| 1 | //===- SLPUtils.cpp - SLP Vectorizer free utility helpers -----------------===// |
| 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 | #include "SLPUtils.h" |
| 10 | |
| 11 | #include "llvm/ADT/STLExtras.h" |
| 12 | #include "llvm/ADT/Sequence.h" |
| 13 | #include "llvm/Analysis/ValueTracking.h" |
| 14 | #include "llvm/Analysis/VectorUtils.h" |
| 15 | #include "llvm/IR/Constants.h" |
| 16 | #include "llvm/IR/DerivedTypes.h" |
| 17 | #include "llvm/IR/Instructions.h" |
| 18 | #include "llvm/IR/IntrinsicInst.h" |
| 19 | #include "llvm/IR/PatternMatch.h" |
| 20 | #include "llvm/Support/Casting.h" |
| 21 | #include "llvm/Support/MathExtras.h" |
| 22 | #include "llvm/Support/raw_ostream.h" |
| 23 | |
| 24 | #include <algorithm> |
| 25 | #include <string> |
| 26 | #include <type_traits> |
| 27 | |
| 28 | using namespace llvm; |
| 29 | using namespace llvm::PatternMatch; |
| 30 | |
| 31 | namespace llvm::slpvectorizer { |
| 32 | |
| 33 | bool isConstant(Value *V) { |
| 34 | return isa<Constant>(Val: V) && !isa<ConstantExpr, GlobalValue>(Val: V); |
| 35 | } |
| 36 | |
| 37 | bool isVectorLikeInstWithConstOps(Value *V) { |
| 38 | auto *I = dyn_cast<Instruction>(Val: V); |
| 39 | // Non-instructions are vector-like only if they are undef. |
| 40 | if (!I) |
| 41 | return isa<UndefValue>(Val: V); |
| 42 | switch (I->getOpcode()) { |
| 43 | case Instruction::ExtractValue: |
| 44 | case Instruction::InsertValue: |
| 45 | return true; |
| 46 | case Instruction::ExtractElement: |
| 47 | return isa<FixedVectorType>(Val: I->getOperand(i: 0)->getType()) && |
| 48 | isConstant(V: I->getOperand(i: 1)); |
| 49 | case Instruction::InsertElement: |
| 50 | return isa<FixedVectorType>(Val: I->getOperand(i: 0)->getType()) && |
| 51 | isConstant(V: I->getOperand(i: 2)); |
| 52 | default: |
| 53 | return false; |
| 54 | } |
| 55 | } |
| 56 | |
| 57 | unsigned getNumElements(Type *Ty) { |
| 58 | assert(!isa<ScalableVectorType>(Ty) && |
| 59 | "ScalableVectorType is not supported." ); |
| 60 | if (isVectorizedTy(Ty)) |
| 61 | return getVectorizedTypeVF(Ty).getFixedValue(); |
| 62 | return 1; |
| 63 | } |
| 64 | |
| 65 | unsigned getPartNumElems(unsigned Size, unsigned NumParts) { |
| 66 | return std::min<unsigned>(a: Size, b: bit_ceil(Value: divideCeil(Numerator: Size, Denominator: NumParts))); |
| 67 | } |
| 68 | |
| 69 | unsigned getNumElems(unsigned Size, unsigned PartNumElems, unsigned Part) { |
| 70 | return std::min<unsigned>(a: PartNumElems, b: Size - Part * PartNumElems); |
| 71 | } |
| 72 | |
| 73 | #if !defined(NDEBUG) |
| 74 | std::string shortBundleName(ArrayRef<Value *> VL, int Idx) { |
| 75 | std::string Result; |
| 76 | raw_string_ostream OS(Result); |
| 77 | if (Idx >= 0) |
| 78 | OS << "Idx: " << Idx << ", " ; |
| 79 | OS << "n=" << VL.size() << " [" << *VL.front() << ", ..]" ; |
| 80 | return Result; |
| 81 | } |
| 82 | #endif |
| 83 | |
| 84 | bool allSameBlock(ArrayRef<Value *> VL) { |
| 85 | auto *It = find_if(Range&: VL, P: IsaPred<Instruction>); |
| 86 | if (It == VL.end()) |
| 87 | return false; |
