LoopVectorize.cpp revision 4bcd5f888fa762613cf8096a79ba7b8a72665de2
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) {} 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 }; 352 353 /// This enum represents the kinds of inductions that we support. 354 enum InductionKind { 355 IK_NoInduction, ///< Not an induction variable. 356 IK_IntInduction, ///< Integer induction variable. Step = 1. 357 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1. 358 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem). 359 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem). 360 }; 361 362 // This enum represents the kind of minmax reduction. 363 enum MinMaxReductionKind { 364 MRK_Invalid, 365 MRK_UIntMin, 366 MRK_UIntMax, 367 MRK_SIntMin, 368 MRK_SIntMax 369 }; 370 371 /// This POD struct holds information about reduction variables. 372 struct ReductionDescriptor { 373 ReductionDescriptor() : StartValue(0), LoopExitInstr(0), 374 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {} 375 376 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K, 377 MinMaxReductionKind MK) 378 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {} 379 380 // The starting value of the reduction. 381 // It does not have to be zero! 382 Value *StartValue; 383 // The instruction who's value is used outside the loop. 384 Instruction *LoopExitInstr; 385 // The kind of the reduction. 386 ReductionKind Kind; 387 // If this a min/max reduction the kind of reduction. 388 MinMaxReductionKind MinMaxKind; 389 }; 390 391 /// This POD struct holds information about a potential reduction operation. 392 struct ReductionInstDesc { 393 ReductionInstDesc(bool IsRedux, Instruction *I) : 394 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {} 395 396 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) : 397 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {} 398 399 // Is this instruction a reduction candidate. 400 bool IsReduction; 401 // The last instruction in a min/max pattern (select of the select(icmp()) 402 // pattern), or the current reduction instruction otherwise. 403 Instruction *PatternLastInst; 404 // If this is a min/max pattern the comparison predicate. 405 MinMaxReductionKind MinMaxKind; 406 }; 407 408 // This POD struct holds information about the memory runtime legality 409 // check that a group of pointers do not overlap. 410 struct RuntimePointerCheck { 411 RuntimePointerCheck() : Need(false) {} 412 413 /// Reset the state of the pointer runtime information. 414 void reset() { 415 Need = false; 416 Pointers.clear(); 417 Starts.clear(); 418 Ends.clear(); 419 } 420 421 /// Insert a pointer and calculate the start and end SCEVs. 422 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr); 423 424 /// This flag indicates if we need to add the runtime check. 425 bool Need; 426 /// Holds the pointers that we need to check. 427 SmallVector<Value*, 2> Pointers; 428 /// Holds the pointer value at the beginning of the loop. 429 SmallVector<const SCEV*, 2> Starts; 430 /// Holds the pointer value at the end of the loop. 431 SmallVector<const SCEV*, 2> Ends; 432 /// Holds the information if this pointer is used for writing to memory. 433 SmallVector<bool, 2> IsWritePtr; 434 }; 435 436 /// A POD for saving information about induction variables. 437 struct InductionInfo { 438 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {} 439 InductionInfo() : StartValue(0), IK(IK_NoInduction) {} 440 /// Start value. 441 Value *StartValue; 442 /// Induction kind. 443 InductionKind IK; 444 }; 445 446 /// ReductionList contains the reduction descriptors for all 447 /// of the reductions that were found in the loop. 448 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList; 449 450 /// InductionList saves induction variables and maps them to the 451 /// induction descriptor. 452 typedef MapVector<PHINode*, InductionInfo> InductionList; 453 454 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their 455 /// respective Store/Load instruction(s) to calculate aliasing. 456 typedef MapVector<Value*, Instruction* > AliasMap; 457 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap; 458 459 /// Returns true if it is legal to vectorize this loop. 460 /// This does not mean that it is profitable to vectorize this 461 /// loop, only that it is legal to do so. 462 bool canVectorize(); 463 464 /// Returns the Induction variable. 465 PHINode *getInduction() { return Induction; } 466 467 /// Returns the reduction variables found in the loop. 468 ReductionList *getReductionVars() { return &Reductions; } 469 470 /// Returns the induction variables found in the loop. 471 InductionList *getInductionVars() { return &Inductions; } 472 473 /// Returns True if V is an induction variable in this loop. 474 bool isInductionVariable(const Value *V); 475 476 /// Return true if the block BB needs to be predicated in order for the loop 477 /// to be vectorized. 478 bool blockNeedsPredication(BasicBlock *BB); 479 480 /// Check if this pointer is consecutive when vectorizing. This happens 481 /// when the last index of the GEP is the induction variable, or that the 482 /// pointer itself is an induction variable. 483 /// This check allows us to vectorize A[idx] into a wide load/store. 484 /// Returns: 485 /// 0 - Stride is unknown or non consecutive. 486 /// 1 - Address is consecutive. 487 /// -1 - Address is consecutive, and decreasing. 488 int isConsecutivePtr(Value *Ptr); 489 490 /// Returns true if the value V is uniform within the loop. 491 bool isUniform(Value *V); 492 493 /// Returns true if this instruction will remain scalar after vectorization. 494 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); } 495 496 /// Returns the information that we collected about runtime memory check. 497 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; } 498 499 /// This function returns the identity element (or neutral element) for 500 /// the operation K. 501 static Constant *getReductionIdentity(ReductionKind K, Type *Tp, 502 MinMaxReductionKind MinMaxK); 503private: 504 /// Check if a single basic block loop is vectorizable. 505 /// At this point we know that this is a loop with a constant trip count 506 /// and we only need to check individual instructions. 507 bool canVectorizeInstrs(); 508 509 /// When we vectorize loops we may change the order in which 510 /// we read and write from memory. This method checks if it is 511 /// legal to vectorize the code, considering only memory constrains. 512 /// Returns true if the loop is vectorizable 513 bool canVectorizeMemory(); 514 515 /// Return true if we can vectorize this loop using the IF-conversion 516 /// transformation. 517 bool canVectorizeWithIfConvert(); 518 519 /// Collect the variables that need to stay uniform after vectorization. 520 void collectLoopUniforms(); 521 522 /// Return true if all of the instructions in the block can be speculatively 523 /// executed. 524 bool blockCanBePredicated(BasicBlock *BB); 525 526 /// Returns True, if 'Phi' is the kind of reduction variable for type 527 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. 528 bool AddReductionVar(PHINode *Phi, ReductionKind Kind); 529 /// Returns a struct describing if the instruction 'I' can be a reduction 530 /// variable of type 'Kind'. If the reduction is a min/max pattern of 531 /// select(icmp()) this function advances the instruction pointer 'I' from the 532 /// compare instruction to the select instruction and stores this pointer in 533 /// 'PatternLastInst' member of the returned struct. 534 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind, 535 ReductionInstDesc &Desc); 536 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 537 /// pattern corresponding to a min(X, Y) or max(X, Y). 538 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I, 539 ReductionInstDesc &Prev); 540 /// Returns the induction kind of Phi. This function may return NoInduction 541 /// if the PHI is not an induction variable. 542 InductionKind isInductionVariable(PHINode *Phi); 543 /// Return true if can compute the address bounds of Ptr within the loop. 544 bool hasComputableBounds(Value *Ptr); 545 /// Return true if there is the chance of write reorder. 546 bool hasPossibleGlobalWriteReorder(Value *Object, 547 Instruction *Inst, 548 AliasMultiMap &WriteObjects, 549 unsigned MaxByteWidth); 550 /// Return the AA location for a load or a store. 551 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst); 552 553 554 /// The loop that we evaluate. 555 Loop *TheLoop; 556 /// Scev analysis. 557 ScalarEvolution *SE; 558 /// DataLayout analysis. 559 DataLayout *DL; 560 /// Dominators. 561 DominatorTree *DT; 562 /// Target Info. 563 TargetTransformInfo *TTI; 564 /// Alias Analysis. 565 AliasAnalysis *AA; 566 /// Target Library Info. 567 TargetLibraryInfo *TLI; 568 569 // --- vectorization state --- // 570 571 /// Holds the integer induction variable. This is the counter of the 572 /// loop. 573 PHINode *Induction; 574 /// Holds the reduction variables. 575 ReductionList Reductions; 576 /// Holds all of the induction variables that we found in the loop. 577 /// Notice that inductions don't need to start at zero and that induction 578 /// variables can be pointers. 579 InductionList Inductions; 580 581 /// Allowed outside users. This holds the reduction 582 /// vars which can be accessed from outside the loop. 583 SmallPtrSet<Value*, 4> AllowedExit; 584 /// This set holds the variables which are known to be uniform after 585 /// vectorization. 586 SmallPtrSet<Instruction*, 4> Uniforms; 587 /// We need to check that all of the pointers in this list are disjoint 588 /// at runtime. 589 RuntimePointerCheck PtrRtCheck; 590}; 591 592/// LoopVectorizationCostModel - estimates the expected speedups due to 593/// vectorization. 594/// In many cases vectorization is not profitable. This can happen because of 595/// a number of reasons. In this class we mainly attempt to predict the 596/// expected speedup/slowdowns due to the supported instruction set. We use the 597/// TargetTransformInfo to query the different backends for the cost of 598/// different operations. 599class LoopVectorizationCostModel { 600public: 601 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI, 602 LoopVectorizationLegality *Legal, 603 const TargetTransformInfo &TTI, 604 DataLayout *DL, const TargetLibraryInfo *TLI) 605 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {} 606 607 /// Information about vectorization costs 608 struct VectorizationFactor { 609 unsigned Width; // Vector width with best cost 610 unsigned Cost; // Cost of the loop with that width 611 }; 612 /// \return The most profitable vectorization factor and the cost of that VF. 613 /// This method checks every power of two up to VF. If UserVF is not ZERO 614 /// then this vectorization factor will be selected if vectorization is 615 /// possible. 616 VectorizationFactor selectVectorizationFactor(bool OptForSize, 617 unsigned UserVF); 618 619 /// \return The size (in bits) of the widest type in the code that 620 /// needs to be vectorized. We ignore values that remain scalar such as 621 /// 64 bit loop indices. 622 unsigned getWidestType(); 623 624 /// \return The most profitable unroll factor. 625 /// If UserUF is non-zero then this method finds the best unroll-factor 626 /// based on register pressure and other parameters. 627 /// VF and LoopCost are the selected vectorization factor and the cost of the 628 /// selected VF. 629 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF, 630 unsigned LoopCost); 631 632 /// \brief A struct that represents some properties of the register usage 633 /// of a loop. 634 struct RegisterUsage { 635 /// Holds the number of loop invariant values that are used in the loop. 636 unsigned LoopInvariantRegs; 637 /// Holds the maximum number of concurrent live intervals in the loop. 638 unsigned MaxLocalUsers; 639 /// Holds the number of instructions in the loop. 640 unsigned NumInstructions; 641 }; 642 643 /// \return information about the register usage of the loop. 644 RegisterUsage calculateRegisterUsage(); 645 646private: 647 /// Returns the expected execution cost. The unit of the cost does 648 /// not matter because we use the 'cost' units to compare different 649 /// vector widths. The cost that is returned is *not* normalized by 650 /// the factor width. 651 unsigned expectedCost(unsigned VF); 652 653 /// Returns the execution time cost of an instruction for a given vector 654 /// width. Vector width of one means scalar. 655 unsigned getInstructionCost(Instruction *I, unsigned VF); 656 657 /// A helper function for converting Scalar types to vector types. 658 /// If the incoming type is void, we return void. If the VF is 1, we return 659 /// the scalar type. 660 static Type* ToVectorTy(Type *Scalar, unsigned VF); 661 662 /// Returns whether the instruction is a load or store and will be a emitted 663 /// as a vector operation. 664 bool isConsecutiveLoadOrStore(Instruction *I); 665 666 /// The loop that we evaluate. 667 Loop *TheLoop; 668 /// Scev analysis. 669 ScalarEvolution *SE; 670 /// Loop Info analysis. 671 LoopInfo *LI; 672 /// Vectorization legality. 673 LoopVectorizationLegality *Legal; 674 /// Vector target information. 675 const TargetTransformInfo &TTI; 676 /// Target data layout information. 677 DataLayout *DL; 678 /// Target Library Info. 679 const TargetLibraryInfo *TLI; 680}; 681 682/// The LoopVectorize Pass. 683struct LoopVectorize : public LoopPass { 684 /// Pass identification, replacement for typeid 685 static char ID; 686 687 explicit LoopVectorize() : LoopPass(ID) { 688 initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); 689 } 690 691 ScalarEvolution *SE; 692 DataLayout *DL; 693 LoopInfo *LI; 694 TargetTransformInfo *TTI; 695 DominatorTree *DT; 696 AliasAnalysis *AA; 697 TargetLibraryInfo *TLI; 698 699 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) { 700 // We only vectorize innermost loops. 701 if (!L->empty()) 702 return false; 703 704 SE = &getAnalysis<ScalarEvolution>(); 705 DL = getAnalysisIfAvailable<DataLayout>(); 706 LI = &getAnalysis<LoopInfo>(); 707 TTI = &getAnalysis<TargetTransformInfo>(); 708 DT = &getAnalysis<DominatorTree>(); 709 AA = getAnalysisIfAvailable<AliasAnalysis>(); 710 TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); 711 712 if (DL == NULL) { 713 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout"); 714 return false; 715 } 716 717 DEBUG(dbgs() << "LV: Checking a loop in \"" << 718 L->getHeader()->getParent()->getName() << "\"\n"); 719 720 // Check if it is legal to vectorize the loop. 721 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI); 722 if (!LVL.canVectorize()) { 723 DEBUG(dbgs() << "LV: Not vectorizing.\n"); 724 return false; 725 } 726 727 // Use the cost model. 728 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI); 729 730 // Check the function attributes to find out if this function should be 731 // optimized for size. 732 Function *F = L->getHeader()->getParent(); 733 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize; 734 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat; 735 unsigned FnIndex = AttributeSet::FunctionIndex; 736 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr); 737 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr); 738 739 if (NoFloat) { 740 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat" 741 "attribute is used.