LoopVectorize.cpp revision 6e4a9c14f6e59b075b3a24825d2dc7dc9e163281
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/MapVector.h" 51#include "llvm/ADT/SmallPtrSet.h" 52#include "llvm/ADT/SmallSet.h" 53#include "llvm/ADT/SmallVector.h" 54#include "llvm/ADT/StringExtras.h" 55#include "llvm/Analysis/AliasAnalysis.h" 56#include "llvm/Analysis/AliasSetTracker.h" 57#include "llvm/Analysis/Dominators.h" 58#include "llvm/Analysis/LoopInfo.h" 59#include "llvm/Analysis/LoopIterator.h" 60#include "llvm/Analysis/LoopPass.h" 61#include "llvm/Analysis/ScalarEvolution.h" 62#include "llvm/Analysis/ScalarEvolutionExpander.h" 63#include "llvm/Analysis/ScalarEvolutionExpressions.h" 64#include "llvm/Analysis/TargetTransformInfo.h" 65#include "llvm/Analysis/ValueTracking.h" 66#include "llvm/Analysis/Verifier.h" 67#include "llvm/IR/Constants.h" 68#include "llvm/IR/DataLayout.h" 69#include "llvm/IR/DerivedTypes.h" 70#include "llvm/IR/Function.h" 71#include "llvm/IR/IRBuilder.h" 72#include "llvm/IR/Instructions.h" 73#include "llvm/IR/IntrinsicInst.h" 74#include "llvm/IR/LLVMContext.h" 75#include "llvm/IR/Module.h" 76#include "llvm/IR/Type.h" 77#include "llvm/IR/Value.h" 78#include "llvm/Pass.h" 79#include "llvm/Support/CommandLine.h" 80#include "llvm/Support/Debug.h" 81#include "llvm/Support/PatternMatch.h" 82#include "llvm/Support/raw_ostream.h" 83#include "llvm/Support/ValueHandle.h" 84#include "llvm/Target/TargetLibraryInfo.h" 85#include "llvm/Transforms/Scalar.h" 86#include "llvm/Transforms/Utils/BasicBlockUtils.h" 87#include "llvm/Transforms/Utils/Local.h" 88#include <algorithm> 89#include <map> 90 91using namespace llvm; 92using namespace llvm::PatternMatch; 93 94static cl::opt<unsigned> 95VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden, 96 cl::desc("Sets the SIMD width. Zero is autoselect.")); 97 98static cl::opt<unsigned> 99VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden, 100 cl::desc("Sets the vectorization unroll count. " 101 "Zero is autoselect.")); 102 103static cl::opt<bool> 104EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, 105 cl::desc("Enable if-conversion during vectorization.")); 106 107/// We don't vectorize loops with a known constant trip count below this number. 108static cl::opt<unsigned> 109TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16), 110 cl::Hidden, 111 cl::desc("Don't vectorize loops with a constant " 112 "trip count that is smaller than this " 113 "value.")); 114 115/// We don't unroll loops with a known constant trip count below this number. 116static const unsigned TinyTripCountUnrollThreshold = 128; 117 118/// When performing memory disambiguation checks at runtime do not make more 119/// than this number of comparisons. 120static const unsigned RuntimeMemoryCheckThreshold = 8; 121 122/// We use a metadata with this name to indicate that a scalar loop was 123/// vectorized and that we don't need to re-vectorize it if we run into it 124/// again. 125static const char* 126AlreadyVectorizedMDName = "llvm.vectorizer.already_vectorized"; 127 128namespace { 129 130// Forward declarations. 131class LoopVectorizationLegality; 132class LoopVectorizationCostModel; 133 134/// InnerLoopVectorizer vectorizes loops which contain only one basic 135/// block to a specified vectorization factor (VF). 136/// This class performs the widening of scalars into vectors, or multiple 137/// scalars. This class also implements the following features: 138/// * It inserts an epilogue loop for handling loops that don't have iteration 139/// counts that are known to be a multiple of the vectorization factor. 140/// * It handles the code generation for reduction variables. 141/// * Scalarization (implementation using scalars) of un-vectorizable 142/// instructions. 143/// InnerLoopVectorizer does not perform any vectorization-legality 144/// checks, and relies on the caller to check for the different legality 145/// aspects. The InnerLoopVectorizer relies on the 146/// LoopVectorizationLegality class to provide information about the induction 147/// and reduction variables that were found to a given vectorization factor. 148class InnerLoopVectorizer { 149public: 150 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, 151 DominatorTree *DT, DataLayout *DL, 152 const TargetLibraryInfo *TLI, unsigned VecWidth, 153 unsigned UnrollFactor) 154 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI), 155 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0), 156 OldInduction(0), WidenMap(UnrollFactor) {} 157 158 // Perform the actual loop widening (vectorization). 159 void vectorize(LoopVectorizationLegality *Legal) { 160 // Create a new empty loop. Unlink the old loop and connect the new one. 161 createEmptyLoop(Legal); 162 // Widen each instruction in the old loop to a new one in the new loop. 163 // Use the Legality module to find the induction and reduction variables. 164 vectorizeLoop(Legal); 165 // Register the new loop and update the analysis passes. 166 updateAnalysis(); 167 } 168 169private: 170 /// A small list of PHINodes. 171 typedef SmallVector<PHINode*, 4> PhiVector; 172 /// When we unroll loops we have multiple vector values for each scalar. 173 /// This data structure holds the unrolled and vectorized values that 174 /// originated from one scalar instruction. 175 typedef SmallVector<Value*, 2> VectorParts; 176 177 /// Add code that checks at runtime if the accessed arrays overlap. 178 /// Returns the comparator value or NULL if no check is needed. 179 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal, 180 Instruction *Loc); 181 /// Create an empty loop, based on the loop ranges of the old loop. 182 void createEmptyLoop(LoopVectorizationLegality *Legal); 183 /// Copy and widen the instructions from the old loop. 184 void vectorizeLoop(LoopVectorizationLegality *Legal); 185 186 /// A helper function that computes the predicate of the block BB, assuming 187 /// that the header block of the loop is set to True. It returns the *entry* 188 /// mask for the block BB. 189 VectorParts createBlockInMask(BasicBlock *BB); 190 /// A helper function that computes the predicate of the edge between SRC 191 /// and DST. 192 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst); 193 194 /// A helper function to vectorize a single BB within the innermost loop. 195 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB, 196 PhiVector *PV); 197 198 /// Insert the new loop to the loop hierarchy and pass manager 199 /// and update the analysis passes. 200 void updateAnalysis(); 201 202 /// This instruction is un-vectorizable. Implement it as a sequence 203 /// of scalars. 204 void scalarizeInstruction(Instruction *Instr); 205 206 /// Vectorize Load and Store instructions, 207 void vectorizeMemoryInstruction(Instruction *Instr, 208 LoopVectorizationLegality *Legal); 209 210 /// Create a broadcast instruction. This method generates a broadcast 211 /// instruction (shuffle) for loop invariant values and for the induction 212 /// value. If this is the induction variable then we extend it to N, N+1, ... 213 /// this is needed because each iteration in the loop corresponds to a SIMD 214 /// element. 215 Value *getBroadcastInstrs(Value *V); 216 217 /// This function adds 0, 1, 2 ... to each vector element, starting at zero. 218 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...). 219 /// The sequence starts at StartIndex. 220 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate); 221 222 /// When we go over instructions in the basic block we rely on previous 223 /// values within the current basic block or on loop invariant values. 224 /// When we widen (vectorize) values we place them in the map. If the values 225 /// are not within the map, they have to be loop invariant, so we simply 226 /// broadcast them into a vector. 227 VectorParts &getVectorValue(Value *V); 228 229 /// Generate a shuffle sequence that will reverse the vector Vec. 230 Value *reverseVector(Value *Vec); 231 232 /// This is a helper class that holds the vectorizer state. It maps scalar 233 /// instructions to vector instructions. When the code is 'unrolled' then 234 /// then a single scalar value is mapped to multiple vector parts. The parts 235 /// are stored in the VectorPart type. 236 struct ValueMap { 237 /// C'tor. UnrollFactor controls the number of vectors ('parts') that 238 /// are mapped. 239 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {} 240 241 /// \return True if 'Key' is saved in the Value Map. 242 bool has(Value *Key) const { return MapStorage.count(Key); } 243 244 /// Initializes a new entry in the map. Sets all of the vector parts to the 245 /// save value in 'Val'. 246 /// \return A reference to a vector with splat values. 247 VectorParts &splat(Value *Key, Value *Val) { 248 VectorParts &Entry = MapStorage[Key]; 249 Entry.assign(UF, Val); 250 return Entry; 251 } 252 253 ///\return A reference to the value that is stored at 'Key'. 254 VectorParts &get(Value *Key) { 255 VectorParts &Entry = MapStorage[Key]; 256 if (Entry.empty()) 257 Entry.resize(UF); 258 assert(Entry.size() == UF); 259 return Entry; 260 } 261 262 private: 263 /// The unroll factor. Each entry in the map stores this number of vector 264 /// elements. 265 unsigned UF; 266 267 /// Map storage. We use std::map and not DenseMap because insertions to a 268 /// dense map invalidates its iterators. 269 std::map<Value *, VectorParts> MapStorage; 270 }; 271 272 /// The original loop. 273 Loop *OrigLoop; 274 /// Scev analysis to use. 275 ScalarEvolution *SE; 276 /// Loop Info. 277 LoopInfo *LI; 278 /// Dominator Tree. 279 DominatorTree *DT; 280 /// Data Layout. 281 DataLayout *DL; 282 /// Target Library Info. 283 const TargetLibraryInfo *TLI; 284 285 /// The vectorization SIMD factor to use. Each vector will have this many 286 /// vector elements. 287 unsigned VF; 288 /// The vectorization unroll factor to use. Each scalar is vectorized to this 289 /// many different vector instructions. 290 unsigned UF; 291 292 /// The builder that we use 293 IRBuilder<> Builder; 294 295 // --- Vectorization state --- 296 297 /// The vector-loop preheader. 298 BasicBlock *LoopVectorPreHeader; 299 /// The scalar-loop preheader. 300 BasicBlock *LoopScalarPreHeader; 301 /// Middle Block between the vector and the scalar. 302 BasicBlock *LoopMiddleBlock; 303 ///The ExitBlock of the scalar loop. 304 BasicBlock *LoopExitBlock; 305 ///The vector loop body. 306 BasicBlock *LoopVectorBody; 307 ///The scalar loop body. 308 BasicBlock *LoopScalarBody; 309 /// A list of all bypass blocks. The first block is the entry of the loop. 310 SmallVector<BasicBlock *, 4> LoopBypassBlocks; 311 312 /// The new Induction variable which was added to the new block. 313 PHINode *Induction; 314 /// The induction variable of the old basic block. 315 PHINode *OldInduction; 316 /// Holds the extended (to the widest induction type) start index. 317 Value *ExtendedIdx; 318 /// Maps scalars to widened vectors. 319 ValueMap WidenMap; 320}; 321 322/// \brief Check if conditionally executed loads are hoistable. 323/// 324/// This class has two functions: isHoistableLoad and canHoistAllLoads. 325/// isHoistableLoad should be called on all load instructions that are executed 326/// conditionally. After all conditional loads are processed, the client should 327/// call canHoistAllLoads to determine if all of the conditional executed loads 328/// have an unconditional memory access to the same memory address in the loop. 329class LoadHoisting { 330 typedef SmallPtrSet<Value *, 8> MemorySet; 331 332 Loop *TheLoop; 333 DominatorTree *DT; 334 MemorySet CondLoadAddrSet; 335 336public: 337 LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {} 338 339 /// \brief Check if the instruction is a load with a identifiable address. 340 bool isHoistableLoad(Instruction *L); 341 342 /// \brief Check if all of the conditional loads are hoistable because there 343 /// exists an unconditional memory access to the same address in the loop. 344 bool canHoistAllLoads(); 345}; 346 347bool LoadHoisting::isHoistableLoad(Instruction *L) { 348 LoadInst *LI = dyn_cast<LoadInst>(L); 349 if (!LI) 350 return false; 351 352 CondLoadAddrSet.insert(LI->getPointerOperand()); 353 return true; 354} 355 356static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) { 357 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) { 358 if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load. 359 Set.insert(LI->getPointerOperand()); 360 else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store. 361 Set.insert(SI->getPointerOperand()); 362 } 363} 364 365bool LoadHoisting::canHoistAllLoads() { 366 // No conditional loads. 367 if (CondLoadAddrSet.empty()) 368 return true; 369 370 MemorySet UncondMemAccesses; 371 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector(); 372 BasicBlock *LoopLatch = TheLoop->getLoopLatch(); 373 374 // Iterate over the unconditional blocks and collect memory access addresses. 375 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) { 376 BasicBlock *BB = LoopBlocks[i]; 377 378 // Ignore conditional blocks. 379 if (BB != LoopLatch && !DT->dominates(BB, LoopLatch)) 380 continue; 381 382 addMemAccesses(BB, UncondMemAccesses); 383 } 384 385 // And make sure there is a matching unconditional access for every 386 // conditional load. 387 for (MemorySet::iterator MI = CondLoadAddrSet.begin(), 388 ME = CondLoadAddrSet.end(); MI != ME; ++MI) 389 if (!UncondMemAccesses.count(*MI)) 390 return false; 391 392 return true; 393} 394 395/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and 396/// to what vectorization factor. 397/// This class does not look at the profitability of vectorization, only the 398/// legality. This class has two main kinds of checks: 399/// * Memory checks - The code in canVectorizeMemory checks if vectorization 400/// will change the order of memory accesses in a way that will change the 401/// correctness of the program. 402/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory 403/// checks for a number of different conditions, such as the availability of a 404/// single induction variable, that all types are supported and vectorize-able, 405/// etc. This code reflects the capabilities of InnerLoopVectorizer. 406/// This class is also used by InnerLoopVectorizer for identifying 407/// induction variable and the different reduction variables. 408class LoopVectorizationLegality { 409public: 410 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL, 411 DominatorTree *DT, TargetTransformInfo* TTI, 412 AliasAnalysis *AA, TargetLibraryInfo *TLI) 413 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI), 414 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false), 415 LoadSpeculation(L, DT) {} 416 417 /// This enum represents the kinds of reductions that we support. 418 enum ReductionKind { 419 RK_NoReduction, ///< Not a reduction. 420 RK_IntegerAdd, ///< Sum of integers. 421 RK_IntegerMult, ///< Product of integers. 422 RK_IntegerOr, ///< Bitwise or logical OR of numbers. 423 RK_IntegerAnd, ///< Bitwise or logical AND of numbers. 424 RK_IntegerXor, ///< Bitwise or logical XOR of numbers. 425 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()). 426 RK_FloatAdd, ///< Sum of floats. 427 RK_FloatMult, ///< Product of floats. 428 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()). 429 }; 430 431 /// This enum represents the kinds of inductions that we support. 432 enum InductionKind { 433 IK_NoInduction, ///< Not an induction variable. 434 IK_IntInduction, ///< Integer induction variable. Step = 1. 435 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1. 436 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem). 437 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem). 438 }; 439 440 // This enum represents the kind of minmax reduction. 441 enum MinMaxReductionKind { 442 MRK_Invalid, 443 MRK_UIntMin, 444 MRK_UIntMax, 445 MRK_SIntMin, 446 MRK_SIntMax, 447 MRK_FloatMin, 448 MRK_FloatMax 449 }; 450 451 /// This POD struct holds information about reduction variables. 452 struct ReductionDescriptor { 453 ReductionDescriptor() : StartValue(0), LoopExitInstr(0), 454 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {} 455 456 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K, 457 MinMaxReductionKind MK) 458 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {} 459 460 // The starting value of the reduction. 461 // It does not have to be zero! 462 TrackingVH<Value> StartValue; 463 // The instruction who's value is used outside the loop. 464 Instruction *LoopExitInstr; 465 // The kind of the reduction. 466 ReductionKind Kind; 467 // If this a min/max reduction the kind of reduction. 468 MinMaxReductionKind MinMaxKind; 469 }; 470 471 /// This POD struct holds information about a potential reduction operation. 472 struct ReductionInstDesc { 473 ReductionInstDesc(bool IsRedux, Instruction *I) : 474 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {} 475 476 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) : 477 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {} 478 479 // Is this instruction a reduction candidate. 480 bool IsReduction; 481 // The last instruction in a min/max pattern (select of the select(icmp()) 482 // pattern), or the current reduction instruction otherwise. 483 Instruction *PatternLastInst; 484 // If this is a min/max pattern the comparison predicate. 485 MinMaxReductionKind MinMaxKind; 486 }; 487 488 // This POD struct holds information about the memory runtime legality 489 // check that a group of pointers do not overlap. 490 struct RuntimePointerCheck { 491 RuntimePointerCheck() : Need(false) {} 492 493 /// Reset the state of the pointer runtime information. 494 void reset() { 495 Need = false; 496 Pointers.clear(); 497 Starts.clear(); 498 Ends.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); 503 504 /// This flag indicates if we need to add the runtime check. 505 bool Need; 506 /// Holds the pointers that we need to check. 507 SmallVector<TrackingVH<Value>, 2> Pointers; 508 /// Holds the pointer value at the beginning of the loop. 509 SmallVector<const SCEV*, 2> Starts; 510 /// Holds the pointer value at the end of the loop. 511 SmallVector<const SCEV*, 2> Ends; 512 /// Holds the information if this pointer is used for writing to memory. 