LoopVectorize.cpp revision f3252b12e02b1fcf01abf0a79b761c53de5985d0
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. Legalization of the IR is done 12// in the codegen. However, the vectorizes uses (will use) the codegen 13// interfaces to generate IR that is likely to result in an optimal binary. 14// 15// The loop vectorizer combines consecutive loop iteration 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/raw_ostream.h" 82#include "llvm/Transforms/Scalar.h" 83#include "llvm/Transforms/Utils/BasicBlockUtils.h" 84#include "llvm/Transforms/Utils/Local.h" 85#include <algorithm> 86#include <map> 87 88using namespace llvm; 89 90static cl::opt<unsigned> 91VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden, 92 cl::desc("Sets the SIMD width. Zero is autoselect.")); 93 94static cl::opt<unsigned> 95VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden, 96 cl::desc("Sets the vectorization unroll count. " 97 "Zero is autoselect.")); 98 99static cl::opt<bool> 100EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, 101 cl::desc("Enable if-conversion during vectorization.")); 102 103/// We don't vectorize loops with a known constant trip count below this number. 104static const unsigned TinyTripCountThreshold = 16; 105 106/// When performing a runtime memory check, do not check more than this 107/// number of pointers. Notice that the check is quadratic! 108static const unsigned RuntimeMemoryCheckThreshold = 4; 109 110/// This is the highest vector width that we try to generate. 111static const unsigned MaxVectorSize = 8; 112 113/// This is the highest Unroll Factor. 114static const unsigned MaxUnrollSize = 4; 115 116namespace { 117 118// Forward declarations. 119class LoopVectorizationLegality; 120class LoopVectorizationCostModel; 121 122/// InnerLoopVectorizer vectorizes loops which contain only one basic 123/// block to a specified vectorization factor (VF). 124/// This class performs the widening of scalars into vectors, or multiple 125/// scalars. This class also implements the following features: 126/// * It inserts an epilogue loop for handling loops that don't have iteration 127/// counts that are known to be a multiple of the vectorization factor. 128/// * It handles the code generation for reduction variables. 129/// * Scalarization (implementation using scalars) of un-vectorizable 130/// instructions. 131/// InnerLoopVectorizer does not perform any vectorization-legality 132/// checks, and relies on the caller to check for the different legality 133/// aspects. The InnerLoopVectorizer relies on the 134/// LoopVectorizationLegality class to provide information about the induction 135/// and reduction variables that were found to a given vectorization factor. 136class InnerLoopVectorizer { 137public: 138 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, 139 DominatorTree *DT, DataLayout *DL, unsigned VecWidth, 140 unsigned UnrollFactor) 141 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), VF(VecWidth), 142 UF(UnrollFactor), Builder(SE->getContext()), Induction(0), 143 OldInduction(0), WidenMap(UnrollFactor) {} 144 145 // Perform the actual loop widening (vectorization). 146 void vectorize(LoopVectorizationLegality *Legal) { 147 // Create a new empty loop. Unlink the old loop and connect the new one. 148 createEmptyLoop(Legal); 149 // Widen each instruction in the old loop to a new one in the new loop. 150 // Use the Legality module to find the induction and reduction variables. 151 vectorizeLoop(Legal); 152 // Register the new loop and update the analysis passes. 153 updateAnalysis(); 154 } 155 156private: 157 /// A small list of PHINodes. 158 typedef SmallVector<PHINode*, 4> PhiVector; 159 /// When we unroll loops we have multiple vector values for each scalar. 160 /// This data structure holds the unrolled and vectorized values that 161 /// originated from one scalar instruction. 162 typedef SmallVector<Value*, 2> VectorParts; 163 164 /// Add code that checks at runtime if the accessed arrays overlap. 165 /// Returns the comparator value or NULL if no check is needed. 166 Value *addRuntimeCheck(LoopVectorizationLegality *Legal, 167 Instruction *Loc); 168 /// Create an empty loop, based on the loop ranges of the old loop. 169 void createEmptyLoop(LoopVectorizationLegality *Legal); 170 /// Copy and widen the instructions from the old loop. 171 void vectorizeLoop(LoopVectorizationLegality *Legal); 172 173 /// A helper function that computes the predicate of the block BB, assuming 174 /// that the header block of the loop is set to True. It returns the *entry* 175 /// mask for the block BB. 176 VectorParts createBlockInMask(BasicBlock *BB); 177 /// A helper function that computes the predicate of the edge between SRC 178 /// and DST. 179 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst); 180 181 /// A helper function to vectorize a single BB within the innermost loop. 182 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB, 183 PhiVector *PV); 184 185 /// Insert the new loop to the loop hierarchy and pass manager 186 /// and update the analysis passes. 187 void updateAnalysis(); 188 189 /// This instruction is un-vectorizable. Implement it as a sequence 190 /// of scalars. 191 void scalarizeInstruction(Instruction *Instr); 192 193 /// Create a broadcast instruction. This method generates a broadcast 194 /// instruction (shuffle) for loop invariant values and for the induction 195 /// value. If this is the induction variable then we extend it to N, N+1, ... 196 /// this is needed because each iteration in the loop corresponds to a SIMD 197 /// element. 198 Value *getBroadcastInstrs(Value *V); 199 200 /// This function adds 0, 1, 2 ... to each vector element, starting at zero. 201 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...). 202 /// The sequence starts at StartIndex. 203 Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate); 204 205 /// When we go over instructions in the basic block we rely on previous 206 /// values within the current basic block or on loop invariant values. 207 /// When we widen (vectorize) values we place them in the map. If the values 208 /// are not within the map, they have to be loop invariant, so we simply 209 /// broadcast them into a vector. 210 VectorParts &getVectorValue(Value *V); 211 212 /// Get a uniform vector of constant integers. We use this to get 213 /// vectors of ones and zeros for the reduction code. 214 Constant* getUniformVector(unsigned Val, Type* ScalarTy); 215 216 /// Generate a shuffle sequence that will reverse the vector Vec. 217 Value *reverseVector(Value *Vec); 218 219 /// This is a helper class that holds the vectorizer state. It maps scalar 220 /// instructions to vector instructions. When the code is 'unrolled' then 221 /// then a single scalar value is mapped to multiple vector parts. The parts 222 /// are stored in the VectorPart type. 223 struct ValueMap { 224 /// C'tor. UnrollFactor controls the number of vectors ('parts') that 225 /// are mapped. 226 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {} 227 228 /// \return True if 'Key' is saved in the Value Map. 229 bool has(Value *Key) { return MapStoreage.count(Key); } 230 231 /// Initializes a new entry in the map. Sets all of the vector parts to the 232 /// save value in 'Val'. 233 /// \return A reference to a vector with splat values. 234 VectorParts &splat(Value *Key, Value *Val) { 235 MapStoreage[Key].clear(); 236 MapStoreage[Key].append(UF, Val); 237 return MapStoreage[Key]; 238 } 239 240 ///\return A reference to the value that is stored at 'Key'. 241 VectorParts &get(Value *Key) { 242 if (!has(Key)) 243 MapStoreage[Key].resize(UF); 244 return MapStoreage[Key]; 245 } 246 247 /// The unroll factor. Each entry in the map stores this number of vector 248 /// elements. 249 unsigned UF; 250 251 /// Map storage. We use std::map and not DenseMap because insertions to a 252 /// dense map invalidates its iterators. 253 std::map<Value*, VectorParts> MapStoreage; 254 }; 255 256 /// The original loop. 257 Loop *OrigLoop; 258 /// Scev analysis to use. 259 ScalarEvolution *SE; 260 /// Loop Info. 261 LoopInfo *LI; 262 /// Dominator Tree. 263 DominatorTree *DT; 264 /// Data Layout. 265 DataLayout *DL; 266 /// The vectorization SIMD factor to use. Each vector will have this many 267 /// vector elements. 268 unsigned VF; 269 /// The vectorization unroll factor to use. Each scalar is vectorized to this 270 /// many different vector instructions. 271 unsigned UF; 272 273 /// The builder that we use 274 IRBuilder<> Builder; 275 276 // --- Vectorization state --- 277 278 /// The vector-loop preheader. 279 BasicBlock *LoopVectorPreHeader; 280 /// The scalar-loop preheader. 281 BasicBlock *LoopScalarPreHeader; 282 /// Middle Block between the vector and the scalar. 283 BasicBlock *LoopMiddleBlock; 284 ///The ExitBlock of the scalar loop. 285 BasicBlock *LoopExitBlock; 286 ///The vector loop body. 287 BasicBlock *LoopVectorBody; 288 ///The scalar loop body. 289 BasicBlock *LoopScalarBody; 290 ///The first bypass block. 291 BasicBlock *LoopBypassBlock; 292 293 /// The new Induction variable which was added to the new block. 294 PHINode *Induction; 295 /// The induction variable of the old basic block. 296 PHINode *OldInduction; 297 /// Maps scalars to widened vectors. 298 ValueMap WidenMap; 299}; 300 301/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and 302/// to what vectorization factor. 303/// This class does not look at the profitability of vectorization, only the 304/// legality. This class has two main kinds of checks: 305/// * Memory checks - The code in canVectorizeMemory checks if vectorization 306/// will change the order of memory accesses in a way that will change the 307/// correctness of the program. 308/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory 309/// checks for a number of different conditions, such as the availability of a 310/// single induction variable, that all types are supported and vectorize-able, 311/// etc. This code reflects the capabilities of InnerLoopVectorizer. 312/// This class is also used by InnerLoopVectorizer for identifying 313/// induction variable and the different reduction variables. 314class LoopVectorizationLegality { 315public: 316 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL, 317 DominatorTree *DT) 318 : TheLoop(L), SE(SE), DL(DL), DT(DT), Induction(0) {} 319 320 /// This enum represents the kinds of reductions that we support. 321 enum ReductionKind { 322 NoReduction, ///< Not a reduction. 323 IntegerAdd, ///< Sum of numbers. 324 IntegerMult, ///< Product of numbers. 325 IntegerOr, ///< Bitwise or logical OR of numbers. 326 IntegerAnd, ///< Bitwise or logical AND of numbers. 327 IntegerXor ///< Bitwise or logical XOR of numbers. 328 }; 329 330 /// This enum represents the kinds of inductions that we support. 331 enum InductionKind { 332 NoInduction, ///< Not an induction variable. 333 IntInduction, ///< Integer induction variable. Step = 1. 334 ReverseIntInduction, ///< Reverse int induction variable. Step = -1. 335 PtrInduction ///< Pointer induction variable. Step = sizeof(elem). 336 }; 337 338 /// This POD struct holds information about reduction variables. 339 struct ReductionDescriptor { 340 ReductionDescriptor() : StartValue(0), LoopExitInstr(0), Kind(NoReduction) { 341 } 342 343 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K) 344 : StartValue(Start), LoopExitInstr(Exit), Kind(K) {} 345 346 // The starting value of the reduction. 347 // It does not have to be zero! 348 Value *StartValue; 349 // The instruction who's value is used outside the loop. 350 Instruction *LoopExitInstr; 351 // The kind of the reduction. 352 ReductionKind Kind; 353 }; 354 355 // This POD struct holds information about the memory runtime legality 356 // check that a group of pointers do not overlap. 357 struct RuntimePointerCheck { 358 RuntimePointerCheck() : Need(false) {} 359 360 /// Reset the state of the pointer runtime information. 361 void reset() { 362 Need = false; 363 Pointers.clear(); 364 Starts.clear(); 365 Ends.clear(); 366 } 367 368 /// Insert a pointer and calculate the start and end SCEVs. 369 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr); 370 371 /// This flag indicates if we need to add the runtime check. 372 bool Need; 373 /// Holds the pointers that we need to check. 374 SmallVector<Value*, 2> Pointers; 375 /// Holds the pointer value at the beginning of the loop. 376 SmallVector<const SCEV*, 2> Starts; 377 /// Holds the pointer value at the end of the loop. 378 SmallVector<const SCEV*, 2> Ends; 379 }; 380 381 /// A POD for saving information about induction variables. 382 struct InductionInfo { 383 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {} 384 InductionInfo() : StartValue(0), IK(NoInduction) {} 385 /// Start value. 386 Value *StartValue; 387 /// Induction kind. 388 InductionKind IK; 389 }; 390 391 /// ReductionList contains the reduction descriptors for all 392 /// of the reductions that were found in the loop. 393 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList; 394 395 /// InductionList saves induction variables and maps them to the 396 /// induction descriptor. 397 typedef MapVector<PHINode*, InductionInfo> InductionList; 398 399 /// Returns true if it is legal to vectorize this loop. 400 /// This does not mean that it is profitable to vectorize this 401 /// loop, only that it is legal to do so. 402 bool canVectorize(); 403 404 /// Returns the Induction variable. 405 PHINode *getInduction() { return Induction; } 406 407 /// Returns the reduction variables found in the loop. 408 ReductionList *getReductionVars() { return &Reductions; } 409 410 /// Returns the induction variables found in the loop. 411 InductionList *getInductionVars() { return &Inductions; } 412 413 /// Returns True if V is an induction variable in this loop. 414 bool isInductionVariable(const Value *V); 415 416 /// Return true if the block BB needs to be predicated in order for the loop 417 /// to be vectorized. 