| 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/SmallBitVector.h" |
| 20 | #include "llvm/ADT/SmallVector.h" |
| 21 | #include "llvm/ADT/Statistic.h" |
| 22 | #include "llvm/Analysis/InstructionSimplify.h" |
| 23 | #include "llvm/Analysis/VectorUtils.h" |
| 24 | #include "llvm/IR/BasicBlock.h" |
| 25 | #include "llvm/IR/Constant.h" |
| 26 | #include "llvm/IR/Constants.h" |
| 27 | #include "llvm/IR/DerivedTypes.h" |
| 28 | #include "llvm/IR/InstrTypes.h" |
| 29 | #include "llvm/IR/Instruction.h" |
| 30 | #include "llvm/IR/Instructions.h" |
| 31 | #include "llvm/IR/Operator.h" |
| 32 | #include "llvm/IR/PatternMatch.h" |
| 33 | #include "llvm/IR/Type.h" |
| 34 | #include "llvm/IR/User.h" |
| 35 | #include "llvm/IR/Value.h" |
| 36 | #include "llvm/Support/Casting.h" |
| 37 | #include "llvm/Support/ErrorHandling.h" |
| 38 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
| 39 | #include <cassert> |
| 40 | #include <cstdint> |
| 41 | #include <iterator> |
| 42 | #include <utility> |
| 43 | |
| 44 | #define DEBUG_TYPE "instcombine" |
| 45 | |
| 46 | using namespace llvm; |
| 47 | using namespace PatternMatch; |
| 48 | |
| 49 | STATISTIC(NumAggregateReconstructionsSimplified, |
| 50 | "Number of aggregate reconstructions turned into reuse of the " |
| 51 | "original aggregate" ); |
| 52 | |
| 53 | /// Return true if the value is cheaper to scalarize than it is to leave as a |
| 54 | /// vector operation. If the extract index \p EI is a constant integer then |
| 55 | /// some operations may be cheap to scalarize. |
| 56 | /// |
| 57 | /// FIXME: It's possible to create more instructions than previously existed. |
| 58 | static bool cheapToScalarize(Value *V, Value *EI) { |
| 59 | ConstantInt *CEI = dyn_cast<ConstantInt>(Val: EI); |
| 60 | |
| 61 | // If we can pick a scalar constant value out of a vector, that is free. |
| 62 | if (auto *C = dyn_cast<Constant>(Val: V)) |
| 63 | return CEI || C->getSplatValue(); |
| 64 | |
| 65 | if (CEI && match(V, P: m_Intrinsic<Intrinsic::stepvector>())) { |
| 66 | ElementCount EC = cast<VectorType>(Val: V->getType())->getElementCount(); |
| 67 | // Index needs to be lower than the minimum size of the vector, because |
| 68 | // for scalable vector, the vector size is known at run time. |
| 69 | return CEI->getValue().ult(RHS: EC.getKnownMinValue()); |
| 70 | } |
| 71 | |
| 72 | // An insertelement to the same constant index as our extract will simplify |
| 73 | // to the scalar inserted element. An insertelement to a different constant |
| 74 | // index is irrelevant to our extract. |
| 75 | if (match(V, P: m_InsertElt(Val: m_Value(), Elt: m_Value(), Idx: m_ConstantInt()))) |
| 76 | return CEI; |
| 77 | |
| 78 | if (match(V, P: m_OneUse(SubPattern: m_Load(Op: m_Value())))) |
| 79 | return true; |
| 80 | |
| 81 | if (match(V, P: m_OneUse(SubPattern: m_UnOp()))) |
| 82 | return true; |
| 83 | |
| 84 | Value *V0, *V1; |
| 85 | if (match(V, P: m_OneUse(SubPattern: m_BinOp(L: m_Value(V&: V0), R: m_Value(V&: V1))))) |
| 86 | if (cheapToScalarize(V: V0, EI) || cheapToScalarize(V: V1, EI)) |
| 87 | return true; |
| 88 | |
| 89 | CmpPredicate UnusedPred; |
| 90 | if (match(V, P: m_OneUse(SubPattern: m_Cmp(Pred&: UnusedPred, L: m_Value(V&: V0), R: m_Value(V&: V1))))) |
| 91 | if (cheapToScalarize(V: V0, EI) || cheapToScalarize(V: V1, EI)) |
| 92 | return true; |
| 93 | |
| 94 | return false; |
| 95 | } |
| 96 | |
| 97 | // If we have a PHI node with a vector type that is only used to feed |
| 98 | // itself and be an operand of extractelement at a constant location, |
| 99 | // try to replace the PHI of the vector type with a PHI of a scalar type. |
| 100 | Instruction *InstCombinerImpl::(ExtractElementInst &EI, |
| 101 | PHINode *PN) { |
| 102 | SmallVector<Instruction *, 2> ; |
| 103 | // The users we want the PHI to have are: |
| 104 | // 1) The EI ExtractElement (we already know this) |
| 105 | // 2) Possibly more ExtractElements with the same index. |
| 106 | // 3) Another operand, which will feed back into the PHI. |
| 107 | Instruction *PHIUser = nullptr; |
| 108 | for (auto *U : PN->users()) { |
| 109 | if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(Val: U)) { |
| 110 | if (EI.getIndexOperand() == EU->getIndexOperand()) |
| 111 | Extracts.push_back(Elt: EU); |
| 112 | else |
| 113 | return nullptr; |
| 114 | } else if (!PHIUser) { |
| 115 | PHIUser = cast<Instruction>(Val: U); |
| 116 | } else { |
| 117 | return nullptr; |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | if (!PHIUser) |
| 122 | return nullptr; |
| 123 | |
| 124 | // Verify that this PHI user has one use, which is the PHI itself, |
| 125 | // and that it is a binary operation which is cheap to scalarize. |
| 126 | // otherwise return nullptr. |
| 127 | if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) || |
| 128 | !(isa<BinaryOperator>(Val: PHIUser)) || |
| 129 | !cheapToScalarize(V: PHIUser, EI: EI.getIndexOperand())) |
| 130 | return nullptr; |
| 131 | |
| 132 | // Create a scalar PHI node that will replace the vector PHI node |
| 133 | // just before the current PHI node. |
| 134 | PHINode *scalarPHI = cast<PHINode>(Val: InsertNewInstWith( |
| 135 | New: PHINode::Create(Ty: EI.getType(), NumReservedValues: PN->getNumIncomingValues(), NameStr: "" ), Old: PN->getIterator())); |
| 136 | // Scalarize each PHI operand. A switch may produce multiple edges from the |
| 137 | // same predecessor; reuse the scalar instruction for duplicate edges. |
| 138 | SmallDenseMap<BasicBlock *, Value *, 4> ScalarizedValues; |
| 139 | for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) { |
| 140 | Value *PHIInVal = PN->getIncomingValue(i); |
| 141 | BasicBlock *inBB = PN->getIncomingBlock(i); |
| 142 | Value *Elt = EI.getIndexOperand(); |
| 143 | |
| 144 | // Reuse scalar value for duplicate edges from the same predecessor. |
| 145 | if (Value *Existing = ScalarizedValues.lookup(Val: inBB)) { |
| 146 | scalarPHI->addIncoming(V: Existing, BB: inBB); |
| 147 | continue; |
| 148 | } |
| 149 | |
| 150 | Value *ScalarVal; |
| 151 | // If the operand is the PHI induction variable: |
| 152 | if (PHIInVal == PHIUser) { |
| 153 | // Scalarize the binary operation. One operand is the |
| 154 | // scalar PHI, and the other is extracted from the other |
| 155 | // vector operand. |
| 156 | BinaryOperator *B0 = cast<BinaryOperator>(Val: PHIUser); |
| 157 | unsigned opId = (B0->getOperand(i_nocapture: 0) == PN) ? 1 : 0; |
| 158 | Value *Op = InsertNewInstWith( |
| 159 | New: ExtractElementInst::Create(Vec: B0->getOperand(i_nocapture: opId), Idx: Elt, |
| 160 | NameStr: B0->getOperand(i_nocapture: opId)->getName() + ".Elt" ), |
| 161 | Old: B0->getIterator()); |
| 162 | // Preserve operand order for binary operation to preserve semantics of |
| 163 | // non-commutative operations. |
| 164 | Value *FirstOp = (B0->getOperand(i_nocapture: 0) == PN) ? scalarPHI : Op; |
| 165 | Value *SecondOp = (B0->getOperand(i_nocapture: 0) == PN) ? Op : scalarPHI; |
| 166 | ScalarVal = InsertNewInstWith(New: BinaryOperator::CreateWithCopiedFlags( |
| 167 | Opc: B0->getOpcode(), V1: FirstOp, V2: SecondOp, CopyO: B0), |
| 168 | Old: B0->getIterator()); |
| 169 | } else { |
| 170 | // Scalarize PHI input: |
| 171 | Instruction *newEI = ExtractElementInst::Create(Vec: PHIInVal, Idx: Elt, NameStr: "" ); |
| 172 | // Insert the new instruction into the predecessor basic block. |
| 173 | Instruction *pos = dyn_cast<Instruction>(Val: PHIInVal); |
| 174 | BasicBlock::iterator InsertPos; |
| 175 | if (pos && !isa<PHINode>(Val: pos)) { |
| 176 | InsertPos = ++pos->getIterator(); |
| 177 | } else { |
| 178 | InsertPos = inBB->getFirstInsertionPt(); |
| 179 | } |
| 180 | |
| 181 | ScalarVal = InsertNewInstWith(New: newEI, Old: InsertPos); |
| 182 | } |
| 183 | |
| 184 | ScalarizedValues[inBB] = ScalarVal; |
| 185 | scalarPHI->addIncoming(V: ScalarVal, BB: inBB); |
| 186 | } |
| 187 | |
| 188 | for (auto *E : Extracts) { |
| 189 | replaceInstUsesWith(I&: *E, V: scalarPHI); |
| 190 | // Add old extract to worklist for DCE. |
| 191 | addToWorklist(I: E); |
| 192 | } |
| 193 | |
| 194 | return &EI; |
| 195 | } |
| 196 | |
| 197 | Instruction *InstCombinerImpl::(ExtractElementInst &Ext) { |
| 198 | Value *X; |
| 199 | uint64_t ExtIndexC; |
| 200 | if (!match(V: Ext.getVectorOperand(), P: m_BitCast(Op: m_Value(V&: X))) || |
| 201 | !match(V: Ext.getIndexOperand(), P: m_ConstantInt(V&: ExtIndexC))) |
| 202 | return nullptr; |
| 203 | |
| 204 | ElementCount NumElts = |
| 205 | cast<VectorType>(Val: Ext.getVectorOperandType())->getElementCount(); |
| 206 | Type *DestTy = Ext.getType(); |
| 207 | unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); |
| 208 | bool IsBigEndian = DL.isBigEndian(); |
| 209 | |
| 210 | // If we are casting an integer to vector and extracting a portion, that is |
| 211 | // a shift-right and truncate. |
| 212 | if (X->getType()->isIntegerTy()) { |
| 213 | assert(isa<FixedVectorType>(Ext.getVectorOperand()->getType()) && |
| 214 | "Expected fixed vector type for bitcast from scalar integer" ); |
| 215 | |
| 216 | // Big endian requires adjusting the extract index since MSB is at index 0. |
| 217 | // LittleEndian: extelt (bitcast i32 X to v4i8), 0 -> trunc i32 X to i8 |
| 218 | // BigEndian: extelt (bitcast i32 X to v4i8), 0 -> trunc i32 (X >> 24) to i8 |
| 219 | if (IsBigEndian) |
| 220 | ExtIndexC = NumElts.getKnownMinValue() - 1 - ExtIndexC; |
| 221 | unsigned ShiftAmountC = ExtIndexC * DestWidth; |
| 222 | if ((!ShiftAmountC || |
| 223 | isDesirableIntType(BitWidth: X->getType()->getPrimitiveSizeInBits())) && |
| 224 | Ext.getVectorOperand()->hasOneUse()) { |
| 225 | if (ShiftAmountC) |
| 226 | X = Builder.CreateLShr(LHS: X, RHS: ShiftAmountC, Name: "extelt.offset" ); |
| 227 | if (DestTy->isFloatingPointTy()) { |
| 228 | Type *DstIntTy = IntegerType::getIntNTy(C&: X->getContext(), N: DestWidth); |
| 229 | Value *Trunc = Builder.CreateTrunc(V: X, DestTy: DstIntTy); |
| 230 | return new BitCastInst(Trunc, DestTy); |
| 231 | } |
| 232 | return new TruncInst(X, DestTy); |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | if (!X->getType()->isVectorTy()) |
| 237 | return nullptr; |
| 238 | |
| 239 | // If this extractelement is using a bitcast from a vector of the same number |
| 240 | // of elements, see if we can find the source element from the source vector: |
| 241 | // extelt (bitcast VecX), IndexC --> bitcast X[IndexC] |
| 242 | auto *SrcTy = cast<VectorType>(Val: X->getType()); |
| 243 | ElementCount NumSrcElts = SrcTy->getElementCount(); |
| 244 | if (NumSrcElts == NumElts) |
| 245 | if (Value *Elt = findScalarElement(V: X, EltNo: ExtIndexC)) |
| 246 | return new BitCastInst(Elt, DestTy); |
| 247 | |
| 248 | assert(NumSrcElts.isScalable() == NumElts.isScalable() && |
| 249 | "Src and Dst must be the same sort of vector type" ); |
| 250 | |
| 251 | // If the source elements are wider than the destination, try to shift and |
| 252 | // truncate a subset of scalar bits of an insert op. |
| 253 | if (NumSrcElts.getKnownMinValue() < NumElts.getKnownMinValue()) { |
| 254 | Value *Scalar; |
| 255 | Value *Vec; |
| 256 | uint64_t InsIndexC; |
| 257 | if (!match(V: X, P: m_InsertElt(Val: m_Value(V&: Vec), Elt: m_Value(V&: Scalar), |
| 258 | Idx: m_ConstantInt(V&: InsIndexC)))) |
| 259 | return nullptr; |
| 260 | |
| 261 | // The extract must be from the subset of vector elements that we inserted |
| 262 | // into. Example: if we inserted element 1 of a <2 x i64> and we are |
| 263 | // extracting an i16 (narrowing ratio = 4), then this extract must be from 1 |
| 264 | // of elements 4-7 of the bitcasted vector. |
| 265 | unsigned NarrowingRatio = |
| 266 | NumElts.getKnownMinValue() / NumSrcElts.getKnownMinValue(); |
| 267 | |
| 268 | if (ExtIndexC / NarrowingRatio != InsIndexC) { |
| 269 | // Remove insertelement, if we don't use the inserted element. |
| 270 | // extractelement (bitcast (insertelement (Vec, b)), a) -> |
| 271 | // extractelement (bitcast (Vec), a) |
| 272 | // FIXME: this should be removed to SimplifyDemandedVectorElts, |
| 273 | // once scale vectors are supported. |
| 274 | if (X->hasOneUse() && Ext.getVectorOperand()->hasOneUse()) { |
| 275 | Value *NewBC = Builder.CreateBitCast(V: Vec, DestTy: Ext.getVectorOperandType()); |
| 276 | return ExtractElementInst::Create(Vec: NewBC, Idx: Ext.getIndexOperand()); |
| 277 | } |
| 278 | return nullptr; |
| 279 | } |
| 280 | |
| 281 | // We are extracting part of the original scalar. How that scalar is |
| 282 | // inserted into the vector depends on the endian-ness. Example: |
| 283 | // Vector Byte Elt Index: 0 1 2 3 4 5 6 7 |
| 284 | // +--+--+--+--+--+--+--+--+ |
| 285 | // inselt <2 x i32> V, <i32> S, 1: |V0|V1|V2|V3|S0|S1|S2|S3| |
| 286 | // extelt <4 x i16> V', 3: | |S2|S3| |
| 287 | // +--+--+--+--+--+--+--+--+ |
| 288 | // If this is little-endian, S2|S3 are the MSB of the 32-bit 'S' value. |
| 289 | // If this is big-endian, S2|S3 are the LSB of the 32-bit 'S' value. |
| 290 | // In this example, we must right-shift little-endian. Big-endian is just a |
| 291 | // truncate. |
| 292 | unsigned Chunk = ExtIndexC % NarrowingRatio; |
| 293 | if (IsBigEndian) |
| 294 | Chunk = NarrowingRatio - 1 - Chunk; |
| 295 | |
| 296 | // Bail out if this is an FP vector to FP vector sequence. That would take |
| 297 | // more instructions than we started with unless there is no shift, and it |
| 298 | // may not be handled as well in the backend. |
| 299 | bool NeedSrcBitcast = SrcTy->getScalarType()->isFloatingPointTy(); |
| 300 | bool NeedDestBitcast = DestTy->isFloatingPointTy(); |
| 301 | if (NeedSrcBitcast && NeedDestBitcast) |
| 302 | return nullptr; |
| 303 | |
| 304 | unsigned SrcWidth = SrcTy->getScalarSizeInBits(); |
| 305 | unsigned ShAmt = Chunk * DestWidth; |
| 306 | |
| 307 | // TODO: This limitation is more strict than necessary. We could sum the |
| 308 | // number of new instructions and subtract the number eliminated to know if |
| 309 | // we can proceed. |
| 310 | if (!X->hasOneUse() || !Ext.getVectorOperand()->hasOneUse()) |
| 311 | if (NeedSrcBitcast || NeedDestBitcast) |
| 312 | return nullptr; |
| 313 | |
| 314 | if (NeedSrcBitcast) { |
| 315 | Type *SrcIntTy = IntegerType::getIntNTy(C&: Scalar->getContext(), N: SrcWidth); |
| 316 | Scalar = Builder.CreateBitCast(V: Scalar, DestTy: SrcIntTy); |
| 317 | } |
| 318 | |
| 319 | if (ShAmt) { |
| 320 | // Bail out if we could end with more instructions than we started with. |
| 321 | if (!Ext.getVectorOperand()->hasOneUse()) |
| 322 | return nullptr; |
| 323 | Scalar = Builder.CreateLShr(LHS: Scalar, RHS: ShAmt); |
| 324 | } |
| 325 | |
| 326 | if (NeedDestBitcast) { |
| 327 | Type *DestIntTy = IntegerType::getIntNTy(C&: Scalar->getContext(), N: DestWidth); |
| 328 | return new BitCastInst(Builder.CreateTrunc(V: Scalar, DestTy: DestIntTy), DestTy); |
| 329 | } |
| 330 | return new TruncInst(Scalar, DestTy); |
| 331 | } |
| 332 | |
| 333 | return nullptr; |
| 334 | } |
| 335 | |
| 336 | /// Find elements of V demanded by UserInstr. If returns false, we were not able |
| 337 | /// to determine all elements. |
| 338 | static bool findDemandedEltsBySingleUser(Value *V, Instruction *UserInstr, |
| 339 | APInt &UnionUsedElts) { |
| 340 | unsigned VWidth = cast<FixedVectorType>(Val: V->getType())->getNumElements(); |
| 341 | |
| 342 | switch (UserInstr->getOpcode()) { |
| 343 | case Instruction::ExtractElement: { |
| 344 | ExtractElementInst *EEI = cast<ExtractElementInst>(Val: UserInstr); |
| 345 | assert(EEI->getVectorOperand() == V); |
| 346 | ConstantInt *EEIIndexC = dyn_cast<ConstantInt>(Val: EEI->getIndexOperand()); |
| 347 | if (EEIIndexC && EEIIndexC->getValue().ult(RHS: VWidth)) { |
| 348 | UnionUsedElts.setBit(EEIIndexC->getZExtValue()); |
| 349 | return true; |
| 350 | } |
| 351 | break; |
| 352 | } |
| 353 | case Instruction::ShuffleVector: { |
| 354 | ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(Val: UserInstr); |
| 355 | unsigned MaskNumElts = |
| 356 | cast<FixedVectorType>(Val: UserInstr->getType())->getNumElements(); |
| 357 | |
| 358 | for (auto I : llvm::seq(Size: MaskNumElts)) { |
| 359 | unsigned MaskVal = Shuffle->getMaskValue(Elt: I); |
| 360 | if (MaskVal == -1u || MaskVal >= 2 * VWidth) |
| 361 | continue; |
| 362 | if (Shuffle->getOperand(i_nocapture: 0) == V && (MaskVal < VWidth)) |
| 363 | UnionUsedElts.setBit(MaskVal); |
| 364 | if (Shuffle->getOperand(i_nocapture: 1) == V && |
| 365 | ((MaskVal >= VWidth) && (MaskVal < 2 * VWidth))) |
| 366 | UnionUsedElts.setBit(MaskVal - VWidth); |
| 367 | } |
| 368 | return true; |
| 369 | } |
| 370 | default: |
| 371 | break; |
| 372 | } |
| 373 | |
| 374 | return false; |
| 375 | } |
| 376 | |
| 377 | /// Find union of elements of V demanded by all its users. |
| 378 | /// If it is known by querying findDemandedEltsBySingleUser that |
| 379 | /// no user demands an element of V, then the corresponding bit |
| 380 | /// remains unset in the returned value. |
| 381 | static APInt findDemandedEltsByAllUsers(Value *V) { |
| 382 | unsigned VWidth = cast<FixedVectorType>(Val: V->getType())->getNumElements(); |
| 383 | |
| 384 | APInt UnionUsedElts(VWidth, 0); |
| 385 | for (const Use &U : V->uses()) { |
| 386 | if (Instruction *I = dyn_cast<Instruction>(Val: U.getUser())) { |
| 387 | if (!findDemandedEltsBySingleUser(V, UserInstr: I, UnionUsedElts)) |
| 388 | return APInt::getAllOnes(numBits: VWidth); |
| 389 | } else { |
| 390 | UnionUsedElts = APInt::getAllOnes(numBits: VWidth); |
| 391 | break; |
| 392 | } |
| 393 | |
| 394 | if (UnionUsedElts.isAllOnes()) |
| 395 | break; |
| 396 | } |
| 397 | |
| 398 | return UnionUsedElts; |
| 399 | } |
| 400 | |
| 401 | /// Given a constant index for a extractelement or insertelement instruction, |
| 402 | /// return it with the canonical type if it isn't already canonical. We |
| 403 | /// arbitrarily pick 64 bit as our canonical type. The actual bitwidth doesn't |
| 404 | /// matter, we just want a consistent type to simplify CSE. |
| 405 | static ConstantInt *getPreferredVectorIndex(ConstantInt *IndexC) { |
| 406 | const unsigned IndexBW = IndexC->getBitWidth(); |
| 407 | if (IndexBW == 64 || IndexC->getValue().getActiveBits() > 64) |
| 408 | return nullptr; |
| 409 | return ConstantInt::get(Context&: IndexC->getContext(), |
| 410 | V: IndexC->getValue().zextOrTrunc(width: 64)); |
| 411 | } |
| 412 | |
| 413 | Instruction *InstCombinerImpl::(ExtractElementInst &EI) { |
| 414 | Value *SrcVec = EI.getVectorOperand(); |
| 415 | Value *Index = EI.getIndexOperand(); |
| 416 | if (Value *V = simplifyExtractElementInst(Vec: SrcVec, Idx: Index, |
| 417 | Q: SQ.getWithInstruction(I: &EI))) |
| 418 | return replaceInstUsesWith(I&: EI, V); |
| 419 | |
| 420 | // extractelt (select %x, %vec1, %vec2), %const -> |
| 421 | // select %x, %vec1[%const], %vec2[%const] |
| 422 | // TODO: Support constant folding of multiple select operands: |
| 423 | // extractelt (select %x, %vec1, %vec2), (select %x, %c1, %c2) |
| 424 | // If the extractelement will for instance try to do out of bounds accesses |
| 425 | // because of the values of %c1 and/or %c2, the sequence could be optimized |
| 426 | // early. This is currently not possible because constant folding will reach |
| 427 | // an unreachable assertion if it doesn't find a constant operand. |
| 428 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: EI.getVectorOperand())) |
| 429 | if (SI->getCondition()->getType()->isIntegerTy() && |
| 430 | isa<Constant>(Val: EI.getIndexOperand())) |
| 431 | if (Instruction *R = FoldOpIntoSelect(Op&: EI, SI)) |
| 432 | return R; |
| 433 | |
| 434 | // If extracting a specified index from the vector, see if we can recursively |
| 435 | // find a previously computed scalar that was inserted into the vector. |
| 436 | auto *IndexC = dyn_cast<ConstantInt>(Val: Index); |
| 437 | bool HasKnownValidIndex = false; |
| 438 | if (IndexC) { |
| 439 | // Canonicalize type of constant indices to i64 to simplify CSE |
| 440 | if (auto *NewIdx = getPreferredVectorIndex(IndexC)) |
| 441 | return replaceOperand(I&: EI, OpNum: 1, V: NewIdx); |
| 442 | |
| 443 | ElementCount EC = EI.getVectorOperandType()->getElementCount(); |
| 444 | unsigned NumElts = EC.getKnownMinValue(); |
| 445 | HasKnownValidIndex = IndexC->getValue().ult(RHS: NumElts); |
| 446 | |
| 447 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: SrcVec)) { |
| 448 | Intrinsic::ID IID = II->getIntrinsicID(); |
| 449 | // Index needs to be lower than the minimum size of the vector, because |
| 450 | // for scalable vector, the vector size is known at run time. |
| 451 | if (IID == Intrinsic::stepvector && IndexC->getValue().ult(RHS: NumElts)) { |
| 452 | Type *Ty = EI.getType(); |
| 453 | unsigned BitWidth = Ty->getIntegerBitWidth(); |
| 454 | Value *Idx; |
| 455 | // Return index when its value does not exceed the allowed limit |
| 456 | // for the element type of the vector. |
| 457 | // TODO: Truncate out-of-range values. |
| 458 | if (IndexC->getValue().getActiveBits() <= BitWidth) |
| 459 | Idx = ConstantInt::get(Ty, V: IndexC->getValue().zextOrTrunc(width: BitWidth)); |
| 460 | else |
| 461 | return nullptr; |
| 462 | return replaceInstUsesWith(I&: EI, V: Idx); |
| 463 | } |
| 464 | } |
| 465 | |
