1//===- InstCombineVectorOps.cpp -------------------------------------------===// 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 file implements instcombine for ExtractElement, InsertElement and 11// ShuffleVector. 12// 13//===----------------------------------------------------------------------===// 14 15#include "InstCombineInternal.h" 16#include "llvm/ADT/DenseMap.h" 17#include "llvm/IR/PatternMatch.h" 18using namespace llvm; 19using namespace PatternMatch; 20 21#define DEBUG_TYPE "instcombine" 22 23/// CheapToScalarize - Return true if the value is cheaper to scalarize than it 24/// is to leave as a vector operation. isConstant indicates whether we're 25/// extracting one known element. If false we're extracting a variable index. 26static bool CheapToScalarize(Value *V, bool isConstant) { 27 if (Constant *C = dyn_cast<Constant>(V)) { 28 if (isConstant) return true; 29 30 // If all elts are the same, we can extract it and use any of the values. 31 if (Constant *Op0 = C->getAggregateElement(0U)) { 32 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e; 33 ++i) 34 if (C->getAggregateElement(i) != Op0) 35 return false; 36 return true; 37 } 38 } 39 Instruction *I = dyn_cast<Instruction>(V); 40 if (!I) return false; 41 42 // Insert element gets simplified to the inserted element or is deleted if 43 // this is constant idx extract element and its a constant idx insertelt. 44 if (I->getOpcode() == Instruction::InsertElement && isConstant && 45 isa<ConstantInt>(I->getOperand(2))) 46 return true; 47 if (I->getOpcode() == Instruction::Load && I->hasOneUse()) 48 return true; 49 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) 50 if (BO->hasOneUse() && 51 (CheapToScalarize(BO->getOperand(0), isConstant) || 52 CheapToScalarize(BO->getOperand(1), isConstant))) 53 return true; 54 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 55 if (CI->hasOneUse() && 56 (CheapToScalarize(CI->getOperand(0), isConstant) || 57 CheapToScalarize(CI->getOperand(1), isConstant))) 58 return true; 59 60 return false; 61} 62 63/// FindScalarElement - Given a vector and an element number, see if the scalar 64/// value is already around as a register, for example if it were inserted then 65/// extracted from the vector. 66static Value *FindScalarElement(Value *V, unsigned EltNo) { 67 assert(V->getType()->isVectorTy() && "Not looking at a vector?"); 68 VectorType *VTy = cast<VectorType>(V->getType()); 69 unsigned Width = VTy->getNumElements(); 70 if (EltNo >= Width) // Out of range access. 71 return UndefValue::get(VTy->getElementType()); 72 73 if (Constant *C = dyn_cast<Constant>(V)) 74 return C->getAggregateElement(EltNo); 75 76 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) { 77 // If this is an insert to a variable element, we don't know what it is. 78 if (!isa<ConstantInt>(III->getOperand(2))) 79 return nullptr; 80 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue(); 81 82 // If this is an insert to the element we are looking for, return the 83 // inserted value. 84 if (EltNo == IIElt) 85 return III->getOperand(1); 86 87 // Otherwise, the insertelement doesn't modify the value, recurse on its 88 // vector input. 89 return FindScalarElement(III->getOperand(0), EltNo); 90 } 91 92 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) { 93 unsigned LHSWidth = SVI->getOperand(0)->getType()->getVectorNumElements(); 94 int InEl = SVI->getMaskValue(EltNo); 95 if (InEl < 0) 96 return UndefValue::get(VTy->getElementType()); 97 if (InEl < (int)LHSWidth) 98 return FindScalarElement(SVI->getOperand(0), InEl); 99 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth); 100 } 101 102 // Extract a value from a vector add operation with a constant zero. 103 Value *Val = nullptr; Constant *Con = nullptr; 104 if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) { 105 if (Con->getAggregateElement(EltNo)->isNullValue()) 106 return FindScalarElement(Val, EltNo); 107 } 108 109 // Otherwise, we don't know. 110 return nullptr; 111} 112 113// If we have a PHI node with a vector type that has only 2 uses: feed 114// itself and be an operand of extractelement at a constant location, 115// try to replace the PHI of the vector type with a PHI of a scalar type. 116Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) { 117 // Verify that the PHI node has exactly 2 uses. Otherwise return NULL. 118 if (!PN->hasNUses(2)) 119 return nullptr; 120 121 // If so, it's known at this point that one operand is PHI and the other is 122 // an extractelement node. Find the PHI user that is not the extractelement 123 // node. 124 auto iu = PN->user_begin(); 125 Instruction *PHIUser = dyn_cast<Instruction>(*iu); 126 if (PHIUser == cast<Instruction>(&EI)) 127 PHIUser = cast<Instruction>(*(++iu)); 128 129 // Verify that this PHI user has one use, which is the PHI itself, 130 // and that it is a binary operation which is cheap to scalarize. 131 // otherwise return NULL. 132 if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) || 133 !(isa<BinaryOperator>(PHIUser)) || !CheapToScalarize(PHIUser, true)) 134 return nullptr; 135 136 // Create a scalar PHI node that will replace the vector PHI node 137 // just before the current PHI node. 138 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith( 139 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN)); 140 // Scalarize each PHI operand. 141 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) { 142 Value *PHIInVal = PN->getIncomingValue(i); 143 BasicBlock *inBB = PN->getIncomingBlock(i); 144 Value *Elt = EI.getIndexOperand(); 145 // If the operand is the PHI induction variable: 146 if (PHIInVal == PHIUser) { 147 // Scalarize the binary operation. Its first operand is the 148 // scalar PHI, and the second operand is extracted from the other 149 // vector operand. 