513 SmallVector<bool, 2> IsWritePtr; 514 }; 515 516 /// A POD for saving information about induction variables. 517 struct InductionInfo { 518 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {} 519 InductionInfo() : StartValue(0), IK(IK_NoInduction) {} 520 /// Start value. 521 TrackingVH<Value> StartValue; 522 /// Induction kind. 523 InductionKind IK; 524 }; 525 526 /// ReductionList contains the reduction descriptors for all 527 /// of the reductions that were found in the loop. 528 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList; 529 530 /// InductionList saves induction variables and maps them to the 531 /// induction descriptor. 532 typedef MapVector<PHINode*, InductionInfo> InductionList; 533 534 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their 535 /// respective Store/Load instruction(s) to calculate aliasing. 536 typedef MapVector<Value*, Instruction* > AliasMap; 537 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap; 538 539 /// Returns true if it is legal to vectorize this loop. 540 /// This does not mean that it is profitable to vectorize this 541 /// loop, only that it is legal to do so. 542 bool canVectorize(); 543 544 /// Returns the Induction variable. 545 PHINode *getInduction() { return Induction; } 546 547 /// Returns the reduction variables found in the loop. 548 ReductionList *getReductionVars() { return &Reductions; } 549 550 /// Returns the induction variables found in the loop. 551 InductionList *getInductionVars() { return &Inductions; } 552 553 /// Returns the widest induction type. 554 Type *getWidestInductionType() { return WidestIndTy; } 555 556 /// Returns True if V is an induction variable in this loop. 557 bool isInductionVariable(const Value *V); 558 559 /// Return true if the block BB needs to be predicated in order for the loop 560 /// to be vectorized. 561 bool blockNeedsPredication(BasicBlock *BB); 562 563 /// Check if this pointer is consecutive when vectorizing. This happens 564 /// when the last index of the GEP is the induction variable, or that the 565 /// pointer itself is an induction variable. 566 /// This check allows us to vectorize A[idx] into a wide load/store. 567 /// Returns: 568 /// 0 - Stride is unknown or non consecutive. 569 /// 1 - Address is consecutive. 570 /// -1 - Address is consecutive, and decreasing. 571 int isConsecutivePtr(Value *Ptr); 572 573 /// Returns true if the value V is uniform within the loop. 574 bool isUniform(Value *V); 575 576 /// Returns true if this instruction will remain scalar after vectorization. 577 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } 578 579 /// Returns the information that we collected about runtime memory check. 580 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; } 581 582 /// This function returns the identity element (or neutral element) for 583 /// the operation K. 584 static Constant *getReductionIdentity(ReductionKind K, Type *Tp); 585private: 586 /// Check if a single basic block loop is vectorizable. 587 /// At this point we know that this is a loop with a constant trip count 588 /// and we only need to check individual instructions. 589 bool canVectorizeInstrs(); 590 591 /// When we vectorize loops we may change the order in which 592 /// we read and write from memory. This method checks if it is 593 /// legal to vectorize the code, considering only memory constrains. 594 /// Returns true if the loop is vectorizable 595 bool canVectorizeMemory(); 596 597 /// Return true if we can vectorize this loop using the IF-conversion 598 /// transformation. 599 bool canVectorizeWithIfConvert(); 600 601 /// Collect the variables that need to stay uniform after vectorization. 602 void collectLoopUniforms(); 603 604 /// Return true if all of the instructions in the block can be speculatively 605 /// executed. 606 bool blockCanBePredicated(BasicBlock *BB); 607 608 /// Returns True, if 'Phi' is the kind of reduction variable for type 609 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. 610 bool AddReductionVar(PHINode *Phi, ReductionKind Kind); 611 /// Returns a struct describing if the instruction 'I' can be a reduction 612 /// variable of type 'Kind'. If the reduction is a min/max pattern of 613 /// select(icmp()) this function advances the instruction pointer 'I' from the 614 /// compare instruction to the select instruction and stores this pointer in 615 /// 'PatternLastInst' member of the returned struct. 616 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind, 617 ReductionInstDesc &Desc); 618 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 619 /// pattern corresponding to a min(X, Y) or max(X, Y). 620 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I, 621 ReductionInstDesc &Prev); 622 /// Returns the induction kind of Phi. This function may return NoInduction 623 /// if the PHI is not an induction variable. 624 InductionKind isInductionVariable(PHINode *Phi); 625 /// Return true if can compute the address bounds of Ptr within the loop. 626 bool hasComputableBounds(Value *Ptr); 627 /// Return true if there is the chance of write reorder. 628 bool hasPossibleGlobalWriteReorder(Value *Object, 629 Instruction *Inst, 630 AliasMultiMap &WriteObjects, 631 unsigned MaxByteWidth); 632 /// Return the AA location for a load or a store. 633 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst); 634 635 636 /// The loop that we evaluate. 637 Loop *TheLoop; 638 /// Scev analysis. 639 ScalarEvolution *SE; 640 /// DataLayout analysis. 641 DataLayout *DL; 642 /// Dominators. 643 DominatorTree *DT; 644 /// Target Info. 645 TargetTransformInfo *TTI; 646 /// Alias Analysis. 647 AliasAnalysis *AA; 648 /// Target Library Info. 649 TargetLibraryInfo *TLI; 650 651 // --- vectorization state --- // 652 653 /// Holds the integer induction variable. This is the counter of the 654 /// loop. 655 PHINode *Induction; 656 /// Holds the reduction variables. 657 ReductionList Reductions; 658 /// Holds all of the induction variables that we found in the loop. 659 /// Notice that inductions don't need to start at zero and that induction 660 /// variables can be pointers. 661 InductionList Inductions; 662 /// Holds the widest induction type encountered. 663 Type *WidestIndTy; 664 665 /// Allowed outside users. This holds the reduction 666 /// vars which can be accessed from outside the loop. 667 SmallPtrSet<Value*, 4> AllowedExit; 668 /// This set holds the variables which are known to be uniform after 669 /// vectorization. 670 SmallPtrSet<Instruction*, 4> Uniforms; 671 /// We need to check that all of the pointers in this list are disjoint 672 /// at runtime. 673 RuntimePointerCheck PtrRtCheck; 674 /// Can we assume the absence of NaNs. 675 bool HasFunNoNaNAttr; 676 677 /// Utility to determine whether loads can be speculated. 678 LoadHoisting LoadSpeculation; 679}; 680 681/// LoopVectorizationCostModel - estimates the expected speedups due to 682/// vectorization. 683/// In many cases vectorization is not profitable. This can happen because of 684/// a number of reasons. In this class we mainly attempt to predict the 685/// expected speedup/slowdowns due to the supported instruction set. We use the 686/// TargetTransformInfo to query the different backends for the cost of 687/// different operations. 688class LoopVectorizationCostModel { 689public: 690 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI, 691 LoopVectorizationLegality *Legal, 692 const TargetTransformInfo &TTI, 693 DataLayout *DL, const TargetLibraryInfo *TLI) 694 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {} 695 696 /// Information about vectorization costs 697 struct VectorizationFactor { 698 unsigned Width; // Vector width with best cost 699 unsigned Cost; // Cost of the loop with that width 700 }; 701 /// \return The most profitable vectorization factor and the cost of that VF. 702 /// This method checks every power of two up to VF. If UserVF is not ZERO 703 /// then this vectorization factor will be selected if vectorization is 704 /// possible. 705 VectorizationFactor selectVectorizationFactor(bool OptForSize, 706 unsigned UserVF); 707 708 /// \return The size (in bits) of the widest type in the code that 709 /// needs to be vectorized. We ignore values that remain scalar such as 710 /// 64 bit loop indices. 711 unsigned getWidestType(); 712 713 /// \return The most profitable unroll factor. 714 /// If UserUF is non-zero then this method finds the best unroll-factor 715 /// based on register pressure and other parameters. 716 /// VF and LoopCost are the selected vectorization factor and the cost of the 717 /// selected VF. 718 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF, 719 unsigned LoopCost); 720 721 /// \brief A struct that represents some properties of the register usage 722 /// of a loop. 723 struct RegisterUsage { 724 /// Holds the number of loop invariant values that are used in the loop. 725 unsigned LoopInvariantRegs; 726 /// Holds the maximum number of concurrent live intervals in the loop. 727 unsigned MaxLocalUsers; 728 /// Holds the number of instructions in the loop. 729 unsigned NumInstructions; 730 }; 731 732 /// \return information about the register usage of the loop. 733 RegisterUsage calculateRegisterUsage(); 734 735private: 736 /// Returns the expected execution cost. The unit of the cost does 737 /// not matter because we use the 'cost' units to compare different 738 /// vector widths. The cost that is returned is *not* normalized by 739 /// the factor width. 740 unsigned expectedCost(unsigned VF); 741 742 /// Returns the execution time cost of an instruction for a given vector 743 /// width. Vector width of one means scalar. 744 unsigned getInstructionCost(Instruction *I, unsigned VF); 745 746 /// A helper function for converting Scalar types to vector types. 747 /// If the incoming type is void, we return void. If the VF is 1, we return 748 /// the scalar type. 749 static Type* ToVectorTy(Type *Scalar, unsigned VF); 750 751 /// Returns whether the instruction is a load or store and will be a emitted 752 /// as a vector operation. 753 bool isConsecutiveLoadOrStore(Instruction *I); 754 755 /// The loop that we evaluate. 756 Loop *TheLoop; 757 /// Scev analysis. 758 ScalarEvolution *SE; 759 /// Loop Info analysis. 760 LoopInfo *LI; 761 /// Vectorization legality. 762 LoopVectorizationLegality *Legal; 763 /// Vector target information. 764 const TargetTransformInfo &TTI; 765 /// Target data layout information. 766 DataLayout *DL; 767 /// Target Library Info. 768 const TargetLibraryInfo *TLI; 769}; 770 771/// The LoopVectorize Pass. 772struct LoopVectorize : public LoopPass { 773 /// Pass identification, replacement for typeid 774 static char ID; 775 776 explicit LoopVectorize() : LoopPass(ID) { 777 initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); 778 } 779 780 ScalarEvolution *SE; 781 DataLayout *DL; 782 LoopInfo *LI; 783 TargetTransformInfo *TTI; 784 DominatorTree *DT; 785 AliasAnalysis *AA; 786 TargetLibraryInfo *TLI; 787 788 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) { 789 // We only vectorize innermost loops. 790 if (!L->empty()) 791 return false; 792 793 SE = &getAnalysis<ScalarEvolution>(); 794 DL = getAnalysisIfAvailable<DataLayout>(); 795 LI = &getAnalysis<LoopInfo>(); 796 TTI = &getAnalysis<TargetTransformInfo>(); 797 DT = &getAnalysis<DominatorTree>(); 798 AA = getAnalysisIfAvailable<AliasAnalysis>(); 799 TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); 800 801 if (DL == NULL) { 802 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout"); 803 return false; 804 } 805 806 DEBUG(dbgs() << "LV: Checking a loop in \"" << 807 L->getHeader()->getParent()->getName() << "\"\n"); 808 809 // Check if it is legal to vectorize the loop. 810 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI); 811 if (!LVL.canVectorize()) { 812 DEBUG(dbgs() << "LV: Not vectorizing.\n"); 813 return false; 814 } 815 816 // Use the cost model. 817 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI); 818 819 // Check the function attributes to find out if this function should be 820 // optimized for size. 821 Function *F = L->getHeader()->getParent(); 822 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize; 823 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat; 824 unsigned FnIndex = AttributeSet::FunctionIndex; 825 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr); 826 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr); 827 828 if (NoFloat) { 829 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" 830 "attribute is used.\n"); 831 return false; 832 } 833 834 // Select the optimal vectorization factor. 835 LoopVectorizationCostModel::VectorizationFactor VF; 836 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor); 837 // Select the unroll factor. 838 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll, 839 VF.Width, VF.Cost); 840 841 if (VF.Width == 1) { 842 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); 843 return false; 844 } 845 846 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<< 847 F->getParent()->getModuleIdentifier()<<"\n"); 848 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n"); 849 850 // If we decided that it is *legal* to vectorize the loop then do it. 851 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF); 852 LB.vectorize(&LVL); 853 854 DEBUG(verifyFunction(*L->getHeader()->getParent())); 855 return true; 856 } 857 858 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 859 LoopPass::getAnalysisUsage(AU); 860 AU.addRequiredID(LoopSimplifyID); 861 AU.addRequiredID(LCSSAID); 862 AU.addRequired<DominatorTree>(); 863 AU.addRequired<LoopInfo>(); 864 AU.addRequired<ScalarEvolution>(); 865 AU.addRequired<TargetTransformInfo>(); 866 AU.addPreserved<LoopInfo>(); 867 AU.addPreserved<DominatorTree>(); 868 } 869 870}; 871 872} // end anonymous namespace 873 874//===----------------------------------------------------------------------===// 875// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and 876// LoopVectorizationCostModel. 877//===----------------------------------------------------------------------===// 878 879void 880LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE, 881 Loop *Lp, Value *Ptr, 882 bool WritePtr) { 883 const SCEV *Sc = SE->getSCEV(Ptr); 884 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); 885 assert(AR && "Invalid addrec expression"); 886 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch()); 887 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE); 888 Pointers.push_back(Ptr); 889 Starts.push_back(AR->getStart()); 890 Ends.push_back(ScEnd); 891 IsWritePtr.push_back(WritePtr); 892} 893 894Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { 895 // Save the current insertion location. 896 Instruction *Loc = Builder.GetInsertPoint(); 897 898 // We need to place the broadcast of invariant variables outside the loop. 899 Instruction *Instr = dyn_cast<Instruction>(V); 900 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody); 901 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr; 902 903 // Place the code for broadcasting invariant variables in the new preheader. 904 if (Invariant) 905 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); 906 907 // Broadcast the scalar into all locations in the vector. 908 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); 909 910 // Restore the builder insertion point. 911 if (Invariant) 912 Builder.SetInsertPoint(Loc); 913 914 return Shuf; 915} 916 917Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx, 918 bool Negate) { 919 assert(Val->getType()->isVectorTy() && "Must be a vector"); 920 assert(Val->getType()->getScalarType()->isIntegerTy() && 921 "Elem must be an integer"); 922 // Create the types. 923 Type *ITy = Val->getType()->getScalarType(); 924 VectorType *Ty = cast<VectorType>(Val->getType()); 925 int VLen = Ty->getNumElements(); 926 SmallVector<Constant*, 8> Indices; 927 928 // Create a vector of consecutive numbers from zero to VF. 929 for (int i = 0; i < VLen; ++i) { 930 int64_t Idx = Negate ? (-i) : i; 931 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate)); 932 } 933 934 // Add the consecutive indices to the vector value. 935 Constant *Cv = ConstantVector::get(Indices); 936 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); 937 return Builder.CreateAdd(Val, Cv, "induction"); 938} 939 940int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { 941 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr"); 942 // Make sure that the pointer does not point to structs. 943 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) 944 return 0; 945 946 // If this value is a pointer induction variable we know it is consecutive. 947 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr); 948 if (Phi && Inductions.count(Phi)) { 949 InductionInfo II = Inductions[Phi]; 950 if (IK_PtrInduction == II.IK) 951 return 1; 952 else if (IK_ReversePtrInduction == II.IK) 953 return -1; 954 } 955 956 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr); 957 if (!Gep) 958 return 0; 959 960 unsigned NumOperands = Gep->getNumOperands(); 961 Value *LastIndex = Gep->getOperand(NumOperands - 1); 962 963 Value *GpPtr = Gep->getPointerOperand(); 964 // If this GEP value is a consecutive pointer induction variable and all of 965 // the indices are constant then we know it is consecutive. We can 966 Phi = dyn_cast<PHINode>(GpPtr); 967 if (Phi && Inductions.count(Phi)) { 968 969 // Make sure that the pointer does not point to structs. 970 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType()); 971 if (GepPtrType->getElementType()->isAggregateType()) 972 return 0; 973 974 // Make sure that all of the index operands are loop invariant. 975 for (unsigned i = 1; i < NumOperands; ++i) 976 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 977 return 0; 978 979 InductionInfo II = Inductions[Phi]; 980 if (IK_PtrInduction == II.IK) 981 return 1; 982 else if (IK_ReversePtrInduction == II.IK) 983 return -1; 984 } 985 986 // Check that all of the gep indices are uniform except for the last. 987 for (unsigned i = 0; i < NumOperands - 1; ++i) 988 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 989 return 0; 990 991 // We can emit wide load/stores only if the last index is the induction 992 // variable. 