| 466 | // InstSimplify should handle cases where the index is invalid. |
| 467 | // For fixed-length vector, it's invalid to extract out-of-range element. |
| 468 | if (!EC.isScalable() && IndexC->getValue().uge(RHS: NumElts)) |
| 469 | return nullptr; |
| 470 | |
| 471 | if (Instruction *I = foldBitcastExtElt(Ext&: EI)) |
| 472 | return I; |
| 473 | |
| 474 | // If there's a vector PHI feeding a scalar use through this extractelement |
| 475 | // instruction, try to scalarize the PHI. |
| 476 | if (auto *Phi = dyn_cast<PHINode>(Val: SrcVec)) |
| 477 | if (Instruction *ScalarPHI = scalarizePHI(EI, PN: Phi)) |
| 478 | return ScalarPHI; |
| 479 | } |
| 480 | |
| 481 | // If SrcVec is a subvector starting at index 0, extract from the |
| 482 | // wider source vector |
| 483 | Value *V; |
| 484 | if (match(V: SrcVec, |
| 485 | P: m_Intrinsic<Intrinsic::vector_extract>(Ops: m_Value(V), Ops: m_Zero()))) |
| 486 | return ExtractElementInst::Create(Vec: V, Idx: Index); |
| 487 | |
| 488 | // TODO come up with a n-ary matcher that subsumes both unary and |
| 489 | // binary matchers. |
| 490 | UnaryOperator *UO; |
| 491 | if (match(V: SrcVec, P: m_UnOp(I&: UO)) && cheapToScalarize(V: SrcVec, EI: Index)) { |
| 492 | // extelt (unop X), Index --> unop (extelt X, Index) |
| 493 | Value *X = UO->getOperand(i_nocapture: 0); |
| 494 | Value *E = Builder.CreateExtractElement(Vec: X, Idx: Index); |
| 495 | return UnaryOperator::CreateWithCopiedFlags(Opc: UO->getOpcode(), V: E, CopyO: UO); |
| 496 | } |
| 497 | |
| 498 | // If the binop is not speculatable, we cannot hoist the extractelement if |
| 499 | // it may make the operand poison. |
| 500 | BinaryOperator *BO; |
| 501 | if (match(V: SrcVec, P: m_BinOp(I&: BO)) && cheapToScalarize(V: SrcVec, EI: Index) && |
| 502 | (HasKnownValidIndex || |
| 503 | isSafeToSpeculativelyExecuteWithVariableReplaced(I: BO))) { |
| 504 | // extelt (binop X, Y), Index --> binop (extelt X, Index), (extelt Y, Index) |
| 505 | Value *X = BO->getOperand(i_nocapture: 0), *Y = BO->getOperand(i_nocapture: 1); |
| 506 | Value *E0 = Builder.CreateExtractElement(Vec: X, Idx: Index); |
| 507 | Value *E1 = Builder.CreateExtractElement(Vec: Y, Idx: Index); |
| 508 | return BinaryOperator::CreateWithCopiedFlags(Opc: BO->getOpcode(), V1: E0, V2: E1, CopyO: BO); |
| 509 | } |
| 510 | |
| 511 | Value *X, *Y; |
| 512 | CmpPredicate Pred; |
| 513 | if (match(V: SrcVec, P: m_Cmp(Pred, L: m_Value(V&: X), R: m_Value(V&: Y))) && |
| 514 | cheapToScalarize(V: SrcVec, EI: Index)) { |
| 515 | // extelt (cmp X, Y), Index --> cmp (extelt X, Index), (extelt Y, Index) |
| 516 | Value *E0 = Builder.CreateExtractElement(Vec: X, Idx: Index); |
| 517 | Value *E1 = Builder.CreateExtractElement(Vec: Y, Idx: Index); |
| 518 | CmpInst *SrcCmpInst = cast<CmpInst>(Val: SrcVec); |
| 519 | return CmpInst::CreateWithCopiedFlags(Op: SrcCmpInst->getOpcode(), Pred, S1: E0, S2: E1, |
| 520 | FlagsSource: SrcCmpInst); |
| 521 | } |
| 522 | |
| 523 | if (auto *I = dyn_cast<Instruction>(Val: SrcVec)) { |
| 524 | if (auto *IE = dyn_cast<InsertElementInst>(Val: I)) { |
| 525 | // instsimplify already handled the case where the indices are constants |
| 526 | // and equal by value, if both are constants, they must not be the same |
| 527 | // value, extract from the pre-inserted value instead. |
| 528 | if (isa<Constant>(Val: IE->getOperand(i_nocapture: 2)) && IndexC) |
| 529 | return replaceOperand(I&: EI, OpNum: 0, V: IE->getOperand(i_nocapture: 0)); |
| 530 | } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: I)) { |
| 531 | auto *VecType = cast<VectorType>(Val: GEP->getType()); |
| 532 | ElementCount EC = VecType->getElementCount(); |
| 533 | uint64_t IdxVal = IndexC ? IndexC->getZExtValue() : 0; |
| 534 | if (IndexC && IdxVal < EC.getKnownMinValue() && GEP->hasOneUse()) { |
| 535 | // Find out why we have a vector result - these are a few examples: |
| 536 | // 1. We have a scalar pointer and a vector of indices, or |
| 537 | // 2. We have a vector of pointers and a scalar index, or |
| 538 | // 3. We have a vector of pointers and a vector of indices, etc. |
| 539 | // Here we only consider combining when there is exactly one vector |
| 540 | // operand, since the optimization is less obviously a win due to |
| 541 | // needing more than one extractelements. |
| 542 | |
| 543 | unsigned VectorOps = |
| 544 | llvm::count_if(Range: GEP->operands(), P: [](const Value *V) { |
| 545 | return isa<VectorType>(Val: V->getType()); |
| 546 | }); |
| 547 | if (VectorOps == 1) { |
| 548 | Value *NewPtr = GEP->getPointerOperand(); |
| 549 | if (isa<VectorType>(Val: NewPtr->getType())) |
| 550 | NewPtr = Builder.CreateExtractElement(Vec: NewPtr, Idx: IndexC); |
| 551 | |
| 552 | SmallVector<Value *> NewOps; |
| 553 | for (unsigned I = 1; I != GEP->getNumOperands(); ++I) { |
| 554 | Value *Op = GEP->getOperand(i_nocapture: I); |
| 555 | if (isa<VectorType>(Val: Op->getType())) |
| 556 | NewOps.push_back(Elt: Builder.CreateExtractElement(Vec: Op, Idx: IndexC)); |
| 557 | else |
| 558 | NewOps.push_back(Elt: Op); |
| 559 | } |
| 560 | |
| 561 | GetElementPtrInst *NewGEP = GetElementPtrInst::Create( |
| 562 | PointeeType: GEP->getSourceElementType(), Ptr: NewPtr, IdxList: NewOps); |
| 563 | NewGEP->setNoWrapFlags(GEP->getNoWrapFlags()); |
| 564 | return NewGEP; |
| 565 | } |
| 566 | } |
| 567 | } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(Val: I)) { |
| 568 | int SplatIndex = getSplatIndex(Mask: SVI->getShuffleMask()); |
| 569 | // We know the all-0 splat must be reading from the first operand, even |
| 570 | // in the case of scalable vectors (vscale is always > 0). |
| 571 | if (SplatIndex == 0) |
| 572 | return ExtractElementInst::Create(Vec: SVI->getOperand(i_nocapture: 0), |
| 573 | Idx: Builder.getInt64(C: 0)); |
| 574 | |
| 575 | if (isa<FixedVectorType>(Val: SVI->getType())) { |
| 576 | std::optional<int> SrcIdx; |
| 577 | // getSplatIndex returns -1 to mean not-found. |
| 578 | if (SplatIndex != -1) |
| 579 | SrcIdx = SplatIndex; |
| 580 | else if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Index)) |
| 581 | SrcIdx = SVI->getMaskValue(Elt: CI->getZExtValue()); |
| 582 | |
| 583 | if (SrcIdx) { |
| 584 | Value *Src; |
| 585 | unsigned LHSWidth = |
| 586 | cast<FixedVectorType>(Val: SVI->getOperand(i_nocapture: 0)->getType()) |
| 587 | ->getNumElements(); |
| 588 | |
| 589 | if (*SrcIdx < 0) |
| 590 | return replaceInstUsesWith(I&: EI, V: PoisonValue::get(T: EI.getType())); |
| 591 | if (*SrcIdx < (int)LHSWidth) |
| 592 | Src = SVI->getOperand(i_nocapture: 0); |
| 593 | else { |
| 594 | *SrcIdx -= LHSWidth; |
| 595 | Src = SVI->getOperand(i_nocapture: 1); |
| 596 | } |
| 597 | Type *Int64Ty = Type::getInt64Ty(C&: EI.getContext()); |
| 598 | return ExtractElementInst::Create( |
| 599 | Vec: Src, Idx: ConstantInt::get(Ty: Int64Ty, V: *SrcIdx, IsSigned: false)); |
| 600 | } |
| 601 | } |
| 602 | } else if (auto *CI = dyn_cast<CastInst>(Val: I)) { |
| 603 | // Canonicalize extractelement(cast) -> cast(extractelement). |
| 604 | // Bitcasts can change the number of vector elements, and they cost |
| 605 | // nothing. |
| 606 | // If the CI has only one use, but that use is inside a loop, this |
| 607 | // canonicalization is not profitable because it would turn a vector |
| 608 | // operation into scalar operations inside the loop. Apply the transform |
| 609 | // when: |
| 610 | // - the index is constant and CI has one use, or |
| 611 | // - the CI and EI are in the same basic block, so the cast won't be sunk |
| 612 | // into a loop. |
| 613 | if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast) && |
| 614 | (EI.getParent() == CI->getParent() || isa<ConstantInt>(Val: Index))) { |
| 615 | Value *EE = Builder.CreateExtractElement(Vec: CI->getOperand(i_nocapture: 0), Idx: Index); |
| 616 | return CastInst::Create(CI->getOpcode(), S: EE, Ty: EI.getType()); |
| 617 | } |
| 618 | } |
| 619 | } |
| 620 | |
| 621 | // Run demanded elements after other transforms as this can drop flags on |
| 622 | // binops. If there's two paths to the same final result, we prefer the |
| 623 | // one which doesn't force us to drop flags. |
| 624 | if (IndexC) { |
| 625 | ElementCount EC = EI.getVectorOperandType()->getElementCount(); |
| 626 | unsigned NumElts = EC.getKnownMinValue(); |
| 627 | // This instruction only demands the single element from the input vector. |
| 628 | // Skip for scalable type, the number of elements is unknown at |
| 629 | // compile-time. |
| 630 | if (!EC.isScalable() && NumElts != 1) { |
| 631 | // If the input vector has a single use, simplify it based on this use |
| 632 | // property. |
| 633 | if (SrcVec->hasOneUse()) { |
| 634 | APInt PoisonElts(NumElts, 0); |
| 635 | APInt DemandedElts(NumElts, 0); |
| 636 | DemandedElts.setBit(IndexC->getZExtValue()); |
| 637 | if (Value *V = |
| 638 | SimplifyDemandedVectorElts(V: SrcVec, DemandedElts, PoisonElts)) |
| 639 | return replaceOperand(I&: EI, OpNum: 0, V); |
| 640 | } else { |
| 641 | // If the input vector has multiple uses, simplify it based on a union |
| 642 | // of all elements used. |
| 643 | APInt DemandedElts = findDemandedEltsByAllUsers(V: SrcVec); |
| 644 | if (!DemandedElts.isAllOnes()) { |
| 645 | APInt PoisonElts(NumElts, 0); |
| 646 | if (Value *V = SimplifyDemandedVectorElts( |
| 647 | V: SrcVec, DemandedElts, PoisonElts, Depth: 0 /* Depth */, |
| 648 | AllowMultipleUsers: true /* AllowMultipleUsers */)) { |
| 649 | if (V != SrcVec) { |
| 650 | Worklist.addValue(V: SrcVec); |
| 651 | SrcVec->replaceAllUsesWith(V); |
| 652 | return &EI; |
| 653 | } |
| 654 | } |
| 655 | } |
| 656 | } |
| 657 | } |
| 658 | } |
| 659 | return nullptr; |
| 660 | } |
| 661 | |
| 662 | /// If V is a shuffle of values that ONLY returns elements from either LHS or |
| 663 | /// RHS, return the shuffle mask and true. Otherwise, return false. |
| 664 | static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS, |
| 665 | SmallVectorImpl<int> &Mask) { |
| 666 | assert(LHS->getType() == RHS->getType() && |
| 667 | "Invalid CollectSingleShuffleElements" ); |
| 668 | unsigned NumElts = cast<FixedVectorType>(Val: V->getType())->getNumElements(); |
| 669 | |
| 670 | if (match(V, P: m_Poison())) { |
| 671 | Mask.assign(NumElts, Elt: -1); |
| 672 | return true; |
| 673 | } |
| 674 | |
| 675 | if (V == LHS) { |
| 676 | for (unsigned i = 0; i != NumElts; ++i) |
| 677 | Mask.push_back(Elt: i); |
| 678 | return true; |
| 679 | } |
| 680 | |
| 681 | if (V == RHS) { |
| 682 | for (unsigned i = 0; i != NumElts; ++i) |
| 683 | Mask.push_back(Elt: i + NumElts); |
| 684 | return true; |
| 685 | } |
| 686 | |
| 687 | if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(Val: V)) { |
| 688 | // If this is an insert of an extract from some other vector, include it. |
| 689 | Value *VecOp = IEI->getOperand(i_nocapture: 0); |
| 690 | Value *ScalarOp = IEI->getOperand(i_nocapture: 1); |
| 691 | Value *IdxOp = IEI->getOperand(i_nocapture: 2); |
| 692 | |
| 693 | if (!isa<ConstantInt>(Val: IdxOp)) |
| 694 | return false; |
| 695 | unsigned InsertedIdx = cast<ConstantInt>(Val: IdxOp)->getZExtValue(); |
| 696 | |
| 697 | if (isa<PoisonValue>(Val: ScalarOp)) { // inserting poison into vector. |
| 698 | // We can handle this if the vector we are inserting into is |
| 699 | // transitively ok. |
| 700 | if (collectSingleShuffleElements(V: VecOp, LHS, RHS, Mask)) { |
| 701 | // If so, update the mask to reflect the inserted poison. |
| 702 | Mask[InsertedIdx] = -1; |
| 703 | return true; |
| 704 | } |
| 705 | } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(Val: ScalarOp)){ |
| 706 | if (isa<ConstantInt>(Val: EI->getOperand(i_nocapture: 1))) { |
| 707 | unsigned = |
| 708 | cast<ConstantInt>(Val: EI->getOperand(i_nocapture: 1))->getZExtValue(); |
| 709 | unsigned NumLHSElts = |
| 710 | cast<FixedVectorType>(Val: LHS->getType())->getNumElements(); |
| 711 | |
| 712 | // This must be extracting from either LHS or RHS. |
| 713 | if (EI->getOperand(i_nocapture: 0) == LHS || EI->getOperand(i_nocapture: 0) == RHS) { |
| 714 | // We can handle this if the vector we are inserting into is |
| 715 | // transitively ok. |
| 716 | if (collectSingleShuffleElements(V: VecOp, LHS, RHS, Mask)) { |
| 717 | // If so, update the mask to reflect the inserted value. |
| 718 | if (EI->getOperand(i_nocapture: 0) == LHS) { |
| 719 | Mask[InsertedIdx % NumElts] = ExtractedIdx; |
| 720 | } else { |
| 721 | assert(EI->getOperand(0) == RHS); |
| 722 | Mask[InsertedIdx % NumElts] = ExtractedIdx + NumLHSElts; |
| 723 | } |
| 724 | return true; |
| 725 | } |
| 726 | } |
| 727 | } |
| 728 | } |
| 729 | } |
| 730 | |
| 731 | return false; |
| 732 | } |
| 733 | |
| 734 | /// If we have insertion into a vector that is wider than the vector that we |
| 735 | /// are extracting from, try to widen the source vector to allow a single |
| 736 | /// shufflevector to replace one or more insert/extract pairs. |
| 737 | static bool (InsertElementInst *InsElt, |
| 738 | ExtractElementInst *ExtElt, |
| 739 | InstCombinerImpl &IC) { |
| 740 | auto *InsVecType = cast<FixedVectorType>(Val: InsElt->getType()); |
| 741 | auto *ExtVecType = cast<FixedVectorType>(Val: ExtElt->getVectorOperandType()); |
| 742 | unsigned NumInsElts = InsVecType->getNumElements(); |
| 743 | unsigned NumExtElts = ExtVecType->getNumElements(); |
| 744 | |
| 745 | // The inserted-to vector must be wider than the extracted-from vector. |
| 746 | if (InsVecType->getElementType() != ExtVecType->getElementType() || |
| 747 | NumExtElts >= NumInsElts) |
| 748 | return false; |
| 749 | |
| 750 | Value *ExtVecOp = ExtElt->getVectorOperand(); |
| 751 | // Bail out on constant vectors. |
| 752 | if (isa<ConstantData>(Val: ExtVecOp)) |
| 753 | return false; |
| 754 | |
| 755 | // Create a shuffle mask to widen the extended-from vector using poison |
| 756 | // values. The mask selects all of the values of the original vector followed |
| 757 | // by as many poison values as needed to create a vector of the same length |
| 758 | // as the inserted-to vector. |
| 759 | SmallVector<int, 16> ExtendMask; |
| 760 | for (unsigned i = 0; i < NumExtElts; ++i) |
| 761 | ExtendMask.push_back(Elt: i); |
| 762 | for (unsigned i = NumExtElts; i < NumInsElts; ++i) |
| 763 | ExtendMask.push_back(Elt: -1); |
| 764 | |
| 765 | auto *ExtVecOpInst = dyn_cast<Instruction>(Val: ExtVecOp); |
| 766 | BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(Val: ExtVecOpInst)) |
| 767 | ? ExtVecOpInst->getParent() |
| 768 | : ExtElt->getParent(); |
| 769 | |
| 770 | // TODO: This restriction matches the basic block check below when creating |
| 771 | // new extractelement instructions. If that limitation is removed, this one |
| 772 | // could also be removed. But for now, we just bail out to ensure that we |
| 773 | // will replace the extractelement instruction that is feeding our |
| 774 | // insertelement instruction. This allows the insertelement to then be |
| 775 | // replaced by a shufflevector. If the insertelement is not replaced, we can |
| 776 | // induce infinite looping because there's an optimization for extractelement |
| 777 | // that will delete our widening shuffle. This would trigger another attempt |
| 778 | // here to create that shuffle, and we spin forever. |
| 779 | if (InsertionBlock != InsElt->getParent()) |
| 780 | return false; |
| 781 | |
| 782 | // TODO: This restriction matches the check in visitInsertElementInst() and |
| 783 | // prevents an infinite loop caused by not turning the extract/insert pair |
| 784 | // into a shuffle. We really should not need either check, but we're lacking |
| 785 | // folds for shufflevectors because we're afraid to generate shuffle masks |
| 786 | // that the backend can't handle. |
| 787 | if (InsElt->hasOneUse() && isa<InsertElementInst>(Val: InsElt->user_back())) |
| 788 | return false; |
| 789 | |
| 790 | auto *WideVec = new ShuffleVectorInst(ExtVecOp, ExtendMask); |
| 791 | |
| 792 | // Insert the new shuffle after the vector operand of the extract is defined |
| 793 | // (as long as it's not a PHI) or at the start of the basic block of the |
| 794 | // extract, so any subsequent extracts in the same basic block can use it. |
| 795 | // TODO: Insert before the earliest ExtractElementInst that is replaced. |
| 796 | if (ExtVecOpInst && !isa<PHINode>(Val: ExtVecOpInst)) |
| 797 | WideVec->insertAfter(InsertPos: ExtVecOpInst->getIterator()); |
| 798 | else |
| 799 | IC.InsertNewInstWith(New: WideVec, Old: ExtElt->getParent()->getFirstInsertionPt()); |
| 800 | |
| 801 | // WideVec is an extension of ExtVecOp to produce a more useful value for |
| 802 | // ExtractElement instructions. If ExtVecOp is an instruction, adopt its |
| 803 | // DebugLoc; if it is not, then this is materializing a constant value, so set |
| 804 | // a CompilerGenerated location. |
| 805 | if (ExtVecOpInst) |
| 806 | WideVec->setDebugLoc(ExtVecOpInst->getDebugLoc()); |
| 807 | else |
| 808 | WideVec->setDebugLoc(DebugLoc::getCompilerGenerated()); |
| 809 | |
| 810 | // Replace extracts from the original narrow vector with extracts from the new |
| 811 | // wide vector. |
| 812 | for (User *U : ExtVecOp->users()) { |
| 813 | ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(Val: U); |
| 814 | if (!OldExt || OldExt->getParent() != WideVec->getParent()) |
| 815 | continue; |
| 816 | auto *NewExt = ExtractElementInst::Create(Vec: WideVec, Idx: OldExt->getOperand(i_nocapture: 1)); |
| 817 | IC.InsertNewInstWith(New: NewExt, Old: OldExt->getIterator()); |
| 818 | IC.replaceInstUsesWith(I&: *OldExt, V: NewExt); |
| 819 | // Add the old extracts to the worklist for DCE. We can't remove the |
| 820 | // extracts directly, because they may still be used by the calling code. |
| 821 | IC.addToWorklist(I: OldExt); |
| 822 | } |
| 823 | |
| 824 | return true; |
| 825 | } |
| 826 | |
| 827 | /// We are building a shuffle to create V, which is a sequence of insertelement, |
| 828 | /// extractelement pairs. If PermittedRHS is set, then we must either use it or |
| 829 | /// not rely on the second vector source. Return a std::pair containing the |
| 830 | /// left and right vectors of the proposed shuffle (or 0), and set the Mask |
| 831 | /// parameter as required. |
| 832 | /// |
| 833 | /// Note: we intentionally don't try to fold earlier shuffles since they have |
| 834 | /// often been chosen carefully to be efficiently implementable on the target. |
| 835 | using ShuffleOps = std::pair<Value *, Value *>; |
| 836 | |
| 837 | static ShuffleOps collectShuffleElements(Value *V, SmallVectorImpl<int> &Mask, |
| 838 | Value *PermittedRHS, |
| 839 | InstCombinerImpl &IC, bool &Rerun) { |
| 840 | assert(V->getType()->isVectorTy() && "Invalid shuffle!" ); |
| 841 | unsigned NumElts = cast<FixedVectorType>(Val: V->getType())->getNumElements(); |
| 842 | |
| 843 | if (match(V, P: m_Poison())) { |
| 844 | Mask.assign(NumElts, Elt: -1); |
| 845 | return std::make_pair( |
| 846 | x: PermittedRHS ? PoisonValue::get(T: PermittedRHS->getType()) : V, y: nullptr); |
| 847 | } |
| 848 | |
| 849 | if (isa<ConstantAggregateZero>(Val: V)) { |
| 850 | Mask.assign(NumElts, Elt: 0); |
| 851 | return std::make_pair(x&: V, y: nullptr); |
| 852 | } |
| 853 | |
| 854 | if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(Val: V)) { |
| 855 | // If this is an insert of an extract from some other vector, include it. |
| 856 | Value *VecOp = IEI->getOperand(i_nocapture: 0); |
| 857 | Value *ScalarOp = IEI->getOperand(i_nocapture: 1); |
| 858 | Value *IdxOp = IEI->getOperand(i_nocapture: 2); |
| 859 | |
| 860 | if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(Val: ScalarOp)) { |
| 861 | if (isa<ConstantInt>(Val: EI->getOperand(i_nocapture: 1)) && isa<ConstantInt>(Val: IdxOp)) { |
| 862 | unsigned = |
| 863 | cast<ConstantInt>(Val: EI->getOperand(i_nocapture: 1))->getZExtValue(); |
| 864 | unsigned InsertedIdx = cast<ConstantInt>(Val: IdxOp)->getZExtValue(); |
| 865 | |
| 866 | // Either the extracted from or inserted into vector must be RHSVec, |
| 867 | // otherwise we'd end up with a shuffle of three inputs. |
| 868 | if (EI->getOperand(i_nocapture: 0) == PermittedRHS || PermittedRHS == nullptr) { |
| 869 | Value *RHS = EI->getOperand(i_nocapture: 0); |
| 870 | ShuffleOps LR = collectShuffleElements(V: VecOp, Mask, PermittedRHS: RHS, IC, Rerun); |
| 871 | assert(LR.second == nullptr || LR.second == RHS); |
| 872 | |
| 873 | if (LR.first->getType() != RHS->getType()) { |
| 874 | // Although we are giving up for now, see if we can create extracts |
| 875 | // that match the inserts for another round of combining. |
| 876 | if (replaceExtractElements(InsElt: IEI, ExtElt: EI, IC)) |
| 877 | Rerun = true; |
| 878 | |
| 879 | // We tried our best, but we can't find anything compatible with RHS |
| 880 | // further up the chain. Return a trivial shuffle. |
| 881 | for (unsigned i = 0; i < NumElts; ++i) |
| 882 | Mask[i] = i; |
| 883 | return std::make_pair(x&: V, y: nullptr); |
| 884 | } |
| 885 | |
| 886 | unsigned NumLHSElts = |
| 887 | cast<FixedVectorType>(Val: RHS->getType())->getNumElements(); |
| 888 | Mask[InsertedIdx % NumElts] = NumLHSElts + ExtractedIdx; |
| 889 | return std::make_pair(x&: LR.first, y&: RHS); |
| 890 | } |
| 891 | |
