LoopVectorize.cpp revision e39b15195a3607ee708be9d105b5fc591b4665dd
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#define LV_NAME "loop-vectorize" 46#define DEBUG_TYPE LV_NAME 47 48#include "llvm/Transforms/Vectorize.h" 49#include "llvm/ADT/DenseMap.h" 50#include "llvm/ADT/EquivalenceClasses.h" 51#include "llvm/ADT/Hashing.h" 52#include "llvm/ADT/MapVector.h" 53#include "llvm/ADT/SetVector.h" 54#include "llvm/ADT/SmallPtrSet.h" 55#include "llvm/ADT/SmallSet.h" 56#include "llvm/ADT/SmallVector.h" 57#include "llvm/ADT/StringExtras.h" 58#include "llvm/Analysis/AliasAnalysis.h" 59#include "llvm/Analysis/Dominators.h" 60#include "llvm/Analysis/LoopInfo.h" 61#include "llvm/Analysis/LoopIterator.h" 62#include "llvm/Analysis/LoopPass.h" 63#include "llvm/Analysis/ScalarEvolution.h" 64#include "llvm/Analysis/ScalarEvolutionExpander.h" 65#include "llvm/Analysis/ScalarEvolutionExpressions.h" 66#include "llvm/Analysis/TargetTransformInfo.h" 67#include "llvm/Analysis/ValueTracking.h" 68#include "llvm/Analysis/Verifier.h" 69#include "llvm/IR/Constants.h" 70#include "llvm/IR/DataLayout.h" 71#include "llvm/IR/DerivedTypes.h" 72#include "llvm/IR/Function.h" 73#include "llvm/IR/IRBuilder.h" 74#include "llvm/IR/Instructions.h" 75#include "llvm/IR/IntrinsicInst.h" 76#include "llvm/IR/LLVMContext.h" 77#include "llvm/IR/Module.h" 78#include "llvm/IR/Type.h" 79#include "llvm/IR/Value.h" 80#include "llvm/Pass.h" 81#include "llvm/Support/CommandLine.h" 82#include "llvm/Support/Debug.h" 83#include "llvm/Support/PatternMatch.h" 84#include "llvm/Support/raw_ostream.h" 85#include "llvm/Support/ValueHandle.h" 86#include "llvm/Target/TargetLibraryInfo.h" 87#include "llvm/Transforms/Scalar.h" 88#include "llvm/Transforms/Utils/BasicBlockUtils.h" 89#include "llvm/Transforms/Utils/Local.h" 90#include <algorithm> 91#include <map> 92 93using namespace llvm; 94using namespace llvm::PatternMatch; 95 96static cl::opt<unsigned> 97VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden, 98 cl::desc("Sets the SIMD width. Zero is autoselect.")); 99 100static cl::opt<unsigned> 101VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden, 102 cl::desc("Sets the vectorization unroll count. " 103 "Zero is autoselect.")); 104 105static cl::opt<bool> 106EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, 107 cl::desc("Enable if-conversion during vectorization.")); 108 109/// We don't vectorize loops with a known constant trip count below this number. 110static cl::opt<unsigned> 111TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16), 112 cl::Hidden, 113 cl::desc("Don't vectorize loops with a constant " 114 "trip count that is smaller than this " 115 "value.")); 116 117/// We don't unroll loops with a known constant trip count below this number. 118static const unsigned TinyTripCountUnrollThreshold = 128; 119 120/// When performing memory disambiguation checks at runtime do not make more 121/// than this number of comparisons. 122static const unsigned RuntimeMemoryCheckThreshold = 8; 123 124/// Maximum simd width. 125static const unsigned MaxVectorWidth = 64; 126 127/// Maximum vectorization unroll count. 128static const unsigned MaxUnrollFactor = 16; 129 130/// The cost of a loop that is considered 'small' by the unroller. 131static const unsigned SmallLoopCost = 20; 132 133namespace { 134 135// Forward declarations. 136class LoopVectorizationLegality; 137class LoopVectorizationCostModel; 138 139/// InnerLoopVectorizer vectorizes loops which contain only one basic 140/// block to a specified vectorization factor (VF). 141/// This class performs the widening of scalars into vectors, or multiple 142/// scalars. This class also implements the following features: 143/// * It inserts an epilogue loop for handling loops that don't have iteration 144/// counts that are known to be a multiple of the vectorization factor. 145/// * It handles the code generation for reduction variables. 146/// * Scalarization (implementation using scalars) of un-vectorizable 147/// instructions. 148/// InnerLoopVectorizer does not perform any vectorization-legality 149/// checks, and relies on the caller to check for the different legality 150/// aspects. The InnerLoopVectorizer relies on the 151/// LoopVectorizationLegality class to provide information about the induction 152/// and reduction variables that were found to a given vectorization factor. 153class InnerLoopVectorizer { 154public: 155 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, 156 DominatorTree *DT, DataLayout *DL, 157 const TargetLibraryInfo *TLI, unsigned VecWidth, 158 unsigned UnrollFactor) 159 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI), 160 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0), 161 OldInduction(0), WidenMap(UnrollFactor) {} 162 163 // Perform the actual loop widening (vectorization). 164 void vectorize(LoopVectorizationLegality *Legal) { 165 // Create a new empty loop. Unlink the old loop and connect the new one. 166 createEmptyLoop(Legal); 167 // Widen each instruction in the old loop to a new one in the new loop. 168 // Use the Legality module to find the induction and reduction variables. 169 vectorizeLoop(Legal); 170 // Register the new loop and update the analysis passes. 171 updateAnalysis(); 172 } 173 174 virtual ~InnerLoopVectorizer() {} 175 176protected: 177 /// A small list of PHINodes. 178 typedef SmallVector<PHINode*, 4> PhiVector; 179 /// When we unroll loops we have multiple vector values for each scalar. 180 /// This data structure holds the unrolled and vectorized values that 181 /// originated from one scalar instruction. 182 typedef SmallVector<Value*, 2> VectorParts; 183 184 // When we if-convert we need create edge masks. We have to cache values so 185 // that we don't end up with exponential recursion/IR. 186 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>, 187 VectorParts> EdgeMaskCache; 188 189 /// Add code that checks at runtime if the accessed arrays overlap. 190 /// Returns the comparator value or NULL if no check is needed. 191 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal, 192 Instruction *Loc); 193 /// Create an empty loop, based on the loop ranges of the old loop. 194 void createEmptyLoop(LoopVectorizationLegality *Legal); 195 /// Copy and widen the instructions from the old loop. 196 virtual void vectorizeLoop(LoopVectorizationLegality *Legal); 197 198 /// \brief The Loop exit block may have single value PHI nodes where the 199 /// incoming value is 'Undef'. While vectorizing we only handled real values 200 /// that were defined inside the loop. Here we fix the 'undef case'. 201 /// See PR14725. 202 void fixLCSSAPHIs(); 203 204 /// A helper function that computes the predicate of the block BB, assuming 205 /// that the header block of the loop is set to True. It returns the *entry* 206 /// mask for the block BB. 207 VectorParts createBlockInMask(BasicBlock *BB); 208 /// A helper function that computes the predicate of the edge between SRC 209 /// and DST. 210 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst); 211 212 /// A helper function to vectorize a single BB within the innermost loop. 213 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB, 214 PhiVector *PV); 215 216 /// Vectorize a single PHINode in a block. This method handles the induction 217 /// variable canonicalization. It supports both VF = 1 for unrolled loops and 218 /// arbitrary length vectors. 219 void widenPHIInstruction(Instruction *PN, VectorParts &Entry, 220 LoopVectorizationLegality *Legal, 221 unsigned UF, unsigned VF, PhiVector *PV); 222 223 /// Insert the new loop to the loop hierarchy and pass manager 224 /// and update the analysis passes. 225 void updateAnalysis(); 226 227 /// This instruction is un-vectorizable. Implement it as a sequence 228 /// of scalars. 229 virtual void scalarizeInstruction(Instruction *Instr); 230 231 /// Vectorize Load and Store instructions, 232 virtual void vectorizeMemoryInstruction(Instruction *Instr, 233 LoopVectorizationLegality *Legal); 234 235 /// Create a broadcast instruction. This method generates a broadcast 236 /// instruction (shuffle) for loop invariant values and for the induction 237 /// value. If this is the induction variable then we extend it to N, N+1, ... 238 /// this is needed because each iteration in the loop corresponds to a SIMD 239 /// element. 240 virtual Value *getBroadcastInstrs(Value *V); 241 242 /// This function adds 0, 1, 2 ... to each vector element, starting at zero. 243 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...). 244 /// The sequence starts at StartIndex. 245 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate); 246 247 /// When we go over instructions in the basic block we rely on previous 248 /// values within the current basic block or on loop invariant values. 249 /// When we widen (vectorize) values we place them in the map. If the values 250 /// are not within the map, they have to be loop invariant, so we simply 251 /// broadcast them into a vector. 252 VectorParts &getVectorValue(Value *V); 253 254 /// Generate a shuffle sequence that will reverse the vector Vec. 255 virtual Value *reverseVector(Value *Vec); 256 257 /// This is a helper class that holds the vectorizer state. It maps scalar 258 /// instructions to vector instructions. When the code is 'unrolled' then 259 /// then a single scalar value is mapped to multiple vector parts. The parts 260 /// are stored in the VectorPart type. 261 struct ValueMap { 262 /// C'tor. UnrollFactor controls the number of vectors ('parts') that 263 /// are mapped. 264 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {} 265 266 /// \return True if 'Key' is saved in the Value Map. 267 bool has(Value *Key) const { return MapStorage.count(Key); } 268 269 /// Initializes a new entry in the map. Sets all of the vector parts to the 270 /// save value in 'Val'. 271 /// \return A reference to a vector with splat values. 272 VectorParts &splat(Value *Key, Value *Val) { 273 VectorParts &Entry = MapStorage[Key]; 274 Entry.assign(UF, Val); 275 return Entry; 276 } 277 278 ///\return A reference to the value that is stored at 'Key'. 279 VectorParts &get(Value *Key) { 280 VectorParts &Entry = MapStorage[Key]; 281 if (Entry.empty()) 282 Entry.resize(UF); 283 assert(Entry.size() == UF); 284 return Entry; 285 } 286 287 private: 288 /// The unroll factor. Each entry in the map stores this number of vector 289 /// elements. 290 unsigned UF; 291 292 /// Map storage. We use std::map and not DenseMap because insertions to a 293 /// dense map invalidates its iterators. 294 std::map<Value *, VectorParts> MapStorage; 295 }; 296 297 /// The original loop. 298 Loop *OrigLoop; 299 /// Scev analysis to use. 300 ScalarEvolution *SE; 301 /// Loop Info. 302 LoopInfo *LI; 303 /// Dominator Tree. 304 DominatorTree *DT; 305 /// Data Layout. 306 DataLayout *DL; 307 /// Target Library Info. 308 const TargetLibraryInfo *TLI; 309 310 /// The vectorization SIMD factor to use. Each vector will have this many 311 /// vector elements. 312 unsigned VF; 313 314protected: 315 /// The vectorization unroll factor to use. Each scalar is vectorized to this 316 /// many different vector instructions. 317 unsigned UF; 318 319 /// The builder that we use 320 IRBuilder<> Builder; 321 322 // --- Vectorization state --- 323 324 /// The vector-loop preheader. 325 BasicBlock *LoopVectorPreHeader; 326 /// The scalar-loop preheader. 327 BasicBlock *LoopScalarPreHeader; 328 /// Middle Block between the vector and the scalar. 329 BasicBlock *LoopMiddleBlock; 330 ///The ExitBlock of the scalar loop. 331 BasicBlock *LoopExitBlock; 332 ///The vector loop body. 333 BasicBlock *LoopVectorBody; 334 ///The scalar loop body. 335 BasicBlock *LoopScalarBody; 336 /// A list of all bypass blocks. The first block is the entry of the loop. 337 SmallVector<BasicBlock *, 4> LoopBypassBlocks; 338 339 /// The new Induction variable which was added to the new block. 340 PHINode *Induction; 341 /// The induction variable of the old basic block. 342 PHINode *OldInduction; 343 /// Holds the extended (to the widest induction type) start index. 344 Value *ExtendedIdx; 345 /// Maps scalars to widened vectors. 346 ValueMap WidenMap; 347 EdgeMaskCache MaskCache; 348}; 349 350class InnerLoopUnroller : public InnerLoopVectorizer { 351public: 352 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, 353 DominatorTree *DT, DataLayout *DL, 354 const TargetLibraryInfo *TLI, unsigned UnrollFactor) : 355 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { } 356 357private: 358 virtual void scalarizeInstruction(Instruction *Instr); 359 virtual void vectorizeMemoryInstruction(Instruction *Instr, 360 LoopVectorizationLegality *Legal); 361 virtual Value *getBroadcastInstrs(Value *V); 362 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate); 363 virtual Value *reverseVector(Value *Vec); 364}; 365 366/// \brief Look for a meaningful debug location on the instruction or it's 367/// operands. 368static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { 369 if (!I) 370 return I; 371 372 DebugLoc Empty; 373 if (I->getDebugLoc() != Empty) 374 return I; 375 376 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) { 377 if (Instruction *OpInst = dyn_cast<Instruction>(*OI)) 378 if (OpInst->getDebugLoc() != Empty) 379 return OpInst; 380 } 381 382 return I; 383} 384 385/// \brief Set the debug location in the builder using the debug location in the 386/// instruction. 387static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) { 388 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) 389 B.SetCurrentDebugLocation(Inst->getDebugLoc()); 390 else 391 B.SetCurrentDebugLocation(DebugLoc()); 392} 393 394/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and 395/// to what vectorization factor. 396/// This class does not look at the profitability of vectorization, only the 397/// legality. This class has two main kinds of checks: 398/// * Memory checks - The code in canVectorizeMemory checks if vectorization 399/// will change the order of memory accesses in a way that will change the 400/// correctness of the program. 401/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory 402/// checks for a number of different conditions, such as the availability of a 403/// single induction variable, that all types are supported and vectorize-able, 404/// etc. This code reflects the capabilities of InnerLoopVectorizer. 405/// This class is also used by InnerLoopVectorizer for identifying 406/// induction variable and the different reduction variables. 407class LoopVectorizationLegality { 408public: 409 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL, 410 DominatorTree *DT, TargetLibraryInfo *TLI) 411 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI), 412 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false), 413 MaxSafeDepDistBytes(-1U) {} 414 415 /// This enum represents the kinds of reductions that we support. 416 enum ReductionKind { 417 RK_NoReduction, ///< Not a reduction. 418 RK_IntegerAdd, ///< Sum of integers. 419 RK_IntegerMult, ///< Product of integers. 420 RK_IntegerOr, ///< Bitwise or logical OR of numbers. 421 RK_IntegerAnd, ///< Bitwise or logical AND of numbers. 422 RK_IntegerXor, ///< Bitwise or logical XOR of numbers. 423 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()). 424 RK_FloatAdd, ///< Sum of floats. 425 RK_FloatMult, ///< Product of floats. 426 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()). 427 }; 428 429 /// This enum represents the kinds of inductions that we support. 430 enum InductionKind { 431 IK_NoInduction, ///< Not an induction variable. 432 IK_IntInduction, ///< Integer induction variable. Step = 1. 433 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1. 434 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem). 435 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem). 436 }; 437 438 // This enum represents the kind of minmax reduction. 439 enum MinMaxReductionKind { 440 MRK_Invalid, 441 MRK_UIntMin, 442 MRK_UIntMax, 443 MRK_SIntMin, 444 MRK_SIntMax, 445 MRK_FloatMin, 446 MRK_FloatMax 447 }; 448 449 /// This struct holds information about reduction variables. 450 struct ReductionDescriptor { 451 ReductionDescriptor() : StartValue(0), LoopExitInstr(0), 452 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {} 453 454 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K, 455 MinMaxReductionKind MK) 456 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {} 457 458 // The starting value of the reduction. 459 // It does not have to be zero! 460 TrackingVH<Value> StartValue; 461 // The instruction who's value is used outside the loop. 462 Instruction *LoopExitInstr; 463 // The kind of the reduction. 464 ReductionKind Kind; 465 // If this a min/max reduction the kind of reduction. 466 MinMaxReductionKind MinMaxKind; 467 }; 468 469 /// This POD struct holds information about a potential reduction operation. 470 struct ReductionInstDesc { 471 ReductionInstDesc(bool IsRedux, Instruction *I) : 472 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {} 473 474 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) : 475 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {} 476 477 // Is this instruction a reduction candidate. 478 bool IsReduction; 479 // The last instruction in a min/max pattern (select of the select(icmp()) 480 // pattern), or the current reduction instruction otherwise. 481 Instruction *PatternLastInst; 482 // If this is a min/max pattern the comparison predicate. 483 MinMaxReductionKind MinMaxKind; 484 }; 485 486 /// This struct holds information about the memory runtime legality 487 /// check that a group of pointers do not overlap. 488 struct RuntimePointerCheck { 489 RuntimePointerCheck() : Need(false) {} 490 491 /// Reset the state of the pointer runtime information. 492 void reset() { 493 Need = false; 494 Pointers.clear(); 495 Starts.clear(); 496 Ends.clear(); 497 IsWritePtr.clear(); 498 DependencySetId.clear(); 499 } 500 501 /// Insert a pointer and calculate the start and end SCEVs. 502 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, 503 unsigned DepSetId); 504 505 /// This flag indicates if we need to add the runtime check. 506 bool Need; 507 /// Holds the pointers that we need to check. 508 SmallVector<TrackingVH<Value>, 2> Pointers; 509 /// Holds the pointer value at the beginning of the loop. 510 SmallVector<const SCEV*, 2> Starts; 511 /// Holds the pointer value at the end of the loop. 512 SmallVector<const SCEV*, 2> Ends; 513 /// Holds the information if this pointer is used for writing to memory. 514 SmallVector<bool, 2> IsWritePtr; 515 /// Holds the id of the set of pointers that could be dependent because of a 516 /// shared underlying object. 517 SmallVector<unsigned, 2> DependencySetId; 518 }; 519 520 /// A struct for saving information about induction variables. 521 struct InductionInfo { 522 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {} 523 InductionInfo() : StartValue(0), IK(IK_NoInduction) {} 524 /// Start value. 525 TrackingVH<Value> StartValue; 526 /// Induction kind. 527 InductionKind IK; 528 }; 529 530 /// ReductionList contains the reduction descriptors for all 531 /// of the reductions that were found in the loop. 532 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList; 533 534 /// InductionList saves induction variables and maps them to the 535 /// induction descriptor. 536 typedef MapVector<PHINode*, InductionInfo> InductionList; 537 538 /// Returns true if it is legal to vectorize this loop. 539 /// This does not mean that it is profitable to vectorize this 540 /// loop, only that it is legal to do so. 541 bool canVectorize(); 542 543 /// Returns the Induction variable. 544 PHINode *getInduction() { return Induction; } 545 546 /// Returns the reduction variables found in the loop. 547 ReductionList *getReductionVars() { return &Reductions; } 548 549 /// Returns the induction variables found in the loop. 550 InductionList *getInductionVars() { return &Inductions; } 551 552 /// Returns the widest induction type. 553 Type *getWidestInductionType() { return WidestIndTy; } 554 555 /// Returns True if V is an induction variable in this loop. 556 bool isInductionVariable(const Value *V); 557 558 /// Return true if the block BB needs to be predicated in order for the loop 559 /// to be vectorized. 560 bool blockNeedsPredication(BasicBlock *BB); 561 562 /// Check if this pointer is consecutive when vectorizing. This happens 563 /// when the last index of the GEP is the induction variable, or that the 564 /// pointer itself is an induction variable. 565 /// This check allows us to vectorize A[idx] into a wide load/store. 566 /// Returns: 567 /// 0 - Stride is unknown or non consecutive. 568 /// 1 - Address is consecutive. 569 /// -1 - Address is consecutive, and decreasing. 570 int isConsecutivePtr(Value *Ptr); 571 572 /// Returns true if the value V is uniform within the loop. 573 bool isUniform(Value *V); 574 575 /// Returns true if this instruction will remain scalar after vectorization. 576 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } 577 578 /// Returns the information that we collected about runtime memory check. 579 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; } 580 581 /// This function returns the identity element (or neutral element) for 582 /// the operation K. 583 static Constant *getReductionIdentity(ReductionKind K, Type *Tp); 584 585 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; } 586 587private: 588 /// Check if a single basic block loop is vectorizable. 589 /// At this point we know that this is a loop with a constant trip count 590 /// and we only need to check individual instructions. 591 bool canVectorizeInstrs(); 592 593 /// When we vectorize loops we may change the order in which 594 /// we read and write from memory. This method checks if it is 595 /// legal to vectorize the code, considering only memory constrains. 596 /// Returns true if the loop is vectorizable 597 bool canVectorizeMemory(); 598 599 /// Return true if we can vectorize this loop using the IF-conversion 600 /// transformation. 601 bool canVectorizeWithIfConvert(); 602 603 /// Collect the variables that need to stay uniform after vectorization. 604 void collectLoopUniforms(); 605 606 /// Return true if all of the instructions in the block can be speculatively 607 /// executed. \p SafePtrs is a list of addresses that are known to be legal 608 /// and we know that we can read from them without segfault. 609 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs); 610 611 /// Returns True, if 'Phi' is the kind of reduction variable for type 612 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. 613 bool AddReductionVar(PHINode *Phi, ReductionKind Kind); 614 /// Returns a struct describing if the instruction 'I' can be a reduction 615 /// variable of type 'Kind'. If the reduction is a min/max pattern of 616 /// select(icmp()) this function advances the instruction pointer 'I' from the 617 /// compare instruction to the select instruction and stores this pointer in 618 /// 'PatternLastInst' member of the returned struct. 619 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind, 620 ReductionInstDesc &Desc); 621 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 622 /// pattern corresponding to a min(X, Y) or max(X, Y). 623 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I, 624 ReductionInstDesc &Prev); 625 /// Returns the induction kind of Phi. This function may return NoInduction 626 /// if the PHI is not an induction variable. 627 InductionKind isInductionVariable(PHINode *Phi); 628 629 /// The loop that we evaluate. 630 Loop *TheLoop; 631 /// Scev analysis. 632 ScalarEvolution *SE; 633 /// DataLayout analysis. 634 DataLayout *DL; 635 /// Dominators. 