| 1 | //===- CIRGenModule.cpp - Per-Module state for CIR generation -------------===// |
| 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 is the internal per-translation-unit state used for CIR translation. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "CIRGenModule.h" |
| 14 | #include "CIRGenCUDARuntime.h" |
| 15 | #include "CIRGenCXXABI.h" |
| 16 | #include "CIRGenConstantEmitter.h" |
| 17 | #include "CIRGenFunction.h" |
| 18 | |
| 19 | #include "mlir/Dialect/LLVMIR/LLVMDialectDecl.h" |
| 20 | #include "mlir/Dialect/OpenMP/Utils/Utils.h" |
| 21 | #include "mlir/IR/SymbolTable.h" |
| 22 | #include "clang/AST/ASTContext.h" |
| 23 | #include "clang/AST/ASTLambda.h" |
| 24 | #include "clang/AST/Attrs.inc" |
| 25 | #include "clang/AST/DeclBase.h" |
| 26 | #include "clang/AST/DeclOpenACC.h" |
| 27 | #include "clang/AST/GlobalDecl.h" |
| 28 | #include "clang/AST/Mangle.h" |
| 29 | #include "clang/AST/RecordLayout.h" |
| 30 | #include "clang/AST/StmtOpenMP.h" |
| 31 | #include "clang/Basic/DiagnosticFrontend.h" |
| 32 | #include "clang/Basic/Module.h" |
| 33 | #include "clang/Basic/SourceManager.h" |
| 34 | #include "clang/CIR/Dialect/IR/CIRAttrs.h" |
| 35 | #include "clang/CIR/Dialect/IR/CIRDialect.h" |
| 36 | #include "clang/CIR/Dialect/IR/CIROpsEnums.h" |
| 37 | #include "clang/CIR/Dialect/IR/CIRTypes.h" |
| 38 | #include "clang/CIR/Interfaces/CIROpInterfaces.h" |
| 39 | #include "clang/CIR/MissingFeatures.h" |
| 40 | #include "clang/CodeGenUtils/ModuleUtils.h" |
| 41 | #include "clang/CodeGenUtils/RecordLayoutUtils.h" |
| 42 | #include "llvm/ADT/STLExtras.h" |
| 43 | #include "llvm/ADT/StringExtras.h" |
| 44 | #include "llvm/ADT/StringRef.h" |
| 45 | #include "llvm/ADT/StringSwitch.h" |
| 46 | #include "llvm/Support/raw_ostream.h" |
| 47 | |
| 48 | #include "CIRGenFunctionInfo.h" |
| 49 | #include "TargetInfo.h" |
| 50 | #include "mlir/Dialect/Ptr/IR/MemorySpaceInterfaces.h" |
| 51 | #include "mlir/IR/Attributes.h" |
| 52 | #include "mlir/IR/BuiltinOps.h" |
| 53 | #include "mlir/IR/Location.h" |
| 54 | #include "mlir/IR/MLIRContext.h" |
| 55 | #include "mlir/IR/Operation.h" |
| 56 | #include "mlir/IR/Verifier.h" |
| 57 | |
| 58 | #include <algorithm> |
| 59 | |
| 60 | using namespace clang; |
| 61 | using namespace clang::CIRGen; |
| 62 | |
| 63 | static CIRGenCXXABI *createCXXABI(CIRGenModule &cgm) { |
| 64 | switch (cgm.getASTContext().getCXXABIKind()) { |
| 65 | case TargetCXXABI::GenericItanium: |
| 66 | case TargetCXXABI::GenericAArch64: |
| 67 | case TargetCXXABI::AppleARM64: |
| 68 | case TargetCXXABI::GenericARM: |
| 69 | return CreateCIRGenItaniumCXXABI(cgm); |
| 70 | case TargetCXXABI::Microsoft: |
| 71 | return CreateCIRGenMicrosoftCXXABI(cgm); |
| 72 | |
| 73 | case TargetCXXABI::Fuchsia: |
| 74 | case TargetCXXABI::iOS: |
| 75 | case TargetCXXABI::WatchOS: |
| 76 | case TargetCXXABI::GenericMIPS: |
| 77 | case TargetCXXABI::WebAssembly: |
| 78 | case TargetCXXABI::XL: |
| 79 | cgm.errorNYI(feature: "createCXXABI: C++ ABI kind" ); |
| 80 | return nullptr; |
| 81 | } |
| 82 | |
| 83 | llvm_unreachable("invalid C++ ABI kind" ); |
| 84 | } |
| 85 | |
| 86 | CIRGenModule::CIRGenModule(mlir::MLIRContext &mlirContext, |
| 87 | clang::ASTContext &astContext, |
| 88 | const clang::CodeGenOptions &cgo, |
| 89 | DiagnosticsEngine &diags) |
| 90 | : builder(mlirContext, *this), astContext(astContext), |
| 91 | langOpts(astContext.getLangOpts()), atomicOpts(astContext.getLangOpts()), |
| 92 | codeGenOpts(cgo), |
| 93 | theModule{mlir::ModuleOp::create(loc: mlir::UnknownLoc::get(context: &mlirContext))}, |
| 94 | diags(diags), target(astContext.getTargetInfo()), |
| 95 | abi(createCXXABI(cgm&: *this)), genTypes(*this), vtables(*this) { |
| 96 | |
| 97 | // Classic CodeGen's CXX20ModuleInits: C++20 named modules get an |
| 98 | // initializer function of their own only under the Itanium mangler, since |
| 99 | // no Microsoft mangling for it has been settled on yet; otherwise they fall |
| 100 | // back to `_GLOBAL__sub_I_` as in classic CodeGen. |
| 101 | cxx20ModuleInits = |
| 102 | langOpts.CPlusPlusModules && |
| 103 | getCXXABI().getMangleContext().getKind() == MangleContext::MK_Itanium; |
| 104 | |
| 105 | // Initialize cached types |
| 106 | voidTy = cir::VoidType::get(ctx: &getMLIRContext()); |
| 107 | voidPtrTy = cir::PointerType::get(pointee: voidTy); |
| 108 | sInt8Ty = cir::IntType::get(context: &getMLIRContext(), width: 8, /*isSigned=*/is_signed: true); |
| 109 | sInt16Ty = cir::IntType::get(context: &getMLIRContext(), width: 16, /*isSigned=*/is_signed: true); |
| 110 | sInt32Ty = cir::IntType::get(context: &getMLIRContext(), width: 32, /*isSigned=*/is_signed: true); |
| 111 | sInt64Ty = cir::IntType::get(context: &getMLIRContext(), width: 64, /*isSigned=*/is_signed: true); |
| 112 | sInt128Ty = cir::IntType::get(context: &getMLIRContext(), width: 128, /*isSigned=*/is_signed: true); |
| 113 | uInt8Ty = cir::IntType::get(context: &getMLIRContext(), width: 8, /*isSigned=*/is_signed: false); |
| 114 | uInt8PtrTy = cir::PointerType::get(pointee: uInt8Ty); |
| 115 | cirAllocaAddressSpace = getTargetCIRGenInfo().getCIRAllocaAddressSpace(); |
| 116 | uInt16Ty = cir::IntType::get(context: &getMLIRContext(), width: 16, /*isSigned=*/is_signed: false); |
| 117 | uInt32Ty = cir::IntType::get(context: &getMLIRContext(), width: 32, /*isSigned=*/is_signed: false); |
| 118 | uInt64Ty = cir::IntType::get(context: &getMLIRContext(), width: 64, /*isSigned=*/is_signed: false); |
| 119 | uInt128Ty = cir::IntType::get(context: &getMLIRContext(), width: 128, /*isSigned=*/is_signed: false); |
| 120 | fP16Ty = cir::FP16Type::get(ctx: &getMLIRContext()); |
| 121 | bFloat16Ty = cir::BF16Type::get(ctx: &getMLIRContext()); |
| 122 | floatTy = cir::SingleType::get(ctx: &getMLIRContext()); |
| 123 | doubleTy = cir::DoubleType::get(ctx: &getMLIRContext()); |
| 124 | fP80Ty = cir::FP80Type::get(ctx: &getMLIRContext()); |
| 125 | fP128Ty = cir::FP128Type::get(ctx: &getMLIRContext()); |
| 126 | |
| 127 | allocaInt8PtrTy = cir::PointerType::get(pointee: uInt8Ty, addrSpace: cirAllocaAddressSpace); |
| 128 | |
| 129 | PointerAlignInBytes = |
| 130 | astContext |
| 131 | .toCharUnitsFromBits( |
| 132 | BitSize: astContext.getTargetInfo().getPointerAlign(AddrSpace: LangAS::Default)) |
| 133 | .getQuantity(); |
| 134 | |
| 135 | const unsigned charSize = target.getCharWidth(); |
| 136 | uCharTy = cir::IntType::get(context: &getMLIRContext(), width: charSize, /*isSigned=*/is_signed: false); |
| 137 | |
| 138 | const unsigned sizeTypeSize = target.getTypeWidth(T: target.getSizeType()); |
| 139 | SizeSizeInBytes = sizeTypeSize / charSize; |
| 140 | // In CIRGenTypeCache, UIntPtrTy and SizeType are fields of the same union |
| 141 | uIntPtrTy = |
| 142 | cir::IntType::get(context: &getMLIRContext(), width: sizeTypeSize, /*isSigned=*/is_signed: false); |
| 143 | ptrDiffTy = |
| 144 | cir::IntType::get(context: &getMLIRContext(), width: sizeTypeSize, /*isSigned=*/is_signed: true); |
| 145 | |
| 146 | std::optional<cir::SourceLanguage> sourceLanguage = getCIRSourceLanguage(); |
| 147 | if (sourceLanguage) |
| 148 | getModule()->setAttr( |
| 149 | name: cir::CIRDialect::getSourceLanguageAttrName(), |
| 150 | value: cir::SourceLanguageAttr::get(context: &mlirContext, value: *sourceLanguage)); |
| 151 | if (langOpts.OpenCL || (langOpts.CUDAIsDevice && getTriple().isSPIRV())) { |
| 152 | // CUDA and HIP use OpenCL 2.0 metadata when targeting SPIR-V. |
| 153 | unsigned version = |
| 154 | langOpts.OpenCL ? langOpts.getOpenCLCompatibleVersion() : 200; |
| 155 | setOpenCLVersionAttr(attrName: cir::CIRDialect::getOpenCLVersionAttrName(), version); |
| 156 | if (langOpts.OpenCLCPlusPlus) |
| 157 | setOpenCLVersionAttr(attrName: cir::CIRDialect::getOpenCLCXXVersionAttrName(), |
| 158 | version: langOpts.OpenCLCPlusPlusVersion); |
| 159 | // SPIR v2.0 s2.12 requires opencl.spir.version. |
| 160 | if (getTriple().isSPIR()) { |
| 161 | unsigned spirMajor = version / 100; |
| 162 | getModule()->setAttr(name: cir::CIRDialect::getOpenCLSPIRVersionAttrName(), |
| 163 | value: cir::OpenCLVersionAttr::get(context: &getMLIRContext(), |
| 164 | majorVersion: spirMajor, |
| 165 | minorVersion: spirMajor > 1 ? 0 : 2)); |
| 166 | } |
| 167 | } |
| 168 | getModule()->setAttr(name: cir::CIRDialect::getTripleAttrName(), |
| 169 | value: builder.getStringAttr(bytes: getTriple().str())); |
| 170 | getModule()->setAttr(name: cir::CIRDialect::getTargetABIAttrName(), |
| 171 | value: builder.getStringAttr(bytes: getTarget().getABI())); |
| 172 | |
| 173 | // Assign the target-cpu and target-features as attributes to the module. This |
| 174 | // will allow subsequent passes to use this to ensure that 'generated' |
| 175 | // functions get the correct behavior for the model. |
| 176 | llvm::StringMap<std::string> defaultCPUAndFeatures; |
| 177 | if (getCPUAndFeaturesAttributes(gd: GlobalDecl(), attrs&: defaultCPUAndFeatures)) |
| 178 | for (const auto &[key, val] : defaultCPUAndFeatures) |
| 179 | getModule()->setAttr(name: key, value: builder.getStringAttr(bytes: val)); |
| 180 | |
| 181 | // If cxxabi was overridden via -fc++-abi, record it. |
| 182 | if (langOpts.CXXABI) |
| 183 | getModule()->setAttr( |
| 184 | name: cir::CIRDialect::getCXXABIAttrName(), |
| 185 | value: builder.getStringAttr(bytes: TargetCXXABI::getSpelling(ABIKind: *langOpts.CXXABI))); |
| 186 | |
| 187 | if (llvm::VersionTuple sdkVersion = getTarget().getSDKVersion(); |
| 188 | !sdkVersion.empty()) |
| 189 | getModule()->setAttr(name: cir::CIRDialect::getSDKVersionAttrName(), |
| 190 | value: builder.getStringAttr(bytes: sdkVersion.getAsString())); |
| 191 | if (langOpts.CUDA && !langOpts.CUID.empty()) |
| 192 | getModule()->setAttr(name: cir::CIRDialect::getCUIDHashAttrName(), |
| 193 | value: builder.getStringAttr(bytes: astContext.getCUIDHash())); |
| 194 | // TODO(CIR): These attributes should eventually be replaced by |
| 195 | // TypeSizeInfoAttr once it is upstreamed. |
| 196 | getModule()->setAttr(name: cir::CIRDialect::getSizeTypeWidthAttrName(), |
| 197 | value: builder.getI32IntegerAttr(value: sizeTypeSize)); |
| 198 | getModule()->setAttr(name: cir::CIRDialect::getIntTypeWidthAttrName(), |
| 199 | value: builder.getI32IntegerAttr(value: target.getIntWidth())); |
| 200 | |
| 201 | // Serialize the lowering-relevant LangOptions and CodeGenOptions onto the |
| 202 | // ModuleOp so a reloaded .cir is self-describing and lowers the same way it |
| 203 | // was compiled, without a live clang::LangOptions or clang::CodeGenOptions. |
| 204 | getModule()->setAttr( |
| 205 | name: cir::CIRDialect::getLoweringOptionsAttrName(), |
| 206 | value: cir::LoweringOptionsAttr::get( |
| 207 | context: &mlirContext, |
| 208 | /*exceptions=*/langOpts.Exceptions, |
| 209 | /*threadsafe_statics=*/langOpts.ThreadsafeStatics, |
| 210 | /*cuda=*/langOpts.CUDA, |
| 211 | /*cuda_is_device=*/langOpts.CUDAIsDevice, |
| 212 | /*hip=*/langOpts.HIP, |
| 213 | /*hlsl=*/langOpts.HLSL, |
| 214 | /*sycl_is_device=*/langOpts.SYCLIsDevice, |
| 215 | /*gpu_rdc=*/langOpts.GPURelocatableDeviceCode, |
| 216 | /*openmp=*/langOpts.OpenMP != 0, |