| 88 | Instruction *I0 = cast<Instruction>(Val: *It); |
| 89 | if (all_of(Range&: VL, P: isVectorLikeInstWithConstOps)) |
| 90 | return true; |
| 91 | |
| 92 | BasicBlock *BB = I0->getParent(); |
| 93 | for (Value *V : iterator_range(It, VL.end())) { |
| 94 | if (isa<PoisonValue>(Val: V)) |
| 95 | continue; |
| 96 | auto *II = dyn_cast<Instruction>(Val: V); |
| 97 | if (!II) |
| 98 | return false; |
| 99 | |
| 100 | if (BB != II->getParent()) |
| 101 | return false; |
| 102 | } |
| 103 | return true; |
| 104 | } |
| 105 | |
| 106 | bool allConstant(ArrayRef<Value *> VL) { |
| 107 | // Constant expressions and globals can't be vectorized like normal integer/FP |
| 108 | // constants. |
| 109 | return all_of(Range&: VL, P: isConstant); |
| 110 | } |
| 111 | |
| 112 | bool isSplat(ArrayRef<Value *> VL) { |
| 113 | Value *FirstNonUndef = nullptr; |
| 114 | for (Value *V : VL) { |
| 115 | if (isa<UndefValue>(Val: V)) |
| 116 | continue; |
| 117 | if (!FirstNonUndef) { |
| 118 | FirstNonUndef = V; |
| 119 | continue; |
| 120 | } |
| 121 | if (V != FirstNonUndef) |
| 122 | return false; |
| 123 | } |
| 124 | return FirstNonUndef != nullptr; |
| 125 | } |
| 126 | |
| 127 | bool isCommutative(const Instruction *I, const Value *ValWithUses, |
| 128 | bool IsCopyable) { |
| 129 | if (auto *Cmp = dyn_cast<CmpInst>(Val: I)) |
| 130 | return Cmp->isCommutative(); |
| 131 | if (auto *BO = dyn_cast<BinaryOperator>(Val: I)) |
| 132 | return BO->isCommutative() || |
| 133 | (BO->getOpcode() == Instruction::Sub && ValWithUses->hasUseList() && |
| 134 | !ValWithUses->hasNUsesOrMore(N: UsesLimit) && |
| 135 | all_of( |
| 136 | Range: ValWithUses->uses(), |
| 137 | P: [&](const Use &U) { |
| 138 | // Commutative, if icmp eq/ne sub, 0 |
| 139 | CmpPredicate Pred; |
| 140 | if (match(V: U.getUser(), |
| 141 | P: m_ICmp(Pred, L: m_Specific(V: U.get()), R: m_Zero())) && |
| 142 | (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) |
| 143 | return true; |
| 144 | // Commutative, if abs(sub nsw, true) or abs(sub, false). |
| 145 | ConstantInt *Flag; |
| 146 | auto *I = dyn_cast<BinaryOperator>(Val: U.get()); |
| 147 | return match(V: U.getUser(), |
| 148 | P: m_Intrinsic<Intrinsic::abs>( |
| 149 | Ops: m_Specific(V: U.get()), Ops: m_ConstantInt(CI&: Flag))) && |
| 150 | ((!IsCopyable && I && !I->hasNoSignedWrap()) || |
| 151 | Flag->isOne()); |
| 152 | })) || |
| 153 | (BO->getOpcode() == Instruction::FSub && ValWithUses->hasUseList() && |
| 154 | !ValWithUses->hasNUsesOrMore(N: UsesLimit) && |
| 155 | all_of(Range: ValWithUses->uses(), P: [](const Use &U) { |
| 156 | return match(V: U.getUser(), |
| 157 | P: m_Intrinsic<Intrinsic::fabs>(Ops: m_Specific(V: U.get()))); |
| 158 | })); |
| 159 | return I->isCommutative(); |
| 160 | } |
| 161 | |
| 162 | bool isCommutative(const Instruction *I) { return isCommutative(I, ValWithUses: I); } |
| 163 | |
| 164 | bool isCommutableOperand(const Instruction *I, Value *ValWithUses, unsigned Op, |
| 165 | bool IsCopyable) { |
| 166 | assert(isCommutative(I, ValWithUses, IsCopyable) && |
| 167 | "The instruction is not commutative." ); |
