalar is unknown, assume that it is a constant or that it is 1724 // loop invariant. Broadcast V and save the value for future uses. 1725 Value *B = getBroadcastInstrs(V); 1726 return WidenMap.splat(V, B); 1727} 1728 1729Value *InnerLoopVectorizer::reverseVector(Value *Vec) { 1730 assert(Vec->getType()->isVectorTy() && "Invalid type"); 1731 SmallVector<Constant*, 8> ShuffleMask; 1732 for (unsigned i = 0; i < VF; ++i) 1733 ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); 1734 1735 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), 1736 ConstantVector::get(ShuffleMask), 1737 "reverse"); 1738} 1739 1740void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) { 1741 // Attempt to issue a wide load. 1742 LoadInst *LI = dyn_cast<LoadInst>(Instr); 1743 StoreInst *SI = dyn_cast<StoreInst>(Instr); 1744 1745 assert((LI || SI) && "Invalid Load/Store instruction"); 1746 1747 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType(); 1748 Type *DataTy = VectorType::get(ScalarDataTy, VF); 1749 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 1750 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment(); 1751 // An alignment of 0 means target abi alignment. We need to use the scalar's 1752 // target abi alignment in such a case. 1753 const DataLayout &DL = Instr->getModule()->getDataLayout(); 1754 if (!Alignment) 1755 Alignment = DL.getABITypeAlignment(ScalarDataTy); 1756 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); 1757 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy); 1758 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF; 1759 1760 if (SI && Legal->blockNeedsPredication(SI->getParent()) && 1761 !Legal->isMaskRequired(SI)) 1762 return scalarizeInstruction(Instr, true); 1763 1764 if (ScalarAllocatedSize != VectorElementSize) 1765 return scalarizeInstruction(Instr); 1766 1767 // If the pointer is loop invariant or if it is non-consecutive, 1768 // scalarize the load. 1769 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 1770 bool Reverse = ConsecutiveStride < 0; 1771 bool UniformLoad = LI && Legal->isUniform(Ptr); 1772 if (!ConsecutiveStride || UniformLoad) 1773 return scalarizeInstruction(Instr); 1774 1775 Constant *Zero = Builder.getInt32(0); 1776 VectorParts &Entry = WidenMap.get(Instr); 1777 1778 // Handle consecutive loads/stores. 1779 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 1780 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) { 1781 setDebugLocFromInst(Builder, Gep); 1782 Value *PtrOperand = Gep->getPointerOperand(); 1783 Value *FirstBasePtr = getVectorValue(PtrOperand)[0]; 1784 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero); 1785 1786 // Create the new GEP with the new induction variable. 1787 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1788 Gep2->setOperand(0, FirstBasePtr); 1789 Gep2->setName("gep.indvar.base"); 1790 Ptr = Builder.Insert(Gep2); 1791 } else if (Gep) { 1792 setDebugLocFromInst(Builder, Gep); 1793 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), 1794 OrigLoop) && "Base ptr must be invariant"); 1795 1796 // The last index does not have to be the induction. It can be 1797 // consecutive and be a function of the index. For example A[I+1]; 1798 unsigned NumOperands = Gep->getNumOperands(); 1799 unsigned InductionOperand = getGEPInductionOperand(Gep); 1800 // Create the new GEP with the new induction variable. 1801 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1802 1803 for (unsigned i = 0; i < NumOperands; ++i) { 1804 Value *GepOperand = Gep->getOperand(i); 1805 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand); 1806 1807 // Update last index or loop invariant instruction anchored in loop. 1808 if (i == InductionOperand || 1809 (GepOperandInst && OrigLoop->contains(GepOperandInst))) { 1810 assert((i == InductionOperand || 1811 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) && 1812 "Must be last index or loop invariant"); 1813 1814 VectorParts &GEPParts = getVectorValue(GepOperand); 1815 Value *Index = GEPParts[0]; 1816 Index = Builder.CreateExtractElement(Index, Zero); 1817 Gep2->setOperand(i, Index); 1818 Gep2->setName("gep.indvar.idx"); 1819 } 1820 } 1821 Ptr = Builder.Insert(Gep2); 1822 } else { 1823 // Use the induction element ptr. 1824 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 1825 setDebugLocFromInst(Builder, Ptr); 1826 VectorParts &PtrVal = getVectorValue(Ptr); 1827 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 1828 } 1829 1830 VectorParts Mask = createBlockInMask(Instr->getParent()); 1831 // Handle Stores: 1832 if (SI) { 1833 assert(!Legal->isUniform(SI->getPointerOperand()) && 1834 "We do not allow storing to uniform addresses"); 1835 setDebugLocFromInst(Builder, SI); 1836 // We don't want to update the value in the map as it might be used in 1837 // another expression. So don't use a reference type for "StoredVal". 1838 VectorParts StoredVal = getVectorValue(SI->getValueOperand()); 1839 1840 for (unsigned Part = 0; Part < UF; ++Part) { 1841 // Calculate the pointer for the specific unroll-part. 1842 Value *PartPtr = 1843 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF)); 1844 1845 if (Reverse) { 1846 // If we store to reverse consecutive memory locations then we need 1847 // to reverse the order of elements in the stored value. 1848 StoredVal[Part] = reverseVector(StoredVal[Part]); 1849 // If the address is consecutive but reversed, then the 1850 // wide store needs to start at the last vector element. 1851 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF)); 1852 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF)); 1853 Mask[Part] = reverseVector(Mask[Part]); 1854 } 1855 1856 Value *VecPtr = Builder.CreateBitCast(PartPtr, 1857 DataTy->getPointerTo(AddressSpace)); 1858 1859 Instruction *NewSI; 1860 if (Legal->isMaskRequired(SI)) 1861 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment, 1862 Mask[Part]); 1863 else 1864 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment); 1865 propagateMetadata(NewSI, SI); 1866 } 1867 return; 1868 } 1869 1870 // Handle loads. 1871 assert(LI && "Must have a load instruction"); 1872 setDebugLocFromInst(Builder, LI); 1873 for (unsigned Part = 0; Part < UF; ++Part) { 1874 // Calculate the pointer for the specific unroll-part. 1875 Value *PartPtr = 1876 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF)); 1877 1878 if (Reverse) { 1879 // If the address is consecutive but reversed, then the 1880 // wide load needs to start at the last vector element. 1881 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF)); 1882 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF)); 1883 Mask[Part] = reverseVector(Mask[Part]); 1884 } 1885 1886 Instruction* NewLI; 1887 Value *VecPtr = Builder.CreateBitCast(PartPtr, 1888 DataTy->getPointerTo(AddressSpace)); 1889 if (Legal->isMaskRequired(LI)) 1890 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part], 1891 UndefValue::get(DataTy), 1892 "wide.masked.load"); 1893 else 1894 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load"); 1895 propagateMetadata(NewLI, LI); 1896 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI; 1897 } 1898} 1899 1900void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) { 1901 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 1902 // Holds vector parameters or scalars, in case of uniform vals. 1903 SmallVector<VectorParts, 4> Params; 1904 1905 setDebugLocFromInst(Builder, Instr); 1906 1907 // Find all of the vectorized parameters. 1908 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1909 Value *SrcOp = Instr->getOperand(op); 1910 1911 // If we are accessing the old induction variable, use the new one. 1912 if (SrcOp == OldInduction) { 1913 Params.push_back(getVectorValue(SrcOp)); 1914 continue; 1915 } 1916 1917 // Try using previously calculated values. 1918 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 1919 1920 // If the src is an instruction that appeared earlier in the basic block 1921 // then it should already be vectorized. 1922 if (SrcInst && OrigLoop->contains(SrcInst)) { 1923 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 1924 // The parameter is a vector value from earlier. 1925 Params.push_back(WidenMap.get(SrcInst)); 1926 } else { 1927 // The parameter is a scalar from outside the loop. Maybe even a constant. 1928 VectorParts Scalars; 1929 Scalars.append(UF, SrcOp); 1930 Params.push_back(Scalars); 1931 } 1932 } 1933 1934 assert(Params.size() == Instr->getNumOperands() && 1935 "Invalid number of operands"); 1936 1937 // Does this instruction return a value ? 1938 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 1939 1940 Value *UndefVec = IsVoidRetTy ? nullptr : 1941 UndefValue::get(VectorType::get(Instr->getType(), VF)); 1942 // Create a new entry in the WidenMap and initialize it to Undef or Null. 1943 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 1944 1945 Instruction *InsertPt = Builder.GetInsertPoint(); 1946 BasicBlock *IfBlock = Builder.GetInsertBlock(); 1947 BasicBlock *CondBlock = nullptr; 1948 1949 VectorParts Cond; 1950 Loop *VectorLp = nullptr; 1951 if (IfPredicateStore) { 1952 assert(Instr->getParent()->getSinglePredecessor() && 1953 "Only support single predecessor blocks"); 1954 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(), 1955 Instr->getParent()); 1956 VectorLp = LI->getLoopFor(IfBlock); 1957 assert(VectorLp && "Must have a loop for this block"); 1958 } 1959 1960 // For each vector unroll 'part': 1961 for (unsigned Part = 0; Part < UF; ++Part) { 1962 // For each scalar that we create: 1963 for (unsigned Width = 0; Width < VF; ++Width) { 1964 1965 // Start if-block. 1966 Value *Cmp = nullptr; 1967 if (IfPredicateStore) { 1968 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width)); 1969 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1)); 1970 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store"); 1971 LoopVectorBody.push_back(CondBlock); 1972 VectorLp->addBasicBlockToLoop(CondBlock, *LI); 1973 // Update Builder with newly created basic block. 1974 Builder.SetInsertPoint(InsertPt); 1975 } 1976 1977 Instruction *Cloned = Instr->clone(); 1978 if (!IsVoidRetTy) 1979 Cloned->setName(Instr->getName() + ".cloned"); 1980 // Replace the operands of the cloned instructions with extracted scalars. 1981 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1982 Value *Op = Params[op][Part]; 1983 // Param is a vector. Need to extract the right lane. 1984 if (Op->getType()->isVectorTy()) 1985 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); 1986 Cloned->setOperand(op, Op); 1987 } 1988 1989 // Place the cloned scalar in the new loop. 1990 Builder.Insert(Cloned); 1991 1992 // If the original scalar returns a value we need to place it in a vector 1993 // so that future users will be able to use it. 1994 if (!IsVoidRetTy) 1995 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, 1996 Builder.getInt32(Width)); 1997 // End if-block. 1998 if (IfPredicateStore) { 1999 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else"); 2000 LoopVectorBody.push_back(NewIfBlock); 2001 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI); 2002 Builder.SetInsertPoint(InsertPt); 2003 Instruction *OldBr = IfBlock->getTerminator(); 2004 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr); 2005 OldBr->eraseFromParent(); 2006 IfBlock = NewIfBlock; 2007 } 2008 } 2009 } 2010} 2011 2012static Instruction *getFirstInst(Instruction *FirstInst, Value *V, 2013 Instruction *Loc) { 2014 if (FirstInst) 2015 return FirstInst; 2016 if (Instruction *I = dyn_cast<Instruction>(V)) 2017 return I->getParent() == Loc->getParent() ? I : nullptr; 2018 return nullptr; 2019} 2020 2021std::pair<Instruction *, Instruction *> 2022InnerLoopVectorizer::addStrideCheck(Instruction *Loc) { 2023 Instruction *tnullptr = nullptr; 2024 if (!Legal->mustCheckStrides()) 2025 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr); 2026 2027 IRBuilder<> ChkBuilder(Loc); 2028 2029 // Emit checks. 2030 Value *Check = nullptr; 2031 Instruction *FirstInst = nullptr; 2032 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(), 2033 SE = Legal->strides_end(); 2034 SI != SE; ++SI) { 2035 Value *Ptr = stripIntegerCast(*SI); 2036 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1), 2037 "stride.chk"); 2038 // Store the first instruction we create. 2039 FirstInst = getFirstInst(FirstInst, C, Loc); 2040 if (Check) 2041 Check = ChkBuilder.CreateOr(Check, C); 2042 else 2043 Check = C; 2044 } 2045 2046 // We have to do this trickery because the IRBuilder might fold the check to a 2047 // constant expression in which case there is no Instruction anchored in a 2048 // the block. 2049 LLVMContext &Ctx = Loc->getContext(); 2050 Instruction *TheCheck = 2051 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx)); 2052 ChkBuilder.Insert(TheCheck, "stride.not.one"); 2053 FirstInst = getFirstInst(FirstInst, TheCheck, Loc); 2054 2055 return std::make_pair(FirstInst, TheCheck); 2056} 2057 2058void InnerLoopVectorizer::createEmptyLoop() { 2059 /* 2060 In this function we generate a new loop. The new loop will contain 2061 the vectorized instructions while the old loop will continue to run the 2062 scalar remainder. 2063 2064 [ ] <-- Back-edge taken count overflow check. 2065 / | 2066 / v 2067 | [ ] <-- vector loop bypass (may consist of multiple blocks). 2068 | / | 2069 | / v 2070 || [ ] <-- vector pre header. 2071 || | 2072 || v 2073 || [ ] \ 2074 || [ ]_| <-- vector loop. 2075 || | 2076 | \ v 2077 | >[ ] <--- middle-block. 2078 | / | 2079 | / v 2080 -|- >[ ] <--- new preheader. 2081 | | 2082 | v 2083 | [ ] \ 2084 | [ ]_| <-- old scalar loop to handle remainder. 2085 \ | 2086 \ v 2087 >[ ] <-- exit block. 2088 ... 2089 */ 2090 2091 BasicBlock *OldBasicBlock = OrigLoop->getHeader(); 2092 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader(); 2093 BasicBlock *ExitBlock = OrigLoop->getExitBlock(); 2094 assert(BypassBlock && "Invalid loop structure"); 2095 assert(ExitBlock && "Must have an exit block"); 2096 2097 // Some loops have a single integer induction variable, while other loops 2098 // don't. One example is c++ iterators that often have multiple pointer 2099 // induction variables. In the code below we also support a case where we 2100 // don't have a single induction variable. 2101 OldInduction = Legal->getInduction(); 2102 Type *IdxTy = Legal->getWidestInductionType(); 2103 2104 // Find the loop boundaries. 2105 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop); 2106 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count"); 2107 2108 // The exit count might have the type of i64 while the phi is i32. This can 2109 // happen if we have an induction variable that is sign extended before the 2110 // compare. The only way that we get a backedge taken count is that the 2111 // induction variable was signed and as such will not overflow. In such a case 2112 // truncation is legal. 2113 if (ExitCount->getType()->getPrimitiveSizeInBits() > 2114 IdxTy->getPrimitiveSizeInBits()) 2115 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy); 2116 2117 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy); 2118 // Get the total trip count from the count by adding 1. 2119 ExitCount = SE->getAddExpr(BackedgeTakeCount, 2120 SE->getConstant(BackedgeTakeCount->getType(), 1)); 2121 2122 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout(); 2123 2124 // Expand the trip count and place the new instructions in the preheader. 2125 // Notice that the pre-header does not change, only the loop body. 2126 SCEVExpander Exp(*SE, DL, "induction"); 2127 2128 // We need to test whether the backedge-taken count is uint##_max. Adding one 2129 // to it will cause overflow and an incorrect loop trip count in the vector 2130 // body. In case of overflow we want to directly jump to the scalar remainder 2131 // loop. 2132 Value *BackedgeCount = 2133 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(), 2134 BypassBlock->getTerminator()); 2135 if (BackedgeCount->getType()->isPointerTy()) 2136 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy, 2137 "backedge.ptrcnt.to.int", 2138 BypassBlock->getTerminator()); 2139 Instruction *CheckBCOverflow = 2140 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount, 2141 Constant::getAllOnesValue(BackedgeCount->getType()), 2142 "backedge.overflow", BypassBlock->getTerminator()); 2143 2144 // The loop index does not have to start at Zero. Find the original start 2145 // value from the induction PHI node. If we don't have an induction variable 2146 // then we know that it starts at zero. 2147 Builder.SetInsertPoint(BypassBlock->getTerminator()); 2148 Value *StartIdx = ExtendedIdx = OldInduction ? 2149 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock), 2150 IdxTy): 2151 ConstantInt::get(IdxTy, 0); 2152 2153 // We need an instruction to anchor the overflow check on. StartIdx needs to 2154 // be defined before the overflow check branch. Because the scalar preheader 2155 // is going to merge the start index and so the overflow branch block needs to 2156 // contain a definition of the start index. 2157 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd( 2158 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor", 2159 BypassBlock->getTerminator()); 2160 2161 // Count holds the overall loop count (N). 