| 217 | /*openmp_is_target_device=*/langOpts.OpenMPIsTargetDevice, |
| 218 | /*clang_abi_compat=*/ |
| 219 | static_cast<int32_t>(langOpts.getClangABICompat()), |
| 220 | long_double_size: static_cast<int32_t>(langOpts.LongDoubleSize), |
| 221 | /*enable_noundef_attrs=*/cgo.EnableNoundefAttrs, |
| 222 | /*no_trapping_math=*/langOpts.getDefaultExceptionMode() == |
| 223 | LangOptions::FPE_Ignore, |
| 224 | /*ssp_buffer_size=*/cgo.SSPBufferSize, |
| 225 | /*relative_cxx_abi_vtables=*/langOpts.RelativeCXXABIVTables)); |
| 226 | |
| 227 | if (cgo.OptimizationLevel > 0 || cgo.OptimizeSize > 0) |
| 228 | getModule()->setAttr(name: cir::CIRDialect::getOptInfoAttrName(), |
| 229 | value: cir::OptInfoAttr::get(context: &mlirContext, |
| 230 | level: cgo.OptimizationLevel, |
| 231 | size: cgo.OptimizeSize)); |
| 232 | |
| 233 | getModule()->setAttr( |
| 234 | name: cir::CIRDialect::getDefaultTlsModelAttrName(), |
| 235 | value: cir::TLSModelAttr::get(context: &mlirContext, value: getDefaultCIRTLSModel())); |
| 236 | |
| 237 | if (langOpts.OpenMP) { |
| 238 | mlir::omp::OffloadModuleOpts ompOpts( |
| 239 | langOpts.OpenMPTargetDebug, langOpts.OpenMPTeamSubscription, |
| 240 | langOpts.OpenMPThreadSubscription, langOpts.OpenMPNoThreadState, |
| 241 | langOpts.OpenMPNoNestedParallelism, langOpts.OpenMPIsTargetDevice, |
| 242 | getTriple().isGPU(), langOpts.OpenMPForceUSM, langOpts.OpenMP, |
| 243 | langOpts.OMPHostIRFile, langOpts.OMPTargetTriples, langOpts.NoGPULib); |
| 244 | mlir::omp::setOffloadModuleInterfaceAttributes(module: getModule(), opts: ompOpts); |
| 245 | mlir::omp::setOpenMPVersionAttribute(module: getModule(), version: langOpts.OpenMP); |
| 246 | } |
| 247 | |
| 248 | if (langOpts.CUDA) |
| 249 | createCUDARuntime(); |
| 250 | if (langOpts.OpenMP) |
| 251 | createOpenMPRuntime(); |
| 252 | |
| 253 | // Set the module name to be the name of the main file. TranslationUnitDecl |
| 254 | // often contains invalid source locations and isn't a reliable source for the |
| 255 | // module location. |
| 256 | FileID mainFileId = astContext.getSourceManager().getMainFileID(); |
| 257 | const FileEntry &mainFile = |
| 258 | *astContext.getSourceManager().getFileEntryForID(FID: mainFileId); |
| 259 | StringRef path = mainFile.tryGetRealPathName(); |
| 260 | if (!path.empty()) { |
| 261 | getModule().setSymName(path); |
| 262 | getModule()->setLoc(mlir::FileLineColLoc::get(context: &mlirContext, fileName: path, |
| 263 | /*line=*/0, |
| 264 | /*column=*/0)); |
| 265 | } |
| 266 | } |
| 267 | |
| 268 | CIRGenModule::~CIRGenModule() = default; |
| 269 | |
| 270 | void CIRGenModule::setOpenCLVersionAttr(StringRef attrName, unsigned version) { |
| 271 | getModule()->setAttr( |
| 272 | name: attrName, value: cir::OpenCLVersionAttr::get(context: &getMLIRContext(), majorVersion: version / 100, |
| 273 | minorVersion: (version % 100) / 10)); |
| 274 | } |
| 275 | |
| 276 | void CIRGenModule::createCUDARuntime() { |
| 277 | cudaRuntime.reset(p: createNVCUDARuntime(cgm&: *this)); |
| 278 | } |
| 279 | |
| 280 | void CIRGenModule::createOpenMPRuntime() { |
| 281 | openMPRuntime = std::make_unique<CIRGenOpenMPRuntime>(args&: *this); |
| 282 | } |
| 283 | |
| 284 | /// FIXME: this could likely be a common helper and not necessarily related |
| 285 | /// with codegen. |
| 286 | /// Return the best known alignment for an unknown pointer to a |
| 287 | /// particular class. |
| 288 | CharUnits CIRGenModule::getClassPointerAlignment(const CXXRecordDecl *rd) { |
| 289 | if (!rd->hasDefinition()) |
| 290 | return CharUnits::One(); // Hopefully won't be used anywhere. |
| 291 | |
| 292 | auto &layout = astContext.getASTRecordLayout(D: rd); |
| 293 | |
| 294 | // If the class is final, then we know that the pointer points to an |
| 295 | // object of that type and can use the full alignment. |
| 296 | if (rd->isEffectivelyFinal()) |
| 297 | return layout.getAlignment(); |
| 298 | |
| 299 | // Otherwise, we have to assume it could be a subclass. |
| 300 | return layout.getNonVirtualAlignment(); |
| 301 | } |
| 302 | |
| 303 | CharUnits CIRGenModule::getNaturalTypeAlignment(QualType t, |
| 304 | LValueBaseInfo *baseInfo, |
| 305 | bool forPointeeType) { |
| 306 | assert(!cir::MissingFeatures::opTBAA()); |
| 307 | |
| 308 | // FIXME: This duplicates logic in ASTContext::getTypeAlignIfKnown, but |
| 309 | // that doesn't return the information we need to compute baseInfo. |
| 310 | |
| 311 | // Honor alignment typedef attributes even on incomplete types. |
| 312 | // We also honor them straight for C++ class types, even as pointees; |
| 313 | // there's an expressivity gap here. |
| 314 | if (const auto *tt = t->getAs<TypedefType>()) { |
| 315 | if (unsigned align = tt->getDecl()->getMaxAlignment()) { |
| 316 | if (baseInfo) |
| 317 | *baseInfo = LValueBaseInfo(AlignmentSource::AttributedType); |
| 318 | return astContext.toCharUnitsFromBits(BitSize: align); |
| 319 | } |
| 320 | } |
| 321 | |
| 322 | bool alignForArray = t->isArrayType(); |
| 323 | |
| 324 | // Analyze the base element type, so we don't get confused by incomplete |
| 325 | // array types. |
| 326 | t = astContext.getBaseElementType(QT: t); |
| 327 | |
| 328 | if (t->isIncompleteType()) { |
| 329 | // We could try to replicate the logic from |
| 330 | // ASTContext::getTypeAlignIfKnown, but nothing uses the alignment if the |
| 331 | // type is incomplete, so it's impossible to test. We could try to reuse |
| 332 | // getTypeAlignIfKnown, but that doesn't return the information we need |
| 333 | // to set baseInfo. So just ignore the possibility that the alignment is |
| 334 | // greater than one. |
| 335 | if (baseInfo) |
| 336 | *baseInfo = LValueBaseInfo(AlignmentSource::Type); |
| 337 | return CharUnits::One(); |
| 338 | } |
| 339 | |
| 340 | if (baseInfo) |
| 341 | *baseInfo = LValueBaseInfo(AlignmentSource::Type); |
| 342 | |
| 343 | CharUnits alignment; |
| 344 | const CXXRecordDecl *rd = nullptr; |
| 345 | if (t.getQualifiers().hasUnaligned()) { |
| 346 | alignment = CharUnits::One(); |
| 347 | } else if (forPointeeType && !alignForArray && |
| 348 | (rd = t->getAsCXXRecordDecl())) { |
| 349 | alignment = getClassPointerAlignment(rd); |
| 350 | } else { |
| 351 | alignment = astContext.getTypeAlignInChars(T: t); |
| 352 | } |
| 353 | |
| 354 | // Cap to the global maximum type alignment unless the alignment |
| 355 | // was somehow explicit on the type. |
| 356 | if (unsigned maxAlign = astContext.getLangOpts().MaxTypeAlign) { |
| 357 | if (alignment.getQuantity() > maxAlign && |
| 358 | !astContext.isAlignmentRequired(T: t)) |
| 359 | alignment = CharUnits::fromQuantity(Quantity: maxAlign); |
| 360 | } |
| 361 | return alignment; |
| 362 | } |
| 363 | |
| 364 | CharUnits |
| 365 | CIRGenModule::getNaturalPointeeTypeAlignment(QualType t, |
| 366 | LValueBaseInfo *baseInfo) { |
| 367 | return getNaturalTypeAlignment(t: t->getPointeeType(), baseInfo, |
| 368 | /*forPointeeType=*/true); |
| 369 | } |
| 370 | |
| 371 | const TargetCIRGenInfo &CIRGenModule::getTargetCIRGenInfo() { |
| 372 | if (theTargetCIRGenInfo) |
| 373 | return *theTargetCIRGenInfo; |
| 374 | |
| 375 | const llvm::Triple &triple = getTarget().getTriple(); |
| 376 | switch (triple.getArch()) { |
| 377 | default: |
| 378 | assert(!cir::MissingFeatures::targetCIRGenInfoArch()); |
| 379 | |
| 380 | // Currently we just fall through to x86_64. |
| 381 | [[fallthrough]]; |
| 382 | |
| 383 | case llvm::Triple::x86_64: { |
| 384 | switch (triple.getOS()) { |
| 385 | default: |
| 386 | assert(!cir::MissingFeatures::targetCIRGenInfoOS()); |
| 387 | |
| 388 | // Currently we just fall through to x86_64. |
| 389 | [[fallthrough]]; |
| 390 | |
| 391 | case llvm::Triple::Linux: |
| 392 | theTargetCIRGenInfo = createX8664TargetCIRGenInfo(cgt&: genTypes); |
| 393 | return *theTargetCIRGenInfo; |
| 394 | } |
| 395 | } |
| 396 | case llvm::Triple::aarch64: |
| 397 | case llvm::Triple::aarch64_32: |
| 398 | case llvm::Triple::aarch64_be: { |
| 399 | theTargetCIRGenInfo = createAArch64TargetCIRGenInfo(cgt&: genTypes); |
| 400 | return *theTargetCIRGenInfo; |
| 401 | } |
| 402 | case llvm::Triple::nvptx: |
| 403 | case llvm::Triple::nvptx64: |
| 404 | theTargetCIRGenInfo = createNVPTXTargetCIRGenInfo(cgt&: genTypes); |
| 405 | return *theTargetCIRGenInfo; |
| 406 | case llvm::Triple::amdgpu: { |
| 407 | theTargetCIRGenInfo = createAMDGPUTargetCIRGenInfo(cgt&: genTypes); |
| 408 | return *theTargetCIRGenInfo; |
| 409 | } |
| 410 | case llvm::Triple::spir: |
| 411 | case llvm::Triple::spir64: |
| 412 | case llvm::Triple::spirv: |
| 413 | case llvm::Triple::spirv32: |
| 414 | case llvm::Triple::spirv64: |
| 415 | theTargetCIRGenInfo = createCommonSPIRTargetCIRGenInfo(cgt&: genTypes); |
| 416 | return *theTargetCIRGenInfo; |
| 417 | case llvm::Triple::systemz: { |
| 418 | // z/OS XPLINK target info is not implemented yet. |
| 419 | // We just default to SystemZ ELF Target for now. |
| 420 | if (triple.isOSzOS()) |
| 421 | assert(!cir::MissingFeatures::targetCIRGenInfoOS()); |
| 422 | theTargetCIRGenInfo = createSystemZTargetCIRGenInfo(cgt&: genTypes); |
| 423 | return *theTargetCIRGenInfo; |
| 424 | } |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | mlir::Location CIRGenModule::getLoc(SourceLocation cLoc) { |
| 429 | assert(cLoc.isValid() && "expected valid source location" ); |
| 430 | const SourceManager &sm = astContext.getSourceManager(); |
| 431 | PresumedLoc pLoc = sm.getPresumedLoc(Loc: cLoc); |
| 432 | StringRef filename = pLoc.getFilename(); |
| 433 | return mlir::FileLineColLoc::get(filename: builder.getStringAttr(bytes: filename), |
| 434 | line: pLoc.getLine(), column: pLoc.getColumn()); |
| 435 | } |
| 436 | |
| 437 | mlir::Location CIRGenModule::getLoc(SourceRange cRange) { |
| 438 | assert(cRange.isValid() && "expected a valid source range" ); |
| 439 | mlir::Location begin = getLoc(cLoc: cRange.getBegin()); |
| 440 | mlir::Location end = getLoc(cLoc: cRange.getEnd()); |
| 441 | mlir::Attribute metadata; |
| 442 | return mlir::FusedLoc::get(locs: {begin, end}, metadata, context: builder.getContext()); |
| 443 | } |
| 444 | |
| 445 | mlir::Operation * |
| 446 | CIRGenModule::getAddrOfGlobal(GlobalDecl gd, ForDefinition_t isForDefinition) { |
| 447 | const Decl *d = gd.getDecl(); |
| 448 | |
| 449 | if (isa<CXXConstructorDecl>(Val: d) || isa<CXXDestructorDecl>(Val: d)) |
| 450 | return getAddrOfCXXStructor(gd, /*FnInfo=*/fnInfo: nullptr, /*FnType=*/fnType: nullptr, |
| 451 | /*DontDefer=*/dontDefer: false, isForDefinition); |
| 452 | |
| 453 | if (isa<CXXMethodDecl>(Val: d)) { |
| 454 | const CIRGenFunctionInfo &fi = |
| 455 | getTypes().arrangeCXXMethodDeclaration(md: cast<CXXMethodDecl>(Val: d)); |
| 456 | cir::FuncType ty = getTypes().getFunctionType(info: fi); |
| 457 | return getAddrOfFunction(gd, funcType: ty, /*ForVTable=*/forVTable: false, /*DontDefer=*/dontDefer: false, |
| 458 | isForDefinition); |
| 459 | } |
| 460 | |
| 461 | if (isa<FunctionDecl>(Val: d)) { |
| 462 | const CIRGenFunctionInfo &fi = getTypes().arrangeGlobalDeclaration(gd); |
| 463 | cir::FuncType ty = getTypes().getFunctionType(info: fi); |
| 464 | return getAddrOfFunction(gd, funcType: ty, /*ForVTable=*/forVTable: false, /*DontDefer=*/dontDefer: false, |