150 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser); 151 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0; 152 Value *Op = InsertNewInstWith( 153 ExtractElementInst::Create(B0->getOperand(opId), Elt, 154 B0->getOperand(opId)->getName() + ".Elt"), 155 *B0); 156 Value *newPHIUser = InsertNewInstWith( 157 BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0); 158 scalarPHI->addIncoming(newPHIUser, inBB); 159 } else { 160 // Scalarize PHI input: 161 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, ""); 162 // Insert the new instruction into the predecessor basic block. 163 Instruction *pos = dyn_cast<Instruction>(PHIInVal); 164 BasicBlock::iterator InsertPos; 165 if (pos && !isa<PHINode>(pos)) { 166 InsertPos = pos; 167 ++InsertPos; 168 } else { 169 InsertPos = inBB->getFirstInsertionPt(); 170 } 171 172 InsertNewInstWith(newEI, *InsertPos); 173 174 scalarPHI->addIncoming(newEI, inBB); 175 } 176 } 177 return ReplaceInstUsesWith(EI, scalarPHI); 178} 179 180Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { 181 // If vector val is constant with all elements the same, replace EI with 182 // that element. We handle a known element # below. 183 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0))) 184 if (CheapToScalarize(C, false)) 185 return ReplaceInstUsesWith(EI, C->getAggregateElement(0U)); 186 187 // If extracting a specified index from the vector, see if we can recursively 188 // find a previously computed scalar that was inserted into the vector. 189 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) { 190 unsigned IndexVal = IdxC->getZExtValue(); 191 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements(); 192 193 // If this is extracting an invalid index, turn this into undef, to avoid 194 // crashing the code below. 195 if (IndexVal >= VectorWidth) 196 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType())); 197 198 // This instruction only demands the single element from the input vector. 199 // If the input vector has a single use, simplify it based on this use 200 // property. 201 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) { 202 APInt UndefElts(VectorWidth, 0); 203 APInt DemandedMask(VectorWidth, 0); 204 DemandedMask.setBit(IndexVal); 205 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask, 206 UndefElts)) { 207 EI.setOperand(0, V); 208 return &EI; 209 } 210 } 211 212 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal)) 213 return ReplaceInstUsesWith(EI, Elt); 214 215 // If the this extractelement is directly using a bitcast from a vector of 216 // the same number of elements, see if we can find the source element from 217 // it. In this case, we will end up needing to bitcast the scalars. 218 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) { 219 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType())) 220 if (VT->getNumElements() == VectorWidth) 221 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal)) 222 return new BitCastInst(Elt, EI.getType()); 223 } 224 225 // If there's a vector PHI feeding a scalar use through this extractelement 226 // instruction, try to scalarize the PHI. 227 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) { 228 Instruction *scalarPHI = scalarizePHI(EI, PN); 229 if (scalarPHI) 230 return scalarPHI; 231 } 232 } 233 234 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) { 235 // Push extractelement into predecessor operation if legal and 236 // profitable to do so 237 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 238 if (I->hasOneUse() && 239 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) { 240 Value *newEI0 = 241 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1), 242 EI.getName()+".lhs"); 243 Value *newEI1 = 244 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1), 245 EI.getName()+".rhs"); 246 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1); 247 } 248 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) { 249 // Extracting the inserted element? 250 if (IE->getOperand(2) == EI.getOperand(1)) 251 return ReplaceInstUsesWith(EI, IE->getOperand(1)); 252 // If the inserted and extracted elements are constants, they must not 253 // be the same value, extract from the pre-inserted value instead. 254 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) { 255 Worklist.AddValue(EI.getOperand(0)); 256 EI.setOperand(0, IE->getOperand(0)); 257 return &EI; 258 } 259 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) { 260 // If this is extracting an element from a shufflevector, figure out where 261 // it came from and extract from the appropriate input element instead. 262 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) { 263 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue()); 264 Value *Src; 265 unsigned LHSWidth = 266 SVI->getOperand(0)->getType()->getVectorNumElements(); 267 268 if (SrcIdx < 0) 269 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType())); 270 if (SrcIdx < (int)LHSWidth) 271 Src = SVI->getOperand(0); 272 else { 273 SrcIdx -= LHSWidth; 274 Src = SVI->getOperand(1); 275 } 276 Type *Int32Ty = Type::getInt32Ty(EI.getContext()); 277 return ExtractElementInst::Create(Src, 278 ConstantInt::get(Int32Ty, 279 SrcIdx, false)); 280 } 281 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 282 // Canonicalize extractelement(cast) -> cast(extractelement) 283 // bitcasts can change the number of vector elements and they cost nothing 284 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) { 285 Value *EE = Builder->CreateExtractElement(CI->getOperand(0), 286 EI.getIndexOperand()); 287 Worklist.AddValue(EE); 288 return CastInst::Create(CI->getOpcode(), EE, EI.getType()); 289 } 290 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 291 if (SI->hasOneUse()) { 292 // TODO: For a select on vectors, it might be useful to do this if it 293 // has multiple extractelement uses. For vector select, that seems to 294 // fight the vectorizer. 