| 1 | //===- SPIR.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 | #include "ABIInfoImpl.h" |
| 10 | #include "HLSLBufferLayoutBuilder.h" |
| 11 | #include "TargetInfo.h" |
| 12 | |
| 13 | using namespace clang; |
| 14 | using namespace clang::CodeGen; |
| 15 | |
| 16 | //===----------------------------------------------------------------------===// |
| 17 | // Base ABI and target codegen info implementation common between SPIR and |
| 18 | // SPIR-V. |
| 19 | //===----------------------------------------------------------------------===// |
| 20 | |
| 21 | namespace { |
| 22 | class CommonSPIRABIInfo : public DefaultABIInfo { |
| 23 | public: |
| 24 | CommonSPIRABIInfo(CodeGenTypes &CGT) : DefaultABIInfo(CGT) { setCCs(); } |
| 25 | |
| 26 | private: |
| 27 | void setCCs(); |
| 28 | }; |
| 29 | |
| 30 | class SPIRVABIInfo : public CommonSPIRABIInfo { |
| 31 | public: |
| 32 | SPIRVABIInfo(CodeGenTypes &CGT) : CommonSPIRABIInfo(CGT) {} |
| 33 | void computeInfo(CGFunctionInfo &FI) const override; |
| 34 | |
| 35 | private: |
| 36 | ABIArgInfo classifyReturnType(QualType RetTy) const; |
| 37 | ABIArgInfo classifyKernelArgumentType(QualType Ty) const; |
| 38 | ABIArgInfo classifyArgumentType(QualType Ty) const; |
| 39 | }; |
| 40 | } // end anonymous namespace |
| 41 | namespace { |
| 42 | class CommonSPIRTargetCodeGenInfo : public TargetCodeGenInfo { |
| 43 | public: |
| 44 | CommonSPIRTargetCodeGenInfo(CodeGen::CodeGenTypes &CGT) |
| 45 | : TargetCodeGenInfo(std::make_unique<CommonSPIRABIInfo>(args&: CGT)) {} |
| 46 | CommonSPIRTargetCodeGenInfo(std::unique_ptr<ABIInfo> ABIInfo) |
| 47 | : TargetCodeGenInfo(std::move(ABIInfo)) {} |
| 48 | |
| 49 | LangAS getASTAllocaAddressSpace() const override { |
| 50 | return getLangASFromTargetAS( |
| 51 | TargetAS: getABIInfo().getDataLayout().getAllocaAddrSpace()); |
| 52 | } |
| 53 | |
| 54 | unsigned getDeviceKernelCallingConv() const override; |
| 55 | llvm::Type *getOpenCLType(CodeGenModule &CGM, const Type *T) const override; |
| 56 | llvm::Type * |
| 57 | getHLSLType(CodeGenModule &CGM, const Type *Ty, |
| 58 | const SmallVector<int32_t> *Packoffsets = nullptr) const override; |
| 59 | llvm::Type *getSPIRVImageTypeFromHLSLResource( |
| 60 | const HLSLAttributedResourceType::Attributes &attributes, |
| 61 | QualType SampledType, CodeGenModule &CGM) const; |
| 62 | void |
| 63 | setOCLKernelStubCallingConvention(const FunctionType *&FT) const override; |
| 64 | }; |
| 65 | class SPIRVTargetCodeGenInfo : public CommonSPIRTargetCodeGenInfo { |
| 66 | public: |
| 67 | SPIRVTargetCodeGenInfo(CodeGen::CodeGenTypes &CGT) |
| 68 | : CommonSPIRTargetCodeGenInfo(std::make_unique<SPIRVABIInfo>(args&: CGT)) {} |
| 69 | void setCUDAKernelCallingConvention(const FunctionType *&FT) const override; |
| 70 | LangAS getGlobalVarAddressSpace(CodeGenModule &CGM, |
| 71 | const VarDecl *D) const override; |
| 72 | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
| 73 | CodeGen::CodeGenModule &M) const override; |
| 74 | llvm::SyncScope::ID getLLVMSyncScopeID(const LangOptions &LangOpts, |
| 75 | SyncScope Scope, |
| 76 | llvm::AtomicOrdering Ordering, |
| 77 | llvm::LLVMContext &Ctx) const override; |
| 78 | bool supportsLibCall() const override { |
