LoopVectorize.cpp revision 4634338655449a6e76b6a948c91fbe5bc736d24b
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 // Get the total trip count from the count by adding 1. 1541 ExitCount = SE->getAddExpr(ExitCount, 1542 SE->getConstant(ExitCount->getType(), 1)); 1543 1544 // Expand the trip count and place the new instructions in the preheader. 1545 // Notice that the pre-header does not change, only the loop body. 1546 SCEVExpander Exp(*SE, "induction"); 1547 1548 // Count holds the overall loop count (N). 1549 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 1550 BypassBlock->getTerminator()); 1551 1552 // The loop index does not have to start at Zero. Find the original start 1553 // value from the induction PHI node. If we don't have an induction variable 1554 // then we know that it starts at zero. 1555 Builder.SetInsertPoint(BypassBlock->getTerminator()); 1556 Value *StartIdx = ExtendedIdx = OldInduction ? 1557 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock), 1558 IdxTy): 1559 ConstantInt::get(IdxTy, 0); 1560 1561 assert(BypassBlock && "Invalid loop structure"); 1562 LoopBypassBlocks.push_back(BypassBlock); 1563 1564 // Split the single block loop into the two loop structure described above. 1565 BasicBlock *VectorPH = 1566 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); 1567 BasicBlock *VecBody = 1568 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 1569 BasicBlock *MiddleBlock = 1570 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 1571 BasicBlock *ScalarPH = 1572 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 1573 1574 // Create and register the new vector loop. 1575 Loop* Lp = new Loop(); 1576 Loop *ParentLoop = OrigLoop->getParentLoop(); 1577 1578 // Insert the new loop into the loop nest and register the new basic blocks 1579 // before calling any utilities such as SCEV that require valid LoopInfo. 1580 if (ParentLoop) { 1581 ParentLoop->addChildLoop(Lp); 1582 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); 1583 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); 1584 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); 1585 } else { 1586 LI->addTopLevelLoop(Lp); 1587 } 1588 Lp->addBasicBlockToLoop(VecBody, LI->getBase()); 1589 1590 // Use this IR builder to create the loop instructions (Phi, Br, Cmp) 1591 // inside the loop. 1592 Builder.SetInsertPoint(VecBody->getFirstNonPHI()); 1593 1594 // Generate the induction variable. 1595 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction)); 1596 Induction = Builder.CreatePHI(IdxTy, 2, "index"); 1597 // The loop step is equal to the vectorization factor (num of SIMD elements) 1598 // times the unroll factor (num of SIMD instructions). 1599 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 1600 1601 // This is the IR builder that we use to add all of the logic for bypassing 1602 // the new vector loop. 1603 IRBuilder<> BypassBuilder(BypassBlock->getTerminator()); 1604 setDebugLocFromInst(BypassBuilder, 1605 getDebugLocFromInstOrOperands(OldInduction)); 1606 1607 // We may need to extend the index in case there is a type mismatch. 1608 // We know that the count starts at zero and does not overflow. 1609 if (Count->getType() != IdxTy) { 1610 // The exit count can be of pointer type. Convert it to the correct 1611 // integer type. 1612 if (ExitCount->getType()->isPointerTy()) 1613 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int"); 1614 else 1615 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast"); 1616 } 1617 1618 // Add the start index to the loop count to get the new end index. 1619 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx"); 1620 1621 // Now we need to generate the expression for N - (N % VF), which is 1622 // the part that the vectorized body will execute. 1623 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf"); 1624 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec"); 1625 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx, 1626 "end.idx.rnd.down"); 1627 1628 // Now, compare the new count to zero. If it is zero skip the vector loop and 1629 // jump to the scalar loop. 1630 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, 1631 "cmp.zero"); 1632 1633 BasicBlock *LastBypassBlock = BypassBlock; 1634 1635 // Generate the code that checks in runtime if arrays overlap. We put the 1636 // checks into a separate block to make the more common case of few elements 1637 // faster. 1638 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal, 1639 BypassBlock->getTerminator()); 1640 if (MemRuntimeCheck) { 1641 // Create a new block containing the memory check. 1642 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck, 1643 "vector.memcheck"); 1644 if (ParentLoop) 1645 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase()); 1646 LoopBypassBlocks.push_back(CheckBlock); 1647 1648 // Replace the branch into the memory check block with a conditional branch 1649 // for the "few elements case". 1650 Instruction *OldTerm = BypassBlock->getTerminator(); 1651 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); 1652 OldTerm->eraseFromParent(); 1653 1654 Cmp = MemRuntimeCheck; 1655 LastBypassBlock = CheckBlock; 1656 } 1657 1658 LastBypassBlock->getTerminator()->eraseFromParent(); 1659 BranchInst::Create(MiddleBlock, VectorPH, Cmp, 1660 LastBypassBlock); 1661 1662 // We are going to resume the execution of the scalar loop. 1663 // Go over all of the induction variables that we found and fix the 1664 // PHIs that are left in the scalar version of the loop. 1665 // The starting values of PHI nodes depend on the counter of the last 1666 // iteration in the vectorized loop. 1667 // If we come from a bypass edge then we need to start from the original 1668 // start value. 1669 1670 // This variable saves the new starting index for the scalar loop. 1671 PHINode *ResumeIndex = 0; 1672 LoopVectorizationLegality::InductionList::iterator I, E; 1673 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 1674 // Set builder to point to last bypass block. 1675 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator()); 1676 for (I = List->begin(), E = List->end(); I != E; ++I) { 1677 PHINode *OrigPhi = I->first; 1678 LoopVectorizationLegality::InductionInfo II = I->second; 1679 1680 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType(); 1681 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val", 1682 MiddleBlock->getTerminator()); 1683 // We might have extended the type of the induction variable but we need a 1684 // truncated version for the scalar loop. 1685 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ? 1686 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val", 1687 MiddleBlock->getTerminator()) : 0; 1688 1689 Value *EndValue = 0; 1690 switch (II.IK) { 1691 case LoopVectorizationLegality::IK_NoInduction: 1692 llvm_unreachable("Unknown induction"); 1693 case LoopVectorizationLegality::IK_IntInduction: { 1694 // Handle the integer induction counter. 1695 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); 1696 1697 // We have the canonical induction variable. 1698 if (OrigPhi == OldInduction) { 1699 // Create a truncated version of the resume value for the scalar loop, 1700 // we might have promoted the type to a larger width. 1701 EndValue = 1702 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType()); 1703 // The new PHI merges the original incoming value, in case of a bypass, 1704 // or the value at the end of the vectorized loop. 1705 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1706 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1707 TruncResumeVal->addIncoming(EndValue, VecBody); 1708 1709 // We know what the end value is. 1710 EndValue = IdxEndRoundDown; 1711 // We also know which PHI node holds it. 1712 ResumeIndex = ResumeVal; 1713 break; 1714 } 1715 1716 // Not the canonical induction variable - add the vector loop count to the 1717 // start value. 1718 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, 1719 II.StartValue->getType(), 1720 "cast.crd"); 1721 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end"); 1722 break; 1723 } 1724 case LoopVectorizationLegality::IK_ReverseIntInduction: { 1725 // Convert the CountRoundDown variable to the PHI size. 1726 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, 1727 II.StartValue->getType(), 1728 "cast.crd"); 1729 // Handle reverse integer induction counter. 1730 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end"); 1731 break; 1732 } 1733 case LoopVectorizationLegality::IK_PtrInduction: { 1734 // For pointer induction variables, calculate the offset using 1735 // the end index. 1736 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown, 1737 "ptr.ind.end"); 1738 break; 1739 } 1740 case LoopVectorizationLegality::IK_ReversePtrInduction: { 1741 // The value at the end of the loop for the reverse pointer is calculated 1742 // by creating a GEP with a negative index starting from the start value. 1743 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0); 1744 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown, 1745 "rev.ind.end"); 1746 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx, 1747 "rev.ptr.ind.end"); 1748 break; 1749 } 1750 }// end of case 1751 1752 // The new PHI merges the original incoming value, in case of a bypass, 1753 // or the value at the end of the vectorized loop. 1754 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) { 1755 if (OrigPhi == OldInduction) 1756 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]); 1757 else 1758 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1759 } 1760 ResumeVal->addIncoming(EndValue, VecBody); 1761 1762 // Fix the scalar body counter (PHI node). 1763 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 1764 // The old inductions phi node in the scalar body needs the truncated value. 1765 if (OrigPhi == OldInduction) 1766 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal); 1767 else 1768 OrigPhi->setIncomingValue(BlockIdx, ResumeVal); 1769 } 1770 1771 // If we are generating a new induction variable then we also need to 1772 // generate the code that calculates the exit value. This value is not 1773 // simply the end of the counter because we may skip the vectorized body 1774 // in case of a runtime check. 1775 if (!OldInduction){ 1776 assert(!ResumeIndex && "Unexpected resume value found"); 1777 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", 1778 MiddleBlock->getTerminator()); 1779 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1780 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]); 1781 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); 1782 } 1783 1784 // Make sure that we found the index where scalar loop needs to continue. 1785 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() && 1786 "Invalid resume Index"); 1787 1788 // Add a check in the middle block to see if we have completed 1789 // all of the iterations in the first vector loop. 1790 // If (N - N%VF) == N, then we *don't* need to run the remainder. 1791 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, 1792 ResumeIndex, "cmp.n", 1793 MiddleBlock->getTerminator()); 1794 1795 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); 1796 // Remove the old terminator. 1797 MiddleBlock->getTerminator()->eraseFromParent(); 1798 1799 // Create i+1 and fill the PHINode. 1800 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); 1801 Induction->addIncoming(StartIdx, VectorPH); 1802 Induction->addIncoming(NextIdx, VecBody); 1803 // Create the compare. 1804 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); 1805 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); 1806 1807 // Now we have two terminators. Remove the old one from the block. 1808 VecBody->getTerminator()->eraseFromParent(); 1809 1810 // Get ready to start creating new instructions into the vectorized body. 1811 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1812 1813 // Save the state. 1814 LoopVectorPreHeader = VectorPH; 1815 LoopScalarPreHeader = ScalarPH; 1816 LoopMiddleBlock = MiddleBlock; 1817 LoopExitBlock = ExitBlock; 1818 LoopVectorBody = VecBody; 1819 LoopScalarBody = OldBasicBlock; 1820 1821 LoopVectorizeHints Hints(Lp, true); 1822 Hints.setAlreadyVectorized(Lp); 1823} 1824 1825/// This function returns the identity element (or neutral element) for 1826/// the operation K. 1827Constant* 1828LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) { 1829 switch (K) { 1830 case RK_IntegerXor: 1831 case RK_IntegerAdd: 1832 case RK_IntegerOr: 1833 // Adding, Xoring, Oring zero to a number does not change it. 1834 return ConstantInt::get(Tp, 0); 1835 case RK_IntegerMult: 1836 // Multiplying a number by 1 does not change it. 1837 return ConstantInt::get(Tp, 1); 1838 case RK_IntegerAnd: 1839 // AND-ing a number with an all-1 value does not change it. 1840 return ConstantInt::get(Tp, -1, true); 1841 case RK_FloatMult: 1842 // Multiplying a number by 1 does not change it. 1843 return ConstantFP::get(Tp, 1.0L); 1844 case RK_FloatAdd: 1845 // Adding zero to a number does not change it. 1846 return ConstantFP::get(Tp, 0.0L); 1847 default: 1848 llvm_unreachable("Unknown reduction kind"); 1849 } 1850} 1851 1852static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I, 1853 Intrinsic::ID ValidIntrinsicID) { 1854 if (I.getNumArgOperands() != 1 || 1855 !I.getArgOperand(0)->getType()->isFloatingPointTy() || 1856 I.getType() != I.getArgOperand(0)->getType() || 1857 !I.onlyReadsMemory()) 1858 return Intrinsic::not_intrinsic; 1859 1860 return ValidIntrinsicID; 1861} 1862 1863static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I, 1864 Intrinsic::ID ValidIntrinsicID) { 1865 if (I.getNumArgOperands() != 2 || 1866 !I.getArgOperand(0)->getType()->isFloatingPointTy() || 1867 !I.getArgOperand(1)->getType()->isFloatingPointTy() || 1868 I.getType() != I.getArgOperand(0)->getType() || 1869 I.getType() != I.getArgOperand(1)->getType() || 1870 !I.onlyReadsMemory()) 1871 return Intrinsic::not_intrinsic; 1872 1873 return ValidIntrinsicID; 1874} 1875 1876 1877static Intrinsic::ID 1878getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) { 1879 // If we have an intrinsic call, check if it is trivially vectorizable. 1880 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) { 1881 switch (II->getIntrinsicID()) { 1882 case Intrinsic::sqrt: 1883 case Intrinsic::sin: 1884 case Intrinsic::cos: 1885 case Intrinsic::exp: 1886 case Intrinsic::exp2: 1887 case Intrinsic::log: 1888 case Intrinsic::log10: 1889 case Intrinsic::log2: 1890 case Intrinsic::fabs: 1891 case Intrinsic::copysign: 1892 case Intrinsic::floor: 1893 case Intrinsic::ceil: 1894 case Intrinsic::trunc: 1895 case Intrinsic::rint: 1896 case Intrinsic::nearbyint: 1897 case Intrinsic::round: 1898 case Intrinsic::pow: 1899 case Intrinsic::fma: 1900 case Intrinsic::fmuladd: 1901 case Intrinsic::lifetime_start: 1902 case Intrinsic::lifetime_end: 1903 return II->getIntrinsicID(); 1904 default: 1905 return Intrinsic::not_intrinsic; 1906 } 1907 } 1908 1909 if (!TLI) 1910 return Intrinsic::not_intrinsic; 1911 1912 LibFunc::Func Func; 1913 Function *F = CI->getCalledFunction(); 1914 // We're going to make assumptions on the semantics of the functions, check 1915 // that the target knows that it's available in this environment and it does 1916 // not have local linkage. 1917 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func)) 1918 return Intrinsic::not_intrinsic; 1919 1920 // Otherwise check if we have a call to a function that can be turned into a 1921 // vector intrinsic. 