| 168 | if (isa<CmpInst>(Val: I)) |
| 169 | return true; |
| 170 | if (auto *BO = dyn_cast<BinaryOperator>(Val: I)) { |
| 171 | switch (BO->getOpcode()) { |
| 172 | case Instruction::Sub: |
| 173 | case Instruction::FSub: |
| 174 | return true; |
| 175 | default: |
| 176 | break; |
| 177 | } |
| 178 | } |
| 179 | return I->isCommutableOperand(Op); |
| 180 | } |
| 181 | |
| 182 | unsigned getNumberOfPotentiallyCommutativeOps(Instruction *I) { |
| 183 | if (isa<IntrinsicInst>(Val: I) && isCommutative(I)) { |
| 184 | // IntrinsicInst::isCommutative returns true if swapping the first "two" |
| 185 | // arguments to the intrinsic produces the same result. |
| 186 | constexpr unsigned IntrinsicNumOperands = 2; |
| 187 | return IntrinsicNumOperands; |
| 188 | } |
| 189 | return I->getNumOperands(); |
| 190 | } |
| 191 | |
| 192 | std::optional<unsigned> getElementIndex(const Value *Inst, unsigned Offset) { |
| 193 | if (auto Index = getInsertExtractIndex<InsertElementInst>(Inst, Offset)) |
| 194 | return Index; |
| 195 | if (auto Index = getInsertExtractIndex<ExtractElementInst>(Inst, Offset)) |
| 196 | return Index; |
| 197 | |
| 198 | unsigned Index = Offset; |
| 199 | |
| 200 | const auto *IV = dyn_cast<InsertValueInst>(Val: Inst); |
| 201 | if (!IV) |
| 202 | return std::nullopt; |
| 203 | |
| 204 | Type *CurrentType = IV->getType(); |
| 205 | for (unsigned I : IV->indices()) { |
| 206 | if (const auto *ST = dyn_cast<StructType>(Val: CurrentType)) { |
| 207 | Index *= ST->getNumElements(); |
| 208 | CurrentType = ST->getElementType(N: I); |
| 209 | } else if (const auto *AT = dyn_cast<ArrayType>(Val: CurrentType)) { |
| 210 | Index *= AT->getNumElements(); |
| 211 | CurrentType = AT->getElementType(); |
| 212 | } else { |
| 213 | return std::nullopt; |
| 214 | } |
| 215 | Index += I; |
| 216 | } |
| 217 | return Index; |
| 218 | } |
| 219 | |
| 220 | bool allSameOpcode(ArrayRef<Value *> VL) { |
| 221 | auto *It = find_if(Range&: VL, P: IsaPred<Instruction>); |
| 222 | if (It == VL.end()) |
| 223 | return true; |
| 224 | Instruction *MainOp = cast<Instruction>(Val: *It); |
| 225 | unsigned Opcode = MainOp->getOpcode(); |
| 226 | bool IsCmpOp = isa<CmpInst>(Val: MainOp); |
| 227 | CmpInst::Predicate BasePred = IsCmpOp ? cast<CmpInst>(Val: MainOp)->getPredicate() |
| 228 | : CmpInst::BAD_ICMP_PREDICATE; |
| 229 | return all_of(Range: make_range(x: It, y: VL.end()), P: [&](Value *V) { |
| 230 | if (auto *CI = dyn_cast<CmpInst>(Val: V)) |
| 231 | return BasePred == CI->getPredicate(); |
| 232 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
| 233 | return I->getOpcode() == Opcode; |
| 234 | return isa<PoisonValue>(Val: V); |
| 235 | }); |
| 236 | } |
| 237 | |
| 238 | std::optional<unsigned> (const Instruction *E) { |
| 239 | unsigned Opcode = E->getOpcode(); |
| 240 | assert((Opcode == Instruction::ExtractElement || |
| 241 | Opcode == Instruction::ExtractValue) && |
| 242 | "Expected extractelement or extractvalue instruction." ); |
| 243 | if (Opcode == Instruction::ExtractElement) { |
