| 1 | //===- AArch64.cpp - AArch64 ABI Implementation ---------------------------===// |
| 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 "llvm/ABI/FunctionInfo.h" |
| 10 | #include "llvm/ABI/TargetInfo.h" |
| 11 | #include "llvm/ABI/Types.h" |
| 12 | #include "llvm/ADT/STLExtras.h" |
| 13 | #include "llvm/ADT/SmallVector.h" |
| 14 | #include "llvm/Support/Casting.h" |
| 15 | #include "llvm/Support/ErrorHandling.h" |
| 16 | #include "llvm/Support/MathExtras.h" |
| 17 | #include "llvm/Support/TypeSize.h" |
| 18 | #include "llvm/Support/WithColor.h" |
| 19 | #include <algorithm> |
| 20 | #include <cstdint> |
| 21 | |
| 22 | namespace llvm { |
| 23 | namespace abi { |
| 24 | |
| 25 | class AArch64TargetInfo : public TargetInfo { |
| 26 | public: |
| 27 | AArch64TargetInfo(TypeBuilder &TB, const AArch64ABIOptions &Opts) |
| 28 | : TargetInfo(TB), Opts(Opts) {} |
| 29 | |
| 30 | const ABICompatInfo &getABICompatInfo() const override { |
| 31 | return Opts.CompatInfo; |
| 32 | } |
| 33 | |
| 34 | void computeInfo(FunctionInfo &FI) const override { |
| 35 | if (!maybeCommonClassifyReturnType(FI)) |
| 36 | FI.getReturnInfo() = |
| 37 | classifyReturnType(RetTy: FI.getReturnType(), IsVariadicFn: FI.isVariadic()); |
| 38 | |
| 39 | unsigned ArgNo = 0; |
| 40 | unsigned NSRN = 0, NPRN = 0; |
| 41 | for (auto &I : FI.arguments()) { |
| 42 | const bool IsNamedArg = |
| 43 | !FI.isVariadic() || ArgNo < FI.getNumRequiredArgs(); |
| 44 | ++ArgNo; |
| 45 | I.Info = classifyArgumentType(Ty: I.ABIType, IsVariadicFn: FI.isVariadic(), IsNamedArg, |
| 46 | CallingConvention: FI.getCallingConvention(), NSRN, NPRN); |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | private: |
| 51 | AArch64ABIOptions Opts; |
| 52 | |
| 53 | ArgInfo classifyReturnType(const Type *RetTy, bool IsVariadicFn) const; |
| 54 | ArgInfo classifyArgumentType(const Type *Ty, bool IsVariadicFn, |
| 55 | bool IsNamedArg, unsigned CallingConvention, |
| 56 | unsigned &NSRN, unsigned &NPRN) const; |
| 57 | |
| 58 | bool isDarwinPCS() const { return Opts.Kind == AArch64ABIKind::DarwinPCS; } |
| 59 | bool isSoftFloat() const { return Opts.Kind == AArch64ABIKind::AAPCSSoft; } |
| 60 | |
| 61 | const VectorType * |
| 62 | convertFixedToScalableVectorType(const VectorType *VT) const; |
| 63 | |
| 64 | ArgInfo coerceIllegalVector(const VectorType *VT, unsigned &NSRN, |
| 65 | unsigned &NPRN) const; |
| 66 | ArgInfo coerceAndExpandPureScalableAggregate( |
| 67 | const Type *Ty, bool IsNamedArg, unsigned NVec, unsigned NPred, |
| 68 | const SmallVectorImpl<const Type *> &UnpaddedCoerceToSeq, unsigned &NSRN, |
| 69 | unsigned &NPRN) const; |
| 70 | |
| 71 | bool isIllegalVectorType(const Type *Ty) const; |
| 72 | |
| 73 | bool passAsAggregateType(const Type *Ty) const; |
| 74 | bool passAsPureScalableType(const Type *Ty, unsigned &NV, unsigned &NP, |
| 75 | SmallVectorImpl<const Type *> &CoerceToSeq) const; |
| 76 | |
| 77 | void flattenType(const Type *Ty, |
| 78 | SmallVectorImpl<const Type *> &Flattened) const; |
| 79 | |
| 80 | bool isHomogeneousAggregateBaseType(const Type *Ty) const override; |
| 81 | bool isHomogeneousAggregateSmallEnough(const Type *Base, |
| 82 | uint64_t Members) const override; |
