LoopVectorize.cpp revision acc47c738d5888f7445275095a174c7f48296449
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 vectorizer 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 iterations into a single 16// 'wide' iteration. After this transformation the index is incremented 17// by the SIMD vector width, and not by one. 18// 19// This pass has three parts: 20// 1. The main loop pass that drives the different parts. 21// 2. LoopVectorizationLegality - A unit that checks for the legality 22// of the vectorization. 23// 3. InnerLoopVectorizer - A unit that performs the actual 24// widening of instructions. 25// 4. LoopVectorizationCostModel - A unit that checks for the profitability 26// of vectorization. It decides on the optimal vector width, which 27// can be one, if vectorization is not profitable. 28// 29//===----------------------------------------------------------------------===// 30// 31// The reduction-variable vectorization is based on the paper: 32// D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. 33// 34// Variable uniformity checks are inspired by: 35// Karrenberg, R. and Hack, S. Whole Function Vectorization. 36// 37// Other ideas/concepts are from: 38// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. 39// 40// S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of 41// Vectorizing Compilers. 42// 43//===----------------------------------------------------------------------===// 44 45#define LV_NAME "loop-vectorize" 46#define DEBUG_TYPE LV_NAME 47 48#include "llvm/Transforms/Vectorize.h" 49#include "llvm/ADT/DenseMap.h" 50#include "llvm/ADT/MapVector.h" 51#include "llvm/ADT/SmallPtrSet.h" 52#include "llvm/ADT/SmallSet.h" 53#include "llvm/ADT/SmallVector.h" 54#include "llvm/ADT/StringExtras.h" 55#include "llvm/Analysis/AliasAnalysis.h" 56#include "llvm/Analysis/AliasSetTracker.h" 57#include "llvm/Analysis/Dominators.h" 58#include "llvm/Analysis/LoopInfo.h" 59#include "llvm/Analysis/LoopIterator.h" 60#include "llvm/Analysis/LoopPass.h" 61#include "llvm/Analysis/ScalarEvolution.h" 62#include "llvm/Analysis/ScalarEvolutionExpander.h" 63#include "llvm/Analysis/ScalarEvolutionExpressions.h" 64#include "llvm/Analysis/TargetTransformInfo.h" 65#include "llvm/Analysis/ValueTracking.h" 66#include "llvm/Analysis/Verifier.h" 67#include "llvm/IR/Constants.h" 68#include "llvm/IR/DataLayout.h" 69#include "llvm/IR/DerivedTypes.h" 70#include "llvm/IR/Function.h" 71#include "llvm/IR/IRBuilder.h" 72#include "llvm/IR/Instructions.h" 73#include "llvm/IR/IntrinsicInst.h" 74#include "llvm/IR/LLVMContext.h" 75#include "llvm/IR/Module.h" 76#include "llvm/IR/Type.h" 77#include "llvm/IR/Value.h" 78#include "llvm/Pass.h" 79#include "llvm/Support/CommandLine.h" 80#include "llvm/Support/Debug.h" 81#include "llvm/Support/PatternMatch.h" 82#include "llvm/Support/raw_ostream.h" 83#include "llvm/Target/TargetLibraryInfo.h" 84#include "llvm/Transforms/Scalar.h" 85#include "llvm/Transforms/Utils/BasicBlockUtils.h" 86#include "llvm/Transforms/Utils/Local.h" 87#include <algorithm> 88#include <map> 89 90using namespace llvm; 91using namespace llvm::PatternMatch; 92 93static cl::opt<unsigned> 94VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden, 95 cl::desc("Sets the SIMD width. Zero is autoselect.")); 96 97static cl::opt<unsigned> 98VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden, 99 cl::desc("Sets the vectorization unroll count. " 100 "Zero is autoselect.")); 101 102static cl::opt<bool> 103EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, 104 cl::desc("Enable if-conversion during vectorization.")); 105 106/// We don't vectorize loops with a known constant trip count below this number. 107static cl::opt<unsigned> 108TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16), 109 cl::Hidden, 110 cl::desc("Don't vectorize loops with a constant " 111 "trip count that is smaller than this " 112 "value.")); 113 114/// We don't unroll loops with a known constant trip count below this number. 115static const unsigned TinyTripCountUnrollThreshold = 128; 116 117/// When performing memory disambiguation checks at runtime do not make more 118/// than this number of comparisons. 119static const unsigned RuntimeMemoryCheckThreshold = 8; 120 121/// We use a metadata with this name to indicate that a scalar loop was 122/// vectorized and that we don't need to re-vectorize it if we run into it 123/// again. 124static const char* 125AlreadyVectorizedMDName = "llvm.vectorizer.already_vectorized"; 126 127namespace { 128 129// Forward declarations. 130class LoopVectorizationLegality; 131class LoopVectorizationCostModel; 132 133/// InnerLoopVectorizer vectorizes loops which contain only one basic 134/// block to a specified vectorization factor (VF). 135/// This class performs the widening of scalars into vectors, or multiple 136/// scalars. This class also implements the following features: 137/// * It inserts an epilogue loop for handling loops that don't have iteration 138/// counts that are known to be a multiple of the vectorization factor. 139/// * It handles the code generation for reduction variables. 140/// * Scalarization (implementation using scalars) of un-vectorizable 141/// instructions. 142/// InnerLoopVectorizer does not perform any vectorization-legality 143/// checks, and relies on the caller to check for the different legality 144/// aspects. The InnerLoopVectorizer relies on the 145/// LoopVectorizationLegality class to provide information about the induction 146/// and reduction variables that were found to a given vectorization factor. 147class InnerLoopVectorizer { 148public: 149 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI, 150 DominatorTree *DT, DataLayout *DL, 151 const TargetLibraryInfo *TLI, unsigned VecWidth, 152 unsigned UnrollFactor) 153 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI), 154 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0), 155 OldInduction(0), WidenMap(UnrollFactor) {} 156 157 // Perform the actual loop widening (vectorization). 158 void vectorize(LoopVectorizationLegality *Legal) { 159 // Create a new empty loop. Unlink the old loop and connect the new one. 160 createEmptyLoop(Legal); 161 // Widen each instruction in the old loop to a new one in the new loop. 162 // Use the Legality module to find the induction and reduction variables. 163 vectorizeLoop(Legal); 164 // Register the new loop and update the analysis passes. 165 updateAnalysis(); 166 } 167 168private: 169 /// A small list of PHINodes. 170 typedef SmallVector<PHINode*, 4> PhiVector; 171 /// When we unroll loops we have multiple vector values for each scalar. 172 /// This data structure holds the unrolled and vectorized values that 173 /// originated from one scalar instruction. 174 typedef SmallVector<Value*, 2> VectorParts; 175 176 /// Add code that checks at runtime if the accessed arrays overlap. 177 /// Returns the comparator value or NULL if no check is needed. 178 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal, 179 Instruction *Loc); 180 /// Create an empty loop, based on the loop ranges of the old loop. 181 void createEmptyLoop(LoopVectorizationLegality *Legal); 182 /// Copy and widen the instructions from the old loop. 183 void vectorizeLoop(LoopVectorizationLegality *Legal); 184 185 /// A helper function that computes the predicate of the block BB, assuming 186 /// that the header block of the loop is set to True. It returns the *entry* 187 /// mask for the block BB. 188 VectorParts createBlockInMask(BasicBlock *BB); 189 /// A helper function that computes the predicate of the edge between SRC 190 /// and DST. 191 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst); 192 193 /// A helper function to vectorize a single BB within the innermost loop. 194 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB, 195 PhiVector *PV); 196 197 /// Insert the new loop to the loop hierarchy and pass manager 198 /// and update the analysis passes. 199 void updateAnalysis(); 200 201 /// This instruction is un-vectorizable. Implement it as a sequence 202 /// of scalars. 203 void scalarizeInstruction(Instruction *Instr); 204 205 /// Vectorize Load and Store instructions, 206 void vectorizeMemoryInstruction(Instruction *Instr, 207 LoopVectorizationLegality *Legal); 208 209 /// Create a broadcast instruction. This method generates a broadcast 210 /// instruction (shuffle) for loop invariant values and for the induction 211 /// value. If this is the induction variable then we extend it to N, N+1, ... 212 /// this is needed because each iteration in the loop corresponds to a SIMD 213 /// element. 214 Value *getBroadcastInstrs(Value *V); 215 216 /// This function adds 0, 1, 2 ... to each vector element, starting at zero. 217 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...). 218 /// The sequence starts at StartIndex. 219 Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate); 220 221 /// When we go over instructions in the basic block we rely on previous 222 /// values within the current basic block or on loop invariant values. 223 /// When we widen (vectorize) values we place them in the map. If the values 224 /// are not within the map, they have to be loop invariant, so we simply 225 /// broadcast them into a vector. 226 VectorParts &getVectorValue(Value *V); 227 228 /// Generate a shuffle sequence that will reverse the vector Vec. 229 Value *reverseVector(Value *Vec); 230 231 /// This is a helper class that holds the vectorizer state. It maps scalar 232 /// instructions to vector instructions. When the code is 'unrolled' then 233 /// then a single scalar value is mapped to multiple vector parts. The parts 234 /// are stored in the VectorPart type. 235 struct ValueMap { 236 /// C'tor. UnrollFactor controls the number of vectors ('parts') that 237 /// are mapped. 238 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {} 239 240 /// \return True if 'Key' is saved in the Value Map. 241 bool has(Value *Key) const { return MapStorage.count(Key); } 242 243 /// Initializes a new entry in the map. Sets all of the vector parts to the 244 /// save value in 'Val'. 245 /// \return A reference to a vector with splat values. 246 VectorParts &splat(Value *Key, Value *Val) { 247 VectorParts &Entry = MapStorage[Key]; 248 Entry.assign(UF, Val); 249 return Entry; 250 } 251 252 ///\return A reference to the value that is stored at 'Key'. 253 VectorParts &get(Value *Key) { 254 VectorParts &Entry = MapStorage[Key]; 255 if (Entry.empty()) 256 Entry.resize(UF); 257 assert(Entry.size() == UF); 258 return Entry; 259 } 260 261 private: 262 /// The unroll factor. Each entry in the map stores this number of vector 263 /// elements. 264 unsigned UF; 265 266 /// Map storage. We use std::map and not DenseMap because insertions to a 267 /// dense map invalidates its iterators. 268 std::map<Value *, VectorParts> MapStorage; 269 }; 270 271 /// The original loop. 272 Loop *OrigLoop; 273 /// Scev analysis to use. 274 ScalarEvolution *SE; 275 /// Loop Info. 276 LoopInfo *LI; 277 /// Dominator Tree. 278 DominatorTree *DT; 279 /// Data Layout. 280 DataLayout *DL; 281 /// Target Library Info. 282 const TargetLibraryInfo *TLI; 283 284 /// The vectorization SIMD factor to use. Each vector will have this many 285 /// vector elements. 286 unsigned VF; 287 /// The vectorization unroll factor to use. Each scalar is vectorized to this 288 /// many different vector instructions. 289 unsigned UF; 290 291 /// The builder that we use 292 IRBuilder<> Builder; 293 294 // --- Vectorization state --- 295 296 /// The vector-loop preheader. 297 BasicBlock *LoopVectorPreHeader; 298 /// The scalar-loop preheader. 299 BasicBlock *LoopScalarPreHeader; 300 /// Middle Block between the vector and the scalar. 301 BasicBlock *LoopMiddleBlock; 302 ///The ExitBlock of the scalar loop. 303 BasicBlock *LoopExitBlock; 304 ///The vector loop body. 305 BasicBlock *LoopVectorBody; 306 ///The scalar loop body. 307 BasicBlock *LoopScalarBody; 308 /// A list of all bypass blocks. The first block is the entry of the loop. 309 SmallVector<BasicBlock *, 4> LoopBypassBlocks; 310 311 /// The new Induction variable which was added to the new block. 312 PHINode *Induction; 313 /// The induction variable of the old basic block. 314 PHINode *OldInduction; 315 /// Maps scalars to widened vectors. 316 ValueMap WidenMap; 317}; 318 319/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and 320/// to what vectorization factor. 321/// This class does not look at the profitability of vectorization, only the 322/// legality. This class has two main kinds of checks: 323/// * Memory checks - The code in canVectorizeMemory checks if vectorization 324/// will change the order of memory accesses in a way that will change the 325/// correctness of the program. 326/// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory 327/// checks for a number of different conditions, such as the availability of a 328/// single induction variable, that all types are supported and vectorize-able, 329/// etc. This code reflects the capabilities of InnerLoopVectorizer. 330/// This class is also used by InnerLoopVectorizer for identifying 331/// induction variable and the different reduction variables. 332class LoopVectorizationLegality { 333public: 334 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL, 335 DominatorTree *DT, TargetTransformInfo* TTI, 336 AliasAnalysis *AA, TargetLibraryInfo *TLI) 337 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI), 338 Induction(0), HasFunNoNaNAttr(false) {} 339 340 /// This enum represents the kinds of reductions that we support. 341 enum ReductionKind { 342 RK_NoReduction, ///< Not a reduction. 343 RK_IntegerAdd, ///< Sum of integers. 344 RK_IntegerMult, ///< Product of integers. 345 RK_IntegerOr, ///< Bitwise or logical OR of numbers. 346 RK_IntegerAnd, ///< Bitwise or logical AND of numbers. 347 RK_IntegerXor, ///< Bitwise or logical XOR of numbers. 348 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()). 349 RK_FloatAdd, ///< Sum of floats. 350 RK_FloatMult, ///< Product of floats. 351 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()). 352 }; 353 354 /// This enum represents the kinds of inductions that we support. 355 enum InductionKind { 356 IK_NoInduction, ///< Not an induction variable. 357 IK_IntInduction, ///< Integer induction variable. Step = 1. 358 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1. 359 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem). 360 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem). 361 }; 362 363 // This enum represents the kind of minmax reduction. 364 enum MinMaxReductionKind { 365 MRK_Invalid, 366 MRK_UIntMin, 367 MRK_UIntMax, 368 MRK_SIntMin, 369 MRK_SIntMax, 370 MRK_FloatMin, 371 MRK_FloatMax 372 }; 373 374 /// This POD struct holds information about reduction variables. 375 struct ReductionDescriptor { 376 ReductionDescriptor() : StartValue(0), LoopExitInstr(0), 377 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {} 378 379 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K, 380 MinMaxReductionKind MK) 381 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {} 382 383 // The starting value of the reduction. 384 // It does not have to be zero! 385 Value *StartValue; 386 // The instruction who's value is used outside the loop. 387 Instruction *LoopExitInstr; 388 // The kind of the reduction. 389 ReductionKind Kind; 390 // If this a min/max reduction the kind of reduction. 391 MinMaxReductionKind MinMaxKind; 392 }; 393 394 /// This POD struct holds information about a potential reduction operation. 395 struct ReductionInstDesc { 396 ReductionInstDesc(bool IsRedux, Instruction *I) : 397 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {} 398 399 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) : 400 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {} 401 402 // Is this instruction a reduction candidate. 403 bool IsReduction; 404 // The last instruction in a min/max pattern (select of the select(icmp()) 405 // pattern), or the current reduction instruction otherwise. 406 Instruction *PatternLastInst; 407 // If this is a min/max pattern the comparison predicate. 408 MinMaxReductionKind MinMaxKind; 409 }; 410 411 // This POD struct holds information about the memory runtime legality 412 // check that a group of pointers do not overlap. 413 struct RuntimePointerCheck { 414 RuntimePointerCheck() : Need(false) {} 415 416 /// Reset the state of the pointer runtime information. 417 void reset() { 418 Need = false; 419 Pointers.clear(); 420 Starts.clear(); 421 Ends.clear(); 422 } 423 424 /// Insert a pointer and calculate the start and end SCEVs. 425 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr); 426 427 /// This flag indicates if we need to add the runtime check. 428 bool Need; 429 /// Holds the pointers that we need to check. 430 SmallVector<Value*, 2> Pointers; 431 /// Holds the pointer value at the beginning of the loop. 432 SmallVector<const SCEV*, 2> Starts; 433 /// Holds the pointer value at the end of the loop. 434 SmallVector<const SCEV*, 2> Ends; 435 /// Holds the information if this pointer is used for writing to memory. 436 SmallVector<bool, 2> IsWritePtr; 437 }; 438 439 /// A POD for saving information about induction variables. 440 struct InductionInfo { 441 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {} 442 InductionInfo() : StartValue(0), IK(IK_NoInduction) {} 443 /// Start value. 