| 244 | auto *CI = dyn_cast<ConstantInt>(Val: E->getOperand(i: 1)); |
| 245 | if (!CI) |
| 246 | return std::nullopt; |
| 247 | // Check if the index is out of bound. We can get the source vector from |
| 248 | // operand 0. |
| 249 | unsigned Idx = CI->getZExtValue(); |
| 250 | auto *EE = cast<ExtractElementInst>(Val: E); |
| 251 | const unsigned VF = getNumElements(Ty: EE->getVectorOperandType()); |
| 252 | if (Idx >= VF) |
| 253 | return std::nullopt; |
| 254 | return Idx; |
| 255 | } |
| 256 | auto *EI = cast<ExtractValueInst>(Val: E); |
| 257 | if (EI->getNumIndices() != 1) |
| 258 | return std::nullopt; |
| 259 | return *EI->idx_begin(); |
| 260 | } |
| 261 | |
| 262 | void inversePermutation(ArrayRef<unsigned> Indices, |
| 263 | SmallVectorImpl<int> &Mask) { |
| 264 | Mask.clear(); |
| 265 | const unsigned E = Indices.size(); |
| 266 | Mask.resize(N: E, NV: PoisonMaskElem); |
| 267 | for (unsigned I = 0; I < E; ++I) |
| 268 | Mask[Indices[I]] = I; |
| 269 | } |
| 270 | |
| 271 | void reorderScalars(SmallVectorImpl<Value *> &Scalars, ArrayRef<int> Mask) { |
| 272 | assert(!Mask.empty() && "Expected non-empty mask." ); |
| 273 | SmallVector<Value *> Prev(Scalars.size(), |
| 274 | PoisonValue::get(T: Scalars.front()->getType())); |
| 275 | Prev.swap(RHS&: Scalars); |
| 276 | for (unsigned I = 0, E = Prev.size(); I < E; ++I) |
| 277 | if (Mask[I] != PoisonMaskElem) |
| 278 | Scalars[Mask[I]] = Prev[I]; |
| 279 | } |
| 280 | |
| 281 | bool allSameType(ArrayRef<Value *> VL) { |
| 282 | assert(!VL.empty() && "Expected non-empty list of values." ); |
| 283 | Type *Ty = VL.consume_front()->getType(); |
| 284 | return all_of(Range&: VL, P: [&](Value *V) { return V->getType() == Ty; }); |
| 285 | } |
| 286 | |
| 287 | template <typename T> |
| 288 | std::optional<unsigned> (const Value *Inst, |
| 289 | unsigned Offset) { |
| 290 | static_assert(std::is_same_v<T, InsertElementInst> || |
| 291 | std::is_same_v<T, ExtractElementInst>, |
| 292 | "unsupported T" ); |
| 293 | const auto *IE = dyn_cast<T>(Inst); |
| 294 | if (!IE) |
| 295 | return std::nullopt; |
| 296 | // InsertElement: result is the vector, index is op 2. |
| 297 | // ExtractElement: result is scalar, vector is op 0, index is op 1. |
| 298 | constexpr bool IsInsert = std::is_same_v<T, InsertElementInst>; |
| 299 | Type *VecTy = IsInsert ? IE->getType() : IE->getOperand(0)->getType(); |
| 300 | const auto *VT = dyn_cast<FixedVectorType>(Val: VecTy); |
| 301 | if (!VT) |
| 302 | return std::nullopt; |
| 303 | const auto *CI = dyn_cast<ConstantInt>(IE->getOperand(IsInsert ? 2 : 1)); |
| 304 | if (!CI) |
| 305 | return std::nullopt; |
| 306 | if (CI->getValue().uge(VT->getNumElements())) |
| 307 | return std::nullopt; |
| 308 | unsigned Index = Offset; |
| 309 | Index *= VT->getNumElements(); |
| 310 | Index += CI->getZExtValue(); |
| 311 | return Index; |
| 312 | } |
| 313 | |
| 314 | // Only these two specializations are used; instantiate them here so the |
| 315 | // definition can stay out of the header. |
| 316 | template std::optional<unsigned> |