\n"); 742 return false; 743 } 744 745 // Select the optimal vectorization factor. 746 LoopVectorizationCostModel::VectorizationFactor VF; 747 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor); 748 // Select the unroll factor. 749 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll, 750 VF.Width, VF.Cost); 751 752 if (VF.Width == 1) { 753 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); 754 return false; 755 } 756 757 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<< 758 F->getParent()->getModuleIdentifier()<<"\n"); 759 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n"); 760 761 // If we decided that it is *legal* to vectorize the loop then do it. 762 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF); 763 LB.vectorize(&LVL); 764 765 DEBUG(verifyFunction(*L->getHeader()->getParent())); 766 return true; 767 } 768 769 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 770 LoopPass::getAnalysisUsage(AU); 771 AU.addRequiredID(LoopSimplifyID); 772 AU.addRequiredID(LCSSAID); 773 AU.addRequired<DominatorTree>(); 774 AU.addRequired<LoopInfo>(); 775 AU.addRequired<ScalarEvolution>(); 776 AU.addRequired<TargetTransformInfo>(); 777 AU.addPreserved<LoopInfo>(); 778 AU.addPreserved<DominatorTree>(); 779 } 780 781}; 782 783} // end anonymous namespace 784 785//===----------------------------------------------------------------------===// 786// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and 787// LoopVectorizationCostModel. 788//===----------------------------------------------------------------------===// 789 790void 791LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE, 792 Loop *Lp, Value *Ptr, 793 bool WritePtr) { 794 const SCEV *Sc = SE->getSCEV(Ptr); 795 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); 796 assert(AR && "Invalid addrec expression"); 797 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch()); 798 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE); 799 Pointers.push_back(Ptr); 800 Starts.push_back(AR->getStart()); 801 Ends.push_back(ScEnd); 802 IsWritePtr.push_back(WritePtr); 803} 804 805Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { 806 // Save the current insertion location. 807 Instruction *Loc = Builder.GetInsertPoint(); 808 809 // We need to place the broadcast of invariant variables outside the loop. 810 Instruction *Instr = dyn_cast<Instruction>(V); 811 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody); 812 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr; 813 814 // Place the code for broadcasting invariant variables in the new preheader. 815 if (Invariant) 816 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); 817 818 // Broadcast the scalar into all locations in the vector. 819 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); 820 821 // Restore the builder insertion point. 822 if (Invariant) 823 Builder.SetInsertPoint(Loc); 824 825 return Shuf; 826} 827 828Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx, 829 bool Negate) { 830 assert(Val->getType()->isVectorTy() && "Must be a vector"); 831 assert(Val->getType()->getScalarType()->isIntegerTy() && 832 "Elem must be an integer"); 833 // Create the types. 834 Type *ITy = Val->getType()->getScalarType(); 835 VectorType *Ty = cast<VectorType>(Val->getType()); 836 int VLen = Ty->getNumElements(); 837 SmallVector<Constant*, 8> Indices; 838 839 // Create a vector of consecutive numbers from zero to VF. 840 for (int i = 0; i < VLen; ++i) { 841 int Idx = Negate ? (-i): i; 842 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx)); 843 } 844 845 // Add the consecutive indices to the vector value. 846 Constant *Cv = ConstantVector::get(Indices); 847 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); 848 return Builder.CreateAdd(Val, Cv, "induction"); 849} 850 851int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) { 852 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr"); 853 // Make sure that the pointer does not point to structs. 854 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) 855 return 0; 856 857 // If this value is a pointer induction variable we know it is consecutive. 858 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr); 859 if (Phi && Inductions.count(Phi)) { 860 InductionInfo II = Inductions[Phi]; 861 if (IK_PtrInduction == II.IK) 862 return 1; 863 else if (IK_ReversePtrInduction == II.IK) 864 return -1; 865 } 866 867 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr); 868 if (!Gep) 869 return 0; 870 871 unsigned NumOperands = Gep->getNumOperands(); 872 Value *LastIndex = Gep->getOperand(NumOperands - 1); 873 874 Value *GpPtr = Gep->getPointerOperand(); 875 // If this GEP value is a consecutive pointer induction variable and all of 876 // the indices are constant then we know it is consecutive. We can 877 Phi = dyn_cast<PHINode>(GpPtr); 878 if (Phi && Inductions.count(Phi)) { 879 880 // Make sure that the pointer does not point to structs. 881 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType()); 882 if (GepPtrType->getElementType()->isAggregateType()) 883 return 0; 884 885 // Make sure that all of the index operands are loop invariant. 886 for (unsigned i = 1; i < NumOperands; ++i) 887 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 888 return 0; 889 890 InductionInfo II = Inductions[Phi]; 891 if (IK_PtrInduction == II.IK) 892 return 1; 893 else if (IK_ReversePtrInduction == II.IK) 894 return -1; 895 } 896 897 // Check that all of the gep indices are uniform except for the last. 898 for (unsigned i = 0; i < NumOperands - 1; ++i) 899 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) 900 return 0; 901 902 // We can emit wide load/stores only if the last index is the induction 903 // variable. 904 const SCEV *Last = SE->getSCEV(LastIndex); 905 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { 906 const SCEV *Step = AR->getStepRecurrence(*SE); 907 908 // The memory is consecutive because the last index is consecutive 909 // and all other indices are loop invariant. 910 if (Step->isOne()) 911 return 1; 912 if (Step->isAllOnesValue()) 913 return -1; 914 } 915 916 return 0; 917} 918 919bool LoopVectorizationLegality::isUniform(Value *V) { 920 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); 921} 922 923InnerLoopVectorizer::VectorParts& 924InnerLoopVectorizer::getVectorValue(Value *V) { 925 assert(V != Induction && "The new induction variable should not be used."); 926 assert(!V->getType()->isVectorTy() && "Can't widen a vector"); 927 928 // If we have this scalar in the map, return it. 929 if (WidenMap.has(V)) 930 return WidenMap.get(V); 931 932 // If this scalar is unknown, assume that it is a constant or that it is 933 // loop invariant. Broadcast V and save the value for future uses. 934 Value *B = getBroadcastInstrs(V); 935 return WidenMap.splat(V, B); 936} 937 938Value *InnerLoopVectorizer::reverseVector(Value *Vec) { 939 assert(Vec->getType()->isVectorTy() && "Invalid type"); 940 SmallVector<Constant*, 8> ShuffleMask; 941 for (unsigned i = 0; i < VF; ++i) 942 ShuffleMask.push_back(Builder.getInt32(VF - i - 1)); 943 944 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()), 945 ConstantVector::get(ShuffleMask), 946 "reverse"); 947} 948 949 950void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr, 951 LoopVectorizationLegality *Legal) { 952 // Attempt to issue a wide load. 953 LoadInst *LI = dyn_cast<LoadInst>(Instr); 954 StoreInst *SI = dyn_cast<StoreInst>(Instr); 955 956 assert((LI || SI) && "Invalid Load/Store instruction"); 957 958 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType(); 959 Type *DataTy = VectorType::get(ScalarDataTy, VF); 960 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand(); 961 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment(); 962 963 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy); 964 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF; 965 966 if (ScalarAllocatedSize != VectorElementSize) 967 return scalarizeInstruction(Instr); 968 969 // If the pointer is loop invariant or if it is non consecutive, 970 // scalarize the load. 971 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 972 bool Reverse = ConsecutiveStride < 0; 973 bool UniformLoad = LI && Legal->isUniform(Ptr); 974 if (!ConsecutiveStride || UniformLoad) 975 return scalarizeInstruction(Instr); 976 977 Constant *Zero = Builder.getInt32(0); 978 VectorParts &Entry = WidenMap.get(Instr); 979 980 // Handle consecutive loads/stores. 981 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); 982 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) { 983 Value *PtrOperand = Gep->getPointerOperand(); 984 Value *FirstBasePtr = getVectorValue(PtrOperand)[0]; 985 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero); 986 987 // Create the new GEP with the new induction variable. 988 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 989 Gep2->setOperand(0, FirstBasePtr); 990 Gep2->setName("gep.indvar.base"); 991 Ptr = Builder.Insert(Gep2); 992 } else if (Gep) { 993 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()), 994 OrigLoop) && "Base ptr must be invariant"); 995 996 // The last index does not have to be the induction. It can be 997 // consecutive and be a function of the index. For example A[I+1]; 998 unsigned NumOperands = Gep->getNumOperands(); 999 1000 Value *LastGepOperand = Gep->getOperand(NumOperands - 1); 1001 VectorParts &GEPParts = getVectorValue(LastGepOperand); 1002 Value *LastIndex = GEPParts[0]; 1003 LastIndex = Builder.CreateExtractElement(LastIndex, Zero); 1004 1005 // Create the new GEP with the new induction variable. 1006 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); 1007 Gep2->setOperand(NumOperands - 1, LastIndex); 1008 Gep2->setName("gep.indvar.idx"); 1009 Ptr = Builder.Insert(Gep2); 1010 } else { 1011 // Use the induction element ptr. 1012 assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); 1013 VectorParts &PtrVal = getVectorValue(Ptr); 1014 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero); 1015 } 1016 1017 // Handle Stores: 1018 if (SI) { 1019 assert(!Legal->isUniform(SI->getPointerOperand()) && 1020 "We do not allow storing to uniform addresses"); 1021 1022 VectorParts &StoredVal = getVectorValue(SI->getValueOperand()); 1023 for (unsigned Part = 0; Part < UF; ++Part) { 1024 // Calculate the pointer for the specific unroll-part. 1025 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1026 1027 if (Reverse) { 1028 // If we store to reverse consecutive memory locations then we need 1029 // to reverse the order of elements in the stored value. 1030 StoredVal[Part] = reverseVector(StoredVal[Part]); 1031 // If the address is consecutive but reversed, then the 1032 // wide store needs to start at the last vector element. 1033 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1034 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1035 } 1036 1037 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); 1038 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment); 1039 } 1040 } 1041 1042 for (unsigned Part = 0; Part < UF; ++Part) { 1043 // Calculate the pointer for the specific unroll-part. 1044 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF)); 1045 1046 if (Reverse) { 1047 // If the address is consecutive but reversed, then the 1048 // wide store needs to start at the last vector element. 1049 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF)); 1050 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF)); 1051 } 1052 1053 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo()); 1054 Value *LI = Builder.CreateLoad(VecPtr, "wide.load"); 1055 cast<LoadInst>(LI)->setAlignment(Alignment); 1056 Entry[Part] = Reverse ? reverseVector(LI) : LI; 1057 } 1058} 1059 1060void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { 1061 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); 1062 // Holds vector parameters or scalars, in case of uniform vals. 1063 SmallVector<VectorParts, 4> Params; 1064 1065 // Find all of the vectorized parameters. 1066 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1067 Value *SrcOp = Instr->getOperand(op); 1068 1069 // If we are accessing the old induction variable, use the new one. 1070 if (SrcOp == OldInduction) { 1071 Params.push_back(getVectorValue(SrcOp)); 1072 continue; 1073 } 1074 1075 // Try using previously calculated values. 1076 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); 1077 1078 // If the src is an instruction that appeared earlier in the basic block 1079 // then it should already be vectorized. 1080 if (SrcInst && OrigLoop->contains(SrcInst)) { 1081 assert(WidenMap.has(SrcInst) && "Source operand is unavailable"); 1082 // The parameter is a vector value from earlier. 1083 Params.push_back(WidenMap.get(SrcInst)); 1084 } else { 1085 // The parameter is a scalar from outside the loop. Maybe even a constant. 1086 VectorParts Scalars; 1087 Scalars.append(UF, SrcOp); 1088 Params.push_back(Scalars); 1089 } 1090 } 1091 1092 assert(Params.size() == Instr->getNumOperands() && 1093 "Invalid number of operands"); 1094 1095 // Does this instruction return a value ? 1096 bool IsVoidRetTy = Instr->getType()->isVoidTy(); 1097 1098 Value *UndefVec = IsVoidRetTy ? 0 : 1099 UndefValue::get(VectorType::get(Instr->getType(), VF)); 1100 // Create a new entry in the WidenMap and initialize it to Undef or Null. 1101 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec); 1102 1103 // For each vector unroll 'part': 1104 for (unsigned Part = 0; Part < UF; ++Part) { 1105 // For each scalar that we create: 1106 for (unsigned Width = 0; Width < VF; ++Width) { 1107 Instruction *Cloned = Instr->clone(); 1108 if (!IsVoidRetTy) 1109 Cloned->setName(Instr->getName() + ".cloned"); 1110 // Replace the operands of the cloned instrucions with extracted scalars. 1111 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { 1112 Value *Op = Params[op][Part]; 1113 // Param is a vector. Need to extract the right lane. 1114 if (Op->getType()->isVectorTy()) 1115 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width)); 1116 Cloned->setOperand(op, Op); 1117 } 1118 1119 // Place the cloned scalar in the new loop. 1120 Builder.Insert(Cloned); 1121 1122 // If the original scalar returns a value we need to place it in a vector 1123 // so that future users will be able to use it. 1124 if (!IsVoidRetTy) 1125 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned, 1126 Builder.getInt32(Width)); 1127 } 1128 } 1129} 1130 1131Instruction * 1132InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal, 1133 Instruction *Loc) { 1134 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck = 1135 Legal->getRuntimePointerCheck(); 1136 1137 if (!PtrRtCheck->Need) 1138 return NULL; 1139 1140 Instruction *MemoryRuntimeCheck = 0; 1141 unsigned NumPointers = PtrRtCheck->Pointers.size(); 1142 SmallVector<Value* , 2> Starts; 1143 SmallVector<Value* , 2> Ends; 1144 1145 SCEVExpander Exp(*SE, "induction"); 1146 1147 // Use this type for pointer arithmetic. 