444 Value *StartValue; 445 /// Induction kind. 446 InductionKind IK; 447 }; 448 449 /// ReductionList contains the reduction descriptors for all 450 /// of the reductions that were found in the loop. 451 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList; 452 453 /// InductionList saves induction variables and maps them to the 454 /// induction descriptor. 455 typedef MapVector<PHINode*, InductionInfo> InductionList; 456 457 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their 458 /// respective Store/Load instruction(s) to calculate aliasing. 459 typedef MapVector<Value*, Instruction* > AliasMap; 460 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap; 461 462 /// Returns true if it is legal to vectorize this loop. 463 /// This does not mean that it is profitable to vectorize this 464 /// loop, only that it is legal to do so. 465 bool canVectorize(); 466 467 /// Returns the Induction variable. 468 PHINode *getInduction() { return Induction; } 469 470 /// Returns the reduction variables found in the loop. 471 ReductionList *getReductionVars() { return &Reductions; } 472 473 /// Returns the induction variables found in the loop. 474 InductionList *getInductionVars() { return &Inductions; } 475 476 /// Returns True if V is an induction variable in this loop. 477 bool isInductionVariable(const Value *V); 478 479 /// Return true if the block BB needs to be predicated in order for the loop 480 /// to be vectorized. 481 bool blockNeedsPredication(BasicBlock *BB); 482 483 /// Check if this pointer is consecutive when vectorizing. This happens 484 /// when the last index of the GEP is the induction variable, or that the 485 /// pointer itself is an induction variable. 486 /// This check allows us to vectorize A[idx] into a wide load/store. 487 /// Returns: 488 /// 0 - Stride is unknown or non consecutive. 489 /// 1 - Address is consecutive. 490 /// -1 - Address is consecutive, and decreasing. 491 int isConsecutivePtr(Value *Ptr); 492 493 /// Returns true if the value V is uniform within the loop. 494 bool isUniform(Value *V); 495 496 /// Returns true if this instruction will remain scalar after vectorization. 497 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } 498 499 /// Returns the information that we collected about runtime memory check. 500 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; } 501 502 /// This function returns the identity element (or neutral element) for 503 /// the operation K. 504 static Constant *getReductionIdentity(ReductionKind K, Type *Tp); 505private: 506 /// Check if a single basic block loop is vectorizable. 507 /// At this point we know that this is a loop with a constant trip count 508 /// and we only need to check individual instructions. 509 bool canVectorizeInstrs(); 510 511 /// When we vectorize loops we may change the order in which 512 /// we read and write from memory. This method checks if it is 513 /// legal to vectorize the code, considering only memory constrains. 514 /// Returns true if the loop is vectorizable 515 bool canVectorizeMemory(); 516 517 /// Return true if we can vectorize this loop using the IF-conversion 518 /// transformation. 519 bool canVectorizeWithIfConvert(); 520 521 /// Collect the variables that need to stay uniform after vectorization. 522 void collectLoopUniforms(); 523 524 /// Return true if all of the instructions in the block can be speculatively 525 /// executed. 526 bool blockCanBePredicated(BasicBlock *BB); 527 528 /// Returns True, if 'Phi' is the kind of reduction variable for type 529 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. 530 bool AddReductionVar(PHINode *Phi, ReductionKind Kind); 531 /// Returns a struct describing if the instruction 'I' can be a reduction 532 /// variable of type 'Kind'. If the reduction is a min/max pattern of 533 /// select(icmp()) this function advances the instruction pointer 'I' from the 534 /// compare instruction to the select instruction and stores this pointer in 535 /// 'PatternLastInst' member of the returned struct. 536 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind, 537 ReductionInstDesc &Desc); 538 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 539 /// pattern corresponding to a min(X, Y) or max(X, Y). 540 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I, 541 ReductionInstDesc &Prev); 542 /// Returns the induction kind of Phi. This function may return NoInduction 543 /// if the PHI is not an induction variable. 544 InductionKind isInductionVariable(PHINode *Phi); 545 /// Return true if can compute the address bounds of Ptr within the loop. 546 bool hasComputableBounds(Value *Ptr); 547 /// Return true if there is the chance of write reorder. 548 bool hasPossibleGlobalWriteReorder(Value *Object, 549 Instruction *Inst, 550 AliasMultiMap &WriteObjects, 551 unsigned MaxByteWidth); 552 /// Return the AA location for a load or a store. 553 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst); 554 555 556 /// The loop that we evaluate. 557 Loop *TheLoop; 558 /// Scev analysis. 559 ScalarEvolution *SE; 560 /// DataLayout analysis. 561 DataLayout *DL; 562 /// Dominators. 563 DominatorTree *DT; 564 /// Target Info. 565 TargetTransformInfo *TTI; 566 /// Alias Analysis. 567 AliasAnalysis *AA; 568 /// Target Library Info. 569 TargetLibraryInfo *TLI; 570 571 // --- vectorization state --- // 572 573 /// Holds the integer induction variable. This is the counter of the 574 /// loop. 575 PHINode *Induction; 576 /// Holds the reduction variables. 577 ReductionList Reductions; 578 /// Holds all of the induction variables that we found in the loop. 579 /// Notice that inductions don't need to start at zero and that induction 580 /// variables can be pointers. 581 InductionList Inductions; 582 583 /// Allowed outside users. This holds the reduction 584 /// vars which can be accessed from outside the loop. 585 SmallPtrSet<Value*, 4> AllowedExit; 586 /// This set holds the variables which are known to be uniform after 587 /// vectorization. 588 SmallPtrSet<Instruction*, 4> Uniforms; 589 /// We need to check that all of the pointers in this list are disjoint 590 /// at runtime. 591 RuntimePointerCheck PtrRtCheck; 592 /// Can we assume the absence of NaNs. 593 bool HasFunNoNaNAttr; 594}; 595 596/// LoopVectorizationCostModel - estimates the expected speedups due to 597/// vectorization. 598/// In many cases vectorization is not profitable. This can happen because of 599/// a number of reasons. In this class we mainly attempt to predict the 600/// expected speedup/slowdowns due to the supported instruction set. We use the 601/// TargetTransformInfo to query the different backends for the cost of 602/// different operations. 603class LoopVectorizationCostModel { 604public: 605 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI, 606 LoopVectorizationLegality *Legal, 607 const TargetTransformInfo &TTI, 608 DataLayout *DL, const TargetLibraryInfo *TLI) 609 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {} 610 611 /// Information about vectorization costs 612 struct VectorizationFactor { 613 unsigned Width; // Vector width with best cost 614 unsigned Cost; // Cost of the loop with that width 615 }; 616 /// \return The most profitable vectorization factor and the cost of that VF. 617 /// This method checks every power of two up to VF. If UserVF is not ZERO 618 /// then this vectorization factor will be selected if vectorization is 619 /// possible. 620 VectorizationFactor selectVectorizationFactor(bool OptForSize, 621 unsigned UserVF); 622 623 /// \return The size (in bits) of the widest type in the code that 624 /// needs to be vectorized. We ignore values that remain scalar such as 625 /// 64 bit loop indices. 626 unsigned getWidestType(); 627 628 /// \return The most profitable unroll factor. 629 /// If UserUF is non-zero then this method finds the best unroll-factor 630 /// based on register pressure and other parameters. 631 /// VF and LoopCost are the selected vectorization factor and the cost of the 632 /// selected VF. 633 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF, 634 unsigned LoopCost); 635 636 /// \brief A struct that represents some properties of the register usage 637 /// of a loop. 638 struct RegisterUsage { 639 /// Holds the number of loop invariant values that are used in the loop. 640 unsigned LoopInvariantRegs; 641 /// Holds the maximum number of concurrent live intervals in the loop. 642 unsigned MaxLocalUsers; 643 /// Holds the number of instructions in the loop. 644 unsigned NumInstructions; 645 }; 646 647 /// \return information about the register usage of the loop. 648 RegisterUsage calculateRegisterUsage(); 649 650private: 651 /// Returns the expected execution cost. The unit of the cost does 652 /// not matter because we use the 'cost' units to compare different 653 /// vector widths. The cost that is returned is *not* normalized by 654 /// the factor width. 655 unsigned expectedCost(unsigned VF); 656 657 /// Returns the execution time cost of an instruction for a given vector 658 /// width. Vector width of one means scalar. 659 unsigned getInstructionCost(Instruction *I, unsigned VF); 660 661 /// A helper function for converting Scalar types to vector types. 662 /// If the incoming type is void, we return void. If the VF is 1, we return 663 /// the scalar type. 664 static Type* ToVectorTy(Type *Scalar, unsigned VF); 665 666 /// Returns whether the instruction is a load or store and will be a emitted 667 /// as a vector operation. 668 bool isConsecutiveLoadOrStore(Instruction *I); 669 670 /// The loop that we evaluate. 671 Loop *TheLoop; 672 /// Scev analysis. 673 ScalarEvolution *SE; 674 /// Loop Info analysis. 675 LoopInfo *LI; 676 /// Vectorization legality. 677 LoopVectorizationLegality *Legal; 678 /// Vector target information. 679 const TargetTransformInfo &TTI; 680 /// Target data layout information. 681 DataLayout *DL; 682 /// Target Library Info. 683 const TargetLibraryInfo *TLI; 684}; 685 686/// The LoopVectorize Pass. 687struct LoopVectorize : public LoopPass { 688 /// Pass identification, replacement for typeid 689 static char ID; 690 691 explicit LoopVectorize() : LoopPass(ID) { 692 initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); 693 } 694 695 ScalarEvolution *SE; 696 DataLayout *DL; 697 LoopInfo *LI; 698 TargetTransformInfo *TTI; 699 DominatorTree *DT; 700 AliasAnalysis *AA; 701 TargetLibraryInfo *TLI; 702 703 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) { 704 // We only vectorize innermost loops. 705 if (!L->empty()) 706 return false; 707 708 SE = &getAnalysis<ScalarEvolution>(); 709 DL = getAnalysisIfAvailable<DataLayout>(); 710 LI = &getAnalysis<LoopInfo>(); 711 TTI = &getAnalysis<TargetTransformInfo>(); 712 DT = &getAnalysis<DominatorTree>(); 713 AA = getAnalysisIfAvailable<AliasAnalysis>(); 714 TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); 715 716 if (DL == NULL) { 717 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout"); 718 return false; 719 } 720 721 DEBUG(dbgs() << "LV: Checking a loop in \"" << 722 L->getHeader()->getParent()->getName() << "\"\n"); 723 724 // Check if it is legal to vectorize the loop. 725 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI); 726 if (!LVL.canVectorize()) { 727 DEBUG(dbgs() << "LV: Not vectorizing.\n"); 728 return false; 729 } 730 731 // Use the cost model. 732 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI); 733 734 // Check the function attributes to find out if this function should be 735 // optimized for size. 736 Function *F = L->getHeader()->getParent(); 737 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize; 738 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat; 739 unsigned FnIndex = AttributeSet::FunctionIndex; 740 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr); 741 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr); 742 743 if (NoFloat) { 744 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" 745 "attribute is used.\n"); 746 return false; 747 } 748 749 // Select the optimal vectorization factor. 750 LoopVectorizationCostModel::VectorizationFactor VF; 751 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor); 752 // Select the unroll factor. 753 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll, 754 VF.Width, VF.Cost); 755 756 if (VF.Width == 1) { 757 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); 758 return false; 759 } 760 761 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<< 762 F->getParent()->getModuleIdentifier()<<"\n"); 763 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n"); 764 765 // If we decided that it is *legal* to vectorize the loop then do it. 766 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF); 767 LB.vectorize(&LVL); 768 769 DEBUG(verifyFunction(*L->getHeader()->getParent())); 770 return true; 771 } 772 773 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 774 LoopPass::getAnalysisUsage(AU); 775 AU.addRequiredID(LoopSimplifyID); 776 AU.addRequiredID(LCSSAID); 777 AU.addRequired<DominatorTree>(); 778 AU.addRequired<LoopInfo>(); 779 AU.addRequired<ScalarEvolution>(); 780 AU.addRequired<TargetTransformInfo>(); 781 AU.addPreserved<LoopInfo>(); 782 AU.addPreserved<DominatorTree>(); 783 } 784 785}; 786 787} // end anonymous namespace 788 789//===----------------------------------------------------------------------===// 790// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and 791// LoopVectorizationCostModel. 792//===----------------------------------------------------------------------===// 793 794void 795LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE, 796 Loop *Lp, Value *Ptr, 797 bool WritePtr) { 798 const SCEV *Sc = SE->getSCEV(Ptr); 799 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); 800 assert(AR && "Invalid addrec expression"); 801 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch()); 802 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE); 803 Pointers.push_back(Ptr); 804 Starts.push_back(AR->getStart()); 805 Ends.push_back(ScEnd); 806 IsWritePtr.push_back(WritePtr); 807} 808 809Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { 810 // Save the current insertion location. 811 Instruction *Loc = Builder.GetInsertPoint(); 812 813 // We need to place the broadcast of invariant variables outside the loop. 814 Instruction *Instr = dyn_cast<Instruction>(V); 815 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody); 816 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr; 817 818 // Place the code for broadcasting invariant variables in the new preheader. 819 if (Invariant) 820 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); 821 822 // Broadcast the scalar into all locations in the vector. 823 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); 824 825 // Restore the builder insertion point. 826 if (Invariant) 827 Builder.SetInsertPoint(Loc); 828 829 return Shuf; 830} 831 832Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx, 833 bool Negate) { 834 assert(Val->getType()->isVectorTy() && "Must be a vector"); 835 assert(Val->getType()->getScalarType()->isIntegerTy() && 836 "Elem must be an integer"); 837 // Create the types. 838 Type *ITy = Val->getType()->getScalarType(); 839 VectorType *Ty = cast<VectorType>(Val->getType()); 840 int VLen = Ty->getNumElements(); 841 SmallVector<Constant*, 8> Indices; 842 843 // Create a vector of consecutive numbers from zero to VF. 844 for (int i = 0; i < VLen; ++i) { 845 int Idx = Negate ? (-i): i; 846 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx)); 847 } 848 849 // Add the consecutive indices to the vector value. 850 Constant *Cv = ConstantVector::get(Indices); 851 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); 852 return Builder.CreateAdd(Val, Cv, "induction"); 853} 854 855int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { 856 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr"); 857 // Make sure that the pointer does not point to structs. 858 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) 859 return 0; 860 861 // If this value is a pointer induction variable we know it is consecutive. 862 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr); 863 if (Phi && Inductions.count(Phi)) { 864 InductionInfo II = Inductions[Phi]; 865 if (IK_PtrInduction == II.IK) 866 return 1; 867 else if (IK_ReversePtrInduction == II.IK) 868 return -1; 869 } 870 871 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr); 872 if (!Gep) 873 return 0; 874 875 unsigned NumOperands = Gep->getNumOperands(); 876 Value *LastIndex = Gep->getOperand(NumOperands - 1); 877 878 Value *GpPtr = Gep->getPointerOperand(); 879 // If this GEP value is a consecutive pointer induction variable and all of 880 // the indices are constant then we know it is consecutive. We can 881 Phi = dyn_cast<PHINode>(GpPtr); 882 if (Phi && Inductions.count(Phi)) { 883 884 // Make sure that the pointer does not point to structs. 885 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType()); 886 if (GepPtrType->getElementType()->isAggregateType()) 887 return 0; 888 889 // Make sure that all of the index operands are loop invariant. 890 for (unsigned i = 1; i < NumOperands; ++i) 891 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 892 return 0; 893 894 InductionInfo II = Inductions[Phi]; 895 if (IK_PtrInduction == II.IK) 896 return 1; 897 else if (IK_ReversePtrInduction == II.IK) 898 return -1; 899 } 900 901 // Check that all of the gep indices are uniform except for the last. 902 for (unsigned i = 0; i < NumOperands - 1; ++i) 903 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 904 return 0; 905 906 // We can emit wide load/stores only if the last index is the induction 907 // variable. 