418 bool blockNeedsPredication(BasicBlock *BB); 419 420 /// Check if this pointer is consecutive when vectorizing. This happens 421 /// when the last index of the GEP is the induction variable, or that the 422 /// pointer itself is an induction variable. 423 /// This check allows us to vectorize A[idx] into a wide load/store. 424 /// Returns: 425 /// 0 - Stride is unknown or non consecutive. 426 /// 1 - Address is consecutive. 427 /// -1 - Address is consecutive, and decreasing. 428 int isConsecutivePtr(Value *Ptr); 429 430 /// Returns true if the value V is uniform within the loop. 431 bool isUniform(Value *V); 432 433 /// Returns true if this instruction will remain scalar after vectorization. 434 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } 435 436 /// Returns the information that we collected about runtime memory check. 437 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; } 438private: 439 /// Check if a single basic block loop is vectorizable. 440 /// At this point we know that this is a loop with a constant trip count 441 /// and we only need to check individual instructions. 442 bool canVectorizeInstrs(); 443 444 /// When we vectorize loops we may change the order in which 445 /// we read and write from memory. This method checks if it is 446 /// legal to vectorize the code, considering only memory constrains. 447 /// Returns true if the loop is vectorizable 448 bool canVectorizeMemory(); 449 450 /// Return true if we can vectorize this loop using the IF-conversion 451 /// transformation. 452 bool canVectorizeWithIfConvert(); 453 454 /// Collect the variables that need to stay uniform after vectorization. 455 void collectLoopUniforms(); 456 457 /// Return true if all of the instructions in the block can be speculatively 458 /// executed. 459 bool blockCanBePredicated(BasicBlock *BB); 460 461 /// Returns True, if 'Phi' is the kind of reduction variable for type 462 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. 463 bool AddReductionVar(PHINode *Phi, ReductionKind Kind); 464 /// Returns true if the instruction I can be a reduction variable of type 465 /// 'Kind'. 466 bool isReductionInstr(Instruction *I, ReductionKind Kind); 467 /// Returns the induction kind of Phi. This function may return NoInduction 468 /// if the PHI is not an induction variable. 469 InductionKind isInductionVariable(PHINode *Phi); 470 /// Return true if can compute the address bounds of Ptr within the loop. 471 bool hasComputableBounds(Value *Ptr); 472 473 /// The loop that we evaluate. 474 Loop *TheLoop; 475 /// Scev analysis. 476 ScalarEvolution *SE; 477 /// DataLayout analysis. 478 DataLayout *DL; 479 // Dominators. 480 DominatorTree *DT; 481 482 // --- vectorization state --- // 483 484 /// Holds the integer induction variable. This is the counter of the 485 /// loop. 486 PHINode *Induction; 487 /// Holds the reduction variables. 488 ReductionList Reductions; 489 /// Holds all of the induction variables that we found in the loop. 490 /// Notice that inductions don't need to start at zero and that induction 491 /// variables can be pointers. 492 InductionList Inductions; 493 494 /// Allowed outside users. This holds the reduction 495 /// vars which can be accessed from outside the loop. 496 SmallPtrSet<Value*, 4> AllowedExit; 497 /// This set holds the variables which are known to be uniform after 498 /// vectorization. 499 SmallPtrSet<Instruction*, 4> Uniforms; 500 /// We need to check that all of the pointers in this list are disjoint 501 /// at runtime. 502 RuntimePointerCheck PtrRtCheck; 503}; 504 505/// LoopVectorizationCostModel - estimates the expected speedups due to 506/// vectorization. 507/// In many cases vectorization is not profitable. This can happen because of 508/// a number of reasons. In this class we mainly attempt to predict the 509/// expected speedup/slowdowns due to the supported instruction set. We use the 510/// TargetTransformInfo to query the different backends for the cost of 511/// different operations. 512class LoopVectorizationCostModel { 513public: 514 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI, 515 LoopVectorizationLegality *Legal, 516 const TargetTransformInfo *TTI) 517 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI) {} 518 519 /// \return The most profitable vectorization factor. 520 /// This method checks every power of two up to VF. If UserVF is not ZERO 521 /// then this vectorization factor will be selected if vectorization is 522 /// possible. 523 unsigned selectVectorizationFactor(bool OptForSize, unsigned UserVF); 524 525 526 /// \return The most profitable unroll factor. 527 /// If UserUF is non-zero then this method finds the best unroll-factor 528 /// based on register pressure and other parameters. 529 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF); 530 531 /// \brief A struct that represents some properties of the register usage 532 /// of a loop. 533 struct RegisterUsage { 534 /// Holds the number of loop invariant values that are used in the loop. 535 unsigned LoopInvariantRegs; 536 /// Holds the maximum number of concurrent live intervals in the loop. 537 unsigned MaxLocalUsers; 538 /// Holds the number of instructions in the loop. 539 unsigned NumInstructions; 540 }; 541 542 /// \return information about the register usage of the loop. 543 RegisterUsage calculateRegisterUsage(); 544 545private: 546 /// Returns the expected execution cost. The unit of the cost does 547 /// not matter because we use the 'cost' units to compare different 548 /// vector widths. The cost that is returned is *not* normalized by 549 /// the factor width. 550 unsigned expectedCost(unsigned VF); 551 552 /// Returns the execution time cost of an instruction for a given vector 553 /// width. Vector width of one means scalar. 554 unsigned getInstructionCost(Instruction *I, unsigned VF); 555 556 /// A helper function for converting Scalar types to vector types. 557 /// If the incoming type is void, we return void. If the VF is 1, we return 558 /// the scalar type. 559 static Type* ToVectorTy(Type *Scalar, unsigned VF); 560 561 /// The loop that we evaluate. 562 Loop *TheLoop; 563 /// Scev analysis. 564 ScalarEvolution *SE; 565 /// Loop Info analysis. 566 LoopInfo *LI; 567 /// Vectorization legality. 568 LoopVectorizationLegality *Legal; 569 /// Vector target information. 570 const TargetTransformInfo *TTI; 571}; 572 573/// The LoopVectorize Pass. 574struct LoopVectorize : public LoopPass { 575 /// Pass identification, replacement for typeid 576 static char ID; 577 578 explicit LoopVectorize() : LoopPass(ID) { 579 initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); 580 } 581 582 ScalarEvolution *SE; 583 DataLayout *DL; 584 LoopInfo *LI; 585 TargetTransformInfo *TTI; 586 DominatorTree *DT; 587 588 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) { 589 // We only vectorize innermost loops. 590 if (!L->empty()) 591 return false; 592 593 SE = &getAnalysis<ScalarEvolution>(); 594 DL = getAnalysisIfAvailable<DataLayout>(); 595 LI = &getAnalysis<LoopInfo>(); 596 TTI = getAnalysisIfAvailable<TargetTransformInfo>(); 597 DT = &getAnalysis<DominatorTree>(); 598 599 DEBUG(dbgs() << "LV: Checking a loop in \"" << 600 L->getHeader()->getParent()->getName() << "\"\n"); 601 602 // Check if it is legal to vectorize the loop. 603 LoopVectorizationLegality LVL(L, SE, DL, DT); 604 if (!LVL.canVectorize()) { 605 DEBUG(dbgs() << "LV: Not vectorizing.\n"); 606 return false; 607 } 608 609 // Use the cost model. 610 LoopVectorizationCostModel CM(L, SE, LI, &LVL, TTI); 611 612 // Check the function attribues to find out if this function should be 613 // optimized for size. 614 Function *F = L->getHeader()->getParent(); 615 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize; 616 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat; 617 unsigned FnIndex = AttributeSet::FunctionIndex; 618 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr); 619 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr); 620 621 if (NoFloat) { 622 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" 623 "attribute is used.\n"); 624 return false; 625 } 626 627 unsigned VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor); 628 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll); 629 630 if (VF == 1) { 631 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); 632 return false; 633 } 634 635 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<< 636 F->getParent()->getModuleIdentifier()<<"\n"); 637 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n"); 638 639 // If we decided that it is *legal* to vectorizer the loop then do it. 640 InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF, UF); 641 LB.vectorize(&LVL); 642 643 DEBUG(verifyFunction(*L->getHeader()->getParent())); 644 return true; 645 } 646 647 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 648 LoopPass::getAnalysisUsage(AU); 649 AU.addRequiredID(LoopSimplifyID); 650 AU.addRequiredID(LCSSAID); 651 AU.addRequired<LoopInfo>(); 652 AU.addRequired<ScalarEvolution>(); 653 AU.addRequired<DominatorTree>(); 654 AU.addPreserved<LoopInfo>(); 655 AU.addPreserved<DominatorTree>(); 656 } 657 658}; 659 660} // end anonymous namespace 661 662//===----------------------------------------------------------------------===// 663// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and 664// LoopVectorizationCostModel. 665//===----------------------------------------------------------------------===// 666 667void 668LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE, 669 Loop *Lp, Value *Ptr) { 670 const SCEV *Sc = SE->getSCEV(Ptr); 671 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); 672 assert(AR && "Invalid addrec expression"); 673 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch()); 674 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE); 675 Pointers.push_back(Ptr); 676 Starts.push_back(AR->getStart()); 677 Ends.push_back(ScEnd); 678} 679 680Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { 681 // Save the current insertion location. 682 Instruction *Loc = Builder.GetInsertPoint(); 683 684 // We need to place the broadcast of invariant variables outside the loop. 685 Instruction *Instr = dyn_cast<Instruction>(V); 686 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody); 687 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr; 688 689 // Place the code for broadcasting invariant variables in the new preheader. 690 if (Invariant) 691 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); 692 693 // Broadcast the scalar into all locations in the vector. 694 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); 695 696 // Restore the builder insertion point. 697 if (Invariant) 698 Builder.SetInsertPoint(Loc); 699 700 return Shuf; 701} 702 703Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx, 704 bool Negate) { 705 assert(Val->getType()->isVectorTy() && "Must be a vector"); 706 assert(Val->getType()->getScalarType()->isIntegerTy() && 707 "Elem must be an integer"); 708 // Create the types. 709 Type *ITy = Val->getType()->getScalarType(); 710 VectorType *Ty = cast<VectorType>(Val->getType()); 711 int VLen = Ty->getNumElements(); 712 SmallVector<Constant*, 8> Indices; 713 714 // Create a vector of consecutive numbers from zero to VF. 715 for (int i = 0; i < VLen; ++i) { 716 int Idx = Negate ? (-i): i; 717 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx)); 718 } 719 720 // Add the consecutive indices to the vector value. 721 Constant *Cv = ConstantVector::get(Indices); 722 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); 723 return Builder.CreateAdd(Val, Cv, "induction"); 724} 725 726int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { 727 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr"); 728 729 // If this value is a pointer induction variable we know it is consecutive. 730 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr); 731 if (Phi && Inductions.count(Phi)) { 732 InductionInfo II = Inductions[Phi]; 733 if (PtrInduction == II.IK) 734 return 1; 735 } 736 737 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr); 738 if (!Gep) 739 return 0; 740 741 unsigned NumOperands = Gep->getNumOperands(); 742 Value *LastIndex = Gep->getOperand(NumOperands - 1); 743 744 // Check that all of the gep indices are uniform except for the last. 745 for (unsigned i = 0; i < NumOperands - 1; ++i) 746 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 747 return 0; 748 749 // We can emit wide load/stores only if the last index is the induction 750 // variable. 751 const SCEV *Last = SE->getSCEV(LastIndex); 752 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { 753 const SCEV *Step = AR->getStepRecurrence(*SE); 754 755 // The memory is consecutive because the last index is consecutive 756 // and all other indices are loop invariant. 757 if (Step->isOne()) 758 return 1; 759 if (Step->isAllOnesValue()) 760 return -1; 761 } 762 763 return 0; 764} 765 766bool LoopVectorizationLegality::isUniform(Value *V) { 767 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); 768} 769 770InnerLoopVectorizer::VectorParts& 771InnerLoopVectorizer::getVectorValue(Value *V) { 772 assert(V != Induction && "The new induction variable should not be used."); 773 assert(!V->getType()->isVectorTy() && "Can't widen a vector"); 774 775 // If we have this scalar in the map, return it. 776 if (WidenMap.has(V)) 777 return WidenMap.get(V); 778 779 // If this scalar is unknown, assume that it is a constant or that it is 780 // loop invariant. Broadcast V and save the value for future uses. 781 Value *B = getBroadcastInstrs(V); 782 WidenMap.splat(V, B); 783 return WidenMap.get(V); 784} 785 786Constant* 787InnerLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) { 788 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true)); 789} 790 791Value *InnerLoopVectorizer::reverseVector(Value *Vec) { 792 assert(Vec->getType()->isVectorTy() && "Invalid type"); 793 SmallVector<Constant*, 8> ShuffleMask; 794 for (unsigned i = 0; i < VF; ++i) 795 ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); 796 797 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), 798 ConstantVector::get(ShuffleMask), 799 "reverse"); 800} 801 802void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { 803 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 804 // Holds vector parameters or scalars, in case of uniform vals. 805 SmallVector<VectorParts, 4> Params; 806 807 // Find all of the vectorized parameters. 