1922 switch (Func) { 1923 default: 1924 break; 1925 case LibFunc::sin: 1926 case LibFunc::sinf: 1927 case LibFunc::sinl: 1928 return checkUnaryFloatSignature(*CI, Intrinsic::sin); 1929 case LibFunc::cos: 1930 case LibFunc::cosf: 1931 case LibFunc::cosl: 1932 return checkUnaryFloatSignature(*CI, Intrinsic::cos); 1933 case LibFunc::exp: 1934 case LibFunc::expf: 1935 case LibFunc::expl: 1936 return checkUnaryFloatSignature(*CI, Intrinsic::exp); 1937 case LibFunc::exp2: 1938 case LibFunc::exp2f: 1939 case LibFunc::exp2l: 1940 return checkUnaryFloatSignature(*CI, Intrinsic::exp2); 1941 case LibFunc::log: 1942 case LibFunc::logf: 1943 case LibFunc::logl: 1944 return checkUnaryFloatSignature(*CI, Intrinsic::log); 1945 case LibFunc::log10: 1946 case LibFunc::log10f: 1947 case LibFunc::log10l: 1948 return checkUnaryFloatSignature(*CI, Intrinsic::log10); 1949 case LibFunc::log2: 1950 case LibFunc::log2f: 1951 case LibFunc::log2l: 1952 return checkUnaryFloatSignature(*CI, Intrinsic::log2); 1953 case LibFunc::fabs: 1954 case LibFunc::fabsf: 1955 case LibFunc::fabsl: 1956 return checkUnaryFloatSignature(*CI, Intrinsic::fabs); 1957 case LibFunc::copysign: 1958 case LibFunc::copysignf: 1959 case LibFunc::copysignl: 1960 return checkBinaryFloatSignature(*CI, Intrinsic::copysign); 1961 case LibFunc::floor: 1962 case LibFunc::floorf: 1963 case LibFunc::floorl: 1964 return checkUnaryFloatSignature(*CI, Intrinsic::floor); 1965 case LibFunc::ceil: 1966 case LibFunc::ceilf: 1967 case LibFunc::ceill: 1968 return checkUnaryFloatSignature(*CI, Intrinsic::ceil); 1969 case LibFunc::trunc: 1970 case LibFunc::truncf: 1971 case LibFunc::truncl: 1972 return checkUnaryFloatSignature(*CI, Intrinsic::trunc); 1973 case LibFunc::rint: 1974 case LibFunc::rintf: 1975 case LibFunc::rintl: 1976 return checkUnaryFloatSignature(*CI, Intrinsic::rint); 1977 case LibFunc::nearbyint: 1978 case LibFunc::nearbyintf: 1979 case LibFunc::nearbyintl: 1980 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint); 1981 case LibFunc::round: 1982 case LibFunc::roundf: 1983 case LibFunc::roundl: 1984 return checkUnaryFloatSignature(*CI, Intrinsic::round); 1985 case LibFunc::pow: 1986 case LibFunc::powf: 1987 case LibFunc::powl: 1988 return checkBinaryFloatSignature(*CI, Intrinsic::pow); 1989 } 1990 1991 return Intrinsic::not_intrinsic; 1992} 1993 1994/// This function translates the reduction kind to an LLVM binary operator. 1995static unsigned 1996getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) { 1997 switch (Kind) { 1998 case LoopVectorizationLegality::RK_IntegerAdd: 1999 return Instruction::Add; 2000 case LoopVectorizationLegality::RK_IntegerMult: 2001 return Instruction::Mul; 2002 case LoopVectorizationLegality::RK_IntegerOr: 2003 return Instruction::Or; 2004 case LoopVectorizationLegality::RK_IntegerAnd: 2005 return Instruction::And; 2006 case LoopVectorizationLegality::RK_IntegerXor: 2007 return Instruction::Xor; 2008 case LoopVectorizationLegality::RK_FloatMult: 2009 return Instruction::FMul; 2010 case LoopVectorizationLegality::RK_FloatAdd: 2011 return Instruction::FAdd; 2012 case LoopVectorizationLegality::RK_IntegerMinMax: 2013 return Instruction::ICmp; 2014 case LoopVectorizationLegality::RK_FloatMinMax: 2015 return Instruction::FCmp; 2016 default: 2017 llvm_unreachable("Unknown reduction operation"); 2018 } 2019} 2020 2021Value *createMinMaxOp(IRBuilder<> &Builder, 2022 LoopVectorizationLegality::MinMaxReductionKind RK, 2023 Value *Left, 2024 Value *Right) { 2025 CmpInst::Predicate P = CmpInst::ICMP_NE; 2026 switch (RK) { 2027 default: 2028 llvm_unreachable("Unknown min/max reduction kind"); 2029 case LoopVectorizationLegality::MRK_UIntMin: 2030 P = CmpInst::ICMP_ULT; 2031 break; 2032 case LoopVectorizationLegality::MRK_UIntMax: 2033 P = CmpInst::ICMP_UGT; 2034 break; 2035 case LoopVectorizationLegality::MRK_SIntMin: 2036 P = CmpInst::ICMP_SLT; 2037 break; 2038 case LoopVectorizationLegality::MRK_SIntMax: 2039 P = CmpInst::ICMP_SGT; 2040 break; 2041 case LoopVectorizationLegality::MRK_FloatMin: 2042 P = CmpInst::FCMP_OLT; 2043 break; 2044 case LoopVectorizationLegality::MRK_FloatMax: 2045 P = CmpInst::FCMP_OGT; 2046 break; 2047 } 2048 2049 Value *Cmp; 2050 if (RK == LoopVectorizationLegality::MRK_FloatMin || 2051 RK == LoopVectorizationLegality::MRK_FloatMax) 2052 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp"); 2053 else 2054 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); 2055 2056 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); 2057 return Select; 2058} 2059 2060namespace { 2061struct CSEDenseMapInfo { 2062 static bool canHandle(Instruction *I) { 2063 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || 2064 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); 2065 } 2066 static inline Instruction *getEmptyKey() { 2067 return DenseMapInfo<Instruction *>::getEmptyKey(); 2068 } 2069 static inline Instruction *getTombstoneKey() { 2070 return DenseMapInfo<Instruction *>::getTombstoneKey(); 2071 } 2072 static unsigned getHashValue(Instruction *I) { 2073 assert(canHandle(I) && "Unknown instruction!"); 2074 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), 2075 I->value_op_end())); 2076 } 2077 static bool isEqual(Instruction *LHS, Instruction *RHS) { 2078 if (LHS == getEmptyKey() || RHS == getEmptyKey() || 2079 LHS == getTombstoneKey() || RHS == getTombstoneKey()) 2080 return LHS == RHS; 2081 return LHS->isIdenticalTo(RHS); 2082 } 2083}; 2084} 2085 2086///\brief Perform cse of induction variable instructions. 2087static void cse(BasicBlock *BB) { 2088 // Perform simple cse. 2089 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; 2090 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { 2091 Instruction *In = I++; 2092 2093 if (!CSEDenseMapInfo::canHandle(In)) 2094 continue; 2095 2096 // Check if we can replace this instruction with any of the 2097 // visited instructions. 2098 if (Instruction *V = CSEMap.lookup(In)) { 2099 In->replaceAllUsesWith(V); 2100 In->eraseFromParent(); 2101 continue; 2102 } 2103 2104 CSEMap[In] = In; 2105 } 2106} 2107 2108void 2109InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { 2110 //===------------------------------------------------===// 2111 // 2112 // Notice: any optimization or new instruction that go 2113 // into the code below should be also be implemented in 2114 // the cost-model. 2115 // 2116 //===------------------------------------------------===// 2117 Constant *Zero = Builder.getInt32(0); 2118 2119 // In order to support reduction variables we need to be able to vectorize 2120 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two 2121 // stages. First, we create a new vector PHI node with no incoming edges. 2122 // We use this value when we vectorize all of the instructions that use the 2123 // PHI. Next, after all of the instructions in the block are complete we 2124 // add the new incoming edges to the PHI. At this point all of the 2125 // instructions in the basic block are vectorized, so we can use them to 2126 // construct the PHI. 2127 PhiVector RdxPHIsToFix; 2128 2129 // Scan the loop in a topological order to ensure that defs are vectorized 2130 // before users. 2131 LoopBlocksDFS DFS(OrigLoop); 2132 DFS.perform(LI); 2133 2134 // Vectorize all of the blocks in the original loop. 2135 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 2136 be = DFS.endRPO(); bb != be; ++bb) 2137 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix); 2138 2139 // At this point every instruction in the original loop is widened to 2140 // a vector form. We are almost done. Now, we need to fix the PHI nodes 2141 // that we vectorized. The PHI nodes are currently empty because we did 2142 // not want to introduce cycles. Notice that the remaining PHI nodes 2143 // that we need to fix are reduction variables. 2144 2145 // Create the 'reduced' values for each of the induction vars. 2146 // The reduced values are the vector values that we scalarize and combine 2147 // after the loop is finished. 2148 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); 2149 it != e; ++it) { 2150 PHINode *RdxPhi = *it; 2151 assert(RdxPhi && "Unable to recover vectorized PHI"); 2152 2153 // Find the reduction variable descriptor. 2154 assert(Legal->getReductionVars()->count(RdxPhi) && 2155 "Unable to find the reduction variable"); 2156 LoopVectorizationLegality::ReductionDescriptor RdxDesc = 2157 (*Legal->getReductionVars())[RdxPhi]; 2158 2159 setDebugLocFromInst(Builder, RdxDesc.StartValue); 2160 2161 // We need to generate a reduction vector from the incoming scalar. 2162 // To do so, we need to generate the 'identity' vector and overide 2163 // one of the elements with the incoming scalar reduction. We need 2164 // to do it in the vector-loop preheader. 2165 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator()); 2166 2167 // This is the vector-clone of the value that leaves the loop. 2168 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr); 2169 Type *VecTy = VectorExit[0]->getType(); 2170 2171 // Find the reduction identity variable. Zero for addition, or, xor, 2172 // one for multiplication, -1 for And. 2173 Value *Identity; 2174 Value *VectorStart; 2175 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax || 2176 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) { 2177 // MinMax reduction have the start value as their identify. 2178 if (VF == 1) { 2179 VectorStart = Identity = RdxDesc.StartValue; 2180 } else { 2181 VectorStart = Identity = Builder.CreateVectorSplat(VF, 2182 RdxDesc.StartValue, 2183 "minmax.ident"); 2184 } 2185 } else { 2186 // Handle other reduction kinds: 2187 Constant *Iden = 2188 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind, 2189 VecTy->getScalarType()); 2190 if (VF == 1) { 2191 Identity = Iden; 2192 // This vector is the Identity vector where the first element is the 2193 // incoming scalar reduction. 2194 VectorStart = RdxDesc.StartValue; 2195 } else { 2196 Identity = ConstantVector::getSplat(VF, Iden); 2197 2198 // This vector is the Identity vector where the first element is the 2199 // incoming scalar reduction. 2200 VectorStart = Builder.CreateInsertElement(Identity, 2201 RdxDesc.StartValue, Zero); 2202 } 2203 } 2204 2205 // Fix the vector-loop phi. 2206 // We created the induction variable so we know that the 2207 // preheader is the first entry. 2208 BasicBlock *VecPreheader = Induction->getIncomingBlock(0); 2209 2210 // Reductions do not have to start at zero. They can start with 2211 // any loop invariant values. 2212 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 2213 BasicBlock *Latch = OrigLoop->getLoopLatch(); 2214 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 2215 VectorParts &Val = getVectorValue(LoopVal); 2216 for (unsigned part = 0; part < UF; ++part) { 2217 // Make sure to add the reduction stat value only to the 2218 // first unroll part. 2219 Value *StartVal = (part == 0) ? VectorStart : Identity; 2220 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader); 2221 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody); 2222 } 2223 2224 // Before each round, move the insertion point right between 2225 // the PHIs and the values we are going to write. 2226 // This allows us to write both PHINodes and the extractelement 2227 // instructions. 2228 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); 2229 2230 VectorParts RdxParts; 2231 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr); 2232 for (unsigned part = 0; part < UF; ++part) { 2233 // This PHINode contains the vectorized reduction variable, or 2234 // the initial value vector, if we bypass the vector loop. 2235 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); 2236 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); 2237 Value *StartVal = (part == 0) ? VectorStart : Identity; 2238 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 2239 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]); 2240 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody); 2241 RdxParts.push_back(NewPhi); 2242 } 2243 2244 // Reduce all of the unrolled parts into a single vector. 2245 Value *ReducedPartRdx = RdxParts[0]; 2246 unsigned Op = getReductionBinOp(RdxDesc.Kind); 2247 setDebugLocFromInst(Builder, ReducedPartRdx); 2248 for (unsigned part = 1; part < UF; ++part) { 2249 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 2250 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op, 2251 RdxParts[part], ReducedPartRdx, 2252 "bin.rdx"); 2253 else 2254 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind, 2255 ReducedPartRdx, RdxParts[part]); 2256 } 2257 2258 if (VF > 1) { 2259 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 2260 // and vector ops, reducing the set of values being computed by half each 2261 // round. 2262 assert(isPowerOf2_32(VF) && 2263 "Reduction emission only supported for pow2 vectors!"); 2264 Value *TmpVec = ReducedPartRdx; 2265 SmallVector<Constant*, 32> ShuffleMask(VF, 0); 2266 for (unsigned i = VF; i != 1; i >>= 1) { 2267 // Move the upper half of the vector to the lower half. 2268 for (unsigned j = 0; j != i/2; ++j) 2269 ShuffleMask[j] = Builder.getInt32(i/2 + j); 2270 2271 // Fill the rest of the mask with undef. 2272 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 2273 UndefValue::get(Builder.getInt32Ty())); 2274 2275 Value *Shuf = 2276 Builder.CreateShuffleVector(TmpVec, 2277 UndefValue::get(TmpVec->getType()), 2278 ConstantVector::get(ShuffleMask), 2279 "rdx.shuf"); 2280 2281 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 2282 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf, 2283 "bin.rdx"); 2284 else 2285 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf); 2286 } 2287 2288 // The result is in the first element of the vector. 2289 ReducedPartRdx = Builder.CreateExtractElement(TmpVec, 2290 Builder.getInt32(0)); 2291 } 2292 2293 // Now, we need to fix the users of the reduction variable 2294 // inside and outside of the scalar remainder loop. 2295 // We know that the loop is in LCSSA form. We need to update the 2296 // PHI nodes in the exit blocks. 2297 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 2298 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 2299 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 2300 if (!LCSSAPhi) break; 2301 2302 // All PHINodes need to have a single entry edge, or two if 2303 // we already fixed them. 2304 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 2305 2306 // We found our reduction value exit-PHI. Update it with the 2307 // incoming bypass edge. 2308 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { 2309 // Add an edge coming from the bypass. 2310 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); 2311 break; 2312 } 2313 }// end of the LCSSA phi scan. 2314 2315 // Fix the scalar loop reduction variable with the incoming reduction sum 2316 // from the vector body and from the backedge value. 2317 int IncomingEdgeBlockIdx = 2318 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 2319 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 2320 // Pick the other block. 2321 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 2322 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx); 2323 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); 2324 }// end of for each redux variable. 2325 2326 fixLCSSAPHIs(); 2327 2328 // Remove redundant induction instructions. 2329 cse(LoopVectorBody); 2330} 2331 2332void InnerLoopVectorizer::fixLCSSAPHIs() { 2333 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 2334 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 2335 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 2336 if (!LCSSAPhi) break; 2337 if (LCSSAPhi->getNumIncomingValues() == 1) 2338 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 2339 LoopMiddleBlock); 2340 } 2341} 2342 2343InnerLoopVectorizer::VectorParts 2344InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 2345 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 2346 "Invalid edge"); 2347 2348 // Look for cached value. 2349 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst); 2350 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge); 2351 if (ECEntryIt != MaskCache.end()) 2352 return ECEntryIt->second; 2353 2354 VectorParts SrcMask = createBlockInMask(Src); 2355 2356 // The terminator has to be a branch inst! 2357 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 2358 assert(BI && "Unexpected terminator found"); 2359 2360 if (BI->isConditional()) { 2361 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 2362 2363 if (BI->getSuccessor(0) != Dst) 2364 for (unsigned part = 0; part < UF; ++part) 2365 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 2366 2367 for (unsigned part = 0; part < UF; ++part) 2368 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 2369 2370 MaskCache[Edge] = EdgeMask; 2371 return EdgeMask; 2372 } 2373 2374 MaskCache[Edge] = SrcMask; 2375 return SrcMask; 2376} 2377 2378InnerLoopVectorizer::VectorParts 2379InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 2380 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 2381 2382 // Loop incoming mask is all-one. 2383 if (OrigLoop->getHeader() == BB) { 2384 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 2385 return getVectorValue(C); 2386 } 2387 2388 // This is the block mask. We OR all incoming edges, and with zero. 2389 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 2390 VectorParts BlockMask = getVectorValue(Zero); 2391 2392 // For each pred: 2393 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 2394 VectorParts EM = createEdgeMask(*it, BB); 2395 for (unsigned part = 0; part < UF; ++part) 2396 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 2397 } 2398 2399 return BlockMask; 2400} 2401 2402void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, 2403 InnerLoopVectorizer::VectorParts &Entry, 2404 LoopVectorizationLegality *Legal, 2405 unsigned UF, unsigned VF, PhiVector *PV) { 2406 PHINode* P = cast<PHINode>(PN); 2407 // Handle reduction variables: 2408 if (Legal->getReductionVars()->count(P)) { 2409 for (unsigned part = 0; part < UF; ++part) { 2410 // This is phase one of vectorizing PHIs. 2411 Type *VecTy = (VF == 1) ? PN->getType() : 2412 VectorType::get(PN->getType(), VF); 2413 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", 2414 LoopVectorBody-> getFirstInsertionPt()); 2415 } 2416 PV->push_back(P); 2417 return; 2418 } 2419 2420 setDebugLocFromInst(Builder, P); 2421 // Check for PHI nodes that are lowered to vector selects. 