908 const SCEV *Last = SE->getSCEV(LastIndex); 909 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { 910 const SCEV *Step = AR->getStepRecurrence(*SE); 911 912 // The memory is consecutive because the last index is consecutive 913 // and all other indices are loop invariant. 914 if (Step->isOne()) 915 return 1; 916 if (Step->isAllOnesValue()) 917 return -1; 918 } 919 920 return 0; 921} 922 923bool LoopVectorizationLegality::isUniform(Value *V) { 924 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); 925} 926 927InnerLoopVectorizer::VectorParts& 928InnerLoopVectorizer::getVectorValue(Value *V) { 929 assert(V != Induction && "The new induction variable should not be used."); 930 assert(!V->getType()->isVectorTy() && "Can't widen a vector"); 931 932 // If we have this scalar in the map, return it. 933 if (WidenMap.has(V)) 934 return WidenMap.get(V); 935 936 // If this scalar is unknown, assume that it is a constant or that it is 937 // loop invariant. Broadcast V and save the value for future uses. 938 Value *B = getBroadcastInstrs(V); 939 return WidenMap.splat(V, B); 940} 941 942Value *InnerLoopVectorizer::reverseVector(Value *Vec) { 943 assert(Vec->getType()->isVectorTy() && "Invalid type"); 944 SmallVector<Constant*, 8> ShuffleMask; 945 for (unsigned i = 0; i < VF; ++i) 946 ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); 947 948 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), 949 ConstantVector::get(ShuffleMask), 950 "reverse"); 951} 952 953 954void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr, 955 LoopVectorizationLegality *Legal) { 956 // Attempt to issue a wide load. 957 LoadInst *LI = dyn_cast<LoadInst>(Instr); 958 StoreInst *SI = dyn_cast<StoreInst>(Instr); 959 960 assert((LI || SI) && "Invalid Load/Store instruction"); 961 962 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType(); 963 Type *DataTy = VectorType::get(ScalarDataTy, VF); 964 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 965 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment(); 966 967 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy); 968 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF; 969 970 if (ScalarAllocatedSize != VectorElementSize) 971 return scalarizeInstruction(Instr); 972 973 // If the pointer is loop invariant or if it is non consecutive, 974 // scalarize the load. 975 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 976 bool Reverse = ConsecutiveStride < 0; 977 bool UniformLoad = LI && Legal->isUniform(Ptr); 978 if (!ConsecutiveStride || UniformLoad) 979 return scalarizeInstruction(Instr); 980 981 Constant *Zero = Builder.getInt32(0); 982 VectorParts &Entry = WidenMap.get(Instr); 983 984 // Handle consecutive loads/stores. 985 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 986 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) { 987 Value *PtrOperand = Gep->getPointerOperand(); 988 Value *FirstBasePtr = getVectorValue(PtrOperand)[0]; 989 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero); 990 991 // Create the new GEP with the new induction variable. 992 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 993 Gep2->setOperand(0, FirstBasePtr); 994 Gep2->setName("gep.indvar.base"); 995 Ptr = Builder.Insert(Gep2); 996 } else if (Gep) { 997 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), 998 OrigLoop) && "Base ptr must be invariant"); 999 1000 // The last index does not have to be the induction. It can be 1001 // consecutive and be a function of the index. For example A[I+1]; 1002 unsigned NumOperands = Gep->getNumOperands(); 1003 1004 Value *LastGepOperand = Gep->getOperand(NumOperands - 1); 1005 VectorParts &GEPParts = getVectorValue(LastGepOperand); 1006 Value *LastIndex = GEPParts[0]; 1007 LastIndex = Builder.CreateExtractElement(LastIndex, Zero); 1008 1009 // Create the new GEP with the new induction variable. 1010 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1011 Gep2->setOperand(NumOperands - 1, LastIndex); 1012 Gep2->setName("gep.indvar.idx"); 1013 Ptr = Builder.Insert(Gep2); 1014 } else { 1015 // Use the induction element ptr. 1016 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 1017 VectorParts &PtrVal = getVectorValue(Ptr); 1018 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 1019 } 1020 1021 // Handle Stores: 1022 if (SI) { 1023 assert(!Legal->isUniform(SI->getPointerOperand()) && 1024 "We do not allow storing to uniform addresses"); 1025 1026 VectorParts &StoredVal = getVectorValue(SI->getValueOperand()); 1027 for (unsigned Part = 0; Part < UF; ++Part) { 1028 // Calculate the pointer for the specific unroll-part. 1029 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1030 1031 if (Reverse) { 1032 // If we store to reverse consecutive memory locations then we need 1033 // to reverse the order of elements in the stored value. 1034 StoredVal[Part] = reverseVector(StoredVal[Part]); 1035 // If the address is consecutive but reversed, then the 1036 // wide store needs to start at the last vector element. 1037 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1038 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1039 } 1040 1041 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); 1042 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment); 1043 } 1044 } 1045 1046 for (unsigned Part = 0; Part < UF; ++Part) { 1047 // Calculate the pointer for the specific unroll-part. 1048 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1049 1050 if (Reverse) { 1051 // If the address is consecutive but reversed, then the 1052 // wide store needs to start at the last vector element. 1053 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1054 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1055 } 1056 1057 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); 1058 Value *LI = Builder.CreateLoad(VecPtr, "wide.load"); 1059 cast<LoadInst>(LI)->setAlignment(Alignment); 1060 Entry[Part] = Reverse ? reverseVector(LI) : LI; 1061 } 1062} 1063 1064void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { 1065 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 1066 // Holds vector parameters or scalars, in case of uniform vals. 1067 SmallVector<VectorParts, 4> Params; 1068 1069 // Find all of the vectorized parameters. 1070 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1071 Value *SrcOp = Instr->getOperand(op); 1072 1073 // If we are accessing the old induction variable, use the new one. 1074 if (SrcOp == OldInduction) { 1075 Params.push_back(getVectorValue(SrcOp)); 1076 continue; 1077 } 1078 1079 // Try using previously calculated values. 1080 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 1081 1082 // If the src is an instruction that appeared earlier in the basic block 1083 // then it should already be vectorized. 1084 if (SrcInst && OrigLoop->contains(SrcInst)) { 1085 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 1086 // The parameter is a vector value from earlier. 1087 Params.push_back(WidenMap.get(SrcInst)); 1088 } else { 1089 // The parameter is a scalar from outside the loop. Maybe even a constant. 1090 VectorParts Scalars; 1091 Scalars.append(UF, SrcOp); 1092 Params.push_back(Scalars); 1093 } 1094 } 1095 1096 assert(Params.size() == Instr->getNumOperands() && 1097 "Invalid number of operands"); 1098 1099 // Does this instruction return a value ? 1100 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 1101 1102 Value *UndefVec = IsVoidRetTy ? 0 : 1103 UndefValue::get(VectorType::get(Instr->getType(), VF)); 1104 // Create a new entry in the WidenMap and initialize it to Undef or Null. 1105 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 1106 1107 // For each vector unroll 'part': 1108 for (unsigned Part = 0; Part < UF; ++Part) { 1109 // For each scalar that we create: 1110 for (unsigned Width = 0; Width < VF; ++Width) { 1111 Instruction *Cloned = Instr->clone(); 1112 if (!IsVoidRetTy) 1113 Cloned->setName(Instr->getName() + ".cloned"); 1114 // Replace the operands of the cloned instrucions with extracted scalars. 1115 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1116 Value *Op = Params[op][Part]; 1117 // Param is a vector. Need to extract the right lane. 1118 if (Op->getType()->isVectorTy()) 1119 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); 1120 Cloned->setOperand(op, Op); 1121 } 1122 1123 // Place the cloned scalar in the new loop. 1124 Builder.Insert(Cloned); 1125 1126 // If the original scalar returns a value we need to place it in a vector 1127 // so that future users will be able to use it. 1128 if (!IsVoidRetTy) 1129 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, 1130 Builder.getInt32(Width)); 1131 } 1132 } 1133} 1134 1135Instruction * 1136InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, 1137 Instruction *Loc) { 1138 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck = 1139 Legal->getRuntimePointerCheck(); 1140 1141 if (!PtrRtCheck->Need) 1142 return NULL; 1143 1144 Instruction *MemoryRuntimeCheck = 0; 1145 unsigned NumPointers = PtrRtCheck->Pointers.size(); 1146 SmallVector<Value* , 2> Starts; 1147 SmallVector<Value* , 2> Ends; 1148 1149 SCEVExpander Exp(*SE, "induction"); 1150 1151 // Use this type for pointer arithmetic. 1152 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0); 1153 1154 for (unsigned i = 0; i < NumPointers; ++i) { 1155 Value *Ptr = PtrRtCheck->Pointers[i]; 1156 const SCEV *Sc = SE->getSCEV(Ptr); 1157 1158 if (SE->isLoopInvariant(Sc, OrigLoop)) { 1159 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" << 1160 *Ptr <<"\n"); 1161 Starts.push_back(Ptr); 1162 Ends.push_back(Ptr); 1163 } else { 1164 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n"); 1165 1166 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc); 1167 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc); 1168 Starts.push_back(Start); 1169 Ends.push_back(End); 1170 } 1171 } 1172 1173 IRBuilder<> ChkBuilder(Loc); 1174 1175 for (unsigned i = 0; i < NumPointers; ++i) { 1176 for (unsigned j = i+1; j < NumPointers; ++j) { 1177 // No need to check if two readonly pointers intersect. 1178 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j]) 1179 continue; 1180 1181 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc"); 1182 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc"); 1183 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc"); 1184 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc"); 1185 1186 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0"); 1187 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1"); 1188 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict"); 1189 if (MemoryRuntimeCheck) 1190 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, 1191 "conflict.rdx"); 1192 1193 MemoryRuntimeCheck = cast<Instruction>(IsConflict); 1194 } 1195 } 1196 1197 return MemoryRuntimeCheck; 1198} 1199 1200void 1201InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { 1202 /* 1203 In this function we generate a new loop. The new loop will contain 1204 the vectorized instructions while the old loop will continue to run the 1205 scalar remainder. 1206 1207 [ ] <-- vector loop bypass (may consist of multiple blocks). 1208 / | 1209 / v 1210 | [ ] <-- vector pre header. 1211 | | 1212 | v 1213 | [ ] \ 1214 | [ ]_| <-- vector loop. 1215 | | 1216 \ v 1217 >[ ] <--- middle-block. 1218 / | 1219 / v 1220 | [ ] <--- new preheader. 1221 | | 1222 | v 1223 | [ ] \ 1224 | [ ]_| <-- old scalar loop to handle remainder. 1225 \ | 1226 \ v 1227 >[ ] <-- exit block. 1228 ... 1229 */ 1230 1231 BasicBlock *OldBasicBlock = OrigLoop->getHeader(); 1232 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader(); 1233 BasicBlock *ExitBlock = OrigLoop->getExitBlock(); 1234 assert(ExitBlock && "Must have an exit block"); 1235 1236 // Mark the old scalar loop with metadata that tells us not to vectorize this 1237 // loop again if we run into it. 1238 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None); 1239 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD); 1240 1241 // Some loops have a single integer induction variable, while other loops 1242 // don't. One example is c++ iterators that often have multiple pointer 1243 // induction variables. In the code below we also support a case where we 1244 // don't have a single induction variable. 1245 OldInduction = Legal->getInduction(); 1246 Type *IdxTy = OldInduction ? OldInduction->getType() : 1247 DL->getIntPtrType(SE->getContext()); 1248 1249 // Find the loop boundaries. 1250 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch()); 1251 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count"); 1252 1253 // Get the total trip count from the count by adding 1. 1254 ExitCount = SE->getAddExpr(ExitCount, 1255 SE->getConstant(ExitCount->getType(), 1)); 1256 1257 // Expand the trip count and place the new instructions in the preheader. 1258 // Notice that the pre-header does not change, only the loop body. 1259 SCEVExpander Exp(*SE, "induction"); 1260 1261 // Count holds the overall loop count (N). 1262 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 1263 BypassBlock->getTerminator()); 1264 1265 // The loop index does not have to start at Zero. Find the original start 1266 // value from the induction PHI node. If we don't have an induction variable 1267 // then we know that it starts at zero. 1268 Value *StartIdx = OldInduction ? 1269 OldInduction->getIncomingValueForBlock(BypassBlock): 1270 ConstantInt::get(IdxTy, 0); 1271 1272 assert(BypassBlock && "Invalid loop structure"); 1273 LoopBypassBlocks.push_back(BypassBlock); 1274 1275 // Split the single block loop into the two loop structure described above. 1276 BasicBlock *VectorPH = 1277 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); 1278 BasicBlock *VecBody = 1279 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 1280 BasicBlock *MiddleBlock = 1281 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 1282 BasicBlock *ScalarPH = 1283 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 1284 1285 // Use this IR builder to create the loop instructions (Phi, Br, Cmp) 1286 // inside the loop. 1287 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1288 1289 // Generate the induction variable. 1290 Induction = Builder.CreatePHI(IdxTy, 2, "index"); 1291 // The loop step is equal to the vectorization factor (num of SIMD elements) 1292 // times the unroll factor (num of SIMD instructions). 1293 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 1294 1295 // This is the IR builder that we use to add all of the logic for bypassing 1296 // the new vector loop. 1297 IRBuilder<> BypassBuilder(BypassBlock->getTerminator()); 1298 1299 // We may need to extend the index in case there is a type mismatch. 1300 // We know that the count starts at zero and does not overflow. 1301 if (Count->getType() != IdxTy) { 1302 // The exit count can be of pointer type. Convert it to the correct 1303 // integer type. 1304 if (ExitCount->getType()->isPointerTy()) 1305 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int"); 1306 else 1307 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast"); 1308 } 1309 1310 // Add the start index to the loop count to get the new end index. 1311 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx"); 1312 1313 // Now we need to generate the expression for N - (N % VF), which is 1314 // the part that the vectorized body will execute. 1315 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf"); 1316 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec"); 1317 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx, 1318 "end.idx.rnd.down"); 1319 1320 // Now, compare the new count to zero. If it is zero skip the vector loop and 1321 // jump to the scalar loop. 1322 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, 1323 "cmp.zero"); 1324 1325 BasicBlock *LastBypassBlock = BypassBlock; 1326 1327 // Generate the code that checks in runtime if arrays overlap. We put the 1328 // checks into a separate block to make the more common case of few elements 1329 // faster. 1330 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal, 1331 BypassBlock->getTerminator()); 1332 if (MemRuntimeCheck) { 1333 // Create a new block containing the memory check. 1334 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck, 1335 "vector.memcheck"); 1336 LoopBypassBlocks.push_back(CheckBlock); 1337 1338 // Replace the branch into the memory check block with a conditional branch 1339 // for the "few elements case". 1340 Instruction *OldTerm = BypassBlock->getTerminator(); 1341 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); 1342 OldTerm->eraseFromParent(); 1343 1344 Cmp = MemRuntimeCheck; 1345 LastBypassBlock = CheckBlock; 1346 } 1347 1348 LastBypassBlock->getTerminator()->eraseFromParent(); 1349 BranchInst::Create(MiddleBlock, VectorPH, Cmp, 1350 LastBypassBlock); 1351 1352 // We are going to resume the execution of the scalar loop. 1353 // Go over all of the induction variables that we found and fix the 1354 // PHIs that are left in the scalar version of the loop. 1355 // The starting values of PHI nodes depend on the counter of the last 1356 // iteration in the vectorized loop. 1357 // If we come from a bypass edge then we need to start from the original 1358 // start value. 