1 //===- InstCombineVectorOps.cpp -------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements instcombine for ExtractElement, InsertElement and 10 // ShuffleVector. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "InstCombineInternal.h" 15 #include "llvm/ADT/APInt.h" 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/VectorUtils.h" 22 #include "llvm/IR/BasicBlock.h" 23 #include "llvm/IR/Constant.h" 24 #include "llvm/IR/Constants.h" 25 #include "llvm/IR/DerivedTypes.h" 26 #include "llvm/IR/InstrTypes.h" 27 #include "llvm/IR/Instruction.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/Operator.h" 30 #include "llvm/IR/PatternMatch.h" 31 #include "llvm/IR/Type.h" 32 #include "llvm/IR/User.h" 33 #include "llvm/IR/Value.h" 34 #include "llvm/Support/Casting.h" 35 #include "llvm/Support/ErrorHandling.h" 36 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" 37 #include <cassert> 38 #include <cstdint> 39 #include <iterator> 40 #include <utility> 41 42 using namespace llvm; 43 using namespace PatternMatch; 44 45 #define DEBUG_TYPE "instcombine" 46 47 /// Return true if the value is cheaper to scalarize than it is to leave as a 48 /// vector operation. IsConstantExtractIndex indicates whether we are extracting 49 /// one known element from a vector constant. 50 /// 51 /// FIXME: It's possible to create more instructions than previously existed. 52 static bool cheapToScalarize(Value *V, bool IsConstantExtractIndex) { 53 // If we can pick a scalar constant value out of a vector, that is free. 54 if (auto *C = dyn_cast<Constant>(V)) 55 return IsConstantExtractIndex || C->getSplatValue(); 56 57 // An insertelement to the same constant index as our extract will simplify 58 // to the scalar inserted element. An insertelement to a different constant 59 // index is irrelevant to our extract. 60 if (match(V, m_InsertElement(m_Value(), m_Value(), m_ConstantInt()))) 61 return IsConstantExtractIndex; 62 63 if (match(V, m_OneUse(m_Load(m_Value())))) 64 return true; 65 66 Value *V0, *V1; 67 if (match(V, m_OneUse(m_BinOp(m_Value(V0), m_Value(V1))))) 68 if (cheapToScalarize(V0, IsConstantExtractIndex) || 69 cheapToScalarize(V1, IsConstantExtractIndex)) 70 return true; 71 72 CmpInst::Predicate UnusedPred; 73 if (match(V, m_OneUse(m_Cmp(UnusedPred, m_Value(V0), m_Value(V1))))) 74 if (cheapToScalarize(V0, IsConstantExtractIndex) || 75 cheapToScalarize(V1, IsConstantExtractIndex)) 76 return true; 77 78 return false; 79 } 80 81 // If we have a PHI node with a vector type that is only used to feed 82 // itself and be an operand of extractelement at a constant location, 83 // try to replace the PHI of the vector type with a PHI of a scalar type. 84 Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) { 85 SmallVector<Instruction *, 2> Extracts; 86 // The users we want the PHI to have are: 87 // 1) The EI ExtractElement (we already know this) 88 // 2) Possibly more ExtractElements with the same index. 89 // 3) Another operand, which will feed back into the PHI. 90 Instruction *PHIUser = nullptr; 91 for (auto U : PN->users()) { 92 if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(U)) { 93 if (EI.getIndexOperand() == EU->getIndexOperand()) 94 Extracts.push_back(EU); 95 else 96 return nullptr; 97 } else if (!PHIUser) { 98 PHIUser = cast<Instruction>(U); 99 } else { 100 return nullptr; 101 } 102 } 103 104 if (!PHIUser) 105 return nullptr; 106 107 // Verify that this PHI user has one use, which is the PHI itself, 108 // and that it is a binary operation which is cheap to scalarize. 109 // otherwise return nullptr. 110 if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) || 111 !(isa<BinaryOperator>(PHIUser)) || !cheapToScalarize(PHIUser, true)) 112 return nullptr; 113 114 // Create a scalar PHI node that will replace the vector PHI node 115 // just before the current PHI node. 116 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith( 117 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN)); 118 // Scalarize each PHI operand. 119 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) { 120 Value *PHIInVal = PN->getIncomingValue(i); 121 BasicBlock *inBB = PN->getIncomingBlock(i); 122 Value *Elt = EI.getIndexOperand(); 123 // If the operand is the PHI induction variable: 124 if (PHIInVal == PHIUser) { 125 // Scalarize the binary operation. Its first operand is the 126 // scalar PHI, and the second operand is extracted from the other 127 // vector operand. 128 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser); 129 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0; 130 Value *Op = InsertNewInstWith( 131 ExtractElementInst::Create(B0->getOperand(opId), Elt, 132 B0->getOperand(opId)->getName() + ".Elt"), 133 *B0); 134 Value *newPHIUser = InsertNewInstWith( 135 BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(), 136 scalarPHI, Op, B0), *B0); 137 scalarPHI->addIncoming(newPHIUser, inBB); 138 } else { 139 // Scalarize PHI input: 140 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, ""); 141 // Insert the new instruction into the predecessor basic block. 142 Instruction *pos = dyn_cast<Instruction>(PHIInVal); 143 BasicBlock::iterator InsertPos; 144 if (pos && !isa<PHINode>(pos)) { 145 InsertPos = ++pos->getIterator(); 146 } else { 147 InsertPos = inBB->getFirstInsertionPt(); 148 } 149 150 InsertNewInstWith(newEI, *InsertPos); 151 152 scalarPHI->addIncoming(newEI, inBB); 153 } 154 } 155 156 for (auto E : Extracts) 157 replaceInstUsesWith(*E, scalarPHI); 158 159 return &EI; 160 } 161 162 static Instruction *foldBitcastExtElt(ExtractElementInst &Ext, 163 InstCombiner::BuilderTy &Builder, 164 bool IsBigEndian) { 165 Value *X; 166 uint64_t ExtIndexC; 167 if (!match(Ext.getVectorOperand(), m_BitCast(m_Value(X))) || 168 !X->getType()->isVectorTy() || 169 !match(Ext.getIndexOperand(), m_ConstantInt(ExtIndexC))) 170 return nullptr; 171 172 // If this extractelement is using a bitcast from a vector of the same number 173 // of elements, see if we can find the source element from the source vector: 174 // extelt (bitcast VecX), IndexC --> bitcast X[IndexC] 175 Type *SrcTy = X->getType(); 176 Type *DestTy = Ext.getType(); 177 unsigned NumSrcElts = SrcTy->getVectorNumElements(); 178 unsigned NumElts = Ext.getVectorOperandType()->getNumElements(); 179 if (NumSrcElts == NumElts) 180 if (Value *Elt = findScalarElement(X, ExtIndexC)) 181 return new BitCastInst(Elt, DestTy); 182 183 // If the source elements are wider than the destination, try to shift and 184 // truncate a subset of scalar bits of an insert op. 185 if (NumSrcElts < NumElts) { 186 Value *Scalar; 187 uint64_t InsIndexC; 188 if (!match(X, m_InsertElement(m_Value(), m_Value(Scalar), 189 m_ConstantInt(InsIndexC)))) 190 return nullptr; 191 192 // The extract must be from the subset of vector elements that we inserted 193 // into. Example: if we inserted element 1 of a <2 x i64> and we are 194 // extracting an i16 (narrowing ratio = 4), then this extract must be from 1 195 // of elements 4-7 of the bitcasted vector. 196 unsigned NarrowingRatio = NumElts / NumSrcElts; 197 if (ExtIndexC / NarrowingRatio != InsIndexC) 198 return nullptr; 199 200 // We are extracting part of the original scalar. How that scalar is 201 // inserted into the vector depends on the endian-ness. Example: 202 // Vector Byte Elt Index: 0 1 2 3 4 5 6 7 203 // +--+--+--+--+--+--+--+--+ 204 // inselt <2 x i32> V, <i32> S, 1: |V0|V1|V2|V3|S0|S1|S2|S3| 205 // extelt <4 x i16> V', 3: | |S2|S3| 206 // +--+--+--+--+--+--+--+--+ 207 // If this is little-endian, S2|S3 are the MSB of the 32-bit 'S' value. 208 // If this is big-endian, S2|S3 are the LSB of the 32-bit 'S' value. 209 // In this example, we must right-shift little-endian. Big-endian is just a 210 // truncate. 211 unsigned Chunk = ExtIndexC % NarrowingRatio; 212 if (IsBigEndian) 213 Chunk = NarrowingRatio - 1 - Chunk; 214 215 // Bail out if this is an FP vector to FP vector sequence. That would take 216 // more instructions than we started with unless there is no shift, and it 217 // may not be handled as well in the backend. 218 bool NeedSrcBitcast = SrcTy->getScalarType()->isFloatingPointTy(); 219 bool NeedDestBitcast = DestTy->isFloatingPointTy(); 220 if (NeedSrcBitcast && NeedDestBitcast) 221 return nullptr; 222 223 unsigned SrcWidth = SrcTy->getScalarSizeInBits(); 224 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 225 unsigned ShAmt = Chunk * DestWidth; 226 227 // TODO: This limitation is more strict than necessary. We could sum the 228 // number of new instructions and subtract the number eliminated to know if 229 // we can proceed. 230 if (!X->hasOneUse() || !Ext.getVectorOperand()->hasOneUse()) 231 if (NeedSrcBitcast || NeedDestBitcast) 232 return nullptr; 233 234 if (NeedSrcBitcast) { 235 Type *SrcIntTy = IntegerType::getIntNTy(Scalar->getContext(), SrcWidth); 236 Scalar = Builder.CreateBitCast(Scalar, SrcIntTy); 237 } 238 239 if (ShAmt) { 240 // Bail out if we could end with more instructions than we started with. 241 if (!Ext.getVectorOperand()->hasOneUse()) 242 return nullptr; 243 Scalar = Builder.CreateLShr(Scalar, ShAmt); 244 } 245 246 if (NeedDestBitcast) { 247 Type *DestIntTy = IntegerType::getIntNTy(Scalar->getContext(), DestWidth); 248 return new BitCastInst(Builder.CreateTrunc(Scalar, DestIntTy), DestTy); 249 } 250 return new TruncInst(Scalar, DestTy); 251 } 252 253 return nullptr; 254 } 255 256 /// Find elements of V demanded by UserInstr. 257 static APInt findDemandedEltsBySingleUser(Value *V, Instruction *UserInstr) { 258 unsigned VWidth = V->getType()->getVectorNumElements(); 259 260 // Conservatively assume that all elements are needed. 261 APInt UsedElts(APInt::getAllOnesValue(VWidth)); 262 263 switch (UserInstr->getOpcode()) { 264 case Instruction::ExtractElement: { 265 ExtractElementInst *EEI = cast<ExtractElementInst>(UserInstr); 266 assert(EEI->getVectorOperand() == V); 267 ConstantInt *EEIIndexC = dyn_cast<ConstantInt>(EEI->getIndexOperand()); 268 if (EEIIndexC && EEIIndexC->getValue().ult(VWidth)) { 269 UsedElts = APInt::getOneBitSet(VWidth, EEIIndexC->getZExtValue()); 270 } 271 break; 272 } 273 case Instruction::ShuffleVector: { 274 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(UserInstr); 275 unsigned MaskNumElts = UserInstr->getType()->getVectorNumElements(); 276 277 UsedElts = APInt(VWidth, 0); 278 for (unsigned i = 0; i < MaskNumElts; i++) { 279 unsigned MaskVal = Shuffle->getMaskValue(i); 280 if (MaskVal == -1u || MaskVal >= 2 * VWidth) 281 continue; 282 if (Shuffle->getOperand(0) == V && (MaskVal < VWidth)) 283 UsedElts.setBit(MaskVal); 284 if (Shuffle->getOperand(1) == V && 285 ((MaskVal >= VWidth) && (MaskVal < 2 * VWidth))) 286 UsedElts.setBit(MaskVal - VWidth); 287 } 288 break; 289 } 290 default: 291 break; 292 } 293 return UsedElts; 294 } 295 296 /// Find union of elements of V demanded by all its users. 297 /// If it is known by querying findDemandedEltsBySingleUser that 298 /// no user demands an element of V, then the corresponding bit 299 /// remains unset in the returned value. 