808 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 809 Value *SrcOp = Instr->getOperand(op); 810 811 // If we are accessing the old induction variable, use the new one. 812 if (SrcOp == OldInduction) { 813 Params.push_back(getVectorValue(SrcOp)); 814 continue; 815 } 816 817 // Try using previously calculated values. 818 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 819 820 // If the src is an instruction that appeared earlier in the basic block 821 // then it should already be vectorized. 822 if (SrcInst && OrigLoop->contains(SrcInst)) { 823 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 824 // The parameter is a vector value from earlier. 825 Params.push_back(WidenMap.get(SrcInst)); 826 } else { 827 // The parameter is a scalar from outside the loop. Maybe even a constant. 828 VectorParts Scalars; 829 Scalars.append(UF, SrcOp); 830 Params.push_back(Scalars); 831 } 832 } 833 834 assert(Params.size() == Instr->getNumOperands() && 835 "Invalid number of operands"); 836 837 // Does this instruction return a value ? 838 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 839 840 Value *UndefVec = IsVoidRetTy ? 0 : 841 UndefValue::get(VectorType::get(Instr->getType(), VF)); 842 // Create a new entry in the WidenMap and initialize it to Undef or Null. 843 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 844 845 // For each scalar that we create: 846 for (unsigned Width = 0; Width < VF; ++Width) { 847 // For each vector unroll 'part': 848 for (unsigned Part = 0; Part < UF; ++Part) { 849 Instruction *Cloned = Instr->clone(); 850 if (!IsVoidRetTy) 851 Cloned->setName(Instr->getName() + ".cloned"); 852 // Replace the operands of the cloned instrucions with extracted scalars. 853 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 854 Value *Op = Params[op][Part]; 855 // Param is a vector. Need to extract the right lane. 856 if (Op->getType()->isVectorTy()) 857 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); 858 Cloned->setOperand(op, Op); 859 } 860 861 // Place the cloned scalar in the new loop. 862 Builder.Insert(Cloned); 863 864 // If the original scalar returns a value we need to place it in a vector 865 // so that future users will be able to use it. 866 if (!IsVoidRetTy) 867 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, 868 Builder.getInt32(Width)); 869 } 870 } 871} 872 873Value* 874InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, 875 Instruction *Loc) { 876 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck = 877 Legal->getRuntimePointerCheck(); 878 879 if (!PtrRtCheck->Need) 880 return NULL; 881 882 Value *MemoryRuntimeCheck = 0; 883 unsigned NumPointers = PtrRtCheck->Pointers.size(); 884 SmallVector<Value* , 2> Starts; 885 SmallVector<Value* , 2> Ends; 886 887 SCEVExpander Exp(*SE, "induction"); 888 889 // Use this type for pointer arithmetic. 890 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0); 891 892 for (unsigned i = 0; i < NumPointers; ++i) { 893 Value *Ptr = PtrRtCheck->Pointers[i]; 894 const SCEV *Sc = SE->getSCEV(Ptr); 895 896 if (SE->isLoopInvariant(Sc, OrigLoop)) { 897 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" << 898 *Ptr <<"\n"); 899 Starts.push_back(Ptr); 900 Ends.push_back(Ptr); 901 } else { 902 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n"); 903 904 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc); 905 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc); 906 Starts.push_back(Start); 907 Ends.push_back(End); 908 } 909 } 910 911 for (unsigned i = 0; i < NumPointers; ++i) { 912 for (unsigned j = i+1; j < NumPointers; ++j) { 913 Instruction::CastOps Op = Instruction::BitCast; 914 Value *Start0 = CastInst::Create(Op, Starts[i], PtrArithTy, "bc", Loc); 915 Value *Start1 = CastInst::Create(Op, Starts[j], PtrArithTy, "bc", Loc); 916 Value *End0 = CastInst::Create(Op, Ends[i], PtrArithTy, "bc", Loc); 917 Value *End1 = CastInst::Create(Op, Ends[j], PtrArithTy, "bc", Loc); 918 919 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE, 920 Start0, End1, "bound0", Loc); 921 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE, 922 Start1, End0, "bound1", Loc); 923 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1, 924 "found.conflict", Loc); 925 if (MemoryRuntimeCheck) 926 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or, 927 MemoryRuntimeCheck, 928 IsConflict, 929 "conflict.rdx", Loc); 930 else 931 MemoryRuntimeCheck = IsConflict; 932 933 } 934 } 935 936 return MemoryRuntimeCheck; 937} 938 939void 940InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { 941 /* 942 In this function we generate a new loop. The new loop will contain 943 the vectorized instructions while the old loop will continue to run the 944 scalar remainder. 945 946 [ ] <-- vector loop bypass. 947 / | 948 / v 949 | [ ] <-- vector pre header. 950 | | 951 | v 952 | [ ] \ 953 | [ ]_| <-- vector loop. 954 | | 955 \ v 956 >[ ] <--- middle-block. 957 / | 958 / v 959 | [ ] <--- new preheader. 960 | | 961 | v 962 | [ ] \ 963 | [ ]_| <-- old scalar loop to handle remainder. 964 \ | 965 \ v 966 >[ ] <-- exit block. 967 ... 968 */ 969 970 BasicBlock *OldBasicBlock = OrigLoop->getHeader(); 971 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader(); 972 BasicBlock *ExitBlock = OrigLoop->getExitBlock(); 973 assert(ExitBlock && "Must have an exit block"); 974 975 // Some loops have a single integer induction variable, while other loops 976 // don't. One example is c++ iterators that often have multiple pointer 977 // induction variables. In the code below we also support a case where we 978 // don't have a single induction variable. 979 OldInduction = Legal->getInduction(); 980 Type *IdxTy = OldInduction ? OldInduction->getType() : 981 DL->getIntPtrType(SE->getContext()); 982 983 // Find the loop boundaries. 984 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch()); 985 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count"); 986 987 // Get the total trip count from the count by adding 1. 988 ExitCount = SE->getAddExpr(ExitCount, 989 SE->getConstant(ExitCount->getType(), 1)); 990 991 // Expand the trip count and place the new instructions in the preheader. 992 // Notice that the pre-header does not change, only the loop body. 993 SCEVExpander Exp(*SE, "induction"); 994 995 // Count holds the overall loop count (N). 996 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 997 BypassBlock->getTerminator()); 998 999 // The loop index does not have to start at Zero. Find the original start 1000 // value from the induction PHI node. If we don't have an induction variable 1001 // then we know that it starts at zero. 1002 Value *StartIdx = OldInduction ? 1003 OldInduction->getIncomingValueForBlock(BypassBlock): 1004 ConstantInt::get(IdxTy, 0); 1005 1006 assert(BypassBlock && "Invalid loop structure"); 1007 1008 // Generate the code that checks in runtime if arrays overlap. 1009 Value *MemoryRuntimeCheck = addRuntimeCheck(Legal, 1010 BypassBlock->getTerminator()); 1011 1012 // Split the single block loop into the two loop structure described above. 1013 BasicBlock *VectorPH = 1014 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); 1015 BasicBlock *VecBody = 1016 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 1017 BasicBlock *MiddleBlock = 1018 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 1019 BasicBlock *ScalarPH = 1020 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 1021 1022 // This is the location in which we add all of the logic for bypassing 1023 // the new vector loop. 1024 Instruction *Loc = BypassBlock->getTerminator(); 1025 1026 // Use this IR builder to create the loop instructions (Phi, Br, Cmp) 1027 // inside the loop. 1028 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1029 1030 // Generate the induction variable. 1031 Induction = Builder.CreatePHI(IdxTy, 2, "index"); 1032 // The loop step is equal to the vectorization factor (num of SIMD elements) 1033 // times the unroll factor (num of SIMD instructions). 1034 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 1035 1036 // We may need to extend the index in case there is a type mismatch. 1037 // We know that the count starts at zero and does not overflow. 1038 if (Count->getType() != IdxTy) { 1039 // The exit count can be of pointer type. Convert it to the correct 1040 // integer type. 1041 if (ExitCount->getType()->isPointerTy()) 1042 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc); 1043 else 1044 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc); 1045 } 1046 1047 // Add the start index to the loop count to get the new end index. 1048 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc); 1049 1050 // Now we need to generate the expression for N - (N % VF), which is 1051 // the part that the vectorized body will execute. 1052 Value *R = BinaryOperator::CreateURem(Count, Step, "n.mod.vf", Loc); 1053 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc); 1054 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx, 1055 "end.idx.rnd.down", Loc); 1056 1057 // Now, compare the new count to zero. If it is zero skip the vector loop and 1058 // jump to the scalar loop. 1059 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, 1060 IdxEndRoundDown, 1061 StartIdx, 1062 "cmp.zero", Loc); 1063 1064 // If we are using memory runtime checks, include them in. 1065 if (MemoryRuntimeCheck) 1066 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck, 1067 "CntOrMem", Loc); 1068 1069 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc); 1070 // Remove the old terminator. 1071 Loc->eraseFromParent(); 1072 1073 // We are going to resume the execution of the scalar loop. 1074 // Go over all of the induction variables that we found and fix the 1075 // PHIs that are left in the scalar version of the loop. 1076 // The starting values of PHI nodes depend on the counter of the last 1077 // iteration in the vectorized loop. 1078 // If we come from a bypass edge then we need to start from the original 1079 // start value. 1080 1081 // This variable saves the new starting index for the scalar loop. 1082 PHINode *ResumeIndex = 0; 1083 LoopVectorizationLegality::InductionList::iterator I, E; 1084 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 1085 for (I = List->begin(), E = List->end(); I != E; ++I) { 1086 PHINode *OrigPhi = I->first; 1087 LoopVectorizationLegality::InductionInfo II = I->second; 1088 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val", 1089 MiddleBlock->getTerminator()); 1090 Value *EndValue = 0; 1091 switch (II.IK) { 1092 case LoopVectorizationLegality::NoInduction: 1093 llvm_unreachable("Unknown induction"); 1094 case LoopVectorizationLegality::IntInduction: { 1095 // Handle the integer induction counter: 1096 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); 1097 assert(OrigPhi == OldInduction && "Unknown integer PHI"); 1098 // We know what the end value is. 1099 EndValue = IdxEndRoundDown; 1100 // We also know which PHI node holds it. 1101 ResumeIndex = ResumeVal; 1102 break; 1103 } 1104 case LoopVectorizationLegality::ReverseIntInduction: { 1105 // Convert the CountRoundDown variable to the PHI size. 1106 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits(); 1107 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits(); 1108 Value *CRD = CountRoundDown; 1109 if (CRDSize > IISize) 1110 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown, 1111 II.StartValue->getType(), 1112 "tr.crd", BypassBlock->getTerminator()); 1113 else if (CRDSize < IISize) 1114 CRD = CastInst::Create(Instruction::SExt, CountRoundDown, 1115 II.StartValue->getType(), 1116 "sext.crd", BypassBlock->getTerminator()); 1117 // Handle reverse integer induction counter: 1118 EndValue = BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end", 1119 BypassBlock->getTerminator()); 1120 break; 1121 } 1122 case LoopVectorizationLegality::PtrInduction: { 1123 // For pointer induction variables, calculate the offset using 1124 // the end index. 1125 EndValue = GetElementPtrInst::Create(II.StartValue, CountRoundDown, 1126 "ptr.ind.end", 1127 BypassBlock->getTerminator()); 1128 break; 1129 } 1130 }// end of case 1131 1132 // The new PHI merges the original incoming value, in case of a bypass, 1133 // or the value at the end of the vectorized loop. 1134 ResumeVal->addIncoming(II.StartValue, BypassBlock); 1135 ResumeVal->addIncoming(EndValue, VecBody); 1136 1137 // Fix the scalar body counter (PHI node). 1138 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 1139 OrigPhi->setIncomingValue(BlockIdx, ResumeVal); 1140 } 1141 1142 // If we are generating a new induction variable then we also need to 1143 // generate the code that calculates the exit value. This value is not 1144 // simply the end of the counter because we may skip the vectorized body 1145 // in case of a runtime check. 1146 if (!OldInduction){ 1147 assert(!ResumeIndex && "Unexpected resume value found"); 1148 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", 1149 MiddleBlock->getTerminator()); 1150 ResumeIndex->addIncoming(StartIdx, BypassBlock); 1151 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); 1152 } 1153 1154 // Make sure that we found the index where scalar loop needs to continue. 