2162 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 2163 BypassBlock->getTerminator()); 2164 2165 LoopBypassBlocks.push_back(BypassBlock); 2166 2167 // Split the single block loop into the two loop structure described above. 2168 BasicBlock *VectorPH = 2169 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); 2170 BasicBlock *VecBody = 2171 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 2172 BasicBlock *MiddleBlock = 2173 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 2174 BasicBlock *ScalarPH = 2175 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 2176 2177 // Create and register the new vector loop. 2178 Loop* Lp = new Loop(); 2179 Loop *ParentLoop = OrigLoop->getParentLoop(); 2180 2181 // Insert the new loop into the loop nest and register the new basic blocks 2182 // before calling any utilities such as SCEV that require valid LoopInfo. 2183 if (ParentLoop) { 2184 ParentLoop->addChildLoop(Lp); 2185 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI); 2186 ParentLoop->addBasicBlockToLoop(VectorPH, *LI); 2187 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI); 2188 } else { 2189 LI->addTopLevelLoop(Lp); 2190 } 2191 Lp->addBasicBlockToLoop(VecBody, *LI); 2192 2193 // Use this IR builder to create the loop instructions (Phi, Br, Cmp) 2194 // inside the loop. 2195 Builder.SetInsertPoint(VecBody->getFirstNonPHI()); 2196 2197 // Generate the induction variable. 2198 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction)); 2199 Induction = Builder.CreatePHI(IdxTy, 2, "index"); 2200 // The loop step is equal to the vectorization factor (num of SIMD elements) 2201 // times the unroll factor (num of SIMD instructions). 2202 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 2203 2204 // This is the IR builder that we use to add all of the logic for bypassing 2205 // the new vector loop. 2206 IRBuilder<> BypassBuilder(BypassBlock->getTerminator()); 2207 setDebugLocFromInst(BypassBuilder, 2208 getDebugLocFromInstOrOperands(OldInduction)); 2209 2210 // We may need to extend the index in case there is a type mismatch. 2211 // We know that the count starts at zero and does not overflow. 2212 if (Count->getType() != IdxTy) { 2213 // The exit count can be of pointer type. Convert it to the correct 2214 // integer type. 2215 if (ExitCount->getType()->isPointerTy()) 2216 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int"); 2217 else 2218 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast"); 2219 } 2220 2221 // Add the start index to the loop count to get the new end index. 2222 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx"); 2223 2224 // Now we need to generate the expression for N - (N % VF), which is 2225 // the part that the vectorized body will execute. 2226 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf"); 2227 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec"); 2228 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx, 2229 "end.idx.rnd.down"); 2230 2231 // Now, compare the new count to zero. If it is zero skip the vector loop and 2232 // jump to the scalar loop. 2233 Value *Cmp = 2234 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero"); 2235 2236 BasicBlock *LastBypassBlock = BypassBlock; 2237 2238 // Generate code to check that the loops trip count that we computed by adding 2239 // one to the backedge-taken count will not overflow. 2240 { 2241 auto PastOverflowCheck = 2242 std::next(BasicBlock::iterator(OverflowCheckAnchor)); 2243 BasicBlock *CheckBlock = 2244 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked"); 2245 if (ParentLoop) 2246 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI); 2247 LoopBypassBlocks.push_back(CheckBlock); 2248 Instruction *OldTerm = LastBypassBlock->getTerminator(); 2249 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm); 2250 OldTerm->eraseFromParent(); 2251 LastBypassBlock = CheckBlock; 2252 } 2253 2254 // Generate the code to check that the strides we assumed to be one are really 2255 // one. We want the new basic block to start at the first instruction in a 2256 // sequence of instructions that form a check. 2257 Instruction *StrideCheck; 2258 Instruction *FirstCheckInst; 2259 std::tie(FirstCheckInst, StrideCheck) = 2260 addStrideCheck(LastBypassBlock->getTerminator()); 2261 if (StrideCheck) { 2262 AddedSafetyChecks = true; 2263 // Create a new block containing the stride check. 2264 BasicBlock *CheckBlock = 2265 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck"); 2266 if (ParentLoop) 2267 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI); 2268 LoopBypassBlocks.push_back(CheckBlock); 2269 2270 // Replace the branch into the memory check block with a conditional branch 2271 // for the "few elements case". 2272 Instruction *OldTerm = LastBypassBlock->getTerminator(); 2273 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); 2274 OldTerm->eraseFromParent(); 2275 2276 Cmp = StrideCheck; 2277 LastBypassBlock = CheckBlock; 2278 } 2279 2280 // Generate the code that checks in runtime if arrays overlap. We put the 2281 // checks into a separate block to make the more common case of few elements 2282 // faster. 2283 Instruction *MemRuntimeCheck; 2284 std::tie(FirstCheckInst, MemRuntimeCheck) = 2285 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator()); 2286 if (MemRuntimeCheck) { 2287 AddedSafetyChecks = true; 2288 // Create a new block containing the memory check. 2289 BasicBlock *CheckBlock = 2290 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck"); 2291 if (ParentLoop) 2292 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI); 2293 LoopBypassBlocks.push_back(CheckBlock); 2294 2295 // Replace the branch into the memory check block with a conditional branch 2296 // for the "few elements case". 2297 Instruction *OldTerm = LastBypassBlock->getTerminator(); 2298 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); 2299 OldTerm->eraseFromParent(); 2300 2301 Cmp = MemRuntimeCheck; 2302 LastBypassBlock = CheckBlock; 2303 } 2304 2305 LastBypassBlock->getTerminator()->eraseFromParent(); 2306 BranchInst::Create(MiddleBlock, VectorPH, Cmp, 2307 LastBypassBlock); 2308 2309 // We are going to resume the execution of the scalar loop. 2310 // Go over all of the induction variables that we found and fix the 2311 // PHIs that are left in the scalar version of the loop. 2312 // The starting values of PHI nodes depend on the counter of the last 2313 // iteration in the vectorized loop. 2314 // If we come from a bypass edge then we need to start from the original 2315 // start value. 2316 2317 // This variable saves the new starting index for the scalar loop. 2318 PHINode *ResumeIndex = nullptr; 2319 LoopVectorizationLegality::InductionList::iterator I, E; 2320 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 2321 // Set builder to point to last bypass block. 2322 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator()); 2323 for (I = List->begin(), E = List->end(); I != E; ++I) { 2324 PHINode *OrigPhi = I->first; 2325 LoopVectorizationLegality::InductionInfo II = I->second; 2326 2327 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType(); 2328 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val", 2329 MiddleBlock->getTerminator()); 2330 // We might have extended the type of the induction variable but we need a 2331 // truncated version for the scalar loop. 2332 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ? 2333 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val", 2334 MiddleBlock->getTerminator()) : nullptr; 2335 2336 // Create phi nodes to merge from the backedge-taken check block. 2337 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val", 2338 ScalarPH->getTerminator()); 2339 BCResumeVal->addIncoming(ResumeVal, MiddleBlock); 2340 2341 PHINode *BCTruncResumeVal = nullptr; 2342 if (OrigPhi == OldInduction) { 2343 BCTruncResumeVal = 2344 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val", 2345 ScalarPH->getTerminator()); 2346 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock); 2347 } 2348 2349 Value *EndValue = nullptr; 2350 switch (II.IK) { 2351 case LoopVectorizationLegality::IK_NoInduction: 2352 llvm_unreachable("Unknown induction"); 2353 case LoopVectorizationLegality::IK_IntInduction: { 2354 // Handle the integer induction counter. 2355 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); 2356 2357 // We have the canonical induction variable. 2358 if (OrigPhi == OldInduction) { 2359 // Create a truncated version of the resume value for the scalar loop, 2360 // we might have promoted the type to a larger width. 2361 EndValue = 2362 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType()); 2363 // The new PHI merges the original incoming value, in case of a bypass, 2364 // or the value at the end of the vectorized loop. 2365 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2366 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 2367 TruncResumeVal->addIncoming(EndValue, VecBody); 2368 2369 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]); 2370 2371 // We know what the end value is. 2372 EndValue = IdxEndRoundDown; 2373 // We also know which PHI node holds it. 2374 ResumeIndex = ResumeVal; 2375 break; 2376 } 2377 2378 // Not the canonical induction variable - add the vector loop count to the 2379 // start value. 2380 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, 2381 II.StartValue->getType(), 2382 "cast.crd"); 2383 EndValue = II.transform(BypassBuilder, CRD); 2384 EndValue->setName("ind.end"); 2385 break; 2386 } 2387 case LoopVectorizationLegality::IK_PtrInduction: { 2388 EndValue = II.transform(BypassBuilder, CountRoundDown); 2389 EndValue->setName("ptr.ind.end"); 2390 break; 2391 } 2392 }// end of case 2393 2394 // The new PHI merges the original incoming value, in case of a bypass, 2395 // or the value at the end of the vectorized loop. 2396 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) { 2397 if (OrigPhi == OldInduction) 2398 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]); 2399 else 2400 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 2401 } 2402 ResumeVal->addIncoming(EndValue, VecBody); 2403 2404 // Fix the scalar body counter (PHI node). 2405 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 2406 2407 // The old induction's phi node in the scalar body needs the truncated 2408 // value. 2409 if (OrigPhi == OldInduction) { 2410 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]); 2411 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal); 2412 } else { 2413 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]); 2414 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal); 2415 } 2416 } 2417 2418 // If we are generating a new induction variable then we also need to 2419 // generate the code that calculates the exit value. This value is not 2420 // simply the end of the counter because we may skip the vectorized body 2421 // in case of a runtime check. 2422 if (!OldInduction){ 2423 assert(!ResumeIndex && "Unexpected resume value found"); 2424 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", 2425 MiddleBlock->getTerminator()); 2426 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2427 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]); 2428 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); 2429 } 2430 2431 // Make sure that we found the index where scalar loop needs to continue. 2432 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() && 2433 "Invalid resume Index"); 2434 2435 // Add a check in the middle block to see if we have completed 2436 // all of the iterations in the first vector loop. 2437 // If (N - N%VF) == N, then we *don't* need to run the remainder. 2438 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, 2439 ResumeIndex, "cmp.n", 2440 MiddleBlock->getTerminator()); 2441 2442 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); 2443 // Remove the old terminator. 2444 MiddleBlock->getTerminator()->eraseFromParent(); 2445 2446 // Create i+1 and fill the PHINode. 2447 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); 2448 Induction->addIncoming(StartIdx, VectorPH); 2449 Induction->addIncoming(NextIdx, VecBody); 2450 // Create the compare. 2451 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); 2452 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); 2453 2454 // Now we have two terminators. Remove the old one from the block. 2455 VecBody->getTerminator()->eraseFromParent(); 2456 2457 // Get ready to start creating new instructions into the vectorized body. 2458 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 2459 2460 // Save the state. 2461 LoopVectorPreHeader = VectorPH; 2462 LoopScalarPreHeader = ScalarPH; 2463 LoopMiddleBlock = MiddleBlock; 2464 LoopExitBlock = ExitBlock; 2465 LoopVectorBody.push_back(VecBody); 2466 LoopScalarBody = OldBasicBlock; 2467 2468 LoopVectorizeHints Hints(Lp, true); 2469 Hints.setAlreadyVectorized(); 2470} 2471 2472/// This function returns the identity element (or neutral element) for 2473/// the operation K. 2474Constant* 2475LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) { 2476 switch (K) { 2477 case RK_IntegerXor: 2478 case RK_IntegerAdd: 2479 case RK_IntegerOr: 2480 // Adding, Xoring, Oring zero to a number does not change it. 2481 return ConstantInt::get(Tp, 0); 2482 case RK_IntegerMult: 2483 // Multiplying a number by 1 does not change it. 2484 return ConstantInt::get(Tp, 1); 2485 case RK_IntegerAnd: 2486 // AND-ing a number with an all-1 value does not change it. 2487 return ConstantInt::get(Tp, -1, true); 2488 case RK_FloatMult: 2489 // Multiplying a number by 1 does not change it. 2490 return ConstantFP::get(Tp, 1.0L); 2491 case RK_FloatAdd: 2492 // Adding zero to a number does not change it. 2493 return ConstantFP::get(Tp, 0.0L); 2494 default: 2495 llvm_unreachable("Unknown reduction kind"); 2496 } 2497} 2498 2499/// This function translates the reduction kind to an LLVM binary operator. 2500static unsigned 2501getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) { 2502 switch (Kind) { 2503 case LoopVectorizationLegality::RK_IntegerAdd: 2504 return Instruction::Add; 2505 case LoopVectorizationLegality::RK_IntegerMult: 2506 return Instruction::Mul; 2507 case LoopVectorizationLegality::RK_IntegerOr: 2508 return Instruction::Or; 2509 case LoopVectorizationLegality::RK_IntegerAnd: 2510 return Instruction::And; 2511 case LoopVectorizationLegality::RK_IntegerXor: 2512 return Instruction::Xor; 2513 case LoopVectorizationLegality::RK_FloatMult: 2514 return Instruction::FMul; 2515 case LoopVectorizationLegality::RK_FloatAdd: 2516 return Instruction::FAdd; 2517 case LoopVectorizationLegality::RK_IntegerMinMax: 2518 return Instruction::ICmp; 2519 case LoopVectorizationLegality::RK_FloatMinMax: 2520 return Instruction::FCmp; 2521 default: 2522 llvm_unreachable("Unknown reduction operation"); 2523 } 2524} 2525 2526static Value *createMinMaxOp(IRBuilder<> &Builder, 2527 LoopVectorizationLegality::MinMaxReductionKind RK, 2528 Value *Left, Value *Right) { 2529 CmpInst::Predicate P = CmpInst::ICMP_NE; 2530 switch (RK) { 2531 default: 2532 llvm_unreachable("Unknown min/max reduction kind"); 2533 case LoopVectorizationLegality::MRK_UIntMin: 2534 P = CmpInst::ICMP_ULT; 2535 break; 2536 case LoopVectorizationLegality::MRK_UIntMax: 2537 P = CmpInst::ICMP_UGT; 2538 break; 2539 case LoopVectorizationLegality::MRK_SIntMin: 2540 P = CmpInst::ICMP_SLT; 2541 break; 2542 case LoopVectorizationLegality::MRK_SIntMax: 2543 P = CmpInst::ICMP_SGT; 2544 break; 2545 case LoopVectorizationLegality::MRK_FloatMin: 2546 P = CmpInst::FCMP_OLT; 2547 break; 2548 case LoopVectorizationLegality::MRK_FloatMax: 2549 P = CmpInst::FCMP_OGT; 2550 break; 2551 } 2552 2553 Value *Cmp; 2554 if (RK == LoopVectorizationLegality::MRK_FloatMin || 2555 RK == LoopVectorizationLegality::MRK_FloatMax) 2556 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp"); 2557 else 2558 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); 2559 2560 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); 2561 return Select; 2562} 2563 2564namespace { 2565struct CSEDenseMapInfo { 2566 static bool canHandle(Instruction *I) { 2567 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 2568 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); 2569 } 2570 static inline Instruction *getEmptyKey() { 2571 return DenseMapInfo<Instruction *>::getEmptyKey(); 2572 } 2573 static inline Instruction *getTombstoneKey() { 2574 return DenseMapInfo<Instruction *>::getTombstoneKey(); 2575 } 2576 static unsigned getHashValue(Instruction *I) { 2577 assert(canHandle(I) && "Unknown instruction!"); 2578 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), 2579 I->value_op_end())); 2580 } 2581 static bool isEqual(Instruction *LHS, Instruction *RHS) { 2582 if (LHS == getEmptyKey() || RHS == getEmptyKey() || 2583 LHS == getTombstoneKey() || RHS == getTombstoneKey()) 2584 return LHS == RHS; 2585 return LHS->isIdenticalTo(RHS); 2586 } 2587}; 2588} 2589 2590/// \brief Check whether this block is a predicated block. 