993 const SCEV *Last = SE->getSCEV(LastIndex); 994 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { 995 const SCEV *Step = AR->getStepRecurrence(*SE); 996 997 // The memory is consecutive because the last index is consecutive 998 // and all other indices are loop invariant. 999 if (Step->isOne()) 1000 return 1; 1001 if (Step->isAllOnesValue()) 1002 return -1; 1003 } 1004 1005 return 0; 1006} 1007 1008bool LoopVectorizationLegality::isUniform(Value *V) { 1009 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); 1010} 1011 1012InnerLoopVectorizer::VectorParts& 1013InnerLoopVectorizer::getVectorValue(Value *V) { 1014 assert(V != Induction && "The new induction variable should not be used."); 1015 assert(!V->getType()->isVectorTy() && "Can't widen a vector"); 1016 1017 // If we have this scalar in the map, return it. 1018 if (WidenMap.has(V)) 1019 return WidenMap.get(V); 1020 1021 // If this scalar is unknown, assume that it is a constant or that it is 1022 // loop invariant. Broadcast V and save the value for future uses. 1023 Value *B = getBroadcastInstrs(V); 1024 return WidenMap.splat(V, B); 1025} 1026 1027Value *InnerLoopVectorizer::reverseVector(Value *Vec) { 1028 assert(Vec->getType()->isVectorTy() && "Invalid type"); 1029 SmallVector<Constant*, 8> ShuffleMask; 1030 for (unsigned i = 0; i < VF; ++i) 1031 ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); 1032 1033 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), 1034 ConstantVector::get(ShuffleMask), 1035 "reverse"); 1036} 1037 1038 1039void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr, 1040 LoopVectorizationLegality *Legal) { 1041 // Attempt to issue a wide load. 1042 LoadInst *LI = dyn_cast<LoadInst>(Instr); 1043 StoreInst *SI = dyn_cast<StoreInst>(Instr); 1044 1045 assert((LI || SI) && "Invalid Load/Store instruction"); 1046 1047 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType(); 1048 Type *DataTy = VectorType::get(ScalarDataTy, VF); 1049 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 1050 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment(); 1051 1052 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy); 1053 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF; 1054 1055 if (ScalarAllocatedSize != VectorElementSize) 1056 return scalarizeInstruction(Instr); 1057 1058 // If the pointer is loop invariant or if it is non consecutive, 1059 // scalarize the load. 1060 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 1061 bool Reverse = ConsecutiveStride < 0; 1062 bool UniformLoad = LI && Legal->isUniform(Ptr); 1063 if (!ConsecutiveStride || UniformLoad) 1064 return scalarizeInstruction(Instr); 1065 1066 Constant *Zero = Builder.getInt32(0); 1067 VectorParts &Entry = WidenMap.get(Instr); 1068 1069 // Handle consecutive loads/stores. 1070 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 1071 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) { 1072 Value *PtrOperand = Gep->getPointerOperand(); 1073 Value *FirstBasePtr = getVectorValue(PtrOperand)[0]; 1074 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero); 1075 1076 // Create the new GEP with the new induction variable. 1077 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1078 Gep2->setOperand(0, FirstBasePtr); 1079 Gep2->setName("gep.indvar.base"); 1080 Ptr = Builder.Insert(Gep2); 1081 } else if (Gep) { 1082 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), 1083 OrigLoop) && "Base ptr must be invariant"); 1084 1085 // The last index does not have to be the induction. It can be 1086 // consecutive and be a function of the index. For example A[I+1]; 1087 unsigned NumOperands = Gep->getNumOperands(); 1088 1089 Value *LastGepOperand = Gep->getOperand(NumOperands - 1); 1090 VectorParts &GEPParts = getVectorValue(LastGepOperand); 1091 Value *LastIndex = GEPParts[0]; 1092 LastIndex = Builder.CreateExtractElement(LastIndex, Zero); 1093 1094 // Create the new GEP with the new induction variable. 1095 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1096 Gep2->setOperand(NumOperands - 1, LastIndex); 1097 Gep2->setName("gep.indvar.idx"); 1098 Ptr = Builder.Insert(Gep2); 1099 } else { 1100 // Use the induction element ptr. 1101 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 1102 VectorParts &PtrVal = getVectorValue(Ptr); 1103 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 1104 } 1105 1106 // Handle Stores: 1107 if (SI) { 1108 assert(!Legal->isUniform(SI->getPointerOperand()) && 1109 "We do not allow storing to uniform addresses"); 1110 1111 VectorParts &StoredVal = getVectorValue(SI->getValueOperand()); 1112 for (unsigned Part = 0; Part < UF; ++Part) { 1113 // Calculate the pointer for the specific unroll-part. 1114 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1115 1116 if (Reverse) { 1117 // If we store to reverse consecutive memory locations then we need 1118 // to reverse the order of elements in the stored value. 1119 StoredVal[Part] = reverseVector(StoredVal[Part]); 1120 // If the address is consecutive but reversed, then the 1121 // wide store needs to start at the last vector element. 1122 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1123 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1124 } 1125 1126 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); 1127 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment); 1128 } 1129 } 1130 1131 for (unsigned Part = 0; Part < UF; ++Part) { 1132 // Calculate the pointer for the specific unroll-part. 1133 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1134 1135 if (Reverse) { 1136 // If the address is consecutive but reversed, then the 1137 // wide store needs to start at the last vector element. 1138 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1139 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1140 } 1141 1142 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); 1143 Value *LI = Builder.CreateLoad(VecPtr, "wide.load"); 1144 cast<LoadInst>(LI)->setAlignment(Alignment); 1145 Entry[Part] = Reverse ? reverseVector(LI) : LI; 1146 } 1147} 1148 1149void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { 1150 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 1151 // Holds vector parameters or scalars, in case of uniform vals. 1152 SmallVector<VectorParts, 4> Params; 1153 1154 // Find all of the vectorized parameters. 1155 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1156 Value *SrcOp = Instr->getOperand(op); 1157 1158 // If we are accessing the old induction variable, use the new one. 1159 if (SrcOp == OldInduction) { 1160 Params.push_back(getVectorValue(SrcOp)); 1161 continue; 1162 } 1163 1164 // Try using previously calculated values. 1165 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 1166 1167 // If the src is an instruction that appeared earlier in the basic block 1168 // then it should already be vectorized. 1169 if (SrcInst && OrigLoop->contains(SrcInst)) { 1170 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 1171 // The parameter is a vector value from earlier. 1172 Params.push_back(WidenMap.get(SrcInst)); 1173 } else { 1174 // The parameter is a scalar from outside the loop. Maybe even a constant. 1175 VectorParts Scalars; 1176 Scalars.append(UF, SrcOp); 1177 Params.push_back(Scalars); 1178 } 1179 } 1180 1181 assert(Params.size() == Instr->getNumOperands() && 1182 "Invalid number of operands"); 1183 1184 // Does this instruction return a value ? 1185 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 1186 1187 Value *UndefVec = IsVoidRetTy ? 0 : 1188 UndefValue::get(VectorType::get(Instr->getType(), VF)); 1189 // Create a new entry in the WidenMap and initialize it to Undef or Null. 1190 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 1191 1192 // For each vector unroll 'part': 1193 for (unsigned Part = 0; Part < UF; ++Part) { 1194 // For each scalar that we create: 1195 for (unsigned Width = 0; Width < VF; ++Width) { 1196 Instruction *Cloned = Instr->clone(); 1197 if (!IsVoidRetTy) 1198 Cloned->setName(Instr->getName() + ".cloned"); 1199 // Replace the operands of the cloned instrucions with extracted scalars. 1200 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1201 Value *Op = Params[op][Part]; 1202 // Param is a vector. Need to extract the right lane. 1203 if (Op->getType()->isVectorTy()) 1204 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); 1205 Cloned->setOperand(op, Op); 1206 } 1207 1208 // Place the cloned scalar in the new loop. 1209 Builder.Insert(Cloned); 1210 1211 // If the original scalar returns a value we need to place it in a vector 1212 // so that future users will be able to use it. 1213 if (!IsVoidRetTy) 1214 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, 1215 Builder.getInt32(Width)); 1216 } 1217 } 1218} 1219 1220Instruction * 1221InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, 1222 Instruction *Loc) { 1223 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck = 1224 Legal->getRuntimePointerCheck(); 1225 1226 if (!PtrRtCheck->Need) 1227 return NULL; 1228 1229 Instruction *MemoryRuntimeCheck = 0; 1230 unsigned NumPointers = PtrRtCheck->Pointers.size(); 1231 SmallVector<Value* , 2> Starts; 1232 SmallVector<Value* , 2> Ends; 1233 1234 SCEVExpander Exp(*SE, "induction"); 1235 1236 // Use this type for pointer arithmetic. 1237 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0); 1238 1239 for (unsigned i = 0; i < NumPointers; ++i) { 1240 Value *Ptr = PtrRtCheck->Pointers[i]; 1241 const SCEV *Sc = SE->getSCEV(Ptr); 1242 1243 if (SE->isLoopInvariant(Sc, OrigLoop)) { 1244 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" << 1245 *Ptr <<"\n"); 1246 Starts.push_back(Ptr); 1247 Ends.push_back(Ptr); 1248 } else { 1249 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n"); 1250 1251 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc); 1252 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc); 1253 Starts.push_back(Start); 1254 Ends.push_back(End); 1255 } 1256 } 1257 1258 IRBuilder<> ChkBuilder(Loc); 1259 1260 for (unsigned i = 0; i < NumPointers; ++i) { 1261 for (unsigned j = i+1; j < NumPointers; ++j) { 1262 // No need to check if two readonly pointers intersect. 1263 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j]) 1264 continue; 1265 1266 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc"); 1267 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc"); 1268 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc"); 1269 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc"); 1270 1271 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0"); 1272 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1"); 1273 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict"); 1274 if (MemoryRuntimeCheck) 1275 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, 1276 "conflict.rdx"); 1277 1278 MemoryRuntimeCheck = cast<Instruction>(IsConflict); 1279 } 1280 } 1281 1282 return MemoryRuntimeCheck; 1283} 1284 1285void 1286InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { 1287 /* 1288 In this function we generate a new loop. The new loop will contain 1289 the vectorized instructions while the old loop will continue to run the 1290 scalar remainder. 1291 1292 [ ] <-- vector loop bypass (may consist of multiple blocks). 1293 / | 1294 / v 1295 | [ ] <-- vector pre header. 1296 | | 1297 | v 1298 | [ ] \ 1299 | [ ]_| <-- vector loop. 1300 | | 1301 \ v 1302 >[ ] <--- middle-block. 1303 / | 1304 / v 1305 | [ ] <--- new preheader. 1306 | | 1307 | v 1308 | [ ] \ 1309 | [ ]_| <-- old scalar loop to handle remainder. 1310 \ | 1311 \ v 1312 >[ ] <-- exit block. 1313 ... 1314 */ 1315 1316 BasicBlock *OldBasicBlock = OrigLoop->getHeader(); 1317 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader(); 1318 BasicBlock *ExitBlock = OrigLoop->getExitBlock(); 1319 assert(ExitBlock && "Must have an exit block"); 1320 1321 // Mark the old scalar loop with metadata that tells us not to vectorize this 1322 // loop again if we run into it. 1323 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None); 1324 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD); 1325 1326 // Some loops have a single integer induction variable, while other loops 1327 // don't. One example is c++ iterators that often have multiple pointer 1328 // induction variables. In the code below we also support a case where we 1329 // don't have a single induction variable. 1330 OldInduction = Legal->getInduction(); 1331 Type *IdxTy = Legal->getWidestInductionType(); 1332 1333 // Find the loop boundaries. 1334 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch()); 1335 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count"); 1336 1337 // Get the total trip count from the count by adding 1. 1338 ExitCount = SE->getAddExpr(ExitCount, 1339 SE->getConstant(ExitCount->getType(), 1)); 1340 1341 // Expand the trip count and place the new instructions in the preheader. 1342 // Notice that the pre-header does not change, only the loop body. 1343 SCEVExpander Exp(*SE, "induction"); 1344 1345 // Count holds the overall loop count (N). 1346 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 1347 BypassBlock->getTerminator()); 1348 1349 // The loop index does not have to start at Zero. Find the original start 1350 // value from the induction PHI node. If we don't have an induction variable 1351 // then we know that it starts at zero. 1352 Builder.SetInsertPoint(BypassBlock->getTerminator()); 1353 Value *StartIdx = ExtendedIdx = OldInduction ? 1354 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock), 1355 IdxTy): 1356 ConstantInt::get(IdxTy, 0); 1357 1358 assert(BypassBlock && "Invalid loop structure"); 1359 LoopBypassBlocks.push_back(BypassBlock); 1360 1361 // Split the single block loop into the two loop structure described above. 1362 BasicBlock *VectorPH = 1363 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); 1364 BasicBlock *VecBody = 1365 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 1366 BasicBlock *MiddleBlock = 1367 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 1368 BasicBlock *ScalarPH = 1369 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 1370 1371 // Use this IR builder to create the loop instructions (Phi, Br, Cmp) 1372 // inside the loop. 1373 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1374 1375 // Generate the induction variable. 1376 Induction = Builder.CreatePHI(IdxTy, 2, "index"); 1377 // The loop step is equal to the vectorization factor (num of SIMD elements) 1378 // times the unroll factor (num of SIMD instructions). 1379 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 1380 1381 // This is the IR builder that we use to add all of the logic for bypassing 1382 // the new vector loop. 1383 IRBuilder<> BypassBuilder(BypassBlock->getTerminator()); 1384 1385 // We may need to extend the index in case there is a type mismatch. 1386 // We know that the count starts at zero and does not overflow. 1387 if (Count->getType() != IdxTy) { 1388 // The exit count can be of pointer type. Convert it to the correct 1389 // integer type. 1390 if (ExitCount->getType()->isPointerTy()) 1391 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int"); 1392 else 1393 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast"); 1394 } 1395 1396 // Add the start index to the loop count to get the new end index. 1397 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx"); 1398 1399 // Now we need to generate the expression for N - (N % VF), which is 1400 // the part that the vectorized body will execute. 1401 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf"); 1402 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec"); 1403 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx, 1404 "end.idx.rnd.down"); 1405 1406 // Now, compare the new count to zero. If it is zero skip the vector loop and 1407 // jump to the scalar loop. 1408 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, 1409 "cmp.zero"); 1410 1411 BasicBlock *LastBypassBlock = BypassBlock; 1412 1413 // Generate the code that checks in runtime if arrays overlap. We put the 1414 // checks into a separate block to make the more common case of few elements 1415 // faster. 1416 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal, 1417 BypassBlock->getTerminator()); 1418 if (MemRuntimeCheck) { 1419 // Create a new block containing the memory check. 1420 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck, 1421 "vector.memcheck"); 1422 LoopBypassBlocks.push_back(CheckBlock); 1423 1424 // Replace the branch into the memory check block with a conditional branch 1425 // for the "few elements case". 1426 Instruction *OldTerm = BypassBlock->getTerminator(); 1427 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); 1428 OldTerm->eraseFromParent(); 1429 1430 Cmp = MemRuntimeCheck; 1431 LastBypassBlock = CheckBlock; 1432 } 1433 1434 LastBypassBlock->getTerminator()->eraseFromParent(); 1435 BranchInst::Create(MiddleBlock, VectorPH, Cmp, 1436 LastBypassBlock); 1437 1438 // We are going to resume the execution of the scalar loop. 1439 // Go over all of the induction variables that we found and fix the 1440 // PHIs that are left in the scalar version of the loop. 1441 // The starting values of PHI nodes depend on the counter of the last 1442 // iteration in the vectorized loop. 1443 // If we come from a bypass edge then we need to start from the original 1444 // start value. 1445 1446 // This variable saves the new starting index for the scalar loop. 1447 PHINode *ResumeIndex = 0; 1448 LoopVectorizationLegality::InductionList::iterator I, E; 1449 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 1450 // Set builder to point to last bypass block. 1451 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator()); 1452 for (I = List->begin(), E = List->end(); I != E; ++I) { 1453 PHINode *OrigPhi = I->first; 1454 LoopVectorizationLegality::InductionInfo II = I->second; 1455 1456 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType(); 1457 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val", 1458 MiddleBlock->getTerminator()); 1459 // We might have extended the type of the induction variable but we need a 1460 // truncated version for the scalar loop. 1461 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ? 