1155 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() && 1156 "Invalid resume Index"); 1157 1158 // Add a check in the middle block to see if we have completed 1159 // all of the iterations in the first vector loop. 1160 // If (N - N%VF) == N, then we *don't* need to run the remainder. 1161 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, 1162 ResumeIndex, "cmp.n", 1163 MiddleBlock->getTerminator()); 1164 1165 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); 1166 // Remove the old terminator. 1167 MiddleBlock->getTerminator()->eraseFromParent(); 1168 1169 // Create i+1 and fill the PHINode. 1170 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); 1171 Induction->addIncoming(StartIdx, VectorPH); 1172 Induction->addIncoming(NextIdx, VecBody); 1173 // Create the compare. 1174 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); 1175 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); 1176 1177 // Now we have two terminators. Remove the old one from the block. 1178 VecBody->getTerminator()->eraseFromParent(); 1179 1180 // Get ready to start creating new instructions into the vectorized body. 1181 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1182 1183 // Create and register the new vector loop. 1184 Loop* Lp = new Loop(); 1185 Loop *ParentLoop = OrigLoop->getParentLoop(); 1186 1187 // Insert the new loop into the loop nest and register the new basic blocks. 1188 if (ParentLoop) { 1189 ParentLoop->addChildLoop(Lp); 1190 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); 1191 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); 1192 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); 1193 } else { 1194 LI->addTopLevelLoop(Lp); 1195 } 1196 1197 Lp->addBasicBlockToLoop(VecBody, LI->getBase()); 1198 1199 // Save the state. 1200 LoopVectorPreHeader = VectorPH; 1201 LoopScalarPreHeader = ScalarPH; 1202 LoopMiddleBlock = MiddleBlock; 1203 LoopExitBlock = ExitBlock; 1204 LoopVectorBody = VecBody; 1205 LoopScalarBody = OldBasicBlock; 1206 LoopBypassBlock = BypassBlock; 1207} 1208 1209/// This function returns the identity element (or neutral element) for 1210/// the operation K. 1211static unsigned 1212getReductionIdentity(LoopVectorizationLegality::ReductionKind K) { 1213 switch (K) { 1214 case LoopVectorizationLegality::IntegerXor: 1215 case LoopVectorizationLegality::IntegerAdd: 1216 case LoopVectorizationLegality::IntegerOr: 1217 // Adding, Xoring, Oring zero to a number does not change it. 1218 return 0; 1219 case LoopVectorizationLegality::IntegerMult: 1220 // Multiplying a number by 1 does not change it. 1221 return 1; 1222 case LoopVectorizationLegality::IntegerAnd: 1223 // AND-ing a number with an all-1 value does not change it. 1224 return -1; 1225 default: 1226 llvm_unreachable("Unknown reduction kind"); 1227 } 1228} 1229 1230static bool 1231isTriviallyVectorizableIntrinsic(Instruction *Inst) { 1232 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst); 1233 if (!II) 1234 return false; 1235 switch (II->getIntrinsicID()) { 1236 case Intrinsic::sqrt: 1237 case Intrinsic::sin: 1238 case Intrinsic::cos: 1239 case Intrinsic::exp: 1240 case Intrinsic::exp2: 1241 case Intrinsic::log: 1242 case Intrinsic::log10: 1243 case Intrinsic::log2: 1244 case Intrinsic::fabs: 1245 case Intrinsic::floor: 1246 case Intrinsic::ceil: 1247 case Intrinsic::trunc: 1248 case Intrinsic::rint: 1249 case Intrinsic::nearbyint: 1250 case Intrinsic::pow: 1251 case Intrinsic::fma: 1252 case Intrinsic::fmuladd: 1253 return true; 1254 default: 1255 return false; 1256 } 1257 return false; 1258} 1259 1260void 1261InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { 1262 //===------------------------------------------------===// 1263 // 1264 // Notice: any optimization or new instruction that go 1265 // into the code below should be also be implemented in 1266 // the cost-model. 1267 // 1268 //===------------------------------------------------===// 1269 BasicBlock &BB = *OrigLoop->getHeader(); 1270 Constant *Zero = 1271 ConstantInt::get(IntegerType::getInt32Ty(BB.getContext()), 0); 1272 1273 // In order to support reduction variables we need to be able to vectorize 1274 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two 1275 // stages. First, we create a new vector PHI node with no incoming edges. 1276 // We use this value when we vectorize all of the instructions that use the 1277 // PHI. Next, after all of the instructions in the block are complete we 1278 // add the new incoming edges to the PHI. At this point all of the 1279 // instructions in the basic block are vectorized, so we can use them to 1280 // construct the PHI. 1281 PhiVector RdxPHIsToFix; 1282 1283 // Scan the loop in a topological order to ensure that defs are vectorized 1284 // before users. 1285 LoopBlocksDFS DFS(OrigLoop); 1286 DFS.perform(LI); 1287 1288 // Vectorize all of the blocks in the original loop. 1289 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 1290 be = DFS.endRPO(); bb != be; ++bb) 1291 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix); 1292 1293 // At this point every instruction in the original loop is widened to 1294 // a vector form. We are almost done. Now, we need to fix the PHI nodes 1295 // that we vectorized. The PHI nodes are currently empty because we did 1296 // not want to introduce cycles. Notice that the remaining PHI nodes 1297 // that we need to fix are reduction variables. 1298 1299 // Create the 'reduced' values for each of the induction vars. 1300 // The reduced values are the vector values that we scalarize and combine 1301 // after the loop is finished. 1302 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); 1303 it != e; ++it) { 1304 PHINode *RdxPhi = *it; 1305 assert(RdxPhi && "Unable to recover vectorized PHI"); 1306 1307 // Find the reduction variable descriptor. 1308 assert(Legal->getReductionVars()->count(RdxPhi) && 1309 "Unable to find the reduction variable"); 1310 LoopVectorizationLegality::ReductionDescriptor RdxDesc = 1311 (*Legal->getReductionVars())[RdxPhi]; 1312 1313 // We need to generate a reduction vector from the incoming scalar. 1314 // To do so, we need to generate the 'identity' vector and overide 1315 // one of the elements with the incoming scalar reduction. We need 1316 // to do it in the vector-loop preheader. 1317 Builder.SetInsertPoint(LoopBypassBlock->getTerminator()); 1318 1319 // This is the vector-clone of the value that leaves the loop. 1320 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr); 1321 Type *VecTy = VectorExit[0]->getType(); 1322 1323 // Find the reduction identity variable. Zero for addition, or, xor, 1324 // one for multiplication, -1 for And. 1325 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind), 1326 VecTy->getScalarType()); 1327 1328 // This vector is the Identity vector where the first element is the 1329 // incoming scalar reduction. 1330 Value *VectorStart = Builder.CreateInsertElement(Identity, 1331 RdxDesc.StartValue, Zero); 1332 1333 // Fix the vector-loop phi. 1334 // We created the induction variable so we know that the 1335 // preheader is the first entry. 1336 BasicBlock *VecPreheader = Induction->getIncomingBlock(0); 1337 1338 // Reductions do not have to start at zero. They can start with 1339 // any loop invariant values. 1340 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 1341 BasicBlock *Latch = OrigLoop->getLoopLatch(); 1342 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 1343 VectorParts &Val = getVectorValue(LoopVal); 1344 for (unsigned part = 0; part < UF; ++part) { 1345 // Make sure to add the reduction stat value only to the 1346 // first unroll part. 1347 Value *StartVal = (part == 0) ? VectorStart : Identity; 1348 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader); 1349 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody); 1350 } 1351 1352 // Before each round, move the insertion point right between 1353 // the PHIs and the values we are going to write. 1354 // This allows us to write both PHINodes and the extractelement 1355 // instructions. 1356 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); 1357 1358 VectorParts RdxParts; 1359 for (unsigned part = 0; part < UF; ++part) { 1360 // This PHINode contains the vectorized reduction variable, or 1361 // the initial value vector, if we bypass the vector loop. 1362 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); 1363 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); 1364 Value *StartVal = (part == 0) ? VectorStart : Identity; 1365 NewPhi->addIncoming(StartVal, LoopBypassBlock); 1366 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody); 1367 RdxParts.push_back(NewPhi); 1368 } 1369 1370 // Reduce all of the unrolled parts into a single vector. 1371 Value *ReducedPartRdx = RdxParts[0]; 1372 for (unsigned part = 1; part < UF; ++part) { 1373 switch (RdxDesc.Kind) { 1374 case LoopVectorizationLegality::IntegerAdd: 1375 ReducedPartRdx = 1376 Builder.CreateAdd(RdxParts[part], ReducedPartRdx, "add.rdx"); 1377 break; 1378 case LoopVectorizationLegality::IntegerMult: 1379 ReducedPartRdx = 1380 Builder.CreateMul(RdxParts[part], ReducedPartRdx, "mul.rdx"); 1381 break; 1382 case LoopVectorizationLegality::IntegerOr: 1383 ReducedPartRdx = 1384 Builder.CreateOr(RdxParts[part], ReducedPartRdx, "or.rdx"); 1385 break; 1386 case LoopVectorizationLegality::IntegerAnd: 1387 ReducedPartRdx = 1388 Builder.CreateAnd(RdxParts[part], ReducedPartRdx, "and.rdx"); 1389 break; 1390 case LoopVectorizationLegality::IntegerXor: 1391 ReducedPartRdx = 1392 Builder.CreateXor(RdxParts[part], ReducedPartRdx, "xor.rdx"); 1393 break; 1394 default: 1395 llvm_unreachable("Unknown reduction operation"); 1396 } 1397 } 1398 1399 1400 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 1401 // and vector ops, reducing the set of values being computed by half each 1402 // round. 1403 assert(isPowerOf2_32(VF) && 1404 "Reduction emission only supported for pow2 vectors!"); 1405 Value *TmpVec = ReducedPartRdx; 1406 SmallVector<Constant*, 32> ShuffleMask(VF, 0); 1407 for (unsigned i = VF; i != 1; i >>= 1) { 1408 // Move the upper half of the vector to the lower half. 1409 for (unsigned j = 0; j != i/2; ++j) 1410 ShuffleMask[j] = Builder.getInt32(i/2 + j); 1411 1412 // Fill the rest of the mask with undef. 1413 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 1414 UndefValue::get(Builder.getInt32Ty())); 1415 1416 Value *Shuf = 1417 Builder.CreateShuffleVector(TmpVec, 1418 UndefValue::get(TmpVec->getType()), 1419 ConstantVector::get(ShuffleMask), 1420 "rdx.shuf"); 1421 1422 // Emit the operation on the shuffled value. 1423 switch (RdxDesc.Kind) { 1424 case LoopVectorizationLegality::IntegerAdd: 1425 TmpVec = Builder.CreateAdd(TmpVec, Shuf, "add.rdx"); 1426 break; 1427 case LoopVectorizationLegality::IntegerMult: 1428 TmpVec = Builder.CreateMul(TmpVec, Shuf, "mul.rdx"); 1429 break; 1430 case LoopVectorizationLegality::IntegerOr: 1431 TmpVec = Builder.CreateOr(TmpVec, Shuf, "or.rdx"); 1432 break; 1433 case LoopVectorizationLegality::IntegerAnd: 1434 TmpVec = Builder.CreateAnd(TmpVec, Shuf, "and.rdx"); 1435 break; 1436 case LoopVectorizationLegality::IntegerXor: 1437 TmpVec = Builder.CreateXor(TmpVec, Shuf, "xor.rdx"); 1438 break; 1439 default: 1440 llvm_unreachable("Unknown reduction operation"); 1441 } 1442 } 1443 1444 // The result is in the first element of the vector. 1445 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); 1446 1447 // Now, we need to fix the users of the reduction variable 1448 // inside and outside of the scalar remainder loop. 1449 // We know that the loop is in LCSSA form. We need to update the 1450 // PHI nodes in the exit blocks. 1451 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 1452 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 1453 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 1454 if (!LCSSAPhi) continue; 1455 1456 // All PHINodes need to have a single entry edge, or two if 1457 // we already fixed them. 1458 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 1459 1460 // We found our reduction value exit-PHI. Update it with the 1461 // incoming bypass edge. 1462 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { 1463 // Add an edge coming from the bypass. 1464 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock); 1465 break; 1466 } 1467 }// end of the LCSSA phi scan. 1468 1469 // Fix the scalar loop reduction variable with the incoming reduction sum 1470 // from the vector body and from the backedge value. 1471 int IncomingEdgeBlockIdx = 1472 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 1473 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 1474 // Pick the other block. 1475 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 1476 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0); 1477 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); 1478 }// end of for each redux variable. 1479 1480 // The Loop exit block may have single value PHI nodes where the incoming 1481 // value is 'undef'. While vectorizing we only handled real values that 1482 // were defined inside the loop. Here we handle the 'undef case'. 1483 // See PR14725. 