| 83 | bool isZeroLengthBitfieldPermittedInHomogeneousAggregate() const override; |
| 84 | bool isPermittedToBeHomogeneousAggregate(const RecordType *RT) const override; |
| 85 | }; |
| 86 | |
| 87 | std::unique_ptr<TargetInfo> |
| 88 | createAArch64TargetInfo(TypeBuilder &TB, const AArch64ABIOptions &Opts) { |
| 89 | return std::make_unique<AArch64TargetInfo>(args&: TB, args: Opts); |
| 90 | } |
| 91 | |
| 92 | static void reportNYI(StringRef Feature) { |
| 93 | WithColor::warning() |
| 94 | << Feature |
| 95 | << " is not yet implemented for AArch64 in the LLVM ABI library.\n" ; |
| 96 | } |
| 97 | |
| 98 | ArgInfo AArch64TargetInfo::classifyReturnType(const Type *RetTy, |
| 99 | bool IsVariadicFn) const { |
| 100 | if (RetTy->isVoid()) |
| 101 | return ArgInfo::getIgnore(); |
| 102 | |
| 103 | if (const auto *VT = dyn_cast<VectorType>(Val: RetTy)) { |
| 104 | if (VT->isFixedLengthSVEData() || VT->isFixedLengthSVEPredicate()) { |
| 105 | unsigned NSRN = 0, NPRN = 0; |
| 106 | return coerceIllegalVector(VT, NSRN, NPRN); |
| 107 | } |
| 108 | |
| 109 | // Large vector types should be returned via memory. |
| 110 | if (VT->getABISizeInBits() > 128) |
| 111 | return getNaturalAlignIndirect(Ty: RetTy, AddrSpace: getAllocaAddrSpace()); |
| 112 | } |
| 113 | |
| 114 | if (!passAsAggregateType(Ty: RetTy)) { |
| 115 | if (const auto *IntTy = dyn_cast<IntegerType>(Val: RetTy)) { |
| 116 | if (IntTy->isBitInt()) |
| 117 | if (RetTy->getSizeInBits().getFixedValue() > 128) |
| 118 | return getNaturalAlignIndirect(Ty: RetTy, AddrSpace: getAllocaAddrSpace()); |
| 119 | |
| 120 | if (isPromotableInteger(IT: IntTy) && isDarwinPCS()) |
| 121 | return ArgInfo::getExtend(T: IntTy); |
| 122 | } |
| 123 | |
| 124 | // Everything not handled above is returned directly. |
| 125 | return ArgInfo::getDirect(); |
| 126 | } |
| 127 | |
| 128 | uint64_t Size = RetTy->getFixedSizeInBitsOrZero(); |
| 129 | if (!RetTy->isSVESizelessType() && (RetTy->isEmptyRecord() || Size == 0)) |
| 130 | return ArgInfo::getIgnore(); |
| 131 | |
| 132 | const Type *Base = nullptr; |
| 133 | uint64_t Members = 0; |
| 134 | if (isHomogeneousAggregate(Ty: RetTy, Base, Members) && |
| 135 | !(Opts.IsILP32 && IsVariadicFn)) { |
| 136 | // Homogeneous Floating-point Aggregates (HFAs) are returned directly. |
| 137 | return ArgInfo::getDirect(); |
| 138 | } |
| 139 | |
| 140 | // In AAPCS return values of a Pure Scalable type are treated as a single |
| 141 | // named argument and passed expanded in registers, or indirectly if there are |
| 142 | // not enough registers. |
| 143 | if (Opts.Kind == AArch64ABIKind::AAPCS) { |
| 144 | unsigned NSRN = 0, NPRN = 0; |
| 145 | unsigned NVec = 0, NPred = 0; |
| 146 | SmallVector<const Type *> UnpaddedCoerceToSeq; |
| 147 | if (passAsPureScalableType(Ty: RetTy, NV&: NVec, NP&: NPred, CoerceToSeq&: UnpaddedCoerceToSeq) && |
| 148 | (NVec + NPred) > 0) |
| 149 | return coerceAndExpandPureScalableAggregate( |
| 150 | Ty: RetTy, /*IsNamedArg=*/true, NVec, NPred, UnpaddedCoerceToSeq, NSRN, |
| 151 | NPRN); |
| 152 | } |
| 153 | |
| 154 | reportNYI(Feature: "Aggregate return type handling" ); |
| 155 | return ArgInfo::getIgnore(); |
| 156 | } |
| 157 | |