1359 1360 // This variable saves the new starting index for the scalar loop. 1361 PHINode *ResumeIndex = 0; 1362 LoopVectorizationLegality::InductionList::iterator I, E; 1363 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 1364 for (I = List->begin(), E = List->end(); I != E; ++I) { 1365 PHINode *OrigPhi = I->first; 1366 LoopVectorizationLegality::InductionInfo II = I->second; 1367 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val", 1368 MiddleBlock->getTerminator()); 1369 Value *EndValue = 0; 1370 switch (II.IK) { 1371 case LoopVectorizationLegality::IK_NoInduction: 1372 llvm_unreachable("Unknown induction"); 1373 case LoopVectorizationLegality::IK_IntInduction: { 1374 // Handle the integer induction counter: 1375 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); 1376 assert(OrigPhi == OldInduction && "Unknown integer PHI"); 1377 // We know what the end value is. 1378 EndValue = IdxEndRoundDown; 1379 // We also know which PHI node holds it. 1380 ResumeIndex = ResumeVal; 1381 break; 1382 } 1383 case LoopVectorizationLegality::IK_ReverseIntInduction: { 1384 // Convert the CountRoundDown variable to the PHI size. 1385 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits(); 1386 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits(); 1387 Value *CRD = CountRoundDown; 1388 if (CRDSize > IISize) 1389 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown, 1390 II.StartValue->getType(), "tr.crd", 1391 LoopBypassBlocks.back()->getTerminator()); 1392 else if (CRDSize < IISize) 1393 CRD = CastInst::Create(Instruction::SExt, CountRoundDown, 1394 II.StartValue->getType(), 1395 "sext.crd", 1396 LoopBypassBlocks.back()->getTerminator()); 1397 // Handle reverse integer induction counter: 1398 EndValue = 1399 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end", 1400 LoopBypassBlocks.back()->getTerminator()); 1401 break; 1402 } 1403 case LoopVectorizationLegality::IK_PtrInduction: { 1404 // For pointer induction variables, calculate the offset using 1405 // the end index. 1406 EndValue = 1407 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end", 1408 LoopBypassBlocks.back()->getTerminator()); 1409 break; 1410 } 1411 case LoopVectorizationLegality::IK_ReversePtrInduction: { 1412 // The value at the end of the loop for the reverse pointer is calculated 1413 // by creating a GEP with a negative index starting from the start value. 1414 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0); 1415 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown, 1416 "rev.ind.end", 1417 LoopBypassBlocks.back()->getTerminator()); 1418 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx, 1419 "rev.ptr.ind.end", 1420 LoopBypassBlocks.back()->getTerminator()); 1421 break; 1422 } 1423 }// end of case 1424 1425 // The new PHI merges the original incoming value, in case of a bypass, 1426 // or the value at the end of the vectorized loop. 1427 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1428 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1429 ResumeVal->addIncoming(EndValue, VecBody); 1430 1431 // Fix the scalar body counter (PHI node). 1432 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 1433 OrigPhi->setIncomingValue(BlockIdx, ResumeVal); 1434 } 1435 1436 // If we are generating a new induction variable then we also need to 1437 // generate the code that calculates the exit value. This value is not 1438 // simply the end of the counter because we may skip the vectorized body 1439 // in case of a runtime check. 1440 if (!OldInduction){ 1441 assert(!ResumeIndex && "Unexpected resume value found"); 1442 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", 1443 MiddleBlock->getTerminator()); 1444 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1445 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]); 1446 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); 1447 } 1448 1449 // Make sure that we found the index where scalar loop needs to continue. 1450 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() && 1451 "Invalid resume Index"); 1452 1453 // Add a check in the middle block to see if we have completed 1454 // all of the iterations in the first vector loop. 1455 // If (N - N%VF) == N, then we *don't* need to run the remainder. 1456 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, 1457 ResumeIndex, "cmp.n", 1458 MiddleBlock->getTerminator()); 1459 1460 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); 1461 // Remove the old terminator. 1462 MiddleBlock->getTerminator()->eraseFromParent(); 1463 1464 // Create i+1 and fill the PHINode. 1465 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); 1466 Induction->addIncoming(StartIdx, VectorPH); 1467 Induction->addIncoming(NextIdx, VecBody); 1468 // Create the compare. 1469 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); 1470 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); 1471 1472 // Now we have two terminators. Remove the old one from the block. 1473 VecBody->getTerminator()->eraseFromParent(); 1474 1475 // Get ready to start creating new instructions into the vectorized body. 1476 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1477 1478 // Create and register the new vector loop. 1479 Loop* Lp = new Loop(); 1480 Loop *ParentLoop = OrigLoop->getParentLoop(); 1481 1482 // Insert the new loop into the loop nest and register the new basic blocks. 1483 if (ParentLoop) { 1484 ParentLoop->addChildLoop(Lp); 1485 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 1486 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase()); 1487 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); 1488 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); 1489 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); 1490 } else { 1491 LI->addTopLevelLoop(Lp); 1492 } 1493 1494 Lp->addBasicBlockToLoop(VecBody, LI->getBase()); 1495 1496 // Save the state. 1497 LoopVectorPreHeader = VectorPH; 1498 LoopScalarPreHeader = ScalarPH; 1499 LoopMiddleBlock = MiddleBlock; 1500 LoopExitBlock = ExitBlock; 1501 LoopVectorBody = VecBody; 1502 LoopScalarBody = OldBasicBlock; 1503} 1504 1505/// This function returns the identity element (or neutral element) for 1506/// the operation K. 1507Constant* 1508LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) { 1509 switch (K) { 1510 case RK_IntegerXor: 1511 case RK_IntegerAdd: 1512 case RK_IntegerOr: 1513 // Adding, Xoring, Oring zero to a number does not change it. 1514 return ConstantInt::get(Tp, 0); 1515 case RK_IntegerMult: 1516 // Multiplying a number by 1 does not change it. 1517 return ConstantInt::get(Tp, 1); 1518 case RK_IntegerAnd: 1519 // AND-ing a number with an all-1 value does not change it. 1520 return ConstantInt::get(Tp, -1, true); 1521 case RK_FloatMult: 1522 // Multiplying a number by 1 does not change it. 1523 return ConstantFP::get(Tp, 1.0L); 1524 case RK_FloatAdd: 1525 // Adding zero to a number does not change it. 1526 return ConstantFP::get(Tp, 0.0L); 1527 default: 1528 llvm_unreachable("Unknown reduction kind"); 1529 } 1530} 1531 1532static Intrinsic::ID 1533getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) { 1534 // If we have an intrinsic call, check if it is trivially vectorizable. 1535 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) { 1536 switch (II->getIntrinsicID()) { 1537 case Intrinsic::sqrt: 1538 case Intrinsic::sin: 1539 case Intrinsic::cos: 1540 case Intrinsic::exp: 1541 case Intrinsic::exp2: 1542 case Intrinsic::log: 1543 case Intrinsic::log10: 1544 case Intrinsic::log2: 1545 case Intrinsic::fabs: 1546 case Intrinsic::floor: 1547 case Intrinsic::ceil: 1548 case Intrinsic::trunc: 1549 case Intrinsic::rint: 1550 case Intrinsic::nearbyint: 1551 case Intrinsic::pow: 1552 case Intrinsic::fma: 1553 case Intrinsic::fmuladd: 1554 return II->getIntrinsicID(); 1555 default: 1556 return Intrinsic::not_intrinsic; 1557 } 1558 } 1559 1560 if (!TLI) 1561 return Intrinsic::not_intrinsic; 1562 1563 LibFunc::Func Func; 1564 Function *F = CI->getCalledFunction(); 1565 // We're going to make assumptions on the semantics of the functions, check 1566 // that the target knows that it's available in this environment. 1567 if (!F || !TLI->getLibFunc(F->getName(), Func)) 1568 return Intrinsic::not_intrinsic; 1569 1570 // Otherwise check if we have a call to a function that can be turned into a 1571 // vector intrinsic. 1572 switch (Func) { 1573 default: 1574 break; 1575 case LibFunc::sin: 1576 case LibFunc::sinf: 1577 case LibFunc::sinl: 1578 return Intrinsic::sin; 1579 case LibFunc::cos: 1580 case LibFunc::cosf: 1581 case LibFunc::cosl: 1582 return Intrinsic::cos; 1583 case LibFunc::exp: 1584 case LibFunc::expf: 1585 case LibFunc::expl: 1586 return Intrinsic::exp; 1587 case LibFunc::exp2: 1588 case LibFunc::exp2f: 1589 case LibFunc::exp2l: 1590 return Intrinsic::exp2; 1591 case LibFunc::log: 1592 case LibFunc::logf: 1593 case LibFunc::logl: 1594 return Intrinsic::log; 1595 case LibFunc::log10: 1596 case LibFunc::log10f: 1597 case LibFunc::log10l: 1598 return Intrinsic::log10; 1599 case LibFunc::log2: 1600 case LibFunc::log2f: 1601 case LibFunc::log2l: 1602 return Intrinsic::log2; 1603 case LibFunc::fabs: 1604 case LibFunc::fabsf: 1605 case LibFunc::fabsl: 1606 return Intrinsic::fabs; 1607 case LibFunc::floor: 1608 case LibFunc::floorf: 1609 case LibFunc::floorl: 1610 return Intrinsic::floor; 1611 case LibFunc::ceil: 1612 case LibFunc::ceilf: 1613 case LibFunc::ceill: 1614 return Intrinsic::ceil; 1615 case LibFunc::trunc: 1616 case LibFunc::truncf: 1617 case LibFunc::truncl: 1618 return Intrinsic::trunc; 1619 case LibFunc::rint: 1620 case LibFunc::rintf: 1621 case LibFunc::rintl: 1622 return Intrinsic::rint; 1623 case LibFunc::nearbyint: 1624 case LibFunc::nearbyintf: 1625 case LibFunc::nearbyintl: 1626 return Intrinsic::nearbyint; 1627 case LibFunc::pow: 1628 case LibFunc::powf: 1629 case LibFunc::powl: 1630 return Intrinsic::pow; 1631 } 1632 1633 return Intrinsic::not_intrinsic; 1634} 1635 1636/// This function translates the reduction kind to an LLVM binary operator. 1637static unsigned 1638getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) { 1639 switch (Kind) { 1640 case LoopVectorizationLegality::RK_IntegerAdd: 1641 return Instruction::Add; 1642 case LoopVectorizationLegality::RK_IntegerMult: 1643 return Instruction::Mul; 1644 case LoopVectorizationLegality::RK_IntegerOr: 1645 return Instruction::Or; 1646 case LoopVectorizationLegality::RK_IntegerAnd: 1647 return Instruction::And; 1648 case LoopVectorizationLegality::RK_IntegerXor: 1649 return Instruction::Xor; 1650 case LoopVectorizationLegality::RK_FloatMult: 1651 return Instruction::FMul; 1652 case LoopVectorizationLegality::RK_FloatAdd: 1653 return Instruction::FAdd; 1654 case LoopVectorizationLegality::RK_IntegerMinMax: 1655 return Instruction::ICmp; 1656 case LoopVectorizationLegality::RK_FloatMinMax: 1657 return Instruction::FCmp; 1658 default: 1659 llvm_unreachable("Unknown reduction operation"); 1660 } 1661} 1662 1663Value *createMinMaxOp(IRBuilder<> &Builder, 1664 LoopVectorizationLegality::MinMaxReductionKind RK, 1665 Value *Left, 1666 Value *Right) { 1667 CmpInst::Predicate P = CmpInst::ICMP_NE; 1668 switch (RK) { 1669 default: 1670 llvm_unreachable("Unknown min/max reduction kind"); 1671 case LoopVectorizationLegality::MRK_UIntMin: 1672 P = CmpInst::ICMP_ULT; 1673 break; 1674 case LoopVectorizationLegality::MRK_UIntMax: 1675 P = CmpInst::ICMP_UGT; 1676 break; 1677 case LoopVectorizationLegality::MRK_SIntMin: 1678 P = CmpInst::ICMP_SLT; 1679 break; 1680 case LoopVectorizationLegality::MRK_SIntMax: 1681 P = CmpInst::ICMP_SGT; 1682 break; 1683 case LoopVectorizationLegality::MRK_FloatMin: 1684 P = CmpInst::FCMP_OLT; 1685 break; 1686 case LoopVectorizationLegality::MRK_FloatMax: 1687 P = CmpInst::FCMP_OGT; 1688 break; 1689 } 1690 1691 Value *Cmp; 1692 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax) 1693 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp"); 1694 else 1695 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); 1696 1697 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); 1698 return Select; 1699} 1700 1701void 1702InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { 1703 //===------------------------------------------------===// 1704 // 1705 // Notice: any optimization or new instruction that go 1706 // into the code below should be also be implemented in 1707 // the cost-model. 1708 // 1709 //===------------------------------------------------===// 1710 Constant *Zero = Builder.getInt32(0); 1711 1712 // In order to support reduction variables we need to be able to vectorize 1713 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two 1714 // stages. First, we create a new vector PHI node with no incoming edges. 1715 // We use this value when we vectorize all of the instructions that use the 1716 // PHI. Next, after all of the instructions in the block are complete we 1717 // add the new incoming edges to the PHI. At this point all of the 1718 // instructions in the basic block are vectorized, so we can use them to 1719 // construct the PHI. 1720 PhiVector RdxPHIsToFix; 1721 1722 // Scan the loop in a topological order to ensure that defs are vectorized 1723 // before users. 1724 LoopBlocksDFS DFS(OrigLoop); 1725 DFS.perform(LI); 1726 1727 // Vectorize all of the blocks in the original loop. 1728 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 1729 be = DFS.endRPO(); bb != be; ++bb) 1730 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix); 1731 1732 // At this point every instruction in the original loop is widened to 1733 // a vector form. We are almost done. Now, we need to fix the PHI nodes 1734 // that we vectorized. The PHI nodes are currently empty because we did 1735 // not want to introduce cycles. Notice that the remaining PHI nodes 1736 // that we need to fix are reduction variables. 1737 1738 // Create the 'reduced' values for each of the induction vars. 1739 // The reduced values are the vector values that we scalarize and combine 1740 // after the loop is finished. 1741 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); 1742 it != e; ++it) { 1743 PHINode *RdxPhi = *it; 1744 assert(RdxPhi && "Unable to recover vectorized PHI"); 1745 1746 // Find the reduction variable descriptor. 1747 assert(Legal->getReductionVars()->count(RdxPhi) && 1748 "Unable to find the reduction variable"); 1749 LoopVectorizationLegality::ReductionDescriptor RdxDesc = 1750 (*Legal->getReductionVars())[RdxPhi]; 1751 1752 // We need to generate a reduction vector from the incoming scalar. 1753 // To do so, we need to generate the 'identity' vector and overide 1754 // one of the elements with the incoming scalar reduction. We need 1755 // to do it in the vector-loop preheader. 1756 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator()); 1757 1758 // This is the vector-clone of the value that leaves the loop. 1759 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr); 1760 Type *VecTy = VectorExit[0]->getType(); 1761 1762 // Find the reduction identity variable. Zero for addition, or, xor, 1763 // one for multiplication, -1 for And. 1764 Value *Identity; 1765 Value *VectorStart; 1766 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax || 1767 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) { 1768 // MinMax reduction have the start value as their identify. 1769 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue, 1770 "minmax.ident"); 1771 } else { 1772 Constant *Iden = 1773 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind, 1774 VecTy->getScalarType()); 1775 Identity = ConstantVector::getSplat(VF, Iden); 1776 1777 // This vector is the Identity vector where the first element is the 1778 // incoming scalar reduction. 1779 VectorStart = Builder.CreateInsertElement(Identity, 1780 RdxDesc.StartValue, Zero); 1781 } 1782 1783 // Fix the vector-loop phi. 1784 // We created the induction variable so we know that the 1785 // preheader is the first entry. 1786 BasicBlock *VecPreheader = Induction->getIncomingBlock(0); 1787 1788 // Reductions do not have to start at zero. They can start with 1789 // any loop invariant values. 1790 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 1791 BasicBlock *Latch = OrigLoop->getLoopLatch(); 1792 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 1793 VectorParts &Val = getVectorValue(LoopVal); 1794 for (unsigned part = 0; part < UF; ++part) { 1795 // Make sure to add the reduction stat value only to the 1796 // first unroll part. 1797 Value *StartVal = (part == 0) ? VectorStart : Identity; 1798 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader); 1799 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody); 1800 } 1801 1802 // Before each round, move the insertion point right between 1803 // the PHIs and the values we are going to write. 1804 // This allows us to write both PHINodes and the extractelement 1805 // instructions. 1806 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); 1807 1808 VectorParts RdxParts; 1809 for (unsigned part = 0; part < UF; ++part) { 1810 // This PHINode contains the vectorized reduction variable, or 1811 // the initial value vector, if we bypass the vector loop. 1812 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); 1813 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); 1814 Value *StartVal = (part == 0) ? VectorStart : Identity; 1815 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1816 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]); 1817 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody); 1818 RdxParts.push_back(NewPhi); 1819 } 1820 1821 // Reduce all of the unrolled parts into a single vector. 