636 DominatorTree *DT; 637 /// Target Library Info. 638 TargetLibraryInfo *TLI; 639 640 // --- vectorization state --- // 641 642 /// Holds the integer induction variable. This is the counter of the 643 /// loop. 644 PHINode *Induction; 645 /// Holds the reduction variables. 646 ReductionList Reductions; 647 /// Holds all of the induction variables that we found in the loop. 648 /// Notice that inductions don't need to start at zero and that induction 649 /// variables can be pointers. 650 InductionList Inductions; 651 /// Holds the widest induction type encountered. 652 Type *WidestIndTy; 653 654 /// Allowed outside users. This holds the reduction 655 /// vars which can be accessed from outside the loop. 656 SmallPtrSet<Value*, 4> AllowedExit; 657 /// This set holds the variables which are known to be uniform after 658 /// vectorization. 659 SmallPtrSet<Instruction*, 4> Uniforms; 660 /// We need to check that all of the pointers in this list are disjoint 661 /// at runtime. 662 RuntimePointerCheck PtrRtCheck; 663 /// Can we assume the absence of NaNs. 664 bool HasFunNoNaNAttr; 665 666 unsigned MaxSafeDepDistBytes; 667}; 668 669/// LoopVectorizationCostModel - estimates the expected speedups due to 670/// vectorization. 671/// In many cases vectorization is not profitable. This can happen because of 672/// a number of reasons. In this class we mainly attempt to predict the 673/// expected speedup/slowdowns due to the supported instruction set. We use the 674/// TargetTransformInfo to query the different backends for the cost of 675/// different operations. 676class LoopVectorizationCostModel { 677public: 678 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI, 679 LoopVectorizationLegality *Legal, 680 const TargetTransformInfo &TTI, 681 DataLayout *DL, const TargetLibraryInfo *TLI) 682 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {} 683 684 /// Information about vectorization costs 685 struct VectorizationFactor { 686 unsigned Width; // Vector width with best cost 687 unsigned Cost; // Cost of the loop with that width 688 }; 689 /// \return The most profitable vectorization factor and the cost of that VF. 690 /// This method checks every power of two up to VF. If UserVF is not ZERO 691 /// then this vectorization factor will be selected if vectorization is 692 /// possible. 693 VectorizationFactor selectVectorizationFactor(bool OptForSize, 694 unsigned UserVF); 695 696 /// \return The size (in bits) of the widest type in the code that 697 /// needs to be vectorized. We ignore values that remain scalar such as 698 /// 64 bit loop indices. 699 unsigned getWidestType(); 700 701 /// \return The most profitable unroll factor. 702 /// If UserUF is non-zero then this method finds the best unroll-factor 703 /// based on register pressure and other parameters. 704 /// VF and LoopCost are the selected vectorization factor and the cost of the 705 /// selected VF. 706 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF, 707 unsigned LoopCost); 708 709 /// \brief A struct that represents some properties of the register usage 710 /// of a loop. 711 struct RegisterUsage { 712 /// Holds the number of loop invariant values that are used in the loop. 713 unsigned LoopInvariantRegs; 714 /// Holds the maximum number of concurrent live intervals in the loop. 715 unsigned MaxLocalUsers; 716 /// Holds the number of instructions in the loop. 717 unsigned NumInstructions; 718 }; 719 720 /// \return information about the register usage of the loop. 721 RegisterUsage calculateRegisterUsage(); 722 723private: 724 /// Returns the expected execution cost. The unit of the cost does 725 /// not matter because we use the 'cost' units to compare different 726 /// vector widths. The cost that is returned is *not* normalized by 727 /// the factor width. 728 unsigned expectedCost(unsigned VF); 729 730 /// Returns the execution time cost of an instruction for a given vector 731 /// width. Vector width of one means scalar. 732 unsigned getInstructionCost(Instruction *I, unsigned VF); 733 734 /// A helper function for converting Scalar types to vector types. 735 /// If the incoming type is void, we return void. If the VF is 1, we return 736 /// the scalar type. 737 static Type* ToVectorTy(Type *Scalar, unsigned VF); 738 739 /// Returns whether the instruction is a load or store and will be a emitted 740 /// as a vector operation. 741 bool isConsecutiveLoadOrStore(Instruction *I); 742 743 /// The loop that we evaluate. 744 Loop *TheLoop; 745 /// Scev analysis. 746 ScalarEvolution *SE; 747 /// Loop Info analysis. 748 LoopInfo *LI; 749 /// Vectorization legality. 750 LoopVectorizationLegality *Legal; 751 /// Vector target information. 752 const TargetTransformInfo &TTI; 753 /// Target data layout information. 754 DataLayout *DL; 755 /// Target Library Info. 756 const TargetLibraryInfo *TLI; 757}; 758 759/// Utility class for getting and setting loop vectorizer hints in the form 760/// of loop metadata. 761struct LoopVectorizeHints { 762 /// Vectorization width. 763 unsigned Width; 764 /// Vectorization unroll factor. 765 unsigned Unroll; 766 767 LoopVectorizeHints(const Loop *L, bool DisableUnrolling) 768 : Width(VectorizationFactor) 769 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll) 770 , LoopID(L->getLoopID()) { 771 getHints(L); 772 // The command line options override any loop metadata except for when 773 // width == 1 which is used to indicate the loop is already vectorized. 774 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1) 775 Width = VectorizationFactor; 776 if (VectorizationUnroll.getNumOccurrences() > 0) 777 Unroll = VectorizationUnroll; 778 779 DEBUG(if (DisableUnrolling && Unroll == 1) 780 dbgs() << "LV: Unrolling disabled by the pass manager\n"); 781 } 782 783 /// Return the loop vectorizer metadata prefix. 784 static StringRef Prefix() { return "llvm.vectorizer."; } 785 786 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) { 787 SmallVector<Value*, 2> Vals; 788 Vals.push_back(MDString::get(Context, Name)); 789 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V)); 790 return MDNode::get(Context, Vals); 791 } 792 793 /// Mark the loop L as already vectorized by setting the width to 1. 794 void setAlreadyVectorized(Loop *L) { 795 LLVMContext &Context = L->getHeader()->getContext(); 796 797 Width = 1; 798 799 // Create a new loop id with one more operand for the already_vectorized 800 // hint. If the loop already has a loop id then copy the existing operands. 801 SmallVector<Value*, 4> Vals(1); 802 if (LoopID) 803 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) 804 Vals.push_back(LoopID->getOperand(i)); 805 806 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width)); 807 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1)); 808 809 MDNode *NewLoopID = MDNode::get(Context, Vals); 810 // Set operand 0 to refer to the loop id itself. 811 NewLoopID->replaceOperandWith(0, NewLoopID); 812 813 L->setLoopID(NewLoopID); 814 if (LoopID) 815 LoopID->replaceAllUsesWith(NewLoopID); 816 817 LoopID = NewLoopID; 818 } 819 820private: 821 MDNode *LoopID; 822 823 /// Find hints specified in the loop metadata. 824 void getHints(const Loop *L) { 825 if (!LoopID) 826 return; 827 828 // First operand should refer to the loop id itself. 829 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 830 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 831 832 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { 833 const MDString *S = 0; 834 SmallVector<Value*, 4> Args; 835 836 // The expected hint is either a MDString or a MDNode with the first 837 // operand a MDString. 838 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) { 839 if (!MD || MD->getNumOperands() == 0) 840 continue; 841 S = dyn_cast<MDString>(MD->getOperand(0)); 842 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i) 843 Args.push_back(MD->getOperand(i)); 844 } else { 845 S = dyn_cast<MDString>(LoopID->getOperand(i)); 846 assert(Args.size() == 0 && "too many arguments for MDString"); 847 } 848 849 if (!S) 850 continue; 851 852 // Check if the hint starts with the vectorizer prefix. 853 StringRef Hint = S->getString(); 854 if (!Hint.startswith(Prefix())) 855 continue; 856 // Remove the prefix. 857 Hint = Hint.substr(Prefix().size(), StringRef::npos); 858 859 if (Args.size() == 1) 860 getHint(Hint, Args[0]); 861 } 862 } 863 864 // Check string hint with one operand. 865 void getHint(StringRef Hint, Value *Arg) { 866 const ConstantInt *C = dyn_cast<ConstantInt>(Arg); 867 if (!C) return; 868 unsigned Val = C->getZExtValue(); 869 870 if (Hint == "width") { 871 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth) 872 Width = Val; 873 else 874 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n"); 875 } else if (Hint == "unroll") { 876 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor) 877 Unroll = Val; 878 else 879 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n"); 880 } else { 881 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n'); 882 } 883 } 884}; 885 886/// The LoopVectorize Pass. 887struct LoopVectorize : public LoopPass { 888 /// Pass identification, replacement for typeid 889 static char ID; 890 891 explicit LoopVectorize(bool NoUnrolling = false) 892 : LoopPass(ID), DisableUnrolling(NoUnrolling) { 893 initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); 894 } 895 896 ScalarEvolution *SE; 897 DataLayout *DL; 898 LoopInfo *LI; 899 TargetTransformInfo *TTI; 900 DominatorTree *DT; 901 TargetLibraryInfo *TLI; 902 bool DisableUnrolling; 903 904 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) { 905 // We only vectorize innermost loops. 906 if (!L->empty()) 907 return false; 908 909 SE = &getAnalysis<ScalarEvolution>(); 910 DL = getAnalysisIfAvailable<DataLayout>(); 911 LI = &getAnalysis<LoopInfo>(); 912 TTI = &getAnalysis<TargetTransformInfo>(); 913 DT = &getAnalysis<DominatorTree>(); 914 TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); 915 916 // If the target claims to have no vector registers don't attempt 917 // vectorization. 918 if (!TTI->getNumberOfRegisters(true)) 919 return false; 920 921 if (DL == NULL) { 922 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout\n"); 923 return false; 924 } 925 926 DEBUG(dbgs() << "LV: Checking a loop in \"" << 927 L->getHeader()->getParent()->getName() << "\"\n"); 928 929 LoopVectorizeHints Hints(L, DisableUnrolling); 930 931 if (Hints.Width == 1 && Hints.Unroll == 1) { 932 DEBUG(dbgs() << "LV: Not vectorizing.\n"); 933 return false; 934 } 935 936 // Check if it is legal to vectorize the loop. 937 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI); 938 if (!LVL.canVectorize()) { 939 DEBUG(dbgs() << "LV: Not vectorizing.\n"); 940 return false; 941 } 942 943 // Use the cost model. 944 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI); 945 946 // Check the function attributes to find out if this function should be 947 // optimized for size. 948 Function *F = L->getHeader()->getParent(); 949 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize; 950 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat; 951 unsigned FnIndex = AttributeSet::FunctionIndex; 952 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr); 953 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr); 954 955 if (NoFloat) { 956 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" 957 "attribute is used.\n"); 958 return false; 959 } 960 961 // Select the optimal vectorization factor. 962 LoopVectorizationCostModel::VectorizationFactor VF; 963 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width); 964 // Select the unroll factor. 965 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width, 966 VF.Cost); 967 968 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<< 969 F->getParent()->getModuleIdentifier() << '\n'); 970 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n'); 971 972 if (VF.Width == 1) { 973 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); 974 if (UF == 1) 975 return false; 976 // We decided not to vectorize, but we may want to unroll. 977 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF); 978 Unroller.vectorize(&LVL); 979 } else { 980 // If we decided that it is *legal* to vectorize the loop then do it. 981 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF); 982 LB.vectorize(&LVL); 983 } 984 985 // Mark the loop as already vectorized to avoid vectorizing again. 986 Hints.setAlreadyVectorized(L); 987 988 DEBUG(verifyFunction(*L->getHeader()->getParent())); 989 return true; 990 } 991 992 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 993 LoopPass::getAnalysisUsage(AU); 994 AU.addRequiredID(LoopSimplifyID); 995 AU.addRequiredID(LCSSAID); 996 AU.addRequired<DominatorTree>(); 997 AU.addRequired<LoopInfo>(); 998 AU.addRequired<ScalarEvolution>(); 999 AU.addRequired<TargetTransformInfo>(); 1000 AU.addPreserved<LoopInfo>(); 1001 AU.addPreserved<DominatorTree>(); 1002 } 1003 1004}; 1005 1006} // end anonymous namespace 1007 1008//===----------------------------------------------------------------------===// 1009// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and 1010// LoopVectorizationCostModel. 1011//===----------------------------------------------------------------------===// 1012 1013void 1014LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE, 1015 Loop *Lp, Value *Ptr, 1016 bool WritePtr, 1017 unsigned DepSetId) { 1018 const SCEV *Sc = SE->getSCEV(Ptr); 1019 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); 1020 assert(AR && "Invalid addrec expression"); 1021 const SCEV *Ex = SE->getBackedgeTakenCount(Lp); 1022 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE); 1023 Pointers.push_back(Ptr); 1024 Starts.push_back(AR->getStart()); 1025 Ends.push_back(ScEnd); 1026 IsWritePtr.push_back(WritePtr); 1027 DependencySetId.push_back(DepSetId); 1028} 1029 1030Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { 1031 // We need to place the broadcast of invariant variables outside the loop. 1032 Instruction *Instr = dyn_cast<Instruction>(V); 1033 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody); 1034 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr; 1035 1036 // Place the code for broadcasting invariant variables in the new preheader. 1037 IRBuilder<>::InsertPointGuard Guard(Builder); 1038 if (Invariant) 1039 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); 1040 1041 // Broadcast the scalar into all locations in the vector. 1042 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); 1043 1044 return Shuf; 1045} 1046 1047Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx, 1048 bool Negate) { 1049 assert(Val->getType()->isVectorTy() && "Must be a vector"); 1050 assert(Val->getType()->getScalarType()->isIntegerTy() && 1051 "Elem must be an integer"); 1052 // Create the types. 1053 Type *ITy = Val->getType()->getScalarType(); 1054 VectorType *Ty = cast<VectorType>(Val->getType()); 1055 int VLen = Ty->getNumElements(); 1056 SmallVector<Constant*, 8> Indices; 1057 1058 // Create a vector of consecutive numbers from zero to VF. 1059 for (int i = 0; i < VLen; ++i) { 1060 int64_t Idx = Negate ? (-i) : i; 1061 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate)); 1062 } 1063 1064 // Add the consecutive indices to the vector value. 1065 Constant *Cv = ConstantVector::get(Indices); 1066 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); 1067 return Builder.CreateAdd(Val, Cv, "induction"); 1068} 1069 1070/// \brief Find the operand of the GEP that should be checked for consecutive 1071/// stores. This ignores trailing indices that have no effect on the final 1072/// pointer. 1073static unsigned getGEPInductionOperand(DataLayout *DL, 1074 const GetElementPtrInst *Gep) { 1075 unsigned LastOperand = Gep->getNumOperands() - 1; 1076 unsigned GEPAllocSize = DL->getTypeAllocSize( 1077 cast<PointerType>(Gep->getType()->getScalarType())->getElementType()); 1078 1079 // Walk backwards and try to peel off zeros. 1080 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) { 1081 // Find the type we're currently indexing into. 1082 gep_type_iterator GEPTI = gep_type_begin(Gep); 1083 std::advance(GEPTI, LastOperand - 1); 1084 1085 // If it's a type with the same allocation size as the result of the GEP we 1086 // can peel off the zero index. 1087 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize) 1088 break; 1089 --LastOperand; 1090 } 1091 1092 return LastOperand; 1093} 1094 1095int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { 1096 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr"); 1097 // Make sure that the pointer does not point to structs. 1098 if (Ptr->getType()->getPointerElementType()->isAggregateType()) 1099 return 0; 1100 1101 // If this value is a pointer induction variable we know it is consecutive. 1102 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr); 1103 if (Phi && Inductions.count(Phi)) { 1104 InductionInfo II = Inductions[Phi]; 1105 if (IK_PtrInduction == II.IK) 1106 return 1; 1107 else if (IK_ReversePtrInduction == II.IK) 1108 return -1; 1109 } 1110 1111 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr); 1112 if (!Gep) 1113 return 0; 1114 1115 unsigned NumOperands = Gep->getNumOperands(); 1116 Value *GpPtr = Gep->getPointerOperand(); 1117 // If this GEP value is a consecutive pointer induction variable and all of 1118 // the indices are constant then we know it is consecutive. We can 1119 Phi = dyn_cast<PHINode>(GpPtr); 1120 if (Phi && Inductions.count(Phi)) { 1121 1122 // Make sure that the pointer does not point to structs. 1123 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType()); 1124 if (GepPtrType->getElementType()->isAggregateType()) 1125 return 0; 1126 1127 // Make sure that all of the index operands are loop invariant. 1128 for (unsigned i = 1; i < NumOperands; ++i) 1129 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 1130 return 0; 1131 1132 InductionInfo II = Inductions[Phi]; 1133 if (IK_PtrInduction == II.IK) 1134 return 1; 1135 else if (IK_ReversePtrInduction == II.IK) 1136 return -1; 1137 } 1138 1139 unsigned InductionOperand = getGEPInductionOperand(DL, Gep); 1140 1141 // Check that all of the gep indices are uniform except for our induction 1142 // operand. 1143 for (unsigned i = 0; i != NumOperands; ++i) 1144 if (i != InductionOperand && 1145 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 1146 return 0; 1147 1148 // We can emit wide load/stores only if the last non-zero index is the 1149 // induction variable. 1150 const SCEV *Last = SE->getSCEV(Gep->getOperand(InductionOperand)); 1151 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { 1152 const SCEV *Step = AR->getStepRecurrence(*SE); 1153 1154 // The memory is consecutive because the last index is consecutive 1155 // and all other indices are loop invariant. 1156 if (Step->isOne()) 1157 return 1; 1158 if (Step->isAllOnesValue()) 1159 return -1; 1160 } 1161 1162 return 0; 1163} 1164 1165bool LoopVectorizationLegality::isUniform(Value *V) { 1166 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); 1167} 1168 1169InnerLoopVectorizer::VectorParts& 1170InnerLoopVectorizer::getVectorValue(Value *V) { 1171 assert(V != Induction && "The new induction variable should not be used."); 1172 assert(!V->getType()->isVectorTy() && "Can't widen a vector"); 1173 1174 // If we have this scalar in the map, return it. 1175 if (WidenMap.has(V)) 1176 return WidenMap.get(V); 1177 1178 // If this scalar is unknown, assume that it is a constant or that it is 1179 // loop invariant. Broadcast V and save the value for future uses. 1180 Value *B = getBroadcastInstrs(V); 1181 return WidenMap.splat(V, B); 1182} 1183 1184Value *InnerLoopVectorizer::reverseVector(Value *Vec) { 1185 assert(Vec->getType()->isVectorTy() && "Invalid type"); 1186 SmallVector<Constant*, 8> ShuffleMask; 1187 for (unsigned i = 0; i < VF; ++i) 1188 ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); 1189 1190 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), 1191 ConstantVector::get(ShuffleMask), 1192 "reverse"); 1193} 1194 1195 1196void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr, 1197 LoopVectorizationLegality *Legal) { 1198 // Attempt to issue a wide load. 1199 LoadInst *LI = dyn_cast<LoadInst>(Instr); 1200 StoreInst *SI = dyn_cast<StoreInst>(Instr); 1201 1202 assert((LI || SI) && "Invalid Load/Store instruction"); 1203 1204 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType(); 1205 Type *DataTy = VectorType::get(ScalarDataTy, VF); 1206 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 1207 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment(); 1208 // An alignment of 0 means target abi alignment. We need to use the scalar's 1209 // target abi alignment in such a case. 1210 if (!Alignment) 1211 Alignment = DL->getABITypeAlignment(ScalarDataTy); 1212 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); 1213 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy); 1214 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF; 1215 1216 if (ScalarAllocatedSize != VectorElementSize) 1217 return scalarizeInstruction(Instr); 1218 1219 // If the pointer is loop invariant or if it is non consecutive, 1220 // scalarize the load. 1221 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 1222 bool Reverse = ConsecutiveStride < 0; 1223 bool UniformLoad = LI && Legal->isUniform(Ptr); 1224 if (!ConsecutiveStride || UniformLoad) 1225 return scalarizeInstruction(Instr); 1226 1227 Constant *Zero = Builder.getInt32(0); 1228 VectorParts &Entry = WidenMap.get(Instr); 1229 1230 // Handle consecutive loads/stores. 1231 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 1232 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) { 1233 setDebugLocFromInst(Builder, Gep); 1234 Value *PtrOperand = Gep->getPointerOperand(); 1235 Value *FirstBasePtr = getVectorValue(PtrOperand)[0]; 1236 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero); 1237 1238 // Create the new GEP with the new induction variable. 1239 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1240 Gep2->setOperand(0, FirstBasePtr); 1241 Gep2->setName("gep.indvar.base"); 1242 Ptr = Builder.Insert(Gep2); 1243 } else if (Gep) { 1244 setDebugLocFromInst(Builder, Gep); 1245 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), 1246 OrigLoop) && "Base ptr must be invariant"); 1247 1248 // The last index does not have to be the induction. It can be 1249 // consecutive and be a function of the index. For example A[I+1]; 1250 unsigned NumOperands = Gep->getNumOperands(); 1251 unsigned InductionOperand = getGEPInductionOperand(DL, Gep); 1252 // Create the new GEP with the new induction variable. 1253 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1254 1255 for (unsigned i = 0; i < NumOperands; ++i) { 1256 Value *GepOperand = Gep->getOperand(i); 1257 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand); 1258 1259 // Update last index or loop invariant instruction anchored in loop. 1260 if (i == InductionOperand || 1261 (GepOperandInst && OrigLoop->contains(GepOperandInst))) { 1262 assert((i == InductionOperand || 1263 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) && 1264 "Must be last index or loop invariant"); 1265 1266 VectorParts &GEPParts = getVectorValue(GepOperand); 1267 Value *Index = GEPParts[0]; 1268 Index = Builder.CreateExtractElement(Index, Zero); 1269 Gep2->setOperand(i, Index); 1270 Gep2->setName("gep.indvar.idx"); 1271 } 1272 } 1273 Ptr = Builder.Insert(Gep2); 1274 } else { 1275 // Use the induction element ptr. 1276 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 1277 setDebugLocFromInst(Builder, Ptr); 1278 VectorParts &PtrVal = getVectorValue(Ptr); 1279 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 1280 } 1281 1282 // Handle Stores: 1283 if (SI) { 1284 assert(!Legal->isUniform(SI->getPointerOperand()) && 1285 "We do not allow storing to uniform addresses"); 1286 setDebugLocFromInst(Builder, SI); 1287 // We don't want to update the value in the map as it might be used in 1288 // another expression. So don't use a reference type for "StoredVal". 