| 465 | isForDefinition); |
| 466 | } |
| 467 | |
| 468 | return getOrCreateCIRGlobal(d: cast<VarDecl>(Val: d), /*ty=*/nullptr, |
| 469 | isForDefinition); |
| 470 | } |
| 471 | |
| 472 | void CIRGenModule::emitGlobalDecl(const clang::GlobalDecl &d) { |
| 473 | // We call getAddrOfGlobal with isForDefinition set to ForDefinition in |
| 474 | // order to get a Value with exactly the type we need, not something that |
| 475 | // might have been created for another decl with the same mangled name but |
| 476 | // different type. |
| 477 | mlir::Operation *op = getAddrOfGlobal(gd: d, isForDefinition: ForDefinition); |
| 478 | |
| 479 | // In case of different address spaces, we may still get a cast, even with |
| 480 | // IsForDefinition equal to ForDefinition. Query mangled names table to get |
| 481 | // GlobalValue. |
| 482 | if (!op) |
| 483 | op = getGlobalValue(ref: getMangledName(gd: d)); |
| 484 | |
| 485 | assert(op && "expected a valid global op" ); |
| 486 | |
| 487 | // Check to see if we've already emitted this. This is necessary for a |
| 488 | // couple of reasons: first, decls can end up in deferred-decls queue |
| 489 | // multiple times, and second, decls can end up with definitions in unusual |
| 490 | // ways (e.g. by an extern inline function acquiring a strong function |
| 491 | // redefinition). Just ignore those cases. |
| 492 | // TODO: Not sure what to map this to for MLIR |
| 493 | mlir::Operation *globalValueOp = op; |
| 494 | if (auto gv = dyn_cast<cir::GetGlobalOp>(Val: op)) { |
| 495 | globalValueOp = getGlobalValue(ref: gv.getName()); |
| 496 | assert(globalValueOp && "expected a valid global op" ); |
| 497 | } |
| 498 | |
| 499 | if (auto cirGlobalValue = |
| 500 | dyn_cast<cir::CIRGlobalValueInterface>(Val: globalValueOp)) |
| 501 | if (!cirGlobalValue.isDeclaration()) |
| 502 | return; |
| 503 | |
| 504 | // If this is OpenMP, check if it is legal to emit this global normally. |
| 505 | assert(!cir::MissingFeatures::openMP()); |
| 506 | |
| 507 | // Otherwise, emit the definition and move on to the next one. |
| 508 | emitGlobalDefinition(gd: d, op); |
| 509 | } |
| 510 | |
| 511 | void CIRGenModule::emitDeferred() { |
| 512 | // Emit code for any potentially referenced deferred decls. Since a previously |
| 513 | // unused static decl may become used during the generation of code for a |
| 514 | // static function, iterate until no changes are made. |
| 515 | |
| 516 | assert(!cir::MissingFeatures::openMP()); |
| 517 | |
| 518 | emitDeferredVTables(); |
| 519 | // Emitting a vtable doesn't directly cause more vtables to |
| 520 | // become deferred, although it can cause functions to be |
| 521 | // emitted that then need those vtables. |
| 522 | assert(deferredVTables.empty()); |
| 523 | |
| 524 | // Emit CUDA/HIP static device variables referenced by host code only. |
| 525 | // Note we should not clear CUDADeviceVarODRUsedByHost since it is still |
| 526 | // needed for further handling. |
| 527 | if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) |
| 528 | llvm::append_range(C&: deferredDeclsToEmit, |
| 529 | R&: getASTContext().CUDADeviceVarODRUsedByHost); |
| 530 | |
| 531 | // Stop if we're out of both deferred vtables and deferred declarations. |
| 532 | if (deferredDeclsToEmit.empty()) |
| 533 | return; |
| 534 | |
| 535 | // Grab the list of decls to emit. If emitGlobalDefinition schedules more |
| 536 | // work, it will not interfere with this. |
| 537 | std::vector<GlobalDecl> curDeclsToEmit; |
| 538 | curDeclsToEmit.swap(x&: deferredDeclsToEmit); |
| 539 | |
| 540 | for (const GlobalDecl &d : curDeclsToEmit) { |
| 541 | // Functions declared with the sycl_kernel_entry_point attribute are |
| 542 | // emitted normally during host compilation. During device compilation, a |
| 543 | // SYCL kernel caller offload entry point function is generated and emitted |
| 544 | // in place of each of these functions. |
| 545 | if (const auto *fd = d.getDecl()->getAsFunction()) { |
| 546 | if (langOpts.SYCLIsDevice && fd->hasAttr<SYCLKernelEntryPointAttr>() && |
| 547 | fd->isDefined()) { |
| 548 | // Functions with an invalid sycl_kernel_entry_point attribute are |
| 549 | // ignored during device compilation. |
| 550 | if (!fd->getAttr<SYCLKernelEntryPointAttr>()->isInvalidAttr()) { |
| 551 | // Generate and emit the SYCL kernel caller function. |
| 552 | emitSYCLKernelCaller(kernelEntryPointFn: fd, ctx&: getASTContext()); |
| 553 | // Recurse to emit any symbols directly or indirectly referenced |
| 554 | // by the SYCL kernel caller function. |
| 555 | emitDeferred(); |
| 556 | } |
| 557 | // Do not emit the sycl_kernel_entry_point attributed function. |
| 558 | continue; |
| 559 | } |
| 560 | } |
| 561 | |
| 562 | emitGlobalDecl(d); |
| 563 | |
| 564 | // If we found out that we need to emit more decls, do that recursively. |
| 565 | // This has the advantage that the decls are emitted in a DFS and related |
| 566 | // ones are close together, which is convenient for testing. |
| 567 | if (!deferredVTables.empty() || !deferredDeclsToEmit.empty()) { |
| 568 | emitDeferred(); |
| 569 | assert(deferredVTables.empty() && deferredDeclsToEmit.empty()); |
| 570 | } |
| 571 | } |
| 572 | } |
| 573 | |
| 574 | void CIRGenModule::printPostfixForExternalizedDecl(llvm::raw_ostream &os, |
| 575 | const Decl *d) { |
| 576 | // ptxas does not allow '.' in symbol names. On the other hand, HIP prefers |
| 577 | // postfix beginning with '.' since the symbol name can be demangled. |
| 578 | if (langOpts.HIP) |
| 579 | os << (isa<VarDecl>(Val: d) ? ".static." : ".intern." ); |
| 580 | else |
| 581 | os << (isa<VarDecl>(Val: d) ? "__static__" : "__intern__" ); |
| 582 | |
| 583 | // If the CUID is not specified we try to generate a unique postfix. |
| 584 | if (getLangOpts().CUID.empty()) { |
| 585 | // TODO: Once we add 'PreprocessorOpts' into CIRGenModule this part can be |
| 586 | // brought in from OG. |
| 587 | errorNYI(d->getSourceRange(), |
| 588 | "printPostfixForExternalizedDecl: CUID is not specified" ); |
| 589 | } else { |
| 590 | os << getASTContext().getCUIDHash(); |
| 591 | } |
| 592 | } |
| 593 | |
| 594 | static bool shouldSkipAliasEmission(const CIRGenModule &cgm, |
| 595 | const ValueDecl *global) { |
| 596 | const LangOptions &langOpts = cgm.getLangOpts(); |
| 597 | if (!langOpts.OpenMPIsTargetDevice && !langOpts.CUDA) |
| 598 | return false; |
| 599 | |
| 600 | const auto *aa = global->getAttr<AliasAttr>(); |
| 601 | GlobalDecl aliaseeGD; |
| 602 | |
| 603 | // Check if the aliasee exists, if the aliasee is not found, skip the alias |
| 604 | // emission. This is executed for both the host and device. |
| 605 | if (!cgm.lookupRepresentativeDecl(mangledName: aa->getAliasee(), gd&: aliaseeGD)) |
| 606 | return true; |
| 607 | |
| 608 | const auto *aliaseeDecl = dyn_cast<ValueDecl>(Val: aliaseeGD.getDecl()); |
| 609 | if (langOpts.OpenMPIsTargetDevice) |
| 610 | return !aliaseeDecl || |
| 611 | !OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD: aliaseeDecl); |
| 612 | |
| 613 | // CUDA / HIP |
| 614 | const bool hasDeviceAttr = global->hasAttr<CUDADeviceAttr>(); |
| 615 | const bool aliaseeHasDeviceAttr = |
| 616 | aliaseeDecl && aliaseeDecl->hasAttr<CUDADeviceAttr>(); |
| 617 | |
| 618 | if (langOpts.CUDAIsDevice) |
| 619 | return !hasDeviceAttr || !aliaseeHasDeviceAttr; |
| 620 | |
| 621 | // CUDA / HIP Host |
| 622 | // we know that the aliasee exists from above, so we know to emit |
| 623 | return false; |
| 624 | } |
| 625 | |
| 626 | void CIRGenModule::emitGlobal(clang::GlobalDecl gd) { |
| 627 | if (const auto *cd = dyn_cast<clang::OpenACCConstructDecl>(Val: gd.getDecl())) { |
| 628 | emitGlobalOpenACCDecl(cd); |
| 629 | return; |
| 630 | } |
| 631 | |
| 632 | const auto *global = cast<ValueDecl>(Val: gd.getDecl()); |
| 633 | |
| 634 | // Weak references don't produce any output by themselves. |
| 635 | if (global->hasAttr<WeakRefAttr>()) |
| 636 | return; |
| 637 | |
| 638 | // If this is an alias definition (which otherwise looks like a declaration) |
| 639 | // emit it now. |
| 640 | if (global->hasAttr<AliasAttr>()) { |
| 641 | if (shouldSkipAliasEmission(cgm: *this, global)) |
| 642 | return; |
| 643 | emitAliasDefinition(gd); |
| 644 | return; |
| 645 | } |
| 646 | |
| 647 | // If this is CUDA, be selective about which declarations we emit. |
| 648 | if (langOpts.CUDA && !getASTContext().shouldEmitCUDADecl(D: global)) |
| 649 | return; |
| 650 | |
| 651 | if (langOpts.OpenMP) { |
| 652 | // If this is OpenMP, check if it is legal to emit this global normally. |
| 653 | if (openMPRuntime && openMPRuntime->emitTargetGlobal(gd)) |
| 654 | return; |
| 655 | if (auto *drd = dyn_cast<OMPDeclareReductionDecl>(Val: global)) { |
| 656 | if (mustBeEmitted(d: global)) |
| 657 | emitOMPDeclareReduction(d: drd); |
| 658 | return; |
| 659 | } |
| 660 | if (auto *dmd = dyn_cast<OMPDeclareMapperDecl>(Val: global)) { |
| 661 | if (mustBeEmitted(d: global)) |
| 662 | emitOMPDeclareMapper(d: dmd); |
| 663 | return; |
| 664 | } |
| 665 | } |
| 666 | |
| 667 | if (const auto *fd = dyn_cast<FunctionDecl>(Val: global)) { |
| 668 | // Update deferred annotations with the latest declaration if the function |
| 669 | // was already used or defined. |
| 670 | if (fd->hasAttr<AnnotateAttr>()) { |
| 671 | StringRef mangledName = getMangledName(gd); |
| 672 | if (getGlobalValue(ref: mangledName)) |
| 673 | deferredAnnotations[mangledName] = fd; |
| 674 | } |
| 675 | if (!fd->doesThisDeclarationHaveABody()) { |
| 676 | if (!fd->doesDeclarationForceExternallyVisibleDefinition() && |
| 677 | (!fd->isMultiVersion() || !getTarget().getTriple().isAArch64())) |
| 678 | return; |
| 679 | |
| 680 | const CIRGenFunctionInfo &fi = getTypes().arrangeGlobalDeclaration(gd); |
| 681 | cir::FuncType ty = getTypes().getFunctionType(info: fi); |
| 682 | getAddrOfFunction(gd, funcType: ty, /*ForVTable=*/forVTable: false, /*DontDefer=*/dontDefer: false); |
| 683 | return; |
| 684 | } |
| 685 | } else { |
| 686 | const auto *vd = cast<VarDecl>(Val: global); |
| 687 | assert(vd->isFileVarDecl() && "Cannot emit local var decl as global." ); |
| 688 | if (vd->isThisDeclarationADefinition() != VarDecl::Definition && |
| 689 | !astContext.isMSStaticDataMemberInlineDefinition(VD: vd)) { |
| 690 | assert(!cir::MissingFeatures::openMP()); |
| 691 | // If this declaration may have caused an inline variable definition to |
| 692 | // change linkage, make sure that it's emitted. |
| 693 | if (astContext.getInlineVariableDefinitionKind(VD: vd) == |
| 694 | ASTContext::InlineVariableDefinitionKind::Strong) |
| 695 | getOrCreateCIRGlobal(d: vd, /*ty=*/{}, isForDefinition: NotForDefinition); |
| 696 | // Otherwise, we can ignore this declaration. The variable will be emitted |
| 697 | // on its first use. |
| 698 | return; |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | // Defer code generation to first use when possible, e.g. if this is an inline |
| 703 | // function. If the global must always be emitted, do it eagerly if possible |
| 704 | // to benefit from cache locality. Deferring code generation is necessary to |