295 296 // If we are extracting an element from a vector select or a select on 297 // vectors, a select on the scalars extracted from the vector arguments. 298 Value *TrueVal = SI->getTrueValue(); 299 Value *FalseVal = SI->getFalseValue(); 300 301 Value *Cond = SI->getCondition(); 302 if (Cond->getType()->isVectorTy()) { 303 Cond = Builder->CreateExtractElement(Cond, 304 EI.getIndexOperand(), 305 Cond->getName() + ".elt"); 306 } 307 308 Value *V1Elem 309 = Builder->CreateExtractElement(TrueVal, 310 EI.getIndexOperand(), 311 TrueVal->getName() + ".elt"); 312 313 Value *V2Elem 314 = Builder->CreateExtractElement(FalseVal, 315 EI.getIndexOperand(), 316 FalseVal->getName() + ".elt"); 317 return SelectInst::Create(Cond, 318 V1Elem, 319 V2Elem, 320 SI->getName() + ".elt"); 321 } 322 } 323 } 324 return nullptr; 325} 326 327/// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns 328/// elements from either LHS or RHS, return the shuffle mask and true. 329/// Otherwise, return false. 330static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS, 331 SmallVectorImpl<Constant*> &Mask) { 332 assert(LHS->getType() == RHS->getType() && 333 "Invalid CollectSingleShuffleElements"); 334 unsigned NumElts = V->getType()->getVectorNumElements(); 335 336 if (isa<UndefValue>(V)) { 337 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 338 return true; 339 } 340 341 if (V == LHS) { 342 for (unsigned i = 0; i != NumElts; ++i) 343 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 344 return true; 345 } 346 347 if (V == RHS) { 348 for (unsigned i = 0; i != NumElts; ++i) 349 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), 350 i+NumElts)); 351 return true; 352 } 353 354 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 355 // If this is an insert of an extract from some other vector, include it. 356 Value *VecOp = IEI->getOperand(0); 357 Value *ScalarOp = IEI->getOperand(1); 358 Value *IdxOp = IEI->getOperand(2); 359 360 if (!isa<ConstantInt>(IdxOp)) 361 return false; 362 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 363 364 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector. 365 // We can handle this if the vector we are inserting into is 366 // transitively ok. 367 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 368 // If so, update the mask to reflect the inserted undef. 369 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext())); 370 return true; 371 } 372 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){ 373 if (isa<ConstantInt>(EI->getOperand(1))) { 374 unsigned ExtractedIdx = 375 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 376 unsigned NumLHSElts = LHS->getType()->getVectorNumElements(); 377 378 // This must be extracting from either LHS or RHS. 379 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) { 380 // We can handle this if the vector we are inserting into is 381 // transitively ok. 382 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 383 // If so, update the mask to reflect the inserted value. 384 if (EI->getOperand(0) == LHS) { 385 Mask[InsertedIdx % NumElts] = 386 ConstantInt::get(Type::getInt32Ty(V->getContext()), 387 ExtractedIdx); 388 } else { 389 assert(EI->getOperand(0) == RHS); 390 Mask[InsertedIdx % NumElts] = 391 ConstantInt::get(Type::getInt32Ty(V->getContext()), 392 ExtractedIdx + NumLHSElts); 393 } 394 return true; 395 } 396 } 397 } 398 } 399 } 400 401 return false; 402} 403 404 405/// We are building a shuffle to create V, which is a sequence of insertelement, 406/// extractelement pairs. If PermittedRHS is set, then we must either use it or 407/// not rely on the second vector source. Return a std::pair containing the 408/// left and right vectors of the proposed shuffle (or 0), and set the Mask 409/// parameter as required. 410/// 411/// Note: we intentionally don't try to fold earlier shuffles since they have 412/// often been chosen carefully to be efficiently implementable on the target. 413typedef std::pair<Value *, Value *> ShuffleOps; 414 415static ShuffleOps CollectShuffleElements(Value *V, 416 SmallVectorImpl<Constant *> &Mask, 417 Value *PermittedRHS) { 418 assert(V->getType()->isVectorTy() && "Invalid shuffle!"); 419 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements(); 420 421 if (isa<UndefValue>(V)) { 422 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 423 return std::make_pair( 424 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr); 425 } 426 427 if (isa<ConstantAggregateZero>(V)) { 428 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0)); 429 return std::make_pair(V, nullptr); 430 } 431 432 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 433 // If this is an insert of an extract from some other vector, include it. 434 Value *VecOp = IEI->getOperand(0); 435 Value *ScalarOp = IEI->getOperand(1); 436 Value *IdxOp = IEI->getOperand(2); 437 438 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { 439 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) { 440 unsigned ExtractedIdx = 441 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 442 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 443 444 // Either the extracted from or inserted into vector must be RHSVec, 445 // otherwise we'd end up with a shuffle of three inputs. 