2591/// Due to if predication of stores we might create a sequence of "if(pred) a[i] 2592/// = ...; " blocks. We start with one vectorized basic block. For every 2593/// conditional block we split this vectorized block. Therefore, every second 2594/// block will be a predicated one. 2595static bool isPredicatedBlock(unsigned BlockNum) { 2596 return BlockNum % 2; 2597} 2598 2599///\brief Perform cse of induction variable instructions. 2600static void cse(SmallVector<BasicBlock *, 4> &BBs) { 2601 // Perform simple cse. 2602 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; 2603 for (unsigned i = 0, e = BBs.size(); i != e; ++i) { 2604 BasicBlock *BB = BBs[i]; 2605 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { 2606 Instruction *In = I++; 2607 2608 if (!CSEDenseMapInfo::canHandle(In)) 2609 continue; 2610 2611 // Check if we can replace this instruction with any of the 2612 // visited instructions. 2613 if (Instruction *V = CSEMap.lookup(In)) { 2614 In->replaceAllUsesWith(V); 2615 In->eraseFromParent(); 2616 continue; 2617 } 2618 // Ignore instructions in conditional blocks. We create "if (pred) a[i] = 2619 // ...;" blocks for predicated stores. Every second block is a predicated 2620 // block. 2621 if (isPredicatedBlock(i)) 2622 continue; 2623 2624 CSEMap[In] = In; 2625 } 2626 } 2627} 2628 2629/// \brief Adds a 'fast' flag to floating point operations. 2630static Value *addFastMathFlag(Value *V) { 2631 if (isa<FPMathOperator>(V)){ 2632 FastMathFlags Flags; 2633 Flags.setUnsafeAlgebra(); 2634 cast<Instruction>(V)->setFastMathFlags(Flags); 2635 } 2636 return V; 2637} 2638 2639/// Estimate the overhead of scalarizing a value. Insert and Extract are set if 2640/// the result needs to be inserted and/or extracted from vectors. 2641static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract, 2642 const TargetTransformInfo &TTI) { 2643 if (Ty->isVoidTy()) 2644 return 0; 2645 2646 assert(Ty->isVectorTy() && "Can only scalarize vectors"); 2647 unsigned Cost = 0; 2648 2649 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) { 2650 if (Insert) 2651 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i); 2652 if (Extract) 2653 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i); 2654 } 2655 2656 return Cost; 2657} 2658 2659// Estimate cost of a call instruction CI if it were vectorized with factor VF. 2660// Return the cost of the instruction, including scalarization overhead if it's 2661// needed. The flag NeedToScalarize shows if the call needs to be scalarized - 2662// i.e. either vector version isn't available, or is too expensive. 2663static unsigned getVectorCallCost(CallInst *CI, unsigned VF, 2664 const TargetTransformInfo &TTI, 2665 const TargetLibraryInfo *TLI, 2666 bool &NeedToScalarize) { 2667 Function *F = CI->getCalledFunction(); 2668 StringRef FnName = CI->getCalledFunction()->getName(); 2669 Type *ScalarRetTy = CI->getType(); 2670 SmallVector<Type *, 4> Tys, ScalarTys; 2671 for (auto &ArgOp : CI->arg_operands()) 2672 ScalarTys.push_back(ArgOp->getType()); 2673 2674 // Estimate cost of scalarized vector call. The source operands are assumed 2675 // to be vectors, so we need to extract individual elements from there, 2676 // execute VF scalar calls, and then gather the result into the vector return 2677 // value. 2678 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys); 2679 if (VF == 1) 2680 return ScalarCallCost; 2681 2682 // Compute corresponding vector type for return value and arguments. 2683 Type *RetTy = ToVectorTy(ScalarRetTy, VF); 2684 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i) 2685 Tys.push_back(ToVectorTy(ScalarTys[i], VF)); 2686 2687 // Compute costs of unpacking argument values for the scalar calls and 2688 // packing the return values to a vector. 2689 unsigned ScalarizationCost = 2690 getScalarizationOverhead(RetTy, true, false, TTI); 2691 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) 2692 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI); 2693 2694 unsigned Cost = ScalarCallCost * VF + ScalarizationCost; 2695 2696 // If we can't emit a vector call for this function, then the currently found 2697 // cost is the cost we need to return. 2698 NeedToScalarize = true; 2699 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin()) 2700 return Cost; 2701 2702 // If the corresponding vector cost is cheaper, return its cost. 2703 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys); 2704 if (VectorCallCost < Cost) { 2705 NeedToScalarize = false; 2706 return VectorCallCost; 2707 } 2708 return Cost; 2709} 2710 2711// Estimate cost of an intrinsic call instruction CI if it were vectorized with 2712// factor VF. Return the cost of the instruction, including scalarization 2713// overhead if it's needed. 2714static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF, 2715 const TargetTransformInfo &TTI, 2716 const TargetLibraryInfo *TLI) { 2717 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 2718 assert(ID && "Expected intrinsic call!"); 2719 2720 Type *RetTy = ToVectorTy(CI->getType(), VF); 2721 SmallVector<Type *, 4> Tys; 2722 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 2723 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 2724 2725 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); 2726} 2727 2728void InnerLoopVectorizer::vectorizeLoop() { 2729 //===------------------------------------------------===// 2730 // 2731 // Notice: any optimization or new instruction that go 2732 // into the code below should be also be implemented in 2733 // the cost-model. 2734 // 2735 //===------------------------------------------------===// 2736 Constant *Zero = Builder.getInt32(0); 2737 2738 // In order to support reduction variables we need to be able to vectorize 2739 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two 2740 // stages. First, we create a new vector PHI node with no incoming edges. 2741 // We use this value when we vectorize all of the instructions that use the 2742 // PHI. Next, after all of the instructions in the block are complete we 2743 // add the new incoming edges to the PHI. At this point all of the 2744 // instructions in the basic block are vectorized, so we can use them to 2745 // construct the PHI. 2746 PhiVector RdxPHIsToFix; 2747 2748 // Scan the loop in a topological order to ensure that defs are vectorized 2749 // before users. 2750 LoopBlocksDFS DFS(OrigLoop); 2751 DFS.perform(LI); 2752 2753 // Vectorize all of the blocks in the original loop. 2754 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 2755 be = DFS.endRPO(); bb != be; ++bb) 2756 vectorizeBlockInLoop(*bb, &RdxPHIsToFix); 2757 2758 // At this point every instruction in the original loop is widened to 2759 // a vector form. We are almost done. Now, we need to fix the PHI nodes 2760 // that we vectorized. The PHI nodes are currently empty because we did 2761 // not want to introduce cycles. Notice that the remaining PHI nodes 2762 // that we need to fix are reduction variables. 2763 2764 // Create the 'reduced' values for each of the induction vars. 2765 // The reduced values are the vector values that we scalarize and combine 2766 // after the loop is finished. 2767 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); 2768 it != e; ++it) { 2769 PHINode *RdxPhi = *it; 2770 assert(RdxPhi && "Unable to recover vectorized PHI"); 2771 2772 // Find the reduction variable descriptor. 2773 assert(Legal->getReductionVars()->count(RdxPhi) && 2774 "Unable to find the reduction variable"); 2775 LoopVectorizationLegality::ReductionDescriptor RdxDesc = 2776 (*Legal->getReductionVars())[RdxPhi]; 2777 2778 setDebugLocFromInst(Builder, RdxDesc.StartValue); 2779 2780 // We need to generate a reduction vector from the incoming scalar. 2781 // To do so, we need to generate the 'identity' vector and override 2782 // one of the elements with the incoming scalar reduction. We need 2783 // to do it in the vector-loop preheader. 2784 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator()); 2785 2786 // This is the vector-clone of the value that leaves the loop. 2787 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr); 2788 Type *VecTy = VectorExit[0]->getType(); 2789 2790 // Find the reduction identity variable. Zero for addition, or, xor, 2791 // one for multiplication, -1 for And. 2792 Value *Identity; 2793 Value *VectorStart; 2794 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax || 2795 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) { 2796 // MinMax reduction have the start value as their identify. 2797 if (VF == 1) { 2798 VectorStart = Identity = RdxDesc.StartValue; 2799 } else { 2800 VectorStart = Identity = Builder.CreateVectorSplat(VF, 2801 RdxDesc.StartValue, 2802 "minmax.ident"); 2803 } 2804 } else { 2805 // Handle other reduction kinds: 2806 Constant *Iden = 2807 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind, 2808 VecTy->getScalarType()); 2809 if (VF == 1) { 2810 Identity = Iden; 2811 // This vector is the Identity vector where the first element is the 2812 // incoming scalar reduction. 2813 VectorStart = RdxDesc.StartValue; 2814 } else { 2815 Identity = ConstantVector::getSplat(VF, Iden); 2816 2817 // This vector is the Identity vector where the first element is the 2818 // incoming scalar reduction. 2819 VectorStart = Builder.CreateInsertElement(Identity, 2820 RdxDesc.StartValue, Zero); 2821 } 2822 } 2823 2824 // Fix the vector-loop phi. 2825 2826 // Reductions do not have to start at zero. They can start with 2827 // any loop invariant values. 2828 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 2829 BasicBlock *Latch = OrigLoop->getLoopLatch(); 2830 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 2831 VectorParts &Val = getVectorValue(LoopVal); 2832 for (unsigned part = 0; part < UF; ++part) { 2833 // Make sure to add the reduction stat value only to the 2834 // first unroll part. 2835 Value *StartVal = (part == 0) ? VectorStart : Identity; 2836 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, 2837 LoopVectorPreHeader); 2838 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], 2839 LoopVectorBody.back()); 2840 } 2841 2842 // Before each round, move the insertion point right between 2843 // the PHIs and the values we are going to write. 2844 // This allows us to write both PHINodes and the extractelement 2845 // instructions. 2846 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); 2847 2848 VectorParts RdxParts; 2849 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr); 2850 for (unsigned part = 0; part < UF; ++part) { 2851 // This PHINode contains the vectorized reduction variable, or 2852 // the initial value vector, if we bypass the vector loop. 2853 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); 2854 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); 2855 Value *StartVal = (part == 0) ? VectorStart : Identity; 2856 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2857 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]); 2858 NewPhi->addIncoming(RdxExitVal[part], 2859 LoopVectorBody.back()); 2860 RdxParts.push_back(NewPhi); 2861 } 2862 2863 // Reduce all of the unrolled parts into a single vector. 2864 Value *ReducedPartRdx = RdxParts[0]; 2865 unsigned Op = getReductionBinOp(RdxDesc.Kind); 2866 setDebugLocFromInst(Builder, ReducedPartRdx); 2867 for (unsigned part = 1; part < UF; ++part) { 2868 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 2869 // Floating point operations had to be 'fast' to enable the reduction. 2870 ReducedPartRdx = addFastMathFlag( 2871 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part], 2872 ReducedPartRdx, "bin.rdx")); 2873 else 2874 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind, 2875 ReducedPartRdx, RdxParts[part]); 2876 } 2877 2878 if (VF > 1) { 2879 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 2880 // and vector ops, reducing the set of values being computed by half each 2881 // round. 2882 assert(isPowerOf2_32(VF) && 2883 "Reduction emission only supported for pow2 vectors!"); 2884 Value *TmpVec = ReducedPartRdx; 2885 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr); 2886 for (unsigned i = VF; i != 1; i >>= 1) { 2887 // Move the upper half of the vector to the lower half. 2888 for (unsigned j = 0; j != i/2; ++j) 2889 ShuffleMask[j] = Builder.getInt32(i/2 + j); 2890 2891 // Fill the rest of the mask with undef. 2892 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 2893 UndefValue::get(Builder.getInt32Ty())); 2894 2895 Value *Shuf = 2896 Builder.CreateShuffleVector(TmpVec, 2897 UndefValue::get(TmpVec->getType()), 2898 ConstantVector::get(ShuffleMask), 2899 "rdx.shuf"); 2900 2901 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 2902 // Floating point operations had to be 'fast' to enable the reduction. 2903 TmpVec = addFastMathFlag(Builder.CreateBinOp( 2904 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx")); 2905 else 2906 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf); 2907 } 2908 2909 // The result is in the first element of the vector. 2910 ReducedPartRdx = Builder.CreateExtractElement(TmpVec, 2911 Builder.getInt32(0)); 2912 } 2913 2914 // Create a phi node that merges control-flow from the backedge-taken check 2915 // block and the middle block. 2916 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx", 2917 LoopScalarPreHeader->getTerminator()); 2918 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]); 2919 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); 2920 2921 // Now, we need to fix the users of the reduction variable 2922 // inside and outside of the scalar remainder loop. 2923 // We know that the loop is in LCSSA form. We need to update the 2924 // PHI nodes in the exit blocks. 2925 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 2926 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 2927 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 2928 if (!LCSSAPhi) break; 2929 2930 // All PHINodes need to have a single entry edge, or two if 2931 // we already fixed them. 2932 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 2933 2934 // We found our reduction value exit-PHI. Update it with the 2935 // incoming bypass edge. 2936 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { 2937 // Add an edge coming from the bypass. 2938 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); 2939 break; 2940 } 2941 }// end of the LCSSA phi scan. 2942 2943 // Fix the scalar loop reduction variable with the incoming reduction sum 2944 // from the vector body and from the backedge value. 2945 int IncomingEdgeBlockIdx = 2946 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 2947 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 2948 // Pick the other block. 2949 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 2950 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); 2951 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); 2952 }// end of for each redux variable. 2953 2954 fixLCSSAPHIs(); 2955 2956 // Remove redundant induction instructions. 2957 cse(LoopVectorBody); 2958} 2959 2960void InnerLoopVectorizer::fixLCSSAPHIs() { 2961 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 2962 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 2963 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 2964 if (!LCSSAPhi) break; 2965 if (LCSSAPhi->getNumIncomingValues() == 1) 2966 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 2967 LoopMiddleBlock); 2968 } 2969} 2970 2971InnerLoopVectorizer::VectorParts 2972InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 2973 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 2974 "Invalid edge"); 2975 2976 // Look for cached value. 2977 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst); 2978 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge); 2979 if (ECEntryIt != MaskCache.end()) 2980 return ECEntryIt->second; 2981 2982 VectorParts SrcMask = createBlockInMask(Src); 2983 2984 // The terminator has to be a branch inst! 2985 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 2986 assert(BI && "Unexpected terminator found"); 2987 2988 if (BI->isConditional()) { 2989 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 2990 2991 if (BI->getSuccessor(0) != Dst) 2992 for (unsigned part = 0; part < UF; ++part) 2993 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 2994 2995 for (unsigned part = 0; part < UF; ++part) 2996 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 2997 2998 MaskCache[Edge] = EdgeMask; 2999 return EdgeMask; 3000 } 3001 3002 MaskCache[Edge] = SrcMask; 3003 return SrcMask; 3004} 3005 3006InnerLoopVectorizer::VectorParts 3007InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 3008 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 3009 3010 // Loop incoming mask is all-one. 3011 if (OrigLoop->getHeader() == BB) { 3012 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 3013 return getVectorValue(C); 3014 } 3015 3016 // This is the block mask. We OR all incoming edges, and with zero. 3017 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 3018 VectorParts BlockMask = getVectorValue(Zero); 3019 3020 // For each pred: 3021 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 3022 VectorParts EM = createEdgeMask(*it, BB); 3023 for (unsigned part = 0; part < UF; ++part) 3024 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 3025 } 3026 3027 return BlockMask; 3028} 3029 3030void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, 3031 InnerLoopVectorizer::VectorParts &Entry, 3032 unsigned UF, unsigned VF, PhiVector *PV) { 3033 PHINode* P = cast<PHINode>(PN); 3034 // Handle reduction variables: 3035 if (Legal->getReductionVars()->count(P)) { 3036 for (unsigned part = 0; part < UF; ++part) { 3037 // This is phase one of vectorizing PHIs. 3038 Type *VecTy = (VF == 1) ? PN->getType() : 3039 VectorType::get(PN->getType(), VF); 3040 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", 3041 LoopVectorBody.back()-> getFirstInsertionPt()); 3042 } 3043 PV->push_back(P); 3044 return; 3045 } 3046 3047 setDebugLocFromInst(Builder, P); 3048 // Check for PHI nodes that are lowered to vector selects. 