| 158 | ArgInfo AArch64TargetInfo::classifyArgumentType( |
| 159 | const Type *Ty, bool IsVariadicFn, bool IsNamedArg, |
| 160 | unsigned CallingConvention, unsigned &NSRN, unsigned &NPRN) const { |
| 161 | Ty = useFirstFieldIfTransparentUnion(Ty); |
| 162 | |
| 163 | // Arm64EC variadic functions classify their arguments with the x86-64 |
| 164 | // rules rather than the AArch64 ones. |
| 165 | if (IsVariadicFn && Opts.IsWindowsArm64EC) { |
| 166 | reportNYI(Feature: "Arm64EC variadic argument handling" ); |
| 167 | return ArgInfo::getIgnore(); |
| 168 | } |
| 169 | |
| 170 | // Handle illegal vector types here. |
| 171 | if (isIllegalVectorType(Ty)) |
| 172 | return coerceIllegalVector(VT: cast<VectorType>(Val: Ty), NSRN, NPRN); |
| 173 | |
| 174 | if (!passAsAggregateType(Ty)) { |
| 175 | if (const auto *IntTy = dyn_cast<IntegerType>(Val: Ty)) { |
| 176 | if (IntTy->isBitInt()) |
| 177 | if (Ty->getSizeInBits().getFixedValue() > 128) |
| 178 | return getNaturalAlignIndirect(Ty, AddrSpace: getAllocaAddrSpace(), |
| 179 | /*ByVal=*/false); |
| 180 | |
| 181 | if (isPromotableInteger(IT: IntTy) && isDarwinPCS()) |
| 182 | return ArgInfo::getExtend(T: IntTy); |
| 183 | } |
| 184 | |
| 185 | // Predicates and svcount_t are passed in a predicate register. Legal |
| 186 | // vectors, SVE data vectors, and floating-point types are passed in a |
| 187 | // SIMD and floating-point register. A tuple occupies one register of the |
| 188 | // appropriate kind per vector it contains. |
| 189 | if (const auto *VT = dyn_cast<VectorType>(Val: Ty)) { |
| 190 | if (VT->isSVEPredicate() || VT->isSVECount()) |
| 191 | NPRN = std::min(a: NPRN + 1, b: 4u); |
| 192 | else |
| 193 | NSRN = std::min(a: NSRN + 1, b: 8u); |
| 194 | } else if (const auto *TT = dyn_cast<TupleType>(Val: Ty)) { |
| 195 | if (TT->getVectorType()->isSVEPredicate()) |
| 196 | NPRN = std::min(a: NPRN + TT->getNumVectors(), b: 4u); |
| 197 | else |
| 198 | NSRN = std::min(a: NSRN + TT->getNumVectors(), b: 8u); |
| 199 | } else if (Ty->isFloat()) { |
| 200 | NSRN = std::min(a: NSRN + 1, b: 8u); |
| 201 | } |
| 202 | |
| 203 | // Everything not handled above is returned directly. |
| 204 | return ArgInfo::getDirect(); |
| 205 | } |
| 206 | |
| 207 | // Structures with either a non-trivial destructor or a non-trivial |
| 208 | // copy constructor are always indirect. |
| 209 | if (auto RecordRAA = getRecordArgABI(Ty)) { |
| 210 | return getNaturalAlignIndirect(Ty, AddrSpace: getAllocaAddrSpace(), |
| 211 | /*ByVal=*/RecordRAA == |
| 212 | RecordArgABI::RAA_DirectInMemory); |
| 213 | } |
| 214 | |
| 215 | // AAPCS64 does not say that empty C records are ignored as arguments, |
| 216 | // but other compilers do so in certain situations, and we copy that behavior. |
| 217 | uint64_t Size = Ty->getFixedSizeInBitsOrZero(); |
| 218 | if (!Ty->isSVESizelessType() && (Ty->isEmptyRecord() || Size == 0)) { |
| 219 | // Darwin overrides the psABI here to ignore all empty records in all modes. |
| 220 | // The ABI explicitly says that an empty class shall be treated as if its |
| 221 | // type were an aggregate with a single member of type unsigned byte. |
| 222 | if (!Opts.IsCXX || isDarwinPCS()) |
| 223 | return ArgInfo::getIgnore(); |