1484 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 1485 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 1486 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 1487 if (!LCSSAPhi) continue; 1488 if (LCSSAPhi->getNumIncomingValues() == 1) 1489 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 1490 LoopMiddleBlock); 1491 } 1492} 1493 1494InnerLoopVectorizer::VectorParts 1495InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 1496 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 1497 "Invalid edge"); 1498 1499 VectorParts SrcMask = createBlockInMask(Src); 1500 1501 // The terminator has to be a branch inst! 1502 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 1503 assert(BI && "Unexpected terminator found"); 1504 1505 if (BI->isConditional()) { 1506 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 1507 1508 if (BI->getSuccessor(0) != Dst) 1509 for (unsigned part = 0; part < UF; ++part) 1510 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 1511 1512 for (unsigned part = 0; part < UF; ++part) 1513 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 1514 return EdgeMask; 1515 } 1516 1517 return SrcMask; 1518} 1519 1520InnerLoopVectorizer::VectorParts 1521InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 1522 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 1523 1524 // Loop incoming mask is all-one. 1525 if (OrigLoop->getHeader() == BB) { 1526 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 1527 return getVectorValue(C); 1528 } 1529 1530 // This is the block mask. We OR all incoming edges, and with zero. 1531 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 1532 VectorParts BlockMask = getVectorValue(Zero); 1533 1534 // For each pred: 1535 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 1536 VectorParts EM = createEdgeMask(*it, BB); 1537 for (unsigned part = 0; part < UF; ++part) 1538 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 1539 } 1540 1541 return BlockMask; 1542} 1543 1544void 1545InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, 1546 BasicBlock *BB, PhiVector *PV) { 1547 Constant *Zero = Builder.getInt32(0); 1548 1549 // For each instruction in the old loop. 1550 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 1551 VectorParts &Entry = WidenMap.get(it); 1552 switch (it->getOpcode()) { 1553 case Instruction::Br: 1554 // Nothing to do for PHIs and BR, since we already took care of the 1555 // loop control flow instructions. 1556 continue; 1557 case Instruction::PHI:{ 1558 PHINode* P = cast<PHINode>(it); 1559 // Handle reduction variables: 1560 if (Legal->getReductionVars()->count(P)) { 1561 for (unsigned part = 0; part < UF; ++part) { 1562 // This is phase one of vectorizing PHIs. 1563 Type *VecTy = VectorType::get(it->getType(), VF); 1564 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", 1565 LoopVectorBody-> getFirstInsertionPt()); 1566 } 1567 PV->push_back(P); 1568 continue; 1569 } 1570 1571 // Check for PHI nodes that are lowered to vector selects. 1572 if (P->getParent() != OrigLoop->getHeader()) { 1573 // We know that all PHIs in non header blocks are converted into 1574 // selects, so we don't have to worry about the insertion order and we 1575 // can just use the builder. 1576 1577 // At this point we generate the predication tree. There may be 1578 // duplications since this is a simple recursive scan, but future 1579 // optimizations will clean it up. 1580 VectorParts Cond = createEdgeMask(P->getIncomingBlock(0), 1581 P->getParent()); 1582 1583 for (unsigned part = 0; part < UF; ++part) { 1584 VectorParts &In0 = getVectorValue(P->getIncomingValue(0)); 1585 VectorParts &In1 = getVectorValue(P->getIncomingValue(1)); 1586 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part], 1587 "predphi"); 1588 } 1589 continue; 1590 } 1591 1592 // This PHINode must be an induction variable. 1593 // Make sure that we know about it. 1594 assert(Legal->getInductionVars()->count(P) && 1595 "Not an induction variable"); 1596 1597 LoopVectorizationLegality::InductionInfo II = 1598 Legal->getInductionVars()->lookup(P); 1599 1600 switch (II.IK) { 1601 case LoopVectorizationLegality::NoInduction: 1602 llvm_unreachable("Unknown induction"); 1603 case LoopVectorizationLegality::IntInduction: { 1604 assert(P == OldInduction && "Unexpected PHI"); 1605 Value *Broadcasted = getBroadcastInstrs(Induction); 1606 // After broadcasting the induction variable we need to make the 1607 // vector consecutive by adding 0, 1, 2 ... 1608 for (unsigned part = 0; part < UF; ++part) 1609 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); 1610 continue; 1611 } 1612 case LoopVectorizationLegality::ReverseIntInduction: 1613 case LoopVectorizationLegality::PtrInduction: 1614 // Handle reverse integer and pointer inductions. 1615 Value *StartIdx = 0; 1616 // If we have a single integer induction variable then use it. 1617 // Otherwise, start counting at zero. 1618 if (OldInduction) { 1619 LoopVectorizationLegality::InductionInfo OldII = 1620 Legal->getInductionVars()->lookup(OldInduction); 1621 StartIdx = OldII.StartValue; 1622 } else { 1623 StartIdx = ConstantInt::get(Induction->getType(), 0); 1624 } 1625 // This is the normalized GEP that starts counting at zero. 1626 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx, 1627 "normalized.idx"); 1628 1629 // Handle the reverse integer induction variable case. 1630 if (LoopVectorizationLegality::ReverseIntInduction == II.IK) { 1631 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType()); 1632 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy, 1633 "resize.norm.idx"); 1634 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI, 1635 "reverse.idx"); 1636 1637 // This is a new value so do not hoist it out. 1638 Value *Broadcasted = getBroadcastInstrs(ReverseInd); 1639 // After broadcasting the induction variable we need to make the 1640 // vector consecutive by adding ... -3, -2, -1, 0. 1641 for (unsigned part = 0; part < UF; ++part) 1642 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true); 1643 continue; 1644 } 1645 1646 // Handle the pointer induction variable case. 1647 assert(P->getType()->isPointerTy() && "Unexpected type."); 1648 1649 // This is the vector of results. Notice that we don't generate 1650 // vector geps because scalar geps result in better code. 1651 for (unsigned part = 0; part < UF; ++part) { 1652 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 1653 for (unsigned int i = 0; i < VF; ++i) { 1654 Constant *Idx = ConstantInt::get(Induction->getType(), 1655 i + part * VF); 1656 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, 1657 "gep.idx"); 1658 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 1659 "next.gep"); 1660 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 1661 Builder.getInt32(i), 1662 "insert.gep"); 1663 } 1664 Entry[part] = VecVal; 1665 } 1666 continue; 1667 } 1668 1669 }// End of PHI. 1670 1671 case Instruction::Add: 1672 case Instruction::FAdd: 1673 case Instruction::Sub: 1674 case Instruction::FSub: 1675 case Instruction::Mul: 1676 case Instruction::FMul: 1677 case Instruction::UDiv: 1678 case Instruction::SDiv: 1679 case Instruction::FDiv: 1680 case Instruction::URem: 1681 case Instruction::SRem: 1682 case Instruction::FRem: 1683 case Instruction::Shl: 1684 case Instruction::LShr: 1685 case Instruction::AShr: 1686 case Instruction::And: 1687 case Instruction::Or: 1688 case Instruction::Xor: { 1689 // Just widen binops. 1690 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 1691 VectorParts &A = getVectorValue(it->getOperand(0)); 1692 VectorParts &B = getVectorValue(it->getOperand(1)); 1693 1694 // Use this vector value for all users of the original instruction. 1695 for (unsigned Part = 0; Part < UF; ++Part) { 1696 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 1697 1698 // Update the NSW, NUW and Exact flags. 1699 BinaryOperator *VecOp = cast<BinaryOperator>(V); 1700 if (isa<OverflowingBinaryOperator>(BinOp)) { 1701 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); 1702 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); 1703 } 1704 if (isa<PossiblyExactOperator>(VecOp)) 1705 VecOp->setIsExact(BinOp->isExact()); 1706 1707 Entry[Part] = V; 1708 } 1709 break; 1710 } 1711 case Instruction::Select: { 1712 // Widen selects. 1713 // If the selector is loop invariant we can create a select 1714 // instruction with a scalar condition. Otherwise, use vector-select. 1715 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), 1716 OrigLoop); 1717 1718 // The condition can be loop invariant but still defined inside the 1719 // loop. This means that we can't just use the original 'cond' value. 1720 // We have to take the 'vectorized' value and pick the first lane. 1721 // Instcombine will make this a no-op. 1722 VectorParts &Cond = getVectorValue(it->getOperand(0)); 1723 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 1724 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 1725 Value *ScalarCond = Builder.CreateExtractElement(Cond[0], 1726 Builder.getInt32(0)); 1727 for (unsigned Part = 0; Part < UF; ++Part) { 1728 Entry[Part] = Builder.CreateSelect( 1729 InvariantCond ? ScalarCond : Cond[Part], 1730 Op0[Part], 1731 Op1[Part]); 1732 } 1733 break; 1734 } 1735 1736 case Instruction::ICmp: 1737 case Instruction::FCmp: { 1738 // Widen compares. Generate vector compares. 1739 bool FCmp = (it->getOpcode() == Instruction::FCmp); 1740 CmpInst *Cmp = dyn_cast<CmpInst>(it); 1741 VectorParts &A = getVectorValue(it->getOperand(0)); 1742 VectorParts &B = getVectorValue(it->getOperand(1)); 1743 for (unsigned Part = 0; Part < UF; ++Part) { 1744 Value *C = 0; 1745 if (FCmp) 1746 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 1747 else 1748 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 1749 Entry[Part] = C; 1750 } 1751 break; 1752 } 1753 1754 case Instruction::Store: { 1755 // Attempt to issue a wide store. 1756 StoreInst *SI = dyn_cast<StoreInst>(it); 1757 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF); 1758 Value *Ptr = SI->getPointerOperand(); 1759 unsigned Alignment = SI->getAlignment(); 1760 1761 assert(!Legal->isUniform(Ptr) && 1762 "We do not allow storing to uniform addresses"); 1763 1764 1765 int Stride = Legal->isConsecutivePtr(Ptr); 1766 bool Reverse = Stride < 0; 1767 if (Stride == 0) { 1768 scalarizeInstruction(it); 1769 break; 1770 } 1771 1772 // Handle consecutive stores. 1773 1774 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 1775 if (Gep) { 1776 // The last index does not have to be the induction. It can be 1777 // consecutive and be a function of the index. For example A[I+1]; 1778 unsigned NumOperands = Gep->getNumOperands(); 1779 1780 Value *LastGepOperand = Gep->getOperand(NumOperands - 1); 1781 VectorParts &GEPParts = getVectorValue(LastGepOperand); 1782 Value *LastIndex = GEPParts[0]; 1783 LastIndex = Builder.CreateExtractElement(LastIndex, Zero); 1784 1785 // Create the new GEP with the new induction variable. 1786 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1787 Gep2->setOperand(NumOperands - 1, LastIndex); 1788 Ptr = Builder.Insert(Gep2); 1789 } else { 1790 // Use the induction element ptr. 1791 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 1792 VectorParts &PtrVal = getVectorValue(Ptr); 1793 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 1794 } 1795 1796 VectorParts &StoredVal = getVectorValue(SI->getValueOperand()); 1797 for (unsigned Part = 0; Part < UF; ++Part) { 1798 // Calculate the pointer for the specific unroll-part. 1799 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1800 1801 if (Reverse) { 1802 // If we store to reverse consecutive memory locations then we need 1803 // to reverse the order of elements in the stored value. 1804 StoredVal[Part] = reverseVector(StoredVal[Part]); 1805 // If the address is consecutive but reversed, then the 1806 // wide store needs to start at the last vector element. 1807 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1808 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1809 } 1810 1811 Value *VecPtr = Builder.CreateBitCast(PartPtr, StTy->getPointerTo()); 1812 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment); 1813 } 1814 break; 1815 } 1816 case Instruction::Load: { 1817 // Attempt to issue a wide load. 1818 LoadInst *LI = dyn_cast<LoadInst>(it); 1819 Type *RetTy = VectorType::get(LI->getType(), VF); 1820 Value *Ptr = LI->getPointerOperand(); 1821 unsigned Alignment = LI->getAlignment(); 1822 1823 // If the pointer is loop invariant or if it is non consecutive, 1824 // scalarize the load. 1825 int Stride = Legal->isConsecutivePtr(Ptr); 1826 bool Reverse = Stride < 0; 1827 if (Legal->isUniform(Ptr) || Stride == 0) { 1828 scalarizeInstruction(it); 1829 break; 1830 } 1831 1832 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 1833 if (Gep) { 1834 // The last index does not have to be the induction. It can be 1835 // consecutive and be a function of the index. For example A[I+1]; 1836 unsigned NumOperands = Gep->getNumOperands(); 1837 1838 Value *LastGepOperand = Gep->getOperand(NumOperands - 1); 1839 VectorParts &GEPParts = getVectorValue(LastGepOperand); 1840 Value *LastIndex = GEPParts[0]; 1841 LastIndex = Builder.CreateExtractElement(LastIndex, Zero); 1842 1843 // Create the new GEP with the new induction variable. 1844 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1845 Gep2->setOperand(NumOperands - 1, LastIndex); 1846 Ptr = Builder.Insert(Gep2); 1847 } else { 1848 // Use the induction element ptr. 1849 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 1850 VectorParts &PtrVal = getVectorValue(Ptr); 1851 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 1852 } 1853 1854 for (unsigned Part = 0; Part < UF; ++Part) { 1855 // Calculate the pointer for the specific unroll-part. 1856 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1857 1858 if (Reverse) { 1859 // If the address is consecutive but reversed, then the 1860 // wide store needs to start at the last vector element. 