1148 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0); 1149 1150 for (unsigned i = 0; i < NumPointers; ++i) { 1151 Value *Ptr = PtrRtCheck->Pointers[i]; 1152 const SCEV *Sc = SE->getSCEV(Ptr); 1153 1154 if (SE->isLoopInvariant(Sc, OrigLoop)) { 1155 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" << 1156 *Ptr <<"\n"); 1157 Starts.push_back(Ptr); 1158 Ends.push_back(Ptr); 1159 } else { 1160 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n"); 1161 1162 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc); 1163 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc); 1164 Starts.push_back(Start); 1165 Ends.push_back(End); 1166 } 1167 } 1168 1169 IRBuilder<> ChkBuilder(Loc); 1170 1171 for (unsigned i = 0; i < NumPointers; ++i) { 1172 for (unsigned j = i+1; j < NumPointers; ++j) { 1173 // No need to check if two readonly pointers intersect. 1174 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j]) 1175 continue; 1176 1177 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc"); 1178 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc"); 1179 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc"); 1180 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc"); 1181 1182 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0"); 1183 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1"); 1184 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict"); 1185 if (MemoryRuntimeCheck) 1186 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, 1187 "conflict.rdx"); 1188 1189 MemoryRuntimeCheck = cast<Instruction>(IsConflict); 1190 } 1191 } 1192 1193 return MemoryRuntimeCheck; 1194} 1195 1196void 1197InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { 1198 /* 1199 In this function we generate a new loop. The new loop will contain 1200 the vectorized instructions while the old loop will continue to run the 1201 scalar remainder. 1202 1203 [ ] <-- vector loop bypass (may consist of multiple blocks). 1204 / | 1205 / v 1206 | [ ] <-- vector pre header. 1207 | | 1208 | v 1209 | [ ] \ 1210 | [ ]_| <-- vector loop. 1211 | | 1212 \ v 1213 >[ ] <--- middle-block. 1214 / | 1215 / v 1216 | [ ] <--- new preheader. 1217 | | 1218 | v 1219 | [ ] \ 1220 | [ ]_| <-- old scalar loop to handle remainder. 1221 \ | 1222 \ v 1223 >[ ] <-- exit block. 1224 ... 1225 */ 1226 1227 BasicBlock *OldBasicBlock = OrigLoop->getHeader(); 1228 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader(); 1229 BasicBlock *ExitBlock = OrigLoop->getExitBlock(); 1230 assert(ExitBlock && "Must have an exit block"); 1231 1232 // Mark the old scalar loop with metadata that tells us not to vectorize this 1233 // loop again if we run into it. 1234 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), ArrayRef<Value*>()); 1235 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD); 1236 1237 // Some loops have a single integer induction variable, while other loops 1238 // don't. One example is c++ iterators that often have multiple pointer 1239 // induction variables. In the code below we also support a case where we 1240 // don't have a single induction variable. 1241 OldInduction = Legal->getInduction(); 1242 Type *IdxTy = OldInduction ? OldInduction->getType() : 1243 DL->getIntPtrType(SE->getContext()); 1244 1245 // Find the loop boundaries. 1246 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch()); 1247 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count"); 1248 1249 // Get the total trip count from the count by adding 1. 1250 ExitCount = SE->getAddExpr(ExitCount, 1251 SE->getConstant(ExitCount->getType(), 1)); 1252 1253 // Expand the trip count and place the new instructions in the preheader. 1254 // Notice that the pre-header does not change, only the loop body. 1255 SCEVExpander Exp(*SE, "induction"); 1256 1257 // Count holds the overall loop count (N). 1258 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), 1259 BypassBlock->getTerminator()); 1260 1261 // The loop index does not have to start at Zero. Find the original start 1262 // value from the induction PHI node. If we don't have an induction variable 1263 // then we know that it starts at zero. 1264 Value *StartIdx = OldInduction ? 1265 OldInduction->getIncomingValueForBlock(BypassBlock): 1266 ConstantInt::get(IdxTy, 0); 1267 1268 assert(BypassBlock && "Invalid loop structure"); 1269 LoopBypassBlocks.push_back(BypassBlock); 1270 1271 // Split the single block loop into the two loop structure described above. 1272 BasicBlock *VectorPH = 1273 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); 1274 BasicBlock *VecBody = 1275 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body"); 1276 BasicBlock *MiddleBlock = 1277 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block"); 1278 BasicBlock *ScalarPH = 1279 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph"); 1280 1281 // Use this IR builder to create the loop instructions (Phi, Br, Cmp) 1282 // inside the loop. 1283 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1284 1285 // Generate the induction variable. 1286 Induction = Builder.CreatePHI(IdxTy, 2, "index"); 1287 // The loop step is equal to the vectorization factor (num of SIMD elements) 1288 // times the unroll factor (num of SIMD instructions). 1289 Constant *Step = ConstantInt::get(IdxTy, VF * UF); 1290 1291 // This is the IR builder that we use to add all of the logic for bypassing 1292 // the new vector loop. 1293 IRBuilder<> BypassBuilder(BypassBlock->getTerminator()); 1294 1295 // We may need to extend the index in case there is a type mismatch. 1296 // We know that the count starts at zero and does not overflow. 1297 if (Count->getType() != IdxTy) { 1298 // The exit count can be of pointer type. Convert it to the correct 1299 // integer type. 1300 if (ExitCount->getType()->isPointerTy()) 1301 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int"); 1302 else 1303 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast"); 1304 } 1305 1306 // Add the start index to the loop count to get the new end index. 1307 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx"); 1308 1309 // Now we need to generate the expression for N - (N % VF), which is 1310 // the part that the vectorized body will execute. 1311 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf"); 1312 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec"); 1313 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx, 1314 "end.idx.rnd.down"); 1315 1316 // Now, compare the new count to zero. If it is zero skip the vector loop and 1317 // jump to the scalar loop. 1318 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, 1319 "cmp.zero"); 1320 1321 BasicBlock *LastBypassBlock = BypassBlock; 1322 1323 // Generate the code that checks in runtime if arrays overlap. We put the 1324 // checks into a separate block to make the more common case of few elements 1325 // faster. 1326 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal, 1327 BypassBlock->getTerminator()); 1328 if (MemRuntimeCheck) { 1329 // Create a new block containing the memory check. 1330 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck, 1331 "vector.memcheck"); 1332 LoopBypassBlocks.push_back(CheckBlock); 1333 1334 // Replace the branch into the memory check block with a conditional branch 1335 // for the "few elements case". 1336 Instruction *OldTerm = BypassBlock->getTerminator(); 1337 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm); 1338 OldTerm->eraseFromParent(); 1339 1340 Cmp = MemRuntimeCheck; 1341 LastBypassBlock = CheckBlock; 1342 } 1343 1344 LastBypassBlock->getTerminator()->eraseFromParent(); 1345 BranchInst::Create(MiddleBlock, VectorPH, Cmp, 1346 LastBypassBlock); 1347 1348 // We are going to resume the execution of the scalar loop. 1349 // Go over all of the induction variables that we found and fix the 1350 // PHIs that are left in the scalar version of the loop. 1351 // The starting values of PHI nodes depend on the counter of the last 1352 // iteration in the vectorized loop. 1353 // If we come from a bypass edge then we need to start from the original 1354 // start value. 1355 1356 // This variable saves the new starting index for the scalar loop. 1357 PHINode *ResumeIndex = 0; 1358 LoopVectorizationLegality::InductionList::iterator I, E; 1359 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars(); 1360 for (I = List->begin(), E = List->end(); I != E; ++I) { 1361 PHINode *OrigPhi = I->first; 1362 LoopVectorizationLegality::InductionInfo II = I->second; 1363 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val", 1364 MiddleBlock->getTerminator()); 1365 Value *EndValue = 0; 1366 switch (II.IK) { 1367 case LoopVectorizationLegality::IK_NoInduction: 1368 llvm_unreachable("Unknown induction"); 1369 case LoopVectorizationLegality::IK_IntInduction: { 1370 // Handle the integer induction counter: 1371 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type"); 1372 assert(OrigPhi == OldInduction && "Unknown integer PHI"); 1373 // We know what the end value is. 1374 EndValue = IdxEndRoundDown; 1375 // We also know which PHI node holds it. 1376 ResumeIndex = ResumeVal; 1377 break; 1378 } 1379 case LoopVectorizationLegality::IK_ReverseIntInduction: { 1380 // Convert the CountRoundDown variable to the PHI size. 1381 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits(); 1382 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits(); 1383 Value *CRD = CountRoundDown; 1384 if (CRDSize > IISize) 1385 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown, 1386 II.StartValue->getType(), "tr.crd", 1387 LoopBypassBlocks.back()->getTerminator()); 1388 else if (CRDSize < IISize) 1389 CRD = CastInst::Create(Instruction::SExt, CountRoundDown, 1390 II.StartValue->getType(), 1391 "sext.crd", 1392 LoopBypassBlocks.back()->getTerminator()); 1393 // Handle reverse integer induction counter: 1394 EndValue = 1395 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end", 1396 LoopBypassBlocks.back()->getTerminator()); 1397 break; 1398 } 1399 case LoopVectorizationLegality::IK_PtrInduction: { 1400 // For pointer induction variables, calculate the offset using 1401 // the end index. 1402 EndValue = 1403 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end", 1404 LoopBypassBlocks.back()->getTerminator()); 1405 break; 1406 } 1407 case LoopVectorizationLegality::IK_ReversePtrInduction: { 1408 // The value at the end of the loop for the reverse pointer is calculated 1409 // by creating a GEP with a negative index starting from the start value. 1410 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0); 1411 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown, 1412 "rev.ind.end", 1413 LoopBypassBlocks.back()->getTerminator()); 1414 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx, 1415 "rev.ptr.ind.end", 1416 LoopBypassBlocks.back()->getTerminator()); 1417 break; 1418 } 1419 }// end of case 1420 1421 // The new PHI merges the original incoming value, in case of a bypass, 1422 // or the value at the end of the vectorized loop. 1423 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1424 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]); 1425 ResumeVal->addIncoming(EndValue, VecBody); 1426 1427 // Fix the scalar body counter (PHI node). 1428 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH); 1429 OrigPhi->setIncomingValue(BlockIdx, ResumeVal); 1430 } 1431 1432 // If we are generating a new induction variable then we also need to 1433 // generate the code that calculates the exit value. This value is not 1434 // simply the end of the counter because we may skip the vectorized body 1435 // in case of a runtime check. 1436 if (!OldInduction){ 1437 assert(!ResumeIndex && "Unexpected resume value found"); 1438 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val", 1439 MiddleBlock->getTerminator()); 1440 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1441 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]); 1442 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); 1443 } 1444 1445 // Make sure that we found the index where scalar loop needs to continue. 1446 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() && 1447 "Invalid resume Index"); 1448 1449 // Add a check in the middle block to see if we have completed 1450 // all of the iterations in the first vector loop. 1451 // If (N - N%VF) == N, then we *don't* need to run the remainder. 1452 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, 1453 ResumeIndex, "cmp.n", 1454 MiddleBlock->getTerminator()); 1455 1456 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); 1457 // Remove the old terminator. 1458 MiddleBlock->getTerminator()->eraseFromParent(); 1459 1460 // Create i+1 and fill the PHINode. 1461 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); 1462 Induction->addIncoming(StartIdx, VectorPH); 1463 Induction->addIncoming(NextIdx, VecBody); 1464 // Create the compare. 1465 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); 1466 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); 1467 1468 // Now we have two terminators. Remove the old one from the block. 1469 VecBody->getTerminator()->eraseFromParent(); 1470 1471 // Get ready to start creating new instructions into the vectorized body. 1472 Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); 1473 1474 // Create and register the new vector loop. 1475 Loop* Lp = new Loop(); 1476 Loop *ParentLoop = OrigLoop->getParentLoop(); 1477 1478 // Insert the new loop into the loop nest and register the new basic blocks. 1479 if (ParentLoop) { 1480 ParentLoop->addChildLoop(Lp); 1481 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 1482 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase()); 1483 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); 1484 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); 1485 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); 1486 } else { 1487 LI->addTopLevelLoop(Lp); 1488 } 1489 1490 Lp->addBasicBlockToLoop(VecBody, LI->getBase()); 1491 1492 // Save the state. 1493 LoopVectorPreHeader = VectorPH; 1494 LoopScalarPreHeader = ScalarPH; 1495 LoopMiddleBlock = MiddleBlock; 1496 LoopExitBlock = ExitBlock; 1497 LoopVectorBody = VecBody; 1498 LoopScalarBody = OldBasicBlock; 1499} 1500 1501/// This function returns the identity element (or neutral element) for 1502/// the operation K. 1503Constant* 1504LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp, 1505 MinMaxReductionKind MinMaxK) { 1506 switch (K) { 1507 case RK_IntegerXor: 1508 case RK_IntegerAdd: 1509 case RK_IntegerOr: 1510 // Adding, Xoring, Oring zero to a number does not change it. 1511 return ConstantInt::get(Tp, 0); 1512 case RK_IntegerMult: 1513 // Multiplying a number by 1 does not change it. 1514 return ConstantInt::get(Tp, 1); 1515 case RK_IntegerAnd: 1516 // AND-ing a number with an all-1 value does not change it. 1517 return ConstantInt::get(Tp, -1, true); 1518 case RK_FloatMult: 1519 // Multiplying a number by 1 does not change it. 1520 return ConstantFP::get(Tp, 1.0L); 1521 case RK_FloatAdd: 1522 // Adding zero to a number does not change it. 1523 return ConstantFP::get(Tp, 0.0L); 1524 case RK_IntegerMinMax: 1525 switch(MinMaxK) { 1526 default: llvm_unreachable("Unknown min/max predicate"); 1527 case MRK_UIntMin: 1528 return ConstantInt::getAllOnesValue(Tp); 1529 case MRK_UIntMax: 1530 return ConstantInt::get(Tp, 0); 1531 case MRK_SIntMin: { 1532 unsigned BitWidth = Tp->getPrimitiveSizeInBits(); 1533 return ConstantInt::get(Tp->getContext(), 1534 APInt::getSignedMaxValue(BitWidth)); 1535 } 1536 case LoopVectorizationLegality::MRK_SIntMax: { 1537 unsigned BitWidth = Tp->getPrimitiveSizeInBits(); 1538 return ConstantInt::get(Tp->getContext(), 1539 APInt::getSignedMinValue(BitWidth)); 1540 } 1541 } 1542 default: 1543 llvm_unreachable("Unknown reduction kind"); 1544 } 1545} 1546 1547static Intrinsic::ID 1548getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) { 1549 // If we have an intrinsic call, check if it is trivially vectorizable. 