300 static APInt findDemandedEltsByAllUsers(Value *V) { 301 unsigned VWidth = V->getType()->getVectorNumElements(); 302 303 APInt UnionUsedElts(VWidth, 0); 304 for (const Use &U : V->uses()) { 305 if (Instruction *I = dyn_cast<Instruction>(U.getUser())) { 306 UnionUsedElts |= findDemandedEltsBySingleUser(V, I); 307 } else { 308 UnionUsedElts = APInt::getAllOnesValue(VWidth); 309 break; 310 } 311 312 if (UnionUsedElts.isAllOnesValue()) 313 break; 314 } 315 316 return UnionUsedElts; 317 } 318 319 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { 320 Value *SrcVec = EI.getVectorOperand(); 321 Value *Index = EI.getIndexOperand(); 322 if (Value *V = SimplifyExtractElementInst(SrcVec, Index, 323 SQ.getWithInstruction(&EI))) 324 return replaceInstUsesWith(EI, V); 325 326 // If extracting a specified index from the vector, see if we can recursively 327 // find a previously computed scalar that was inserted into the vector. 328 auto *IndexC = dyn_cast<ConstantInt>(Index); 329 if (IndexC) { 330 unsigned NumElts = EI.getVectorOperandType()->getNumElements(); 331 332 // InstSimplify should handle cases where the index is invalid. 333 if (!IndexC->getValue().ule(NumElts)) 334 return nullptr; 335 336 // This instruction only demands the single element from the input vector. 337 if (NumElts != 1) { 338 // If the input vector has a single use, simplify it based on this use 339 // property. 340 if (SrcVec->hasOneUse()) { 341 APInt UndefElts(NumElts, 0); 342 APInt DemandedElts(NumElts, 0); 343 DemandedElts.setBit(IndexC->getZExtValue()); 344 if (Value *V = 345 SimplifyDemandedVectorElts(SrcVec, DemandedElts, UndefElts)) { 346 EI.setOperand(0, V); 347 return &EI; 348 } 349 } else { 350 // If the input vector has multiple uses, simplify it based on a union 351 // of all elements used. 352 APInt DemandedElts = findDemandedEltsByAllUsers(SrcVec); 353 if (!DemandedElts.isAllOnesValue()) { 354 APInt UndefElts(NumElts, 0); 355 if (Value *V = SimplifyDemandedVectorElts( 356 SrcVec, DemandedElts, UndefElts, 0 /* Depth */, 357 true /* AllowMultipleUsers */)) { 358 if (V != SrcVec) { 359 SrcVec->replaceAllUsesWith(V); 360 return &EI; 361 } 362 } 363 } 364 } 365 } 366 if (Instruction *I = foldBitcastExtElt(EI, Builder, DL.isBigEndian())) 367 return I; 368 369 // If there's a vector PHI feeding a scalar use through this extractelement 370 // instruction, try to scalarize the PHI. 371 if (auto *Phi = dyn_cast<PHINode>(SrcVec)) 372 if (Instruction *ScalarPHI = scalarizePHI(EI, Phi)) 373 return ScalarPHI; 374 } 375 376 BinaryOperator *BO; 377 if (match(SrcVec, m_BinOp(BO)) && cheapToScalarize(SrcVec, IndexC)) { 378 // extelt (binop X, Y), Index --> binop (extelt X, Index), (extelt Y, Index) 379 Value *X = BO->getOperand(0), *Y = BO->getOperand(1); 380 Value *E0 = Builder.CreateExtractElement(X, Index); 381 Value *E1 = Builder.CreateExtractElement(Y, Index); 382 return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(), E0, E1, BO); 383 } 384 385 Value *X, *Y; 386 CmpInst::Predicate Pred; 387 if (match(SrcVec, m_Cmp(Pred, m_Value(X), m_Value(Y))) && 388 cheapToScalarize(SrcVec, IndexC)) { 389 // extelt (cmp X, Y), Index --> cmp (extelt X, Index), (extelt Y, Index) 390 Value *E0 = Builder.CreateExtractElement(X, Index); 391 Value *E1 = Builder.CreateExtractElement(Y, Index); 392 return CmpInst::Create(cast<CmpInst>(SrcVec)->getOpcode(), Pred, E0, E1); 393 } 394 395 if (auto *I = dyn_cast<Instruction>(SrcVec)) { 396 if (auto *IE = dyn_cast<InsertElementInst>(I)) { 397 // Extracting the inserted element? 398 if (IE->getOperand(2) == Index) 399 return replaceInstUsesWith(EI, IE->getOperand(1)); 400 // If the inserted and extracted elements are constants, they must not 401 // be the same value, extract from the pre-inserted value instead. 402 if (isa<Constant>(IE->getOperand(2)) && IndexC) { 403 Worklist.AddValue(SrcVec); 404 EI.setOperand(0, IE->getOperand(0)); 405 return &EI; 406 } 407 } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(I)) { 408 // If this is extracting an element from a shufflevector, figure out where 409 // it came from and extract from the appropriate input element instead. 410 if (auto *Elt = dyn_cast<ConstantInt>(Index)) { 411 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue()); 412 Value *Src; 413 unsigned LHSWidth = 414 SVI->getOperand(0)->getType()->getVectorNumElements(); 415 416 if (SrcIdx < 0) 417 return replaceInstUsesWith(EI, UndefValue::get(EI.getType())); 418 if (SrcIdx < (int)LHSWidth) 419 Src = SVI->getOperand(0); 420 else { 421 SrcIdx -= LHSWidth; 422 Src = SVI->getOperand(1); 423 } 424 Type *Int32Ty = Type::getInt32Ty(EI.getContext()); 425 return ExtractElementInst::Create(Src, 426 ConstantInt::get(Int32Ty, 427 SrcIdx, false)); 428 } 429 } else if (auto *CI = dyn_cast<CastInst>(I)) { 430 // Canonicalize extractelement(cast) -> cast(extractelement). 431 // Bitcasts can change the number of vector elements, and they cost 432 // nothing. 433 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) { 434 Value *EE = Builder.CreateExtractElement(CI->getOperand(0), Index); 435 Worklist.AddValue(EE); 436 return CastInst::Create(CI->getOpcode(), EE, EI.getType()); 437 } 438 } 439 } 440 return nullptr; 441 } 442 443 /// If V is a shuffle of values that ONLY returns elements from either LHS or 444 /// RHS, return the shuffle mask and true. Otherwise, return false. 445 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS, 446 SmallVectorImpl<Constant*> &Mask) { 447 assert(LHS->getType() == RHS->getType() && 448 "Invalid CollectSingleShuffleElements"); 449 unsigned NumElts = V->getType()->getVectorNumElements(); 450 451 if (isa<UndefValue>(V)) { 452 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 453 return true; 454 } 455 456 if (V == LHS) { 457 for (unsigned i = 0; i != NumElts; ++i) 458 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 459 return true; 460 } 461 462 if (V == RHS) { 463 for (unsigned i = 0; i != NumElts; ++i) 464 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), 465 i+NumElts)); 466 return true; 467 } 468 469 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 470 // If this is an insert of an extract from some other vector, include it. 471 Value *VecOp = IEI->getOperand(0); 472 Value *ScalarOp = IEI->getOperand(1); 473 Value *IdxOp = IEI->getOperand(2); 474 475 if (!isa<ConstantInt>(IdxOp)) 476 return false; 477 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 478 479 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector. 480 // We can handle this if the vector we are inserting into is 481 // transitively ok. 482 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 483 // If so, update the mask to reflect the inserted undef. 484 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext())); 485 return true; 486 } 487 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){ 488 if (isa<ConstantInt>(EI->getOperand(1))) { 489 unsigned ExtractedIdx = 490 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 491 unsigned NumLHSElts = LHS->getType()->getVectorNumElements(); 492 493 // This must be extracting from either LHS or RHS. 494 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) { 495 // We can handle this if the vector we are inserting into is 496 // transitively ok. 497 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 498 // If so, update the mask to reflect the inserted value. 499 if (EI->getOperand(0) == LHS) { 500 Mask[InsertedIdx % NumElts] = 501 ConstantInt::get(Type::getInt32Ty(V->getContext()), 502 ExtractedIdx); 503 } else { 504 assert(EI->getOperand(0) == RHS); 505 Mask[InsertedIdx % NumElts] = 506 ConstantInt::get(Type::getInt32Ty(V->getContext()), 507 ExtractedIdx + NumLHSElts); 508 } 509 return true; 510 } 511 } 512 } 513 } 514 } 515 516 return false; 517 } 518 519 /// If we have insertion into a vector that is wider than the vector that we 520 /// are extracting from, try to widen the source vector to allow a single 521 /// shufflevector to replace one or more insert/extract pairs. 522 static void replaceExtractElements(InsertElementInst *InsElt, 523 ExtractElementInst *ExtElt, 524 InstCombiner &IC) { 525 VectorType *InsVecType = InsElt->getType(); 526 VectorType *ExtVecType = ExtElt->getVectorOperandType(); 527 unsigned NumInsElts = InsVecType->getVectorNumElements(); 528 unsigned NumExtElts = ExtVecType->getVectorNumElements(); 529 530 // The inserted-to vector must be wider than the extracted-from vector. 531 if (InsVecType->getElementType() != ExtVecType->getElementType() || 532 NumExtElts >= NumInsElts) 533 return; 534 535 // Create a shuffle mask to widen the extended-from vector using undefined 536 // values. The mask selects all of the values of the original vector followed 537 // by as many undefined values as needed to create a vector of the same length 538 // as the inserted-to vector. 539 SmallVector<Constant *, 16> ExtendMask; 540 IntegerType *IntType = Type::getInt32Ty(InsElt->getContext()); 541 for (unsigned i = 0; i < NumExtElts; ++i) 542 ExtendMask.push_back(ConstantInt::get(IntType, i)); 543 for (unsigned i = NumExtElts; i < NumInsElts; ++i) 544 ExtendMask.push_back(UndefValue::get(IntType)); 545 546 Value *ExtVecOp = ExtElt->getVectorOperand(); 547 auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp); 548 BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst)) 549 ? ExtVecOpInst->getParent() 550 : ExtElt->getParent(); 551 552 // TODO: This restriction matches the basic block check below when creating 553 // new extractelement instructions. If that limitation is removed, this one 554 // could also be removed. But for now, we just bail out to ensure that we 555 // will replace the extractelement instruction that is feeding our 556 // insertelement instruction. This allows the insertelement to then be 557 // replaced by a shufflevector. If the insertelement is not replaced, we can 558 // induce infinite looping because there's an optimization for extractelement 559 // that will delete our widening shuffle. This would trigger another attempt 560 // here to create that shuffle, and we spin forever. 561 if (InsertionBlock != InsElt->getParent()) 562 return; 563 564 // TODO: This restriction matches the check in visitInsertElementInst() and 565 // prevents an infinite loop caused by not turning the extract/insert pair 566 // into a shuffle. We really should not need either check, but we're lacking 567 // folds for shufflevectors because we're afraid to generate shuffle masks 568 // that the backend can't handle. 569 if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back())) 570 return; 571 572 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType), 573 ConstantVector::get(ExtendMask)); 574 575 // Insert the new shuffle after the vector operand of the extract is defined 576 // (as long as it's not a PHI) or at the start of the basic block of the 577 // extract, so any subsequent extracts in the same basic block can use it. 578 // TODO: Insert before the earliest ExtractElementInst that is replaced. 579 if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst)) 580 WideVec->insertAfter(ExtVecOpInst); 581 else 582 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt()); 583 584 // Replace extracts from the original narrow vector with extracts from the new 585 // wide vector. 