2422 if (P->getParent() != OrigLoop->getHeader()) { 2423 // We know that all PHIs in non header blocks are converted into 2424 // selects, so we don't have to worry about the insertion order and we 2425 // can just use the builder. 2426 // At this point we generate the predication tree. There may be 2427 // duplications since this is a simple recursive scan, but future 2428 // optimizations will clean it up. 2429 2430 unsigned NumIncoming = P->getNumIncomingValues(); 2431 2432 // Generate a sequence of selects of the form: 2433 // SELECT(Mask3, In3, 2434 // SELECT(Mask2, In2, 2435 // ( ...))) 2436 for (unsigned In = 0; In < NumIncoming; In++) { 2437 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), 2438 P->getParent()); 2439 VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); 2440 2441 for (unsigned part = 0; part < UF; ++part) { 2442 // We might have single edge PHIs (blocks) - use an identity 2443 // 'select' for the first PHI operand. 2444 if (In == 0) 2445 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2446 In0[part]); 2447 else 2448 // Select between the current value and the previous incoming edge 2449 // based on the incoming mask. 2450 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2451 Entry[part], "predphi"); 2452 } 2453 } 2454 return; 2455 } 2456 2457 // This PHINode must be an induction variable. 2458 // Make sure that we know about it. 2459 assert(Legal->getInductionVars()->count(P) && 2460 "Not an induction variable"); 2461 2462 LoopVectorizationLegality::InductionInfo II = 2463 Legal->getInductionVars()->lookup(P); 2464 2465 switch (II.IK) { 2466 case LoopVectorizationLegality::IK_NoInduction: 2467 llvm_unreachable("Unknown induction"); 2468 case LoopVectorizationLegality::IK_IntInduction: { 2469 assert(P->getType() == II.StartValue->getType() && "Types must match"); 2470 Type *PhiTy = P->getType(); 2471 Value *Broadcasted; 2472 if (P == OldInduction) { 2473 // Handle the canonical induction variable. We might have had to 2474 // extend the type. 2475 Broadcasted = Builder.CreateTrunc(Induction, PhiTy); 2476 } else { 2477 // Handle other induction variables that are now based on the 2478 // canonical one. 2479 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx, 2480 "normalized.idx"); 2481 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy); 2482 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx, 2483 "offset.idx"); 2484 } 2485 Broadcasted = getBroadcastInstrs(Broadcasted); 2486 // After broadcasting the induction variable we need to make the vector 2487 // consecutive by adding 0, 1, 2, etc. 2488 for (unsigned part = 0; part < UF; ++part) 2489 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); 2490 return; 2491 } 2492 case LoopVectorizationLegality::IK_ReverseIntInduction: 2493 case LoopVectorizationLegality::IK_PtrInduction: 2494 case LoopVectorizationLegality::IK_ReversePtrInduction: 2495 // Handle reverse integer and pointer inductions. 2496 Value *StartIdx = ExtendedIdx; 2497 // This is the normalized GEP that starts counting at zero. 2498 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx, 2499 "normalized.idx"); 2500 2501 // Handle the reverse integer induction variable case. 2502 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) { 2503 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType()); 2504 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy, 2505 "resize.norm.idx"); 2506 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI, 2507 "reverse.idx"); 2508 2509 // This is a new value so do not hoist it out. 2510 Value *Broadcasted = getBroadcastInstrs(ReverseInd); 2511 // After broadcasting the induction variable we need to make the 2512 // vector consecutive by adding ... -3, -2, -1, 0. 2513 for (unsigned part = 0; part < UF; ++part) 2514 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part, 2515 true); 2516 return; 2517 } 2518 2519 // Handle the pointer induction variable case. 2520 assert(P->getType()->isPointerTy() && "Unexpected type."); 2521 2522 // Is this a reverse induction ptr or a consecutive induction ptr. 2523 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction == 2524 II.IK); 2525 2526 // This is the vector of results. Notice that we don't generate 2527 // vector geps because scalar geps result in better code. 2528 for (unsigned part = 0; part < UF; ++part) { 2529 if (VF == 1) { 2530 int EltIndex = (part) * (Reverse ? -1 : 1); 2531 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 2532 Value *GlobalIdx; 2533 if (Reverse) 2534 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); 2535 else 2536 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); 2537 2538 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 2539 "next.gep"); 2540 Entry[part] = SclrGep; 2541 continue; 2542 } 2543 2544 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 2545 for (unsigned int i = 0; i < VF; ++i) { 2546 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1); 2547 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 2548 Value *GlobalIdx; 2549 if (!Reverse) 2550 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); 2551 else 2552 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); 2553 2554 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 2555 "next.gep"); 2556 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 2557 Builder.getInt32(i), 2558 "insert.gep"); 2559 } 2560 Entry[part] = VecVal; 2561 } 2562 return; 2563 } 2564} 2565 2566void 2567InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, 2568 BasicBlock *BB, PhiVector *PV) { 2569 // For each instruction in the old loop. 2570 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 2571 VectorParts &Entry = WidenMap.get(it); 2572 switch (it->getOpcode()) { 2573 case Instruction::Br: 2574 // Nothing to do for PHIs and BR, since we already took care of the 2575 // loop control flow instructions. 2576 continue; 2577 case Instruction::PHI:{ 2578 // Vectorize PHINodes. 2579 widenPHIInstruction(it, Entry, Legal, UF, VF, PV); 2580 continue; 2581 }// End of PHI. 2582 2583 case Instruction::Add: 2584 case Instruction::FAdd: 2585 case Instruction::Sub: 2586 case Instruction::FSub: 2587 case Instruction::Mul: 2588 case Instruction::FMul: 2589 case Instruction::UDiv: 2590 case Instruction::SDiv: 2591 case Instruction::FDiv: 2592 case Instruction::URem: 2593 case Instruction::SRem: 2594 case Instruction::FRem: 2595 case Instruction::Shl: 2596 case Instruction::LShr: 2597 case Instruction::AShr: 2598 case Instruction::And: 2599 case Instruction::Or: 2600 case Instruction::Xor: { 2601 // Just widen binops. 2602 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 2603 setDebugLocFromInst(Builder, BinOp); 2604 VectorParts &A = getVectorValue(it->getOperand(0)); 2605 VectorParts &B = getVectorValue(it->getOperand(1)); 2606 2607 // Use this vector value for all users of the original instruction. 2608 for (unsigned Part = 0; Part < UF; ++Part) { 2609 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 2610 2611 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef. 2612 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V); 2613 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) { 2614 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); 2615 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); 2616 } 2617 if (VecOp && isa<PossiblyExactOperator>(VecOp)) 2618 VecOp->setIsExact(BinOp->isExact()); 2619 2620 Entry[Part] = V; 2621 } 2622 break; 2623 } 2624 case Instruction::Select: { 2625 // Widen selects. 2626 // If the selector is loop invariant we can create a select 2627 // instruction with a scalar condition. Otherwise, use vector-select. 2628 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), 2629 OrigLoop); 2630 setDebugLocFromInst(Builder, it); 2631 2632 // The condition can be loop invariant but still defined inside the 2633 // loop. This means that we can't just use the original 'cond' value. 2634 // We have to take the 'vectorized' value and pick the first lane. 2635 // Instcombine will make this a no-op. 2636 VectorParts &Cond = getVectorValue(it->getOperand(0)); 2637 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 2638 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 2639 2640 Value *ScalarCond = (VF == 1) ? Cond[0] : 2641 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0)); 2642 2643 for (unsigned Part = 0; Part < UF; ++Part) { 2644 Entry[Part] = Builder.CreateSelect( 2645 InvariantCond ? ScalarCond : Cond[Part], 2646 Op0[Part], 2647 Op1[Part]); 2648 } 2649 break; 2650 } 2651 2652 case Instruction::ICmp: 2653 case Instruction::FCmp: { 2654 // Widen compares. Generate vector compares. 2655 bool FCmp = (it->getOpcode() == Instruction::FCmp); 2656 CmpInst *Cmp = dyn_cast<CmpInst>(it); 2657 setDebugLocFromInst(Builder, it); 2658 VectorParts &A = getVectorValue(it->getOperand(0)); 2659 VectorParts &B = getVectorValue(it->getOperand(1)); 2660 for (unsigned Part = 0; Part < UF; ++Part) { 2661 Value *C = 0; 2662 if (FCmp) 2663 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 2664 else 2665 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 2666 Entry[Part] = C; 2667 } 2668 break; 2669 } 2670 2671 case Instruction::Store: 2672 case Instruction::Load: 2673 vectorizeMemoryInstruction(it, Legal); 2674 break; 2675 case Instruction::ZExt: 2676 case Instruction::SExt: 2677 case Instruction::FPToUI: 2678 case Instruction::FPToSI: 2679 case Instruction::FPExt: 2680 case Instruction::PtrToInt: 2681 case Instruction::IntToPtr: 2682 case Instruction::SIToFP: 2683 case Instruction::UIToFP: 2684 case Instruction::Trunc: 2685 case Instruction::FPTrunc: 2686 case Instruction::BitCast: { 2687 CastInst *CI = dyn_cast<CastInst>(it); 2688 setDebugLocFromInst(Builder, it); 2689 /// Optimize the special case where the source is the induction 2690 /// variable. Notice that we can only optimize the 'trunc' case 2691 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 2692 /// c. other casts depend on pointer size. 2693 if (CI->getOperand(0) == OldInduction && 2694 it->getOpcode() == Instruction::Trunc) { 2695 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 2696 CI->getType()); 2697 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 2698 for (unsigned Part = 0; Part < UF; ++Part) 2699 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false); 2700 break; 2701 } 2702 /// Vectorize casts. 2703 Type *DestTy = (VF == 1) ? CI->getType() : 2704 VectorType::get(CI->getType(), VF); 2705 2706 VectorParts &A = getVectorValue(it->getOperand(0)); 2707 for (unsigned Part = 0; Part < UF; ++Part) 2708 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 2709 break; 2710 } 2711 2712 case Instruction::Call: { 2713 // Ignore dbg intrinsics. 2714 if (isa<DbgInfoIntrinsic>(it)) 2715 break; 2716 setDebugLocFromInst(Builder, it); 2717 2718 Module *M = BB->getParent()->getParent(); 2719 CallInst *CI = cast<CallInst>(it); 2720 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 2721 assert(ID && "Not an intrinsic call!"); 2722 switch (ID) { 2723 case Intrinsic::lifetime_end: 2724 case Intrinsic::lifetime_start: 2725 scalarizeInstruction(it); 2726 break; 2727 default: 2728 for (unsigned Part = 0; Part < UF; ++Part) { 2729 SmallVector<Value *, 4> Args; 2730 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 2731 VectorParts &Arg = getVectorValue(CI->getArgOperand(i)); 2732 Args.push_back(Arg[Part]); 2733 } 2734 Type *Tys[] = {CI->getType()}; 2735 if (VF > 1) 2736 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF); 2737 2738 Function *F = Intrinsic::getDeclaration(M, ID, Tys); 2739 Entry[Part] = Builder.CreateCall(F, Args); 2740 } 2741 break; 2742 } 2743 break; 2744 } 2745 2746 default: 2747 // All other instructions are unsupported. Scalarize them. 2748 scalarizeInstruction(it); 2749 break; 2750 }// end of switch. 2751 }// end of for_each instr. 2752} 2753 2754void InnerLoopVectorizer::updateAnalysis() { 2755 // Forget the original basic block. 2756 SE->forgetLoop(OrigLoop); 2757 2758 // Update the dominator tree information. 2759 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && 2760 "Entry does not dominate exit."); 2761 2762 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2763 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); 2764 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); 2765 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); 2766 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front()); 2767 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); 2768 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 2769 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); 2770 2771 DEBUG(DT->verifyAnalysis()); 2772} 2773 2774bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 2775 if (!EnableIfConversion) 2776 return false; 2777 2778 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 2779 2780 // A list of pointers that we can safely read and write to. 2781 SmallPtrSet<Value *, 8> SafePointes; 2782 2783 // Collect safe addresses. 2784 for (Loop::block_iterator BI = TheLoop->block_begin(), 2785 BE = TheLoop->block_end(); BI != BE; ++BI) { 2786 BasicBlock *BB = *BI; 2787 2788 if (blockNeedsPredication(BB)) 2789 continue; 2790 2791 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 2792 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 2793 SafePointes.insert(LI->getPointerOperand()); 2794 else if (StoreInst *SI = dyn_cast<StoreInst>(I)) 2795 SafePointes.insert(SI->getPointerOperand()); 2796 } 2797 } 2798 2799 // Collect the blocks that need predication. 2800 for (Loop::block_iterator BI = TheLoop->block_begin(), 2801 BE = TheLoop->block_end(); BI != BE; ++BI) { 2802 BasicBlock *BB = *BI; 2803 2804 // We don't support switch statements inside loops. 2805 if (!isa<BranchInst>(BB->getTerminator())) 2806 return false; 2807 2808 // We must be able to predicate all blocks that need to be predicated. 2809 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes)) 2810 return false; 2811 } 2812 2813 // We can if-convert this loop. 2814 return true; 2815} 2816 2817bool LoopVectorizationLegality::canVectorize() { 2818 // We must have a loop in canonical form. Loops with indirectbr in them cannot 2819 // be canonicalized. 2820 if (!TheLoop->getLoopPreheader()) 2821 return false; 2822 2823 // We can only vectorize innermost loops. 2824 if (TheLoop->getSubLoopsVector().size()) 2825 return false; 2826 2827 // We must have a single backedge. 2828 if (TheLoop->getNumBackEdges() != 1) 2829 return false; 2830 2831 // We must have a single exiting block. 2832 if (!TheLoop->getExitingBlock()) 2833 return false; 2834 2835 // We need to have a loop header. 2836 DEBUG(dbgs() << "LV: Found a loop: " << 2837 TheLoop->getHeader()->getName() << '\n'); 2838 2839 // Check if we can if-convert non single-bb loops. 2840 unsigned NumBlocks = TheLoop->getNumBlocks(); 2841 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 2842 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 2843 return false; 2844 } 2845 2846 // ScalarEvolution needs to be able to find the exit count. 2847 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop); 2848 if (ExitCount == SE->getCouldNotCompute()) { 2849 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 2850 return false; 2851 } 2852 2853 // Do not loop-vectorize loops with a tiny trip count. 2854 BasicBlock *Latch = TheLoop->getLoopLatch(); 2855 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch); 2856 if (TC > 0u && TC < TinyTripCountVectorThreshold) { 2857 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << 2858 "This loop is not worth vectorizing.