| 705 | // avoid adding initializers to external declarations. |
| 706 | if (mustBeEmitted(d: global) && mayBeEmittedEagerly(d: global)) { |
| 707 | // Make sure we don't double-emit this declaration. |
| 708 | if (auto existing = dyn_cast_if_present<cir::CIRGlobalValueInterface>( |
| 709 | Val: getGlobalValue(ref: getMangledName(gd))); |
| 710 | existing && !existing.isDeclaration()) |
| 711 | return; |
| 712 | |
| 713 | // Emit the definition if it can't be deferred. |
| 714 | emitGlobalDefinition(gd); |
| 715 | return; |
| 716 | } |
| 717 | |
| 718 | // If we're deferring emission of a C++ variable with an initializer, remember |
| 719 | // the order in which it appeared on the file. |
| 720 | assert(!cir::MissingFeatures::deferredCXXGlobalInit()); |
| 721 | |
| 722 | llvm::StringRef mangledName = getMangledName(gd); |
| 723 | if (getGlobalValue(ref: mangledName) != nullptr) { |
| 724 | // The value has already been used and should therefore be emitted. |
| 725 | addDeferredDeclToEmit(GD: gd); |
| 726 | } else if (mustBeEmitted(d: global)) { |
| 727 | // The value must be emitted, but cannot be emitted eagerly. |
| 728 | assert(!mayBeEmittedEagerly(global)); |
| 729 | addDeferredDeclToEmit(GD: gd); |
| 730 | } else { |
| 731 | // Otherwise, remember that we saw a deferred decl with this name. The first |
| 732 | // use of the mangled name will cause it to move into deferredDeclsToEmit. |
| 733 | deferredDecls[mangledName] = gd; |
| 734 | } |
| 735 | } |
| 736 | |
| 737 | void CIRGenModule::emitGlobalFunctionDefinition(clang::GlobalDecl gd, |
| 738 | mlir::Operation *op) { |
| 739 | auto const *funcDecl = cast<FunctionDecl>(Val: gd.getDecl()); |
| 740 | const CIRGenFunctionInfo &fi = getTypes().arrangeGlobalDeclaration(gd); |
| 741 | cir::FuncType funcType = getTypes().getFunctionType(info: fi); |
| 742 | cir::FuncOp funcOp = dyn_cast_if_present<cir::FuncOp>(Val: op); |
| 743 | if (!funcOp || funcOp.getFunctionType() != funcType) { |
| 744 | funcOp = getAddrOfFunction(gd, funcType, /*ForVTable=*/forVTable: false, |
| 745 | /*DontDefer=*/dontDefer: true, isForDefinition: ForDefinition); |
| 746 | } |
| 747 | |
| 748 | // Already emitted. |
| 749 | if (!funcOp.isDeclaration()) |
| 750 | return; |
| 751 | |
| 752 | setFunctionLinkage(gd, f: funcOp); |
| 753 | setGVProperties(op: funcOp, d: funcDecl); |
| 754 | assert(!cir::MissingFeatures::opFuncMaybeHandleStaticInExternC()); |
| 755 | maybeSetTrivialComdat(d: *funcDecl, op: funcOp); |
| 756 | assert(!cir::MissingFeatures::setLLVMFunctionFEnvAttributes()); |
| 757 | |
| 758 | CIRGenFunction cgf(*this, builder); |
| 759 | curCGF = &cgf; |
| 760 | { |
| 761 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 762 | cgf.generateCode(gd, fn: funcOp, funcType); |
| 763 | } |
| 764 | curCGF = nullptr; |
| 765 | |
| 766 | setNonAliasAttributes(gd, op: funcOp); |
| 767 | setCIRFunctionAttributesForDefinition(decl: funcDecl, f: funcOp); |
| 768 | |
| 769 | auto getPriority = [this](const auto *attr) -> int { |
| 770 | Expr *e = attr->getPriority(); |
| 771 | if (e) |
| 772 | return e->EvaluateKnownConstInt(Ctx: this->getASTContext()).getExtValue(); |
| 773 | return attr->DefaultPriority; |
| 774 | }; |
| 775 | |
| 776 | if (const ConstructorAttr *ca = funcDecl->getAttr<ConstructorAttr>()) |
| 777 | addGlobalCtor(ctor: funcOp, priority: getPriority(ca)); |
| 778 | if (const DestructorAttr *da = funcDecl->getAttr<DestructorAttr>()) |
| 779 | addGlobalDtor(dtor: funcOp, priority: getPriority(da)); |
| 780 | |
| 781 | if (funcDecl->getAttr<AnnotateAttr>()) |
| 782 | deferredAnnotations[getMangledName(gd)] = funcDecl; |
| 783 | |
| 784 | if (getLangOpts().OpenMP && funcDecl->hasAttr<OMPDeclareTargetDeclAttr>()) |
| 785 | getOpenMPRuntime().emitDeclareTargetFunction(fd: funcDecl, funcOp); |
| 786 | } |
| 787 | |
| 788 | /// Track functions to be called before main() runs. |
| 789 | void CIRGenModule::addGlobalCtor(cir::FuncOp ctor, |
| 790 | std::optional<int> priority) { |
| 791 | assert(!cir::MissingFeatures::globalCtorLexOrder()); |
| 792 | assert(!cir::MissingFeatures::globalCtorAssociatedData()); |
| 793 | |
| 794 | // Traditional LLVM codegen directly adds the function to the list of global |
| 795 | // ctors. In CIR we just add a global_ctor attribute to the function. The |
| 796 | // global list is created in LoweringPrepare. |
| 797 | // |
| 798 | // FIXME(from traditional LLVM): Type coercion of void()* types. |
| 799 | ctor.setGlobalCtorPriority(priority); |
| 800 | } |
| 801 | |
| 802 | /// Add a function to the list that will be called when the module is unloaded. |
| 803 | void CIRGenModule::addGlobalDtor(cir::FuncOp dtor, |
| 804 | std::optional<int> priority) { |
| 805 | if (codeGenOpts.RegisterGlobalDtorsWithAtExit && |
| 806 | (!getASTContext().getTargetInfo().getTriple().isOSAIX())) |
| 807 | errorNYI(loc: dtor.getLoc(), feature: "registerGlobalDtorsWithAtExit" ); |
| 808 | |
| 809 | // FIXME(from traditional LLVM): Type coercion of void()* types. |
| 810 | dtor.setGlobalDtorPriority(priority); |
| 811 | } |
| 812 | |
| 813 | void CIRGenModule::handleCXXStaticMemberVarInstantiation(VarDecl *vd) { |
| 814 | VarDecl::DefinitionKind dk = vd->isThisDeclarationADefinition(); |
| 815 | if ((dk == VarDecl::Definition && vd->hasAttr<DLLImportAttr>()) || |
| 816 | (langOpts.CUDA && !getASTContext().shouldEmitCUDADecl(D: vd))) |
| 817 | return; |
| 818 | |
| 819 | TemplateSpecializationKind tsk = vd->getTemplateSpecializationKind(); |
| 820 | // If we have a definition, this might be a deferred decl. If the |
| 821 | // instantiation is explicit, make sure we emit it at the end. |
| 822 | if (vd->getDefinition() && tsk == TSK_ExplicitInstantiationDefinition) |
| 823 | getOrCreateCIRGlobal(d: vd, /*ty=*/{}, isForDefinition: NotForDefinition); |
| 824 | |
| 825 | emitTopLevelDecl(decl: vd); |
| 826 | } |
| 827 | |
| 828 | mlir::Operation *CIRGenModule::getGlobalValue(StringRef name) { |
| 829 | auto it = symbolLookupCache.find(Key: name); |
| 830 | return it != symbolLookupCache.end() ? it->second : nullptr; |
| 831 | } |
| 832 | |
| 833 | cir::GlobalOp |
| 834 | CIRGenModule::createGlobalOp(mlir::Location loc, StringRef name, mlir::Type t, |
| 835 | bool isConstant, |
| 836 | mlir::ptr::MemorySpaceAttrInterface addrSpace, |
| 837 | mlir::Operation *insertPoint) { |
| 838 | cir::GlobalOp g; |
| 839 | CIRGenBuilderTy &builder = getBuilder(); |
| 840 | |
| 841 | { |
| 842 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 843 | |
| 844 | // If an insertion point is provided, we're replacing an existing global, |
| 845 | // otherwise, create the new global immediately after the last gloabl we |
| 846 | // emitted. |
| 847 | if (insertPoint) { |
| 848 | builder.setInsertionPoint(insertPoint); |
| 849 | } else { |
| 850 | // Group global operations together at the top of the module. |
| 851 | if (lastGlobalOp) |
| 852 | builder.setInsertionPointAfter(lastGlobalOp); |
| 853 | else |
| 854 | builder.setInsertionPointToStart(getModule().getBody()); |
| 855 | } |
| 856 | |
| 857 | g = cir::GlobalOp::create(builder, location: loc, sym_name: name, sym_type: t, isConstant, addrSpace); |
| 858 | if (!insertPoint) |
| 859 | lastGlobalOp = g; |
| 860 | |
| 861 | // Default to private until we can judge based on the initializer, |
| 862 | // since MLIR doesn't allow public declarations. |
| 863 | mlir::SymbolTable::setSymbolVisibility( |
| 864 | symbol: g, vis: mlir::SymbolTable::Visibility::Private); |
| 865 | } |
| 866 | symbolLookupCache[g.getSymNameAttr()] = g; |
| 867 | return g; |
| 868 | } |
| 869 | |
| 870 | void CIRGenModule::setCommonAttributes(GlobalDecl gd, mlir::Operation *gv) { |
| 871 | const Decl *d = gd.getDecl(); |
| 872 | if (isa_and_nonnull<NamedDecl>(Val: d)) |
| 873 | setGVProperties(op: gv, d: dyn_cast<NamedDecl>(Val: d)); |
| 874 | assert(!cir::MissingFeatures::defaultVisibility()); |
| 875 | |
| 876 | if (auto gvi = mlir::dyn_cast<cir::CIRGlobalValueInterface>(Val: gv)) { |
| 877 | if (d && d->hasAttr<UsedAttr>()) |
| 878 | addUsedOrCompilerUsedGlobal(gv: gvi); |
| 879 | |
| 880 | if (const auto *vd = dyn_cast_if_present<VarDecl>(Val: d); |
| 881 | vd && ((codeGenOpts.KeepPersistentStorageVariables && |
| 882 | (vd->getStorageDuration() == SD_Static || |
| 883 | vd->getStorageDuration() == SD_Thread)) || |
| 884 | (codeGenOpts.KeepStaticConsts && |
| 885 | vd->getStorageDuration() == SD_Static && |
| 886 | vd->getType().isConstQualified()))) |
| 887 | addUsedOrCompilerUsedGlobal(gv: gvi); |
| 888 | } |
| 889 | } |
| 890 | |
| 891 | /// Get the feature delta from the default feature map for the given target CPU. |
| 892 | static std::vector<std::string> |
| 893 | getFeatureDeltaFromDefault(const CIRGenModule &cgm, llvm::StringRef targetCPU, |
| 894 | llvm::StringMap<bool> &featureMap) { |
| 895 | llvm::StringMap<bool> defaultFeatureMap; |
| 896 | cgm.getTarget().initFeatureMap( |
| 897 | Features&: defaultFeatureMap, Diags&: cgm.getASTContext().getDiagnostics(), CPU: targetCPU, FeatureVec: {}); |
| 898 | |
| 899 | std::vector<std::string> delta; |
| 900 | for (const auto &[k, v] : featureMap) { |
| 901 | auto defaultIt = defaultFeatureMap.find(Key: k); |
| 902 | if (defaultIt == defaultFeatureMap.end() || defaultIt->getValue() != v) |
| 903 | delta.push_back(x: (v ? "+" : "-" ) + k.str()); |
| 904 | } |
| 905 | |
| 906 | return delta; |
| 907 | } |
| 908 | |
| 909 | bool CIRGenModule::getCPUAndFeaturesAttributes( |
| 910 | GlobalDecl gd, llvm::StringMap<std::string> &attrs, |
| 911 | bool setTargetFeatures) { |
| 912 | // Add target-cpu and target-features attributes to functions. If |
| 913 | // we have a decl for the function and it has a target attribute then |
| 914 | // parse that and add it to the feature set. |
| 915 | llvm::StringRef targetCPU = getTarget().getTargetOpts().CPU; |
| 916 | llvm::StringRef tuneCPU = getTarget().getTargetOpts().TuneCPU; |
| 917 | std::vector<std::string> features; |
| 918 | // `fd` may be null when emitting attributes for globals that don't have a |
| 919 | // FunctionDecl. The AMDGPU branch below handles |
| 920 | // the null case via initFeatureMap. |
| 921 | const auto *fd = dyn_cast_or_null<FunctionDecl>(Val: gd.getDecl()); |
| 922 | fd = fd ? fd->getMostRecentDecl() : fd; |
| 923 | const auto *td = fd ? fd->getAttr<TargetAttr>() : nullptr; |
| 924 | const auto *tv = fd ? fd->getAttr<TargetVersionAttr>() : nullptr; |
| 925 | assert((!td || !tv) && "both target_version and target specified" ); |
| 926 | const auto *sd = fd ? fd->getAttr<CPUSpecificAttr>() : nullptr; |
| 927 | const auto *tc = fd ? fd->getAttr<TargetClonesAttr>() : nullptr; |
| 928 | bool addedAttr = false; |
| 929 | if (td || tv || sd || tc) { |
| 930 | llvm::StringMap<bool> featureMap; |
| 931 | astContext.getFunctionFeatureMap(FeatureMap&: featureMap, GD: gd); |
| 932 | |
| 933 | // Now add the target-cpu and target-features to the function. |
| 934 | // While we populated the feature map above, we still need to |
| 935 | // get and parse the target/target_clones attribute so we can |
| 936 | // get the cpu for the function. |