1550 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) { 1551 switch (II->getIntrinsicID()) { 1552 case Intrinsic::sqrt: 1553 case Intrinsic::sin: 1554 case Intrinsic::cos: 1555 case Intrinsic::exp: 1556 case Intrinsic::exp2: 1557 case Intrinsic::log: 1558 case Intrinsic::log10: 1559 case Intrinsic::log2: 1560 case Intrinsic::fabs: 1561 case Intrinsic::floor: 1562 case Intrinsic::ceil: 1563 case Intrinsic::trunc: 1564 case Intrinsic::rint: 1565 case Intrinsic::nearbyint: 1566 case Intrinsic::pow: 1567 case Intrinsic::fma: 1568 case Intrinsic::fmuladd: 1569 return II->getIntrinsicID(); 1570 default: 1571 return Intrinsic::not_intrinsic; 1572 } 1573 } 1574 1575 if (!TLI) 1576 return Intrinsic::not_intrinsic; 1577 1578 LibFunc::Func Func; 1579 Function *F = CI->getCalledFunction(); 1580 // We're going to make assumptions on the semantics of the functions, check 1581 // that the target knows that it's available in this environment. 1582 if (!F || !TLI->getLibFunc(F->getName(), Func)) 1583 return Intrinsic::not_intrinsic; 1584 1585 // Otherwise check if we have a call to a function that can be turned into a 1586 // vector intrinsic. 1587 switch (Func) { 1588 default: 1589 break; 1590 case LibFunc::sin: 1591 case LibFunc::sinf: 1592 case LibFunc::sinl: 1593 return Intrinsic::sin; 1594 case LibFunc::cos: 1595 case LibFunc::cosf: 1596 case LibFunc::cosl: 1597 return Intrinsic::cos; 1598 case LibFunc::exp: 1599 case LibFunc::expf: 1600 case LibFunc::expl: 1601 return Intrinsic::exp; 1602 case LibFunc::exp2: 1603 case LibFunc::exp2f: 1604 case LibFunc::exp2l: 1605 return Intrinsic::exp2; 1606 case LibFunc::log: 1607 case LibFunc::logf: 1608 case LibFunc::logl: 1609 return Intrinsic::log; 1610 case LibFunc::log10: 1611 case LibFunc::log10f: 1612 case LibFunc::log10l: 1613 return Intrinsic::log10; 1614 case LibFunc::log2: 1615 case LibFunc::log2f: 1616 case LibFunc::log2l: 1617 return Intrinsic::log2; 1618 case LibFunc::fabs: 1619 case LibFunc::fabsf: 1620 case LibFunc::fabsl: 1621 return Intrinsic::fabs; 1622 case LibFunc::floor: 1623 case LibFunc::floorf: 1624 case LibFunc::floorl: 1625 return Intrinsic::floor; 1626 case LibFunc::ceil: 1627 case LibFunc::ceilf: 1628 case LibFunc::ceill: 1629 return Intrinsic::ceil; 1630 case LibFunc::trunc: 1631 case LibFunc::truncf: 1632 case LibFunc::truncl: 1633 return Intrinsic::trunc; 1634 case LibFunc::rint: 1635 case LibFunc::rintf: 1636 case LibFunc::rintl: 1637 return Intrinsic::rint; 1638 case LibFunc::nearbyint: 1639 case LibFunc::nearbyintf: 1640 case LibFunc::nearbyintl: 1641 return Intrinsic::nearbyint; 1642 case LibFunc::pow: 1643 case LibFunc::powf: 1644 case LibFunc::powl: 1645 return Intrinsic::pow; 1646 } 1647 1648 return Intrinsic::not_intrinsic; 1649} 1650 1651/// This function translates the reduction kind to an LLVM binary operator. 1652static unsigned 1653getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) { 1654 switch (Kind) { 1655 case LoopVectorizationLegality::RK_IntegerAdd: 1656 return Instruction::Add; 1657 case LoopVectorizationLegality::RK_IntegerMult: 1658 return Instruction::Mul; 1659 case LoopVectorizationLegality::RK_IntegerOr: 1660 return Instruction::Or; 1661 case LoopVectorizationLegality::RK_IntegerAnd: 1662 return Instruction::And; 1663 case LoopVectorizationLegality::RK_IntegerXor: 1664 return Instruction::Xor; 1665 case LoopVectorizationLegality::RK_FloatMult: 1666 return Instruction::FMul; 1667 case LoopVectorizationLegality::RK_FloatAdd: 1668 return Instruction::FAdd; 1669 case LoopVectorizationLegality::RK_IntegerMinMax: 1670 return Instruction::ICmp; 1671 default: 1672 llvm_unreachable("Unknown reduction operation"); 1673 } 1674} 1675 1676Value *createMinMaxOp(IRBuilder<> &Builder, 1677 LoopVectorizationLegality::MinMaxReductionKind RK, 1678 Value *Left, 1679 Value *Right) { 1680 CmpInst::Predicate P = CmpInst::ICMP_NE; 1681 switch (RK) { 1682 default: 1683 llvm_unreachable("Unknown min/max reduction kind"); 1684 case LoopVectorizationLegality::MRK_UIntMin: 1685 P = CmpInst::ICMP_ULT; 1686 break; 1687 case LoopVectorizationLegality::MRK_UIntMax: 1688 P = CmpInst::ICMP_UGT; 1689 break; 1690 case LoopVectorizationLegality::MRK_SIntMin: 1691 P = CmpInst::ICMP_SLT; 1692 break; 1693 case LoopVectorizationLegality::MRK_SIntMax: 1694 P = CmpInst::ICMP_SGT; 1695 } 1696 Value *Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); 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 Constant *Iden = 1765 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind, 1766 VecTy->getScalarType(), 1767 RdxDesc.MinMaxKind); 1768 Constant *Identity = ConstantVector::getSplat(VF, Iden); 1769 1770 // This vector is the Identity vector where the first element is the 1771 // incoming scalar reduction. 1772 Value *VectorStart = Builder.CreateInsertElement(Identity, 1773 RdxDesc.StartValue, Zero); 1774 1775 // Fix the vector-loop phi. 1776 // We created the induction variable so we know that the 1777 // preheader is the first entry. 1778 BasicBlock *VecPreheader = Induction->getIncomingBlock(0); 1779 1780 // Reductions do not have to start at zero. They can start with 1781 // any loop invariant values. 1782 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi); 1783 BasicBlock *Latch = OrigLoop->getLoopLatch(); 1784 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch); 1785 VectorParts &Val = getVectorValue(LoopVal); 1786 for (unsigned part = 0; part < UF; ++part) { 1787 // Make sure to add the reduction stat value only to the 1788 // first unroll part. 1789 Value *StartVal = (part == 0) ? VectorStart : Identity; 1790 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader); 1791 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody); 1792 } 1793 1794 // Before each round, move the insertion point right between 1795 // the PHIs and the values we are going to write. 1796 // This allows us to write both PHINodes and the extractelement 1797 // instructions. 1798 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); 1799 1800 VectorParts RdxParts; 1801 for (unsigned part = 0; part < UF; ++part) { 1802 // This PHINode contains the vectorized reduction variable, or 1803 // the initial value vector, if we bypass the vector loop. 1804 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr); 1805 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); 1806 Value *StartVal = (part == 0) ? VectorStart : Identity; 1807 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) 1808 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]); 1809 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody); 1810 RdxParts.push_back(NewPhi); 1811 } 1812 1813 // Reduce all of the unrolled parts into a single vector. 1814 Value *ReducedPartRdx = RdxParts[0]; 1815 unsigned Op = getReductionBinOp(RdxDesc.Kind); 1816 for (unsigned part = 1; part < UF; ++part) { 1817 if (Op != Instruction::ICmp) 1818 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op, 1819 RdxParts[part], ReducedPartRdx, 1820 "bin.rdx"); 1821 else 1822 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind, 1823 ReducedPartRdx, RdxParts[part]); 1824 } 1825 1826 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 1827 // and vector ops, reducing the set of values being computed by half each 1828 // round. 1829 assert(isPowerOf2_32(VF) && 1830 "Reduction emission only supported for pow2 vectors!"); 1831 Value *TmpVec = ReducedPartRdx; 1832 SmallVector<Constant*, 32> ShuffleMask(VF, 0); 1833 for (unsigned i = VF; i != 1; i >>= 1) { 1834 // Move the upper half of the vector to the lower half. 1835 for (unsigned j = 0; j != i/2; ++j) 1836 ShuffleMask[j] = Builder.getInt32(i/2 + j); 1837 1838 // Fill the rest of the mask with undef. 1839 std::fill(&ShuffleMask[i/2], ShuffleMask.end(), 1840 UndefValue::get(Builder.getInt32Ty())); 1841 1842 Value *Shuf = 1843 Builder.CreateShuffleVector(TmpVec, 1844 UndefValue::get(TmpVec->getType()), 1845 ConstantVector::get(ShuffleMask), 1846 "rdx.shuf"); 1847 1848 if (Op != Instruction::ICmp) 1849 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf, 1850 "bin.rdx"); 1851 else 1852 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf); 1853 } 1854 1855 // The result is in the first element of the vector. 1856 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); 1857 1858 // Now, we need to fix the users of the reduction variable 1859 // inside and outside of the scalar remainder loop. 1860 // We know that the loop is in LCSSA form. We need to update the 1861 // PHI nodes in the exit blocks. 1862 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 1863 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 1864 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 1865 if (!LCSSAPhi) continue; 1866 1867 // All PHINodes need to have a single entry edge, or two if 1868 // we already fixed them. 1869 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); 1870 1871 // We found our reduction value exit-PHI. Update it with the 1872 // incoming bypass edge. 1873 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { 1874 // Add an edge coming from the bypass. 1875 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock); 1876 break; 1877 } 1878 }// end of the LCSSA phi scan. 1879 1880 // Fix the scalar loop reduction variable with the incoming reduction sum 1881 // from the vector body and from the backedge value. 1882 int IncomingEdgeBlockIdx = 1883 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch()); 1884 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index"); 1885 // Pick the other block. 1886 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); 1887 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0); 1888 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); 1889 }// end of for each redux variable. 1890 1891 // The Loop exit block may have single value PHI nodes where the incoming 1892 // value is 'undef'. While vectorizing we only handled real values that 1893 // were defined inside the loop. Here we handle the 'undef case'. 1894 // See PR14725. 1895 for (BasicBlock::iterator LEI = LoopExitBlock->begin(), 1896 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { 1897 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); 1898 if (!LCSSAPhi) continue; 1899 if (LCSSAPhi->getNumIncomingValues() == 1) 1900 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()), 1901 LoopMiddleBlock); 1902 } 1903} 1904 1905InnerLoopVectorizer::VectorParts 1906InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { 1907 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && 1908 "Invalid edge"); 1909 1910 VectorParts SrcMask = createBlockInMask(Src); 1911 1912 // The terminator has to be a branch inst! 1913 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); 1914 assert(BI && "Unexpected terminator found"); 1915 1916 if (BI->isConditional()) { 1917 VectorParts EdgeMask = getVectorValue(BI->getCondition()); 1918 1919 if (BI->getSuccessor(0) != Dst) 1920 for (unsigned part = 0; part < UF; ++part) 1921 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]); 1922 1923 for (unsigned part = 0; part < UF; ++part) 1924 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]); 1925 return EdgeMask; 1926 } 1927 1928 return SrcMask; 1929} 1930 1931InnerLoopVectorizer::VectorParts 1932InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { 1933 assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); 1934 1935 // Loop incoming mask is all-one. 1936 if (OrigLoop->getHeader() == BB) { 1937 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1); 1938 return getVectorValue(C); 1939 } 1940 1941 // This is the block mask. We OR all incoming edges, and with zero. 1942 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); 1943 VectorParts BlockMask = getVectorValue(Zero); 1944 1945 // For each pred: 1946 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) { 1947 VectorParts EM = createEdgeMask(*it, BB); 1948 for (unsigned part = 0; part < UF; ++part) 1949 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]); 1950 } 1951 1952 return BlockMask; 1953} 1954 1955void 1956InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal, 1957 BasicBlock *BB, PhiVector *PV) { 1958 // For each instruction in the old loop. 1959 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 1960 VectorParts &Entry = WidenMap.get(it); 1961 switch (it->getOpcode()) { 1962 case Instruction::Br: 1963 // Nothing to do for PHIs and BR, since we already took care of the 1964 // loop control flow instructions. 1965 continue; 1966 case Instruction::PHI:{ 1967 PHINode* P = cast<PHINode>(it); 1968 // Handle reduction variables: 1969 if (Legal->getReductionVars()->count(P)) { 1970 for (unsigned part = 0; part < UF; ++part) { 1971 // This is phase one of vectorizing PHIs. 1972 Type *VecTy = VectorType::get(it->getType(), VF); 1973 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi", 1974 LoopVectorBody-> getFirstInsertionPt()); 1975 } 1976 PV->push_back(P); 1977 continue; 1978 } 1979 1980 // Check for PHI nodes that are lowered to vector selects. 1981 if (P->getParent() != OrigLoop->getHeader()) { 1982 // We know that all PHIs in non header blocks are converted into 1983 // selects, so we don't have to worry about the insertion order and we 1984 // can just use the builder. 1985 // At this point we generate the predication tree. There may be 1986 // duplications since this is a simple recursive scan, but future 1987 // optimizations will clean it up. 1988 1989 unsigned NumIncoming = P->getNumIncomingValues(); 1990 assert(NumIncoming > 1 && "Invalid PHI"); 1991 1992 // Generate a sequence of selects of the form: 1993 // SELECT(Mask3, In3, 1994 // SELECT(Mask2, In2, 1995 // ( ...))) 1996 for (unsigned In = 0; In < NumIncoming; In++) { 1997 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In), 1998 P->getParent()); 1999 VectorParts &In0 = getVectorValue(P->getIncomingValue(In)); 2000 2001 for (unsigned part = 0; part < UF; ++part) { 2002 // We don't need to 'select' the first PHI operand because it is 2003 // the default value if all of the other masks don't match. 2004 if (In == 0) 2005 Entry[part] = In0[part]; 2006 else 2007 // Select between the current value and the previous incoming edge 2008 // based on the incoming mask. 2009 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], 2010 Entry[part], "predphi"); 2011 } 2012 } 2013 continue; 2014 } 2015 2016 // This PHINode must be an induction variable. 