3049 if (P->getParent() != OrigLoop->getHeader()) { 3050 // We know that all PHIs in non-header blocks are converted into 3051 // selects, so we don't have to worry about the insertion order and we 3052 // can just use the builder. 3053 // At this point we generate the predication tree. There may be 3054 // duplications since this is a simple recursive scan, but future 3055 // optimizations will clean it up. 3056 3057 unsigned NumIncoming = P->getNumIncomingValues(); 3058 3059 // Generate a sequence of selects of the form: 3060 // SELECT(Mask3, In3, 3061 // SELECT(Mask2, In2, 3062 // ( ...))) 3063 for (unsigned In = 0; In < NumIncoming; In++) { 3064 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), 3065 P->getParent()); 3066 VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); 3067 3068 for (unsigned part = 0; part < UF; ++part) { 3069 // We might have single edge PHIs (blocks) - use an identity 3070 // 'select' for the first PHI operand. 3071 if (In == 0) 3072 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 3073 In0[part]); 3074 else 3075 // Select between the current value and the previous incoming edge 3076 // based on the incoming mask. 3077 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 3078 Entry[part], "predphi"); 3079 } 3080 } 3081 return; 3082 } 3083 3084 // This PHINode must be an induction variable. 3085 // Make sure that we know about it. 3086 assert(Legal->getInductionVars()->count(P) && 3087 "Not an induction variable"); 3088 3089 LoopVectorizationLegality::InductionInfo II = 3090 Legal->getInductionVars()->lookup(P); 3091 3092 // FIXME: The newly created binary instructions should contain nsw/nuw flags, 3093 // which can be found from the original scalar operations. 3094 switch (II.IK) { 3095 case LoopVectorizationLegality::IK_NoInduction: 3096 llvm_unreachable("Unknown induction"); 3097 case LoopVectorizationLegality::IK_IntInduction: { 3098 assert(P->getType() == II.StartValue->getType() && "Types must match"); 3099 Type *PhiTy = P->getType(); 3100 Value *Broadcasted; 3101 if (P == OldInduction) { 3102 // Handle the canonical induction variable. We might have had to 3103 // extend the type. 3104 Broadcasted = Builder.CreateTrunc(Induction, PhiTy); 3105 } else { 3106 // Handle other induction variables that are now based on the 3107 // canonical one. 3108 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx, 3109 "normalized.idx"); 3110 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy); 3111 Broadcasted = II.transform(Builder, NormalizedIdx); 3112 Broadcasted->setName("offset.idx"); 3113 } 3114 Broadcasted = getBroadcastInstrs(Broadcasted); 3115 // After broadcasting the induction variable we need to make the vector 3116 // consecutive by adding 0, 1, 2, etc. 3117 for (unsigned part = 0; part < UF; ++part) 3118 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue); 3119 return; 3120 } 3121 case LoopVectorizationLegality::IK_PtrInduction: 3122 // Handle the pointer induction variable case. 3123 assert(P->getType()->isPointerTy() && "Unexpected type."); 3124 // This is the normalized GEP that starts counting at zero. 3125 Value *NormalizedIdx = 3126 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx"); 3127 // This is the vector of results. Notice that we don't generate 3128 // vector geps because scalar geps result in better code. 3129 for (unsigned part = 0; part < UF; ++part) { 3130 if (VF == 1) { 3131 int EltIndex = part; 3132 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 3133 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx); 3134 Value *SclrGep = II.transform(Builder, GlobalIdx); 3135 SclrGep->setName("next.gep"); 3136 Entry[part] = SclrGep; 3137 continue; 3138 } 3139 3140 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 3141 for (unsigned int i = 0; i < VF; ++i) { 3142 int EltIndex = i + part * VF; 3143 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 3144 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx); 3145 Value *SclrGep = II.transform(Builder, GlobalIdx); 3146 SclrGep->setName("next.gep"); 3147 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 3148 Builder.getInt32(i), 3149 "insert.gep"); 3150 } 3151 Entry[part] = VecVal; 3152 } 3153 return; 3154 } 3155} 3156 3157void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) { 3158 // For each instruction in the old loop. 3159 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3160 VectorParts &Entry = WidenMap.get(it); 3161 switch (it->getOpcode()) { 3162 case Instruction::Br: 3163 // Nothing to do for PHIs and BR, since we already took care of the 3164 // loop control flow instructions. 3165 continue; 3166 case Instruction::PHI: { 3167 // Vectorize PHINodes. 3168 widenPHIInstruction(it, Entry, UF, VF, PV); 3169 continue; 3170 }// End of PHI. 3171 3172 case Instruction::Add: 3173 case Instruction::FAdd: 3174 case Instruction::Sub: 3175 case Instruction::FSub: 3176 case Instruction::Mul: 3177 case Instruction::FMul: 3178 case Instruction::UDiv: 3179 case Instruction::SDiv: 3180 case Instruction::FDiv: 3181 case Instruction::URem: 3182 case Instruction::SRem: 3183 case Instruction::FRem: 3184 case Instruction::Shl: 3185 case Instruction::LShr: 3186 case Instruction::AShr: 3187 case Instruction::And: 3188 case Instruction::Or: 3189 case Instruction::Xor: { 3190 // Just widen binops. 3191 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 3192 setDebugLocFromInst(Builder, BinOp); 3193 VectorParts &A = getVectorValue(it->getOperand(0)); 3194 VectorParts &B = getVectorValue(it->getOperand(1)); 3195 3196 // Use this vector value for all users of the original instruction. 3197 for (unsigned Part = 0; Part < UF; ++Part) { 3198 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 3199 3200 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V)) 3201 VecOp->copyIRFlags(BinOp); 3202 3203 Entry[Part] = V; 3204 } 3205 3206 propagateMetadata(Entry, it); 3207 break; 3208 } 3209 case Instruction::Select: { 3210 // Widen selects. 3211 // If the selector is loop invariant we can create a select 3212 // instruction with a scalar condition. Otherwise, use vector-select. 3213 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), 3214 OrigLoop); 3215 setDebugLocFromInst(Builder, it); 3216 3217 // The condition can be loop invariant but still defined inside the 3218 // loop. This means that we can't just use the original 'cond' value. 3219 // We have to take the 'vectorized' value and pick the first lane. 3220 // Instcombine will make this a no-op. 3221 VectorParts &Cond = getVectorValue(it->getOperand(0)); 3222 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 3223 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 3224 3225 Value *ScalarCond = (VF == 1) ? Cond[0] : 3226 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0)); 3227 3228 for (unsigned Part = 0; Part < UF; ++Part) { 3229 Entry[Part] = Builder.CreateSelect( 3230 InvariantCond ? ScalarCond : Cond[Part], 3231 Op0[Part], 3232 Op1[Part]); 3233 } 3234 3235 propagateMetadata(Entry, it); 3236 break; 3237 } 3238 3239 case Instruction::ICmp: 3240 case Instruction::FCmp: { 3241 // Widen compares. Generate vector compares. 3242 bool FCmp = (it->getOpcode() == Instruction::FCmp); 3243 CmpInst *Cmp = dyn_cast<CmpInst>(it); 3244 setDebugLocFromInst(Builder, it); 3245 VectorParts &A = getVectorValue(it->getOperand(0)); 3246 VectorParts &B = getVectorValue(it->getOperand(1)); 3247 for (unsigned Part = 0; Part < UF; ++Part) { 3248 Value *C = nullptr; 3249 if (FCmp) 3250 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 3251 else 3252 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 3253 Entry[Part] = C; 3254 } 3255 3256 propagateMetadata(Entry, it); 3257 break; 3258 } 3259 3260 case Instruction::Store: 3261 case Instruction::Load: 3262 vectorizeMemoryInstruction(it); 3263 break; 3264 case Instruction::ZExt: 3265 case Instruction::SExt: 3266 case Instruction::FPToUI: 3267 case Instruction::FPToSI: 3268 case Instruction::FPExt: 3269 case Instruction::PtrToInt: 3270 case Instruction::IntToPtr: 3271 case Instruction::SIToFP: 3272 case Instruction::UIToFP: 3273 case Instruction::Trunc: 3274 case Instruction::FPTrunc: 3275 case Instruction::BitCast: { 3276 CastInst *CI = dyn_cast<CastInst>(it); 3277 setDebugLocFromInst(Builder, it); 3278 /// Optimize the special case where the source is the induction 3279 /// variable. Notice that we can only optimize the 'trunc' case 3280 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 3281 /// c. other casts depend on pointer size. 3282 if (CI->getOperand(0) == OldInduction && 3283 it->getOpcode() == Instruction::Trunc) { 3284 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 3285 CI->getType()); 3286 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 3287 LoopVectorizationLegality::InductionInfo II = 3288 Legal->getInductionVars()->lookup(OldInduction); 3289 Constant *Step = 3290 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue()); 3291 for (unsigned Part = 0; Part < UF; ++Part) 3292 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step); 3293 propagateMetadata(Entry, it); 3294 break; 3295 } 3296 /// Vectorize casts. 3297 Type *DestTy = (VF == 1) ? CI->getType() : 3298 VectorType::get(CI->getType(), VF); 3299 3300 VectorParts &A = getVectorValue(it->getOperand(0)); 3301 for (unsigned Part = 0; Part < UF; ++Part) 3302 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 3303 propagateMetadata(Entry, it); 3304 break; 3305 } 3306 3307 case Instruction::Call: { 3308 // Ignore dbg intrinsics. 3309 if (isa<DbgInfoIntrinsic>(it)) 3310 break; 3311 setDebugLocFromInst(Builder, it); 3312 3313 Module *M = BB->getParent()->getParent(); 3314 CallInst *CI = cast<CallInst>(it); 3315 3316 StringRef FnName = CI->getCalledFunction()->getName(); 3317 Function *F = CI->getCalledFunction(); 3318 Type *RetTy = ToVectorTy(CI->getType(), VF); 3319 SmallVector<Type *, 4> Tys; 3320 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 3321 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 3322 3323 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 3324 if (ID && 3325 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || 3326 ID == Intrinsic::lifetime_start)) { 3327 scalarizeInstruction(it); 3328 break; 3329 } 3330 // The flag shows whether we use Intrinsic or a usual Call for vectorized 3331 // version of the instruction. 3332 // Is it beneficial to perform intrinsic call compared to lib call? 3333 bool NeedToScalarize; 3334 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize); 3335 bool UseVectorIntrinsic = 3336 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost; 3337 if (!UseVectorIntrinsic && NeedToScalarize) { 3338 scalarizeInstruction(it); 3339 break; 3340 } 3341 3342 for (unsigned Part = 0; Part < UF; ++Part) { 3343 SmallVector<Value *, 4> Args; 3344 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 3345 Value *Arg = CI->getArgOperand(i); 3346 // Some intrinsics have a scalar argument - don't replace it with a 3347 // vector. 3348 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) { 3349 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i)); 3350 Arg = VectorArg[Part]; 3351 } 3352 Args.push_back(Arg); 3353 } 3354 3355 Function *VectorF; 3356 if (UseVectorIntrinsic) { 3357 // Use vector version of the intrinsic. 3358 Type *TysForDecl[] = {CI->getType()}; 3359 if (VF > 1) 3360 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF); 3361 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl); 3362 } else { 3363 // Use vector version of the library call. 3364 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF); 3365 assert(!VFnName.empty() && "Vector function name is empty."); 3366 VectorF = M->getFunction(VFnName); 3367 if (!VectorF) { 3368 // Generate a declaration 3369 FunctionType *FTy = FunctionType::get(RetTy, Tys, false); 3370 VectorF = 3371 Function::Create(FTy, Function::ExternalLinkage, VFnName, M); 3372 VectorF->copyAttributesFrom(F); 3373 } 3374 } 3375 assert(VectorF && "Can't create vector function."); 3376 Entry[Part] = Builder.CreateCall(VectorF, Args); 3377 } 3378 3379 propagateMetadata(Entry, it); 3380 break; 3381 } 3382 3383 default: 3384 // All other instructions are unsupported. Scalarize them. 3385 scalarizeInstruction(it); 3386 break; 3387 }// end of switch. 3388 }// end of for_each instr. 3389} 3390 3391void InnerLoopVectorizer::updateAnalysis() { 3392 // Forget the original basic block. 3393 SE->forgetLoop(OrigLoop); 3394 3395 // Update the dominator tree information. 3396 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && 3397 "Entry does not dominate exit."); 3398 3399 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 3400 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); 3401 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); 3402 3403 // Due to if predication of stores we might create a sequence of "if(pred) 3404 // a[i] = ...; " blocks. 3405 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) { 3406 if (i == 0) 3407 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader); 3408 else if (isPredicatedBlock(i)) { 3409 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]); 3410 } else { 3411 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]); 3412 } 3413 } 3414 3415 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]); 3416 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]); 3417 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 3418 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]); 3419 3420 DEBUG(DT->verifyDomTree()); 3421} 3422 3423/// \brief Check whether it is safe to if-convert this phi node. 3424/// 3425/// Phi nodes with constant expressions that can trap are not safe to if 3426/// convert. 3427static bool canIfConvertPHINodes(BasicBlock *BB) { 3428 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 3429 PHINode *Phi = dyn_cast<PHINode>(I); 3430 if (!Phi) 3431 return true; 3432 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p) 3433 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p))) 3434 if (C->canTrap()) 3435 return false; 3436 } 3437 return true; 3438} 3439 3440bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 3441 if (!EnableIfConversion) { 3442 emitAnalysis(VectorizationReport() << "if-conversion is disabled"); 3443 return false; 3444 } 3445 3446 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 3447 3448 // A list of pointers that we can safely read and write to. 3449 SmallPtrSet<Value *, 8> SafePointes; 3450 3451 // Collect safe addresses. 3452 for (Loop::block_iterator BI = TheLoop->block_begin(), 3453 BE = TheLoop->block_end(); BI != BE; ++BI) { 3454 BasicBlock *BB = *BI; 3455 3456 if (blockNeedsPredication(BB)) 3457 continue; 3458 3459 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 3460 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 3461 SafePointes.insert(LI->getPointerOperand()); 3462 else if (StoreInst *SI = dyn_cast<StoreInst>(I)) 3463 SafePointes.insert(SI->getPointerOperand()); 3464 } 3465 } 3466 3467 // Collect the blocks that need predication. 3468 BasicBlock *Header = TheLoop->getHeader(); 3469 for (Loop::block_iterator BI = TheLoop->block_begin(), 3470 BE = TheLoop->block_end(); BI != BE; ++BI) { 3471 BasicBlock *BB = *BI; 3472 3473 // We don't support switch statements inside loops. 3474 if (!isa<BranchInst>(BB->getTerminator())) { 3475 emitAnalysis(VectorizationReport(BB->getTerminator()) 3476 << "loop contains a switch statement"); 3477 return false; 3478 } 3479 3480 // We must be able to predicate all blocks that need to be predicated. 3481 if (blockNeedsPredication(BB)) { 3482 if (!blockCanBePredicated(BB, SafePointes)) { 3483 emitAnalysis(VectorizationReport(BB->getTerminator()) 3484 << "control flow cannot be substituted for a select"); 3485 return false; 3486 } 3487 } else if (BB != Header && !canIfConvertPHINodes(BB)) { 3488 emitAnalysis(VectorizationReport(BB->getTerminator()) 3489 << "control flow cannot be substituted for a select"); 3490 return false; 3491 } 3492 } 3493 3494 // We can if-convert this loop. 3495 return true; 3496} 3497 3498bool LoopVectorizationLegality::canVectorize() { 3499 // We must have a loop in canonical form. Loops with indirectbr in them cannot 3500 // be canonicalized. 