| 937 | llvm::StringRef featureStr = td ? td->getFeaturesStr() : llvm::StringRef(); |
| 938 | if (tc && (getTriple().isOSAIX() || getTriple().isX86())) |
| 939 | featureStr = tc->getFeatureStr(Index: gd.getMultiVersionIndex()); |
| 940 | if (!featureStr.empty()) { |
| 941 | clang::ParsedTargetAttr parsedAttr = |
| 942 | getTarget().parseTargetAttr(Str: featureStr); |
| 943 | if (!parsedAttr.CPU.empty() && |
| 944 | getTarget().isValidCPUName(Name: parsedAttr.CPU)) { |
| 945 | targetCPU = parsedAttr.CPU; |
| 946 | tuneCPU = "" ; // Clear the tune CPU. |
| 947 | } |
| 948 | if (!parsedAttr.Tune.empty() && |
| 949 | getTarget().isValidCPUName(Name: parsedAttr.Tune)) |
| 950 | tuneCPU = parsedAttr.Tune; |
| 951 | } |
| 952 | |
| 953 | if (sd) { |
| 954 | // Apply the given CPU name as the 'tune-cpu' so that the optimizer can |
| 955 | // favor this processor. |
| 956 | tuneCPU = sd->getCPUName(Index: gd.getMultiVersionIndex())->getName(); |
| 957 | } |
| 958 | |
| 959 | // For AMDGPU, only emit delta features (features that differ from the |
| 960 | // target CPU's defaults). Other targets might want to follow a similar |
| 961 | // pattern. |
| 962 | if (getTarget().getTriple().isAMDGPU()) { |
| 963 | features = getFeatureDeltaFromDefault(cgm: *this, targetCPU, featureMap); |
| 964 | } else { |
| 965 | // Produce the canonical string for this set of features. |
| 966 | features.reserve(n: features.size() + featureMap.size()); |
| 967 | for (const auto &entry : featureMap) |
| 968 | features.push_back(x: (entry.getValue() ? "+" : "-" ) + |
| 969 | entry.getKey().str()); |
| 970 | } |
| 971 | } else { |
| 972 | // Just add the existing target cpu and target features to the function. |
| 973 | if (setTargetFeatures && getTarget().getTriple().isAMDGPU()) { |
| 974 | llvm::StringMap<bool> featureMap; |
| 975 | if (fd) |
| 976 | astContext.getFunctionFeatureMap(FeatureMap&: featureMap, GD: gd); |
| 977 | else |
| 978 | getTarget().initFeatureMap(Features&: featureMap, Diags&: astContext.getDiagnostics(), |
| 979 | CPU: targetCPU, |
| 980 | FeatureVec: getTarget().getTargetOpts().Features); |
| 981 | features = getFeatureDeltaFromDefault(cgm: *this, targetCPU, featureMap); |
| 982 | } else { |
| 983 | features = getTarget().getTargetOpts().Features; |
| 984 | } |
| 985 | } |
| 986 | |
| 987 | if (!targetCPU.empty()) { |
| 988 | attrs[cir::CIRDialect::getTargetCPUAttrName()] = targetCPU.str(); |
| 989 | addedAttr = true; |
| 990 | } |
| 991 | if (!tuneCPU.empty()) { |
| 992 | attrs[cir::CIRDialect::getTuneCPUAttrName()] = tuneCPU.str(); |
| 993 | addedAttr = true; |
| 994 | } |
| 995 | if (!features.empty() && setTargetFeatures) { |
| 996 | llvm::erase_if(C&: features, P: [&](const std::string &f) { |
| 997 | assert(!f.empty() && (f[0] == '+' || f[0] == '-') && |
| 998 | "feature string must start with '+' or '-'" ); |
| 999 | return getTarget().isReadOnlyFeature(Feature: f.substr(pos: 1)); |
| 1000 | }); |
| 1001 | llvm::sort(C&: features); |
| 1002 | attrs[cir::CIRDialect::getTargetFeaturesAttrName()] = |
| 1003 | llvm::join(R&: features, Separator: "," ); |
| 1004 | addedAttr = true; |
| 1005 | } |
| 1006 | // TODO(cir): add metadata for AArch64 Function Multi Versioning. |
| 1007 | assert(!cir::MissingFeatures::opFuncMultiVersioning()); |
| 1008 | return addedAttr; |
| 1009 | } |
| 1010 | |
| 1011 | void CIRGenModule::setNonAliasAttributes(GlobalDecl gd, mlir::Operation *op) { |
| 1012 | setCommonAttributes(gd, gv: op); |
| 1013 | |
| 1014 | const Decl *d = gd.getDecl(); |
| 1015 | if (d) { |
| 1016 | if (auto gvi = mlir::dyn_cast<cir::CIRGlobalValueInterface>(Val: op)) { |
| 1017 | if (const auto *sa = d->getAttr<SectionAttr>()) |
| 1018 | gvi.setSection(builder.getStringAttr(bytes: sa->getName())); |
| 1019 | if (d->hasAttr<RetainAttr>()) |
| 1020 | addUsedGlobal(gv: gvi); |
| 1021 | |
| 1022 | if (auto func = dyn_cast<cir::FuncOp>(Val: op)) { |
| 1023 | llvm::StringMap<std::string> attrs; |
| 1024 | if (getCPUAndFeaturesAttributes(gd, attrs)) { |
| 1025 | // TODO(cir): Classic codegen also removes fmv-features here, which |
| 1026 | // CIR does not emit yet. |
| 1027 | // |
| 1028 | // getCPUAndFeaturesAttributes reads the most recent declaration, so |
| 1029 | // its result supersedes anything an earlier one wrote. Clear first: |
| 1030 | // setAttr alone would leave a name this call no longer produces. |
| 1031 | for (llvm::StringRef name : |
| 1032 | {cir::CIRDialect::getTargetCPUAttrName(), |
| 1033 | cir::CIRDialect::getTuneCPUAttrName(), |
| 1034 | cir::CIRDialect::getTargetFeaturesAttrName()}) |
| 1035 | func->removeAttr(name); |
| 1036 | for (const auto &[key, val] : attrs) |
| 1037 | func->setAttr(name: key, value: builder.getStringAttr(bytes: val)); |
| 1038 | } |
| 1039 | } |
| 1040 | } |
| 1041 | } |
| 1042 | |
| 1043 | assert(!cir::MissingFeatures::opGlobalPragmaClangSection()); |
| 1044 | getTargetCIRGenInfo().setTargetAttributes(decl: gd.getDecl(), global: op, module&: *this); |
| 1045 | } |
| 1046 | |
| 1047 | std::optional<cir::SourceLanguage> CIRGenModule::getCIRSourceLanguage() const { |
| 1048 | using ClangStd = clang::LangStandard; |
| 1049 | using CIRLang = cir::SourceLanguage; |
| 1050 | auto opts = getLangOpts(); |
| 1051 | |
| 1052 | if (opts.OpenCLCPlusPlus) |
| 1053 | return CIRLang::OpenCLCXX; |
| 1054 | if (opts.OpenCL) |
| 1055 | return CIRLang::OpenCLC; |
| 1056 | if (opts.CPlusPlus) |
| 1057 | return CIRLang::CXX; |
| 1058 | if (opts.C99 || opts.C11 || opts.C17 || opts.C23 || opts.C2y || |
| 1059 | opts.LangStd == ClangStd::lang_c89 || |
| 1060 | opts.LangStd == ClangStd::lang_gnu89) |
| 1061 | return CIRLang::C; |
| 1062 | |
| 1063 | // TODO(cir): support remaining source languages. |
| 1064 | assert(!cir::MissingFeatures::sourceLanguageCases()); |
| 1065 | errorNYI(feature: "CIR does not yet support the given source language" ); |
| 1066 | return std::nullopt; |
| 1067 | } |
| 1068 | |
| 1069 | LangAS CIRGenModule::getGlobalVarAddressSpace(const VarDecl *d) { |
| 1070 | if (langOpts.OpenCL) { |
| 1071 | LangAS as = d ? d->getType().getAddressSpace() : LangAS::opencl_global; |
| 1072 | assert(as == LangAS::opencl_global || as == LangAS::opencl_global_device || |
| 1073 | as == LangAS::opencl_global_host || as == LangAS::opencl_constant || |
| 1074 | as == LangAS::opencl_local || as >= LangAS::FirstTargetAddressSpace); |
| 1075 | return as; |
| 1076 | } |
| 1077 | |
| 1078 | if (langOpts.SYCLIsDevice && |
| 1079 | (!d || d->getType().getAddressSpace() == LangAS::Default)) |
| 1080 | return LangAS::sycl_global; |
| 1081 | |
| 1082 | if (langOpts.CUDA && langOpts.CUDAIsDevice) { |
| 1083 | if (d) { |
| 1084 | if (d->hasAttr<CUDAConstantAttr>()) |
| 1085 | return LangAS::cuda_constant; |
| 1086 | if (d->hasAttr<CUDASharedAttr>()) |
| 1087 | return LangAS::cuda_shared; |
| 1088 | if (d->hasAttr<CUDADeviceAttr>()) |
| 1089 | return LangAS::cuda_device; |
| 1090 | if (d->getType().isConstQualified()) |
| 1091 | return LangAS::cuda_constant; |
| 1092 | } |
| 1093 | return LangAS::cuda_device; |
| 1094 | } |
| 1095 | |
| 1096 | if (langOpts.OpenMP && d && d->hasAttr<OMPAllocateDeclAttr>()) |
| 1097 | errorNYI(d->getSourceRange(), "OpenMP global address space" ); |
| 1098 | |
| 1099 | return getTargetCIRGenInfo().getGlobalVarAddressSpace(cgm&: *this, d); |
| 1100 | } |
| 1101 | |
| 1102 | static void setLinkageForGV(cir::GlobalOp &gv, const NamedDecl *nd) { |
| 1103 | // Set linkage and visibility in case we never see a definition. |
| 1104 | LinkageInfo lv = nd->getLinkageAndVisibility(); |
| 1105 | // Don't set internal linkage on declarations. |
| 1106 | // "extern_weak" is overloaded in LLVM; we probably should have |
| 1107 | // separate linkage types for this. |
| 1108 | if (isExternallyVisible(L: lv.getLinkage()) && |
| 1109 | (nd->hasAttr<WeakAttr>() || nd->isWeakImported())) |
| 1110 | gv.setLinkage(cir::GlobalLinkageKind::ExternalWeakLinkage); |
| 1111 | } |
| 1112 | |
| 1113 | static void setLinkageForFunction(CIRGenModule &cgm, cir::FuncOp &func, |
| 1114 | const NamedDecl *nd) { |
| 1115 | // Mirrors CodeGenModule::setLinkageForGV for function declarations. |
| 1116 | LinkageInfo lv = nd->getLinkageAndVisibility(); |
| 1117 | if (isExternallyVisible(L: lv.getLinkage()) && |
| 1118 | (nd->hasAttr<WeakAttr>() || nd->isWeakImported())) { |
| 1119 | auto linkage = cir::GlobalLinkageKind::ExternalWeakLinkage; |
| 1120 | func.setLinkage(linkage); |
| 1121 | func.setLinkageAttr( |
| 1122 | cir::GlobalLinkageKindAttr::get(context: &cgm.getMLIRContext(), value: linkage)); |
| 1123 | // Declarations must keep 'private' MLIR visibility; only update for defs. |
| 1124 | if (!func.isDeclaration()) |
| 1125 | mlir::SymbolTable::setSymbolVisibility( |
| 1126 | symbol: func, vis: cgm.getMLIRVisibilityFromCIRLinkage(GLK: linkage)); |
| 1127 | } |
| 1128 | } |
| 1129 | |
| 1130 | static llvm::SmallVector<int64_t> indexesOfArrayAttr(mlir::ArrayAttr indexes) { |
| 1131 | llvm::SmallVector<int64_t> inds; |
| 1132 | for (mlir::Attribute i : indexes) { |
| 1133 | auto ind = mlir::cast<mlir::IntegerAttr>(Val&: i); |
| 1134 | inds.push_back(Elt: ind.getValue().getSExtValue()); |
| 1135 | } |
| 1136 | return inds; |
| 1137 | } |
| 1138 | |
| 1139 | static bool isViewOnGlobal(cir::GlobalOp glob, cir::GlobalViewAttr view) { |
| 1140 | return view.getSymbol().getValue() == glob.getSymName(); |
| 1141 | } |
| 1142 | |
| 1143 | static mlir::Attribute createNewGlobalView(CIRGenModule &cgm, |
| 1144 | cir::GlobalOp newGlob, |
| 1145 | cir::GlobalViewAttr attr, |
| 1146 | mlir::Type oldTy) { |
| 1147 | // If the attribute does not require indexes or it is not a global view on |
| 1148 | // the global we're replacing, keep the original attribute. |
| 1149 | if (!attr.getIndices() || !isViewOnGlobal(glob: newGlob, view: attr)) |
| 1150 | return attr; |
| 1151 | |
| 1152 | llvm::SmallVector<int64_t> oldInds = indexesOfArrayAttr(indexes: attr.getIndices()); |
| 1153 | llvm::SmallVector<int64_t> newInds; |
| 1154 | CIRGenBuilderTy &bld = cgm.getBuilder(); |
| 1155 | const cir::CIRDataLayout &layout = cgm.getDataLayout(); |
| 1156 | mlir::Type newTy = newGlob.getSymType(); |
| 1157 | |
| 1158 | uint64_t offset = |
| 1159 | bld.computeOffsetFromGlobalViewIndices(layout, ty: oldTy, indices: oldInds); |
| 1160 | if (!bld.computeGlobalViewIndicesFromFlatOffset(offset, ty: newTy, layout, |
| 1161 | indices&: newInds)) |
| 1162 | return cir::GlobalOffsetAttr::get(type: attr.getType(), symbol: attr.getSymbol(), |
| 1163 | offset: static_cast<int64_t>(offset)); |
| 1164 | |
| 1165 | cir::PointerType newPtrTy; |
| 1166 | |
| 1167 | if (isa<cir::RecordType>(Val: oldTy)) |
| 1168 | newPtrTy = cir::PointerType::get(pointee: newTy); |
| 1169 | else if (isa<cir::ArrayType>(Val: oldTy)) |
| 1170 | newPtrTy = cast<cir::PointerType>(Val: attr.getType()); |
| 1171 | |
| 1172 | if (newPtrTy) |
| 1173 | return bld.getGlobalViewAttr(type: newPtrTy, globalOp: newGlob, indices: newInds, |
| 1174 | addressPoint: attr.getAddressPoint()); |
| 1175 | |
| 1176 | // This may be unreachable in practice, but keep it as errorNYI while CIR |
| 1177 | // is still under development. |
| 1178 | cgm.errorNYI(feature: "Unhandled type in createNewGlobalView" ); |