446 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) { 447 Value *RHS = EI->getOperand(0); 448 ShuffleOps LR = CollectShuffleElements(VecOp, Mask, RHS); 449 assert(LR.second == nullptr || LR.second == RHS); 450 451 if (LR.first->getType() != RHS->getType()) { 452 // We tried our best, but we can't find anything compatible with RHS 453 // further up the chain. Return a trivial shuffle. 454 for (unsigned i = 0; i < NumElts; ++i) 455 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i); 456 return std::make_pair(V, nullptr); 457 } 458 459 unsigned NumLHSElts = RHS->getType()->getVectorNumElements(); 460 Mask[InsertedIdx % NumElts] = 461 ConstantInt::get(Type::getInt32Ty(V->getContext()), 462 NumLHSElts+ExtractedIdx); 463 return std::make_pair(LR.first, RHS); 464 } 465 466 if (VecOp == PermittedRHS) { 467 // We've gone as far as we can: anything on the other side of the 468 // extractelement will already have been converted into a shuffle. 469 unsigned NumLHSElts = 470 EI->getOperand(0)->getType()->getVectorNumElements(); 471 for (unsigned i = 0; i != NumElts; ++i) 472 Mask.push_back(ConstantInt::get( 473 Type::getInt32Ty(V->getContext()), 474 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i)); 475 return std::make_pair(EI->getOperand(0), PermittedRHS); 476 } 477 478 // If this insertelement is a chain that comes from exactly these two 479 // vectors, return the vector and the effective shuffle. 480 if (EI->getOperand(0)->getType() == PermittedRHS->getType() && 481 CollectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS, 482 Mask)) 483 return std::make_pair(EI->getOperand(0), PermittedRHS); 484 } 485 } 486 } 487 488 // Otherwise, can't do anything fancy. Return an identity vector. 489 for (unsigned i = 0; i != NumElts; ++i) 490 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 491 return std::make_pair(V, nullptr); 492} 493 494/// Try to find redundant insertvalue instructions, like the following ones: 495/// %0 = insertvalue { i8, i32 } undef, i8 %x, 0 496/// %1 = insertvalue { i8, i32 } %0, i8 %y, 0 497/// Here the second instruction inserts values at the same indices, as the 498/// first one, making the first one redundant. 499/// It should be transformed to: 500/// %0 = insertvalue { i8, i32 } undef, i8 %y, 0 501Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) { 502 bool IsRedundant = false; 503 ArrayRef<unsigned int> FirstIndices = I.getIndices(); 504 505 // If there is a chain of insertvalue instructions (each of them except the 506 // last one has only one use and it's another insertvalue insn from this 507 // chain), check if any of the 'children' uses the same indices as the first 508 // instruction. In this case, the first one is redundant. 509 Value *V = &I; 510 unsigned Depth = 0; 511 while (V->hasOneUse() && Depth < 10) { 512 User *U = V->user_back(); 513 auto UserInsInst = dyn_cast<InsertValueInst>(U); 514 if (!UserInsInst || U->getOperand(0) != V) 515 break; 516 if (UserInsInst->getIndices() == FirstIndices) { 517 IsRedundant = true; 518 break; 519 } 520 V = UserInsInst; 521 Depth++; 522 } 523 524 if (IsRedundant) 525 return ReplaceInstUsesWith(I, I.getOperand(0)); 526 return nullptr; 527} 528 529Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) { 530 Value *VecOp = IE.getOperand(0); 531 Value *ScalarOp = IE.getOperand(1); 532 Value *IdxOp = IE.getOperand(2); 533 534 // Inserting an undef or into an undefined place, remove this. 535 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp)) 536 ReplaceInstUsesWith(IE, VecOp); 537 538 // If the inserted element was extracted from some other vector, and if the 539 // indexes are constant, try to turn this into a shufflevector operation. 540 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { 541 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) { 542 unsigned NumInsertVectorElts = IE.getType()->getNumElements(); 543 unsigned NumExtractVectorElts = 544 EI->getOperand(0)->getType()->getVectorNumElements(); 545 unsigned ExtractedIdx = 546 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 547 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 548 549 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract. 550 return ReplaceInstUsesWith(IE, VecOp); 551 552 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert. 553 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType())); 554 555 // If we are extracting a value from a vector, then inserting it right 556 // back into the same place, just use the input vector. 557 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx) 558 return ReplaceInstUsesWith(IE, VecOp); 559 560 // If this insertelement isn't used by some other insertelement, turn it 561 // (and any insertelements it points to), into one big shuffle. 562 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) { 563 SmallVector<Constant*, 16> Mask; 564 ShuffleOps LR = CollectShuffleElements(&IE, Mask, nullptr); 565 566 // The proposed shuffle may be trivial, in which case we shouldn't 567 // perform the combine. 568 if (LR.first != &IE && LR.second != &IE) { 569 // We now have a shuffle of LHS, RHS, Mask. 570 if (LR.second == nullptr) 571 LR.second = UndefValue::get(LR.first->getType()); 572 return new ShuffleVectorInst(LR.first, LR.second, 573 ConstantVector::get(Mask)); 574 } 575 } 576 } 577 } 578 579 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements(); 580 APInt UndefElts(VWidth, 0); 581 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 582 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) { 583 if (V != &IE) 584 return ReplaceInstUsesWith(IE, V); 585 return &IE; 586 } 587 588 return nullptr; 589} 590 591/// Return true if we can evaluate the specified expression tree if the vector 592/// elements were shuffled in a different order. 