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/Analysis/InstructionSimplify.h" 18 #include "llvm/Analysis/VectorUtils.h" 19 #include "llvm/IR/PatternMatch.h" 20 using namespace llvm; 21 using namespace PatternMatch; 22 23 #define DEBUG_TYPE "instcombine" 24 25 /// Return true if the value is cheaper to scalarize than it is to leave as a 26 /// vector operation. isConstant indicates whether we're extracting one known 27 /// element. If false we're extracting a variable index. 28 static bool cheapToScalarize(Value *V, bool isConstant) { 29 if (Constant *C = dyn_cast<Constant>(V)) { 30 if (isConstant) return true; 31 32 // If all elts are the same, we can extract it and use any of the values. 33 if (Constant *Op0 = C->getAggregateElement(0U)) { 34 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e; 35 ++i) 36 if (C->getAggregateElement(i) != Op0) 37 return false; 38 return true; 39 } 40 } 41 Instruction *I = dyn_cast<Instruction>(V); 42 if (!I) return false; 43 44 // Insert element gets simplified to the inserted element or is deleted if 45 // this is constant idx extract element and its a constant idx insertelt. 46 if (I->getOpcode() == Instruction::InsertElement && isConstant && 47 isa<ConstantInt>(I->getOperand(2))) 48 return true; 49 if (I->getOpcode() == Instruction::Load && I->hasOneUse()) 50 return true; 51 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) 52 if (BO->hasOneUse() && 53 (cheapToScalarize(BO->getOperand(0), isConstant) || 54 cheapToScalarize(BO->getOperand(1), isConstant))) 55 return true; 56 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 57 if (CI->hasOneUse() && 58 (cheapToScalarize(CI->getOperand(0), isConstant) || 59 cheapToScalarize(CI->getOperand(1), isConstant))) 60 return true; 61 62 return false; 63 } 64 65 // If we have a PHI node with a vector type that is only used to feed 66 // itself and be an operand of extractelement at a constant location, 67 // try to replace the PHI of the vector type with a PHI of a scalar type. 68 Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) { 69 SmallVector<Instruction *, 2> Extracts; 70 // The users we want the PHI to have are: 71 // 1) The EI ExtractElement (we already know this) 72 // 2) Possibly more ExtractElements with the same index. 73 // 3) Another operand, which will feed back into the PHI. 74 Instruction *PHIUser = nullptr; 75 for (auto U : PN->users()) { 76 if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(U)) { 77 if (EI.getIndexOperand() == EU->getIndexOperand()) 78 Extracts.push_back(EU); 79 else 80 return nullptr; 81 } else if (!PHIUser) { 82 PHIUser = cast<Instruction>(U); 83 } else { 84 return nullptr; 85 } 86 } 87 88 if (!PHIUser) 89 return nullptr; 90 91 // Verify that this PHI user has one use, which is the PHI itself, 92 // and that it is a binary operation which is cheap to scalarize. 93 // otherwise return NULL. 94 if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) || 95 !(isa<BinaryOperator>(PHIUser)) || !cheapToScalarize(PHIUser, true)) 96 return nullptr; 97 98 // Create a scalar PHI node that will replace the vector PHI node 99 // just before the current PHI node. 100 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith( 101 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN)); 102 // Scalarize each PHI operand. 103 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) { 104 Value *PHIInVal = PN->getIncomingValue(i); 105 BasicBlock *inBB = PN->getIncomingBlock(i); 106 Value *Elt = EI.getIndexOperand(); 107 // If the operand is the PHI induction variable: 108 if (PHIInVal == PHIUser) { 109 // Scalarize the binary operation. Its first operand is the 110 // scalar PHI, and the second operand is extracted from the other 111 // vector operand. 112 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser); 113 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0; 114 Value *Op = InsertNewInstWith( 115 ExtractElementInst::Create(B0->getOperand(opId), Elt, 116 B0->getOperand(opId)->getName() + ".Elt"), 117 *B0); 118 Value *newPHIUser = InsertNewInstWith( 119 BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(), 120 scalarPHI, Op, B0), *B0); 121 scalarPHI->addIncoming(newPHIUser, inBB); 122 } else { 123 // Scalarize PHI input: 124 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, ""); 125 // Insert the new instruction into the predecessor basic block. 126 Instruction *pos = dyn_cast<Instruction>(PHIInVal); 127 BasicBlock::iterator InsertPos; 128 if (pos && !isa<PHINode>(pos)) { 129 InsertPos = ++pos->getIterator(); 130 } else { 131 InsertPos = inBB->getFirstInsertionPt(); 132 } 133 134 InsertNewInstWith(newEI, *InsertPos); 135 136 scalarPHI->addIncoming(newEI, inBB); 137 } 138 } 139 140 for (auto E : Extracts) 141 replaceInstUsesWith(*E, scalarPHI); 142 143 return &EI; 144 } 145 146 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { 147 if (Value *V = SimplifyExtractElementInst( 148 EI.getVectorOperand(), EI.getIndexOperand(), DL, TLI, DT, AC)) 149 return replaceInstUsesWith(EI, V); 150 151 // If vector val is constant with all elements the same, replace EI with 152 // that element. We handle a known element # below. 153 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0))) 154 if (cheapToScalarize(C, false)) 155 return replaceInstUsesWith(EI, C->getAggregateElement(0U)); 156 157 // If extracting a specified index from the vector, see if we can recursively 158 // find a previously computed scalar that was inserted into the vector. 159 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) { 160 unsigned IndexVal = IdxC->getZExtValue(); 161 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements(); 162 163 // InstSimplify handles cases where the index is invalid. 164 assert(IndexVal < VectorWidth); 165 166 // This instruction only demands the single element from the input vector. 167 // If the input vector has a single use, simplify it based on this use 168 // property. 