| 1179 | return {}; |
| 1180 | } |
| 1181 | |
| 1182 | static mlir::Attribute getNewInitValue(CIRGenModule &cgm, cir::GlobalOp newGlob, |
| 1183 | mlir::Type oldTy, |
| 1184 | mlir::Attribute oldInit) { |
| 1185 | if (auto oldView = mlir::dyn_cast<cir::GlobalViewAttr>(Val&: oldInit)) |
| 1186 | return createNewGlobalView(cgm, newGlob, attr: oldView, oldTy); |
| 1187 | |
| 1188 | // A byte offset from a symbol doesn't depend on the symbol's type, so it |
| 1189 | // remains valid when the global is replaced. |
| 1190 | if (mlir::isa<cir::GlobalOffsetAttr>(Val: oldInit)) |
| 1191 | return oldInit; |
| 1192 | |
| 1193 | auto getNewInitElements = |
| 1194 | [&](mlir::ArrayAttr oldElements) -> mlir::ArrayAttr { |
| 1195 | llvm::SmallVector<mlir::Attribute> newElements; |
| 1196 | for (mlir::Attribute elt : oldElements) { |
| 1197 | if (auto view = mlir::dyn_cast<cir::GlobalViewAttr>(Val&: elt)) |
| 1198 | newElements.push_back(Elt: createNewGlobalView(cgm, newGlob, attr: view, oldTy)); |
| 1199 | else if (mlir::isa<cir::ConstArrayAttr, cir::ConstRecordAttr>(Val: elt)) |
| 1200 | newElements.push_back(Elt: getNewInitValue(cgm, newGlob, oldTy, oldInit: elt)); |
| 1201 | else |
| 1202 | newElements.push_back(Elt: elt); |
| 1203 | } |
| 1204 | return mlir::ArrayAttr::get(context: cgm.getBuilder().getContext(), value: newElements); |
| 1205 | }; |
| 1206 | |
| 1207 | if (auto oldArray = mlir::dyn_cast<cir::ConstArrayAttr>(Val&: oldInit)) { |
| 1208 | // ConstArrayAttr::verify guarantees the elements are either an ArrayAttr or |
| 1209 | // a StringAttr. A StringAttr is a string-literal initializer: raw 8-bit |
| 1210 | // character bytes with no nested global references, so there is nothing to |
| 1211 | // rewrite and it is returned unchanged. The ArrayAttr case recurses to |
| 1212 | // rewrite any nested global views. |
| 1213 | mlir::Attribute oldElts = oldArray.getElts(); |
| 1214 | if (mlir::isa<mlir::StringAttr>(Val: oldElts)) |
| 1215 | return oldInit; |
| 1216 | mlir::Attribute newElements = |
| 1217 | getNewInitElements(mlir::cast<mlir::ArrayAttr>(Val&: oldElts)); |
| 1218 | return cgm.getBuilder().getConstArray( |
| 1219 | attrs: newElements, arrayTy: mlir::cast<cir::ArrayType>(Val: oldArray.getType())); |
| 1220 | } |
| 1221 | if (auto oldRecord = mlir::dyn_cast<cir::ConstRecordAttr>(Val&: oldInit)) { |
| 1222 | mlir::ArrayAttr newMembers = getNewInitElements(oldRecord.getMembers()); |
| 1223 | auto recordTy = mlir::cast<cir::RecordType>(Val: oldRecord.getType()); |
| 1224 | return cgm.getBuilder().getConstRecordOrZeroAttr(arrayAttr: newMembers, recordTy); |
| 1225 | } |
| 1226 | |
| 1227 | // This may be unreachable in practice, but keep it as errorNYI while CIR |
| 1228 | // is still under development. |
| 1229 | cgm.errorNYI(feature: "Unhandled type in getNewInitValue" ); |
| 1230 | return {}; |
| 1231 | } |
| 1232 | |
| 1233 | // We want to replace a global value, but because of CIR's typed pointers, |
| 1234 | // we need to update the existing uses to reflect the new type, not just replace |
| 1235 | // them directly. |
| 1236 | void CIRGenModule::replaceGlobal(cir::GlobalOp oldGV, cir::GlobalOp newGV) { |
| 1237 | assert(oldGV.getSymName() == newGV.getSymName() && "symbol names must match" ); |
| 1238 | |
| 1239 | mlir::Type oldTy = oldGV.getSymType(); |
| 1240 | mlir::Type newTy = newGV.getSymType(); |
| 1241 | |
| 1242 | assert(!cir::MissingFeatures::addressSpace()); |
| 1243 | |
| 1244 | // If the type didn't change, why are we here? |
| 1245 | assert(oldTy != newTy && "expected type change in replaceGlobal" ); |
| 1246 | |
| 1247 | // Visit all uses and add handling to fix up the types. |
| 1248 | std::optional<mlir::SymbolTable::UseRange> oldSymUses = |
| 1249 | oldGV.getSymbolUses(from: getModule()); |
| 1250 | for (mlir::SymbolTable::SymbolUse use : *oldSymUses) { |
| 1251 | mlir::Operation *userOp = use.getUser(); |
| 1252 | assert( |
| 1253 | (mlir::isa<cir::GetGlobalOp, cir::GlobalOp, cir::ConstantOp>(userOp)) && |
| 1254 | "Unexpected user for global op" ); |
| 1255 | |
| 1256 | if (auto getGlobalOp = dyn_cast<cir::GetGlobalOp>(Val: use.getUser())) { |
| 1257 | mlir::Value useOpResultValue = getGlobalOp.getAddr(); |
| 1258 | useOpResultValue.setType(cir::PointerType::get(pointee: newTy)); |
| 1259 | |
| 1260 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 1261 | builder.setInsertionPointAfter(getGlobalOp); |
| 1262 | mlir::Type ptrTy = builder.getPointerTo(ty: oldTy); |
| 1263 | mlir::Value cast = |
| 1264 | builder.createBitcast(loc: getGlobalOp->getLoc(), src: useOpResultValue, newTy: ptrTy); |
| 1265 | useOpResultValue.replaceAllUsesExcept(newValue: cast, exceptedUser: cast.getDefiningOp()); |
| 1266 | } else if (auto glob = dyn_cast<cir::GlobalOp>(Val: userOp)) { |
| 1267 | if (auto init = glob.getInitialValue()) { |
| 1268 | mlir::Attribute nw = getNewInitValue(cgm&: *this, newGlob: newGV, oldTy, oldInit: init.value()); |
| 1269 | glob.setInitialValueAttr(nw); |
| 1270 | } |
| 1271 | } else if (auto c = dyn_cast<cir::ConstantOp>(Val: userOp)) { |
| 1272 | mlir::Attribute init = getNewInitValue(cgm&: *this, newGlob: newGV, oldTy, oldInit: c.getValue()); |
| 1273 | auto typedAttr = mlir::cast<mlir::TypedAttr>(Val&: init); |
| 1274 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 1275 | builder.setInsertionPointAfter(c); |
| 1276 | auto newUser = cir::ConstantOp::create(builder, location: c.getLoc(), value: typedAttr); |
| 1277 | c.replaceAllUsesWith(op: newUser.getOperation()); |
| 1278 | c.erase(); |
| 1279 | } |
| 1280 | } |
| 1281 | |
| 1282 | // If the old global is being tracked as the most-recently-created global, |
| 1283 | // update it so that subsequent globals are not inserted after a (now |
| 1284 | // erased) operation, which would leave them detached from the module. |
| 1285 | if (lastGlobalOp == oldGV) |
| 1286 | lastGlobalOp = newGV; |
| 1287 | if (getLangOpts().CUDA) |
| 1288 | getCUDARuntime().handleGlobalReplace(oldGV, newGV); |
| 1289 | eraseGlobalSymbol(op: oldGV); |
| 1290 | oldGV.erase(); |
| 1291 | } |
| 1292 | |
| 1293 | /// If the specified mangled name is not in the module, |
| 1294 | /// create and return an mlir GlobalOp with the specified type (TODO(cir): |
| 1295 | /// address space). |
| 1296 | /// |
| 1297 | /// TODO(cir): |
| 1298 | /// 1. If there is something in the module with the specified name, return |
| 1299 | /// it potentially bitcasted to the right type. |
| 1300 | /// |
| 1301 | /// 2. If \p d is non-null, it specifies a decl that correspond to this. This |
| 1302 | /// is used to set the attributes on the global when it is first created. |
| 1303 | /// |
| 1304 | /// 3. If \p isForDefinition is true, it is guaranteed that an actual global |
| 1305 | /// with type \p ty will be returned, not conversion of a variable with the same |
| 1306 | /// mangled name but some other type. |
| 1307 | cir::GlobalOp CIRGenModule::getOrCreateCIRGlobal( |
| 1308 | StringRef mangledName, mlir::Type ty, LangAS langAS, mlir::Location loc, |
| 1309 | const VarDecl *d, ForDefinition_t isForDefinition) { |
| 1310 | |
| 1311 | // Lookup the entry, lazily creating it if necessary. |
| 1312 | cir::GlobalOp entry; |
| 1313 | if (mlir::Operation *v = getGlobalValue(name: mangledName)) { |
| 1314 | if (!isa<cir::GlobalOp>(Val: v)) { |
| 1315 | errorNYI(loc, feature: "getOrCreateCIRGlobal: global with non-GlobalOp type" ); |
| 1316 | return nullptr; |
| 1317 | } |
| 1318 | entry = cast<cir::GlobalOp>(Val: v); |
| 1319 | } |
| 1320 | |
| 1321 | if (entry) { |
| 1322 | mlir::ptr::MemorySpaceAttrInterface entryCIRAS = entry.getAddrSpaceAttr(); |
| 1323 | |
| 1324 | // A weakref's target is only weak until something else names it. |
| 1325 | if (weakRefReferences.erase(Ptr: entry) && d && !d->hasAttr<WeakAttr>()) |
| 1326 | entry.setLinkage(cir::GlobalLinkageKind::ExternalLinkage); |
| 1327 | |
| 1328 | assert(!cir::MissingFeatures::setDLLStorageClass()); |
| 1329 | |
| 1330 | if (langOpts.OpenMP && !langOpts.OpenMPSimd && d) |
| 1331 | getOpenMPRuntime().registerTargetGlobalVariable(vd: d); |
| 1332 | |
| 1333 | if (entry.getSymType() == ty && |
| 1334 | cir::normalizeDefaultAddressSpace(addrSpace: entryCIRAS) == |
| 1335 | cir::normalizeDefaultAddressSpace(addrSpace: getGlobalCIRAddressSpace(as: langAS))) |
| 1336 | return entry; |
| 1337 | |
| 1338 | // If there are two attempts to define the same mangled name, issue an |
| 1339 | // error. |
| 1340 | // |
| 1341 | // TODO(cir): look at mlir::GlobalValue::isDeclaration for all aspects of |
| 1342 | // recognizing the global as a declaration, for now only check if |
| 1343 | // initializer is present. |
| 1344 | if (isForDefinition && !entry.isDeclaration()) { |
| 1345 | errorNYI(loc, feature: "getOrCreateCIRGlobal: global with conflicting type" ); |
| 1346 | } |
| 1347 | |
| 1348 | // Address space check removed because it is unnecessary because CIR records |
| 1349 | // address space info in types. |
| 1350 | |
| 1351 | // (If global is requested for a definition, we always need to create a new |
| 1352 | // global, not just return a bitcast.) |
| 1353 | if (!isForDefinition) |
| 1354 | return entry; |
| 1355 | } |
| 1356 | |
| 1357 | // Calculate constant storage flag before creating the global. This was moved |
| 1358 | // from after the global creation to ensure the constant flag is set correctly |
| 1359 | // at creation time, matching the logic used in emitCXXGlobalVarDeclInit. |
| 1360 | bool isConstant = false; |
| 1361 | if (d) { |
| 1362 | QualType declType = d->getType(); |
| 1363 | |
| 1364 | // Classic codegen doesn't try to exclude ctor or dtor here, but has a FIXME |
| 1365 | // to try to do a better job. So this bit of code does slightly more effort |
| 1366 | // to get a more accurate answer. We can try to exclude ctor, but only when |
| 1367 | // the type is complete, as otherwise we have to check for |
| 1368 | // fields(particularly whether they are mutable). |
| 1369 | bool excludeCtor = !declType->isIncompleteType(); |
| 1370 | bool needsDtor = |
| 1371 | d->needsDestruction(Ctx: astContext) == QualType::DK_cxx_destructor; |
| 1372 | |
| 1373 | isConstant = declType.isConstantStorage(Ctx: astContext, ExcludeCtor: excludeCtor, |
| 1374 | /*ExcludeDtor=*/!needsDtor); |
| 1375 | } |
| 1376 | |
| 1377 | mlir::ptr::MemorySpaceAttrInterface declCIRAS = |
| 1378 | getGlobalCIRAddressSpace(as: getGlobalVarAddressSpace(d)); |
| 1379 | |
| 1380 | // mlir::SymbolTable::Visibility::Public is the default, no need to explicitly |
| 1381 | // mark it as such. |
| 1382 | cir::GlobalOp gv = createGlobalOp(loc, name: mangledName, t: ty, isConstant, addrSpace: declCIRAS, |
| 1383 | /*insertPoint=*/entry.getOperation()); |
| 1384 | |
| 1385 | // If we already created a global with the same mangled name (but different |
| 1386 | // type) before, remove it from its parent. |
| 1387 | if (entry) |
| 1388 | replaceGlobal(oldGV: entry, newGV: gv); |
| 1389 | |
| 1390 | // This is the first use or definition of a mangled name. If there is a |
| 1391 | // deferred decl with this name, remember that we need to emit it at the end |