2017 // Make sure that we know about it. 2018 assert(Legal->getInductionVars()->count(P) && 2019 "Not an induction variable"); 2020 2021 LoopVectorizationLegality::InductionInfo II = 2022 Legal->getInductionVars()->lookup(P); 2023 2024 switch (II.IK) { 2025 case LoopVectorizationLegality::IK_NoInduction: 2026 llvm_unreachable("Unknown induction"); 2027 case LoopVectorizationLegality::IK_IntInduction: { 2028 assert(P == OldInduction && "Unexpected PHI"); 2029 Value *Broadcasted = getBroadcastInstrs(Induction); 2030 // After broadcasting the induction variable we need to make the 2031 // vector consecutive by adding 0, 1, 2 ... 2032 for (unsigned part = 0; part < UF; ++part) 2033 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false); 2034 continue; 2035 } 2036 case LoopVectorizationLegality::IK_ReverseIntInduction: 2037 case LoopVectorizationLegality::IK_PtrInduction: 2038 case LoopVectorizationLegality::IK_ReversePtrInduction: 2039 // Handle reverse integer and pointer inductions. 2040 Value *StartIdx = 0; 2041 // If we have a single integer induction variable then use it. 2042 // Otherwise, start counting at zero. 2043 if (OldInduction) { 2044 LoopVectorizationLegality::InductionInfo OldII = 2045 Legal->getInductionVars()->lookup(OldInduction); 2046 StartIdx = OldII.StartValue; 2047 } else { 2048 StartIdx = ConstantInt::get(Induction->getType(), 0); 2049 } 2050 // This is the normalized GEP that starts counting at zero. 2051 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx, 2052 "normalized.idx"); 2053 2054 // Handle the reverse integer induction variable case. 2055 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) { 2056 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType()); 2057 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy, 2058 "resize.norm.idx"); 2059 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI, 2060 "reverse.idx"); 2061 2062 // This is a new value so do not hoist it out. 2063 Value *Broadcasted = getBroadcastInstrs(ReverseInd); 2064 // After broadcasting the induction variable we need to make the 2065 // vector consecutive by adding ... -3, -2, -1, 0. 2066 for (unsigned part = 0; part < UF; ++part) 2067 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true); 2068 continue; 2069 } 2070 2071 // Handle the pointer induction variable case. 2072 assert(P->getType()->isPointerTy() && "Unexpected type."); 2073 2074 // Is this a reverse induction ptr or a consecutive induction ptr. 2075 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction == 2076 II.IK); 2077 2078 // This is the vector of results. Notice that we don't generate 2079 // vector geps because scalar geps result in better code. 2080 for (unsigned part = 0; part < UF; ++part) { 2081 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); 2082 for (unsigned int i = 0; i < VF; ++i) { 2083 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1); 2084 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex); 2085 Value *GlobalIdx; 2086 if (!Reverse) 2087 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx"); 2088 else 2089 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx"); 2090 2091 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, 2092 "next.gep"); 2093 VecVal = Builder.CreateInsertElement(VecVal, SclrGep, 2094 Builder.getInt32(i), 2095 "insert.gep"); 2096 } 2097 Entry[part] = VecVal; 2098 } 2099 continue; 2100 } 2101 2102 }// End of PHI. 2103 2104 case Instruction::Add: 2105 case Instruction::FAdd: 2106 case Instruction::Sub: 2107 case Instruction::FSub: 2108 case Instruction::Mul: 2109 case Instruction::FMul: 2110 case Instruction::UDiv: 2111 case Instruction::SDiv: 2112 case Instruction::FDiv: 2113 case Instruction::URem: 2114 case Instruction::SRem: 2115 case Instruction::FRem: 2116 case Instruction::Shl: 2117 case Instruction::LShr: 2118 case Instruction::AShr: 2119 case Instruction::And: 2120 case Instruction::Or: 2121 case Instruction::Xor: { 2122 // Just widen binops. 2123 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); 2124 VectorParts &A = getVectorValue(it->getOperand(0)); 2125 VectorParts &B = getVectorValue(it->getOperand(1)); 2126 2127 // Use this vector value for all users of the original instruction. 2128 for (unsigned Part = 0; Part < UF; ++Part) { 2129 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); 2130 2131 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef. 2132 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V); 2133 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) { 2134 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); 2135 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); 2136 } 2137 if (VecOp && isa<PossiblyExactOperator>(VecOp)) 2138 VecOp->setIsExact(BinOp->isExact()); 2139 2140 Entry[Part] = V; 2141 } 2142 break; 2143 } 2144 case Instruction::Select: { 2145 // Widen selects. 2146 // If the selector is loop invariant we can create a select 2147 // instruction with a scalar condition. Otherwise, use vector-select. 2148 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)), 2149 OrigLoop); 2150 2151 // The condition can be loop invariant but still defined inside the 2152 // loop. This means that we can't just use the original 'cond' value. 2153 // We have to take the 'vectorized' value and pick the first lane. 2154 // Instcombine will make this a no-op. 2155 VectorParts &Cond = getVectorValue(it->getOperand(0)); 2156 VectorParts &Op0 = getVectorValue(it->getOperand(1)); 2157 VectorParts &Op1 = getVectorValue(it->getOperand(2)); 2158 Value *ScalarCond = Builder.CreateExtractElement(Cond[0], 2159 Builder.getInt32(0)); 2160 for (unsigned Part = 0; Part < UF; ++Part) { 2161 Entry[Part] = Builder.CreateSelect( 2162 InvariantCond ? ScalarCond : Cond[Part], 2163 Op0[Part], 2164 Op1[Part]); 2165 } 2166 break; 2167 } 2168 2169 case Instruction::ICmp: 2170 case Instruction::FCmp: { 2171 // Widen compares. Generate vector compares. 2172 bool FCmp = (it->getOpcode() == Instruction::FCmp); 2173 CmpInst *Cmp = dyn_cast<CmpInst>(it); 2174 VectorParts &A = getVectorValue(it->getOperand(0)); 2175 VectorParts &B = getVectorValue(it->getOperand(1)); 2176 for (unsigned Part = 0; Part < UF; ++Part) { 2177 Value *C = 0; 2178 if (FCmp) 2179 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]); 2180 else 2181 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]); 2182 Entry[Part] = C; 2183 } 2184 break; 2185 } 2186 2187 case Instruction::Store: 2188 case Instruction::Load: 2189 vectorizeMemoryInstruction(it, Legal); 2190 break; 2191 case Instruction::ZExt: 2192 case Instruction::SExt: 2193 case Instruction::FPToUI: 2194 case Instruction::FPToSI: 2195 case Instruction::FPExt: 2196 case Instruction::PtrToInt: 2197 case Instruction::IntToPtr: 2198 case Instruction::SIToFP: 2199 case Instruction::UIToFP: 2200 case Instruction::Trunc: 2201 case Instruction::FPTrunc: 2202 case Instruction::BitCast: { 2203 CastInst *CI = dyn_cast<CastInst>(it); 2204 /// Optimize the special case where the source is the induction 2205 /// variable. Notice that we can only optimize the 'trunc' case 2206 /// because: a. FP conversions lose precision, b. sext/zext may wrap, 2207 /// c. other casts depend on pointer size. 2208 if (CI->getOperand(0) == OldInduction && 2209 it->getOpcode() == Instruction::Trunc) { 2210 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, 2211 CI->getType()); 2212 Value *Broadcasted = getBroadcastInstrs(ScalarCast); 2213 for (unsigned Part = 0; Part < UF; ++Part) 2214 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false); 2215 break; 2216 } 2217 /// Vectorize casts. 2218 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF); 2219 2220 VectorParts &A = getVectorValue(it->getOperand(0)); 2221 for (unsigned Part = 0; Part < UF; ++Part) 2222 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy); 2223 break; 2224 } 2225 2226 case Instruction::Call: { 2227 // Ignore dbg intrinsics. 2228 if (isa<DbgInfoIntrinsic>(it)) 2229 break; 2230 2231 Module *M = BB->getParent()->getParent(); 2232 CallInst *CI = cast<CallInst>(it); 2233 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 2234 assert(ID && "Not an intrinsic call!"); 2235 for (unsigned Part = 0; Part < UF; ++Part) { 2236 SmallVector<Value*, 4> Args; 2237 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) { 2238 VectorParts &Arg = getVectorValue(CI->getArgOperand(i)); 2239 Args.push_back(Arg[Part]); 2240 } 2241 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) }; 2242 Function *F = Intrinsic::getDeclaration(M, ID, Tys); 2243 Entry[Part] = Builder.CreateCall(F, Args); 2244 } 2245 break; 2246 } 2247 2248 default: 2249 // All other instructions are unsupported. Scalarize them. 2250 scalarizeInstruction(it); 2251 break; 2252 }// end of switch. 2253 }// end of for_each instr. 2254} 2255 2256void InnerLoopVectorizer::updateAnalysis() { 2257 // Forget the original basic block. 2258 SE->forgetLoop(OrigLoop); 2259 2260 // Update the dominator tree information. 2261 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) && 2262 "Entry does not dominate exit."); 2263 2264 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) 2265 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]); 2266 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back()); 2267 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); 2268 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front()); 2269 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); 2270 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); 2271 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); 2272 2273 DEBUG(DT->verifyAnalysis()); 2274} 2275 2276bool LoopVectorizationLegality::canVectorizeWithIfConvert() { 2277 if (!EnableIfConversion) 2278 return false; 2279 2280 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable"); 2281 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector(); 2282 2283 // Collect the blocks that need predication. 2284 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) { 2285 BasicBlock *BB = LoopBlocks[i]; 2286 2287 // We don't support switch statements inside loops. 2288 if (!isa<BranchInst>(BB->getTerminator())) 2289 return false; 2290 2291 // We must be able to predicate all blocks that need to be predicated. 2292 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB)) 2293 return false; 2294 } 2295 2296 // We can if-convert this loop. 2297 return true; 2298} 2299 2300bool LoopVectorizationLegality::canVectorize() { 2301 assert(TheLoop->getLoopPreheader() && "No preheader!!"); 2302 2303 // We can only vectorize innermost loops. 2304 if (TheLoop->getSubLoopsVector().size()) 2305 return false; 2306 2307 // We must have a single backedge. 2308 if (TheLoop->getNumBackEdges() != 1) 2309 return false; 2310 2311 // We must have a single exiting block. 2312 if (!TheLoop->getExitingBlock()) 2313 return false; 2314 2315 unsigned NumBlocks = TheLoop->getNumBlocks(); 2316 2317 // Check if we can if-convert non single-bb loops. 2318 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { 2319 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n"); 2320 return false; 2321 } 2322 2323 // We need to have a loop header. 2324 BasicBlock *Latch = TheLoop->getLoopLatch(); 2325 DEBUG(dbgs() << "LV: Found a loop: " << 2326 TheLoop->getHeader()->getName() << "\n"); 2327 2328 // ScalarEvolution needs to be able to find the exit count. 2329 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch); 2330 if (ExitCount == SE->getCouldNotCompute()) { 2331 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); 2332 return false; 2333 } 2334 2335 // Do not loop-vectorize loops with a tiny trip count. 2336 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch); 2337 if (TC > 0u && TC < TinyTripCountVectorThreshold) { 2338 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << 2339 "This loop is not worth vectorizing.\n"); 2340 return false; 2341 } 2342 2343 // Check if we can vectorize the instructions and CFG in this loop. 2344 if (!canVectorizeInstrs()) { 2345 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n"); 2346 return false; 2347 } 2348 2349 // Go over each instruction and look at memory deps. 2350 if (!canVectorizeMemory()) { 2351 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n"); 2352 return false; 2353 } 2354 2355 // Collect all of the variables that remain uniform after vectorization. 2356 collectLoopUniforms(); 2357 2358 DEBUG(dbgs() << "LV: We can vectorize this loop" << 2359 (PtrRtCheck.Need ? " (with a runtime bound check)" : "") 2360 <<"!\n"); 2361 2362 // Okay! We can vectorize. At this point we don't have any other mem analysis 2363 // which may limit our maximum vectorization factor, so just return true with 2364 // no restrictions. 2365 return true; 2366} 2367 2368bool LoopVectorizationLegality::canVectorizeInstrs() { 2369 BasicBlock *PreHeader = TheLoop->getLoopPreheader(); 2370 BasicBlock *Header = TheLoop->getHeader(); 2371 2372 // If we marked the scalar loop as "already vectorized" then no need 2373 // to vectorize it again. 2374 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) { 2375 DEBUG(dbgs() << "LV: This loop was vectorized before\n"); 2376 return false; 2377 } 2378 2379 // For each block in the loop. 2380 for (Loop::block_iterator bb = TheLoop->block_begin(), 2381 be = TheLoop->block_end(); bb != be; ++bb) { 2382 2383 // Scan the instructions in the block and look for hazards. 2384 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2385 ++it) { 2386 2387 if (PHINode *Phi = dyn_cast<PHINode>(it)) { 2388 // Check that this PHI type is allowed. 2389 if (!Phi->getType()->isIntegerTy() && 2390 !Phi->getType()->isFloatingPointTy() && 2391 !Phi->getType()->isPointerTy()) { 2392 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n"); 2393 return false; 2394 } 2395 2396 // If this PHINode is not in the header block, then we know that we 2397 // can convert it to select during if-conversion. No need to check if 2398 // the PHIs in this block are induction or reduction variables. 2399 if (*bb != Header) 2400 continue; 2401 2402 // We only allow if-converted PHIs with more than two incoming values. 2403 if (Phi->getNumIncomingValues() != 2) { 2404 DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); 2405 return false; 2406 } 2407 2408 // This is the value coming from the preheader. 2409 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader); 2410 // Check if this is an induction variable. 2411 InductionKind IK = isInductionVariable(Phi); 2412 2413 if (IK_NoInduction != IK) { 2414 // Int inductions are special because we only allow one IV. 2415 if (IK == IK_IntInduction) { 2416 if (Induction) { 2417 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n"); 2418 return false; 2419 } 2420 Induction = Phi; 2421 } 2422 2423 DEBUG(dbgs() << "LV: Found an induction variable.\n"); 2424 Inductions[Phi] = InductionInfo(StartValue, IK); 2425 continue; 2426 } 2427 2428 if (AddReductionVar(Phi, RK_IntegerAdd)) { 2429 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); 2430 continue; 2431 } 2432 if (AddReductionVar(Phi, RK_IntegerMult)) { 2433 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); 2434 continue; 2435 } 2436 if (AddReductionVar(Phi, RK_IntegerOr)) { 2437 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); 2438 continue; 2439 } 2440 if (AddReductionVar(Phi, RK_IntegerAnd)) { 2441 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); 2442 continue; 2443 } 2444 if (AddReductionVar(Phi, RK_IntegerXor)) { 2445 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); 2446 continue; 2447 } 2448 if (AddReductionVar(Phi, RK_IntegerMinMax)) { 2449 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n"); 2450 continue; 2451 } 2452 if (AddReductionVar(Phi, RK_FloatMult)) { 2453 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n"); 2454 continue; 2455 } 2456 if (AddReductionVar(Phi, RK_FloatAdd)) { 2457 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n"); 2458 continue; 2459 } 2460 2461 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); 2462 return false; 2463 }// end of PHI handling 2464 2465 // We still don't handle functions. However, we can ignore dbg intrinsic 2466 // calls and we do handle certain intrinsic and libm functions. 