SLPVectorizer.cpp revision a0d44fe4cd92c11466b82af4f5089af845a2eeb5
1//===- SLPVectorizer.cpp - A bottom up SLP 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// This pass implements the Bottom Up SLP vectorizer. It detects consecutive 10// stores that can be put together into vector-stores. Next, it attempts to 11// construct vectorizable tree using the use-def chains. If a profitable tree 12// was found, the SLP vectorizer performs vectorization on the tree. 13// 14// The pass is inspired by the work described in the paper: 15// "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. 16// 17//===----------------------------------------------------------------------===// 18#define SV_NAME "slp-vectorizer" 19#define DEBUG_TYPE "SLP" 20 21#include "llvm/Transforms/Vectorize.h" 22#include "llvm/ADT/MapVector.h" 23#include "llvm/ADT/PostOrderIterator.h" 24#include "llvm/ADT/SetVector.h" 25#include "llvm/Analysis/AliasAnalysis.h" 26#include "llvm/Analysis/ScalarEvolution.h" 27#include "llvm/Analysis/ScalarEvolutionExpressions.h" 28#include "llvm/Analysis/TargetTransformInfo.h" 29#include "llvm/Analysis/ValueTracking.h" 30#include "llvm/Analysis/Verifier.h" 31#include "llvm/Analysis/LoopInfo.h" 32#include "llvm/IR/DataLayout.h" 33#include "llvm/IR/Instructions.h" 34#include "llvm/IR/IntrinsicInst.h" 35#include "llvm/IR/IRBuilder.h" 36#include "llvm/IR/Module.h" 37#include "llvm/IR/Type.h" 38#include "llvm/IR/Value.h" 39#include "llvm/Pass.h" 40#include "llvm/Support/CommandLine.h" 41#include "llvm/Support/Debug.h" 42#include "llvm/Support/raw_ostream.h" 43#include <algorithm> 44#include <map> 45 46using namespace llvm; 47 48static cl::opt<int> 49 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, 50 cl::desc("Only vectorize if you gain more than this " 51 "number ")); 52 53static cl::opt<bool> 54ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden, 55 cl::desc("Attempt to vectorize horizontal reductions")); 56 57static cl::opt<bool> ShouldStartVectorizeHorAtStore( 58 "slp-vectorize-hor-store", cl::init(false), cl::Hidden, 59 cl::desc( 60 "Attempt to vectorize horizontal reductions feeding into a store")); 61 62namespace { 63 64static const unsigned MinVecRegSize = 128; 65 66static const unsigned RecursionMaxDepth = 12; 67 68/// A helper class for numbering instructions in multiple blocks. 69/// Numbers start at zero for each basic block. 70struct BlockNumbering { 71 72 BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {} 73 74 BlockNumbering() : BB(0), Valid(false) {} 75 76 void numberInstructions() { 77 unsigned Loc = 0; 78 InstrIdx.clear(); 79 InstrVec.clear(); 80 // Number the instructions in the block. 81 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 82 InstrIdx[it] = Loc++; 83 InstrVec.push_back(it); 84 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation"); 85 } 86 Valid = true; 87 } 88 89 int getIndex(Instruction *I) { 90 assert(I->getParent() == BB && "Invalid instruction"); 91 if (!Valid) 92 numberInstructions(); 93 assert(InstrIdx.count(I) && "Unknown instruction"); 94 return InstrIdx[I]; 95 } 96 97 Instruction *getInstruction(unsigned loc) { 98 if (!Valid) 99 numberInstructions(); 100 assert(InstrVec.size() > loc && "Invalid Index"); 101 return InstrVec[loc]; 102 } 103 104 void forget() { Valid = false; } 105 106private: 107 /// The block we are numbering. 108 BasicBlock *BB; 109 /// Is the block numbered. 110 bool Valid; 111 /// Maps instructions to numbers and back. 112 SmallDenseMap<Instruction *, int> InstrIdx; 113 /// Maps integers to Instructions. 114 SmallVector<Instruction *, 32> InstrVec; 115}; 116 117/// \returns the parent basic block if all of the instructions in \p VL 118/// are in the same block or null otherwise. 119static BasicBlock *getSameBlock(ArrayRef<Value *> VL) { 120 Instruction *I0 = dyn_cast<Instruction>(VL[0]); 121 if (!I0) 122 return 0; 123 BasicBlock *BB = I0->getParent(); 124 for (int i = 1, e = VL.size(); i < e; i++) { 125 Instruction *I = dyn_cast<Instruction>(VL[i]); 126 if (!I) 127 return 0; 128 129 if (BB != I->getParent()) 130 return 0; 131 } 132 return BB; 133} 134 135/// \returns True if all of the values in \p VL are constants. 136static bool allConstant(ArrayRef<Value *> VL) { 137 for (unsigned i = 0, e = VL.size(); i < e; ++i) 138 if (!isa<Constant>(VL[i])) 139 return false; 140 return true; 141} 142 143/// \returns True if all of the values in \p VL are identical. 144static bool isSplat(ArrayRef<Value *> VL) { 145 for (unsigned i = 1, e = VL.size(); i < e; ++i) 146 if (VL[i] != VL[0]) 147 return false; 148 return true; 149} 150 151/// \returns The opcode if all of the Instructions in \p VL have the same 152/// opcode, or zero. 153static unsigned getSameOpcode(ArrayRef<Value *> VL) { 154 Instruction *I0 = dyn_cast<Instruction>(VL[0]); 155 if (!I0) 156 return 0; 157 unsigned Opcode = I0->getOpcode(); 158 for (int i = 1, e = VL.size(); i < e; i++) { 159 Instruction *I = dyn_cast<Instruction>(VL[i]); 160 if (!I || Opcode != I->getOpcode()) 161 return 0; 162 } 163 return Opcode; 164} 165 166/// \returns \p I after propagating metadata from \p VL. 167static Instruction *propagateMetadata(Instruction *I, ArrayRef<Value *> VL) { 168 Instruction *I0 = cast<Instruction>(VL[0]); 169 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata; 170 I0->getAllMetadataOtherThanDebugLoc(Metadata); 171 172 for (unsigned i = 0, n = Metadata.size(); i != n; ++i) { 173 unsigned Kind = Metadata[i].first; 174 MDNode *MD = Metadata[i].second; 175 176 for (int i = 1, e = VL.size(); MD && i != e; i++) { 177 Instruction *I = cast<Instruction>(VL[i]); 178 MDNode *IMD = I->getMetadata(Kind); 179 180 switch (Kind) { 181 default: 182 MD = 0; // Remove unknown metadata 183 break; 184 case LLVMContext::MD_tbaa: 185 MD = MDNode::getMostGenericTBAA(MD, IMD); 186 break; 187 case LLVMContext::MD_fpmath: 188 MD = MDNode::getMostGenericFPMath(MD, IMD); 189 break; 190 } 191 } 192 I->setMetadata(Kind, MD); 193 } 194 return I; 195} 196 197/// \returns The type that all of the values in \p VL have or null if there 198/// are different types. 199static Type* getSameType(ArrayRef<Value *> VL) { 200 Type *Ty = VL[0]->getType(); 201 for (int i = 1, e = VL.size(); i < e; i++) 202 if (VL[i]->getType() != Ty) 203 return 0; 204 205 return Ty; 206} 207 208/// \returns True if the ExtractElement instructions in VL can be vectorized 209/// to use the original vector. 210static bool CanReuseExtract(ArrayRef<Value *> VL) { 211 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode"); 212 // Check if all of the extracts come from the same vector and from the 213 // correct offset. 214 Value *VL0 = VL[0]; 215 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0); 216 Value *Vec = E0->getOperand(0); 217 218 // We have to extract from the same vector type. 219 unsigned NElts = Vec->getType()->getVectorNumElements(); 220 221 if (NElts != VL.size()) 222 return false; 223 224 // Check that all of the indices extract from the correct offset. 225 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1)); 226 if (!CI || CI->getZExtValue()) 227 return false; 228 229 for (unsigned i = 1, e = VL.size(); i < e; ++i) { 230 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]); 231 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1)); 232 233 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec) 234 return false; 235 } 236 237 return true; 238} 239 240static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, 241 SmallVectorImpl<Value *> &Left, 242 SmallVectorImpl<Value *> &Right) { 243 244 SmallVector<Value *, 16> OrigLeft, OrigRight; 245 246 bool AllSameOpcodeLeft = true; 247 bool AllSameOpcodeRight = true; 248 for (unsigned i = 0, e = VL.size(); i != e; ++i) { 249 Instruction *I = cast<Instruction>(VL[i]); 250 Value *V0 = I->getOperand(0); 251 Value *V1 = I->getOperand(1); 252 253 OrigLeft.push_back(V0); 254 OrigRight.push_back(V1); 255 256 Instruction *I0 = dyn_cast<Instruction>(V0); 257 Instruction *I1 = dyn_cast<Instruction>(V1); 258 259 // Check whether all operands on one side have the same opcode. In this case 260 // we want to preserve the original order and not make things worse by 261 // reordering. 262 AllSameOpcodeLeft = I0; 263 AllSameOpcodeRight = I1; 264 265 if (i && AllSameOpcodeLeft) { 266 if(Instruction *P0 = dyn_cast<Instruction>(OrigLeft[i-1])) { 267 if(P0->getOpcode() != I0->getOpcode()) 268 AllSameOpcodeLeft = false; 269 } else 270 AllSameOpcodeLeft = false; 271 } 272 if (i && AllSameOpcodeRight) { 273 if(Instruction *P1 = dyn_cast<Instruction>(OrigRight[i-1])) { 274 if(P1->getOpcode() != I1->getOpcode()) 275 AllSameOpcodeRight = false; 276 } else 277 AllSameOpcodeRight = false; 278 } 279 280 // Sort two opcodes. In the code below we try to preserve the ability to use 281 // broadcast of values instead of individual inserts. 282 // vl1 = load 283 // vl2 = phi 284 // vr1 = load 285 // vr2 = vr2 286 // = vl1 x vr1 287 // = vl2 x vr2 288 // If we just sorted according to opcode we would leave the first line in 289 // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load). 290 // = vl1 x vr1 291 // = vr2 x vl2 292 // Because vr2 and vr1 are from the same load we loose the opportunity of a 293 // broadcast for the packed right side in the backend: we have [vr1, vl2] 294 // instead of [vr1, vr2=vr1]. 295 if (I0 && I1) { 296 if(!i && I0->getOpcode() > I1->getOpcode()) { 297 Left.push_back(I1); 298 Right.push_back(I0); 299 } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) { 300 // Try not to destroy a broad cast for no apparent benefit. 301 Left.push_back(I1); 302 Right.push_back(I0); 303 } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) { 304 // Try preserve broadcasts. 305 Left.push_back(I1); 306 Right.push_back(I0); 307 } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) { 308 // Try preserve broadcasts. 309 Left.push_back(I1); 310 Right.push_back(I0); 311 } else { 312 Left.push_back(I0); 313 Right.push_back(I1); 314 } 315 continue; 316 } 317 // One opcode, put the instruction on the right. 318 if (I0) { 319 Left.push_back(V1); 320 Right.push_back(I0); 321 continue; 322 } 323 Left.push_back(V0); 324 Right.push_back(V1); 325 } 326 327 bool LeftBroadcast = isSplat(Left); 328 bool RightBroadcast = isSplat(Right); 329 330 // Don't reorder if the operands where good to begin with. 331 if (!(LeftBroadcast || RightBroadcast) && 332 (AllSameOpcodeRight || AllSameOpcodeLeft)) { 333 Left = OrigLeft; 334 Right = OrigRight; 335 } 336} 337 338/// Bottom Up SLP Vectorizer. 339class BoUpSLP { 340public: 341 typedef SmallVector<Value *, 8> ValueList; 342 typedef SmallVector<Instruction *, 16> InstrList; 343 typedef SmallPtrSet<Value *, 16> ValueSet; 344 typedef SmallVector<StoreInst *, 8> StoreList; 345 346 BoUpSLP(Function *Func, ScalarEvolution *Se, DataLayout *Dl, 347 TargetTransformInfo *Tti, AliasAnalysis *Aa, LoopInfo *Li, 348 DominatorTree *Dt) : 349 F(Func), SE(Se), DL(Dl), TTI(Tti), AA(Aa), LI(Li), DT(Dt), 350 Builder(Se->getContext()) { 351 // Setup the block numbering utility for all of the blocks in the 352 // function. 