1289 VectorParts StoredVal = getVectorValue(SI->getValueOperand()); 1290 1291 for (unsigned Part = 0; Part < UF; ++Part) { 1292 // Calculate the pointer for the specific unroll-part. 1293 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1294 1295 if (Reverse) { 1296 // If we store to reverse consecutive memory locations then we need 1297 // to reverse the order of elements in the stored value. 1298 StoredVal[Part] = reverseVector(StoredVal[Part]); 1299 // If the address is consecutive but reversed, then the 1300 // wide store needs to start at the last vector element. 1301 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1302 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1303 } 1304 1305 Value *VecPtr = Builder.CreateBitCast(PartPtr, 1306 DataTy->getPointerTo(AddressSpace)); 1307 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment); 1308 } 1309 return; 1310 } 1311 1312 // Handle loads. 1313 assert(LI && "Must have a load instruction"); 1314 setDebugLocFromInst(Builder, LI); 1315 for (unsigned Part = 0; Part < UF; ++Part) { 1316 // Calculate the pointer for the specific unroll-part. 1317 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1318 1319 if (Reverse) { 1320 // If the address is consecutive but reversed, then the 1321 // wide store needs to start at the last vector element. 1322 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1323 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1324 } 1325 1326 Value *VecPtr = Builder.CreateBitCast(PartPtr, 1327 DataTy->getPointerTo(AddressSpace)); 1328 Value *LI = Builder.CreateLoad(VecPtr, "wide.load"); 1329 cast<LoadInst>(LI)->setAlignment(Alignment); 1330 Entry[Part] = Reverse ? reverseVector(LI) : LI; 1331 } 1332} 1333 1334void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { 1335 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 1336 // Holds vector parameters or scalars, in case of uniform vals. 1337 SmallVector<VectorParts, 4> Params; 1338 1339 setDebugLocFromInst(Builder, Instr); 1340 1341 // Find all of the vectorized parameters. 1342 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1343 Value *SrcOp = Instr->getOperand(op); 1344 1345 // If we are accessing the old induction variable, use the new one. 1346 if (SrcOp == OldInduction) { 1347 Params.push_back(getVectorValue(SrcOp)); 1348 continue; 1349 } 1350 1351 // Try using previously calculated values. 1352 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 1353 1354 // If the src is an instruction that appeared earlier in the basic block 1355 // then it should already be vectorized. 1356 if (SrcInst && OrigLoop->contains(SrcInst)) { 1357 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 1358 // The parameter is a vector value from earlier. 1359 Params.push_back(WidenMap.get(SrcInst)); 1360 } else { 1361 // The parameter is a scalar from outside the loop. Maybe even a constant. 1362 VectorParts Scalars; 1363 Scalars.append(UF, SrcOp); 1364 Params.push_back(Scalars); 1365 } 1366 } 1367 1368 assert(Params.size() == Instr->getNumOperands() && 1369 "Invalid number of operands"); 1370 1371 // Does this instruction return a value ? 1372 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 1373 1374 Value *UndefVec = IsVoidRetTy ? 0 : 1375 UndefValue::get(VectorType::get(Instr->getType(), VF)); 1376 // Create a new entry in the WidenMap and initialize it to Undef or Null. 1377 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 1378 1379 // For each vector unroll 'part': 1380 for (unsigned Part = 0; Part < UF; ++Part) { 1381 // For each scalar that we create: 1382 for (unsigned Width = 0; Width < VF; ++Width) { 1383 Instruction *Cloned = Instr->clone(); 1384 if (!IsVoidRetTy) 1385 Cloned->setName(Instr->getName() + ".cloned"); 1386 // Replace the operands of the cloned instructions with extracted scalars. 1387 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1388 Value *Op = Params[op][Part]; 1389 // Param is a vector. Need to extract the right lane. 1390 if (Op->getType()->isVectorTy()) 1391 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); 1392 Cloned->setOperand(op, Op); 1393 } 1394 1395 // Place the cloned scalar in the new loop. 1396 Builder.Insert(Cloned); 1397 1398 // If the original scalar returns a value we need to place it in a vector 1399 // so that future users will be able to use it. 1400 if (!IsVoidRetTy) 1401 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, 1402 Builder.getInt32(Width)); 1403 } 1404 } 1405} 1406 1407Instruction * 1408InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, 1409 Instruction *Loc) { 1410 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck = 1411 Legal->getRuntimePointerCheck(); 1412 1413 if (!PtrRtCheck->Need) 1414 return NULL; 1415 1416 unsigned NumPointers = PtrRtCheck->Pointers.size(); 1417 SmallVector<TrackingVH<Value> , 2> Starts; 1418 SmallVector<TrackingVH<Value> , 2> Ends; 1419 1420 LLVMContext &Ctx = Loc->getContext(); 1421 SCEVExpander Exp(*SE, "induction"); 1422 1423 for (unsigned i = 0; i < NumPointers; ++i) { 1424 Value *Ptr = PtrRtCheck->Pointers[i]; 1425 const SCEV *Sc = SE->getSCEV(Ptr); 1426 1427 if (SE->isLoopInvariant(Sc, OrigLoop)) { 1428 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" << 1429 *Ptr <<"\n"); 1430 Starts.push_back(Ptr); 1431 Ends.push_back(Ptr); 1432 } else { 1433 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n'); 1434 unsigned AS = Ptr->getType()->getPointerAddressSpace(); 1435 1436 // Use this type for pointer arithmetic. 1437 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS); 1438 1439 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc); 1440 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc); 1441 Starts.push_back(Start); 1442 Ends.push_back(End); 1443 } 1444 } 1445 1446 IRBuilder<> ChkBuilder(Loc); 1447 // Our instructions might fold to a constant. 1448 Value *MemoryRuntimeCheck = 0; 1449 for (unsigned i = 0; i < NumPointers; ++i) { 1450 for (unsigned j = i+1; j < NumPointers; ++j) { 1451 // No need to check if two readonly pointers intersect. 1452 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j]) 1453 continue; 1454 1455 // Only need to check pointers between two different dependency sets. 1456 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j]) 1457 continue; 1458 1459 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace(); 1460 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace(); 1461 1462 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) && 1463 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) && 1464 "Trying to bounds check pointers with different address spaces"); 1465 1466 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0); 1467 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1); 1468 1469 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc"); 1470 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc"); 1471 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc"); 1472 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc"); 1473 1474 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0"); 1475 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1"); 1476 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict"); 1477 if (MemoryRuntimeCheck) 1478 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, 1479 "conflict.rdx"); 1480 MemoryRuntimeCheck = IsConflict; 1481 } 1482 } 1483 1484 // We have to do this trickery because the IRBuilder might fold the check to a 1485 // constant expression in which case there is no Instruction anchored in a 1486 // the block. 1487 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck, 1488 ConstantInt::getTrue(Ctx)); 1489 ChkBuilder.Insert(Check, "memcheck.conflict"); 1490 return Check; 1491} 1492 1493void 1494InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { 1495 /* 1496 In this function we generate a new loop. The new loop will contain 1497 the vectorized instructions while the old loop will continue to run the 1498 scalar remainder. 1499 1500 [ ] <-- vector loop bypass (may consist of multiple blocks). 1501 / | 1502 / v 1503 | [ ] <-- vector pre header. 1504 | | 1505 | v 1506 | [ ] \ 1507 | [ ]_| <-- vector loop. 1508 | | 1509 \ v 1510 >[ ] <--- middle-block. 1511 / | 1512 / v 1513 | [ ] <--- new preheader. 1514 | | 1515 | v 1516 | [ ] \ 1517 | [ ]_| <-- old scalar loop to handle remainder. 1518 \ | 1519 \ v 1520 >[ ] <-- exit block. 1521 ... 1522 */ 1523 1524 BasicBlock *OldBasicBlock = OrigLoop->getHeader(); 1525 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader(); 1526 BasicBlock *ExitBlock = OrigLoop->getExitBlock(); 1527 assert(ExitBlock && "Must have an exit block"); 1528 1529 // Some loops have a single integer induction variable, while other loops 1530 // don't. One example is c++ iterators that often have multiple pointer 1531 // induction variables. In the code below we also support a case where we 1532 // don't have a single induction variable. 1533 OldInduction = Legal->getInduction(); 1534 Type *IdxTy = Legal->getWidestInductionType(); 1535 1536 // Find the loop boundaries. 1537 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop); 1538 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count"); 1539 1540 // The exit count might have the type of i64 while the phi is i32. This can 1541 // happen if we have an induction variable that is sign extended before the 1542 // compare. The only way that we get a backedge taken count is that the 1543 // induction variable was signed and as such will not overflow. In such a case 1544 // truncation is legal. 1545 if (ExitCount->getType()->getPrimitiveSizeInBits() > 1546 IdxTy->getPrimitiveSizeInBits()) 1547 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy); 1548 1549 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy); 1550 // Get the total trip count from the count by adding 1. 1551 ExitCount = SE->getAddExpr(ExitCount, 1552 SE->getConstant(ExitCount->getType(), 1)); 1553 1554 // Expand the trip count and place the new instructions in the preheader. 1555 // Notice that the pre-header does not change, only the loop body. 1556 SCEVExpander Exp(*SE, "induction"); 1557 1558 // Count holds the overall loop count (N). 1559 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 1560 BypassBlock->getTerminator()); 1561 1562 // The loop index does not have to start at Zero. Find the original start 1563 // value from the induction PHI node. If we don't have an induction variable 1564 // then we know that it starts at zero. 1565 Builder.SetInsertPoint(BypassBlock->getTerminator()); 1566 Value *StartIdx = ExtendedIdx = OldInduction ? 1567 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock), 1568 IdxTy): 1569 ConstantInt::get(IdxTy, 0); 1570 1571 assert(BypassBlock && "Invalid loop structure"); 1572 LoopBypassBlocks.push_back(BypassBlock); 1573 1574 // Split the single block loop into the two loop structure described above. 1575 BasicBlock *VectorPH = 1576 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); 1577 BasicBlock *VecBody = 1578 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 1579 BasicBlock *MiddleBlock = 1580 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 1581 BasicBlock *ScalarPH = 1582 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 1583 1584 // Create and register the new vector loop. 1585 Loop* Lp = new Loop(); 1586 Loop *ParentLoop = OrigLoop->getParentLoop(); 1587 1588 // Insert the new loop into the loop nest and register the new basic blocks 1589 // before calling any utilities such as SCEV that require valid LoopInfo. 1590 if (ParentLoop) { 1591 ParentLoop->addChildLoop(Lp); 1592 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); 1593 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); 1594 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); 1595 } else { 1596 LI->addTopLevelLoop(Lp); 1597 } 1598 Lp->addBasicBlockToLoop(VecBody, LI->getBase()); 1599 1600 // Use this IR builder to create the loop instructions (Phi, Br, Cmp) 1601 // inside the loop. 1602 Builder.SetInsertPoint(VecBody->getFirstNonPHI()); 1603 1604 // Generate the induction variable. 1605 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction)); 1606 Induction = Builder.CreatePHI(IdxTy, 2, "index"); 1607 // The loop step is equal to the vectorization factor (num of SIMD elements) 1608 // times the unroll factor (num of SIMD instructions). 1609 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 1610 1611 // This is the IR builder that we use to add all of the logic for bypassing 1612 // the new vector loop. 1613 IRBuilder<> BypassBuilder(BypassBlock->getTerminator()); 1614 setDebugLocFromInst(BypassBuilder, 1615 getDebugLocFromInstOrOperands(OldInduction)); 1616 1617 // We may need to extend the index in case there is a type mismatch. 1618 // We know that the count starts at zero and does not overflow. 1619 if (Count->getType() != IdxTy) { 1620 // The exit count can be of pointer type. Convert it to the correct 1621 // integer type. 1622 if (ExitCount->getType()->isPointerTy()) 1623 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int"); 1624 else 1625 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast"); 1626 } 1627 1628 // Add the start index to the loop count to get the new end index. 1629 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx"); 1630 1631 // Now we need to generate the expression for N - (N % VF), which is 1632 // the part that the vectorized body will execute. 1633 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf"); 1634 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec"); 1635 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx, 1636 "end.idx.rnd.down"); 1637 1638 // Now, compare the new count to zero. If it is zero skip the vector loop and 1639 // jump to the scalar loop. 1640 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, 1641 "cmp.zero"); 1642 1643 BasicBlock *LastBypassBlock = BypassBlock; 1644 1645 // Generate the code that checks in runtime if arrays overlap. We put the 1646 // checks into a separate block to make the more common case of few elements 1647 // faster. 1648 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal, 1649 BypassBlock->getTerminator()); 1650 if (MemRuntimeCheck) { 1651 // Create a new block containing the memory check. 1652 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck, 1653 "vector.memcheck"); 1654 if (ParentLoop) 1655 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase()); 1656 LoopBypassBlocks.push_back(CheckBlock); 1657 1658 // Replace the branch into the memory check block with a conditional branch 1659 // for the "few elements case". 1660 Instruction *OldTerm = BypassBlock->getTerminator(); 1661 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); 1662 OldTerm->eraseFromParent(); 1663 1664 Cmp = MemRuntimeCheck; 1665 LastBypassBlock = CheckBlock; 1666 } 1667 1668 LastBypassBlock->getTerminator()->eraseFromParent(); 1669 BranchInst::Create(MiddleBlock, VectorPH, Cmp, 1670 LastBypassBlock); 1671 1672 // We are going to resume the execution of the scalar loop. 1673 // Go over all of the induction variables that we found and fix the 1674 // PHIs that are left in the scalar version of the loop. 1675 // The starting values of PHI nodes depend on the counter of the last 1676 // iteration in the vectorized loop. 1677 // If we come from a bypass edge then we need to start from the original 1678 // start value. 1679 1680 // This variable saves the new starting index for the scalar loop. 1681 PHINode *ResumeIndex = 0; 1682 LoopVectorizationLegality::InductionList::iterator I, E; 1683 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 1684 // Set builder to point to last bypass block. 1685 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator()); 1686 for (I = List->begin(), E = List->end(); I != E; ++I) { 1687 PHINode *OrigPhi = I->first; 1688 LoopVectorizationLegality::InductionInfo II = I->second; 1689 1690 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType(); 1691 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val", 1692 MiddleBlock->getTerminator()); 1693 // We might have extended the type of the induction variable but we need a 1694 // truncated version for the scalar loop. 1695 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ? 1696 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val", 1697 MiddleBlock->getTerminator()) : 0; 1698 1699 Value *EndValue = 0; 1700 switch (II.IK) { 1701 case LoopVectorizationLegality::IK_NoInduction: 1702 llvm_unreachable("Unknown induction"); 1703 case LoopVectorizationLegality::IK_IntInduction: { 1704 // Handle the integer induction counter. 1705 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); 1706 1707 // We have the canonical induction variable. 1708 if (OrigPhi == OldInduction) { 1709 // Create a truncated version of the resume value for the scalar loop, 1710 // we might have promoted the type to a larger width. 1711 EndValue = 1712 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType()); 1713 // The new PHI merges the original incoming value, in case of a bypass, 1714 // or the value at the end of the vectorized loop. 1715 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1716 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1717 TruncResumeVal->addIncoming(EndValue, VecBody); 1718 1719 // We know what the end value is. 1720 EndValue = IdxEndRoundDown; 1721 // We also know which PHI node holds it. 1722 ResumeIndex = ResumeVal; 1723 break; 1724 } 1725 1726 // Not the canonical induction variable - add the vector loop count to the 1727 // start value. 1728 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, 1729 II.StartValue->getType(), 1730 "cast.crd"); 1731 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end"); 1732 break; 1733 } 1734 case LoopVectorizationLegality::IK_ReverseIntInduction: { 1735 // Convert the CountRoundDown variable to the PHI size. 1736 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, 1737 II.StartValue->getType(), 1738 "cast.crd"); 1739 // Handle reverse integer induction counter. 1740 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end"); 1741 break; 1742 } 1743 case LoopVectorizationLegality::IK_PtrInduction: { 1744 // For pointer induction variables, calculate the offset using 1745 // the end index. 1746 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown, 1747 "ptr.ind.end"); 1748 break; 1749 } 1750 case LoopVectorizationLegality::IK_ReversePtrInduction: { 1751 // The value at the end of the loop for the reverse pointer is calculated 1752 // by creating a GEP with a negative index starting from the start value. 1753 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0); 1754 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown, 1755 "rev.ind.end"); 1756 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx, 1757 "rev.ptr.ind.end"); 1758 break; 1759 } 1760 }// end of case 1761 1762 // The new PHI merges the original incoming value, in case of a bypass, 1763 // or the value at the end of the vectorized loop. 1764 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) { 1765 if (OrigPhi == OldInduction) 1766 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]); 1767 else 1768 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1769 } 1770 ResumeVal->addIncoming(EndValue, VecBody); 1771 1772 // Fix the scalar body counter (PHI node). 1773 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 1774 // The old inductions phi node in the scalar body needs the truncated value. 1775 if (OrigPhi == OldInduction) 1776 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal); 1777 else 1778 OrigPhi->setIncomingValue(BlockIdx, ResumeVal); 1779 } 1780 1781 // If we are generating a new induction variable then we also need to 1782 // generate the code that calculates the exit value. This value is not 1783 // simply the end of the counter because we may skip the vectorized body 1784 // in case of a runtime check. 1785 if (!OldInduction){ 1786 assert(!ResumeIndex && "Unexpected resume value found"); 1787 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", 1788 MiddleBlock->getTerminator()); 1789 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1790 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]); 1791 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); 1792 } 1793 1794 // Make sure that we found the index where scalar loop needs to continue. 1795 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() && 1796 "Invalid resume Index"); 1797 1798 // Add a check in the middle block to see if we have completed 1799 // all of the iterations in the first vector loop. 1800 // If (N - N%VF) == N, then we *don't* need to run the remainder. 1801 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, 1802 ResumeIndex, "cmp.n", 1803 MiddleBlock->getTerminator()); 1804 1805 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); 1806 // Remove the old terminator. 1807 MiddleBlock->getTerminator()->eraseFromParent(); 1808 1809 // Create i+1 and fill the PHINode. 1810 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); 1811 Induction->addIncoming(StartIdx, VectorPH); 1812 Induction->addIncoming(NextIdx, VecBody); 1813 // Create the compare. 1814 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); 1815 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); 1816 1817 // Now we have two terminators. Remove the old one from the block. 1818 VecBody->getTerminator()->eraseFromParent(); 1819 1820 // Get ready to start creating new instructions into the vectorized body. 1821 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1822 1823 // Save the state. 1824 LoopVectorPreHeader = VectorPH; 1825 LoopScalarPreHeader = ScalarPH; 1826 LoopMiddleBlock = MiddleBlock; 1827 LoopExitBlock = ExitBlock; 1828 LoopVectorBody = VecBody; 1829 LoopScalarBody = OldBasicBlock; 1830 1831 LoopVectorizeHints Hints(Lp, true); 1832 Hints.setAlreadyVectorized(Lp); 1833} 1834 1835/// This function returns the identity element (or neutral element) for 1836/// the operation K. 1837Constant* 1838LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) { 1839 switch (K) { 1840 case RK_IntegerXor: 1841 case RK_IntegerAdd: 1842 case RK_IntegerOr: 1843 // Adding, Xoring, Oring zero to a number does not change it. 1844 return ConstantInt::get(Tp, 0); 1845 case RK_IntegerMult: 1846 // Multiplying a number by 1 does not change it. 1847 return ConstantInt::get(Tp, 1); 1848 case RK_IntegerAnd: 1849 // AND-ing a number with an all-1 value does not change it. 1850 return ConstantInt::get(Tp, -1, true); 1851 case RK_FloatMult: 1852 // Multiplying a number by 1 does not change it. 1853 return ConstantFP::get(Tp, 1.0L); 1854 case RK_FloatAdd: 1855 // Adding zero to a number does not change it. 1856 return ConstantFP::get(Tp, 0.0L); 1857 default: 1858 llvm_unreachable("Unknown reduction kind"); 1859 } 1860} 1861 1862static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I, 1863 Intrinsic::ID ValidIntrinsicID) { 1864 if (I.getNumArgOperands() != 1 || 1865 !I.getArgOperand(0)->getType()->isFloatingPointTy() || 1866 I.getType() != I.getArgOperand(0)->getType() || 1867 !I.onlyReadsMemory()) 1868 return Intrinsic::not_intrinsic; 1869 1870 return ValidIntrinsicID; 1871} 1872 1873static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I, 1874 Intrinsic::ID ValidIntrinsicID) { 1875 if (I.getNumArgOperands() != 2 || 1876 !I.getArgOperand(0)->getType()->isFloatingPointTy() || 1877 !I.getArgOperand(1)->getType()->isFloatingPointTy() || 1878 I.getType() != I.getArgOperand(0)->getType() || 1879 I.getType() != I.getArgOperand(1)->getType() || 1880 !I.onlyReadsMemory()) 1881 return Intrinsic::not_intrinsic; 1882 1883 return ValidIntrinsicID; 1884} 1885 1886 1887static Intrinsic::ID 1888getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) { 1889 // If we have an intrinsic call, check if it is trivially vectorizable. 