\n"); 2859 return false; 2860 } 2861 2862 // Check if we can vectorize the instructions and CFG in this loop. 2863 if (!canVectorizeInstrs()) { 2864 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 2865 return false; 2866 } 2867 2868 // Go over each instruction and look at memory deps. 2869 if (!canVectorizeMemory()) { 2870 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 2871 return false; 2872 } 2873 2874 // Collect all of the variables that remain uniform after vectorization. 2875 collectLoopUniforms(); 2876 2877 DEBUG(dbgs() << "LV: We can vectorize this loop" << 2878 (PtrRtCheck.Need ? " (with a runtime bound check)" : "") 2879 <<"!\n"); 2880 2881 // Okay! We can vectorize. At this point we don't have any other mem analysis 2882 // which may limit our maximum vectorization factor, so just return true with 2883 // no restrictions. 2884 return true; 2885} 2886 2887static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) { 2888 if (Ty->isPointerTy()) 2889 return DL.getIntPtrType(Ty); 2890 2891 return Ty; 2892} 2893 2894static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) { 2895 Ty0 = convertPointerToIntegerType(DL, Ty0); 2896 Ty1 = convertPointerToIntegerType(DL, Ty1); 2897 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits()) 2898 return Ty0; 2899 return Ty1; 2900} 2901 2902/// \brief Check that the instruction has outside loop users and is not an 2903/// identified reduction variable. 2904static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst, 2905 SmallPtrSet<Value *, 4> &Reductions) { 2906 // Reduction instructions are allowed to have exit users. All other 2907 // instructions must not have external users. 2908 if (!Reductions.count(Inst)) 2909 //Check that all of the users of the loop are inside the BB. 2910 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end(); 2911 I != E; ++I) { 2912 Instruction *U = cast<Instruction>(*I); 2913 // This user may be a reduction exit value. 2914 if (!TheLoop->contains(U)) { 2915 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n'); 2916 return true; 2917 } 2918 } 2919 return false; 2920} 2921 2922bool LoopVectorizationLegality::canVectorizeInstrs() { 2923 BasicBlock *PreHeader = TheLoop->getLoopPreheader(); 2924 BasicBlock *Header = TheLoop->getHeader(); 2925 2926 // Look for the attribute signaling the absence of NaNs. 2927 Function &F = *Header->getParent(); 2928 if (F.hasFnAttribute("no-nans-fp-math")) 2929 HasFunNoNaNAttr = F.getAttributes().getAttribute( 2930 AttributeSet::FunctionIndex, 2931 "no-nans-fp-math").getValueAsString() == "true"; 2932 2933 // For each block in the loop. 2934 for (Loop::block_iterator bb = TheLoop->block_begin(), 2935 be = TheLoop->block_end(); bb != be; ++bb) { 2936 2937 // Scan the instructions in the block and look for hazards. 2938 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2939 ++it) { 2940 2941 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 2942 Type *PhiTy = Phi->getType(); 2943 // Check that this PHI type is allowed. 2944 if (!PhiTy->isIntegerTy() && 2945 !PhiTy->isFloatingPointTy() && 2946 !PhiTy->isPointerTy()) { 2947 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 2948 return false; 2949 } 2950 2951 // If this PHINode is not in the header block, then we know that we 2952 // can convert it to select during if-conversion. No need to check if 2953 // the PHIs in this block are induction or reduction variables. 2954 if (*bb != Header) { 2955 // Check that this instruction has no outside users or is an 2956 // identified reduction value with an outside user. 2957 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit)) 2958 continue; 2959 return false; 2960 } 2961 2962 // We only allow if-converted PHIs with more than two incoming values. 2963 if (Phi->getNumIncomingValues() != 2) { 2964 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 2965 return false; 2966 } 2967 2968 // This is the value coming from the preheader. 2969 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); 2970 // Check if this is an induction variable. 2971 InductionKind IK = isInductionVariable(Phi); 2972 2973 if (IK_NoInduction != IK) { 2974 // Get the widest type. 2975 if (!WidestIndTy) 2976 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy); 2977 else 2978 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy); 2979 2980 // Int inductions are special because we only allow one IV. 2981 if (IK == IK_IntInduction) { 2982 // Use the phi node with the widest type as induction. Use the last 2983 // one if there are multiple (no good reason for doing this other 2984 // than it is expedient). 2985 if (!Induction || PhiTy == WidestIndTy) 2986 Induction = Phi; 2987 } 2988 2989 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 2990 Inductions[Phi] = InductionInfo(StartValue, IK); 2991 2992 // Until we explicitly handle the case of an induction variable with 2993 // an outside loop user we have to give up vectorizing this loop. 2994 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) 2995 return false; 2996 2997 continue; 2998 } 2999 3000 if (AddReductionVar(Phi, RK_IntegerAdd)) { 3001 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); 3002 continue; 3003 } 3004 if (AddReductionVar(Phi, RK_IntegerMult)) { 3005 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); 3006 continue; 3007 } 3008 if (AddReductionVar(Phi, RK_IntegerOr)) { 3009 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); 3010 continue; 3011 } 3012 if (AddReductionVar(Phi, RK_IntegerAnd)) { 3013 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); 3014 continue; 3015 } 3016 if (AddReductionVar(Phi, RK_IntegerXor)) { 3017 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); 3018 continue; 3019 } 3020 if (AddReductionVar(Phi, RK_IntegerMinMax)) { 3021 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n"); 3022 continue; 3023 } 3024 if (AddReductionVar(Phi, RK_FloatMult)) { 3025 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n"); 3026 continue; 3027 } 3028 if (AddReductionVar(Phi, RK_FloatAdd)) { 3029 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n"); 3030 continue; 3031 } 3032 if (AddReductionVar(Phi, RK_FloatMinMax)) { 3033 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi << 3034 "\n"); 3035 continue; 3036 } 3037 3038 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 3039 return false; 3040 }// end of PHI handling 3041 3042 // We still don't handle functions. However, we can ignore dbg intrinsic 3043 // calls and we do handle certain intrinsic and libm functions. 3044 CallInst *CI = dyn_cast<CallInst>(it); 3045 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) { 3046 DEBUG(dbgs() << "LV: Found a call site.\n"); 3047 return false; 3048 } 3049 3050 // Check that the instruction return type is vectorizable. 3051 // Also, we can't vectorize extractelement instructions. 3052 if ((!VectorType::isValidElementType(it->getType()) && 3053 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) { 3054 DEBUG(dbgs() << "LV: Found unvectorizable type.\n"); 3055 return false; 3056 } 3057 3058 // Check that the stored type is vectorizable. 3059 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 3060 Type *T = ST->getValueOperand()->getType(); 3061 if (!VectorType::isValidElementType(T)) 3062 return false; 3063 } 3064 3065 // Reduction instructions are allowed to have exit users. 3066 // All other instructions must not have external users. 3067 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) 3068 return false; 3069 3070 } // next instr. 3071 3072 } 3073 3074 if (!Induction) { 3075 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 3076 if (Inductions.empty()) 3077 return false; 3078 } 3079 3080 return true; 3081} 3082 3083void LoopVectorizationLegality::collectLoopUniforms() { 3084 // We now know that the loop is vectorizable! 3085 // Collect variables that will remain uniform after vectorization. 3086 std::vector<Value*> Worklist; 3087 BasicBlock *Latch = TheLoop->getLoopLatch(); 3088 3089 // Start with the conditional branch and walk up the block. 3090 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 3091 3092 while (Worklist.size()) { 3093 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 3094 Worklist.pop_back(); 3095 3096 // Look at instructions inside this loop. 3097 // Stop when reaching PHI nodes. 3098 // TODO: we need to follow values all over the loop, not only in this block. 3099 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 3100 continue; 3101 3102 // This is a known uniform. 3103 Uniforms.insert(I); 3104 3105 // Insert all operands. 3106 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end()); 3107 } 3108} 3109 3110namespace { 3111/// \brief Analyses memory accesses in a loop. 3112/// 3113/// Checks whether run time pointer checks are needed and builds sets for data 3114/// dependence checking. 3115class AccessAnalysis { 3116public: 3117 /// \brief Read or write access location. 3118 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 3119 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet; 3120 3121 /// \brief Set of potential dependent memory accesses. 3122 typedef EquivalenceClasses<MemAccessInfo> DepCandidates; 3123 3124 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) : 3125 DL(Dl), DepCands(DA), AreAllWritesIdentified(true), 3126 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {} 3127 3128 /// \brief Register a load and whether it is only read from. 3129 void addLoad(Value *Ptr, bool IsReadOnly) { 3130 Accesses.insert(MemAccessInfo(Ptr, false)); 3131 if (IsReadOnly) 3132 ReadOnlyPtr.insert(Ptr); 3133 } 3134 3135 /// \brief Register a store. 3136 void addStore(Value *Ptr) { 3137 Accesses.insert(MemAccessInfo(Ptr, true)); 3138 } 3139 3140 /// \brief Check whether we can check the pointers at runtime for 3141 /// non-intersection. 3142 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck, 3143 unsigned &NumComparisons, ScalarEvolution *SE, 3144 Loop *TheLoop, bool ShouldCheckStride = false); 3145 3146 /// \brief Goes over all memory accesses, checks whether a RT check is needed 3147 /// and builds sets of dependent accesses. 3148 void buildDependenceSets() { 3149 // Process read-write pointers first. 3150 processMemAccesses(false); 3151 // Next, process read pointers. 3152 processMemAccesses(true); 3153 } 3154 3155 bool isRTCheckNeeded() { return IsRTCheckNeeded; } 3156 3157 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); } 3158 void resetDepChecks() { CheckDeps.clear(); } 3159 3160 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; } 3161 3162private: 3163 typedef SetVector<MemAccessInfo> PtrAccessSet; 3164 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap; 3165 3166 /// \brief Go over all memory access or only the deferred ones if 3167 /// \p UseDeferred is true and check whether runtime pointer checks are needed 3168 /// and build sets of dependency check candidates. 3169 void processMemAccesses(bool UseDeferred); 3170 3171 /// Set of all accesses. 3172 PtrAccessSet Accesses; 3173 3174 /// Set of access to check after all writes have been processed. 3175 PtrAccessSet DeferredAccesses; 3176 3177 /// Map of pointers to last access encountered. 3178 UnderlyingObjToAccessMap ObjToLastAccess; 3179 3180 /// Set of accesses that need a further dependence check. 3181 MemAccessInfoSet CheckDeps; 3182 3183 /// Set of pointers that are read only. 3184 SmallPtrSet<Value*, 16> ReadOnlyPtr; 3185 3186 /// Set of underlying objects already written to. 3187 SmallPtrSet<Value*, 16> WriteObjects; 3188 3189 DataLayout *DL; 3190 3191 /// Sets of potentially dependent accesses - members of one set share an 3192 /// underlying pointer. The set "CheckDeps" identfies which sets really need a 3193 /// dependence check. 3194 DepCandidates &DepCands; 3195 3196 bool AreAllWritesIdentified; 3197 bool AreAllReadsIdentified; 3198 bool IsRTCheckNeeded; 3199}; 3200 3201} // end anonymous namespace 3202 3203/// \brief Check whether a pointer can participate in a runtime bounds check. 3204static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) { 3205 const SCEV *PtrScev = SE->getSCEV(Ptr); 3206 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 3207 if (!AR) 3208 return false; 3209 3210 return AR->isAffine(); 3211} 3212 3213/// \brief Check the stride of the pointer and ensure that it does not wrap in 3214/// the address space. 3215static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr, 3216 const Loop *Lp); 3217 3218bool AccessAnalysis::canCheckPtrAtRT( 3219 LoopVectorizationLegality::RuntimePointerCheck &RtCheck, 3220 unsigned &NumComparisons, ScalarEvolution *SE, 3221 Loop *TheLoop, bool ShouldCheckStride) { 3222 // Find pointers with computable bounds. We are going to use this information 3223 // to place a runtime bound check. 3224 unsigned NumReadPtrChecks = 0; 3225 unsigned NumWritePtrChecks = 0; 3226 bool CanDoRT = true; 3227 3228 bool IsDepCheckNeeded = isDependencyCheckNeeded(); 3229 // We assign consecutive id to access from different dependence sets. 3230 // Accesses within the same set don't need a runtime check. 3231 unsigned RunningDepId = 1; 3232 DenseMap<Value *, unsigned> DepSetId; 3233 3234 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end(); 3235 AI != AE; ++AI) { 3236 const MemAccessInfo &Access = *AI; 3237 Value *Ptr = Access.getPointer(); 3238 bool IsWrite = Access.getInt(); 3239 3240 // Just add write checks if we have both. 3241 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true))) 3242 continue; 3243 3244 if (IsWrite) 3245 ++NumWritePtrChecks; 3246 else 3247 ++NumReadPtrChecks; 3248 3249 if (hasComputableBounds(SE, Ptr) && 3250 // When we run after a failing dependency check we have to make sure we 3251 // don't have wrapping pointers. 3252 (!ShouldCheckStride || isStridedPtr(SE, DL, Ptr, TheLoop) == 1)) { 3253 // The id of the dependence set. 3254 unsigned DepId; 3255 3256 if (IsDepCheckNeeded) { 3257 Value *Leader = DepCands.getLeaderValue(Access).getPointer(); 3258 unsigned &LeaderId = DepSetId[Leader]; 3259 if (!LeaderId) 3260 LeaderId = RunningDepId++; 3261 DepId = LeaderId; 3262 } else 3263 // Each access has its own dependence set. 3264 DepId = RunningDepId++; 3265 3266 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId); 3267 3268 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n'); 3269 } else { 3270 CanDoRT = false; 3271 } 3272 } 3273 3274 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2) 3275 NumComparisons = 0; // Only one dependence set. 3276 else { 3277 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks + 3278 NumWritePtrChecks - 1)); 3279 } 3280 3281 // If the pointers that we would use for the bounds comparison have different 3282 // address spaces, assume the values aren't directly comparable, so we can't 3283 // use them for the runtime check. We also have to assume they could 3284 // overlap. In the future there should be metadata for whether address spaces 3285 // are disjoint. 3286 unsigned NumPointers = RtCheck.Pointers.size(); 3287 for (unsigned i = 0; i < NumPointers; ++i) { 3288 for (unsigned j = i + 1; j < NumPointers; ++j) { 3289 // Only need to check pointers between two different dependency sets. 3290 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j]) 3291 continue; 3292 3293 Value *PtrI = RtCheck.Pointers[i]; 3294 Value *PtrJ = RtCheck.Pointers[j]; 3295 3296 unsigned ASi = PtrI->getType()->getPointerAddressSpace(); 3297 unsigned ASj = PtrJ->getType()->getPointerAddressSpace(); 3298 if (ASi != ASj) { 3299 DEBUG(dbgs() << "LV: Runtime check would require comparison between" 3300 " different address spaces\n"); 3301 return false; 3302 } 3303 } 3304 } 3305 3306 return CanDoRT; 3307} 3308 3309static bool isFunctionScopeIdentifiedObject(Value *Ptr) { 3310 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr); 3311} 3312 3313void AccessAnalysis::processMemAccesses(bool UseDeferred) { 3314 // We process the set twice: first we process read-write pointers, last we 3315 // process read-only pointers. This allows us to skip dependence tests for 3316 // read-only pointers. 3317 3318 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses; 3319 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) { 3320 const MemAccessInfo &Access = *AI; 3321 Value *Ptr = Access.getPointer(); 3322 bool IsWrite = Access.getInt(); 3323 3324 DepCands.insert(Access); 3325 3326 // Memorize read-only pointers for later processing and skip them in the 3327 // first round (they need to be checked after we have seen all write 3328 // pointers). Note: we also mark pointer that are not consecutive as 3329 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the 3330 // second check for "!IsWrite". 