1462 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val", 1463 MiddleBlock->getTerminator()) : 0; 1464 1465 Value *EndValue = 0; 1466 switch (II.IK) { 1467 case LoopVectorizationLegality::IK_NoInduction: 1468 llvm_unreachable("Unknown induction"); 1469 case LoopVectorizationLegality::IK_IntInduction: { 1470 // Handle the integer induction counter. 1471 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); 1472 1473 // We have the canonical induction variable. 1474 if (OrigPhi == OldInduction) { 1475 // Create a truncated version of the resume value for the scalar loop, 1476 // we might have promoted the type to a larger width. 1477 EndValue = 1478 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType()); 1479 // The new PHI merges the original incoming value, in case of a bypass, 1480 // or the value at the end of the vectorized loop. 1481 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1482 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1483 TruncResumeVal->addIncoming(EndValue, VecBody); 1484 1485 // We know what the end value is. 1486 EndValue = IdxEndRoundDown; 1487 // We also know which PHI node holds it. 1488 ResumeIndex = ResumeVal; 1489 break; 1490 } 1491 1492 // Not the canonical induction variable - add the vector loop count to the 1493 // start value. 1494 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, 1495 II.StartValue->getType(), 1496 "cast.crd"); 1497 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end"); 1498 break; 1499 } 1500 case LoopVectorizationLegality::IK_ReverseIntInduction: { 1501 // Convert the CountRoundDown variable to the PHI size. 1502 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown, 1503 II.StartValue->getType(), 1504 "cast.crd"); 1505 // Handle reverse integer induction counter. 1506 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end"); 1507 break; 1508 } 1509 case LoopVectorizationLegality::IK_PtrInduction: { 1510 // For pointer induction variables, calculate the offset using 1511 // the end index. 1512 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown, 1513 "ptr.ind.end"); 1514 break; 1515 } 1516 case LoopVectorizationLegality::IK_ReversePtrInduction: { 1517 // The value at the end of the loop for the reverse pointer is calculated 1518 // by creating a GEP with a negative index starting from the start value. 1519 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0); 1520 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown, 1521 "rev.ind.end"); 1522 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx, 1523 "rev.ptr.ind.end"); 1524 break; 1525 } 1526 }// end of case 1527 1528 // The new PHI merges the original incoming value, in case of a bypass, 1529 // or the value at the end of the vectorized loop. 1530 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) { 1531 if (OrigPhi == OldInduction) 1532 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]); 1533 else 1534 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1535 } 1536 ResumeVal->addIncoming(EndValue, VecBody); 1537 1538 // Fix the scalar body counter (PHI node). 1539 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 1540 // The old inductions phi node in the scalar body needs the truncated value. 1541 if (OrigPhi == OldInduction) 1542 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal); 1543 else 1544 OrigPhi->setIncomingValue(BlockIdx, ResumeVal); 1545 } 1546 1547 // If we are generating a new induction variable then we also need to 1548 // generate the code that calculates the exit value. This value is not 1549 // simply the end of the counter because we may skip the vectorized body 1550 // in case of a runtime check. 1551 if (!OldInduction){ 1552 assert(!ResumeIndex && "Unexpected resume value found"); 1553 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", 1554 MiddleBlock->getTerminator()); 1555 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1556 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]); 1557 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); 1558 } 1559 1560 // Make sure that we found the index where scalar loop needs to continue. 1561 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() && 1562 "Invalid resume Index"); 1563 1564 // Add a check in the middle block to see if we have completed 1565 // all of the iterations in the first vector loop. 1566 // If (N - N%VF) == N, then we *don't* need to run the remainder. 1567 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, 1568 ResumeIndex, "cmp.n", 1569 MiddleBlock->getTerminator()); 1570 1571 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); 1572 // Remove the old terminator. 1573 MiddleBlock->getTerminator()->eraseFromParent(); 1574 1575 // Create i+1 and fill the PHINode. 1576 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); 1577 Induction->addIncoming(StartIdx, VectorPH); 1578 Induction->addIncoming(NextIdx, VecBody); 1579 // Create the compare. 1580 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); 1581 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); 1582 1583 // Now we have two terminators. Remove the old one from the block. 1584 VecBody->getTerminator()->eraseFromParent(); 1585 1586 // Get ready to start creating new instructions into the vectorized body. 1587 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1588 1589 // Create and register the new vector loop. 1590 Loop* Lp = new Loop(); 1591 Loop *ParentLoop = OrigLoop->getParentLoop(); 1592 1593 // Insert the new loop into the loop nest and register the new basic blocks. 1594 if (ParentLoop) { 1595 ParentLoop->addChildLoop(Lp); 1596 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 1597 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase()); 1598 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); 1599 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); 1600 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); 1601 } else { 1602 LI->addTopLevelLoop(Lp); 1603 } 1604 1605 Lp->addBasicBlockToLoop(VecBody, LI->getBase()); 1606 1607 // Save the state. 1608 LoopVectorPreHeader = VectorPH; 1609 LoopScalarPreHeader = ScalarPH; 1610 LoopMiddleBlock = MiddleBlock; 1611 LoopExitBlock = ExitBlock; 1612 LoopVectorBody = VecBody; 1613 LoopScalarBody = OldBasicBlock; 1614} 1615 1616/// This function returns the identity element (or neutral element) for 1617/// the operation K. 1618Constant* 1619LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) { 1620 switch (K) { 1621 case RK_IntegerXor: 1622 case RK_IntegerAdd: 1623 case RK_IntegerOr: 1624 // Adding, Xoring, Oring zero to a number does not change it. 1625 return ConstantInt::get(Tp, 0); 1626 case RK_IntegerMult: 1627 // Multiplying a number by 1 does not change it. 1628 return ConstantInt::get(Tp, 1); 1629 case RK_IntegerAnd: 1630 // AND-ing a number with an all-1 value does not change it. 1631 return ConstantInt::get(Tp, -1, true); 1632 case RK_FloatMult: 1633 // Multiplying a number by 1 does not change it. 1634 return ConstantFP::get(Tp, 1.0L); 1635 case RK_FloatAdd: 1636 // Adding zero to a number does not change it. 1637 return ConstantFP::get(Tp, 0.0L); 1638 default: 1639 llvm_unreachable("Unknown reduction kind"); 1640 } 1641} 1642 1643static Intrinsic::ID 1644getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) { 1645 // If we have an intrinsic call, check if it is trivially vectorizable. 1646 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) { 1647 switch (II->getIntrinsicID()) { 1648 case Intrinsic::sqrt: 1649 case Intrinsic::sin: 1650 case Intrinsic::cos: 1651 case Intrinsic::exp: 1652 case Intrinsic::exp2: 1653 case Intrinsic::log: 1654 case Intrinsic::log10: 1655 case Intrinsic::log2: 1656 case Intrinsic::fabs: 1657 case Intrinsic::floor: 1658 case Intrinsic::ceil: 1659 case Intrinsic::trunc: 1660 case Intrinsic::rint: 1661 case Intrinsic::nearbyint: 1662 case Intrinsic::pow: 1663 case Intrinsic::fma: 1664 case Intrinsic::fmuladd: 1665 return II->getIntrinsicID(); 1666 default: 1667 return Intrinsic::not_intrinsic; 1668 } 1669 } 1670 1671 if (!TLI) 1672 return Intrinsic::not_intrinsic; 1673 1674 LibFunc::Func Func; 1675 Function *F = CI->getCalledFunction(); 1676 // We're going to make assumptions on the semantics of the functions, check 1677 // that the target knows that it's available in this environment. 1678 if (!F || !TLI->getLibFunc(F->getName(), Func)) 1679 return Intrinsic::not_intrinsic; 1680 1681 // Otherwise check if we have a call to a function that can be turned into a 1682 // vector intrinsic. 1683 switch (Func) { 1684 default: 1685 break; 1686 case LibFunc::sin: 1687 case LibFunc::sinf: 1688 case LibFunc::sinl: 1689 return Intrinsic::sin; 1690 case LibFunc::cos: 1691 case LibFunc::cosf: 1692 case LibFunc::cosl: 1693 return Intrinsic::cos; 1694 case LibFunc::exp: 1695 case LibFunc::expf: 1696 case LibFunc::expl: 1697 return Intrinsic::exp; 1698 case LibFunc::exp2: 1699 case LibFunc::exp2f: 1700 case LibFunc::exp2l: 1701 return Intrinsic::exp2; 1702 case LibFunc::log: 1703 case LibFunc::logf: 1704 case LibFunc::logl: 1705 return Intrinsic::log; 1706 case LibFunc::log10: 1707 case LibFunc::log10f: 1708 case LibFunc::log10l: 1709 return Intrinsic::log10; 1710 case LibFunc::log2: 1711 case LibFunc::log2f: 1712 case LibFunc::log2l: 1713 return Intrinsic::log2; 1714 case LibFunc::fabs: 1715 case LibFunc::fabsf: 1716 case LibFunc::fabsl: 1717 return Intrinsic::fabs; 1718 case LibFunc::floor: 1719 case LibFunc::floorf: 1720 case LibFunc::floorl: 1721 return Intrinsic::floor; 1722 case LibFunc::ceil: 1723 case LibFunc::ceilf: 1724 case LibFunc::ceill: 1725 return Intrinsic::ceil; 1726 case LibFunc::trunc: 1727 case LibFunc::truncf: 1728 case LibFunc::truncl: 1729 return Intrinsic::trunc; 1730 case LibFunc::rint: 1731 case LibFunc::rintf: 1732 case LibFunc::rintl: 1733 return Intrinsic::rint; 1734 case LibFunc::nearbyint: 1735 case LibFunc::nearbyintf: 1736 case LibFunc::nearbyintl: 1737 return Intrinsic::nearbyint; 1738 case LibFunc::pow: 1739 case LibFunc::powf: 1740 case LibFunc::powl: 1741 return Intrinsic::pow; 1742 } 1743 1744 return Intrinsic::not_intrinsic; 1745} 1746 1747/// This function translates the reduction kind to an LLVM binary operator. 1748static unsigned 1749getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) { 1750 switch (Kind) { 1751 case LoopVectorizationLegality::RK_IntegerAdd: 1752 return Instruction::Add; 1753 case LoopVectorizationLegality::RK_IntegerMult: 1754 return Instruction::Mul; 1755 case LoopVectorizationLegality::RK_IntegerOr: 1756 return Instruction::Or; 1757 case LoopVectorizationLegality::RK_IntegerAnd: 1758 return Instruction::And; 1759 case LoopVectorizationLegality::RK_IntegerXor: 1760 return Instruction::Xor; 1761 case LoopVectorizationLegality::RK_FloatMult: 1762 return Instruction::FMul; 1763 case LoopVectorizationLegality::RK_FloatAdd: 1764 return Instruction::FAdd; 1765 case LoopVectorizationLegality::RK_IntegerMinMax: 1766 return Instruction::ICmp; 1767 case LoopVectorizationLegality::RK_FloatMinMax: 1768 return Instruction::FCmp; 1769 default: 1770 llvm_unreachable("Unknown reduction operation"); 1771 } 1772} 1773 1774Value *createMinMaxOp(IRBuilder<> &Builder, 1775 LoopVectorizationLegality::MinMaxReductionKind RK, 1776 Value *Left, 1777 Value *Right) { 1778 CmpInst::Predicate P = CmpInst::ICMP_NE; 1779 switch (RK) { 1780 default: 1781 llvm_unreachable("Unknown min/max reduction kind"); 1782 case LoopVectorizationLegality::MRK_UIntMin: 1783 P = CmpInst::ICMP_ULT; 1784 break; 1785 case LoopVectorizationLegality::MRK_UIntMax: 1786 P = CmpInst::ICMP_UGT; 1787 break; 1788 case LoopVectorizationLegality::MRK_SIntMin: 1789 P = CmpInst::ICMP_SLT; 1790 break; 1791 case LoopVectorizationLegality::MRK_SIntMax: 1792 P = CmpInst::ICMP_SGT; 1793 break; 1794 case LoopVectorizationLegality::MRK_FloatMin: 1795 P = CmpInst::FCMP_OLT; 1796 break; 1797 case LoopVectorizationLegality::MRK_FloatMax: 1798 P = CmpInst::FCMP_OGT; 1799 break; 1800 } 1801 1802 Value *Cmp; 1803 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax) 1804 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp"); 1805 else 1806 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); 1807 1808 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); 1809 return Select; 1810} 1811 1812void 1813InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { 1814 //===------------------------------------------------===// 1815 // 1816 // Notice: any optimization or new instruction that go 1817 // into the code below should be also be implemented in 1818 // the cost-model. 1819 // 1820 //===------------------------------------------------===// 1821 Constant *Zero = Builder.getInt32(0); 1822 1823 // In order to support reduction variables we need to be able to vectorize 1824 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two 1825 // stages. First, we create a new vector PHI node with no incoming edges. 1826 // We use this value when we vectorize all of the instructions that use the 1827 // PHI. Next, after all of the instructions in the block are complete we 1828 // add the new incoming edges to the PHI. At this point all of the 1829 // instructions in the basic block are vectorized, so we can use them to 1830 // construct the PHI. 1831 PhiVector RdxPHIsToFix; 1832 1833 // Scan the loop in a topological order to ensure that defs are vectorized 1834 // before users. 1835 LoopBlocksDFS DFS(OrigLoop); 1836 DFS.perform(LI); 1837 1838 // Vectorize all of the blocks in the original loop. 1839 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 1840 be = DFS.endRPO(); bb != be; ++bb) 1841 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix); 1842 1843 // At this point every instruction in the original loop is widened to 1844 // a vector form. We are almost done. Now, we need to fix the PHI nodes 1845 // that we vectorized. The PHI nodes are currently empty because we did 1846 // not want to introduce cycles. Notice that the remaining PHI nodes 1847 // that we need to fix are reduction variables. 1848 1849 // Create the 'reduced' values for each of the induction vars. 1850 // The reduced values are the vector values that we scalarize and combine 1851 // after the loop is finished. 1852 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); 1853 it != e; ++it) { 1854 PHINode *RdxPhi = *it; 1855 assert(RdxPhi && "Unable to recover vectorized PHI"); 1856 1857 // Find the reduction variable descriptor. 1858 assert(Legal->getReductionVars()->count(RdxPhi) && 1859 "Unable to find the reduction variable"); 1860 LoopVectorizationLegality::ReductionDescriptor RdxDesc = 1861 (*Legal->getReductionVars())[RdxPhi]; 1862 1863 // We need to generate a reduction vector from the incoming scalar. 1864 // To do so, we need to generate the 'identity' vector and overide 1865 // one of the elements with the incoming scalar reduction. We need 1866 // to do it in the vector-loop preheader. 1867 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator()); 1868 1869 // This is the vector-clone of the value that leaves the loop. 1870 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr); 1871 Type *VecTy = VectorExit[0]->getType(); 1872 1873 // Find the reduction identity variable. Zero for addition, or, xor, 1874 // one for multiplication, -1 for And. 1875 Value *Identity; 1876 Value *VectorStart; 1877 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax || 1878 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) { 1879 // MinMax reduction have the start value as their identify. 1880 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue, 1881 "minmax.ident"); 1882 } else { 1883 Constant *Iden = 1884 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind, 1885 VecTy->getScalarType()); 1886 Identity = ConstantVector::getSplat(VF, Iden); 1887 1888 // This vector is the Identity vector where the first element is the 1889 // incoming scalar reduction. 1890 VectorStart = Builder.CreateInsertElement(Identity, 1891 RdxDesc.StartValue, Zero); 1892 } 1893 1894 // Fix the vector-loop phi. 1895 // We created the induction variable so we know that the 1896 // preheader is the first entry. 1897 BasicBlock *VecPreheader = Induction->getIncomingBlock(0); 1898 1899 // Reductions do not have to start at zero. They can start with 1900 // any loop invariant values. 1901 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 1902 BasicBlock *Latch = OrigLoop->getLoopLatch(); 1903 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 1904 VectorParts &Val = getVectorValue(LoopVal); 1905 for (unsigned part = 0; part < UF; ++part) { 1906 // Make sure to add the reduction stat value only to the 1907 // first unroll part. 1908 Value *StartVal = (part == 0) ? VectorStart : Identity; 1909 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader); 1910 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody); 1911 } 1912 1913 // Before each round, move the insertion point right between 1914 // the PHIs and the values we are going to write. 1915 // This allows us to write both PHINodes and the extractelement 1916 // instructions. 1917 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); 1918 1919 VectorParts RdxParts; 1920 for (unsigned part = 0; part < UF; ++part) { 1921 // This PHINode contains the vectorized reduction variable, or 1922 // the initial value vector, if we bypass the vector loop. 1923 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); 1924 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); 1925 Value *StartVal = (part == 0) ? VectorStart : Identity; 1926 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1927 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]); 1928 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody); 1929 RdxParts.push_back(NewPhi); 1930 } 1931 1932 // Reduce all of the unrolled parts into a single vector. 