169 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) { 170 APInt UndefElts(VectorWidth, 0); 171 APInt DemandedMask(VectorWidth, 0); 172 DemandedMask.setBit(IndexVal); 173 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask, 174 UndefElts)) { 175 EI.setOperand(0, V); 176 return &EI; 177 } 178 } 179 180 // If this extractelement is directly using a bitcast from a vector of 181 // the same number of elements, see if we can find the source element from 182 // it. In this case, we will end up needing to bitcast the scalars. 183 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) { 184 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType())) 185 if (VT->getNumElements() == VectorWidth) 186 if (Value *Elt = findScalarElement(BCI->getOperand(0), IndexVal)) 187 return new BitCastInst(Elt, EI.getType()); 188 } 189 190 // If there's a vector PHI feeding a scalar use through this extractelement 191 // instruction, try to scalarize the PHI. 192 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) { 193 Instruction *scalarPHI = scalarizePHI(EI, PN); 194 if (scalarPHI) 195 return scalarPHI; 196 } 197 } 198 199 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) { 200 // Push extractelement into predecessor operation if legal and 201 // profitable to do so. 202 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 203 if (I->hasOneUse() && 204 cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) { 205 Value *newEI0 = 206 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1), 207 EI.getName()+".lhs"); 208 Value *newEI1 = 209 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1), 210 EI.getName()+".rhs"); 211 return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(), 212 newEI0, newEI1, BO); 213 } 214 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) { 215 // Extracting the inserted element? 216 if (IE->getOperand(2) == EI.getOperand(1)) 217 return replaceInstUsesWith(EI, IE->getOperand(1)); 218 // If the inserted and extracted elements are constants, they must not 219 // be the same value, extract from the pre-inserted value instead. 220 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) { 221 Worklist.AddValue(EI.getOperand(0)); 222 EI.setOperand(0, IE->getOperand(0)); 223 return &EI; 224 } 225 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) { 226 // If this is extracting an element from a shufflevector, figure out where 227 // it came from and extract from the appropriate input element instead. 228 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) { 229 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue()); 230 Value *Src; 231 unsigned LHSWidth = 232 SVI->getOperand(0)->getType()->getVectorNumElements(); 233 234 if (SrcIdx < 0) 235 return replaceInstUsesWith(EI, UndefValue::get(EI.getType())); 236 if (SrcIdx < (int)LHSWidth) 237 Src = SVI->getOperand(0); 238 else { 239 SrcIdx -= LHSWidth; 240 Src = SVI->getOperand(1); 241 } 242 Type *Int32Ty = Type::getInt32Ty(EI.getContext()); 243 return ExtractElementInst::Create(Src, 244 ConstantInt::get(Int32Ty, 245 SrcIdx, false)); 246 } 247 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 248 // Canonicalize extractelement(cast) -> cast(extractelement). 249 // Bitcasts can change the number of vector elements, and they cost 250 // nothing. 251 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) { 252 Value *EE = Builder->CreateExtractElement(CI->getOperand(0), 253 EI.getIndexOperand()); 254 Worklist.AddValue(EE); 255 return CastInst::Create(CI->getOpcode(), EE, EI.getType()); 256 } 257 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 258 if (SI->hasOneUse()) { 259 // TODO: For a select on vectors, it might be useful to do this if it 260 // has multiple extractelement uses. For vector select, that seems to 261 // fight the vectorizer. 262 263 // If we are extracting an element from a vector select or a select on 264 // vectors, create a select on the scalars extracted from the vector 265 // arguments. 266 Value *TrueVal = SI->getTrueValue(); 267 Value *FalseVal = SI->getFalseValue(); 268 269 Value *Cond = SI->getCondition(); 270 if (Cond->getType()->isVectorTy()) { 271 Cond = Builder->CreateExtractElement(Cond, 272 EI.getIndexOperand(), 273 Cond->getName() + ".elt"); 274 } 275 276 Value *V1Elem 277 = Builder->CreateExtractElement(TrueVal, 278 EI.getIndexOperand(), 279 TrueVal->getName() + ".elt"); 280 281 Value *V2Elem 282 = Builder->CreateExtractElement(FalseVal, 283 EI.getIndexOperand(), 284 FalseVal->getName() + ".elt"); 285 return SelectInst::Create(Cond, 286 V1Elem, 287 V2Elem, 288 SI->getName() + ".elt"); 289 } 290 } 291 } 292 return nullptr; 293 } 294 295 /// If V is a shuffle of values that ONLY returns elements from either LHS or 296 /// RHS, return the shuffle mask and true. Otherwise, return false. 297 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS, 298 SmallVectorImpl<Constant*> &Mask) { 299 assert(LHS->getType() == RHS->getType() && 300 "Invalid CollectSingleShuffleElements"); 301 unsigned NumElts = V->getType()->getVectorNumElements(); 302 303 if (isa<UndefValue>(V)) { 304 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 305 return true; 306 } 307 308 if (V == LHS) { 309 for (unsigned i = 0; i != NumElts; ++i) 310 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 311 return true; 312 } 313 314 if (V == RHS) { 315 for (unsigned i = 0; i != NumElts; ++i) 316 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), 317 i+NumElts)); 318 return true; 319 } 320 321 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 322 // If this is an insert of an extract from some other vector, include it. 323 Value *VecOp = IEI->getOperand(0); 324 Value *ScalarOp = IEI->getOperand(1); 325 Value *IdxOp = IEI->getOperand(2); 326 327 if (!isa<ConstantInt>(IdxOp)) 328 return false; 329 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 330 331 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector. 