| 892 | if (VecOp == PermittedRHS) { |
| 893 | // We've gone as far as we can: anything on the other side of the |
| 894 | // extractelement will already have been converted into a shuffle. |
| 895 | unsigned NumLHSElts = |
| 896 | cast<FixedVectorType>(Val: EI->getOperand(i_nocapture: 0)->getType()) |
| 897 | ->getNumElements(); |
| 898 | for (unsigned i = 0; i != NumElts; ++i) |
| 899 | Mask.push_back(Elt: i == InsertedIdx ? ExtractedIdx : NumLHSElts + i); |
| 900 | return std::make_pair(x: EI->getOperand(i_nocapture: 0), y&: PermittedRHS); |
| 901 | } |
| 902 | |
| 903 | // If this insertelement is a chain that comes from exactly these two |
| 904 | // vectors, return the vector and the effective shuffle. |
| 905 | if (EI->getOperand(i_nocapture: 0)->getType() == PermittedRHS->getType() && |
| 906 | collectSingleShuffleElements(V: IEI, LHS: EI->getOperand(i_nocapture: 0), RHS: PermittedRHS, |
| 907 | Mask)) |
| 908 | return std::make_pair(x: EI->getOperand(i_nocapture: 0), y&: PermittedRHS); |
| 909 | } |
| 910 | } |
| 911 | } |
| 912 | |
| 913 | // Otherwise, we can't do anything fancy. Return an identity vector. |
| 914 | for (unsigned i = 0; i != NumElts; ++i) |
| 915 | Mask.push_back(Elt: i); |
| 916 | return std::make_pair(x&: V, y: nullptr); |
| 917 | } |
| 918 | |
| 919 | /// Look for chain of insertvalue's that fully define an aggregate, and trace |
| 920 | /// back the values inserted, see if they are all were extractvalue'd from |
| 921 | /// the same source aggregate from the exact same element indexes. |
| 922 | /// If they were, just reuse the source aggregate. |
| 923 | /// This potentially deals with PHI indirections. |
| 924 | Instruction *InstCombinerImpl::foldAggregateConstructionIntoAggregateReuse( |
| 925 | InsertValueInst &OrigIVI) { |
| 926 | Type *AggTy = OrigIVI.getType(); |
| 927 | unsigned NumAggElts; |
| 928 | switch (AggTy->getTypeID()) { |
| 929 | case Type::StructTyID: |
| 930 | NumAggElts = AggTy->getStructNumElements(); |
| 931 | break; |
| 932 | case Type::ArrayTyID: |
| 933 | NumAggElts = AggTy->getArrayNumElements(); |
| 934 | break; |
| 935 | default: |
| 936 | llvm_unreachable("Unhandled aggregate type?" ); |
| 937 | } |
| 938 | |
| 939 | // Arbitrary aggregate size cut-off. Motivation for limit of 2 is to be able |
| 940 | // to handle clang C++ exception struct (which is hardcoded as {i8*, i32}), |
| 941 | // FIXME: any interesting patterns to be caught with larger limit? |
| 942 | assert(NumAggElts > 0 && "Aggregate should have elements." ); |
| 943 | if (NumAggElts > 2) |
| 944 | return nullptr; |
| 945 | |
| 946 | static constexpr auto NotFound = std::nullopt; |
| 947 | static constexpr auto FoundMismatch = nullptr; |
| 948 | |
| 949 | // Try to find a value of each element of an aggregate. |
| 950 | // FIXME: deal with more complex, not one-dimensional, aggregate types |
| 951 | SmallVector<std::optional<Instruction *>, 2> AggElts(NumAggElts, NotFound); |
| 952 | |
| 953 | // Do we know values for each element of the aggregate? |
| 954 | auto KnowAllElts = [&AggElts]() { |
| 955 | return !llvm::is_contained(Range&: AggElts, Element: NotFound); |
| 956 | }; |
| 957 | |
| 958 | int Depth = 0; |
| 959 | |
| 960 | // Arbitrary `insertvalue` visitation depth limit. Let's be okay with |
| 961 | // every element being overwritten twice, which should never happen. |
| 962 | static const int DepthLimit = 2 * NumAggElts; |
| 963 | |
| 964 | // Recurse up the chain of `insertvalue` aggregate operands until either we've |
| 965 | // reconstructed full initializer or can't visit any more `insertvalue`'s. |
| 966 | for (InsertValueInst *CurrIVI = &OrigIVI; |
| 967 | Depth < DepthLimit && CurrIVI && !KnowAllElts(); |
| 968 | CurrIVI = dyn_cast<InsertValueInst>(Val: CurrIVI->getAggregateOperand()), |
| 969 | ++Depth) { |
| 970 | auto *InsertedValue = |
| 971 | dyn_cast<Instruction>(Val: CurrIVI->getInsertedValueOperand()); |
| 972 | if (!InsertedValue) |
| 973 | return nullptr; // Inserted value must be produced by an instruction. |
| 974 | |
| 975 | ArrayRef<unsigned int> Indices = CurrIVI->getIndices(); |
| 976 | |
| 977 | // Don't bother with more than single-level aggregates. |
| 978 | if (Indices.size() != 1) |
| 979 | return nullptr; // FIXME: deal with more complex aggregates? |
| 980 | |
| 981 | // Now, we may have already previously recorded the value for this element |
| 982 | // of an aggregate. If we did, that means the CurrIVI will later be |
| 983 | // overwritten with the already-recorded value. But if not, let's record it! |
| 984 | std::optional<Instruction *> &Elt = AggElts[Indices.front()]; |
| 985 | Elt = Elt.value_or(u&: InsertedValue); |
| 986 | |
| 987 | // FIXME: should we handle chain-terminating undef base operand? |
| 988 | } |
| 989 | |
| 990 | // Was that sufficient to deduce the full initializer for the aggregate? |
| 991 | if (!KnowAllElts()) |
| 992 | return nullptr; // Give up then. |
| 993 | |
| 994 | // We now want to find the source[s] of the aggregate elements we've found. |
| 995 | // And with "source" we mean the original aggregate[s] from which |
| 996 | // the inserted elements were extracted. This may require PHI translation. |
| 997 | |
| 998 | enum class AggregateDescription { |
| 999 | /// When analyzing the value that was inserted into an aggregate, we did |
| 1000 | /// not manage to find defining `extractvalue` instruction to analyze. |
| 1001 | NotFound, |
| 1002 | /// When analyzing the value that was inserted into an aggregate, we did |
| 1003 | /// manage to find defining `extractvalue` instruction[s], and everything |
| 1004 | /// matched perfectly - aggregate type, element insertion/extraction index. |
| 1005 | Found, |
| 1006 | /// When analyzing the value that was inserted into an aggregate, we did |
| 1007 | /// manage to find defining `extractvalue` instruction, but there was |
| 1008 | /// a mismatch: either the source type from which the extraction was didn't |
| 1009 | /// match the aggregate type into which the insertion was, |
| 1010 | /// or the extraction/insertion channels mismatched, |
| 1011 | /// or different elements had different source aggregates. |
| 1012 | FoundMismatch |
| 1013 | }; |
| 1014 | auto Describe = [](std::optional<Value *> SourceAggregate) { |
| 1015 | if (SourceAggregate == NotFound) |
| 1016 | return AggregateDescription::NotFound; |
| 1017 | if (*SourceAggregate == FoundMismatch) |
| 1018 | return AggregateDescription::FoundMismatch; |
| 1019 | return AggregateDescription::Found; |
| 1020 | }; |
| 1021 | |
| 1022 | // If an aggregate element is defined in UseBB, we can't use it in PredBB. |
| 1023 | bool EltDefinedInUseBB = false; |
| 1024 | |
| 1025 | // Given the value \p Elt that was being inserted into element \p EltIdx of an |
| 1026 | // aggregate AggTy, see if \p Elt was originally defined by an |
| 1027 | // appropriate extractvalue (same element index, same aggregate type). |
| 1028 | // If found, return the source aggregate from which the extraction was. |
| 1029 | // If \p PredBB is provided, does PHI translation of an \p Elt first. |
| 1030 | auto FindSourceAggregate = |
| 1031 | [&](Instruction *Elt, unsigned EltIdx, std::optional<BasicBlock *> UseBB, |
| 1032 | std::optional<BasicBlock *> PredBB) -> std::optional<Value *> { |
| 1033 | // For now(?), only deal with, at most, a single level of PHI indirection. |
| 1034 | if (UseBB && PredBB) { |
| 1035 | Elt = dyn_cast<Instruction>(Val: Elt->DoPHITranslation(CurBB: *UseBB, PredBB: *PredBB)); |
| 1036 | if (Elt && Elt->getParent() == *UseBB) |
| 1037 | EltDefinedInUseBB = true; |
| 1038 | } |
| 1039 | // FIXME: deal with multiple levels of PHI indirection? |
| 1040 | |
| 1041 | // Did we find an extraction? |
| 1042 | auto *EVI = dyn_cast_or_null<ExtractValueInst>(Val: Elt); |
| 1043 | if (!EVI) |
| 1044 | return NotFound; |
| 1045 | |
| 1046 | Value *SourceAggregate = EVI->getAggregateOperand(); |
| 1047 | |
| 1048 | // Is the extraction from the same type into which the insertion was? |
| 1049 | if (SourceAggregate->getType() != AggTy) |
| 1050 | return FoundMismatch; |
| 1051 | // And the element index doesn't change between extraction and insertion? |
| 1052 | if (EVI->getNumIndices() != 1 || EltIdx != EVI->getIndices().front()) |
| 1053 | return FoundMismatch; |
| 1054 | |
| 1055 | return SourceAggregate; // AggregateDescription::Found |
| 1056 | }; |
| 1057 | |
| 1058 | // Given elements AggElts that were constructing an aggregate OrigIVI, |
| 1059 | // see if we can find appropriate source aggregate for each of the elements, |
| 1060 | // and see it's the same aggregate for each element. If so, return it. |
| 1061 | auto FindCommonSourceAggregate = |
| 1062 | [&](std::optional<BasicBlock *> UseBB, |
| 1063 | std::optional<BasicBlock *> PredBB) -> std::optional<Value *> { |
| 1064 | std::optional<Value *> SourceAggregate; |
| 1065 | |
| 1066 | for (auto I : enumerate(First&: AggElts)) { |
| 1067 | assert(Describe(SourceAggregate) != AggregateDescription::FoundMismatch && |
| 1068 | "We don't store nullptr in SourceAggregate!" ); |
| 1069 | assert((Describe(SourceAggregate) == AggregateDescription::Found) == |
| 1070 | (I.index() != 0) && |
| 1071 | "SourceAggregate should be valid after the first element," ); |
| 1072 | |
| 1073 | // For this element, is there a plausible source aggregate? |
| 1074 | // FIXME: we could special-case undef element, IFF we know that in the |
| 1075 | // source aggregate said element isn't poison. |
| 1076 | std::optional<Value *> SourceAggregateForElement = |
| 1077 | FindSourceAggregate(*I.value(), I.index(), UseBB, PredBB); |
| 1078 | |
| 1079 | // Okay, what have we found? Does that correlate with previous findings? |
| 1080 | |
| 1081 | // Regardless of whether or not we have previously found source |
| 1082 | // aggregate for previous elements (if any), if we didn't find one for |
| 1083 | // this element, passthrough whatever we have just found. |
| 1084 | if (Describe(SourceAggregateForElement) != AggregateDescription::Found) |
| 1085 | return SourceAggregateForElement; |
| 1086 | |
| 1087 | // Okay, we have found source aggregate for this element. |
| 1088 | // Let's see what we already know from previous elements, if any. |
| 1089 | switch (Describe(SourceAggregate)) { |
| 1090 | case AggregateDescription::NotFound: |
| 1091 | // This is apparently the first element that we have examined. |
| 1092 | SourceAggregate = SourceAggregateForElement; // Record the aggregate! |
| 1093 | continue; // Great, now look at next element. |
| 1094 | case AggregateDescription::Found: |
| 1095 | // We have previously already successfully examined other elements. |
| 1096 | // Is this the same source aggregate we've found for other elements? |
| 1097 | if (*SourceAggregateForElement != *SourceAggregate) |
| 1098 | return FoundMismatch; |
| 1099 | continue; // Still the same aggregate, look at next element. |
| 1100 | case AggregateDescription::FoundMismatch: |
| 1101 | llvm_unreachable("Can't happen. We would have early-exited then." ); |
| 1102 | }; |
| 1103 | } |
| 1104 | |
| 1105 | assert(Describe(SourceAggregate) == AggregateDescription::Found && |
| 1106 | "Must be a valid Value" ); |
| 1107 | return *SourceAggregate; |
| 1108 | }; |
| 1109 | |
| 1110 | std::optional<Value *> SourceAggregate; |
| 1111 | |
| 1112 | // Can we find the source aggregate without looking at predecessors? |
| 1113 | SourceAggregate = FindCommonSourceAggregate(/*UseBB=*/std::nullopt, |
| 1114 | /*PredBB=*/std::nullopt); |
| 1115 | if (Describe(SourceAggregate) != AggregateDescription::NotFound) { |
| 1116 | if (Describe(SourceAggregate) == AggregateDescription::FoundMismatch) |
| 1117 | return nullptr; // Conflicting source aggregates! |
| 1118 | ++NumAggregateReconstructionsSimplified; |
| 1119 | return replaceInstUsesWith(I&: OrigIVI, V: *SourceAggregate); |
| 1120 | } |
| 1121 | |
| 1122 | // Okay, apparently we need to look at predecessors. |
| 1123 | |
| 1124 | // We should be smart about picking the "use" basic block, which will be the |
| 1125 | // merge point for aggregate, where we'll insert the final PHI that will be |
| 1126 | // used instead of OrigIVI. Basic block of OrigIVI is *not* the right choice. |
| 1127 | // We should look in which blocks each of the AggElts is being defined, |
| 1128 | // they all should be defined in the same basic block. |
| 1129 | BasicBlock *UseBB = nullptr; |
| 1130 | |
| 1131 | for (const std::optional<Instruction *> &I : AggElts) { |
| 1132 | BasicBlock *BB = (*I)->getParent(); |
| 1133 | // If it's the first instruction we've encountered, record the basic block. |
| 1134 | if (!UseBB) { |
| 1135 | UseBB = BB; |
| 1136 | continue; |
| 1137 | } |
| 1138 | // Otherwise, this must be the same basic block we've seen previously. |
| 1139 | if (UseBB != BB) |
| 1140 | return nullptr; |
| 1141 | } |
| 1142 | |
| 1143 | // If *all* of the elements are basic-block-independent, meaning they are |
| 1144 | // either function arguments, or constant expressions, then if we didn't |
| 1145 | // handle them without predecessor-aware handling, we won't handle them now. |
| 1146 | if (!UseBB) |
| 1147 | return nullptr; |
| 1148 | |
| 1149 | // If we didn't manage to find source aggregate without looking at |
| 1150 | // predecessors, and there are no predecessors to look at, then we're done. |
| 1151 | if (pred_empty(BB: UseBB)) |
| 1152 | return nullptr; |
| 1153 | |
| 1154 | // Arbitrary predecessor count limit. |
| 1155 | static const int PredCountLimit = 64; |
| 1156 | |
| 1157 | // Cache the (non-uniqified!) list of predecessors in a vector, |
| 1158 | // checking the limit at the same time for efficiency. |
| 1159 | SmallVector<BasicBlock *, 4> Preds; // May have duplicates! |
| 1160 | for (BasicBlock *Pred : predecessors(BB: UseBB)) { |
| 1161 | // Don't bother if there are too many predecessors. |
| 1162 | if (Preds.size() >= PredCountLimit) // FIXME: only count duplicates once? |
| 1163 | return nullptr; |
| 1164 | Preds.emplace_back(Args&: Pred); |
| 1165 | } |
| 1166 | |
| 1167 | // For each predecessor, what is the source aggregate, |
| 1168 | // from which all the elements were originally extracted from? |
| 1169 | // Note that we want for the map to have stable iteration order! |
| 1170 | SmallMapVector<BasicBlock *, Value *, 4> SourceAggregates; |
| 1171 | bool FoundSrcAgg = false; |
| 1172 | for (BasicBlock *Pred : Preds) { |
| 1173 | std::pair<decltype(SourceAggregates)::iterator, bool> IV = |
| 1174 | SourceAggregates.try_emplace(Key: Pred); |
| 1175 | // Did we already evaluate this predecessor? |
| 1176 | if (!IV.second) |
| 1177 | continue; |
| 1178 | |
| 1179 | // Let's hope that when coming from predecessor Pred, all elements of the |
| 1180 | // aggregate produced by OrigIVI must have been originally extracted from |
| 1181 | // the same aggregate. Is that so? Can we find said original aggregate? |
| 1182 | SourceAggregate = FindCommonSourceAggregate(UseBB, Pred); |
| 1183 | if (Describe(SourceAggregate) == AggregateDescription::Found) { |
| 1184 | FoundSrcAgg = true; |
| 1185 | IV.first->second = *SourceAggregate; |
| 1186 | } else { |
| 1187 | // If UseBB is the single successor of Pred, we can add InsertValue to |
| 1188 | // Pred. |
| 1189 | auto *BI = dyn_cast<UncondBrInst>(Val: Pred->getTerminator()); |
| 1190 | if (!BI) |
| 1191 | return nullptr; |
| 1192 | } |
| 1193 | } |
| 1194 | |
| 1195 | if (!FoundSrcAgg) |
| 1196 | return nullptr; |
| 1197 | |
| 1198 | // Do some sanity check if we need to add insertvalue into predecessors. |
| 1199 | auto OrigBB = OrigIVI.getParent(); |
| 1200 | for (auto &It : SourceAggregates) { |
| 1201 | if (Describe(It.second) == AggregateDescription::Found) |
| 1202 | continue; |
| 1203 | |
| 1204 | // Element is defined in UseBB, so it can't be used in predecessors. |
| 1205 | if (EltDefinedInUseBB) |
| 1206 | return nullptr; |
| 1207 | |
| 1208 | // Do this transformation cross loop boundary may cause dead loop. So we |
| 1209 | // should avoid this situation. But LoopInfo is not generally available, we |
| 1210 | // must be conservative here. |
| 1211 | // If OrigIVI is in UseBB and it's the only successor of PredBB, PredBB |
| 1212 | // can't be in inner loop. |
| 1213 | if (UseBB != OrigBB) |
| 1214 | return nullptr; |
| 1215 | |
| 1216 | // Avoid constructing constant aggregate because constant value may expose |
| 1217 | // more optimizations. |
| 1218 | bool ConstAgg = true; |
| 1219 | for (auto Val : AggElts) { |
| 1220 | Value *Elt = (*Val)->DoPHITranslation(CurBB: UseBB, PredBB: It.first); |
| 1221 | if (!isa<Constant>(Val: Elt)) { |
| 1222 | ConstAgg = false; |
| 1223 | break; |
| 1224 | } |
| 1225 | } |
| 1226 | if (ConstAgg) |
| 1227 | return nullptr; |
| 1228 | } |
| 1229 | |
| 1230 | // For predecessors without appropriate source aggregate, create one in the |
| 1231 | // predecessor. |
| 1232 | for (auto &It : SourceAggregates) { |
| 1233 | if (Describe(It.second) == AggregateDescription::Found) |
| 1234 | continue; |
| 1235 | |
| 1236 | BasicBlock *Pred = It.first; |
| 1237 | Builder.SetInsertPoint(Pred->getTerminator()); |
| 1238 | Value *V = PoisonValue::get(T: AggTy); |
| 1239 | for (auto [Idx, Val] : enumerate(First&: AggElts)) { |
| 1240 | Value *Elt = (*Val)->DoPHITranslation(CurBB: UseBB, PredBB: Pred); |
| 1241 | V = Builder.CreateInsertValue(Agg: V, Val: Elt, Idxs: Idx); |
| 1242 | } |
| 1243 | |
| 1244 | It.second = V; |
| 1245 | } |
| 1246 | |
| 1247 | // All good! Now we just need to thread the source aggregates here. |
| 1248 | // Note that we have to insert the new PHI here, ourselves, because we can't |
| 1249 | // rely on InstCombinerImpl::run() inserting it into the right basic block. |
| 1250 | // Note that the same block can be a predecessor more than once, |
| 1251 | // and we need to preserve that invariant for the PHI node. |
| 1252 | BuilderTy::InsertPointGuard Guard(Builder); |
| 1253 | Builder.SetInsertPoint(TheBB: UseBB, IP: UseBB->getFirstNonPHIIt()); |
| 1254 | auto *PHI = |
| 1255 | Builder.CreatePHI(Ty: AggTy, NumReservedValues: Preds.size(), Name: OrigIVI.getName() + ".merged" ); |
| 1256 | for (BasicBlock *Pred : Preds) |
| 1257 | PHI->addIncoming(V: SourceAggregates[Pred], BB: Pred); |
| 1258 | |
| 1259 | ++NumAggregateReconstructionsSimplified; |
| 1260 | return replaceInstUsesWith(I&: OrigIVI, V: PHI); |
| 1261 | } |
| 1262 | |
| 1263 | /// Try to find redundant insertvalue instructions, like the following ones: |
| 1264 | /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0 |
| 1265 | /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0 |
| 1266 | /// Here the second instruction inserts values at the same indices, as the |
| 1267 | /// first one, making the first one redundant. |
| 1268 | /// It should be transformed to: |
| 1269 | /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0 |
| 1270 | Instruction *InstCombinerImpl::visitInsertValueInst(InsertValueInst &I) { |
| 1271 | if (Value *V = simplifyInsertValueInst( |
| 1272 | Agg: I.getAggregateOperand(), Val: I.getInsertedValueOperand(), Idxs: I.getIndices(), |
| 1273 | Q: SQ.getWithInstruction(I: &I))) |
| 1274 | return replaceInstUsesWith(I, V); |
| 1275 | |
| 1276 | bool IsRedundant = false; |
| 1277 | ArrayRef<unsigned int> FirstIndices = I.getIndices(); |
| 1278 | |
| 1279 | // If there is a chain of insertvalue instructions (each of them except the |
| 1280 | // last one has only one use and it's another insertvalue insn from this |
| 1281 | // chain), check if any of the 'children' uses the same indices as the first |
| 1282 | // instruction. In this case, the first one is redundant. |
| 1283 | Value *V = &I; |
| 1284 | unsigned Depth = 0; |
| 1285 | while (V->hasOneUse() && Depth < 10) { |
| 1286 | User *U = V->user_back(); |
| 1287 | auto UserInsInst = dyn_cast<InsertValueInst>(Val: U); |
| 1288 | if (!UserInsInst || U->getOperand(i: 0) != V) |
| 1289 | break; |
| 1290 | if (UserInsInst->getIndices() == FirstIndices) { |
| 1291 | IsRedundant = true; |
| 1292 | break; |
| 1293 | } |
| 1294 | V = UserInsInst; |
| 1295 | Depth++; |
| 1296 | } |
| 1297 | |
| 1298 | if (IsRedundant) |
| 1299 | return replaceInstUsesWith(I, V: I.getOperand(i_nocapture: 0)); |
| 1300 | |
| 1301 | if (Instruction *NewI = foldAggregateConstructionIntoAggregateReuse(OrigIVI&: I)) |
| 1302 | return NewI; |
| 1303 | |
| 1304 | return nullptr; |
| 1305 | } |
| 1306 | |
| 1307 | static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) { |
| 1308 | // Can not analyze scalable type, the number of elements is not a compile-time |
| 1309 | // constant. |
| 1310 | if (isa<ScalableVectorType>(Val: Shuf.getOperand(i_nocapture: 0)->getType())) |
| 1311 | return false; |
| 1312 | |
| 1313 | int MaskSize = Shuf.getShuffleMask().size(); |
| 1314 | int VecSize = |
| 1315 | cast<FixedVectorType>(Val: Shuf.getOperand(i_nocapture: 0)->getType())->getNumElements(); |
| 1316 | |
| 1317 | // A vector select does not change the size of the operands. |
| 1318 | if (MaskSize != VecSize) |
| 1319 | return false; |
| 1320 | |
| 1321 | // Each mask element must be undefined or choose a vector element from one of |
| 1322 | // the source operands without crossing vector lanes. |
| 1323 | for (int i = 0; i != MaskSize; ++i) { |