1933 Value *ReducedPartRdx = RdxParts[0]; 1934 unsigned Op = getReductionBinOp(RdxDesc.Kind); 1935 for (unsigned part = 1; part < UF; ++part) { 1936 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 1937 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op, 1938 RdxParts[part], ReducedPartRdx, 1939 "bin.rdx"); 1940 else 1941 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind, 1942 ReducedPartRdx, RdxParts[part]); 1943 } 1944 1945 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 1946 // and vector ops, reducing the set of values being computed by half each 1947 // round. 1948 assert(isPowerOf2_32(VF) && 1949 "Reduction emission only supported for pow2 vectors!"); 1950 Value *TmpVec = ReducedPartRdx; 1951 SmallVector<Constant*, 32> ShuffleMask(VF, 0); 1952 for (unsigned i = VF; i != 1; i >>= 1) { 1953 // Move the upper half of the vector to the lower half. 1954 for (unsigned j = 0; j != i/2; ++j) 1955 ShuffleMask[j] = Builder.getInt32(i/2 + j); 1956 1957 // Fill the rest of the mask with undef. 1958 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 1959 UndefValue::get(Builder.getInt32Ty())); 1960 1961 Value *Shuf = 1962 Builder.CreateShuffleVector(TmpVec, 1963 UndefValue::get(TmpVec->getType()), 1964 ConstantVector::get(ShuffleMask), 1965 "rdx.shuf"); 1966 1967 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 1968 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf, 1969 "bin.rdx"); 1970 else 1971 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf); 1972 } 1973 1974 // The result is in the first element of the vector. 1975 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); 1976 1977 // Now, we need to fix the users of the reduction variable 1978 // inside and outside of the scalar remainder loop. 1979 // We know that the loop is in LCSSA form. We need to update the 1980 // PHI nodes in the exit blocks. 1981 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 1982 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 1983 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 1984 if (!LCSSAPhi) continue; 1985 1986 // All PHINodes need to have a single entry edge, or two if 1987 // we already fixed them. 1988 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 1989 1990 // We found our reduction value exit-PHI. Update it with the 1991 // incoming bypass edge. 1992 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { 1993 // Add an edge coming from the bypass. 1994 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock); 1995 break; 1996 } 1997 }// end of the LCSSA phi scan. 1998 1999 // Fix the scalar loop reduction variable with the incoming reduction sum 2000 // from the vector body and from the backedge value. 2001 int IncomingEdgeBlockIdx = 2002 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 2003 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 2004 // Pick the other block. 2005 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 2006 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0); 2007 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); 2008 }// end of for each redux variable. 2009 2010 // The Loop exit block may have single value PHI nodes where the incoming 2011 // value is 'undef'. While vectorizing we only handled real values that 2012 // were defined inside the loop. Here we handle the 'undef case'. 2013 // See PR14725. 2014 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 2015 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 2016 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 2017 if (!LCSSAPhi) continue; 2018 if (LCSSAPhi->getNumIncomingValues() == 1) 2019 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 2020 LoopMiddleBlock); 2021 } 2022} 2023 2024InnerLoopVectorizer::VectorParts 2025InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 2026 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 2027 "Invalid edge"); 2028 2029 VectorParts SrcMask = createBlockInMask(Src); 2030 2031 // The terminator has to be a branch inst! 2032 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 2033 assert(BI && "Unexpected terminator found"); 2034 2035 if (BI->isConditional()) { 2036 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 2037 2038 if (BI->getSuccessor(0) != Dst) 2039 for (unsigned part = 0; part < UF; ++part) 2040 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 2041 2042 for (unsigned part = 0; part < UF; ++part) 2043 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 2044 return EdgeMask; 2045 } 2046 2047 return SrcMask; 2048} 2049 2050InnerLoopVectorizer::VectorParts 2051InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 2052 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 2053 2054 // Loop incoming mask is all-one. 2055 if (OrigLoop->getHeader() == BB) { 2056 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 2057 return getVectorValue(C); 2058 } 2059 2060 // This is the block mask. We OR all incoming edges, and with zero. 2061 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 2062 VectorParts BlockMask = getVectorValue(Zero); 2063 2064 // For each pred: 2065 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 2066 VectorParts EM = createEdgeMask(*it, BB); 2067 for (unsigned part = 0; part < UF; ++part) 2068 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 2069 } 2070 2071 return BlockMask; 2072} 2073 2074void 2075InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, 2076 BasicBlock *BB, PhiVector *PV) { 2077 // For each instruction in the old loop. 2078 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 2079 VectorParts &Entry = WidenMap.get(it); 2080 switch (it->getOpcode()) { 2081 case Instruction::Br: 2082 // Nothing to do for PHIs and BR, since we already took care of the 2083 // loop control flow instructions. 2084 continue; 2085 case Instruction::PHI:{ 2086 PHINode* P = cast<PHINode>(it); 2087 // Handle reduction variables: 2088 if (Legal->getReductionVars()->count(P)) { 2089 for (unsigned part = 0; part < UF; ++part) { 2090 // This is phase one of vectorizing PHIs. 2091 Type *VecTy = VectorType::get(it->getType(), VF); 2092 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", 2093 LoopVectorBody-> getFirstInsertionPt()); 2094 } 2095 PV->push_back(P); 2096 continue; 2097 } 2098 2099 // Check for PHI nodes that are lowered to vector selects. 2100 if (P->getParent() != OrigLoop->getHeader()) { 2101 // We know that all PHIs in non header blocks are converted into 2102 // selects, so we don't have to worry about the insertion order and we 2103 // can just use the builder. 2104 // At this point we generate the predication tree. There may be 2105 // duplications since this is a simple recursive scan, but future 2106 // optimizations will clean it up. 2107 2108 unsigned NumIncoming = P->getNumIncomingValues(); 2109 2110 // Generate a sequence of selects of the form: 2111 // SELECT(Mask3, In3, 2112 // SELECT(Mask2, In2, 2113 // ( ...))) 2114 for (unsigned In = 0; In < NumIncoming; In++) { 2115 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), 2116 P->getParent()); 2117 VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); 2118 2119 for (unsigned part = 0; part < UF; ++part) { 2120 // We might have single edge PHIs (blocks) - use an identity 2121 // 'select' for the first PHI operand. 2122 if (In == 0) 2123 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2124 In0[part]); 2125 else 2126 // Select between the current value and the previous incoming edge 2127 // based on the incoming mask. 2128 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2129 Entry[part], "predphi"); 2130 } 2131 } 2132 continue; 2133 } 2134 2135 // This PHINode must be an induction variable. 2136 // Make sure that we know about it. 2137 assert(Legal->getInductionVars()->count(P) && 2138 "Not an induction variable"); 2139 2140 LoopVectorizationLegality::InductionInfo II = 2141 Legal->getInductionVars()->lookup(P); 2142 2143 switch (II.IK) { 2144 case LoopVectorizationLegality::IK_NoInduction: 2145 llvm_unreachable("Unknown induction"); 2146 case LoopVectorizationLegality::IK_IntInduction: { 2147 assert(P->getType() == II.StartValue->getType() && "Types must match"); 2148 Type *PhiTy = P->getType(); 2149 Value *Broadcasted; 2150 if (P == OldInduction) { 2151 // Handle the canonical induction variable. We might have had to 2152 // extend the type. 2153 Broadcasted = Builder.CreateTrunc(Induction, PhiTy); 2154 } else { 2155 // Handle other induction variables that are now based on the 2156 // canonical one. 2157 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx, 2158 "normalized.idx"); 2159 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy); 2160 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx, 2161 "offset.idx"); 2162 } 2163 Broadcasted = getBroadcastInstrs(Broadcasted); 2164 // After broadcasting the induction variable we need to make the vector 2165 // consecutive by adding 0, 1, 2, etc. 2166 for (unsigned part = 0; part < UF; ++part) 2167 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); 2168 continue; 2169 } 2170 case LoopVectorizationLegality::IK_ReverseIntInduction: 2171 case LoopVectorizationLegality::IK_PtrInduction: 2172 case LoopVectorizationLegality::IK_ReversePtrInduction: 2173 // Handle reverse integer and pointer inductions. 2174 Value *StartIdx = ExtendedIdx; 2175 // This is the normalized GEP that starts counting at zero. 2176 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx, 2177 "normalized.idx"); 2178 2179 // Handle the reverse integer induction variable case. 2180 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) { 2181 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType()); 2182 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy, 2183 "resize.norm.idx"); 2184 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI, 2185 "reverse.idx"); 2186 2187 // This is a new value so do not hoist it out. 2188 Value *Broadcasted = getBroadcastInstrs(ReverseInd); 2189 // After broadcasting the induction variable we need to make the 2190 // vector consecutive by adding ... -3, -2, -1, 0. 2191 for (unsigned part = 0; part < UF; ++part) 2192 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part, 2193 true); 2194 continue; 2195 } 2196 2197 // Handle the pointer induction variable case. 2198 assert(P->getType()->isPointerTy() && "Unexpected type."); 2199 2200 // Is this a reverse induction ptr or a consecutive induction ptr. 2201 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction == 2202 II.IK); 2203 2204 // This is the vector of results. Notice that we don't generate 2205 // vector geps because scalar geps result in better code. 2206 for (unsigned part = 0; part < UF; ++part) { 2207 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 2208 for (unsigned int i = 0; i < VF; ++i) { 2209 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1); 2210 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 2211 Value *GlobalIdx; 2212 if (!Reverse) 2213 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); 2214 else 2215 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); 2216 2217 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 2218 "next.gep"); 2219 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 2220 Builder.getInt32(i), 2221 "insert.gep"); 2222 } 2223 Entry[part] = VecVal; 2224 } 2225 continue; 2226 } 2227 2228 }// End of PHI. 2229 2230 case Instruction::Add: 2231 case Instruction::FAdd: 2232 case Instruction::Sub: 2233 case Instruction::FSub: 2234 case Instruction::Mul: 2235 case Instruction::FMul: 2236 case Instruction::UDiv: 2237 case Instruction::SDiv: 2238 case Instruction::FDiv: 2239 case Instruction::URem: 2240 case Instruction::SRem: 2241 case Instruction::FRem: 2242 case Instruction::Shl: 2243 case Instruction::LShr: 2244 case Instruction::AShr: 2245 case Instruction::And: 2246 case Instruction::Or: 2247 case Instruction::Xor: { 2248 // Just widen binops. 2249 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 2250 VectorParts &A = getVectorValue(it->getOperand(0)); 2251 VectorParts &B = getVectorValue(it->getOperand(1)); 2252 2253 // Use this vector value for all users of the original instruction. 2254 for (unsigned Part = 0; Part < UF; ++Part) { 2255 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 2256 2257 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef. 2258 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V); 2259 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) { 2260 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); 2261 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); 2262 } 2263 if (VecOp && isa<PossiblyExactOperator>(VecOp)) 2264 VecOp->setIsExact(BinOp->isExact()); 2265 2266 Entry[Part] = V; 2267 } 2268 break; 2269 } 2270 case Instruction::Select: { 2271 // Widen selects. 2272 // If the selector is loop invariant we can create a select 2273 // instruction with a scalar condition. Otherwise, use vector-select. 2274 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), 2275 OrigLoop); 2276 2277 // The condition can be loop invariant but still defined inside the 2278 // loop. This means that we can't just use the original 'cond' value. 2279 // We have to take the 'vectorized' value and pick the first lane. 2280 // Instcombine will make this a no-op. 2281 VectorParts &Cond = getVectorValue(it->getOperand(0)); 2282 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 2283 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 2284 Value *ScalarCond = Builder.CreateExtractElement(Cond[0], 2285 Builder.getInt32(0)); 2286 for (unsigned Part = 0; Part < UF; ++Part) { 2287 Entry[Part] = Builder.CreateSelect( 2288 InvariantCond ? ScalarCond : Cond[Part], 2289 Op0[Part], 2290 Op1[Part]); 2291 } 2292 break; 2293 } 2294 2295 case Instruction::ICmp: 2296 case Instruction::FCmp: { 2297 // Widen compares. Generate vector compares. 2298 bool FCmp = (it->getOpcode() == Instruction::FCmp); 2299 CmpInst *Cmp = dyn_cast<CmpInst>(it); 2300 VectorParts &A = getVectorValue(it->getOperand(0)); 2301 VectorParts &B = getVectorValue(it->getOperand(1)); 2302 for (unsigned Part = 0; Part < UF; ++Part) { 2303 Value *C = 0; 2304 if (FCmp) 2305 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 2306 else 2307 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 2308 Entry[Part] = C; 2309 } 2310 break; 2311 } 2312 2313 case Instruction::Store: 2314 case Instruction::Load: 2315 vectorizeMemoryInstruction(it, Legal); 2316 break; 2317 case Instruction::ZExt: 2318 case Instruction::SExt: 2319 case Instruction::FPToUI: 2320 case Instruction::FPToSI: 2321 case Instruction::FPExt: 2322 case Instruction::PtrToInt: 2323 case Instruction::IntToPtr: 2324 case Instruction::SIToFP: 2325 case Instruction::UIToFP: 2326 case Instruction::Trunc: 2327 case Instruction::FPTrunc: 2328 case Instruction::BitCast: { 2329 CastInst *CI = dyn_cast<CastInst>(it); 2330 /// Optimize the special case where the source is the induction 2331 /// variable. Notice that we can only optimize the 'trunc' case 2332 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 2333 /// c. other casts depend on pointer size. 2334 if (CI->getOperand(0) == OldInduction && 2335 it->getOpcode() == Instruction::Trunc) { 2336 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 2337 CI->getType()); 2338 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 2339 for (unsigned Part = 0; Part < UF; ++Part) 2340 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false); 2341 break; 2342 } 2343 /// Vectorize casts. 2344 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF); 2345 2346 VectorParts &A = getVectorValue(it->getOperand(0)); 2347 for (unsigned Part = 0; Part < UF; ++Part) 2348 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 2349 break; 2350 } 2351 2352 case Instruction::Call: { 2353 // Ignore dbg intrinsics. 2354 if (isa<DbgInfoIntrinsic>(it)) 2355 break; 2356 2357 Module *M = BB->getParent()->getParent(); 2358 CallInst *CI = cast<CallInst>(it); 2359 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 2360 assert(ID && "Not an intrinsic call!"); 2361 for (unsigned Part = 0; Part < UF; ++Part) { 2362 SmallVector<Value*, 4> Args; 2363 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 2364 VectorParts &Arg = getVectorValue(CI->getArgOperand(i)); 2365 Args.push_back(Arg[Part]); 2366 } 2367 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) }; 2368 Function *F = Intrinsic::getDeclaration(M, ID, Tys); 2369 Entry[Part] = Builder.CreateCall(F, Args); 2370 } 2371 break; 2372 } 2373 2374 default: 2375 // All other instructions are unsupported. Scalarize them. 2376 scalarizeInstruction(it); 2377 break; 2378 }// end of switch. 2379 }// end of for_each instr. 2380} 2381 2382void InnerLoopVectorizer::updateAnalysis() { 2383 // Forget the original basic block. 2384 SE->forgetLoop(OrigLoop); 2385 2386 // Update the dominator tree information. 2387 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && 2388 "Entry does not dominate exit."); 2389 2390 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2391 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); 2392 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); 2393 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); 2394 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front()); 2395 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); 2396 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 2397 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); 2398 2399 DEBUG(DT->verifyAnalysis()); 2400} 2401 2402bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 2403 if (!EnableIfConversion) 2404 return false; 2405 2406 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 2407 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector(); 2408 2409 // Collect the blocks that need predication. 2410 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) { 2411 BasicBlock *BB = LoopBlocks[i]; 2412 2413 // We don't support switch statements inside loops. 2414 if (!isa<BranchInst>(BB->getTerminator())) 2415 return false; 2416 2417 // We must be able to predicate all blocks that need to be predicated. 2418 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB)) 2419 return false; 2420 } 2421 2422 // Check that we can actually speculate the hoistable loads. 