1890 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) { 1891 switch (II->getIntrinsicID()) { 1892 case Intrinsic::sqrt: 1893 case Intrinsic::sin: 1894 case Intrinsic::cos: 1895 case Intrinsic::exp: 1896 case Intrinsic::exp2: 1897 case Intrinsic::log: 1898 case Intrinsic::log10: 1899 case Intrinsic::log2: 1900 case Intrinsic::fabs: 1901 case Intrinsic::copysign: 1902 case Intrinsic::floor: 1903 case Intrinsic::ceil: 1904 case Intrinsic::trunc: 1905 case Intrinsic::rint: 1906 case Intrinsic::nearbyint: 1907 case Intrinsic::round: 1908 case Intrinsic::pow: 1909 case Intrinsic::fma: 1910 case Intrinsic::fmuladd: 1911 case Intrinsic::lifetime_start: 1912 case Intrinsic::lifetime_end: 1913 return II->getIntrinsicID(); 1914 default: 1915 return Intrinsic::not_intrinsic; 1916 } 1917 } 1918 1919 if (!TLI) 1920 return Intrinsic::not_intrinsic; 1921 1922 LibFunc::Func Func; 1923 Function *F = CI->getCalledFunction(); 1924 // We're going to make assumptions on the semantics of the functions, check 1925 // that the target knows that it's available in this environment and it does 1926 // not have local linkage. 1927 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func)) 1928 return Intrinsic::not_intrinsic; 1929 1930 // Otherwise check if we have a call to a function that can be turned into a 1931 // vector intrinsic. 1932 switch (Func) { 1933 default: 1934 break; 1935 case LibFunc::sin: 1936 case LibFunc::sinf: 1937 case LibFunc::sinl: 1938 return checkUnaryFloatSignature(*CI, Intrinsic::sin); 1939 case LibFunc::cos: 1940 case LibFunc::cosf: 1941 case LibFunc::cosl: 1942 return checkUnaryFloatSignature(*CI, Intrinsic::cos); 1943 case LibFunc::exp: 1944 case LibFunc::expf: 1945 case LibFunc::expl: 1946 return checkUnaryFloatSignature(*CI, Intrinsic::exp); 1947 case LibFunc::exp2: 1948 case LibFunc::exp2f: 1949 case LibFunc::exp2l: 1950 return checkUnaryFloatSignature(*CI, Intrinsic::exp2); 1951 case LibFunc::log: 1952 case LibFunc::logf: 1953 case LibFunc::logl: 1954 return checkUnaryFloatSignature(*CI, Intrinsic::log); 1955 case LibFunc::log10: 1956 case LibFunc::log10f: 1957 case LibFunc::log10l: 1958 return checkUnaryFloatSignature(*CI, Intrinsic::log10); 1959 case LibFunc::log2: 1960 case LibFunc::log2f: 1961 case LibFunc::log2l: 1962 return checkUnaryFloatSignature(*CI, Intrinsic::log2); 1963 case LibFunc::fabs: 1964 case LibFunc::fabsf: 1965 case LibFunc::fabsl: 1966 return checkUnaryFloatSignature(*CI, Intrinsic::fabs); 1967 case LibFunc::copysign: 1968 case LibFunc::copysignf: 1969 case LibFunc::copysignl: 1970 return checkBinaryFloatSignature(*CI, Intrinsic::copysign); 1971 case LibFunc::floor: 1972 case LibFunc::floorf: 1973 case LibFunc::floorl: 1974 return checkUnaryFloatSignature(*CI, Intrinsic::floor); 1975 case LibFunc::ceil: 1976 case LibFunc::ceilf: 1977 case LibFunc::ceill: 1978 return checkUnaryFloatSignature(*CI, Intrinsic::ceil); 1979 case LibFunc::trunc: 1980 case LibFunc::truncf: 1981 case LibFunc::truncl: 1982 return checkUnaryFloatSignature(*CI, Intrinsic::trunc); 1983 case LibFunc::rint: 1984 case LibFunc::rintf: 1985 case LibFunc::rintl: 1986 return checkUnaryFloatSignature(*CI, Intrinsic::rint); 1987 case LibFunc::nearbyint: 1988 case LibFunc::nearbyintf: 1989 case LibFunc::nearbyintl: 1990 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint); 1991 case LibFunc::round: 1992 case LibFunc::roundf: 1993 case LibFunc::roundl: 1994 return checkUnaryFloatSignature(*CI, Intrinsic::round); 1995 case LibFunc::pow: 1996 case LibFunc::powf: 1997 case LibFunc::powl: 1998 return checkBinaryFloatSignature(*CI, Intrinsic::pow); 1999 } 2000 2001 return Intrinsic::not_intrinsic; 2002} 2003 2004/// This function translates the reduction kind to an LLVM binary operator. 2005static unsigned 2006getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) { 2007 switch (Kind) { 2008 case LoopVectorizationLegality::RK_IntegerAdd: 2009 return Instruction::Add; 2010 case LoopVectorizationLegality::RK_IntegerMult: 2011 return Instruction::Mul; 2012 case LoopVectorizationLegality::RK_IntegerOr: 2013 return Instruction::Or; 2014 case LoopVectorizationLegality::RK_IntegerAnd: 2015 return Instruction::And; 2016 case LoopVectorizationLegality::RK_IntegerXor: 2017 return Instruction::Xor; 2018 case LoopVectorizationLegality::RK_FloatMult: 2019 return Instruction::FMul; 2020 case LoopVectorizationLegality::RK_FloatAdd: 2021 return Instruction::FAdd; 2022 case LoopVectorizationLegality::RK_IntegerMinMax: 2023 return Instruction::ICmp; 2024 case LoopVectorizationLegality::RK_FloatMinMax: 2025 return Instruction::FCmp; 2026 default: 2027 llvm_unreachable("Unknown reduction operation"); 2028 } 2029} 2030 2031Value *createMinMaxOp(IRBuilder<> &Builder, 2032 LoopVectorizationLegality::MinMaxReductionKind RK, 2033 Value *Left, 2034 Value *Right) { 2035 CmpInst::Predicate P = CmpInst::ICMP_NE; 2036 switch (RK) { 2037 default: 2038 llvm_unreachable("Unknown min/max reduction kind"); 2039 case LoopVectorizationLegality::MRK_UIntMin: 2040 P = CmpInst::ICMP_ULT; 2041 break; 2042 case LoopVectorizationLegality::MRK_UIntMax: 2043 P = CmpInst::ICMP_UGT; 2044 break; 2045 case LoopVectorizationLegality::MRK_SIntMin: 2046 P = CmpInst::ICMP_SLT; 2047 break; 2048 case LoopVectorizationLegality::MRK_SIntMax: 2049 P = CmpInst::ICMP_SGT; 2050 break; 2051 case LoopVectorizationLegality::MRK_FloatMin: 2052 P = CmpInst::FCMP_OLT; 2053 break; 2054 case LoopVectorizationLegality::MRK_FloatMax: 2055 P = CmpInst::FCMP_OGT; 2056 break; 2057 } 2058 2059 Value *Cmp; 2060 if (RK == LoopVectorizationLegality::MRK_FloatMin || 2061 RK == LoopVectorizationLegality::MRK_FloatMax) 2062 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp"); 2063 else 2064 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); 2065 2066 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); 2067 return Select; 2068} 2069 2070namespace { 2071struct CSEDenseMapInfo { 2072 static bool canHandle(Instruction *I) { 2073 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 2074 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); 2075 } 2076 static inline Instruction *getEmptyKey() { 2077 return DenseMapInfo<Instruction *>::getEmptyKey(); 2078 } 2079 static inline Instruction *getTombstoneKey() { 2080 return DenseMapInfo<Instruction *>::getTombstoneKey(); 2081 } 2082 static unsigned getHashValue(Instruction *I) { 2083 assert(canHandle(I) && "Unknown instruction!"); 2084 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), 2085 I->value_op_end())); 2086 } 2087 static bool isEqual(Instruction *LHS, Instruction *RHS) { 2088 if (LHS == getEmptyKey() || RHS == getEmptyKey() || 2089 LHS == getTombstoneKey() || RHS == getTombstoneKey()) 2090 return LHS == RHS; 2091 return LHS->isIdenticalTo(RHS); 2092 } 2093}; 2094} 2095 2096///\brief Perform cse of induction variable instructions. 2097static void cse(BasicBlock *BB) { 2098 // Perform simple cse. 2099 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; 2100 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { 2101 Instruction *In = I++; 2102 2103 if (!CSEDenseMapInfo::canHandle(In)) 2104 continue; 2105 2106 // Check if we can replace this instruction with any of the 2107 // visited instructions. 2108 if (Instruction *V = CSEMap.lookup(In)) { 2109 In->replaceAllUsesWith(V); 2110 In->eraseFromParent(); 2111 continue; 2112 } 2113 2114 CSEMap[In] = In; 2115 } 2116} 2117 2118void 2119InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { 2120 //===------------------------------------------------===// 2121 // 2122 // Notice: any optimization or new instruction that go 2123 // into the code below should be also be implemented in 2124 // the cost-model. 2125 // 2126 //===------------------------------------------------===// 2127 Constant *Zero = Builder.getInt32(0); 2128 2129 // In order to support reduction variables we need to be able to vectorize 2130 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two 2131 // stages. First, we create a new vector PHI node with no incoming edges. 2132 // We use this value when we vectorize all of the instructions that use the 2133 // PHI. Next, after all of the instructions in the block are complete we 2134 // add the new incoming edges to the PHI. At this point all of the 2135 // instructions in the basic block are vectorized, so we can use them to 2136 // construct the PHI. 2137 PhiVector RdxPHIsToFix; 2138 2139 // Scan the loop in a topological order to ensure that defs are vectorized 2140 // before users. 2141 LoopBlocksDFS DFS(OrigLoop); 2142 DFS.perform(LI); 2143 2144 // Vectorize all of the blocks in the original loop. 2145 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 2146 be = DFS.endRPO(); bb != be; ++bb) 2147 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix); 2148 2149 // At this point every instruction in the original loop is widened to 2150 // a vector form. We are almost done. Now, we need to fix the PHI nodes 2151 // that we vectorized. The PHI nodes are currently empty because we did 2152 // not want to introduce cycles. Notice that the remaining PHI nodes 2153 // that we need to fix are reduction variables. 2154 2155 // Create the 'reduced' values for each of the induction vars. 2156 // The reduced values are the vector values that we scalarize and combine 2157 // after the loop is finished. 2158 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); 2159 it != e; ++it) { 2160 PHINode *RdxPhi = *it; 2161 assert(RdxPhi && "Unable to recover vectorized PHI"); 2162 2163 // Find the reduction variable descriptor. 2164 assert(Legal->getReductionVars()->count(RdxPhi) && 2165 "Unable to find the reduction variable"); 2166 LoopVectorizationLegality::ReductionDescriptor RdxDesc = 2167 (*Legal->getReductionVars())[RdxPhi]; 2168 2169 setDebugLocFromInst(Builder, RdxDesc.StartValue); 2170 2171 // We need to generate a reduction vector from the incoming scalar. 2172 // To do so, we need to generate the 'identity' vector and overide 2173 // one of the elements with the incoming scalar reduction. We need 2174 // to do it in the vector-loop preheader. 2175 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator()); 2176 2177 // This is the vector-clone of the value that leaves the loop. 2178 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr); 2179 Type *VecTy = VectorExit[0]->getType(); 2180 2181 // Find the reduction identity variable. Zero for addition, or, xor, 2182 // one for multiplication, -1 for And. 2183 Value *Identity; 2184 Value *VectorStart; 2185 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax || 2186 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) { 2187 // MinMax reduction have the start value as their identify. 2188 if (VF == 1) { 2189 VectorStart = Identity = RdxDesc.StartValue; 2190 } else { 2191 VectorStart = Identity = Builder.CreateVectorSplat(VF, 2192 RdxDesc.StartValue, 2193 "minmax.ident"); 2194 } 2195 } else { 2196 // Handle other reduction kinds: 2197 Constant *Iden = 2198 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind, 2199 VecTy->getScalarType()); 2200 if (VF == 1) { 2201 Identity = Iden; 2202 // This vector is the Identity vector where the first element is the 2203 // incoming scalar reduction. 2204 VectorStart = RdxDesc.StartValue; 2205 } else { 2206 Identity = ConstantVector::getSplat(VF, Iden); 2207 2208 // This vector is the Identity vector where the first element is the 2209 // incoming scalar reduction. 2210 VectorStart = Builder.CreateInsertElement(Identity, 2211 RdxDesc.StartValue, Zero); 2212 } 2213 } 2214 2215 // Fix the vector-loop phi. 2216 // We created the induction variable so we know that the 2217 // preheader is the first entry. 2218 BasicBlock *VecPreheader = Induction->getIncomingBlock(0); 2219 2220 // Reductions do not have to start at zero. They can start with 2221 // any loop invariant values. 2222 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 2223 BasicBlock *Latch = OrigLoop->getLoopLatch(); 2224 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 2225 VectorParts &Val = getVectorValue(LoopVal); 2226 for (unsigned part = 0; part < UF; ++part) { 2227 // Make sure to add the reduction stat value only to the 2228 // first unroll part. 2229 Value *StartVal = (part == 0) ? VectorStart : Identity; 2230 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader); 2231 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody); 2232 } 2233 2234 // Before each round, move the insertion point right between 2235 // the PHIs and the values we are going to write. 2236 // This allows us to write both PHINodes and the extractelement 2237 // instructions. 2238 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); 2239 2240 VectorParts RdxParts; 2241 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr); 2242 for (unsigned part = 0; part < UF; ++part) { 2243 // This PHINode contains the vectorized reduction variable, or 2244 // the initial value vector, if we bypass the vector loop. 2245 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); 2246 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); 2247 Value *StartVal = (part == 0) ? VectorStart : Identity; 2248 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 2249 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]); 2250 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody); 2251 RdxParts.push_back(NewPhi); 2252 } 2253 2254 // Reduce all of the unrolled parts into a single vector. 2255 Value *ReducedPartRdx = RdxParts[0]; 2256 unsigned Op = getReductionBinOp(RdxDesc.Kind); 2257 setDebugLocFromInst(Builder, ReducedPartRdx); 2258 for (unsigned part = 1; part < UF; ++part) { 2259 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 2260 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op, 2261 RdxParts[part], ReducedPartRdx, 2262 "bin.rdx"); 2263 else 2264 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind, 2265 ReducedPartRdx, RdxParts[part]); 2266 } 2267 2268 if (VF > 1) { 2269 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 2270 // and vector ops, reducing the set of values being computed by half each 2271 // round. 2272 assert(isPowerOf2_32(VF) && 2273 "Reduction emission only supported for pow2 vectors!"); 2274 Value *TmpVec = ReducedPartRdx; 2275 SmallVector<Constant*, 32> ShuffleMask(VF, 0); 2276 for (unsigned i = VF; i != 1; i >>= 1) { 2277 // Move the upper half of the vector to the lower half. 2278 for (unsigned j = 0; j != i/2; ++j) 2279 ShuffleMask[j] = Builder.getInt32(i/2 + j); 2280 2281 // Fill the rest of the mask with undef. 2282 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 2283 UndefValue::get(Builder.getInt32Ty())); 2284 2285 Value *Shuf = 2286 Builder.CreateShuffleVector(TmpVec, 2287 UndefValue::get(TmpVec->getType()), 2288 ConstantVector::get(ShuffleMask), 2289 "rdx.shuf"); 2290 2291 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 2292 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf, 2293 "bin.rdx"); 2294 else 2295 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf); 2296 } 2297 2298 // The result is in the first element of the vector. 2299 ReducedPartRdx = Builder.CreateExtractElement(TmpVec, 2300 Builder.getInt32(0)); 2301 } 2302 2303 // Now, we need to fix the users of the reduction variable 2304 // inside and outside of the scalar remainder loop. 2305 // We know that the loop is in LCSSA form. We need to update the 2306 // PHI nodes in the exit blocks. 2307 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 2308 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 2309 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 2310 if (!LCSSAPhi) break; 2311 2312 // All PHINodes need to have a single entry edge, or two if 2313 // we already fixed them. 2314 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 2315 2316 // We found our reduction value exit-PHI. Update it with the 2317 // incoming bypass edge. 2318 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { 2319 // Add an edge coming from the bypass. 2320 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); 2321 break; 2322 } 2323 }// end of the LCSSA phi scan. 2324 2325 // Fix the scalar loop reduction variable with the incoming reduction sum 2326 // from the vector body and from the backedge value. 2327 int IncomingEdgeBlockIdx = 2328 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 2329 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 2330 // Pick the other block. 2331 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 2332 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx); 2333 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); 2334 }// end of for each redux variable. 2335 2336 fixLCSSAPHIs(); 2337 2338 // Remove redundant induction instructions. 2339 cse(LoopVectorBody); 2340} 2341 2342void InnerLoopVectorizer::fixLCSSAPHIs() { 2343 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 2344 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 2345 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 2346 if (!LCSSAPhi) break; 2347 if (LCSSAPhi->getNumIncomingValues() == 1) 2348 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 2349 LoopMiddleBlock); 2350 } 2351} 2352 2353InnerLoopVectorizer::VectorParts 2354InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 2355 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 2356 "Invalid edge"); 2357 2358 // Look for cached value. 2359 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst); 2360 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge); 2361 if (ECEntryIt != MaskCache.end()) 2362 return ECEntryIt->second; 2363 2364 VectorParts SrcMask = createBlockInMask(Src); 2365 2366 // The terminator has to be a branch inst! 2367 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 2368 assert(BI && "Unexpected terminator found"); 2369 2370 if (BI->isConditional()) { 2371 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 2372 2373 if (BI->getSuccessor(0) != Dst) 2374 for (unsigned part = 0; part < UF; ++part) 2375 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 2376 2377 for (unsigned part = 0; part < UF; ++part) 2378 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 2379 2380 MaskCache[Edge] = EdgeMask; 2381 return EdgeMask; 2382 } 2383 2384 MaskCache[Edge] = SrcMask; 2385 return SrcMask; 2386} 2387 2388InnerLoopVectorizer::VectorParts 2389InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 2390 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 2391 2392 // Loop incoming mask is all-one. 2393 if (OrigLoop->getHeader() == BB) { 2394 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 2395 return getVectorValue(C); 2396 } 2397 2398 // This is the block mask. We OR all incoming edges, and with zero. 2399 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 2400 VectorParts BlockMask = getVectorValue(Zero); 2401 2402 // For each pred: 2403 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 2404 VectorParts EM = createEdgeMask(*it, BB); 2405 for (unsigned part = 0; part < UF; ++part) 2406 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 2407 } 2408 2409 return BlockMask; 2410} 2411 2412void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, 2413 InnerLoopVectorizer::VectorParts &Entry, 2414 LoopVectorizationLegality *Legal, 2415 unsigned UF, unsigned VF, PhiVector *PV) { 2416 PHINode* P = cast<PHINode>(PN); 2417 // Handle reduction variables: 2418 if (Legal->getReductionVars()->count(P)) { 2419 for (unsigned part = 0; part < UF; ++part) { 2420 // This is phase one of vectorizing PHIs. 2421 Type *VecTy = (VF == 1) ? PN->getType() : 2422 VectorType::get(PN->getType(), VF); 2423 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", 2424 LoopVectorBody-> getFirstInsertionPt()); 2425 } 2426 PV->push_back(P); 2427 return; 2428 } 2429 2430 setDebugLocFromInst(Builder, P); 2431 // Check for PHI nodes that are lowered to vector selects. 2432 if (P->getParent() != OrigLoop->getHeader()) { 2433 // We know that all PHIs in non header blocks are converted into 2434 // selects, so we don't have to worry about the insertion order and we 2435 // can just use the builder. 2436 // At this point we generate the predication tree. There may be 2437 // duplications since this is a simple recursive scan, but future 2438 // optimizations will clean it up. 2439 2440 unsigned NumIncoming = P->getNumIncomingValues(); 2441 2442 // Generate a sequence of selects of the form: 2443 // SELECT(Mask3, In3, 2444 // SELECT(Mask2, In2, 2445 // ( ...))) 2446 for (unsigned In = 0; In < NumIncoming; In++) { 2447 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), 2448 P->getParent()); 2449 VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); 2450 2451 for (unsigned part = 0; part < UF; ++part) { 2452 // We might have single edge PHIs (blocks) - use an identity 2453 // 'select' for the first PHI operand. 2454 if (In == 0) 2455 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2456 In0[part]); 2457 else 2458 // Select between the current value and the previous incoming edge 2459 // based on the incoming mask. 2460 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2461 Entry[part], "predphi"); 2462 } 2463 } 2464 return; 2465 } 2466 2467 // This PHINode must be an induction variable. 2468 // Make sure that we know about it. 2469 assert(Legal->getInductionVars()->count(P) && 2470 "Not an induction variable"); 2471 2472 LoopVectorizationLegality::InductionInfo II = 2473 Legal->getInductionVars()->lookup(P); 2474 2475 switch (II.IK) { 2476 case LoopVectorizationLegality::IK_NoInduction: 2477 llvm_unreachable("Unknown induction"); 2478 case LoopVectorizationLegality::IK_IntInduction: { 2479 assert(P->getType() == II.StartValue->getType() && "Types must match"); 2480 Type *PhiTy = P->getType(); 2481 Value *Broadcasted; 2482 if (P == OldInduction) { 2483 // Handle the canonical induction variable. We might have had to 2484 // extend the type. 2485 Broadcasted = Builder.CreateTrunc(Induction, PhiTy); 2486 } else { 2487 // Handle other induction variables that are now based on the 2488 // canonical one. 2489 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx, 2490 "normalized.idx"); 2491 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy); 2492 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx, 2493 "offset.idx"); 2494 } 2495 Broadcasted = getBroadcastInstrs(Broadcasted); 2496 // After broadcasting the induction variable we need to make the vector 2497 // consecutive by adding 0, 1, 2, etc. 2498 for (unsigned part = 0; part < UF; ++part) 2499 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); 2500 return; 2501 } 2502 case LoopVectorizationLegality::IK_ReverseIntInduction: 2503 case LoopVectorizationLegality::IK_PtrInduction: 2504 case LoopVectorizationLegality::IK_ReversePtrInduction: 2505 // Handle reverse integer and pointer inductions. 2506 Value *StartIdx = ExtendedIdx; 2507 // This is the normalized GEP that starts counting at zero. 2508 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx, 2509 "normalized.idx"); 2510 2511 // Handle the reverse integer induction variable case. 2512 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) { 2513 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType()); 2514 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy, 2515 "resize.norm.idx"); 2516 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI, 2517 "reverse.idx"); 2518 2519 // This is a new value so do not hoist it out. 2520 Value *Broadcasted = getBroadcastInstrs(ReverseInd); 2521 // After broadcasting the induction variable we need to make the 2522 // vector consecutive by adding ... -3, -2, -1, 0. 