353 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) { 354 BasicBlock *BB = it; 355 BlocksNumbers[BB] = BlockNumbering(BB); 356 } 357 } 358 359 /// \brief Vectorize the tree that starts with the elements in \p VL. 360 /// Returns the vectorized root. 361 Value *vectorizeTree(); 362 363 /// \returns the vectorization cost of the subtree that starts at \p VL. 364 /// A negative number means that this is profitable. 365 int getTreeCost(); 366 367 /// Construct a vectorizable tree that starts at \p Roots and is possibly 368 /// used by a reduction of \p RdxOps. 369 void buildTree(ArrayRef<Value *> Roots, ValueSet *RdxOps = 0); 370 371 /// Clear the internal data structures that are created by 'buildTree'. 372 void deleteTree() { 373 RdxOps = 0; 374 VectorizableTree.clear(); 375 ScalarToTreeEntry.clear(); 376 MustGather.clear(); 377 ExternalUses.clear(); 378 MemBarrierIgnoreList.clear(); 379 } 380 381 /// \returns true if the memory operations A and B are consecutive. 382 bool isConsecutiveAccess(Value *A, Value *B); 383 384 /// \brief Perform LICM and CSE on the newly generated gather sequences. 385 void optimizeGatherSequence(); 386private: 387 struct TreeEntry; 388 389 /// \returns the cost of the vectorizable entry. 390 int getEntryCost(TreeEntry *E); 391 392 /// This is the recursive part of buildTree. 393 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth); 394 395 /// Vectorize a single entry in the tree. 396 Value *vectorizeTree(TreeEntry *E); 397 398 /// Vectorize a single entry in the tree, starting in \p VL. 399 Value *vectorizeTree(ArrayRef<Value *> VL); 400 401 /// \returns the pointer to the vectorized value if \p VL is already 402 /// vectorized, or NULL. They may happen in cycles. 403 Value *alreadyVectorized(ArrayRef<Value *> VL) const; 404 405 /// \brief Take the pointer operand from the Load/Store instruction. 406 /// \returns NULL if this is not a valid Load/Store instruction. 407 static Value *getPointerOperand(Value *I); 408 409 /// \brief Take the address space operand from the Load/Store instruction. 410 /// \returns -1 if this is not a valid Load/Store instruction. 411 static unsigned getAddressSpaceOperand(Value *I); 412 413 /// \returns the scalarization cost for this type. Scalarization in this 414 /// context means the creation of vectors from a group of scalars. 415 int getGatherCost(Type *Ty); 416 417 /// \returns the scalarization cost for this list of values. Assuming that 418 /// this subtree gets vectorized, we may need to extract the values from the 419 /// roots. This method calculates the cost of extracting the values. 420 int getGatherCost(ArrayRef<Value *> VL); 421 422 /// \returns the AA location that is being access by the instruction. 423 AliasAnalysis::Location getLocation(Instruction *I); 424 425 /// \brief Checks if it is possible to sink an instruction from 426 /// \p Src to \p Dst. 427 /// \returns the pointer to the barrier instruction if we can't sink. 428 Value *getSinkBarrier(Instruction *Src, Instruction *Dst); 429 430 /// \returns the index of the last instruction in the BB from \p VL. 431 int getLastIndex(ArrayRef<Value *> VL); 432 433 /// \returns the Instruction in the bundle \p VL. 434 Instruction *getLastInstruction(ArrayRef<Value *> VL); 435 436 /// \brief Set the Builder insert point to one after the last instruction in 437 /// the bundle 438 void setInsertPointAfterBundle(ArrayRef<Value *> VL); 439 440 /// \returns a vector from a collection of scalars in \p VL. 441 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty); 442 443 /// \returns whether the VectorizableTree is fully vectoriable and will 444 /// be beneficial even the tree height is tiny. 445 bool isFullyVectorizableTinyTree(); 446 447 struct TreeEntry { 448 TreeEntry() : Scalars(), VectorizedValue(0), LastScalarIndex(0), 449 NeedToGather(0) {} 450 451 /// \returns true if the scalars in VL are equal to this entry. 452 bool isSame(ArrayRef<Value *> VL) const { 453 assert(VL.size() == Scalars.size() && "Invalid size"); 454 return std::equal(VL.begin(), VL.end(), Scalars.begin()); 455 } 456 457 /// A vector of scalars. 458 ValueList Scalars; 459 460 /// The Scalars are vectorized into this value. It is initialized to Null. 461 Value *VectorizedValue; 462 463 /// The index in the basic block of the last scalar. 464 int LastScalarIndex; 465 466 /// Do we need to gather this sequence ? 467 bool NeedToGather; 468 }; 469 470 /// Create a new VectorizableTree entry. 471 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) { 472 VectorizableTree.push_back(TreeEntry()); 473 int idx = VectorizableTree.size() - 1; 474 TreeEntry *Last = &VectorizableTree[idx]; 475 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); 476 Last->NeedToGather = !Vectorized; 477 if (Vectorized) { 478 Last->LastScalarIndex = getLastIndex(VL); 479 for (int i = 0, e = VL.size(); i != e; ++i) { 480 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!"); 481 ScalarToTreeEntry[VL[i]] = idx; 482 } 483 } else { 484 Last->LastScalarIndex = 0; 485 MustGather.insert(VL.begin(), VL.end()); 486 } 487 return Last; 488 } 489 490 /// -- Vectorization State -- 491 /// Holds all of the tree entries. 492 std::vector<TreeEntry> VectorizableTree; 493 494 /// Maps a specific scalar to its tree entry. 495 SmallDenseMap<Value*, int> ScalarToTreeEntry; 496 497 /// A list of scalars that we found that we need to keep as scalars. 498 ValueSet MustGather; 499 500 /// This POD struct describes one external user in the vectorized tree. 501 struct ExternalUser { 502 ExternalUser (Value *S, llvm::User *U, int L) : 503 Scalar(S), User(U), Lane(L){}; 504 // Which scalar in our function. 505 Value *Scalar; 506 // Which user that uses the scalar. 507 llvm::User *User; 508 // Which lane does the scalar belong to. 509 int Lane; 510 }; 511 typedef SmallVector<ExternalUser, 16> UserList; 512 513 /// A list of values that need to extracted out of the tree. 514 /// This list holds pairs of (Internal Scalar : External User). 515 UserList ExternalUses; 516 517 /// A list of instructions to ignore while sinking 518 /// memory instructions. This map must be reset between runs of getCost. 519 ValueSet MemBarrierIgnoreList; 520 521 /// Holds all of the instructions that we gathered. 522 SetVector<Instruction *> GatherSeq; 523 /// A list of blocks that we are going to CSE. 524 SmallSet<BasicBlock *, 8> CSEBlocks; 525 526 /// Numbers instructions in different blocks. 527 DenseMap<BasicBlock *, BlockNumbering> BlocksNumbers; 528 529 /// Reduction operators. 530 ValueSet *RdxOps; 531 532 // Analysis and block reference. 533 Function *F; 534 ScalarEvolution *SE; 535 DataLayout *DL; 536 TargetTransformInfo *TTI; 537 AliasAnalysis *AA; 538 LoopInfo *LI; 539 DominatorTree *DT; 540 /// Instruction builder to construct the vectorized tree. 541 IRBuilder<> Builder; 542}; 543 544void BoUpSLP::buildTree(ArrayRef<Value *> Roots, ValueSet *Rdx) { 545 deleteTree(); 546 RdxOps = Rdx; 547 if (!getSameType(Roots)) 548 return; 549 buildTree_rec(Roots, 0); 550 551 // Collect the values that we need to extract from the tree. 552 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) { 553 TreeEntry *Entry = &VectorizableTree[EIdx]; 554 555 // For each lane: 556 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { 557 Value *Scalar = Entry->Scalars[Lane]; 558 559 // No need to handle users of gathered values. 560 if (Entry->NeedToGather) 561 continue; 562 563 for (Value::use_iterator User = Scalar->use_begin(), 564 UE = Scalar->use_end(); User != UE; ++User) { 565 DEBUG(dbgs() << "SLP: Checking user:" << **User << ".\n"); 566 567 bool Gathered = MustGather.count(*User); 568 569 // Skip in-tree scalars that become vectors. 570 if (ScalarToTreeEntry.count(*User) && !Gathered) { 571 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << 572 **User << ".\n"); 573 int Idx = ScalarToTreeEntry[*User]; (void) Idx; 574 assert(!VectorizableTree[Idx].NeedToGather && "Bad state"); 575 continue; 576 } 577 Instruction *UserInst = dyn_cast<Instruction>(*User); 578 if (!UserInst) 579 continue; 580 581 // Ignore uses that are part of the reduction. 582 if (Rdx && std::find(Rdx->begin(), Rdx->end(), UserInst) != Rdx->end()) 583 continue; 584 585 DEBUG(dbgs() << "SLP: Need to extract:" << **User << " from lane " << 586 Lane << " from " << *Scalar << ".\n"); 587 ExternalUses.push_back(ExternalUser(Scalar, *User, Lane)); 588 } 589 } 590 } 591} 592 593 594void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) { 595 bool SameTy = getSameType(VL); (void)SameTy; 596 assert(SameTy && "Invalid types!"); 597 598 if (Depth == RecursionMaxDepth) { 599 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n"); 600 newTreeEntry(VL, false); 601 return; 602 } 603 604 // Don't handle vectors. 605 if (VL[0]->getType()->isVectorTy()) { 606 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n"); 607 newTreeEntry(VL, false); 608 return; 609 } 610 611 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) 612 if (SI->getValueOperand()->getType()->isVectorTy()) { 613 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n"); 614 newTreeEntry(VL, false); 615 return; 616 } 617 618 // If all of the operands are identical or constant we have a simple solution. 619 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) || 620 !getSameOpcode(VL)) { 621 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n"); 622 newTreeEntry(VL, false); 623 return; 624 } 625 626 // We now know that this is a vector of instructions of the same type from 627 // the same block. 628 629 // Check if this is a duplicate of another entry. 630 if (ScalarToTreeEntry.count(VL[0])) { 631 int Idx = ScalarToTreeEntry[VL[0]]; 632 TreeEntry *E = &VectorizableTree[Idx]; 633 for (unsigned i = 0, e = VL.size(); i != e; ++i) { 634 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n"); 635 if (E->Scalars[i] != VL[i]) { 636 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n"); 637 newTreeEntry(VL, false); 638 return; 639 } 640 } 641 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n"); 642 return; 643 } 644 645 // Check that none of the instructions in the bundle are already in the tree. 646 for (unsigned i = 0, e = VL.size(); i != e; ++i) { 647 if (ScalarToTreeEntry.count(VL[i])) { 648 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] << 649 ") is already in tree.\n"); 650 newTreeEntry(VL, false); 651 return; 652 } 653 } 654 655 // If any of the scalars appears in the table OR it is marked as a value that 656 // needs to stat scalar then we need to gather the scalars. 657 for (unsigned i = 0, e = VL.size(); i != e; ++i) { 658 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) { 659 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n"); 660 newTreeEntry(VL, false); 661 return; 662 } 663 } 664 665 // Check that all of the users of the scalars that we want to vectorize are 666 // schedulable. 