| 224 | |
| 225 | // In C++ mode, arguments which have sizeof() == 0 (which are non-standard |
| 226 | // C++) are ignored. This isn't defined by any standard, so we copy GCC's |
| 227 | // behaviour here. |
| 228 | if (Size == 0) |
| 229 | return ArgInfo::getIgnore(); |
| 230 | } |
| 231 | |
| 232 | // Homogeneous Floating-point Aggregates (HFAs) need to be expanded. |
| 233 | const Type *Base = nullptr; |
| 234 | uint64_t Members = 0; |
| 235 | bool IsWin64 = Opts.Kind == AArch64ABIKind::Win64 || |
| 236 | CallingConvention == llvm::CallingConv::Win64; |
| 237 | bool IsWinVariadic = IsWin64 && IsVariadicFn; |
| 238 | // In variadic functions on Windows, all composite types are treated alike, |
| 239 | // no special handling of HFAs/HVAs. |
| 240 | if (!IsWinVariadic && isHomogeneousAggregate(Ty, Base, Members)) { |
| 241 | NSRN = std::min(a: NSRN + Members, b: uint64_t(8)); |
| 242 | uint64_t BaseAllocSizeInBits = Base->getTypeAllocSize().getFixedValue() * 8; |
| 243 | const Type *CoerceTy = |
| 244 | TB.getArrayType(ElementType: Base, NumElements: Members, SizeInBits: Members * BaseAllocSizeInBits); |
| 245 | if (Opts.Kind != AArch64ABIKind::AAPCS) |
| 246 | return ArgInfo::getDirect(T: CoerceTy); |
| 247 | |
| 248 | // For HFAs/HVAs, cap the argument alignment to 16, otherwise |
| 249 | // set it to 8 according to the AAPCS64 document. |
| 250 | unsigned TyAlign = Ty->getUnadjustedAlignment().value(); |
| 251 | TyAlign = (TyAlign >= 16) ? 16 : 8; |
| 252 | return ArgInfo::getDirect(T: CoerceTy, /*Offset=*/0, Align: llvm::Align(TyAlign)); |
| 253 | } |
| 254 | |
| 255 | // In AAPCS, named arguments of a Pure Scalable Type are passed expanded |
| 256 | // in registers, or indirectly if there are not enough registers. |
| 257 | if (Opts.Kind == AArch64ABIKind::AAPCS) { |
| 258 | unsigned NVec = 0, NPred = 0; |
| 259 | SmallVector<const Type *> UnpaddedCoerceToSeq; |
| 260 | if (passAsPureScalableType(Ty, NV&: NVec, NP&: NPred, CoerceToSeq&: UnpaddedCoerceToSeq) && |
| 261 | (NVec + NPred) > 0) |
| 262 | return coerceAndExpandPureScalableAggregate( |
| 263 | Ty, IsNamedArg, NVec, NPred, UnpaddedCoerceToSeq, NSRN, NPRN); |
| 264 | } |
| 265 | |
| 266 | reportNYI(Feature: "Aggregate argument type handling" ); |
| 267 | return ArgInfo::getIgnore(); |
| 268 | } |
| 269 | |
| 270 | bool AArch64TargetInfo::passAsAggregateType(const Type *Ty) const { |
| 271 | if (Opts.Kind == AArch64ABIKind::AAPCS && Ty->isSVESizelessType()) { |
| 272 | // svcount_t and the single-vector types occupy a register of their own, |
| 273 | // so only the data and predicate tuples are passed as aggregates. |
| 274 | const auto *TupleTy = dyn_cast<TupleType>(Val: Ty); |
| 275 | assert((!TupleTy || TupleTy->getNumVectors() > 1) && |
| 276 | "unexpected single vector tuple" ); |
| 277 | return TupleTy && !TupleTy->getVectorType()->isSVECount(); |
| 278 | } |
| 279 | return isAggregateTypeForABI(Ty); |
| 280 | } |
| 281 | |
| 282 | /// Returns the scalable vector type that \p VT, a fixed-length SVE vector, |
| 283 | /// is passed as. A scalable SVE vector holds 128 bits per granule, so the |
| 284 | /// scalable element count is 128 divided by the element size, regardless of |