2523 for (unsigned part = 0; part < UF; ++part) 2524 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part, 2525 true); 2526 return; 2527 } 2528 2529 // Handle the pointer induction variable case. 2530 assert(P->getType()->isPointerTy() && "Unexpected type."); 2531 2532 // Is this a reverse induction ptr or a consecutive induction ptr. 2533 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction == 2534 II.IK); 2535 2536 // This is the vector of results. Notice that we don't generate 2537 // vector geps because scalar geps result in better code. 2538 for (unsigned part = 0; part < UF; ++part) { 2539 if (VF == 1) { 2540 int EltIndex = (part) * (Reverse ? -1 : 1); 2541 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 2542 Value *GlobalIdx; 2543 if (Reverse) 2544 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); 2545 else 2546 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); 2547 2548 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 2549 "next.gep"); 2550 Entry[part] = SclrGep; 2551 continue; 2552 } 2553 2554 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 2555 for (unsigned int i = 0; i < VF; ++i) { 2556 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1); 2557 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 2558 Value *GlobalIdx; 2559 if (!Reverse) 2560 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); 2561 else 2562 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); 2563 2564 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 2565 "next.gep"); 2566 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 2567 Builder.getInt32(i), 2568 "insert.gep"); 2569 } 2570 Entry[part] = VecVal; 2571 } 2572 return; 2573 } 2574} 2575 2576void 2577InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, 2578 BasicBlock *BB, PhiVector *PV) { 2579 // For each instruction in the old loop. 2580 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 2581 VectorParts &Entry = WidenMap.get(it); 2582 switch (it->getOpcode()) { 2583 case Instruction::Br: 2584 // Nothing to do for PHIs and BR, since we already took care of the 2585 // loop control flow instructions. 2586 continue; 2587 case Instruction::PHI:{ 2588 // Vectorize PHINodes. 2589 widenPHIInstruction(it, Entry, Legal, UF, VF, PV); 2590 continue; 2591 }// End of PHI. 2592 2593 case Instruction::Add: 2594 case Instruction::FAdd: 2595 case Instruction::Sub: 2596 case Instruction::FSub: 2597 case Instruction::Mul: 2598 case Instruction::FMul: 2599 case Instruction::UDiv: 2600 case Instruction::SDiv: 2601 case Instruction::FDiv: 2602 case Instruction::URem: 2603 case Instruction::SRem: 2604 case Instruction::FRem: 2605 case Instruction::Shl: 2606 case Instruction::LShr: 2607 case Instruction::AShr: 2608 case Instruction::And: 2609 case Instruction::Or: 2610 case Instruction::Xor: { 2611 // Just widen binops. 2612 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 2613 setDebugLocFromInst(Builder, BinOp); 2614 VectorParts &A = getVectorValue(it->getOperand(0)); 2615 VectorParts &B = getVectorValue(it->getOperand(1)); 2616 2617 // Use this vector value for all users of the original instruction. 2618 for (unsigned Part = 0; Part < UF; ++Part) { 2619 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 2620 2621 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef. 2622 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V); 2623 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) { 2624 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); 2625 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); 2626 } 2627 if (VecOp && isa<PossiblyExactOperator>(VecOp)) 2628 VecOp->setIsExact(BinOp->isExact()); 2629 2630 Entry[Part] = V; 2631 } 2632 break; 2633 } 2634 case Instruction::Select: { 2635 // Widen selects. 2636 // If the selector is loop invariant we can create a select 2637 // instruction with a scalar condition. Otherwise, use vector-select. 2638 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), 2639 OrigLoop); 2640 setDebugLocFromInst(Builder, it); 2641 2642 // The condition can be loop invariant but still defined inside the 2643 // loop. This means that we can't just use the original 'cond' value. 2644 // We have to take the 'vectorized' value and pick the first lane. 2645 // Instcombine will make this a no-op. 2646 VectorParts &Cond = getVectorValue(it->getOperand(0)); 2647 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 2648 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 2649 2650 Value *ScalarCond = (VF == 1) ? Cond[0] : 2651 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0)); 2652 2653 for (unsigned Part = 0; Part < UF; ++Part) { 2654 Entry[Part] = Builder.CreateSelect( 2655 InvariantCond ? ScalarCond : Cond[Part], 2656 Op0[Part], 2657 Op1[Part]); 2658 } 2659 break; 2660 } 2661 2662 case Instruction::ICmp: 2663 case Instruction::FCmp: { 2664 // Widen compares. Generate vector compares. 2665 bool FCmp = (it->getOpcode() == Instruction::FCmp); 2666 CmpInst *Cmp = dyn_cast<CmpInst>(it); 2667 setDebugLocFromInst(Builder, it); 2668 VectorParts &A = getVectorValue(it->getOperand(0)); 2669 VectorParts &B = getVectorValue(it->getOperand(1)); 2670 for (unsigned Part = 0; Part < UF; ++Part) { 2671 Value *C = 0; 2672 if (FCmp) 2673 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 2674 else 2675 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 2676 Entry[Part] = C; 2677 } 2678 break; 2679 } 2680 2681 case Instruction::Store: 2682 case Instruction::Load: 2683 vectorizeMemoryInstruction(it, Legal); 2684 break; 2685 case Instruction::ZExt: 2686 case Instruction::SExt: 2687 case Instruction::FPToUI: 2688 case Instruction::FPToSI: 2689 case Instruction::FPExt: 2690 case Instruction::PtrToInt: 2691 case Instruction::IntToPtr: 2692 case Instruction::SIToFP: 2693 case Instruction::UIToFP: 2694 case Instruction::Trunc: 2695 case Instruction::FPTrunc: 2696 case Instruction::BitCast: { 2697 CastInst *CI = dyn_cast<CastInst>(it); 2698 setDebugLocFromInst(Builder, it); 2699 /// Optimize the special case where the source is the induction 2700 /// variable. Notice that we can only optimize the 'trunc' case 2701 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 2702 /// c. other casts depend on pointer size. 2703 if (CI->getOperand(0) == OldInduction && 2704 it->getOpcode() == Instruction::Trunc) { 2705 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 2706 CI->getType()); 2707 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 2708 for (unsigned Part = 0; Part < UF; ++Part) 2709 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false); 2710 break; 2711 } 2712 /// Vectorize casts. 2713 Type *DestTy = (VF == 1) ? CI->getType() : 2714 VectorType::get(CI->getType(), VF); 2715 2716 VectorParts &A = getVectorValue(it->getOperand(0)); 2717 for (unsigned Part = 0; Part < UF; ++Part) 2718 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 2719 break; 2720 } 2721 2722 case Instruction::Call: { 2723 // Ignore dbg intrinsics. 2724 if (isa<DbgInfoIntrinsic>(it)) 2725 break; 2726 setDebugLocFromInst(Builder, it); 2727 2728 Module *M = BB->getParent()->getParent(); 2729 CallInst *CI = cast<CallInst>(it); 2730 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 2731 assert(ID && "Not an intrinsic call!"); 2732 switch (ID) { 2733 case Intrinsic::lifetime_end: 2734 case Intrinsic::lifetime_start: 2735 scalarizeInstruction(it); 2736 break; 2737 default: 2738 for (unsigned Part = 0; Part < UF; ++Part) { 2739 SmallVector<Value *, 4> Args; 2740 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 2741 VectorParts &Arg = getVectorValue(CI->getArgOperand(i)); 2742 Args.push_back(Arg[Part]); 2743 } 2744 Type *Tys[] = {CI->getType()}; 2745 if (VF > 1) 2746 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF); 2747 2748 Function *F = Intrinsic::getDeclaration(M, ID, Tys); 2749 Entry[Part] = Builder.CreateCall(F, Args); 2750 } 2751 break; 2752 } 2753 break; 2754 } 2755 2756 default: 2757 // All other instructions are unsupported. Scalarize them. 2758 scalarizeInstruction(it); 2759 break; 2760 }// end of switch. 2761 }// end of for_each instr. 2762} 2763 2764void InnerLoopVectorizer::updateAnalysis() { 2765 // Forget the original basic block. 2766 SE->forgetLoop(OrigLoop); 2767 2768 // Update the dominator tree information. 2769 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && 2770 "Entry does not dominate exit."); 2771 2772 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2773 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); 2774 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); 2775 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); 2776 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front()); 2777 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); 2778 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 2779 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); 2780 2781 DEBUG(DT->verifyAnalysis()); 2782} 2783 2784/// \brief Check whether it is safe to if-convert this phi node. 2785/// 2786/// Phi nodes with constant expressions that can trap are not safe to if 2787/// convert. 2788static bool canIfConvertPHINodes(BasicBlock *BB) { 2789 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 2790 PHINode *Phi = dyn_cast<PHINode>(I); 2791 if (!Phi) 2792 return true; 2793 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p) 2794 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p))) 2795 if (C->canTrap()) 2796 return false; 2797 } 2798 return true; 2799} 2800 2801bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 2802 if (!EnableIfConversion) 2803 return false; 2804 2805 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 2806 2807 // A list of pointers that we can safely read and write to. 2808 SmallPtrSet<Value *, 8> SafePointes; 2809 2810 // Collect safe addresses. 2811 for (Loop::block_iterator BI = TheLoop->block_begin(), 2812 BE = TheLoop->block_end(); BI != BE; ++BI) { 2813 BasicBlock *BB = *BI; 2814 2815 if (blockNeedsPredication(BB)) 2816 continue; 2817 2818 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 2819 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 2820 SafePointes.insert(LI->getPointerOperand()); 2821 else if (StoreInst *SI = dyn_cast<StoreInst>(I)) 2822 SafePointes.insert(SI->getPointerOperand()); 2823 } 2824 } 2825 2826 // Collect the blocks that need predication. 2827 BasicBlock *Header = TheLoop->getHeader(); 2828 for (Loop::block_iterator BI = TheLoop->block_begin(), 2829 BE = TheLoop->block_end(); BI != BE; ++BI) { 2830 BasicBlock *BB = *BI; 2831 2832 // We don't support switch statements inside loops. 2833 if (!isa<BranchInst>(BB->getTerminator())) 2834 return false; 2835 2836 // We must be able to predicate all blocks that need to be predicated. 2837 if (blockNeedsPredication(BB)) { 2838 if (!blockCanBePredicated(BB, SafePointes)) 2839 return false; 2840 } else if (BB != Header && !canIfConvertPHINodes(BB)) 2841 return false; 2842 2843 } 2844 2845 // We can if-convert this loop. 2846 return true; 2847} 2848 2849bool LoopVectorizationLegality::canVectorize() { 2850 // We must have a loop in canonical form. Loops with indirectbr in them cannot 2851 // be canonicalized. 2852 if (!TheLoop->getLoopPreheader()) 2853 return false; 2854 2855 // We can only vectorize innermost loops. 2856 if (TheLoop->getSubLoopsVector().size()) 2857 return false; 2858 2859 // We must have a single backedge. 2860 if (TheLoop->getNumBackEdges() != 1) 2861 return false; 2862 2863 // We must have a single exiting block. 2864 if (!TheLoop->getExitingBlock()) 2865 return false; 2866 2867 // We need to have a loop header. 2868 DEBUG(dbgs() << "LV: Found a loop: " << 2869 TheLoop->getHeader()->getName() << '\n'); 2870 2871 // Check if we can if-convert non single-bb loops. 2872 unsigned NumBlocks = TheLoop->getNumBlocks(); 2873 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 2874 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 2875 return false; 2876 } 2877 2878 // ScalarEvolution needs to be able to find the exit count. 2879 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop); 2880 if (ExitCount == SE->getCouldNotCompute()) { 2881 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 2882 return false; 2883 } 2884 2885 // Do not loop-vectorize loops with a tiny trip count. 2886 BasicBlock *Latch = TheLoop->getLoopLatch(); 2887 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch); 2888 if (TC > 0u && TC < TinyTripCountVectorThreshold) { 2889 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << 2890 "This loop is not worth vectorizing.\n"); 2891 return false; 2892 } 2893 2894 // Check if we can vectorize the instructions and CFG in this loop. 2895 if (!canVectorizeInstrs()) { 2896 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 2897 return false; 2898 } 2899 2900 // Go over each instruction and look at memory deps. 2901 if (!canVectorizeMemory()) { 2902 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 2903 return false; 2904 } 2905 2906 // Collect all of the variables that remain uniform after vectorization. 2907 collectLoopUniforms(); 2908 2909 DEBUG(dbgs() << "LV: We can vectorize this loop" << 2910 (PtrRtCheck.Need ? " (with a runtime bound check)" : "") 2911 <<"!\n"); 2912 2913 // Okay! We can vectorize. At this point we don't have any other mem analysis 2914 // which may limit our maximum vectorization factor, so just return true with 2915 // no restrictions. 2916 return true; 2917} 2918 2919static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) { 2920 if (Ty->isPointerTy()) 2921 return DL.getIntPtrType(Ty); 2922 2923 // It is possible that char's or short's overflow when we ask for the loop's 2924 // trip count, work around this by changing the type size. 2925 if (Ty->getScalarSizeInBits() < 32) 2926 return Type::getInt32Ty(Ty->getContext()); 2927 2928 return Ty; 2929} 2930 2931static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) { 2932 Ty0 = convertPointerToIntegerType(DL, Ty0); 2933 Ty1 = convertPointerToIntegerType(DL, Ty1); 2934 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits()) 2935 return Ty0; 2936 return Ty1; 2937} 2938 2939/// \brief Check that the instruction has outside loop users and is not an 2940/// identified reduction variable. 2941static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst, 2942 SmallPtrSet<Value *, 4> &Reductions) { 2943 // Reduction instructions are allowed to have exit users. All other 2944 // instructions must not have external users. 2945 if (!Reductions.count(Inst)) 2946 //Check that all of the users of the loop are inside the BB. 2947 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end(); 2948 I != E; ++I) { 2949 Instruction *U = cast<Instruction>(*I); 2950 // This user may be a reduction exit value. 2951 if (!TheLoop->contains(U)) { 2952 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n'); 2953 return true; 2954 } 2955 } 2956 return false; 2957} 2958 2959bool LoopVectorizationLegality::canVectorizeInstrs() { 2960 BasicBlock *PreHeader = TheLoop->getLoopPreheader(); 2961 BasicBlock *Header = TheLoop->getHeader(); 2962 2963 // Look for the attribute signaling the absence of NaNs. 2964 Function &F = *Header->getParent(); 2965 if (F.hasFnAttribute("no-nans-fp-math")) 2966 HasFunNoNaNAttr = F.getAttributes().getAttribute( 2967 AttributeSet::FunctionIndex, 2968 "no-nans-fp-math").getValueAsString() == "true"; 2969 2970 // For each block in the loop. 2971 for (Loop::block_iterator bb = TheLoop->block_begin(), 2972 be = TheLoop->block_end(); bb != be; ++bb) { 2973 2974 // Scan the instructions in the block and look for hazards. 2975 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2976 ++it) { 2977 2978 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 2979 Type *PhiTy = Phi->getType(); 2980 // Check that this PHI type is allowed. 2981 if (!PhiTy->isIntegerTy() && 2982 !PhiTy->isFloatingPointTy() && 2983 !PhiTy->isPointerTy()) { 2984 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 2985 return false; 2986 } 2987 2988 // If this PHINode is not in the header block, then we know that we 2989 // can convert it to select during if-conversion. No need to check if 2990 // the PHIs in this block are induction or reduction variables. 2991 if (*bb != Header) { 2992 // Check that this instruction has no outside users or is an 2993 // identified reduction value with an outside user. 2994 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit)) 2995 continue; 2996 return false; 2997 } 2998 2999 // We only allow if-converted PHIs with more than two incoming values. 3000 if (Phi->getNumIncomingValues() != 2) { 3001 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 3002 return false; 3003 } 3004 3005 // This is the value coming from the preheader. 3006 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); 3007 // Check if this is an induction variable. 3008 InductionKind IK = isInductionVariable(Phi); 3009 3010 if (IK_NoInduction != IK) { 3011 // Get the widest type. 3012 if (!WidestIndTy) 3013 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy); 3014 else 3015 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy); 3016 3017 // Int inductions are special because we only allow one IV. 3018 if (IK == IK_IntInduction) { 3019 // Use the phi node with the widest type as induction. Use the last 3020 // one if there are multiple (no good reason for doing this other 3021 // than it is expedient). 3022 if (!Induction || PhiTy == WidestIndTy) 3023 Induction = Phi; 3024 } 3025 3026 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 3027 Inductions[Phi] = InductionInfo(StartValue, IK); 3028 3029 // Until we explicitly handle the case of an induction variable with 3030 // an outside loop user we have to give up vectorizing this loop. 3031 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) 3032 return false; 3033 3034 continue; 3035 } 3036 3037 if (AddReductionVar(Phi, RK_IntegerAdd)) { 3038 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); 3039 continue; 3040 } 3041 if (AddReductionVar(Phi, RK_IntegerMult)) { 3042 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); 3043 continue; 3044 } 3045 if (AddReductionVar(Phi, RK_IntegerOr)) { 3046 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); 3047 continue; 3048 } 3049 if (AddReductionVar(Phi, RK_IntegerAnd)) { 3050 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); 3051 continue; 3052 } 3053 if (AddReductionVar(Phi, RK_IntegerXor)) { 3054 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); 3055 continue; 3056 } 3057 if (AddReductionVar(Phi, RK_IntegerMinMax)) { 3058 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n"); 3059 continue; 3060 } 3061 if (AddReductionVar(Phi, RK_FloatMult)) { 3062 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n"); 3063 continue; 3064 } 3065 if (AddReductionVar(Phi, RK_FloatAdd)) { 3066 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n"); 3067 continue; 3068 } 3069 if (AddReductionVar(Phi, RK_FloatMinMax)) { 3070 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi << 3071 "\n"); 3072 continue; 3073 } 3074 3075 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 3076 return false; 3077 }// end of PHI handling 3078 3079 // We still don't handle functions. However, we can ignore dbg intrinsic 3080 // calls and we do handle certain intrinsic and libm functions. 3081 CallInst *CI = dyn_cast<CallInst>(it); 3082 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) { 3083 DEBUG(dbgs() << "LV: Found a call site.\n"); 3084 return false; 3085 } 3086 3087 // Check that the instruction return type is vectorizable. 3088 // Also, we can't vectorize extractelement instructions. 3089 if ((!VectorType::isValidElementType(it->getType()) && 3090 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) { 3091 DEBUG(dbgs() << "LV: Found unvectorizable type.\n"); 3092 return false; 3093 } 3094 3095 // Check that the stored type is vectorizable. 3096 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 3097 Type *T = ST->getValueOperand()->getType(); 3098 if (!VectorType::isValidElementType(T)) 3099 return false; 3100 } 3101 3102 // Reduction instructions are allowed to have exit users. 3103 // All other instructions must not have external users. 3104 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) 3105 return false; 3106 3107 } // next instr. 3108 3109 } 3110 3111 if (!Induction) { 3112 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 3113 if (Inductions.empty()) 3114 return false; 3115 } 3116 3117 return true; 3118} 3119 3120void LoopVectorizationLegality::collectLoopUniforms() { 3121 // We now know that the loop is vectorizable! 3122 // Collect variables that will remain uniform after vectorization. 3123 std::vector<Value*> Worklist; 3124 BasicBlock *Latch = TheLoop->getLoopLatch(); 3125 3126 // Start with the conditional branch and walk up the block. 3127 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 3128 3129 while (Worklist.size()) { 3130 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 3131 Worklist.pop_back(); 3132 3133 // Look at instructions inside this loop. 3134 // Stop when reaching PHI nodes. 3135 // TODO: we need to follow values all over the loop, not only in this block. 3136 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 3137 continue; 3138 3139 // This is a known uniform. 3140 Uniforms.insert(I); 3141 3142 // Insert all operands. 3143 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end()); 3144 } 3145} 3146 3147namespace { 3148/// \brief Analyses memory accesses in a loop. 3149/// 3150/// Checks whether run time pointer checks are needed and builds sets for data 3151/// dependence checking. 3152class AccessAnalysis { 3153public: 3154 /// \brief Read or write access location. 3155 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 3156 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet; 3157 3158 /// \brief Set of potential dependent memory accesses. 3159 typedef EquivalenceClasses<MemAccessInfo> DepCandidates; 3160 3161 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) : 3162 DL(Dl), DepCands(DA), AreAllWritesIdentified(true), 3163 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {} 3164 3165 /// \brief Register a load and whether it is only read from. 3166 void addLoad(Value *Ptr, bool IsReadOnly) { 3167 Accesses.insert(MemAccessInfo(Ptr, false)); 3168 if (IsReadOnly) 3169 ReadOnlyPtr.insert(Ptr); 3170 } 3171 3172 /// \brief Register a store. 3173 void addStore(Value *Ptr) { 3174 Accesses.insert(MemAccessInfo(Ptr, true)); 3175 } 3176 3177 /// \brief Check whether we can check the pointers at runtime for 3178 /// non-intersection. 3179 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck, 3180 unsigned &NumComparisons, ScalarEvolution *SE, 3181 Loop *TheLoop, bool ShouldCheckStride = false); 3182 3183 /// \brief Goes over all memory accesses, checks whether a RT check is needed 3184 /// and builds sets of dependent accesses. 3185 void buildDependenceSets() { 3186 // Process read-write pointers first. 3187 processMemAccesses(false); 3188 // Next, process read pointers. 3189 processMemAccesses(true); 3190 } 3191 3192 bool isRTCheckNeeded() { return IsRTCheckNeeded; } 3193 3194 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); } 3195 void resetDepChecks() { CheckDeps.clear(); } 3196 3197 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; } 3198 3199private: 3200 typedef SetVector<MemAccessInfo> PtrAccessSet; 3201 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap; 3202 3203 /// \brief Go over all memory access or only the deferred ones if 3204 /// \p UseDeferred is true and check whether runtime pointer checks are needed 3205 /// and build sets of dependency check candidates. 3206 void processMemAccesses(bool UseDeferred); 3207 3208 /// Set of all accesses. 3209 PtrAccessSet Accesses; 3210 3211 /// Set of access to check after all writes have been processed. 3212 PtrAccessSet DeferredAccesses; 3213 3214 /// Map of pointers to last access encountered. 