| 79 | return getABIInfo().getTarget().getTriple().getVendor() != |
| 80 | llvm::Triple::AMD; |
| 81 | } |
| 82 | }; |
| 83 | |
| 84 | inline StringRef mapClangSyncScopeToLLVM(SyncScope Scope) { |
| 85 | switch (Scope) { |
| 86 | case SyncScope::HIPSingleThread: |
| 87 | case SyncScope::SingleScope: |
| 88 | return "singlethread" ; |
| 89 | case SyncScope::HIPWavefront: |
| 90 | case SyncScope::OpenCLSubGroup: |
| 91 | case SyncScope::WavefrontScope: |
| 92 | return "subgroup" ; |
| 93 | case SyncScope::HIPWorkgroup: |
| 94 | case SyncScope::OpenCLWorkGroup: |
| 95 | case SyncScope::WorkgroupScope: |
| 96 | return "workgroup" ; |
| 97 | case SyncScope::HIPAgent: |
| 98 | case SyncScope::OpenCLDevice: |
| 99 | case SyncScope::DeviceScope: |
| 100 | return "device" ; |
| 101 | case SyncScope::SystemScope: |
| 102 | case SyncScope::HIPSystem: |
| 103 | case SyncScope::OpenCLAllSVMDevices: |
| 104 | return "" ; |
| 105 | } |
| 106 | return "" ; |
| 107 | } |
| 108 | } // End anonymous namespace. |
| 109 | |
| 110 | void CommonSPIRABIInfo::setCCs() { |
| 111 | assert(getRuntimeCC() == llvm::CallingConv::C); |
| 112 | RuntimeCC = llvm::CallingConv::SPIR_FUNC; |
| 113 | } |
| 114 | |
| 115 | ABIArgInfo SPIRVABIInfo::classifyReturnType(QualType RetTy) const { |
| 116 | if (getTarget().getTriple().getVendor() != llvm::Triple::AMD) |
| 117 | return DefaultABIInfo::classifyReturnType(RetTy); |
| 118 | if (!isAggregateTypeForABI(T: RetTy) || getRecordArgABI(T: RetTy, CXXABI&: getCXXABI())) |
| 119 | return DefaultABIInfo::classifyReturnType(RetTy); |
| 120 | |
| 121 | if (const RecordType *RT = RetTy->getAs<RecordType>()) { |
| 122 | const RecordDecl *RD = RT->getDecl(); |
| 123 | if (RD->hasFlexibleArrayMember()) |
| 124 | return DefaultABIInfo::classifyReturnType(RetTy); |
| 125 | } |
| 126 | |
| 127 | // TODO: The AMDGPU ABI is non-trivial to represent in SPIR-V; in order to |
| 128 | // avoid encoding various architecture specific bits here we return everything |
| 129 | // as direct to retain type info for things like aggregates, for later perusal |
| 130 | // when translating back to LLVM/lowering in the BE. This is also why we |
| 131 | // disable flattening as the outcomes can mismatch between SPIR-V and AMDGPU. |
| 132 | // This will be revisited / optimised in the future. |
| 133 | return ABIArgInfo::getDirect(T: CGT.ConvertType(T: RetTy), Offset: 0u, Padding: nullptr, CanBeFlattened: false); |
| 134 | } |
| 135 | |
| 136 | ABIArgInfo SPIRVABIInfo::classifyKernelArgumentType(QualType Ty) const { |
| 137 | if (getContext().getLangOpts().CUDAIsDevice) { |
| 138 | // Coerce pointer arguments with default address space to CrossWorkGroup |
| 139 | // pointers for HIPSPV/CUDASPV. When the language mode is HIP/CUDA, the |
| 140 | // SPIRTargetInfo maps cuda_device to SPIR-V's CrossWorkGroup address space. |
| 141 | llvm::Type *LTy = CGT.ConvertType(T: Ty); |
| 142 | auto DefaultAS = getContext().getTargetAddressSpace(AS: LangAS::Default); |
| 143 | auto GlobalAS = getContext().getTargetAddressSpace(AS: LangAS::cuda_device); |
| 144 | auto *PtrTy = llvm::dyn_cast<llvm::PointerType>(Val: LTy); |
| 145 | if (PtrTy && PtrTy->getAddressSpace() == DefaultAS) { |
| 146 | LTy = llvm::PointerType::get(C&: PtrTy->getContext(), AddressSpace: GlobalAS); |