2467 CallInst *CI = dyn_cast<CallInst>(it); 2468 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) { 2469 DEBUG(dbgs() << "LV: Found a call site.\n"); 2470 return false; 2471 } 2472 2473 // Check that the instruction return type is vectorizable. 2474 if (!VectorType::isValidElementType(it->getType()) && 2475 !it->getType()->isVoidTy()) { 2476 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n"); 2477 return false; 2478 } 2479 2480 // Check that the stored type is vectorizable. 2481 if (StoreInst *ST = dyn_cast<StoreInst>(it)) { 2482 Type *T = ST->getValueOperand()->getType(); 2483 if (!VectorType::isValidElementType(T)) 2484 return false; 2485 } 2486 2487 // Reduction instructions are allowed to have exit users. 2488 // All other instructions must not have external users. 2489 if (!AllowedExit.count(it)) 2490 //Check that all of the users of the loop are inside the BB. 2491 for (Value::use_iterator I = it->use_begin(), E = it->use_end(); 2492 I != E; ++I) { 2493 Instruction *U = cast<Instruction>(*I); 2494 // This user may be a reduction exit value. 2495 if (!TheLoop->contains(U)) { 2496 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n"); 2497 return false; 2498 } 2499 } 2500 } // next instr. 2501 2502 } 2503 2504 if (!Induction) { 2505 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n"); 2506 assert(getInductionVars()->size() && "No induction variables"); 2507 } 2508 2509 return true; 2510} 2511 2512void LoopVectorizationLegality::collectLoopUniforms() { 2513 // We now know that the loop is vectorizable! 2514 // Collect variables that will remain uniform after vectorization. 2515 std::vector<Value*> Worklist; 2516 BasicBlock *Latch = TheLoop->getLoopLatch(); 2517 2518 // Start with the conditional branch and walk up the block. 2519 Worklist.push_back(Latch->getTerminator()->getOperand(0)); 2520 2521 while (Worklist.size()) { 2522 Instruction *I = dyn_cast<Instruction>(Worklist.back()); 2523 Worklist.pop_back(); 2524 2525 // Look at instructions inside this loop. 2526 // Stop when reaching PHI nodes. 2527 // TODO: we need to follow values all over the loop, not only in this block. 2528 if (!I || !TheLoop->contains(I) || isa<PHINode>(I)) 2529 continue; 2530 2531 // This is a known uniform. 2532 Uniforms.insert(I); 2533 2534 // Insert all operands. 2535 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) { 2536 Worklist.push_back(I->getOperand(i)); 2537 } 2538 } 2539} 2540 2541AliasAnalysis::Location 2542LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) { 2543 if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) 2544 return AA->getLocation(Store); 2545 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) 2546 return AA->getLocation(Load); 2547 2548 llvm_unreachable("Should be either load or store instruction"); 2549} 2550 2551bool 2552LoopVectorizationLegality::hasPossibleGlobalWriteReorder( 2553 Value *Object, 2554 Instruction *Inst, 2555 AliasMultiMap& WriteObjects, 2556 unsigned MaxByteWidth) { 2557 2558 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst); 2559 2560 std::vector<Instruction*>::iterator 2561 it = WriteObjects[Object].begin(), 2562 end = WriteObjects[Object].end(); 2563 2564 for (; it != end; ++it) { 2565 Instruction* I = *it; 2566 if (I == Inst) 2567 continue; 2568 2569 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I); 2570 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth), 2571 ThatLoc.getWithNewSize(MaxByteWidth))) 2572 return true; 2573 } 2574 return false; 2575} 2576 2577bool LoopVectorizationLegality::canVectorizeMemory() { 2578 2579 typedef SmallVector<Value*, 16> ValueVector; 2580 typedef SmallPtrSet<Value*, 16> ValueSet; 2581 // Holds the Load and Store *instructions*. 2582 ValueVector Loads; 2583 ValueVector Stores; 2584 PtrRtCheck.Pointers.clear(); 2585 PtrRtCheck.Need = false; 2586 2587 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); 2588 2589 // For each block. 2590 for (Loop::block_iterator bb = TheLoop->block_begin(), 2591 be = TheLoop->block_end(); bb != be; ++bb) { 2592 2593 // Scan the BB and collect legal loads and stores. 2594 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 2595 ++it) { 2596 2597 // If this is a load, save it. If this instruction can read from memory 2598 // but is not a load, then we quit. Notice that we don't handle function 2599 // calls that read or write. 2600 if (it->mayReadFromMemory()) { 2601 LoadInst *Ld = dyn_cast<LoadInst>(it); 2602 if (!Ld) return false; 2603 if (!Ld->isSimple() && !IsAnnotatedParallel) { 2604 DEBUG(dbgs() << "LV: Found a non-simple load.\n"); 2605 return false; 2606 } 2607 Loads.push_back(Ld); 2608 continue; 2609 } 2610 2611 // Save 'store' instructions. Abort if other instructions write to memory. 2612 if (it->mayWriteToMemory()) { 2613 StoreInst *St = dyn_cast<StoreInst>(it); 2614 if (!St) return false; 2615 if (!St->isSimple() && !IsAnnotatedParallel) { 2616 DEBUG(dbgs() << "LV: Found a non-simple store.\n"); 2617 return false; 2618 } 2619 Stores.push_back(St); 2620 } 2621 } // next instr. 2622 } // next block. 2623 2624 // Now we have two lists that hold the loads and the stores. 2625 // Next, we find the pointers that they use. 2626 2627 // Check if we see any stores. If there are no stores, then we don't 2628 // care if the pointers are *restrict*. 2629 if (!Stores.size()) { 2630 DEBUG(dbgs() << "LV: Found a read-only loop!\n"); 2631 return true; 2632 } 2633 2634 // Holds the read and read-write *pointers* that we find. These maps hold 2635 // unique values for pointers (so no need for multi-map). 2636 AliasMap Reads; 2637 AliasMap ReadWrites; 2638 2639 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects 2640 // multiple times on the same object. If the ptr is accessed twice, once 2641 // for read and once for write, it will only appear once (on the write 2642 // list). This is okay, since we are going to check for conflicts between 2643 // writes and between reads and writes, but not between reads and reads. 2644 ValueSet Seen; 2645 2646 ValueVector::iterator I, IE; 2647 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) { 2648 StoreInst *ST = cast<StoreInst>(*I); 2649 Value* Ptr = ST->getPointerOperand(); 2650 2651 if (isUniform(Ptr)) { 2652 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); 2653 return false; 2654 } 2655 2656 // If we did *not* see this pointer before, insert it to 2657 // the read-write list. At this phase it is only a 'write' list. 2658 if (Seen.insert(Ptr)) 2659 ReadWrites.insert(std::make_pair(Ptr, ST)); 2660 } 2661 2662 if (IsAnnotatedParallel) { 2663 DEBUG(dbgs() 2664 << "LV: A loop annotated parallel, ignore memory dependency " 2665 << "checks.\n"); 2666 return true; 2667 } 2668 2669 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { 2670 LoadInst *LD = cast<LoadInst>(*I); 2671 Value* Ptr = LD->getPointerOperand(); 2672 // If we did *not* see this pointer before, insert it to the 2673 // read list. If we *did* see it before, then it is already in 2674 // the read-write list. This allows us to vectorize expressions 2675 // such as A[i] += x; Because the address of A[i] is a read-write 2676 // pointer. This only works if the index of A[i] is consecutive. 2677 // If the address of i is unknown (for example A[B[i]]) then we may 2678 // read a few words, modify, and write a few words, and some of the 2679 // words may be written to the same address. 2680 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr)) 2681 Reads.insert(std::make_pair(Ptr, LD)); 2682 } 2683 2684 // If we write (or read-write) to a single destination and there are no 2685 // other reads in this loop then is it safe to vectorize. 2686 if (ReadWrites.size() == 1 && Reads.size() == 0) { 2687 DEBUG(dbgs() << "LV: Found a write-only loop!\n"); 2688 return true; 2689 } 2690 2691 unsigned NumReadPtrs = 0; 2692 unsigned NumWritePtrs = 0; 2693 2694 // Find pointers with computable bounds. We are going to use this information 2695 // to place a runtime bound check. 2696 bool CanDoRT = true; 2697 AliasMap::iterator MI, ME; 2698 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { 2699 Value *V = (*MI).first; 2700 if (hasComputableBounds(V)) { 2701 PtrRtCheck.insert(SE, TheLoop, V, true); 2702 NumWritePtrs++; 2703 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); 2704 } else { 2705 CanDoRT = false; 2706 break; 2707 } 2708 } 2709 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { 2710 Value *V = (*MI).first; 2711 if (hasComputableBounds(V)) { 2712 PtrRtCheck.insert(SE, TheLoop, V, false); 2713 NumReadPtrs++; 2714 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n"); 2715 } else { 2716 CanDoRT = false; 2717 break; 2718 } 2719 } 2720 2721 // Check that we did not collect too many pointers or found a 2722 // unsizeable pointer. 2723 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1)); 2724 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n"); 2725 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) { 2726 PtrRtCheck.reset(); 2727 CanDoRT = false; 2728 } 2729 2730 if (CanDoRT) { 2731 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n"); 2732 } 2733 2734 bool NeedRTCheck = false; 2735 2736 // Biggest vectorized access possible, vector width * unroll factor. 2737 // TODO: We're being very pessimistic here, find a way to know the 2738 // real access width before getting here. 2739 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) * 2740 TTI->getMaximumUnrollFactor(); 2741 // Now that the pointers are in two lists (Reads and ReadWrites), we 2742 // can check that there are no conflicts between each of the writes and 2743 // between the writes to the reads. 2744 // Note that WriteObjects duplicates the stores (indexed now by underlying 2745 // objects) to avoid pointing to elements inside ReadWrites. 2746 // TODO: Maybe create a new type where they can interact without duplication. 2747 AliasMultiMap WriteObjects; 2748 ValueVector TempObjects; 2749 2750 // Check that the read-writes do not conflict with other read-write 2751 // pointers. 2752 bool AllWritesIdentified = true; 2753 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) { 2754 Value *Val = (*MI).first; 2755 Instruction *Inst = (*MI).second; 2756 2757 GetUnderlyingObjects(Val, TempObjects, DL); 2758 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); 2759 UI != UE; ++UI) { 2760 if (!isIdentifiedObject(*UI)) { 2761 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n"); 2762 NeedRTCheck = true; 2763 AllWritesIdentified = false; 2764 } 2765 2766 // Never seen it before, can't alias. 2767 if (WriteObjects[*UI].empty()) { 2768 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n"); 2769 WriteObjects[*UI].push_back(Inst); 2770 continue; 2771 } 2772 // Direct alias found. 2773 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) { 2774 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2775 << **UI <<"\n"); 2776 return false; 2777 } 2778 DEBUG(dbgs() << "LV: Found a conflicting global value:" 2779 << **UI <<"\n"); 2780 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n"); 2781 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); 2782 2783 // If global alias, make sure they do alias. 2784 if (hasPossibleGlobalWriteReorder(*UI, 2785 Inst, 2786 WriteObjects, 2787 MaxByteWidth)) { 2788 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2789 << *UI <<"\n"); 2790 return false; 2791 } 2792 2793 // Didn't alias, insert into map for further reference. 2794 WriteObjects[*UI].push_back(Inst); 2795 } 2796 TempObjects.clear(); 2797 } 2798 2799 /// Check that the reads don't conflict with the read-writes. 2800 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) { 2801 Value *Val = (*MI).first; 2802 GetUnderlyingObjects(Val, TempObjects, DL); 2803 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end(); 2804 UI != UE; ++UI) { 2805 // If all of the writes are identified then we don't care if the read 2806 // pointer is identified or not. 2807 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) { 2808 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n"); 2809 NeedRTCheck = true; 2810 } 2811 2812 // Never seen it before, can't alias. 2813 if (WriteObjects[*UI].empty()) 2814 continue; 2815 // Direct alias found. 2816 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) { 2817 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" 2818 << **UI <<"\n"); 2819 return false; 2820 } 2821 DEBUG(dbgs() << "LV: Found a global value: " 2822 << **UI <<"\n"); 2823 Instruction *Inst = (*MI).second; 2824 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n"); 2825 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n"); 2826 2827 // If global alias, make sure they do alias. 2828 if (hasPossibleGlobalWriteReorder(*UI, 2829 Inst, 2830 WriteObjects, 2831 MaxByteWidth)) { 2832 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" 2833 << *UI <<"\n"); 2834 return false; 2835 } 2836 } 2837 TempObjects.clear(); 2838 } 2839 2840 PtrRtCheck.Need = NeedRTCheck; 2841 if (NeedRTCheck && !CanDoRT) { 2842 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " << 2843 "the array bounds.\n"); 2844 PtrRtCheck.reset(); 2845 return false; 2846 } 2847 2848 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") << 2849 " need a runtime memory check.\n"); 2850 return true; 2851} 2852 2853bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, 2854 ReductionKind Kind) { 2855 if (Phi->getNumIncomingValues() != 2) 2856 return false; 2857 2858 // Reduction variables are only found in the loop header block. 2859 if (Phi->getParent() != TheLoop->getHeader()) 2860 return false; 2861 2862 // Obtain the reduction start value from the value that comes from the loop 2863 // preheader. 2864 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader()); 2865 2866 // ExitInstruction is the single value which is used outside the loop. 2867 // We only allow for a single reduction value to be used outside the loop. 2868 // This includes users of the reduction, variables (which form a cycle 2869 // which ends in the phi node). 2870 Instruction *ExitInstruction = 0; 2871 // Indicates that we found a binary operation in our scan. 2872 bool FoundBinOp = false; 2873 2874 // Iter is our iterator. We start with the PHI node and scan for all of the 2875 // users of this instruction. All users must be instructions that can be 2876 // used as reduction variables (such as ADD). We may have a single 2877 // out-of-block user. The cycle must end with the original PHI. 2878 Instruction *Iter = Phi; 2879 2880 // To recognize min/max patterns formed by a icmp select sequence, we store 2881 // the number of instruction we saw from the recognized min/max pattern, 2882 // such that we don't stop when we see the phi has two uses (one by the select 2883 // and one by the icmp) and to make sure we only see exactly the two 2884 // instructions. 