332 // We can handle this if the vector we are inserting into is 333 // transitively ok. 334 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 335 // If so, update the mask to reflect the inserted undef. 336 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext())); 337 return true; 338 } 339 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){ 340 if (isa<ConstantInt>(EI->getOperand(1))) { 341 unsigned ExtractedIdx = 342 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 343 unsigned NumLHSElts = LHS->getType()->getVectorNumElements(); 344 345 // This must be extracting from either LHS or RHS. 346 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) { 347 // We can handle this if the vector we are inserting into is 348 // transitively ok. 349 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 350 // If so, update the mask to reflect the inserted value. 351 if (EI->getOperand(0) == LHS) { 352 Mask[InsertedIdx % NumElts] = 353 ConstantInt::get(Type::getInt32Ty(V->getContext()), 354 ExtractedIdx); 355 } else { 356 assert(EI->getOperand(0) == RHS); 357 Mask[InsertedIdx % NumElts] = 358 ConstantInt::get(Type::getInt32Ty(V->getContext()), 359 ExtractedIdx + NumLHSElts); 360 } 361 return true; 362 } 363 } 364 } 365 } 366 } 367 368 return false; 369 } 370 371 /// If we have insertion into a vector that is wider than the vector that we 372 /// are extracting from, try to widen the source vector to allow a single 373 /// shufflevector to replace one or more insert/extract pairs. 374 static void replaceExtractElements(InsertElementInst *InsElt, 375 ExtractElementInst *ExtElt, 376 InstCombiner &IC) { 377 VectorType *InsVecType = InsElt->getType(); 378 VectorType *ExtVecType = ExtElt->getVectorOperandType(); 379 unsigned NumInsElts = InsVecType->getVectorNumElements(); 380 unsigned NumExtElts = ExtVecType->getVectorNumElements(); 381 382 // The inserted-to vector must be wider than the extracted-from vector. 383 if (InsVecType->getElementType() != ExtVecType->getElementType() || 384 NumExtElts >= NumInsElts) 385 return; 386 387 // Create a shuffle mask to widen the extended-from vector using undefined 388 // values. The mask selects all of the values of the original vector followed 389 // by as many undefined values as needed to create a vector of the same length 390 // as the inserted-to vector. 391 SmallVector<Constant *, 16> ExtendMask; 392 IntegerType *IntType = Type::getInt32Ty(InsElt->getContext()); 393 for (unsigned i = 0; i < NumExtElts; ++i) 394 ExtendMask.push_back(ConstantInt::get(IntType, i)); 395 for (unsigned i = NumExtElts; i < NumInsElts; ++i) 396 ExtendMask.push_back(UndefValue::get(IntType)); 397 398 Value *ExtVecOp = ExtElt->getVectorOperand(); 399 auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp); 400 BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst)) 401 ? ExtVecOpInst->getParent() 402 : ExtElt->getParent(); 403 404 // TODO: This restriction matches the basic block check below when creating 405 // new extractelement instructions. If that limitation is removed, this one 406 // could also be removed. But for now, we just bail out to ensure that we 407 // will replace the extractelement instruction that is feeding our 408 // insertelement instruction. This allows the insertelement to then be 409 // replaced by a shufflevector. If the insertelement is not replaced, we can 410 // induce infinite looping because there's an optimization for extractelement 411 // that will delete our widening shuffle. This would trigger another attempt 412 // here to create that shuffle, and we spin forever. 413 if (InsertionBlock != InsElt->getParent()) 414 return; 415 416 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType), 417 ConstantVector::get(ExtendMask)); 418 419 // Insert the new shuffle after the vector operand of the extract is defined 420 // (as long as it's not a PHI) or at the start of the basic block of the 421 // extract, so any subsequent extracts in the same basic block can use it. 422 // TODO: Insert before the earliest ExtractElementInst that is replaced. 423 if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst)) 424 WideVec->insertAfter(ExtVecOpInst); 425 else 426 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt()); 427 428 // Replace extracts from the original narrow vector with extracts from the new 429 // wide vector. 430 for (User *U : ExtVecOp->users()) { 431 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U); 432 if (!OldExt || OldExt->getParent() != WideVec->getParent()) 433 continue; 434 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1)); 435 NewExt->insertAfter(WideVec); 436 IC.replaceInstUsesWith(*OldExt, NewExt); 437 } 438 } 439 440 /// We are building a shuffle to create V, which is a sequence of insertelement, 441 /// extractelement pairs. If PermittedRHS is set, then we must either use it or 442 /// not rely on the second vector source. Return a std::pair containing the 443 /// left and right vectors of the proposed shuffle (or 0), and set the Mask 444 /// parameter as required. 445 /// 446 /// Note: we intentionally don't try to fold earlier shuffles since they have 447 /// often been chosen carefully to be efficiently implementable on the target. 