| 147 | return ABIArgInfo::getDirect(T: LTy, Offset: 0, Padding: nullptr, CanBeFlattened: false); |
| 148 | } |
| 149 | |
| 150 | if (isAggregateTypeForABI(T: Ty)) { |
| 151 | if (getTarget().getTriple().getVendor() == llvm::Triple::AMD) |
| 152 | // TODO: The AMDGPU kernel ABI passes aggregates byref, which is not |
| 153 | // currently expressible in SPIR-V; SPIR-V passes aggregates byval, |
| 154 | // which the AMDGPU kernel ABI does not allow. Passing aggregates as |
| 155 | // direct works around this impedance mismatch, as it retains type info |
| 156 | // and can be correctly handled, post reverse-translation, by the AMDGPU |
| 157 | // BE, which has to support this CC for legacy OpenCL purposes. It can |
| 158 | // be brittle and does lead to performance degradation in certain |
| 159 | // pathological cases. This will be revisited / optimised in the future, |
| 160 | // once a way to deal with the byref/byval impedance mismatch is |
| 161 | // identified. |
| 162 | return ABIArgInfo::getDirect(T: LTy, Offset: 0, Padding: nullptr, CanBeFlattened: false); |
| 163 | // Force copying aggregate type in kernel arguments by value when |
| 164 | // compiling CUDA targeting SPIR-V. This is required for the object |
| 165 | // copied to be valid on the device. |
| 166 | // This behavior follows the CUDA spec |
| 167 | // https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#global-function-argument-processing, |
| 168 | // and matches the NVPTX implementation. TODO: hardcoding to 0 should be |
| 169 | // revisited if HIPSPV / byval starts making use of the AS of an indirect |
| 170 | // arg. |
| 171 | return getNaturalAlignIndirect(Ty, /*AddrSpace=*/0, /*byval=*/ByVal: true); |
| 172 | } |
| 173 | } |
| 174 | return classifyArgumentType(Ty); |
| 175 | } |
| 176 | |
| 177 | ABIArgInfo SPIRVABIInfo::classifyArgumentType(QualType Ty) const { |
| 178 | if (getTarget().getTriple().getVendor() != llvm::Triple::AMD) |
| 179 | return DefaultABIInfo::classifyArgumentType(RetTy: Ty); |
| 180 | if (!isAggregateTypeForABI(T: Ty)) |
| 181 | return DefaultABIInfo::classifyArgumentType(RetTy: Ty); |
| 182 | |
| 183 | // Records with non-trivial destructors/copy-constructors should not be |
| 184 | // passed by value. |
| 185 | if (auto RAA = getRecordArgABI(T: Ty, CXXABI&: getCXXABI())) |
| 186 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace(), |
| 187 | ByVal: RAA == CGCXXABI::RAA_DirectInMemory); |
| 188 | |
| 189 | if (const RecordType *RT = Ty->getAs<RecordType>()) { |
| 190 | const RecordDecl *RD = RT->getDecl(); |
| 191 | if (RD->hasFlexibleArrayMember()) |
| 192 | return DefaultABIInfo::classifyArgumentType(RetTy: Ty); |
| 193 | } |
| 194 | |
| 195 | return ABIArgInfo::getDirect(T: CGT.ConvertType(T: Ty), Offset: 0u, Padding: nullptr, CanBeFlattened: false); |
| 196 | } |
| 197 | |
| 198 | void SPIRVABIInfo::computeInfo(CGFunctionInfo &FI) const { |
| 199 | // The logic is same as in DefaultABIInfo with an exception on the kernel |
| 200 | // arguments handling. |
| 201 | llvm::CallingConv::ID CC = FI.getCallingConvention(); |
| 202 | |
| 203 | if (!getCXXABI().classifyReturnType(FI)) |
| 204 | FI.getReturnInfo() = classifyReturnType(RetTy: FI.getReturnType()); |
| 205 | |
| 206 | for (auto &I : FI.arguments()) { |