667 Instruction *VL0 = cast<Instruction>(VL[0]); 668 int MyLastIndex = getLastIndex(VL); 669 BasicBlock *BB = cast<Instruction>(VL0)->getParent(); 670 671 for (unsigned i = 0, e = VL.size(); i != e; ++i) { 672 Instruction *Scalar = cast<Instruction>(VL[i]); 673 DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n"); 674 for (Value::use_iterator U = Scalar->use_begin(), UE = Scalar->use_end(); 675 U != UE; ++U) { 676 DEBUG(dbgs() << "SLP: \tUser " << **U << ". \n"); 677 Instruction *User = dyn_cast<Instruction>(*U); 678 if (!User) { 679 DEBUG(dbgs() << "SLP: Gathering due unknown user. \n"); 680 newTreeEntry(VL, false); 681 return; 682 } 683 684 // We don't care if the user is in a different basic block. 685 BasicBlock *UserBlock = User->getParent(); 686 if (UserBlock != BB) { 687 DEBUG(dbgs() << "SLP: User from a different basic block " 688 << *User << ". \n"); 689 continue; 690 } 691 692 // If this is a PHINode within this basic block then we can place the 693 // extract wherever we want. 694 if (isa<PHINode>(*User)) { 695 DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *User << ". \n"); 696 continue; 697 } 698 699 // Check if this is a safe in-tree user. 700 if (ScalarToTreeEntry.count(User)) { 701 int Idx = ScalarToTreeEntry[User]; 702 int VecLocation = VectorizableTree[Idx].LastScalarIndex; 703 if (VecLocation <= MyLastIndex) { 704 DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n"); 705 newTreeEntry(VL, false); 706 return; 707 } 708 DEBUG(dbgs() << "SLP: In-tree user (" << *User << ") at #" << 709 VecLocation << " vector value (" << *Scalar << ") at #" 710 << MyLastIndex << ".\n"); 711 continue; 712 } 713 714 // This user is part of the reduction. 715 if (RdxOps && RdxOps->count(User)) 716 continue; 717 718 // Make sure that we can schedule this unknown user. 719 BlockNumbering &BN = BlocksNumbers[BB]; 720 int UserIndex = BN.getIndex(User); 721 if (UserIndex < MyLastIndex) { 722 723 DEBUG(dbgs() << "SLP: Can't schedule extractelement for " 724 << *User << ". \n"); 725 newTreeEntry(VL, false); 726 return; 727 } 728 } 729 } 730 731 // Check that every instructions appears once in this bundle. 732 for (unsigned i = 0, e = VL.size(); i < e; ++i) 733 for (unsigned j = i+1; j < e; ++j) 734 if (VL[i] == VL[j]) { 735 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n"); 736 newTreeEntry(VL, false); 737 return; 738 } 739 740 // Check that instructions in this bundle don't reference other instructions. 741 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4. 742 for (unsigned i = 0, e = VL.size(); i < e; ++i) { 743 for (Value::use_iterator U = VL[i]->use_begin(), UE = VL[i]->use_end(); 744 U != UE; ++U) { 745 for (unsigned j = 0; j < e; ++j) { 746 if (i != j && *U == VL[j]) { 747 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << **U << ". \n"); 748 newTreeEntry(VL, false); 749 return; 750 } 751 } 752 } 753 } 754 755 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n"); 756 757 unsigned Opcode = getSameOpcode(VL); 758 759 // Check if it is safe to sink the loads or the stores. 760 if (Opcode == Instruction::Load || Opcode == Instruction::Store) { 761 Instruction *Last = getLastInstruction(VL); 762 763 for (unsigned i = 0, e = VL.size(); i < e; ++i) { 764 if (VL[i] == Last) 765 continue; 766 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last); 767 if (Barrier) { 768 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last 769 << "\n because of " << *Barrier << ". Gathering.\n"); 770 newTreeEntry(VL, false); 771 return; 772 } 773 } 774 } 775 776 switch (Opcode) { 777 case Instruction::PHI: { 778 PHINode *PH = dyn_cast<PHINode>(VL0); 779 780 // Check for terminator values (e.g. invoke). 781 for (unsigned j = 0; j < VL.size(); ++j) 782 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { 783 TerminatorInst *Term = dyn_cast<TerminatorInst>(cast<PHINode>(VL[j])->getIncomingValue(i)); 784 if (Term) { 785 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n"); 786 newTreeEntry(VL, false); 787 return; 788 } 789 } 790 791 newTreeEntry(VL, true); 792 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n"); 793 794 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { 795 ValueList Operands; 796 // Prepare the operand vector. 797 for (unsigned j = 0; j < VL.size(); ++j) 798 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValue(i)); 799 800 buildTree_rec(Operands, Depth + 1); 801 } 802 return; 803 } 804 case Instruction::ExtractElement: { 805 bool Reuse = CanReuseExtract(VL); 806 if (Reuse) { 807 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n"); 808 } 809 newTreeEntry(VL, Reuse); 810 return; 811 } 812 case Instruction::Load: { 813 // Check if the loads are consecutive or of we need to swizzle them. 814 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) { 815 LoadInst *L = cast<LoadInst>(VL[i]); 816 if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) { 817 newTreeEntry(VL, false); 818 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n"); 819 return; 820 } 821 } 822 newTreeEntry(VL, true); 823 DEBUG(dbgs() << "SLP: added a vector of loads.\n"); 824 return; 825 } 826 case Instruction::ZExt: 827 case Instruction::SExt: 828 case Instruction::FPToUI: 829 case Instruction::FPToSI: 830 case Instruction::FPExt: 831 case Instruction::PtrToInt: 832 case Instruction::IntToPtr: 833 case Instruction::SIToFP: 834 case Instruction::UIToFP: 835 case Instruction::Trunc: 836 case Instruction::FPTrunc: 837 case Instruction::BitCast: { 838 Type *SrcTy = VL0->getOperand(0)->getType(); 839 for (unsigned i = 0; i < VL.size(); ++i) { 840 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType(); 841 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) { 842 newTreeEntry(VL, false); 843 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n"); 844 return; 845 } 846 } 847 newTreeEntry(VL, true); 848 DEBUG(dbgs() << "SLP: added a vector of casts.\n"); 849 850 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { 851 ValueList Operands; 852 // Prepare the operand vector. 853 for (unsigned j = 0; j < VL.size(); ++j) 854 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i)); 855 856 buildTree_rec(Operands, Depth+1); 857 } 858 return; 859 } 860 case Instruction::ICmp: 861 case Instruction::FCmp: { 862 // Check that all of the compares have the same predicate. 863 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate(); 864 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType(); 865 for (unsigned i = 1, e = VL.size(); i < e; ++i) { 866 CmpInst *Cmp = cast<CmpInst>(VL[i]); 867 if (Cmp->getPredicate() != P0 || 868 Cmp->getOperand(0)->getType() != ComparedTy) { 869 newTreeEntry(VL, false); 870 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n"); 871 return; 872 } 873 } 874 875 newTreeEntry(VL, true); 876 DEBUG(dbgs() << "SLP: added a vector of compares.\n"); 877 878 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { 879 ValueList Operands; 880 // Prepare the operand vector. 881 for (unsigned j = 0; j < VL.size(); ++j) 882 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i)); 883 884 buildTree_rec(Operands, Depth+1); 885 } 886 return; 887 } 888 case Instruction::Select: 889 case Instruction::Add: 890 case Instruction::FAdd: 891 case Instruction::Sub: 892 case Instruction::FSub: 893 case Instruction::Mul: 894 case Instruction::FMul: 895 case Instruction::UDiv: 896 case Instruction::SDiv: 897 case Instruction::FDiv: 898 case Instruction::URem: 899 case Instruction::SRem: 900 case Instruction::FRem: 901 case Instruction::Shl: 902 case Instruction::LShr: 903 case Instruction::AShr: 904 case Instruction::And: 905 case Instruction::Or: 906 case Instruction::Xor: { 907 newTreeEntry(VL, true); 908 DEBUG(dbgs() << "SLP: added a vector of bin op.\n"); 909 910 // Sort operands of the instructions so that each side is more likely to 911 // have the same opcode. 912 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { 913 ValueList Left, Right; 914 reorderInputsAccordingToOpcode(VL, Left, Right); 915 buildTree_rec(Left, Depth + 1); 916 buildTree_rec(Right, Depth + 1); 917 return; 918 } 919 920 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { 921 ValueList Operands; 922 // Prepare the operand vector. 923 for (unsigned j = 0; j < VL.size(); ++j) 924 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i)); 925 926 buildTree_rec(Operands, Depth+1); 927 } 928 return; 929 } 930 case Instruction::Store: { 931 // Check if the stores are consecutive or of we need to swizzle them. 932 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) 933 if (!isConsecutiveAccess(VL[i], VL[i + 1])) { 934 newTreeEntry(VL, false); 935 DEBUG(dbgs() << "SLP: Non consecutive store.\n"); 936 return; 937 } 938 939 newTreeEntry(VL, true); 940 DEBUG(dbgs() << "SLP: added a vector of stores.\n"); 941 942 ValueList Operands; 943 for (unsigned j = 0; j < VL.size(); ++j) 944 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0)); 945 946 // We can ignore these values because we are sinking them down. 947 MemBarrierIgnoreList.insert(VL.begin(), VL.end()); 948 buildTree_rec(Operands, Depth + 1); 949 return; 950 } 951 default: 952 newTreeEntry(VL, false); 953 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n"); 954 return; 955 } 956} 957 958int BoUpSLP::getEntryCost(TreeEntry *E) { 959 ArrayRef<Value*> VL = E->Scalars; 960 961 Type *ScalarTy = VL[0]->getType(); 962 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) 963 ScalarTy = SI->getValueOperand()->getType(); 964 VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); 965 966 if (E->NeedToGather) { 967 if (allConstant(VL)) 968 return 0; 969 if (isSplat(VL)) { 970 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0); 971 } 972 return getGatherCost(E->Scalars); 973 } 974 975 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) && 976 "Invalid VL"); 977 Instruction *VL0 = cast<Instruction>(VL[0]); 978 unsigned Opcode = VL0->getOpcode(); 979 switch (Opcode) { 980 case Instruction::PHI: { 981 return 0; 982 } 983 case Instruction::ExtractElement: { 984 if (CanReuseExtract(VL)) 985 return 0; 986 return getGatherCost(VecTy); 987 } 988 case Instruction::ZExt: 989 case Instruction::SExt: 990 case Instruction::FPToUI: 991 case Instruction::FPToSI: 992 case Instruction::FPExt: 993 case Instruction::PtrToInt: 994 case Instruction::IntToPtr: 995 case Instruction::SIToFP: 996 case Instruction::UIToFP: 997 case Instruction::Trunc: 998 case Instruction::FPTrunc: 999 case Instruction::BitCast: { 1000 Type *SrcTy = VL0->getOperand(0)->getType(); 1001 1002 // Calculate the cost of this instruction. 