| 1324 | int Elt = Shuf.getMaskValue(Elt: i); |
| 1325 | if (Elt != -1 && Elt != i && Elt != i + VecSize) |
| 1326 | return false; |
| 1327 | } |
| 1328 | |
| 1329 | return true; |
| 1330 | } |
| 1331 | |
| 1332 | /// Turn a chain of inserts that splats a value into an insert + shuffle: |
| 1333 | /// insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ... -> |
| 1334 | /// shufflevector(insertelt(X, %k, 0), poison, zero) |
| 1335 | static Instruction *foldInsSequenceIntoSplat(InsertElementInst &InsElt) { |
| 1336 | // We are interested in the last insert in a chain. So if this insert has a |
| 1337 | // single user and that user is an insert, bail. |
| 1338 | if (InsElt.hasOneUse() && isa<InsertElementInst>(Val: InsElt.user_back())) |
| 1339 | return nullptr; |
| 1340 | |
| 1341 | VectorType *VecTy = InsElt.getType(); |
| 1342 | // Can not handle scalable type, the number of elements is not a compile-time |
| 1343 | // constant. |
| 1344 | if (isa<ScalableVectorType>(Val: VecTy)) |
| 1345 | return nullptr; |
| 1346 | unsigned NumElements = cast<FixedVectorType>(Val: VecTy)->getNumElements(); |
| 1347 | |
| 1348 | // Do not try to do this for a one-element vector, since that's a nop, |
| 1349 | // and will cause an inf-loop. |
| 1350 | if (NumElements == 1) |
| 1351 | return nullptr; |
| 1352 | |
| 1353 | Value *SplatVal = InsElt.getOperand(i_nocapture: 1); |
| 1354 | InsertElementInst *CurrIE = &InsElt; |
| 1355 | SmallBitVector ElementPresent(NumElements, false); |
| 1356 | InsertElementInst *FirstIE = nullptr; |
| 1357 | |
| 1358 | // Walk the chain backwards, keeping track of which indices we inserted into, |
| 1359 | // until we hit something that isn't an insert of the splatted value. |
| 1360 | while (CurrIE) { |
| 1361 | auto *Idx = dyn_cast<ConstantInt>(Val: CurrIE->getOperand(i_nocapture: 2)); |
| 1362 | if (!Idx || CurrIE->getOperand(i_nocapture: 1) != SplatVal) |
| 1363 | return nullptr; |
| 1364 | |
| 1365 | auto *NextIE = dyn_cast<InsertElementInst>(Val: CurrIE->getOperand(i_nocapture: 0)); |
| 1366 | // Check none of the intermediate steps have any additional uses, except |
| 1367 | // for the root insertelement instruction, which can be re-used, if it |
| 1368 | // inserts at position 0. |
| 1369 | if (CurrIE != &InsElt && |
| 1370 | (!CurrIE->hasOneUse() && (NextIE != nullptr || !Idx->isZero()))) |
| 1371 | return nullptr; |
| 1372 | |
| 1373 | ElementPresent[Idx->getZExtValue()] = true; |
| 1374 | FirstIE = CurrIE; |
| 1375 | CurrIE = NextIE; |
| 1376 | } |
| 1377 | |
| 1378 | // If this is just a single insertelement (not a sequence), we are done. |
| 1379 | if (FirstIE == &InsElt) |
| 1380 | return nullptr; |
| 1381 | |
| 1382 | // If we are not inserting into a poison vector, make sure we've seen an |
| 1383 | // insert into every element. |
| 1384 | // TODO: If the base vector is not undef, it might be better to create a splat |
| 1385 | // and then a select-shuffle (blend) with the base vector. |
| 1386 | if (!match(V: FirstIE->getOperand(i_nocapture: 0), P: m_Poison())) |
| 1387 | if (!ElementPresent.all()) |
| 1388 | return nullptr; |
| 1389 | |
| 1390 | // Create the insert + shuffle. |
| 1391 | Type *Int64Ty = Type::getInt64Ty(C&: InsElt.getContext()); |
| 1392 | PoisonValue *PoisonVec = PoisonValue::get(T: VecTy); |
| 1393 | Constant *Zero = ConstantInt::get(Ty: Int64Ty, V: 0); |
| 1394 | if (!cast<ConstantInt>(Val: FirstIE->getOperand(i_nocapture: 2))->isZero()) |
| 1395 | FirstIE = InsertElementInst::Create(Vec: PoisonVec, NewElt: SplatVal, Idx: Zero, NameStr: "" , |
| 1396 | InsertBefore: InsElt.getIterator()); |
| 1397 | |
| 1398 | // Splat from element 0, but replace absent elements with poison in the mask. |
| 1399 | SmallVector<int, 16> Mask(NumElements, 0); |
| 1400 | for (unsigned i = 0; i != NumElements; ++i) |
| 1401 | if (!ElementPresent[i]) |
| 1402 | Mask[i] = -1; |
| 1403 | |
| 1404 | return new ShuffleVectorInst(FirstIE, Mask); |
| 1405 | } |
| 1406 | |
| 1407 | /// Try to fold an insert element into an existing splat shuffle by changing |
| 1408 | /// the shuffle's mask to include the index of this insert element. |
| 1409 | static Instruction *foldInsEltIntoSplat(InsertElementInst &InsElt) { |
| 1410 | // Check if the vector operand of this insert is a canonical splat shuffle. |
| 1411 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: InsElt.getOperand(i_nocapture: 0)); |
| 1412 | if (!Shuf || !Shuf->isZeroEltSplat()) |
| 1413 | return nullptr; |
| 1414 | |
| 1415 | // Bail out early if shuffle is scalable type. The number of elements in |
| 1416 | // shuffle mask is unknown at compile-time. |
| 1417 | if (isa<ScalableVectorType>(Val: Shuf->getType())) |
| 1418 | return nullptr; |
| 1419 | |
| 1420 | // Check for a constant insertion index. |
| 1421 | uint64_t IdxC; |
| 1422 | if (!match(V: InsElt.getOperand(i_nocapture: 2), P: m_ConstantInt(V&: IdxC))) |
| 1423 | return nullptr; |
| 1424 | |
| 1425 | // Check if the splat shuffle's input is the same as this insert's scalar op. |
| 1426 | Value *X = InsElt.getOperand(i_nocapture: 1); |
| 1427 | Value *Op0 = Shuf->getOperand(i_nocapture: 0); |
| 1428 | if (!match(V: Op0, P: m_InsertElt(Val: m_Undef(), Elt: m_Specific(V: X), Idx: m_ZeroInt()))) |
| 1429 | return nullptr; |
| 1430 | |
| 1431 | // Replace the shuffle mask element at the index of this insert with a zero. |
| 1432 | // For example: |
| 1433 | // inselt (shuf (inselt undef, X, 0), _, <0,undef,0,undef>), X, 1 |
| 1434 | // --> shuf (inselt undef, X, 0), poison, <0,0,0,undef> |
| 1435 | unsigned NumMaskElts = |
| 1436 | cast<FixedVectorType>(Val: Shuf->getType())->getNumElements(); |
| 1437 | SmallVector<int, 16> NewMask(NumMaskElts); |
| 1438 | for (unsigned i = 0; i != NumMaskElts; ++i) |
| 1439 | NewMask[i] = i == IdxC ? 0 : Shuf->getMaskValue(Elt: i); |
| 1440 | |
| 1441 | return new ShuffleVectorInst(Op0, NewMask); |
| 1442 | } |
| 1443 | |
| 1444 | /// Try to fold an extract+insert element into an existing identity shuffle by |
| 1445 | /// changing the shuffle's mask to include the index of this insert element. |
| 1446 | static Instruction *foldInsEltIntoIdentityShuffle(InsertElementInst &InsElt) { |
| 1447 | // Check if the vector operand of this insert is an identity shuffle. |
| 1448 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: InsElt.getOperand(i_nocapture: 0)); |
| 1449 | if (!Shuf || !match(V: Shuf->getOperand(i_nocapture: 1), P: m_Poison()) || |
| 1450 | !(Shuf->isIdentityWithExtract() || Shuf->isIdentityWithPadding())) |
| 1451 | return nullptr; |
| 1452 | |
| 1453 | // Bail out early if shuffle is scalable type. The number of elements in |
| 1454 | // shuffle mask is unknown at compile-time. |
| 1455 | if (isa<ScalableVectorType>(Val: Shuf->getType())) |
| 1456 | return nullptr; |
| 1457 | |
| 1458 | // Check for a constant insertion index. |
| 1459 | uint64_t IdxC; |
| 1460 | if (!match(V: InsElt.getOperand(i_nocapture: 2), P: m_ConstantInt(V&: IdxC))) |
| 1461 | return nullptr; |
| 1462 | |
| 1463 | // Check if this insert's scalar op is extracted from the identity shuffle's |
| 1464 | // input vector. |
| 1465 | Value *Scalar = InsElt.getOperand(i_nocapture: 1); |
| 1466 | Value *X = Shuf->getOperand(i_nocapture: 0); |
| 1467 | if (!match(V: Scalar, P: m_ExtractElt(Val: m_Specific(V: X), Idx: m_SpecificInt(V: IdxC)))) |
| 1468 | return nullptr; |
| 1469 | |
| 1470 | // Replace the shuffle mask element at the index of this extract+insert with |
| 1471 | // that same index value. |
| 1472 | // For example: |
| 1473 | // inselt (shuf X, IdMask), (extelt X, IdxC), IdxC --> shuf X, IdMask' |
| 1474 | unsigned NumMaskElts = |
| 1475 | cast<FixedVectorType>(Val: Shuf->getType())->getNumElements(); |
| 1476 | SmallVector<int, 16> NewMask(NumMaskElts); |
| 1477 | ArrayRef<int> OldMask = Shuf->getShuffleMask(); |
| 1478 | for (unsigned i = 0; i != NumMaskElts; ++i) { |
| 1479 | if (i != IdxC) { |
| 1480 | // All mask elements besides the inserted element remain the same. |
| 1481 | NewMask[i] = OldMask[i]; |
| 1482 | } else if (OldMask[i] == (int)IdxC) { |
| 1483 | // If the mask element was already set, there's nothing to do |
| 1484 | // (demanded elements analysis may unset it later). |
| 1485 | return nullptr; |
| 1486 | } else { |
| 1487 | assert(OldMask[i] == PoisonMaskElem && |
| 1488 | "Unexpected shuffle mask element for identity shuffle" ); |
| 1489 | NewMask[i] = IdxC; |
| 1490 | } |
| 1491 | } |
| 1492 | |
| 1493 | return new ShuffleVectorInst(X, Shuf->getOperand(i_nocapture: 1), NewMask); |
| 1494 | } |
| 1495 | |
| 1496 | /// If we have an insertelement instruction feeding into another insertelement |
| 1497 | /// and the 2nd is inserting a constant into the vector, canonicalize that |
| 1498 | /// constant insertion before the insertion of a variable: |
| 1499 | /// |
| 1500 | /// insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 --> |
| 1501 | /// insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1 |
| 1502 | /// |
| 1503 | /// This has the potential of eliminating the 2nd insertelement instruction |
| 1504 | /// via constant folding of the scalar constant into a vector constant. |
| 1505 | static Instruction *hoistInsEltConst(InsertElementInst &InsElt2, |
| 1506 | InstCombiner::BuilderTy &Builder) { |
| 1507 | auto *InsElt1 = dyn_cast<InsertElementInst>(Val: InsElt2.getOperand(i_nocapture: 0)); |
| 1508 | if (!InsElt1 || !InsElt1->hasOneUse()) |
| 1509 | return nullptr; |
| 1510 | |
| 1511 | Value *X, *Y; |
| 1512 | Constant *ScalarC; |
| 1513 | ConstantInt *IdxC1, *IdxC2; |
| 1514 | if (match(V: InsElt1->getOperand(i_nocapture: 0), P: m_Value(V&: X)) && |
| 1515 | match(V: InsElt1->getOperand(i_nocapture: 1), P: m_Value(V&: Y)) && !isa<Constant>(Val: Y) && |
| 1516 | match(V: InsElt1->getOperand(i_nocapture: 2), P: m_ConstantInt(CI&: IdxC1)) && |
| 1517 | match(V: InsElt2.getOperand(i_nocapture: 1), P: m_Constant(C&: ScalarC)) && |
| 1518 | match(V: InsElt2.getOperand(i_nocapture: 2), P: m_ConstantInt(CI&: IdxC2)) && IdxC1 != IdxC2) { |
| 1519 | Value *NewInsElt1 = Builder.CreateInsertElement(Vec: X, NewElt: ScalarC, Idx: IdxC2); |
| 1520 | return InsertElementInst::Create(Vec: NewInsElt1, NewElt: Y, Idx: IdxC1); |
| 1521 | } |
| 1522 | |
| 1523 | return nullptr; |
| 1524 | } |
| 1525 | |
| 1526 | /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex |
| 1527 | /// --> shufflevector X, CVec', Mask' |
| 1528 | static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) { |
| 1529 | auto *Inst = dyn_cast<Instruction>(Val: InsElt.getOperand(i_nocapture: 0)); |
| 1530 | // Bail out if the parent has more than one use. In that case, we'd be |
| 1531 | // replacing the insertelt with a shuffle, and that's not a clear win. |
| 1532 | if (!Inst || !Inst->hasOneUse()) |
| 1533 | return nullptr; |
| 1534 | if (auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: InsElt.getOperand(i_nocapture: 0))) { |
| 1535 | // The shuffle must have a constant vector operand. The insertelt must have |
| 1536 | // a constant scalar being inserted at a constant position in the vector. |
| 1537 | Constant *ShufConstVec, *InsEltScalar; |
| 1538 | uint64_t InsEltIndex; |
| 1539 | if (!match(V: Shuf->getOperand(i_nocapture: 1), P: m_Constant(C&: ShufConstVec)) || |
| 1540 | !match(V: InsElt.getOperand(i_nocapture: 1), P: m_Constant(C&: InsEltScalar)) || |
| 1541 | !match(V: InsElt.getOperand(i_nocapture: 2), P: m_ConstantInt(V&: InsEltIndex))) |
| 1542 | return nullptr; |
| 1543 | |
| 1544 | // Adding an element to an arbitrary shuffle could be expensive, but a |
| 1545 | // shuffle that selects elements from vectors without crossing lanes is |
| 1546 | // assumed cheap. |
| 1547 | // If we're just adding a constant into that shuffle, it will still be |
| 1548 | // cheap. |
| 1549 | if (!isShuffleEquivalentToSelect(Shuf&: *Shuf)) |
| 1550 | return nullptr; |
| 1551 | |
| 1552 | // From the above 'select' check, we know that the mask has the same number |
| 1553 | // of elements as the vector input operands. We also know that each constant |
| 1554 | // input element is used in its lane and can not be used more than once by |
| 1555 | // the shuffle. Therefore, replace the constant in the shuffle's constant |
| 1556 | // vector with the insertelt constant. Replace the constant in the shuffle's |
| 1557 | // mask vector with the insertelt index plus the length of the vector |
| 1558 | // (because the constant vector operand of a shuffle is always the 2nd |
| 1559 | // operand). |
| 1560 | ArrayRef<int> Mask = Shuf->getShuffleMask(); |
| 1561 | unsigned NumElts = Mask.size(); |
| 1562 | SmallVector<Constant *, 16> NewShufElts(NumElts); |
| 1563 | SmallVector<int, 16> NewMaskElts(NumElts); |
| 1564 | for (unsigned I = 0; I != NumElts; ++I) { |
| 1565 | if (I == InsEltIndex) { |
| 1566 | NewShufElts[I] = InsEltScalar; |
| 1567 | NewMaskElts[I] = InsEltIndex + NumElts; |
| 1568 | } else { |
| 1569 | // Copy over the existing values. |
| 1570 | NewShufElts[I] = ShufConstVec->getAggregateElement(Elt: I); |
| 1571 | NewMaskElts[I] = Mask[I]; |
| 1572 | } |
| 1573 | |
| 1574 | // Bail if we failed to find an element. |
| 1575 | if (!NewShufElts[I]) |
| 1576 | return nullptr; |
| 1577 | } |
| 1578 | |
| 1579 | // Create new operands for a shuffle that includes the constant of the |
| 1580 | // original insertelt. The old shuffle will be dead now. |
| 1581 | return new ShuffleVectorInst(Shuf->getOperand(i_nocapture: 0), |
| 1582 | ConstantVector::get(V: NewShufElts), NewMaskElts); |
| 1583 | } else if (auto *IEI = dyn_cast<InsertElementInst>(Val: Inst)) { |
| 1584 | // Transform sequences of insertelements ops with constant data/indexes into |
| 1585 | // a single shuffle op. |
| 1586 | // Can not handle scalable type, the number of elements needed to create |
| 1587 | // shuffle mask is not a compile-time constant. |
| 1588 | if (isa<ScalableVectorType>(Val: InsElt.getType())) |
| 1589 | return nullptr; |
| 1590 | unsigned NumElts = |
| 1591 | cast<FixedVectorType>(Val: InsElt.getType())->getNumElements(); |
| 1592 | |
| 1593 | uint64_t InsertIdx[2]; |
| 1594 | Constant *Val[2]; |
| 1595 | if (!match(V: InsElt.getOperand(i_nocapture: 2), P: m_ConstantInt(V&: InsertIdx[0])) || |
| 1596 | !match(V: InsElt.getOperand(i_nocapture: 1), P: m_Constant(C&: Val[0])) || |
| 1597 | !match(V: IEI->getOperand(i_nocapture: 2), P: m_ConstantInt(V&: InsertIdx[1])) || |
| 1598 | !match(V: IEI->getOperand(i_nocapture: 1), P: m_Constant(C&: Val[1]))) |
| 1599 | return nullptr; |
| 1600 | SmallVector<Constant *, 16> Values(NumElts); |
| 1601 | SmallVector<int, 16> Mask(NumElts); |
| 1602 | auto ValI = std::begin(arr&: Val); |
| 1603 | // Generate new constant vector and mask. |
| 1604 | // We have 2 values/masks from the insertelements instructions. Insert them |
| 1605 | // into new value/mask vectors. |
| 1606 | for (uint64_t I : InsertIdx) { |
| 1607 | if (!Values[I]) { |
| 1608 | Values[I] = *ValI; |
| 1609 | Mask[I] = NumElts + I; |
| 1610 | } |
| 1611 | ++ValI; |
| 1612 | } |
| 1613 | // Remaining values are filled with 'poison' values. |
| 1614 | for (unsigned I = 0; I < NumElts; ++I) { |
| 1615 | if (!Values[I]) { |
| 1616 | Values[I] = PoisonValue::get(T: InsElt.getType()->getElementType()); |
| 1617 | Mask[I] = I; |
| 1618 | } |
| 1619 | } |
| 1620 | // Create new operands for a shuffle that includes the constant of the |
| 1621 | // original insertelt. |
| 1622 | return new ShuffleVectorInst(IEI->getOperand(i_nocapture: 0), |
| 1623 | ConstantVector::get(V: Values), Mask); |
| 1624 | } |
| 1625 | return nullptr; |
| 1626 | } |
| 1627 | |
| 1628 | /// If both the base vector and the inserted element are extended from the same |
| 1629 | /// type, do the insert element in the narrow source type followed by extend. |
| 1630 | /// TODO: This can be extended to include other cast opcodes, but particularly |
| 1631 | /// if we create a wider insertelement, make sure codegen is not harmed. |
| 1632 | static Instruction *narrowInsElt(InsertElementInst &InsElt, |
| 1633 | InstCombiner::BuilderTy &Builder) { |
| 1634 | // We are creating a vector extend. If the original vector extend has another |
| 1635 | // use, that would mean we end up with 2 vector extends, so avoid that. |
| 1636 | // TODO: We could ease the use-clause to "if at least one op has one use" |
| 1637 | // (assuming that the source types match - see next TODO comment). |
| 1638 | Value *Vec = InsElt.getOperand(i_nocapture: 0); |
| 1639 | if (!Vec->hasOneUse()) |
| 1640 | return nullptr; |
| 1641 | |
| 1642 | Value *Scalar = InsElt.getOperand(i_nocapture: 1); |
| 1643 | Value *X, *Y; |
| 1644 | CastInst::CastOps CastOpcode; |
| 1645 | if (match(V: Vec, P: m_FPExt(Op: m_Value(V&: X))) && match(V: Scalar, P: m_FPExt(Op: m_Value(V&: Y)))) |
| 1646 | CastOpcode = Instruction::FPExt; |
| 1647 | else if (match(V: Vec, P: m_SExt(Op: m_Value(V&: X))) && match(V: Scalar, P: m_SExt(Op: m_Value(V&: Y)))) |
| 1648 | CastOpcode = Instruction::SExt; |
| 1649 | else if (match(V: Vec, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Scalar, P: m_ZExt(Op: m_Value(V&: Y)))) |
| 1650 | CastOpcode = Instruction::ZExt; |
| 1651 | else |
| 1652 | return nullptr; |
| 1653 | |
| 1654 | // TODO: We can allow mismatched types by creating an intermediate cast. |
| 1655 | if (X->getType()->getScalarType() != Y->getType()) |
| 1656 | return nullptr; |
| 1657 | |
| 1658 | // inselt (ext X), (ext Y), Index --> ext (inselt X, Y, Index) |
| 1659 | Value *NewInsElt = Builder.CreateInsertElement(Vec: X, NewElt: Y, Idx: InsElt.getOperand(i_nocapture: 2)); |
| 1660 | return CastInst::Create(CastOpcode, S: NewInsElt, Ty: InsElt.getType()); |
| 1661 | } |
| 1662 | |
| 1663 | /// If we are inserting 2 halves of a value into adjacent elements of a vector, |
| 1664 | /// try to convert to a single insert with appropriate bitcasts. |
| 1665 | static Instruction *foldTruncInsEltPair(InsertElementInst &InsElt, |
| 1666 | bool IsBigEndian, |
| 1667 | InstCombiner::BuilderTy &Builder) { |
| 1668 | Value *VecOp = InsElt.getOperand(i_nocapture: 0); |
| 1669 | Value *ScalarOp = InsElt.getOperand(i_nocapture: 1); |
| 1670 | Value *IndexOp = InsElt.getOperand(i_nocapture: 2); |
| 1671 | |
| 1672 | // Pattern depends on endian because we expect lower index is inserted first. |
| 1673 | // Big endian: |
| 1674 | // inselt (inselt BaseVec, (trunc (lshr X, BW/2), Index0), (trunc X), Index1 |
| 1675 | // Little endian: |
| 1676 | // inselt (inselt BaseVec, (trunc X), Index0), (trunc (lshr X, BW/2)), Index1 |
| 1677 | // Note: It is not safe to do this transform with an arbitrary base vector |
| 1678 | // because the bitcast of that vector to fewer/larger elements could |
| 1679 | // allow poison to spill into an element that was not poison before. |
| 1680 | // TODO: Detect smaller fractions of the scalar. |
| 1681 | // TODO: One-use checks are conservative. |
| 1682 | auto *VTy = dyn_cast<FixedVectorType>(Val: InsElt.getType()); |
| 1683 | Value *Scalar0, *BaseVec; |
| 1684 | uint64_t Index0, Index1; |
| 1685 | if (!VTy || (VTy->getNumElements() & 1) || |
| 1686 | !match(V: IndexOp, P: m_ConstantInt(V&: Index1)) || |
| 1687 | !match(V: VecOp, P: m_InsertElt(Val: m_Value(V&: BaseVec), Elt: m_Value(V&: Scalar0), |
| 1688 | Idx: m_ConstantInt(V&: Index0))) || |
| 1689 | !match(V: BaseVec, P: m_Undef())) |
| 1690 | return nullptr; |
| 1691 | |
| 1692 | // The first insert must be to the index one less than this one, and |
| 1693 | // the first insert must be to an even index. |
| 1694 | if (Index0 + 1 != Index1 || Index0 & 1) |
| 1695 | return nullptr; |
| 1696 | |
| 1697 | // For big endian, the high half of the value should be inserted first. |
| 1698 | // For little endian, the low half of the value should be inserted first. |
| 1699 | Value *X; |
| 1700 | uint64_t ShAmt; |
| 1701 | if (IsBigEndian) { |
| 1702 | if (!match(V: ScalarOp, P: m_Trunc(Op: m_Value(V&: X))) || |
| 1703 | !match(V: Scalar0, P: m_Trunc(Op: m_LShr(L: m_Specific(V: X), R: m_ConstantInt(V&: ShAmt))))) |
| 1704 | return nullptr; |
| 1705 | } else { |
| 1706 | if (!match(V: Scalar0, P: m_Trunc(Op: m_Value(V&: X))) || |
| 1707 | !match(V: ScalarOp, P: m_Trunc(Op: m_LShr(L: m_Specific(V: X), R: m_ConstantInt(V&: ShAmt))))) |
| 1708 | return nullptr; |
| 1709 | } |
| 1710 | |
| 1711 | Type *SrcTy = X->getType(); |
| 1712 | unsigned ScalarWidth = SrcTy->getScalarSizeInBits(); |
| 1713 | unsigned VecEltWidth = VTy->getScalarSizeInBits(); |
| 1714 | if (ScalarWidth != VecEltWidth * 2 || ShAmt != VecEltWidth) |
| 1715 | return nullptr; |
| 1716 | |
| 1717 | // Bitcast the base vector to a vector type with the source element type. |
| 1718 | Type *CastTy = FixedVectorType::get(ElementType: SrcTy, NumElts: VTy->getNumElements() / 2); |
| 1719 | Value *CastBaseVec = Builder.CreateBitCast(V: BaseVec, DestTy: CastTy); |
| 1720 | |
| 1721 | // Scale the insert index for a vector with half as many elements. |
| 1722 | // bitcast (inselt (bitcast BaseVec), X, NewIndex) |
| 1723 | uint64_t NewIndex = IsBigEndian ? Index1 / 2 : Index0 / 2; |
| 1724 | Value *NewInsert = Builder.CreateInsertElement(Vec: CastBaseVec, NewElt: X, Idx: NewIndex); |
| 1725 | return new BitCastInst(NewInsert, VTy); |
| 1726 | } |
| 1727 | |