| 1392 | // of the file. |
| 1393 | auto ddi = deferredDecls.find(x: mangledName); |
| 1394 | if (ddi != deferredDecls.end()) { |
| 1395 | // Move the potentially referenced deferred decl to the DeferredDeclsToEmit |
| 1396 | // list, and remove it from DeferredDecls (since we don't need it anymore). |
| 1397 | addDeferredDeclToEmit(GD: ddi->second); |
| 1398 | deferredDecls.erase(position: ddi); |
| 1399 | } |
| 1400 | |
| 1401 | // Handle things which are present even on external declarations. |
| 1402 | if (d) { |
| 1403 | if (langOpts.OpenMP && !langOpts.OpenMPSimd) |
| 1404 | getOpenMPRuntime().registerTargetGlobalVariable(vd: d); |
| 1405 | |
| 1406 | gv.setAlignmentAttr(getSize(size: astContext.getDeclAlign(D: d))); |
| 1407 | |
| 1408 | setLinkageForGV(gv, nd: d); |
| 1409 | |
| 1410 | if (d->getTLSKind()) |
| 1411 | setTLSMode(op: gv, d: *d); |
| 1412 | |
| 1413 | setGVProperties(op: gv, d); |
| 1414 | |
| 1415 | // If required by the ABI, treat declarations of static data members with |
| 1416 | // inline initializers as definitions. |
| 1417 | if (astContext.isMSStaticDataMemberInlineDefinition(VD: d)) |
| 1418 | errorNYI(d->getSourceRange(), |
| 1419 | "getOrCreateCIRGlobal: MS static data member inline definition" ); |
| 1420 | |
| 1421 | // Emit section information for extern variables. |
| 1422 | if (d->hasExternalStorage()) { |
| 1423 | if (const SectionAttr *sa = d->getAttr<SectionAttr>()) |
| 1424 | gv.setSectionAttr(builder.getStringAttr(bytes: sa->getName())); |
| 1425 | } |
| 1426 | |
| 1427 | // Handle XCore specific ABI requirements. |
| 1428 | if (getTriple().getArch() == llvm::Triple::xcore) |
| 1429 | errorNYI(d->getSourceRange(), |
| 1430 | "getOrCreateCIRGlobal: XCore specific ABI requirements" ); |
| 1431 | |
| 1432 | // Check if we a have a const declaration with an initializer, we may be |
| 1433 | // able to emit it as available_externally to expose it's value to the |
| 1434 | // optimizer. |
| 1435 | if (getLangOpts().CPlusPlus && gv.isPublic() && |
| 1436 | d->getType().isConstQualified() && gv.isDeclaration() && |
| 1437 | !d->hasDefinition() && d->hasInit() && !d->hasAttr<DLLImportAttr>()) |
| 1438 | errorNYI( |
| 1439 | d->getSourceRange(), |
| 1440 | "getOrCreateCIRGlobal: external const declaration with initializer" ); |
| 1441 | } |
| 1442 | |
| 1443 | if (d && |
| 1444 | d->isThisDeclarationADefinition(astContext) == VarDecl::DeclarationOnly) { |
| 1445 | getTargetCIRGenInfo().setTargetAttributes(decl: d, global: gv, module&: *this); |
| 1446 | // TODO(cir): set target attributes |
| 1447 | // External HIP managed variables needed to be recorded for transformation |
| 1448 | // in both device and host compilations. |
| 1449 | if (getLangOpts().CUDA && d && d->hasAttr<HIPManagedAttr>() && |
| 1450 | d->hasExternalStorage()) |
| 1451 | getCUDARuntime().handleVarRegistration(vd: d, var: gv); |
| 1452 | } |
| 1453 | |
| 1454 | assert(!cir::MissingFeatures::addressSpace()); |
| 1455 | return gv; |
| 1456 | } |
| 1457 | |
| 1458 | mlir::Operation *CIRGenModule::getWeakRefReference(const ValueDecl *vd) { |
| 1459 | const auto *aa = vd->getAttr<AliasAttr>(); |
| 1460 | assert(aa && "No alias?" ); |
| 1461 | |
| 1462 | // See if there is already something with the target's name in the module. |
| 1463 | if (mlir::Operation *entry = getGlobalValue(name: aa->getAliasee())) { |
| 1464 | // Classic codegen is able to just treat whatever value it gets as the |
| 1465 | // 'right' type. This needs to introduce SOME sort of conversion it seems. |
| 1466 | // For now, reject any cases that don't match. |
| 1467 | if (const auto *fd = dyn_cast<FunctionDecl>(Val: vd)) { |
| 1468 | auto fn = dyn_cast<cir::FuncOp>(Val: entry); |
| 1469 | |
| 1470 | if (fn && fn.getFunctionType() == convertType(type: fd->getType())) |
| 1471 | return fn; |
| 1472 | } else if (auto gv = dyn_cast<cir::GlobalOp>(Val: entry)) { |
| 1473 | if (gv.getSymType() == getTypes().convertTypeForMem(vd->getType())) |
| 1474 | return gv; |
| 1475 | } |
| 1476 | |
| 1477 | errorNYI(vd->getSourceRange(), |
| 1478 | "weakref target already declared with a different type" ); |
| 1479 | return nullptr; |
| 1480 | } |
| 1481 | |
| 1482 | mlir::Operation *aliasee; |
| 1483 | if (const auto *fd = dyn_cast<FunctionDecl>(Val: vd)) { |
| 1484 | cir::FuncOp fn = |
| 1485 | getOrCreateCIRFunction(mangledName: aa->getAliasee(), funcType: convertType(type: fd->getType()), |
| 1486 | gd: GlobalDecl(fd), /*forVTable=*/false); |
| 1487 | fn.setLinkage(cir::GlobalLinkageKind::ExternalWeakLinkage); |
| 1488 | aliasee = fn; |
| 1489 | } else { |
| 1490 | cir::GlobalOp gv = getOrCreateCIRGlobal( |
| 1491 | mangledName: aa->getAliasee(), ty: getTypes().convertTypeForMem(vd->getType()), |
| 1492 | langAS: LangAS::Default, loc: getLoc(cRange: vd->getSourceRange()), /*d=*/nullptr, |
| 1493 | isForDefinition: NotForDefinition); |
| 1494 | gv.setLinkage(cir::GlobalLinkageKind::ExternalWeakLinkage); |
| 1495 | aliasee = gv; |
| 1496 | } |
| 1497 | |
| 1498 | weakRefReferences.insert(Ptr: aliasee); |
| 1499 | return aliasee; |
| 1500 | } |
| 1501 | |
| 1502 | cir::GlobalOp |
| 1503 | CIRGenModule::getOrCreateCIRGlobal(const VarDecl *d, mlir::Type ty, |
| 1504 | ForDefinition_t isForDefinition) { |
| 1505 | assert(d->hasGlobalStorage() && "Not a global variable" ); |
| 1506 | QualType astTy = d->getType(); |
| 1507 | if (!ty) |
| 1508 | ty = getTypes().convertTypeForMem(astTy); |
| 1509 | |
| 1510 | StringRef mangledName = getMangledName(gd: d); |
| 1511 | return getOrCreateCIRGlobal(mangledName, ty, langAS: getGlobalVarAddressSpace(d), |
| 1512 | loc: getLoc(cRange: d->getSourceRange()), d, isForDefinition); |
| 1513 | } |
| 1514 | |
| 1515 | /// Return the mlir::Value for the address of the given global variable. If |
| 1516 | /// \p ty is non-null and if the global doesn't exist, then it will be created |
| 1517 | /// with the specified type instead of whatever the normal requested type would |
| 1518 | /// be. If \p isForDefinition is true, it is guaranteed that an actual global |
| 1519 | /// with type \p ty will be returned, not conversion of a variable with the same |
| 1520 | /// mangled name but some other type. |
| 1521 | mlir::Value CIRGenModule::getAddrOfGlobalVar(const VarDecl *d, mlir::Type ty, |
| 1522 | ForDefinition_t isForDefinition) { |
| 1523 | assert(d->hasGlobalStorage() && "Not a global variable" ); |
| 1524 | QualType astTy = d->getType(); |
| 1525 | if (!ty) |
| 1526 | ty = getTypes().convertTypeForMem(astTy); |
| 1527 | |
| 1528 | bool tlsAccess = d->getTLSKind() != VarDecl::TLS_None; |
| 1529 | cir::GlobalOp g = getOrCreateCIRGlobal(d, ty, isForDefinition); |
| 1530 | mlir::Type ptrTy = builder.getPointerTo(ty: g.getSymType(), as: g.getAddrSpaceAttr()); |
| 1531 | mlir::Value addr = cir::GetGlobalOp::create( |
| 1532 | builder, location: getLoc(cRange: d->getSourceRange()), addr: ptrTy, name: g.getSymNameAttr(), |
| 1533 | tls: tlsAccess, |
| 1534 | /*static_local=*/g.getDynamicInitGuard().has_value() && |
| 1535 | d->isLocalVarDecl()); |
| 1536 | return castGlobalToDeclAddrSpace(addr, vd: *d); |
| 1537 | } |
| 1538 | |
| 1539 | mlir::Value CIRGenModule::castGlobalToDeclAddrSpace(mlir::Value addr, |
| 1540 | const VarDecl &vd) { |
| 1541 | // A global may live in a different address space than its declared type, |
| 1542 | // e.g. a CUDA __shared__ variable. Like classic CodeGen, cast once where |
| 1543 | // the address is formed so every user sees the declared type. |
| 1544 | auto ptrTy = mlir::cast<cir::PointerType>(Val: addr.getType()); |
| 1545 | mlir::ptr::MemorySpaceAttrInterface declAS = |
| 1546 | getTypes().getPointerAddressSpace(pointeeTy: vd.getType()); |
| 1547 | if (ptrTy.getAddrSpace() == declAS) |
| 1548 | return addr; |
| 1549 | return builder.createAddrSpaceCast( |
| 1550 | src: addr, newTy: builder.getPointerTo(ty: ptrTy.getPointee(), as: declAS)); |
| 1551 | } |
| 1552 | |
| 1553 | cir::GlobalViewAttr CIRGenModule::getAddrOfGlobalVarAttr(const VarDecl *d) { |
| 1554 | assert(d->hasGlobalStorage() && "Not a global variable" ); |
| 1555 | mlir::Type ty = getTypes().convertTypeForMem(d->getType()); |
| 1556 | |
| 1557 | cir::GlobalOp globalOp = getOrCreateCIRGlobal(d, ty, isForDefinition: NotForDefinition); |
| 1558 | cir::PointerType ptrTy = |
| 1559 | builder.getPointerTo(ty: globalOp.getSymType(), as: globalOp.getAddrSpaceAttr()); |
| 1560 | return builder.getGlobalViewAttr(type: ptrTy, globalOp); |
| 1561 | } |
| 1562 | |
| 1563 | void CIRGenModule::addUsedGlobal(cir::CIRGlobalValueInterface gv) { |
| 1564 | assert((mlir::isa<cir::FuncOp>(gv.getOperation()) || |
| 1565 | !gv.isDeclarationForLinker()) && |
| 1566 | "Only globals with definition can force usage." ); |
| 1567 | llvmUsed.emplace_back(args&: gv); |
| 1568 | } |
| 1569 | |
| 1570 | void CIRGenModule::addCompilerUsedGlobal(cir::CIRGlobalValueInterface gv) { |
| 1571 | assert(!gv.isDeclarationForLinker() && |
| 1572 | "Only globals with definition can force usage." ); |
| 1573 | llvmCompilerUsed.emplace_back(args&: gv); |
| 1574 | } |
| 1575 | |
| 1576 | void CIRGenModule::addUsedOrCompilerUsedGlobal( |
| 1577 | cir::CIRGlobalValueInterface gv) { |
| 1578 | assert((mlir::isa<cir::FuncOp>(gv.getOperation()) || |
| 1579 | !gv.isDeclarationForLinker()) && |
| 1580 | "Only globals with definition can force usage." ); |
| 1581 | if (getTriple().isOSBinFormatELF()) |
| 1582 | llvmCompilerUsed.emplace_back(args&: gv); |
| 1583 | else |
| 1584 | llvmUsed.emplace_back(args&: gv); |
| 1585 | } |
| 1586 | |
| 1587 | static void emitUsed(CIRGenModule &cgm, StringRef name, |
| 1588 | std::vector<cir::CIRGlobalValueInterface> &list) { |
| 1589 | if (list.empty()) |
| 1590 | return; |
| 1591 | |
| 1592 | CIRGenBuilderTy &builder = cgm.getBuilder(); |
| 1593 | mlir::Location loc = builder.getUnknownLoc(); |
| 1594 | cir::PointerType voidPtrTy = builder.getVoidPtrTy( |
| 1595 | as: cgm.getTypes().getPointerAddressSpace(pointeeTy: cgm.getASTContext().VoidTy)); |
| 1596 | |
| 1597 | llvm::SmallVector<mlir::Attribute> usedArray; |
| 1598 | usedArray.resize(N: list.size()); |
| 1599 | for (auto [i, op] : llvm::enumerate(First&: list)) { |
| 1600 | usedArray[i] = cir::GlobalViewAttr::get( |
| 1601 | type: voidPtrTy, symbol: mlir::FlatSymbolRefAttr::get(value: op.getNameAttr())); |
| 1602 | } |
| 1603 | |
| 1604 | cir::ArrayType arrayTy = cir::ArrayType::get(elementType: voidPtrTy, size: usedArray.size()); |
| 1605 | |
| 1606 | cir::ConstArrayAttr initAttr = cir::ConstArrayAttr::get( |
| 1607 | type: arrayTy, elts: mlir::ArrayAttr::get(context: &cgm.getMLIRContext(), value: usedArray)); |
| 1608 | |
| 1609 | cir::GlobalOp gv = cgm.createGlobalOp( |
| 1610 | loc, name, t: arrayTy, /*isConstant=*/false, |
| 1611 | addrSpace: cgm.getDataLayout().getDefaultGlobalsAddrSpace(ctx: &cgm.getMLIRContext())); |
| 1612 | gv.setLinkage(cir::GlobalLinkageKind::AppendingLinkage); |
| 1613 | gv.setInitialValueAttr(initAttr); |
| 1614 | gv.setSectionAttr(builder.getStringAttr(bytes: "llvm.metadata" )); |
| 1615 | } |
| 1616 | |
| 1617 | void CIRGenModule::emitLLVMUsed() { |
| 1618 | emitUsed(cgm&: *this, name: "llvm.used" , list&: llvmUsed); |