| 207 | if (CC == llvm::CallingConv::SPIR_KERNEL) { |
| 208 | I.info = classifyKernelArgumentType(Ty: I.type); |
| 209 | } else { |
| 210 | I.info = classifyArgumentType(Ty: I.type); |
| 211 | } |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | namespace clang { |
| 216 | namespace CodeGen { |
| 217 | void computeSPIRKernelABIInfo(CodeGenModule &CGM, CGFunctionInfo &FI) { |
| 218 | if (CGM.getTarget().getTriple().isSPIRV()) |
| 219 | SPIRVABIInfo(CGM.getTypes()).computeInfo(FI); |
| 220 | else |
| 221 | CommonSPIRABIInfo(CGM.getTypes()).computeInfo(FI); |
| 222 | } |
| 223 | } |
| 224 | } |
| 225 | |
| 226 | unsigned CommonSPIRTargetCodeGenInfo::getDeviceKernelCallingConv() const { |
| 227 | return llvm::CallingConv::SPIR_KERNEL; |
| 228 | } |
| 229 | |
| 230 | void SPIRVTargetCodeGenInfo::setCUDAKernelCallingConvention( |
| 231 | const FunctionType *&FT) const { |
| 232 | // Convert HIP kernels to SPIR-V kernels. |
| 233 | if (getABIInfo().getContext().getLangOpts().HIP) { |
| 234 | FT = getABIInfo().getContext().adjustFunctionType( |
| 235 | Fn: FT, EInfo: FT->getExtInfo().withCallingConv(cc: CC_DeviceKernel)); |
| 236 | return; |
| 237 | } |
| 238 | } |
| 239 | |
| 240 | void CommonSPIRTargetCodeGenInfo::setOCLKernelStubCallingConvention( |
| 241 | const FunctionType *&FT) const { |
| 242 | FT = getABIInfo().getContext().adjustFunctionType( |
| 243 | Fn: FT, EInfo: FT->getExtInfo().withCallingConv(cc: CC_SpirFunction)); |
| 244 | } |
| 245 | |
| 246 | LangAS |
| 247 | SPIRVTargetCodeGenInfo::getGlobalVarAddressSpace(CodeGenModule &CGM, |
| 248 | const VarDecl *D) const { |
| 249 | assert(!CGM.getLangOpts().OpenCL && |
| 250 | !(CGM.getLangOpts().CUDA && CGM.getLangOpts().CUDAIsDevice) && |
| 251 | "Address space agnostic languages only" ); |
| 252 | // If we're here it means that we're using the SPIRDefIsGen ASMap, hence for |
| 253 | // the global AS we can rely on either cuda_device or sycl_global to be |
| 254 | // correct; however, since this is not a CUDA Device context, we use |
| 255 | // sycl_global to prevent confusion with the assertion. |
| 256 | LangAS DefaultGlobalAS = getLangASFromTargetAS( |
| 257 | TargetAS: CGM.getContext().getTargetAddressSpace(AS: LangAS::sycl_global)); |
| 258 | if (!D) |
| 259 | return DefaultGlobalAS; |
| 260 | |
| 261 | LangAS AddrSpace = D->getType().getAddressSpace(); |
| 262 | if (AddrSpace != LangAS::Default) |
| 263 | return AddrSpace; |
| 264 | |
| 265 | return DefaultGlobalAS; |
| 266 | } |
| 267 | |
| 268 | void SPIRVTargetCodeGenInfo::setTargetAttributes( |
| 269 | const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &M) const { |
| 270 | if (!M.getLangOpts().HIP || |
| 271 | M.getTarget().getTriple().getVendor() != llvm::Triple::AMD) |
| 272 | return; |
| 273 | if (GV->isDeclaration()) |
| 274 | return; |
| 275 | |
| 276 | auto F = dyn_cast<llvm::Function>(Val: GV); |
| 277 | if (!F) |
| 278 | return; |
| 279 | |
| 280 | auto FD = dyn_cast_or_null<FunctionDecl>(Val: D); |
| 281 | if (!FD) |
| 282 | return; |
| 283 | if (!FD->hasAttr<CUDAGlobalAttr>()) |
| 284 | return; |
| 285 | |
| 286 | unsigned N = M.getLangOpts().GPUMaxThreadsPerBlock; |
| 287 | if (auto FlatWGS = FD->getAttr<AMDGPUFlatWorkGroupSizeAttr>()) |
| 288 | N = FlatWGS->getMax()->EvaluateKnownConstInt(Ctx: M.getContext()).getExtValue(); |