3501 if (!TheLoop->getLoopPreheader()) { 3502 emitAnalysis( 3503 VectorizationReport() << 3504 "loop control flow is not understood by vectorizer"); 3505 return false; 3506 } 3507 3508 // We can only vectorize innermost loops. 3509 if (!TheLoop->getSubLoopsVector().empty()) { 3510 emitAnalysis(VectorizationReport() << "loop is not the innermost loop"); 3511 return false; 3512 } 3513 3514 // We must have a single backedge. 3515 if (TheLoop->getNumBackEdges() != 1) { 3516 emitAnalysis( 3517 VectorizationReport() << 3518 "loop control flow is not understood by vectorizer"); 3519 return false; 3520 } 3521 3522 // We must have a single exiting block. 3523 if (!TheLoop->getExitingBlock()) { 3524 emitAnalysis( 3525 VectorizationReport() << 3526 "loop control flow is not understood by vectorizer"); 3527 return false; 3528 } 3529 3530 // We only handle bottom-tested loops, i.e. loop in which the condition is 3531 // checked at the end of each iteration. With that we can assume that all 3532 // instructions in the loop are executed the same number of times. 3533 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { 3534 emitAnalysis( 3535 VectorizationReport() << 3536 "loop control flow is not understood by vectorizer"); 3537 return false; 3538 } 3539 3540 // We need to have a loop header. 3541 DEBUG(dbgs() << "LV: Found a loop: " << 3542 TheLoop->getHeader()->getName() << '\n'); 3543 3544 // Check if we can if-convert non-single-bb loops. 3545 unsigned NumBlocks = TheLoop->getNumBlocks(); 3546 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 3547 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 3548 return false; 3549 } 3550 3551 // ScalarEvolution needs to be able to find the exit count. 3552 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop); 3553 if (ExitCount == SE->getCouldNotCompute()) { 3554 emitAnalysis(VectorizationReport() << 3555 "could not determine number of loop iterations"); 3556 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 3557 return false; 3558 } 3559 3560 // Check if we can vectorize the instructions and CFG in this loop. 3561 if (!canVectorizeInstrs()) { 3562 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 3563 return false; 3564 } 3565 3566 // Go over each instruction and look at memory deps. 3567 if (!canVectorizeMemory()) { 3568 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 3569 return false; 3570 } 3571 3572 // Collect all of the variables that remain uniform after vectorization. 3573 collectLoopUniforms(); 3574 3575 DEBUG(dbgs() << "LV: We can vectorize this loop" << 3576 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" : 3577 "") 3578 <<"!\n"); 3579 3580 // Okay! We can vectorize. At this point we don't have any other mem analysis 3581 // which may limit our maximum vectorization factor, so just return true with 3582 // no restrictions. 3583 return true; 3584} 3585 3586static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) { 3587 if (Ty->isPointerTy()) 3588 return DL.getIntPtrType(Ty); 3589 3590 // It is possible that char's or short's overflow when we ask for the loop's 3591 // trip count, work around this by changing the type size. 3592 if (Ty->getScalarSizeInBits() < 32) 3593 return Type::getInt32Ty(Ty->getContext()); 3594 3595 return Ty; 3596} 3597 3598static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) { 3599 Ty0 = convertPointerToIntegerType(DL, Ty0); 3600 Ty1 = convertPointerToIntegerType(DL, Ty1); 3601 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits()) 3602 return Ty0; 3603 return Ty1; 3604} 3605 3606/// \brief Check that the instruction has outside loop users and is not an 3607/// identified reduction variable. 3608static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst, 3609 SmallPtrSetImpl<Value *> &Reductions) { 3610 // Reduction instructions are allowed to have exit users. All other 3611 // instructions must not have external users. 3612 if (!Reductions.count(Inst)) 3613 //Check that all of the users of the loop are inside the BB. 3614 for (User *U : Inst->users()) { 3615 Instruction *UI = cast<Instruction>(U); 3616 // This user may be a reduction exit value. 3617 if (!TheLoop->contains(UI)) { 3618 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n'); 3619 return true; 3620 } 3621 } 3622 return false; 3623} 3624 3625bool LoopVectorizationLegality::canVectorizeInstrs() { 3626 BasicBlock *PreHeader = TheLoop->getLoopPreheader(); 3627 BasicBlock *Header = TheLoop->getHeader(); 3628 3629 // Look for the attribute signaling the absence of NaNs. 3630 Function &F = *Header->getParent(); 3631 const DataLayout &DL = F.getParent()->getDataLayout(); 3632 if (F.hasFnAttribute("no-nans-fp-math")) 3633 HasFunNoNaNAttr = 3634 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true"; 3635 3636 // For each block in the loop. 3637 for (Loop::block_iterator bb = TheLoop->block_begin(), 3638 be = TheLoop->block_end(); bb != be; ++bb) { 3639 3640 // Scan the instructions in the block and look for hazards. 3641 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 3642 ++it) { 3643 3644 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 3645 Type *PhiTy = Phi->getType(); 3646 // Check that this PHI type is allowed. 3647 if (!PhiTy->isIntegerTy() && 3648 !PhiTy->isFloatingPointTy() && 3649 !PhiTy->isPointerTy()) { 3650 emitAnalysis(VectorizationReport(it) 3651 << "loop control flow is not understood by vectorizer"); 3652 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 3653 return false; 3654 } 3655 3656 // If this PHINode is not in the header block, then we know that we 3657 // can convert it to select during if-conversion. No need to check if 3658 // the PHIs in this block are induction or reduction variables. 3659 if (*bb != Header) { 3660 // Check that this instruction has no outside users or is an 3661 // identified reduction value with an outside user. 3662 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit)) 3663 continue; 3664 emitAnalysis(VectorizationReport(it) << 3665 "value could not be identified as " 3666 "an induction or reduction variable"); 3667 return false; 3668 } 3669 3670 // We only allow if-converted PHIs with exactly two incoming values. 3671 if (Phi->getNumIncomingValues() != 2) { 3672 emitAnalysis(VectorizationReport(it) 3673 << "control flow not understood by vectorizer"); 3674 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 3675 return false; 3676 } 3677 3678 // This is the value coming from the preheader. 3679 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); 3680 ConstantInt *StepValue = nullptr; 3681 // Check if this is an induction variable. 3682 InductionKind IK = isInductionVariable(Phi, StepValue); 3683 3684 if (IK_NoInduction != IK) { 3685 // Get the widest type. 3686 if (!WidestIndTy) 3687 WidestIndTy = convertPointerToIntegerType(DL, PhiTy); 3688 else 3689 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy); 3690 3691 // Int inductions are special because we only allow one IV. 3692 if (IK == IK_IntInduction && StepValue->isOne()) { 3693 // Use the phi node with the widest type as induction. Use the last 3694 // one if there are multiple (no good reason for doing this other 3695 // than it is expedient). 3696 if (!Induction || PhiTy == WidestIndTy) 3697 Induction = Phi; 3698 } 3699 3700 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 3701 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue); 3702 3703 // Until we explicitly handle the case of an induction variable with 3704 // an outside loop user we have to give up vectorizing this loop. 3705 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) { 3706 emitAnalysis(VectorizationReport(it) << 3707 "use of induction value outside of the " 3708 "loop is not handled by vectorizer"); 3709 return false; 3710 } 3711 3712 continue; 3713 } 3714 3715 if (AddReductionVar(Phi, RK_IntegerAdd)) { 3716 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); 3717 continue; 3718 } 3719 if (AddReductionVar(Phi, RK_IntegerMult)) { 3720 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); 3721 continue; 3722 } 3723 if (AddReductionVar(Phi, RK_IntegerOr)) { 3724 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); 3725 continue; 3726 } 3727 if (AddReductionVar(Phi, RK_IntegerAnd)) { 3728 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); 3729 continue; 3730 } 3731 if (AddReductionVar(Phi, RK_IntegerXor)) { 3732 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); 3733 continue; 3734 } 3735 if (AddReductionVar(Phi, RK_IntegerMinMax)) { 3736 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n"); 3737 continue; 3738 } 3739 if (AddReductionVar(Phi, RK_FloatMult)) { 3740 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n"); 3741 continue; 3742 } 3743 if (AddReductionVar(Phi, RK_FloatAdd)) { 3744 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n"); 3745 continue; 3746 } 3747 if (AddReductionVar(Phi, RK_FloatMinMax)) { 3748 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi << 3749 "\n"); 3750 continue; 3751 } 3752 3753 emitAnalysis(VectorizationReport(it) << 3754 "value that could not be identified as " 3755 "reduction is used outside the loop"); 3756 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 3757 return false; 3758 }// end of PHI handling 3759 3760 // We handle calls that: 3761 // * Are debug info intrinsics. 3762 // * Have a mapping to an IR intrinsic. 3763 // * Have a vector version available. 3764 CallInst *CI = dyn_cast<CallInst>(it); 3765 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) && 3766 !(CI->getCalledFunction() && TLI && 3767 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) { 3768 emitAnalysis(VectorizationReport(it) << 3769 "call instruction cannot be vectorized"); 3770 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n"); 3771 return false; 3772 } 3773 3774 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the 3775 // second argument is the same (i.e. loop invariant) 3776 if (CI && 3777 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) { 3778 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) { 3779 emitAnalysis(VectorizationReport(it) 3780 << "intrinsic instruction cannot be vectorized"); 3781 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n"); 3782 return false; 3783 } 3784 } 3785 3786 // Check that the instruction return type is vectorizable. 3787 // Also, we can't vectorize extractelement instructions. 3788 if ((!VectorType::isValidElementType(it->getType()) && 3789 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) { 3790 emitAnalysis(VectorizationReport(it) 3791 << "instruction return type cannot be vectorized"); 3792 DEBUG(dbgs() << "LV: Found unvectorizable type.\n"); 3793 return false; 3794 } 3795 3796 // Check that the stored type is vectorizable. 3797 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 3798 Type *T = ST->getValueOperand()->getType(); 3799 if (!VectorType::isValidElementType(T)) { 3800 emitAnalysis(VectorizationReport(ST) << 3801 "store instruction cannot be vectorized"); 3802 return false; 3803 } 3804 if (EnableMemAccessVersioning) 3805 collectStridedAccess(ST); 3806 } 3807 3808 if (EnableMemAccessVersioning) 3809 if (LoadInst *LI = dyn_cast<LoadInst>(it)) 3810 collectStridedAccess(LI); 3811 3812 // Reduction instructions are allowed to have exit users. 3813 // All other instructions must not have external users. 3814 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) { 3815 emitAnalysis(VectorizationReport(it) << 3816 "value cannot be used outside the loop"); 3817 return false; 3818 } 3819 3820 } // next instr. 3821 3822 } 3823 3824 if (!Induction) { 3825 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 3826 if (Inductions.empty()) { 3827 emitAnalysis(VectorizationReport() 3828 << "loop induction variable could not be identified"); 3829 return false; 3830 } 3831 } 3832 3833 return true; 3834} 3835 3836///\brief Remove GEPs whose indices but the last one are loop invariant and 3837/// return the induction operand of the gep pointer. 3838static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) { 3839 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 3840 if (!GEP) 3841 return Ptr; 3842 3843 unsigned InductionOperand = getGEPInductionOperand(GEP); 3844 3845 // Check that all of the gep indices are uniform except for our induction 3846 // operand. 3847 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) 3848 if (i != InductionOperand && 3849 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp)) 3850 return Ptr; 3851 return GEP->getOperand(InductionOperand); 3852} 3853 3854///\brief Look for a cast use of the passed value. 3855static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) { 3856 Value *UniqueCast = nullptr; 3857 for (User *U : Ptr->users()) { 3858 CastInst *CI = dyn_cast<CastInst>(U); 3859 if (CI && CI->getType() == Ty) { 3860 if (!UniqueCast) 3861 UniqueCast = CI; 3862 else 3863 return nullptr; 3864 } 3865 } 3866 return UniqueCast; 3867} 3868 3869///\brief Get the stride of a pointer access in a loop. 3870/// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a 3871/// pointer to the Value, or null otherwise. 3872static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) { 3873 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 3874 if (!PtrTy || PtrTy->isAggregateType()) 3875 return nullptr; 3876 3877 // Try to remove a gep instruction to make the pointer (actually index at this 3878 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the 3879 // pointer, otherwise, we are analyzing the index. 3880 Value *OrigPtr = Ptr; 3881 3882 // The size of the pointer access. 3883 int64_t PtrAccessSize = 1; 3884 3885 Ptr = stripGetElementPtr(Ptr, SE, Lp); 3886 const SCEV *V = SE->getSCEV(Ptr); 3887 3888 if (Ptr != OrigPtr) 3889 // Strip off casts. 3890 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) 3891 V = C->getOperand(); 3892 3893 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V); 3894 if (!S) 3895 return nullptr; 3896 3897 V = S->getStepRecurrence(*SE); 3898 if (!V) 3899 return nullptr; 3900 3901 // Strip off the size of access multiplication if we are still analyzing the 3902 // pointer. 3903 if (OrigPtr == Ptr) { 3904 const DataLayout &DL = Lp->getHeader()->getModule()->getDataLayout(); 3905 DL.getTypeAllocSize(PtrTy->getElementType()); 3906 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) { 3907 if (M->getOperand(0)->getSCEVType() != scConstant) 3908 return nullptr; 3909 3910 const APInt &APStepVal = 3911 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue(); 3912 3913 // Huge step value - give up. 3914 if (APStepVal.getBitWidth() > 64) 3915 return nullptr; 3916 3917 int64_t StepVal = APStepVal.getSExtValue(); 3918 if (PtrAccessSize != StepVal) 3919 return nullptr; 3920 V = M->getOperand(1); 3921 } 3922 } 3923 3924 // Strip off casts. 3925 Type *StripedOffRecurrenceCast = nullptr; 3926 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) { 3927 StripedOffRecurrenceCast = C->getType(); 3928 V = C->getOperand(); 3929 } 3930 3931 // Look for the loop invariant symbolic value. 3932 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V); 3933 if (!U) 3934 return nullptr; 3935 3936 Value *Stride = U->getValue(); 3937 if (!Lp->isLoopInvariant(Stride)) 3938 return nullptr; 3939 3940 // If we have stripped off the recurrence cast we have to make sure that we 3941 // return the value that is used in this loop so that we can replace it later. 3942 if (StripedOffRecurrenceCast) 3943 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast); 3944 3945 return Stride; 3946} 3947 3948void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) { 3949 Value *Ptr = nullptr; 3950 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess)) 3951 Ptr = LI->getPointerOperand(); 3952 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess)) 3953 Ptr = SI->getPointerOperand(); 3954 else 3955 return; 3956 3957 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop); 3958 if (!Stride) 3959 return; 3960 3961 DEBUG(dbgs() << "LV: Found a strided access that we can version"); 3962 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n"); 3963 Strides[Ptr] = Stride; 3964 StrideSet.insert(Stride); 3965} 3966 3967void LoopVectorizationLegality::collectLoopUniforms() { 3968 // We now know that the loop is vectorizable! 3969 // Collect variables that will remain uniform after vectorization. 3970 std::vector<Value*> Worklist; 3971 BasicBlock *Latch = TheLoop->getLoopLatch(); 3972 3973 // Start with the conditional branch and walk up the block. 3974 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 3975 3976 // Also add all consecutive pointer values; these values will be uniform 3977 // after vectorization (and subsequent cleanup) and, until revectorization is 3978 // supported, all dependencies must also be uniform. 3979 for (Loop::block_iterator B = TheLoop->block_begin(), 3980 BE = TheLoop->block_end(); B != BE; ++B) 3981 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end(); 3982 I != IE; ++I) 3983 if (I->getType()->isPointerTy() && isConsecutivePtr(I)) 3984 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end()); 3985 3986 while (!Worklist.empty()) { 3987 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 3988 Worklist.pop_back(); 3989 3990 // Look at instructions inside this loop. 3991 // Stop when reaching PHI nodes. 