3215 UnderlyingObjToAccessMap ObjToLastAccess; 3216 3217 /// Set of accesses that need a further dependence check. 3218 MemAccessInfoSet CheckDeps; 3219 3220 /// Set of pointers that are read only. 3221 SmallPtrSet<Value*, 16> ReadOnlyPtr; 3222 3223 /// Set of underlying objects already written to. 3224 SmallPtrSet<Value*, 16> WriteObjects; 3225 3226 DataLayout *DL; 3227 3228 /// Sets of potentially dependent accesses - members of one set share an 3229 /// underlying pointer. The set "CheckDeps" identfies which sets really need a 3230 /// dependence check. 3231 DepCandidates &DepCands; 3232 3233 bool AreAllWritesIdentified; 3234 bool AreAllReadsIdentified; 3235 bool IsRTCheckNeeded; 3236}; 3237 3238} // end anonymous namespace 3239 3240/// \brief Check whether a pointer can participate in a runtime bounds check. 3241static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) { 3242 const SCEV *PtrScev = SE->getSCEV(Ptr); 3243 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 3244 if (!AR) 3245 return false; 3246 3247 return AR->isAffine(); 3248} 3249 3250/// \brief Check the stride of the pointer and ensure that it does not wrap in 3251/// the address space. 3252static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr, 3253 const Loop *Lp); 3254 3255bool AccessAnalysis::canCheckPtrAtRT( 3256 LoopVectorizationLegality::RuntimePointerCheck &RtCheck, 3257 unsigned &NumComparisons, ScalarEvolution *SE, 3258 Loop *TheLoop, bool ShouldCheckStride) { 3259 // Find pointers with computable bounds. We are going to use this information 3260 // to place a runtime bound check. 3261 unsigned NumReadPtrChecks = 0; 3262 unsigned NumWritePtrChecks = 0; 3263 bool CanDoRT = true; 3264 3265 bool IsDepCheckNeeded = isDependencyCheckNeeded(); 3266 // We assign consecutive id to access from different dependence sets. 3267 // Accesses within the same set don't need a runtime check. 3268 unsigned RunningDepId = 1; 3269 DenseMap<Value *, unsigned> DepSetId; 3270 3271 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end(); 3272 AI != AE; ++AI) { 3273 const MemAccessInfo &Access = *AI; 3274 Value *Ptr = Access.getPointer(); 3275 bool IsWrite = Access.getInt(); 3276 3277 // Just add write checks if we have both. 3278 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true))) 3279 continue; 3280 3281 if (IsWrite) 3282 ++NumWritePtrChecks; 3283 else 3284 ++NumReadPtrChecks; 3285 3286 if (hasComputableBounds(SE, Ptr) && 3287 // When we run after a failing dependency check we have to make sure we 3288 // don't have wrapping pointers. 3289 (!ShouldCheckStride || isStridedPtr(SE, DL, Ptr, TheLoop) == 1)) { 3290 // The id of the dependence set. 3291 unsigned DepId; 3292 3293 if (IsDepCheckNeeded) { 3294 Value *Leader = DepCands.getLeaderValue(Access).getPointer(); 3295 unsigned &LeaderId = DepSetId[Leader]; 3296 if (!LeaderId) 3297 LeaderId = RunningDepId++; 3298 DepId = LeaderId; 3299 } else 3300 // Each access has its own dependence set. 3301 DepId = RunningDepId++; 3302 3303 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId); 3304 3305 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n'); 3306 } else { 3307 CanDoRT = false; 3308 } 3309 } 3310 3311 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2) 3312 NumComparisons = 0; // Only one dependence set. 3313 else { 3314 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks + 3315 NumWritePtrChecks - 1)); 3316 } 3317 3318 // If the pointers that we would use for the bounds comparison have different 3319 // address spaces, assume the values aren't directly comparable, so we can't 3320 // use them for the runtime check. We also have to assume they could 3321 // overlap. In the future there should be metadata for whether address spaces 3322 // are disjoint. 3323 unsigned NumPointers = RtCheck.Pointers.size(); 3324 for (unsigned i = 0; i < NumPointers; ++i) { 3325 for (unsigned j = i + 1; j < NumPointers; ++j) { 3326 // Only need to check pointers between two different dependency sets. 3327 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j]) 3328 continue; 3329 3330 Value *PtrI = RtCheck.Pointers[i]; 3331 Value *PtrJ = RtCheck.Pointers[j]; 3332 3333 unsigned ASi = PtrI->getType()->getPointerAddressSpace(); 3334 unsigned ASj = PtrJ->getType()->getPointerAddressSpace(); 3335 if (ASi != ASj) { 3336 DEBUG(dbgs() << "LV: Runtime check would require comparison between" 3337 " different address spaces\n"); 3338 return false; 3339 } 3340 } 3341 } 3342 3343 return CanDoRT; 3344} 3345 3346static bool isFunctionScopeIdentifiedObject(Value *Ptr) { 3347 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr); 3348} 3349 3350void AccessAnalysis::processMemAccesses(bool UseDeferred) { 3351 // We process the set twice: first we process read-write pointers, last we 3352 // process read-only pointers. This allows us to skip dependence tests for 3353 // read-only pointers. 3354 3355 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses; 3356 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) { 3357 const MemAccessInfo &Access = *AI; 3358 Value *Ptr = Access.getPointer(); 3359 bool IsWrite = Access.getInt(); 3360 3361 DepCands.insert(Access); 3362 3363 // Memorize read-only pointers for later processing and skip them in the 3364 // first round (they need to be checked after we have seen all write 3365 // pointers). Note: we also mark pointer that are not consecutive as 3366 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the 3367 // second check for "!IsWrite". 3368 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite; 3369 if (!UseDeferred && IsReadOnlyPtr) { 3370 DeferredAccesses.insert(Access); 3371 continue; 3372 } 3373 3374 bool NeedDepCheck = false; 3375 // Check whether there is the possiblity of dependency because of underlying 3376 // objects being the same. 3377 typedef SmallVector<Value*, 16> ValueVector; 3378 ValueVector TempObjects; 3379 GetUnderlyingObjects(Ptr, TempObjects, DL); 3380 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end(); 3381 UI != UE; ++UI) { 3382 Value *UnderlyingObj = *UI; 3383 3384 // If this is a write then it needs to be an identified object. If this a 3385 // read and all writes (so far) are identified function scope objects we 3386 // don't need an identified underlying object but only an Argument (the 3387 // next write is going to invalidate this assumption if it is 3388 // unidentified). 3389 // This is a micro-optimization for the case where all writes are 3390 // identified and we have one argument pointer. 3391 // Otherwise, we do need a runtime check. 3392 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) || 3393 (!IsWrite && (!AreAllWritesIdentified || 3394 !isa<Argument>(UnderlyingObj)) && 3395 !isIdentifiedObject(UnderlyingObj))) { 3396 DEBUG(dbgs() << "LV: Found an unidentified " << 3397 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj << 3398 "\n"); 3399 IsRTCheckNeeded = (IsRTCheckNeeded || 3400 !isIdentifiedObject(UnderlyingObj) || 3401 !AreAllReadsIdentified); 3402 3403 if (IsWrite) 3404 AreAllWritesIdentified = false; 3405 if (!IsWrite) 3406 AreAllReadsIdentified = false; 3407 } 3408 3409 // If this is a write - check other reads and writes for conflicts. If 3410 // this is a read only check other writes for conflicts (but only if there 3411 // is no other write to the ptr - this is an optimization to catch "a[i] = 3412 // a[i] + " without having to do a dependence check). 3413 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj)) 3414 NeedDepCheck = true; 3415 3416 if (IsWrite) 3417 WriteObjects.insert(UnderlyingObj); 3418 3419 // Create sets of pointers connected by shared underlying objects. 3420 UnderlyingObjToAccessMap::iterator Prev = 3421 ObjToLastAccess.find(UnderlyingObj); 3422 if (Prev != ObjToLastAccess.end()) 3423 DepCands.unionSets(Access, Prev->second); 3424 3425 ObjToLastAccess[UnderlyingObj] = Access; 3426 } 3427 3428 if (NeedDepCheck) 3429 CheckDeps.insert(Access); 3430 } 3431} 3432 3433namespace { 3434/// \brief Checks memory dependences among accesses to the same underlying 3435/// object to determine whether there vectorization is legal or not (and at 3436/// which vectorization factor). 3437/// 3438/// This class works under the assumption that we already checked that memory 3439/// locations with different underlying pointers are "must-not alias". 3440/// We use the ScalarEvolution framework to symbolically evalutate access 3441/// functions pairs. Since we currently don't restructure the loop we can rely 3442/// on the program order of memory accesses to determine their safety. 3443/// At the moment we will only deem accesses as safe for: 3444/// * A negative constant distance assuming program order. 3445/// 3446/// Safe: tmp = a[i + 1]; OR a[i + 1] = x; 3447/// a[i] = tmp; y = a[i]; 3448/// 3449/// The latter case is safe because later checks guarantuee that there can't 3450/// be a cycle through a phi node (that is, we check that "x" and "y" is not 3451/// the same variable: a header phi can only be an induction or a reduction, a 3452/// reduction can't have a memory sink, an induction can't have a memory 3453/// source). This is important and must not be violated (or we have to 3454/// resort to checking for cycles through memory). 3455/// 3456/// * A positive constant distance assuming program order that is bigger 3457/// than the biggest memory access. 3458/// 3459/// tmp = a[i] OR b[i] = x 3460/// a[i+2] = tmp y = b[i+2]; 3461/// 3462/// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively. 3463/// 3464/// * Zero distances and all accesses have the same size. 3465/// 3466class MemoryDepChecker { 3467public: 3468 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 3469 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet; 3470 3471 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) 3472 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0), 3473 ShouldRetryWithRuntimeCheck(false) {} 3474 3475 /// \brief Register the location (instructions are given increasing numbers) 3476 /// of a write access. 3477 void addAccess(StoreInst *SI) { 3478 Value *Ptr = SI->getPointerOperand(); 3479 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx); 3480 InstMap.push_back(SI); 3481 ++AccessIdx; 3482 } 3483 3484 /// \brief Register the location (instructions are given increasing numbers) 3485 /// of a write access. 3486 void addAccess(LoadInst *LI) { 3487 Value *Ptr = LI->getPointerOperand(); 3488 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx); 3489 InstMap.push_back(LI); 3490 ++AccessIdx; 3491 } 3492 3493 /// \brief Check whether the dependencies between the accesses are safe. 3494 /// 3495 /// Only checks sets with elements in \p CheckDeps. 3496 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets, 3497 MemAccessInfoSet &CheckDeps); 3498 3499 /// \brief The maximum number of bytes of a vector register we can vectorize 3500 /// the accesses safely with. 3501 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; } 3502 3503 /// \brief In same cases when the dependency check fails we can still 3504 /// vectorize the loop with a dynamic array access check. 3505 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; } 3506 3507private: 3508 ScalarEvolution *SE; 3509 DataLayout *DL; 3510 const Loop *InnermostLoop; 3511 3512 /// \brief Maps access locations (ptr, read/write) to program order. 3513 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses; 3514 3515 /// \brief Memory access instructions in program order. 3516 SmallVector<Instruction *, 16> InstMap; 3517 3518 /// \brief The program order index to be used for the next instruction. 3519 unsigned AccessIdx; 3520 3521 // We can access this many bytes in parallel safely. 3522 unsigned MaxSafeDepDistBytes; 3523 3524 /// \brief If we see a non constant dependence distance we can still try to 3525 /// vectorize this loop with runtime checks. 3526 bool ShouldRetryWithRuntimeCheck; 3527 3528 /// \brief Check whether there is a plausible dependence between the two 3529 /// accesses. 3530 /// 3531 /// Access \p A must happen before \p B in program order. The two indices 3532 /// identify the index into the program order map. 3533 /// 3534 /// This function checks whether there is a plausible dependence (or the 3535 /// absence of such can't be proved) between the two accesses. If there is a 3536 /// plausible dependence but the dependence distance is bigger than one 3537 /// element access it records this distance in \p MaxSafeDepDistBytes (if this 3538 /// distance is smaller than any other distance encountered so far). 3539 /// Otherwise, this function returns true signaling a possible dependence. 3540 bool isDependent(const MemAccessInfo &A, unsigned AIdx, 3541 const MemAccessInfo &B, unsigned BIdx); 3542 3543 /// \brief Check whether the data dependence could prevent store-load 3544 /// forwarding. 3545 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize); 3546}; 3547 3548} // end anonymous namespace 3549 3550static bool isInBoundsGep(Value *Ptr) { 3551 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) 3552 return GEP->isInBounds(); 3553 return false; 3554} 3555 3556/// \brief Check whether the access through \p Ptr has a constant stride. 3557static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr, 3558 const Loop *Lp) { 3559 const Type *Ty = Ptr->getType(); 3560 assert(Ty->isPointerTy() && "Unexpected non ptr"); 3561 3562 // Make sure that the pointer does not point to aggregate types. 3563 const PointerType *PtrTy = cast<PointerType>(Ty); 3564 if (PtrTy->getElementType()->isAggregateType()) { 3565 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr << 3566 "\n"); 3567 return 0; 3568 } 3569 3570 const SCEV *PtrScev = SE->getSCEV(Ptr); 3571 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 3572 if (!AR) { 3573 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer " 3574 << *Ptr << " SCEV: " << *PtrScev << "\n"); 3575 return 0; 3576 } 3577 3578 // The accesss function must stride over the innermost loop. 3579 if (Lp != AR->getLoop()) { 3580 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " << 3581 *Ptr << " SCEV: " << *PtrScev << "\n"); 3582 } 3583 3584 // The address calculation must not wrap. Otherwise, a dependence could be 3585 // inverted. 3586 // An inbounds getelementptr that is a AddRec with a unit stride 3587 // cannot wrap per definition. The unit stride requirement is checked later. 3588 // An getelementptr without an inbounds attribute and unit stride would have 3589 // to access the pointer value "0" which is undefined behavior in address 3590 // space 0, therefore we can also vectorize this case. 3591 bool IsInBoundsGEP = isInBoundsGep(Ptr); 3592 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask); 3593 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0; 3594 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) { 3595 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space " 3596 << *Ptr << " SCEV: " << *PtrScev << "\n"); 3597 return 0; 3598 } 3599 3600 // Check the step is constant. 3601 const SCEV *Step = AR->getStepRecurrence(*SE); 3602 3603 // Calculate the pointer stride and check if it is consecutive. 3604 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 3605 if (!C) { 3606 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr << 3607 " SCEV: " << *PtrScev << "\n"); 3608 return 0; 3609 } 3610 3611 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType()); 3612 const APInt &APStepVal = C->getValue()->getValue(); 3613 3614 // Huge step value - give up. 3615 if (APStepVal.getBitWidth() > 64) 3616 return 0; 3617 3618 int64_t StepVal = APStepVal.getSExtValue(); 3619 3620 // Strided access. 3621 int64_t Stride = StepVal / Size; 3622 int64_t Rem = StepVal % Size; 3623 if (Rem) 3624 return 0; 3625 3626 // If the SCEV could wrap but we have an inbounds gep with a unit stride we 3627 // know we can't "wrap around the address space". In case of address space 3628 // zero we know that this won't happen without triggering undefined behavior. 3629 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) && 3630 Stride != 1 && Stride != -1) 3631 return 0; 3632 3633 return Stride; 3634} 3635 3636bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance, 3637 unsigned TypeByteSize) { 3638 // If loads occur at a distance that is not a multiple of a feasible vector 3639 // factor store-load forwarding does not take place. 3640 // Positive dependences might cause troubles because vectorizing them might 3641 // prevent store-load forwarding making vectorized code run a lot slower. 3642 // a[i] = a[i-3] ^ a[i-8]; 3643 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and 3644 // hence on your typical architecture store-load forwarding does not take 3645 // place. Vectorizing in such cases does not make sense. 3646 // Store-load forwarding distance. 3647 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize; 3648 // Maximum vector factor. 3649 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize; 3650 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues) 3651 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes; 3652 3653 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues; 3654 vf *= 2) { 3655 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) { 3656 MaxVFWithoutSLForwardIssues = (vf >>=1); 3657 break; 3658 } 3659 } 3660 3661 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) { 3662 DEBUG(dbgs() << "LV: Distance " << Distance << 3663 " that could cause a store-load forwarding conflict\n"); 3664 return true; 3665 } 3666 3667 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes && 3668 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize) 3669 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues; 3670 return false; 3671} 3672 3673bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx, 3674 const MemAccessInfo &B, unsigned BIdx) { 3675 assert (AIdx < BIdx && "Must pass arguments in program order"); 3676 3677 Value *APtr = A.getPointer(); 3678 Value *BPtr = B.getPointer(); 3679 bool AIsWrite = A.getInt(); 3680 bool BIsWrite = B.getInt(); 3681 3682 // Two reads are independent. 3683 if (!AIsWrite && !BIsWrite) 3684 return false; 3685 3686 const SCEV *AScev = SE->getSCEV(APtr); 3687 const SCEV *BScev = SE->getSCEV(BPtr); 3688 3689 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop); 3690 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop); 3691 3692 const SCEV *Src = AScev; 3693 const SCEV *Sink = BScev; 3694 3695 // If the induction step is negative we have to invert source and sink of the 3696 // dependence. 3697 if (StrideAPtr < 0) { 3698 //Src = BScev; 3699 //Sink = AScev; 3700 std::swap(APtr, BPtr); 3701 std::swap(Src, Sink); 3702 std::swap(AIsWrite, BIsWrite); 3703 std::swap(AIdx, BIdx); 3704 std::swap(StrideAPtr, StrideBPtr); 3705 } 3706 3707 const SCEV *Dist = SE->getMinusSCEV(Sink, Src); 3708 3709 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink 3710 << "(Induction step: " << StrideAPtr << ")\n"); 3711 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to " 3712 << *InstMap[BIdx] << ": " << *Dist << "\n"); 3713 3714 // Need consecutive accesses. We don't want to vectorize 3715 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in 3716 // the address space. 3717 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){ 3718 DEBUG(dbgs() << "Non-consecutive pointer access\n"); 3719 return true; 3720 } 3721 3722 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist); 3723 if (!C) { 3724 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n"); 3725 ShouldRetryWithRuntimeCheck = true; 3726 return true; 3727 } 3728 3729 Type *ATy = APtr->getType()->getPointerElementType(); 3730 Type *BTy = BPtr->getType()->getPointerElementType(); 3731 unsigned TypeByteSize = DL->getTypeAllocSize(ATy); 3732 3733 // Negative distances are not plausible dependencies. 3734 const APInt &Val = C->getValue()->getValue(); 3735 if (Val.isNegative()) { 3736 bool IsTrueDataDependence = (AIsWrite && !BIsWrite); 3737 if (IsTrueDataDependence && 3738 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) || 3739 ATy != BTy)) 3740 return true; 3741 3742 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n"); 3743 return false; 3744 } 3745 3746 // Write to the same location with the same size. 3747 // Could be improved to assert type sizes are the same (i32 == float, etc). 3748 if (Val == 0) { 3749 if (ATy == BTy) 3750 return false; 3751 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n"); 3752 return true; 3753 } 3754 3755 assert(Val.isStrictlyPositive() && "Expect a positive value"); 3756 3757 // Positive distance bigger than max vectorization factor. 3758 if (ATy != BTy) { 3759 DEBUG(dbgs() << 3760 "LV: ReadWrite-Write positive dependency with different types\n"); 3761 return false; 3762 } 3763 3764 unsigned Distance = (unsigned) Val.getZExtValue(); 3765 3766 // Bail out early if passed-in parameters make vectorization not feasible. 3767 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1; 3768 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1; 3769 3770 // The distance must be bigger than the size needed for a vectorized version 3771 // of the operation and the size of the vectorized operation must not be 3772 // bigger than the currrent maximum size. 3773 if (Distance < 2*TypeByteSize || 3774 2*TypeByteSize > MaxSafeDepDistBytes || 3775 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) { 3776 DEBUG(dbgs() << "LV: Failure because of Positive distance " 3777 << Val.getSExtValue() << '\n'); 3778 return true; 3779 } 3780 3781 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ? 3782 Distance : MaxSafeDepDistBytes; 3783 3784 bool IsTrueDataDependence = (!AIsWrite && BIsWrite); 3785 if (IsTrueDataDependence && 3786 couldPreventStoreLoadForward(Distance, TypeByteSize)) 3787 return true; 3788 3789 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() << 3790 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n'); 3791 3792 return false; 3793} 3794 3795bool 3796MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets, 3797 MemAccessInfoSet &CheckDeps) { 3798 3799 MaxSafeDepDistBytes = -1U; 3800 while (!CheckDeps.empty()) { 3801 MemAccessInfo CurAccess = *CheckDeps.begin(); 3802 3803 // Get the relevant memory access set. 3804 EquivalenceClasses<MemAccessInfo>::iterator I = 3805 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess)); 3806 3807 // Check accesses within this set. 3808 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE; 3809 AI = AccessSets.member_begin(I), AE = AccessSets.member_end(); 3810 3811 // Check every access pair. 3812 while (AI != AE) { 3813 CheckDeps.erase(*AI); 3814 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI); 3815 while (OI != AE) { 3816 // Check every accessing instruction pair in program order. 3817 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(), 3818 I1E = Accesses[*AI].end(); I1 != I1E; ++I1) 3819 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(), 3820 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) { 3821 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2)) 3822 return false; 3823 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1)) 3824 return false; 3825 } 3826 ++OI; 3827 } 3828 AI++; 3829 } 3830 } 3831 return true; 3832} 3833 3834bool LoopVectorizationLegality::canVectorizeMemory() { 3835 3836 typedef SmallVector<Value*, 16> ValueVector; 3837 typedef SmallPtrSet<Value*, 16> ValueSet; 3838 3839 // Holds the Load and Store *instructions*. 3840 ValueVector Loads; 3841 ValueVector Stores; 3842 3843 // Holds all the different accesses in the loop. 3844 unsigned NumReads = 0; 3845 unsigned NumReadWrites = 0; 3846 3847 PtrRtCheck.Pointers.clear(); 3848 PtrRtCheck.Need = false; 3849 3850 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); 3851 MemoryDepChecker DepChecker(SE, DL, TheLoop); 3852 3853 // For each block. 