| 289 | |
| 290 | // We encode the maximum flat WG size in the first component of the 3D |
| 291 | // max_work_group_size attribute, which will get reverse translated into the |
| 292 | // original AMDGPU attribute when targeting AMDGPU. |
| 293 | auto Int32Ty = llvm::IntegerType::getInt32Ty(C&: M.getLLVMContext()); |
| 294 | llvm::Metadata *AttrMDArgs[] = { |
| 295 | llvm::ConstantAsMetadata::get(C: llvm::ConstantInt::get(Ty: Int32Ty, V: N)), |
| 296 | llvm::ConstantAsMetadata::get(C: llvm::ConstantInt::get(Ty: Int32Ty, V: 1)), |
| 297 | llvm::ConstantAsMetadata::get(C: llvm::ConstantInt::get(Ty: Int32Ty, V: 1))}; |
| 298 | |
| 299 | F->setMetadata(Kind: "max_work_group_size" , |
| 300 | Node: llvm::MDNode::get(Context&: M.getLLVMContext(), MDs: AttrMDArgs)); |
| 301 | } |
| 302 | |
| 303 | llvm::SyncScope::ID |
| 304 | SPIRVTargetCodeGenInfo::getLLVMSyncScopeID(const LangOptions &, SyncScope Scope, |
| 305 | llvm::AtomicOrdering, |
| 306 | llvm::LLVMContext &Ctx) const { |
| 307 | return Ctx.getOrInsertSyncScopeID(SSN: mapClangSyncScopeToLLVM(Scope)); |
| 308 | } |
| 309 | |
| 310 | /// Construct a SPIR-V target extension type for the given OpenCL image type. |
| 311 | static llvm::Type *getSPIRVImageType(llvm::LLVMContext &Ctx, StringRef BaseType, |
| 312 | StringRef OpenCLName, |
| 313 | unsigned AccessQualifier) { |
| 314 | // These parameters compare to the operands of OpTypeImage (see |
| 315 | // https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpTypeImage |
| 316 | // for more details). The first 6 integer parameters all default to 0, and |
| 317 | // will be changed to 1 only for the image type(s) that set the parameter to |
| 318 | // one. The 7th integer parameter is the access qualifier, which is tacked on |
| 319 | // at the end. |
| 320 | SmallVector<unsigned, 7> IntParams = {0, 0, 0, 0, 0, 0}; |
| 321 | |
| 322 | // Choose the dimension of the image--this corresponds to the Dim enum in |
| 323 | // SPIR-V (first integer parameter of OpTypeImage). |
| 324 | if (OpenCLName.starts_with(Prefix: "image2d" )) |
| 325 | IntParams[0] = 1; // 1D |
| 326 | else if (OpenCLName.starts_with(Prefix: "image3d" )) |
| 327 | IntParams[0] = 2; // 2D |
| 328 | else if (OpenCLName == "image1d_buffer" ) |
| 329 | IntParams[0] = 5; // Buffer |
| 330 | else |
| 331 | assert(OpenCLName.starts_with("image1d" ) && "Unknown image type" ); |
| 332 | |
| 333 | // Set the other integer parameters of OpTypeImage if necessary. Note that the |
| 334 | // OpenCL image types don't provide any information for the Sampled or |
| 335 | // Image Format parameters. |
| 336 | if (OpenCLName.contains(Other: "_depth" )) |
| 337 | IntParams[1] = 1; |
| 338 | if (OpenCLName.contains(Other: "_array" )) |
| 339 | IntParams[2] = 1; |
| 340 | if (OpenCLName.contains(Other: "_msaa" )) |
| 341 | IntParams[3] = 1; |
| 342 | |
| 343 | // Access qualifier |
| 344 | IntParams.push_back(Elt: AccessQualifier); |
| 345 | |
| 346 | return llvm::TargetExtType::get(Context&: Ctx, Name: BaseType, Types: {llvm::Type::getVoidTy(C&: Ctx)}, |
| 347 | Ints: IntParams); |
| 348 | } |
| 349 | |
| 350 | llvm::Type *CommonSPIRTargetCodeGenInfo::getOpenCLType(CodeGenModule &CGM, |
| 351 | const Type *Ty) const { |