1003 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(), 1004 VL0->getType(), SrcTy); 1005 1006 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size()); 1007 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy); 1008 return VecCost - ScalarCost; 1009 } 1010 case Instruction::FCmp: 1011 case Instruction::ICmp: 1012 case Instruction::Select: 1013 case Instruction::Add: 1014 case Instruction::FAdd: 1015 case Instruction::Sub: 1016 case Instruction::FSub: 1017 case Instruction::Mul: 1018 case Instruction::FMul: 1019 case Instruction::UDiv: 1020 case Instruction::SDiv: 1021 case Instruction::FDiv: 1022 case Instruction::URem: 1023 case Instruction::SRem: 1024 case Instruction::FRem: 1025 case Instruction::Shl: 1026 case Instruction::LShr: 1027 case Instruction::AShr: 1028 case Instruction::And: 1029 case Instruction::Or: 1030 case Instruction::Xor: { 1031 // Calculate the cost of this instruction. 1032 int ScalarCost = 0; 1033 int VecCost = 0; 1034 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp || 1035 Opcode == Instruction::Select) { 1036 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size()); 1037 ScalarCost = VecTy->getNumElements() * 1038 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty()); 1039 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy); 1040 } else { 1041 // Certain instructions can be cheaper to vectorize if they have a 1042 // constant second vector operand. 1043 TargetTransformInfo::OperandValueKind Op1VK = 1044 TargetTransformInfo::OK_AnyValue; 1045 TargetTransformInfo::OperandValueKind Op2VK = 1046 TargetTransformInfo::OK_UniformConstantValue; 1047 1048 // Check whether all second operands are constant. 1049 for (unsigned i = 0; i < VL.size(); ++i) 1050 if (!isa<ConstantInt>(cast<Instruction>(VL[i])->getOperand(1))) { 1051 Op2VK = TargetTransformInfo::OK_AnyValue; 1052 break; 1053 } 1054 1055 ScalarCost = 1056 VecTy->getNumElements() * 1057 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK); 1058 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK); 1059 } 1060 return VecCost - ScalarCost; 1061 } 1062 case Instruction::Load: { 1063 // Cost of wide load - cost of scalar loads. 1064 int ScalarLdCost = VecTy->getNumElements() * 1065 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0); 1066 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0); 1067 return VecLdCost - ScalarLdCost; 1068 } 1069 case Instruction::Store: { 1070 // We know that we can merge the stores. Calculate the cost. 1071 int ScalarStCost = VecTy->getNumElements() * 1072 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0); 1073 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0); 1074 return VecStCost - ScalarStCost; 1075 } 1076 default: 1077 llvm_unreachable("Unknown instruction"); 1078 } 1079} 1080 1081bool BoUpSLP::isFullyVectorizableTinyTree() { 1082 DEBUG(dbgs() << "SLP: Check whether the tree with height " << 1083 VectorizableTree.size() << " is fully vectorizable .\n"); 1084 1085 // We only handle trees of height 2. 1086 if (VectorizableTree.size() != 2) 1087 return false; 1088 1089 // Gathering cost would be too much for tiny trees. 1090 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather) 1091 return false; 1092 1093 return true; 1094} 1095 1096int BoUpSLP::getTreeCost() { 1097 int Cost = 0; 1098 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " << 1099 VectorizableTree.size() << ".\n"); 1100 1101 // We only vectorize tiny trees if it is fully vectorizable. 1102 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) { 1103 if (!VectorizableTree.size()) { 1104 assert(!ExternalUses.size() && "We should not have any external users"); 1105 } 1106 return INT_MAX; 1107 } 1108 1109 unsigned BundleWidth = VectorizableTree[0].Scalars.size(); 1110 1111 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) { 1112 int C = getEntryCost(&VectorizableTree[i]); 1113 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " 1114 << *VectorizableTree[i].Scalars[0] << " .\n"); 1115 Cost += C; 1116 } 1117 1118 int ExtractCost = 0; 1119 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end(); 1120 I != E; ++I) { 1121 1122 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth); 1123 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, 1124 I->Lane); 1125 } 1126 1127 1128 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n"); 1129 return Cost + ExtractCost; 1130} 1131 1132int BoUpSLP::getGatherCost(Type *Ty) { 1133 int Cost = 0; 1134 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i) 1135 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i); 1136 return Cost; 1137} 1138 1139int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) { 1140 // Find the type of the operands in VL. 1141 Type *ScalarTy = VL[0]->getType(); 1142 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) 1143 ScalarTy = SI->getValueOperand()->getType(); 1144 VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); 1145 // Find the cost of inserting/extracting values from the vector. 1146 return getGatherCost(VecTy); 1147} 1148 1149AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) { 1150 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 1151 return AA->getLocation(SI); 1152 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1153 return AA->getLocation(LI); 1154 return AliasAnalysis::Location(); 1155} 1156 1157Value *BoUpSLP::getPointerOperand(Value *I) { 1158 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1159 return LI->getPointerOperand(); 1160 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 1161 return SI->getPointerOperand(); 1162 return 0; 1163} 1164 1165unsigned BoUpSLP::getAddressSpaceOperand(Value *I) { 1166 if (LoadInst *L = dyn_cast<LoadInst>(I)) 1167 return L->getPointerAddressSpace(); 1168 if (StoreInst *S = dyn_cast<StoreInst>(I)) 1169 return S->getPointerAddressSpace(); 1170 return -1; 1171} 1172 1173bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) { 1174 Value *PtrA = getPointerOperand(A); 1175 Value *PtrB = getPointerOperand(B); 1176 unsigned ASA = getAddressSpaceOperand(A); 1177 unsigned ASB = getAddressSpaceOperand(B); 1178 1179 // Check that the address spaces match and that the pointers are valid. 1180 if (!PtrA || !PtrB || (ASA != ASB)) 1181 return false; 1182 1183 // Make sure that A and B are different pointers of the same type. 1184 if (PtrA == PtrB || PtrA->getType() != PtrB->getType()) 1185 return false; 1186 1187 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA); 1188 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType(); 1189 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty)); 1190 1191 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0); 1192 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA); 1193 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB); 1194 1195 APInt OffsetDelta = OffsetB - OffsetA; 1196 1197 // Check if they are based on the same pointer. That makes the offsets 1198 // sufficient. 1199 if (PtrA == PtrB) 1200 return OffsetDelta == Size; 1201 1202 // Compute the necessary base pointer delta to have the necessary final delta 1203 // equal to the size. 1204 APInt BaseDelta = Size - OffsetDelta; 1205 1206 // Otherwise compute the distance with SCEV between the base pointers. 1207 const SCEV *PtrSCEVA = SE->getSCEV(PtrA); 1208 const SCEV *PtrSCEVB = SE->getSCEV(PtrB); 1209 const SCEV *C = SE->getConstant(BaseDelta); 1210 const SCEV *X = SE->getAddExpr(PtrSCEVA, C); 1211 return X == PtrSCEVB; 1212} 1213 1214Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) { 1215 assert(Src->getParent() == Dst->getParent() && "Not the same BB"); 1216 BasicBlock::iterator I = Src, E = Dst; 1217 /// Scan all of the instruction from SRC to DST and check if 1218 /// the source may alias. 1219 for (++I; I != E; ++I) { 1220 // Ignore store instructions that are marked as 'ignore'. 1221 if (MemBarrierIgnoreList.count(I)) 1222 continue; 1223 if (Src->mayWriteToMemory()) /* Write */ { 1224 if (!I->mayReadOrWriteMemory()) 1225 continue; 1226 } else /* Read */ { 1227 if (!I->mayWriteToMemory()) 1228 continue; 1229 } 1230 AliasAnalysis::Location A = getLocation(&*I); 1231 AliasAnalysis::Location B = getLocation(Src); 1232 1233 if (!A.Ptr || !B.Ptr || AA->alias(A, B)) 1234 return I; 1235 } 1236 return 0; 1237} 1238 1239int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) { 1240 BasicBlock *BB = cast<Instruction>(VL[0])->getParent(); 1241 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block"); 1242 BlockNumbering &BN = BlocksNumbers[BB]; 1243 1244 int MaxIdx = BN.getIndex(BB->getFirstNonPHI()); 1245 for (unsigned i = 0, e = VL.size(); i < e; ++i) 1246 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i]))); 1247 return MaxIdx; 1248} 1249 1250Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) { 1251 BasicBlock *BB = cast<Instruction>(VL[0])->getParent(); 1252 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block"); 1253 BlockNumbering &BN = BlocksNumbers[BB]; 1254 1255 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0])); 1256 for (unsigned i = 1, e = VL.size(); i < e; ++i) 1257 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i]))); 1258 Instruction *I = BN.getInstruction(MaxIdx); 1259 assert(I && "bad location"); 1260 return I; 1261} 1262 1263void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) { 1264 Instruction *VL0 = cast<Instruction>(VL[0]); 1265 Instruction *LastInst = getLastInstruction(VL); 1266 BasicBlock::iterator NextInst = LastInst; 1267 ++NextInst; 1268 Builder.SetInsertPoint(VL0->getParent(), NextInst); 1269 Builder.SetCurrentDebugLocation(VL0->getDebugLoc()); 1270} 1271 1272Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) { 1273 Value *Vec = UndefValue::get(Ty); 1274 // Generate the 'InsertElement' instruction. 1275 for (unsigned i = 0; i < Ty->getNumElements(); ++i) { 1276 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i)); 1277 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) { 1278 GatherSeq.insert(Insrt); 1279 CSEBlocks.insert(Insrt->getParent()); 1280 1281 // Add to our 'need-to-extract' list. 1282 if (ScalarToTreeEntry.count(VL[i])) { 1283 int Idx = ScalarToTreeEntry[VL[i]]; 1284 TreeEntry *E = &VectorizableTree[Idx]; 1285 // Find which lane we need to extract. 1286 int FoundLane = -1; 1287 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) { 1288 // Is this the lane of the scalar that we are looking for ? 