| 317 | getInsertExtractIndex<InsertElementInst>(const Value *, unsigned); |
| 318 | template std::optional<unsigned> |
| 319 | getInsertExtractIndex<ExtractElementInst>(const Value *, unsigned); |
| 320 | |
| 321 | bool areAllOperandsNonInsts(Value *V) { |
| 322 | auto *I = dyn_cast<Instruction>(Val: V); |
| 323 | if (!I) |
| 324 | return true; |
| 325 | return !mayHaveNonDefUseDependency(I: *I) && |
| 326 | all_of(Range: I->operands(), P: [I](Value *V) { |
| 327 | auto *IO = dyn_cast<Instruction>(Val: V); |
| 328 | if (!IO) |
| 329 | return true; |
| 330 | return isa<PHINode>(Val: IO) || IO->getParent() != I->getParent(); |
| 331 | }); |
| 332 | } |
| 333 | |
| 334 | bool isUsedOutsideBlock(Value *V) { |
| 335 | auto *I = dyn_cast<Instruction>(Val: V); |
| 336 | if (!I) |
| 337 | return true; |
| 338 | // Limits the number of uses to save compile time. |
| 339 | return !I->mayReadOrWriteMemory() && !I->hasNUsesOrMore(N: UsesLimit) && |
| 340 | all_of(Range: I->users(), P: [I](User *U) { |
| 341 | auto *IU = dyn_cast<Instruction>(Val: U); |
| 342 | if (!IU) |
| 343 | return true; |
| 344 | return IU->getParent() != I->getParent() || isa<PHINode>(Val: IU); |
| 345 | }); |
| 346 | } |
| 347 | |
| 348 | bool doesNotNeedToBeScheduled(Value *V) { |
| 349 | return areAllOperandsNonInsts(V) && isUsedOutsideBlock(V); |
| 350 | } |
| 351 | |
| 352 | bool doesNotNeedToSchedule(ArrayRef<Value *> VL) { |
| 353 | return !VL.empty() && |
| 354 | (all_of(Range&: VL, P: isUsedOutsideBlock) || all_of(Range&: VL, P: areAllOperandsNonInsts)); |
| 355 | } |
| 356 | |
| 357 | void transformScalarShuffleIndiciesToVector(unsigned VecTyNumElements, |
| 358 | SmallVectorImpl<int> &Mask) { |
| 359 | // The ShuffleBuilder implementation use shufflevector to splat an "element". |
| 360 | // But the element have different meaning for SLP (scalar) and REVEC |
| 361 | // (vector). We need to expand Mask into masks which shufflevector can use |
| 362 | // directly. |
| 363 | SmallVector<int> NewMask(Mask.size() * VecTyNumElements); |
| 364 | for (unsigned I : seq<unsigned>(Size: Mask.size())) |
| 365 | for (auto [J, MaskV] : enumerate(First: MutableArrayRef(NewMask).slice( |
| 366 | N: I * VecTyNumElements, M: VecTyNumElements))) |
| 367 | MaskV = Mask[I] == PoisonMaskElem ? PoisonMaskElem |
| 368 | : Mask[I] * VecTyNumElements + J; |
| 369 | Mask.swap(RHS&: NewMask); |
| 370 | } |
| 371 | |
| 372 | unsigned getShufflevectorNumGroups(ArrayRef<Value *> VL) { |
| 373 | if (VL.empty()) |
| 374 | return 0; |
| 375 | if (!all_of(Range&: VL, P: IsaPred<ShuffleVectorInst>)) |
| 376 | return 0; |
| 377 | auto *SV = cast<ShuffleVectorInst>(Val: VL.front()); |
| 378 | unsigned SVNumElements = |
| 379 | cast<FixedVectorType>(Val: SV->getOperand(i_nocapture: 0)->getType())->getNumElements(); |
| 380 | unsigned ShuffleMaskSize = SV->getShuffleMask().size(); |
| 381 | if (SVNumElements % ShuffleMaskSize != 0) |
| 382 | return 0; |
| 383 | unsigned GroupSize = SVNumElements / ShuffleMaskSize; |
| 384 | if (GroupSize == 0 || (VL.size() % GroupSize) != 0) |
| 385 | return 0; |