| 352 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); |
| 353 | if (auto *PipeTy = dyn_cast<PipeType>(Val: Ty)) |
| 354 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.Pipe" , Types: {}, |
| 355 | Ints: {!PipeTy->isReadOnly()}); |
| 356 | if (auto *BuiltinTy = dyn_cast<BuiltinType>(Val: Ty)) { |
| 357 | enum AccessQualifier : unsigned { AQ_ro = 0, AQ_wo = 1, AQ_rw = 2 }; |
| 358 | switch (BuiltinTy->getKind()) { |
| 359 | #define IMAGE_TYPE(ImgType, Id, SingletonId, Access, Suffix) \ |
| 360 | case BuiltinType::Id: \ |
| 361 | return getSPIRVImageType(Ctx, "spirv.Image", #ImgType, AQ_##Suffix); |
| 362 | #include "clang/Basic/OpenCLImageTypes.def" |
| 363 | case BuiltinType::OCLSampler: |
| 364 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.Sampler" ); |
| 365 | case BuiltinType::OCLEvent: |
| 366 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.Event" ); |
| 367 | case BuiltinType::OCLClkEvent: |
| 368 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.DeviceEvent" ); |
| 369 | case BuiltinType::OCLQueue: |
| 370 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.Queue" ); |
| 371 | case BuiltinType::OCLReserveID: |
| 372 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.ReserveId" ); |
| 373 | #define INTEL_SUBGROUP_AVC_TYPE(Name, Id) \ |
| 374 | case BuiltinType::OCLIntelSubgroupAVC##Id: \ |
| 375 | return llvm::TargetExtType::get(Ctx, "spirv.Avc" #Id "INTEL"); |
| 376 | #include "clang/Basic/OpenCLExtensionTypes.def" |
| 377 | default: |
| 378 | return nullptr; |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | return nullptr; |
| 383 | } |
| 384 | |
| 385 | // Gets a spirv.IntegralConstant or spirv.Literal. If IntegralType is present, |
| 386 | // returns an IntegralConstant, otherwise returns a Literal. |
| 387 | static llvm::Type *getInlineSpirvConstant(CodeGenModule &CGM, |
| 388 | llvm::Type *IntegralType, |
| 389 | llvm::APInt Value) { |
| 390 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); |
| 391 | |
| 392 | // Convert the APInt value to an array of uint32_t words |
| 393 | llvm::SmallVector<uint32_t> Words; |
| 394 | |
| 395 | while (Value.ugt(RHS: 0)) { |
| 396 | uint32_t Word = Value.trunc(width: 32).getZExtValue(); |
| 397 | Value.lshrInPlace(ShiftAmt: 32); |
| 398 | |
| 399 | Words.push_back(Elt: Word); |
| 400 | } |
| 401 | if (Words.size() == 0) |
| 402 | Words.push_back(Elt: 0); |
| 403 | |
| 404 | if (IntegralType) |
| 405 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.IntegralConstant" , |
| 406 | Types: {IntegralType}, Ints: Words); |
| 407 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.Literal" , Types: {}, Ints: Words); |
| 408 | } |
| 409 | |
| 410 | static llvm::Type *getInlineSpirvType(CodeGenModule &CGM, |
| 411 | const HLSLInlineSpirvType *SpirvType) { |
| 412 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); |
| 413 | |
| 414 | llvm::SmallVector<llvm::Type *> Operands; |
| 415 | |
| 416 | for (auto &Operand : SpirvType->getOperands()) { |
| 417 | using SpirvOperandKind = SpirvOperand::SpirvOperandKind; |
| 418 | |
| 419 | llvm::Type *Result = nullptr; |
| 420 | switch (Operand.getKind()) { |
| 421 | case SpirvOperandKind::ConstantId: { |
| 422 | llvm::Type *IntegralType = |
| 423 | CGM.getTypes().ConvertType(T: Operand.getResultType()); |