593static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask, 594 unsigned Depth = 5) { 595 // We can always reorder the elements of a constant. 596 if (isa<Constant>(V)) 597 return true; 598 599 // We won't reorder vector arguments. No IPO here. 600 Instruction *I = dyn_cast<Instruction>(V); 601 if (!I) return false; 602 603 // Two users may expect different orders of the elements. Don't try it. 604 if (!I->hasOneUse()) 605 return false; 606 607 if (Depth == 0) return false; 608 609 switch (I->getOpcode()) { 610 case Instruction::Add: 611 case Instruction::FAdd: 612 case Instruction::Sub: 613 case Instruction::FSub: 614 case Instruction::Mul: 615 case Instruction::FMul: 616 case Instruction::UDiv: 617 case Instruction::SDiv: 618 case Instruction::FDiv: 619 case Instruction::URem: 620 case Instruction::SRem: 621 case Instruction::FRem: 622 case Instruction::Shl: 623 case Instruction::LShr: 624 case Instruction::AShr: 625 case Instruction::And: 626 case Instruction::Or: 627 case Instruction::Xor: 628 case Instruction::ICmp: 629 case Instruction::FCmp: 630 case Instruction::Trunc: 631 case Instruction::ZExt: 632 case Instruction::SExt: 633 case Instruction::FPToUI: 634 case Instruction::FPToSI: 635 case Instruction::UIToFP: 636 case Instruction::SIToFP: 637 case Instruction::FPTrunc: 638 case Instruction::FPExt: 639 case Instruction::GetElementPtr: { 640 for (int i = 0, e = I->getNumOperands(); i != e; ++i) { 641 if (!CanEvaluateShuffled(I->getOperand(i), Mask, Depth-1)) 642 return false; 643 } 644 return true; 645 } 646 case Instruction::InsertElement: { 647 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2)); 648 if (!CI) return false; 649 int ElementNumber = CI->getLimitedValue(); 650 651 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement' 652 // can't put an element into multiple indices. 653 bool SeenOnce = false; 654 for (int i = 0, e = Mask.size(); i != e; ++i) { 655 if (Mask[i] == ElementNumber) { 656 if (SeenOnce) 657 return false; 658 SeenOnce = true; 659 } 660 } 661 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1); 662 } 663 } 664 return false; 665} 666 667/// Rebuild a new instruction just like 'I' but with the new operands given. 668/// In the event of type mismatch, the type of the operands is correct. 669static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) { 670 // We don't want to use the IRBuilder here because we want the replacement 671 // instructions to appear next to 'I', not the builder's insertion point. 672 switch (I->getOpcode()) { 673 case Instruction::Add: 674 case Instruction::FAdd: 675 case Instruction::Sub: 676 case Instruction::FSub: 677 case Instruction::Mul: 678 case Instruction::FMul: 679 case Instruction::UDiv: 680 case Instruction::SDiv: 681 case Instruction::FDiv: 682 case Instruction::URem: 683 case Instruction::SRem: 684 case Instruction::FRem: 685 case Instruction::Shl: 686 case Instruction::LShr: 687 case Instruction::AShr: 688 case Instruction::And: 689 case Instruction::Or: 690 case Instruction::Xor: { 691 BinaryOperator *BO = cast<BinaryOperator>(I); 692 assert(NewOps.size() == 2 && "binary operator with #ops != 2"); 693 BinaryOperator *New = 694 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(), 695 NewOps[0], NewOps[1], "", BO); 696 if (isa<OverflowingBinaryOperator>(BO)) { 697 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap()); 698 New->setHasNoSignedWrap(BO->hasNoSignedWrap()); 699 } 700 if (isa<PossiblyExactOperator>(BO)) { 701 New->setIsExact(BO->isExact()); 702 } 703 if (isa<FPMathOperator>(BO)) 704 New->copyFastMathFlags(I); 705 return New; 706 } 707 case Instruction::ICmp: 708 assert(NewOps.size() == 2 && "icmp with #ops != 2"); 709 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(), 710 NewOps[0], NewOps[1]); 711 case Instruction::FCmp: 712 assert(NewOps.size() == 2 && "fcmp with #ops != 2"); 713 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(), 714 NewOps[0], NewOps[1]); 715 case Instruction::Trunc: 716 case Instruction::ZExt: 717 case Instruction::SExt: 718 case Instruction::FPToUI: 719 case Instruction::FPToSI: 720 case Instruction::UIToFP: 721 case Instruction::SIToFP: 722 case Instruction::FPTrunc: 723 case Instruction::FPExt: { 724 // It's possible that the mask has a different number of elements from 725 // the original cast. We recompute the destination type to match the mask. 726 Type *DestTy = 727 VectorType::get(I->getType()->getScalarType(), 728 NewOps[0]->getType()->getVectorNumElements()); 729 assert(NewOps.size() == 1 && "cast with #ops != 1"); 730 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy, 731 "", I); 732 } 733 case Instruction::GetElementPtr: { 734 Value *Ptr = NewOps[0]; 735 ArrayRef<Value*> Idx = NewOps.slice(1); 736 GetElementPtrInst *GEP = GetElementPtrInst::Create( 737 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I); 738 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds()); 739 return GEP; 740 } 741 } 742 llvm_unreachable("failed to rebuild vector instructions"); 743} 744 745Value * 746InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) { 747 // Mask.size() does not need to be equal to the number of vector elements. 