| 285 | /// how many elements the fixed-length type has. |
| 286 | const VectorType *AArch64TargetInfo::convertFixedToScalableVectorType( |
| 287 | const VectorType *VT) const { |
| 288 | // TODO: Verify that this correctly handles MFloat8 when we decide on a |
| 289 | // mapping for that type. |
| 290 | |
| 291 | if (VT->isFixedLengthSVEPredicate()) |
| 292 | return TB.getScalablePredicateOrCountVectorType(ABIAlign: Align(2), |
| 293 | Kind: VectorKind::SVEPredicate); |
| 294 | |
| 295 | assert(VT->isFixedLengthSVEData() && "expected a fixed-length SVE vector!" ); |
| 296 | |
| 297 | const Type *EltTy = VT->getElementType(); |
| 298 | uint64_t EltBits = EltTy->getSizeInBits().getFixedValue(); |
| 299 | assert(EltBits >= 8 && EltBits <= 64 && isPowerOf2_64(EltBits) && |
| 300 | "unexpected element type for SVE data vector!" ); |
| 301 | |
| 302 | return TB.getVectorType(ElementType: EltTy, NumElements: ElementCount::getScalable(MinVal: 128 / EltBits), |
| 303 | Align: llvm::Align(16), VecKind: VectorKind::SVEData); |
| 304 | } |
| 305 | |
| 306 | ArgInfo AArch64TargetInfo::coerceIllegalVector(const VectorType *VT, |
| 307 | unsigned &NSRN, |
| 308 | unsigned &NPRN) const { |
| 309 | if (VT->isFixedLengthSVEPredicate()) { |
| 310 | // Fixed-length predicates are described with 8-bit elements, but they are |
| 311 | // passed in a predicate register as a scalable vector of 16 one-bit |
| 312 | // elements. |
| 313 | assert(isa<IntegerType>(VT->getElementType()) && |
| 314 | VT->getElementType()->getSizeInBits().getFixedValue() == 8 && |
| 315 | "unexpected element type for SVE predicate!" ); |
| 316 | NPRN = std::min(a: NPRN + 1, b: 4u); |
| 317 | return ArgInfo::getDirect(T: TB.getScalablePredicateOrCountVectorType( |
| 318 | ABIAlign: Align(2), Kind: VectorKind::SVEPredicate)); |
| 319 | } |
| 320 | |
| 321 | if (VT->isFixedLengthSVEData()) { |
| 322 | NSRN = std::min(a: NSRN + 1, b: 8u); |
| 323 | return ArgInfo::getDirect(T: convertFixedToScalableVectorType(VT)); |
| 324 | } |
| 325 | |
| 326 | uint64_t Size = VT->getABISizeInBits(); |
| 327 | // Android promotes <2 x i8> to i16, not i32 |
| 328 | if (Opts.IsAndroidOrOHOS && (Size <= 16)) { |
| 329 | auto *ResType = TB.getIntegerType(BitWidth: 16, Align: llvm::Align(2), /*Signed=*/false); |
| 330 | return ArgInfo::getDirect(T: ResType); |
| 331 | } |
| 332 | const Type *I32 = TB.getIntegerType(BitWidth: 32, Align: llvm::Align(4), /*Signed=*/false); |
| 333 | if (Size <= 32) |
| 334 | return ArgInfo::getDirect(T: I32); |
| 335 | if (Size == 64) { |
| 336 | NSRN = std::min(a: NSRN + 1, b: 8u); |
| 337 | return ArgInfo::getDirect( |
| 338 | T: TB.getVectorType(ElementType: I32, NumElements: ElementCount::getFixed(MinVal: 2), Align: llvm::Align(8))); |
| 339 | } |
| 340 | if (Size == 128) { |
| 341 | NSRN = std::min(a: NSRN + 1, b: 8u); |
| 342 | return ArgInfo::getDirect( |
| 343 | T: TB.getVectorType(ElementType: I32, NumElements: ElementCount::getFixed(MinVal: 4), Align: llvm::Align(16))); |
| 344 | } |
| 345 | |
| 346 | return getNaturalAlignIndirect(Ty: VT, AddrSpace: getAllocaAddrSpace(), /*ByVal=*/false); |
| 347 | } |
| 348 | |
| 349 | bool AArch64TargetInfo::isIllegalVectorType(const Type *Ty) const { |