448 typedef std::pair<Value *, Value *> ShuffleOps; 449 450 static ShuffleOps collectShuffleElements(Value *V, 451 SmallVectorImpl<Constant *> &Mask, 452 Value *PermittedRHS, 453 InstCombiner &IC) { 454 assert(V->getType()->isVectorTy() && "Invalid shuffle!"); 455 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements(); 456 457 if (isa<UndefValue>(V)) { 458 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 459 return std::make_pair( 460 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr); 461 } 462 463 if (isa<ConstantAggregateZero>(V)) { 464 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0)); 465 return std::make_pair(V, nullptr); 466 } 467 468 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 469 // If this is an insert of an extract from some other vector, include it. 470 Value *VecOp = IEI->getOperand(0); 471 Value *ScalarOp = IEI->getOperand(1); 472 Value *IdxOp = IEI->getOperand(2); 473 474 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { 475 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) { 476 unsigned ExtractedIdx = 477 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 478 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 479 480 // Either the extracted from or inserted into vector must be RHSVec, 481 // otherwise we'd end up with a shuffle of three inputs. 482 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) { 483 Value *RHS = EI->getOperand(0); 484 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC); 485 assert(LR.second == nullptr || LR.second == RHS); 486 487 if (LR.first->getType() != RHS->getType()) { 488 // Although we are giving up for now, see if we can create extracts 489 // that match the inserts for another round of combining. 490 replaceExtractElements(IEI, EI, IC); 491 492 // We tried our best, but we can't find anything compatible with RHS 493 // further up the chain. Return a trivial shuffle. 494 for (unsigned i = 0; i < NumElts; ++i) 495 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i); 496 return std::make_pair(V, nullptr); 497 } 498 499 unsigned NumLHSElts = RHS->getType()->getVectorNumElements(); 500 Mask[InsertedIdx % NumElts] = 501 ConstantInt::get(Type::getInt32Ty(V->getContext()), 502 NumLHSElts+ExtractedIdx); 503 return std::make_pair(LR.first, RHS); 504 } 505 506 if (VecOp == PermittedRHS) { 507 // We've gone as far as we can: anything on the other side of the 508 // extractelement will already have been converted into a shuffle. 509 unsigned NumLHSElts = 510 EI->getOperand(0)->getType()->getVectorNumElements(); 511 for (unsigned i = 0; i != NumElts; ++i) 512 Mask.push_back(ConstantInt::get( 513 Type::getInt32Ty(V->getContext()), 514 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i)); 515 return std::make_pair(EI->getOperand(0), PermittedRHS); 516 } 517 518 // If this insertelement is a chain that comes from exactly these two 519 // vectors, return the vector and the effective shuffle. 520 if (EI->getOperand(0)->getType() == PermittedRHS->getType() && 521 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS, 522 Mask)) 523 return std::make_pair(EI->getOperand(0), PermittedRHS); 524 } 525 } 526 } 527 528 // Otherwise, we can't do anything fancy. Return an identity vector. 529 for (unsigned i = 0; i != NumElts; ++i) 530 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 531 return std::make_pair(V, nullptr); 532 } 533 534 /// Try to find redundant insertvalue instructions, like the following ones: 535 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0 536 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0 537 /// Here the second instruction inserts values at the same indices, as the 538 /// first one, making the first one redundant. 539 /// It should be transformed to: 540 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0 541 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) { 542 bool IsRedundant = false; 543 ArrayRef<unsigned int> FirstIndices = I.getIndices(); 544 545 // If there is a chain of insertvalue instructions (each of them except the 546 // last one has only one use and it's another insertvalue insn from this 547 // chain), check if any of the 'children' uses the same indices as the first 548 // instruction. In this case, the first one is redundant. 549 Value *V = &I; 550 unsigned Depth = 0; 551 while (V->hasOneUse() && Depth < 10) { 552 User *U = V->user_back(); 553 auto UserInsInst = dyn_cast<InsertValueInst>(U); 554 if (!UserInsInst || U->getOperand(0) != V) 555 break; 556 if (UserInsInst->getIndices() == FirstIndices) { 557 IsRedundant = true; 558 break; 559 } 560 V = UserInsInst; 561 Depth++; 562 } 563 564 if (IsRedundant) 565 return replaceInstUsesWith(I, I.getOperand(0)); 566 return nullptr; 567 } 568 569 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) { 570 Value *VecOp = IE.getOperand(0); 571 Value *ScalarOp = IE.getOperand(1); 572 Value *IdxOp = IE.getOperand(2); 573 574 // Inserting an undef or into an undefined place, remove this. 575 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp)) 576 replaceInstUsesWith(IE, VecOp); 577 578 // If the inserted element was extracted from some other vector, and if the 579 // indexes are constant, try to turn this into a shufflevector operation. 580 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { 581 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) { 582 unsigned NumInsertVectorElts = IE.getType()->getNumElements(); 583 unsigned NumExtractVectorElts = 584 EI->getOperand(0)->getType()->getVectorNumElements(); 585 unsigned ExtractedIdx = 586 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 587 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 588 589 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract. 590 return replaceInstUsesWith(IE, VecOp); 591 592 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert. 593 return replaceInstUsesWith(IE, UndefValue::get(IE.getType())); 594 595 // If we are extracting a value from a vector, then inserting it right 596 // back into the same place, just use the input vector. 597 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx) 598 return replaceInstUsesWith(IE, VecOp); 599 600 // If this insertelement isn't used by some other insertelement, turn it 601 // (and any insertelements it points to), into one big shuffle. 602 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) { 603 SmallVector<Constant*, 16> Mask; 604 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this); 605 606 // The proposed shuffle may be trivial, in which case we shouldn't 607 // perform the combine. 