748 749 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements"); 750 if (isa<UndefValue>(V)) { 751 return UndefValue::get(VectorType::get(V->getType()->getScalarType(), 752 Mask.size())); 753 } 754 if (isa<ConstantAggregateZero>(V)) { 755 return ConstantAggregateZero::get( 756 VectorType::get(V->getType()->getScalarType(), 757 Mask.size())); 758 } 759 if (Constant *C = dyn_cast<Constant>(V)) { 760 SmallVector<Constant *, 16> MaskValues; 761 for (int i = 0, e = Mask.size(); i != e; ++i) { 762 if (Mask[i] == -1) 763 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty())); 764 else 765 MaskValues.push_back(Builder->getInt32(Mask[i])); 766 } 767 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()), 768 ConstantVector::get(MaskValues)); 769 } 770 771 Instruction *I = cast<Instruction>(V); 772 switch (I->getOpcode()) { 773 case Instruction::Add: 774 case Instruction::FAdd: 775 case Instruction::Sub: 776 case Instruction::FSub: 777 case Instruction::Mul: 778 case Instruction::FMul: 779 case Instruction::UDiv: 780 case Instruction::SDiv: 781 case Instruction::FDiv: 782 case Instruction::URem: 783 case Instruction::SRem: 784 case Instruction::FRem: 785 case Instruction::Shl: 786 case Instruction::LShr: 787 case Instruction::AShr: 788 case Instruction::And: 789 case Instruction::Or: 790 case Instruction::Xor: 791 case Instruction::ICmp: 792 case Instruction::FCmp: 793 case Instruction::Trunc: 794 case Instruction::ZExt: 795 case Instruction::SExt: 796 case Instruction::FPToUI: 797 case Instruction::FPToSI: 798 case Instruction::UIToFP: 799 case Instruction::SIToFP: 800 case Instruction::FPTrunc: 801 case Instruction::FPExt: 802 case Instruction::Select: 803 case Instruction::GetElementPtr: { 804 SmallVector<Value*, 8> NewOps; 805 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements()); 806 for (int i = 0, e = I->getNumOperands(); i != e; ++i) { 807 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask); 808 NewOps.push_back(V); 809 NeedsRebuild |= (V != I->getOperand(i)); 810 } 811 if (NeedsRebuild) { 812 return BuildNew(I, NewOps); 813 } 814 return I; 815 } 816 case Instruction::InsertElement: { 817 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue(); 818 819 // The insertelement was inserting at Element. Figure out which element 820 // that becomes after shuffling. The answer is guaranteed to be unique 821 // by CanEvaluateShuffled. 822 bool Found = false; 823 int Index = 0; 824 for (int e = Mask.size(); Index != e; ++Index) { 825 if (Mask[Index] == Element) { 826 Found = true; 827 break; 828 } 829 } 830 831 // If element is not in Mask, no need to handle the operand 1 (element to 832 // be inserted). Just evaluate values in operand 0 according to Mask. 833 if (!Found) 834 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask); 835 836 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask); 837 return InsertElementInst::Create(V, I->getOperand(1), 838 Builder->getInt32(Index), "", I); 839 } 840 } 841 llvm_unreachable("failed to reorder elements of vector instruction!"); 842} 843 844static void RecognizeIdentityMask(const SmallVectorImpl<int> &Mask, 845 bool &isLHSID, bool &isRHSID) { 846 isLHSID = isRHSID = true; 847 848 for (unsigned i = 0, e = Mask.size(); i != e; ++i) { 849 if (Mask[i] < 0) continue; // Ignore undef values. 850 // Is this an identity shuffle of the LHS value? 851 isLHSID &= (Mask[i] == (int)i); 852 853 // Is this an identity shuffle of the RHS value? 854 isRHSID &= (Mask[i]-e == i); 855 } 856} 857 858// Returns true if the shuffle is extracting a contiguous range of values from 859// LHS, for example: 860// +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ 861// Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP| 862// Shuffles to: |EE|FF|GG|HH| 863// +--+--+--+--+ 864static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI, 865 SmallVector<int, 16> &Mask) { 866 unsigned LHSElems = 867 cast<VectorType>(SVI.getOperand(0)->getType())->getNumElements(); 868 unsigned MaskElems = Mask.size(); 869 unsigned BegIdx = Mask.front(); 870 unsigned EndIdx = Mask.back(); 871 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1) 872 return false; 873 for (unsigned I = 0; I != MaskElems; ++I) 874 if (static_cast<unsigned>(Mask[I]) != BegIdx + I) 875 return false; 876 return true; 877} 878 879Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { 880 Value *LHS = SVI.getOperand(0); 881 Value *RHS = SVI.getOperand(1); 882 SmallVector<int, 16> Mask = SVI.getShuffleMask(); 883 Type *Int32Ty = Type::getInt32Ty(SVI.getContext()); 884 885 bool MadeChange = false; 886 887 // Undefined shuffle mask -> undefined value. 888 if (isa<UndefValue>(SVI.getOperand(2))) 889 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType())); 890 891 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements(); 892 893 APInt UndefElts(VWidth, 0); 894 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 895 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) { 896 if (V != &SVI) 897 return ReplaceInstUsesWith(SVI, V); 898 LHS = SVI.getOperand(0); 899 RHS = SVI.getOperand(1); 900 MadeChange = true; 901 } 902 903 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements(); 904 905 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask') 906 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask'). 907 if (LHS == RHS || isa<UndefValue>(LHS)) { 908 if (isa<UndefValue>(LHS) && LHS == RHS) { 909 // shuffle(undef,undef,mask) -> undef. 910 Value *Result = (VWidth == LHSWidth) 911 ? LHS : UndefValue::get(SVI.getType()); 912 return ReplaceInstUsesWith(SVI, Result); 913 } 914 915 // Remap any references to RHS to use LHS. 916 SmallVector<Constant*, 16> Elts; 917 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) { 918 if (Mask[i] < 0) { 919 Elts.push_back(UndefValue::get(Int32Ty)); 920 continue; 921 } 922 923 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) || 924 (Mask[i] < (int)e && isa<UndefValue>(LHS))) { 925 Mask[i] = -1; // Turn into undef. 926 Elts.push_back(UndefValue::get(Int32Ty)); 927 } else { 928 Mask[i] = Mask[i] % e; // Force to LHS. 929 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i])); 930 } 931 } 932 SVI.setOperand(0, SVI.getOperand(1)); 933 SVI.setOperand(1, UndefValue::get(RHS->getType())); 934 SVI.setOperand(2, ConstantVector::get(Elts)); 935 LHS = SVI.getOperand(0); 936 RHS = SVI.getOperand(1); 937 MadeChange = true; 938 } 939 940 if (VWidth == LHSWidth) { 941 // Analyze the shuffle, are the LHS or RHS and identity shuffles? 