| 386 | unsigned NumGroup = 0; |
| 387 | for (size_t I = 0, E = VL.size(); I != E; I += GroupSize) { |
| 388 | auto *SV = cast<ShuffleVectorInst>(Val: VL[I]); |
| 389 | Value *Src = SV->getOperand(i_nocapture: 0); |
| 390 | ArrayRef<Value *> Group = VL.slice(N: I, M: GroupSize); |
| 391 | SmallBitVector ExpectedIndex(GroupSize); |
| 392 | if (!all_of(Range&: Group, P: [&](Value *V) { |
| 393 | auto *SV = cast<ShuffleVectorInst>(Val: V); |
| 394 | // From the same source. |
| 395 | if (SV->getOperand(i_nocapture: 0) != Src) |
| 396 | return false; |
| 397 | int Index; |
| 398 | if (!SV->isExtractSubvectorMask(Index)) |
| 399 | return false; |
| 400 | ExpectedIndex.set(Index / ShuffleMaskSize); |
| 401 | return true; |
| 402 | })) |
| 403 | return 0; |
| 404 | if (!ExpectedIndex.all()) |
| 405 | return 0; |
| 406 | ++NumGroup; |
| 407 | } |
| 408 | assert(NumGroup == (VL.size() / GroupSize) && "Unexpected number of groups" ); |
| 409 | return NumGroup; |
| 410 | } |
| 411 | |
| 412 | SmallVector<int> calculateShufflevectorMask(ArrayRef<Value *> VL) { |
| 413 | assert(getShufflevectorNumGroups(VL) && "Not supported shufflevector usage." ); |
| 414 | auto *SV = cast<ShuffleVectorInst>(Val: VL.front()); |
| 415 | unsigned SVNumElements = |
| 416 | cast<FixedVectorType>(Val: SV->getOperand(i_nocapture: 0)->getType())->getNumElements(); |
| 417 | SmallVector<int> Mask; |
| 418 | unsigned AccumulateLength = 0; |
| 419 | for (Value *V : VL) { |
| 420 | auto *SV = cast<ShuffleVectorInst>(Val: V); |
| 421 | for (int M : SV->getShuffleMask()) |
| 422 | Mask.push_back(Elt: M == PoisonMaskElem ? PoisonMaskElem |
| 423 | : AccumulateLength + M); |
| 424 | AccumulateLength += SVNumElements; |
| 425 | } |
| 426 | return Mask; |
| 427 | } |
| 428 | |
| 429 | SmallBitVector buildUseMask(int VF, ArrayRef<int> Mask, UseMask MaskArg) { |
| 430 | SmallBitVector UseMask(VF, true); |
| 431 | for (auto [Idx, Value] : enumerate(First&: Mask)) { |
| 432 | if (Value == PoisonMaskElem) { |
| 433 | if (MaskArg == UseMask::UndefsAsMask) |
| 434 | UseMask.reset(Idx); |
| 435 | continue; |
| 436 | } |
| 437 | if (MaskArg == UseMask::FirstArg && Value < VF) |
| 438 | UseMask.reset(Idx: Value); |
| 439 | else if (MaskArg == UseMask::SecondArg && Value >= VF) |
| 440 | UseMask.reset(Idx: Value - VF); |
| 441 | } |
| 442 | return UseMask; |
| 443 | } |
| 444 | |
| 445 | template <bool IsPoisonOnly> |
| 446 | SmallBitVector isUndefVector(const Value *V, const SmallBitVector &UseMask) { |
| 447 | SmallBitVector Res(UseMask.empty() ? 1 : UseMask.size(), true); |
| 448 | using T = std::conditional_t<IsPoisonOnly, PoisonValue, UndefValue>; |
| 449 | if (isa<T>(V)) |
| 450 | return Res; |
| 451 | auto *VecTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
| 452 | if (!VecTy) |
| 453 | return Res.reset(); |
| 454 | auto *C = dyn_cast<Constant>(Val: V); |
| 455 | if (!C) { |
| 456 | if (!UseMask.empty()) { |
| 457 | const Value *Base = V; |
| 458 | while (auto *II = dyn_cast<InsertElementInst>(Val: Base)) { |
| 459 | Base = II->getOperand(i_nocapture: 0); |