1822 Value *ReducedPartRdx = RdxParts[0]; 1823 unsigned Op = getReductionBinOp(RdxDesc.Kind); 1824 for (unsigned part = 1; part < UF; ++part) { 1825 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 1826 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op, 1827 RdxParts[part], ReducedPartRdx, 1828 "bin.rdx"); 1829 else 1830 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind, 1831 ReducedPartRdx, RdxParts[part]); 1832 } 1833 1834 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 1835 // and vector ops, reducing the set of values being computed by half each 1836 // round. 1837 assert(isPowerOf2_32(VF) && 1838 "Reduction emission only supported for pow2 vectors!"); 1839 Value *TmpVec = ReducedPartRdx; 1840 SmallVector<Constant*, 32> ShuffleMask(VF, 0); 1841 for (unsigned i = VF; i != 1; i >>= 1) { 1842 // Move the upper half of the vector to the lower half. 1843 for (unsigned j = 0; j != i/2; ++j) 1844 ShuffleMask[j] = Builder.getInt32(i/2 + j); 1845 1846 // Fill the rest of the mask with undef. 1847 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 1848 UndefValue::get(Builder.getInt32Ty())); 1849 1850 Value *Shuf = 1851 Builder.CreateShuffleVector(TmpVec, 1852 UndefValue::get(TmpVec->getType()), 1853 ConstantVector::get(ShuffleMask), 1854 "rdx.shuf"); 1855 1856 if (Op != Instruction::ICmp && Op != Instruction::FCmp) 1857 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf, 1858 "bin.rdx"); 1859 else 1860 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf); 1861 } 1862 1863 // The result is in the first element of the vector. 1864 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); 1865 1866 // Now, we need to fix the users of the reduction variable 1867 // inside and outside of the scalar remainder loop. 1868 // We know that the loop is in LCSSA form. We need to update the 1869 // PHI nodes in the exit blocks. 1870 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 1871 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 1872 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 1873 if (!LCSSAPhi) continue; 1874 1875 // All PHINodes need to have a single entry edge, or two if 1876 // we already fixed them. 1877 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 1878 1879 // We found our reduction value exit-PHI. Update it with the 1880 // incoming bypass edge. 1881 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { 1882 // Add an edge coming from the bypass. 1883 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock); 1884 break; 1885 } 1886 }// end of the LCSSA phi scan. 1887 1888 // Fix the scalar loop reduction variable with the incoming reduction sum 1889 // from the vector body and from the backedge value. 1890 int IncomingEdgeBlockIdx = 1891 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 1892 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 1893 // Pick the other block. 1894 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 1895 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0); 1896 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); 1897 }// end of for each redux variable. 1898 1899 // The Loop exit block may have single value PHI nodes where the incoming 1900 // value is 'undef'. While vectorizing we only handled real values that 1901 // were defined inside the loop. Here we handle the 'undef case'. 1902 // See PR14725. 1903 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 1904 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 1905 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 1906 if (!LCSSAPhi) continue; 1907 if (LCSSAPhi->getNumIncomingValues() == 1) 1908 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 1909 LoopMiddleBlock); 1910 } 1911} 1912 1913InnerLoopVectorizer::VectorParts 1914InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 1915 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 1916 "Invalid edge"); 1917 1918 VectorParts SrcMask = createBlockInMask(Src); 1919 1920 // The terminator has to be a branch inst! 1921 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 1922 assert(BI && "Unexpected terminator found"); 1923 1924 if (BI->isConditional()) { 1925 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 1926 1927 if (BI->getSuccessor(0) != Dst) 1928 for (unsigned part = 0; part < UF; ++part) 1929 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 1930 1931 for (unsigned part = 0; part < UF; ++part) 1932 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 1933 return EdgeMask; 1934 } 1935 1936 return SrcMask; 1937} 1938 1939InnerLoopVectorizer::VectorParts 1940InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 1941 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 1942 1943 // Loop incoming mask is all-one. 1944 if (OrigLoop->getHeader() == BB) { 1945 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 1946 return getVectorValue(C); 1947 } 1948 1949 // This is the block mask. We OR all incoming edges, and with zero. 1950 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 1951 VectorParts BlockMask = getVectorValue(Zero); 1952 1953 // For each pred: 1954 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 1955 VectorParts EM = createEdgeMask(*it, BB); 1956 for (unsigned part = 0; part < UF; ++part) 1957 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 1958 } 1959 1960 return BlockMask; 1961} 1962 1963void 1964InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, 1965 BasicBlock *BB, PhiVector *PV) { 1966 // For each instruction in the old loop. 1967 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 1968 VectorParts &Entry = WidenMap.get(it); 1969 switch (it->getOpcode()) { 1970 case Instruction::Br: 1971 // Nothing to do for PHIs and BR, since we already took care of the 1972 // loop control flow instructions. 1973 continue; 1974 case Instruction::PHI:{ 1975 PHINode* P = cast<PHINode>(it); 1976 // Handle reduction variables: 1977 if (Legal->getReductionVars()->count(P)) { 1978 for (unsigned part = 0; part < UF; ++part) { 1979 // This is phase one of vectorizing PHIs. 1980 Type *VecTy = VectorType::get(it->getType(), VF); 1981 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", 1982 LoopVectorBody-> getFirstInsertionPt()); 1983 } 1984 PV->push_back(P); 1985 continue; 1986 } 1987 1988 // Check for PHI nodes that are lowered to vector selects. 1989 if (P->getParent() != OrigLoop->getHeader()) { 1990 // We know that all PHIs in non header blocks are converted into 1991 // selects, so we don't have to worry about the insertion order and we 1992 // can just use the builder. 1993 // At this point we generate the predication tree. There may be 1994 // duplications since this is a simple recursive scan, but future 1995 // optimizations will clean it up. 1996 1997 unsigned NumIncoming = P->getNumIncomingValues(); 1998 assert(NumIncoming > 1 && "Invalid PHI"); 1999 2000 // Generate a sequence of selects of the form: 2001 // SELECT(Mask3, In3, 2002 // SELECT(Mask2, In2, 2003 // ( ...))) 2004 for (unsigned In = 0; In < NumIncoming; In++) { 2005 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), 2006 P->getParent()); 2007 VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); 2008 2009 for (unsigned part = 0; part < UF; ++part) { 2010 // We don't need to 'select' the first PHI operand because it is 2011 // the default value if all of the other masks don't match. 2012 if (In == 0) 2013 Entry[part] = In0[part]; 2014 else 2015 // Select between the current value and the previous incoming edge 2016 // based on the incoming mask. 2017 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2018 Entry[part], "predphi"); 2019 } 2020 } 2021 continue; 2022 } 2023 2024 // This PHINode must be an induction variable. 2025 // Make sure that we know about it. 2026 assert(Legal->getInductionVars()->count(P) && 2027 "Not an induction variable"); 2028 2029 LoopVectorizationLegality::InductionInfo II = 2030 Legal->getInductionVars()->lookup(P); 2031 2032 switch (II.IK) { 2033 case LoopVectorizationLegality::IK_NoInduction: 2034 llvm_unreachable("Unknown induction"); 2035 case LoopVectorizationLegality::IK_IntInduction: { 2036 assert(P == OldInduction && "Unexpected PHI"); 2037 Value *Broadcasted = getBroadcastInstrs(Induction); 2038 // After broadcasting the induction variable we need to make the 2039 // vector consecutive by adding 0, 1, 2 ... 2040 for (unsigned part = 0; part < UF; ++part) 2041 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); 2042 continue; 2043 } 2044 case LoopVectorizationLegality::IK_ReverseIntInduction: 2045 case LoopVectorizationLegality::IK_PtrInduction: 2046 case LoopVectorizationLegality::IK_ReversePtrInduction: 2047 // Handle reverse integer and pointer inductions. 2048 Value *StartIdx = 0; 2049 // If we have a single integer induction variable then use it. 2050 // Otherwise, start counting at zero. 2051 if (OldInduction) { 2052 LoopVectorizationLegality::InductionInfo OldII = 2053 Legal->getInductionVars()->lookup(OldInduction); 2054 StartIdx = OldII.StartValue; 2055 } else { 2056 StartIdx = ConstantInt::get(Induction->getType(), 0); 2057 } 2058 // This is the normalized GEP that starts counting at zero. 2059 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx, 2060 "normalized.idx"); 2061 2062 // Handle the reverse integer induction variable case. 2063 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) { 2064 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType()); 2065 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy, 2066 "resize.norm.idx"); 2067 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI, 2068 "reverse.idx"); 2069 2070 // This is a new value so do not hoist it out. 2071 Value *Broadcasted = getBroadcastInstrs(ReverseInd); 2072 // After broadcasting the induction variable we need to make the 2073 // vector consecutive by adding ... -3, -2, -1, 0. 2074 for (unsigned part = 0; part < UF; ++part) 2075 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true); 2076 continue; 2077 } 2078 2079 // Handle the pointer induction variable case. 2080 assert(P->getType()->isPointerTy() && "Unexpected type."); 2081 2082 // Is this a reverse induction ptr or a consecutive induction ptr. 2083 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction == 2084 II.IK); 2085 2086 // This is the vector of results. Notice that we don't generate 2087 // vector geps because scalar geps result in better code. 2088 for (unsigned part = 0; part < UF; ++part) { 2089 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 2090 for (unsigned int i = 0; i < VF; ++i) { 2091 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1); 2092 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 2093 Value *GlobalIdx; 2094 if (!Reverse) 2095 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); 2096 else 2097 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); 2098 2099 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 2100 "next.gep"); 2101 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 2102 Builder.getInt32(i), 2103 "insert.gep"); 2104 } 2105 Entry[part] = VecVal; 2106 } 2107 continue; 2108 } 2109 2110 }// End of PHI. 2111 2112 case Instruction::Add: 2113 case Instruction::FAdd: 2114 case Instruction::Sub: 2115 case Instruction::FSub: 2116 case Instruction::Mul: 2117 case Instruction::FMul: 2118 case Instruction::UDiv: 2119 case Instruction::SDiv: 2120 case Instruction::FDiv: 2121 case Instruction::URem: 2122 case Instruction::SRem: 2123 case Instruction::FRem: 2124 case Instruction::Shl: 2125 case Instruction::LShr: 2126 case Instruction::AShr: 2127 case Instruction::And: 2128 case Instruction::Or: 2129 case Instruction::Xor: { 2130 // Just widen binops. 2131 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 2132 VectorParts &A = getVectorValue(it->getOperand(0)); 2133 VectorParts &B = getVectorValue(it->getOperand(1)); 2134 2135 // Use this vector value for all users of the original instruction. 2136 for (unsigned Part = 0; Part < UF; ++Part) { 2137 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 2138 2139 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef. 2140 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V); 2141 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) { 2142 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); 2143 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); 2144 } 2145 if (VecOp && isa<PossiblyExactOperator>(VecOp)) 2146 VecOp->setIsExact(BinOp->isExact()); 2147 2148 Entry[Part] = V; 2149 } 2150 break; 2151 } 2152 case Instruction::Select: { 2153 // Widen selects. 2154 // If the selector is loop invariant we can create a select 2155 // instruction with a scalar condition. Otherwise, use vector-select. 2156 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), 2157 OrigLoop); 2158 2159 // The condition can be loop invariant but still defined inside the 2160 // loop. This means that we can't just use the original 'cond' value. 2161 // We have to take the 'vectorized' value and pick the first lane. 2162 // Instcombine will make this a no-op. 2163 VectorParts &Cond = getVectorValue(it->getOperand(0)); 2164 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 2165 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 2166 Value *ScalarCond = Builder.CreateExtractElement(Cond[0], 2167 Builder.getInt32(0)); 2168 for (unsigned Part = 0; Part < UF; ++Part) { 2169 Entry[Part] = Builder.CreateSelect( 2170 InvariantCond ? ScalarCond : Cond[Part], 2171 Op0[Part], 2172 Op1[Part]); 2173 } 2174 break; 2175 } 2176 2177 case Instruction::ICmp: 2178 case Instruction::FCmp: { 2179 // Widen compares. Generate vector compares. 2180 bool FCmp = (it->getOpcode() == Instruction::FCmp); 2181 CmpInst *Cmp = dyn_cast<CmpInst>(it); 2182 VectorParts &A = getVectorValue(it->getOperand(0)); 2183 VectorParts &B = getVectorValue(it->getOperand(1)); 2184 for (unsigned Part = 0; Part < UF; ++Part) { 2185 Value *C = 0; 2186 if (FCmp) 2187 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 2188 else 2189 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 2190 Entry[Part] = C; 2191 } 2192 break; 2193 } 2194 2195 case Instruction::Store: 2196 case Instruction::Load: 2197 vectorizeMemoryInstruction(it, Legal); 2198 break; 2199 case Instruction::ZExt: 2200 case Instruction::SExt: 2201 case Instruction::FPToUI: 2202 case Instruction::FPToSI: 2203 case Instruction::FPExt: 2204 case Instruction::PtrToInt: 2205 case Instruction::IntToPtr: 2206 case Instruction::SIToFP: 2207 case Instruction::UIToFP: 2208 case Instruction::Trunc: 2209 case Instruction::FPTrunc: 2210 case Instruction::BitCast: { 2211 CastInst *CI = dyn_cast<CastInst>(it); 2212 /// Optimize the special case where the source is the induction 2213 /// variable. Notice that we can only optimize the 'trunc' case 2214 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 2215 /// c. other casts depend on pointer size. 2216 if (CI->getOperand(0) == OldInduction && 2217 it->getOpcode() == Instruction::Trunc) { 2218 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 2219 CI->getType()); 2220 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 2221 for (unsigned Part = 0; Part < UF; ++Part) 2222 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false); 2223 break; 2224 } 2225 /// Vectorize casts. 2226 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF); 2227 2228 VectorParts &A = getVectorValue(it->getOperand(0)); 2229 for (unsigned Part = 0; Part < UF; ++Part) 2230 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 2231 break; 2232 } 2233 2234 case Instruction::Call: { 2235 // Ignore dbg intrinsics. 2236 if (isa<DbgInfoIntrinsic>(it)) 2237 break; 2238 2239 Module *M = BB->getParent()->getParent(); 2240 CallInst *CI = cast<CallInst>(it); 2241 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 2242 assert(ID && "Not an intrinsic call!"); 2243 for (unsigned Part = 0; Part < UF; ++Part) { 2244 SmallVector<Value*, 4> Args; 2245 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 2246 VectorParts &Arg = getVectorValue(CI->getArgOperand(i)); 2247 Args.push_back(Arg[Part]); 2248 } 2249 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) }; 2250 Function *F = Intrinsic::getDeclaration(M, ID, Tys); 2251 Entry[Part] = Builder.CreateCall(F, Args); 2252 } 2253 break; 2254 } 2255 2256 default: 2257 // All other instructions are unsupported. Scalarize them. 2258 scalarizeInstruction(it); 2259 break; 2260 }// end of switch. 2261 }// end of for_each instr. 2262} 2263 2264void InnerLoopVectorizer::updateAnalysis() { 2265 // Forget the original basic block. 2266 SE->forgetLoop(OrigLoop); 2267 2268 // Update the dominator tree information. 2269 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && 2270 "Entry does not dominate exit."); 2271 2272 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2273 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); 2274 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); 2275 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); 2276 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front()); 2277 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); 2278 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 2279 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); 2280 2281 DEBUG(DT->verifyAnalysis()); 2282} 2283 2284bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 2285 if (!EnableIfConversion) 2286 return false; 2287 2288 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 2289 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector(); 2290 2291 // Collect the blocks that need predication. 