586 for (User *U : ExtVecOp->users()) { 587 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U); 588 if (!OldExt || OldExt->getParent() != WideVec->getParent()) 589 continue; 590 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1)); 591 NewExt->insertAfter(OldExt); 592 IC.replaceInstUsesWith(*OldExt, NewExt); 593 } 594 } 595 596 /// We are building a shuffle to create V, which is a sequence of insertelement, 597 /// extractelement pairs. If PermittedRHS is set, then we must either use it or 598 /// not rely on the second vector source. Return a std::pair containing the 599 /// left and right vectors of the proposed shuffle (or 0), and set the Mask 600 /// parameter as required. 601 /// 602 /// Note: we intentionally don't try to fold earlier shuffles since they have 603 /// often been chosen carefully to be efficiently implementable on the target. 604 using ShuffleOps = std::pair<Value *, Value *>; 605 606 static ShuffleOps collectShuffleElements(Value *V, 607 SmallVectorImpl<Constant *> &Mask, 608 Value *PermittedRHS, 609 InstCombiner &IC) { 610 assert(V->getType()->isVectorTy() && "Invalid shuffle!"); 611 unsigned NumElts = V->getType()->getVectorNumElements(); 612 613 if (isa<UndefValue>(V)) { 614 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 615 return std::make_pair( 616 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr); 617 } 618 619 if (isa<ConstantAggregateZero>(V)) { 620 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0)); 621 return std::make_pair(V, nullptr); 622 } 623 624 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 625 // If this is an insert of an extract from some other vector, include it. 626 Value *VecOp = IEI->getOperand(0); 627 Value *ScalarOp = IEI->getOperand(1); 628 Value *IdxOp = IEI->getOperand(2); 629 630 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { 631 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) { 632 unsigned ExtractedIdx = 633 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 634 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 635 636 // Either the extracted from or inserted into vector must be RHSVec, 637 // otherwise we'd end up with a shuffle of three inputs. 638 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) { 639 Value *RHS = EI->getOperand(0); 640 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC); 641 assert(LR.second == nullptr || LR.second == RHS); 642 643 if (LR.first->getType() != RHS->getType()) { 644 // Although we are giving up for now, see if we can create extracts 645 // that match the inserts for another round of combining. 646 replaceExtractElements(IEI, EI, IC); 647 648 // We tried our best, but we can't find anything compatible with RHS 649 // further up the chain. Return a trivial shuffle. 650 for (unsigned i = 0; i < NumElts; ++i) 651 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i); 652 return std::make_pair(V, nullptr); 653 } 654 655 unsigned NumLHSElts = RHS->getType()->getVectorNumElements(); 656 Mask[InsertedIdx % NumElts] = 657 ConstantInt::get(Type::getInt32Ty(V->getContext()), 658 NumLHSElts+ExtractedIdx); 659 return std::make_pair(LR.first, RHS); 660 } 661 662 if (VecOp == PermittedRHS) { 663 // We've gone as far as we can: anything on the other side of the 664 // extractelement will already have been converted into a shuffle. 665 unsigned NumLHSElts = 666 EI->getOperand(0)->getType()->getVectorNumElements(); 667 for (unsigned i = 0; i != NumElts; ++i) 668 Mask.push_back(ConstantInt::get( 669 Type::getInt32Ty(V->getContext()), 670 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i)); 671 return std::make_pair(EI->getOperand(0), PermittedRHS); 672 } 673 674 // If this insertelement is a chain that comes from exactly these two 675 // vectors, return the vector and the effective shuffle. 676 if (EI->getOperand(0)->getType() == PermittedRHS->getType() && 677 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS, 678 Mask)) 679 return std::make_pair(EI->getOperand(0), PermittedRHS); 680 } 681 } 682 } 683 684 // Otherwise, we can't do anything fancy. Return an identity vector. 685 for (unsigned i = 0; i != NumElts; ++i) 686 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 687 return std::make_pair(V, nullptr); 688 } 689 690 /// Try to find redundant insertvalue instructions, like the following ones: 691 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0 692 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0 693 /// Here the second instruction inserts values at the same indices, as the 694 /// first one, making the first one redundant. 695 /// It should be transformed to: 696 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0 697 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) { 698 bool IsRedundant = false; 699 ArrayRef<unsigned int> FirstIndices = I.getIndices(); 700 701 // If there is a chain of insertvalue instructions (each of them except the 702 // last one has only one use and it's another insertvalue insn from this 703 // chain), check if any of the 'children' uses the same indices as the first 704 // instruction. In this case, the first one is redundant. 705 Value *V = &I; 706 unsigned Depth = 0; 707 while (V->hasOneUse() && Depth < 10) { 708 User *U = V->user_back(); 709 auto UserInsInst = dyn_cast<InsertValueInst>(U); 710 if (!UserInsInst || U->getOperand(0) != V) 711 break; 712 if (UserInsInst->getIndices() == FirstIndices) { 713 IsRedundant = true; 714 break; 715 } 716 V = UserInsInst; 717 Depth++; 718 } 719 720 if (IsRedundant) 721 return replaceInstUsesWith(I, I.getOperand(0)); 722 return nullptr; 723 } 724 725 static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) { 726 int MaskSize = Shuf.getMask()->getType()->getVectorNumElements(); 727 int VecSize = Shuf.getOperand(0)->getType()->getVectorNumElements(); 728 729 // A vector select does not change the size of the operands. 730 if (MaskSize != VecSize) 731 return false; 732 733 // Each mask element must be undefined or choose a vector element from one of 734 // the source operands without crossing vector lanes. 735 for (int i = 0; i != MaskSize; ++i) { 736 int Elt = Shuf.getMaskValue(i); 737 if (Elt != -1 && Elt != i && Elt != i + VecSize) 738 return false; 739 } 740 741 return true; 742 } 743 744 /// Turn a chain of inserts that splats a value into an insert + shuffle: 745 /// insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ... -> 746 /// shufflevector(insertelt(X, %k, 0), undef, zero) 747 static Instruction *foldInsSequenceIntoSplat(InsertElementInst &InsElt) { 748 // We are interested in the last insert in a chain. So if this insert has a 749 // single user and that user is an insert, bail. 750 if (InsElt.hasOneUse() && isa<InsertElementInst>(InsElt.user_back())) 751 return nullptr; 752 753 auto *VecTy = cast<VectorType>(InsElt.getType()); 754 unsigned NumElements = VecTy->getNumElements(); 755 756 // Do not try to do this for a one-element vector, since that's a nop, 757 // and will cause an inf-loop. 758 if (NumElements == 1) 759 return nullptr; 760 761 Value *SplatVal = InsElt.getOperand(1); 762 InsertElementInst *CurrIE = &InsElt; 763 SmallVector<bool, 16> ElementPresent(NumElements, false); 764 InsertElementInst *FirstIE = nullptr; 765 766 // Walk the chain backwards, keeping track of which indices we inserted into, 767 // until we hit something that isn't an insert of the splatted value. 768 while (CurrIE) { 769 auto *Idx = dyn_cast<ConstantInt>(CurrIE->getOperand(2)); 770 if (!Idx || CurrIE->getOperand(1) != SplatVal) 771 return nullptr; 772 773 auto *NextIE = dyn_cast<InsertElementInst>(CurrIE->getOperand(0)); 774 // Check none of the intermediate steps have any additional uses, except 775 // for the root insertelement instruction, which can be re-used, if it 776 // inserts at position 0. 777 if (CurrIE != &InsElt && 778 (!CurrIE->hasOneUse() && (NextIE != nullptr || !Idx->isZero()))) 779 return nullptr; 780 781 ElementPresent[Idx->getZExtValue()] = true; 782 FirstIE = CurrIE; 783 CurrIE = NextIE; 784 } 785 786 // If this is just a single insertelement (not a sequence), we are done. 787 if (FirstIE == &InsElt) 788 return nullptr; 789 790 // If we are not inserting into an undef vector, make sure we've seen an 791 // insert into every element. 792 // TODO: If the base vector is not undef, it might be better to create a splat 793 // and then a select-shuffle (blend) with the base vector. 794 if (!isa<UndefValue>(FirstIE->getOperand(0))) 795 if (any_of(ElementPresent, [](bool Present) { return !Present; })) 796 return nullptr; 797 798 // Create the insert + shuffle. 799 Type *Int32Ty = Type::getInt32Ty(InsElt.getContext()); 800 UndefValue *UndefVec = UndefValue::get(VecTy); 801 Constant *Zero = ConstantInt::get(Int32Ty, 0); 802 if (!cast<ConstantInt>(FirstIE->getOperand(2))->isZero()) 803 FirstIE = InsertElementInst::Create(UndefVec, SplatVal, Zero, "", &InsElt); 804 805 // Splat from element 0, but replace absent elements with undef in the mask. 806 SmallVector<Constant *, 16> Mask(NumElements, Zero); 807 for (unsigned i = 0; i != NumElements; ++i) 808 if (!ElementPresent[i]) 809 Mask[i] = UndefValue::get(Int32Ty); 810 811 return new ShuffleVectorInst(FirstIE, UndefVec, ConstantVector::get(Mask)); 812 } 813 814 /// Try to fold an insert element into an existing splat shuffle by changing 815 /// the shuffle's mask to include the index of this insert element. 816 static Instruction *foldInsEltIntoSplat(InsertElementInst &InsElt) { 817 // Check if the vector operand of this insert is a canonical splat shuffle. 818 auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0)); 819 if (!Shuf || !Shuf->isZeroEltSplat()) 820 return nullptr; 821 822 // Check for a constant insertion index. 823 uint64_t IdxC; 824 if (!match(InsElt.getOperand(2), m_ConstantInt(IdxC))) 825 return nullptr; 826 827 // Check if the splat shuffle's input is the same as this insert's scalar op. 828 Value *X = InsElt.getOperand(1); 829 Value *Op0 = Shuf->getOperand(0); 830 if (!match(Op0, m_InsertElement(m_Undef(), m_Specific(X), m_ZeroInt()))) 831 return nullptr; 832 833 // Replace the shuffle mask element at the index of this insert with a zero. 834 // For example: 835 // inselt (shuf (inselt undef, X, 0), undef, <0,undef,0,undef>), X, 1 836 // --> shuf (inselt undef, X, 0), undef, <0,0,0,undef> 837 unsigned NumMaskElts = Shuf->getType()->getVectorNumElements(); 838 SmallVector<Constant *, 16> NewMaskVec(NumMaskElts); 839 Type *I32Ty = IntegerType::getInt32Ty(Shuf->getContext()); 840 Constant *Zero = ConstantInt::getNullValue(I32Ty); 841 for (unsigned i = 0; i != NumMaskElts; ++i) 842 NewMaskVec[i] = i == IdxC ? Zero : Shuf->getMask()->getAggregateElement(i); 843 844 Constant *NewMask = ConstantVector::get(NewMaskVec); 845 return new ShuffleVectorInst(Op0, UndefValue::get(Op0->getType()), NewMask); 846 } 847 848 /// Try to fold an extract+insert element into an existing identity shuffle by 849 /// changing the shuffle's mask to include the index of this insert element. 850 static Instruction *foldInsEltIntoIdentityShuffle(InsertElementInst &InsElt) { 851 // Check if the vector operand of this insert is an identity shuffle. 852 auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0)); 853 if (!Shuf || !isa<UndefValue>(Shuf->getOperand(1)) || 854 !(Shuf->isIdentityWithExtract() || Shuf->isIdentityWithPadding())) 855 return nullptr; 856 857 // Check for a constant insertion index. 