3331 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite; 3332 if (!UseDeferred && IsReadOnlyPtr) { 3333 DeferredAccesses.insert(Access); 3334 continue; 3335 } 3336 3337 bool NeedDepCheck = false; 3338 // Check whether there is the possiblity of dependency because of underlying 3339 // objects being the same. 3340 typedef SmallVector<Value*, 16> ValueVector; 3341 ValueVector TempObjects; 3342 GetUnderlyingObjects(Ptr, TempObjects, DL); 3343 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end(); 3344 UI != UE; ++UI) { 3345 Value *UnderlyingObj = *UI; 3346 3347 // If this is a write then it needs to be an identified object. If this a 3348 // read and all writes (so far) are identified function scope objects we 3349 // don't need an identified underlying object but only an Argument (the 3350 // next write is going to invalidate this assumption if it is 3351 // unidentified). 3352 // This is a micro-optimization for the case where all writes are 3353 // identified and we have one argument pointer. 3354 // Otherwise, we do need a runtime check. 3355 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) || 3356 (!IsWrite && (!AreAllWritesIdentified || 3357 !isa<Argument>(UnderlyingObj)) && 3358 !isIdentifiedObject(UnderlyingObj))) { 3359 DEBUG(dbgs() << "LV: Found an unidentified " << 3360 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj << 3361 "\n"); 3362 IsRTCheckNeeded = (IsRTCheckNeeded || 3363 !isIdentifiedObject(UnderlyingObj) || 3364 !AreAllReadsIdentified); 3365 3366 if (IsWrite) 3367 AreAllWritesIdentified = false; 3368 if (!IsWrite) 3369 AreAllReadsIdentified = false; 3370 } 3371 3372 // If this is a write - check other reads and writes for conflicts. If 3373 // this is a read only check other writes for conflicts (but only if there 3374 // is no other write to the ptr - this is an optimization to catch "a[i] = 3375 // a[i] + " without having to do a dependence check). 3376 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj)) 3377 NeedDepCheck = true; 3378 3379 if (IsWrite) 3380 WriteObjects.insert(UnderlyingObj); 3381 3382 // Create sets of pointers connected by shared underlying objects. 3383 UnderlyingObjToAccessMap::iterator Prev = 3384 ObjToLastAccess.find(UnderlyingObj); 3385 if (Prev != ObjToLastAccess.end()) 3386 DepCands.unionSets(Access, Prev->second); 3387 3388 ObjToLastAccess[UnderlyingObj] = Access; 3389 } 3390 3391 if (NeedDepCheck) 3392 CheckDeps.insert(Access); 3393 } 3394} 3395 3396namespace { 3397/// \brief Checks memory dependences among accesses to the same underlying 3398/// object to determine whether there vectorization is legal or not (and at 3399/// which vectorization factor). 3400/// 3401/// This class works under the assumption that we already checked that memory 3402/// locations with different underlying pointers are "must-not alias". 3403/// We use the ScalarEvolution framework to symbolically evalutate access 3404/// functions pairs. Since we currently don't restructure the loop we can rely 3405/// on the program order of memory accesses to determine their safety. 3406/// At the moment we will only deem accesses as safe for: 3407/// * A negative constant distance assuming program order. 3408/// 3409/// Safe: tmp = a[i + 1]; OR a[i + 1] = x; 3410/// a[i] = tmp; y = a[i]; 3411/// 3412/// The latter case is safe because later checks guarantuee that there can't 3413/// be a cycle through a phi node (that is, we check that "x" and "y" is not 3414/// the same variable: a header phi can only be an induction or a reduction, a 3415/// reduction can't have a memory sink, an induction can't have a memory 3416/// source). This is important and must not be violated (or we have to 3417/// resort to checking for cycles through memory). 3418/// 3419/// * A positive constant distance assuming program order that is bigger 3420/// than the biggest memory access. 3421/// 3422/// tmp = a[i] OR b[i] = x 3423/// a[i+2] = tmp y = b[i+2]; 3424/// 3425/// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively. 3426/// 3427/// * Zero distances and all accesses have the same size. 3428/// 3429class MemoryDepChecker { 3430public: 3431 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 3432 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet; 3433 3434 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) 3435 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0), 3436 ShouldRetryWithRuntimeCheck(false) {} 3437 3438 /// \brief Register the location (instructions are given increasing numbers) 3439 /// of a write access. 3440 void addAccess(StoreInst *SI) { 3441 Value *Ptr = SI->getPointerOperand(); 3442 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx); 3443 InstMap.push_back(SI); 3444 ++AccessIdx; 3445 } 3446 3447 /// \brief Register the location (instructions are given increasing numbers) 3448 /// of a write access. 3449 void addAccess(LoadInst *LI) { 3450 Value *Ptr = LI->getPointerOperand(); 3451 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx); 3452 InstMap.push_back(LI); 3453 ++AccessIdx; 3454 } 3455 3456 /// \brief Check whether the dependencies between the accesses are safe. 3457 /// 3458 /// Only checks sets with elements in \p CheckDeps. 3459 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets, 3460 MemAccessInfoSet &CheckDeps); 3461 3462 /// \brief The maximum number of bytes of a vector register we can vectorize 3463 /// the accesses safely with. 3464 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; } 3465 3466 /// \brief In same cases when the dependency check fails we can still 3467 /// vectorize the loop with a dynamic array access check. 3468 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; } 3469 3470private: 3471 ScalarEvolution *SE; 3472 DataLayout *DL; 3473 const Loop *InnermostLoop; 3474 3475 /// \brief Maps access locations (ptr, read/write) to program order. 3476 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses; 3477 3478 /// \brief Memory access instructions in program order. 3479 SmallVector<Instruction *, 16> InstMap; 3480 3481 /// \brief The program order index to be used for the next instruction. 3482 unsigned AccessIdx; 3483 3484 // We can access this many bytes in parallel safely. 3485 unsigned MaxSafeDepDistBytes; 3486 3487 /// \brief If we see a non constant dependence distance we can still try to 3488 /// vectorize this loop with runtime checks. 3489 bool ShouldRetryWithRuntimeCheck; 3490 3491 /// \brief Check whether there is a plausible dependence between the two 3492 /// accesses. 3493 /// 3494 /// Access \p A must happen before \p B in program order. The two indices 3495 /// identify the index into the program order map. 3496 /// 3497 /// This function checks whether there is a plausible dependence (or the 3498 /// absence of such can't be proved) between the two accesses. If there is a 3499 /// plausible dependence but the dependence distance is bigger than one 3500 /// element access it records this distance in \p MaxSafeDepDistBytes (if this 3501 /// distance is smaller than any other distance encountered so far). 3502 /// Otherwise, this function returns true signaling a possible dependence. 3503 bool isDependent(const MemAccessInfo &A, unsigned AIdx, 3504 const MemAccessInfo &B, unsigned BIdx); 3505 3506 /// \brief Check whether the data dependence could prevent store-load 3507 /// forwarding. 3508 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize); 3509}; 3510 3511} // end anonymous namespace 3512 3513static bool isInBoundsGep(Value *Ptr) { 3514 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) 3515 return GEP->isInBounds(); 3516 return false; 3517} 3518 3519/// \brief Check whether the access through \p Ptr has a constant stride. 3520static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr, 3521 const Loop *Lp) { 3522 const Type *Ty = Ptr->getType(); 3523 assert(Ty->isPointerTy() && "Unexpected non ptr"); 3524 3525 // Make sure that the pointer does not point to aggregate types. 3526 const PointerType *PtrTy = cast<PointerType>(Ty); 3527 if (PtrTy->getElementType()->isAggregateType()) { 3528 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr << 3529 "\n"); 3530 return 0; 3531 } 3532 3533 const SCEV *PtrScev = SE->getSCEV(Ptr); 3534 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 3535 if (!AR) { 3536 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer " 3537 << *Ptr << " SCEV: " << *PtrScev << "\n"); 3538 return 0; 3539 } 3540 3541 // The accesss function must stride over the innermost loop. 3542 if (Lp != AR->getLoop()) { 3543 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " << 3544 *Ptr << " SCEV: " << *PtrScev << "\n"); 3545 } 3546 3547 // The address calculation must not wrap. Otherwise, a dependence could be 3548 // inverted. 3549 // An inbounds getelementptr that is a AddRec with a unit stride 3550 // cannot wrap per definition. The unit stride requirement is checked later. 3551 // An getelementptr without an inbounds attribute and unit stride would have 3552 // to access the pointer value "0" which is undefined behavior in address 3553 // space 0, therefore we can also vectorize this case. 3554 bool IsInBoundsGEP = isInBoundsGep(Ptr); 3555 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask); 3556 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0; 3557 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) { 3558 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space " 3559 << *Ptr << " SCEV: " << *PtrScev << "\n"); 3560 return 0; 3561 } 3562 3563 // Check the step is constant. 3564 const SCEV *Step = AR->getStepRecurrence(*SE); 3565 3566 // Calculate the pointer stride and check if it is consecutive. 3567 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 3568 if (!C) { 3569 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr << 3570 " SCEV: " << *PtrScev << "\n"); 3571 return 0; 3572 } 3573 3574 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType()); 3575 const APInt &APStepVal = C->getValue()->getValue(); 3576 3577 // Huge step value - give up. 3578 if (APStepVal.getBitWidth() > 64) 3579 return 0; 3580 3581 int64_t StepVal = APStepVal.getSExtValue(); 3582 3583 // Strided access. 3584 int64_t Stride = StepVal / Size; 3585 int64_t Rem = StepVal % Size; 3586 if (Rem) 3587 return 0; 3588 3589 // If the SCEV could wrap but we have an inbounds gep with a unit stride we 3590 // know we can't "wrap around the address space". In case of address space 3591 // zero we know that this won't happen without triggering undefined behavior. 3592 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) && 3593 Stride != 1 && Stride != -1) 3594 return 0; 3595 3596 return Stride; 3597} 3598 3599bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance, 3600 unsigned TypeByteSize) { 3601 // If loads occur at a distance that is not a multiple of a feasible vector 3602 // factor store-load forwarding does not take place. 3603 // Positive dependences might cause troubles because vectorizing them might 3604 // prevent store-load forwarding making vectorized code run a lot slower. 3605 // a[i] = a[i-3] ^ a[i-8]; 3606 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and 3607 // hence on your typical architecture store-load forwarding does not take 3608 // place. Vectorizing in such cases does not make sense. 3609 // Store-load forwarding distance. 3610 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize; 3611 // Maximum vector factor. 3612 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize; 3613 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues) 3614 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes; 3615 3616 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues; 3617 vf *= 2) { 3618 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) { 3619 MaxVFWithoutSLForwardIssues = (vf >>=1); 3620 break; 3621 } 3622 } 3623 3624 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) { 3625 DEBUG(dbgs() << "LV: Distance " << Distance << 3626 " that could cause a store-load forwarding conflict\n"); 3627 return true; 3628 } 3629 3630 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes && 3631 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize) 3632 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues; 3633 return false; 3634} 3635 3636bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx, 3637 const MemAccessInfo &B, unsigned BIdx) { 3638 assert (AIdx < BIdx && "Must pass arguments in program order"); 3639 3640 Value *APtr = A.getPointer(); 3641 Value *BPtr = B.getPointer(); 3642 bool AIsWrite = A.getInt(); 3643 bool BIsWrite = B.getInt(); 3644 3645 // Two reads are independent. 3646 if (!AIsWrite && !BIsWrite) 3647 return false; 3648 3649 const SCEV *AScev = SE->getSCEV(APtr); 3650 const SCEV *BScev = SE->getSCEV(BPtr); 3651 3652 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop); 3653 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop); 3654 3655 const SCEV *Src = AScev; 3656 const SCEV *Sink = BScev; 3657 3658 // If the induction step is negative we have to invert source and sink of the 3659 // dependence. 3660 if (StrideAPtr < 0) { 3661 //Src = BScev; 3662 //Sink = AScev; 3663 std::swap(APtr, BPtr); 3664 std::swap(Src, Sink); 3665 std::swap(AIsWrite, BIsWrite); 3666 std::swap(AIdx, BIdx); 3667 std::swap(StrideAPtr, StrideBPtr); 3668 } 3669 3670 const SCEV *Dist = SE->getMinusSCEV(Sink, Src); 3671 3672 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink 3673 << "(Induction step: " << StrideAPtr << ")\n"); 3674 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to " 3675 << *InstMap[BIdx] << ": " << *Dist << "\n"); 3676 3677 // Need consecutive accesses. We don't want to vectorize 3678 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in 3679 // the address space. 3680 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){ 3681 DEBUG(dbgs() << "Non-consecutive pointer access\n"); 3682 return true; 3683 } 3684 3685 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist); 3686 if (!C) { 3687 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n"); 3688 ShouldRetryWithRuntimeCheck = true; 3689 return true; 3690 } 3691 3692 Type *ATy = APtr->getType()->getPointerElementType(); 3693 Type *BTy = BPtr->getType()->getPointerElementType(); 3694 unsigned TypeByteSize = DL->getTypeAllocSize(ATy); 3695 3696 // Negative distances are not plausible dependencies. 3697 const APInt &Val = C->getValue()->getValue(); 3698 if (Val.isNegative()) { 3699 bool IsTrueDataDependence = (AIsWrite && !BIsWrite); 3700 if (IsTrueDataDependence && 3701 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) || 3702 ATy != BTy)) 3703 return true; 3704 3705 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n"); 3706 return false; 3707 } 3708 3709 // Write to the same location with the same size. 3710 // Could be improved to assert type sizes are the same (i32 == float, etc). 3711 if (Val == 0) { 3712 if (ATy == BTy) 3713 return false; 3714 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n"); 3715 return true; 3716 } 3717 3718 assert(Val.isStrictlyPositive() && "Expect a positive value"); 3719 3720 // Positive distance bigger than max vectorization factor. 3721 if (ATy != BTy) { 3722 DEBUG(dbgs() << 3723 "LV: ReadWrite-Write positive dependency with different types\n"); 3724 return false; 3725 } 3726 3727 unsigned Distance = (unsigned) Val.getZExtValue(); 3728 3729 // Bail out early if passed-in parameters make vectorization not feasible. 3730 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1; 3731 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1; 3732 3733 // The distance must be bigger than the size needed for a vectorized version 3734 // of the operation and the size of the vectorized operation must not be 3735 // bigger than the currrent maximum size. 3736 if (Distance < 2*TypeByteSize || 3737 2*TypeByteSize > MaxSafeDepDistBytes || 3738 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) { 3739 DEBUG(dbgs() << "LV: Failure because of Positive distance " 3740 << Val.getSExtValue() << '\n'); 3741 return true; 3742 } 3743 3744 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ? 