1861 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1862 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1863 } 1864 1865 Value *VecPtr = Builder.CreateBitCast(PartPtr, RetTy->getPointerTo()); 1866 Value *LI = Builder.CreateLoad(VecPtr, "wide.load"); 1867 cast<LoadInst>(LI)->setAlignment(Alignment); 1868 Entry[Part] = Reverse ? reverseVector(LI) : LI; 1869 } 1870 break; 1871 } 1872 case Instruction::ZExt: 1873 case Instruction::SExt: 1874 case Instruction::FPToUI: 1875 case Instruction::FPToSI: 1876 case Instruction::FPExt: 1877 case Instruction::PtrToInt: 1878 case Instruction::IntToPtr: 1879 case Instruction::SIToFP: 1880 case Instruction::UIToFP: 1881 case Instruction::Trunc: 1882 case Instruction::FPTrunc: 1883 case Instruction::BitCast: { 1884 CastInst *CI = dyn_cast<CastInst>(it); 1885 /// Optimize the special case where the source is the induction 1886 /// variable. Notice that we can only optimize the 'trunc' case 1887 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 1888 /// c. other casts depend on pointer size. 1889 if (CI->getOperand(0) == OldInduction && 1890 it->getOpcode() == Instruction::Trunc) { 1891 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 1892 CI->getType()); 1893 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 1894 for (unsigned Part = 0; Part < UF; ++Part) 1895 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false); 1896 break; 1897 } 1898 /// Vectorize casts. 1899 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF); 1900 1901 VectorParts &A = getVectorValue(it->getOperand(0)); 1902 for (unsigned Part = 0; Part < UF; ++Part) 1903 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 1904 break; 1905 } 1906 1907 case Instruction::Call: { 1908 assert(isTriviallyVectorizableIntrinsic(it)); 1909 Module *M = BB->getParent()->getParent(); 1910 IntrinsicInst *II = cast<IntrinsicInst>(it); 1911 Intrinsic::ID ID = II->getIntrinsicID(); 1912 for (unsigned Part = 0; Part < UF; ++Part) { 1913 SmallVector<Value*, 4> Args; 1914 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) { 1915 VectorParts &Arg = getVectorValue(II->getArgOperand(i)); 1916 Args.push_back(Arg[Part]); 1917 } 1918 Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) }; 1919 Function *F = Intrinsic::getDeclaration(M, ID, Tys); 1920 Entry[Part] = Builder.CreateCall(F, Args); 1921 } 1922 break; 1923 } 1924 1925 default: 1926 // All other instructions are unsupported. Scalarize them. 1927 scalarizeInstruction(it); 1928 break; 1929 }// end of switch. 1930 }// end of for_each instr. 1931} 1932 1933void InnerLoopVectorizer::updateAnalysis() { 1934 // Forget the original basic block. 1935 SE->forgetLoop(OrigLoop); 1936 1937 // Update the dominator tree information. 1938 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) && 1939 "Entry does not dominate exit."); 1940 1941 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock); 1942 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); 1943 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock); 1944 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); 1945 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 1946 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); 1947 1948 DEBUG(DT->verifyAnalysis()); 1949} 1950 1951bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 1952 if (!EnableIfConversion) 1953 return false; 1954 1955 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 1956 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector(); 1957 1958 // Collect the blocks that need predication. 1959 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) { 1960 BasicBlock *BB = LoopBlocks[i]; 1961 1962 // We don't support switch statements inside loops. 1963 if (!isa<BranchInst>(BB->getTerminator())) 1964 return false; 1965 1966 // We must have at most two predecessors because we need to convert 1967 // all PHIs to selects. 1968 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB)); 1969 if (Preds > 2) 1970 return false; 1971 1972 // We must be able to predicate all blocks that need to be predicated. 1973 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB)) 1974 return false; 1975 } 1976 1977 // We can if-convert this loop. 1978 return true; 1979} 1980 1981bool LoopVectorizationLegality::canVectorize() { 1982 assert(TheLoop->getLoopPreheader() && "No preheader!!"); 1983 1984 // We can only vectorize innermost loops. 1985 if (TheLoop->getSubLoopsVector().size()) 1986 return false; 1987 1988 // We must have a single backedge. 1989 if (TheLoop->getNumBackEdges() != 1) 1990 return false; 1991 1992 // We must have a single exiting block. 1993 if (!TheLoop->getExitingBlock()) 1994 return false; 1995 1996 unsigned NumBlocks = TheLoop->getNumBlocks(); 1997 1998 // Check if we can if-convert non single-bb loops. 1999 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 2000 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 2001 return false; 2002 } 2003 2004 // We need to have a loop header. 2005 BasicBlock *Latch = TheLoop->getLoopLatch(); 2006 DEBUG(dbgs() << "LV: Found a loop: " << 2007 TheLoop->getHeader()->getName() << "\n"); 2008 2009 // ScalarEvolution needs to be able to find the exit count. 2010 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch); 2011 if (ExitCount == SE->getCouldNotCompute()) { 2012 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 2013 return false; 2014 } 2015 2016 // Do not loop-vectorize loops with a tiny trip count. 2017 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch); 2018 if (TC > 0u && TC < TinyTripCountThreshold) { 2019 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << 2020 "This loop is not worth vectorizing.\n"); 2021 return false; 2022 } 2023 2024 // Check if we can vectorize the instructions and CFG in this loop. 2025 if (!canVectorizeInstrs()) { 2026 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 2027 return false; 2028 } 2029 2030 // Go over each instruction and look at memory deps. 2031 if (!canVectorizeMemory()) { 2032 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 2033 return false; 2034 } 2035 2036 // Collect all of the variables that remain uniform after vectorization. 2037 collectLoopUniforms(); 2038 2039 DEBUG(dbgs() << "LV: We can vectorize this loop" << 2040 (PtrRtCheck.Need ? " (with a runtime bound check)" : "") 2041 <<"!\n"); 2042 2043 // Okay! We can vectorize. At this point we don't have any other mem analysis 2044 // which may limit our maximum vectorization factor, so just return true with 2045 // no restrictions. 2046 return true; 2047} 2048 2049bool LoopVectorizationLegality::canVectorizeInstrs() { 2050 BasicBlock *PreHeader = TheLoop->getLoopPreheader(); 2051 BasicBlock *Header = TheLoop->getHeader(); 2052 2053 // For each block in the loop. 2054 for (Loop::block_iterator bb = TheLoop->block_begin(), 2055 be = TheLoop->block_end(); bb != be; ++bb) { 2056 2057 // Scan the instructions in the block and look for hazards. 2058 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2059 ++it) { 2060 2061 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 2062 // This should not happen because the loop should be normalized. 2063 if (Phi->getNumIncomingValues() != 2) { 2064 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 2065 return false; 2066 } 2067 2068 // Check that this PHI type is allowed. 2069 if (!Phi->getType()->isIntegerTy() && 2070 !Phi->getType()->isPointerTy()) { 2071 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 2072 return false; 2073 } 2074 2075 // If this PHINode is not in the header block, then we know that we 2076 // can convert it to select during if-conversion. No need to check if 2077 // the PHIs in this block are induction or reduction variables. 2078 if (*bb != Header) 2079 continue; 2080 2081 // This is the value coming from the preheader. 2082 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); 2083 // Check if this is an induction variable. 2084 InductionKind IK = isInductionVariable(Phi); 2085 2086 if (NoInduction != IK) { 2087 // Int inductions are special because we only allow one IV. 2088 if (IK == IntInduction) { 2089 if (Induction) { 2090 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n"); 2091 return false; 2092 } 2093 Induction = Phi; 2094 } 2095 2096 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 2097 Inductions[Phi] = InductionInfo(StartValue, IK); 2098 continue; 2099 } 2100 2101 if (AddReductionVar(Phi, IntegerAdd)) { 2102 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); 2103 continue; 2104 } 2105 if (AddReductionVar(Phi, IntegerMult)) { 2106 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); 2107 continue; 2108 } 2109 if (AddReductionVar(Phi, IntegerOr)) { 2110 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); 2111 continue; 2112 } 2113 if (AddReductionVar(Phi, IntegerAnd)) { 2114 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); 2115 continue; 2116 } 2117 if (AddReductionVar(Phi, IntegerXor)) { 2118 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); 2119 continue; 2120 } 2121 2122 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 2123 return false; 2124 }// end of PHI handling 2125 2126 // We still don't handle functions. 2127 CallInst *CI = dyn_cast<CallInst>(it); 2128 if (CI && !isTriviallyVectorizableIntrinsic(it)) { 2129 DEBUG(dbgs() << "LV: Found a call site.\n"); 2130 return false; 2131 } 2132 2133 // Check that the instruction return type is vectorizable. 2134 if (!VectorType::isValidElementType(it->getType()) && 2135 !it->getType()->isVoidTy()) { 2136 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n"); 2137 return false; 2138 } 2139 2140 // Check that the stored type is vectorizable. 2141 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 2142 Type *T = ST->getValueOperand()->getType(); 2143 if (!VectorType::isValidElementType(T)) 2144 return false; 2145 } 2146 2147 // Reduction instructions are allowed to have exit users. 2148 // All other instructions must not have external users. 2149 if (!AllowedExit.count(it)) 2150 //Check that all of the users of the loop are inside the BB. 2151 for (Value::use_iterator I = it->use_begin(), E = it->use_end(); 2152 I != E; ++I) { 2153 Instruction *U = cast<Instruction>(*I); 2154 // This user may be a reduction exit value. 2155 if (!TheLoop->contains(U)) { 2156 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n"); 2157 return false; 2158 } 2159 } 2160 } // next instr. 2161 2162 } 2163 2164 if (!Induction) { 2165 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 2166 assert(getInductionVars()->size() && "No induction variables"); 2167 } 2168 2169 return true; 2170} 2171 2172void LoopVectorizationLegality::collectLoopUniforms() { 2173 // We now know that the loop is vectorizable! 2174 // Collect variables that will remain uniform after vectorization. 2175 std::vector<Value*> Worklist; 2176 BasicBlock *Latch = TheLoop->getLoopLatch(); 2177 2178 // Start with the conditional branch and walk up the block. 2179 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 2180 2181 while (Worklist.size()) { 2182 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 2183 Worklist.pop_back(); 2184 2185 // Look at instructions inside this loop. 2186 // Stop when reaching PHI nodes. 2187 // TODO: we need to follow values all over the loop, not only in this block. 2188 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 2189 continue; 2190 2191 // This is a known uniform. 2192 Uniforms.insert(I); 2193 2194 // Insert all operands. 2195 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) { 2196 Worklist.push_back(I->getOperand(i)); 2197 } 2198 } 2199} 2200 2201bool LoopVectorizationLegality::canVectorizeMemory() { 2202 typedef SmallVector<Value*, 16> ValueVector; 2203 typedef SmallPtrSet<Value*, 16> ValueSet; 2204 // Holds the Load and Store *instructions*. 2205 ValueVector Loads; 2206 ValueVector Stores; 2207 PtrRtCheck.Pointers.clear(); 2208 PtrRtCheck.Need = false; 2209 2210 // For each block. 2211 for (Loop::block_iterator bb = TheLoop->block_begin(), 2212 be = TheLoop->block_end(); bb != be; ++bb) { 2213 2214 // Scan the BB and collect legal loads and stores. 2215 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2216 ++it) { 2217 2218 // If this is a load, save it. If this instruction can read from memory 2219 // but is not a load, then we quit. Notice that we don't handle function 2220 // calls that read or write. 2221 if (it->mayReadFromMemory()) { 2222 LoadInst *Ld = dyn_cast<LoadInst>(it); 2223 if (!Ld) return false; 2224 if (!Ld->isSimple()) { 2225 DEBUG(dbgs() << "LV: Found a non-simple load.\n"); 2226 return false; 2227 } 2228 Loads.push_back(Ld); 2229 continue; 2230 } 2231 2232 // Save 'store' instructions. Abort if other instructions write to memory. 2233 if (it->mayWriteToMemory()) { 2234 StoreInst *St = dyn_cast<StoreInst>(it); 2235 if (!St) return false; 2236 if (!St->isSimple()) { 2237 DEBUG(dbgs() << "LV: Found a non-simple store.\n"); 2238 return false; 2239 } 2240 Stores.push_back(St); 2241 } 2242 } // next instr. 2243 } // next block. 2244 2245 // Now we have two lists that hold the loads and the stores. 2246 // Next, we find the pointers that they use. 2247 2248 // Check if we see any stores. If there are no stores, then we don't 2249 // care if the pointers are *restrict*. 2250 if (!Stores.size()) { 2251 DEBUG(dbgs() << "LV: Found a read-only loop!\n"); 2252 return true; 2253 } 2254 2255 // Holds the read and read-write *pointers* that we find. 2256 ValueVector Reads; 2257 ValueVector ReadWrites; 2258 2259 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects 2260 // multiple times on the same object. If the ptr is accessed twice, once 2261 // for read and once for write, it will only appear once (on the write 2262 // list). This is okay, since we are going to check for conflicts between 2263 // writes and between reads and writes, but not between reads and reads. 2264 ValueSet Seen; 2265 2266 ValueVector::iterator I, IE; 2267 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) { 2268 StoreInst *ST = cast<StoreInst>(*I); 2269 Value* Ptr = ST->getPointerOperand(); 2270 2271 if (isUniform(Ptr)) { 2272 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 2273 return false; 2274 } 2275 2276 // If we did *not* see this pointer before, insert it to 2277 // the read-write list. At this phase it is only a 'write' list. 2278 if (Seen.insert(Ptr)) 2279 ReadWrites.push_back(Ptr); 2280 } 2281 2282 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { 2283 LoadInst *LD = cast<LoadInst>(*I); 2284 Value* Ptr = LD->getPointerOperand(); 2285 // If we did *not* see this pointer before, insert it to the 2286 // read list. If we *did* see it before, then it is already in 2287 // the read-write list. This allows us to vectorize expressions 2288 // such as A[i] += x; Because the address of A[i] is a read-write 2289 // pointer. This only works if the index of A[i] is consecutive. 