| 350 | if (const auto *VT = dyn_cast<VectorType>(Val: Ty)) { |
| 351 | // Check whether VT is a fixed-length SVE vector. These types are |
| 352 | // represented as scalable vectors in function args/return and must be |
| 353 | // coerced from fixed vectors. |
| 354 | if (VT->isFixedLengthSVEData() || VT->isFixedLengthSVEPredicate()) |
| 355 | return true; |
| 356 | |
| 357 | // Scalable SVE types are legal. |
| 358 | if (VT->isScalable()) |
| 359 | return false; |
| 360 | |
| 361 | // Check whether VT is legal. |
| 362 | unsigned NumElements = VT->getNumElements().getFixedValue(); |
| 363 | uint64_t Size = VT->getABISizeInBits(); |
| 364 | // NumElements should be power of 2. |
| 365 | if (!isPowerOf2_32(Value: NumElements)) |
| 366 | return true; |
| 367 | |
| 368 | // arm64_32 has to be compatible with the ARM logic here, which allows huge |
| 369 | // vectors for some reason. |
| 370 | if (Opts.IsILP32 && Opts.IsMachO) |
| 371 | return Size <= 32; |
| 372 | |
| 373 | return Size != 64 && (Size != 128 || NumElements == 1); |
| 374 | } |
| 375 | return false; |
| 376 | } |
| 377 | |
| 378 | // Expand a memory type into a sequence with an element for each non-record, |
| 379 | // non-array member of the type, with the exception of the padding types, which |
| 380 | // are retained. |
| 381 | void AArch64TargetInfo::flattenType( |
| 382 | const Type *Ty, SmallVectorImpl<const Type *> &Flattened) const { |
| 383 | if (ArgInfo::isPaddingForCoerceAndExpand(Ty)) { |
| 384 | Flattened.push_back(Elt: Ty); |
| 385 | return; |
| 386 | } |
| 387 | |
| 388 | if (const auto *AT = dyn_cast<ArrayType>(Val: Ty)) { |
| 389 | uint64_t NElt = AT->getNumElements(); |
| 390 | if (NElt == 0) |
| 391 | return; |
| 392 | |
| 393 | SmallVector<const Type *, 4> EltFlattened; |
| 394 | flattenType(Ty: AT->getElementType(), Flattened&: EltFlattened); |
| 395 | |
| 396 | for (uint64_t I = 0; I < NElt; ++I) |
| 397 | llvm::append_range(C&: Flattened, R&: EltFlattened); |
| 398 | return; |
| 399 | } |
| 400 | |
| 401 | if (const auto *RT = dyn_cast<RecordType>(Val: Ty)) { |
| 402 | for (const FieldInfo &Field : RT->getFields()) |
| 403 | flattenType(Ty: Field.FieldType, Flattened); |
| 404 | return; |
| 405 | } |
| 406 | |
| 407 | Flattened.push_back(Elt: Ty); |
| 408 | } |
| 409 | |
| 410 | ArgInfo AArch64TargetInfo::coerceAndExpandPureScalableAggregate( |
| 411 | const Type *Ty, bool IsNamedArg, unsigned NVec, unsigned NPred, |
| 412 | const SmallVectorImpl<const Type *> &UnpaddedCoerceToSeq, unsigned &NSRN, |
| 413 | unsigned &NPRN) const { |
| 414 | // An unnamed argument, or one that does not fit in the remaining Z or P |
| 415 | // registers, is passed indirectly and does not consume those registers. |
| 416 | if (!IsNamedArg || NSRN + NVec > 8 || NPRN + NPred > 4) |
| 417 | return getNaturalAlignIndirect(Ty, AddrSpace: getAllocaAddrSpace(), /*ByVal=*/false); |
| 418 | |
| 419 | NSRN += NVec; |
| 420 | NPRN += NPred; |
| 421 | |
| 422 | // A sizeless SVE tuple is already one register per member. |
| 423 | if (Ty->isSVESizelessType()) |
| 424 | return ArgInfo::getDirect(); |
| 425 | |
| 426 | assert(!UnpaddedCoerceToSeq.empty() && "pure scalable type has no members" ); |