3745 Distance : MaxSafeDepDistBytes; 3746 3747 bool IsTrueDataDependence = (!AIsWrite && BIsWrite); 3748 if (IsTrueDataDependence && 3749 couldPreventStoreLoadForward(Distance, TypeByteSize)) 3750 return true; 3751 3752 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() << 3753 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n'); 3754 3755 return false; 3756} 3757 3758bool 3759MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets, 3760 MemAccessInfoSet &CheckDeps) { 3761 3762 MaxSafeDepDistBytes = -1U; 3763 while (!CheckDeps.empty()) { 3764 MemAccessInfo CurAccess = *CheckDeps.begin(); 3765 3766 // Get the relevant memory access set. 3767 EquivalenceClasses<MemAccessInfo>::iterator I = 3768 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess)); 3769 3770 // Check accesses within this set. 3771 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE; 3772 AI = AccessSets.member_begin(I), AE = AccessSets.member_end(); 3773 3774 // Check every access pair. 3775 while (AI != AE) { 3776 CheckDeps.erase(*AI); 3777 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI); 3778 while (OI != AE) { 3779 // Check every accessing instruction pair in program order. 3780 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(), 3781 I1E = Accesses[*AI].end(); I1 != I1E; ++I1) 3782 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(), 3783 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) { 3784 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2)) 3785 return false; 3786 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1)) 3787 return false; 3788 } 3789 ++OI; 3790 } 3791 AI++; 3792 } 3793 } 3794 return true; 3795} 3796 3797bool LoopVectorizationLegality::canVectorizeMemory() { 3798 3799 typedef SmallVector<Value*, 16> ValueVector; 3800 typedef SmallPtrSet<Value*, 16> ValueSet; 3801 3802 // Holds the Load and Store *instructions*. 3803 ValueVector Loads; 3804 ValueVector Stores; 3805 3806 // Holds all the different accesses in the loop. 3807 unsigned NumReads = 0; 3808 unsigned NumReadWrites = 0; 3809 3810 PtrRtCheck.Pointers.clear(); 3811 PtrRtCheck.Need = false; 3812 3813 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); 3814 MemoryDepChecker DepChecker(SE, DL, TheLoop); 3815 3816 // For each block. 3817 for (Loop::block_iterator bb = TheLoop->block_begin(), 3818 be = TheLoop->block_end(); bb != be; ++bb) { 3819 3820 // Scan the BB and collect legal loads and stores. 3821 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 3822 ++it) { 3823 3824 // If this is a load, save it. If this instruction can read from memory 3825 // but is not a load, then we quit. Notice that we don't handle function 3826 // calls that read or write. 3827 if (it->mayReadFromMemory()) { 3828 // Many math library functions read the rounding mode. We will only 3829 // vectorize a loop if it contains known function calls that don't set 3830 // the flag. Therefore, it is safe to ignore this read from memory. 3831 CallInst *Call = dyn_cast<CallInst>(it); 3832 if (Call && getIntrinsicIDForCall(Call, TLI)) 3833 continue; 3834 3835 LoadInst *Ld = dyn_cast<LoadInst>(it); 3836 if (!Ld) return false; 3837 if (!Ld->isSimple() && !IsAnnotatedParallel) { 3838 DEBUG(dbgs() << "LV: Found a non-simple load.\n"); 3839 return false; 3840 } 3841 Loads.push_back(Ld); 3842 DepChecker.addAccess(Ld); 3843 continue; 3844 } 3845 3846 // Save 'store' instructions. Abort if other instructions write to memory. 3847 if (it->mayWriteToMemory()) { 3848 StoreInst *St = dyn_cast<StoreInst>(it); 3849 if (!St) return false; 3850 if (!St->isSimple() && !IsAnnotatedParallel) { 3851 DEBUG(dbgs() << "LV: Found a non-simple store.\n"); 3852 return false; 3853 } 3854 Stores.push_back(St); 3855 DepChecker.addAccess(St); 3856 } 3857 } // Next instr. 3858 } // Next block. 3859 3860 // Now we have two lists that hold the loads and the stores. 3861 // Next, we find the pointers that they use. 3862 3863 // Check if we see any stores. If there are no stores, then we don't 3864 // care if the pointers are *restrict*. 3865 if (!Stores.size()) { 3866 DEBUG(dbgs() << "LV: Found a read-only loop!\n"); 3867 return true; 3868 } 3869 3870 AccessAnalysis::DepCandidates DependentAccesses; 3871 AccessAnalysis Accesses(DL, DependentAccesses); 3872 3873 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects 3874 // multiple times on the same object. If the ptr is accessed twice, once 3875 // for read and once for write, it will only appear once (on the write 3876 // list). This is okay, since we are going to check for conflicts between 3877 // writes and between reads and writes, but not between reads and reads. 3878 ValueSet Seen; 3879 3880 ValueVector::iterator I, IE; 3881 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) { 3882 StoreInst *ST = cast<StoreInst>(*I); 3883 Value* Ptr = ST->getPointerOperand(); 3884 3885 if (isUniform(Ptr)) { 3886 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 3887 return false; 3888 } 3889 3890 // If we did *not* see this pointer before, insert it to the read-write 3891 // list. At this phase it is only a 'write' list. 3892 if (Seen.insert(Ptr)) { 3893 ++NumReadWrites; 3894 Accesses.addStore(Ptr); 3895 } 3896 } 3897 3898 if (IsAnnotatedParallel) { 3899 DEBUG(dbgs() 3900 << "LV: A loop annotated parallel, ignore memory dependency " 3901 << "checks.\n"); 3902 return true; 3903 } 3904 3905 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { 3906 LoadInst *LD = cast<LoadInst>(*I); 3907 Value* Ptr = LD->getPointerOperand(); 3908 // If we did *not* see this pointer before, insert it to the 3909 // read list. If we *did* see it before, then it is already in 3910 // the read-write list. This allows us to vectorize expressions 3911 // such as A[i] += x; Because the address of A[i] is a read-write 3912 // pointer. This only works if the index of A[i] is consecutive. 3913 // If the address of i is unknown (for example A[B[i]]) then we may 3914 // read a few words, modify, and write a few words, and some of the 3915 // words may be written to the same address. 3916 bool IsReadOnlyPtr = false; 3917 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) { 3918 ++NumReads; 3919 IsReadOnlyPtr = true; 3920 } 3921 Accesses.addLoad(Ptr, IsReadOnlyPtr); 3922 } 3923 3924 // If we write (or read-write) to a single destination and there are no 3925 // other reads in this loop then is it safe to vectorize. 3926 if (NumReadWrites == 1 && NumReads == 0) { 3927 DEBUG(dbgs() << "LV: Found a write-only loop!\n"); 3928 return true; 3929 } 3930 3931 // Build dependence sets and check whether we need a runtime pointer bounds 3932 // check. 3933 Accesses.buildDependenceSets(); 3934 bool NeedRTCheck = Accesses.isRTCheckNeeded(); 3935 3936 // Find pointers with computable bounds. We are going to use this information 3937 // to place a runtime bound check. 3938 unsigned NumComparisons = 0; 3939 bool CanDoRT = false; 3940 if (NeedRTCheck) 3941 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop); 3942 3943 3944 DEBUG(dbgs() << "LV: We need to do " << NumComparisons << 3945 " pointer comparisons.\n"); 3946 3947 // If we only have one set of dependences to check pointers among we don't 3948 // need a runtime check. 3949 if (NumComparisons == 0 && NeedRTCheck) 3950 NeedRTCheck = false; 3951 3952 // Check that we did not collect too many pointers or found an unsizeable 3953 // pointer. 3954 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { 3955 PtrRtCheck.reset(); 3956 CanDoRT = false; 3957 } 3958 3959 if (CanDoRT) { 3960 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n"); 3961 } 3962 3963 if (NeedRTCheck && !CanDoRT) { 3964 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " << 3965 "the array bounds.\n"); 3966 PtrRtCheck.reset(); 3967 return false; 3968 } 3969 3970 PtrRtCheck.Need = NeedRTCheck; 3971 3972 bool CanVecMem = true; 3973 if (Accesses.isDependencyCheckNeeded()) { 3974 DEBUG(dbgs() << "LV: Checking memory dependencies\n"); 3975 CanVecMem = DepChecker.areDepsSafe(DependentAccesses, 3976 Accesses.getDependenciesToCheck()); 3977 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes(); 3978 3979 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) { 3980 DEBUG(dbgs() << "LV: Retrying with memory checks\n"); 3981 NeedRTCheck = true; 3982 3983 // Clear the dependency checks. We assume they are not needed. 3984 Accesses.resetDepChecks(); 3985 3986 PtrRtCheck.reset(); 3987 PtrRtCheck.Need = true; 3988 3989 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, 3990 TheLoop, true); 3991 // Check that we did not collect too many pointers or found an unsizeable 3992 // pointer. 3993 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { 3994 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n"); 3995 PtrRtCheck.reset(); 3996 return false; 3997 } 3998 3999 CanVecMem = true; 4000 } 4001 } 4002 4003 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") << 4004 " need a runtime memory check.\n"); 4005 4006 return CanVecMem; 4007} 4008 4009static bool hasMultipleUsesOf(Instruction *I, 4010 SmallPtrSet<Instruction *, 8> &Insts) { 4011 unsigned NumUses = 0; 4012 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) { 4013 if (Insts.count(dyn_cast<Instruction>(*Use))) 4014 ++NumUses; 4015 if (NumUses > 1) 4016 return true; 4017 } 4018 4019 return false; 4020} 4021 4022static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) { 4023 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) 4024 if (!Set.count(dyn_cast<Instruction>(*Use))) 4025 return false; 4026 return true; 4027} 4028 4029bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, 4030 ReductionKind Kind) { 4031 if (Phi->getNumIncomingValues() != 2) 4032 return false; 4033 4034 // Reduction variables are only found in the loop header block. 4035 if (Phi->getParent() != TheLoop->getHeader()) 4036 return false; 4037 4038 // Obtain the reduction start value from the value that comes from the loop 4039 // preheader. 4040 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 4041 4042 // ExitInstruction is the single value which is used outside the loop. 4043 // We only allow for a single reduction value to be used outside the loop. 4044 // This includes users of the reduction, variables (which form a cycle 4045 // which ends in the phi node). 4046 Instruction *ExitInstruction = 0; 4047 // Indicates that we found a reduction operation in our scan. 4048 bool FoundReduxOp = false; 4049 4050 // We start with the PHI node and scan for all of the users of this 4051 // instruction. All users must be instructions that can be used as reduction 4052 // variables (such as ADD). We must have a single out-of-block user. The cycle 4053 // must include the original PHI. 4054 bool FoundStartPHI = false; 4055 4056 // To recognize min/max patterns formed by a icmp select sequence, we store 4057 // the number of instruction we saw from the recognized min/max pattern, 4058 // to make sure we only see exactly the two instructions. 4059 unsigned NumCmpSelectPatternInst = 0; 4060 ReductionInstDesc ReduxDesc(false, 0); 4061 4062 SmallPtrSet<Instruction *, 8> VisitedInsts; 4063 SmallVector<Instruction *, 8> Worklist; 4064 Worklist.push_back(Phi); 4065 VisitedInsts.insert(Phi); 4066 4067 // A value in the reduction can be used: 4068 // - By the reduction: 4069 // - Reduction operation: 4070 // - One use of reduction value (safe). 4071 // - Multiple use of reduction value (not safe). 4072 // - PHI: 4073 // - All uses of the PHI must be the reduction (safe). 4074 // - Otherwise, not safe. 4075 // - By one instruction outside of the loop (safe). 4076 // - By further instructions outside of the loop (not safe). 4077 // - By an instruction that is not part of the reduction (not safe). 4078 // This is either: 4079 // * An instruction type other than PHI or the reduction operation. 4080 // * A PHI in the header other than the initial PHI. 4081 while (!Worklist.empty()) { 4082 Instruction *Cur = Worklist.back(); 4083 Worklist.pop_back(); 4084 4085 // No Users. 4086 // If the instruction has no users then this is a broken chain and can't be 4087 // a reduction variable. 4088 if (Cur->use_empty()) 4089 return false; 4090 4091 bool IsAPhi = isa<PHINode>(Cur); 4092 4093 // A header PHI use other than the original PHI. 4094 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent()) 4095 return false; 4096 4097 // Reductions of instructions such as Div, and Sub is only possible if the 4098 // LHS is the reduction variable. 4099 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) && 4100 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) && 4101 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0)))) 4102 return false; 4103 4104 // Any reduction instruction must be of one of the allowed kinds. 4105 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc); 4106 if (!ReduxDesc.IsReduction) 4107 return false; 4108 4109 // A reduction operation must only have one use of the reduction value. 4110 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax && 4111 hasMultipleUsesOf(Cur, VisitedInsts)) 4112 return false; 4113 4114 // All inputs to a PHI node must be a reduction value. 4115 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts)) 4116 return false; 4117 4118 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) || 4119 isa<SelectInst>(Cur))) 4120 ++NumCmpSelectPatternInst; 4121 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) || 4122 isa<SelectInst>(Cur))) 4123 ++NumCmpSelectPatternInst; 4124 4125 // Check whether we found a reduction operator. 4126 FoundReduxOp |= !IsAPhi; 4127 4128 // Process users of current instruction. Push non PHI nodes after PHI nodes 4129 // onto the stack. This way we are going to have seen all inputs to PHI 4130 // nodes once we get to them. 4131 SmallVector<Instruction *, 8> NonPHIs; 4132 SmallVector<Instruction *, 8> PHIs; 4133 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E; 4134 ++UI) { 4135 Instruction *Usr = cast<Instruction>(*UI); 4136 4137 // Check if we found the exit user. 4138 BasicBlock *Parent = Usr->getParent(); 4139 if (!TheLoop->contains(Parent)) { 4140 // Exit if you find multiple outside users or if the header phi node is 4141 // being used. In this case the user uses the value of the previous 4142 // iteration, in which case we would loose "VF-1" iterations of the 4143 // reduction operation if we vectorize. 4144 if (ExitInstruction != 0 || Cur == Phi) 4145 return false; 4146 4147 // The instruction used by an outside user must be the last instruction 4148 // before we feed back to the reduction phi. Otherwise, we loose VF-1 4149 // operations on the value. 4150 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end()) 4151 return false; 4152 4153 ExitInstruction = Cur; 4154 continue; 4155 } 4156 4157 // Process instructions only once (termination). 4158 if (VisitedInsts.insert(Usr)) { 4159 if (isa<PHINode>(Usr)) 4160 PHIs.push_back(Usr); 4161 else 4162 NonPHIs.push_back(Usr); 4163 } 4164 // Remember that we completed the cycle. 4165 if (Usr == Phi) 4166 FoundStartPHI = true; 4167 } 4168 Worklist.append(PHIs.begin(), PHIs.end()); 4169 Worklist.append(NonPHIs.begin(), NonPHIs.end()); 4170 } 4171 4172 // This means we have seen one but not the other instruction of the 4173 // pattern or more than just a select and cmp. 4174 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) && 4175 NumCmpSelectPatternInst != 2) 4176 return false; 4177 4178 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction) 4179 return false; 4180 4181 // We found a reduction var if we have reached the original phi node and we 4182 // only have a single instruction with out-of-loop users. 4183 4184 // This instruction is allowed to have out-of-loop users. 4185 AllowedExit.insert(ExitInstruction); 4186 4187 // Save the description of this reduction variable. 4188 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind, 4189 ReduxDesc.MinMaxKind); 4190 Reductions[Phi] = RD; 4191 // We've ended the cycle. This is a reduction variable if we have an 4192 // outside user and it has a binary op. 4193 4194 return true; 4195} 4196 4197/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 4198/// pattern corresponding to a min(X, Y) or max(X, Y). 4199LoopVectorizationLegality::ReductionInstDesc 4200LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, 4201 ReductionInstDesc &Prev) { 4202 4203 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) && 4204 "Expect a select instruction"); 4205 Instruction *Cmp = 0; 4206 SelectInst *Select = 0; 4207 4208 // We must handle the select(cmp()) as a single instruction. Advance to the 4209 // select. 