2885 unsigned NumICmpSelectPatternInst = 0; 2886 ReductionInstDesc ReduxDesc(false, 0); 2887 2888 // Avoid cycles in the chain. 2889 SmallPtrSet<Instruction *, 8> VisitedInsts; 2890 while (VisitedInsts.insert(Iter)) { 2891 // If the instruction has no users then this is a broken 2892 // chain and can't be a reduction variable. 2893 if (Iter->use_empty()) 2894 return false; 2895 2896 // Did we find a user inside this loop already ? 2897 bool FoundInBlockUser = false; 2898 // Did we reach the initial PHI node already ? 2899 bool FoundStartPHI = false; 2900 2901 // Is this a bin op ? 2902 FoundBinOp |= !isa<PHINode>(Iter); 2903 2904 // For each of the *users* of iter. 2905 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end(); 2906 it != e; ++it) { 2907 Instruction *U = cast<Instruction>(*it); 2908 // We already know that the PHI is a user. 2909 if (U == Phi) { 2910 FoundStartPHI = true; 2911 continue; 2912 } 2913 2914 // Check if we found the exit user. 2915 BasicBlock *Parent = U->getParent(); 2916 if (!TheLoop->contains(Parent)) { 2917 // Exit if you find multiple outside users. 2918 if (ExitInstruction != 0) 2919 return false; 2920 ExitInstruction = Iter; 2921 } 2922 2923 // We allow in-loop PHINodes which are not the original reduction PHI 2924 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE 2925 // structure) then don't skip this PHI. 2926 if (isa<PHINode>(Iter) && isa<PHINode>(U) && 2927 U->getParent() != TheLoop->getHeader() && 2928 TheLoop->contains(U) && 2929 Iter->hasNUsesOrMore(2)) 2930 continue; 2931 2932 // We can't have multiple inside users except for a combination of 2933 // icmp/select both using the phi. 2934 if (FoundInBlockUser && !NumICmpSelectPatternInst) 2935 return false; 2936 FoundInBlockUser = true; 2937 2938 // Any reduction instr must be of one of the allowed kinds. 2939 ReduxDesc = isReductionInstr(U, Kind, ReduxDesc); 2940 if (!ReduxDesc.IsReduction) 2941 return false; 2942 2943 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(U) || 2944 isa<SelectInst>(U))) 2945 ++NumICmpSelectPatternInst; 2946 2947 // Reductions of instructions such as Div, and Sub is only 2948 // possible if the LHS is the reduction variable. 2949 if (!U->isCommutative() && !isa<PHINode>(U) && !isa<SelectInst>(U) && 2950 !isa<ICmpInst>(U) && U->getOperand(0) != Iter) 2951 return false; 2952 2953 Iter = ReduxDesc.PatternLastInst; 2954 } 2955 2956 // This means we have seen one but not the other instruction of the 2957 // pattern or more than just a select and cmp. 2958 if (Kind == RK_IntegerMinMax && NumICmpSelectPatternInst != 2) 2959 return false; 2960 2961 // We found a reduction var if we have reached the original 2962 // phi node and we only have a single instruction with out-of-loop 2963 // users. 2964 if (FoundStartPHI) { 2965 // This instruction is allowed to have out-of-loop users. 2966 AllowedExit.insert(ExitInstruction); 2967 2968 // Save the description of this reduction variable. 2969 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind, 2970 ReduxDesc.MinMaxKind); 2971 Reductions[Phi] = RD; 2972 // We've ended the cycle. This is a reduction variable if we have an 2973 // outside user and it has a binary op. 2974 return FoundBinOp && ExitInstruction; 2975 } 2976 } 2977 2978 return false; 2979} 2980 2981/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction 2982/// pattern corresponding to a min(X, Y) or max(X, Y). 2983LoopVectorizationLegality::ReductionInstDesc 2984LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, ReductionInstDesc &Prev) { 2985 2986 assert((isa<ICmpInst>(I) || isa<SelectInst>(I)) && 2987 "Expect a select instruction"); 2988 ICmpInst *Cmp = 0; 2989 SelectInst *Select = 0; 2990 2991 // We must handle the select(cmp()) as a single instruction. Advance to the 2992 // select. 2993 if ((Cmp = dyn_cast<ICmpInst>(I))) { 2994 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin()))) 2995 return ReductionInstDesc(false, I); 2996 return ReductionInstDesc(Select, Prev.MinMaxKind); 2997 } 2998 2999 // Only handle single use cases for now. 3000 if (!(Select = dyn_cast<SelectInst>(I))) 3001 return ReductionInstDesc(false, I); 3002 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0)))) 3003 return ReductionInstDesc(false, I); 3004 if (!Cmp->hasOneUse()) 3005 return ReductionInstDesc(false, I); 3006 3007 Value *CmpLeft = Cmp->getOperand(0); 3008 Value *CmpRight = Cmp->getOperand(1); 3009 3010 // Look for a min/max pattern. 3011 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3012 return ReductionInstDesc(Select, MRK_UIntMin); 3013 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3014 return ReductionInstDesc(Select, MRK_UIntMax); 3015 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3016 return ReductionInstDesc(Select, MRK_SIntMax); 3017 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select)) 3018 return ReductionInstDesc(Select, MRK_SIntMin); 3019 3020 return ReductionInstDesc(false, I); 3021} 3022 3023LoopVectorizationLegality::ReductionInstDesc 3024LoopVectorizationLegality::isReductionInstr(Instruction *I, 3025 ReductionKind Kind, 3026 ReductionInstDesc &Prev) { 3027 bool FP = I->getType()->isFloatingPointTy(); 3028 bool FastMath = (FP && I->isCommutative() && I->isAssociative()); 3029 switch (I->getOpcode()) { 3030 default: 3031 return ReductionInstDesc(false, I); 3032 case Instruction::PHI: 3033 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd)) 3034 return ReductionInstDesc(false, I); 3035 return ReductionInstDesc(I, Prev.MinMaxKind); 3036 case Instruction::Sub: 3037 case Instruction::Add: 3038 return ReductionInstDesc(Kind == RK_IntegerAdd, I); 3039 case Instruction::Mul: 3040 return ReductionInstDesc(Kind == RK_IntegerMult, I); 3041 case Instruction::And: 3042 return ReductionInstDesc(Kind == RK_IntegerAnd, I); 3043 case Instruction::Or: 3044 return ReductionInstDesc(Kind == RK_IntegerOr, I); 3045 case Instruction::Xor: 3046 return ReductionInstDesc(Kind == RK_IntegerXor, I); 3047 case Instruction::FMul: 3048 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I); 3049 case Instruction::FAdd: 3050 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I); 3051 case Instruction::ICmp: 3052 case Instruction::Select: 3053 if (Kind != RK_IntegerMinMax) 3054 return ReductionInstDesc(false, I); 3055 return isMinMaxSelectCmpPattern(I, Prev); 3056 } 3057} 3058 3059LoopVectorizationLegality::InductionKind 3060LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { 3061 Type *PhiTy = Phi->getType(); 3062 // We only handle integer and pointer inductions variables. 3063 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy()) 3064 return IK_NoInduction; 3065 3066 // Check that the PHI is consecutive. 3067 const SCEV *PhiScev = SE->getSCEV(Phi); 3068 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 3069 if (!AR) { 3070 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); 3071 return IK_NoInduction; 3072 } 3073 const SCEV *Step = AR->getStepRecurrence(*SE); 3074 3075 // Integer inductions need to have a stride of one. 3076 if (PhiTy->isIntegerTy()) { 3077 if (Step->isOne()) 3078 return IK_IntInduction; 3079 if (Step->isAllOnesValue()) 3080 return IK_ReverseIntInduction; 3081 return IK_NoInduction; 3082 } 3083 3084 // Calculate the pointer stride and check if it is consecutive. 3085 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 3086 if (!C) 3087 return IK_NoInduction; 3088 3089 assert(PhiTy->isPointerTy() && "The PHI must be a pointer"); 3090 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType()); 3091 if (C->getValue()->equalsInt(Size)) 3092 return IK_PtrInduction; 3093 else if (C->getValue()->equalsInt(0 - Size)) 3094 return IK_ReversePtrInduction; 3095 3096 return IK_NoInduction; 3097} 3098 3099bool LoopVectorizationLegality::isInductionVariable(const Value *V) { 3100 Value *In0 = const_cast<Value*>(V); 3101 PHINode *PN = dyn_cast_or_null<PHINode>(In0); 3102 if (!PN) 3103 return false; 3104 3105 return Inductions.count(PN); 3106} 3107 3108bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) { 3109 assert(TheLoop->contains(BB) && "Unknown block used"); 3110 3111 // Blocks that do not dominate the latch need predication. 3112 BasicBlock* Latch = TheLoop->getLoopLatch(); 3113 return !DT->dominates(BB, Latch); 3114} 3115 3116bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) { 3117 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3118 // We don't predicate loads/stores at the moment. 3119 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow()) 3120 return false; 3121 3122 // The instructions below can trap. 3123 switch (it->getOpcode()) { 3124 default: continue; 3125 case Instruction::UDiv: 3126 case Instruction::SDiv: 3127 case Instruction::URem: 3128 case Instruction::SRem: 3129 return false; 3130 } 3131 } 3132 3133 return true; 3134} 3135 3136bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) { 3137 const SCEV *PhiScev = SE->getSCEV(Ptr); 3138 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); 3139 if (!AR) 3140 return false; 3141 3142 return AR->isAffine(); 3143} 3144 3145LoopVectorizationCostModel::VectorizationFactor 3146LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize, 3147 unsigned UserVF) { 3148 // Width 1 means no vectorize 3149 VectorizationFactor Factor = { 1U, 0U }; 3150 if (OptForSize && Legal->getRuntimePointerCheck()->Need) { 3151 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n"); 3152 return Factor; 3153 } 3154 3155 // Find the trip count. 3156 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch()); 3157 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n"); 3158 3159 unsigned WidestType = getWidestType(); 3160 unsigned WidestRegister = TTI.getRegisterBitWidth(true); 3161 unsigned MaxVectorSize = WidestRegister / WidestType; 3162 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n"); 3163 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n"); 3164 3165 if (MaxVectorSize == 0) { 3166 DEBUG(dbgs() << "LV: The target has no vector registers.\n"); 3167 MaxVectorSize = 1; 3168 } 3169 3170 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements" 3171 " into one vector!"); 3172 3173 unsigned VF = MaxVectorSize; 3174 3175 // If we optimize the program for size, avoid creating the tail loop. 3176 if (OptForSize) { 3177 // If we are unable to calculate the trip count then don't try to vectorize. 3178 if (TC < 2) { 3179 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 3180 return Factor; 3181 } 3182 3183 // Find the maximum SIMD width that can fit within the trip count. 3184 VF = TC % MaxVectorSize; 3185 3186 if (VF == 0) 3187 VF = MaxVectorSize; 3188 3189 // If the trip count that we found modulo the vectorization factor is not 3190 // zero then we require a tail. 3191 if (VF < 2) { 3192 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n"); 3193 return Factor; 3194 } 3195 } 3196 3197 if (UserVF != 0) { 3198 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two"); 3199 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n"); 3200 3201 Factor.Width = UserVF; 3202 return Factor; 3203 } 3204 3205 float Cost = expectedCost(1); 3206 unsigned Width = 1; 3207 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n"); 3208 for (unsigned i=2; i <= VF; i*=2) { 3209 // Notice that the vector loop needs to be executed less times, so 3210 // we need to divide the cost of the vector loops by the width of 3211 // the vector elements. 3212 float VectorCost = expectedCost(i) / (float)i; 3213 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " << 3214 (int)VectorCost << ".\n"); 3215 if (VectorCost < Cost) { 3216 Cost = VectorCost; 3217 Width = i; 3218 } 3219 } 3220 3221 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n"); 3222 Factor.Width = Width; 3223 Factor.Cost = Width * Cost; 3224 return Factor; 3225} 3226 3227unsigned LoopVectorizationCostModel::getWidestType() { 3228 unsigned MaxWidth = 8; 3229 3230 // For each block. 3231 for (Loop::block_iterator bb = TheLoop->block_begin(), 3232 be = TheLoop->block_end(); bb != be; ++bb) { 3233 BasicBlock *BB = *bb; 3234 3235 // For each instruction in the loop. 3236 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3237 Type *T = it->getType(); 3238 3239 // Only examine Loads, Stores and PHINodes. 3240 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it)) 3241 continue; 3242 3243 // Examine PHI nodes that are reduction variables. 3244 if (PHINode *PN = dyn_cast<PHINode>(it)) 3245 if (!Legal->getReductionVars()->count(PN)) 3246 continue; 3247 3248 // Examine the stored values. 3249 if (StoreInst *ST = dyn_cast<StoreInst>(it)) 3250 T = ST->getValueOperand()->getType(); 3251 3252 // Ignore loaded pointer types and stored pointer types that are not 3253 // consecutive. However, we do want to take consecutive stores/loads of 3254 // pointer vectors into account. 3255 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it)) 3256 continue; 3257 3258 MaxWidth = std::max(MaxWidth, 3259 (unsigned)DL->getTypeSizeInBits(T->getScalarType())); 3260 } 3261 } 3262 3263 return MaxWidth; 3264} 3265 3266unsigned 3267LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize, 3268 unsigned UserUF, 3269 unsigned VF, 3270 unsigned LoopCost) { 3271 3272 // -- The unroll heuristics -- 3273 // We unroll the loop in order to expose ILP and reduce the loop overhead. 3274 // There are many micro-architectural considerations that we can't predict 3275 // at this level. For example frontend pressure (on decode or fetch) due to 3276 // code size, or the number and capabilities of the execution ports. 3277 // 3278 // We use the following heuristics to select the unroll factor: 3279 // 1. If the code has reductions the we unroll in order to break the cross 3280 // iteration dependency. 3281 // 2. If the loop is really small then we unroll in order to reduce the loop 3282 // overhead. 3283 // 3. We don't unroll if we think that we will spill registers to memory due 3284 // to the increased register pressure. 3285 3286 // Use the user preference, unless 'auto' is selected. 3287 if (UserUF != 0) 3288 return UserUF; 3289 3290 // When we optimize for size we don't unroll. 3291 if (OptForSize) 3292 return 1; 3293 3294 // Do not unroll loops with a relatively small trip count. 3295 unsigned TC = SE->getSmallConstantTripCount(TheLoop, 3296 TheLoop->getLoopLatch()); 3297 if (TC > 1 && TC < TinyTripCountUnrollThreshold) 3298 return 1; 3299 3300 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true); 3301 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters << 3302 " vector registers\n"); 3303 3304 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage(); 3305 // We divide by these constants so assume that we have at least one 3306 // instruction that uses at least one register. 3307 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U); 3308 R.NumInstructions = std::max(R.NumInstructions, 1U); 3309 3310 // We calculate the unroll factor using the following formula. 3311 // Subtract the number of loop invariants from the number of available 3312 // registers. These registers are used by all of the unrolled instances. 