| 1728 | Instruction *InstCombinerImpl::visitInsertElementInst(InsertElementInst &IE) { |
| 1729 | Value *VecOp = IE.getOperand(i_nocapture: 0); |
| 1730 | Value *ScalarOp = IE.getOperand(i_nocapture: 1); |
| 1731 | Value *IdxOp = IE.getOperand(i_nocapture: 2); |
| 1732 | |
| 1733 | if (auto *V = simplifyInsertElementInst( |
| 1734 | Vec: VecOp, Elt: ScalarOp, Idx: IdxOp, Q: SQ.getWithInstruction(I: &IE))) |
| 1735 | return replaceInstUsesWith(I&: IE, V); |
| 1736 | |
| 1737 | // Canonicalize type of constant indices to i64 to simplify CSE |
| 1738 | if (auto *IndexC = dyn_cast<ConstantInt>(Val: IdxOp)) { |
| 1739 | if (auto *NewIdx = getPreferredVectorIndex(IndexC)) |
| 1740 | return replaceOperand(I&: IE, OpNum: 2, V: NewIdx); |
| 1741 | |
| 1742 | Value *BaseVec, *OtherScalar; |
| 1743 | uint64_t OtherIndexVal; |
| 1744 | if (match(V: VecOp, P: m_OneUse(SubPattern: m_InsertElt(Val: m_Value(V&: BaseVec), |
| 1745 | Elt: m_Value(V&: OtherScalar), |
| 1746 | Idx: m_ConstantInt(V&: OtherIndexVal)))) && |
| 1747 | !isa<Constant>(Val: OtherScalar) && OtherIndexVal > IndexC->getZExtValue()) { |
| 1748 | Value *NewIns = Builder.CreateInsertElement(Vec: BaseVec, NewElt: ScalarOp, Idx: IdxOp); |
| 1749 | return InsertElementInst::Create(Vec: NewIns, NewElt: OtherScalar, |
| 1750 | Idx: Builder.getInt64(C: OtherIndexVal)); |
| 1751 | } |
| 1752 | } |
| 1753 | |
| 1754 | // If the scalar is bitcast and inserted into undef, do the insert in the |
| 1755 | // source type followed by bitcast. |
| 1756 | // TODO: Generalize for insert into any constant, not just undef? |
| 1757 | Value *ScalarSrc; |
| 1758 | if (match(V: VecOp, P: m_Undef()) && |
| 1759 | match(V: ScalarOp, P: m_OneUse(SubPattern: m_BitCast(Op: m_Value(V&: ScalarSrc)))) && |
| 1760 | (ScalarSrc->getType()->isIntegerTy() || |
| 1761 | ScalarSrc->getType()->isFloatingPointTy())) { |
| 1762 | // inselt undef, (bitcast ScalarSrc), IdxOp --> |
| 1763 | // bitcast (inselt undef, ScalarSrc, IdxOp) |
| 1764 | Type *ScalarTy = ScalarSrc->getType(); |
| 1765 | Type *VecTy = VectorType::get(ElementType: ScalarTy, EC: IE.getType()->getElementCount()); |
| 1766 | Constant *NewUndef = isa<PoisonValue>(Val: VecOp) ? PoisonValue::get(T: VecTy) |
| 1767 | : UndefValue::get(T: VecTy); |
| 1768 | Value *NewInsElt = Builder.CreateInsertElement(Vec: NewUndef, NewElt: ScalarSrc, Idx: IdxOp); |
| 1769 | return new BitCastInst(NewInsElt, IE.getType()); |
| 1770 | } |
| 1771 | |
| 1772 | // If the vector and scalar are both bitcast from the same element type, do |
| 1773 | // the insert in that source type followed by bitcast. |
| 1774 | Value *VecSrc; |
| 1775 | if (match(V: VecOp, P: m_BitCast(Op: m_Value(V&: VecSrc))) && |
| 1776 | match(V: ScalarOp, P: m_BitCast(Op: m_Value(V&: ScalarSrc))) && |
| 1777 | (VecOp->hasOneUse() || ScalarOp->hasOneUse()) && |
| 1778 | VecSrc->getType()->isVectorTy() && !ScalarSrc->getType()->isVectorTy() && |
| 1779 | cast<VectorType>(Val: VecSrc->getType())->getElementType() == |
| 1780 | ScalarSrc->getType()) { |
| 1781 | // inselt (bitcast VecSrc), (bitcast ScalarSrc), IdxOp --> |
| 1782 | // bitcast (inselt VecSrc, ScalarSrc, IdxOp) |
| 1783 | Value *NewInsElt = Builder.CreateInsertElement(Vec: VecSrc, NewElt: ScalarSrc, Idx: IdxOp); |
| 1784 | return new BitCastInst(NewInsElt, IE.getType()); |
| 1785 | } |
| 1786 | |
| 1787 | // If the inserted element was extracted from some other fixed-length vector |
| 1788 | // and both indexes are valid constants, try to turn this into a shuffle. |
| 1789 | // Can not handle scalable vector type, the number of elements needed to |
| 1790 | // create shuffle mask is not a compile-time constant. |
| 1791 | uint64_t InsertedIdx, ; |
| 1792 | Value *ExtVecOp; |
| 1793 | if (isa<FixedVectorType>(Val: IE.getType()) && |
| 1794 | match(V: IdxOp, P: m_ConstantInt(V&: InsertedIdx)) && |
| 1795 | match(V: ScalarOp, |
| 1796 | P: m_ExtractElt(Val: m_Value(V&: ExtVecOp), Idx: m_ConstantInt(V&: ExtractedIdx))) && |
| 1797 | isa<FixedVectorType>(Val: ExtVecOp->getType()) && |
| 1798 | ExtractedIdx < |
| 1799 | cast<FixedVectorType>(Val: ExtVecOp->getType())->getNumElements()) { |
| 1800 | // TODO: Looking at the user(s) to determine if this insert is a |
| 1801 | // fold-to-shuffle opportunity does not match the usual instcombine |
| 1802 | // constraints. We should decide if the transform is worthy based only |
| 1803 | // on this instruction and its operands, but that may not work currently. |
| 1804 | // |
| 1805 | // Here, we are trying to avoid creating shuffles before reaching |
| 1806 | // the end of a chain of extract-insert pairs. This is complicated because |
| 1807 | // we do not generally form arbitrary shuffle masks in instcombine |
| 1808 | // (because those may codegen poorly), but collectShuffleElements() does |
| 1809 | // exactly that. |
| 1810 | // |
| 1811 | // The rules for determining what is an acceptable target-independent |
| 1812 | // shuffle mask are fuzzy because they evolve based on the backend's |
| 1813 | // capabilities and real-world impact. |
| 1814 | auto isShuffleRootCandidate = [](InsertElementInst &Insert) { |
| 1815 | if (!Insert.hasOneUse()) |
| 1816 | return true; |
| 1817 | auto *InsertUser = dyn_cast<InsertElementInst>(Val: Insert.user_back()); |
| 1818 | if (!InsertUser) |
| 1819 | return true; |
| 1820 | return false; |
| 1821 | }; |
| 1822 | |
| 1823 | // Try to form a shuffle from a chain of extract-insert ops. |
| 1824 | if (isShuffleRootCandidate(IE)) { |
| 1825 | bool Rerun = true; |
| 1826 | while (Rerun) { |
| 1827 | Rerun = false; |
| 1828 | |
| 1829 | SmallVector<int, 16> Mask; |
| 1830 | ShuffleOps LR = |
| 1831 | collectShuffleElements(V: &IE, Mask, PermittedRHS: nullptr, IC&: *this, Rerun); |
| 1832 | |
| 1833 | // The proposed shuffle may be trivial, in which case we shouldn't |
| 1834 | // perform the combine. |
| 1835 | if (LR.first != &IE && LR.second != &IE) { |
| 1836 | // We now have a shuffle of LHS, RHS, Mask. |
| 1837 | if (LR.second == nullptr) |
| 1838 | LR.second = PoisonValue::get(T: LR.first->getType()); |
| 1839 | return new ShuffleVectorInst(LR.first, LR.second, Mask); |
| 1840 | } |
| 1841 | } |
| 1842 | } |
| 1843 | } |
| 1844 | |
| 1845 | if (auto VecTy = dyn_cast<FixedVectorType>(Val: VecOp->getType())) { |
| 1846 | unsigned VWidth = VecTy->getNumElements(); |
| 1847 | APInt PoisonElts(VWidth, 0); |
| 1848 | APInt AllOnesEltMask(APInt::getAllOnes(numBits: VWidth)); |
| 1849 | if (Value *V = SimplifyDemandedVectorElts(V: &IE, DemandedElts: AllOnesEltMask, |
| 1850 | PoisonElts)) { |
| 1851 | if (V != &IE) |
| 1852 | return replaceInstUsesWith(I&: IE, V); |
| 1853 | return &IE; |
| 1854 | } |
| 1855 | } |
| 1856 | |
| 1857 | if (Instruction *Shuf = foldConstantInsEltIntoShuffle(InsElt&: IE)) |
| 1858 | return Shuf; |
| 1859 | |
| 1860 | if (Instruction *NewInsElt = hoistInsEltConst(InsElt2&: IE, Builder)) |
| 1861 | return NewInsElt; |
| 1862 | |
| 1863 | if (Instruction *Broadcast = foldInsSequenceIntoSplat(InsElt&: IE)) |
| 1864 | return Broadcast; |
| 1865 | |
| 1866 | if (Instruction *Splat = foldInsEltIntoSplat(InsElt&: IE)) |
| 1867 | return Splat; |
| 1868 | |
| 1869 | if (Instruction *IdentityShuf = foldInsEltIntoIdentityShuffle(InsElt&: IE)) |
| 1870 | return IdentityShuf; |
| 1871 | |
| 1872 | if (Instruction *Ext = narrowInsElt(InsElt&: IE, Builder)) |
| 1873 | return Ext; |
| 1874 | |
| 1875 | if (Instruction *Ext = foldTruncInsEltPair(InsElt&: IE, IsBigEndian: DL.isBigEndian(), Builder)) |
| 1876 | return Ext; |
| 1877 | |
| 1878 | return nullptr; |
| 1879 | } |
| 1880 | |
| 1881 | /// Return true if we can evaluate the specified expression tree if the vector |
| 1882 | /// elements were shuffled in a different order. |
| 1883 | static bool canEvaluateShuffled(Value *V, ArrayRef<int> Mask, |
| 1884 | unsigned Depth = 5) { |
| 1885 | // We can always reorder the elements of a constant. |
| 1886 | if (isa<Constant>(Val: V)) |
| 1887 | return true; |
| 1888 | |
| 1889 | // We won't reorder vector arguments. No IPO here. |
| 1890 | Instruction *I = dyn_cast<Instruction>(Val: V); |
| 1891 | if (!I) return false; |
| 1892 | |
| 1893 | // Two users may expect different orders of the elements. Don't try it. |
| 1894 | if (!I->hasOneUse()) |
| 1895 | return false; |
| 1896 | |
| 1897 | if (Depth == 0) return false; |
| 1898 | |
| 1899 | switch (I->getOpcode()) { |
| 1900 | case Instruction::UDiv: |
| 1901 | case Instruction::SDiv: |
| 1902 | case Instruction::URem: |
| 1903 | case Instruction::SRem: |
| 1904 | // Propagating an undefined shuffle mask element to integer div/rem is not |
| 1905 | // allowed because those opcodes can create immediate undefined behavior |
| 1906 | // from an undefined element in an operand. |
| 1907 | if (llvm::is_contained(Range&: Mask, Element: -1)) |
| 1908 | return false; |
| 1909 | [[fallthrough]]; |
| 1910 | case Instruction::Add: |
| 1911 | case Instruction::FAdd: |
| 1912 | case Instruction::Sub: |
| 1913 | case Instruction::FSub: |
| 1914 | case Instruction::Mul: |
| 1915 | case Instruction::FMul: |
| 1916 | case Instruction::FDiv: |
| 1917 | case Instruction::FRem: |
| 1918 | case Instruction::Shl: |
| 1919 | case Instruction::LShr: |
| 1920 | case Instruction::AShr: |
| 1921 | case Instruction::And: |
| 1922 | case Instruction::Or: |
| 1923 | case Instruction::Xor: |
| 1924 | case Instruction::ICmp: |
| 1925 | case Instruction::FCmp: |
| 1926 | case Instruction::Trunc: |
| 1927 | case Instruction::ZExt: |
| 1928 | case Instruction::SExt: |
| 1929 | case Instruction::FPToUI: |
| 1930 | case Instruction::FPToSI: |
| 1931 | case Instruction::UIToFP: |
| 1932 | case Instruction::SIToFP: |
| 1933 | case Instruction::FPTrunc: |
| 1934 | case Instruction::FPExt: |
| 1935 | case Instruction::GetElementPtr: { |
| 1936 | // Bail out if we would create longer vector ops. We could allow creating |
| 1937 | // longer vector ops, but that may result in more expensive codegen. |
| 1938 | Type *ITy = I->getType(); |
| 1939 | if (ITy->isVectorTy() && |
| 1940 | Mask.size() > cast<FixedVectorType>(Val: ITy)->getNumElements()) |
| 1941 | return false; |
| 1942 | for (Value *Operand : I->operands()) { |
| 1943 | if (!canEvaluateShuffled(V: Operand, Mask, Depth: Depth - 1)) |
| 1944 | return false; |
| 1945 | } |
| 1946 | return true; |
| 1947 | } |
| 1948 | case Instruction::InsertElement: { |
| 1949 | ConstantInt *CI = dyn_cast<ConstantInt>(Val: I->getOperand(i: 2)); |
| 1950 | if (!CI) return false; |
| 1951 | int ElementNumber = CI->getLimitedValue(); |
| 1952 | |
| 1953 | // Verify that 'CI' does not occur twice in Mask. A single 'insertelement' |
| 1954 | // can't put an element into multiple indices. |
| 1955 | bool SeenOnce = false; |
| 1956 | for (int I : Mask) { |
| 1957 | if (I == ElementNumber) { |
| 1958 | if (SeenOnce) |
| 1959 | return false; |
| 1960 | SeenOnce = true; |
| 1961 | } |
| 1962 | } |
| 1963 | return canEvaluateShuffled(V: I->getOperand(i: 0), Mask, Depth: Depth - 1); |
| 1964 | } |
| 1965 | } |
| 1966 | return false; |
| 1967 | } |
| 1968 | |
| 1969 | /// Rebuild a new instruction just like 'I' but with the new operands given. |
| 1970 | /// In the event of type mismatch, the type of the operands is correct. |
| 1971 | static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps, |
| 1972 | IRBuilderBase &Builder) { |
| 1973 | Builder.SetInsertPoint(I); |
| 1974 | switch (I->getOpcode()) { |
| 1975 | case Instruction::Add: |
| 1976 | case Instruction::FAdd: |
| 1977 | case Instruction::Sub: |
| 1978 | case Instruction::FSub: |
| 1979 | case Instruction::Mul: |
| 1980 | case Instruction::FMul: |
| 1981 | case Instruction::UDiv: |
| 1982 | case Instruction::SDiv: |
| 1983 | case Instruction::FDiv: |
| 1984 | case Instruction::URem: |
| 1985 | case Instruction::SRem: |
| 1986 | case Instruction::FRem: |
| 1987 | case Instruction::Shl: |
| 1988 | case Instruction::LShr: |
| 1989 | case Instruction::AShr: |
| 1990 | case Instruction::And: |
| 1991 | case Instruction::Or: |
| 1992 | case Instruction::Xor: { |
| 1993 | BinaryOperator *BO = cast<BinaryOperator>(Val: I); |
| 1994 | assert(NewOps.size() == 2 && "binary operator with #ops != 2" ); |
| 1995 | Value *New = Builder.CreateBinOp(Opc: cast<BinaryOperator>(Val: I)->getOpcode(), |
| 1996 | LHS: NewOps[0], RHS: NewOps[1]); |
| 1997 | if (auto *NewI = dyn_cast<Instruction>(Val: New)) { |
| 1998 | if (isa<OverflowingBinaryOperator>(Val: BO)) { |
| 1999 | NewI->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap()); |
| 2000 | NewI->setHasNoSignedWrap(BO->hasNoSignedWrap()); |
| 2001 | } |
| 2002 | if (isa<PossiblyExactOperator>(Val: BO)) { |
| 2003 | NewI->setIsExact(BO->isExact()); |
| 2004 | } |
| 2005 | if (isa<FPMathOperator>(Val: BO)) |
| 2006 | NewI->copyFastMathFlags(I); |
| 2007 | } |
| 2008 | return New; |
| 2009 | } |
| 2010 | case Instruction::ICmp: { |
| 2011 | assert(NewOps.size() == 2 && "icmp with #ops != 2" ); |
| 2012 | Value *New = Builder.CreateICmp(P: cast<ICmpInst>(Val: I)->getPredicate(), |
| 2013 | LHS: NewOps[0], RHS: NewOps[1]); |
| 2014 | if (auto *NewI = dyn_cast<Instruction>(Val: New)) |
| 2015 | NewI->copyIRFlags(V: I); |
| 2016 | return New; |
| 2017 | } |
| 2018 | case Instruction::FCmp: |
| 2019 | assert(NewOps.size() == 2 && "fcmp with #ops != 2" ); |
| 2020 | return Builder.CreateFCmpFMF(P: cast<FCmpInst>(Val: I)->getPredicate(), LHS: NewOps[0], |
| 2021 | RHS: NewOps[1], FMFSource: I); |
| 2022 | case Instruction::Trunc: |
| 2023 | case Instruction::ZExt: |
| 2024 | case Instruction::SExt: |
| 2025 | case Instruction::FPToUI: |
| 2026 | case Instruction::FPToSI: |
| 2027 | case Instruction::UIToFP: |
| 2028 | case Instruction::SIToFP: |
| 2029 | case Instruction::FPTrunc: |
| 2030 | case Instruction::FPExt: { |
| 2031 | // It's possible that the mask has a different number of elements from |
| 2032 | // the original cast. We recompute the destination type to match the mask. |
| 2033 | Type *DestTy = VectorType::get( |
| 2034 | ElementType: I->getType()->getScalarType(), |
| 2035 | EC: cast<VectorType>(Val: NewOps[0]->getType())->getElementCount()); |
| 2036 | assert(NewOps.size() == 1 && "cast with #ops != 1" ); |
| 2037 | Value *New = |
| 2038 | Builder.CreateCast(Op: cast<CastInst>(Val: I)->getOpcode(), V: NewOps[0], DestTy); |
| 2039 | if (auto *NewI = dyn_cast<Instruction>(Val: New)) |
| 2040 | NewI->copyIRFlags(V: I); |
| 2041 | return New; |
| 2042 | } |
| 2043 | case Instruction::GetElementPtr: { |
| 2044 | Value *Ptr = NewOps[0]; |
| 2045 | ArrayRef<Value*> Idx = NewOps.slice(N: 1); |
| 2046 | return Builder.CreateGEP(Ty: cast<GEPOperator>(Val: I)->getSourceElementType(), |
| 2047 | Ptr, IdxList: Idx, Name: "" , |
| 2048 | NW: cast<GEPOperator>(Val: I)->getNoWrapFlags()); |
| 2049 | } |
| 2050 | } |
| 2051 | llvm_unreachable("failed to rebuild vector instructions" ); |
| 2052 | } |
| 2053 | |
| 2054 | static Value *evaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask, |
| 2055 | IRBuilderBase &Builder) { |
| 2056 | // Mask.size() does not need to be equal to the number of vector elements. |
| 2057 | |
| 2058 | assert(V->getType()->isVectorTy() && "can't reorder non-vector elements" ); |
| 2059 | Type *EltTy = V->getType()->getScalarType(); |
| 2060 | |
| 2061 | if (isa<PoisonValue>(Val: V)) |
| 2062 | return PoisonValue::get(T: FixedVectorType::get(ElementType: EltTy, NumElts: Mask.size())); |
| 2063 | |
| 2064 | if (match(V, P: m_Undef())) |
| 2065 | return UndefValue::get(T: FixedVectorType::get(ElementType: EltTy, NumElts: Mask.size())); |
| 2066 | |
| 2067 | if (isa<ConstantAggregateZero>(Val: V)) |
| 2068 | return ConstantAggregateZero::get(Ty: FixedVectorType::get(ElementType: EltTy, NumElts: Mask.size())); |
| 2069 | |
| 2070 | if (Constant *C = dyn_cast<Constant>(Val: V)) |
| 2071 | return ConstantExpr::getShuffleVector(V1: C, V2: PoisonValue::get(T: C->getType()), |
| 2072 | Mask); |
| 2073 | |
| 2074 | Instruction *I = cast<Instruction>(Val: V); |
| 2075 | switch (I->getOpcode()) { |
| 2076 | case Instruction::Add: |
| 2077 | case Instruction::FAdd: |
| 2078 | case Instruction::Sub: |
| 2079 | case Instruction::FSub: |
| 2080 | case Instruction::Mul: |
| 2081 | case Instruction::FMul: |
| 2082 | case Instruction::UDiv: |
| 2083 | case Instruction::SDiv: |
| 2084 | case Instruction::FDiv: |
| 2085 | case Instruction::URem: |
| 2086 | case Instruction::SRem: |
| 2087 | case Instruction::FRem: |
| 2088 | case Instruction::Shl: |
| 2089 | case Instruction::LShr: |
| 2090 | case Instruction::AShr: |
| 2091 | case Instruction::And: |
| 2092 | case Instruction::Or: |
| 2093 | case Instruction::Xor: |
| 2094 | case Instruction::ICmp: |
| 2095 | case Instruction::FCmp: |
| 2096 | case Instruction::Trunc: |
| 2097 | case Instruction::ZExt: |
| 2098 | case Instruction::SExt: |
| 2099 | case Instruction::FPToUI: |
| 2100 | case Instruction::FPToSI: |
| 2101 | case Instruction::UIToFP: |
| 2102 | case Instruction::SIToFP: |
| 2103 | case Instruction::FPTrunc: |
| 2104 | case Instruction::FPExt: |
| 2105 | case Instruction::Select: |
| 2106 | case Instruction::GetElementPtr: { |
| 2107 | SmallVector<Value*, 8> NewOps; |
| 2108 | bool NeedsRebuild = |
| 2109 | (Mask.size() != |
| 2110 | cast<FixedVectorType>(Val: I->getType())->getNumElements()); |
| 2111 | for (int i = 0, e = I->getNumOperands(); i != e; ++i) { |
| 2112 | Value *V; |
| 2113 | // Recursively call evaluateInDifferentElementOrder on vector arguments |
| 2114 | // as well. E.g. GetElementPtr may have scalar operands even if the |
| 2115 | // return value is a vector, so we need to examine the operand type. |
| 2116 | if (I->getOperand(i)->getType()->isVectorTy()) |
| 2117 | V = evaluateInDifferentElementOrder(V: I->getOperand(i), Mask, Builder); |
| 2118 | else |
| 2119 | V = I->getOperand(i); |
| 2120 | NewOps.push_back(Elt: V); |
| 2121 | NeedsRebuild |= (V != I->getOperand(i)); |
| 2122 | } |
| 2123 | if (NeedsRebuild) |
| 2124 | return buildNew(I, NewOps, Builder); |
| 2125 | return I; |
| 2126 | } |
| 2127 | case Instruction::InsertElement: { |
| 2128 | int Element = cast<ConstantInt>(Val: I->getOperand(i: 2))->getLimitedValue(); |
| 2129 | |
| 2130 | // The insertelement was inserting at Element. Figure out which element |
| 2131 | // that becomes after shuffling. The answer is guaranteed to be unique |
| 2132 | // by CanEvaluateShuffled. |
| 2133 | bool Found = false; |
| 2134 | int Index = 0; |
| 2135 | for (int e = Mask.size(); Index != e; ++Index) { |
| 2136 | if (Mask[Index] == Element) { |
| 2137 | Found = true; |
| 2138 | break; |
| 2139 | } |
| 2140 | } |
| 2141 | |
| 2142 | // If element is not in Mask, no need to handle the operand 1 (element to |
| 2143 | // be inserted). Just evaluate values in operand 0 according to Mask. |
| 2144 | if (!Found) |
| 2145 | return evaluateInDifferentElementOrder(V: I->getOperand(i: 0), Mask, Builder); |
| 2146 | |
| 2147 | Value *V = evaluateInDifferentElementOrder(V: I->getOperand(i: 0), Mask, |
| 2148 | Builder); |
| 2149 | Builder.SetInsertPoint(I); |
| 2150 | return Builder.CreateInsertElement(Vec: V, NewElt: I->getOperand(i: 1), Idx: Index); |
| 2151 | } |
| 2152 | } |
| 2153 | llvm_unreachable("failed to reorder elements of vector instruction!" ); |
| 2154 | } |
| 2155 | |
| 2156 | // Returns true if the shuffle is extracting a contiguous range of values from |
| 2157 | // LHS, for example: |
| 2158 | // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ |
| 2159 | // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP| |
| 2160 | // Shuffles to: |EE|FF|GG|HH| |
| 2161 | // +--+--+--+--+ |
| 2162 | static bool (ShuffleVectorInst &SVI, |
| 2163 | ArrayRef<int> Mask) { |
| 2164 | unsigned LHSElems = |
| 2165 | cast<FixedVectorType>(Val: SVI.getOperand(i_nocapture: 0)->getType())->getNumElements(); |
| 2166 | unsigned MaskElems = Mask.size(); |
| 2167 | unsigned BegIdx = Mask.front(); |
| 2168 | unsigned EndIdx = Mask.back(); |
| 2169 | if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1) |
| 2170 | return false; |
| 2171 | for (unsigned I = 0; I != MaskElems; ++I) |
| 2172 | if (static_cast<unsigned>(Mask[I]) != BegIdx + I) |
| 2173 | return false; |
| 2174 | return true; |
| 2175 | } |
| 2176 | |
| 2177 | /// These are the ingredients in an alternate form binary operator as described |
| 2178 | /// below. |
| 2179 | struct BinopElts { |
| 2180 | BinaryOperator::BinaryOps Opcode; |
| 2181 | Value *Op0; |
| 2182 | Value *Op1; |
| 2183 | BinopElts(BinaryOperator::BinaryOps Opc = (BinaryOperator::BinaryOps)0, |
| 2184 | Value *V0 = nullptr, Value *V1 = nullptr) : |
| 2185 | Opcode(Opc), Op0(V0), Op1(V1) {} |
| 2186 | operator bool() const { return Opcode != 0; } |
| 2187 | }; |
| 2188 | |
| 2189 | /// Binops may be transformed into binops with different opcodes and operands. |
| 2190 | /// Reverse the usual canonicalization to enable folds with the non-canonical |
| 2191 | /// form of the binop. If a transform is possible, return the elements of the |
| 2192 | /// new binop. If not, return invalid elements. |
| 2193 | static BinopElts getAlternateBinop(BinaryOperator *BO, const DataLayout &DL) { |
| 2194 | Value *BO0 = BO->getOperand(i_nocapture: 0), *BO1 = BO->getOperand(i_nocapture: 1); |
| 2195 | Type *Ty = BO->getType(); |