2292 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) { 2293 BasicBlock *BB = LoopBlocks[i]; 2294 2295 // We don't support switch statements inside loops. 2296 if (!isa<BranchInst>(BB->getTerminator())) 2297 return false; 2298 2299 // We must be able to predicate all blocks that need to be predicated. 2300 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB)) 2301 return false; 2302 } 2303 2304 // We can if-convert this loop. 2305 return true; 2306} 2307 2308bool LoopVectorizationLegality::canVectorize() { 2309 assert(TheLoop->getLoopPreheader() && "No preheader!!"); 2310 2311 // We can only vectorize innermost loops. 2312 if (TheLoop->getSubLoopsVector().size()) 2313 return false; 2314 2315 // We must have a single backedge. 2316 if (TheLoop->getNumBackEdges() != 1) 2317 return false; 2318 2319 // We must have a single exiting block. 2320 if (!TheLoop->getExitingBlock()) 2321 return false; 2322 2323 unsigned NumBlocks = TheLoop->getNumBlocks(); 2324 2325 // Check if we can if-convert non single-bb loops. 2326 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 2327 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 2328 return false; 2329 } 2330 2331 // We need to have a loop header. 2332 BasicBlock *Latch = TheLoop->getLoopLatch(); 2333 DEBUG(dbgs() << "LV: Found a loop: " << 2334 TheLoop->getHeader()->getName() << "\n"); 2335 2336 // ScalarEvolution needs to be able to find the exit count. 2337 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch); 2338 if (ExitCount == SE->getCouldNotCompute()) { 2339 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 2340 return false; 2341 } 2342 2343 // Do not loop-vectorize loops with a tiny trip count. 2344 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch); 2345 if (TC > 0u && TC < TinyTripCountVectorThreshold) { 2346 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << 2347 "This loop is not worth vectorizing.\n"); 2348 return false; 2349 } 2350 2351 // Check if we can vectorize the instructions and CFG in this loop. 2352 if (!canVectorizeInstrs()) { 2353 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 2354 return false; 2355 } 2356 2357 // Go over each instruction and look at memory deps. 2358 if (!canVectorizeMemory()) { 2359 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 2360 return false; 2361 } 2362 2363 // Collect all of the variables that remain uniform after vectorization. 2364 collectLoopUniforms(); 2365 2366 DEBUG(dbgs() << "LV: We can vectorize this loop" << 2367 (PtrRtCheck.Need ? " (with a runtime bound check)" : "") 2368 <<"!\n"); 2369 2370 // Okay! We can vectorize. At this point we don't have any other mem analysis 2371 // which may limit our maximum vectorization factor, so just return true with 2372 // no restrictions. 2373 return true; 2374} 2375 2376bool LoopVectorizationLegality::canVectorizeInstrs() { 2377 BasicBlock *PreHeader = TheLoop->getLoopPreheader(); 2378 BasicBlock *Header = TheLoop->getHeader(); 2379 2380 // If we marked the scalar loop as "already vectorized" then no need 2381 // to vectorize it again. 2382 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) { 2383 DEBUG(dbgs() << "LV: This loop was vectorized before\n"); 2384 return false; 2385 } 2386 2387 // Look for the attribute signaling the absence of NaNs. 2388 Function &F = *Header->getParent(); 2389 if (F.hasFnAttribute("no-nans-fp-math")) 2390 HasFunNoNaNAttr = F.getAttributes().getAttribute( 2391 AttributeSet::FunctionIndex, 2392 "no-nans-fp-math").getValueAsString() == "true"; 2393 2394 // For each block in the loop. 2395 for (Loop::block_iterator bb = TheLoop->block_begin(), 2396 be = TheLoop->block_end(); bb != be; ++bb) { 2397 2398 // Scan the instructions in the block and look for hazards. 2399 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2400 ++it) { 2401 2402 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 2403 // Check that this PHI type is allowed. 2404 if (!Phi->getType()->isIntegerTy() && 2405 !Phi->getType()->isFloatingPointTy() && 2406 !Phi->getType()->isPointerTy()) { 2407 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 2408 return false; 2409 } 2410 2411 // If this PHINode is not in the header block, then we know that we 2412 // can convert it to select during if-conversion. No need to check if 2413 // the PHIs in this block are induction or reduction variables. 2414 if (*bb != Header) 2415 continue; 2416 2417 // We only allow if-converted PHIs with more than two incoming values. 2418 if (Phi->getNumIncomingValues() != 2) { 2419 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 2420 return false; 2421 } 2422 2423 // This is the value coming from the preheader. 2424 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); 2425 // Check if this is an induction variable. 2426 InductionKind IK = isInductionVariable(Phi); 2427 2428 if (IK_NoInduction != IK) { 2429 // Int inductions are special because we only allow one IV. 2430 if (IK == IK_IntInduction) { 2431 if (Induction) { 2432 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n"); 2433 return false; 2434 } 2435 Induction = Phi; 2436 } 2437 2438 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 2439 Inductions[Phi] = InductionInfo(StartValue, IK); 2440 continue; 2441 } 2442 2443 if (AddReductionVar(Phi, RK_IntegerAdd)) { 2444 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); 2445 continue; 2446 } 2447 if (AddReductionVar(Phi, RK_IntegerMult)) { 2448 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); 2449 continue; 2450 } 2451 if (AddReductionVar(Phi, RK_IntegerOr)) { 2452 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); 2453 continue; 2454 } 2455 if (AddReductionVar(Phi, RK_IntegerAnd)) { 2456 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); 2457 continue; 2458 } 2459 if (AddReductionVar(Phi, RK_IntegerXor)) { 2460 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); 2461 continue; 2462 } 2463 if (AddReductionVar(Phi, RK_IntegerMinMax)) { 2464 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n"); 2465 continue; 2466 } 2467 if (AddReductionVar(Phi, RK_FloatMult)) { 2468 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n"); 2469 continue; 2470 } 2471 if (AddReductionVar(Phi, RK_FloatAdd)) { 2472 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n"); 2473 continue; 2474 } 2475 if (AddReductionVar(Phi, RK_FloatMinMax)) { 2476 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n"); 2477 continue; 2478 } 2479 2480 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 2481 return false; 2482 }// end of PHI handling 2483 2484 // We still don't handle functions. However, we can ignore dbg intrinsic 2485 // calls and we do handle certain intrinsic and libm functions. 2486 CallInst *CI = dyn_cast<CallInst>(it); 2487 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) { 2488 DEBUG(dbgs() << "LV: Found a call site.\n"); 2489 return false; 2490 } 2491 2492 // Check that the instruction return type is vectorizable. 2493 if (!VectorType::isValidElementType(it->getType()) && 2494 !it->getType()->isVoidTy()) { 2495 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n"); 2496 return false; 2497 } 2498 2499 // Check that the stored type is vectorizable. 2500 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 2501 Type *T = ST->getValueOperand()->getType(); 2502 if (!VectorType::isValidElementType(T)) 2503 return false; 2504 } 2505 2506 // Reduction instructions are allowed to have exit users. 2507 // All other instructions must not have external users. 2508 if (!AllowedExit.count(it)) 2509 //Check that all of the users of the loop are inside the BB. 2510 for (Value::use_iterator I = it->use_begin(), E = it->use_end(); 2511 I != E; ++I) { 2512 Instruction *U = cast<Instruction>(*I); 2513 // This user may be a reduction exit value. 2514 if (!TheLoop->contains(U)) { 2515 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n"); 2516 return false; 2517 } 2518 } 2519 } // next instr. 2520 2521 } 2522 2523 if (!Induction) { 2524 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 2525 assert(getInductionVars()->size() && "No induction variables"); 2526 } 2527 2528 return true; 2529} 2530 2531void LoopVectorizationLegality::collectLoopUniforms() { 2532 // We now know that the loop is vectorizable! 2533 // Collect variables that will remain uniform after vectorization. 2534 std::vector<Value*> Worklist; 2535 BasicBlock *Latch = TheLoop->getLoopLatch(); 2536 2537 // Start with the conditional branch and walk up the block. 2538 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 2539 2540 while (Worklist.size()) { 2541 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 2542 Worklist.pop_back(); 2543 2544 // Look at instructions inside this loop. 2545 // Stop when reaching PHI nodes. 2546 // TODO: we need to follow values all over the loop, not only in this block. 2547 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 2548 continue; 2549 2550 // This is a known uniform. 2551 Uniforms.insert(I); 2552 2553 // Insert all operands. 2554 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) { 2555 Worklist.push_back(I->getOperand(i)); 2556 } 2557 } 2558} 2559 2560AliasAnalysis::Location 2561LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) { 2562 if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) 2563 return AA->getLocation(Store); 2564 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) 2565 return AA->getLocation(Load); 2566 2567 llvm_unreachable("Should be either load or store instruction"); 2568} 2569 2570bool 2571LoopVectorizationLegality::hasPossibleGlobalWriteReorder( 2572 Value *Object, 2573 Instruction *Inst, 2574 AliasMultiMap& WriteObjects, 2575 unsigned MaxByteWidth) { 2576 2577 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst); 2578 2579 std::vector<Instruction*>::iterator 2580 it = WriteObjects[Object].begin(), 2581 end = WriteObjects[Object].end(); 2582 2583 for (; it != end; ++it) { 2584 Instruction* I = *it; 2585 if (I == Inst) 2586 continue; 2587 2588 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I); 2589 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth), 2590 ThatLoc.getWithNewSize(MaxByteWidth))) 2591 return true; 2592 } 2593 return false; 2594} 2595 2596bool LoopVectorizationLegality::canVectorizeMemory() { 2597 2598 typedef SmallVector<Value*, 16> ValueVector; 2599 typedef SmallPtrSet<Value*, 16> ValueSet; 2600 // Holds the Load and Store *instructions*. 2601 ValueVector Loads; 2602 ValueVector Stores; 2603 PtrRtCheck.Pointers.clear(); 2604 PtrRtCheck.Need = false; 2605 2606 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); 2607 2608 // For each block. 2609 for (Loop::block_iterator bb = TheLoop->block_begin(), 2610 be = TheLoop->block_end(); bb != be; ++bb) { 2611 2612 // Scan the BB and collect legal loads and stores. 2613 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2614 ++it) { 2615 2616 // If this is a load, save it. If this instruction can read from memory 2617 // but is not a load, then we quit. Notice that we don't handle function 2618 // calls that read or write. 2619 if (it->mayReadFromMemory()) { 2620 LoadInst *Ld = dyn_cast<LoadInst>(it); 2621 if (!Ld) return false; 2622 if (!Ld->isSimple() && !IsAnnotatedParallel) { 2623 DEBUG(dbgs() << "LV: Found a non-simple load.\n"); 2624 return false; 2625 } 2626 Loads.push_back(Ld); 2627 continue; 2628 } 2629 2630 // Save 'store' instructions. Abort if other instructions write to memory. 2631 if (it->mayWriteToMemory()) { 2632 StoreInst *St = dyn_cast<StoreInst>(it); 2633 if (!St) return false; 2634 if (!St->isSimple() && !IsAnnotatedParallel) { 2635 DEBUG(dbgs() << "LV: Found a non-simple store.\n"); 2636 return false; 2637 } 2638 Stores.push_back(St); 2639 } 2640 } // next instr. 2641 } // next block. 2642 2643 // Now we have two lists that hold the loads and the stores. 2644 // Next, we find the pointers that they use. 2645 2646 // Check if we see any stores. If there are no stores, then we don't 2647 // care if the pointers are *restrict*. 2648 if (!Stores.size()) { 2649 DEBUG(dbgs() << "LV: Found a read-only loop!\n"); 2650 return true; 2651 } 2652 2653 // Holds the read and read-write *pointers* that we find. These maps hold 2654 // unique values for pointers (so no need for multi-map). 2655 AliasMap Reads; 2656 AliasMap ReadWrites; 2657 2658 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects 2659 // multiple times on the same object. If the ptr is accessed twice, once 2660 // for read and once for write, it will only appear once (on the write 2661 // list). This is okay, since we are going to check for conflicts between 2662 // writes and between reads and writes, but not between reads and reads. 2663 ValueSet Seen; 2664 2665 ValueVector::iterator I, IE; 2666 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) { 2667 StoreInst *ST = cast<StoreInst>(*I); 2668 Value* Ptr = ST->getPointerOperand(); 2669 2670 if (isUniform(Ptr)) { 2671 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 2672 return false; 2673 } 2674 2675 // If we did *not* see this pointer before, insert it to 2676 // the read-write list. At this phase it is only a 'write' list. 2677 if (Seen.insert(Ptr)) 2678 ReadWrites.insert(std::make_pair(Ptr, ST)); 2679 } 2680 2681 if (IsAnnotatedParallel) { 2682 DEBUG(dbgs() 2683 << "LV: A loop annotated parallel, ignore memory dependency " 2684 << "checks.\n"); 2685 return true; 2686 } 2687 2688 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { 2689 LoadInst *LD = cast<LoadInst>(*I); 2690 Value* Ptr = LD->getPointerOperand(); 2691 // If we did *not* see this pointer before, insert it to the 2692 // read list. If we *did* see it before, then it is already in 2693 // the read-write list. This allows us to vectorize expressions 2694 // such as A[i] += x; Because the address of A[i] is a read-write 2695 // pointer. This only works if the index of A[i] is consecutive. 2696 // If the address of i is unknown (for example A[B[i]]) then we may 2697 // read a few words, modify, and write a few words, and some of the 2698 // words may be written to the same address. 2699 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr)) 2700 Reads.insert(std::make_pair(Ptr, LD)); 2701 } 2702 2703 // If we write (or read-write) to a single destination and there are no 2704 // other reads in this loop then is it safe to vectorize. 2705 if (ReadWrites.size() == 1 && Reads.size() == 0) { 2706 DEBUG(dbgs() << "LV: Found a write-only loop!\n"); 2707 return true; 2708 } 2709 2710 unsigned NumReadPtrs = 0; 2711 unsigned NumWritePtrs = 0; 2712 2713 // Find pointers with computable bounds. We are going to use this information 2714 // to place a runtime bound check. 2715 bool CanDoRT = true; 2716 AliasMap::iterator MI, ME; 2717 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { 2718 Value *V = (*MI).first; 2719 if (hasComputableBounds(V)) { 2720 PtrRtCheck.insert(SE, TheLoop, V, true); 2721 NumWritePtrs++; 2722 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); 2723 } else { 2724 CanDoRT = false; 2725 break; 2726 } 2727 } 2728 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { 2729 Value *V = (*MI).first; 2730 if (hasComputableBounds(V)) { 2731 PtrRtCheck.insert(SE, TheLoop, V, false); 2732 NumReadPtrs++; 2733 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); 2734 } else { 2735 CanDoRT = false; 2736 break; 2737 } 2738 } 2739 2740 // Check that we did not collect too many pointers or found a 2741 // unsizeable pointer. 2742 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1)); 2743 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n"); 2744 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { 2745 PtrRtCheck.reset(); 2746 CanDoRT = false; 2747 } 2748 2749 if (CanDoRT) { 2750 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n"); 2751 } 2752 2753 bool NeedRTCheck = false; 2754 2755 // Biggest vectorized access possible, vector width * unroll factor. 2756 // TODO: We're being very pessimistic here, find a way to know the 2757 // real access width before getting here. 2758 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) * 2759 TTI->getMaximumUnrollFactor(); 2760 // Now that the pointers are in two lists (Reads and ReadWrites), we 2761 // can check that there are no conflicts between each of the writes and 2762 // between the writes to the reads. 2763 // Note that WriteObjects duplicates the stores (indexed now by underlying 2764 // objects) to avoid pointing to elements inside ReadWrites. 2765 // TODO: Maybe create a new type where they can interact without duplication. 2766 AliasMultiMap WriteObjects; 2767 ValueVector TempObjects; 2768 2769 // Check that the read-writes do not conflict with other read-write 2770 // pointers. 2771 bool AllWritesIdentified = true; 2772 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { 2773 Value *Val = (*MI).first; 2774 Instruction *Inst = (*MI).second; 2775 2776 GetUnderlyingObjects(Val, TempObjects, DL); 2777 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); 2778 UI != UE; ++UI) { 2779 if (!isIdentifiedObject(*UI)) { 2780 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n"); 2781 NeedRTCheck = true; 2782 AllWritesIdentified = false; 2783 } 2784 2785 // Never seen it before, can't alias. 2786 if (WriteObjects[*UI].empty()) { 2787 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n"); 2788 WriteObjects[*UI].push_back(Inst); 2789 continue; 2790 } 2791 // Direct alias found. 