4210 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) { 4211 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin()))) 4212 return ReductionInstDesc(false, I); 4213 return ReductionInstDesc(Select, Prev.MinMaxKind); 4214 } 4215 4216 // Only handle single use cases for now. 4217 if (!(Select = dyn_cast<SelectInst>(I))) 4218 return ReductionInstDesc(false, I); 4219 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) && 4220 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0)))) 4221 return ReductionInstDesc(false, I); 4222 if (!Cmp->hasOneUse()) 4223 return ReductionInstDesc(false, I); 4224 4225 Value *CmpLeft; 4226 Value *CmpRight; 4227 4228 // Look for a min/max pattern. 4229 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4230 return ReductionInstDesc(Select, MRK_UIntMin); 4231 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4232 return ReductionInstDesc(Select, MRK_UIntMax); 4233 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4234 return ReductionInstDesc(Select, MRK_SIntMax); 4235 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4236 return ReductionInstDesc(Select, MRK_SIntMin); 4237 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4238 return ReductionInstDesc(Select, MRK_FloatMin); 4239 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4240 return ReductionInstDesc(Select, MRK_FloatMax); 4241 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4242 return ReductionInstDesc(Select, MRK_FloatMin); 4243 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 4244 return ReductionInstDesc(Select, MRK_FloatMax); 4245 4246 return ReductionInstDesc(false, I); 4247} 4248 4249LoopVectorizationLegality::ReductionInstDesc 4250LoopVectorizationLegality::isReductionInstr(Instruction *I, 4251 ReductionKind Kind, 4252 ReductionInstDesc &Prev) { 4253 bool FP = I->getType()->isFloatingPointTy(); 4254 bool FastMath = (FP && I->isCommutative() && I->isAssociative()); 4255 switch (I->getOpcode()) { 4256 default: 4257 return ReductionInstDesc(false, I); 4258 case Instruction::PHI: 4259 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd && 4260 Kind != RK_FloatMinMax)) 4261 return ReductionInstDesc(false, I); 4262 return ReductionInstDesc(I, Prev.MinMaxKind); 4263 case Instruction::Sub: 4264 case Instruction::Add: 4265 return ReductionInstDesc(Kind == RK_IntegerAdd, I); 4266 case Instruction::Mul: 4267 return ReductionInstDesc(Kind == RK_IntegerMult, I); 4268 case Instruction::And: 4269 return ReductionInstDesc(Kind == RK_IntegerAnd, I); 4270 case Instruction::Or: 4271 return ReductionInstDesc(Kind == RK_IntegerOr, I); 4272 case Instruction::Xor: 4273 return ReductionInstDesc(Kind == RK_IntegerXor, I); 4274 case Instruction::FMul: 4275 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I); 4276 case Instruction::FAdd: 4277 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I); 4278 case Instruction::FCmp: 4279 case Instruction::ICmp: 4280 case Instruction::Select: 4281 if (Kind != RK_IntegerMinMax && 4282 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax)) 4283 return ReductionInstDesc(false, I); 4284 return isMinMaxSelectCmpPattern(I, Prev); 4285 } 4286} 4287 4288LoopVectorizationLegality::InductionKind 4289LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { 4290 Type *PhiTy = Phi->getType(); 4291 // We only handle integer and pointer inductions variables. 4292 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 4293 return IK_NoInduction; 4294 4295 // Check that the PHI is consecutive. 4296 const SCEV *PhiScev = SE->getSCEV(Phi); 4297 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 4298 if (!AR) { 4299 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 4300 return IK_NoInduction; 4301 } 4302 const SCEV *Step = AR->getStepRecurrence(*SE); 4303 4304 // Integer inductions need to have a stride of one. 4305 if (PhiTy->isIntegerTy()) { 4306 if (Step->isOne()) 4307 return IK_IntInduction; 4308 if (Step->isAllOnesValue()) 4309 return IK_ReverseIntInduction; 4310 return IK_NoInduction; 4311 } 4312 4313 // Calculate the pointer stride and check if it is consecutive. 4314 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 4315 if (!C) 4316 return IK_NoInduction; 4317 4318 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 4319 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType()); 4320 if (C->getValue()->equalsInt(Size)) 4321 return IK_PtrInduction; 4322 else if (C->getValue()->equalsInt(0 - Size)) 4323 return IK_ReversePtrInduction; 4324 4325 return IK_NoInduction; 4326} 4327 4328bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 4329 Value *In0 = const_cast<Value*>(V); 4330 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 4331 if (!PN) 4332 return false; 4333 4334 return Inductions.count(PN); 4335} 4336 4337bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 4338 assert(TheLoop->contains(BB) && "Unknown block used"); 4339 4340 // Blocks that do not dominate the latch need predication. 4341 BasicBlock* Latch = TheLoop->getLoopLatch(); 4342 return !DT->dominates(BB, Latch); 4343} 4344 4345bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB, 4346 SmallPtrSet<Value *, 8>& SafePtrs) { 4347 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4348 // We might be able to hoist the load. 4349 if (it->mayReadFromMemory()) { 4350 LoadInst *LI = dyn_cast<LoadInst>(it); 4351 if (!LI || !SafePtrs.count(LI->getPointerOperand())) 4352 return false; 4353 } 4354 4355 // We don't predicate stores at the moment. 4356 if (it->mayWriteToMemory() || it->mayThrow()) 4357 return false; 4358 4359 // The instructions below can trap. 4360 switch (it->getOpcode()) { 4361 default: continue; 4362 case Instruction::UDiv: 4363 case Instruction::SDiv: 4364 case Instruction::URem: 4365 case Instruction::SRem: 4366 return false; 4367 } 4368 } 4369 4370 return true; 4371} 4372 4373LoopVectorizationCostModel::VectorizationFactor 4374LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize, 4375 unsigned UserVF) { 4376 // Width 1 means no vectorize 4377 VectorizationFactor Factor = { 1U, 0U }; 4378 if (OptForSize && Legal->getRuntimePointerCheck()->Need) { 4379 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); 4380 return Factor; 4381 } 4382 4383 // Find the trip count. 4384 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch()); 4385 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n'); 4386 4387 unsigned WidestType = getWidestType(); 4388 unsigned WidestRegister = TTI.getRegisterBitWidth(true); 4389 unsigned MaxSafeDepDist = -1U; 4390 if (Legal->getMaxSafeDepDistBytes() != -1U) 4391 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8; 4392 WidestRegister = ((WidestRegister < MaxSafeDepDist) ? 4393 WidestRegister : MaxSafeDepDist); 4394 unsigned MaxVectorSize = WidestRegister / WidestType; 4395 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n"); 4396 DEBUG(dbgs() << "LV: The Widest register is: " 4397 << WidestRegister << " bits.\n"); 4398 4399 if (MaxVectorSize == 0) { 4400 DEBUG(dbgs() << "LV: The target has no vector registers.\n"); 4401 MaxVectorSize = 1; 4402 } 4403 4404 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements" 4405 " into one vector!"); 4406 4407 unsigned VF = MaxVectorSize; 4408 4409 // If we optimize the program for size, avoid creating the tail loop. 4410 if (OptForSize) { 4411 // If we are unable to calculate the trip count then don't try to vectorize. 4412 if (TC < 2) { 4413 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 4414 return Factor; 4415 } 4416 4417 // Find the maximum SIMD width that can fit within the trip count. 4418 VF = TC % MaxVectorSize; 4419 4420 if (VF == 0) 4421 VF = MaxVectorSize; 4422 4423 // If the trip count that we found modulo the vectorization factor is not 4424 // zero then we require a tail. 4425 if (VF < 2) { 4426 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 4427 return Factor; 4428 } 4429 } 4430 4431 if (UserVF != 0) { 4432 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 4433 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n"); 4434 4435 Factor.Width = UserVF; 4436 return Factor; 4437 } 4438 4439 float Cost = expectedCost(1); 4440 unsigned Width = 1; 4441 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n"); 4442 for (unsigned i=2; i <= VF; i*=2) { 4443 // Notice that the vector loop needs to be executed less times, so 4444 // we need to divide the cost of the vector loops by the width of 4445 // the vector elements. 4446 float VectorCost = expectedCost(i) / (float)i; 4447 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " << 4448 (int)VectorCost << ".\n"); 4449 if (VectorCost < Cost) { 4450 Cost = VectorCost; 4451 Width = i; 4452 } 4453 } 4454 4455 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n"); 4456 Factor.Width = Width; 4457 Factor.Cost = Width * Cost; 4458 return Factor; 4459} 4460 4461unsigned LoopVectorizationCostModel::getWidestType() { 4462 unsigned MaxWidth = 8; 4463 4464 // For each block. 4465 for (Loop::block_iterator bb = TheLoop->block_begin(), 4466 be = TheLoop->block_end(); bb != be; ++bb) { 4467 BasicBlock *BB = *bb; 4468 4469 // For each instruction in the loop. 4470 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4471 Type *T = it->getType(); 4472 4473 // Only examine Loads, Stores and PHINodes. 4474 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it)) 4475 continue; 4476 4477 // Examine PHI nodes that are reduction variables. 4478 if (PHINode *PN = dyn_cast<PHINode>(it)) 4479 if (!Legal->getReductionVars()->count(PN)) 4480 continue; 4481 4482 // Examine the stored values. 4483 if (StoreInst *ST = dyn_cast<StoreInst>(it)) 4484 T = ST->getValueOperand()->getType(); 4485 4486 // Ignore loaded pointer types and stored pointer types that are not 4487 // consecutive. However, we do want to take consecutive stores/loads of 4488 // pointer vectors into account. 4489 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it)) 4490 continue; 4491 4492 MaxWidth = std::max(MaxWidth, 4493 (unsigned)DL->getTypeSizeInBits(T->getScalarType())); 4494 } 4495 } 4496 4497 return MaxWidth; 4498} 4499 4500unsigned 4501LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, 4502 unsigned UserUF, 4503 unsigned VF, 4504 unsigned LoopCost) { 4505 4506 // -- The unroll heuristics -- 4507 // We unroll the loop in order to expose ILP and reduce the loop overhead. 4508 // There are many micro-architectural considerations that we can't predict 4509 // at this level. For example frontend pressure (on decode or fetch) due to 4510 // code size, or the number and capabilities of the execution ports. 4511 // 4512 // We use the following heuristics to select the unroll factor: 4513 // 1. If the code has reductions the we unroll in order to break the cross 4514 // iteration dependency. 4515 // 2. If the loop is really small then we unroll in order to reduce the loop 4516 // overhead. 4517 // 3. We don't unroll if we think that we will spill registers to memory due 4518 // to the increased register pressure. 4519 4520 // Use the user preference, unless 'auto' is selected. 4521 if (UserUF != 0) 4522 return UserUF; 4523 4524 // When we optimize for size we don't unroll. 4525 if (OptForSize) 4526 return 1; 4527 4528 // We used the distance for the unroll factor. 4529 if (Legal->getMaxSafeDepDistBytes() != -1U) 4530 return 1; 4531 4532 // Do not unroll loops with a relatively small trip count. 4533 unsigned TC = SE->getSmallConstantTripCount(TheLoop, 4534 TheLoop->getLoopLatch()); 4535 if (TC > 1 && TC < TinyTripCountUnrollThreshold) 4536 return 1; 4537 4538 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true); 4539 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters << 4540 " vector registers\n"); 4541 4542 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); 4543 // We divide by these constants so assume that we have at least one 4544 // instruction that uses at least one register. 4545 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 4546 R.NumInstructions = std::max(R.NumInstructions, 1U); 4547 4548 // We calculate the unroll factor using the following formula. 4549 // Subtract the number of loop invariants from the number of available 4550 // registers. These registers are used by all of the unrolled instances. 4551 // Next, divide the remaining registers by the number of registers that is 4552 // required by the loop, in order to estimate how many parallel instances 4553 // fit without causing spills. 4554 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers; 4555 4556 // Clamp the unroll factor ranges to reasonable factors. 4557 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor(); 4558 4559 // If we did not calculate the cost for VF (because the user selected the VF) 4560 // then we calculate the cost of VF here. 4561 if (LoopCost == 0) 4562 LoopCost = expectedCost(VF); 4563 4564 // Clamp the calculated UF to be between the 1 and the max unroll factor 4565 // that the target allows. 4566 if (UF > MaxUnrollSize) 4567 UF = MaxUnrollSize; 4568 else if (UF < 1) 4569 UF = 1; 4570 4571 bool HasReductions = Legal->getReductionVars()->size(); 4572 4573 // Decide if we want to unroll if we decided that it is legal to vectorize 4574 // but not profitable. 4575 if (VF == 1) { 4576 if (TheLoop->getNumBlocks() > 1 || !HasReductions || 4577 LoopCost > SmallLoopCost) 4578 return 1; 4579 4580 return UF; 4581 } 4582 4583 if (HasReductions) { 4584 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n"); 4585 return UF; 4586 } 4587 4588 // We want to unroll tiny loops in order to reduce the loop overhead. 4589 // We assume that the cost overhead is 1 and we use the cost model 4590 // to estimate the cost of the loop and unroll until the cost of the 4591 // loop overhead is about 5% of the cost of the loop. 4592 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n'); 4593 if (LoopCost < SmallLoopCost) { 4594 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n"); 4595 unsigned NewUF = SmallLoopCost / (LoopCost + 1); 4596 return std::min(NewUF, UF); 4597 } 4598 4599 DEBUG(dbgs() << "LV: Not Unrolling.\n"); 4600 return 1; 4601} 4602 4603LoopVectorizationCostModel::RegisterUsage 4604LoopVectorizationCostModel::calculateRegisterUsage() { 4605 // This function calculates the register usage by measuring the highest number 4606 // of values that are alive at a single location. Obviously, this is a very 4607 // rough estimation. We scan the loop in a topological order in order and 4608 // assign a number to each instruction. We use RPO to ensure that defs are 4609 // met before their users. We assume that each instruction that has in-loop 4610 // users starts an interval. We record every time that an in-loop value is 4611 // used, so we have a list of the first and last occurrences of each 4612 // instruction. Next, we transpose this data structure into a multi map that 4613 // holds the list of intervals that *end* at a specific location. This multi 4614 // map allows us to perform a linear search. We scan the instructions linearly 4615 // and record each time that a new interval starts, by placing it in a set. 4616 // If we find this value in the multi-map then we remove it from the set. 4617 // The max register usage is the maximum size of the set. 4618 // We also search for instructions that are defined outside the loop, but are 4619 // used inside the loop. We need this number separately from the max-interval 4620 // usage number because when we unroll, loop-invariant values do not take 4621 // more register. 4622 LoopBlocksDFS DFS(TheLoop); 4623 DFS.perform(LI); 4624 4625 RegisterUsage R; 4626 R.NumInstructions = 0; 4627 4628 // Each 'key' in the map opens a new interval. The values 4629 // of the map are the index of the 'last seen' usage of the 4630 // instruction that is the key. 4631 typedef DenseMap<Instruction*, unsigned> IntervalMap; 4632 // Maps instruction to its index. 4633 DenseMap<unsigned, Instruction*> IdxToInstr; 4634 // Marks the end of each interval. 4635 IntervalMap EndPoint; 4636 // Saves the list of instruction indices that are used in the loop. 4637 SmallSet<Instruction*, 8> Ends; 4638 // Saves the list of values that are used in the loop but are 4639 // defined outside the loop, such as arguments and constants. 