| 2196 | switch (BO->getOpcode()) { |
| 2197 | case Instruction::Shl: { |
| 2198 | // shl X, C --> mul X, (1 << C) |
| 2199 | Constant *C; |
| 2200 | if (match(V: BO1, P: m_ImmConstant(C))) { |
| 2201 | Constant *ShlOne = ConstantFoldBinaryOpOperands( |
| 2202 | Opcode: Instruction::Shl, LHS: ConstantInt::get(Ty, V: 1), RHS: C, DL); |
| 2203 | assert(ShlOne && "Constant folding of immediate constants failed" ); |
| 2204 | return {Instruction::Mul, BO0, ShlOne}; |
| 2205 | } |
| 2206 | break; |
| 2207 | } |
| 2208 | case Instruction::Or: { |
| 2209 | // or disjoin X, C --> add X, C |
| 2210 | if (cast<PossiblyDisjointInst>(Val: BO)->isDisjoint()) |
| 2211 | return {Instruction::Add, BO0, BO1}; |
| 2212 | break; |
| 2213 | } |
| 2214 | case Instruction::Sub: |
| 2215 | // sub 0, X --> mul X, -1 |
| 2216 | if (match(V: BO0, P: m_ZeroInt())) |
| 2217 | return {Instruction::Mul, BO1, ConstantInt::getAllOnesValue(Ty)}; |
| 2218 | break; |
| 2219 | default: |
| 2220 | break; |
| 2221 | } |
| 2222 | return {}; |
| 2223 | } |
| 2224 | |
| 2225 | /// A select shuffle of a select shuffle with a shared operand can be reduced |
| 2226 | /// to a single select shuffle. This is an obvious improvement in IR, and the |
| 2227 | /// backend is expected to lower select shuffles efficiently. |
| 2228 | static Instruction *foldSelectShuffleOfSelectShuffle(ShuffleVectorInst &Shuf) { |
| 2229 | assert(Shuf.isSelect() && "Must have select-equivalent shuffle" ); |
| 2230 | |
| 2231 | Value *Op0 = Shuf.getOperand(i_nocapture: 0), *Op1 = Shuf.getOperand(i_nocapture: 1); |
| 2232 | SmallVector<int, 16> Mask; |
| 2233 | Shuf.getShuffleMask(Result&: Mask); |
| 2234 | unsigned NumElts = Mask.size(); |
| 2235 | |
| 2236 | // Canonicalize a select shuffle with common operand as Op1. |
| 2237 | auto *ShufOp = dyn_cast<ShuffleVectorInst>(Val: Op0); |
| 2238 | if (ShufOp && ShufOp->isSelect() && |
| 2239 | (ShufOp->getOperand(i_nocapture: 0) == Op1 || ShufOp->getOperand(i_nocapture: 1) == Op1)) { |
| 2240 | std::swap(a&: Op0, b&: Op1); |
| 2241 | ShuffleVectorInst::commuteShuffleMask(Mask, InVecNumElts: NumElts); |
| 2242 | } |
| 2243 | |
| 2244 | ShufOp = dyn_cast<ShuffleVectorInst>(Val: Op1); |
| 2245 | if (!ShufOp || !ShufOp->isSelect() || |
| 2246 | (ShufOp->getOperand(i_nocapture: 0) != Op0 && ShufOp->getOperand(i_nocapture: 1) != Op0)) |
| 2247 | return nullptr; |
| 2248 | |
| 2249 | Value *X = ShufOp->getOperand(i_nocapture: 0), *Y = ShufOp->getOperand(i_nocapture: 1); |
| 2250 | SmallVector<int, 16> Mask1; |
| 2251 | ShufOp->getShuffleMask(Result&: Mask1); |
| 2252 | assert(Mask1.size() == NumElts && "Vector size changed with select shuffle" ); |
| 2253 | |
| 2254 | // Canonicalize common operand (Op0) as X (first operand of first shuffle). |
| 2255 | if (Y == Op0) { |
| 2256 | std::swap(a&: X, b&: Y); |
| 2257 | ShuffleVectorInst::commuteShuffleMask(Mask: Mask1, InVecNumElts: NumElts); |
| 2258 | } |
| 2259 | |
| 2260 | // If the mask chooses from X (operand 0), it stays the same. |
| 2261 | // If the mask chooses from the earlier shuffle, the other mask value is |
| 2262 | // transferred to the combined select shuffle: |
| 2263 | // shuf X, (shuf X, Y, M1), M --> shuf X, Y, M' |
| 2264 | SmallVector<int, 16> NewMask(NumElts); |
| 2265 | for (unsigned i = 0; i != NumElts; ++i) |
| 2266 | NewMask[i] = Mask[i] < (signed)NumElts ? Mask[i] : Mask1[i]; |
| 2267 | |
| 2268 | // A select mask with undef elements might look like an identity mask. |
| 2269 | assert((ShuffleVectorInst::isSelectMask(NewMask, NumElts) || |
| 2270 | ShuffleVectorInst::isIdentityMask(NewMask, NumElts)) && |
| 2271 | "Unexpected shuffle mask" ); |
| 2272 | return new ShuffleVectorInst(X, Y, NewMask); |
| 2273 | } |
| 2274 | |
| 2275 | static Instruction *foldSelectShuffleWith1Binop(ShuffleVectorInst &Shuf, |
| 2276 | const SimplifyQuery &SQ) { |
| 2277 | assert(Shuf.isSelect() && "Must have select-equivalent shuffle" ); |
| 2278 | |
| 2279 | // Are we shuffling together some value and that same value after it has been |
| 2280 | // modified by a binop with a constant? |
| 2281 | Value *Op0 = Shuf.getOperand(i_nocapture: 0), *Op1 = Shuf.getOperand(i_nocapture: 1); |
| 2282 | Constant *C; |
| 2283 | bool Op0IsBinop; |
| 2284 | if (match(V: Op0, P: m_BinOp(L: m_Specific(V: Op1), R: m_Constant(C)))) |
| 2285 | Op0IsBinop = true; |
| 2286 | else if (match(V: Op1, P: m_BinOp(L: m_Specific(V: Op0), R: m_Constant(C)))) |
| 2287 | Op0IsBinop = false; |
| 2288 | else |
| 2289 | return nullptr; |
| 2290 | |
| 2291 | // The identity constant for a binop leaves a variable operand unchanged. For |
| 2292 | // a vector, this is a splat of something like 0, -1, or 1. |
| 2293 | // If there's no identity constant for this binop, we're done. |
| 2294 | auto *BO = cast<BinaryOperator>(Val: Op0IsBinop ? Op0 : Op1); |
| 2295 | BinaryOperator::BinaryOps BOpcode = BO->getOpcode(); |
| 2296 | Constant *IdC = ConstantExpr::getBinOpIdentity(Opcode: BOpcode, Ty: Shuf.getType(), AllowRHSConstant: true); |
| 2297 | if (!IdC) |
| 2298 | return nullptr; |
| 2299 | |
| 2300 | Value *X = Op0IsBinop ? Op1 : Op0; |
| 2301 | |
| 2302 | // Prevent folding in the case the non-binop operand might have NaN values. |
| 2303 | // If X can have NaN elements then we have that the floating point math |
| 2304 | // operation in the transformed code may not preserve the exact NaN |
| 2305 | // bit-pattern -- e.g. `fadd sNaN, 0.0 -> qNaN`. |
| 2306 | // This makes the transformation incorrect since the original program would |
| 2307 | // have preserved the exact NaN bit-pattern. |
| 2308 | // Avoid the folding if X can have NaN elements. |
| 2309 | bool IsFloatingPointTy = |
| 2310 | Shuf.getType()->getElementType()->isFloatingPointTy(); |
| 2311 | if (IsFloatingPointTy && !isKnownNeverNaN(V: X, SQ)) |
| 2312 | return nullptr; |
| 2313 | |
| 2314 | // Shuffle identity constants into the lanes that return the original value. |
| 2315 | // Example: shuf (mul X, {-1,-2,-3,-4}), X, {0,5,6,3} --> mul X, {-1,1,1,-4} |
| 2316 | // Example: shuf X, (add X, {-1,-2,-3,-4}), {0,1,6,7} --> add X, {0,0,-3,-4} |
| 2317 | // The existing binop constant vector remains in the same operand position. |
| 2318 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
| 2319 | Constant *NewC = Op0IsBinop ? ConstantExpr::getShuffleVector(V1: C, V2: IdC, Mask) : |
| 2320 | ConstantExpr::getShuffleVector(V1: IdC, V2: C, Mask); |
| 2321 | |
| 2322 | bool MightCreatePoisonOrUB = |
| 2323 | is_contained(Range&: Mask, Element: PoisonMaskElem) && |
| 2324 | (Instruction::isIntDivRem(Opcode: BOpcode) || Instruction::isShift(Opcode: BOpcode)); |
| 2325 | if (MightCreatePoisonOrUB) |
| 2326 | NewC = InstCombiner::getSafeVectorConstantForBinop(Opcode: BOpcode, In: NewC, IsRHSConstant: true); |
| 2327 | |
| 2328 | // shuf (bop X, C), X, M --> bop X, C' |
| 2329 | // shuf X, (bop X, C), M --> bop X, C' |
| 2330 | BinaryOperator *NewBO = BinaryOperator::Create(Op: BOpcode, S1: X, S2: NewC); |
| 2331 | NewBO->copyIRFlags(V: BO); |
| 2332 | |
| 2333 | // Drop noinf FMF if X can be Inf. If X can have Inf elements and noinf FMF is |
| 2334 | // set, the transformation may generate poison where the original program |
| 2335 | // would preserve the Inf value. |
| 2336 | if (IsFloatingPointTy && NewBO->hasNoInfs() && !isKnownNeverInfinity(V: X, SQ)) |
| 2337 | NewBO->setHasNoInfs(false); |
| 2338 | |
| 2339 | // An undef shuffle mask element may propagate as an undef constant element in |
| 2340 | // the new binop. That would produce poison where the original code might not. |
| 2341 | // If we already made a safe constant, then there's no danger. |
| 2342 | if (is_contained(Range&: Mask, Element: PoisonMaskElem) && !MightCreatePoisonOrUB) |
| 2343 | NewBO->dropPoisonGeneratingFlags(); |
| 2344 | return NewBO; |
| 2345 | } |
| 2346 | |
| 2347 | /// If we have an insert of a scalar to a non-zero element of an undefined |
| 2348 | /// vector and then shuffle that value, that's the same as inserting to the zero |
| 2349 | /// element and shuffling. Splatting from the zero element is recognized as the |
| 2350 | /// canonical form of splat. |
| 2351 | static Instruction *canonicalizeInsertSplat(ShuffleVectorInst &Shuf, |
| 2352 | InstCombiner::BuilderTy &Builder) { |
| 2353 | Value *Op0 = Shuf.getOperand(i_nocapture: 0), *Op1 = Shuf.getOperand(i_nocapture: 1); |
| 2354 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
| 2355 | Value *X; |
| 2356 | uint64_t IndexC; |
| 2357 | |
| 2358 | // Match a shuffle that is a splat to a non-zero element. |
| 2359 | if (!match(V: Op0, P: m_OneUse(SubPattern: m_InsertElt(Val: m_Poison(), Elt: m_Value(V&: X), |
| 2360 | Idx: m_ConstantInt(V&: IndexC)))) || |
| 2361 | !match(V: Op1, P: m_Poison()) || match(Mask, P: m_ZeroMask()) || IndexC == 0) |
| 2362 | return nullptr; |
| 2363 | |
| 2364 | // Insert into element 0 of a poison vector. |
| 2365 | PoisonValue *PoisonVec = PoisonValue::get(T: Shuf.getType()); |
| 2366 | Value *NewIns = Builder.CreateInsertElement(Vec: PoisonVec, NewElt: X, Idx: (uint64_t)0); |
| 2367 | |
| 2368 | // Splat from element 0. Any mask element that is poison remains poison. |
| 2369 | // For example: |
| 2370 | // shuf (inselt poison, X, 2), _, <2,2,undef> |
| 2371 | // --> shuf (inselt poison, X, 0), poison, <0,0,undef> |
| 2372 | unsigned NumMaskElts = |
| 2373 | cast<FixedVectorType>(Val: Shuf.getType())->getNumElements(); |
| 2374 | SmallVector<int, 16> NewMask(NumMaskElts, 0); |
| 2375 | for (unsigned i = 0; i != NumMaskElts; ++i) |
| 2376 | if (Mask[i] == PoisonMaskElem) |
| 2377 | NewMask[i] = Mask[i]; |
| 2378 | |
| 2379 | return new ShuffleVectorInst(NewIns, NewMask); |
| 2380 | } |
| 2381 | |
| 2382 | /// Try to fold shuffles that are the equivalent of a vector select. |
| 2383 | Instruction *InstCombinerImpl::foldSelectShuffle(ShuffleVectorInst &Shuf) { |
| 2384 | if (!Shuf.isSelect()) |
| 2385 | return nullptr; |
| 2386 | |
| 2387 | // Canonicalize to choose from operand 0 first unless operand 1 is undefined. |
| 2388 | // Commuting undef to operand 0 conflicts with another canonicalization. |
| 2389 | unsigned NumElts = cast<FixedVectorType>(Val: Shuf.getType())->getNumElements(); |
| 2390 | if (!match(V: Shuf.getOperand(i_nocapture: 1), P: m_Undef()) && |
| 2391 | Shuf.getMaskValue(Elt: 0) >= (int)NumElts) { |
| 2392 | // TODO: Can we assert that both operands of a shuffle-select are not undef |
| 2393 | // (otherwise, it would have been folded by instsimplify? |
| 2394 | Shuf.commute(); |
| 2395 | return &Shuf; |
| 2396 | } |
| 2397 | |
| 2398 | if (Instruction *I = foldSelectShuffleOfSelectShuffle(Shuf)) |
| 2399 | return I; |
| 2400 | |
| 2401 | if (Instruction *I = foldSelectShuffleWith1Binop( |
| 2402 | Shuf, SQ: getSimplifyQuery().getWithInstruction(I: &Shuf))) |
| 2403 | return I; |
| 2404 | |
| 2405 | BinaryOperator *B0, *B1; |
| 2406 | if (!match(V: Shuf.getOperand(i_nocapture: 0), P: m_BinOp(I&: B0)) || |
| 2407 | !match(V: Shuf.getOperand(i_nocapture: 1), P: m_BinOp(I&: B1))) |
| 2408 | return nullptr; |
| 2409 | |
| 2410 | // If one operand is "0 - X", allow that to be viewed as "X * -1" |
| 2411 | // (ConstantsAreOp1) by getAlternateBinop below. If the neg is not paired |
| 2412 | // with a multiply, we will exit because C0/C1 will not be set. |
| 2413 | Value *X, *Y; |
| 2414 | Constant *C0 = nullptr, *C1 = nullptr; |
| 2415 | bool ConstantsAreOp1; |
| 2416 | if (match(V: B0, P: m_BinOp(L: m_Constant(C&: C0), R: m_Value(V&: X))) && |
| 2417 | match(V: B1, P: m_BinOp(L: m_Constant(C&: C1), R: m_Value(V&: Y)))) |
| 2418 | ConstantsAreOp1 = false; |
| 2419 | else if (match(V: B0, P: m_CombineOr(Ps: m_BinOp(L: m_Value(V&: X), R: m_Constant(C&: C0)), |
| 2420 | Ps: m_Neg(V: m_Value(V&: X)))) && |
| 2421 | match(V: B1, P: m_CombineOr(Ps: m_BinOp(L: m_Value(V&: Y), R: m_Constant(C&: C1)), |
| 2422 | Ps: m_Neg(V: m_Value(V&: Y))))) |
| 2423 | ConstantsAreOp1 = true; |
| 2424 | else |
| 2425 | return nullptr; |
| 2426 | |
| 2427 | // We need matching binops to fold the lanes together. |
| 2428 | BinaryOperator::BinaryOps Opc0 = B0->getOpcode(); |
| 2429 | BinaryOperator::BinaryOps Opc1 = B1->getOpcode(); |
| 2430 | bool DropNSW = false; |
| 2431 | if (ConstantsAreOp1 && Opc0 != Opc1) { |
| 2432 | // TODO: We drop "nsw" if shift is converted into multiply because it may |
| 2433 | // not be correct when the shift amount is BitWidth - 1. We could examine |
| 2434 | // each vector element to determine if it is safe to keep that flag. |
| 2435 | if (Opc0 == Instruction::Shl || Opc1 == Instruction::Shl) |
| 2436 | DropNSW = true; |
| 2437 | if (BinopElts AltB0 = getAlternateBinop(BO: B0, DL)) { |
| 2438 | assert(isa<Constant>(AltB0.Op1) && "Expecting constant with alt binop" ); |
| 2439 | Opc0 = AltB0.Opcode; |
| 2440 | C0 = cast<Constant>(Val: AltB0.Op1); |
| 2441 | } else if (BinopElts AltB1 = getAlternateBinop(BO: B1, DL)) { |
| 2442 | assert(isa<Constant>(AltB1.Op1) && "Expecting constant with alt binop" ); |
| 2443 | Opc1 = AltB1.Opcode; |
| 2444 | C1 = cast<Constant>(Val: AltB1.Op1); |
| 2445 | } |
| 2446 | } |
| 2447 | |
| 2448 | if (Opc0 != Opc1 || !C0 || !C1) |
| 2449 | return nullptr; |
| 2450 | |
| 2451 | // The opcodes must be the same. Use a new name to make that clear. |
| 2452 | BinaryOperator::BinaryOps BOpc = Opc0; |
| 2453 | |
| 2454 | // Select the constant elements needed for the single binop. |
| 2455 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
| 2456 | Constant *NewC = ConstantExpr::getShuffleVector(V1: C0, V2: C1, Mask); |
| 2457 | |
| 2458 | // We are moving a binop after a shuffle. When a shuffle has an undefined |
| 2459 | // mask element, the result is undefined, but it is not poison or undefined |
| 2460 | // behavior. That is not necessarily true for div/rem/shift. |
| 2461 | bool MightCreatePoisonOrUB = |
| 2462 | is_contained(Range&: Mask, Element: PoisonMaskElem) && |
| 2463 | (Instruction::isIntDivRem(Opcode: BOpc) || Instruction::isShift(Opcode: BOpc)); |
| 2464 | if (MightCreatePoisonOrUB) |
| 2465 | NewC = InstCombiner::getSafeVectorConstantForBinop(Opcode: BOpc, In: NewC, |
| 2466 | IsRHSConstant: ConstantsAreOp1); |
| 2467 | |
| 2468 | Value *V; |
| 2469 | if (X == Y) { |
| 2470 | // Remove a binop and the shuffle by rearranging the constant: |
| 2471 | // shuffle (op V, C0), (op V, C1), M --> op V, C' |
| 2472 | // shuffle (op C0, V), (op C1, V), M --> op C', V |
| 2473 | V = X; |
| 2474 | } else { |
| 2475 | // If there are 2 different variable operands, we must create a new shuffle |
| 2476 | // (select) first, so check uses to ensure that we don't end up with more |
| 2477 | // instructions than we started with. |
| 2478 | if (!B0->hasOneUse() && !B1->hasOneUse()) |
| 2479 | return nullptr; |
| 2480 | |
| 2481 | // If we use the original shuffle mask and op1 is *variable*, we would be |
| 2482 | // putting an undef into operand 1 of div/rem/shift. This is either UB or |
| 2483 | // poison. We do not have to guard against UB when *constants* are op1 |
| 2484 | // because safe constants guarantee that we do not overflow sdiv/srem (and |
| 2485 | // there's no danger for other opcodes). |
| 2486 | // TODO: To allow this case, create a new shuffle mask with no undefs. |
| 2487 | if (MightCreatePoisonOrUB && !ConstantsAreOp1) |
| 2488 | return nullptr; |
| 2489 | |
| 2490 | // Note: In general, we do not create new shuffles in InstCombine because we |
| 2491 | // do not know if a target can lower an arbitrary shuffle optimally. In this |
| 2492 | // case, the shuffle uses the existing mask, so there is no additional risk. |
| 2493 | |
| 2494 | // Select the variable vectors first, then perform the binop: |
| 2495 | // shuffle (op X, C0), (op Y, C1), M --> op (shuffle X, Y, M), C' |
| 2496 | // shuffle (op C0, X), (op C1, Y), M --> op C', (shuffle X, Y, M) |
| 2497 | V = Builder.CreateShuffleVector(V1: X, V2: Y, Mask); |
| 2498 | } |
| 2499 | |
| 2500 | Value *NewBO = ConstantsAreOp1 ? Builder.CreateBinOp(Opc: BOpc, LHS: V, RHS: NewC) : |
| 2501 | Builder.CreateBinOp(Opc: BOpc, LHS: NewC, RHS: V); |
| 2502 | |
| 2503 | // Flags are intersected from the 2 source binops. But there are 2 exceptions: |
| 2504 | // 1. If we changed an opcode, poison conditions might have changed. |
| 2505 | // 2. If the shuffle had undef mask elements, the new binop might have undefs |
| 2506 | // where the original code did not. But if we already made a safe constant, |
| 2507 | // then there's no danger. |
| 2508 | if (auto *NewI = dyn_cast<Instruction>(Val: NewBO)) { |
| 2509 | NewI->copyIRFlags(V: B0); |
| 2510 | NewI->andIRFlags(V: B1); |
| 2511 | if (DropNSW) |
| 2512 | NewI->setHasNoSignedWrap(false); |
| 2513 | if (is_contained(Range&: Mask, Element: PoisonMaskElem) && !MightCreatePoisonOrUB) |
| 2514 | NewI->dropPoisonGeneratingFlags(); |
| 2515 | } |
| 2516 | return replaceInstUsesWith(I&: Shuf, V: NewBO); |
| 2517 | } |
| 2518 | |
| 2519 | /// Convert a narrowing shuffle of a bitcasted vector into a vector truncate. |
| 2520 | /// Example (little endian): |
| 2521 | /// shuf (bitcast <4 x i16> X to <8 x i8>), <0, 2, 4, 6> --> trunc X to <4 x i8> |
| 2522 | static Instruction *foldTruncShuffle(ShuffleVectorInst &Shuf, |
| 2523 | bool IsBigEndian) { |
| 2524 | // This must be a bitcasted shuffle of 1 vector integer operand. |
| 2525 | Type *DestType = Shuf.getType(); |
| 2526 | Value *X; |
| 2527 | if (!match(V: Shuf.getOperand(i_nocapture: 0), P: m_BitCast(Op: m_Value(V&: X))) || |
| 2528 | !match(V: Shuf.getOperand(i_nocapture: 1), P: m_Poison()) || !DestType->isIntOrIntVectorTy()) |
| 2529 | return nullptr; |
| 2530 | |
| 2531 | // The source type must have the same number of elements as the shuffle, |
| 2532 | // and the source element type must be larger than the shuffle element type. |
| 2533 | Type *SrcType = X->getType(); |
| 2534 | if (!SrcType->isVectorTy() || !SrcType->isIntOrIntVectorTy() || |
| 2535 | cast<FixedVectorType>(Val: SrcType)->getNumElements() != |
| 2536 | cast<FixedVectorType>(Val: DestType)->getNumElements() || |
| 2537 | SrcType->getScalarSizeInBits() % DestType->getScalarSizeInBits() != 0) |
| 2538 | return nullptr; |
| 2539 | |
| 2540 | assert(Shuf.changesLength() && !Shuf.increasesLength() && |
| 2541 | "Expected a shuffle that decreases length" ); |
| 2542 | |
| 2543 | // Last, check that the mask chooses the correct low bits for each narrow |
| 2544 | // element in the result. |
| 2545 | uint64_t TruncRatio = |
| 2546 | SrcType->getScalarSizeInBits() / DestType->getScalarSizeInBits(); |
| 2547 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
| 2548 | for (unsigned i = 0, e = Mask.size(); i != e; ++i) { |
| 2549 | if (Mask[i] == PoisonMaskElem) |
| 2550 | continue; |
| 2551 | uint64_t LSBIndex = IsBigEndian ? (i + 1) * TruncRatio - 1 : i * TruncRatio; |
| 2552 | assert(LSBIndex <= INT32_MAX && "Overflowed 32-bits" ); |
| 2553 | if (Mask[i] != (int)LSBIndex) |
| 2554 | return nullptr; |
| 2555 | } |
| 2556 | |
| 2557 | return new TruncInst(X, DestType); |
| 2558 | } |
| 2559 | |
| 2560 | /// Match a shuffle-select-shuffle pattern where the shuffles are widening and |
| 2561 | /// narrowing (concatenating with poison and extracting back to the original |
| 2562 | /// length). This allows replacing the wide select with a narrow select. |
| 2563 | static Instruction *narrowVectorSelect(ShuffleVectorInst &Shuf, |
| 2564 | InstCombiner::BuilderTy &Builder) { |
| 2565 | // This must be a narrowing identity shuffle. It extracts the 1st N elements |
| 2566 | // of the 1st vector operand of a shuffle. |
| 2567 | if (!match(V: Shuf.getOperand(i_nocapture: 1), P: m_Poison()) || !Shuf.isIdentityWithExtract()) |
| 2568 | return nullptr; |
| 2569 | |
| 2570 | // The vector being shuffled must be a vector select that we can eliminate. |
| 2571 | // TODO: The one-use requirement could be eased if X and/or Y are constants. |
| 2572 | Value *Cond, *X, *Y; |
| 2573 | if (!match(V: Shuf.getOperand(i_nocapture: 0), |
| 2574 | P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: X), R: m_Value(V&: Y))))) |
| 2575 | return nullptr; |
| 2576 | |
| 2577 | // We need a narrow condition value. It must be extended with poison elements |
| 2578 | // and have the same number of elements as this shuffle. |
| 2579 | unsigned NarrowNumElts = |
| 2580 | cast<FixedVectorType>(Val: Shuf.getType())->getNumElements(); |
| 2581 | Value *NarrowCond; |
| 2582 | if (!match(V: Cond, P: m_OneUse(SubPattern: m_Shuffle(v1: m_Value(V&: NarrowCond), v2: m_Poison()))) || |
| 2583 | cast<FixedVectorType>(Val: NarrowCond->getType())->getNumElements() != |
| 2584 | NarrowNumElts || |
| 2585 | !cast<ShuffleVectorInst>(Val: Cond)->isIdentityWithPadding()) |
| 2586 | return nullptr; |
| 2587 | |
| 2588 | // shuf (sel (shuf NarrowCond, poison, WideMask), X, Y), poison, NarrowMask) |
| 2589 | // --> |
| 2590 | // sel NarrowCond, (shuf X, poison, NarrowMask), (shuf Y, poison, NarrowMask) |
| 2591 | Value *NarrowX = Builder.CreateShuffleVector(V: X, Mask: Shuf.getShuffleMask()); |
| 2592 | Value *NarrowY = Builder.CreateShuffleVector(V: Y, Mask: Shuf.getShuffleMask()); |
| 2593 | return SelectInst::Create(C: NarrowCond, S1: NarrowX, S2: NarrowY); |
| 2594 | } |
| 2595 | |
| 2596 | /// Canonicalize FP negate/abs after shuffle. |