2792 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) { 2793 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2794 << **UI <<"\n"); 2795 return false; 2796 } 2797 DEBUG(dbgs() << "LV: Found a conflicting global value:" 2798 << **UI <<"\n"); 2799 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n"); 2800 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); 2801 2802 // If global alias, make sure they do alias. 2803 if (hasPossibleGlobalWriteReorder(*UI, 2804 Inst, 2805 WriteObjects, 2806 MaxByteWidth)) { 2807 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI 2808 << "\n"); 2809 return false; 2810 } 2811 2812 // Didn't alias, insert into map for further reference. 2813 WriteObjects[*UI].push_back(Inst); 2814 } 2815 TempObjects.clear(); 2816 } 2817 2818 /// Check that the reads don't conflict with the read-writes. 2819 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { 2820 Value *Val = (*MI).first; 2821 GetUnderlyingObjects(Val, TempObjects, DL); 2822 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); 2823 UI != UE; ++UI) { 2824 // If all of the writes are identified then we don't care if the read 2825 // pointer is identified or not. 2826 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) { 2827 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n"); 2828 NeedRTCheck = true; 2829 } 2830 2831 // Never seen it before, can't alias. 2832 if (WriteObjects[*UI].empty()) 2833 continue; 2834 // Direct alias found. 2835 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) { 2836 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2837 << **UI <<"\n"); 2838 return false; 2839 } 2840 DEBUG(dbgs() << "LV: Found a global value: " 2841 << **UI <<"\n"); 2842 Instruction *Inst = (*MI).second; 2843 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n"); 2844 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); 2845 2846 // If global alias, make sure they do alias. 2847 if (hasPossibleGlobalWriteReorder(*UI, 2848 Inst, 2849 WriteObjects, 2850 MaxByteWidth)) { 2851 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI 2852 << "\n"); 2853 return false; 2854 } 2855 } 2856 TempObjects.clear(); 2857 } 2858 2859 PtrRtCheck.Need = NeedRTCheck; 2860 if (NeedRTCheck && !CanDoRT) { 2861 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " << 2862 "the array bounds.\n"); 2863 PtrRtCheck.reset(); 2864 return false; 2865 } 2866 2867 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") << 2868 " need a runtime memory check.\n"); 2869 return true; 2870} 2871 2872bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, 2873 ReductionKind Kind) { 2874 if (Phi->getNumIncomingValues() != 2) 2875 return false; 2876 2877 // Reduction variables are only found in the loop header block. 2878 if (Phi->getParent() != TheLoop->getHeader()) 2879 return false; 2880 2881 // Obtain the reduction start value from the value that comes from the loop 2882 // preheader. 2883 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 2884 2885 // ExitInstruction is the single value which is used outside the loop. 2886 // We only allow for a single reduction value to be used outside the loop. 2887 // This includes users of the reduction, variables (which form a cycle 2888 // which ends in the phi node). 2889 Instruction *ExitInstruction = 0; 2890 // Indicates that we found a binary operation in our scan. 2891 bool FoundBinOp = false; 2892 2893 // Iter is our iterator. We start with the PHI node and scan for all of the 2894 // users of this instruction. All users must be instructions that can be 2895 // used as reduction variables (such as ADD). We may have a single 2896 // out-of-block user. The cycle must end with the original PHI. 2897 Instruction *Iter = Phi; 2898 2899 // To recognize min/max patterns formed by a icmp select sequence, we store 2900 // the number of instruction we saw from the recognized min/max pattern, 2901 // such that we don't stop when we see the phi has two uses (one by the select 2902 // and one by the icmp) and to make sure we only see exactly the two 2903 // instructions. 2904 unsigned NumCmpSelectPatternInst = 0; 2905 ReductionInstDesc ReduxDesc(false, 0); 2906 2907 // Avoid cycles in the chain. 2908 SmallPtrSet<Instruction *, 8> VisitedInsts; 2909 while (VisitedInsts.insert(Iter)) { 2910 // If the instruction has no users then this is a broken 2911 // chain and can't be a reduction variable. 2912 if (Iter->use_empty()) 2913 return false; 2914 2915 // Did we find a user inside this loop already ? 2916 bool FoundInBlockUser = false; 2917 // Did we reach the initial PHI node already ? 2918 bool FoundStartPHI = false; 2919 2920 // Is this a bin op ? 2921 FoundBinOp |= !isa<PHINode>(Iter); 2922 2923 // For each of the *users* of iter. 2924 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end(); 2925 it != e; ++it) { 2926 Instruction *U = cast<Instruction>(*it); 2927 // We already know that the PHI is a user. 2928 if (U == Phi) { 2929 FoundStartPHI = true; 2930 continue; 2931 } 2932 2933 // Check if we found the exit user. 2934 BasicBlock *Parent = U->getParent(); 2935 if (!TheLoop->contains(Parent)) { 2936 // Exit if you find multiple outside users. 2937 if (ExitInstruction != 0) 2938 return false; 2939 ExitInstruction = Iter; 2940 } 2941 2942 // We allow in-loop PHINodes which are not the original reduction PHI 2943 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE 2944 // structure) then don't skip this PHI. 2945 if (isa<PHINode>(Iter) && isa<PHINode>(U) && 2946 U->getParent() != TheLoop->getHeader() && 2947 TheLoop->contains(U) && 2948 Iter->hasNUsesOrMore(2)) 2949 continue; 2950 2951 // We can't have multiple inside users except for a combination of 2952 // icmp/select both using the phi. 2953 if (FoundInBlockUser && !NumCmpSelectPatternInst) 2954 return false; 2955 FoundInBlockUser = true; 2956 2957 // Any reduction instr must be of one of the allowed kinds. 2958 ReduxDesc = isReductionInstr(U, Kind, ReduxDesc); 2959 if (!ReduxDesc.IsReduction) 2960 return false; 2961 2962 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(U) || isa<SelectInst>(U))) 2963 ++NumCmpSelectPatternInst; 2964 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(U) || isa<SelectInst>(U))) 2965 ++NumCmpSelectPatternInst; 2966 2967 // Reductions of instructions such as Div, and Sub is only 2968 // possible if the LHS is the reduction variable. 2969 if (!U->isCommutative() && !isa<PHINode>(U) && !isa<SelectInst>(U) && 2970 !isa<ICmpInst>(U) && !isa<FCmpInst>(U) && U->getOperand(0) != Iter) 2971 return false; 2972 2973 Iter = ReduxDesc.PatternLastInst; 2974 } 2975 2976 // This means we have seen one but not the other instruction of the 2977 // pattern or more than just a select and cmp. 2978 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) && 2979 NumCmpSelectPatternInst != 2) 2980 return false; 2981 2982 // We found a reduction var if we have reached the original 2983 // phi node and we only have a single instruction with out-of-loop 2984 // users. 2985 if (FoundStartPHI) { 2986 // This instruction is allowed to have out-of-loop users. 2987 AllowedExit.insert(ExitInstruction); 2988 2989 // Save the description of this reduction variable. 2990 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind, 2991 ReduxDesc.MinMaxKind); 2992 Reductions[Phi] = RD; 2993 // We've ended the cycle. This is a reduction variable if we have an 2994 // outside user and it has a binary op. 2995 return FoundBinOp && ExitInstruction; 2996 } 2997 } 2998 2999 return false; 3000} 3001 3002/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 3003/// pattern corresponding to a min(X, Y) or max(X, Y). 3004LoopVectorizationLegality::ReductionInstDesc 3005LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, 3006 ReductionInstDesc &Prev) { 3007 3008 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) && 3009 "Expect a select instruction"); 3010 Instruction *Cmp = 0; 3011 SelectInst *Select = 0; 3012 3013 // We must handle the select(cmp()) as a single instruction. Advance to the 3014 // select. 3015 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) { 3016 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin()))) 3017 return ReductionInstDesc(false, I); 3018 return ReductionInstDesc(Select, Prev.MinMaxKind); 3019 } 3020 3021 // Only handle single use cases for now. 3022 if (!(Select = dyn_cast<SelectInst>(I))) 3023 return ReductionInstDesc(false, I); 3024 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) && 3025 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0)))) 3026 return ReductionInstDesc(false, I); 3027 if (!Cmp->hasOneUse()) 3028 return ReductionInstDesc(false, I); 3029 3030 Value *CmpLeft; 3031 Value *CmpRight; 3032 3033 // Look for a min/max pattern. 3034 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3035 return ReductionInstDesc(Select, MRK_UIntMin); 3036 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3037 return ReductionInstDesc(Select, MRK_UIntMax); 3038 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3039 return ReductionInstDesc(Select, MRK_SIntMax); 3040 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3041 return ReductionInstDesc(Select, MRK_SIntMin); 3042 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3043 return ReductionInstDesc(Select, MRK_FloatMin); 3044 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3045 return ReductionInstDesc(Select, MRK_FloatMax); 3046 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3047 return ReductionInstDesc(Select, MRK_FloatMin); 3048 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3049 return ReductionInstDesc(Select, MRK_FloatMax); 3050 3051 return ReductionInstDesc(false, I); 3052} 3053 3054LoopVectorizationLegality::ReductionInstDesc 3055LoopVectorizationLegality::isReductionInstr(Instruction *I, 3056 ReductionKind Kind, 3057 ReductionInstDesc &Prev) { 3058 bool FP = I->getType()->isFloatingPointTy(); 3059 bool FastMath = (FP && I->isCommutative() && I->isAssociative()); 3060 switch (I->getOpcode()) { 3061 default: 3062 return ReductionInstDesc(false, I); 3063 case Instruction::PHI: 3064 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd && 3065 Kind != RK_FloatMinMax)) 3066 return ReductionInstDesc(false, I); 3067 return ReductionInstDesc(I, Prev.MinMaxKind); 3068 case Instruction::Sub: 3069 case Instruction::Add: 3070 return ReductionInstDesc(Kind == RK_IntegerAdd, I); 3071 case Instruction::Mul: 3072 return ReductionInstDesc(Kind == RK_IntegerMult, I); 3073 case Instruction::And: 3074 return ReductionInstDesc(Kind == RK_IntegerAnd, I); 3075 case Instruction::Or: 3076 return ReductionInstDesc(Kind == RK_IntegerOr, I); 3077 case Instruction::Xor: 3078 return ReductionInstDesc(Kind == RK_IntegerXor, I); 3079 case Instruction::FMul: 3080 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I); 3081 case Instruction::FAdd: 3082 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I); 3083 case Instruction::FCmp: 3084 case Instruction::ICmp: 3085 case Instruction::Select: 3086 if (Kind != RK_IntegerMinMax && 3087 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax)) 3088 return ReductionInstDesc(false, I); 3089 return isMinMaxSelectCmpPattern(I, Prev); 3090 } 3091} 3092 3093LoopVectorizationLegality::InductionKind 3094LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { 3095 Type *PhiTy = Phi->getType(); 3096 // We only handle integer and pointer inductions variables. 3097 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 3098 return IK_NoInduction; 3099 3100 // Check that the PHI is consecutive. 3101 const SCEV *PhiScev = SE->getSCEV(Phi); 3102 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 3103 if (!AR) { 3104 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 3105 return IK_NoInduction; 3106 } 3107 const SCEV *Step = AR->getStepRecurrence(*SE); 3108 3109 // Integer inductions need to have a stride of one. 3110 if (PhiTy->isIntegerTy()) { 3111 if (Step->isOne()) 3112 return IK_IntInduction; 3113 if (Step->isAllOnesValue()) 3114 return IK_ReverseIntInduction; 3115 return IK_NoInduction; 3116 } 3117 3118 // Calculate the pointer stride and check if it is consecutive. 3119 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 3120 if (!C) 3121 return IK_NoInduction; 3122 3123 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 3124 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType()); 3125 if (C->getValue()->equalsInt(Size)) 3126 return IK_PtrInduction; 3127 else if (C->getValue()->equalsInt(0 - Size)) 3128 return IK_ReversePtrInduction; 3129 3130 return IK_NoInduction; 3131} 3132 3133bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 3134 Value *In0 = const_cast<Value*>(V); 3135 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 3136 if (!PN) 3137 return false; 3138 3139 return Inductions.count(PN); 3140} 3141 3142bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 3143 assert(TheLoop->contains(BB) && "Unknown block used"); 3144 3145 // Blocks that do not dominate the latch need predication. 3146 BasicBlock* Latch = TheLoop->getLoopLatch(); 3147 return !DT->dominates(BB, Latch); 3148} 3149 3150bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { 3151 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3152 // We don't predicate loads/stores at the moment. 3153 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow()) 3154 return false; 3155 3156 // The instructions below can trap. 3157 switch (it->getOpcode()) { 3158 default: continue; 3159 case Instruction::UDiv: 3160 case Instruction::SDiv: 3161 case Instruction::URem: 3162 case Instruction::SRem: 3163 return false; 3164 } 3165 } 3166 3167 return true; 3168} 3169 3170bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) { 3171 const SCEV *PhiScev = SE->getSCEV(Ptr); 3172 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 3173 if (!AR) 3174 return false; 3175 3176 return AR->isAffine(); 3177} 3178 3179LoopVectorizationCostModel::VectorizationFactor 3180LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize, 3181 unsigned UserVF) { 3182 // Width 1 means no vectorize 3183 VectorizationFactor Factor = { 1U, 0U }; 3184 if (OptForSize && Legal->getRuntimePointerCheck()->Need) { 3185 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); 3186 return Factor; 3187 } 3188 3189 // Find the trip count. 3190 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch()); 3191 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n"); 3192 3193 unsigned WidestType = getWidestType(); 3194 unsigned WidestRegister = TTI.getRegisterBitWidth(true); 3195 unsigned MaxVectorSize = WidestRegister / WidestType; 3196 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n"); 3197 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n"); 3198 3199 if (MaxVectorSize == 0) { 3200 DEBUG(dbgs() << "LV: The target has no vector registers.\n"); 3201 MaxVectorSize = 1; 3202 } 3203 3204 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements" 3205 " into one vector!"); 3206 3207 unsigned VF = MaxVectorSize; 3208 3209 // If we optimize the program for size, avoid creating the tail loop. 3210 if (OptForSize) { 3211 // If we are unable to calculate the trip count then don't try to vectorize. 3212 if (TC < 2) { 3213 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 3214 return Factor; 3215 } 3216 3217 // Find the maximum SIMD width that can fit within the trip count. 3218 VF = TC % MaxVectorSize; 3219 3220 if (VF == 0) 3221 VF = MaxVectorSize; 3222 3223 // If the trip count that we found modulo the vectorization factor is not 3224 // zero then we require a tail. 3225 if (VF < 2) { 3226 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 3227 return Factor; 3228 } 3229 } 3230 3231 if (UserVF != 0) { 3232 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 3233 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n"); 3234 3235 Factor.Width = UserVF; 3236 return Factor; 3237 } 3238 3239 float Cost = expectedCost(1); 3240 unsigned Width = 1; 3241 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n"); 3242 for (unsigned i=2; i <= VF; i*=2) { 3243 // Notice that the vector loop needs to be executed less times, so 3244 // we need to divide the cost of the vector loops by the width of 3245 // the vector elements. 3246 float VectorCost = expectedCost(i) / (float)i; 3247 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " << 3248 (int)VectorCost << ".\n"); 3249 if (VectorCost < Cost) { 3250 Cost = VectorCost; 3251 Width = i; 3252 } 3253 } 3254 3255 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n"); 3256 Factor.Width = Width; 3257 Factor.Cost = Width * Cost; 3258 return Factor; 3259} 3260 3261unsigned LoopVectorizationCostModel::getWidestType() { 3262 unsigned MaxWidth = 8; 3263 3264 // For each block. 3265 for (Loop::block_iterator bb = TheLoop->block_begin(), 3266 be = TheLoop->block_end(); bb != be; ++bb) { 3267 BasicBlock *BB = *bb; 3268 3269 // For each instruction in the loop. 3270 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3271 Type *T = it->getType(); 3272 3273 // Only examine Loads, Stores and PHINodes. 3274 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it)) 3275 continue; 3276 3277 // Examine PHI nodes that are reduction variables. 3278 if (PHINode *PN = dyn_cast<PHINode>(it)) 3279 if (!Legal->getReductionVars()->count(PN)) 3280 continue; 3281 3282 // Examine the stored values. 3283 if (StoreInst *ST = dyn_cast<StoreInst>(it)) 3284 T = ST->getValueOperand()->getType(); 3285 3286 // Ignore loaded pointer types and stored pointer types that are not 3287 // consecutive. However, we do want to take consecutive stores/loads of 3288 // pointer vectors into account. 