1289 if (E->Scalars[Lane] == VL[i]) { 1290 FoundLane = Lane; 1291 break; 1292 } 1293 } 1294 assert(FoundLane >= 0 && "Could not find the correct lane"); 1295 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane)); 1296 } 1297 } 1298 } 1299 1300 return Vec; 1301} 1302 1303Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const { 1304 SmallDenseMap<Value*, int>::const_iterator Entry 1305 = ScalarToTreeEntry.find(VL[0]); 1306 if (Entry != ScalarToTreeEntry.end()) { 1307 int Idx = Entry->second; 1308 const TreeEntry *En = &VectorizableTree[Idx]; 1309 if (En->isSame(VL) && En->VectorizedValue) 1310 return En->VectorizedValue; 1311 } 1312 return 0; 1313} 1314 1315Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { 1316 if (ScalarToTreeEntry.count(VL[0])) { 1317 int Idx = ScalarToTreeEntry[VL[0]]; 1318 TreeEntry *E = &VectorizableTree[Idx]; 1319 if (E->isSame(VL)) 1320 return vectorizeTree(E); 1321 } 1322 1323 Type *ScalarTy = VL[0]->getType(); 1324 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) 1325 ScalarTy = SI->getValueOperand()->getType(); 1326 VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); 1327 1328 return Gather(VL, VecTy); 1329} 1330 1331Value *BoUpSLP::vectorizeTree(TreeEntry *E) { 1332 IRBuilder<>::InsertPointGuard Guard(Builder); 1333 1334 if (E->VectorizedValue) { 1335 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n"); 1336 return E->VectorizedValue; 1337 } 1338 1339 Instruction *VL0 = cast<Instruction>(E->Scalars[0]); 1340 Type *ScalarTy = VL0->getType(); 1341 if (StoreInst *SI = dyn_cast<StoreInst>(VL0)) 1342 ScalarTy = SI->getValueOperand()->getType(); 1343 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size()); 1344 1345 if (E->NeedToGather) { 1346 setInsertPointAfterBundle(E->Scalars); 1347 return Gather(E->Scalars, VecTy); 1348 } 1349 1350 unsigned Opcode = VL0->getOpcode(); 1351 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode"); 1352 1353 switch (Opcode) { 1354 case Instruction::PHI: { 1355 PHINode *PH = dyn_cast<PHINode>(VL0); 1356 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); 1357 Builder.SetCurrentDebugLocation(PH->getDebugLoc()); 1358 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); 1359 E->VectorizedValue = NewPhi; 1360 1361 // PHINodes may have multiple entries from the same block. We want to 1362 // visit every block once. 1363 SmallSet<BasicBlock*, 4> VisitedBBs; 1364 1365 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { 1366 ValueList Operands; 1367 BasicBlock *IBB = PH->getIncomingBlock(i); 1368 1369 if (!VisitedBBs.insert(IBB)) { 1370 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); 1371 continue; 1372 } 1373 1374 // Prepare the operand vector. 1375 for (unsigned j = 0; j < E->Scalars.size(); ++j) 1376 Operands.push_back(cast<PHINode>(E->Scalars[j])-> 1377 getIncomingValueForBlock(IBB)); 1378 1379 Builder.SetInsertPoint(IBB->getTerminator()); 1380 Builder.SetCurrentDebugLocation(PH->getDebugLoc()); 1381 Value *Vec = vectorizeTree(Operands); 1382 NewPhi->addIncoming(Vec, IBB); 1383 } 1384 1385 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && 1386 "Invalid number of incoming values"); 1387 return NewPhi; 1388 } 1389 1390 case Instruction::ExtractElement: { 1391 if (CanReuseExtract(E->Scalars)) { 1392 Value *V = VL0->getOperand(0); 1393 E->VectorizedValue = V; 1394 return V; 1395 } 1396 return Gather(E->Scalars, VecTy); 1397 } 1398 case Instruction::ZExt: 1399 case Instruction::SExt: 1400 case Instruction::FPToUI: 1401 case Instruction::FPToSI: 1402 case Instruction::FPExt: 1403 case Instruction::PtrToInt: 1404 case Instruction::IntToPtr: 1405 case Instruction::SIToFP: 1406 case Instruction::UIToFP: 1407 case Instruction::Trunc: 1408 case Instruction::FPTrunc: 1409 case Instruction::BitCast: { 1410 ValueList INVL; 1411 for (int i = 0, e = E->Scalars.size(); i < e; ++i) 1412 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0)); 1413 1414 setInsertPointAfterBundle(E->Scalars); 1415 1416 Value *InVec = vectorizeTree(INVL); 1417 1418 if (Value *V = alreadyVectorized(E->Scalars)) 1419 return V; 1420 1421 CastInst *CI = dyn_cast<CastInst>(VL0); 1422 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); 1423 E->VectorizedValue = V; 1424 return V; 1425 } 1426 case Instruction::FCmp: 1427 case Instruction::ICmp: { 1428 ValueList LHSV, RHSV; 1429 for (int i = 0, e = E->Scalars.size(); i < e; ++i) { 1430 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0)); 1431 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1)); 1432 } 1433 1434 setInsertPointAfterBundle(E->Scalars); 1435 1436 Value *L = vectorizeTree(LHSV); 1437 Value *R = vectorizeTree(RHSV); 1438 1439 if (Value *V = alreadyVectorized(E->Scalars)) 1440 return V; 1441 1442 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate(); 1443 Value *V; 1444 if (Opcode == Instruction::FCmp) 1445 V = Builder.CreateFCmp(P0, L, R); 1446 else 1447 V = Builder.CreateICmp(P0, L, R); 1448 1449 E->VectorizedValue = V; 1450 return V; 1451 } 1452 case Instruction::Select: { 1453 ValueList TrueVec, FalseVec, CondVec; 1454 for (int i = 0, e = E->Scalars.size(); i < e; ++i) { 1455 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0)); 1456 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1)); 1457 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2)); 1458 } 1459 1460 setInsertPointAfterBundle(E->Scalars); 1461 1462 Value *Cond = vectorizeTree(CondVec); 1463 Value *True = vectorizeTree(TrueVec); 1464 Value *False = vectorizeTree(FalseVec); 1465 1466 if (Value *V = alreadyVectorized(E->Scalars)) 1467 return V; 1468 1469 Value *V = Builder.CreateSelect(Cond, True, False); 1470 E->VectorizedValue = V; 1471 return V; 1472 } 1473 case Instruction::Add: 1474 case Instruction::FAdd: 1475 case Instruction::Sub: 1476 case Instruction::FSub: 1477 case Instruction::Mul: 1478 case Instruction::FMul: 1479 case Instruction::UDiv: 1480 case Instruction::SDiv: 1481 case Instruction::FDiv: 1482 case Instruction::URem: 1483 case Instruction::SRem: 1484 case Instruction::FRem: 1485 case Instruction::Shl: 1486 case Instruction::LShr: 1487 case Instruction::AShr: 1488 case Instruction::And: 1489 case Instruction::Or: 1490 case Instruction::Xor: { 1491 ValueList LHSVL, RHSVL; 1492 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) 1493 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL); 1494 else 1495 for (int i = 0, e = E->Scalars.size(); i < e; ++i) { 1496 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0)); 1497 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1)); 1498 } 1499 1500 setInsertPointAfterBundle(E->Scalars); 1501 1502 Value *LHS = vectorizeTree(LHSVL); 1503 Value *RHS = vectorizeTree(RHSVL); 1504 1505 if (LHS == RHS && isa<Instruction>(LHS)) { 1506 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order"); 1507 } 1508 1509 if (Value *V = alreadyVectorized(E->Scalars)) 1510 return V; 1511 1512 BinaryOperator *BinOp = cast<BinaryOperator>(VL0); 1513 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS); 1514 E->VectorizedValue = V; 1515 1516 if (Instruction *I = dyn_cast<Instruction>(V)) 1517 return propagateMetadata(I, E->Scalars); 1518 1519 return V; 1520 } 1521 case Instruction::Load: { 1522 // Loads are inserted at the head of the tree because we don't want to 1523 // sink them all the way down past store instructions. 1524 setInsertPointAfterBundle(E->Scalars); 1525 1526 LoadInst *LI = cast<LoadInst>(VL0); 1527 unsigned AS = LI->getPointerAddressSpace(); 1528 1529 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(), 1530 VecTy->getPointerTo(AS)); 1531 unsigned Alignment = LI->getAlignment(); 1532 LI = Builder.CreateLoad(VecPtr); 1533 LI->setAlignment(Alignment); 1534 E->VectorizedValue = LI; 1535 return propagateMetadata(LI, E->Scalars); 1536 } 1537 case Instruction::Store: { 1538 StoreInst *SI = cast<StoreInst>(VL0); 1539 unsigned Alignment = SI->getAlignment(); 1540 unsigned AS = SI->getPointerAddressSpace(); 1541 1542 ValueList ValueOp; 1543 for (int i = 0, e = E->Scalars.size(); i < e; ++i) 1544 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand()); 1545 1546 setInsertPointAfterBundle(E->Scalars); 1547 1548 Value *VecValue = vectorizeTree(ValueOp); 1549 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(), 1550 VecTy->getPointerTo(AS)); 1551 StoreInst *S = Builder.CreateStore(VecValue, VecPtr); 1552 S->setAlignment(Alignment); 1553 E->VectorizedValue = S; 1554 return propagateMetadata(S, E->Scalars); 1555 } 1556 default: 1557 llvm_unreachable("unknown inst"); 1558 } 1559 return 0; 1560} 1561 1562Value *BoUpSLP::vectorizeTree() { 1563 Builder.SetInsertPoint(F->getEntryBlock().begin()); 1564 vectorizeTree(&VectorizableTree[0]); 1565 1566 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n"); 1567 1568 // Extract all of the elements with the external uses. 1569 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end(); 1570 it != e; ++it) { 1571 Value *Scalar = it->Scalar; 1572 llvm::User *User = it->User; 1573 1574 // Skip users that we already RAUW. This happens when one instruction 1575 // has multiple uses of the same value. 1576 if (std::find(Scalar->use_begin(), Scalar->use_end(), User) == 1577 Scalar->use_end()) 1578 continue; 1579 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar"); 1580 1581 int Idx = ScalarToTreeEntry[Scalar]; 1582 TreeEntry *E = &VectorizableTree[Idx]; 1583 assert(!E->NeedToGather && "Extracting from a gather list"); 1584 1585 Value *Vec = E->VectorizedValue; 1586 assert(Vec && "Can't find vectorizable value"); 1587 1588 Value *Lane = Builder.getInt32(it->Lane); 1589 // Generate extracts for out-of-tree users. 1590 // Find the insertion point for the extractelement lane. 1591 if (PHINode *PN = dyn_cast<PHINode>(Vec)) { 1592 Builder.SetInsertPoint(PN->getParent()->getFirstInsertionPt()); 1593 Value *Ex = Builder.CreateExtractElement(Vec, Lane); 1594 CSEBlocks.insert(PN->getParent()); 1595 User->replaceUsesOfWith(Scalar, Ex); 1596 } else if (isa<Instruction>(Vec)){ 1597 if (PHINode *PH = dyn_cast<PHINode>(User)) { 1598 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { 1599 if (PH->getIncomingValue(i) == Scalar) { 1600 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); 1601 Value *Ex = Builder.CreateExtractElement(Vec, Lane); 1602 CSEBlocks.insert(PH->getIncomingBlock(i)); 1603 PH->setOperand(i, Ex); 1604 } 1605 } 1606 } else { 1607 Builder.SetInsertPoint(cast<Instruction>(User)); 1608 Value *Ex = Builder.CreateExtractElement(Vec, Lane); 1609 CSEBlocks.insert(cast<Instruction>(User)->getParent()); 1610 User->replaceUsesOfWith(Scalar, Ex); 1611 } 1612 } else { 1613 Builder.SetInsertPoint(F->getEntryBlock().begin()); 1614 Value *Ex = Builder.CreateExtractElement(Vec, Lane); 1615 CSEBlocks.insert(&F->getEntryBlock()); 1616 User->replaceUsesOfWith(Scalar, Ex); 1617 } 1618 1619 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n"); 1620 } 1621 1622 // For each vectorized value: 1623 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) { 1624 TreeEntry *Entry = &VectorizableTree[EIdx]; 1625 1626 // For each lane: 1627 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { 1628 Value *Scalar = Entry->Scalars[Lane]; 1629 1630 // No need to handle users of gathered values. 1631 if (Entry->NeedToGather) 1632 continue; 1633 1634 assert(Entry->VectorizedValue && "Can't find vectorizable value"); 1635 1636 Type *Ty = Scalar->getType(); 1637 if (!Ty->isVoidTy()) { 1638 for (Value::use_iterator User = Scalar->use_begin(), 1639 UE = Scalar->use_end(); User != UE; ++User) { 1640 DEBUG(dbgs() << "SLP: \tvalidating user:" << **User << ".\n"); 1641 assert(!MustGather.count(*User) && 1642 "Replacing gathered value with undef"); 1643 1644 assert((ScalarToTreeEntry.count(*User) || 1645 // It is legal to replace the reduction users by undef. 