858 uint64_t IdxC; 859 if (!match(InsElt.getOperand(2), m_ConstantInt(IdxC))) 860 return nullptr; 861 862 // Check if this insert's scalar op is extracted from the identity shuffle's 863 // input vector. 864 Value *Scalar = InsElt.getOperand(1); 865 Value *X = Shuf->getOperand(0); 866 if (!match(Scalar, m_ExtractElement(m_Specific(X), m_SpecificInt(IdxC)))) 867 return nullptr; 868 869 // Replace the shuffle mask element at the index of this extract+insert with 870 // that same index value. 871 // For example: 872 // inselt (shuf X, IdMask), (extelt X, IdxC), IdxC --> shuf X, IdMask' 873 unsigned NumMaskElts = Shuf->getType()->getVectorNumElements(); 874 SmallVector<Constant *, 16> NewMaskVec(NumMaskElts); 875 Type *I32Ty = IntegerType::getInt32Ty(Shuf->getContext()); 876 Constant *NewMaskEltC = ConstantInt::get(I32Ty, IdxC); 877 Constant *OldMask = Shuf->getMask(); 878 for (unsigned i = 0; i != NumMaskElts; ++i) { 879 if (i != IdxC) { 880 // All mask elements besides the inserted element remain the same. 881 NewMaskVec[i] = OldMask->getAggregateElement(i); 882 } else if (OldMask->getAggregateElement(i) == NewMaskEltC) { 883 // If the mask element was already set, there's nothing to do 884 // (demanded elements analysis may unset it later). 885 return nullptr; 886 } else { 887 assert(isa<UndefValue>(OldMask->getAggregateElement(i)) && 888 "Unexpected shuffle mask element for identity shuffle"); 889 NewMaskVec[i] = NewMaskEltC; 890 } 891 } 892 893 Constant *NewMask = ConstantVector::get(NewMaskVec); 894 return new ShuffleVectorInst(X, Shuf->getOperand(1), NewMask); 895 } 896 897 /// If we have an insertelement instruction feeding into another insertelement 898 /// and the 2nd is inserting a constant into the vector, canonicalize that 899 /// constant insertion before the insertion of a variable: 900 /// 901 /// insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 --> 902 /// insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1 903 /// 904 /// This has the potential of eliminating the 2nd insertelement instruction 905 /// via constant folding of the scalar constant into a vector constant. 906 static Instruction *hoistInsEltConst(InsertElementInst &InsElt2, 907 InstCombiner::BuilderTy &Builder) { 908 auto *InsElt1 = dyn_cast<InsertElementInst>(InsElt2.getOperand(0)); 909 if (!InsElt1 || !InsElt1->hasOneUse()) 910 return nullptr; 911 912 Value *X, *Y; 913 Constant *ScalarC; 914 ConstantInt *IdxC1, *IdxC2; 915 if (match(InsElt1->getOperand(0), m_Value(X)) && 916 match(InsElt1->getOperand(1), m_Value(Y)) && !isa<Constant>(Y) && 917 match(InsElt1->getOperand(2), m_ConstantInt(IdxC1)) && 918 match(InsElt2.getOperand(1), m_Constant(ScalarC)) && 919 match(InsElt2.getOperand(2), m_ConstantInt(IdxC2)) && IdxC1 != IdxC2) { 920 Value *NewInsElt1 = Builder.CreateInsertElement(X, ScalarC, IdxC2); 921 return InsertElementInst::Create(NewInsElt1, Y, IdxC1); 922 } 923 924 return nullptr; 925 } 926 927 /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex 928 /// --> shufflevector X, CVec', Mask' 929 static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) { 930 auto *Inst = dyn_cast<Instruction>(InsElt.getOperand(0)); 931 // Bail out if the parent has more than one use. In that case, we'd be 932 // replacing the insertelt with a shuffle, and that's not a clear win. 933 if (!Inst || !Inst->hasOneUse()) 934 return nullptr; 935 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0))) { 936 // The shuffle must have a constant vector operand. The insertelt must have 937 // a constant scalar being inserted at a constant position in the vector. 938 Constant *ShufConstVec, *InsEltScalar; 939 uint64_t InsEltIndex; 940 if (!match(Shuf->getOperand(1), m_Constant(ShufConstVec)) || 941 !match(InsElt.getOperand(1), m_Constant(InsEltScalar)) || 942 !match(InsElt.getOperand(2), m_ConstantInt(InsEltIndex))) 943 return nullptr; 944 945 // Adding an element to an arbitrary shuffle could be expensive, but a 946 // shuffle that selects elements from vectors without crossing lanes is 947 // assumed cheap. 948 // If we're just adding a constant into that shuffle, it will still be 949 // cheap. 950 if (!isShuffleEquivalentToSelect(*Shuf)) 951 return nullptr; 952 953 // From the above 'select' check, we know that the mask has the same number 954 // of elements as the vector input operands. We also know that each constant 955 // input element is used in its lane and can not be used more than once by 956 // the shuffle. Therefore, replace the constant in the shuffle's constant 957 // vector with the insertelt constant. Replace the constant in the shuffle's 958 // mask vector with the insertelt index plus the length of the vector 959 // (because the constant vector operand of a shuffle is always the 2nd 960 // operand). 961 Constant *Mask = Shuf->getMask(); 962 unsigned NumElts = Mask->getType()->getVectorNumElements(); 963 SmallVector<Constant *, 16> NewShufElts(NumElts); 964 SmallVector<Constant *, 16> NewMaskElts(NumElts); 965 for (unsigned I = 0; I != NumElts; ++I) { 966 if (I == InsEltIndex) { 967 NewShufElts[I] = InsEltScalar; 968 Type *Int32Ty = Type::getInt32Ty(Shuf->getContext()); 969 NewMaskElts[I] = ConstantInt::get(Int32Ty, InsEltIndex + NumElts); 970 } else { 971 // Copy over the existing values. 972 NewShufElts[I] = ShufConstVec->getAggregateElement(I); 973 NewMaskElts[I] = Mask->getAggregateElement(I); 974 } 975 } 976 977 // Create new operands for a shuffle that includes the constant of the 978 // original insertelt. The old shuffle will be dead now. 979 return new ShuffleVectorInst(Shuf->getOperand(0), 980 ConstantVector::get(NewShufElts), 981 ConstantVector::get(NewMaskElts)); 982 } else if (auto *IEI = dyn_cast<InsertElementInst>(Inst)) { 983 // Transform sequences of insertelements ops with constant data/indexes into 984 // a single shuffle op. 985 unsigned NumElts = InsElt.getType()->getNumElements(); 986 987 uint64_t InsertIdx[2]; 988 Constant *Val[2]; 989 if (!match(InsElt.getOperand(2), m_ConstantInt(InsertIdx[0])) || 990 !match(InsElt.getOperand(1), m_Constant(Val[0])) || 991 !match(IEI->getOperand(2), m_ConstantInt(InsertIdx[1])) || 992 !match(IEI->getOperand(1), m_Constant(Val[1]))) 993 return nullptr; 994 SmallVector<Constant *, 16> Values(NumElts); 995 SmallVector<Constant *, 16> Mask(NumElts); 996 auto ValI = std::begin(Val); 997 // Generate new constant vector and mask. 998 // We have 2 values/masks from the insertelements instructions. Insert them 999 // into new value/mask vectors. 1000 for (uint64_t I : InsertIdx) { 1001 if (!Values[I]) { 1002 assert(!Mask[I]); 1003 Values[I] = *ValI; 1004 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), 1005 NumElts + I); 1006 } 1007 ++ValI; 1008 } 1009 // Remaining values are filled with 'undef' values. 1010 for (unsigned I = 0; I < NumElts; ++I) { 1011 if (!Values[I]) { 1012 assert(!Mask[I]); 1013 Values[I] = UndefValue::get(InsElt.getType()->getElementType()); 1014 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), I); 1015 } 1016 } 1017 // Create new operands for a shuffle that includes the constant of the 1018 // original insertelt. 1019 return new ShuffleVectorInst(IEI->getOperand(0), 1020 ConstantVector::get(Values), 1021 ConstantVector::get(Mask)); 1022 } 1023 return nullptr; 1024 } 1025 1026 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) { 1027 Value *VecOp = IE.getOperand(0); 1028 Value *ScalarOp = IE.getOperand(1); 1029 Value *IdxOp = IE.getOperand(2); 1030 1031 if (auto *V = SimplifyInsertElementInst( 1032 VecOp, ScalarOp, IdxOp, SQ.getWithInstruction(&IE))) 1033 return replaceInstUsesWith(IE, V); 1034 1035 // If the vector and scalar are both bitcast from the same element type, do 1036 // the insert in that source type followed by bitcast. 1037 Value *VecSrc, *ScalarSrc; 1038 if (match(VecOp, m_BitCast(m_Value(VecSrc))) && 1039 match(ScalarOp, m_BitCast(m_Value(ScalarSrc))) && 1040 (VecOp->hasOneUse() || ScalarOp->hasOneUse()) && 1041 VecSrc->getType()->isVectorTy() && !ScalarSrc->getType()->isVectorTy() && 1042 VecSrc->getType()->getVectorElementType() == ScalarSrc->getType()) { 1043 // inselt (bitcast VecSrc), (bitcast ScalarSrc), IdxOp --> 1044 // bitcast (inselt VecSrc, ScalarSrc, IdxOp) 1045 Value *NewInsElt = Builder.CreateInsertElement(VecSrc, ScalarSrc, IdxOp); 1046 return new BitCastInst(NewInsElt, IE.getType()); 1047 } 1048 1049 // If the inserted element was extracted from some other vector and both 1050 // indexes are valid constants, try to turn this into a shuffle. 1051 uint64_t InsertedIdx, ExtractedIdx; 1052 Value *ExtVecOp; 1053 if (match(IdxOp, m_ConstantInt(InsertedIdx)) && 1054 match(ScalarOp, m_ExtractElement(m_Value(ExtVecOp), 1055 m_ConstantInt(ExtractedIdx))) && 1056 ExtractedIdx < ExtVecOp->getType()->getVectorNumElements()) { 1057 // TODO: Looking at the user(s) to determine if this insert is a 1058 // fold-to-shuffle opportunity does not match the usual instcombine 1059 // constraints. We should decide if the transform is worthy based only 1060 // on this instruction and its operands, but that may not work currently. 1061 // 1062 // Here, we are trying to avoid creating shuffles before reaching 1063 // the end of a chain of extract-insert pairs. This is complicated because 1064 // we do not generally form arbitrary shuffle masks in instcombine 1065 // (because those may codegen poorly), but collectShuffleElements() does 1066 // exactly that. 1067 // 1068 // The rules for determining what is an acceptable target-independent 1069 // shuffle mask are fuzzy because they evolve based on the backend's 1070 // capabilities and real-world impact. 1071 auto isShuffleRootCandidate = [](InsertElementInst &Insert) { 1072 if (!Insert.hasOneUse()) 1073 return true; 1074 auto *InsertUser = dyn_cast<InsertElementInst>(Insert.user_back()); 1075 if (!InsertUser) 1076 return true; 1077 return false; 1078 }; 1079 1080 // Try to form a shuffle from a chain of extract-insert ops. 1081 if (isShuffleRootCandidate(IE)) { 1082 SmallVector<Constant*, 16> Mask; 1083 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this); 1084 1085 // The proposed shuffle may be trivial, in which case we shouldn't 1086 // perform the combine. 1087 if (LR.first != &IE && LR.second != &IE) { 1088 // We now have a shuffle of LHS, RHS, Mask. 1089 if (LR.second == nullptr) 1090 LR.second = UndefValue::get(LR.first->getType()); 1091 return new ShuffleVectorInst(LR.first, LR.second, 1092 ConstantVector::get(Mask)); 1093 } 1094 } 1095 } 1096 1097 unsigned VWidth = VecOp->getType()->getVectorNumElements(); 1098 APInt UndefElts(VWidth, 0); 1099 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 1100 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) { 1101 if (V != &IE) 1102 return replaceInstUsesWith(IE, V); 1103 return &IE; 1104 } 1105 1106 if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE)) 1107 return Shuf; 1108 1109 if (Instruction *NewInsElt = hoistInsEltConst(IE, Builder)) 1110 return NewInsElt; 1111 1112 if (Instruction *Broadcast = foldInsSequenceIntoSplat(IE)) 1113 return Broadcast; 1114 1115 if (Instruction *Splat = foldInsEltIntoSplat(IE)) 1116 return Splat; 1117 1118 if (Instruction *IdentityShuf = foldInsEltIntoIdentityShuffle(IE)) 1119 return IdentityShuf; 1120 1121 return nullptr; 1122 } 1123 1124 /// Return true if we can evaluate the specified expression tree if the vector 1125 /// elements were shuffled in a different order. 