2290 // If the address of i is unknown (for example A[B[i]]) then we may 2291 // read a few words, modify, and write a few words, and some of the 2292 // words may be written to the same address. 2293 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr)) 2294 Reads.push_back(Ptr); 2295 } 2296 2297 // If we write (or read-write) to a single destination and there are no 2298 // other reads in this loop then is it safe to vectorize. 2299 if (ReadWrites.size() == 1 && Reads.size() == 0) { 2300 DEBUG(dbgs() << "LV: Found a write-only loop!\n"); 2301 return true; 2302 } 2303 2304 // Find pointers with computable bounds. We are going to use this information 2305 // to place a runtime bound check. 2306 bool CanDoRT = true; 2307 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) 2308 if (hasComputableBounds(*I)) { 2309 PtrRtCheck.insert(SE, TheLoop, *I); 2310 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n"); 2311 } else { 2312 CanDoRT = false; 2313 break; 2314 } 2315 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) 2316 if (hasComputableBounds(*I)) { 2317 PtrRtCheck.insert(SE, TheLoop, *I); 2318 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n"); 2319 } else { 2320 CanDoRT = false; 2321 break; 2322 } 2323 2324 // Check that we did not collect too many pointers or found a 2325 // unsizeable pointer. 2326 if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) { 2327 PtrRtCheck.reset(); 2328 CanDoRT = false; 2329 } 2330 2331 if (CanDoRT) { 2332 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n"); 2333 } 2334 2335 bool NeedRTCheck = false; 2336 2337 // Now that the pointers are in two lists (Reads and ReadWrites), we 2338 // can check that there are no conflicts between each of the writes and 2339 // between the writes to the reads. 2340 ValueSet WriteObjects; 2341 ValueVector TempObjects; 2342 2343 // Check that the read-writes do not conflict with other read-write 2344 // pointers. 2345 bool AllWritesIdentified = true; 2346 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) { 2347 GetUnderlyingObjects(*I, TempObjects, DL); 2348 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end(); 2349 it != e; ++it) { 2350 if (!isIdentifiedObject(*it)) { 2351 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n"); 2352 NeedRTCheck = true; 2353 AllWritesIdentified = false; 2354 } 2355 if (!WriteObjects.insert(*it)) { 2356 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2357 << **it <<"\n"); 2358 return false; 2359 } 2360 } 2361 TempObjects.clear(); 2362 } 2363 2364 /// Check that the reads don't conflict with the read-writes. 2365 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) { 2366 GetUnderlyingObjects(*I, TempObjects, DL); 2367 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end(); 2368 it != e; ++it) { 2369 // If all of the writes are identified then we don't care if the read 2370 // pointer is identified or not. 2371 if (!AllWritesIdentified && !isIdentifiedObject(*it)) { 2372 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n"); 2373 NeedRTCheck = true; 2374 } 2375 if (WriteObjects.count(*it)) { 2376 DEBUG(dbgs() << "LV: Found a possible read/write reorder:" 2377 << **it <<"\n"); 2378 return false; 2379 } 2380 } 2381 TempObjects.clear(); 2382 } 2383 2384 PtrRtCheck.Need = NeedRTCheck; 2385 if (NeedRTCheck && !CanDoRT) { 2386 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " << 2387 "the array bounds.\n"); 2388 PtrRtCheck.reset(); 2389 return false; 2390 } 2391 2392 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") << 2393 " need a runtime memory check.\n"); 2394 return true; 2395} 2396 2397bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, 2398 ReductionKind Kind) { 2399 if (Phi->getNumIncomingValues() != 2) 2400 return false; 2401 2402 // Reduction variables are only found in the loop header block. 2403 if (Phi->getParent() != TheLoop->getHeader()) 2404 return false; 2405 2406 // Obtain the reduction start value from the value that comes from the loop 2407 // preheader. 2408 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 2409 2410 // ExitInstruction is the single value which is used outside the loop. 2411 // We only allow for a single reduction value to be used outside the loop. 2412 // This includes users of the reduction, variables (which form a cycle 2413 // which ends in the phi node). 2414 Instruction *ExitInstruction = 0; 2415 2416 // Iter is our iterator. We start with the PHI node and scan for all of the 2417 // users of this instruction. All users must be instructions that can be 2418 // used as reduction variables (such as ADD). We may have a single 2419 // out-of-block user. The cycle must end with the original PHI. 2420 Instruction *Iter = Phi; 2421 while (true) { 2422 // If the instruction has no users then this is a broken 2423 // chain and can't be a reduction variable. 2424 if (Iter->use_empty()) 2425 return false; 2426 2427 // Did we find a user inside this loop already ? 2428 bool FoundInBlockUser = false; 2429 // Did we reach the initial PHI node already ? 2430 bool FoundStartPHI = false; 2431 2432 // For each of the *users* of iter. 2433 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end(); 2434 it != e; ++it) { 2435 Instruction *U = cast<Instruction>(*it); 2436 // We already know that the PHI is a user. 2437 if (U == Phi) { 2438 FoundStartPHI = true; 2439 continue; 2440 } 2441 2442 // Check if we found the exit user. 2443 BasicBlock *Parent = U->getParent(); 2444 if (!TheLoop->contains(Parent)) { 2445 // Exit if you find multiple outside users. 2446 if (ExitInstruction != 0) 2447 return false; 2448 ExitInstruction = Iter; 2449 } 2450 2451 // We allow in-loop PHINodes which are not the original reduction PHI 2452 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE 2453 // structure) then don't skip this PHI. 2454 if (isa<PHINode>(Iter) && isa<PHINode>(U) && 2455 U->getParent() != TheLoop->getHeader() && 2456 TheLoop->contains(U) && 2457 Iter->getNumUses() > 1) 2458 continue; 2459 2460 // We can't have multiple inside users. 2461 if (FoundInBlockUser) 2462 return false; 2463 FoundInBlockUser = true; 2464 2465 // Any reduction instr must be of one of the allowed kinds. 2466 if (!isReductionInstr(U, Kind)) 2467 return false; 2468 2469 // Reductions of instructions such as Div, and Sub is only 2470 // possible if the LHS is the reduction variable. 2471 if (!U->isCommutative() && U->getOperand(0) != Iter) 2472 return false; 2473 2474 Iter = U; 2475 } 2476 2477 // We found a reduction var if we have reached the original 2478 // phi node and we only have a single instruction with out-of-loop 2479 // users. 2480 if (FoundStartPHI && ExitInstruction) { 2481 // This instruction is allowed to have out-of-loop users. 2482 AllowedExit.insert(ExitInstruction); 2483 2484 // Save the description of this reduction variable. 2485 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind); 2486 Reductions[Phi] = RD; 2487 return true; 2488 } 2489 2490 // If we've reached the start PHI but did not find an outside user then 2491 // this is dead code. Abort. 2492 if (FoundStartPHI) 2493 return false; 2494 } 2495} 2496 2497bool 2498LoopVectorizationLegality::isReductionInstr(Instruction *I, 2499 ReductionKind Kind) { 2500 switch (I->getOpcode()) { 2501 default: 2502 return false; 2503 case Instruction::PHI: 2504 // possibly. 2505 return true; 2506 case Instruction::Sub: 2507 case Instruction::Add: 2508 return Kind == IntegerAdd; 2509 case Instruction::SDiv: 2510 case Instruction::UDiv: 2511 case Instruction::Mul: 2512 return Kind == IntegerMult; 2513 case Instruction::And: 2514 return Kind == IntegerAnd; 2515 case Instruction::Or: 2516 return Kind == IntegerOr; 2517 case Instruction::Xor: 2518 return Kind == IntegerXor; 2519 } 2520} 2521 2522LoopVectorizationLegality::InductionKind 2523LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { 2524 Type *PhiTy = Phi->getType(); 2525 // We only handle integer and pointer inductions variables. 2526 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 2527 return NoInduction; 2528 2529 // Check that the PHI is consecutive and starts at zero. 2530 const SCEV *PhiScev = SE->getSCEV(Phi); 2531 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 2532 if (!AR) { 2533 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 2534 return NoInduction; 2535 } 2536 const SCEV *Step = AR->getStepRecurrence(*SE); 2537 2538 // Integer inductions need to have a stride of one. 2539 if (PhiTy->isIntegerTy()) { 2540 if (Step->isOne()) 2541 return IntInduction; 2542 if (Step->isAllOnesValue()) 2543 return ReverseIntInduction; 2544 return NoInduction; 2545 } 2546 2547 // Calculate the pointer stride and check if it is consecutive. 2548 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 2549 if (!C) 2550 return NoInduction; 2551 2552 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 2553 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType()); 2554 if (C->getValue()->equalsInt(Size)) 2555 return PtrInduction; 2556 2557 return NoInduction; 2558} 2559 2560bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 2561 Value *In0 = const_cast<Value*>(V); 2562 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 2563 if (!PN) 2564 return false; 2565 2566 return Inductions.count(PN); 2567} 2568 2569bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 2570 assert(TheLoop->contains(BB) && "Unknown block used"); 2571 2572 // Blocks that do not dominate the latch need predication. 2573 BasicBlock* Latch = TheLoop->getLoopLatch(); 2574 return !DT->dominates(BB, Latch); 2575} 2576 2577bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { 2578 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 2579 // We don't predicate loads/stores at the moment. 2580 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow()) 2581 return false; 2582 2583 // The instructions below can trap. 2584 switch (it->getOpcode()) { 2585 default: continue; 2586 case Instruction::UDiv: 2587 case Instruction::SDiv: 2588 case Instruction::URem: 2589 case Instruction::SRem: 2590 return false; 2591 } 2592 } 2593 2594 return true; 2595} 2596 2597bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) { 2598 const SCEV *PhiScev = SE->getSCEV(Ptr); 2599 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 2600 if (!AR) 2601 return false; 2602 2603 return AR->isAffine(); 2604} 2605 2606unsigned 2607LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize, 2608 unsigned UserVF) { 2609 if (OptForSize && Legal->getRuntimePointerCheck()->Need) { 2610 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); 2611 return 1; 2612 } 2613 2614 // Find the trip count. 2615 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch()); 2616 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n"); 2617 2618 unsigned VF = MaxVectorSize; 2619 2620 // If we optimize the program for size, avoid creating the tail loop. 2621 if (OptForSize) { 2622 // If we are unable to calculate the trip count then don't try to vectorize. 2623 if (TC < 2) { 2624 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 2625 return 1; 2626 } 2627 2628 // Find the maximum SIMD width that can fit within the trip count. 2629 VF = TC % MaxVectorSize; 2630 2631 if (VF == 0) 2632 VF = MaxVectorSize; 2633 2634 // If the trip count that we found modulo the vectorization factor is not 2635 // zero then we require a tail. 2636 if (VF < 2) { 2637 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 2638 return 1; 2639 } 2640 } 2641 2642 if (UserVF != 0) { 2643 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 2644 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n"); 2645 2646 return UserVF; 2647 } 2648 2649 if (!TTI) { 2650 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n"); 2651 return 1; 2652 } 2653 2654 float Cost = expectedCost(1); 2655 unsigned Width = 1; 2656 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n"); 2657 for (unsigned i=2; i <= VF; i*=2) { 2658 // Notice that the vector loop needs to be executed less times, so 2659 // we need to divide the cost of the vector loops by the width of 2660 // the vector elements. 2661 float VectorCost = expectedCost(i) / (float)i; 2662 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " << 2663 (int)VectorCost << ".\n"); 2664 if (VectorCost < Cost) { 2665 Cost = VectorCost; 2666 Width = i; 2667 } 2668 } 2669 2670 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n"); 2671 return Width; 2672} 2673 2674unsigned 2675LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, 2676 unsigned UserUF) { 2677 // Use the user preference, unless 'auto' is selected. 2678 if (UserUF != 0) 2679 return UserUF; 2680 2681 // When we optimize for size we don't unroll. 2682 if (OptForSize) 2683 return 1; 2684 2685 unsigned TargetVectorRegisters = TTI->getNumberOfRegisters(true); 2686 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters << 2687 " vector registers\n"); 2688 2689 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); 2690 // We divide by these constants so assume that we have at least one 2691 // instruction that uses at least one register. 2692 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 2693 R.NumInstructions = std::max(R.NumInstructions, 1U); 2694 2695 // We calculate the unroll factor using the following formula. 2696 // Subtract the number of loop invariants from the number of available 2697 // registers. These registers are used by all of the unrolled instances. 