3289 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it)) 3290 continue; 3291 3292 MaxWidth = std::max(MaxWidth, 3293 (unsigned)DL->getTypeSizeInBits(T->getScalarType())); 3294 } 3295 } 3296 3297 return MaxWidth; 3298} 3299 3300unsigned 3301LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, 3302 unsigned UserUF, 3303 unsigned VF, 3304 unsigned LoopCost) { 3305 3306 // -- The unroll heuristics -- 3307 // We unroll the loop in order to expose ILP and reduce the loop overhead. 3308 // There are many micro-architectural considerations that we can't predict 3309 // at this level. For example frontend pressure (on decode or fetch) due to 3310 // code size, or the number and capabilities of the execution ports. 3311 // 3312 // We use the following heuristics to select the unroll factor: 3313 // 1. If the code has reductions the we unroll in order to break the cross 3314 // iteration dependency. 3315 // 2. If the loop is really small then we unroll in order to reduce the loop 3316 // overhead. 3317 // 3. We don't unroll if we think that we will spill registers to memory due 3318 // to the increased register pressure. 3319 3320 // Use the user preference, unless 'auto' is selected. 3321 if (UserUF != 0) 3322 return UserUF; 3323 3324 // When we optimize for size we don't unroll. 3325 if (OptForSize) 3326 return 1; 3327 3328 // Do not unroll loops with a relatively small trip count. 3329 unsigned TC = SE->getSmallConstantTripCount(TheLoop, 3330 TheLoop->getLoopLatch()); 3331 if (TC > 1 && TC < TinyTripCountUnrollThreshold) 3332 return 1; 3333 3334 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true); 3335 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters << 3336 " vector registers\n"); 3337 3338 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); 3339 // We divide by these constants so assume that we have at least one 3340 // instruction that uses at least one register. 3341 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 3342 R.NumInstructions = std::max(R.NumInstructions, 1U); 3343 3344 // We calculate the unroll factor using the following formula. 3345 // Subtract the number of loop invariants from the number of available 3346 // registers. These registers are used by all of the unrolled instances. 3347 // Next, divide the remaining registers by the number of registers that is 3348 // required by the loop, in order to estimate how many parallel instances 3349 // fit without causing spills. 3350 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers; 3351 3352 // Clamp the unroll factor ranges to reasonable factors. 3353 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor(); 3354 3355 // If we did not calculate the cost for VF (because the user selected the VF) 3356 // then we calculate the cost of VF here. 3357 if (LoopCost == 0) 3358 LoopCost = expectedCost(VF); 3359 3360 // Clamp the calculated UF to be between the 1 and the max unroll factor 3361 // that the target allows. 3362 if (UF > MaxUnrollSize) 3363 UF = MaxUnrollSize; 3364 else if (UF < 1) 3365 UF = 1; 3366 3367 if (Legal->getReductionVars()->size()) { 3368 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n"); 3369 return UF; 3370 } 3371 3372 // We want to unroll tiny loops in order to reduce the loop overhead. 3373 // We assume that the cost overhead is 1 and we use the cost model 3374 // to estimate the cost of the loop and unroll until the cost of the 3375 // loop overhead is about 5% of the cost of the loop. 3376 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n"); 3377 if (LoopCost < 20) { 3378 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n"); 3379 unsigned NewUF = 20/LoopCost + 1; 3380 return std::min(NewUF, UF); 3381 } 3382 3383 DEBUG(dbgs() << "LV: Not Unrolling. \n"); 3384 return 1; 3385} 3386 3387LoopVectorizationCostModel::RegisterUsage 3388LoopVectorizationCostModel::calculateRegisterUsage() { 3389 // This function calculates the register usage by measuring the highest number 3390 // of values that are alive at a single location. Obviously, this is a very 3391 // rough estimation. We scan the loop in a topological order in order and 3392 // assign a number to each instruction. We use RPO to ensure that defs are 3393 // met before their users. We assume that each instruction that has in-loop 3394 // users starts an interval. We record every time that an in-loop value is 3395 // used, so we have a list of the first and last occurrences of each 3396 // instruction. Next, we transpose this data structure into a multi map that 3397 // holds the list of intervals that *end* at a specific location. This multi 3398 // map allows us to perform a linear search. We scan the instructions linearly 3399 // and record each time that a new interval starts, by placing it in a set. 3400 // If we find this value in the multi-map then we remove it from the set. 3401 // The max register usage is the maximum size of the set. 3402 // We also search for instructions that are defined outside the loop, but are 3403 // used inside the loop. We need this number separately from the max-interval 3404 // usage number because when we unroll, loop-invariant values do not take 3405 // more register. 3406 LoopBlocksDFS DFS(TheLoop); 3407 DFS.perform(LI); 3408 3409 RegisterUsage R; 3410 R.NumInstructions = 0; 3411 3412 // Each 'key' in the map opens a new interval. The values 3413 // of the map are the index of the 'last seen' usage of the 3414 // instruction that is the key. 3415 typedef DenseMap<Instruction*, unsigned> IntervalMap; 3416 // Maps instruction to its index. 3417 DenseMap<unsigned, Instruction*> IdxToInstr; 3418 // Marks the end of each interval. 3419 IntervalMap EndPoint; 3420 // Saves the list of instruction indices that are used in the loop. 3421 SmallSet<Instruction*, 8> Ends; 3422 // Saves the list of values that are used in the loop but are 3423 // defined outside the loop, such as arguments and constants. 3424 SmallPtrSet<Value*, 8> LoopInvariants; 3425 3426 unsigned Index = 0; 3427 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 3428 be = DFS.endRPO(); bb != be; ++bb) { 3429 R.NumInstructions += (*bb)->size(); 3430 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 3431 ++it) { 3432 Instruction *I = it; 3433 IdxToInstr[Index++] = I; 3434 3435 // Save the end location of each USE. 3436 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 3437 Value *U = I->getOperand(i); 3438 Instruction *Instr = dyn_cast<Instruction>(U); 3439 3440 // Ignore non-instruction values such as arguments, constants, etc. 3441 if (!Instr) continue; 3442 3443 // If this instruction is outside the loop then record it and continue. 3444 if (!TheLoop->contains(Instr)) { 3445 LoopInvariants.insert(Instr); 3446 continue; 3447 } 3448 3449 // Overwrite previous end points. 3450 EndPoint[Instr] = Index; 3451 Ends.insert(Instr); 3452 } 3453 } 3454 } 3455 3456 // Saves the list of intervals that end with the index in 'key'. 3457 typedef SmallVector<Instruction*, 2> InstrList; 3458 DenseMap<unsigned, InstrList> TransposeEnds; 3459 3460 // Transpose the EndPoints to a list of values that end at each index. 3461 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 3462 it != e; ++it) 3463 TransposeEnds[it->second].push_back(it->first); 3464 3465 SmallSet<Instruction*, 8> OpenIntervals; 3466 unsigned MaxUsage = 0; 3467 3468 3469 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 3470 for (unsigned int i = 0; i < Index; ++i) { 3471 Instruction *I = IdxToInstr[i]; 3472 // Ignore instructions that are never used within the loop. 3473 if (!Ends.count(I)) continue; 3474 3475 // Remove all of the instructions that end at this location. 3476 InstrList &List = TransposeEnds[i]; 3477 for (unsigned int j=0, e = List.size(); j < e; ++j) 3478 OpenIntervals.erase(List[j]); 3479 3480 // Count the number of live interals. 3481 MaxUsage = std::max(MaxUsage, OpenIntervals.size()); 3482 3483 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << 3484 OpenIntervals.size() <<"\n"); 3485 3486 // Add the current instruction to the list of open intervals. 3487 OpenIntervals.insert(I); 3488 } 3489 3490 unsigned Invariant = LoopInvariants.size(); 3491 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n"); 3492 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n"); 3493 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n"); 3494 3495 R.LoopInvariantRegs = Invariant; 3496 R.MaxLocalUsers = MaxUsage; 3497 return R; 3498} 3499 3500unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 3501 unsigned Cost = 0; 3502 3503 // For each block. 3504 for (Loop::block_iterator bb = TheLoop->block_begin(), 3505 be = TheLoop->block_end(); bb != be; ++bb) { 3506 unsigned BlockCost = 0; 3507 BasicBlock *BB = *bb; 3508 3509 // For each instruction in the old loop. 3510 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3511 // Skip dbg intrinsics. 3512 if (isa<DbgInfoIntrinsic>(it)) 3513 continue; 3514 3515 unsigned C = getInstructionCost(it, VF); 3516 Cost += C; 3517 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " << 3518 VF << " For instruction: "<< *it << "\n"); 3519 } 3520 3521 // We assume that if-converted blocks have a 50% chance of being executed. 3522 // When the code is scalar then some of the blocks are avoided due to CF. 3523 // When the code is vectorized we execute all code paths. 3524 if (Legal->blockNeedsPredication(*bb) && VF == 1) 3525 BlockCost /= 2; 3526 3527 Cost += BlockCost; 3528 } 3529 3530 return Cost; 3531} 3532 3533unsigned 3534LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 3535 // If we know that this instruction will remain uniform, check the cost of 3536 // the scalar version. 3537 if (Legal->isUniformAfterVectorization(I)) 3538 VF = 1; 3539 3540 Type *RetTy = I->getType(); 3541 Type *VectorTy = ToVectorTy(RetTy, VF); 3542 3543 // TODO: We need to estimate the cost of intrinsic calls. 3544 switch (I->getOpcode()) { 3545 case Instruction::GetElementPtr: 3546 // We mark this instruction as zero-cost because the cost of GEPs in 3547 // vectorized code depends on whether the corresponding memory instruction 3548 // is scalarized or not. Therefore, we handle GEPs with the memory 3549 // instruction cost. 3550 return 0; 3551 case Instruction::Br: { 3552 return TTI.getCFInstrCost(I->getOpcode()); 3553 } 3554 case Instruction::PHI: 3555 //TODO: IF-converted IFs become selects. 3556 return 0; 3557 case Instruction::Add: 3558 case Instruction::FAdd: 3559 case Instruction::Sub: 3560 case Instruction::FSub: 3561 case Instruction::Mul: 3562 case Instruction::FMul: 3563 case Instruction::UDiv: 3564 case Instruction::SDiv: 3565 case Instruction::FDiv: 3566 case Instruction::URem: 3567 case Instruction::SRem: 3568 case Instruction::FRem: 3569 case Instruction::Shl: 3570 case Instruction::LShr: 3571 case Instruction::AShr: 3572 case Instruction::And: 3573 case Instruction::Or: 3574 case Instruction::Xor: { 3575 // Certain instructions can be cheaper to vectorize if they have a constant 3576 // second vector operand. One example of this are shifts on x86. 3577 TargetTransformInfo::OperandValueKind Op1VK = 3578 TargetTransformInfo::OK_AnyValue; 3579 TargetTransformInfo::OperandValueKind Op2VK = 3580 TargetTransformInfo::OK_AnyValue; 3581 3582 if (isa<ConstantInt>(I->getOperand(1))) 3583 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 3584 3585 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK); 3586 } 3587 case Instruction::Select: { 3588 SelectInst *SI = cast<SelectInst>(I); 3589 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 3590 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 3591 Type *CondTy = SI->getCondition()->getType(); 3592 if (!ScalarCond) 3593 CondTy = VectorType::get(CondTy, VF); 3594 3595 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 3596 } 3597 case Instruction::ICmp: 3598 case Instruction::FCmp: { 3599 Type *ValTy = I->getOperand(0)->getType(); 3600 VectorTy = ToVectorTy(ValTy, VF); 3601 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); 3602 } 3603 case Instruction::Store: 3604 case Instruction::Load: { 3605 StoreInst *SI = dyn_cast<StoreInst>(I); 3606 LoadInst *LI = dyn_cast<LoadInst>(I); 3607 Type *ValTy = (SI ? SI->getValueOperand()->getType() : 3608 LI->getType()); 3609 VectorTy = ToVectorTy(ValTy, VF); 3610 3611 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); 3612 unsigned AS = SI ? SI->getPointerAddressSpace() : 3613 LI->getPointerAddressSpace(); 3614 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); 3615 // We add the cost of address computation here instead of with the gep 3616 // instruction because only here we know whether the operation is 3617 // scalarized. 3618 if (VF == 1) 3619 return TTI.getAddressComputationCost(VectorTy) + 3620 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 3621 3622 // Scalarized loads/stores. 3623 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 3624 bool Reverse = ConsecutiveStride < 0; 3625 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy); 3626 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF; 3627 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { 3628 unsigned Cost = 0; 3629 // The cost of extracting from the value vector and pointer vector. 3630 Type *PtrTy = ToVectorTy(Ptr->getType(), VF); 3631 for (unsigned i = 0; i < VF; ++i) { 3632 // The cost of extracting the pointer operand. 3633 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); 3634 // In case of STORE, the cost of ExtractElement from the vector. 3635 // In case of LOAD, the cost of InsertElement into the returned 3636 // vector. 3637 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : 3638 Instruction::InsertElement, 3639 VectorTy, i); 3640 } 3641 3642 // The cost of the scalar loads/stores. 3643 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType()); 3644 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), 3645 Alignment, AS); 3646 return Cost; 3647 } 3648 3649 // Wide load/stores. 3650 unsigned Cost = TTI.getAddressComputationCost(VectorTy); 3651 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 3652 3653 if (Reverse) 3654 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, 3655 VectorTy, 0); 3656 return Cost; 3657 } 3658 case Instruction::ZExt: 3659 case Instruction::SExt: 3660 case Instruction::FPToUI: 3661 case Instruction::FPToSI: 3662 case Instruction::FPExt: 3663 case Instruction::PtrToInt: 3664 case Instruction::IntToPtr: 3665 case Instruction::SIToFP: 3666 case Instruction::UIToFP: 3667 case Instruction::Trunc: 3668 case Instruction::FPTrunc: 3669 case Instruction::BitCast: { 3670 // We optimize the truncation of induction variable. 3671 // The cost of these is the same as the scalar operation. 3672 if (I->getOpcode() == Instruction::Trunc && 3673 Legal->isInductionVariable(I->getOperand(0))) 3674 return TTI.getCastInstrCost(I->getOpcode(), I->getType(), 3675 I->getOperand(0)->getType()); 3676 3677 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); 3678 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 3679 } 3680 case Instruction::Call: { 3681 CallInst *CI = cast<CallInst>(I); 3682 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 3683 assert(ID && "Not an intrinsic call!"); 3684 Type *RetTy = ToVectorTy(CI->getType(), VF); 3685 SmallVector<Type*, 4> Tys; 3686 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 3687 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 3688 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); 3689 } 3690 default: { 3691 // We are scalarizing the instruction. Return the cost of the scalar 3692 // instruction, plus the cost of insert and extract into vector 3693 // elements, times the vector width. 3694 unsigned Cost = 0; 3695 3696 if (!RetTy->isVoidTy() && VF != 1) { 3697 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, 3698 VectorTy); 3699 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, 3700 VectorTy); 3701 3702 // The cost of inserting the results plus extracting each one of the 3703 // operands. 3704 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 3705 } 3706 3707 // The cost of executing VF copies of the scalar instruction. This opcode 3708 // is unknown. Assume that it is the same as 'mul'. 3709 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); 3710 return Cost; 3711 } 3712 }// end of switch. 3713} 3714 3715Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) { 3716 if (Scalar->isVoidTy() || VF == 1) 3717 return Scalar; 3718 return VectorType::get(Scalar, VF); 3719} 3720 3721char LoopVectorize::ID = 0; 3722static const char lv_name[] = "Loop Vectorization"; 3723INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 3724INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3725INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3726INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3727INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 3728INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 3729 3730namespace llvm { 3731 Pass *createLoopVectorizePass() { 3732 return new LoopVectorize(); 3733 } 3734} 3735 3736bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { 3737 // Check for a store. 3738 if (StoreInst *ST = dyn_cast<StoreInst>(Inst)) 3739 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; 3740 3741 // Check for a load. 3742 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 3743 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; 3744 3745 return false; 3746} 3747