608 if (LR.first != &IE && LR.second != &IE) { 609 // We now have a shuffle of LHS, RHS, Mask. 610 if (LR.second == nullptr) 611 LR.second = UndefValue::get(LR.first->getType()); 612 return new ShuffleVectorInst(LR.first, LR.second, 613 ConstantVector::get(Mask)); 614 } 615 } 616 } 617 } 618 619 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements(); 620 APInt UndefElts(VWidth, 0); 621 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 622 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) { 623 if (V != &IE) 624 return replaceInstUsesWith(IE, V); 625 return &IE; 626 } 627 628 return nullptr; 629 } 630 631 /// Return true if we can evaluate the specified expression tree if the vector 632 /// elements were shuffled in a different order. 633 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask, 634 unsigned Depth = 5) { 635 // We can always reorder the elements of a constant. 636 if (isa<Constant>(V)) 637 return true; 638 639 // We won't reorder vector arguments. No IPO here. 640 Instruction *I = dyn_cast<Instruction>(V); 641 if (!I) return false; 642 643 // Two users may expect different orders of the elements. Don't try it. 644 if (!I->hasOneUse()) 645 return false; 646 647 if (Depth == 0) return false; 648 649 switch (I->getOpcode()) { 650 case Instruction::Add: 651 case Instruction::FAdd: 652 case Instruction::Sub: 653 case Instruction::FSub: 654 case Instruction::Mul: 655 case Instruction::FMul: 656 case Instruction::UDiv: 657 case Instruction::SDiv: 658 case Instruction::FDiv: 659 case Instruction::URem: 660 case Instruction::SRem: 661 case Instruction::FRem: 662 case Instruction::Shl: 663 case Instruction::LShr: 664 case Instruction::AShr: 665 case Instruction::And: 666 case Instruction::Or: 667 case Instruction::Xor: 668 case Instruction::ICmp: 669 case Instruction::FCmp: 670 case Instruction::Trunc: 671 case Instruction::ZExt: 672 case Instruction::SExt: 673 case Instruction::FPToUI: 674 case Instruction::FPToSI: 675 case Instruction::UIToFP: 676 case Instruction::SIToFP: 677 case Instruction::FPTrunc: 678 case Instruction::FPExt: 679 case Instruction::GetElementPtr: { 680 for (Value *Operand : I->operands()) { 681 if (!CanEvaluateShuffled(Operand, Mask, Depth-1)) 682 return false; 683 } 684 return true; 685 } 686 case Instruction::InsertElement: { 687 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2)); 688 if (!CI) return false; 689 int ElementNumber = CI->getLimitedValue(); 690 691 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement' 692 // can't put an element into multiple indices. 693 bool SeenOnce = false; 694 for (int i = 0, e = Mask.size(); i != e; ++i) { 695 if (Mask[i] == ElementNumber) { 696 if (SeenOnce) 697 return false; 698 SeenOnce = true; 699 } 700 } 701 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1); 702 } 703 } 704 return false; 705 } 706 707 /// Rebuild a new instruction just like 'I' but with the new operands given. 708 /// In the event of type mismatch, the type of the operands is correct. 709 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) { 710 // We don't want to use the IRBuilder here because we want the replacement 711 // instructions to appear next to 'I', not the builder's insertion point. 712 switch (I->getOpcode()) { 713 case Instruction::Add: 714 case Instruction::FAdd: 715 case Instruction::Sub: 716 case Instruction::FSub: 717 case Instruction::Mul: 718 case Instruction::FMul: 719 case Instruction::UDiv: 720 case Instruction::SDiv: 721 case Instruction::FDiv: 722 case Instruction::URem: 723 case Instruction::SRem: 724 case Instruction::FRem: 725 case Instruction::Shl: 726 case Instruction::LShr: 727 case Instruction::AShr: 728 case Instruction::And: 729 case Instruction::Or: 730 case Instruction::Xor: { 731 BinaryOperator *BO = cast<BinaryOperator>(I); 732 assert(NewOps.size() == 2 && "binary operator with #ops != 2"); 733 BinaryOperator *New = 734 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(), 735 NewOps[0], NewOps[1], "", BO); 736 if (isa<OverflowingBinaryOperator>(BO)) { 737 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap()); 738 New->setHasNoSignedWrap(BO->hasNoSignedWrap()); 739 } 740 if (isa<PossiblyExactOperator>(BO)) { 741 New->setIsExact(BO->isExact()); 742 } 743 if (isa<FPMathOperator>(BO)) 744 New->copyFastMathFlags(I); 745 return New; 746 } 747 case Instruction::ICmp: 748 assert(NewOps.size() == 2 && "icmp with #ops != 2"); 749 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(), 750 NewOps[0], NewOps[1]); 751 case Instruction::FCmp: 752 assert(NewOps.size() == 2 && "fcmp with #ops != 2"); 753 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(), 754 NewOps[0], NewOps[1]); 755 case Instruction::Trunc: 756 case Instruction::ZExt: 757 case Instruction::SExt: 758 case Instruction::FPToUI: 759 case Instruction::FPToSI: 760 case Instruction::UIToFP: 761 case Instruction::SIToFP: 762 case Instruction::FPTrunc: 763 case Instruction::FPExt: { 764 // It's possible that the mask has a different number of elements from 765 // the original cast. We recompute the destination type to match the mask. 766 Type *DestTy = 767 VectorType::get(I->getType()->getScalarType(), 768 NewOps[0]->getType()->getVectorNumElements()); 769 assert(NewOps.size() == 1 && "cast with #ops != 1"); 770 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy, 771 "", I); 772 } 773 case Instruction::GetElementPtr: { 774 Value *Ptr = NewOps[0]; 775 ArrayRef<Value*> Idx = NewOps.slice(1); 776 GetElementPtrInst *GEP = GetElementPtrInst::Create( 777 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I); 778 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds()); 779 return GEP; 780 } 781 } 782 llvm_unreachable("failed to rebuild vector instructions"); 783 } 784 785 Value * 786 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) { 787 // Mask.size() does not need to be equal to the number of vector elements. 