1646 (RdxOps && RdxOps->count(*User))) && 1647 "Replacing out-of-tree value with undef"); 1648 } 1649 Value *Undef = UndefValue::get(Ty); 1650 Scalar->replaceAllUsesWith(Undef); 1651 } 1652 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n"); 1653 cast<Instruction>(Scalar)->eraseFromParent(); 1654 } 1655 } 1656 1657 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) { 1658 BlocksNumbers[it].forget(); 1659 } 1660 Builder.ClearInsertionPoint(); 1661 1662 return VectorizableTree[0].VectorizedValue; 1663} 1664 1665class DTCmp { 1666 const DominatorTree *DT; 1667 1668public: 1669 DTCmp(const DominatorTree *DT) : DT(DT) {} 1670 bool operator()(const BasicBlock *A, const BasicBlock *B) const { 1671 return DT->properlyDominates(A, B); 1672 } 1673}; 1674 1675void BoUpSLP::optimizeGatherSequence() { 1676 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size() 1677 << " gather sequences instructions.\n"); 1678 // LICM InsertElementInst sequences. 1679 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(), 1680 e = GatherSeq.end(); it != e; ++it) { 1681 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it); 1682 1683 if (!Insert) 1684 continue; 1685 1686 // Check if this block is inside a loop. 1687 Loop *L = LI->getLoopFor(Insert->getParent()); 1688 if (!L) 1689 continue; 1690 1691 // Check if it has a preheader. 1692 BasicBlock *PreHeader = L->getLoopPreheader(); 1693 if (!PreHeader) 1694 continue; 1695 1696 // If the vector or the element that we insert into it are 1697 // instructions that are defined in this basic block then we can't 1698 // hoist this instruction. 1699 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0)); 1700 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1)); 1701 if (CurrVec && L->contains(CurrVec)) 1702 continue; 1703 if (NewElem && L->contains(NewElem)) 1704 continue; 1705 1706 // We can hoist this instruction. Move it to the pre-header. 1707 Insert->moveBefore(PreHeader->getTerminator()); 1708 } 1709 1710 // Sort blocks by domination. This ensures we visit a block after all blocks 1711 // dominating it are visited. 1712 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end()); 1713 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(), DTCmp(DT)); 1714 1715 // Perform O(N^2) search over the gather sequences and merge identical 1716 // instructions. TODO: We can further optimize this scan if we split the 1717 // instructions into different buckets based on the insert lane. 1718 SmallVector<Instruction *, 16> Visited; 1719 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(), 1720 E = CSEWorkList.end(); 1721 I != E; ++I) { 1722 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *llvm::prior(I))) && 1723 "Worklist not sorted properly!"); 1724 BasicBlock *BB = *I; 1725 // For all instructions in blocks containing gather sequences: 1726 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) { 1727 Instruction *In = it++; 1728 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In)) 1729 continue; 1730 1731 // Check if we can replace this instruction with any of the 1732 // visited instructions. 1733 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(), 1734 ve = Visited.end(); 1735 v != ve; ++v) { 1736 if (In->isIdenticalTo(*v) && 1737 DT->dominates((*v)->getParent(), In->getParent())) { 1738 In->replaceAllUsesWith(*v); 1739 In->eraseFromParent(); 1740 In = 0; 1741 break; 1742 } 1743 } 1744 if (In) { 1745 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end()); 1746 Visited.push_back(In); 1747 } 1748 } 1749 } 1750 CSEBlocks.clear(); 1751 GatherSeq.clear(); 1752} 1753 1754/// The SLPVectorizer Pass. 1755struct SLPVectorizer : public FunctionPass { 1756 typedef SmallVector<StoreInst *, 8> StoreList; 1757 typedef MapVector<Value *, StoreList> StoreListMap; 1758 1759 /// Pass identification, replacement for typeid 1760 static char ID; 1761 1762 explicit SLPVectorizer() : FunctionPass(ID) { 1763 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); 1764 } 1765 1766 ScalarEvolution *SE; 1767 DataLayout *DL; 1768 TargetTransformInfo *TTI; 1769 AliasAnalysis *AA; 1770 LoopInfo *LI; 1771 DominatorTree *DT; 1772 1773 virtual bool runOnFunction(Function &F) { 1774 SE = &getAnalysis<ScalarEvolution>(); 1775 DL = getAnalysisIfAvailable<DataLayout>(); 1776 TTI = &getAnalysis<TargetTransformInfo>(); 1777 AA = &getAnalysis<AliasAnalysis>(); 1778 LI = &getAnalysis<LoopInfo>(); 1779 DT = &getAnalysis<DominatorTree>(); 1780 1781 StoreRefs.clear(); 1782 bool Changed = false; 1783 1784 // If the target claims to have no vector registers don't attempt 1785 // vectorization. 1786 if (!TTI->getNumberOfRegisters(true)) 1787 return false; 1788 1789 // Must have DataLayout. We can't require it because some tests run w/o 1790 // triple. 1791 if (!DL) 1792 return false; 1793 1794 // Don't vectorize when the attribute NoImplicitFloat is used. 1795 if (F.hasFnAttribute(Attribute::NoImplicitFloat)) 1796 return false; 1797 1798 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n"); 1799 1800 // Use the bollom up slp vectorizer to construct chains that start with 1801 // he store instructions. 1802 BoUpSLP R(&F, SE, DL, TTI, AA, LI, DT); 1803 1804 // Scan the blocks in the function in post order. 1805 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()), 1806 e = po_end(&F.getEntryBlock()); it != e; ++it) { 1807 BasicBlock *BB = *it; 1808 1809 // Vectorize trees that end at stores. 1810 if (unsigned count = collectStores(BB, R)) { 1811 (void)count; 1812 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n"); 1813 Changed |= vectorizeStoreChains(R); 1814 } 1815 1816 // Vectorize trees that end at reductions. 1817 Changed |= vectorizeChainsInBlock(BB, R); 1818 } 1819 1820 if (Changed) { 1821 R.optimizeGatherSequence(); 1822 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n"); 1823 DEBUG(verifyFunction(F)); 1824 } 1825 return Changed; 1826 } 1827 1828 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 1829 FunctionPass::getAnalysisUsage(AU); 1830 AU.addRequired<ScalarEvolution>(); 1831 AU.addRequired<AliasAnalysis>(); 1832 AU.addRequired<TargetTransformInfo>(); 1833 AU.addRequired<LoopInfo>(); 1834 AU.addRequired<DominatorTree>(); 1835 AU.addPreserved<LoopInfo>(); 1836 AU.addPreserved<DominatorTree>(); 1837 AU.setPreservesCFG(); 1838 } 1839 1840private: 1841 1842 /// \brief Collect memory references and sort them according to their base 1843 /// object. We sort the stores to their base objects to reduce the cost of the 1844 /// quadratic search on the stores. TODO: We can further reduce this cost 1845 /// if we flush the chain creation every time we run into a memory barrier. 1846 unsigned collectStores(BasicBlock *BB, BoUpSLP &R); 1847 1848 /// \brief Try to vectorize a chain that starts at two arithmetic instrs. 1849 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R); 1850 1851 /// \brief Try to vectorize a list of operands. 1852 /// \returns true if a value was vectorized. 1853 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R); 1854 1855 /// \brief Try to vectorize a chain that may start at the operands of \V; 1856 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R); 1857 1858 /// \brief Vectorize the stores that were collected in StoreRefs. 1859 bool vectorizeStoreChains(BoUpSLP &R); 1860 1861 /// \brief Scan the basic block and look for patterns that are likely to start 1862 /// a vectorization chain. 1863 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R); 1864 1865 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold, 1866 BoUpSLP &R); 1867 1868 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold, 1869 BoUpSLP &R); 1870private: 1871 StoreListMap StoreRefs; 1872}; 1873 1874/// \brief Check that the Values in the slice in VL array are still existant in 1875/// the WeakVH array. 1876/// Vectorization of part of the VL array may cause later values in the VL array 1877/// to become invalid. We track when this has happened in the WeakVH array. 1878static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL, 1879 SmallVectorImpl<WeakVH> &VH, 1880 unsigned SliceBegin, 1881 unsigned SliceSize) { 1882 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i) 1883 if (VH[i] != VL[i]) 1884 return true; 1885 1886 return false; 1887} 1888 1889bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain, 1890 int CostThreshold, BoUpSLP &R) { 1891 unsigned ChainLen = Chain.size(); 1892 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen 1893 << "\n"); 1894 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType(); 1895 unsigned Sz = DL->getTypeSizeInBits(StoreTy); 1896 unsigned VF = MinVecRegSize / Sz; 1897 1898 if (!isPowerOf2_32(Sz) || VF < 2) 1899 return false; 1900 1901 // Keep track of values that were delete by vectorizing in the loop below. 1902 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end()); 1903 1904 bool Changed = false; 1905 // Look for profitable vectorizable trees at all offsets, starting at zero. 1906 for (unsigned i = 0, e = ChainLen; i < e; ++i) { 1907 if (i + VF > e) 1908 break; 1909 1910 // Check that a previous iteration of this loop did not delete the Value. 1911 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF)) 1912 continue; 1913 1914 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i 1915 << "\n"); 1916 ArrayRef<Value *> Operands = Chain.slice(i, VF); 1917 1918 R.buildTree(Operands); 1919 1920 int Cost = R.getTreeCost(); 1921 1922 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n"); 1923 if (Cost < CostThreshold) { 1924 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n"); 1925 R.vectorizeTree(); 1926 1927 // Move to the next bundle. 1928 i += VF - 1; 1929 Changed = true; 1930 } 1931 } 1932 1933 return Changed; 1934} 1935 1936bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores, 1937 int costThreshold, BoUpSLP &R) { 1938 SetVector<Value *> Heads, Tails; 1939 SmallDenseMap<Value *, Value *> ConsecutiveChain; 1940 1941 // We may run into multiple chains that merge into a single chain. We mark the 1942 // stores that we vectorized so that we don't visit the same store twice. 1943 BoUpSLP::ValueSet VectorizedStores; 1944 bool Changed = false; 1945 1946 // Do a quadratic search on all of the given stores and find 1947 // all of the pairs of stores that follow each other. 1948 for (unsigned i = 0, e = Stores.size(); i < e; ++i) { 1949 for (unsigned j = 0; j < e; ++j) { 1950 if (i == j) 1951 continue; 1952 1953 if (R.isConsecutiveAccess(Stores[i], Stores[j])) { 1954 Tails.insert(Stores[j]); 1955 Heads.insert(Stores[i]); 1956 ConsecutiveChain[Stores[i]] = Stores[j]; 1957 } 1958 } 1959 } 1960 1961 // For stores that start but don't end a link in the chain: 1962 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end(); 1963 it != e; ++it) { 1964 if (Tails.count(*it)) 1965 continue; 1966 1967 // We found a store instr that starts a chain. Now follow the chain and try 1968 // to vectorize it. 1969 BoUpSLP::ValueList Operands; 1970 Value *I = *it; 1971 // Collect the chain into a list. 1972 while (Tails.count(I) || Heads.count(I)) { 1973 if (VectorizedStores.count(I)) 1974 break; 1975 Operands.push_back(I); 1976 // Move to the next value in the chain. 