| 2597 | static Instruction *foldShuffleOfUnaryOps(ShuffleVectorInst &Shuf, |
| 2598 | InstCombiner::BuilderTy &Builder) { |
| 2599 | auto *S0 = dyn_cast<Instruction>(Val: Shuf.getOperand(i_nocapture: 0)); |
| 2600 | Value *X; |
| 2601 | if (!S0 || !match(V: S0, P: m_CombineOr(Ps: m_FNeg(X: m_Value(V&: X)), Ps: m_FAbs(Op0: m_Value(V&: X))))) |
| 2602 | return nullptr; |
| 2603 | |
| 2604 | bool IsFNeg = S0->getOpcode() == Instruction::FNeg; |
| 2605 | |
| 2606 | // Match 2-input (binary) shuffle. |
| 2607 | auto *S1 = dyn_cast<Instruction>(Val: Shuf.getOperand(i_nocapture: 1)); |
| 2608 | Value *Y; |
| 2609 | if (!S1 || !match(V: S1, P: m_CombineOr(Ps: m_FNeg(X: m_Value(V&: Y)), Ps: m_FAbs(Op0: m_Value(V&: Y)))) || |
| 2610 | S0->getOpcode() != S1->getOpcode() || |
| 2611 | (!S0->hasOneUse() && !S1->hasOneUse())) |
| 2612 | return nullptr; |
| 2613 | |
| 2614 | // shuf (fneg/fabs X), (fneg/fabs Y), Mask --> fneg/fabs (shuf X, Y, Mask) |
| 2615 | Value *NewShuf = Builder.CreateShuffleVector(V1: X, V2: Y, Mask: Shuf.getShuffleMask()); |
| 2616 | Instruction *NewF; |
| 2617 | if (IsFNeg) { |
| 2618 | NewF = UnaryOperator::CreateFNeg(V: NewShuf); |
| 2619 | } else { |
| 2620 | Function *FAbs = Intrinsic::getOrInsertDeclaration( |
| 2621 | M: Shuf.getModule(), id: Intrinsic::fabs, OverloadTys: Shuf.getType()); |
| 2622 | NewF = CallInst::Create(Func: FAbs, Args: {NewShuf}); |
| 2623 | } |
| 2624 | NewF->copyIRFlags(V: S0); |
| 2625 | NewF->andIRFlags(V: S1); |
| 2626 | return NewF; |
| 2627 | } |
| 2628 | |
| 2629 | /// Canonicalize casts after shuffle. |
| 2630 | static Instruction *foldCastShuffle(ShuffleVectorInst &Shuf, |
| 2631 | InstCombiner::BuilderTy &Builder) { |
| 2632 | auto *Cast0 = dyn_cast<CastInst>(Val: Shuf.getOperand(i_nocapture: 0)); |
| 2633 | if (!Cast0) |
| 2634 | return nullptr; |
| 2635 | |
| 2636 | // TODO: Allow other opcodes? That would require easing the type restrictions |
| 2637 | // below here. |
| 2638 | CastInst::CastOps CastOpcode = Cast0->getOpcode(); |
| 2639 | switch (CastOpcode) { |
| 2640 | case Instruction::SExt: |
| 2641 | case Instruction::ZExt: |
| 2642 | case Instruction::FPToSI: |
| 2643 | case Instruction::FPToUI: |
| 2644 | case Instruction::SIToFP: |
| 2645 | case Instruction::UIToFP: |
| 2646 | break; |
| 2647 | default: |
| 2648 | return nullptr; |
| 2649 | } |
| 2650 | |
| 2651 | VectorType *CastSrcTy = cast<VectorType>(Val: Cast0->getSrcTy()); |
| 2652 | VectorType *ShufTy = Shuf.getType(); |
| 2653 | VectorType *ShufOpTy = cast<VectorType>(Val: Shuf.getOperand(i_nocapture: 0)->getType()); |
| 2654 | |
| 2655 | // TODO: Allow length-increasing shuffles? |
| 2656 | if (ShufTy->getElementCount().getKnownMinValue() > |
| 2657 | ShufOpTy->getElementCount().getKnownMinValue()) |
| 2658 | return nullptr; |
| 2659 | |
| 2660 | // shuffle (cast X), Poison, identity-with-extract-mask --> |
| 2661 | // cast (shuffle X, Poison, identity-with-extract-mask). |
| 2662 | if (isa<PoisonValue>(Val: Shuf.getOperand(i_nocapture: 1)) && Cast0->hasOneUse() && |
| 2663 | Shuf.isIdentityWithExtract()) { |
| 2664 | auto *NewIns = Builder.CreateShuffleVector(V1: Cast0->getOperand(i_nocapture: 0), |
| 2665 | V2: PoisonValue::get(T: CastSrcTy), |
| 2666 | Mask: Shuf.getShuffleMask()); |
| 2667 | return CastInst::Create(Cast0->getOpcode(), S: NewIns, Ty: Shuf.getType()); |
| 2668 | } |
| 2669 | |
| 2670 | auto *Cast1 = dyn_cast<CastInst>(Val: Shuf.getOperand(i_nocapture: 1)); |
| 2671 | // Do we have 2 matching cast operands? |
| 2672 | if (!Cast1 || Cast0->getOpcode() != Cast1->getOpcode() || |
| 2673 | Cast0->getSrcTy() != Cast1->getSrcTy()) |
| 2674 | return nullptr; |
| 2675 | |
| 2676 | // TODO: Allow element-size-decreasing casts (ex: fptosi float to i8)? |
| 2677 | assert(isa<FixedVectorType>(CastSrcTy) && isa<FixedVectorType>(ShufOpTy) && |
| 2678 | "Expected fixed vector operands for casts and binary shuffle" ); |
| 2679 | if (CastSrcTy->getPrimitiveSizeInBits() > ShufOpTy->getPrimitiveSizeInBits()) |
| 2680 | return nullptr; |
| 2681 | |
| 2682 | // At least one of the operands must have only one use (the shuffle). |
| 2683 | if (!Cast0->hasOneUse() && !Cast1->hasOneUse()) |
| 2684 | return nullptr; |
| 2685 | |
| 2686 | // shuffle (cast X), (cast Y), Mask --> cast (shuffle X, Y, Mask) |
| 2687 | Value *X = Cast0->getOperand(i_nocapture: 0); |
| 2688 | Value *Y = Cast1->getOperand(i_nocapture: 0); |
| 2689 | Value *NewShuf = Builder.CreateShuffleVector(V1: X, V2: Y, Mask: Shuf.getShuffleMask()); |
| 2690 | return CastInst::Create(CastOpcode, S: NewShuf, Ty: ShufTy); |
| 2691 | } |
| 2692 | |
| 2693 | /// Try to fold an extract subvector operation. |
| 2694 | static Instruction *(ShuffleVectorInst &Shuf) { |
| 2695 | Value *Op0 = Shuf.getOperand(i_nocapture: 0), *Op1 = Shuf.getOperand(i_nocapture: 1); |
| 2696 | if (!Shuf.isIdentityWithExtract() || !match(V: Op1, P: m_Poison())) |
| 2697 | return nullptr; |
| 2698 | |
| 2699 | // Check if we are extracting all bits of an inserted scalar: |
| 2700 | // extract-subvec (bitcast (inselt ?, X, 0) --> bitcast X to subvec type |
| 2701 | Value *X; |
| 2702 | if (match(V: Op0, P: m_BitCast(Op: m_InsertElt(Val: m_Value(), Elt: m_Value(V&: X), Idx: m_Zero()))) && |
| 2703 | X->getType()->getPrimitiveSizeInBits() == |
| 2704 | Shuf.getType()->getPrimitiveSizeInBits()) |
| 2705 | return new BitCastInst(X, Shuf.getType()); |
| 2706 | |
| 2707 | // Try to combine 2 shuffles into 1 shuffle by concatenating a shuffle mask. |
| 2708 | Value *Y; |
| 2709 | ArrayRef<int> Mask; |
| 2710 | if (!match(V: Op0, P: m_Shuffle(v1: m_Value(V&: X), v2: m_Value(V&: Y), mask: m_Mask(Mask)))) |
| 2711 | return nullptr; |
| 2712 | |
| 2713 | // Be conservative with shuffle transforms. If we can't kill the 1st shuffle, |
| 2714 | // then combining may result in worse codegen. |
| 2715 | if (!Op0->hasOneUse()) |
| 2716 | return nullptr; |
| 2717 | |
| 2718 | // We are extracting a subvector from a shuffle. Remove excess elements from |
| 2719 | // the 1st shuffle mask to eliminate the extract. |
| 2720 | // |
| 2721 | // This transform is conservatively limited to identity extracts because we do |
| 2722 | // not allow arbitrary shuffle mask creation as a target-independent transform |
| 2723 | // (because we can't guarantee that will lower efficiently). |
| 2724 | // |
| 2725 | // If the extracting shuffle has an poison mask element, it transfers to the |
| 2726 | // new shuffle mask. Otherwise, copy the original mask element. Example: |
| 2727 | // shuf (shuf X, Y, <C0, C1, C2, poison, C4>), poison, <0, poison, 2, 3> --> |
| 2728 | // shuf X, Y, <C0, poison, C2, poison> |
| 2729 | unsigned NumElts = cast<FixedVectorType>(Val: Shuf.getType())->getNumElements(); |
| 2730 | SmallVector<int, 16> NewMask(NumElts); |
| 2731 | assert(NumElts < Mask.size() && |
| 2732 | "Identity with extract must have less elements than its inputs" ); |
| 2733 | |
| 2734 | for (unsigned i = 0; i != NumElts; ++i) { |
| 2735 | int = Shuf.getMaskValue(Elt: i); |
| 2736 | int MaskElt = Mask[i]; |
| 2737 | NewMask[i] = ExtractMaskElt == PoisonMaskElem ? ExtractMaskElt : MaskElt; |
| 2738 | } |
| 2739 | return new ShuffleVectorInst(X, Y, NewMask); |
| 2740 | } |
| 2741 | |
| 2742 | /// Try to replace a shuffle with an insertelement or try to replace a shuffle |
| 2743 | /// operand with the operand of an insertelement. |
| 2744 | static Instruction *foldShuffleWithInsert(ShuffleVectorInst &Shuf, |
| 2745 | InstCombinerImpl &IC) { |
| 2746 | Value *V0 = Shuf.getOperand(i_nocapture: 0), *V1 = Shuf.getOperand(i_nocapture: 1); |
| 2747 | SmallVector<int, 16> Mask; |
| 2748 | Shuf.getShuffleMask(Result&: Mask); |
| 2749 | |
| 2750 | int NumElts = Mask.size(); |
| 2751 | int InpNumElts = cast<FixedVectorType>(Val: V0->getType())->getNumElements(); |
| 2752 | |
| 2753 | // This is a specialization of a fold in SimplifyDemandedVectorElts. We may |
| 2754 | // not be able to handle it there if the insertelement has >1 use. |
| 2755 | // If the shuffle has an insertelement operand but does not choose the |
| 2756 | // inserted scalar element from that value, then we can replace that shuffle |
| 2757 | // operand with the source vector of the insertelement. |
| 2758 | Value *X; |
| 2759 | uint64_t IdxC; |
| 2760 | if (match(V: V0, P: m_InsertElt(Val: m_Value(V&: X), Elt: m_Value(), Idx: m_ConstantInt(V&: IdxC)))) { |
| 2761 | // shuf (inselt X, ?, IdxC), ?, Mask --> shuf X, ?, Mask |
| 2762 | if (!is_contained(Range&: Mask, Element: (int)IdxC)) |
| 2763 | return IC.replaceOperand(I&: Shuf, OpNum: 0, V: X); |
| 2764 | } |
| 2765 | if (match(V: V1, P: m_InsertElt(Val: m_Value(V&: X), Elt: m_Value(), Idx: m_ConstantInt(V&: IdxC)))) { |
| 2766 | // Offset the index constant by the vector width because we are checking for |
| 2767 | // accesses to the 2nd vector input of the shuffle. |
| 2768 | IdxC += InpNumElts; |
| 2769 | // shuf ?, (inselt X, ?, IdxC), Mask --> shuf ?, X, Mask |
| 2770 | if (!is_contained(Range&: Mask, Element: (int)IdxC)) |
| 2771 | return IC.replaceOperand(I&: Shuf, OpNum: 1, V: X); |
| 2772 | } |
| 2773 | // For the rest of the transform, the shuffle must not change vector sizes. |
| 2774 | // TODO: This restriction could be removed if the insert has only one use |
| 2775 | // (because the transform would require a new length-changing shuffle). |
| 2776 | if (NumElts != InpNumElts) |
| 2777 | return nullptr; |
| 2778 | |
| 2779 | // shuffle (insert ?, Scalar, IndexC), V1, Mask --> insert V1, Scalar, IndexC' |
| 2780 | auto isShufflingScalarIntoOp1 = [&](Value *&Scalar, ConstantInt *&IndexC) { |
| 2781 | // We need an insertelement with a constant index. |
| 2782 | if (!match(V: V0, P: m_InsertElt(Val: m_Value(), Elt: m_Value(V&: Scalar), |
| 2783 | Idx: m_ConstantInt(CI&: IndexC)))) |
| 2784 | return false; |
| 2785 | |
| 2786 | // Test the shuffle mask to see if it splices the inserted scalar into the |
| 2787 | // operand 1 vector of the shuffle. |
| 2788 | int NewInsIndex = -1; |
| 2789 | for (int i = 0; i != NumElts; ++i) { |
| 2790 | // Ignore undef mask elements. |
| 2791 | if (Mask[i] == -1) |
| 2792 | continue; |
| 2793 | |
| 2794 | // The shuffle takes elements of operand 1 without lane changes. |
| 2795 | if (Mask[i] == NumElts + i) |
| 2796 | continue; |
| 2797 | |
| 2798 | // The shuffle must choose the inserted scalar exactly once. |
| 2799 | if (NewInsIndex != -1 || Mask[i] != IndexC->getSExtValue()) |
| 2800 | return false; |
| 2801 | |
| 2802 | // The shuffle is placing the inserted scalar into element i. |
| 2803 | NewInsIndex = i; |
| 2804 | } |
| 2805 | |
| 2806 | assert(NewInsIndex != -1 && "Did not fold shuffle with unused operand?" ); |
| 2807 | |
| 2808 | // Index is updated to the potentially translated insertion lane. |
| 2809 | IndexC = ConstantInt::get(Ty: IndexC->getIntegerType(), V: NewInsIndex); |
| 2810 | return true; |
| 2811 | }; |
| 2812 | |
| 2813 | // If the shuffle is unnecessary, insert the scalar operand directly into |
| 2814 | // operand 1 of the shuffle. Example: |
| 2815 | // shuffle (insert ?, S, 1), V1, <1, 5, 6, 7> --> insert V1, S, 0 |
| 2816 | Value *Scalar; |
| 2817 | ConstantInt *IndexC; |
| 2818 | if (isShufflingScalarIntoOp1(Scalar, IndexC)) |
| 2819 | return InsertElementInst::Create(Vec: V1, NewElt: Scalar, Idx: IndexC); |
| 2820 | |
| 2821 | // Try again after commuting shuffle. Example: |
| 2822 | // shuffle V0, (insert ?, S, 0), <0, 1, 2, 4> --> |
| 2823 | // shuffle (insert ?, S, 0), V0, <4, 5, 6, 0> --> insert V0, S, 3 |
| 2824 | std::swap(a&: V0, b&: V1); |
| 2825 | ShuffleVectorInst::commuteShuffleMask(Mask, InVecNumElts: NumElts); |
| 2826 | if (isShufflingScalarIntoOp1(Scalar, IndexC)) |
| 2827 | return InsertElementInst::Create(Vec: V1, NewElt: Scalar, Idx: IndexC); |
| 2828 | |
| 2829 | return nullptr; |
| 2830 | } |
| 2831 | |
| 2832 | static Instruction *foldIdentityPaddedShuffles(ShuffleVectorInst &Shuf) { |
| 2833 | // Match the operands as identity with padding (also known as concatenation |
| 2834 | // with undef) shuffles of the same source type. The backend is expected to |
| 2835 | // recreate these concatenations from a shuffle of narrow operands. |
| 2836 | auto *Shuffle0 = dyn_cast<ShuffleVectorInst>(Val: Shuf.getOperand(i_nocapture: 0)); |
| 2837 | auto *Shuffle1 = dyn_cast<ShuffleVectorInst>(Val: Shuf.getOperand(i_nocapture: 1)); |
| 2838 | if (!Shuffle0 || !Shuffle0->isIdentityWithPadding() || |
| 2839 | !Shuffle1 || !Shuffle1->isIdentityWithPadding()) |
| 2840 | return nullptr; |
| 2841 | |
| 2842 | // We limit this transform to power-of-2 types because we expect that the |
| 2843 | // backend can convert the simplified IR patterns to identical nodes as the |
| 2844 | // original IR. |
| 2845 | // TODO: If we can verify the same behavior for arbitrary types, the |
| 2846 | // power-of-2 checks can be removed. |
| 2847 | Value *X = Shuffle0->getOperand(i_nocapture: 0); |
| 2848 | Value *Y = Shuffle1->getOperand(i_nocapture: 0); |
| 2849 | if (X->getType() != Y->getType() || |
| 2850 | !isPowerOf2_32(Value: cast<FixedVectorType>(Val: Shuf.getType())->getNumElements()) || |
| 2851 | !isPowerOf2_32( |
| 2852 | Value: cast<FixedVectorType>(Val: Shuffle0->getType())->getNumElements()) || |
| 2853 | !isPowerOf2_32(Value: cast<FixedVectorType>(Val: X->getType())->getNumElements()) || |
| 2854 | match(V: X, P: m_Undef()) || match(V: Y, P: m_Undef())) |
| 2855 | return nullptr; |
| 2856 | assert(match(Shuffle0->getOperand(1), m_Undef()) && |
| 2857 | match(Shuffle1->getOperand(1), m_Undef()) && |
| 2858 | "Unexpected operand for identity shuffle" ); |
| 2859 | |
| 2860 | // This is a shuffle of 2 widening shuffles. We can shuffle the narrow source |
| 2861 | // operands directly by adjusting the shuffle mask to account for the narrower |
| 2862 | // types: |
| 2863 | // shuf (widen X), (widen Y), Mask --> shuf X, Y, Mask' |
| 2864 | int NarrowElts = cast<FixedVectorType>(Val: X->getType())->getNumElements(); |
| 2865 | int WideElts = cast<FixedVectorType>(Val: Shuffle0->getType())->getNumElements(); |
| 2866 | assert(WideElts > NarrowElts && "Unexpected types for identity with padding" ); |
| 2867 | |
| 2868 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
| 2869 | SmallVector<int, 16> NewMask(Mask.size(), -1); |
| 2870 | for (int i = 0, e = Mask.size(); i != e; ++i) { |
| 2871 | if (Mask[i] == -1) |
| 2872 | continue; |
| 2873 | |
| 2874 | // If this shuffle is choosing an undef element from 1 of the sources, that |
| 2875 | // element is undef. |
| 2876 | if (Mask[i] < WideElts) { |
| 2877 | if (Shuffle0->getMaskValue(Elt: Mask[i]) == -1) |
| 2878 | continue; |
| 2879 | } else { |
| 2880 | if (Shuffle1->getMaskValue(Elt: Mask[i] - WideElts) == -1) |
| 2881 | continue; |
| 2882 | } |
| 2883 | |
| 2884 | // If this shuffle is choosing from the 1st narrow op, the mask element is |
| 2885 | // the same. If this shuffle is choosing from the 2nd narrow op, the mask |
| 2886 | // element is offset down to adjust for the narrow vector widths. |
| 2887 | if (Mask[i] < WideElts) { |
| 2888 | assert(Mask[i] < NarrowElts && "Unexpected shuffle mask" ); |
| 2889 | NewMask[i] = Mask[i]; |
| 2890 | } else { |
| 2891 | assert(Mask[i] < (WideElts + NarrowElts) && "Unexpected shuffle mask" ); |
| 2892 | NewMask[i] = Mask[i] - (WideElts - NarrowElts); |
| 2893 | } |
| 2894 | } |
| 2895 | return new ShuffleVectorInst(X, Y, NewMask); |
| 2896 | } |
| 2897 | |
| 2898 | // Splatting the first element of the result of a BinOp, where any of the |
| 2899 | // BinOp's operands are the result of a first element splat can be simplified to |
| 2900 | // splatting the first element of the result of the BinOp |
| 2901 | Instruction *InstCombinerImpl::simplifyBinOpSplats(ShuffleVectorInst &SVI) { |
| 2902 | if (!match(V: SVI.getOperand(i_nocapture: 1), P: m_Poison()) || |
| 2903 | !match(Mask: SVI.getShuffleMask(), P: m_ZeroMask()) || |
| 2904 | !SVI.getOperand(i_nocapture: 0)->hasOneUse()) |
| 2905 | return nullptr; |
| 2906 | |
| 2907 | Value *Op0 = SVI.getOperand(i_nocapture: 0); |
| 2908 | Value *X, *Y; |
| 2909 | if (!match(V: Op0, P: m_BinOp(L: m_Shuffle(v1: m_Value(V&: X), v2: m_Poison(), mask: m_ZeroMask()), |
| 2910 | R: m_Value(V&: Y))) && |
| 2911 | !match(V: Op0, P: m_BinOp(L: m_Value(V&: X), |
| 2912 | R: m_Shuffle(v1: m_Value(V&: Y), v2: m_Poison(), mask: m_ZeroMask())))) |
| 2913 | return nullptr; |
| 2914 | if (X->getType() != Y->getType()) |
| 2915 | return nullptr; |
| 2916 | |
| 2917 | auto *BinOp = cast<BinaryOperator>(Val: Op0); |
| 2918 | if (!isSafeToSpeculativelyExecuteWithVariableReplaced(I: BinOp)) |
| 2919 | return nullptr; |
| 2920 | |
| 2921 | Value *NewBO = Builder.CreateBinOp(Opc: BinOp->getOpcode(), LHS: X, RHS: Y); |
| 2922 | if (auto NewBOI = dyn_cast<Instruction>(Val: NewBO)) |
| 2923 | NewBOI->copyIRFlags(V: BinOp); |
| 2924 | |
| 2925 | return new ShuffleVectorInst(NewBO, SVI.getShuffleMask()); |
| 2926 | } |
| 2927 | |
| 2928 | Instruction *InstCombinerImpl::visitShuffleVectorInst(ShuffleVectorInst &SVI) { |
| 2929 | Value *LHS = SVI.getOperand(i_nocapture: 0); |
| 2930 | Value *RHS = SVI.getOperand(i_nocapture: 1); |
| 2931 | SimplifyQuery ShufQuery = SQ.getWithInstruction(I: &SVI); |
| 2932 | if (auto *V = simplifyShuffleVectorInst(Op0: LHS, Op1: RHS, Mask: SVI.getShuffleMask(), |
| 2933 | RetTy: SVI.getType(), Q: ShufQuery)) |
| 2934 | return replaceInstUsesWith(I&: SVI, V); |
| 2935 | |
| 2936 | if (Instruction *I = simplifyBinOpSplats(SVI)) |
| 2937 | return I; |
| 2938 | |
| 2939 | // Canonicalize splat shuffle to use poison RHS. Handle this explicitly in |
| 2940 | // order to support scalable vectors. |
| 2941 | if (match(Mask: SVI.getShuffleMask(), P: m_ZeroMask()) && !isa<PoisonValue>(Val: RHS)) |
| 2942 | return replaceOperand(I&: SVI, OpNum: 1, V: PoisonValue::get(T: RHS->getType())); |
| 2943 | |
| 2944 | if (isa<ScalableVectorType>(Val: LHS->getType())) |
| 2945 | return nullptr; |
| 2946 | |
| 2947 | unsigned VWidth = cast<FixedVectorType>(Val: SVI.getType())->getNumElements(); |
| 2948 | unsigned LHSWidth = cast<FixedVectorType>(Val: LHS->getType())->getNumElements(); |
| 2949 | |
| 2950 | // shuffle (bitcast X), (bitcast Y), Mask --> bitcast (shuffle X, Y, Mask) |
| 2951 | // |
| 2952 | // if X and Y are of the same (vector) type, and the element size is not |
| 2953 | // changed by the bitcasts, we can distribute the bitcasts through the |
| 2954 | // shuffle, hopefully reducing the number of instructions. We make sure that |
| 2955 | // at least one bitcast only has one use, so we don't *increase* the number of |
| 2956 | // instructions here. |
| 2957 | Value *X, *Y; |
| 2958 | if (match(V: LHS, P: m_BitCast(Op: m_Value(V&: X))) && match(V: RHS, P: m_BitCast(Op: m_Value(V&: Y))) && |
| 2959 | X->getType()->isVectorTy() && X->getType() == Y->getType() && |
| 2960 | X->getType()->getScalarSizeInBits() == |
| 2961 | SVI.getType()->getScalarSizeInBits() && |
| 2962 | (LHS->hasOneUse() || RHS->hasOneUse())) { |
| 2963 | Value *V = Builder.CreateShuffleVector(V1: X, V2: Y, Mask: SVI.getShuffleMask(), |
| 2964 | Name: SVI.getName() + ".uncasted" ); |
| 2965 | return new BitCastInst(V, SVI.getType()); |
| 2966 | } |
| 2967 | |
| 2968 | ArrayRef<int> Mask = SVI.getShuffleMask(); |
| 2969 | |
| 2970 | // Peek through a bitcasted shuffle operand by scaling the mask. If the |
| 2971 | // simulated shuffle can simplify, then this shuffle is unnecessary: |
| 2972 | // shuf (bitcast X), undef, Mask --> bitcast X' |
| 2973 | // TODO: This could be extended to allow length-changing shuffles. |
| 2974 | // The transform might also be obsoleted if we allowed canonicalization |
| 2975 | // of bitcasted shuffles. |
| 2976 | if (match(V: LHS, P: m_BitCast(Op: m_Value(V&: X))) && match(V: RHS, P: m_Undef()) && |
| 2977 | X->getType()->isVectorTy() && VWidth == LHSWidth) { |
| 2978 | // Try to create a scaled mask constant. |
| 2979 | auto *XType = cast<FixedVectorType>(Val: X->getType()); |
| 2980 | unsigned XNumElts = XType->getNumElements(); |
| 2981 | SmallVector<int, 16> ScaledMask; |
| 2982 | if (scaleShuffleMaskElts(NumDstElts: XNumElts, Mask, ScaledMask)) { |
| 2983 | // If the shuffled source vector simplifies, cast that value to this |
| 2984 | // shuffle's type. |
| 2985 | if (auto *V = simplifyShuffleVectorInst(Op0: X, Op1: UndefValue::get(T: XType), |
| 2986 | Mask: ScaledMask, RetTy: XType, Q: ShufQuery)) |
| 2987 | return BitCastInst::Create(Instruction::BitCast, S: V, Ty: SVI.getType()); |
| 2988 | } |
| 2989 | } |
| 2990 | |
| 2991 | // shuffle x, x, mask --> shuffle x, undef, mask' |
| 2992 | if (LHS == RHS) { |
| 2993 | assert(!match(RHS, m_Undef()) && |
| 2994 | "Shuffle with 2 undef ops not simplified?" ); |
| 2995 | return new ShuffleVectorInst(LHS, createUnaryMask(Mask, NumElts: LHSWidth)); |
| 2996 | } |
| 2997 | |
| 2998 | // shuffle undef, x, mask --> shuffle x, undef, mask' |
| 2999 | if (match(V: LHS, P: m_Undef())) { |
| 3000 | SVI.commute(); |
| 3001 | return &SVI; |
| 3002 | } |
| 3003 | |
| 3004 | if (Instruction *I = canonicalizeInsertSplat(Shuf&: SVI, Builder)) |
| 3005 | return I; |
| 3006 | |
| 3007 | if (Instruction *I = foldSelectShuffle(Shuf&: SVI)) |
| 3008 | return I; |
| 3009 | |