| 424 | llvm::APInt Value = Operand.getValue(); |
| 425 | |
| 426 | Result = getInlineSpirvConstant(CGM, IntegralType, Value); |
| 427 | break; |
| 428 | } |
| 429 | case SpirvOperandKind::Literal: { |
| 430 | llvm::APInt Value = Operand.getValue(); |
| 431 | Result = getInlineSpirvConstant(CGM, IntegralType: nullptr, Value); |
| 432 | break; |
| 433 | } |
| 434 | case SpirvOperandKind::TypeId: { |
| 435 | QualType TypeOperand = Operand.getResultType(); |
| 436 | if (auto *RT = TypeOperand->getAs<RecordType>()) { |
| 437 | auto *RD = RT->getDecl(); |
| 438 | assert(RD->isCompleteDefinition() && |
| 439 | "Type completion should have been required in Sema" ); |
| 440 | |
| 441 | const FieldDecl *HandleField = RD->findFirstNamedDataMember(); |
| 442 | if (HandleField) { |
| 443 | QualType ResourceType = HandleField->getType(); |
| 444 | if (ResourceType->getAs<HLSLAttributedResourceType>()) { |
| 445 | TypeOperand = ResourceType; |
| 446 | } |
| 447 | } |
| 448 | } |
| 449 | Result = CGM.getTypes().ConvertType(T: TypeOperand); |
| 450 | break; |
| 451 | } |
| 452 | default: |
| 453 | llvm_unreachable("HLSLInlineSpirvType had invalid operand!" ); |
| 454 | break; |
| 455 | } |
| 456 | |
| 457 | assert(Result); |
| 458 | Operands.push_back(Elt: Result); |
| 459 | } |
| 460 | |
| 461 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.Type" , Types: Operands, |
| 462 | Ints: {SpirvType->getOpcode(), SpirvType->getSize(), |
| 463 | SpirvType->getAlignment()}); |
| 464 | } |
| 465 | |
| 466 | llvm::Type *CommonSPIRTargetCodeGenInfo::getHLSLType( |
| 467 | CodeGenModule &CGM, const Type *Ty, |
| 468 | const SmallVector<int32_t> *Packoffsets) const { |
| 469 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); |
| 470 | |
| 471 | if (auto *SpirvType = dyn_cast<HLSLInlineSpirvType>(Val: Ty)) |
| 472 | return getInlineSpirvType(CGM, SpirvType); |
| 473 | |
| 474 | auto *ResType = dyn_cast<HLSLAttributedResourceType>(Val: Ty); |
| 475 | if (!ResType) |
| 476 | return nullptr; |
| 477 | |
| 478 | const HLSLAttributedResourceType::Attributes &ResAttrs = ResType->getAttrs(); |
| 479 | switch (ResAttrs.ResourceClass) { |
| 480 | case llvm::dxil::ResourceClass::UAV: |
| 481 | case llvm::dxil::ResourceClass::SRV: { |
| 482 | // TypedBuffer and RawBuffer both need element type |
| 483 | QualType ContainedTy = ResType->getContainedType(); |
| 484 | if (ContainedTy.isNull()) |
| 485 | return nullptr; |
| 486 | |
| 487 | assert(!ResAttrs.IsROV && |
| 488 | "Rasterizer order views not implemented for SPIR-V yet" ); |
| 489 | |
| 490 | if (!ResAttrs.RawBuffer) { |
| 491 | // convert element type |
| 492 | return getSPIRVImageTypeFromHLSLResource(attributes: ResAttrs, SampledType: ContainedTy, CGM); |
| 493 | } |
| 494 | |
| 495 | llvm::Type *ElemType = CGM.getTypes().ConvertTypeForMem(T: ContainedTy); |
| 496 | llvm::ArrayType *RuntimeArrayType = llvm::ArrayType::get(ElementType: ElemType, NumElements: 0); |
| 497 | uint32_t StorageClass = /* StorageBuffer storage class */ 12; |
| 498 | bool IsWritable = ResAttrs.ResourceClass == llvm::dxil::ResourceClass::UAV; |
| 499 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.VulkanBuffer" , |
| 500 | Types: {RuntimeArrayType}, |
| 501 | Ints: {StorageClass, IsWritable}); |