3854 for (Loop::block_iterator bb = TheLoop->block_begin(), 3855 be = TheLoop->block_end(); bb != be; ++bb) { 3856 3857 // Scan the BB and collect legal loads and stores. 3858 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 3859 ++it) { 3860 3861 // If this is a load, save it. If this instruction can read from memory 3862 // but is not a load, then we quit. Notice that we don't handle function 3863 // calls that read or write. 3864 if (it->mayReadFromMemory()) { 3865 // Many math library functions read the rounding mode. We will only 3866 // vectorize a loop if it contains known function calls that don't set 3867 // the flag. Therefore, it is safe to ignore this read from memory. 3868 CallInst *Call = dyn_cast<CallInst>(it); 3869 if (Call && getIntrinsicIDForCall(Call, TLI)) 3870 continue; 3871 3872 LoadInst *Ld = dyn_cast<LoadInst>(it); 3873 if (!Ld) return false; 3874 if (!Ld->isSimple() && !IsAnnotatedParallel) { 3875 DEBUG(dbgs() << "LV: Found a non-simple load.\n"); 3876 return false; 3877 } 3878 Loads.push_back(Ld); 3879 DepChecker.addAccess(Ld); 3880 continue; 3881 } 3882 3883 // Save 'store' instructions. Abort if other instructions write to memory. 3884 if (it->mayWriteToMemory()) { 3885 StoreInst *St = dyn_cast<StoreInst>(it); 3886 if (!St) return false; 3887 if (!St->isSimple() && !IsAnnotatedParallel) { 3888 DEBUG(dbgs() << "LV: Found a non-simple store.\n"); 3889 return false; 3890 } 3891 Stores.push_back(St); 3892 DepChecker.addAccess(St); 3893 } 3894 } // Next instr. 3895 } // Next block. 3896 3897 // Now we have two lists that hold the loads and the stores. 3898 // Next, we find the pointers that they use. 3899 3900 // Check if we see any stores. If there are no stores, then we don't 3901 // care if the pointers are *restrict*. 3902 if (!Stores.size()) { 3903 DEBUG(dbgs() << "LV: Found a read-only loop!\n"); 3904 return true; 3905 } 3906 3907 AccessAnalysis::DepCandidates DependentAccesses; 3908 AccessAnalysis Accesses(DL, DependentAccesses); 3909 3910 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects 3911 // multiple times on the same object. If the ptr is accessed twice, once 3912 // for read and once for write, it will only appear once (on the write 3913 // list). This is okay, since we are going to check for conflicts between 3914 // writes and between reads and writes, but not between reads and reads. 3915 ValueSet Seen; 3916 3917 ValueVector::iterator I, IE; 3918 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) { 3919 StoreInst *ST = cast<StoreInst>(*I); 3920 Value* Ptr = ST->getPointerOperand(); 3921 3922 if (isUniform(Ptr)) { 3923 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 3924 return false; 3925 } 3926 3927 // If we did *not* see this pointer before, insert it to the read-write 3928 // list. At this phase it is only a 'write' list. 3929 if (Seen.insert(Ptr)) { 3930 ++NumReadWrites; 3931 Accesses.addStore(Ptr); 3932 } 3933 } 3934 3935 if (IsAnnotatedParallel) { 3936 DEBUG(dbgs() 3937 << "LV: A loop annotated parallel, ignore memory dependency " 3938 << "checks.\n"); 3939 return true; 3940 } 3941 3942 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { 3943 LoadInst *LD = cast<LoadInst>(*I); 3944 Value* Ptr = LD->getPointerOperand(); 3945 // If we did *not* see this pointer before, insert it to the 3946 // read list. If we *did* see it before, then it is already in 3947 // the read-write list. This allows us to vectorize expressions 3948 // such as A[i] += x; Because the address of A[i] is a read-write 3949 // pointer. This only works if the index of A[i] is consecutive. 3950 // If the address of i is unknown (for example A[B[i]]) then we may 3951 // read a few words, modify, and write a few words, and some of the 3952 // words may be written to the same address. 3953 bool IsReadOnlyPtr = false; 3954 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) { 3955 ++NumReads; 3956 IsReadOnlyPtr = true; 3957 } 3958 Accesses.addLoad(Ptr, IsReadOnlyPtr); 3959 } 3960 3961 // If we write (or read-write) to a single destination and there are no 3962 // other reads in this loop then is it safe to vectorize. 3963 if (NumReadWrites == 1 && NumReads == 0) { 3964 DEBUG(dbgs() << "LV: Found a write-only loop!\n"); 3965 return true; 3966 } 3967 3968 // Build dependence sets and check whether we need a runtime pointer bounds 3969 // check. 3970 Accesses.buildDependenceSets(); 3971 bool NeedRTCheck = Accesses.isRTCheckNeeded(); 3972 3973 // Find pointers with computable bounds. We are going to use this information 3974 // to place a runtime bound check. 3975 unsigned NumComparisons = 0; 3976 bool CanDoRT = false; 3977 if (NeedRTCheck) 3978 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop); 3979 3980 3981 DEBUG(dbgs() << "LV: We need to do " << NumComparisons << 3982 " pointer comparisons.\n"); 3983 3984 // If we only have one set of dependences to check pointers among we don't 3985 // need a runtime check. 3986 if (NumComparisons == 0 && NeedRTCheck) 3987 NeedRTCheck = false; 3988 3989 // Check that we did not collect too many pointers or found an unsizeable 3990 // pointer. 3991 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { 3992 PtrRtCheck.reset(); 3993 CanDoRT = false; 3994 } 3995 3996 if (CanDoRT) { 3997 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n"); 3998 } 3999 4000 if (NeedRTCheck && !CanDoRT) { 4001 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " << 4002 "the array bounds.\n"); 4003 PtrRtCheck.reset(); 4004 return false; 4005 } 4006 4007 PtrRtCheck.Need = NeedRTCheck; 4008 4009 bool CanVecMem = true; 4010 if (Accesses.isDependencyCheckNeeded()) { 4011 DEBUG(dbgs() << "LV: Checking memory dependencies\n"); 4012 CanVecMem = DepChecker.areDepsSafe(DependentAccesses, 4013 Accesses.getDependenciesToCheck()); 4014 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes(); 4015 4016 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) { 4017 DEBUG(dbgs() << "LV: Retrying with memory checks\n"); 4018 NeedRTCheck = true; 4019 4020 // Clear the dependency checks. We assume they are not needed. 4021 Accesses.resetDepChecks(); 4022 4023 PtrRtCheck.reset(); 4024 PtrRtCheck.Need = true; 4025 4026 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, 4027 TheLoop, true); 4028 // Check that we did not collect too many pointers or found an unsizeable 4029 // pointer. 4030 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { 4031 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n"); 4032 PtrRtCheck.reset(); 4033 return false; 4034 } 4035 4036 CanVecMem = true; 4037 } 4038 } 4039 4040 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") << 4041 " need a runtime memory check.\n"); 4042 4043 return CanVecMem; 4044} 4045 4046static bool hasMultipleUsesOf(Instruction *I, 4047 SmallPtrSet<Instruction *, 8> &Insts) { 4048 unsigned NumUses = 0; 4049 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) { 4050 if (Insts.count(dyn_cast<Instruction>(*Use))) 4051 ++NumUses; 4052 if (NumUses > 1) 4053 return true; 4054 } 4055 4056 return false; 4057} 4058 4059static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) { 4060 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) 4061 if (!Set.count(dyn_cast<Instruction>(*Use))) 4062 return false; 4063 return true; 4064} 4065 4066bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, 4067 ReductionKind Kind) { 4068 if (Phi->getNumIncomingValues() != 2) 4069 return false; 4070 4071 // Reduction variables are only found in the loop header block. 4072 if (Phi->getParent() != TheLoop->getHeader()) 4073 return false; 4074 4075 // Obtain the reduction start value from the value that comes from the loop 4076 // preheader. 4077 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 4078 4079 // ExitInstruction is the single value which is used outside the loop. 4080 // We only allow for a single reduction value to be used outside the loop. 4081 // This includes users of the reduction, variables (which form a cycle 4082 // which ends in the phi node). 4083 Instruction *ExitInstruction = 0; 4084 // Indicates that we found a reduction operation in our scan. 4085 bool FoundReduxOp = false; 4086 4087 // We start with the PHI node and scan for all of the users of this 4088 // instruction. All users must be instructions that can be used as reduction 4089 // variables (such as ADD). We must have a single out-of-block user. The cycle 4090 // must include the original PHI. 4091 bool FoundStartPHI = false; 4092 4093 // To recognize min/max patterns formed by a icmp select sequence, we store 4094 // the number of instruction we saw from the recognized min/max pattern, 4095 // to make sure we only see exactly the two instructions. 4096 unsigned NumCmpSelectPatternInst = 0; 4097 ReductionInstDesc ReduxDesc(false, 0); 4098 4099 SmallPtrSet<Instruction *, 8> VisitedInsts; 4100 SmallVector<Instruction *, 8> Worklist; 4101 Worklist.push_back(Phi); 4102 VisitedInsts.insert(Phi); 4103 4104 // A value in the reduction can be used: 4105 // - By the reduction: 4106 // - Reduction operation: 4107 // - One use of reduction value (safe). 4108 // - Multiple use of reduction value (not safe). 4109 // - PHI: 4110 // - All uses of the PHI must be the reduction (safe). 4111 // - Otherwise, not safe. 4112 // - By one instruction outside of the loop (safe). 4113 // - By further instructions outside of the loop (not safe). 4114 // - By an instruction that is not part of the reduction (not safe). 4115 // This is either: 4116 // * An instruction type other than PHI or the reduction operation. 4117 // * A PHI in the header other than the initial PHI. 4118 while (!Worklist.empty()) { 4119 Instruction *Cur = Worklist.back(); 4120 Worklist.pop_back(); 4121 4122 // No Users. 4123 // If the instruction has no users then this is a broken chain and can't be 4124 // a reduction variable. 4125 if (Cur->use_empty()) 4126 return false; 4127 4128 bool IsAPhi = isa<PHINode>(Cur); 4129 4130 // A header PHI use other than the original PHI. 4131 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent()) 4132 return false; 4133 4134 // Reductions of instructions such as Div, and Sub is only possible if the 4135 // LHS is the reduction variable. 4136 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) && 4137 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) && 4138 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0)))) 4139 return false; 4140 4141 // Any reduction instruction must be of one of the allowed kinds. 4142 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc); 4143 if (!ReduxDesc.IsReduction) 4144 return false; 4145 4146 // A reduction operation must only have one use of the reduction value. 4147 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax && 4148 hasMultipleUsesOf(Cur, VisitedInsts)) 4149 return false; 4150 4151 // All inputs to a PHI node must be a reduction value. 4152 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts)) 4153 return false; 4154 4155 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) || 4156 isa<SelectInst>(Cur))) 4157 ++NumCmpSelectPatternInst; 4158 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) || 4159 isa<SelectInst>(Cur))) 4160 ++NumCmpSelectPatternInst; 4161 4162 // Check whether we found a reduction operator. 4163 FoundReduxOp |= !IsAPhi; 4164 4165 // Process users of current instruction. Push non PHI nodes after PHI nodes 4166 // onto the stack. This way we are going to have seen all inputs to PHI 4167 // nodes once we get to them. 4168 SmallVector<Instruction *, 8> NonPHIs; 4169 SmallVector<Instruction *, 8> PHIs; 4170 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E; 4171 ++UI) { 4172 Instruction *Usr = cast<Instruction>(*UI); 4173 4174 // Check if we found the exit user. 4175 BasicBlock *Parent = Usr->getParent(); 4176 if (!TheLoop->contains(Parent)) { 4177 // Exit if you find multiple outside users or if the header phi node is 4178 // being used. In this case the user uses the value of the previous 4179 // iteration, in which case we would loose "VF-1" iterations of the 4180 // reduction operation if we vectorize. 4181 if (ExitInstruction != 0 || Cur == Phi) 4182 return false; 4183 4184 // The instruction used by an outside user must be the last instruction 4185 // before we feed back to the reduction phi. Otherwise, we loose VF-1 4186 // operations on the value. 4187 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end()) 4188 return false; 4189 4190 ExitInstruction = Cur; 4191 continue; 4192 } 4193 4194 // Process instructions only once (termination). 4195 if (VisitedInsts.insert(Usr)) { 4196 if (isa<PHINode>(Usr)) 4197 PHIs.push_back(Usr); 4198 else 4199 NonPHIs.push_back(Usr); 4200 } 4201 // Remember that we completed the cycle. 4202 if (Usr == Phi) 4203 FoundStartPHI = true; 4204 } 4205 Worklist.append(PHIs.begin(), PHIs.end()); 4206 Worklist.append(NonPHIs.begin(), NonPHIs.end()); 4207 } 4208 4209 // This means we have seen one but not the other instruction of the 4210 // pattern or more than just a select and cmp. 4211 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) && 4212 NumCmpSelectPatternInst != 2) 4213 return false; 4214 4215 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction) 4216 return false; 4217 4218 // We found a reduction var if we have reached the original phi node and we 4219 // only have a single instruction with out-of-loop users. 4220 4221 // This instruction is allowed to have out-of-loop users. 4222 AllowedExit.insert(ExitInstruction); 4223 4224 // Save the description of this reduction variable. 4225 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind, 4226 ReduxDesc.MinMaxKind); 4227 Reductions[Phi] = RD; 4228 // We've ended the cycle. This is a reduction variable if we have an 4229 // outside user and it has a binary op. 4230 4231 return true; 4232} 4233 4234/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 4235/// pattern corresponding to a min(X, Y) or max(X, Y). 4236LoopVectorizationLegality::ReductionInstDesc 4237LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, 4238 ReductionInstDesc &Prev) { 4239 4240 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) && 4241 "Expect a select instruction"); 4242 Instruction *Cmp = 0; 4243 SelectInst *Select = 0; 4244 4245 // We must handle the select(cmp()) as a single instruction. Advance to the 4246 // select. 4247 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) { 4248 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin()))) 4249 return ReductionInstDesc(false, I); 4250 return ReductionInstDesc(Select, Prev.MinMaxKind); 4251 } 4252 4253 // Only handle single use cases for now. 4254 if (!(Select = dyn_cast<SelectInst>(I))) 4255 return ReductionInstDesc(false, I); 4256 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) && 4257 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0)))) 4258 return ReductionInstDesc(false, I); 4259 if (!Cmp->hasOneUse()) 4260 return ReductionInstDesc(false, I); 4261 4262 Value *CmpLeft; 4263 Value *CmpRight; 4264 4265 // Look for a min/max pattern. 4266 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4267 return ReductionInstDesc(Select, MRK_UIntMin); 4268 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4269 return ReductionInstDesc(Select, MRK_UIntMax); 4270 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4271 return ReductionInstDesc(Select, MRK_SIntMax); 4272 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4273 return ReductionInstDesc(Select, MRK_SIntMin); 4274 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4275 return ReductionInstDesc(Select, MRK_FloatMin); 4276 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4277 return ReductionInstDesc(Select, MRK_FloatMax); 4278 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4279 return ReductionInstDesc(Select, MRK_FloatMin); 4280 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4281 return ReductionInstDesc(Select, MRK_FloatMax); 4282 4283 return ReductionInstDesc(false, I); 4284} 4285 4286LoopVectorizationLegality::ReductionInstDesc 4287LoopVectorizationLegality::isReductionInstr(Instruction *I, 4288 ReductionKind Kind, 4289 ReductionInstDesc &Prev) { 4290 bool FP = I->getType()->isFloatingPointTy(); 4291 bool FastMath = (FP && I->isCommutative() && I->isAssociative()); 4292 switch (I->getOpcode()) { 4293 default: 4294 return ReductionInstDesc(false, I); 4295 case Instruction::PHI: 4296 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd && 4297 Kind != RK_FloatMinMax)) 4298 return ReductionInstDesc(false, I); 4299 return ReductionInstDesc(I, Prev.MinMaxKind); 4300 case Instruction::Sub: 4301 case Instruction::Add: 4302 return ReductionInstDesc(Kind == RK_IntegerAdd, I); 4303 case Instruction::Mul: 4304 return ReductionInstDesc(Kind == RK_IntegerMult, I); 4305 case Instruction::And: 4306 return ReductionInstDesc(Kind == RK_IntegerAnd, I); 4307 case Instruction::Or: 4308 return ReductionInstDesc(Kind == RK_IntegerOr, I); 4309 case Instruction::Xor: 4310 return ReductionInstDesc(Kind == RK_IntegerXor, I); 4311 case Instruction::FMul: 4312 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I); 4313 case Instruction::FAdd: 4314 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I); 4315 case Instruction::FCmp: 4316 case Instruction::ICmp: 4317 case Instruction::Select: 4318 if (Kind != RK_IntegerMinMax && 4319 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax)) 4320 return ReductionInstDesc(false, I); 4321 return isMinMaxSelectCmpPattern(I, Prev); 4322 } 4323} 4324 4325LoopVectorizationLegality::InductionKind 4326LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { 4327 Type *PhiTy = Phi->getType(); 4328 // We only handle integer and pointer inductions variables. 4329 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 4330 return IK_NoInduction; 4331 4332 // Check that the PHI is consecutive. 4333 const SCEV *PhiScev = SE->getSCEV(Phi); 4334 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 4335 if (!AR) { 4336 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 4337 return IK_NoInduction; 4338 } 4339 const SCEV *Step = AR->getStepRecurrence(*SE); 4340 4341 // Integer inductions need to have a stride of one. 4342 if (PhiTy->isIntegerTy()) { 4343 if (Step->isOne()) 4344 return IK_IntInduction; 4345 if (Step->isAllOnesValue()) 4346 return IK_ReverseIntInduction; 4347 return IK_NoInduction; 4348 } 4349 4350 // Calculate the pointer stride and check if it is consecutive. 4351 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 4352 if (!C) 4353 return IK_NoInduction; 4354 4355 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 4356 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType()); 4357 if (C->getValue()->equalsInt(Size)) 4358 return IK_PtrInduction; 4359 else if (C->getValue()->equalsInt(0 - Size)) 4360 return IK_ReversePtrInduction; 4361 4362 return IK_NoInduction; 4363} 4364 4365bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 4366 Value *In0 = const_cast<Value*>(V); 4367 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 4368 if (!PN) 4369 return false; 4370 4371 return Inductions.count(PN); 4372} 4373 4374bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 4375 assert(TheLoop->contains(BB) && "Unknown block used"); 4376 4377 // Blocks that do not dominate the latch need predication. 4378 BasicBlock* Latch = TheLoop->getLoopLatch(); 4379 return !DT->dominates(BB, Latch); 4380} 4381 4382bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB, 4383 SmallPtrSet<Value *, 8>& SafePtrs) { 4384 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4385 // We might be able to hoist the load. 4386 if (it->mayReadFromMemory()) { 4387 LoadInst *LI = dyn_cast<LoadInst>(it); 4388 if (!LI || !SafePtrs.count(LI->getPointerOperand())) 4389 return false; 4390 } 4391 4392 // We don't predicate stores at the moment. 4393 if (it->mayWriteToMemory() || it->mayThrow()) 4394 return false; 4395 4396 // Check that we don't have a constant expression that can trap as operand. 4397 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end(); 4398 OI != OE; ++OI) { 4399 if (Constant *C = dyn_cast<Constant>(*OI)) 4400 if (C->canTrap()) 4401 return false; 4402 } 4403 4404 // The instructions below can trap. 4405 switch (it->getOpcode()) { 4406 default: continue; 4407 case Instruction::UDiv: 4408 case Instruction::SDiv: 4409 case Instruction::URem: 4410 case Instruction::SRem: 4411 return false; 4412 } 4413 } 4414 4415 return true; 4416} 4417 4418LoopVectorizationCostModel::VectorizationFactor 4419LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize, 4420 unsigned UserVF) { 4421 // Width 1 means no vectorize 4422 VectorizationFactor Factor = { 1U, 0U }; 4423 if (OptForSize && Legal->getRuntimePointerCheck()->Need) { 4424 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); 4425 return Factor; 4426 } 4427 4428 // Find the trip count. 4429 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch()); 4430 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n'); 4431 4432 unsigned WidestType = getWidestType(); 4433 unsigned WidestRegister = TTI.getRegisterBitWidth(true); 4434 unsigned MaxSafeDepDist = -1U; 4435 if (Legal->getMaxSafeDepDistBytes() != -1U) 4436 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8; 4437 WidestRegister = ((WidestRegister < MaxSafeDepDist) ? 4438 WidestRegister : MaxSafeDepDist); 4439 unsigned MaxVectorSize = WidestRegister / WidestType; 4440 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n"); 4441 DEBUG(dbgs() << "LV: The Widest register is: " 4442 << WidestRegister << " bits.\n"); 4443 4444 if (MaxVectorSize == 0) { 4445 DEBUG(dbgs() << "LV: The target has no vector registers.\n"); 4446 MaxVectorSize = 1; 4447 } 4448 4449 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements" 4450 " into one vector!"); 4451 4452 unsigned VF = MaxVectorSize; 4453 4454 // If we optimize the program for size, avoid creating the tail loop. 4455 if (OptForSize) { 4456 // If we are unable to calculate the trip count then don't try to vectorize. 4457 if (TC < 2) { 4458 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 4459 return Factor; 4460 } 4461 4462 // Find the maximum SIMD width that can fit within the trip count. 4463 VF = TC % MaxVectorSize; 4464 4465 if (VF == 0) 4466 VF = MaxVectorSize; 4467 4468 // If the trip count that we found modulo the vectorization factor is not 4469 // zero then we require a tail. 4470 if (VF < 2) { 4471 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 4472 return Factor; 4473 } 4474 } 4475 4476 if (UserVF != 0) { 4477 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 4478 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n"); 4479 4480 Factor.Width = UserVF; 4481 return Factor; 4482 } 4483 4484 float Cost = expectedCost(1); 4485 unsigned Width = 1; 4486 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n"); 4487 for (unsigned i=2; i <= VF; i*=2) { 4488 // Notice that the vector loop needs to be executed less times, so 4489 // we need to divide the cost of the vector loops by the width of 4490 // the vector elements. 4491 float VectorCost = expectedCost(i) / (float)i; 4492 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " << 4493 (int)VectorCost << ".\n"); 4494 if (VectorCost < Cost) { 4495 Cost = VectorCost; 4496 Width = i; 4497 } 4498 } 4499 4500 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n"); 4501 Factor.Width = Width; 4502 Factor.Cost = Width * Cost; 4503 return Factor; 4504} 4505 4506unsigned LoopVectorizationCostModel::getWidestType() { 4507 unsigned MaxWidth = 8; 4508 4509 // For each block. 4510 for (Loop::block_iterator bb = TheLoop->block_begin(), 4511 be = TheLoop->block_end(); bb != be; ++bb) { 4512 BasicBlock *BB = *bb; 4513 4514 // For each instruction in the loop. 4515 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4516 Type *T = it->getType(); 4517 4518 // Only examine Loads, Stores and PHINodes. 