1126 static bool canEvaluateShuffled(Value *V, ArrayRef<int> Mask, 1127 unsigned Depth = 5) { 1128 // We can always reorder the elements of a constant. 1129 if (isa<Constant>(V)) 1130 return true; 1131 1132 // We won't reorder vector arguments. No IPO here. 1133 Instruction *I = dyn_cast<Instruction>(V); 1134 if (!I) return false; 1135 1136 // Two users may expect different orders of the elements. Don't try it. 1137 if (!I->hasOneUse()) 1138 return false; 1139 1140 if (Depth == 0) return false; 1141 1142 switch (I->getOpcode()) { 1143 case Instruction::UDiv: 1144 case Instruction::SDiv: 1145 case Instruction::URem: 1146 case Instruction::SRem: 1147 // Propagating an undefined shuffle mask element to integer div/rem is not 1148 // allowed because those opcodes can create immediate undefined behavior 1149 // from an undefined element in an operand. 1150 if (llvm::any_of(Mask, [](int M){ return M == -1; })) 1151 return false; 1152 LLVM_FALLTHROUGH; 1153 case Instruction::Add: 1154 case Instruction::FAdd: 1155 case Instruction::Sub: 1156 case Instruction::FSub: 1157 case Instruction::Mul: 1158 case Instruction::FMul: 1159 case Instruction::FDiv: 1160 case Instruction::FRem: 1161 case Instruction::Shl: 1162 case Instruction::LShr: 1163 case Instruction::AShr: 1164 case Instruction::And: 1165 case Instruction::Or: 1166 case Instruction::Xor: 1167 case Instruction::ICmp: 1168 case Instruction::FCmp: 1169 case Instruction::Trunc: 1170 case Instruction::ZExt: 1171 case Instruction::SExt: 1172 case Instruction::FPToUI: 1173 case Instruction::FPToSI: 1174 case Instruction::UIToFP: 1175 case Instruction::SIToFP: 1176 case Instruction::FPTrunc: 1177 case Instruction::FPExt: 1178 case Instruction::GetElementPtr: { 1179 // Bail out if we would create longer vector ops. We could allow creating 1180 // longer vector ops, but that may result in more expensive codegen. 1181 Type *ITy = I->getType(); 1182 if (ITy->isVectorTy() && Mask.size() > ITy->getVectorNumElements()) 1183 return false; 1184 for (Value *Operand : I->operands()) { 1185 if (!canEvaluateShuffled(Operand, Mask, Depth - 1)) 1186 return false; 1187 } 1188 return true; 1189 } 1190 case Instruction::InsertElement: { 1191 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2)); 1192 if (!CI) return false; 1193 int ElementNumber = CI->getLimitedValue(); 1194 1195 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement' 1196 // can't put an element into multiple indices. 1197 bool SeenOnce = false; 1198 for (int i = 0, e = Mask.size(); i != e; ++i) { 1199 if (Mask[i] == ElementNumber) { 1200 if (SeenOnce) 1201 return false; 1202 SeenOnce = true; 1203 } 1204 } 1205 return canEvaluateShuffled(I->getOperand(0), Mask, Depth - 1); 1206 } 1207 } 1208 return false; 1209 } 1210 1211 /// Rebuild a new instruction just like 'I' but with the new operands given. 1212 /// In the event of type mismatch, the type of the operands is correct. 1213 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) { 1214 // We don't want to use the IRBuilder here because we want the replacement 1215 // instructions to appear next to 'I', not the builder's insertion point. 1216 switch (I->getOpcode()) { 1217 case Instruction::Add: 1218 case Instruction::FAdd: 1219 case Instruction::Sub: 1220 case Instruction::FSub: 1221 case Instruction::Mul: 1222 case Instruction::FMul: 1223 case Instruction::UDiv: 1224 case Instruction::SDiv: 1225 case Instruction::FDiv: 1226 case Instruction::URem: 1227 case Instruction::SRem: 1228 case Instruction::FRem: 1229 case Instruction::Shl: 1230 case Instruction::LShr: 1231 case Instruction::AShr: 1232 case Instruction::And: 1233 case Instruction::Or: 1234 case Instruction::Xor: { 1235 BinaryOperator *BO = cast<BinaryOperator>(I); 1236 assert(NewOps.size() == 2 && "binary operator with #ops != 2"); 1237 BinaryOperator *New = 1238 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(), 1239 NewOps[0], NewOps[1], "", BO); 1240 if (isa<OverflowingBinaryOperator>(BO)) { 1241 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap()); 1242 New->setHasNoSignedWrap(BO->hasNoSignedWrap()); 1243 } 1244 if (isa<PossiblyExactOperator>(BO)) { 1245 New->setIsExact(BO->isExact()); 1246 } 1247 if (isa<FPMathOperator>(BO)) 1248 New->copyFastMathFlags(I); 1249 return New; 1250 } 1251 case Instruction::ICmp: 1252 assert(NewOps.size() == 2 && "icmp with #ops != 2"); 1253 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(), 1254 NewOps[0], NewOps[1]); 1255 case Instruction::FCmp: 1256 assert(NewOps.size() == 2 && "fcmp with #ops != 2"); 1257 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(), 1258 NewOps[0], NewOps[1]); 1259 case Instruction::Trunc: 1260 case Instruction::ZExt: 1261 case Instruction::SExt: 1262 case Instruction::FPToUI: 1263 case Instruction::FPToSI: 1264 case Instruction::UIToFP: 1265 case Instruction::SIToFP: 1266 case Instruction::FPTrunc: 1267 case Instruction::FPExt: { 1268 // It's possible that the mask has a different number of elements from 1269 // the original cast. We recompute the destination type to match the mask. 1270 Type *DestTy = 1271 VectorType::get(I->getType()->getScalarType(), 1272 NewOps[0]->getType()->getVectorNumElements()); 1273 assert(NewOps.size() == 1 && "cast with #ops != 1"); 1274 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy, 1275 "", I); 1276 } 1277 case Instruction::GetElementPtr: { 1278 Value *Ptr = NewOps[0]; 1279 ArrayRef<Value*> Idx = NewOps.slice(1); 1280 GetElementPtrInst *GEP = GetElementPtrInst::Create( 1281 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I); 1282 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds()); 1283 return GEP; 1284 } 1285 } 1286 llvm_unreachable("failed to rebuild vector instructions"); 1287 } 1288 1289 static Value *evaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) { 1290 // Mask.size() does not need to be equal to the number of vector elements. 1291 1292 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements"); 1293 Type *EltTy = V->getType()->getScalarType(); 1294 Type *I32Ty = IntegerType::getInt32Ty(V->getContext()); 1295 if (isa<UndefValue>(V)) 1296 return UndefValue::get(VectorType::get(EltTy, Mask.size())); 1297 1298 if (isa<ConstantAggregateZero>(V)) 1299 return ConstantAggregateZero::get(VectorType::get(EltTy, Mask.size())); 1300 1301 if (Constant *C = dyn_cast<Constant>(V)) { 1302 SmallVector<Constant *, 16> MaskValues; 1303 for (int i = 0, e = Mask.size(); i != e; ++i) { 1304 if (Mask[i] == -1) 1305 MaskValues.push_back(UndefValue::get(I32Ty)); 1306 else 1307 MaskValues.push_back(ConstantInt::get(I32Ty, Mask[i])); 1308 } 1309 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()), 1310 ConstantVector::get(MaskValues)); 1311 } 1312 1313 Instruction *I = cast<Instruction>(V); 1314 switch (I->getOpcode()) { 1315 case Instruction::Add: 1316 case Instruction::FAdd: 1317 case Instruction::Sub: 1318 case Instruction::FSub: 1319 case Instruction::Mul: 1320 case Instruction::FMul: 1321 case Instruction::UDiv: 1322 case Instruction::SDiv: 1323 case Instruction::FDiv: 1324 case Instruction::URem: 1325 case Instruction::SRem: 1326 case Instruction::FRem: 1327 case Instruction::Shl: 1328 case Instruction::LShr: 1329 case Instruction::AShr: 1330 case Instruction::And: 1331 case Instruction::Or: 1332 case Instruction::Xor: 1333 case Instruction::ICmp: 1334 case Instruction::FCmp: 1335 case Instruction::Trunc: 1336 case Instruction::ZExt: 1337 case Instruction::SExt: 1338 case Instruction::FPToUI: 1339 case Instruction::FPToSI: 1340 case Instruction::UIToFP: 1341 case Instruction::SIToFP: 1342 case Instruction::FPTrunc: 1343 case Instruction::FPExt: 1344 case Instruction::Select: 1345 case Instruction::GetElementPtr: { 1346 SmallVector<Value*, 8> NewOps; 1347 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements()); 1348 for (int i = 0, e = I->getNumOperands(); i != e; ++i) { 1349 Value *V; 1350 // Recursively call evaluateInDifferentElementOrder on vector arguments 1351 // as well. E.g. GetElementPtr may have scalar operands even if the 1352 // return value is a vector, so we need to examine the operand type. 1353 if (I->getOperand(i)->getType()->isVectorTy()) 1354 V = evaluateInDifferentElementOrder(I->getOperand(i), Mask); 1355 else 1356 V = I->getOperand(i); 1357 NewOps.push_back(V); 1358 NeedsRebuild |= (V != I->getOperand(i)); 1359 } 1360 if (NeedsRebuild) { 1361 return buildNew(I, NewOps); 1362 } 1363 return I; 1364 } 1365 case Instruction::InsertElement: { 1366 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue(); 1367 1368 // The insertelement was inserting at Element. Figure out which element 1369 // that becomes after shuffling. The answer is guaranteed to be unique 1370 // by CanEvaluateShuffled. 1371 bool Found = false; 1372 int Index = 0; 1373 for (int e = Mask.size(); Index != e; ++Index) { 1374 if (Mask[Index] == Element) { 1375 Found = true; 1376 break; 1377 } 1378 } 1379 1380 // If element is not in Mask, no need to handle the operand 1 (element to 1381 // be inserted). Just evaluate values in operand 0 according to Mask. 1382 if (!Found) 1383 return evaluateInDifferentElementOrder(I->getOperand(0), Mask); 1384 1385 Value *V = evaluateInDifferentElementOrder(I->getOperand(0), Mask); 1386 return InsertElementInst::Create(V, I->getOperand(1), 1387 ConstantInt::get(I32Ty, Index), "", I); 1388 } 1389 } 1390 llvm_unreachable("failed to reorder elements of vector instruction!"); 1391 } 1392 1393 // Returns true if the shuffle is extracting a contiguous range of values from 1394 // LHS, for example: 1395 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ 1396 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP| 1397 // Shuffles to: |EE|FF|GG|HH| 1398 // +--+--+--+--+ 1399 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI, 1400 SmallVector<int, 16> &Mask) { 1401 unsigned LHSElems = SVI.getOperand(0)->getType()->getVectorNumElements(); 1402 unsigned MaskElems = Mask.size(); 1403 unsigned BegIdx = Mask.front(); 1404 unsigned EndIdx = Mask.back(); 1405 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1) 1406 return false; 1407 for (unsigned I = 0; I != MaskElems; ++I) 1408 if (static_cast<unsigned>(Mask[I]) != BegIdx + I) 1409 return false; 1410 return true; 1411 } 1412 1413 /// These are the ingredients in an alternate form binary operator as described 1414 /// below. 1415 struct BinopElts { 1416 BinaryOperator::BinaryOps Opcode; 1417 Value *Op0; 1418 Value *Op1; 1419 BinopElts(BinaryOperator::BinaryOps Opc = (BinaryOperator::BinaryOps)0, 1420 Value *V0 = nullptr, Value *V1 = nullptr) : 1421 Opcode(Opc), Op0(V0), Op1(V1) {} 1422 operator bool() const { return Opcode != 0; } 1423 }; 1424 1425 /// Binops may be transformed into binops with different opcodes and operands. 1426 /// Reverse the usual canonicalization to enable folds with the non-canonical 1427 /// form of the binop. If a transform is possible, return the elements of the 1428 /// new binop. If not, return invalid elements. 