| 460 | if (isa<T>(II->getOperand(i_nocapture: 1))) |
| 461 | continue; |
| 462 | std::optional<unsigned> Idx = getElementIndex(Inst: II); |
| 463 | if (!Idx) { |
| 464 | Res.reset(); |
| 465 | return Res; |
| 466 | } |
| 467 | if (*Idx < UseMask.size() && !UseMask.test(Idx: *Idx)) |
| 468 | Res.reset(Idx: *Idx); |
| 469 | } |
| 470 | // TODO: Add analysis for shuffles here too. |
| 471 | if (V == Base) { |
| 472 | Res.reset(); |
| 473 | } else { |
| 474 | SmallBitVector SubMask(UseMask.size(), false); |
| 475 | Res &= isUndefVector<IsPoisonOnly>(Base, SubMask); |
| 476 | } |
| 477 | } else { |
| 478 | Res.reset(); |
| 479 | } |
| 480 | return Res; |
| 481 | } |
| 482 | for (unsigned I = 0, E = VecTy->getNumElements(); I != E; ++I) { |
| 483 | if (Constant *Elem = C->getAggregateElement(Elt: I)) |
| 484 | if (!isa<T>(Elem) && |
| 485 | (UseMask.empty() || (I < UseMask.size() && !UseMask.test(Idx: I)))) |
| 486 | Res.reset(Idx: I); |
| 487 | } |
| 488 | return Res; |
| 489 | } |
| 490 | |
| 491 | template SmallBitVector isUndefVector<false>(const Value *, |
| 492 | const SmallBitVector &); |
| 493 | template SmallBitVector isUndefVector<true>(const Value *, |
| 494 | const SmallBitVector &); |
| 495 | |
| 496 | bool (Value *Scalar, Instruction *UserInst, |
| 497 | TargetLibraryInfo *TLI, |
| 498 | const TargetTransformInfo *TTI) { |
| 499 | if (!UserInst) |
| 500 | return false; |
| 501 | unsigned Opcode = UserInst->getOpcode(); |
| 502 | switch (Opcode) { |
| 503 | case Instruction::Load: { |
| 504 | LoadInst *LI = cast<LoadInst>(Val: UserInst); |
| 505 | return (LI->getPointerOperand() == Scalar); |
| 506 | } |
| 507 | case Instruction::Store: { |
| 508 | StoreInst *SI = cast<StoreInst>(Val: UserInst); |
| 509 | return (SI->getPointerOperand() == Scalar); |
| 510 | } |
| 511 | case Instruction::Call: { |
| 512 | CallInst *CI = cast<CallInst>(Val: UserInst); |
| 513 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); |
| 514 | return any_of(Range: enumerate(First: CI->args()), P: [&](auto &&Arg) { |
| 515 | return isVectorIntrinsicWithScalarOpAtArg(ID, Arg.index(), TTI) && |
| 516 | Arg.value().get() == Scalar; |
| 517 | }); |
| 518 | } |
| 519 | default: |
| 520 | return false; |
| 521 | } |
| 522 | } |
| 523 | |
| 524 | MemoryLocation getLocation(Instruction *I) { |
| 525 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) |
| 526 | return MemoryLocation::get(SI); |
| 527 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) |
| 528 | return MemoryLocation::get(LI); |
| 529 | return MemoryLocation(); |
| 530 | } |
| 531 | |
| 532 | bool isSimple(Instruction *I) { |
| 533 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) |
| 534 | return LI->isSimple(); |
| 535 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) |
| 536 | return SI->isSimple(); |
| 537 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: I)) |
| 538 | return !MI->isVolatile(); |
| 539 | return true; |
| 540 | } |
| 541 | |
| 542 | void addMask(SmallVectorImpl<int> &Mask, ArrayRef<int> SubMask, |
| 543 | bool ExtendingManyInputs) { |