| 3010 | if (Instruction *I = foldTruncShuffle(Shuf&: SVI, IsBigEndian: DL.isBigEndian())) |
| 3011 | return I; |
| 3012 | |
| 3013 | if (Instruction *I = narrowVectorSelect(Shuf&: SVI, Builder)) |
| 3014 | return I; |
| 3015 | |
| 3016 | if (Instruction *I = foldShuffleOfUnaryOps(Shuf&: SVI, Builder)) |
| 3017 | return I; |
| 3018 | |
| 3019 | if (Instruction *I = foldCastShuffle(Shuf&: SVI, Builder)) |
| 3020 | return I; |
| 3021 | |
| 3022 | APInt PoisonElts(VWidth, 0); |
| 3023 | APInt AllOnesEltMask(APInt::getAllOnes(numBits: VWidth)); |
| 3024 | if (Value *V = SimplifyDemandedVectorElts(V: &SVI, DemandedElts: AllOnesEltMask, PoisonElts)) { |
| 3025 | if (V != &SVI) |
| 3026 | return replaceInstUsesWith(I&: SVI, V); |
| 3027 | return &SVI; |
| 3028 | } |
| 3029 | |
| 3030 | if (Instruction *I = foldIdentityExtractShuffle(Shuf&: SVI)) |
| 3031 | return I; |
| 3032 | |
| 3033 | // These transforms have the potential to lose undef knowledge, so they are |
| 3034 | // intentionally placed after SimplifyDemandedVectorElts(). |
| 3035 | if (Instruction *I = foldShuffleWithInsert(Shuf&: SVI, IC&: *this)) |
| 3036 | return I; |
| 3037 | if (Instruction *I = foldIdentityPaddedShuffles(Shuf&: SVI)) |
| 3038 | return I; |
| 3039 | |
| 3040 | if (match(V: RHS, P: m_Constant())) { |
| 3041 | if (auto *SI = dyn_cast<SelectInst>(Val: LHS)) { |
| 3042 | // We cannot do this fold for elementwise select since ShuffleVector is |
| 3043 | // not elementwise. |
| 3044 | if (SI->getCondition()->getType()->isIntegerTy() && |
| 3045 | (isa<PoisonValue>(Val: RHS) || |
| 3046 | isGuaranteedNotToBePoison(V: SI->getCondition()))) { |
| 3047 | if (Instruction *I = FoldOpIntoSelect(Op&: SVI, SI)) |
| 3048 | return I; |
| 3049 | } |
| 3050 | } |
| 3051 | if (auto *PN = dyn_cast<PHINode>(Val: LHS)) { |
| 3052 | if (Instruction *I = foldOpIntoPhi(I&: SVI, PN, /*AllowMultipleUses=*/true)) |
| 3053 | return I; |
| 3054 | } |
| 3055 | } |
| 3056 | |
| 3057 | if (match(V: RHS, P: m_Poison()) && canEvaluateShuffled(V: LHS, Mask)) { |
| 3058 | Value *V = evaluateInDifferentElementOrder(V: LHS, Mask, Builder); |
| 3059 | return replaceInstUsesWith(I&: SVI, V); |
| 3060 | } |
| 3061 | |
| 3062 | // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to |
| 3063 | // a non-vector type. We can instead bitcast the original vector followed by |
| 3064 | // an extract of the desired element: |
| 3065 | // |
| 3066 | // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef, |
| 3067 | // <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| 3068 | // %1 = bitcast <4 x i8> %sroa to i32 |
| 3069 | // Becomes: |
| 3070 | // %bc = bitcast <16 x i8> %in to <4 x i32> |
| 3071 | // %ext = extractelement <4 x i32> %bc, i32 0 |
| 3072 | // |
| 3073 | // If the shuffle is extracting a contiguous range of values from the input |
| 3074 | // vector then each use which is a bitcast of the extracted size can be |
| 3075 | // replaced. This will work if the vector types are compatible, and the begin |
| 3076 | // index is aligned to a value in the casted vector type. If the begin index |
| 3077 | // isn't aligned then we can shuffle the original vector (keeping the same |
| 3078 | // vector type) before extracting. |
| 3079 | // |
| 3080 | // This code will bail out if the target type is fundamentally incompatible |
| 3081 | // with vectors of the source type. |
| 3082 | // |
| 3083 | // Example of <16 x i8>, target type i32: |
| 3084 | // Index range [4,8): v-----------v Will work. |
| 3085 | // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ |
| 3086 | // <16 x i8>: | | | | | | | | | | | | | | | | | |
| 3087 | // <4 x i32>: | | | | | |
| 3088 | // +-----------+-----------+-----------+-----------+ |
| 3089 | // Index range [6,10): ^-----------^ Needs an extra shuffle. |
| 3090 | // Target type i40: ^--------------^ Won't work, bail. |
| 3091 | bool MadeChange = false; |
| 3092 | if (isShuffleExtractingFromLHS(SVI, Mask)) { |
| 3093 | Value *V = LHS; |
| 3094 | unsigned MaskElems = Mask.size(); |
| 3095 | auto *SrcTy = cast<FixedVectorType>(Val: V->getType()); |
| 3096 | unsigned VecBitWidth = DL.getTypeSizeInBits(Ty: SrcTy); |
| 3097 | unsigned SrcElemBitWidth = DL.getTypeSizeInBits(Ty: SrcTy->getElementType()); |
| 3098 | assert(SrcElemBitWidth && "vector elements must have a bitwidth" ); |
| 3099 | unsigned SrcNumElems = SrcTy->getNumElements(); |
| 3100 | SmallVector<BitCastInst *, 8> BCs; |
| 3101 | DenseMap<Type *, Value *> NewBCs; |
| 3102 | for (User *U : SVI.users()) |
| 3103 | if (BitCastInst *BC = dyn_cast<BitCastInst>(Val: U)) { |
| 3104 | // Only visit bitcasts that weren't previously handled. |
| 3105 | if (BC->use_empty()) |
| 3106 | continue; |
| 3107 | // Prefer to combine bitcasts of bitcasts before attempting this fold. |
| 3108 | if (BC->hasOneUse()) { |
| 3109 | auto *BC2 = dyn_cast<BitCastInst>(Val: BC->user_back()); |
| 3110 | if (BC2 && isEliminableCastPair(CI1: BC, CI2: BC2)) |
| 3111 | continue; |
| 3112 | } |
| 3113 | BCs.push_back(Elt: BC); |
| 3114 | } |
| 3115 | for (BitCastInst *BC : BCs) { |
| 3116 | unsigned BegIdx = Mask.front(); |
| 3117 | Type *TgtTy = BC->getDestTy(); |
| 3118 | unsigned TgtElemBitWidth = DL.getTypeSizeInBits(Ty: TgtTy); |
| 3119 | if (!TgtElemBitWidth) |
| 3120 | continue; |
| 3121 | unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth; |
| 3122 | bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth; |
| 3123 | bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth); |
| 3124 | if (!VecBitWidthsEqual) |
| 3125 | continue; |
| 3126 | if (!VectorType::isValidElementType(ElemTy: TgtTy)) |
| 3127 | continue; |
| 3128 | auto *CastSrcTy = FixedVectorType::get(ElementType: TgtTy, NumElts: TgtNumElems); |
| 3129 | if (!BegIsAligned) { |
| 3130 | // Shuffle the input so [0,NumElements) contains the output, and |
| 3131 | // [NumElems,SrcNumElems) is undef. |
| 3132 | SmallVector<int, 16> ShuffleMask(SrcNumElems, -1); |
| 3133 | for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I) |
| 3134 | ShuffleMask[I] = Idx; |
| 3135 | V = Builder.CreateShuffleVector(V, Mask: ShuffleMask, |
| 3136 | Name: SVI.getName() + ".extract" ); |
| 3137 | BegIdx = 0; |
| 3138 | } |
| 3139 | unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth; |
| 3140 | assert(SrcElemsPerTgtElem); |
| 3141 | BegIdx /= SrcElemsPerTgtElem; |
| 3142 | auto [It, Inserted] = NewBCs.try_emplace(Key: CastSrcTy); |
| 3143 | if (Inserted) |
| 3144 | It->second = Builder.CreateBitCast(V, DestTy: CastSrcTy, Name: SVI.getName() + ".bc" ); |
| 3145 | auto *Ext = Builder.CreateExtractElement(Vec: It->second, Idx: BegIdx, |
| 3146 | Name: SVI.getName() + ".extract" ); |
| 3147 | // The shufflevector isn't being replaced: the bitcast that used it |
| 3148 | // is. InstCombine will visit the newly-created instructions. |
| 3149 | replaceInstUsesWith(I&: *BC, V: Ext); |
| 3150 | MadeChange = true; |
| 3151 | } |
| 3152 | } |
| 3153 | |
| 3154 | // If the LHS is a shufflevector itself, see if we can combine it with this |
| 3155 | // one without producing an unusual shuffle. |
| 3156 | // Cases that might be simplified: |
| 3157 | // 1. |
| 3158 | // x1=shuffle(v1,v2,mask1) |
| 3159 | // x=shuffle(x1,undef,mask) |
| 3160 | // ==> |
| 3161 | // x=shuffle(v1,undef,newMask) |
| 3162 | // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1 |
| 3163 | // 2. |
| 3164 | // x1=shuffle(v1,undef,mask1) |
| 3165 | // x=shuffle(x1,x2,mask) |
| 3166 | // where v1.size() == mask1.size() |
| 3167 | // ==> |
| 3168 | // x=shuffle(v1,x2,newMask) |
| 3169 | // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i] |
| 3170 | // 3. |
| 3171 | // x2=shuffle(v2,undef,mask2) |
| 3172 | // x=shuffle(x1,x2,mask) |
| 3173 | // where v2.size() == mask2.size() |
| 3174 | // ==> |
| 3175 | // x=shuffle(x1,v2,newMask) |
| 3176 | // newMask[i] = (mask[i] < x1.size()) |
| 3177 | // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size() |
| 3178 | // 4. |
| 3179 | // x1=shuffle(v1,undef,mask1) |
| 3180 | // x2=shuffle(v2,undef,mask2) |
| 3181 | // x=shuffle(x1,x2,mask) |
| 3182 | // where v1.size() == v2.size() |
| 3183 | // ==> |
| 3184 | // x=shuffle(v1,v2,newMask) |
| 3185 | // newMask[i] = (mask[i] < x1.size()) |
| 3186 | // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size() |
| 3187 | // |
| 3188 | // Here we are really conservative: |
| 3189 | // we are absolutely afraid of producing a shuffle mask not in the input |
| 3190 | // program, because the code gen may not be smart enough to turn a merged |
| 3191 | // shuffle into two specific shuffles: it may produce worse code. As such, |
| 3192 | // we only merge two shuffles if the result is either a splat or one of the |
| 3193 | // input shuffle masks. In this case, merging the shuffles just removes |
| 3194 | // one instruction, which we know is safe. This is good for things like |
| 3195 | // turning: (splat(splat)) -> splat, or |
| 3196 | // merge(V[0..n], V[n+1..2n]) -> V[0..2n] |
| 3197 | ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(Val: LHS); |
| 3198 | ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(Val: RHS); |
| 3199 | if (LHSShuffle) |
| 3200 | if (!match(V: LHSShuffle->getOperand(i_nocapture: 1), P: m_Poison()) && |
| 3201 | !match(V: RHS, P: m_Poison())) |
| 3202 | LHSShuffle = nullptr; |
| 3203 | if (RHSShuffle) |
| 3204 | if (!match(V: RHSShuffle->getOperand(i_nocapture: 1), P: m_Poison())) |
| 3205 | RHSShuffle = nullptr; |
| 3206 | if (!LHSShuffle && !RHSShuffle) |
| 3207 | return MadeChange ? &SVI : nullptr; |
| 3208 | |
| 3209 | Value* LHSOp0 = nullptr; |
| 3210 | Value* LHSOp1 = nullptr; |
| 3211 | Value* RHSOp0 = nullptr; |
| 3212 | unsigned LHSOp0Width = 0; |
| 3213 | unsigned RHSOp0Width = 0; |
| 3214 | if (LHSShuffle) { |
| 3215 | LHSOp0 = LHSShuffle->getOperand(i_nocapture: 0); |
| 3216 | LHSOp1 = LHSShuffle->getOperand(i_nocapture: 1); |
| 3217 | LHSOp0Width = cast<FixedVectorType>(Val: LHSOp0->getType())->getNumElements(); |
| 3218 | } |
| 3219 | if (RHSShuffle) { |
| 3220 | RHSOp0 = RHSShuffle->getOperand(i_nocapture: 0); |
| 3221 | RHSOp0Width = cast<FixedVectorType>(Val: RHSOp0->getType())->getNumElements(); |
| 3222 | } |
| 3223 | Value* newLHS = LHS; |
| 3224 | Value* newRHS = RHS; |
| 3225 | if (LHSShuffle) { |
| 3226 | // case 1 |
| 3227 | if (match(V: RHS, P: m_Poison())) { |
| 3228 | newLHS = LHSOp0; |
| 3229 | newRHS = LHSOp1; |
| 3230 | } |
| 3231 | // case 2 or 4 |
| 3232 | else if (LHSOp0Width == LHSWidth) { |
| 3233 | newLHS = LHSOp0; |
| 3234 | } |
| 3235 | } |
| 3236 | // case 3 or 4 |
| 3237 | if (RHSShuffle && RHSOp0Width == LHSWidth) { |
| 3238 | newRHS = RHSOp0; |
| 3239 | } |
| 3240 | // case 4 |
| 3241 | if (LHSOp0 == RHSOp0) { |
| 3242 | newLHS = LHSOp0; |
| 3243 | newRHS = nullptr; |
| 3244 | } |
| 3245 | |
| 3246 | if (newLHS == LHS && newRHS == RHS) |
| 3247 | return MadeChange ? &SVI : nullptr; |
| 3248 | |
| 3249 | ArrayRef<int> LHSMask; |
| 3250 | ArrayRef<int> RHSMask; |
| 3251 | if (newLHS != LHS) |
| 3252 | LHSMask = LHSShuffle->getShuffleMask(); |
| 3253 | if (RHSShuffle && newRHS != RHS) |
| 3254 | RHSMask = RHSShuffle->getShuffleMask(); |
| 3255 | |
| 3256 | unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth; |
| 3257 | SmallVector<int, 16> newMask; |
| 3258 | bool isSplat = true; |
| 3259 | int SplatElt = -1; |
| 3260 | // Create a new mask for the new ShuffleVectorInst so that the new |
| 3261 | // ShuffleVectorInst is equivalent to the original one. |
| 3262 | for (unsigned i = 0; i < VWidth; ++i) { |
| 3263 | int eltMask; |
| 3264 | if (Mask[i] < 0) { |
| 3265 | // This element is a poison value. |
| 3266 | eltMask = -1; |
| 3267 | } else if (Mask[i] < (int)LHSWidth) { |
| 3268 | // This element is from left hand side vector operand. |
| 3269 | // |
| 3270 | // If LHS is going to be replaced (case 1, 2, or 4), calculate the |
| 3271 | // new mask value for the element. |
| 3272 | if (newLHS != LHS) { |
| 3273 | eltMask = LHSMask[Mask[i]]; |
| 3274 | // If the value selected is an poison value, explicitly specify it |
| 3275 | // with a -1 mask value. |
| 3276 | if (eltMask >= (int)LHSOp0Width && isa<PoisonValue>(Val: LHSOp1)) |
| 3277 | eltMask = -1; |
| 3278 | } else |
| 3279 | eltMask = Mask[i]; |
| 3280 | } else { |
| 3281 | // This element is from right hand side vector operand |
| 3282 | // |
| 3283 | // If the value selected is a poison value, explicitly specify it |
| 3284 | // with a -1 mask value. (case 1) |
| 3285 | if (match(V: RHS, P: m_Poison())) |
| 3286 | eltMask = -1; |
| 3287 | // If RHS is going to be replaced (case 3 or 4), calculate the |
| 3288 | // new mask value for the element. |
| 3289 | else if (newRHS != RHS) { |
| 3290 | eltMask = RHSMask[Mask[i]-LHSWidth]; |
| 3291 | // If the value selected is an poison value, explicitly specify it |
| 3292 | // with a -1 mask value. |
| 3293 | if (eltMask >= (int)RHSOp0Width) { |
| 3294 | assert(match(RHSShuffle->getOperand(1), m_Poison()) && |
| 3295 | "should have been check above" ); |
| 3296 | eltMask = -1; |
| 3297 | } |
| 3298 | } else |
| 3299 | eltMask = Mask[i]-LHSWidth; |
| 3300 | |
| 3301 | // If LHS's width is changed, shift the mask value accordingly. |
| 3302 | // If newRHS == nullptr, i.e. LHSOp0 == RHSOp0, we want to remap any |
| 3303 | // references from RHSOp0 to LHSOp0, so we don't need to shift the mask. |
| 3304 | // If newRHS == newLHS, we want to remap any references from newRHS to |
| 3305 | // newLHS so that we can properly identify splats that may occur due to |
| 3306 | // obfuscation across the two vectors. |
| 3307 | if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS) |
| 3308 | eltMask += newLHSWidth; |
| 3309 | } |
| 3310 | |
| 3311 | // Check if this could still be a splat. |
| 3312 | if (eltMask >= 0) { |
| 3313 | if (SplatElt >= 0 && SplatElt != eltMask) |
| 3314 | isSplat = false; |
| 3315 | SplatElt = eltMask; |
| 3316 | } |
| 3317 | |
| 3318 | newMask.push_back(Elt: eltMask); |
| 3319 | } |
| 3320 | |
| 3321 | // If the result mask is equal to one of the original shuffle masks, |
| 3322 | // or is a splat, do the replacement. |
| 3323 | if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) { |
| 3324 | if (!newRHS) |
| 3325 | newRHS = PoisonValue::get(T: newLHS->getType()); |
| 3326 | return new ShuffleVectorInst(newLHS, newRHS, newMask); |
| 3327 | } |
| 3328 | |
| 3329 | return MadeChange ? &SVI : nullptr; |
| 3330 | } |
| 3331 | |
| 3332 | /// Given the following de-interleaving shufflevectors and the consuming zexts: |
| 3333 | /// ``` |
| 3334 | /// %f0 = shufflevector <8 x i32> %v, <4 x i32> <i32 0, i32 2, i32 4, i32 6> |
| 3335 | /// %f1 = shufflevector <8 x i32> %v, <4 x i32> <i32 1, i32 3, i32 5, i32 7> |
| 3336 | /// %z0 = zext <4 x i32> %f0 to <4 x i64> |
| 3337 | /// %z1 = zext <4 x i32> %f1 to <4 x i64> |
| 3338 | /// ``` |
| 3339 | /// We can actually bitcast the input value, `%v` first before replacing zexts |
| 3340 | /// with simple arithmetics on this new bitcast: |
| 3341 | /// ``` |
| 3342 | /// %bc = bitcast <8 x i32> %v to <4 x i64> |
| 3343 | // %z0 = and <4 x i64> %bc, splat (i64 4294967295) |
| 3344 | // %z1 = lshr <4 x i64> %bc, splat (i64 32) |
| 3345 | /// ``` |
| 3346 | /// This transformation is almost always benefitial as shufflevector is more |
| 3347 | /// expensive than normal arithmetics. |
| 3348 | Instruction * |
| 3349 | InstCombinerImpl::(ZExtInst &RootZExt) { |
| 3350 | // This pattern involves bitcast that is not compatible with big endian. |
| 3351 | if (DL.isBigEndian()) |
| 3352 | return nullptr; |
| 3353 | |
| 3354 | // The actual value that got de-interleaved. |
| 3355 | Value *DIV; |
| 3356 | |
| 3357 | using namespace PatternMatch; |
| 3358 | Instruction *SVI = nullptr, *DI = nullptr; |
| 3359 | if (!match( |
| 3360 | V: &RootZExt, |
| 3361 | P: m_ZExt(Op: m_CombineOr( |
| 3362 | Ps: m_ExtractValue(V: m_Instruction(I&: DI, P: m_Deinterleave2(Op: m_Value(V&: DIV)))), |
| 3363 | Ps: m_Instruction(I&: SVI, P: m_Shuffle(v1: m_Value(), v2: m_Value())))))) |
| 3364 | return nullptr; |
| 3365 | |
| 3366 | auto isDeinterleaveShuffle = |
| 3367 | [](Instruction *I) -> std::pair<Value *, unsigned> { |
| 3368 | Value *V; |
| 3369 | ArrayRef<int> ShuffleMask; |
| 3370 | unsigned Index; |
| 3371 | if (match(V: I, P: m_Shuffle(v1: m_Value(V), v2: m_Undef(), mask: m_Mask(ShuffleMask))) && |
| 3372 | isa<FixedVectorType>(Val: V->getType())) { |
| 3373 | unsigned NumInputElements = |
| 3374 | cast<VectorType>(Val: V->getType())->getElementCount().getFixedValue(); |
| 3375 | if (ShuffleVectorInst::isDeInterleaveMaskOfFactor(Mask: ShuffleMask, Factor: 2, |
| 3376 | Index) && |
| 3377 | Index < 2 && |
| 3378 | ShuffleVectorInst::isSingleSourceMask(Mask: ShuffleMask, |
| 3379 | NumSrcElts: NumInputElements) && |
| 3380 | ShuffleMask.size() * 2 == NumInputElements) |
| 3381 | return {V, Index}; |
| 3382 | } |
| 3383 | return {nullptr, UINT_MAX}; |
| 3384 | }; |
| 3385 | |
| 3386 | // Validate either the shufflevector or the vector.deinterleave2 and obtain |
| 3387 | // the value they're de-interleaving. |
| 3388 | if (SVI) { |
| 3389 | // We will find other shufflevectors later. |
| 3390 | DIV = isDeinterleaveShuffle(SVI).first; |
| 3391 | if (!DIV) |
| 3392 | return nullptr; |
| 3393 | } else { |
| 3394 | // We should already capture the value that got de-interleaved (i.e. DIV). |
| 3395 | assert(DI && DIV); |
| 3396 | if (!all_of(Range: DI->users(), P: [](User *Usr) -> bool { |
| 3397 | auto *EV = dyn_cast<ExtractValueInst>(Val: Usr); |
| 3398 | return EV && EV->getNumIndices() == 1; |
| 3399 | })) |
| 3400 | return nullptr; |
| 3401 | } |
| 3402 | |
| 3403 | auto *InputVecTy = dyn_cast<VectorType>(Val: DIV->getType()); |
| 3404 | if (!InputVecTy) |
| 3405 | return nullptr; |
| 3406 | auto *InElementTy = dyn_cast<IntegerType>(Val: InputVecTy->getElementType()); |
| 3407 | if (!InElementTy) |
| 3408 | return nullptr; |
| 3409 | if (!InputVecTy->getElementCount().isKnownEven()) |
| 3410 | return nullptr; |
| 3411 | |
| 3412 | // {Field instruction, Field index} |
| 3413 | SmallVector<std::pair<Instruction *, unsigned>, 4> Fields; |
| 3414 | if (SVI) { |
| 3415 | for (auto *Usr : DIV->users()) { |
| 3416 | auto *FieldI = dyn_cast<Instruction>(Val: Usr); |
| 3417 | if (!FieldI) |
| 3418 | continue; |
| 3419 | auto [V, Index] = isDeinterleaveShuffle(FieldI); |
| 3420 | if (V != DIV) |
| 3421 | continue; |
| 3422 | assert(Index < 2); |
| 3423 | // Find the earliest field extraction instruction. |
| 3424 | if (FieldI->getParent() != SVI->getParent()) |
| 3425 | continue; |
| 3426 | if (FieldI != SVI && FieldI->comesBefore(Other: SVI)) |
| 3427 | SVI = FieldI; |
| 3428 | Fields.push_back(Elt: {FieldI, Index}); |
| 3429 | } |
| 3430 | } else { |
| 3431 | // llvm.vector.deinterleave2. |
| 3432 | for (User *Field : DI->users()) { |
| 3433 | auto *FieldI = cast<ExtractValueInst>(Val: Field); |
| 3434 | unsigned FieldIdx = *FieldI->idx_begin(); |
| 3435 | assert(FieldIdx < 2); |
| 3436 | Fields.push_back(Elt: {FieldI, FieldIdx}); |
| 3437 | } |
| 3438 | } |
| 3439 | |
| 3440 | // {field to be replaced, field index} |
| 3441 | SmallVector<std::pair<ZExtInst *, unsigned>, 4> FieldReplacements; |
| 3442 | // We commit the transformation only if all the field users can be replaced, |
| 3443 | // otherwise the primary de-interleaving construction, regardless of |
| 3444 | // llvm.vector.deinterleave2 or shufflevectors, will still be there. |
| 3445 | for (auto [Field, FieldIdx] : Fields) { |
| 3446 | for (User *FieldUsr : Field->users()) { |
| 3447 | auto *ZExt = dyn_cast<ZExtInst>(Val: FieldUsr); |
| 3448 | if (!ZExt) |
| 3449 | return nullptr; |
| 3450 | // Only if it's doubling the element size. |
| 3451 | if (ZExt->getDestTy() != ZExt->getSrcTy()->getExtendedType()) |
| 3452 | return nullptr; |
| 3453 | FieldReplacements.push_back(Elt: {ZExt, FieldIdx}); |
| 3454 | } |
| 3455 | } |
| 3456 | |
| 3457 | // This will insert replacement instructions before all the fields users. |
| 3458 | Builder.SetInsertPoint(DI ? DI : SVI); |
| 3459 | |
| 3460 | // Double the element size but half the vector length. |
| 3461 | auto *BitcastedTy = VectorType::getExtendedElementVectorType(VTy: InputVecTy); |
| 3462 | BitcastedTy = VectorType::getHalfElementsVectorType(VTy: BitcastedTy); |
| 3463 | // Since we're going to "merge" lanes via bitcast, we need to freeze any |
| 3464 | // potential poison lanes first. |
| 3465 | Value *Freeze = Builder.CreateFreeze(V: DIV); |
| 3466 | Value *Bitcast = Builder.CreateBitCast(V: Freeze, DestTy: BitcastedTy); |
| 3467 | unsigned InElementBitWidth = InElementTy->getBitWidth(); |
| 3468 | auto Mask = APInt::getLowBitsSet(numBits: InElementBitWidth * 2, loBitsSet: InElementBitWidth); |
| 3469 | Value *NewField0 = Builder.CreateAnd(LHS: Bitcast, RHS: Mask); |
| 3470 | Value *NewField1 = Builder.CreateLShr(LHS: Bitcast, RHS: InElementBitWidth); |
| 3471 | |
| 3472 | for (auto [I, Idx] : FieldReplacements) { |
| 3473 | assert(Idx < 2 && "unsupported field index" ); |
| 3474 | replaceInstUsesWith(I&: *I, V: Idx ? NewField1 : NewField0); |
| 3475 | // Make sure the old ZExt are in the worklist so that they |
| 3476 | // can be removed in the following iterations. |
| 3477 | addToWorklist(I); |
| 3478 | } |
| 3479 | |
| 3480 | return &RootZExt; |
| 3481 | } |
| 3482 | |