2698 // Next, divide the remaining registers by the number of registers that is 2699 // required by the loop, in order to estimate how many parallel instances 2700 // fit without causing spills. 2701 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers; 2702 2703 // We don't want to unroll the loops to the point where they do not fit into 2704 // the decoded cache. Assume that we only allow 32 IR instructions. 2705 UF = std::min(UF, (32 / R.NumInstructions)); 2706 2707 // Clamp the unroll factor ranges to reasonable factors. 2708 if (UF > MaxUnrollSize) 2709 UF = MaxUnrollSize; 2710 else if (UF < 1) 2711 UF = 1; 2712 2713 return UF; 2714} 2715 2716LoopVectorizationCostModel::RegisterUsage 2717LoopVectorizationCostModel::calculateRegisterUsage() { 2718 // This function calculates the register usage by measuring the highest number 2719 // of values that are alive at a single location. Obviously, this is a very 2720 // rough estimation. We scan the loop in a topological order in order and 2721 // assign a number to each instruction. We use RPO to ensure that defs are 2722 // met before their users. We assume that each instruction that has in-loop 2723 // users starts an interval. We record every time that an in-loop value is 2724 // used, so we have a list of the first and last occurrences of each 2725 // instruction. Next, we transpose this data structure into a multi map that 2726 // holds the list of intervals that *end* at a specific location. This multi 2727 // map allows us to perform a linear search. We scan the instructions linearly 2728 // and record each time that a new interval starts, by placing it in a set. 2729 // If we find this value in the multi-map then we remove it from the set. 2730 // The max register usage is the maximum size of the set. 2731 // We also search for instructions that are defined outside the loop, but are 2732 // used inside the loop. We need this number separately from the max-interval 2733 // usage number because when we unroll, loop-invariant values do not take 2734 // more register. 2735 LoopBlocksDFS DFS(TheLoop); 2736 DFS.perform(LI); 2737 2738 RegisterUsage R; 2739 R.NumInstructions = 0; 2740 2741 // Each 'key' in the map opens a new interval. The values 2742 // of the map are the index of the 'last seen' usage of the 2743 // instruction that is the key. 2744 typedef DenseMap<Instruction*, unsigned> IntervalMap; 2745 // Maps instruction to its index. 2746 DenseMap<unsigned, Instruction*> IdxToInstr; 2747 // Marks the end of each interval. 2748 IntervalMap EndPoint; 2749 // Saves the list of instruction indices that are used in the loop. 2750 SmallSet<Instruction*, 8> Ends; 2751 // Saves the list of values that are used in the loop but are 2752 // defined outside the loop, such as arguments and constants. 2753 SmallPtrSet<Value*, 8> LoopInvariants; 2754 2755 unsigned Index = 0; 2756 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 2757 be = DFS.endRPO(); bb != be; ++bb) { 2758 R.NumInstructions += (*bb)->size(); 2759 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2760 ++it) { 2761 Instruction *I = it; 2762 IdxToInstr[Index++] = I; 2763 2764 // Save the end location of each USE. 2765 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 2766 Value *U = I->getOperand(i); 2767 Instruction *Instr = dyn_cast<Instruction>(U); 2768 2769 // Ignore non-instruction values such as arguments, constants, etc. 2770 if (!Instr) continue; 2771 2772 // If this instruction is outside the loop then record it and continue. 2773 if (!TheLoop->contains(Instr)) { 2774 LoopInvariants.insert(Instr); 2775 continue; 2776 } 2777 2778 // Overwrite previous end points. 2779 EndPoint[Instr] = Index; 2780 Ends.insert(Instr); 2781 } 2782 } 2783 } 2784 2785 // Saves the list of intervals that end with the index in 'key'. 2786 typedef SmallVector<Instruction*, 2> InstrList; 2787 DenseMap<unsigned, InstrList> TransposeEnds; 2788 2789 // Transpose the EndPoints to a list of values that end at each index. 2790 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 2791 it != e; ++it) 2792 TransposeEnds[it->second].push_back(it->first); 2793 2794 SmallSet<Instruction*, 8> OpenIntervals; 2795 unsigned MaxUsage = 0; 2796 2797 2798 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 2799 for (unsigned int i = 0; i < Index; ++i) { 2800 Instruction *I = IdxToInstr[i]; 2801 // Ignore instructions that are never used within the loop. 2802 if (!Ends.count(I)) continue; 2803 2804 // Remove all of the instructions that end at this location. 2805 InstrList &List = TransposeEnds[i]; 2806 for (unsigned int j=0, e = List.size(); j < e; ++j) 2807 OpenIntervals.erase(List[j]); 2808 2809 // Count the number of live interals. 2810 MaxUsage = std::max(MaxUsage, OpenIntervals.size()); 2811 2812 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << 2813 OpenIntervals.size() <<"\n"); 2814 2815 // Add the current instruction to the list of open intervals. 2816 OpenIntervals.insert(I); 2817 } 2818 2819 unsigned Invariant = LoopInvariants.size(); 2820 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n"); 2821 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n"); 2822 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n"); 2823 2824 R.LoopInvariantRegs = Invariant; 2825 R.MaxLocalUsers = MaxUsage; 2826 return R; 2827} 2828 2829unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 2830 unsigned Cost = 0; 2831 2832 // For each block. 2833 for (Loop::block_iterator bb = TheLoop->block_begin(), 2834 be = TheLoop->block_end(); bb != be; ++bb) { 2835 unsigned BlockCost = 0; 2836 BasicBlock *BB = *bb; 2837 2838 // For each instruction in the old loop. 2839 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 2840 unsigned C = getInstructionCost(it, VF); 2841 Cost += C; 2842 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " << 2843 VF << " For instruction: "<< *it << "\n"); 2844 } 2845 2846 // We assume that if-converted blocks have a 50% chance of being executed. 2847 // When the code is scalar then some of the blocks are avoided due to CF. 2848 // When the code is vectorized we execute all code paths. 2849 if (Legal->blockNeedsPredication(*bb) && VF == 1) 2850 BlockCost /= 2; 2851 2852 Cost += BlockCost; 2853 } 2854 2855 return Cost; 2856} 2857 2858unsigned 2859LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 2860 assert(TTI && "Invalid vector target transformation info"); 2861 2862 // If we know that this instruction will remain uniform, check the cost of 2863 // the scalar version. 2864 if (Legal->isUniformAfterVectorization(I)) 2865 VF = 1; 2866 2867 Type *RetTy = I->getType(); 2868 Type *VectorTy = ToVectorTy(RetTy, VF); 2869 2870 // TODO: We need to estimate the cost of intrinsic calls. 2871 switch (I->getOpcode()) { 2872 case Instruction::GetElementPtr: 2873 // We mark this instruction as zero-cost because scalar GEPs are usually 2874 // lowered to the intruction addressing mode. At the moment we don't 2875 // generate vector geps. 2876 return 0; 2877 case Instruction::Br: { 2878 return TTI->getCFInstrCost(I->getOpcode()); 2879 } 2880 case Instruction::PHI: 2881 //TODO: IF-converted IFs become selects. 2882 return 0; 2883 case Instruction::Add: 2884 case Instruction::FAdd: 2885 case Instruction::Sub: 2886 case Instruction::FSub: 2887 case Instruction::Mul: 2888 case Instruction::FMul: 2889 case Instruction::UDiv: 2890 case Instruction::SDiv: 2891 case Instruction::FDiv: 2892 case Instruction::URem: 2893 case Instruction::SRem: 2894 case Instruction::FRem: 2895 case Instruction::Shl: 2896 case Instruction::LShr: 2897 case Instruction::AShr: 2898 case Instruction::And: 2899 case Instruction::Or: 2900 case Instruction::Xor: 2901 return TTI->getArithmeticInstrCost(I->getOpcode(), VectorTy); 2902 case Instruction::Select: { 2903 SelectInst *SI = cast<SelectInst>(I); 2904 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 2905 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 2906 Type *CondTy = SI->getCondition()->getType(); 2907 if (ScalarCond) 2908 CondTy = VectorType::get(CondTy, VF); 2909 2910 return TTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 2911 } 2912 case Instruction::ICmp: 2913 case Instruction::FCmp: { 2914 Type *ValTy = I->getOperand(0)->getType(); 2915 VectorTy = ToVectorTy(ValTy, VF); 2916 return TTI->getCmpSelInstrCost(I->getOpcode(), VectorTy); 2917 } 2918 case Instruction::Store: { 2919 StoreInst *SI = cast<StoreInst>(I); 2920 Type *ValTy = SI->getValueOperand()->getType(); 2921 VectorTy = ToVectorTy(ValTy, VF); 2922 2923 if (VF == 1) 2924 return TTI->getMemoryOpCost(I->getOpcode(), VectorTy, 2925 SI->getAlignment(), 2926 SI->getPointerAddressSpace()); 2927 2928 // Scalarized stores. 2929 int Stride = Legal->isConsecutivePtr(SI->getPointerOperand()); 2930 bool Reverse = Stride < 0; 2931 if (0 == Stride) { 2932 unsigned Cost = 0; 2933 2934 // The cost of extracting from the value vector and pointer vector. 2935 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF); 2936 for (unsigned i = 0; i < VF; ++i) { 2937 Cost += TTI->getVectorInstrCost(Instruction::ExtractElement, 2938 VectorTy, i); 2939 Cost += TTI->getVectorInstrCost(Instruction::ExtractElement, 2940 PtrTy, i); 2941 } 2942 2943 // The cost of the scalar stores. 2944 Cost += VF * TTI->getMemoryOpCost(I->getOpcode(), 2945 ValTy->getScalarType(), 2946 SI->getAlignment(), 2947 SI->getPointerAddressSpace()); 2948 return Cost; 2949 } 2950 2951 // Wide stores. 2952 unsigned Cost = TTI->getMemoryOpCost(I->getOpcode(), VectorTy, 2953 SI->getAlignment(), 2954 SI->getPointerAddressSpace()); 2955 if (Reverse) 2956 Cost += TTI->getShuffleCost(TargetTransformInfo::SK_Reverse, 2957 VectorTy, 0); 2958 return Cost; 2959 } 2960 case Instruction::Load: { 2961 LoadInst *LI = cast<LoadInst>(I); 2962 2963 if (VF == 1) 2964 return TTI->getMemoryOpCost(I->getOpcode(), VectorTy, 2965 LI->getAlignment(), 2966 LI->getPointerAddressSpace()); 2967 2968 // Scalarized loads. 2969 int Stride = Legal->isConsecutivePtr(LI->getPointerOperand()); 2970 bool Reverse = Stride < 0; 2971 if (0 == Stride) { 2972 unsigned Cost = 0; 2973 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF); 2974 2975 // The cost of extracting from the pointer vector. 2976 for (unsigned i = 0; i < VF; ++i) 2977 Cost += TTI->getVectorInstrCost(Instruction::ExtractElement, 2978 PtrTy, i); 2979 2980 // The cost of inserting data to the result vector. 2981 for (unsigned i = 0; i < VF; ++i) 2982 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, 2983 VectorTy, i); 2984 2985 // The cost of the scalar stores. 2986 Cost += VF * TTI->getMemoryOpCost(I->getOpcode(), 2987 RetTy->getScalarType(), 2988 LI->getAlignment(), 2989 LI->getPointerAddressSpace()); 2990 return Cost; 2991 } 2992 2993 // Wide loads. 2994 unsigned Cost = TTI->getMemoryOpCost(I->getOpcode(), VectorTy, 2995 LI->getAlignment(), 2996 LI->getPointerAddressSpace()); 2997 if (Reverse) 2998 Cost += TTI->getShuffleCost(TargetTransformInfo::SK_Reverse, 2999 VectorTy, 0); 3000 return Cost; 3001 } 3002 case Instruction::ZExt: 3003 case Instruction::SExt: 3004 case Instruction::FPToUI: 3005 case Instruction::FPToSI: 3006 case Instruction::FPExt: 3007 case Instruction::PtrToInt: 3008 case Instruction::IntToPtr: 3009 case Instruction::SIToFP: 3010 case Instruction::UIToFP: 3011 case Instruction::Trunc: 3012 case Instruction::FPTrunc: 3013 case Instruction::BitCast: { 3014 // We optimize the truncation of induction variable. 3015 // The cost of these is the same as the scalar operation. 3016 if (I->getOpcode() == Instruction::Trunc && 3017 Legal->isInductionVariable(I->getOperand(0))) 3018 return TTI->getCastInstrCost(I->getOpcode(), I->getType(), 3019 I->getOperand(0)->getType()); 3020 3021 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); 3022 return TTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 3023 } 3024 case Instruction::Call: { 3025 assert(isTriviallyVectorizableIntrinsic(I)); 3026 IntrinsicInst *II = cast<IntrinsicInst>(I); 3027 Type *RetTy = ToVectorTy(II->getType(), VF); 3028 SmallVector<Type*, 4> Tys; 3029 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) 3030 Tys.push_back(ToVectorTy(II->getArgOperand(i)->getType(), VF)); 3031 return TTI->getIntrinsicInstrCost(II->getIntrinsicID(), RetTy, Tys); 3032 } 3033 default: { 3034 // We are scalarizing the instruction. Return the cost of the scalar 3035 // instruction, plus the cost of insert and extract into vector 3036 // elements, times the vector width. 3037 unsigned Cost = 0; 3038 3039 if (!RetTy->isVoidTy() && VF != 1) { 3040 unsigned InsCost = TTI->getVectorInstrCost(Instruction::InsertElement, 3041 VectorTy); 3042 unsigned ExtCost = TTI->getVectorInstrCost(Instruction::ExtractElement, 3043 VectorTy); 3044 3045 // The cost of inserting the results plus extracting each one of the 3046 // operands. 3047 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 3048 } 3049 3050 // The cost of executing VF copies of the scalar instruction. This opcode 3051 // is unknown. Assume that it is the same as 'mul'. 3052 Cost += VF * TTI->getArithmeticInstrCost(Instruction::Mul, VectorTy); 3053 return Cost; 3054 } 3055 }// end of switch. 3056} 3057 3058Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) { 3059 if (Scalar->isVoidTy() || VF == 1) 3060 return Scalar; 3061 return VectorType::get(Scalar, VF); 3062} 3063 3064char LoopVectorize::ID = 0; 3065static const char lv_name[] = "Loop Vectorization"; 3066INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 3067INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3068INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3069INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 3070INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 3071 3072namespace llvm { 3073 Pass *createLoopVectorizePass() { 3074 return new LoopVectorize(); 3075 } 3076} 3077 3078 3079