3992 // TODO: we need to follow values all over the loop, not only in this block. 3993 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 3994 continue; 3995 3996 // This is a known uniform. 3997 Uniforms.insert(I); 3998 3999 // Insert all operands. 4000 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end()); 4001 } 4002} 4003 4004bool LoopVectorizationLegality::canVectorizeMemory() { 4005 LAI = &LAA->getInfo(TheLoop, Strides); 4006 auto &OptionalReport = LAI->getReport(); 4007 if (OptionalReport) 4008 emitAnalysis(VectorizationReport(*OptionalReport)); 4009 if (!LAI->canVectorizeMemory()) 4010 return false; 4011 4012 if (LAI->hasStoreToLoopInvariantAddress()) { 4013 emitAnalysis( 4014 VectorizationReport() 4015 << "write to a loop invariant address could not be vectorized"); 4016 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 4017 return false; 4018 } 4019 4020 if (LAI->getNumRuntimePointerChecks() > 4021 VectorizerParams::RuntimeMemoryCheckThreshold) { 4022 emitAnalysis(VectorizationReport() 4023 << LAI->getNumRuntimePointerChecks() << " exceeds limit of " 4024 << VectorizerParams::RuntimeMemoryCheckThreshold 4025 << " dependent memory operations checked at runtime"); 4026 DEBUG(dbgs() << "LV: Too many memory checks needed.\n"); 4027 return false; 4028 } 4029 return true; 4030} 4031 4032static bool hasMultipleUsesOf(Instruction *I, 4033 SmallPtrSetImpl<Instruction *> &Insts) { 4034 unsigned NumUses = 0; 4035 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) { 4036 if (Insts.count(dyn_cast<Instruction>(*Use))) 4037 ++NumUses; 4038 if (NumUses > 1) 4039 return true; 4040 } 4041 4042 return false; 4043} 4044 4045static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) { 4046 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) 4047 if (!Set.count(dyn_cast<Instruction>(*Use))) 4048 return false; 4049 return true; 4050} 4051 4052bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, 4053 ReductionKind Kind) { 4054 if (Phi->getNumIncomingValues() != 2) 4055 return false; 4056 4057 // Reduction variables are only found in the loop header block. 4058 if (Phi->getParent() != TheLoop->getHeader()) 4059 return false; 4060 4061 // Obtain the reduction start value from the value that comes from the loop 4062 // preheader. 4063 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 4064 4065 // ExitInstruction is the single value which is used outside the loop. 4066 // We only allow for a single reduction value to be used outside the loop. 4067 // This includes users of the reduction, variables (which form a cycle 4068 // which ends in the phi node). 4069 Instruction *ExitInstruction = nullptr; 4070 // Indicates that we found a reduction operation in our scan. 4071 bool FoundReduxOp = false; 4072 4073 // We start with the PHI node and scan for all of the users of this 4074 // instruction. All users must be instructions that can be used as reduction 4075 // variables (such as ADD). We must have a single out-of-block user. The cycle 4076 // must include the original PHI. 4077 bool FoundStartPHI = false; 4078 4079 // To recognize min/max patterns formed by a icmp select sequence, we store 4080 // the number of instruction we saw from the recognized min/max pattern, 4081 // to make sure we only see exactly the two instructions. 4082 unsigned NumCmpSelectPatternInst = 0; 4083 ReductionInstDesc ReduxDesc(false, nullptr); 4084 4085 SmallPtrSet<Instruction *, 8> VisitedInsts; 4086 SmallVector<Instruction *, 8> Worklist; 4087 Worklist.push_back(Phi); 4088 VisitedInsts.insert(Phi); 4089 4090 // A value in the reduction can be used: 4091 // - By the reduction: 4092 // - Reduction operation: 4093 // - One use of reduction value (safe). 4094 // - Multiple use of reduction value (not safe). 4095 // - PHI: 4096 // - All uses of the PHI must be the reduction (safe). 4097 // - Otherwise, not safe. 4098 // - By one instruction outside of the loop (safe). 4099 // - By further instructions outside of the loop (not safe). 4100 // - By an instruction that is not part of the reduction (not safe). 4101 // This is either: 4102 // * An instruction type other than PHI or the reduction operation. 4103 // * A PHI in the header other than the initial PHI. 4104 while (!Worklist.empty()) { 4105 Instruction *Cur = Worklist.back(); 4106 Worklist.pop_back(); 4107 4108 // No Users. 4109 // If the instruction has no users then this is a broken chain and can't be 4110 // a reduction variable. 4111 if (Cur->use_empty()) 4112 return false; 4113 4114 bool IsAPhi = isa<PHINode>(Cur); 4115 4116 // A header PHI use other than the original PHI. 4117 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent()) 4118 return false; 4119 4120 // Reductions of instructions such as Div, and Sub is only possible if the 4121 // LHS is the reduction variable. 4122 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) && 4123 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) && 4124 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0)))) 4125 return false; 4126 4127 // Any reduction instruction must be of one of the allowed kinds. 4128 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc); 4129 if (!ReduxDesc.IsReduction) 4130 return false; 4131 4132 // A reduction operation must only have one use of the reduction value. 4133 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax && 4134 hasMultipleUsesOf(Cur, VisitedInsts)) 4135 return false; 4136 4137 // All inputs to a PHI node must be a reduction value. 4138 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts)) 4139 return false; 4140 4141 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) || 4142 isa<SelectInst>(Cur))) 4143 ++NumCmpSelectPatternInst; 4144 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) || 4145 isa<SelectInst>(Cur))) 4146 ++NumCmpSelectPatternInst; 4147 4148 // Check whether we found a reduction operator. 4149 FoundReduxOp |= !IsAPhi; 4150 4151 // Process users of current instruction. Push non-PHI nodes after PHI nodes 4152 // onto the stack. This way we are going to have seen all inputs to PHI 4153 // nodes once we get to them. 4154 SmallVector<Instruction *, 8> NonPHIs; 4155 SmallVector<Instruction *, 8> PHIs; 4156 for (User *U : Cur->users()) { 4157 Instruction *UI = cast<Instruction>(U); 4158 4159 // Check if we found the exit user. 4160 BasicBlock *Parent = UI->getParent(); 4161 if (!TheLoop->contains(Parent)) { 4162 // Exit if you find multiple outside users or if the header phi node is 4163 // being used. In this case the user uses the value of the previous 4164 // iteration, in which case we would loose "VF-1" iterations of the 4165 // reduction operation if we vectorize. 4166 if (ExitInstruction != nullptr || Cur == Phi) 4167 return false; 4168 4169 // The instruction used by an outside user must be the last instruction 4170 // before we feed back to the reduction phi. Otherwise, we loose VF-1 4171 // operations on the value. 4172 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end()) 4173 return false; 4174 4175 ExitInstruction = Cur; 4176 continue; 4177 } 4178 4179 // Process instructions only once (termination). Each reduction cycle 4180 // value must only be used once, except by phi nodes and min/max 4181 // reductions which are represented as a cmp followed by a select. 4182 ReductionInstDesc IgnoredVal(false, nullptr); 4183 if (VisitedInsts.insert(UI).second) { 4184 if (isa<PHINode>(UI)) 4185 PHIs.push_back(UI); 4186 else 4187 NonPHIs.push_back(UI); 4188 } else if (!isa<PHINode>(UI) && 4189 ((!isa<FCmpInst>(UI) && 4190 !isa<ICmpInst>(UI) && 4191 !isa<SelectInst>(UI)) || 4192 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction)) 4193 return false; 4194 4195 // Remember that we completed the cycle. 4196 if (UI == Phi) 4197 FoundStartPHI = true; 4198 } 4199 Worklist.append(PHIs.begin(), PHIs.end()); 4200 Worklist.append(NonPHIs.begin(), NonPHIs.end()); 4201 } 4202 4203 // This means we have seen one but not the other instruction of the 4204 // pattern or more than just a select and cmp. 4205 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) && 4206 NumCmpSelectPatternInst != 2) 4207 return false; 4208 4209 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction) 4210 return false; 4211 4212 // We found a reduction var if we have reached the original phi node and we 4213 // only have a single instruction with out-of-loop users. 4214 4215 // This instruction is allowed to have out-of-loop users. 4216 AllowedExit.insert(ExitInstruction); 4217 4218 // Save the description of this reduction variable. 4219 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind, 4220 ReduxDesc.MinMaxKind); 4221 Reductions[Phi] = RD; 4222 // We've ended the cycle. This is a reduction variable if we have an 4223 // outside user and it has a binary op. 4224 4225 return true; 4226} 4227 4228/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 4229/// pattern corresponding to a min(X, Y) or max(X, Y). 4230LoopVectorizationLegality::ReductionInstDesc 4231LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, 4232 ReductionInstDesc &Prev) { 4233 4234 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) && 4235 "Expect a select instruction"); 4236 Instruction *Cmp = nullptr; 4237 SelectInst *Select = nullptr; 4238 4239 // We must handle the select(cmp()) as a single instruction. Advance to the 4240 // select. 4241 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) { 4242 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin()))) 4243 return ReductionInstDesc(false, I); 4244 return ReductionInstDesc(Select, Prev.MinMaxKind); 4245 } 4246 4247 // Only handle single use cases for now. 4248 if (!(Select = dyn_cast<SelectInst>(I))) 4249 return ReductionInstDesc(false, I); 4250 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) && 4251 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0)))) 4252 return ReductionInstDesc(false, I); 4253 if (!Cmp->hasOneUse()) 4254 return ReductionInstDesc(false, I); 4255 4256 Value *CmpLeft; 4257 Value *CmpRight; 4258 4259 // Look for a min/max pattern. 4260 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4261 return ReductionInstDesc(Select, MRK_UIntMin); 4262 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4263 return ReductionInstDesc(Select, MRK_UIntMax); 4264 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4265 return ReductionInstDesc(Select, MRK_SIntMax); 4266 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4267 return ReductionInstDesc(Select, MRK_SIntMin); 4268 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4269 return ReductionInstDesc(Select, MRK_FloatMin); 4270 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4271 return ReductionInstDesc(Select, MRK_FloatMax); 4272 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4273 return ReductionInstDesc(Select, MRK_FloatMin); 4274 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4275 return ReductionInstDesc(Select, MRK_FloatMax); 4276 4277 return ReductionInstDesc(false, I); 4278} 4279 4280LoopVectorizationLegality::ReductionInstDesc 4281LoopVectorizationLegality::isReductionInstr(Instruction *I, 4282 ReductionKind Kind, 4283 ReductionInstDesc &Prev) { 4284 bool FP = I->getType()->isFloatingPointTy(); 4285 bool FastMath = FP && I->hasUnsafeAlgebra(); 4286 switch (I->getOpcode()) { 4287 default: 4288 return ReductionInstDesc(false, I); 4289 case Instruction::PHI: 4290 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd && 4291 Kind != RK_FloatMinMax)) 4292 return ReductionInstDesc(false, I); 4293 return ReductionInstDesc(I, Prev.MinMaxKind); 4294 case Instruction::Sub: 4295 case Instruction::Add: 4296 return ReductionInstDesc(Kind == RK_IntegerAdd, I); 4297 case Instruction::Mul: 4298 return ReductionInstDesc(Kind == RK_IntegerMult, I); 4299 case Instruction::And: 4300 return ReductionInstDesc(Kind == RK_IntegerAnd, I); 4301 case Instruction::Or: 4302 return ReductionInstDesc(Kind == RK_IntegerOr, I); 4303 case Instruction::Xor: 4304 return ReductionInstDesc(Kind == RK_IntegerXor, I); 4305 case Instruction::FMul: 4306 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I); 4307 case Instruction::FSub: 4308 case Instruction::FAdd: 4309 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I); 4310 case Instruction::FCmp: 4311 case Instruction::ICmp: 4312 case Instruction::Select: 4313 if (Kind != RK_IntegerMinMax && 4314 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax)) 4315 return ReductionInstDesc(false, I); 4316 return isMinMaxSelectCmpPattern(I, Prev); 4317 } 4318} 4319 4320bool llvm::isInductionPHI(PHINode *Phi, ScalarEvolution *SE, 4321 ConstantInt *&StepValue) { 4322 Type *PhiTy = Phi->getType(); 4323 // We only handle integer and pointer inductions variables. 4324 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 4325 return false; 4326 4327 // Check that the PHI is consecutive. 4328 const SCEV *PhiScev = SE->getSCEV(Phi); 4329 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 4330 if (!AR) { 4331 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 4332 return false; 4333 } 4334 4335 const SCEV *Step = AR->getStepRecurrence(*SE); 4336 // Calculate the pointer stride and check if it is consecutive. 4337 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 4338 if (!C) 4339 return false; 4340 4341 ConstantInt *CV = C->getValue(); 4342 if (PhiTy->isIntegerTy()) { 4343 StepValue = CV; 4344 return true; 4345 } 4346 4347 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 4348 Type *PointerElementType = PhiTy->getPointerElementType(); 4349 // The pointer stride cannot be determined if the pointer element type is not 4350 // sized. 4351 if (!PointerElementType->isSized()) 4352 return false; 4353 4354 const DataLayout &DL = Phi->getModule()->getDataLayout(); 4355 int64_t Size = static_cast<int64_t>(DL.getTypeAllocSize(PointerElementType)); 4356 int64_t CVSize = CV->getSExtValue(); 4357 if (CVSize % Size) 4358 return false; 4359 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size); 4360 return true; 4361} 4362 4363LoopVectorizationLegality::InductionKind 4364LoopVectorizationLegality::isInductionVariable(PHINode *Phi, 4365 ConstantInt *&StepValue) { 4366 if (!isInductionPHI(Phi, SE, StepValue)) 4367 return IK_NoInduction; 4368 4369 Type *PhiTy = Phi->getType(); 4370 // Found an Integer induction variable. 4371 if (PhiTy->isIntegerTy()) 4372 return IK_IntInduction; 4373 // Found an Pointer induction variable. 4374 return IK_PtrInduction; 4375} 4376 4377bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 4378 Value *In0 = const_cast<Value*>(V); 4379 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 4380 if (!PN) 4381 return false; 4382 4383 return Inductions.count(PN); 4384} 4385 4386bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 4387 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT); 4388} 4389 4390bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB, 4391 SmallPtrSetImpl<Value *> &SafePtrs) { 4392 4393 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4394 // Check that we don't have a constant expression that can trap as operand. 4395 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end(); 4396 OI != OE; ++OI) { 4397 if (Constant *C = dyn_cast<Constant>(*OI)) 4398 if (C->canTrap()) 4399 return false; 4400 } 4401 // We might be able to hoist the load. 4402 if (it->mayReadFromMemory()) { 4403 LoadInst *LI = dyn_cast<LoadInst>(it); 4404 if (!LI) 4405 return false; 4406 if (!SafePtrs.count(LI->getPointerOperand())) { 4407 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) { 4408 MaskedOp.insert(LI); 4409 continue; 4410 } 4411 return false; 4412 } 4413 } 4414 4415 // We don't predicate stores at the moment. 4416 if (it->mayWriteToMemory()) { 4417 StoreInst *SI = dyn_cast<StoreInst>(it); 4418 // We only support predication of stores in basic blocks with one 4419 // predecessor. 4420 if (!SI) 4421 return false; 4422 4423 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0); 4424 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor(); 4425 4426 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr || 4427 !isSinglePredecessor) { 4428 // Build a masked store if it is legal for the target, otherwise scalarize 4429 // the block. 4430 bool isLegalMaskedOp = 4431 isLegalMaskedStore(SI->getValueOperand()->getType(), 4432 SI->getPointerOperand()); 4433 if (isLegalMaskedOp) { 4434 --NumPredStores; 4435 MaskedOp.insert(SI); 4436 continue; 4437 } 4438 return false; 4439 } 4440 } 4441 if (it->mayThrow()) 4442 return false; 4443 4444 // The instructions below can trap. 4445 switch (it->getOpcode()) { 4446 default: continue; 4447 case Instruction::UDiv: 4448 case Instruction::SDiv: 4449 case Instruction::URem: 4450 case Instruction::SRem: 4451 return false; 4452 } 4453 } 4454 4455 return true; 4456} 4457 4458LoopVectorizationCostModel::VectorizationFactor 4459LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) { 4460 // Width 1 means no vectorize 4461 VectorizationFactor Factor = { 1U, 0U }; 4462 if (OptForSize && Legal->getRuntimePointerCheck()->Need) { 4463 emitAnalysis(VectorizationReport() << 4464 "runtime pointer checks needed. Enable vectorization of this " 4465 "loop with '#pragma clang loop vectorize(enable)' when " 4466 "compiling with -Os"); 4467 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); 4468 return Factor; 4469 } 4470 4471 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) { 4472 emitAnalysis(VectorizationReport() << 4473 "store that is conditionally executed prevents vectorization"); 4474 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n"); 4475 return Factor; 4476 } 4477 4478 // Find the trip count. 