2423 if (!LoadSpeculation.canHoistAllLoads()) 2424 return false; 2425 2426 // We can if-convert this loop. 2427 return true; 2428} 2429 2430bool LoopVectorizationLegality::canVectorize() { 2431 assert(TheLoop->getLoopPreheader() && "No preheader!!"); 2432 2433 // We can only vectorize innermost loops. 2434 if (TheLoop->getSubLoopsVector().size()) 2435 return false; 2436 2437 // We must have a single backedge. 2438 if (TheLoop->getNumBackEdges() != 1) 2439 return false; 2440 2441 // We must have a single exiting block. 2442 if (!TheLoop->getExitingBlock()) 2443 return false; 2444 2445 unsigned NumBlocks = TheLoop->getNumBlocks(); 2446 2447 // Check if we can if-convert non single-bb loops. 2448 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 2449 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 2450 return false; 2451 } 2452 2453 // We need to have a loop header. 2454 BasicBlock *Latch = TheLoop->getLoopLatch(); 2455 DEBUG(dbgs() << "LV: Found a loop: " << 2456 TheLoop->getHeader()->getName() << "\n"); 2457 2458 // ScalarEvolution needs to be able to find the exit count. 2459 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch); 2460 if (ExitCount == SE->getCouldNotCompute()) { 2461 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 2462 return false; 2463 } 2464 2465 // Do not loop-vectorize loops with a tiny trip count. 2466 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch); 2467 if (TC > 0u && TC < TinyTripCountVectorThreshold) { 2468 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << 2469 "This loop is not worth vectorizing.\n"); 2470 return false; 2471 } 2472 2473 // Check if we can vectorize the instructions and CFG in this loop. 2474 if (!canVectorizeInstrs()) { 2475 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 2476 return false; 2477 } 2478 2479 // Go over each instruction and look at memory deps. 2480 if (!canVectorizeMemory()) { 2481 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 2482 return false; 2483 } 2484 2485 // Collect all of the variables that remain uniform after vectorization. 2486 collectLoopUniforms(); 2487 2488 DEBUG(dbgs() << "LV: We can vectorize this loop" << 2489 (PtrRtCheck.Need ? " (with a runtime bound check)" : "") 2490 <<"!\n"); 2491 2492 // Okay! We can vectorize. At this point we don't have any other mem analysis 2493 // which may limit our maximum vectorization factor, so just return true with 2494 // no restrictions. 2495 return true; 2496} 2497 2498static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) { 2499 if (Ty->isPointerTy()) 2500 return DL.getIntPtrType(Ty->getContext()); 2501 return Ty; 2502} 2503 2504static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) { 2505 Ty0 = convertPointerToIntegerType(DL, Ty0); 2506 Ty1 = convertPointerToIntegerType(DL, Ty1); 2507 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits()) 2508 return Ty0; 2509 return Ty1; 2510} 2511 2512bool LoopVectorizationLegality::canVectorizeInstrs() { 2513 BasicBlock *PreHeader = TheLoop->getLoopPreheader(); 2514 BasicBlock *Header = TheLoop->getHeader(); 2515 2516 // If we marked the scalar loop as "already vectorized" then no need 2517 // to vectorize it again. 2518 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) { 2519 DEBUG(dbgs() << "LV: This loop was vectorized before\n"); 2520 return false; 2521 } 2522 2523 // Look for the attribute signaling the absence of NaNs. 2524 Function &F = *Header->getParent(); 2525 if (F.hasFnAttribute("no-nans-fp-math")) 2526 HasFunNoNaNAttr = F.getAttributes().getAttribute( 2527 AttributeSet::FunctionIndex, 2528 "no-nans-fp-math").getValueAsString() == "true"; 2529 2530 // For each block in the loop. 2531 for (Loop::block_iterator bb = TheLoop->block_begin(), 2532 be = TheLoop->block_end(); bb != be; ++bb) { 2533 2534 // Scan the instructions in the block and look for hazards. 2535 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2536 ++it) { 2537 2538 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 2539 Type *PhiTy = Phi->getType(); 2540 // Check that this PHI type is allowed. 2541 if (!PhiTy->isIntegerTy() && 2542 !PhiTy->isFloatingPointTy() && 2543 !PhiTy->isPointerTy()) { 2544 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 2545 return false; 2546 } 2547 2548 // If this PHINode is not in the header block, then we know that we 2549 // can convert it to select during if-conversion. No need to check if 2550 // the PHIs in this block are induction or reduction variables. 2551 if (*bb != Header) 2552 continue; 2553 2554 // We only allow if-converted PHIs with more than two incoming values. 2555 if (Phi->getNumIncomingValues() != 2) { 2556 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 2557 return false; 2558 } 2559 2560 // This is the value coming from the preheader. 2561 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); 2562 // Check if this is an induction variable. 2563 InductionKind IK = isInductionVariable(Phi); 2564 2565 if (IK_NoInduction != IK) { 2566 // Get the widest type. 2567 if (!WidestIndTy) 2568 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy); 2569 else 2570 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy); 2571 2572 // Int inductions are special because we only allow one IV. 2573 if (IK == IK_IntInduction) { 2574 // Use the phi node with the widest type as induction. Use the last 2575 // one if there are multiple (no good reason for doing this other 2576 // than it is expedient). 2577 if (!Induction || PhiTy == WidestIndTy) 2578 Induction = Phi; 2579 } 2580 2581 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 2582 Inductions[Phi] = InductionInfo(StartValue, IK); 2583 continue; 2584 } 2585 2586 if (AddReductionVar(Phi, RK_IntegerAdd)) { 2587 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); 2588 continue; 2589 } 2590 if (AddReductionVar(Phi, RK_IntegerMult)) { 2591 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); 2592 continue; 2593 } 2594 if (AddReductionVar(Phi, RK_IntegerOr)) { 2595 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); 2596 continue; 2597 } 2598 if (AddReductionVar(Phi, RK_IntegerAnd)) { 2599 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); 2600 continue; 2601 } 2602 if (AddReductionVar(Phi, RK_IntegerXor)) { 2603 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); 2604 continue; 2605 } 2606 if (AddReductionVar(Phi, RK_IntegerMinMax)) { 2607 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n"); 2608 continue; 2609 } 2610 if (AddReductionVar(Phi, RK_FloatMult)) { 2611 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n"); 2612 continue; 2613 } 2614 if (AddReductionVar(Phi, RK_FloatAdd)) { 2615 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n"); 2616 continue; 2617 } 2618 if (AddReductionVar(Phi, RK_FloatMinMax)) { 2619 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n"); 2620 continue; 2621 } 2622 2623 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 2624 return false; 2625 }// end of PHI handling 2626 2627 // We still don't handle functions. However, we can ignore dbg intrinsic 2628 // calls and we do handle certain intrinsic and libm functions. 2629 CallInst *CI = dyn_cast<CallInst>(it); 2630 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) { 2631 DEBUG(dbgs() << "LV: Found a call site.\n"); 2632 return false; 2633 } 2634 2635 // Check that the instruction return type is vectorizable. 2636 if (!VectorType::isValidElementType(it->getType()) && 2637 !it->getType()->isVoidTy()) { 2638 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n"); 2639 return false; 2640 } 2641 2642 // Check that the stored type is vectorizable. 2643 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 2644 Type *T = ST->getValueOperand()->getType(); 2645 if (!VectorType::isValidElementType(T)) 2646 return false; 2647 } 2648 2649 // Reduction instructions are allowed to have exit users. 2650 // All other instructions must not have external users. 2651 if (!AllowedExit.count(it)) 2652 //Check that all of the users of the loop are inside the BB. 2653 for (Value::use_iterator I = it->use_begin(), E = it->use_end(); 2654 I != E; ++I) { 2655 Instruction *U = cast<Instruction>(*I); 2656 // This user may be a reduction exit value. 2657 if (!TheLoop->contains(U)) { 2658 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n"); 2659 return false; 2660 } 2661 } 2662 } // next instr. 2663 2664 } 2665 2666 if (!Induction) { 2667 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 2668 if (Inductions.empty()) 2669 return false; 2670 } 2671 2672 return true; 2673} 2674 2675void LoopVectorizationLegality::collectLoopUniforms() { 2676 // We now know that the loop is vectorizable! 2677 // Collect variables that will remain uniform after vectorization. 2678 std::vector<Value*> Worklist; 2679 BasicBlock *Latch = TheLoop->getLoopLatch(); 2680 2681 // Start with the conditional branch and walk up the block. 2682 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 2683 2684 while (Worklist.size()) { 2685 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 2686 Worklist.pop_back(); 2687 2688 // Look at instructions inside this loop. 2689 // Stop when reaching PHI nodes. 2690 // TODO: we need to follow values all over the loop, not only in this block. 2691 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 2692 continue; 2693 2694 // This is a known uniform. 2695 Uniforms.insert(I); 2696 2697 // Insert all operands. 2698 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end()); 2699 } 2700} 2701 2702AliasAnalysis::Location 2703LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) { 2704 if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) 2705 return AA->getLocation(Store); 2706 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) 2707 return AA->getLocation(Load); 2708 2709 llvm_unreachable("Should be either load or store instruction"); 2710} 2711 2712bool 2713LoopVectorizationLegality::hasPossibleGlobalWriteReorder( 2714 Value *Object, 2715 Instruction *Inst, 2716 AliasMultiMap& WriteObjects, 2717 unsigned MaxByteWidth) { 2718 2719 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst); 2720 2721 std::vector<Instruction*>::iterator 2722 it = WriteObjects[Object].begin(), 2723 end = WriteObjects[Object].end(); 2724 2725 for (; it != end; ++it) { 2726 Instruction* I = *it; 2727 if (I == Inst) 2728 continue; 2729 2730 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I); 2731 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth), 2732 ThatLoc.getWithNewSize(MaxByteWidth))) 2733 return true; 2734 } 2735 return false; 2736} 2737 2738bool LoopVectorizationLegality::canVectorizeMemory() { 2739 2740 typedef SmallVector<Value*, 16> ValueVector; 2741 typedef SmallPtrSet<Value*, 16> ValueSet; 2742 // Holds the Load and Store *instructions*. 2743 ValueVector Loads; 2744 ValueVector Stores; 2745 PtrRtCheck.Pointers.clear(); 2746 PtrRtCheck.Need = false; 2747 2748 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); 2749 2750 // For each block. 2751 for (Loop::block_iterator bb = TheLoop->block_begin(), 2752 be = TheLoop->block_end(); bb != be; ++bb) { 2753 2754 // Scan the BB and collect legal loads and stores. 2755 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2756 ++it) { 2757 2758 // If this is a load, save it. If this instruction can read from memory 2759 // but is not a load, then we quit. Notice that we don't handle function 2760 // calls that read or write. 2761 if (it->mayReadFromMemory()) { 2762 LoadInst *Ld = dyn_cast<LoadInst>(it); 2763 if (!Ld) return false; 2764 if (!Ld->isSimple() && !IsAnnotatedParallel) { 2765 DEBUG(dbgs() << "LV: Found a non-simple load.\n"); 2766 return false; 2767 } 2768 Loads.push_back(Ld); 2769 continue; 2770 } 2771 2772 // Save 'store' instructions. Abort if other instructions write to memory. 2773 if (it->mayWriteToMemory()) { 2774 StoreInst *St = dyn_cast<StoreInst>(it); 2775 if (!St) return false; 2776 if (!St->isSimple() && !IsAnnotatedParallel) { 2777 DEBUG(dbgs() << "LV: Found a non-simple store.\n"); 2778 return false; 2779 } 2780 Stores.push_back(St); 2781 } 2782 } // next instr. 2783 } // next block. 2784 2785 // Now we have two lists that hold the loads and the stores. 2786 // Next, we find the pointers that they use. 2787 2788 // Check if we see any stores. If there are no stores, then we don't 2789 // care if the pointers are *restrict*. 2790 if (!Stores.size()) { 2791 DEBUG(dbgs() << "LV: Found a read-only loop!\n"); 2792 return true; 2793 } 2794 2795 // Holds the read and read-write *pointers* that we find. These maps hold 2796 // unique values for pointers (so no need for multi-map). 2797 AliasMap Reads; 2798 AliasMap ReadWrites; 2799 2800 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects 2801 // multiple times on the same object. If the ptr is accessed twice, once 2802 // for read and once for write, it will only appear once (on the write 2803 // list). This is okay, since we are going to check for conflicts between 2804 // writes and between reads and writes, but not between reads and reads. 2805 ValueSet Seen; 2806 2807 ValueVector::iterator I, IE; 2808 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) { 2809 StoreInst *ST = cast<StoreInst>(*I); 2810 Value* Ptr = ST->getPointerOperand(); 2811 2812 if (isUniform(Ptr)) { 2813 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 2814 return false; 2815 } 2816 2817 // If we did *not* see this pointer before, insert it to 2818 // the read-write list. At this phase it is only a 'write' list. 2819 if (Seen.insert(Ptr)) 2820 ReadWrites.insert(std::make_pair(Ptr, ST)); 2821 } 2822 2823 if (IsAnnotatedParallel) { 2824 DEBUG(dbgs() 2825 << "LV: A loop annotated parallel, ignore memory dependency " 2826 << "checks.\n"); 2827 return true; 2828 } 2829 2830 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { 2831 LoadInst *LD = cast<LoadInst>(*I); 2832 Value* Ptr = LD->getPointerOperand(); 2833 // If we did *not* see this pointer before, insert it to the 2834 // read list. If we *did* see it before, then it is already in 2835 // the read-write list. This allows us to vectorize expressions 2836 // such as A[i] += x; Because the address of A[i] is a read-write 2837 // pointer. This only works if the index of A[i] is consecutive. 2838 // If the address of i is unknown (for example A[B[i]]) then we may 2839 // read a few words, modify, and write a few words, and some of the 2840 // words may be written to the same address. 2841 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr)) 2842 Reads.insert(std::make_pair(Ptr, LD)); 2843 } 2844 2845 // If we write (or read-write) to a single destination and there are no 2846 // other reads in this loop then is it safe to vectorize. 2847 if (ReadWrites.size() == 1 && Reads.size() == 0) { 2848 DEBUG(dbgs() << "LV: Found a write-only loop!\n"); 2849 return true; 2850 } 2851 2852 unsigned NumReadPtrs = 0; 2853 unsigned NumWritePtrs = 0; 2854 2855 // Find pointers with computable bounds. We are going to use this information 2856 // to place a runtime bound check. 2857 bool CanDoRT = true; 2858 AliasMap::iterator MI, ME; 2859 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { 2860 Value *V = (*MI).first; 2861 if (hasComputableBounds(V)) { 2862 PtrRtCheck.insert(SE, TheLoop, V, true); 2863 NumWritePtrs++; 2864 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); 2865 } else { 2866 CanDoRT = false; 2867 break; 2868 } 2869 } 2870 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { 2871 Value *V = (*MI).first; 2872 if (hasComputableBounds(V)) { 2873 PtrRtCheck.insert(SE, TheLoop, V, false); 2874 NumReadPtrs++; 2875 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); 2876 } else { 2877 CanDoRT = false; 2878 break; 2879 } 2880 } 2881 2882 // Check that we did not collect too many pointers or found a 2883 // unsizeable pointer. 2884 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1)); 2885 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n"); 2886 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { 2887 PtrRtCheck.reset(); 2888 CanDoRT = false; 2889 } 2890 2891 if (CanDoRT) { 2892 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n"); 2893 } 2894 2895 bool NeedRTCheck = false; 2896 2897 // Biggest vectorized access possible, vector width * unroll factor. 2898 // TODO: We're being very pessimistic here, find a way to know the 2899 // real access width before getting here. 2900 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) * 2901 TTI->getMaximumUnrollFactor(); 2902 // Now that the pointers are in two lists (Reads and ReadWrites), we 2903 // can check that there are no conflicts between each of the writes and 2904 // between the writes to the reads. 2905 // Note that WriteObjects duplicates the stores (indexed now by underlying 2906 // objects) to avoid pointing to elements inside ReadWrites. 2907 // TODO: Maybe create a new type where they can interact without duplication. 2908 AliasMultiMap WriteObjects; 2909 ValueVector TempObjects; 2910 2911 // Check that the read-writes do not conflict with other read-write 2912 // pointers. 2913 bool AllWritesIdentified = true; 2914 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { 2915 Value *Val = (*MI).first; 2916 Instruction *Inst = (*MI).second; 2917 2918 GetUnderlyingObjects(Val, TempObjects, DL); 2919 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); 2920 UI != UE; ++UI) { 2921 if (!isIdentifiedObject(*UI)) { 2922 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n"); 2923 NeedRTCheck = true; 2924 AllWritesIdentified = false; 2925 } 2926 2927 // Never seen it before, can't alias. 2928 if (WriteObjects[*UI].empty()) { 2929 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n"); 2930 WriteObjects[*UI].push_back(Inst); 2931 continue; 2932 } 2933 // Direct alias found. 2934 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) { 2935 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2936 << **UI <<"\n"); 2937 return false; 2938 } 2939 DEBUG(dbgs() << "LV: Found a conflicting global value:" 2940 << **UI <<"\n"); 2941 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n"); 2942 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); 2943 2944 // If global alias, make sure they do alias. 