| 544 | if (SubMask.empty()) |
| 545 | return; |
| 546 | assert( |
| 547 | (!ExtendingManyInputs || SubMask.size() > Mask.size() || |
| 548 | // Check if input scalars were extended to match the size of other node. |
| 549 | (SubMask.size() == Mask.size() && Mask.back() == PoisonMaskElem)) && |
| 550 | "SubMask with many inputs support must be larger than the mask." ); |
| 551 | if (Mask.empty()) { |
| 552 | Mask.append(in_start: SubMask.begin(), in_end: SubMask.end()); |
| 553 | return; |
| 554 | } |
| 555 | SmallVector<int> NewMask(SubMask.size(), PoisonMaskElem); |
| 556 | int TermValue = std::min(a: Mask.size(), b: SubMask.size()); |
| 557 | for (int I = 0, E = SubMask.size(); I < E; ++I) { |
| 558 | if (SubMask[I] == PoisonMaskElem || |
| 559 | (!ExtendingManyInputs && |
| 560 | (SubMask[I] >= TermValue || Mask[SubMask[I]] >= TermValue))) |
| 561 | continue; |
| 562 | NewMask[I] = Mask[SubMask[I]]; |
| 563 | } |
| 564 | Mask.swap(RHS&: NewMask); |
| 565 | } |
| 566 | |
| 567 | void fixupOrderingIndices(MutableArrayRef<unsigned> Order) { |
| 568 | const size_t Sz = Order.size(); |
| 569 | SmallBitVector UnusedIndices(Sz, /*t=*/true); |
| 570 | SmallBitVector MaskedIndices(Sz); |
| 571 | for (unsigned I = 0; I < Sz; ++I) { |
| 572 | if (Order[I] < Sz) |
| 573 | UnusedIndices.reset(Idx: Order[I]); |
| 574 | else |
| 575 | MaskedIndices.set(I); |
| 576 | } |
| 577 | if (MaskedIndices.none()) |
| 578 | return; |
| 579 | assert(UnusedIndices.count() == MaskedIndices.count() && |
| 580 | "Non-synced masked/available indices." ); |
| 581 | int Idx = UnusedIndices.find_first(); |
| 582 | int MIdx = MaskedIndices.find_first(); |
| 583 | while (MIdx >= 0) { |
| 584 | assert(Idx >= 0 && "Indices must be synced." ); |
| 585 | Order[MIdx] = Idx; |
| 586 | Idx = UnusedIndices.find_next(Prev: Idx); |
| 587 | MIdx = MaskedIndices.find_next(Prev: MIdx); |
| 588 | } |
| 589 | } |
| 590 | |
| 591 | SmallBitVector getAltInstrMask(ArrayRef<Value *> VL, Type *ScalarTy, |
| 592 | unsigned Opcode0, unsigned Opcode1) { |
| 593 | unsigned ScalarTyNumElements = getNumElements(Ty: ScalarTy); |
| 594 | SmallBitVector OpcodeMask(VL.size() * ScalarTyNumElements, false); |
| 595 | for (unsigned Lane : seq<unsigned>(Size: VL.size())) { |
| 596 | if (isa<PoisonValue>(Val: VL[Lane])) |
| 597 | continue; |
| 598 | if (cast<Instruction>(Val: VL[Lane])->getOpcode() == Opcode1) |
| 599 | OpcodeMask.set(I: Lane * ScalarTyNumElements, |
| 600 | E: Lane * ScalarTyNumElements + ScalarTyNumElements); |
| 601 | } |
| 602 | return OpcodeMask; |
| 603 | } |
| 604 | |
| 605 | SmallVector<Constant *> replicateMask(ArrayRef<Constant *> Val, unsigned VF) { |
| 606 | assert(none_of(Val, [](Constant *C) { return C->getType()->isVectorTy(); }) && |
| 607 | "Expected scalar constants." ); |
| 608 | SmallVector<Constant *> NewVal(Val.size() * VF); |
| 609 | for (auto [I, V] : enumerate(First&: Val)) |
| 610 | std::fill_n(first: NewVal.begin() + I * VF, n: VF, value: V); |
| 611 | return NewVal; |
| 612 | } |
| 613 | |
| 614 | } // namespace llvm::slpvectorizer |
| 615 | |