788 789 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements"); 790 if (isa<UndefValue>(V)) { 791 return UndefValue::get(VectorType::get(V->getType()->getScalarType(), 792 Mask.size())); 793 } 794 if (isa<ConstantAggregateZero>(V)) { 795 return ConstantAggregateZero::get( 796 VectorType::get(V->getType()->getScalarType(), 797 Mask.size())); 798 } 799 if (Constant *C = dyn_cast<Constant>(V)) { 800 SmallVector<Constant *, 16> MaskValues; 801 for (int i = 0, e = Mask.size(); i != e; ++i) { 802 if (Mask[i] == -1) 803 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty())); 804 else 805 MaskValues.push_back(Builder->getInt32(Mask[i])); 806 } 807 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()), 808 ConstantVector::get(MaskValues)); 809 } 810 811 Instruction *I = cast<Instruction>(V); 812 switch (I->getOpcode()) { 813 case Instruction::Add: 814 case Instruction::FAdd: 815 case Instruction::Sub: 816 case Instruction::FSub: 817 case Instruction::Mul: 818 case Instruction::FMul: 819 case Instruction::UDiv: 820 case Instruction::SDiv: 821 case Instruction::FDiv: 822 case Instruction::URem: 823 case Instruction::SRem: 824 case Instruction::FRem: 825 case Instruction::Shl: 826 case Instruction::LShr: 827 case Instruction::AShr: 828 case Instruction::And: 829 case Instruction::Or: 830 case Instruction::Xor: 831 case Instruction::ICmp: 832 case Instruction::FCmp: 833 case Instruction::Trunc: 834 case Instruction::ZExt: 835 case Instruction::SExt: 836 case Instruction::FPToUI: 837 case Instruction::FPToSI: 838 case Instruction::UIToFP: 839 case Instruction::SIToFP: 840 case Instruction::FPTrunc: 841 case Instruction::FPExt: 842 case Instruction::Select: 843 case Instruction::GetElementPtr: { 844 SmallVector<Value*, 8> NewOps; 845 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements()); 846 for (int i = 0, e = I->getNumOperands(); i != e; ++i) { 847 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask); 848 NewOps.push_back(V); 849 NeedsRebuild |= (V != I->getOperand(i)); 850 } 851 if (NeedsRebuild) { 852 return buildNew(I, NewOps); 853 } 854 return I; 855 } 856 case Instruction::InsertElement: { 857 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue(); 858 859 // The insertelement was inserting at Element. Figure out which element 860 // that becomes after shuffling. The answer is guaranteed to be unique 861 // by CanEvaluateShuffled. 862 bool Found = false; 863 int Index = 0; 864 for (int e = Mask.size(); Index != e; ++Index) { 865 if (Mask[Index] == Element) { 866 Found = true; 867 break; 868 } 869 } 870 871 // If element is not in Mask, no need to handle the operand 1 (element to 872 // be inserted). Just evaluate values in operand 0 according to Mask. 873 if (!Found) 874 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask); 875 876 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask); 877 return InsertElementInst::Create(V, I->getOperand(1), 878 Builder->getInt32(Index), "", I); 879 } 880 } 881 llvm_unreachable("failed to reorder elements of vector instruction!"); 882 } 883 884 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask, 885 bool &isLHSID, bool &isRHSID) { 886 isLHSID = isRHSID = true; 887 888 for (unsigned i = 0, e = Mask.size(); i != e; ++i) { 889 if (Mask[i] < 0) continue; // Ignore undef values. 890 // Is this an identity shuffle of the LHS value? 891 isLHSID &= (Mask[i] == (int)i); 892 893 // Is this an identity shuffle of the RHS value? 894 isRHSID &= (Mask[i]-e == i); 895 } 896 } 897 898 // Returns true if the shuffle is extracting a contiguous range of values from 899 // LHS, for example: 900 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ 901 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP| 902 // Shuffles to: |EE|FF|GG|HH| 903 // +--+--+--+--+ 904 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI, 905 SmallVector<int, 16> &Mask) { 906 unsigned LHSElems = 907 cast<VectorType>(SVI.getOperand(0)->getType())->getNumElements(); 908 unsigned MaskElems = Mask.size(); 909 unsigned BegIdx = Mask.front(); 910 unsigned EndIdx = Mask.back(); 911 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1) 912 return false; 913 for (unsigned I = 0; I != MaskElems; ++I) 914 if (static_cast<unsigned>(Mask[I]) != BegIdx + I) 915 return false; 916 return true; 917 } 918 919 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { 920 Value *LHS = SVI.getOperand(0); 921 Value *RHS = SVI.getOperand(1); 922 SmallVector<int, 16> Mask = SVI.getShuffleMask(); 923 Type *Int32Ty = Type::getInt32Ty(SVI.getContext()); 924 925 bool MadeChange = false; 926 927 // Undefined shuffle mask -> undefined value. 928 if (isa<UndefValue>(SVI.getOperand(2))) 929 return replaceInstUsesWith(SVI, UndefValue::get(SVI.getType())); 930 931 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements(); 932 933 APInt UndefElts(VWidth, 0); 934 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 935 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) { 936 if (V != &SVI) 937 return replaceInstUsesWith(SVI, V); 938 LHS = SVI.getOperand(0); 939 RHS = SVI.getOperand(1); 940 MadeChange = true; 941 } 942 943 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements(); 944 945 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask') 946 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask'). 947 if (LHS == RHS || isa<UndefValue>(LHS)) { 948 if (isa<UndefValue>(LHS) && LHS == RHS) { 949 // shuffle(undef,undef,mask) -> undef. 950 Value *Result = (VWidth == LHSWidth) 951 ? LHS : UndefValue::get(SVI.getType()); 952 return replaceInstUsesWith(SVI, Result); 953 } 954 955 // Remap any references to RHS to use LHS. 956 SmallVector<Constant*, 16> Elts; 957 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) { 958 if (Mask[i] < 0) { 959 Elts.push_back(UndefValue::get(Int32Ty)); 960 continue; 961 } 962 963 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) || 964 (Mask[i] < (int)e && isa<UndefValue>(LHS))) { 965 Mask[i] = -1; // Turn into undef. 