4640 SmallPtrSet<Value*, 8> LoopInvariants; 4641 4642 unsigned Index = 0; 4643 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 4644 be = DFS.endRPO(); bb != be; ++bb) { 4645 R.NumInstructions += (*bb)->size(); 4646 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 4647 ++it) { 4648 Instruction *I = it; 4649 IdxToInstr[Index++] = I; 4650 4651 // Save the end location of each USE. 4652 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 4653 Value *U = I->getOperand(i); 4654 Instruction *Instr = dyn_cast<Instruction>(U); 4655 4656 // Ignore non-instruction values such as arguments, constants, etc. 4657 if (!Instr) continue; 4658 4659 // If this instruction is outside the loop then record it and continue. 4660 if (!TheLoop->contains(Instr)) { 4661 LoopInvariants.insert(Instr); 4662 continue; 4663 } 4664 4665 // Overwrite previous end points. 4666 EndPoint[Instr] = Index; 4667 Ends.insert(Instr); 4668 } 4669 } 4670 } 4671 4672 // Saves the list of intervals that end with the index in 'key'. 4673 typedef SmallVector<Instruction*, 2> InstrList; 4674 DenseMap<unsigned, InstrList> TransposeEnds; 4675 4676 // Transpose the EndPoints to a list of values that end at each index. 4677 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 4678 it != e; ++it) 4679 TransposeEnds[it->second].push_back(it->first); 4680 4681 SmallSet<Instruction*, 8> OpenIntervals; 4682 unsigned MaxUsage = 0; 4683 4684 4685 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 4686 for (unsigned int i = 0; i < Index; ++i) { 4687 Instruction *I = IdxToInstr[i]; 4688 // Ignore instructions that are never used within the loop. 4689 if (!Ends.count(I)) continue; 4690 4691 // Remove all of the instructions that end at this location. 4692 InstrList &List = TransposeEnds[i]; 4693 for (unsigned int j=0, e = List.size(); j < e; ++j) 4694 OpenIntervals.erase(List[j]); 4695 4696 // Count the number of live interals. 4697 MaxUsage = std::max(MaxUsage, OpenIntervals.size()); 4698 4699 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << 4700 OpenIntervals.size() << '\n'); 4701 4702 // Add the current instruction to the list of open intervals. 4703 OpenIntervals.insert(I); 4704 } 4705 4706 unsigned Invariant = LoopInvariants.size(); 4707 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n'); 4708 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n'); 4709 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n'); 4710 4711 R.LoopInvariantRegs = Invariant; 4712 R.MaxLocalUsers = MaxUsage; 4713 return R; 4714} 4715 4716unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 4717 unsigned Cost = 0; 4718 4719 // For each block. 4720 for (Loop::block_iterator bb = TheLoop->block_begin(), 4721 be = TheLoop->block_end(); bb != be; ++bb) { 4722 unsigned BlockCost = 0; 4723 BasicBlock *BB = *bb; 4724 4725 // For each instruction in the old loop. 4726 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 4727 // Skip dbg intrinsics. 4728 if (isa<DbgInfoIntrinsic>(it)) 4729 continue; 4730 4731 unsigned C = getInstructionCost(it, VF); 4732 BlockCost += C; 4733 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " << 4734 VF << " For instruction: " << *it << '\n'); 4735 } 4736 4737 // We assume that if-converted blocks have a 50% chance of being executed. 4738 // When the code is scalar then some of the blocks are avoided due to CF. 4739 // When the code is vectorized we execute all code paths. 4740 if (VF == 1 && Legal->blockNeedsPredication(*bb)) 4741 BlockCost /= 2; 4742 4743 Cost += BlockCost; 4744 } 4745 4746 return Cost; 4747} 4748 4749/// \brief Check whether the address computation for a non-consecutive memory 4750/// access looks like an unlikely candidate for being merged into the indexing 4751/// mode. 4752/// 4753/// We look for a GEP which has one index that is an induction variable and all 4754/// other indices are loop invariant. If the stride of this access is also 4755/// within a small bound we decide that this address computation can likely be 4756/// merged into the addressing mode. 4757/// In all other cases, we identify the address computation as complex. 4758static bool isLikelyComplexAddressComputation(Value *Ptr, 4759 LoopVectorizationLegality *Legal, 4760 ScalarEvolution *SE, 4761 const Loop *TheLoop) { 4762 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 4763 if (!Gep) 4764 return true; 4765 4766 // We are looking for a gep with all loop invariant indices except for one 4767 // which should be an induction variable. 4768 unsigned NumOperands = Gep->getNumOperands(); 4769 for (unsigned i = 1; i < NumOperands; ++i) { 4770 Value *Opd = Gep->getOperand(i); 4771 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && 4772 !Legal->isInductionVariable(Opd)) 4773 return true; 4774 } 4775 4776 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step 4777 // can likely be merged into the address computation. 4778 unsigned MaxMergeDistance = 64; 4779 4780 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr)); 4781 if (!AddRec) 4782 return true; 4783 4784 // Check the step is constant. 4785 const SCEV *Step = AddRec->getStepRecurrence(*SE); 4786 // Calculate the pointer stride and check if it is consecutive. 4787 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 4788 if (!C) 4789 return true; 4790 4791 const APInt &APStepVal = C->getValue()->getValue(); 4792 4793 // Huge step value - give up. 4794 if (APStepVal.getBitWidth() > 64) 4795 return true; 4796 4797 int64_t StepVal = APStepVal.getSExtValue(); 4798 4799 return StepVal > MaxMergeDistance; 4800} 4801 4802unsigned 4803LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 4804 // If we know that this instruction will remain uniform, check the cost of 4805 // the scalar version. 4806 if (Legal->isUniformAfterVectorization(I)) 4807 VF = 1; 4808 4809 Type *RetTy = I->getType(); 4810 Type *VectorTy = ToVectorTy(RetTy, VF); 4811 4812 // TODO: We need to estimate the cost of intrinsic calls. 4813 switch (I->getOpcode()) { 4814 case Instruction::GetElementPtr: 4815 // We mark this instruction as zero-cost because the cost of GEPs in 4816 // vectorized code depends on whether the corresponding memory instruction 4817 // is scalarized or not. Therefore, we handle GEPs with the memory 4818 // instruction cost. 4819 return 0; 4820 case Instruction::Br: { 4821 return TTI.getCFInstrCost(I->getOpcode()); 4822 } 4823 case Instruction::PHI: 4824 //TODO: IF-converted IFs become selects. 4825 return 0; 4826 case Instruction::Add: 4827 case Instruction::FAdd: 4828 case Instruction::Sub: 4829 case Instruction::FSub: 4830 case Instruction::Mul: 4831 case Instruction::FMul: 4832 case Instruction::UDiv: 4833 case Instruction::SDiv: 4834 case Instruction::FDiv: 4835 case Instruction::URem: 4836 case Instruction::SRem: 4837 case Instruction::FRem: 4838 case Instruction::Shl: 4839 case Instruction::LShr: 4840 case Instruction::AShr: 4841 case Instruction::And: 4842 case Instruction::Or: 4843 case Instruction::Xor: { 4844 // Certain instructions can be cheaper to vectorize if they have a constant 4845 // second vector operand. One example of this are shifts on x86. 4846 TargetTransformInfo::OperandValueKind Op1VK = 4847 TargetTransformInfo::OK_AnyValue; 4848 TargetTransformInfo::OperandValueKind Op2VK = 4849 TargetTransformInfo::OK_AnyValue; 4850 4851 if (isa<ConstantInt>(I->getOperand(1))) 4852 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 4853 4854 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK); 4855 } 4856 case Instruction::Select: { 4857 SelectInst *SI = cast<SelectInst>(I); 4858 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 4859 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 4860 Type *CondTy = SI->getCondition()->getType(); 4861 if (!ScalarCond) 4862 CondTy = VectorType::get(CondTy, VF); 4863 4864 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 4865 } 4866 case Instruction::ICmp: 4867 case Instruction::FCmp: { 4868 Type *ValTy = I->getOperand(0)->getType(); 4869 VectorTy = ToVectorTy(ValTy, VF); 4870 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); 4871 } 4872 case Instruction::Store: 4873 case Instruction::Load: { 4874 StoreInst *SI = dyn_cast<StoreInst>(I); 4875 LoadInst *LI = dyn_cast<LoadInst>(I); 4876 Type *ValTy = (SI ? SI->getValueOperand()->getType() : 4877 LI->getType()); 4878 VectorTy = ToVectorTy(ValTy, VF); 4879 4880 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); 4881 unsigned AS = SI ? SI->getPointerAddressSpace() : 4882 LI->getPointerAddressSpace(); 4883 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); 4884 // We add the cost of address computation here instead of with the gep 4885 // instruction because only here we know whether the operation is 4886 // scalarized. 4887 if (VF == 1) 4888 return TTI.getAddressComputationCost(VectorTy) + 4889 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 4890 4891 // Scalarized loads/stores. 4892 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 4893 bool Reverse = ConsecutiveStride < 0; 4894 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy); 4895 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF; 4896 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { 4897 bool IsComplexComputation = 4898 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop); 4899 unsigned Cost = 0; 4900 // The cost of extracting from the value vector and pointer vector. 4901 Type *PtrTy = ToVectorTy(Ptr->getType(), VF); 4902 for (unsigned i = 0; i < VF; ++i) { 4903 // The cost of extracting the pointer operand. 4904 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); 4905 // In case of STORE, the cost of ExtractElement from the vector. 4906 // In case of LOAD, the cost of InsertElement into the returned 4907 // vector. 4908 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : 4909 Instruction::InsertElement, 4910 VectorTy, i); 4911 } 4912 4913 // The cost of the scalar loads/stores. 4914 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation); 4915 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), 4916 Alignment, AS); 4917 return Cost; 4918 } 4919 4920 // Wide load/stores. 4921 unsigned Cost = TTI.getAddressComputationCost(VectorTy); 4922 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 4923 4924 if (Reverse) 4925 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, 4926 VectorTy, 0); 4927 return Cost; 4928 } 4929 case Instruction::ZExt: 4930 case Instruction::SExt: 4931 case Instruction::FPToUI: 4932 case Instruction::FPToSI: 4933 case Instruction::FPExt: 4934 case Instruction::PtrToInt: 4935 case Instruction::IntToPtr: 4936 case Instruction::SIToFP: 4937 case Instruction::UIToFP: 4938 case Instruction::Trunc: 4939 case Instruction::FPTrunc: 4940 case Instruction::BitCast: { 4941 // We optimize the truncation of induction variable. 4942 // The cost of these is the same as the scalar operation. 4943 if (I->getOpcode() == Instruction::Trunc && 4944 Legal->isInductionVariable(I->getOperand(0))) 4945 return TTI.getCastInstrCost(I->getOpcode(), I->getType(), 4946 I->getOperand(0)->getType()); 4947 4948 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); 4949 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 4950 } 4951 case Instruction::Call: { 4952 CallInst *CI = cast<CallInst>(I); 4953 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 4954 assert(ID && "Not an intrinsic call!"); 4955 Type *RetTy = ToVectorTy(CI->getType(), VF); 4956 SmallVector<Type*, 4> Tys; 4957 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 4958 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 4959 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); 4960 } 4961 default: { 4962 // We are scalarizing the instruction. Return the cost of the scalar 4963 // instruction, plus the cost of insert and extract into vector 4964 // elements, times the vector width. 4965 unsigned Cost = 0; 4966 4967 if (!RetTy->isVoidTy() && VF != 1) { 4968 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, 4969 VectorTy); 4970 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, 4971 VectorTy); 4972 4973 // The cost of inserting the results plus extracting each one of the 4974 // operands. 4975 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 4976 } 4977 4978 // The cost of executing VF copies of the scalar instruction. This opcode 4979 // is unknown. Assume that it is the same as 'mul'. 4980 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); 4981 return Cost; 4982 } 4983 }// end of switch. 4984} 4985 4986Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) { 4987 if (Scalar->isVoidTy() || VF == 1) 4988 return Scalar; 4989 return VectorType::get(Scalar, VF); 4990} 4991 4992char LoopVectorize::ID = 0; 4993static const char lv_name[] = "Loop Vectorization"; 4994INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 4995INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 4996INITIALIZE_PASS_DEPENDENCY(DominatorTree) 4997INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4998INITIALIZE_PASS_DEPENDENCY(LCSSA) 4999INITIALIZE_PASS_DEPENDENCY(LoopInfo) 5000INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 5001INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 5002 5003namespace llvm { 5004 Pass *createLoopVectorizePass(bool NoUnrolling) { 5005 return new LoopVectorize(NoUnrolling); 5006 } 5007} 5008 5009bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { 5010 // Check for a store. 5011 if (StoreInst *ST = dyn_cast<StoreInst>(Inst)) 5012 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; 5013 5014 // Check for a load. 5015 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 5016 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; 5017 5018 return false; 5019} 5020 5021 5022void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) { 5023 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 5024 // Holds vector parameters or scalars, in case of uniform vals. 5025 SmallVector<VectorParts, 4> Params; 5026 5027 setDebugLocFromInst(Builder, Instr); 5028 5029 // Find all of the vectorized parameters. 5030 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5031 Value *SrcOp = Instr->getOperand(op); 5032 5033 // If we are accessing the old induction variable, use the new one. 5034 if (SrcOp == OldInduction) { 5035 Params.push_back(getVectorValue(SrcOp)); 5036 continue; 5037 } 5038 5039 // Try using previously calculated values. 5040 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 5041 5042 // If the src is an instruction that appeared earlier in the basic block 5043 // then it should already be vectorized. 5044 if (SrcInst && OrigLoop->contains(SrcInst)) { 5045 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 5046 // The parameter is a vector value from earlier. 5047 Params.push_back(WidenMap.get(SrcInst)); 5048 } else { 5049 // The parameter is a scalar from outside the loop. Maybe even a constant. 5050 VectorParts Scalars; 5051 Scalars.append(UF, SrcOp); 5052 Params.push_back(Scalars); 5053 } 5054 } 5055 5056 assert(Params.size() == Instr->getNumOperands() && 5057 "Invalid number of operands"); 5058 5059 // Does this instruction return a value ? 5060 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 5061 5062 Value *UndefVec = IsVoidRetTy ? 0 : 5063 UndefValue::get(Instr->getType()); 5064 // Create a new entry in the WidenMap and initialize it to Undef or Null. 5065 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 5066 5067 // For each vector unroll 'part': 5068 for (unsigned Part = 0; Part < UF; ++Part) { 5069 // For each scalar that we create: 5070 5071 Instruction *Cloned = Instr->clone(); 5072 if (!IsVoidRetTy) 5073 Cloned->setName(Instr->getName() + ".cloned"); 5074 // Replace the operands of the cloned instructions with extracted scalars. 5075 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 5076 Value *Op = Params[op][Part]; 5077 Cloned->setOperand(op, Op); 5078 } 5079 5080 // Place the cloned scalar in the new loop. 5081 Builder.Insert(Cloned); 5082 5083 // If the original scalar returns a value we need to place it in a vector 5084 // so that future users will be able to use it. 5085 if (!IsVoidRetTy) 5086 VecResults[Part] = Cloned; 5087 } 5088} 5089 5090void 5091InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr, 5092 LoopVectorizationLegality*) { 5093 return scalarizeInstruction(Instr); 5094} 5095 5096Value *InnerLoopUnroller::reverseVector(Value *Vec) { 5097 return Vec; 5098} 5099 5100Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { 5101 return V; 5102} 5103 5104Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx, 5105 bool Negate) { 5106 // When unrolling and the VF is 1, we only need to add a simple scalar. 5107 Type *ITy = Val->getType(); 5108 assert(!ITy->isVectorTy() && "Val must be a scalar"); 5109 Constant *C = ConstantInt::get(ITy, StartIdx, Negate); 5110 return Builder.CreateAdd(Val, C, "induction"); 5111} 5112 5113