2945 if (hasPossibleGlobalWriteReorder(*UI, 2946 Inst, 2947 WriteObjects, 2948 MaxByteWidth)) { 2949 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI 2950 << "\n"); 2951 return false; 2952 } 2953 2954 // Didn't alias, insert into map for further reference. 2955 WriteObjects[*UI].push_back(Inst); 2956 } 2957 TempObjects.clear(); 2958 } 2959 2960 /// Check that the reads don't conflict with the read-writes. 2961 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { 2962 Value *Val = (*MI).first; 2963 GetUnderlyingObjects(Val, TempObjects, DL); 2964 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); 2965 UI != UE; ++UI) { 2966 // If all of the writes are identified then we don't care if the read 2967 // pointer is identified or not. 2968 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) { 2969 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n"); 2970 NeedRTCheck = true; 2971 } 2972 2973 // Never seen it before, can't alias. 2974 if (WriteObjects[*UI].empty()) 2975 continue; 2976 // Direct alias found. 2977 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) { 2978 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2979 << **UI <<"\n"); 2980 return false; 2981 } 2982 DEBUG(dbgs() << "LV: Found a global value: " 2983 << **UI <<"\n"); 2984 Instruction *Inst = (*MI).second; 2985 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n"); 2986 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); 2987 2988 // If global alias, make sure they do alias. 2989 if (hasPossibleGlobalWriteReorder(*UI, 2990 Inst, 2991 WriteObjects, 2992 MaxByteWidth)) { 2993 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI 2994 << "\n"); 2995 return false; 2996 } 2997 } 2998 TempObjects.clear(); 2999 } 3000 3001 PtrRtCheck.Need = NeedRTCheck; 3002 if (NeedRTCheck && !CanDoRT) { 3003 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " << 3004 "the array bounds.\n"); 3005 PtrRtCheck.reset(); 3006 return false; 3007 } 3008 3009 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") << 3010 " need a runtime memory check.\n"); 3011 return true; 3012} 3013 3014static bool hasMultipleUsesOf(Instruction *I, 3015 SmallPtrSet<Instruction *, 8> &Insts) { 3016 unsigned NumUses = 0; 3017 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) { 3018 if (Insts.count(dyn_cast<Instruction>(*Use))) 3019 ++NumUses; 3020 if (NumUses > 1) 3021 return true; 3022 } 3023 3024 return false; 3025} 3026 3027static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) { 3028 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) 3029 if (!Set.count(dyn_cast<Instruction>(*Use))) 3030 return false; 3031 return true; 3032} 3033 3034bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, 3035 ReductionKind Kind) { 3036 if (Phi->getNumIncomingValues() != 2) 3037 return false; 3038 3039 // Reduction variables are only found in the loop header block. 3040 if (Phi->getParent() != TheLoop->getHeader()) 3041 return false; 3042 3043 // Obtain the reduction start value from the value that comes from the loop 3044 // preheader. 3045 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 3046 3047 // ExitInstruction is the single value which is used outside the loop. 3048 // We only allow for a single reduction value to be used outside the loop. 3049 // This includes users of the reduction, variables (which form a cycle 3050 // which ends in the phi node). 3051 Instruction *ExitInstruction = 0; 3052 // Indicates that we found a reduction operation in our scan. 3053 bool FoundReduxOp = false; 3054 3055 // We start with the PHI node and scan for all of the users of this 3056 // instruction. All users must be instructions that can be used as reduction 3057 // variables (such as ADD). We must have a single out-of-block user. The cycle 3058 // must include the original PHI. 3059 bool FoundStartPHI = false; 3060 3061 // To recognize min/max patterns formed by a icmp select sequence, we store 3062 // the number of instruction we saw from the recognized min/max pattern, 3063 // to make sure we only see exactly the two instructions. 3064 unsigned NumCmpSelectPatternInst = 0; 3065 ReductionInstDesc ReduxDesc(false, 0); 3066 3067 SmallPtrSet<Instruction *, 8> VisitedInsts; 3068 SmallVector<Instruction *, 8> Worklist; 3069 Worklist.push_back(Phi); 3070 VisitedInsts.insert(Phi); 3071 3072 // A value in the reduction can be used: 3073 // - By the reduction: 3074 // - Reduction operation: 3075 // - One use of reduction value (safe). 3076 // - Multiple use of reduction value (not safe). 3077 // - PHI: 3078 // - All uses of the PHI must be the reduction (safe). 3079 // - Otherwise, not safe. 3080 // - By one instruction outside of the loop (safe). 3081 // - By further instructions outside of the loop (not safe). 3082 // - By an instruction that is not part of the reduction (not safe). 3083 // This is either: 3084 // * An instruction type other than PHI or the reduction operation. 3085 // * A PHI in the header other than the initial PHI. 3086 while (!Worklist.empty()) { 3087 Instruction *Cur = Worklist.back(); 3088 Worklist.pop_back(); 3089 3090 // No Users. 3091 // If the instruction has no users then this is a broken chain and can't be 3092 // a reduction variable. 3093 if (Cur->use_empty()) 3094 return false; 3095 3096 bool IsAPhi = isa<PHINode>(Cur); 3097 3098 // A header PHI use other than the original PHI. 3099 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent()) 3100 return false; 3101 3102 // Reductions of instructions such as Div, and Sub is only possible if the 3103 // LHS is the reduction variable. 3104 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) && 3105 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) && 3106 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0)))) 3107 return false; 3108 3109 // Any reduction instruction must be of one of the allowed kinds. 3110 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc); 3111 if (!ReduxDesc.IsReduction) 3112 return false; 3113 3114 // A reduction operation must only have one use of the reduction value. 3115 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax && 3116 hasMultipleUsesOf(Cur, VisitedInsts)) 3117 return false; 3118 3119 // All inputs to a PHI node must be a reduction value. 3120 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts)) 3121 return false; 3122 3123 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) || 3124 isa<SelectInst>(Cur))) 3125 ++NumCmpSelectPatternInst; 3126 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) || 3127 isa<SelectInst>(Cur))) 3128 ++NumCmpSelectPatternInst; 3129 3130 // Check whether we found a reduction operator. 3131 FoundReduxOp |= !IsAPhi; 3132 3133 // Process users of current instruction. Push non PHI nodes after PHI nodes 3134 // onto the stack. This way we are going to have seen all inputs to PHI 3135 // nodes once we get to them. 3136 SmallVector<Instruction *, 8> NonPHIs; 3137 SmallVector<Instruction *, 8> PHIs; 3138 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E; 3139 ++UI) { 3140 Instruction *Usr = cast<Instruction>(*UI); 3141 3142 // Check if we found the exit user. 3143 BasicBlock *Parent = Usr->getParent(); 3144 if (!TheLoop->contains(Parent)) { 3145 // Exit if you find multiple outside users. 3146 if (ExitInstruction != 0) 3147 return false; 3148 ExitInstruction = Cur; 3149 continue; 3150 } 3151 3152 // Process instructions only once (termination). 3153 if (VisitedInsts.insert(Usr)) { 3154 if (isa<PHINode>(Usr)) 3155 PHIs.push_back(Usr); 3156 else 3157 NonPHIs.push_back(Usr); 3158 } 3159 // Remember that we completed the cycle. 3160 if (Usr == Phi) 3161 FoundStartPHI = true; 3162 } 3163 Worklist.append(PHIs.begin(), PHIs.end()); 3164 Worklist.append(NonPHIs.begin(), NonPHIs.end()); 3165 } 3166 3167 // This means we have seen one but not the other instruction of the 3168 // pattern or more than just a select and cmp. 3169 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) && 3170 NumCmpSelectPatternInst != 2) 3171 return false; 3172 3173 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction) 3174 return false; 3175 3176 // We found a reduction var if we have reached the original phi node and we 3177 // only have a single instruction with out-of-loop users. 3178 3179 // This instruction is allowed to have out-of-loop users. 3180 AllowedExit.insert(ExitInstruction); 3181 3182 // Save the description of this reduction variable. 3183 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind, 3184 ReduxDesc.MinMaxKind); 3185 Reductions[Phi] = RD; 3186 // We've ended the cycle. This is a reduction variable if we have an 3187 // outside user and it has a binary op. 3188 3189 return true; 3190} 3191 3192/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 3193/// pattern corresponding to a min(X, Y) or max(X, Y). 3194LoopVectorizationLegality::ReductionInstDesc 3195LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, 3196 ReductionInstDesc &Prev) { 3197 3198 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) && 3199 "Expect a select instruction"); 3200 Instruction *Cmp = 0; 3201 SelectInst *Select = 0; 3202 3203 // We must handle the select(cmp()) as a single instruction. Advance to the 3204 // select. 3205 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) { 3206 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin()))) 3207 return ReductionInstDesc(false, I); 3208 return ReductionInstDesc(Select, Prev.MinMaxKind); 3209 } 3210 3211 // Only handle single use cases for now. 3212 if (!(Select = dyn_cast<SelectInst>(I))) 3213 return ReductionInstDesc(false, I); 3214 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) && 3215 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0)))) 3216 return ReductionInstDesc(false, I); 3217 if (!Cmp->hasOneUse()) 3218 return ReductionInstDesc(false, I); 3219 3220 Value *CmpLeft; 3221 Value *CmpRight; 3222 3223 // Look for a min/max pattern. 3224 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3225 return ReductionInstDesc(Select, MRK_UIntMin); 3226 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3227 return ReductionInstDesc(Select, MRK_UIntMax); 3228 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3229 return ReductionInstDesc(Select, MRK_SIntMax); 3230 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3231 return ReductionInstDesc(Select, MRK_SIntMin); 3232 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3233 return ReductionInstDesc(Select, MRK_FloatMin); 3234 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3235 return ReductionInstDesc(Select, MRK_FloatMax); 3236 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3237 return ReductionInstDesc(Select, MRK_FloatMin); 3238 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3239 return ReductionInstDesc(Select, MRK_FloatMax); 3240 3241 return ReductionInstDesc(false, I); 3242} 3243 3244LoopVectorizationLegality::ReductionInstDesc 3245LoopVectorizationLegality::isReductionInstr(Instruction *I, 3246 ReductionKind Kind, 3247 ReductionInstDesc &Prev) { 3248 bool FP = I->getType()->isFloatingPointTy(); 3249 bool FastMath = (FP && I->isCommutative() && I->isAssociative()); 3250 switch (I->getOpcode()) { 3251 default: 3252 return ReductionInstDesc(false, I); 3253 case Instruction::PHI: 3254 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd && 3255 Kind != RK_FloatMinMax)) 3256 return ReductionInstDesc(false, I); 3257 return ReductionInstDesc(I, Prev.MinMaxKind); 3258 case Instruction::Sub: 3259 case Instruction::Add: 3260 return ReductionInstDesc(Kind == RK_IntegerAdd, I); 3261 case Instruction::Mul: 3262 return ReductionInstDesc(Kind == RK_IntegerMult, I); 3263 case Instruction::And: 3264 return ReductionInstDesc(Kind == RK_IntegerAnd, I); 3265 case Instruction::Or: 3266 return ReductionInstDesc(Kind == RK_IntegerOr, I); 3267 case Instruction::Xor: 3268 return ReductionInstDesc(Kind == RK_IntegerXor, I); 3269 case Instruction::FMul: 3270 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I); 3271 case Instruction::FAdd: 3272 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I); 3273 case Instruction::FCmp: 3274 case Instruction::ICmp: 3275 case Instruction::Select: 3276 if (Kind != RK_IntegerMinMax && 3277 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax)) 3278 return ReductionInstDesc(false, I); 3279 return isMinMaxSelectCmpPattern(I, Prev); 3280 } 3281} 3282 3283LoopVectorizationLegality::InductionKind 3284LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { 3285 Type *PhiTy = Phi->getType(); 3286 // We only handle integer and pointer inductions variables. 3287 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 3288 return IK_NoInduction; 3289 3290 // Check that the PHI is consecutive. 3291 const SCEV *PhiScev = SE->getSCEV(Phi); 3292 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 3293 if (!AR) { 3294 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 3295 return IK_NoInduction; 3296 } 3297 const SCEV *Step = AR->getStepRecurrence(*SE); 3298 3299 // Integer inductions need to have a stride of one. 3300 if (PhiTy->isIntegerTy()) { 3301 if (Step->isOne()) 3302 return IK_IntInduction; 3303 if (Step->isAllOnesValue()) 3304 return IK_ReverseIntInduction; 3305 return IK_NoInduction; 3306 } 3307 3308 // Calculate the pointer stride and check if it is consecutive. 3309 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 3310 if (!C) 3311 return IK_NoInduction; 3312 3313 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 3314 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType()); 3315 if (C->getValue()->equalsInt(Size)) 3316 return IK_PtrInduction; 3317 else if (C->getValue()->equalsInt(0 - Size)) 3318 return IK_ReversePtrInduction; 3319 3320 return IK_NoInduction; 3321} 3322 3323bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 3324 Value *In0 = const_cast<Value*>(V); 3325 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 3326 if (!PN) 3327 return false; 3328 3329 return Inductions.count(PN); 3330} 3331 3332bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 3333 assert(TheLoop->contains(BB) && "Unknown block used"); 3334 3335 // Blocks that do not dominate the latch need predication. 3336 BasicBlock* Latch = TheLoop->getLoopLatch(); 3337 return !DT->dominates(BB, Latch); 3338} 3339 3340bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { 3341 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3342 // We might be able to hoist the load. 3343 if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it)) 3344 return false; 3345 3346 // We don't predicate stores at the moment. 3347 if (it->mayWriteToMemory() || it->mayThrow()) 3348 return false; 3349 3350 // The instructions below can trap. 3351 switch (it->getOpcode()) { 3352 default: continue; 3353 case Instruction::UDiv: 3354 case Instruction::SDiv: 3355 case Instruction::URem: 3356 case Instruction::SRem: 3357 return false; 3358 } 3359 } 3360 3361 return true; 3362} 3363 3364bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) { 3365 const SCEV *PhiScev = SE->getSCEV(Ptr); 3366 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 3367 if (!AR) 3368 return false; 3369 3370 return AR->isAffine(); 3371} 3372 3373LoopVectorizationCostModel::VectorizationFactor 3374LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize, 3375 unsigned UserVF) { 3376 // Width 1 means no vectorize 3377 VectorizationFactor Factor = { 1U, 0U }; 3378 if (OptForSize && Legal->getRuntimePointerCheck()->Need) { 3379 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); 3380 return Factor; 3381 } 3382 3383 // Find the trip count. 3384 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch()); 3385 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n"); 3386 3387 unsigned WidestType = getWidestType(); 3388 unsigned WidestRegister = TTI.getRegisterBitWidth(true); 3389 unsigned MaxVectorSize = WidestRegister / WidestType; 3390 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n"); 3391 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n"); 3392 3393 if (MaxVectorSize == 0) { 3394 DEBUG(dbgs() << "LV: The target has no vector registers.\n"); 3395 MaxVectorSize = 1; 3396 } 3397 3398 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements" 3399 " into one vector!"); 3400 3401 unsigned VF = MaxVectorSize; 3402 3403 // If we optimize the program for size, avoid creating the tail loop. 3404 if (OptForSize) { 3405 // If we are unable to calculate the trip count then don't try to vectorize. 3406 if (TC < 2) { 3407 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 3408 return Factor; 3409 } 3410 3411 // Find the maximum SIMD width that can fit within the trip count. 3412 VF = TC % MaxVectorSize; 3413 3414 if (VF == 0) 3415 VF = MaxVectorSize; 3416 3417 // If the trip count that we found modulo the vectorization factor is not 3418 // zero then we require a tail. 3419 if (VF < 2) { 3420 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 3421 return Factor; 3422 } 3423 } 3424 3425 if (UserVF != 0) { 3426 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 3427 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n"); 3428 3429 Factor.Width = UserVF; 3430 return Factor; 3431 } 3432 3433 float Cost = expectedCost(1); 3434 unsigned Width = 1; 3435 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n"); 3436 for (unsigned i=2; i <= VF; i*=2) { 3437 // Notice that the vector loop needs to be executed less times, so 3438 // we need to divide the cost of the vector loops by the width of 3439 // the vector elements. 3440 float VectorCost = expectedCost(i) / (float)i; 3441 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " << 3442 (int)VectorCost << ".\n"); 3443 if (VectorCost < Cost) { 3444 Cost = VectorCost; 3445 Width = i; 3446 } 3447 } 3448 3449 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n"); 3450 Factor.Width = Width; 3451 Factor.Cost = Width * Cost; 3452 return Factor; 3453} 3454 3455unsigned LoopVectorizationCostModel::getWidestType() { 3456 unsigned MaxWidth = 8; 3457 3458 // For each block. 