966 Elts.push_back(UndefValue::get(Int32Ty)); 967 } else { 968 Mask[i] = Mask[i] % e; // Force to LHS. 969 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i])); 970 } 971 } 972 SVI.setOperand(0, SVI.getOperand(1)); 973 SVI.setOperand(1, UndefValue::get(RHS->getType())); 974 SVI.setOperand(2, ConstantVector::get(Elts)); 975 LHS = SVI.getOperand(0); 976 RHS = SVI.getOperand(1); 977 MadeChange = true; 978 } 979 980 if (VWidth == LHSWidth) { 981 // Analyze the shuffle, are the LHS or RHS and identity shuffles? 982 bool isLHSID, isRHSID; 983 recognizeIdentityMask(Mask, isLHSID, isRHSID); 984 985 // Eliminate identity shuffles. 986 if (isLHSID) return replaceInstUsesWith(SVI, LHS); 987 if (isRHSID) return replaceInstUsesWith(SVI, RHS); 988 } 989 990 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) { 991 Value *V = EvaluateInDifferentElementOrder(LHS, Mask); 992 return replaceInstUsesWith(SVI, V); 993 } 994 995 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to 996 // a non-vector type. We can instead bitcast the original vector followed by 997 // an extract of the desired element: 998 // 999 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef, 1000 // <4 x i32> <i32 0, i32 1, i32 2, i32 3> 1001 // %1 = bitcast <4 x i8> %sroa to i32 1002 // Becomes: 1003 // %bc = bitcast <16 x i8> %in to <4 x i32> 1004 // %ext = extractelement <4 x i32> %bc, i32 0 1005 // 1006 // If the shuffle is extracting a contiguous range of values from the input 1007 // vector then each use which is a bitcast of the extracted size can be 1008 // replaced. This will work if the vector types are compatible, and the begin 1009 // index is aligned to a value in the casted vector type. If the begin index 1010 // isn't aligned then we can shuffle the original vector (keeping the same 1011 // vector type) before extracting. 1012 // 1013 // This code will bail out if the target type is fundamentally incompatible 1014 // with vectors of the source type. 1015 // 1016 // Example of <16 x i8>, target type i32: 1017 // Index range [4,8): v-----------v Will work. 1018 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ 1019 // <16 x i8>: | | | | | | | | | | | | | | | | | 1020 // <4 x i32>: | | | | | 1021 // +-----------+-----------+-----------+-----------+ 1022 // Index range [6,10): ^-----------^ Needs an extra shuffle. 1023 // Target type i40: ^--------------^ Won't work, bail. 1024 if (isShuffleExtractingFromLHS(SVI, Mask)) { 1025 Value *V = LHS; 1026 unsigned MaskElems = Mask.size(); 1027 unsigned BegIdx = Mask.front(); 1028 VectorType *SrcTy = cast<VectorType>(V->getType()); 1029 unsigned VecBitWidth = SrcTy->getBitWidth(); 1030 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType()); 1031 assert(SrcElemBitWidth && "vector elements must have a bitwidth"); 1032 unsigned SrcNumElems = SrcTy->getNumElements(); 1033 SmallVector<BitCastInst *, 8> BCs; 1034 DenseMap<Type *, Value *> NewBCs; 1035 for (User *U : SVI.users()) 1036 if (BitCastInst *BC = dyn_cast<BitCastInst>(U)) 1037 if (!BC->use_empty()) 1038 // Only visit bitcasts that weren't previously handled. 1039 BCs.push_back(BC); 1040 for (BitCastInst *BC : BCs) { 1041 Type *TgtTy = BC->getDestTy(); 1042 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy); 1043 if (!TgtElemBitWidth) 1044 continue; 1045 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth; 1046 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth; 1047 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth); 1048 if (!VecBitWidthsEqual) 1049 continue; 1050 if (!VectorType::isValidElementType(TgtTy)) 1051 continue; 1052 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems); 1053 if (!BegIsAligned) { 1054 // Shuffle the input so [0,NumElements) contains the output, and 1055 // [NumElems,SrcNumElems) is undef. 1056 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems, 1057 UndefValue::get(Int32Ty)); 1058 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I) 1059 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx); 1060 V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()), 1061 ConstantVector::get(ShuffleMask), 1062 SVI.getName() + ".extract"); 1063 BegIdx = 0; 1064 } 1065 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth; 1066 assert(SrcElemsPerTgtElem); 1067 BegIdx /= SrcElemsPerTgtElem; 1068 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end(); 1069 auto *NewBC = 1070 BCAlreadyExists 1071 ? NewBCs[CastSrcTy] 1072 : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc"); 1073 if (!BCAlreadyExists) 1074 NewBCs[CastSrcTy] = NewBC; 1075 auto *Ext = Builder->CreateExtractElement( 1076 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract"); 1077 // The shufflevector isn't being replaced: the bitcast that used it 1078 // is. InstCombine will visit the newly-created instructions. 1079 replaceInstUsesWith(*BC, Ext); 1080 MadeChange = true; 1081 } 1082 } 1083 1084 // If the LHS is a shufflevector itself, see if we can combine it with this 1085 // one without producing an unusual shuffle. 1086 // Cases that might be simplified: 1087 // 1. 1088 // x1=shuffle(v1,v2,mask1) 1089 // x=shuffle(x1,undef,mask) 1090 // ==> 1091 // x=shuffle(v1,undef,newMask) 1092 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1 1093 // 2. 1094 // x1=shuffle(v1,undef,mask1) 1095 // x=shuffle(x1,x2,mask) 1096 // where v1.size() == mask1.size() 1097 // ==> 1098 // x=shuffle(v1,x2,newMask) 1099 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i] 1100 // 3. 1101 // x2=shuffle(v2,undef,mask2) 1102 // x=shuffle(x1,x2,mask) 1103 // where v2.size() == mask2.size() 1104 // ==> 1105 // x=shuffle(x1,v2,newMask) 1106 // newMask[i] = (mask[i] < x1.size()) 1107 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size() 1108 // 4. 