| 502 | } |
| 503 | case llvm::dxil::ResourceClass::CBuffer: { |
| 504 | QualType ContainedTy = ResType->getContainedType(); |
| 505 | if (ContainedTy.isNull() || !ContainedTy->isStructureType()) |
| 506 | return nullptr; |
| 507 | |
| 508 | llvm::Type *BufferLayoutTy = |
| 509 | HLSLBufferLayoutBuilder(CGM, "spirv.Layout" ) |
| 510 | .createLayoutType(StructType: ContainedTy->getAsStructureType(), Packoffsets); |
| 511 | uint32_t StorageClass = /* Uniform storage class */ 2; |
| 512 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.VulkanBuffer" , Types: {BufferLayoutTy}, |
| 513 | Ints: {StorageClass, false}); |
| 514 | break; |
| 515 | } |
| 516 | case llvm::dxil::ResourceClass::Sampler: |
| 517 | return llvm::TargetExtType::get(Context&: Ctx, Name: "spirv.Sampler" ); |
| 518 | } |
| 519 | return nullptr; |
| 520 | } |
| 521 | |
| 522 | llvm::Type *CommonSPIRTargetCodeGenInfo::getSPIRVImageTypeFromHLSLResource( |
| 523 | const HLSLAttributedResourceType::Attributes &attributes, QualType Ty, |
| 524 | CodeGenModule &CGM) const { |
| 525 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); |
| 526 | |
| 527 | Ty = Ty->getCanonicalTypeUnqualified(); |
| 528 | if (const VectorType *V = dyn_cast<VectorType>(Val&: Ty)) |
| 529 | Ty = V->getElementType(); |
| 530 | assert(!Ty->isVectorType() && "We still have a vector type." ); |
| 531 | |
| 532 | llvm::Type *SampledType = CGM.getTypes().ConvertTypeForMem(T: Ty); |
| 533 | |
| 534 | assert((SampledType->isIntegerTy() || SampledType->isFloatingPointTy()) && |
| 535 | "The element type for a SPIR-V resource must be a scalar integer or " |
| 536 | "floating point type." ); |
| 537 | |
| 538 | // These parameters correspond to the operands to the OpTypeImage SPIR-V |
| 539 | // instruction. See |
| 540 | // https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpTypeImage. |
| 541 | SmallVector<unsigned, 6> IntParams(6, 0); |
| 542 | |
| 543 | const char *Name = |
| 544 | Ty->isSignedIntegerType() ? "spirv.SignedImage" : "spirv.Image" ; |
| 545 | |
| 546 | // Dim |
| 547 | // For now we assume everything is a buffer. |
| 548 | IntParams[0] = 5; |
| 549 | |
| 550 | // Depth |
| 551 | // HLSL does not indicate if it is a depth texture or not, so we use unknown. |
| 552 | IntParams[1] = 2; |
| 553 | |
| 554 | // Arrayed |
| 555 | IntParams[2] = 0; |
| 556 | |
| 557 | // MS |
| 558 | IntParams[3] = 0; |
| 559 | |
| 560 | // Sampled |
| 561 | IntParams[4] = |
| 562 | attributes.ResourceClass == llvm::dxil::ResourceClass::UAV ? 2 : 1; |
| 563 | |
| 564 | // Image format. |
| 565 | // Setting to unknown for now. |
| 566 | IntParams[5] = 0; |
| 567 | |
| 568 | llvm::TargetExtType *ImageType = |
| 569 | llvm::TargetExtType::get(Context&: Ctx, Name, Types: {SampledType}, Ints: IntParams); |
| 570 | return ImageType; |
| 571 | } |
| 572 | |
| 573 | std::unique_ptr<TargetCodeGenInfo> |
| 574 | CodeGen::createCommonSPIRTargetCodeGenInfo(CodeGenModule &CGM) { |
| 575 | return std::make_unique<CommonSPIRTargetCodeGenInfo>(args&: CGM.getTypes()); |
| 576 | } |
| 577 | |
| 578 | std::unique_ptr<TargetCodeGenInfo> |
| 579 | CodeGen::createSPIRVTargetCodeGenInfo(CodeGenModule &CGM) { |
| 580 | return std::make_unique<SPIRVTargetCodeGenInfo>(args&: CGM.getTypes()); |
| 581 | } |
| 582 | |