942 bool isLHSID, isRHSID; 943 RecognizeIdentityMask(Mask, isLHSID, isRHSID); 944 945 // Eliminate identity shuffles. 946 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS); 947 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS); 948 } 949 950 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) { 951 Value *V = EvaluateInDifferentElementOrder(LHS, Mask); 952 return ReplaceInstUsesWith(SVI, V); 953 } 954 955 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to 956 // a non-vector type. We can instead bitcast the original vector followed by 957 // an extract of the desired element: 958 // 959 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef, 960 // <4 x i32> <i32 0, i32 1, i32 2, i32 3> 961 // %1 = bitcast <4 x i8> %sroa to i32 962 // Becomes: 963 // %bc = bitcast <16 x i8> %in to <4 x i32> 964 // %ext = extractelement <4 x i32> %bc, i32 0 965 // 966 // If the shuffle is extracting a contiguous range of values from the input 967 // vector then each use which is a bitcast of the extracted size can be 968 // replaced. This will work if the vector types are compatible, and the begin 969 // index is aligned to a value in the casted vector type. If the begin index 970 // isn't aligned then we can shuffle the original vector (keeping the same 971 // vector type) before extracting. 972 // 973 // This code will bail out if the target type is fundamentally incompatible 974 // with vectors of the source type. 975 // 976 // Example of <16 x i8>, target type i32: 977 // Index range [4,8): v-----------v Will work. 978 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ 979 // <16 x i8>: | | | | | | | | | | | | | | | | | 980 // <4 x i32>: | | | | | 981 // +-----------+-----------+-----------+-----------+ 982 // Index range [6,10): ^-----------^ Needs an extra shuffle. 983 // Target type i40: ^--------------^ Won't work, bail. 984 if (isShuffleExtractingFromLHS(SVI, Mask)) { 985 Value *V = LHS; 986 unsigned MaskElems = Mask.size(); 987 unsigned BegIdx = Mask.front(); 988 VectorType *SrcTy = cast<VectorType>(V->getType()); 989 unsigned VecBitWidth = SrcTy->getBitWidth(); 990 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType()); 991 assert(SrcElemBitWidth && "vector elements must have a bitwidth"); 992 unsigned SrcNumElems = SrcTy->getNumElements(); 993 SmallVector<BitCastInst *, 8> BCs; 994 DenseMap<Type *, Value *> NewBCs; 995 for (User *U : SVI.users()) 996 if (BitCastInst *BC = dyn_cast<BitCastInst>(U)) 997 if (!BC->use_empty()) 998 // Only visit bitcasts that weren't previously handled. 999 BCs.push_back(BC); 1000 for (BitCastInst *BC : BCs) { 1001 Type *TgtTy = BC->getDestTy(); 1002 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy); 1003 if (!TgtElemBitWidth) 1004 continue; 1005 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth; 1006 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth; 1007 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth); 1008 if (!VecBitWidthsEqual) 1009 continue; 1010 if (!VectorType::isValidElementType(TgtTy)) 1011 continue; 1012 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems); 1013 if (!BegIsAligned) { 1014 // Shuffle the input so [0,NumElements) contains the output, and 1015 // [NumElems,SrcNumElems) is undef. 1016 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems, 1017 UndefValue::get(Int32Ty)); 1018 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I) 1019 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx); 1020 V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()), 1021 ConstantVector::get(ShuffleMask), 1022 SVI.getName() + ".extract"); 1023 BegIdx = 0; 1024 } 1025 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth; 1026 assert(SrcElemsPerTgtElem); 1027 BegIdx /= SrcElemsPerTgtElem; 1028 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end(); 1029 auto *NewBC = 1030 BCAlreadyExists 1031 ? NewBCs[CastSrcTy] 1032 : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc"); 1033 if (!BCAlreadyExists) 1034 NewBCs[CastSrcTy] = NewBC; 1035 auto *Ext = Builder->CreateExtractElement( 1036 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract"); 1037 // The shufflevector isn't being replaced: the bitcast that used it 1038 // is. InstCombine will visit the newly-created instructions. 1039 ReplaceInstUsesWith(*BC, Ext); 1040 MadeChange = true; 1041 } 1042 } 1043 1044 // If the LHS is a shufflevector itself, see if we can combine it with this 1045 // one without producing an unusual shuffle. 1046 // Cases that might be simplified: 1047 // 1. 1048 // x1=shuffle(v1,v2,mask1) 1049 // x=shuffle(x1,undef,mask) 1050 // ==> 1051 // x=shuffle(v1,undef,newMask) 1052 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1 1053 // 2. 1054 // x1=shuffle(v1,undef,mask1) 1055 // x=shuffle(x1,x2,mask) 1056 // where v1.size() == mask1.size() 1057 // ==> 1058 // x=shuffle(v1,x2,newMask) 1059 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i] 1060 // 3. 1061 // x2=shuffle(v2,undef,mask2) 1062 // x=shuffle(x1,x2,mask) 1063 // where v2.size() == mask2.size() 1064 // ==> 1065 // x=shuffle(x1,v2,newMask) 1066 // newMask[i] = (mask[i] < x1.size()) 1067 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size() 1068 // 4. 