1429 static BinopElts getAlternateBinop(BinaryOperator *BO, const DataLayout &DL) { 1430 Value *BO0 = BO->getOperand(0), *BO1 = BO->getOperand(1); 1431 Type *Ty = BO->getType(); 1432 switch (BO->getOpcode()) { 1433 case Instruction::Shl: { 1434 // shl X, C --> mul X, (1 << C) 1435 Constant *C; 1436 if (match(BO1, m_Constant(C))) { 1437 Constant *ShlOne = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C); 1438 return { Instruction::Mul, BO0, ShlOne }; 1439 } 1440 break; 1441 } 1442 case Instruction::Or: { 1443 // or X, C --> add X, C (when X and C have no common bits set) 1444 const APInt *C; 1445 if (match(BO1, m_APInt(C)) && MaskedValueIsZero(BO0, *C, DL)) 1446 return { Instruction::Add, BO0, BO1 }; 1447 break; 1448 } 1449 default: 1450 break; 1451 } 1452 return {}; 1453 } 1454 1455 static Instruction *foldSelectShuffleWith1Binop(ShuffleVectorInst &Shuf) { 1456 assert(Shuf.isSelect() && "Must have select-equivalent shuffle"); 1457 1458 // Are we shuffling together some value and that same value after it has been 1459 // modified by a binop with a constant? 1460 Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1); 1461 Constant *C; 1462 bool Op0IsBinop; 1463 if (match(Op0, m_BinOp(m_Specific(Op1), m_Constant(C)))) 1464 Op0IsBinop = true; 1465 else if (match(Op1, m_BinOp(m_Specific(Op0), m_Constant(C)))) 1466 Op0IsBinop = false; 1467 else 1468 return nullptr; 1469 1470 // The identity constant for a binop leaves a variable operand unchanged. For 1471 // a vector, this is a splat of something like 0, -1, or 1. 1472 // If there's no identity constant for this binop, we're done. 1473 auto *BO = cast<BinaryOperator>(Op0IsBinop ? Op0 : Op1); 1474 BinaryOperator::BinaryOps BOpcode = BO->getOpcode(); 1475 Constant *IdC = ConstantExpr::getBinOpIdentity(BOpcode, Shuf.getType(), true); 1476 if (!IdC) 1477 return nullptr; 1478 1479 // Shuffle identity constants into the lanes that return the original value. 1480 // Example: shuf (mul X, {-1,-2,-3,-4}), X, {0,5,6,3} --> mul X, {-1,1,1,-4} 1481 // Example: shuf X, (add X, {-1,-2,-3,-4}), {0,1,6,7} --> add X, {0,0,-3,-4} 1482 // The existing binop constant vector remains in the same operand position. 1483 Constant *Mask = Shuf.getMask(); 1484 Constant *NewC = Op0IsBinop ? ConstantExpr::getShuffleVector(C, IdC, Mask) : 1485 ConstantExpr::getShuffleVector(IdC, C, Mask); 1486 1487 bool MightCreatePoisonOrUB = 1488 Mask->containsUndefElement() && 1489 (Instruction::isIntDivRem(BOpcode) || Instruction::isShift(BOpcode)); 1490 if (MightCreatePoisonOrUB) 1491 NewC = getSafeVectorConstantForBinop(BOpcode, NewC, true); 1492 1493 // shuf (bop X, C), X, M --> bop X, C' 1494 // shuf X, (bop X, C), M --> bop X, C' 1495 Value *X = Op0IsBinop ? Op1 : Op0; 1496 Instruction *NewBO = BinaryOperator::Create(BOpcode, X, NewC); 1497 NewBO->copyIRFlags(BO); 1498 1499 // An undef shuffle mask element may propagate as an undef constant element in 1500 // the new binop. That would produce poison where the original code might not. 1501 // If we already made a safe constant, then there's no danger. 1502 if (Mask->containsUndefElement() && !MightCreatePoisonOrUB) 1503 NewBO->dropPoisonGeneratingFlags(); 1504 return NewBO; 1505 } 1506 1507 /// If we have an insert of a scalar to a non-zero element of an undefined 1508 /// vector and then shuffle that value, that's the same as inserting to the zero 1509 /// element and shuffling. Splatting from the zero element is recognized as the 1510 /// canonical form of splat. 1511 static Instruction *canonicalizeInsertSplat(ShuffleVectorInst &Shuf, 1512 InstCombiner::BuilderTy &Builder) { 1513 Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1); 1514 Constant *Mask = Shuf.getMask(); 1515 Value *X; 1516 uint64_t IndexC; 1517 1518 // Match a shuffle that is a splat to a non-zero element. 1519 if (!match(Op0, m_OneUse(m_InsertElement(m_Undef(), m_Value(X), 1520 m_ConstantInt(IndexC)))) || 1521 !match(Op1, m_Undef()) || match(Mask, m_ZeroInt()) || IndexC == 0) 1522 return nullptr; 1523 1524 // Insert into element 0 of an undef vector. 1525 UndefValue *UndefVec = UndefValue::get(Shuf.getType()); 1526 Constant *Zero = Builder.getInt32(0); 1527 Value *NewIns = Builder.CreateInsertElement(UndefVec, X, Zero); 1528 1529 // Splat from element 0. Any mask element that is undefined remains undefined. 1530 // For example: 1531 // shuf (inselt undef, X, 2), undef, <2,2,undef> 1532 // --> shuf (inselt undef, X, 0), undef, <0,0,undef> 1533 unsigned NumMaskElts = Shuf.getType()->getVectorNumElements(); 1534 SmallVector<Constant *, 16> NewMask(NumMaskElts, Zero); 1535 for (unsigned i = 0; i != NumMaskElts; ++i) 1536 if (isa<UndefValue>(Mask->getAggregateElement(i))) 1537 NewMask[i] = Mask->getAggregateElement(i); 1538 1539 return new ShuffleVectorInst(NewIns, UndefVec, ConstantVector::get(NewMask)); 1540 } 1541 1542 /// Try to fold shuffles that are the equivalent of a vector select. 1543 static Instruction *foldSelectShuffle(ShuffleVectorInst &Shuf, 1544 InstCombiner::BuilderTy &Builder, 1545 const DataLayout &DL) { 1546 if (!Shuf.isSelect()) 1547 return nullptr; 1548 1549 // Canonicalize to choose from operand 0 first unless operand 1 is undefined. 1550 // Commuting undef to operand 0 conflicts with another canonicalization. 1551 unsigned NumElts = Shuf.getType()->getVectorNumElements(); 1552 if (!isa<UndefValue>(Shuf.getOperand(1)) && 1553 Shuf.getMaskValue(0) >= (int)NumElts) { 1554 // TODO: Can we assert that both operands of a shuffle-select are not undef 1555 // (otherwise, it would have been folded by instsimplify? 1556 Shuf.commute(); 1557 return &Shuf; 1558 } 1559 1560 if (Instruction *I = foldSelectShuffleWith1Binop(Shuf)) 1561 return I; 1562 1563 BinaryOperator *B0, *B1; 1564 if (!match(Shuf.getOperand(0), m_BinOp(B0)) || 1565 !match(Shuf.getOperand(1), m_BinOp(B1))) 1566 return nullptr; 1567 1568 Value *X, *Y; 1569 Constant *C0, *C1; 1570 bool ConstantsAreOp1; 1571 if (match(B0, m_BinOp(m_Value(X), m_Constant(C0))) && 1572 match(B1, m_BinOp(m_Value(Y), m_Constant(C1)))) 1573 ConstantsAreOp1 = true; 1574 else if (match(B0, m_BinOp(m_Constant(C0), m_Value(X))) && 1575 match(B1, m_BinOp(m_Constant(C1), m_Value(Y)))) 1576 ConstantsAreOp1 = false; 1577 else 1578 return nullptr; 1579 1580 // We need matching binops to fold the lanes together. 1581 BinaryOperator::BinaryOps Opc0 = B0->getOpcode(); 1582 BinaryOperator::BinaryOps Opc1 = B1->getOpcode(); 1583 bool DropNSW = false; 1584 if (ConstantsAreOp1 && Opc0 != Opc1) { 1585 // TODO: We drop "nsw" if shift is converted into multiply because it may 1586 // not be correct when the shift amount is BitWidth - 1. We could examine 1587 // each vector element to determine if it is safe to keep that flag. 1588 if (Opc0 == Instruction::Shl || Opc1 == Instruction::Shl) 1589 DropNSW = true; 1590 if (BinopElts AltB0 = getAlternateBinop(B0, DL)) { 1591 assert(isa<Constant>(AltB0.Op1) && "Expecting constant with alt binop"); 1592 Opc0 = AltB0.Opcode; 1593 C0 = cast<Constant>(AltB0.Op1); 1594 } else if (BinopElts AltB1 = getAlternateBinop(B1, DL)) { 1595 assert(isa<Constant>(AltB1.Op1) && "Expecting constant with alt binop"); 1596 Opc1 = AltB1.Opcode; 1597 C1 = cast<Constant>(AltB1.Op1); 1598 } 1599 } 1600 1601 if (Opc0 != Opc1) 1602 return nullptr; 1603 1604 // The opcodes must be the same. Use a new name to make that clear. 1605 BinaryOperator::BinaryOps BOpc = Opc0; 1606 1607 // Select the constant elements needed for the single binop. 1608 Constant *Mask = Shuf.getMask(); 1609 Constant *NewC = ConstantExpr::getShuffleVector(C0, C1, Mask); 1610 1611 // We are moving a binop after a shuffle. When a shuffle has an undefined 1612 // mask element, the result is undefined, but it is not poison or undefined 1613 // behavior. That is not necessarily true for div/rem/shift. 1614 bool MightCreatePoisonOrUB = 1615 Mask->containsUndefElement() && 1616 (Instruction::isIntDivRem(BOpc) || Instruction::isShift(BOpc)); 1617 if (MightCreatePoisonOrUB) 1618 NewC = getSafeVectorConstantForBinop(BOpc, NewC, ConstantsAreOp1); 1619 1620 Value *V; 1621 if (X == Y) { 1622 // Remove a binop and the shuffle by rearranging the constant: 1623 // shuffle (op V, C0), (op V, C1), M --> op V, C' 1624 // shuffle (op C0, V), (op C1, V), M --> op C', V 1625 V = X; 1626 } else { 1627 // If there are 2 different variable operands, we must create a new shuffle 1628 // (select) first, so check uses to ensure that we don't end up with more 1629 // instructions than we started with. 1630 if (!B0->hasOneUse() && !B1->hasOneUse()) 1631 return nullptr; 1632 1633 // If we use the original shuffle mask and op1 is *variable*, we would be 1634 // putting an undef into operand 1 of div/rem/shift. This is either UB or 1635 // poison. We do not have to guard against UB when *constants* are op1 1636 // because safe constants guarantee that we do not overflow sdiv/srem (and 1637 // there's no danger for other opcodes). 1638 // TODO: To allow this case, create a new shuffle mask with no undefs. 1639 if (MightCreatePoisonOrUB && !ConstantsAreOp1) 1640 return nullptr; 1641 1642 // Note: In general, we do not create new shuffles in InstCombine because we 1643 // do not know if a target can lower an arbitrary shuffle optimally. In this 1644 // case, the shuffle uses the existing mask, so there is no additional risk. 1645 1646 // Select the variable vectors first, then perform the binop: 1647 // shuffle (op X, C0), (op Y, C1), M --> op (shuffle X, Y, M), C' 1648 // shuffle (op C0, X), (op C1, Y), M --> op C', (shuffle X, Y, M) 1649 V = Builder.CreateShuffleVector(X, Y, Mask); 1650 } 1651 1652 Instruction *NewBO = ConstantsAreOp1 ? BinaryOperator::Create(BOpc, V, NewC) : 1653 BinaryOperator::Create(BOpc, NewC, V); 1654 1655 // Flags are intersected from the 2 source binops. But there are 2 exceptions: 1656 // 1. If we changed an opcode, poison conditions might have changed. 1657 // 2. If the shuffle had undef mask elements, the new binop might have undefs 1658 // where the original code did not. But if we already made a safe constant, 1659 // then there's no danger. 1660 NewBO->copyIRFlags(B0); 1661 NewBO->andIRFlags(B1); 1662 if (DropNSW) 1663 NewBO->setHasNoSignedWrap(false); 1664 if (Mask->containsUndefElement() && !MightCreatePoisonOrUB) 1665 NewBO->dropPoisonGeneratingFlags(); 1666 return NewBO; 1667 } 1668 1669 /// Match a shuffle-select-shuffle pattern where the shuffles are widening and 1670 /// narrowing (concatenating with undef and extracting back to the original 1671 /// length). This allows replacing the wide select with a narrow select. 1672 static Instruction *narrowVectorSelect(ShuffleVectorInst &Shuf, 1673 InstCombiner::BuilderTy &Builder) { 1674 // This must be a narrowing identity shuffle. It extracts the 1st N elements 1675 // of the 1st vector operand of a shuffle. 1676 if (!match(Shuf.getOperand(1), m_Undef()) || !Shuf.isIdentityWithExtract()) 1677 return nullptr; 1678 1679 // The vector being shuffled must be a vector select that we can eliminate. 1680 // TODO: The one-use requirement could be eased if X and/or Y are constants. 1681 Value *Cond, *X, *Y; 1682 if (!match(Shuf.getOperand(0), 1683 m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) 1684 return nullptr; 1685 1686 // We need a narrow condition value. It must be extended with undef elements 1687 // and have the same number of elements as this shuffle. 