1977 I = ConsecutiveChain[I]; 1978 } 1979 1980 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R); 1981 1982 // Mark the vectorized stores so that we don't vectorize them again. 1983 if (Vectorized) 1984 VectorizedStores.insert(Operands.begin(), Operands.end()); 1985 Changed |= Vectorized; 1986 } 1987 1988 return Changed; 1989} 1990 1991 1992unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) { 1993 unsigned count = 0; 1994 StoreRefs.clear(); 1995 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { 1996 StoreInst *SI = dyn_cast<StoreInst>(it); 1997 if (!SI) 1998 continue; 1999 2000 // Don't touch volatile stores. 2001 if (!SI->isSimple()) 2002 continue; 2003 2004 // Check that the pointer points to scalars. 2005 Type *Ty = SI->getValueOperand()->getType(); 2006 if (Ty->isAggregateType() || Ty->isVectorTy()) 2007 return 0; 2008 2009 // Find the base pointer. 2010 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL); 2011 2012 // Save the store locations. 2013 StoreRefs[Ptr].push_back(SI); 2014 count++; 2015 } 2016 return count; 2017} 2018 2019bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { 2020 if (!A || !B) 2021 return false; 2022 Value *VL[] = { A, B }; 2023 return tryToVectorizeList(VL, R); 2024} 2025 2026bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R) { 2027 if (VL.size() < 2) 2028 return false; 2029 2030 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n"); 2031 2032 // Check that all of the parts are scalar instructions of the same type. 2033 Instruction *I0 = dyn_cast<Instruction>(VL[0]); 2034 if (!I0) 2035 return false; 2036 2037 unsigned Opcode0 = I0->getOpcode(); 2038 2039 Type *Ty0 = I0->getType(); 2040 unsigned Sz = DL->getTypeSizeInBits(Ty0); 2041 unsigned VF = MinVecRegSize / Sz; 2042 2043 for (int i = 0, e = VL.size(); i < e; ++i) { 2044 Type *Ty = VL[i]->getType(); 2045 if (Ty->isAggregateType() || Ty->isVectorTy()) 2046 return false; 2047 Instruction *Inst = dyn_cast<Instruction>(VL[i]); 2048 if (!Inst || Inst->getOpcode() != Opcode0) 2049 return false; 2050 } 2051 2052 bool Changed = false; 2053 2054 // Keep track of values that were delete by vectorizing in the loop below. 2055 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end()); 2056 2057 for (unsigned i = 0, e = VL.size(); i < e; ++i) { 2058 unsigned OpsWidth = 0; 2059 2060 if (i + VF > e) 2061 OpsWidth = e - i; 2062 else 2063 OpsWidth = VF; 2064 2065 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2) 2066 break; 2067 2068 // Check that a previous iteration of this loop did not delete the Value. 2069 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth)) 2070 continue; 2071 2072 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations " 2073 << "\n"); 2074 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth); 2075 2076 R.buildTree(Ops); 2077 int Cost = R.getTreeCost(); 2078 2079 if (Cost < -SLPCostThreshold) { 2080 DEBUG(dbgs() << "SLP: Vectorizing pair at cost:" << Cost << ".\n"); 2081 R.vectorizeTree(); 2082 2083 // Move to the next bundle. 2084 i += VF - 1; 2085 Changed = true; 2086 } 2087 } 2088 2089 return Changed; 2090} 2091 2092bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) { 2093 if (!V) 2094 return false; 2095 2096 // Try to vectorize V. 2097 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R)) 2098 return true; 2099 2100 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0)); 2101 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1)); 2102 // Try to skip B. 2103 if (B && B->hasOneUse()) { 2104 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); 2105 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); 2106 if (tryToVectorizePair(A, B0, R)) { 2107 B->moveBefore(V); 2108 return true; 2109 } 2110 if (tryToVectorizePair(A, B1, R)) { 2111 B->moveBefore(V); 2112 return true; 2113 } 2114 } 2115 2116 // Try to skip A. 2117 if (A && A->hasOneUse()) { 2118 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); 2119 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); 2120 if (tryToVectorizePair(A0, B, R)) { 2121 A->moveBefore(V); 2122 return true; 2123 } 2124 if (tryToVectorizePair(A1, B, R)) { 2125 A->moveBefore(V); 2126 return true; 2127 } 2128 } 2129 return 0; 2130} 2131 2132/// \brief Generate a shuffle mask to be used in a reduction tree. 2133/// 2134/// \param VecLen The length of the vector to be reduced. 2135/// \param NumEltsToRdx The number of elements that should be reduced in the 2136/// vector. 2137/// \param IsPairwise Whether the reduction is a pairwise or splitting 2138/// reduction. A pairwise reduction will generate a mask of 2139/// <0,2,...> or <1,3,..> while a splitting reduction will generate 2140/// <2,3, undef,undef> for a vector of 4 and NumElts = 2. 2141/// \param IsLeft True will generate a mask of even elements, odd otherwise. 2142static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx, 2143 bool IsPairwise, bool IsLeft, 2144 IRBuilder<> &Builder) { 2145 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask"); 2146 2147 SmallVector<Constant *, 32> ShuffleMask( 2148 VecLen, UndefValue::get(Builder.getInt32Ty())); 2149 2150 if (IsPairwise) 2151 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right). 2152 for (unsigned i = 0; i != NumEltsToRdx; ++i) 2153 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft); 2154 else 2155 // Move the upper half of the vector to the lower half. 2156 for (unsigned i = 0; i != NumEltsToRdx; ++i) 2157 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i); 2158 2159 return ConstantVector::get(ShuffleMask); 2160} 2161 2162 2163/// Model horizontal reductions. 2164/// 2165/// A horizontal reduction is a tree of reduction operations (currently add and 2166/// fadd) that has operations that can be put into a vector as its leaf. 2167/// For example, this tree: 2168/// 2169/// mul mul mul mul 2170/// \ / \ / 2171/// + + 2172/// \ / 2173/// + 2174/// This tree has "mul" as its reduced values and "+" as its reduction 2175/// operations. A reduction might be feeding into a store or a binary operation 2176/// feeding a phi. 2177/// ... 2178/// \ / 2179/// + 2180/// | 2181/// phi += 2182/// 2183/// Or: 2184/// ... 2185/// \ / 2186/// + 2187/// | 2188/// *p = 2189/// 2190class HorizontalReduction { 2191 SmallPtrSet<Value *, 16> ReductionOps; 2192 SmallVector<Value *, 32> ReducedVals; 2193 2194 BinaryOperator *ReductionRoot; 2195 PHINode *ReductionPHI; 2196 2197 /// The opcode of the reduction. 2198 unsigned ReductionOpcode; 2199 /// The opcode of the values we perform a reduction on. 2200 unsigned ReducedValueOpcode; 2201 /// The width of one full horizontal reduction operation. 2202 unsigned ReduxWidth; 2203 /// Should we model this reduction as a pairwise reduction tree or a tree that 2204 /// splits the vector in halves and adds those halves. 2205 bool IsPairwiseReduction; 2206 2207public: 2208 HorizontalReduction() 2209 : ReductionRoot(0), ReductionPHI(0), ReductionOpcode(0), 2210 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {} 2211 2212 /// \brief Try to find a reduction tree. 2213 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B, 2214 DataLayout *DL) { 2215 assert((!Phi || 2216 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) && 2217 "Thi phi needs to use the binary operator"); 2218 2219 // We could have a initial reductions that is not an add. 2220 // r *= v1 + v2 + v3 + v4 2221 // In such a case start looking for a tree rooted in the first '+'. 2222 if (Phi) { 2223 if (B->getOperand(0) == Phi) { 2224 Phi = 0; 2225 B = dyn_cast<BinaryOperator>(B->getOperand(1)); 2226 } else if (B->getOperand(1) == Phi) { 2227 Phi = 0; 2228 B = dyn_cast<BinaryOperator>(B->getOperand(0)); 2229 } 2230 } 2231 2232 if (!B) 2233 return false; 2234 2235 Type *Ty = B->getType(); 2236 if (Ty->isVectorTy()) 2237 return false; 2238 2239 ReductionOpcode = B->getOpcode(); 2240 ReducedValueOpcode = 0; 2241 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty); 2242 ReductionRoot = B; 2243 ReductionPHI = Phi; 2244 2245 if (ReduxWidth < 4) 2246 return false; 2247 2248 // We currently only support adds. 2249 if (ReductionOpcode != Instruction::Add && 2250 ReductionOpcode != Instruction::FAdd) 2251 return false; 2252 2253 // Post order traverse the reduction tree starting at B. We only handle true 2254 // trees containing only binary operators. 2255 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack; 2256 Stack.push_back(std::make_pair(B, 0)); 2257 while (!Stack.empty()) { 2258 BinaryOperator *TreeN = Stack.back().first; 2259 unsigned EdgeToVist = Stack.back().second++; 2260 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode; 2261 2262 // Only handle trees in the current basic block. 2263 if (TreeN->getParent() != B->getParent()) 2264 return false; 2265 2266 // Each tree node needs to have one user except for the ultimate 2267 // reduction. 2268 if (!TreeN->hasOneUse() && TreeN != B) 2269 return false; 2270 2271 // Postorder vist. 2272 if (EdgeToVist == 2 || IsReducedValue) { 2273 if (IsReducedValue) { 2274 // Make sure that the opcodes of the operations that we are going to 2275 // reduce match. 2276 if (!ReducedValueOpcode) 2277 ReducedValueOpcode = TreeN->getOpcode(); 2278 else if (ReducedValueOpcode != TreeN->getOpcode()) 2279 return false; 2280 ReducedVals.push_back(TreeN); 2281 } else { 2282 // We need to be able to reassociate the adds. 2283 if (!TreeN->isAssociative()) 2284 return false; 2285 ReductionOps.insert(TreeN); 2286 } 2287 // Retract. 2288 Stack.pop_back(); 2289 continue; 2290 } 2291 2292 // Visit left or right. 2293 Value *NextV = TreeN->getOperand(EdgeToVist); 2294 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV); 2295 if (Next) 2296 Stack.push_back(std::make_pair(Next, 0)); 2297 else if (NextV != Phi) 2298 return false; 2299 } 2300 return true; 2301 } 2302 2303 /// \brief Attempt to vectorize the tree found by 2304 /// matchAssociativeReduction. 2305 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { 2306 if (ReducedVals.empty()) 2307 return false; 2308 2309 unsigned NumReducedVals = ReducedVals.size(); 2310 if (NumReducedVals < ReduxWidth) 2311 return false; 2312 2313 Value *VectorizedTree = 0; 2314 IRBuilder<> Builder(ReductionRoot); 2315 FastMathFlags Unsafe; 2316 Unsafe.setUnsafeAlgebra(); 2317 Builder.SetFastMathFlags(Unsafe); 2318 unsigned i = 0; 2319 2320 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) { 2321 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth); 2322 V.buildTree(ValsToReduce, &ReductionOps); 2323 2324 // Estimate cost. 2325 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]); 2326 if (Cost >= -SLPCostThreshold) 2327 break; 2328 2329 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost 2330 << ". (HorRdx)\n"); 2331 2332 // Vectorize a tree. 2333 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc(); 2334 Value *VectorizedRoot = V.vectorizeTree(); 2335 2336 // Emit a reduction. 