1069 // x1=shuffle(v1,undef,mask1) 1070 // x2=shuffle(v2,undef,mask2) 1071 // x=shuffle(x1,x2,mask) 1072 // where v1.size() == v2.size() 1073 // ==> 1074 // x=shuffle(v1,v2,newMask) 1075 // newMask[i] = (mask[i] < x1.size()) 1076 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size() 1077 // 1078 // Here we are really conservative: 1079 // we are absolutely afraid of producing a shuffle mask not in the input 1080 // program, because the code gen may not be smart enough to turn a merged 1081 // shuffle into two specific shuffles: it may produce worse code. As such, 1082 // we only merge two shuffles if the result is either a splat or one of the 1083 // input shuffle masks. In this case, merging the shuffles just removes 1084 // one instruction, which we know is safe. This is good for things like 1085 // turning: (splat(splat)) -> splat, or 1086 // merge(V[0..n], V[n+1..2n]) -> V[0..2n] 1087 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS); 1088 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS); 1089 if (LHSShuffle) 1090 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS)) 1091 LHSShuffle = nullptr; 1092 if (RHSShuffle) 1093 if (!isa<UndefValue>(RHSShuffle->getOperand(1))) 1094 RHSShuffle = nullptr; 1095 if (!LHSShuffle && !RHSShuffle) 1096 return MadeChange ? &SVI : nullptr; 1097 1098 Value* LHSOp0 = nullptr; 1099 Value* LHSOp1 = nullptr; 1100 Value* RHSOp0 = nullptr; 1101 unsigned LHSOp0Width = 0; 1102 unsigned RHSOp0Width = 0; 1103 if (LHSShuffle) { 1104 LHSOp0 = LHSShuffle->getOperand(0); 1105 LHSOp1 = LHSShuffle->getOperand(1); 1106 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements(); 1107 } 1108 if (RHSShuffle) { 1109 RHSOp0 = RHSShuffle->getOperand(0); 1110 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements(); 1111 } 1112 Value* newLHS = LHS; 1113 Value* newRHS = RHS; 1114 if (LHSShuffle) { 1115 // case 1 1116 if (isa<UndefValue>(RHS)) { 1117 newLHS = LHSOp0; 1118 newRHS = LHSOp1; 1119 } 1120 // case 2 or 4 1121 else if (LHSOp0Width == LHSWidth) { 1122 newLHS = LHSOp0; 1123 } 1124 } 1125 // case 3 or 4 1126 if (RHSShuffle && RHSOp0Width == LHSWidth) { 1127 newRHS = RHSOp0; 1128 } 1129 // case 4 1130 if (LHSOp0 == RHSOp0) { 1131 newLHS = LHSOp0; 1132 newRHS = nullptr; 1133 } 1134 1135 if (newLHS == LHS && newRHS == RHS) 1136 return MadeChange ? &SVI : nullptr; 1137 1138 SmallVector<int, 16> LHSMask; 1139 SmallVector<int, 16> RHSMask; 1140 if (newLHS != LHS) 1141 LHSMask = LHSShuffle->getShuffleMask(); 1142 if (RHSShuffle && newRHS != RHS) 1143 RHSMask = RHSShuffle->getShuffleMask(); 1144 1145 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth; 1146 SmallVector<int, 16> newMask; 1147 bool isSplat = true; 1148 int SplatElt = -1; 1149 // Create a new mask for the new ShuffleVectorInst so that the new 1150 // ShuffleVectorInst is equivalent to the original one. 1151 for (unsigned i = 0; i < VWidth; ++i) { 1152 int eltMask; 1153 if (Mask[i] < 0) { 1154 // This element is an undef value. 1155 eltMask = -1; 1156 } else if (Mask[i] < (int)LHSWidth) { 1157 // This element is from left hand side vector operand. 1158 // 1159 // If LHS is going to be replaced (case 1, 2, or 4), calculate the 1160 // new mask value for the element. 1161 if (newLHS != LHS) { 1162 eltMask = LHSMask[Mask[i]]; 1163 // If the value selected is an undef value, explicitly specify it 1164 // with a -1 mask value. 1165 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1)) 1166 eltMask = -1; 1167 } else 1168 eltMask = Mask[i]; 1169 } else { 1170 // This element is from right hand side vector operand 1171 // 1172 // If the value selected is an undef value, explicitly specify it 1173 // with a -1 mask value. (case 1) 1174 if (isa<UndefValue>(RHS)) 1175 eltMask = -1; 1176 // If RHS is going to be replaced (case 3 or 4), calculate the 1177 // new mask value for the element. 1178 else if (newRHS != RHS) { 1179 eltMask = RHSMask[Mask[i]-LHSWidth]; 1180 // If the value selected is an undef value, explicitly specify it 1181 // with a -1 mask value. 1182 if (eltMask >= (int)RHSOp0Width) { 1183 assert(isa<UndefValue>(RHSShuffle->getOperand(1)) 1184 && "should have been check above"); 1185 eltMask = -1; 1186 } 1187 } else 1188 eltMask = Mask[i]-LHSWidth; 1189 1190 // If LHS's width is changed, shift the mask value accordingly. 1191 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any 1192 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask. 1193 // If newRHS == newLHS, we want to remap any references from newRHS to 1194 // newLHS so that we can properly identify splats that may occur due to 1195 // obfuscation across the two vectors. 1196 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS) 1197 eltMask += newLHSWidth; 1198 } 1199 1200 // Check if this could still be a splat. 1201 if (eltMask >= 0) { 1202 if (SplatElt >= 0 && SplatElt != eltMask) 1203 isSplat = false; 1204 SplatElt = eltMask; 1205 } 1206 1207 newMask.push_back(eltMask); 1208 } 1209 1210 // If the result mask is equal to one of the original shuffle masks, 1211 // or is a splat, do the replacement. 1212 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) { 1213 SmallVector<Constant*, 16> Elts; 1214 for (unsigned i = 0, e = newMask.size(); i != e; ++i) { 1215 if (newMask[i] < 0) { 1216 Elts.push_back(UndefValue::get(Int32Ty)); 1217 } else { 1218 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i])); 1219 } 1220 } 1221 if (!newRHS) 1222 newRHS = UndefValue::get(newLHS->getType()); 1223 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts)); 1224 } 1225 1226 // If the result mask is an identity, replace uses of this instruction with 1227 // corresponding argument. 1228 bool isLHSID, isRHSID; 1229 RecognizeIdentityMask(newMask, isLHSID, isRHSID); 1230 if (isLHSID && VWidth == LHSOp0Width) return ReplaceInstUsesWith(SVI, newLHS); 1231 if (isRHSID && VWidth == RHSOp0Width) return ReplaceInstUsesWith(SVI, newRHS); 1232 1233 return MadeChange ? &SVI : nullptr; 1234} 1235