1109 // x1=shuffle(v1,undef,mask1) 1110 // x2=shuffle(v2,undef,mask2) 1111 // x=shuffle(x1,x2,mask) 1112 // where v1.size() == v2.size() 1113 // ==> 1114 // x=shuffle(v1,v2,newMask) 1115 // newMask[i] = (mask[i] < x1.size()) 1116 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size() 1117 // 1118 // Here we are really conservative: 1119 // we are absolutely afraid of producing a shuffle mask not in the input 1120 // program, because the code gen may not be smart enough to turn a merged 1121 // shuffle into two specific shuffles: it may produce worse code. As such, 1122 // we only merge two shuffles if the result is either a splat or one of the 1123 // input shuffle masks. In this case, merging the shuffles just removes 1124 // one instruction, which we know is safe. This is good for things like 1125 // turning: (splat(splat)) -> splat, or 1126 // merge(V[0..n], V[n+1..2n]) -> V[0..2n] 1127 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS); 1128 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS); 1129 if (LHSShuffle) 1130 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS)) 1131 LHSShuffle = nullptr; 1132 if (RHSShuffle) 1133 if (!isa<UndefValue>(RHSShuffle->getOperand(1))) 1134 RHSShuffle = nullptr; 1135 if (!LHSShuffle && !RHSShuffle) 1136 return MadeChange ? &SVI : nullptr; 1137 1138 Value* LHSOp0 = nullptr; 1139 Value* LHSOp1 = nullptr; 1140 Value* RHSOp0 = nullptr; 1141 unsigned LHSOp0Width = 0; 1142 unsigned RHSOp0Width = 0; 1143 if (LHSShuffle) { 1144 LHSOp0 = LHSShuffle->getOperand(0); 1145 LHSOp1 = LHSShuffle->getOperand(1); 1146 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements(); 1147 } 1148 if (RHSShuffle) { 1149 RHSOp0 = RHSShuffle->getOperand(0); 1150 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements(); 1151 } 1152 Value* newLHS = LHS; 1153 Value* newRHS = RHS; 1154 if (LHSShuffle) { 1155 // case 1 1156 if (isa<UndefValue>(RHS)) { 1157 newLHS = LHSOp0; 1158 newRHS = LHSOp1; 1159 } 1160 // case 2 or 4 1161 else if (LHSOp0Width == LHSWidth) { 1162 newLHS = LHSOp0; 1163 } 1164 } 1165 // case 3 or 4 1166 if (RHSShuffle && RHSOp0Width == LHSWidth) { 1167 newRHS = RHSOp0; 1168 } 1169 // case 4 1170 if (LHSOp0 == RHSOp0) { 1171 newLHS = LHSOp0; 1172 newRHS = nullptr; 1173 } 1174 1175 if (newLHS == LHS && newRHS == RHS) 1176 return MadeChange ? &SVI : nullptr; 1177 1178 SmallVector<int, 16> LHSMask; 1179 SmallVector<int, 16> RHSMask; 1180 if (newLHS != LHS) 1181 LHSMask = LHSShuffle->getShuffleMask(); 1182 if (RHSShuffle && newRHS != RHS) 1183 RHSMask = RHSShuffle->getShuffleMask(); 1184 1185 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth; 1186 SmallVector<int, 16> newMask; 1187 bool isSplat = true; 1188 int SplatElt = -1; 1189 // Create a new mask for the new ShuffleVectorInst so that the new 1190 // ShuffleVectorInst is equivalent to the original one. 1191 for (unsigned i = 0; i < VWidth; ++i) { 1192 int eltMask; 1193 if (Mask[i] < 0) { 1194 // This element is an undef value. 1195 eltMask = -1; 1196 } else if (Mask[i] < (int)LHSWidth) { 1197 // This element is from left hand side vector operand. 1198 // 1199 // If LHS is going to be replaced (case 1, 2, or 4), calculate the 1200 // new mask value for the element. 1201 if (newLHS != LHS) { 1202 eltMask = LHSMask[Mask[i]]; 1203 // If the value selected is an undef value, explicitly specify it 1204 // with a -1 mask value. 1205 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1)) 1206 eltMask = -1; 1207 } else 1208 eltMask = Mask[i]; 1209 } else { 1210 // This element is from right hand side vector operand 1211 // 1212 // If the value selected is an undef value, explicitly specify it 1213 // with a -1 mask value. (case 1) 1214 if (isa<UndefValue>(RHS)) 1215 eltMask = -1; 1216 // If RHS is going to be replaced (case 3 or 4), calculate the 1217 // new mask value for the element. 1218 else if (newRHS != RHS) { 1219 eltMask = RHSMask[Mask[i]-LHSWidth]; 1220 // If the value selected is an undef value, explicitly specify it 1221 // with a -1 mask value. 1222 if (eltMask >= (int)RHSOp0Width) { 1223 assert(isa<UndefValue>(RHSShuffle->getOperand(1)) 1224 && "should have been check above"); 1225 eltMask = -1; 1226 } 1227 } else 1228 eltMask = Mask[i]-LHSWidth; 1229 1230 // If LHS's width is changed, shift the mask value accordingly. 1231 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any 1232 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask. 1233 // If newRHS == newLHS, we want to remap any references from newRHS to 1234 // newLHS so that we can properly identify splats that may occur due to 1235 // obfuscation across the two vectors. 1236 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS) 1237 eltMask += newLHSWidth; 1238 } 1239 1240 // Check if this could still be a splat. 1241 if (eltMask >= 0) { 1242 if (SplatElt >= 0 && SplatElt != eltMask) 1243 isSplat = false; 1244 SplatElt = eltMask; 1245 } 1246 1247 newMask.push_back(eltMask); 1248 } 1249 1250 // If the result mask is equal to one of the original shuffle masks, 1251 // or is a splat, do the replacement. 1252 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) { 1253 SmallVector<Constant*, 16> Elts; 1254 for (unsigned i = 0, e = newMask.size(); i != e; ++i) { 1255 if (newMask[i] < 0) { 1256 Elts.push_back(UndefValue::get(Int32Ty)); 1257 } else { 1258 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i])); 1259 } 1260 } 1261 if (!newRHS) 1262 newRHS = UndefValue::get(newLHS->getType()); 1263 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts)); 1264 } 1265 1266 // If the result mask is an identity, replace uses of this instruction with 1267 // corresponding argument. 1268 bool isLHSID, isRHSID; 1269 recognizeIdentityMask(newMask, isLHSID, isRHSID); 1270 if (isLHSID && VWidth == LHSOp0Width) return replaceInstUsesWith(SVI, newLHS); 1271 if (isRHSID && VWidth == RHSOp0Width) return replaceInstUsesWith(SVI, newRHS); 1272 1273 return MadeChange ? &SVI : nullptr; 1274 } 1275