| 427 | const Type *UnpaddedCoerceToType = |
| 428 | UnpaddedCoerceToSeq.size() == 1 |
| 429 | ? UnpaddedCoerceToSeq[0] |
| 430 | : getStructOfTypes(Elems: UnpaddedCoerceToSeq, /*Packed=*/true); |
| 431 | |
| 432 | SmallVector<const Type *, 8> CoerceToSeq; |
| 433 | flattenType(Ty: convertTypeForMem(Ty), Flattened&: CoerceToSeq); |
| 434 | return ArgInfo::getCoerceAndExpand( |
| 435 | CoerceToType: getStructOfTypes(Elems: CoerceToSeq, /*Packed=*/false), UnpaddedCoerceToType); |
| 436 | } |
| 437 | |
| 438 | // A Pure Scalable Type (AAPCS64) is passed in Z and P registers. On success |
| 439 | // NVec and NPred are how many of each it needs, and CoerceToSeq has one |
| 440 | // scalable vector per register. A sequence longer than 12 is rejected so the |
| 441 | // caller treats the type as a large composite. |
| 442 | bool AArch64TargetInfo::passAsPureScalableType( |
| 443 | const Type *Ty, unsigned &NVec, unsigned &NPred, |
| 444 | SmallVectorImpl<const Type *> &CoerceToSeq) const { |
| 445 | if (const auto *AT = dyn_cast<ArrayType>(Val: Ty)) { |
| 446 | if (AT->isMatrixType()) |
| 447 | return false; |
| 448 | |
| 449 | uint64_t NElt = AT->getNumElements(); |
| 450 | if (NElt == 0) |
| 451 | return false; |
| 452 | |
| 453 | unsigned NV = 0, NP = 0; |
| 454 | SmallVector<const Type *, 4> EltCoerceToSeq; |
| 455 | if (!passAsPureScalableType(Ty: AT->getElementType(), NVec&: NV, NPred&: NP, CoerceToSeq&: EltCoerceToSeq)) |
| 456 | return false; |
| 457 | |
| 458 | if (CoerceToSeq.size() + NElt * EltCoerceToSeq.size() > 12) |
| 459 | return false; |
| 460 | |
| 461 | for (uint64_t I = 0; I < NElt; ++I) |
| 462 | llvm::append_range(C&: CoerceToSeq, R&: EltCoerceToSeq); |
| 463 | |
| 464 | NVec += NElt * NV; |
| 465 | NPred += NElt * NP; |
| 466 | return true; |
| 467 | } |
| 468 | |
| 469 | if (const auto *RT = dyn_cast<RecordType>(Val: Ty)) { |
| 470 | if (getRecordArgABI(RT) != RAA_Default) |
| 471 | return false; |
| 472 | // Pure scalable types are never unions and never contain unions. |
| 473 | if (RT->isUnion()) |
| 474 | return false; |
| 475 | |
| 476 | // A flexible array member is lowered as a zero-length array, which the |
| 477 | // field walk below skips. The member disqualifies a pure scalable type. |
| 478 | if (RT->hasFlexibleArrayMember()) |
| 479 | return false; |
| 480 | |
| 481 | // Direct virtual bases are not in getBaseClasses(). A record that has |
| 482 | // one cannot be passed in registers, and getRecordArgABI rejected it |
| 483 | // above. |
| 484 | for (const FieldInfo &Base : RT->getBaseClasses()) { |
| 485 | if (Base.FieldType->isEmptyRecord()) |
| 486 | continue; |
| 487 | if (!passAsPureScalableType(Ty: Base.FieldType, NVec, NPred, CoerceToSeq)) |
| 488 | return false; |
| 489 | } |
| 490 | for (const FieldInfo &Field : RT->getFields()) { |
| 491 | if (Field.isEmpty()) |
| 492 | continue; |
| 493 | if (!passAsPureScalableType(Ty: Field.FieldType, NVec, NPred, CoerceToSeq)) |
| 494 | return false; |
| 495 | } |
| 496 | return true; |
| 497 | } |
| 498 | |
| 499 | if (const auto *TT = dyn_cast<TupleType>(Val: Ty)) { |
| 500 | const VectorType *VT = TT->getVectorType(); |
| 501 | if (!VT->isScalable() || VT->isSVECount()) |