2337 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder); 2338 if (VectorizedTree) { 2339 Builder.SetCurrentDebugLocation(Loc); 2340 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree, 2341 ReducedSubTree, "bin.rdx"); 2342 } else 2343 VectorizedTree = ReducedSubTree; 2344 } 2345 2346 if (VectorizedTree) { 2347 // Finish the reduction. 2348 for (; i < NumReducedVals; ++i) { 2349 Builder.SetCurrentDebugLocation( 2350 cast<Instruction>(ReducedVals[i])->getDebugLoc()); 2351 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree, 2352 ReducedVals[i]); 2353 } 2354 // Update users. 2355 if (ReductionPHI) { 2356 assert(ReductionRoot != NULL && "Need a reduction operation"); 2357 ReductionRoot->setOperand(0, VectorizedTree); 2358 ReductionRoot->setOperand(1, ReductionPHI); 2359 } else 2360 ReductionRoot->replaceAllUsesWith(VectorizedTree); 2361 } 2362 return VectorizedTree != 0; 2363 } 2364 2365private: 2366 2367 /// \brief Calcuate the cost of a reduction. 2368 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) { 2369 Type *ScalarTy = FirstReducedVal->getType(); 2370 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth); 2371 2372 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true); 2373 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false); 2374 2375 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost; 2376 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost; 2377 2378 int ScalarReduxCost = 2379 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy); 2380 2381 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost 2382 << " for reduction that starts with " << *FirstReducedVal 2383 << " (It is a " 2384 << (IsPairwiseReduction ? "pairwise" : "splitting") 2385 << " reduction)\n"); 2386 2387 return VecReduxCost - ScalarReduxCost; 2388 } 2389 2390 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L, 2391 Value *R, const Twine &Name = "") { 2392 if (Opcode == Instruction::FAdd) 2393 return Builder.CreateFAdd(L, R, Name); 2394 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name); 2395 } 2396 2397 /// \brief Emit a horizontal reduction of the vectorized value. 2398 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) { 2399 assert(VectorizedValue && "Need to have a vectorized tree node"); 2400 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue); 2401 assert(isPowerOf2_32(ReduxWidth) && 2402 "We only handle power-of-two reductions for now"); 2403 2404 Value *TmpVec = ValToReduce; 2405 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) { 2406 if (IsPairwiseReduction) { 2407 Value *LeftMask = 2408 createRdxShuffleMask(ReduxWidth, i, true, true, Builder); 2409 Value *RightMask = 2410 createRdxShuffleMask(ReduxWidth, i, true, false, Builder); 2411 2412 Value *LeftShuf = Builder.CreateShuffleVector( 2413 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l"); 2414 Value *RightShuf = Builder.CreateShuffleVector( 2415 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask), 2416 "rdx.shuf.r"); 2417 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf, 2418 "bin.rdx"); 2419 } else { 2420 Value *UpperHalf = 2421 createRdxShuffleMask(ReduxWidth, i, false, false, Builder); 2422 Value *Shuf = Builder.CreateShuffleVector( 2423 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf"); 2424 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx"); 2425 } 2426 } 2427 2428 // The result is in the first element of the vector. 2429 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); 2430 } 2431}; 2432 2433/// \brief Recognize construction of vectors like 2434/// %ra = insertelement <4 x float> undef, float %s0, i32 0 2435/// %rb = insertelement <4 x float> %ra, float %s1, i32 1 2436/// %rc = insertelement <4 x float> %rb, float %s2, i32 2 2437/// %rd = insertelement <4 x float> %rc, float %s3, i32 3 2438/// 2439/// Returns true if it matches 2440/// 2441static bool findBuildVector(InsertElementInst *IE, 2442 SmallVectorImpl<Value *> &Ops) { 2443 if (!isa<UndefValue>(IE->getOperand(0))) 2444 return false; 2445 2446 while (true) { 2447 Ops.push_back(IE->getOperand(1)); 2448 2449 if (IE->use_empty()) 2450 return false; 2451 2452 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->use_back()); 2453 if (!NextUse) 2454 return true; 2455 2456 // If this isn't the final use, make sure the next insertelement is the only 2457 // use. It's OK if the final constructed vector is used multiple times 2458 if (!IE->hasOneUse()) 2459 return false; 2460 2461 IE = NextUse; 2462 } 2463 2464 return false; 2465} 2466 2467static bool PhiTypeSorterFunc(Value *V, Value *V2) { 2468 return V->getType() < V2->getType(); 2469} 2470 2471bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { 2472 bool Changed = false; 2473 SmallVector<Value *, 4> Incoming; 2474 SmallSet<Value *, 16> VisitedInstrs; 2475 2476 bool HaveVectorizedPhiNodes = true; 2477 while (HaveVectorizedPhiNodes) { 2478 HaveVectorizedPhiNodes = false; 2479 2480 // Collect the incoming values from the PHIs. 2481 Incoming.clear(); 2482 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie; 2483 ++instr) { 2484 PHINode *P = dyn_cast<PHINode>(instr); 2485 if (!P) 2486 break; 2487 2488 if (!VisitedInstrs.count(P)) 2489 Incoming.push_back(P); 2490 } 2491 2492 // Sort by type. 2493 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc); 2494 2495 // Try to vectorize elements base on their type. 2496 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(), 2497 E = Incoming.end(); 2498 IncIt != E;) { 2499 2500 // Look for the next elements with the same type. 2501 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt; 2502 while (SameTypeIt != E && 2503 (*SameTypeIt)->getType() == (*IncIt)->getType()) { 2504 VisitedInstrs.insert(*SameTypeIt); 2505 ++SameTypeIt; 2506 } 2507 2508 // Try to vectorize them. 2509 unsigned NumElts = (SameTypeIt - IncIt); 2510 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"); 2511 if (NumElts > 1 && 2512 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) { 2513 // Success start over because instructions might have been changed. 2514 HaveVectorizedPhiNodes = true; 2515 Changed = true; 2516 break; 2517 } 2518 2519 // Start over at the next instruction of a differnt type (or the end). 2520 IncIt = SameTypeIt; 2521 } 2522 } 2523 2524 VisitedInstrs.clear(); 2525 2526 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) { 2527 // We may go through BB multiple times so skip the one we have checked. 2528 if (!VisitedInstrs.insert(it)) 2529 continue; 2530 2531 if (isa<DbgInfoIntrinsic>(it)) 2532 continue; 2533 2534 // Try to vectorize reductions that use PHINodes. 2535 if (PHINode *P = dyn_cast<PHINode>(it)) { 2536 // Check that the PHI is a reduction PHI. 2537 if (P->getNumIncomingValues() != 2) 2538 return Changed; 2539 Value *Rdx = 2540 (P->getIncomingBlock(0) == BB 2541 ? (P->getIncomingValue(0)) 2542 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) : 0)); 2543 // Check if this is a Binary Operator. 2544 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx); 2545 if (!BI) 2546 continue; 2547 2548 // Try to match and vectorize a horizontal reduction. 2549 HorizontalReduction HorRdx; 2550 if (ShouldVectorizeHor && 2551 HorRdx.matchAssociativeReduction(P, BI, DL) && 2552 HorRdx.tryToReduce(R, TTI)) { 2553 Changed = true; 2554 it = BB->begin(); 2555 e = BB->end(); 2556 continue; 2557 } 2558 2559 Value *Inst = BI->getOperand(0); 2560 if (Inst == P) 2561 Inst = BI->getOperand(1); 2562 2563 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) { 2564 // We would like to start over since some instructions are deleted 2565 // and the iterator may become invalid value. 2566 Changed = true; 2567 it = BB->begin(); 2568 e = BB->end(); 2569 continue; 2570 } 2571 2572 continue; 2573 } 2574 2575 // Try to vectorize horizontal reductions feeding into a store. 2576 if (ShouldStartVectorizeHorAtStore) 2577 if (StoreInst *SI = dyn_cast<StoreInst>(it)) 2578 if (BinaryOperator *BinOp = 2579 dyn_cast<BinaryOperator>(SI->getValueOperand())) { 2580 HorizontalReduction HorRdx; 2581 if (((HorRdx.matchAssociativeReduction(0, BinOp, DL) && 2582 HorRdx.tryToReduce(R, TTI)) || 2583 tryToVectorize(BinOp, R))) { 2584 Changed = true; 2585 it = BB->begin(); 2586 e = BB->end(); 2587 continue; 2588 } 2589 } 2590 2591 // Try to vectorize trees that start at compare instructions. 2592 if (CmpInst *CI = dyn_cast<CmpInst>(it)) { 2593 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) { 2594 Changed = true; 2595 // We would like to start over since some instructions are deleted 2596 // and the iterator may become invalid value. 2597 it = BB->begin(); 2598 e = BB->end(); 2599 continue; 2600 } 2601 2602 for (int i = 0; i < 2; ++i) { 2603 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) { 2604 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) { 2605 Changed = true; 2606 // We would like to start over since some instructions are deleted 2607 // and the iterator may become invalid value. 2608 it = BB->begin(); 2609 e = BB->end(); 2610 } 2611 } 2612 } 2613 continue; 2614 } 2615 2616 // Try to vectorize trees that start at insertelement instructions. 2617 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(it)) { 2618 SmallVector<Value *, 8> Ops; 2619 if (!findBuildVector(IE, Ops)) 2620 continue; 2621 2622 if (tryToVectorizeList(Ops, R)) { 2623 Changed = true; 2624 it = BB->begin(); 2625 e = BB->end(); 2626 } 2627 2628 continue; 2629 } 2630 } 2631 2632 return Changed; 2633} 2634 2635bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) { 2636 bool Changed = false; 2637 // Attempt to sort and vectorize each of the store-groups. 2638 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end(); 2639 it != e; ++it) { 2640 if (it->second.size() < 2) 2641 continue; 2642 2643 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " 2644 << it->second.size() << ".\n"); 2645 2646 // Process the stores in chunks of 16. 2647 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) { 2648 unsigned Len = std::min<unsigned>(CE - CI, 16); 2649 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len); 2650 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R); 2651 } 2652 } 2653 return Changed; 2654} 2655 2656} // end anonymous namespace 2657 2658char SLPVectorizer::ID = 0; 2659static const char lv_name[] = "SLP Vectorizer"; 2660INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false) 2661INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 2662INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 2663INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 2664INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 2665INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false) 2666 2667namespace llvm { 2668Pass *createSLPVectorizerPass() { return new SLPVectorizer(); } 2669} 2670