3313 // Next, divide the remaining registers by the number of registers that is 3314 // required by the loop, in order to estimate how many parallel instances 3315 // fit without causing spills. 3316 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers; 3317 3318 // Clamp the unroll factor ranges to reasonable factors. 3319 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor(); 3320 3321 // If we did not calculate the cost for VF (because the user selected the VF) 3322 // then we calculate the cost of VF here. 3323 if (LoopCost == 0) 3324 LoopCost = expectedCost(VF); 3325 3326 // Clamp the calculated UF to be between the 1 and the max unroll factor 3327 // that the target allows. 3328 if (UF > MaxUnrollSize) 3329 UF = MaxUnrollSize; 3330 else if (UF < 1) 3331 UF = 1; 3332 3333 if (Legal->getReductionVars()->size()) { 3334 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n"); 3335 return UF; 3336 } 3337 3338 // We want to unroll tiny loops in order to reduce the loop overhead. 3339 // We assume that the cost overhead is 1 and we use the cost model 3340 // to estimate the cost of the loop and unroll until the cost of the 3341 // loop overhead is about 5% of the cost of the loop. 3342 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n"); 3343 if (LoopCost < 20) { 3344 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n"); 3345 unsigned NewUF = 20/LoopCost + 1; 3346 return std::min(NewUF, UF); 3347 } 3348 3349 DEBUG(dbgs() << "LV: Not Unrolling. \n"); 3350 return 1; 3351} 3352 3353LoopVectorizationCostModel::RegisterUsage 3354LoopVectorizationCostModel::calculateRegisterUsage() { 3355 // This function calculates the register usage by measuring the highest number 3356 // of values that are alive at a single location. Obviously, this is a very 3357 // rough estimation. We scan the loop in a topological order in order and 3358 // assign a number to each instruction. We use RPO to ensure that defs are 3359 // met before their users. We assume that each instruction that has in-loop 3360 // users starts an interval. We record every time that an in-loop value is 3361 // used, so we have a list of the first and last occurrences of each 3362 // instruction. Next, we transpose this data structure into a multi map that 3363 // holds the list of intervals that *end* at a specific location. This multi 3364 // map allows us to perform a linear search. We scan the instructions linearly 3365 // and record each time that a new interval starts, by placing it in a set. 3366 // If we find this value in the multi-map then we remove it from the set. 3367 // The max register usage is the maximum size of the set. 3368 // We also search for instructions that are defined outside the loop, but are 3369 // used inside the loop. We need this number separately from the max-interval 3370 // usage number because when we unroll, loop-invariant values do not take 3371 // more register. 3372 LoopBlocksDFS DFS(TheLoop); 3373 DFS.perform(LI); 3374 3375 RegisterUsage R; 3376 R.NumInstructions = 0; 3377 3378 // Each 'key' in the map opens a new interval. The values 3379 // of the map are the index of the 'last seen' usage of the 3380 // instruction that is the key. 3381 typedef DenseMap<Instruction*, unsigned> IntervalMap; 3382 // Maps instruction to its index. 3383 DenseMap<unsigned, Instruction*> IdxToInstr; 3384 // Marks the end of each interval. 3385 IntervalMap EndPoint; 3386 // Saves the list of instruction indices that are used in the loop. 3387 SmallSet<Instruction*, 8> Ends; 3388 // Saves the list of values that are used in the loop but are 3389 // defined outside the loop, such as arguments and constants. 3390 SmallPtrSet<Value*, 8> LoopInvariants; 3391 3392 unsigned Index = 0; 3393 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), 3394 be = DFS.endRPO(); bb != be; ++bb) { 3395 R.NumInstructions += (*bb)->size(); 3396 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e; 3397 ++it) { 3398 Instruction *I = it; 3399 IdxToInstr[Index++] = I; 3400 3401 // Save the end location of each USE. 3402 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 3403 Value *U = I->getOperand(i); 3404 Instruction *Instr = dyn_cast<Instruction>(U); 3405 3406 // Ignore non-instruction values such as arguments, constants, etc. 3407 if (!Instr) continue; 3408 3409 // If this instruction is outside the loop then record it and continue. 3410 if (!TheLoop->contains(Instr)) { 3411 LoopInvariants.insert(Instr); 3412 continue; 3413 } 3414 3415 // Overwrite previous end points. 3416 EndPoint[Instr] = Index; 3417 Ends.insert(Instr); 3418 } 3419 } 3420 } 3421 3422 // Saves the list of intervals that end with the index in 'key'. 3423 typedef SmallVector<Instruction*, 2> InstrList; 3424 DenseMap<unsigned, InstrList> TransposeEnds; 3425 3426 // Transpose the EndPoints to a list of values that end at each index. 3427 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end(); 3428 it != e; ++it) 3429 TransposeEnds[it->second].push_back(it->first); 3430 3431 SmallSet<Instruction*, 8> OpenIntervals; 3432 unsigned MaxUsage = 0; 3433 3434 3435 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n"); 3436 for (unsigned int i = 0; i < Index; ++i) { 3437 Instruction *I = IdxToInstr[i]; 3438 // Ignore instructions that are never used within the loop. 3439 if (!Ends.count(I)) continue; 3440 3441 // Remove all of the instructions that end at this location. 3442 InstrList &List = TransposeEnds[i]; 3443 for (unsigned int j=0, e = List.size(); j < e; ++j) 3444 OpenIntervals.erase(List[j]); 3445 3446 // Count the number of live interals. 3447 MaxUsage = std::max(MaxUsage, OpenIntervals.size()); 3448 3449 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " << 3450 OpenIntervals.size() <<"\n"); 3451 3452 // Add the current instruction to the list of open intervals. 3453 OpenIntervals.insert(I); 3454 } 3455 3456 unsigned Invariant = LoopInvariants.size(); 3457 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n"); 3458 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n"); 3459 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n"); 3460 3461 R.LoopInvariantRegs = Invariant; 3462 R.MaxLocalUsers = MaxUsage; 3463 return R; 3464} 3465 3466unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { 3467 unsigned Cost = 0; 3468 3469 // For each block. 3470 for (Loop::block_iterator bb = TheLoop->block_begin(), 3471 be = TheLoop->block_end(); bb != be; ++bb) { 3472 unsigned BlockCost = 0; 3473 BasicBlock *BB = *bb; 3474 3475 // For each instruction in the old loop. 3476 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 3477 // Skip dbg intrinsics. 3478 if (isa<DbgInfoIntrinsic>(it)) 3479 continue; 3480 3481 unsigned C = getInstructionCost(it, VF); 3482 Cost += C; 3483 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " << 3484 VF << " For instruction: "<< *it << "\n"); 3485 } 3486 3487 // We assume that if-converted blocks have a 50% chance of being executed. 3488 // When the code is scalar then some of the blocks are avoided due to CF. 3489 // When the code is vectorized we execute all code paths. 3490 if (Legal->blockNeedsPredication(*bb) && VF == 1) 3491 BlockCost /= 2; 3492 3493 Cost += BlockCost; 3494 } 3495 3496 return Cost; 3497} 3498 3499unsigned 3500LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { 3501 // If we know that this instruction will remain uniform, check the cost of 3502 // the scalar version. 3503 if (Legal->isUniformAfterVectorization(I)) 3504 VF = 1; 3505 3506 Type *RetTy = I->getType(); 3507 Type *VectorTy = ToVectorTy(RetTy, VF); 3508 3509 // TODO: We need to estimate the cost of intrinsic calls. 3510 switch (I->getOpcode()) { 3511 case Instruction::GetElementPtr: 3512 // We mark this instruction as zero-cost because the cost of GEPs in 3513 // vectorized code depends on whether the corresponding memory instruction 3514 // is scalarized or not. Therefore, we handle GEPs with the memory 3515 // instruction cost. 3516 return 0; 3517 case Instruction::Br: { 3518 return TTI.getCFInstrCost(I->getOpcode()); 3519 } 3520 case Instruction::PHI: 3521 //TODO: IF-converted IFs become selects. 3522 return 0; 3523 case Instruction::Add: 3524 case Instruction::FAdd: 3525 case Instruction::Sub: 3526 case Instruction::FSub: 3527 case Instruction::Mul: 3528 case Instruction::FMul: 3529 case Instruction::UDiv: 3530 case Instruction::SDiv: 3531 case Instruction::FDiv: 3532 case Instruction::URem: 3533 case Instruction::SRem: 3534 case Instruction::FRem: 3535 case Instruction::Shl: 3536 case Instruction::LShr: 3537 case Instruction::AShr: 3538 case Instruction::And: 3539 case Instruction::Or: 3540 case Instruction::Xor: { 3541 // Certain instructions can be cheaper to vectorize if they have a constant 3542 // second vector operand. One example of this are shifts on x86. 3543 TargetTransformInfo::OperandValueKind Op1VK = 3544 TargetTransformInfo::OK_AnyValue; 3545 TargetTransformInfo::OperandValueKind Op2VK = 3546 TargetTransformInfo::OK_AnyValue; 3547 3548 if (isa<ConstantInt>(I->getOperand(1))) 3549 Op2VK = TargetTransformInfo::OK_UniformConstantValue; 3550 3551 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK); 3552 } 3553 case Instruction::Select: { 3554 SelectInst *SI = cast<SelectInst>(I); 3555 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); 3556 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); 3557 Type *CondTy = SI->getCondition()->getType(); 3558 if (!ScalarCond) 3559 CondTy = VectorType::get(CondTy, VF); 3560 3561 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); 3562 } 3563 case Instruction::ICmp: 3564 case Instruction::FCmp: { 3565 Type *ValTy = I->getOperand(0)->getType(); 3566 VectorTy = ToVectorTy(ValTy, VF); 3567 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy); 3568 } 3569 case Instruction::Store: 3570 case Instruction::Load: { 3571 StoreInst *SI = dyn_cast<StoreInst>(I); 3572 LoadInst *LI = dyn_cast<LoadInst>(I); 3573 Type *ValTy = (SI ? SI->getValueOperand()->getType() : 3574 LI->getType()); 3575 VectorTy = ToVectorTy(ValTy, VF); 3576 3577 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment(); 3578 unsigned AS = SI ? SI->getPointerAddressSpace() : 3579 LI->getPointerAddressSpace(); 3580 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand(); 3581 // We add the cost of address computation here instead of with the gep 3582 // instruction because only here we know whether the operation is 3583 // scalarized. 3584 if (VF == 1) 3585 return TTI.getAddressComputationCost(VectorTy) + 3586 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 3587 3588 // Scalarized loads/stores. 3589 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr); 3590 bool Reverse = ConsecutiveStride < 0; 3591 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy); 3592 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF; 3593 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) { 3594 unsigned Cost = 0; 3595 // The cost of extracting from the value vector and pointer vector. 3596 Type *PtrTy = ToVectorTy(Ptr->getType(), VF); 3597 for (unsigned i = 0; i < VF; ++i) { 3598 // The cost of extracting the pointer operand. 3599 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i); 3600 // In case of STORE, the cost of ExtractElement from the vector. 3601 // In case of LOAD, the cost of InsertElement into the returned 3602 // vector. 3603 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement : 3604 Instruction::InsertElement, 3605 VectorTy, i); 3606 } 3607 3608 // The cost of the scalar loads/stores. 3609 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType()); 3610 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), 3611 Alignment, AS); 3612 return Cost; 3613 } 3614 3615 // Wide load/stores. 3616 unsigned Cost = TTI.getAddressComputationCost(VectorTy); 3617 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS); 3618 3619 if (Reverse) 3620 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, 3621 VectorTy, 0); 3622 return Cost; 3623 } 3624 case Instruction::ZExt: 3625 case Instruction::SExt: 3626 case Instruction::FPToUI: 3627 case Instruction::FPToSI: 3628 case Instruction::FPExt: 3629 case Instruction::PtrToInt: 3630 case Instruction::IntToPtr: 3631 case Instruction::SIToFP: 3632 case Instruction::UIToFP: 3633 case Instruction::Trunc: 3634 case Instruction::FPTrunc: 3635 case Instruction::BitCast: { 3636 // We optimize the truncation of induction variable. 3637 // The cost of these is the same as the scalar operation. 3638 if (I->getOpcode() == Instruction::Trunc && 3639 Legal->isInductionVariable(I->getOperand(0))) 3640 return TTI.getCastInstrCost(I->getOpcode(), I->getType(), 3641 I->getOperand(0)->getType()); 3642 3643 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); 3644 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); 3645 } 3646 case Instruction::Call: { 3647 CallInst *CI = cast<CallInst>(I); 3648 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); 3649 assert(ID && "Not an intrinsic call!"); 3650 Type *RetTy = ToVectorTy(CI->getType(), VF); 3651 SmallVector<Type*, 4> Tys; 3652 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) 3653 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF)); 3654 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys); 3655 } 3656 default: { 3657 // We are scalarizing the instruction. Return the cost of the scalar 3658 // instruction, plus the cost of insert and extract into vector 3659 // elements, times the vector width. 3660 unsigned Cost = 0; 3661 3662 if (!RetTy->isVoidTy() && VF != 1) { 3663 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement, 3664 VectorTy); 3665 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement, 3666 VectorTy); 3667 3668 // The cost of inserting the results plus extracting each one of the 3669 // operands. 3670 Cost += VF * (InsCost + ExtCost * I->getNumOperands()); 3671 } 3672 3673 // The cost of executing VF copies of the scalar instruction. This opcode 3674 // is unknown. Assume that it is the same as 'mul'. 3675 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy); 3676 return Cost; 3677 } 3678 }// end of switch. 3679} 3680 3681Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) { 3682 if (Scalar->isVoidTy() || VF == 1) 3683 return Scalar; 3684 return VectorType::get(Scalar, VF); 3685} 3686 3687char LoopVectorize::ID = 0; 3688static const char lv_name[] = "Loop Vectorization"; 3689INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) 3690INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 3691INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 3692INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 3693INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 3694INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) 3695 3696namespace llvm { 3697 Pass *createLoopVectorizePass() { 3698 return new LoopVectorize(); 3699 } 3700} 3701 3702bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { 3703 // Check for a store. 3704 if (StoreInst *ST = dyn_cast<StoreInst>(Inst)) 3705 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0; 3706 3707 // Check for a load. 3708 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 3709 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0; 3710 3711 return false; 3712} 3713