4479 unsigned TC = SE->getSmallConstantTripCount(TheLoop); 4480 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n'); 4481 4482 unsigned WidestType = getWidestType(); 4483 unsigned WidestRegister = TTI.getRegisterBitWidth(true); 4484 unsigned MaxSafeDepDist = -1U; 4485 if (Legal->getMaxSafeDepDistBytes() != -1U) 4486 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8; 4487 WidestRegister = ((WidestRegister < MaxSafeDepDist) ? 4488 WidestRegister : MaxSafeDepDist); 4489 unsigned MaxVectorSize = WidestRegister / WidestType; 4490 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n"); 4491 DEBUG(dbgs() << "LV: The Widest register is: " 4492 << WidestRegister << " bits.\n"); 4493 4494 if (MaxVectorSize == 0) { 4495 DEBUG(dbgs() << "LV: The target has no vector registers.\n"); 4496 MaxVectorSize = 1; 4497 } 4498 4499 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements" 4500 " into one vector!"); 4501 4502 unsigned VF = MaxVectorSize; 4503 4504 // If we optimize the program for size, avoid creating the tail loop. 4505 if (OptForSize) { 4506 // If we are unable to calculate the trip count then don't try to vectorize. 4507 if (TC < 2) { 4508 emitAnalysis 4509 (VectorizationReport() << 4510 "unable to calculate the loop count due to complex control flow"); 4511 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 4512 return Factor; 4513 } 4514 4515 // Find the maximum SIMD width that can fit within the trip count. 4516 VF = TC % MaxVectorSize; 4517 4518 if (VF == 0) 4519 VF = MaxVectorSize; 4520 4521 // If the trip count that we found modulo the vectorization factor is not 4522 // zero then we require a tail. 4523 if (VF < 2) { 4524 emitAnalysis(VectorizationReport() << 4525 "cannot optimize for size and vectorize at the " 4526 "same time. Enable vectorization of this loop " 4527 "with '#pragma clang loop vectorize(enable)' " 4528 "when compiling with -Os"); 4529 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 4530 return Factor; 4531 } 4532 } 4533 4534 int UserVF = Hints->getWidth(); 4535 if (UserVF != 0) { 4536 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 4537 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n"); 4538 4539 Factor.Width = UserVF; 4540 return Factor; 4541 } 4542 4543 float Cost = expectedCost(1); 4544#ifndef NDEBUG 4545 const float ScalarCost = Cost; 4546#endif /* NDEBUG */ 4547 unsigned Width = 1; 4548 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n"); 4549 4550 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; 4551 // Ignore scalar width, because the user explicitly wants vectorization. 4552 if (ForceVectorization && VF > 1) { 4553 Width = 2; 4554 Cost = expectedCost(Width) / (float)Width; 4555 } 4556 4557 for (unsigned i=2; i <= VF; i*=2) { 4558 // Notice that the vector loop needs to be executed less times, so 4559 // we need to divide the cost of the vector loops by the width of 4560 // the vector elements. 4561 float VectorCost = expectedCost(i) / (float)i; 4562 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " << 4563 (int)VectorCost << ".\n"); 4564 if (VectorCost < Cost) { 4565 Cost = VectorCost; 4566 Width = i; 4567 } 4568 } 4569 4570 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs() 4571 << "LV: Vectorization seems to be not beneficial, " 4572 << "but was forced by a user.\n"); 4573 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n"); 4574 Factor.Width = Width; 4575 Factor.Cost = Width * Cost; 4576 return Factor; 4577} 4578 4579unsigned LoopVectorizationCostModel::getWidestType() { 4580 unsigned MaxWidth = 8; 4581 const DataLayout &DL = TheFunction->getParent()->getDataLayout(); 4582 4583 // For each block. 4584 for (Loop::block_iterator bb = TheLoop->block_begin(), 4585 be = TheLoop->block_end(); bb != be; ++bb) { 4586 BasicBlock *BB = *bb; 4587 4588 // For each instruction in the loop. 4589 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4590 Type *T = it->getType(); 4591 4592 // Ignore ephemeral values. 4593 if (EphValues.count(it)) 4594 continue; 4595 4596 // Only examine Loads, Stores and PHINodes. 4597 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it)) 4598 continue; 4599 4600 // Examine PHI nodes that are reduction variables. 4601 if (PHINode *PN = dyn_cast<PHINode>(it)) 4602 if (!Legal->getReductionVars()->count(PN)) 4603 continue; 4604 4605 // Examine the stored values. 4606 if (StoreInst *ST = dyn_cast<StoreInst>(it)) 4607 T = ST->getValueOperand()->getType(); 4608 4609 // Ignore loaded pointer types and stored pointer types that are not 4610 // consecutive. However, we do want to take consecutive stores/loads of 4611 // pointer vectors into account. 4612 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it)) 4613 continue; 4614 4615 MaxWidth = std::max(MaxWidth, 4616 (unsigned)DL.getTypeSizeInBits(T->getScalarType())); 4617 } 4618 } 4619 4620 return MaxWidth; 4621} 4622 4623unsigned 4624LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, 4625 unsigned VF, 4626 unsigned LoopCost) { 4627 4628 // -- The unroll heuristics -- 4629 // We unroll the loop in order to expose ILP and reduce the loop overhead. 4630 // There are many micro-architectural considerations that we can't predict 4631 // at this level. For example, frontend pressure (on decode or fetch) due to 4632 // code size, or the number and capabilities of the execution ports. 4633 // 4634 // We use the following heuristics to select the unroll factor: 4635 // 1. If the code has reductions, then we unroll in order to break the cross 4636 // iteration dependency. 4637 // 2. If the loop is really small, then we unroll in order to reduce the loop 4638 // overhead. 4639 // 3. We don't unroll if we think that we will spill registers to memory due 4640 // to the increased register pressure. 4641 4642 // Use the user preference, unless 'auto' is selected. 4643 int UserUF = Hints->getInterleave(); 4644 if (UserUF != 0) 4645 return UserUF; 4646 4647 // When we optimize for size, we don't unroll. 4648 if (OptForSize) 4649 return 1; 4650 4651 // We used the distance for the unroll factor. 4652 if (Legal->getMaxSafeDepDistBytes() != -1U) 4653 return 1; 4654 4655 // Do not unroll loops with a relatively small trip count. 4656 unsigned TC = SE->getSmallConstantTripCount(TheLoop); 4657 if (TC > 1 && TC < TinyTripCountUnrollThreshold) 4658 return 1; 4659 4660 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1); 4661 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters << 4662 " registers\n"); 4663 4664 if (VF == 1) { 4665 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) 4666 TargetNumRegisters = ForceTargetNumScalarRegs; 4667 } else { 4668 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) 4669 TargetNumRegisters = ForceTargetNumVectorRegs; 4670 } 4671 4672 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); 4673 // We divide by these constants so assume that we have at least one 4674 // instruction that uses at least one register. 4675 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 4676 R.NumInstructions = std::max(R.NumInstructions, 1U); 4677 4678 // We calculate the unroll factor using the following formula. 4679 // Subtract the number of loop invariants from the number of available 4680 // registers. These registers are used by all of the unrolled instances. 4681 // Next, divide the remaining registers by the number of registers that is 4682 // required by the loop, in order to estimate how many parallel instances 4683 // fit without causing spills. All of this is rounded down if necessary to be 4684 // a power of two. We want power of two unroll factors to simplify any 4685 // addressing operations or alignment considerations. 4686 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) / 4687 R.MaxLocalUsers); 4688 4689 // Don't count the induction variable as unrolled. 4690 if (EnableIndVarRegisterHeur) 4691 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) / 4692 std::max(1U, (R.MaxLocalUsers - 1))); 4693 4694 // Clamp the unroll factor ranges to reasonable factors. 4695 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor(); 4696 4697 // Check if the user has overridden the unroll max. 4698 if (VF == 1) { 4699 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) 4700 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor; 4701 } else { 4702 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) 4703 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor; 4704 } 4705 4706 // If we did not calculate the cost for VF (because the user selected the VF) 4707 // then we calculate the cost of VF here. 4708 if (LoopCost == 0) 4709 LoopCost = expectedCost(VF); 4710 4711 // Clamp the calculated UF to be between the 1 and the max unroll factor 4712 // that the target allows. 4713 if (UF > MaxInterleaveSize) 4714 UF = MaxInterleaveSize; 4715 else if (UF < 1) 4716 UF = 1; 4717 4718 // Unroll if we vectorized this loop and there is a reduction that could 4719 // benefit from unrolling. 4720 if (VF > 1 && Legal->getReductionVars()->size()) { 4721 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n"); 4722 return UF; 4723 } 4724 4725 // Note that if we've already vectorized the loop we will have done the 4726 // runtime check and so unrolling won't require further checks. 4727 bool UnrollingRequiresRuntimePointerCheck = 4728 (VF == 1 && Legal->getRuntimePointerCheck()->Need); 4729 4730 // We want to unroll small loops in order to reduce the loop overhead and 4731 // potentially expose ILP opportunities. 4732 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n'); 4733 if (!UnrollingRequiresRuntimePointerCheck && 4734 LoopCost < SmallLoopCost) { 4735 // We assume that the cost overhead is 1 and we use the cost model 4736 // to estimate the cost of the loop and unroll until the cost of the 4737 // loop overhead is about 5% of the cost of the loop. 4738 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost)); 4739 4740 // Unroll until store/load ports (estimated by max unroll factor) are 4741 // saturated. 4742 unsigned NumStores = Legal->getNumStores(); 4743 unsigned NumLoads = Legal->getNumLoads(); 4744 unsigned StoresUF = UF / (NumStores ? NumStores : 1); 4745 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1); 4746 4747 // If we have a scalar reduction (vector reductions are already dealt with 4748 // by this point), we can increase the critical path length if the loop 4749 // we're unrolling is inside another loop. Limit, by default to 2, so the 4750 // critical path only gets increased by one reduction operation. 4751 if (Legal->getReductionVars()->size() && 4752 TheLoop->getLoopDepth() > 1) { 4753 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF); 4754 SmallUF = std::min(SmallUF, F); 4755 StoresUF = std::min(StoresUF, F); 4756 LoadsUF = std::min(LoadsUF, F); 4757 } 4758 4759 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) { 4760 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n"); 4761 return std::max(StoresUF, LoadsUF); 4762 } 4763 4764 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n"); 4765 return SmallUF; 4766 } 4767 4768 // Unroll if this is a large loop (small loops are already dealt with by this 4769 // point) that could benefit from interleaved unrolling. 4770 bool HasReductions = (Legal->getReductionVars()->size() > 0); 4771 if (TTI.enableAggressiveInterleaving(HasReductions)) { 4772 DEBUG(dbgs() << "LV: Unrolling to expose ILP.\n"); 4773 return UF; 4774 } 4775 4776 DEBUG(dbgs() << "LV: Not Unrolling.\n"); 4777 return 1; 4778} 4779 4780LoopVectorizationCostModel::RegisterUsage 4781LoopVectorizationCostModel::calculateRegisterUsage() { 4782 // This function calculates the register usage by measuring the highest number 4783 // of values that are alive at a single location. Obviously, this is a very 4784 // rough estimation. We scan the loop in a topological order in order and 4785 // assign a number to each instruction. We use RPO to ensure that defs are 4786 // met before their users. We assume that each instruction that has in-loop 4787 // users starts an interval. We record every time that an in-loop value is 4788 // used, so we have a list of the first and last occurrences of each 4789 // instruction. Next, we transpose this data structure into a multi map that 4790 // holds the list of intervals that *end* at a specific location. This multi 4791 // map allows us to perform a linear search. We scan the instructions linearly 4792 // and record each time that a new interval starts, by placing it in a set. 4793 // If we find this value in the multi-map then we remove it from the set. 4794 // The max register usage is the maximum size of the set. 4795 // We also search for instructions that are defined outside the loop, but are 4796 // used inside the loop. We need this number separately from the max-interval 4797 // usage number because when we unroll, loop-invariant values do not take 4798 // more register. 4799 LoopBlocksDFS DFS(TheLoop); 4800 DFS.perform(LI); 4801 4802 RegisterUsage R; 4803 R.NumInstructions = 0; 4804 4805 // Each 'key' in the map opens a new interval. The values 4806 // of the map are the index of the 'last seen' usage of the 4807 // instruction that is the key. 4808 typedef DenseMap<Instruction*, unsigned> IntervalMap; 4809 // Maps instruction to its index. 4810 DenseMap<unsigned, Instruction*> IdxToInstr; 4811 // Marks the end of each interval. 4812 IntervalMap EndPoint; 4813 // Saves the list of instruction indices that are used in the loop. 4814 SmallSet<Instruction*, 8> Ends; 4815 // Saves the list of values that are used in the loop but are 4816 // defined outside the loop, such as arguments and constants. 4817 SmallPtrSet<Value*, 8> LoopInvariants; 4818 4819 unsigned Index = 0; 4820 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 4821 be = DFS.endRPO(); bb != be; ++bb) { 4822 R.NumInstructions += (*bb)->size(); 4823 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 4824 ++it) { 4825 Instruction *I = it; 4826 IdxToInstr[Index++] = I; 4827 4828 // Save the end location of each USE. 4829 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 4830 Value *U = I->getOperand(i); 4831 Instruction *Instr = dyn_cast<Instruction>(U); 4832 4833 // Ignore non-instruction values such as arguments, constants, etc. 4834 if (!Instr) continue; 4835 4836 // If this instruction is outside the loop then record it and continue. 4837 if (!TheLoop->contains(Instr)) { 4838 LoopInvariants.insert(Instr); 4839 continue; 4840 } 4841 4842 // Overwrite previous end points. 4843 EndPoint[Instr] = Index; 4844 Ends.insert(Instr); 4845 } 4846 } 4847 } 4848 4849 // Saves the list of intervals that end with the index in 'key'. 4850 typedef SmallVector<Instruction*, 2> InstrList; 4851 DenseMap<unsigned, InstrList> TransposeEnds; 4852 4853 // Transpose the EndPoints to a list of values that end at each index. 4854 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 4855 it != e; ++it) 4856 TransposeEnds[it->second].push_back(it->first); 4857 4858 SmallSet<Instruction*, 8> OpenIntervals; 4859 unsigned MaxUsage = 0; 4860 4861 4862 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 4863 for (unsigned int i = 0; i < Index; ++i) { 4864 Instruction *I = IdxToInstr[i]; 4865 // Ignore instructions that are never used within the loop. 4866 if (!Ends.count(I)) continue; 4867 4868 // Ignore ephemeral values. 4869 if (EphValues.count(I)) 4870 continue; 4871 4872 // Remove all of the instructions that end at this location. 4873 InstrList &List = TransposeEnds[i]; 4874 for (unsigned int j=0, e = List.size(); j < e; ++j) 4875 OpenIntervals.erase(List[j]); 4876 4877 // Count the number of live interals. 4878 MaxUsage = std::max(MaxUsage, OpenIntervals.size()); 4879 4880 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << 4881 OpenIntervals.size() << '\n'); 4882 4883 // Add the current instruction to the list of open intervals. 4884 OpenIntervals.insert(I); 4885 } 4886 4887 unsigned Invariant = LoopInvariants.size(); 4888 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n'); 4889 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n'); 4890 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n'); 4891 4892 R.LoopInvariantRegs = Invariant; 4893 R.MaxLocalUsers = MaxUsage; 4894 return R; 4895} 4896 4897unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 4898 unsigned Cost = 0; 4899 4900 // For each block. 4901 for (Loop::block_iterator bb = TheLoop->block_begin(), 4902 be = TheLoop->block_end(); bb != be; ++bb) { 4903 unsigned BlockCost = 0; 4904 BasicBlock *BB = *bb; 4905 4906 // For each instruction in the old loop. 4907 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4908 // Skip dbg intrinsics. 4909 if (isa<DbgInfoIntrinsic>(it)) 4910 continue; 4911 4912 // Ignore ephemeral values. 