| 502 | return false; |
| 503 | |
| 504 | unsigned N = TT->getNumVectors(); |
| 505 | if (CoerceToSeq.size() + N > 12) |
| 506 | return false; |
| 507 | |
| 508 | bool IsPred = VT->isSVEPredicate(); |
| 509 | if (!IsPred && !VT->isSVEData()) |
| 510 | return false; |
| 511 | if (IsPred) |
| 512 | NPred += N; |
| 513 | else |
| 514 | NVec += N; |
| 515 | |
| 516 | for (unsigned I = 0; I < N; ++I) |
| 517 | CoerceToSeq.push_back(Elt: VT); |
| 518 | return true; |
| 519 | } |
| 520 | |
| 521 | if (const auto *VT = dyn_cast<VectorType>(Val: Ty)) { |
| 522 | const Type *Coerced = nullptr; |
| 523 | bool IsPred = false; |
| 524 | if (VT->isFixedLengthSVEPredicate()) { |
| 525 | IsPred = true; |
| 526 | Coerced = convertFixedToScalableVectorType(VT); |
| 527 | } else if (VT->isFixedLengthSVEData()) { |
| 528 | Coerced = convertFixedToScalableVectorType(VT); |
| 529 | } else if (VT->isScalable() && VT->isSVEPredicate()) { |
| 530 | IsPred = true; |
| 531 | Coerced = VT; |
| 532 | } else if (VT->isScalable() && VT->isSVEData()) { |
| 533 | Coerced = VT; |
| 534 | } else { |
| 535 | return false; |
| 536 | } |
| 537 | |
| 538 | if (CoerceToSeq.size() + 1 > 12) |
| 539 | return false; |
| 540 | |
| 541 | if (IsPred) |
| 542 | ++NPred; |
| 543 | else |
| 544 | ++NVec; |
| 545 | CoerceToSeq.push_back(Elt: Coerced); |
| 546 | return true; |
| 547 | } |
| 548 | |
| 549 | return false; |
| 550 | } |
| 551 | |
| 552 | bool AArch64TargetInfo::isHomogeneousAggregateBaseType(const Type *Ty) const { |
| 553 | // Soft-float ABI: no types are homogeneous aggregates. |
| 554 | if (isSoftFloat()) |
| 555 | return false; |
| 556 | |
| 557 | // Homogeneous aggregates for AAPCS64 must have base types of a floating |
| 558 | // point type or a short-vector type. |
| 559 | if (Ty->isFloat()) |
| 560 | return true; |
| 561 | |
| 562 | if (const auto *VT = dyn_cast<VectorType>(Val: Ty)) { |
| 563 | if (VT->isScalable() || VT->isSVEData() || VT->isSVEPredicate()) |
| 564 | return false; |
| 565 | |
| 566 | uint64_t VecSize = VT->getABISizeInBits(); |
| 567 | if (VecSize == 64 || VecSize == 128) |
| 568 | return true; |
| 569 | } |
| 570 | return false; |
| 571 | } |
| 572 | |
| 573 | bool AArch64TargetInfo::isHomogeneousAggregateSmallEnough( |
| 574 | const Type * /*Base*/, uint64_t Members) const { |
| 575 | return Members <= 4; |
| 576 | } |
| 577 | |
| 578 | bool AArch64TargetInfo::isZeroLengthBitfieldPermittedInHomogeneousAggregate() |
| 579 | const { |
| 580 | // AAPCS64 applies homogeneity to the output of the data layout decision, so |
| 581 | // zero-length bitfields do not affect homogeneity. |
| 582 | return true; |
| 583 | } |
| 584 | |
| 585 | bool AArch64TargetInfo::isPermittedToBeHomogeneousAggregate( |
| 586 | const RecordType *RT) const { |
| 587 | if (Opts.IsMicrosoftCXXABI && RT->isCXXRecord()) { |
| 588 | // This won't always return false, but we don't have enough information to |
| 589 | // perform the full check correctly yet. |
| 590 | reportNYI(Feature: "MicrosoftCXXABI homogeneous record classification" ); |
| 591 | return false; |
| 592 | } |
| 593 | |
| 594 | return true; |
| 595 | } |
| 596 | |
| 597 | } // namespace abi |
| 598 | } // namespace llvm |
| 599 | |