4519 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it)) 4520 continue; 4521 4522 // Examine PHI nodes that are reduction variables. 4523 if (PHINode *PN = dyn_cast<PHINode>(it)) 4524 if (!Legal->getReductionVars()->count(PN)) 4525 continue; 4526 4527 // Examine the stored values. 4528 if (StoreInst *ST = dyn_cast<StoreInst>(it)) 4529 T = ST->getValueOperand()->getType(); 4530 4531 // Ignore loaded pointer types and stored pointer types that are not 4532 // consecutive. However, we do want to take consecutive stores/loads of 4533 // pointer vectors into account. 4534 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it)) 4535 continue; 4536 4537 MaxWidth = std::max(MaxWidth, 4538 (unsigned)DL->getTypeSizeInBits(T->getScalarType())); 4539 } 4540 } 4541 4542 return MaxWidth; 4543} 4544 4545unsigned 4546LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, 4547 unsigned UserUF, 4548 unsigned VF, 4549 unsigned LoopCost) { 4550 4551 // -- The unroll heuristics -- 4552 // We unroll the loop in order to expose ILP and reduce the loop overhead. 4553 // There are many micro-architectural considerations that we can't predict 4554 // at this level. For example frontend pressure (on decode or fetch) due to 4555 // code size, or the number and capabilities of the execution ports. 4556 // 4557 // We use the following heuristics to select the unroll factor: 4558 // 1. If the code has reductions the we unroll in order to break the cross 4559 // iteration dependency. 4560 // 2. If the loop is really small then we unroll in order to reduce the loop 4561 // overhead. 4562 // 3. We don't unroll if we think that we will spill registers to memory due 4563 // to the increased register pressure. 4564 4565 // Use the user preference, unless 'auto' is selected. 4566 if (UserUF != 0) 4567 return UserUF; 4568 4569 // When we optimize for size we don't unroll. 4570 if (OptForSize) 4571 return 1; 4572 4573 // We used the distance for the unroll factor. 4574 if (Legal->getMaxSafeDepDistBytes() != -1U) 4575 return 1; 4576 4577 // Do not unroll loops with a relatively small trip count. 4578 unsigned TC = SE->getSmallConstantTripCount(TheLoop, 4579 TheLoop->getLoopLatch()); 4580 if (TC > 1 && TC < TinyTripCountUnrollThreshold) 4581 return 1; 4582 4583 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true); 4584 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters << 4585 " vector registers\n"); 4586 4587 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); 4588 // We divide by these constants so assume that we have at least one 4589 // instruction that uses at least one register. 4590 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 4591 R.NumInstructions = std::max(R.NumInstructions, 1U); 4592 4593 // We calculate the unroll factor using the following formula. 4594 // Subtract the number of loop invariants from the number of available 4595 // registers. These registers are used by all of the unrolled instances. 4596 // Next, divide the remaining registers by the number of registers that is 4597 // required by the loop, in order to estimate how many parallel instances 4598 // fit without causing spills. 4599 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers; 4600 4601 // Clamp the unroll factor ranges to reasonable factors. 4602 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor(); 4603 4604 // If we did not calculate the cost for VF (because the user selected the VF) 4605 // then we calculate the cost of VF here. 4606 if (LoopCost == 0) 4607 LoopCost = expectedCost(VF); 4608 4609 // Clamp the calculated UF to be between the 1 and the max unroll factor 4610 // that the target allows. 4611 if (UF > MaxUnrollSize) 4612 UF = MaxUnrollSize; 4613 else if (UF < 1) 4614 UF = 1; 4615 4616 bool HasReductions = Legal->getReductionVars()->size(); 4617 4618 // Decide if we want to unroll if we decided that it is legal to vectorize 4619 // but not profitable. 4620 if (VF == 1) { 4621 if (TheLoop->getNumBlocks() > 1 || !HasReductions || 4622 LoopCost > SmallLoopCost) 4623 return 1; 4624 4625 return UF; 4626 } 4627 4628 if (HasReductions) { 4629 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n"); 4630 return UF; 4631 } 4632 4633 // We want to unroll tiny loops in order to reduce the loop overhead. 4634 // We assume that the cost overhead is 1 and we use the cost model 4635 // to estimate the cost of the loop and unroll until the cost of the 4636 // loop overhead is about 5% of the cost of the loop. 4637 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n'); 4638 if (LoopCost < SmallLoopCost) { 4639 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n"); 4640 unsigned NewUF = SmallLoopCost / (LoopCost + 1); 4641 return std::min(NewUF, UF); 4642 } 4643 4644 DEBUG(dbgs() << "LV: Not Unrolling.\n"); 4645 return 1; 4646} 4647 4648LoopVectorizationCostModel::RegisterUsage 4649LoopVectorizationCostModel::calculateRegisterUsage() { 4650 // This function calculates the register usage by measuring the highest number 4651 // of values that are alive at a single location. Obviously, this is a very 4652 // rough estimation. We scan the loop in a topological order in order and 4653 // assign a number to each instruction. We use RPO to ensure that defs are 4654 // met before their users. We assume that each instruction that has in-loop 4655 // users starts an interval. We record every time that an in-loop value is 4656 // used, so we have a list of the first and last occurrences of each 4657 // instruction. Next, we transpose this data structure into a multi map that 4658 // holds the list of intervals that *end* at a specific location. This multi 4659 // map allows us to perform a linear search. We scan the instructions linearly 4660 // and record each time that a new interval starts, by placing it in a set. 4661 // If we find this value in the multi-map then we remove it from the set. 4662 // The max register usage is the maximum size of the set. 4663 // We also search for instructions that are defined outside the loop, but are 4664 // used inside the loop. We need this number separately from the max-interval 4665 // usage number because when we unroll, loop-invariant values do not take 4666 // more register. 4667 LoopBlocksDFS DFS(TheLoop); 4668 DFS.perform(LI); 4669 4670 RegisterUsage R; 4671 R.NumInstructions = 0; 4672 4673 // Each 'key' in the map opens a new interval. The values 4674 // of the map are the index of the 'last seen' usage of the 4675 // instruction that is the key. 4676 typedef DenseMap<Instruction*, unsigned> IntervalMap; 4677 // Maps instruction to its index. 4678 DenseMap<unsigned, Instruction*> IdxToInstr; 4679 // Marks the end of each interval. 4680 IntervalMap EndPoint; 4681 // Saves the list of instruction indices that are used in the loop. 4682 SmallSet<Instruction*, 8> Ends; 4683 // Saves the list of values that are used in the loop but are 4684 // defined outside the loop, such as arguments and constants. 4685 SmallPtrSet<Value*, 8> LoopInvariants; 4686 4687 unsigned Index = 0; 4688 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 4689 be = DFS.endRPO(); bb != be; ++bb) { 4690 R.NumInstructions += (*bb)->size(); 4691 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 4692 ++it) { 4693 Instruction *I = it; 4694 IdxToInstr[Index++] = I; 4695 4696 // Save the end location of each USE. 4697 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 4698 Value *U = I->getOperand(i); 4699 Instruction *Instr = dyn_cast<Instruction>(U); 4700 4701 // Ignore non-instruction values such as arguments, constants, etc. 4702 if (!Instr) continue; 4703 4704 // If this instruction is outside the loop then record it and continue. 4705 if (!TheLoop->contains(Instr)) { 4706 LoopInvariants.insert(Instr); 4707 continue; 4708 } 4709 4710 // Overwrite previous end points. 4711 EndPoint[Instr] = Index; 4712 Ends.insert(Instr); 4713 } 4714 } 4715 } 4716 4717 // Saves the list of intervals that end with the index in 'key'. 4718 typedef SmallVector<Instruction*, 2> InstrList; 4719 DenseMap<unsigned, InstrList> TransposeEnds; 4720 4721 // Transpose the EndPoints to a list of values that end at each index. 4722 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 4723 it != e; ++it) 4724 TransposeEnds[it->second].push_back(it->first); 4725 4726 SmallSet<Instruction*, 8> OpenIntervals; 4727 unsigned MaxUsage = 0; 4728 4729 4730 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 4731 for (unsigned int i = 0; i < Index; ++i) { 4732 Instruction *I = IdxToInstr[i]; 4733 // Ignore instructions that are never used within the loop. 4734 if (!Ends.count(I)) continue; 4735 4736 // Remove all of the instructions that end at this location. 4737 InstrList &List = TransposeEnds[i]; 4738 for (unsigned int j=0, e = List.size(); j < e; ++j) 4739 OpenIntervals.erase(List[j]); 4740 4741 // Count the number of live interals. 4742 MaxUsage = std::max(MaxUsage, OpenIntervals.size()); 4743 4744 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << 4745 OpenIntervals.size() << '\n'); 4746 4747 // Add the current instruction to the list of open intervals. 4748 OpenIntervals.insert(I); 4749 } 4750 4751 unsigned Invariant = LoopInvariants.size(); 4752 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n'); 4753 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n'); 4754 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n'); 4755 4756 R.LoopInvariantRegs = Invariant; 4757 R.MaxLocalUsers = MaxUsage; 4758 return R; 4759} 4760 4761unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 4762 unsigned Cost = 0; 4763 4764 // For each block. 4765 for (Loop::block_iterator bb = TheLoop->block_begin(), 4766 be = TheLoop->block_end(); bb != be; ++bb) { 4767 unsigned BlockCost = 0; 4768 BasicBlock *BB = *bb; 4769 4770 // For each instruction in the old loop. 4771 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4772 // Skip dbg intrinsics. 4773 if (isa<DbgInfoIntrinsic>(it)) 4774 continue; 4775 4776 unsigned C = getInstructionCost(it, VF); 4777 BlockCost += C; 4778 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " << 4779 VF << " For instruction: " << *it << '\n'); 4780 } 4781 4782 // We assume that if-converted blocks have a 50% chance of being executed. 4783 // When the code is scalar then some of the blocks are avoided due to CF. 4784 // When the code is vectorized we execute all code paths. 4785 if (VF == 1 && Legal->blockNeedsPredication(*bb)) 4786 BlockCost /= 2; 4787 4788 Cost += BlockCost; 4789 } 4790 4791 return Cost; 4792} 4793 4794/// \brief Check whether the address computation for a non-consecutive memory 4795/// access looks like an unlikely candidate for being merged into the indexing 4796/// mode. 4797/// 4798/// We look for a GEP which has one index that is an induction variable and all 4799/// other indices are loop invariant. If the stride of this access is also 4800/// within a small bound we decide that this address computation can likely be 4801/// merged into the addressing mode. 4802/// In all other cases, we identify the address computation as complex. 4803static bool isLikelyComplexAddressComputation(Value *Ptr, 4804 LoopVectorizationLegality *Legal, 4805 ScalarEvolution *SE, 4806 const Loop *TheLoop) { 4807 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 4808 if (!Gep) 4809 return true; 4810 4811 // We are looking for a gep with all loop invariant indices except for one 4812 // which should be an induction variable. 4813 unsigned NumOperands = Gep->getNumOperands(); 4814 for (unsigned i = 1; i < NumOperands; ++i) { 4815 Value *Opd = Gep->getOperand(i); 4816 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && 4817 !Legal->isInductionVariable(Opd)) 4818 return true; 4819 } 4820 4821 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step 4822 // can likely be merged into the address computation. 4823 unsigned MaxMergeDistance = 64; 4824 4825 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr)); 4826 if (!AddRec) 4827 return true; 4828 4829 // Check the step is constant. 4830 const SCEV *Step = AddRec->getStepRecurrence(*SE); 4831 // Calculate the pointer stride and check if it is consecutive. 4832 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 4833 if (!C) 4834 return true; 4835 4836 const APInt &APStepVal = C->getValue()->getValue(); 4837 4838 // Huge step value - give up. 4839 if (APStepVal.getBitWidth() > 64) 4840 return true; 4841 4842 int64_t StepVal = APStepVal.getSExtValue(); 4843 4844 return StepVal > MaxMergeDistance; 4845} 4846 4847unsigned 4848LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 4849 // If we know that this instruction will remain uniform, check the cost of 4850 // the scalar version. 4851 if (Legal->isUniformAfterVectorization(I)) 4852 VF = 1; 4853 4854 Type *RetTy = I->getType(); 4855 Type *VectorTy = ToVectorTy(RetTy, VF); 4856 4857 // TODO: We need to estimate the cost of intrinsic calls. 4858 switch (I->getOpcode()) { 4859 case Instruction::GetElementPtr: 4860 // We mark this instruction as zero-cost because the cost of GEPs in 4861 // vectorized code depends on whether the corresponding memory instruction 4862 // is scalarized or not. Therefore, we handle GEPs with the memory 4863 // instruction cost. 4864 return 0; 4865 case Instruction::Br: { 4866 return TTI.getCFInstrCost(I->getOpcode()); 4867 } 4868 case Instruction::PHI: 4869 //TODO: IF-converted IFs become selects. 4870 return 0; 4871 case Instruction::Add: 4872 case Instruction::FAdd: 4873 case Instruction::Sub: 4874 case Instruction::FSub: 4875 case Instruction::Mul: 4876 case Instruction::FMul: 4877 case Instruction::UDiv: 4878 case Instruction::SDiv: 4879 case Instruction::FDiv: 4880 case Instruction::URem: 4881 case Instruction::SRem: 4882 case Instruction::FRem: 4883 case Instruction::Shl: 4884 case Instruction::LShr: 4885 case Instruction::AShr: 4886 case Instruction::And: 4887 case Instruction::Or: 4888 case Instruction::Xor: { 4889 // Certain instructions can be cheaper to vectorize if they have a constant 4890 // second vector operand. One example of this are shifts on x86. 4891 TargetTransformInfo::OperandValueKind Op1VK = 4892 TargetTransformInfo::OK_AnyValue; 4893 TargetTransformInfo::OperandValueKind Op2VK = 4894 TargetTransformInfo::OK_AnyValue; 4895 4896 if (isa<ConstantInt>(I->getOperand(1))) 4897 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 4898 4899 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK); 4900 } 4901 case Instruction::Select: { 4902 SelectInst *SI = cast<SelectInst>(I); 4903 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 4904 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 4905 Type *CondTy = SI->getCondition()->getType(); 4906 if (!ScalarCond) 4907 CondTy = VectorType::get(CondTy, VF); 4908 4909 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 4910 } 4911 case Instruction::ICmp: 4912 case Instruction::FCmp: { 4913 Type *ValTy = I->getOperand(0)->getType(); 4914 VectorTy = ToVectorTy(ValTy, VF); 4915 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); 4916 } 4917 case Instruction::Store: 4918 case Instruction::Load: { 4919 StoreInst *SI = dyn_cast<StoreInst>(I); 4920 LoadInst *LI = dyn_cast<LoadInst>(I); 4921 Type *ValTy = (SI ? SI->getValueOperand()->getType() : 4922 LI->getType()); 4923 VectorTy = ToVectorTy(ValTy, VF); 4924 4925 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); 4926 unsigned AS = SI ? SI->getPointerAddressSpace() : 4927 LI->getPointerAddressSpace(); 4928 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); 4929 // We add the cost of address computation here instead of with the gep 4930 // instruction because only here we know whether the operation is 4931 // scalarized. 4932 if (VF == 1) 4933 return TTI.getAddressComputationCost(VectorTy) + 4934 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 4935 4936 // Scalarized loads/stores. 4937 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 4938 bool Reverse = ConsecutiveStride < 0; 4939 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy); 4940 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF; 4941 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { 4942 bool IsComplexComputation = 4943 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop); 4944 unsigned Cost = 0; 4945 // The cost of extracting from the value vector and pointer vector. 4946 Type *PtrTy = ToVectorTy(Ptr->getType(), VF); 4947 for (unsigned i = 0; i < VF; ++i) { 4948 // The cost of extracting the pointer operand. 4949 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); 4950 // In case of STORE, the cost of ExtractElement from the vector. 4951 // In case of LOAD, the cost of InsertElement into the returned 4952 // vector. 4953 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : 4954 Instruction::InsertElement, 4955 VectorTy, i); 4956 } 4957 4958 // The cost of the scalar loads/stores. 4959 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation); 4960 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), 4961 Alignment, AS); 4962 return Cost; 4963 } 4964 4965 // Wide load/stores. 4966 unsigned Cost = TTI.getAddressComputationCost(VectorTy); 4967 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 4968 4969 if (Reverse) 4970 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, 4971 VectorTy, 0); 4972 return Cost; 4973 } 4974 case Instruction::ZExt: 4975 case Instruction::SExt: 4976 case Instruction::FPToUI: 4977 case Instruction::FPToSI: 4978 case Instruction::FPExt: 4979 case Instruction::PtrToInt: 4980 case Instruction::IntToPtr: 4981 case Instruction::SIToFP: 4982 case Instruction::UIToFP: 4983 case Instruction::Trunc: 4984 case Instruction::FPTrunc: 4985 case Instruction::BitCast: { 4986 // We optimize the truncation of induction variable. 4987 // The cost of these is the same as the scalar operation. 4988 if (I->getOpcode() == Instruction::Trunc && 4989 Legal->isInductionVariable(I->getOperand(0))) 4990 return TTI.getCastInstrCost(I->getOpcode(), I->getType(), 4991 I->getOperand(0)->getType()); 4992 4993 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); 4994 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 4995 } 4996 case Instruction::Call: { 4997 CallInst *CI = cast<CallInst>(I); 4998 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 4999 assert(ID && "Not an intrinsic call!"); 5000 Type *RetTy = ToVectorTy(CI->getType(), VF); 5001 SmallVector<Type*, 4> Tys; 5002 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 5003 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 5004 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); 5005 } 5006 default: { 5007 // We are scalarizing the instruction. Return the cost of the scalar 5008 // instruction, plus the cost of insert and extract into vector 5009 // elements, times the vector width. 5010 unsigned Cost = 0; 5011 5012 if (!RetTy->isVoidTy() && VF != 1) { 5013 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, 5014 VectorTy); 5015 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, 5016 VectorTy); 5017 5018 // The cost of inserting the results plus extracting each one of the 5019 // operands. 5020 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 5021 } 5022 5023 // The cost of executing VF copies of the scalar instruction. This opcode 5024 // is unknown. Assume that it is the same as 'mul'. 5025 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); 5026 return Cost; 5027 } 5028 }// end of switch. 5029} 5030 5031Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) { 5032 if (Scalar->isVoidTy() || VF == 1) 5033 return Scalar; 5034 return VectorType::get(Scalar, VF); 5035} 5036 5037char LoopVectorize::ID = 0; 5038static const char lv_name[] = "Loop Vectorization"; 5039INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 5040INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 5041INITIALIZE_PASS_DEPENDENCY(DominatorTree) 5042INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 5043INITIALIZE_PASS_DEPENDENCY(LCSSA) 5044INITIALIZE_PASS_DEPENDENCY(LoopInfo) 5045INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 5046INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 5047 5048namespace llvm { 5049 Pass *createLoopVectorizePass(bool NoUnrolling) { 5050 return new LoopVectorize(NoUnrolling); 5051 } 5052} 5053 5054bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { 5055 // Check for a store. 5056 if (StoreInst *ST = dyn_cast<StoreInst>(Inst)) 5057 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; 5058 5059 // Check for a load. 5060 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 5061 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; 5062 5063 return false; 5064} 5065 5066 5067void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) { 5068 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 5069 // Holds vector parameters or scalars, in case of uniform vals. 5070 SmallVector<VectorParts, 4> Params; 5071 5072 setDebugLocFromInst(Builder, Instr); 5073 5074 // Find all of the vectorized parameters. 5075 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5076 Value *SrcOp = Instr->getOperand(op); 5077 5078 // If we are accessing the old induction variable, use the new one. 5079 if (SrcOp == OldInduction) { 5080 Params.push_back(getVectorValue(SrcOp)); 5081 continue; 5082 } 5083 5084 // Try using previously calculated values. 5085 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 5086 5087 // If the src is an instruction that appeared earlier in the basic block 5088 // then it should already be vectorized. 5089 if (SrcInst && OrigLoop->contains(SrcInst)) { 5090 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 5091 // The parameter is a vector value from earlier. 5092 Params.push_back(WidenMap.get(SrcInst)); 5093 } else { 5094 // The parameter is a scalar from outside the loop. Maybe even a constant. 5095 VectorParts Scalars; 5096 Scalars.append(UF, SrcOp); 5097 Params.push_back(Scalars); 5098 } 5099 } 5100 5101 assert(Params.size() == Instr->getNumOperands() && 5102 "Invalid number of operands"); 5103 5104 // Does this instruction return a value ? 5105 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 5106 5107 Value *UndefVec = IsVoidRetTy ? 0 : 5108 UndefValue::get(Instr->getType()); 5109 // Create a new entry in the WidenMap and initialize it to Undef or Null. 5110 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 5111 5112 // For each vector unroll 'part': 5113 for (unsigned Part = 0; Part < UF; ++Part) { 5114 // For each scalar that we create: 5115 5116 Instruction *Cloned = Instr->clone(); 5117 if (!IsVoidRetTy) 5118 Cloned->setName(Instr->getName() + ".cloned"); 5119 // Replace the operands of the cloned instructions with extracted scalars. 5120 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5121 Value *Op = Params[op][Part]; 5122 Cloned->setOperand(op, Op); 5123 } 5124 5125 // Place the cloned scalar in the new loop. 5126 Builder.Insert(Cloned); 5127 5128 // If the original scalar returns a value we need to place it in a vector 5129 // so that future users will be able to use it. 5130 if (!IsVoidRetTy) 5131 VecResults[Part] = Cloned; 5132 } 5133} 5134 5135void 5136InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr, 5137 LoopVectorizationLegality*) { 5138 return scalarizeInstruction(Instr); 5139} 5140 5141Value *InnerLoopUnroller::reverseVector(Value *Vec) { 5142 return Vec; 5143} 5144 5145Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { 5146 return V; 5147} 5148 5149Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx, 5150 bool Negate) { 5151 // When unrolling and the VF is 1, we only need to add a simple scalar. 5152 Type *ITy = Val->getType(); 5153 assert(!ITy->isVectorTy() && "Val must be a scalar"); 5154 Constant *C = ConstantInt::get(ITy, StartIdx, Negate); 5155 return Builder.CreateAdd(Val, C, "induction"); 5156} 5157 5158