4913 if (EphValues.count(it)) 4914 continue; 4915 4916 unsigned C = getInstructionCost(it, VF); 4917 4918 // Check if we should override the cost. 4919 if (ForceTargetInstructionCost.getNumOccurrences() > 0) 4920 C = ForceTargetInstructionCost; 4921 4922 BlockCost += C; 4923 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " << 4924 VF << " For instruction: " << *it << '\n'); 4925 } 4926 4927 // We assume that if-converted blocks have a 50% chance of being executed. 4928 // When the code is scalar then some of the blocks are avoided due to CF. 4929 // When the code is vectorized we execute all code paths. 4930 if (VF == 1 && Legal->blockNeedsPredication(*bb)) 4931 BlockCost /= 2; 4932 4933 Cost += BlockCost; 4934 } 4935 4936 return Cost; 4937} 4938 4939/// \brief Check whether the address computation for a non-consecutive memory 4940/// access looks like an unlikely candidate for being merged into the indexing 4941/// mode. 4942/// 4943/// We look for a GEP which has one index that is an induction variable and all 4944/// other indices are loop invariant. If the stride of this access is also 4945/// within a small bound we decide that this address computation can likely be 4946/// merged into the addressing mode. 4947/// In all other cases, we identify the address computation as complex. 4948static bool isLikelyComplexAddressComputation(Value *Ptr, 4949 LoopVectorizationLegality *Legal, 4950 ScalarEvolution *SE, 4951 const Loop *TheLoop) { 4952 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 4953 if (!Gep) 4954 return true; 4955 4956 // We are looking for a gep with all loop invariant indices except for one 4957 // which should be an induction variable. 4958 unsigned NumOperands = Gep->getNumOperands(); 4959 for (unsigned i = 1; i < NumOperands; ++i) { 4960 Value *Opd = Gep->getOperand(i); 4961 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && 4962 !Legal->isInductionVariable(Opd)) 4963 return true; 4964 } 4965 4966 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step 4967 // can likely be merged into the address computation. 4968 unsigned MaxMergeDistance = 64; 4969 4970 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr)); 4971 if (!AddRec) 4972 return true; 4973 4974 // Check the step is constant. 4975 const SCEV *Step = AddRec->getStepRecurrence(*SE); 4976 // Calculate the pointer stride and check if it is consecutive. 4977 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 4978 if (!C) 4979 return true; 4980 4981 const APInt &APStepVal = C->getValue()->getValue(); 4982 4983 // Huge step value - give up. 4984 if (APStepVal.getBitWidth() > 64) 4985 return true; 4986 4987 int64_t StepVal = APStepVal.getSExtValue(); 4988 4989 return StepVal > MaxMergeDistance; 4990} 4991 4992static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { 4993 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1))) 4994 return true; 4995 return false; 4996} 4997 4998unsigned 4999LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 5000 // If we know that this instruction will remain uniform, check the cost of 5001 // the scalar version. 5002 if (Legal->isUniformAfterVectorization(I)) 5003 VF = 1; 5004 5005 Type *RetTy = I->getType(); 5006 Type *VectorTy = ToVectorTy(RetTy, VF); 5007 5008 // TODO: We need to estimate the cost of intrinsic calls. 5009 switch (I->getOpcode()) { 5010 case Instruction::GetElementPtr: 5011 // We mark this instruction as zero-cost because the cost of GEPs in 5012 // vectorized code depends on whether the corresponding memory instruction 5013 // is scalarized or not. Therefore, we handle GEPs with the memory 5014 // instruction cost. 5015 return 0; 5016 case Instruction::Br: { 5017 return TTI.getCFInstrCost(I->getOpcode()); 5018 } 5019 case Instruction::PHI: 5020 //TODO: IF-converted IFs become selects. 5021 return 0; 5022 case Instruction::Add: 5023 case Instruction::FAdd: 5024 case Instruction::Sub: 5025 case Instruction::FSub: 5026 case Instruction::Mul: 5027 case Instruction::FMul: 5028 case Instruction::UDiv: 5029 case Instruction::SDiv: 5030 case Instruction::FDiv: 5031 case Instruction::URem: 5032 case Instruction::SRem: 5033 case Instruction::FRem: 5034 case Instruction::Shl: 5035 case Instruction::LShr: 5036 case Instruction::AShr: 5037 case Instruction::And: 5038 case Instruction::Or: 5039 case Instruction::Xor: { 5040 // Since we will replace the stride by 1 the multiplication should go away. 5041 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) 5042 return 0; 5043 // Certain instructions can be cheaper to vectorize if they have a constant 5044 // second vector operand. One example of this are shifts on x86. 5045 TargetTransformInfo::OperandValueKind Op1VK = 5046 TargetTransformInfo::OK_AnyValue; 5047 TargetTransformInfo::OperandValueKind Op2VK = 5048 TargetTransformInfo::OK_AnyValue; 5049 TargetTransformInfo::OperandValueProperties Op1VP = 5050 TargetTransformInfo::OP_None; 5051 TargetTransformInfo::OperandValueProperties Op2VP = 5052 TargetTransformInfo::OP_None; 5053 Value *Op2 = I->getOperand(1); 5054 5055 // Check for a splat of a constant or for a non uniform vector of constants. 5056 if (isa<ConstantInt>(Op2)) { 5057 ConstantInt *CInt = cast<ConstantInt>(Op2); 5058 if (CInt && CInt->getValue().isPowerOf2()) 5059 Op2VP = TargetTransformInfo::OP_PowerOf2; 5060 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 5061 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) { 5062 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; 5063 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue(); 5064 if (SplatValue) { 5065 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue); 5066 if (CInt && CInt->getValue().isPowerOf2()) 5067 Op2VP = TargetTransformInfo::OP_PowerOf2; 5068 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 5069 } 5070 } 5071 5072 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK, 5073 Op1VP, Op2VP); 5074 } 5075 case Instruction::Select: { 5076 SelectInst *SI = cast<SelectInst>(I); 5077 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 5078 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 5079 Type *CondTy = SI->getCondition()->getType(); 5080 if (!ScalarCond) 5081 CondTy = VectorType::get(CondTy, VF); 5082 5083 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 5084 } 5085 case Instruction::ICmp: 5086 case Instruction::FCmp: { 5087 Type *ValTy = I->getOperand(0)->getType(); 5088 VectorTy = ToVectorTy(ValTy, VF); 5089 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); 5090 } 5091 case Instruction::Store: 5092 case Instruction::Load: { 5093 StoreInst *SI = dyn_cast<StoreInst>(I); 5094 LoadInst *LI = dyn_cast<LoadInst>(I); 5095 Type *ValTy = (SI ? SI->getValueOperand()->getType() : 5096 LI->getType()); 5097 VectorTy = ToVectorTy(ValTy, VF); 5098 5099 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); 5100 unsigned AS = SI ? SI->getPointerAddressSpace() : 5101 LI->getPointerAddressSpace(); 5102 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); 5103 // We add the cost of address computation here instead of with the gep 5104 // instruction because only here we know whether the operation is 5105 // scalarized. 5106 if (VF == 1) 5107 return TTI.getAddressComputationCost(VectorTy) + 5108 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 5109 5110 // Scalarized loads/stores. 5111 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 5112 bool Reverse = ConsecutiveStride < 0; 5113 const DataLayout &DL = I->getModule()->getDataLayout(); 5114 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy); 5115 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF; 5116 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { 5117 bool IsComplexComputation = 5118 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop); 5119 unsigned Cost = 0; 5120 // The cost of extracting from the value vector and pointer vector. 5121 Type *PtrTy = ToVectorTy(Ptr->getType(), VF); 5122 for (unsigned i = 0; i < VF; ++i) { 5123 // The cost of extracting the pointer operand. 5124 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); 5125 // In case of STORE, the cost of ExtractElement from the vector. 5126 // In case of LOAD, the cost of InsertElement into the returned 5127 // vector. 5128 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : 5129 Instruction::InsertElement, 5130 VectorTy, i); 5131 } 5132 5133 // The cost of the scalar loads/stores. 5134 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation); 5135 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), 5136 Alignment, AS); 5137 return Cost; 5138 } 5139 5140 // Wide load/stores. 5141 unsigned Cost = TTI.getAddressComputationCost(VectorTy); 5142 if (Legal->isMaskRequired(I)) 5143 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, 5144 AS); 5145 else 5146 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 5147 5148 if (Reverse) 5149 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, 5150 VectorTy, 0); 5151 return Cost; 5152 } 5153 case Instruction::ZExt: 5154 case Instruction::SExt: 5155 case Instruction::FPToUI: 5156 case Instruction::FPToSI: 5157 case Instruction::FPExt: 5158 case Instruction::PtrToInt: 5159 case Instruction::IntToPtr: 5160 case Instruction::SIToFP: 5161 case Instruction::UIToFP: 5162 case Instruction::Trunc: 5163 case Instruction::FPTrunc: 5164 case Instruction::BitCast: { 5165 // We optimize the truncation of induction variable. 5166 // The cost of these is the same as the scalar operation. 5167 if (I->getOpcode() == Instruction::Trunc && 5168 Legal->isInductionVariable(I->getOperand(0))) 5169 return TTI.getCastInstrCost(I->getOpcode(), I->getType(), 5170 I->getOperand(0)->getType()); 5171 5172 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); 5173 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 5174 } 5175 case Instruction::Call: { 5176 bool NeedToScalarize; 5177 CallInst *CI = cast<CallInst>(I); 5178 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize); 5179 if (getIntrinsicIDForCall(CI, TLI)) 5180 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI)); 5181 return CallCost; 5182 } 5183 default: { 5184 // We are scalarizing the instruction. Return the cost of the scalar 5185 // instruction, plus the cost of insert and extract into vector 5186 // elements, times the vector width. 5187 unsigned Cost = 0; 5188 5189 if (!RetTy->isVoidTy() && VF != 1) { 5190 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, 5191 VectorTy); 5192 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, 5193 VectorTy); 5194 5195 // The cost of inserting the results plus extracting each one of the 5196 // operands. 5197 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 5198 } 5199 5200 // The cost of executing VF copies of the scalar instruction. This opcode 5201 // is unknown. Assume that it is the same as 'mul'. 5202 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); 5203 return Cost; 5204 } 5205 }// end of switch. 5206} 5207 5208char LoopVectorize::ID = 0; 5209static const char lv_name[] = "Loop Vectorization"; 5210INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 5211INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 5212INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 5213INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 5214INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo) 5215INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 5216INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 5217INITIALIZE_PASS_DEPENDENCY(LCSSA) 5218INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 5219INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 5220INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis) 5221INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 5222 5223namespace llvm { 5224 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) { 5225 return new LoopVectorize(NoUnrolling, AlwaysVectorize); 5226 } 5227} 5228 5229bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { 5230 // Check for a store. 5231 if (StoreInst *ST = dyn_cast<StoreInst>(Inst)) 5232 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; 5233 5234 // Check for a load. 5235 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 5236 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; 5237 5238 return false; 5239} 5240 5241 5242void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr, 5243 bool IfPredicateStore) { 5244 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 5245 // Holds vector parameters or scalars, in case of uniform vals. 5246 SmallVector<VectorParts, 4> Params; 5247 5248 setDebugLocFromInst(Builder, Instr); 5249 5250 // Find all of the vectorized parameters. 5251 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5252 Value *SrcOp = Instr->getOperand(op); 5253 5254 // If we are accessing the old induction variable, use the new one. 5255 if (SrcOp == OldInduction) { 5256 Params.push_back(getVectorValue(SrcOp)); 5257 continue; 5258 } 5259 5260 // Try using previously calculated values. 5261 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 5262 5263 // If the src is an instruction that appeared earlier in the basic block 5264 // then it should already be vectorized. 5265 if (SrcInst && OrigLoop->contains(SrcInst)) { 5266 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 5267 // The parameter is a vector value from earlier. 5268 Params.push_back(WidenMap.get(SrcInst)); 5269 } else { 5270 // The parameter is a scalar from outside the loop. Maybe even a constant. 5271 VectorParts Scalars; 5272 Scalars.append(UF, SrcOp); 5273 Params.push_back(Scalars); 5274 } 5275 } 5276 5277 assert(Params.size() == Instr->getNumOperands() && 5278 "Invalid number of operands"); 5279 5280 // Does this instruction return a value ? 5281 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 5282 5283 Value *UndefVec = IsVoidRetTy ? nullptr : 5284 UndefValue::get(Instr->getType()); 5285 // Create a new entry in the WidenMap and initialize it to Undef or Null. 5286 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 5287 5288 Instruction *InsertPt = Builder.GetInsertPoint(); 5289 BasicBlock *IfBlock = Builder.GetInsertBlock(); 5290 BasicBlock *CondBlock = nullptr; 5291 5292 VectorParts Cond; 5293 Loop *VectorLp = nullptr; 5294 if (IfPredicateStore) { 5295 assert(Instr->getParent()->getSinglePredecessor() && 5296 "Only support single predecessor blocks"); 5297 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(), 5298 Instr->getParent()); 5299 VectorLp = LI->getLoopFor(IfBlock); 5300 assert(VectorLp && "Must have a loop for this block"); 5301 } 5302 5303 // For each vector unroll 'part': 5304 for (unsigned Part = 0; Part < UF; ++Part) { 5305 // For each scalar that we create: 5306 5307 // Start an "if (pred) a[i] = ..." block. 5308 Value *Cmp = nullptr; 5309 if (IfPredicateStore) { 5310 if (Cond[Part]->getType()->isVectorTy()) 5311 Cond[Part] = 5312 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0)); 5313 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part], 5314 ConstantInt::get(Cond[Part]->getType(), 1)); 5315 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store"); 5316 LoopVectorBody.push_back(CondBlock); 5317 VectorLp->addBasicBlockToLoop(CondBlock, *LI); 5318 // Update Builder with newly created basic block. 5319 Builder.SetInsertPoint(InsertPt); 5320 } 5321 5322 Instruction *Cloned = Instr->clone(); 5323 if (!IsVoidRetTy) 5324 Cloned->setName(Instr->getName() + ".cloned"); 5325 // Replace the operands of the cloned instructions with extracted scalars. 5326 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5327 Value *Op = Params[op][Part]; 5328 Cloned->setOperand(op, Op); 5329 } 5330 5331 // Place the cloned scalar in the new loop. 5332 Builder.Insert(Cloned); 5333 5334 // If the original scalar returns a value we need to place it in a vector 5335 // so that future users will be able to use it. 5336 if (!IsVoidRetTy) 5337 VecResults[Part] = Cloned; 5338 5339 // End if-block. 5340 if (IfPredicateStore) { 5341 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else"); 5342 LoopVectorBody.push_back(NewIfBlock); 5343 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI); 5344 Builder.SetInsertPoint(InsertPt); 5345 Instruction *OldBr = IfBlock->getTerminator(); 5346 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr); 5347 OldBr->eraseFromParent(); 5348 IfBlock = NewIfBlock; 5349 } 5350 } 5351} 5352 5353void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) { 5354 StoreInst *SI = dyn_cast<StoreInst>(Instr); 5355 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent())); 5356 5357 return scalarizeInstruction(Instr, IfPredicateStore); 5358} 5359 5360Value *InnerLoopUnroller::reverseVector(Value *Vec) { 5361 return Vec; 5362} 5363 5364Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { 5365 return V; 5366} 5367 5368Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) { 5369 // When unrolling and the VF is 1, we only need to add a simple scalar. 5370 Type *ITy = Val->getType(); 5371 assert(!ITy->isVectorTy() && "Val must be a scalar"); 5372 Constant *C = ConstantInt::get(ITy, StartIdx); 5373 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction"); 5374} 5375