3459 for (Loop::block_iterator bb = TheLoop->block_begin(), 3460 be = TheLoop->block_end(); bb != be; ++bb) { 3461 BasicBlock *BB = *bb; 3462 3463 // For each instruction in the loop. 3464 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3465 Type *T = it->getType(); 3466 3467 // Only examine Loads, Stores and PHINodes. 3468 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it)) 3469 continue; 3470 3471 // Examine PHI nodes that are reduction variables. 3472 if (PHINode *PN = dyn_cast<PHINode>(it)) 3473 if (!Legal->getReductionVars()->count(PN)) 3474 continue; 3475 3476 // Examine the stored values. 3477 if (StoreInst *ST = dyn_cast<StoreInst>(it)) 3478 T = ST->getValueOperand()->getType(); 3479 3480 // Ignore loaded pointer types and stored pointer types that are not 3481 // consecutive. However, we do want to take consecutive stores/loads of 3482 // pointer vectors into account. 3483 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it)) 3484 continue; 3485 3486 MaxWidth = std::max(MaxWidth, 3487 (unsigned)DL->getTypeSizeInBits(T->getScalarType())); 3488 } 3489 } 3490 3491 return MaxWidth; 3492} 3493 3494unsigned 3495LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, 3496 unsigned UserUF, 3497 unsigned VF, 3498 unsigned LoopCost) { 3499 3500 // -- The unroll heuristics -- 3501 // We unroll the loop in order to expose ILP and reduce the loop overhead. 3502 // There are many micro-architectural considerations that we can't predict 3503 // at this level. For example frontend pressure (on decode or fetch) due to 3504 // code size, or the number and capabilities of the execution ports. 3505 // 3506 // We use the following heuristics to select the unroll factor: 3507 // 1. If the code has reductions the we unroll in order to break the cross 3508 // iteration dependency. 3509 // 2. If the loop is really small then we unroll in order to reduce the loop 3510 // overhead. 3511 // 3. We don't unroll if we think that we will spill registers to memory due 3512 // to the increased register pressure. 3513 3514 // Use the user preference, unless 'auto' is selected. 3515 if (UserUF != 0) 3516 return UserUF; 3517 3518 // When we optimize for size we don't unroll. 3519 if (OptForSize) 3520 return 1; 3521 3522 // Do not unroll loops with a relatively small trip count. 3523 unsigned TC = SE->getSmallConstantTripCount(TheLoop, 3524 TheLoop->getLoopLatch()); 3525 if (TC > 1 && TC < TinyTripCountUnrollThreshold) 3526 return 1; 3527 3528 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true); 3529 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters << 3530 " vector registers\n"); 3531 3532 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); 3533 // We divide by these constants so assume that we have at least one 3534 // instruction that uses at least one register. 3535 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 3536 R.NumInstructions = std::max(R.NumInstructions, 1U); 3537 3538 // We calculate the unroll factor using the following formula. 3539 // Subtract the number of loop invariants from the number of available 3540 // registers. These registers are used by all of the unrolled instances. 3541 // Next, divide the remaining registers by the number of registers that is 3542 // required by the loop, in order to estimate how many parallel instances 3543 // fit without causing spills. 3544 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers; 3545 3546 // Clamp the unroll factor ranges to reasonable factors. 3547 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor(); 3548 3549 // If we did not calculate the cost for VF (because the user selected the VF) 3550 // then we calculate the cost of VF here. 3551 if (LoopCost == 0) 3552 LoopCost = expectedCost(VF); 3553 3554 // Clamp the calculated UF to be between the 1 and the max unroll factor 3555 // that the target allows. 3556 if (UF > MaxUnrollSize) 3557 UF = MaxUnrollSize; 3558 else if (UF < 1) 3559 UF = 1; 3560 3561 if (Legal->getReductionVars()->size()) { 3562 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n"); 3563 return UF; 3564 } 3565 3566 // We want to unroll tiny loops in order to reduce the loop overhead. 3567 // We assume that the cost overhead is 1 and we use the cost model 3568 // to estimate the cost of the loop and unroll until the cost of the 3569 // loop overhead is about 5% of the cost of the loop. 3570 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n"); 3571 if (LoopCost < 20) { 3572 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n"); 3573 unsigned NewUF = 20/LoopCost + 1; 3574 return std::min(NewUF, UF); 3575 } 3576 3577 DEBUG(dbgs() << "LV: Not Unrolling. \n"); 3578 return 1; 3579} 3580 3581LoopVectorizationCostModel::RegisterUsage 3582LoopVectorizationCostModel::calculateRegisterUsage() { 3583 // This function calculates the register usage by measuring the highest number 3584 // of values that are alive at a single location. Obviously, this is a very 3585 // rough estimation. We scan the loop in a topological order in order and 3586 // assign a number to each instruction. We use RPO to ensure that defs are 3587 // met before their users. We assume that each instruction that has in-loop 3588 // users starts an interval. We record every time that an in-loop value is 3589 // used, so we have a list of the first and last occurrences of each 3590 // instruction. Next, we transpose this data structure into a multi map that 3591 // holds the list of intervals that *end* at a specific location. This multi 3592 // map allows us to perform a linear search. We scan the instructions linearly 3593 // and record each time that a new interval starts, by placing it in a set. 3594 // If we find this value in the multi-map then we remove it from the set. 3595 // The max register usage is the maximum size of the set. 3596 // We also search for instructions that are defined outside the loop, but are 3597 // used inside the loop. We need this number separately from the max-interval 3598 // usage number because when we unroll, loop-invariant values do not take 3599 // more register. 3600 LoopBlocksDFS DFS(TheLoop); 3601 DFS.perform(LI); 3602 3603 RegisterUsage R; 3604 R.NumInstructions = 0; 3605 3606 // Each 'key' in the map opens a new interval. The values 3607 // of the map are the index of the 'last seen' usage of the 3608 // instruction that is the key. 3609 typedef DenseMap<Instruction*, unsigned> IntervalMap; 3610 // Maps instruction to its index. 3611 DenseMap<unsigned, Instruction*> IdxToInstr; 3612 // Marks the end of each interval. 3613 IntervalMap EndPoint; 3614 // Saves the list of instruction indices that are used in the loop. 3615 SmallSet<Instruction*, 8> Ends; 3616 // Saves the list of values that are used in the loop but are 3617 // defined outside the loop, such as arguments and constants. 3618 SmallPtrSet<Value*, 8> LoopInvariants; 3619 3620 unsigned Index = 0; 3621 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 3622 be = DFS.endRPO(); bb != be; ++bb) { 3623 R.NumInstructions += (*bb)->size(); 3624 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 3625 ++it) { 3626 Instruction *I = it; 3627 IdxToInstr[Index++] = I; 3628 3629 // Save the end location of each USE. 3630 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 3631 Value *U = I->getOperand(i); 3632 Instruction *Instr = dyn_cast<Instruction>(U); 3633 3634 // Ignore non-instruction values such as arguments, constants, etc. 3635 if (!Instr) continue; 3636 3637 // If this instruction is outside the loop then record it and continue. 3638 if (!TheLoop->contains(Instr)) { 3639 LoopInvariants.insert(Instr); 3640 continue; 3641 } 3642 3643 // Overwrite previous end points. 3644 EndPoint[Instr] = Index; 3645 Ends.insert(Instr); 3646 } 3647 } 3648 } 3649 3650 // Saves the list of intervals that end with the index in 'key'. 3651 typedef SmallVector<Instruction*, 2> InstrList; 3652 DenseMap<unsigned, InstrList> TransposeEnds; 3653 3654 // Transpose the EndPoints to a list of values that end at each index. 3655 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 3656 it != e; ++it) 3657 TransposeEnds[it->second].push_back(it->first); 3658 3659 SmallSet<Instruction*, 8> OpenIntervals; 3660 unsigned MaxUsage = 0; 3661 3662 3663 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 3664 for (unsigned int i = 0; i < Index; ++i) { 3665 Instruction *I = IdxToInstr[i]; 3666 // Ignore instructions that are never used within the loop. 3667 if (!Ends.count(I)) continue; 3668 3669 // Remove all of the instructions that end at this location. 3670 InstrList &List = TransposeEnds[i]; 3671 for (unsigned int j=0, e = List.size(); j < e; ++j) 3672 OpenIntervals.erase(List[j]); 3673 3674 // Count the number of live interals. 3675 MaxUsage = std::max(MaxUsage, OpenIntervals.size()); 3676 3677 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << 3678 OpenIntervals.size() <<"\n"); 3679 3680 // Add the current instruction to the list of open intervals. 3681 OpenIntervals.insert(I); 3682 } 3683 3684 unsigned Invariant = LoopInvariants.size(); 3685 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n"); 3686 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n"); 3687 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n"); 3688 3689 R.LoopInvariantRegs = Invariant; 3690 R.MaxLocalUsers = MaxUsage; 3691 return R; 3692} 3693 3694unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 3695 unsigned Cost = 0; 3696 3697 // For each block. 3698 for (Loop::block_iterator bb = TheLoop->block_begin(), 3699 be = TheLoop->block_end(); bb != be; ++bb) { 3700 unsigned BlockCost = 0; 3701 BasicBlock *BB = *bb; 3702 3703 // For each instruction in the old loop. 3704 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3705 // Skip dbg intrinsics. 3706 if (isa<DbgInfoIntrinsic>(it)) 3707 continue; 3708 3709 unsigned C = getInstructionCost(it, VF); 3710 Cost += C; 3711 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " << 3712 VF << " For instruction: "<< *it << "\n"); 3713 } 3714 3715 // We assume that if-converted blocks have a 50% chance of being executed. 3716 // When the code is scalar then some of the blocks are avoided due to CF. 3717 // When the code is vectorized we execute all code paths. 3718 if (Legal->blockNeedsPredication(*bb) && VF == 1) 3719 BlockCost /= 2; 3720 3721 Cost += BlockCost; 3722 } 3723 3724 return Cost; 3725} 3726 3727unsigned 3728LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 3729 // If we know that this instruction will remain uniform, check the cost of 3730 // the scalar version. 3731 if (Legal->isUniformAfterVectorization(I)) 3732 VF = 1; 3733 3734 Type *RetTy = I->getType(); 3735 Type *VectorTy = ToVectorTy(RetTy, VF); 3736 3737 // TODO: We need to estimate the cost of intrinsic calls. 3738 switch (I->getOpcode()) { 3739 case Instruction::GetElementPtr: 3740 // We mark this instruction as zero-cost because the cost of GEPs in 3741 // vectorized code depends on whether the corresponding memory instruction 3742 // is scalarized or not. Therefore, we handle GEPs with the memory 3743 // instruction cost. 3744 return 0; 3745 case Instruction::Br: { 3746 return TTI.getCFInstrCost(I->getOpcode()); 3747 } 3748 case Instruction::PHI: 3749 //TODO: IF-converted IFs become selects. 3750 return 0; 3751 case Instruction::Add: 3752 case Instruction::FAdd: 3753 case Instruction::Sub: 3754 case Instruction::FSub: 3755 case Instruction::Mul: 3756 case Instruction::FMul: 3757 case Instruction::UDiv: 3758 case Instruction::SDiv: 3759 case Instruction::FDiv: 3760 case Instruction::URem: 3761 case Instruction::SRem: 3762 case Instruction::FRem: 3763 case Instruction::Shl: 3764 case Instruction::LShr: 3765 case Instruction::AShr: 3766 case Instruction::And: 3767 case Instruction::Or: 3768 case Instruction::Xor: { 3769 // Certain instructions can be cheaper to vectorize if they have a constant 3770 // second vector operand. One example of this are shifts on x86. 3771 TargetTransformInfo::OperandValueKind Op1VK = 3772 TargetTransformInfo::OK_AnyValue; 3773 TargetTransformInfo::OperandValueKind Op2VK = 3774 TargetTransformInfo::OK_AnyValue; 3775 3776 if (isa<ConstantInt>(I->getOperand(1))) 3777 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 3778 3779 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK); 3780 } 3781 case Instruction::Select: { 3782 SelectInst *SI = cast<SelectInst>(I); 3783 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 3784 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 3785 Type *CondTy = SI->getCondition()->getType(); 3786 if (!ScalarCond) 3787 CondTy = VectorType::get(CondTy, VF); 3788 3789 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 3790 } 3791 case Instruction::ICmp: 3792 case Instruction::FCmp: { 3793 Type *ValTy = I->getOperand(0)->getType(); 3794 VectorTy = ToVectorTy(ValTy, VF); 3795 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); 3796 } 3797 case Instruction::Store: 3798 case Instruction::Load: { 3799 StoreInst *SI = dyn_cast<StoreInst>(I); 3800 LoadInst *LI = dyn_cast<LoadInst>(I); 3801 Type *ValTy = (SI ? SI->getValueOperand()->getType() : 3802 LI->getType()); 3803 VectorTy = ToVectorTy(ValTy, VF); 3804 3805 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); 3806 unsigned AS = SI ? SI->getPointerAddressSpace() : 3807 LI->getPointerAddressSpace(); 3808 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); 3809 // We add the cost of address computation here instead of with the gep 3810 // instruction because only here we know whether the operation is 3811 // scalarized. 3812 if (VF == 1) 3813 return TTI.getAddressComputationCost(VectorTy) + 3814 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 3815 3816 // Scalarized loads/stores. 3817 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 3818 bool Reverse = ConsecutiveStride < 0; 3819 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy); 3820 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF; 3821 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { 3822 unsigned Cost = 0; 3823 // The cost of extracting from the value vector and pointer vector. 3824 Type *PtrTy = ToVectorTy(Ptr->getType(), VF); 3825 for (unsigned i = 0; i < VF; ++i) { 3826 // The cost of extracting the pointer operand. 3827 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); 3828 // In case of STORE, the cost of ExtractElement from the vector. 3829 // In case of LOAD, the cost of InsertElement into the returned 3830 // vector. 3831 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : 3832 Instruction::InsertElement, 3833 VectorTy, i); 3834 } 3835 3836 // The cost of the scalar loads/stores. 3837 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType()); 3838 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), 3839 Alignment, AS); 3840 return Cost; 3841 } 3842 3843 // Wide load/stores. 3844 unsigned Cost = TTI.getAddressComputationCost(VectorTy); 3845 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 3846 3847 if (Reverse) 3848 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, 3849 VectorTy, 0); 3850 return Cost; 3851 } 3852 case Instruction::ZExt: 3853 case Instruction::SExt: 3854 case Instruction::FPToUI: 3855 case Instruction::FPToSI: 3856 case Instruction::FPExt: 3857 case Instruction::PtrToInt: 3858 case Instruction::IntToPtr: 3859 case Instruction::SIToFP: 3860 case Instruction::UIToFP: 3861 case Instruction::Trunc: 3862 case Instruction::FPTrunc: 3863 case Instruction::BitCast: { 3864 // We optimize the truncation of induction variable. 3865 // The cost of these is the same as the scalar operation. 3866 if (I->getOpcode() == Instruction::Trunc && 3867 Legal->isInductionVariable(I->getOperand(0))) 3868 return TTI.getCastInstrCost(I->getOpcode(), I->getType(), 3869 I->getOperand(0)->getType()); 3870 3871 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); 3872 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 3873 } 3874 case Instruction::Call: { 3875 CallInst *CI = cast<CallInst>(I); 3876 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 3877 assert(ID && "Not an intrinsic call!"); 3878 Type *RetTy = ToVectorTy(CI->getType(), VF); 3879 SmallVector<Type*, 4> Tys; 3880 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 3881 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 3882 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); 3883 } 3884 default: { 3885 // We are scalarizing the instruction. Return the cost of the scalar 3886 // instruction, plus the cost of insert and extract into vector 3887 // elements, times the vector width. 3888 unsigned Cost = 0; 3889 3890 if (!RetTy->isVoidTy() && VF != 1) { 3891 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, 3892 VectorTy); 3893 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, 3894 VectorTy); 3895 3896 // The cost of inserting the results plus extracting each one of the 3897 // operands. 3898 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 3899 } 3900 3901 // The cost of executing VF copies of the scalar instruction. This opcode 3902 // is unknown. Assume that it is the same as 'mul'. 3903 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); 3904 return Cost; 3905 } 3906 }// end of switch. 3907} 3908 3909Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) { 3910 if (Scalar->isVoidTy() || VF == 1) 3911 return Scalar; 3912 return VectorType::get(Scalar, VF); 3913} 3914 3915char LoopVectorize::ID = 0; 3916static const char lv_name[] = "Loop Vectorization"; 3917INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 3918INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3919INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3920INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3921INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 3922INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 3923 3924namespace llvm { 3925 Pass *createLoopVectorizePass() { 3926 return new LoopVectorize(); 3927 } 3928} 3929 3930bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { 3931 // Check for a store. 3932 if (StoreInst *ST = dyn_cast<StoreInst>(Inst)) 3933 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; 3934 3935 // Check for a load. 3936 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 3937 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; 3938 3939 return false; 3940} 3941