1688 unsigned NarrowNumElts = Shuf.getType()->getVectorNumElements(); 1689 Value *NarrowCond; 1690 if (!match(Cond, m_OneUse(m_ShuffleVector(m_Value(NarrowCond), m_Undef(), 1691 m_Constant()))) || 1692 NarrowCond->getType()->getVectorNumElements() != NarrowNumElts || 1693 !cast<ShuffleVectorInst>(Cond)->isIdentityWithPadding()) 1694 return nullptr; 1695 1696 // shuf (sel (shuf NarrowCond, undef, WideMask), X, Y), undef, NarrowMask) --> 1697 // sel NarrowCond, (shuf X, undef, NarrowMask), (shuf Y, undef, NarrowMask) 1698 Value *Undef = UndefValue::get(X->getType()); 1699 Value *NarrowX = Builder.CreateShuffleVector(X, Undef, Shuf.getMask()); 1700 Value *NarrowY = Builder.CreateShuffleVector(Y, Undef, Shuf.getMask()); 1701 return SelectInst::Create(NarrowCond, NarrowX, NarrowY); 1702 } 1703 1704 /// Try to combine 2 shuffles into 1 shuffle by concatenating a shuffle mask. 1705 static Instruction *foldIdentityExtractShuffle(ShuffleVectorInst &Shuf) { 1706 Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1); 1707 if (!Shuf.isIdentityWithExtract() || !isa<UndefValue>(Op1)) 1708 return nullptr; 1709 1710 Value *X, *Y; 1711 Constant *Mask; 1712 if (!match(Op0, m_ShuffleVector(m_Value(X), m_Value(Y), m_Constant(Mask)))) 1713 return nullptr; 1714 1715 // Be conservative with shuffle transforms. If we can't kill the 1st shuffle, 1716 // then combining may result in worse codegen. 1717 if (!Op0->hasOneUse()) 1718 return nullptr; 1719 1720 // We are extracting a subvector from a shuffle. Remove excess elements from 1721 // the 1st shuffle mask to eliminate the extract. 1722 // 1723 // This transform is conservatively limited to identity extracts because we do 1724 // not allow arbitrary shuffle mask creation as a target-independent transform 1725 // (because we can't guarantee that will lower efficiently). 1726 // 1727 // If the extracting shuffle has an undef mask element, it transfers to the 1728 // new shuffle mask. Otherwise, copy the original mask element. Example: 1729 // shuf (shuf X, Y, <C0, C1, C2, undef, C4>), undef, <0, undef, 2, 3> --> 1730 // shuf X, Y, <C0, undef, C2, undef> 1731 unsigned NumElts = Shuf.getType()->getVectorNumElements(); 1732 SmallVector<Constant *, 16> NewMask(NumElts); 1733 assert(NumElts < Mask->getType()->getVectorNumElements() && 1734 "Identity with extract must have less elements than its inputs"); 1735 1736 for (unsigned i = 0; i != NumElts; ++i) { 1737 Constant *ExtractMaskElt = Shuf.getMask()->getAggregateElement(i); 1738 Constant *MaskElt = Mask->getAggregateElement(i); 1739 NewMask[i] = isa<UndefValue>(ExtractMaskElt) ? ExtractMaskElt : MaskElt; 1740 } 1741 return new ShuffleVectorInst(X, Y, ConstantVector::get(NewMask)); 1742 } 1743 1744 /// Try to replace a shuffle with an insertelement or try to replace a shuffle 1745 /// operand with the operand of an insertelement. 1746 static Instruction *foldShuffleWithInsert(ShuffleVectorInst &Shuf) { 1747 Value *V0 = Shuf.getOperand(0), *V1 = Shuf.getOperand(1); 1748 SmallVector<int, 16> Mask = Shuf.getShuffleMask(); 1749 1750 // The shuffle must not change vector sizes. 1751 // TODO: This restriction could be removed if the insert has only one use 1752 // (because the transform would require a new length-changing shuffle). 1753 int NumElts = Mask.size(); 1754 if (NumElts != (int)(V0->getType()->getVectorNumElements())) 1755 return nullptr; 1756 1757 // This is a specialization of a fold in SimplifyDemandedVectorElts. We may 1758 // not be able to handle it there if the insertelement has >1 use. 1759 // If the shuffle has an insertelement operand but does not choose the 1760 // inserted scalar element from that value, then we can replace that shuffle 1761 // operand with the source vector of the insertelement. 1762 Value *X; 1763 uint64_t IdxC; 1764 if (match(V0, m_InsertElement(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) { 1765 // shuf (inselt X, ?, IdxC), ?, Mask --> shuf X, ?, Mask 1766 if (none_of(Mask, [IdxC](int MaskElt) { return MaskElt == (int)IdxC; })) { 1767 Shuf.setOperand(0, X); 1768 return &Shuf; 1769 } 1770 } 1771 if (match(V1, m_InsertElement(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) { 1772 // Offset the index constant by the vector width because we are checking for 1773 // accesses to the 2nd vector input of the shuffle. 1774 IdxC += NumElts; 1775 // shuf ?, (inselt X, ?, IdxC), Mask --> shuf ?, X, Mask 1776 if (none_of(Mask, [IdxC](int MaskElt) { return MaskElt == (int)IdxC; })) { 1777 Shuf.setOperand(1, X); 1778 return &Shuf; 1779 } 1780 } 1781 1782 // shuffle (insert ?, Scalar, IndexC), V1, Mask --> insert V1, Scalar, IndexC' 1783 auto isShufflingScalarIntoOp1 = [&](Value *&Scalar, ConstantInt *&IndexC) { 1784 // We need an insertelement with a constant index. 1785 if (!match(V0, m_InsertElement(m_Value(), m_Value(Scalar), 1786 m_ConstantInt(IndexC)))) 1787 return false; 1788 1789 // Test the shuffle mask to see if it splices the inserted scalar into the 1790 // operand 1 vector of the shuffle. 1791 int NewInsIndex = -1; 1792 for (int i = 0; i != NumElts; ++i) { 1793 // Ignore undef mask elements. 1794 if (Mask[i] == -1) 1795 continue; 1796 1797 // The shuffle takes elements of operand 1 without lane changes. 1798 if (Mask[i] == NumElts + i) 1799 continue; 1800 1801 // The shuffle must choose the inserted scalar exactly once. 1802 if (NewInsIndex != -1 || Mask[i] != IndexC->getSExtValue()) 1803 return false; 1804 1805 // The shuffle is placing the inserted scalar into element i. 1806 NewInsIndex = i; 1807 } 1808 1809 assert(NewInsIndex != -1 && "Did not fold shuffle with unused operand?"); 1810 1811 // Index is updated to the potentially translated insertion lane. 1812 IndexC = ConstantInt::get(IndexC->getType(), NewInsIndex); 1813 return true; 1814 }; 1815 1816 // If the shuffle is unnecessary, insert the scalar operand directly into 1817 // operand 1 of the shuffle. Example: 1818 // shuffle (insert ?, S, 1), V1, <1, 5, 6, 7> --> insert V1, S, 0 1819 Value *Scalar; 1820 ConstantInt *IndexC; 1821 if (isShufflingScalarIntoOp1(Scalar, IndexC)) 1822 return InsertElementInst::Create(V1, Scalar, IndexC); 1823 1824 // Try again after commuting shuffle. Example: 1825 // shuffle V0, (insert ?, S, 0), <0, 1, 2, 4> --> 1826 // shuffle (insert ?, S, 0), V0, <4, 5, 6, 0> --> insert V0, S, 3 1827 std::swap(V0, V1); 1828 ShuffleVectorInst::commuteShuffleMask(Mask, NumElts); 1829 if (isShufflingScalarIntoOp1(Scalar, IndexC)) 1830 return InsertElementInst::Create(V1, Scalar, IndexC); 1831 1832 return nullptr; 1833 } 1834 1835 static Instruction *foldIdentityPaddedShuffles(ShuffleVectorInst &Shuf) { 1836 // Match the operands as identity with padding (also known as concatenation 1837 // with undef) shuffles of the same source type. The backend is expected to 1838 // recreate these concatenations from a shuffle of narrow operands. 1839 auto *Shuffle0 = dyn_cast<ShuffleVectorInst>(Shuf.getOperand(0)); 1840 auto *Shuffle1 = dyn_cast<ShuffleVectorInst>(Shuf.getOperand(1)); 1841 if (!Shuffle0 || !Shuffle0->isIdentityWithPadding() || 1842 !Shuffle1 || !Shuffle1->isIdentityWithPadding()) 1843 return nullptr; 1844 1845 // We limit this transform to power-of-2 types because we expect that the 1846 // backend can convert the simplified IR patterns to identical nodes as the 1847 // original IR. 1848 // TODO: If we can verify the same behavior for arbitrary types, the 1849 // power-of-2 checks can be removed. 1850 Value *X = Shuffle0->getOperand(0); 1851 Value *Y = Shuffle1->getOperand(0); 1852 if (X->getType() != Y->getType() || 1853 !isPowerOf2_32(Shuf.getType()->getVectorNumElements()) || 1854 !isPowerOf2_32(Shuffle0->getType()->getVectorNumElements()) || 1855 !isPowerOf2_32(X->getType()->getVectorNumElements()) || 1856 isa<UndefValue>(X) || isa<UndefValue>(Y)) 1857 return nullptr; 1858 assert(isa<UndefValue>(Shuffle0->getOperand(1)) && 1859 isa<UndefValue>(Shuffle1->getOperand(1)) && 1860 "Unexpected operand for identity shuffle"); 1861 1862 // This is a shuffle of 2 widening shuffles. We can shuffle the narrow source 1863 // operands directly by adjusting the shuffle mask to account for the narrower 1864 // types: 1865 // shuf (widen X), (widen Y), Mask --> shuf X, Y, Mask' 1866 int NarrowElts = X->getType()->getVectorNumElements(); 1867 int WideElts = Shuffle0->getType()->getVectorNumElements(); 1868 assert(WideElts > NarrowElts && "Unexpected types for identity with padding"); 1869 1870 Type *I32Ty = IntegerType::getInt32Ty(Shuf.getContext()); 1871 SmallVector<int, 16> Mask = Shuf.getShuffleMask(); 1872 SmallVector<Constant *, 16> NewMask(Mask.size(), UndefValue::get(I32Ty)); 1873 for (int i = 0, e = Mask.size(); i != e; ++i) { 1874 if (Mask[i] == -1) 1875 continue; 1876 1877 // If this shuffle is choosing an undef element from 1 of the sources, that 1878 // element is undef. 1879 if (Mask[i] < WideElts) { 1880 if (Shuffle0->getMaskValue(Mask[i]) == -1) 1881 continue; 1882 } else { 1883 if (Shuffle1->getMaskValue(Mask[i] - WideElts) == -1) 1884 continue; 1885 } 1886 1887 // If this shuffle is choosing from the 1st narrow op, the mask element is 1888 // the same. If this shuffle is choosing from the 2nd narrow op, the mask 1889 // element is offset down to adjust for the narrow vector widths. 1890 if (Mask[i] < WideElts) { 1891 assert(Mask[i] < NarrowElts && "Unexpected shuffle mask"); 1892 NewMask[i] = ConstantInt::get(I32Ty, Mask[i]); 1893 } else { 1894 assert(Mask[i] < (WideElts + NarrowElts) && "Unexpected shuffle mask"); 1895 NewMask[i] = ConstantInt::get(I32Ty, Mask[i] - (WideElts - NarrowElts)); 1896 } 1897 } 1898 return new ShuffleVectorInst(X, Y, ConstantVector::get(NewMask)); 1899 } 1900 1901 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { 1902 Value *LHS = SVI.getOperand(0); 1903 Value *RHS = SVI.getOperand(1); 1904 if (auto *V = SimplifyShuffleVectorInst( 1905 LHS, RHS, SVI.getMask(), SVI.getType(), SQ.getWithInstruction(&SVI))) 1906 return replaceInstUsesWith(SVI, V); 1907 1908 // shuffle x, x, mask --> shuffle x, undef, mask' 1909 unsigned VWidth = SVI.getType()->getVectorNumElements(); 1910 unsigned LHSWidth = LHS->getType()->getVectorNumElements(); 1911 SmallVector<int, 16> Mask = SVI.getShuffleMask(); 1912 Type *Int32Ty = Type::getInt32Ty(SVI.getContext()); 1913 if (LHS == RHS) { 1914 assert(!isa<UndefValue>(RHS) && "Shuffle with 2 undef ops not simplified?"); 1915 // Remap any references to RHS to use LHS. 1916 SmallVector<Constant*, 16> Elts; 1917 for (unsigned i = 0; i != VWidth; ++i) { 1918 // Propagate undef elements or force mask to LHS. 1919 if (Mask[i] < 0) 1920 Elts.push_back(UndefValue::get(Int32Ty)); 1921 else 1922 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i] % LHSWidth)); 1923 } 1924 SVI.setOperand(0, SVI.getOperand(1)); 1925 SVI.setOperand(1, UndefValue::get(RHS->getType())); 1926 SVI.setOperand(2, ConstantVector::get(Elts)); 1927 return &SVI; 1928 } 1929 1930 // shuffle undef, x, mask --> shuffle x, undef, mask' 1931 if (isa<UndefValue>(LHS)) { 1932 SVI.commute(); 1933 return &SVI; 1934 } 1935 1936 if (Instruction *I = canonicalizeInsertSplat(SVI, Builder)) 1937 return I; 1938 1939 if (Instruction *I = foldSelectShuffle(SVI, Builder, DL)) 1940 return I; 1941 1942 if (Instruction *I = narrowVectorSelect(SVI, Builder)) 1943 return I; 1944 1945 APInt UndefElts(VWidth, 0); 1946 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 1947 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) { 1948 if (V != &SVI) 1949 return replaceInstUsesWith(SVI, V); 1950 return &SVI; 1951 } 1952 1953 if (Instruction *I = foldIdentityExtractShuffle(SVI)) 1954 return I; 1955 1956 // These transforms have the potential to lose undef knowledge, so they are 1957 // intentionally placed after SimplifyDemandedVectorElts(). 1958 if (Instruction *I = foldShuffleWithInsert(SVI)) 1959 return I; 1960 if (Instruction *I = foldIdentityPaddedShuffles(SVI)) 1961 return I; 1962 1963 if (isa<UndefValue>(RHS) && canEvaluateShuffled(LHS, Mask)) { 1964 Value *V = evaluateInDifferentElementOrder(LHS, Mask); 1965 return replaceInstUsesWith(SVI, V); 1966 } 1967 1968 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to 1969 // a non-vector type. We can instead bitcast the original vector followed by 1970 // an extract of the desired element: 1971 // 1972 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef, 1973 // <4 x i32> <i32 0, i32 1, i32 2, i32 3> 1974 // %1 = bitcast <4 x i8> %sroa to i32 1975 // Becomes: 1976 // %bc = bitcast <16 x i8> %in to <4 x i32> 1977 // %ext = extractelement <4 x i32> %bc, i32 0 1978 // 1979 // If the shuffle is extracting a contiguous range of values from the input 1980 // vector then each use which is a bitcast of the extracted size can be 1981 // replaced. This will work if the vector types are compatible, and the begin 1982 // index is aligned to a value in the casted vector type. If the begin index 1983 // isn't aligned then we can shuffle the original vector (keeping the same 1984 // vector type) before extracting. 1985 // 1986 // This code will bail out if the target type is fundamentally incompatible 1987 // with vectors of the source type. 1988 // 1989 // Example of <16 x i8>, target type i32: 1990 // Index range [4,8): v-----------v Will work. 1991 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ 1992 // <16 x i8>: | | | | | | | | | | | | | | | | | 1993 // <4 x i32>: | | | | | 1994 // +-----------+-----------+-----------+-----------+ 1995 // Index range [6,10): ^-----------^ Needs an extra shuffle. 1996 // Target type i40: ^--------------^ Won't work, bail. 1997 bool MadeChange = false; 1998 if (isShuffleExtractingFromLHS(SVI, Mask)) { 1999 Value *V = LHS; 2000 unsigned MaskElems = Mask.size(); 2001 VectorType *SrcTy = cast<VectorType>(V->getType()); 2002 unsigned VecBitWidth = SrcTy->getBitWidth(); 2003 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType()); 2004 assert(SrcElemBitWidth && "vector elements must have a bitwidth"); 2005 unsigned SrcNumElems = SrcTy->getNumElements(); 2006 SmallVector<BitCastInst *, 8> BCs; 2007 DenseMap<Type *, Value *> NewBCs; 2008 for (User *U : SVI.users()) 2009 if (BitCastInst *BC = dyn_cast<BitCastInst>(U)) 2010 if (!BC->use_empty()) 2011 // Only visit bitcasts that weren't previously handled. 2012 BCs.push_back(BC); 2013 for (BitCastInst *BC : BCs) { 2014 unsigned BegIdx = Mask.front(); 2015 Type *TgtTy = BC->getDestTy(); 2016 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy); 2017 if (!TgtElemBitWidth) 2018 continue; 2019 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth; 2020 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth; 2021 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth); 2022 if (!VecBitWidthsEqual) 2023 continue; 2024 if (!VectorType::isValidElementType(TgtTy)) 2025 continue; 2026 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems); 2027 if (!BegIsAligned) { 2028 // Shuffle the input so [0,NumElements) contains the output, and 2029 // [NumElems,SrcNumElems) is undef. 2030 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems, 2031 UndefValue::get(Int32Ty)); 2032 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I) 2033 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx); 2034 V = Builder.CreateShuffleVector(V, UndefValue::get(V->getType()), 2035 ConstantVector::get(ShuffleMask), 2036 SVI.getName() + ".extract"); 2037 BegIdx = 0; 2038 } 2039 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth; 2040 assert(SrcElemsPerTgtElem); 2041 BegIdx /= SrcElemsPerTgtElem; 2042 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end(); 2043 auto *NewBC = 2044 BCAlreadyExists 2045 ? NewBCs[CastSrcTy] 2046 : Builder.CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc"); 2047 if (!BCAlreadyExists) 2048 NewBCs[CastSrcTy] = NewBC; 2049 auto *Ext = Builder.CreateExtractElement( 2050 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract"); 2051 // The shufflevector isn't being replaced: the bitcast that used it 2052 // is. InstCombine will visit the newly-created instructions. 2053 replaceInstUsesWith(*BC, Ext); 2054 MadeChange = true; 2055 } 2056 } 2057 2058 // If the LHS is a shufflevector itself, see if we can combine it with this 2059 // one without producing an unusual shuffle. 2060 // Cases that might be simplified: 2061 // 1. 2062 // x1=shuffle(v1,v2,mask1) 2063 // x=shuffle(x1,undef,mask) 2064 // ==> 2065 // x=shuffle(v1,undef,newMask) 2066 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1 2067 // 2. 2068 // x1=shuffle(v1,undef,mask1) 2069 // x=shuffle(x1,x2,mask) 2070 // where v1.size() == mask1.size() 2071 // ==> 2072 // x=shuffle(v1,x2,newMask) 2073 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i] 2074 // 3. 2075 // x2=shuffle(v2,undef,mask2) 2076 // x=shuffle(x1,x2,mask) 2077 // where v2.size() == mask2.size() 2078 // ==> 2079 // x=shuffle(x1,v2,newMask) 2080 // newMask[i] = (mask[i] < x1.size()) 2081 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size() 2082 // 4. 2083 // x1=shuffle(v1,undef,mask1) 2084 // x2=shuffle(v2,undef,mask2) 2085 // x=shuffle(x1,x2,mask) 2086 // where v1.size() == v2.size() 2087 // ==> 2088 // x=shuffle(v1,v2,newMask) 2089 // newMask[i] = (mask[i] < x1.size()) 2090 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size() 2091 // 2092 // Here we are really conservative: 2093 // we are absolutely afraid of producing a shuffle mask not in the input 2094 // program, because the code gen may not be smart enough to turn a merged 2095 // shuffle into two specific shuffles: it may produce worse code. As such, 2096 // we only merge two shuffles if the result is either a splat or one of the 2097 // input shuffle masks. In this case, merging the shuffles just removes 2098 // one instruction, which we know is safe. This is good for things like 2099 // turning: (splat(splat)) -> splat, or 2100 // merge(V[0..n], V[n+1..2n]) -> V[0..2n] 2101 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS); 2102 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS); 2103 if (LHSShuffle) 2104 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS)) 2105 LHSShuffle = nullptr; 2106 if (RHSShuffle) 2107 if (!isa<UndefValue>(RHSShuffle->getOperand(1))) 2108 RHSShuffle = nullptr; 2109 if (!LHSShuffle && !RHSShuffle) 2110 return MadeChange ? &SVI : nullptr; 2111 2112 Value* LHSOp0 = nullptr; 2113 Value* LHSOp1 = nullptr; 2114 Value* RHSOp0 = nullptr; 2115 unsigned LHSOp0Width = 0; 2116 unsigned RHSOp0Width = 0; 2117 if (LHSShuffle) { 2118 LHSOp0 = LHSShuffle->getOperand(0); 2119 LHSOp1 = LHSShuffle->getOperand(1); 2120 LHSOp0Width = LHSOp0->getType()->getVectorNumElements(); 2121 } 2122 if (RHSShuffle) { 2123 RHSOp0 = RHSShuffle->getOperand(0); 2124 RHSOp0Width = RHSOp0->getType()->getVectorNumElements(); 2125 } 2126 Value* newLHS = LHS; 2127 Value* newRHS = RHS; 2128 if (LHSShuffle) { 2129 // case 1 2130 if (isa<UndefValue>(RHS)) { 2131 newLHS = LHSOp0; 2132 newRHS = LHSOp1; 2133 } 2134 // case 2 or 4 2135 else if (LHSOp0Width == LHSWidth) { 2136 newLHS = LHSOp0; 2137 } 2138 } 2139 // case 3 or 4 2140 if (RHSShuffle && RHSOp0Width == LHSWidth) { 2141 newRHS = RHSOp0; 2142 } 2143 // case 4 2144 if (LHSOp0 == RHSOp0) { 2145 newLHS = LHSOp0; 2146 newRHS = nullptr; 2147 } 2148 2149 if (newLHS == LHS && newRHS == RHS) 2150 return MadeChange ? &SVI : nullptr; 2151 2152 SmallVector<int, 16> LHSMask; 2153 SmallVector<int, 16> RHSMask; 2154 if (newLHS != LHS) 2155 LHSMask = LHSShuffle->getShuffleMask(); 2156 if (RHSShuffle && newRHS != RHS) 2157 RHSMask = RHSShuffle->getShuffleMask(); 2158 2159 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth; 2160 SmallVector<int, 16> newMask; 2161 bool isSplat = true; 2162 int SplatElt = -1; 2163 // Create a new mask for the new ShuffleVectorInst so that the new 2164 // ShuffleVectorInst is equivalent to the original one. 2165 for (unsigned i = 0; i < VWidth; ++i) { 2166 int eltMask; 2167 if (Mask[i] < 0) { 2168 // This element is an undef value. 2169 eltMask = -1; 2170 } else if (Mask[i] < (int)LHSWidth) { 2171 // This element is from left hand side vector operand. 2172 // 2173 // If LHS is going to be replaced (case 1, 2, or 4), calculate the 2174 // new mask value for the element. 2175 if (newLHS != LHS) { 2176 eltMask = LHSMask[Mask[i]]; 2177 // If the value selected is an undef value, explicitly specify it 2178 // with a -1 mask value. 2179 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1)) 2180 eltMask = -1; 2181 } else 2182 eltMask = Mask[i]; 2183 } else { 2184 // This element is from right hand side vector operand 2185 // 2186 // If the value selected is an undef value, explicitly specify it 2187 // with a -1 mask value. (case 1) 2188 if (isa<UndefValue>(RHS)) 2189 eltMask = -1; 2190 // If RHS is going to be replaced (case 3 or 4), calculate the 2191 // new mask value for the element. 2192 else if (newRHS != RHS) { 2193 eltMask = RHSMask[Mask[i]-LHSWidth]; 2194 // If the value selected is an undef value, explicitly specify it 2195 // with a -1 mask value. 2196 if (eltMask >= (int)RHSOp0Width) { 2197 assert(isa<UndefValue>(RHSShuffle->getOperand(1)) 2198 && "should have been check above"); 2199 eltMask = -1; 2200 } 2201 } else 2202 eltMask = Mask[i]-LHSWidth; 2203 2204 // If LHS's width is changed, shift the mask value accordingly. 2205 // If newRHS == nullptr, i.e. LHSOp0 == RHSOp0, we want to remap any 2206 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask. 2207 // If newRHS == newLHS, we want to remap any references from newRHS to 2208 // newLHS so that we can properly identify splats that may occur due to 2209 // obfuscation across the two vectors. 2210 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS) 2211 eltMask += newLHSWidth; 2212 } 2213 2214 // Check if this could still be a splat. 2215 if (eltMask >= 0) { 2216 if (SplatElt >= 0 && SplatElt != eltMask) 2217 isSplat = false; 2218 SplatElt = eltMask; 2219 } 2220 2221 newMask.push_back(eltMask); 2222 } 2223 2224 // If the result mask is equal to one of the original shuffle masks, 2225 // or is a splat, do the replacement. 2226 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) { 2227 SmallVector<Constant*, 16> Elts; 2228 for (unsigned i = 0, e = newMask.size(); i != e; ++i) { 2229 if (newMask[i] < 0) { 2230 Elts.push_back(UndefValue::get(Int32Ty)); 2231 } else { 2232 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i])); 2233 } 2234 } 2235 if (!newRHS) 2236 newRHS = UndefValue::get(newLHS->getType()); 2237 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts)); 2238 } 2239 2240 return MadeChange ? &SVI : nullptr; 2241 } 2242