| 1619 | emitUsed(cgm&: *this, name: "llvm.compiler.used" , list&: llvmCompilerUsed); |
| 1620 | } |
| 1621 | |
| 1622 | void CIRGenModule::emitGlobalVarDefinition(const clang::VarDecl *vd, |
| 1623 | bool isTentative) { |
| 1624 | // OpenCL global variables of sampler type are translated to function calls, |
| 1625 | // therefore no need to be translated. |
| 1626 | if (getLangOpts().OpenCL && vd->getType()->isSamplerT()) |
| 1627 | return; |
| 1628 | |
| 1629 | if (getLangOpts().OpenMPIsTargetDevice) { |
| 1630 | errorNYI(vd->getSourceRange(), |
| 1631 | "emitGlobalVarDefinition: emit OpenMP global variable" ); |
| 1632 | return; |
| 1633 | } |
| 1634 | |
| 1635 | // Whether the definition of the variable is available externally. |
| 1636 | // If yes, we shouldn't emit the GloablCtor and GlobalDtor for the variable |
| 1637 | // since this is the job for its original source. |
| 1638 | bool isDefinitionAvailableExternally = |
| 1639 | astContext.GetGVALinkageForVariable(VD: vd) == GVA_AvailableExternally; |
| 1640 | |
| 1641 | // It is useless to emit the definition for an available_externally variable |
| 1642 | // which can't be marked as const. |
| 1643 | if (isDefinitionAvailableExternally && |
| 1644 | (!vd->hasConstantInitialization() || |
| 1645 | // TODO: Update this when we have interface to check constexpr |
| 1646 | // destructor. |
| 1647 | vd->needsDestruction(Ctx: astContext) || |
| 1648 | !vd->getType().isConstantStorage(Ctx: astContext, ExcludeCtor: true, ExcludeDtor: true))) |
| 1649 | return; |
| 1650 | |
| 1651 | mlir::Attribute init; |
| 1652 | bool needsGlobalCtor = false; |
| 1653 | bool needsGlobalDtor = |
| 1654 | !isDefinitionAvailableExternally && |
| 1655 | vd->needsDestruction(Ctx: astContext) == QualType::DK_cxx_destructor; |
| 1656 | const VarDecl *initDecl; |
| 1657 | const Expr *initExpr = vd->getAnyInitializer(D&: initDecl); |
| 1658 | |
| 1659 | std::optional<ConstantEmitter> emitter; |
| 1660 | |
| 1661 | // CUDA E.2.4.1 "__shared__ variables cannot have an initialization |
| 1662 | // as part of their declaration." Sema has already checked for |
| 1663 | // error cases, so we just need to set Init to PoisonValue. |
| 1664 | bool isCUDASharedVar = |
| 1665 | getLangOpts().CUDAIsDevice && vd->hasAttr<CUDASharedAttr>(); |
| 1666 | // Shadows of initialized device-side global variables are also left |
| 1667 | // undefined. |
| 1668 | // Managed Variables should be initialized on both host side and device side. |
| 1669 | bool isCUDAShadowVar = |
| 1670 | !getLangOpts().CUDAIsDevice && !vd->hasAttr<HIPManagedAttr>() && |
| 1671 | (vd->hasAttr<CUDAConstantAttr>() || vd->hasAttr<CUDADeviceAttr>() || |
| 1672 | vd->hasAttr<CUDASharedAttr>()); |
| 1673 | bool isCUDADeviceShadowVar = |
| 1674 | getLangOpts().CUDAIsDevice && !vd->hasAttr<HIPManagedAttr>() && |
| 1675 | (vd->getType()->isCUDADeviceBuiltinSurfaceType() || |
| 1676 | vd->getType()->isCUDADeviceBuiltinTextureType()); |
| 1677 | |
| 1678 | if (getLangOpts().CUDA && |
| 1679 | (isCUDASharedVar || isCUDAShadowVar || isCUDADeviceShadowVar)) { |
| 1680 | init = cir::UndefAttr::get(type: convertType(type: vd->getType())); |
| 1681 | } else if (vd->hasAttr<LoaderUninitializedAttr>()) { |
| 1682 | errorNYI(vd->getSourceRange(), |
| 1683 | "emitGlobalVarDefinition: loader uninitialized attribute" ); |
| 1684 | } else if (!initExpr) { |
| 1685 | // This is a tentative definition; tentative definitions are |
| 1686 | // implicitly initialized with { 0 }. |
| 1687 | // |
| 1688 | // Note that tentative definitions are only emitted at the end of |
| 1689 | // a translation unit, so they should never have incomplete |
| 1690 | // type. In addition, EmitTentativeDefinition makes sure that we |
| 1691 | // never attempt to emit a tentative definition if a real one |
| 1692 | // exists. A use may still exists, however, so we still may need |
| 1693 | // to do a RAUW. |
| 1694 | assert(!vd->getType()->isIncompleteType() && "Unexpected incomplete type" ); |
| 1695 | init = emitNullConstantAttr(t: vd->getType()); |
| 1696 | } else { |
| 1697 | emitter.emplace(args&: *this); |
| 1698 | mlir::Attribute initializer = emitter->tryEmitForInitializer(d: *initDecl); |
| 1699 | if (!initializer) { |
| 1700 | QualType qt = initExpr->getType(); |
| 1701 | if (vd->getType()->isReferenceType()) |
| 1702 | qt = vd->getType(); |
| 1703 | |
| 1704 | if (getLangOpts().CPlusPlus) { |
| 1705 | if (initDecl->hasFlexibleArrayInit(Ctx: astContext)) |
| 1706 | errorNYI(vd->getSourceRange(), |
| 1707 | "emitGlobalVarDefinition: flexible array initializer" ); |
| 1708 | init = emitNullConstantAttr(t: qt); |
| 1709 | if (!isDefinitionAvailableExternally) |
| 1710 | needsGlobalCtor = true; |
| 1711 | } else { |
| 1712 | errorNYI(vd->getSourceRange(), |
| 1713 | "emitGlobalVarDefinition: static initializer" ); |
| 1714 | } |
| 1715 | } else { |
| 1716 | init = initializer; |
| 1717 | // We don't need an initializer, so remove the entry for the delayed |
| 1718 | // initializer position (just in case this entry was delayed) if we |
| 1719 | // also don't need to register a destructor. |
| 1720 | assert(!cir::MissingFeatures::deferredCXXGlobalInit()); |
| 1721 | } |
| 1722 | } |
| 1723 | |
| 1724 | mlir::Type initType; |
| 1725 | if (mlir::isa<mlir::SymbolRefAttr>(Val: init)) { |
| 1726 | errorNYI( |
| 1727 | vd->getSourceRange(), |
| 1728 | "emitGlobalVarDefinition: global initializer is a symbol reference" ); |
| 1729 | return; |
| 1730 | } else { |
| 1731 | assert(mlir::isa<mlir::TypedAttr>(init) && "This should have a type" ); |
| 1732 | auto typedInitAttr = mlir::cast<mlir::TypedAttr>(Val&: init); |
| 1733 | initType = typedInitAttr.getType(); |
| 1734 | } |
| 1735 | assert(!mlir::isa<mlir::NoneType>(initType) && "Should have a type by now" ); |
| 1736 | |
| 1737 | cir::GlobalOp gv = |
| 1738 | getOrCreateCIRGlobal(d: vd, ty: initType, isForDefinition: ForDefinition_t(!isTentative)); |
| 1739 | // TODO(cir): Strip off pointer casts from Entry if we get them? |
| 1740 | |
| 1741 | if (!gv || gv.getSymType() != initType) { |
| 1742 | errorNYI(vd->getSourceRange(), |
| 1743 | "emitGlobalVarDefinition: global initializer with type mismatch" ); |
| 1744 | return; |
| 1745 | } |
| 1746 | |
| 1747 | assert(!cir::MissingFeatures::maybeHandleStaticInExternC()); |
| 1748 | |
| 1749 | if (vd->hasAttr<AnnotateAttr>()) |
| 1750 | addGlobalAnnotations(d: vd, gv); |
| 1751 | |
| 1752 | // Set CIR's linkage type as appropriate. |
| 1753 | cir::GlobalLinkageKind linkage = getCIRLinkageVarDefinition(vd); |
| 1754 | |
| 1755 | // CUDA B.2.1 "The __device__ qualifier declares a variable that resides on |
| 1756 | // the device. [...]" |
| 1757 | // CUDA B.2.2 "The __constant__ qualifier, optionally used together with |
| 1758 | // __device__, declares a variable that: [...] |
| 1759 | // Is accessible from all the threads within the grid and from the host |
| 1760 | // through the runtime library (cudaGetSymbolAddress() / cudaGetSymbolSize() |
| 1761 | // / cudaMemcpyToSymbol() / cudaMemcpyFromSymbol())." |
| 1762 | if (langOpts.CUDA) { |
| 1763 | if (langOpts.CUDAIsDevice) { |
| 1764 | // __shared__ variables is not marked as externally initialized, |
| 1765 | // because they must not be initialized. |
| 1766 | if (linkage != cir::GlobalLinkageKind::InternalLinkage && |
| 1767 | !vd->isConstexpr() && !vd->getType().isConstQualified() && |
| 1768 | (vd->hasAttr<CUDADeviceAttr>() || vd->hasAttr<CUDAConstantAttr>() || |
| 1769 | vd->getType()->isCUDADeviceBuiltinSurfaceType() || |
| 1770 | vd->getType()->isCUDADeviceBuiltinTextureType())) { |
| 1771 | gv->setAttr(name: cir::CUDAExternallyInitializedAttr::getMnemonic(), |
| 1772 | value: cir::CUDAExternallyInitializedAttr::get(ctx: &getMLIRContext())); |
| 1773 | } |
| 1774 | } else { |
| 1775 | // Adjust linkage of shadow variables in host compilation |
| 1776 | getCUDARuntime().internalizeDeviceSideVar(d: vd, linkage); |
| 1777 | } |
| 1778 | getCUDARuntime().handleVarRegistration(vd, var: gv); |
| 1779 | } |
| 1780 | |
| 1781 | // Set initializer and finalize emission |
| 1782 | CIRGenModule::setInitializer(op&: gv, value: init); |
| 1783 | if (emitter) |
| 1784 | emitter->finalize(gv); |
| 1785 | |
| 1786 | // If it is safe to mark the global 'constant', do so now. |
| 1787 | // Use the same logic as classic codegen EmitGlobalVarDefinition. |
| 1788 | gv.setConstant((vd->hasAttr<CUDAConstantAttr>() && langOpts.CUDAIsDevice) || |
| 1789 | (!needsGlobalCtor && !needsGlobalDtor && |
| 1790 | vd->getType().isConstantStorage(Ctx: astContext, |
| 1791 | /*ExcludeCtor=*/true, |
| 1792 | /*ExcludeDtor=*/true))); |
| 1793 | // If it is in a read-only section, mark it 'constant'. |
| 1794 | if (const SectionAttr *sa = vd->getAttr<SectionAttr>()) { |
| 1795 | const ASTContext::SectionInfo &si = astContext.SectionInfos[sa->getName()]; |
| 1796 | if ((si.SectionFlags & ASTContext::PSF_Write) == 0) |
| 1797 | gv.setConstant(true); |
| 1798 | } |
| 1799 | |
| 1800 | // The global may have been created without a declaration, e.g. as the target |
| 1801 | // of a weakref, so make sure the alignment comes from the definition. |
| 1802 | gv.setAlignmentAttr(getSize(size: astContext.getDeclAlign(D: vd))); |
| 1803 | |
| 1804 | // Set CIR linkage and DLL storage class. |
| 1805 | gv.setLinkage(linkage); |
| 1806 | // FIXME(cir): setLinkage should likely set MLIR's visibility automatically. |
| 1807 | gv.setVisibility(getMLIRVisibilityFromCIRLinkage(GLK: linkage)); |
| 1808 | assert(!cir::MissingFeatures::opGlobalDLLImportExport()); |
| 1809 | if (linkage == cir::GlobalLinkageKind::CommonLinkage) { |
| 1810 | // common vars aren't constant even if declared const. |
| 1811 | gv.setConstant(false); |
| 1812 | // Tentative definition of global variables may be initialized with |
| 1813 | // non-zero null pointers. In this case they should have weak linkage |
| 1814 | // since common linkage must have zero initializer and must not have |
| 1815 | // explicit section therefore cannot have non-zero initial value. |
| 1816 | std::optional<mlir::Attribute> initializer = gv.getInitialValue(); |
| 1817 | if (initializer && !getBuilder().isNullValue(attr: *initializer)) |
| 1818 | gv.setLinkage(cir::GlobalLinkageKind::WeakAnyLinkage); |
| 1819 | } |
| 1820 | |
| 1821 | setNonAliasAttributes(gd: vd, op: gv); |
| 1822 | |
| 1823 | if (vd->getTLSKind() && !vd->isStaticLocal()) |
| 1824 | setTLSMode(op: gv, d: *vd); |
| 1825 | |
| 1826 | maybeSetTrivialComdat(d: *vd, op: gv); |
| 1827 | |
| 1828 | // Emit the initializer function if necessary. |
| 1829 | if (needsGlobalCtor || needsGlobalDtor) { |
| 1830 | // We need to make sure that these are emitted 'in order' of definition, |
| 1831 | // rather than reference, so make sure this is now 'last', if it needs a |
| 1832 | // global ctor/dtor, so that initialization happens in the correct order. |
| 1833 | // Other uses of 'lastGlobalOp' just make sure we emit all our globals |
| 1834 | // before functions/etc, so it might have ended up out of order/referenced |
| 1835 | // earlier, so this is a 'move to the end' once we see definition thing |
| 1836 | // happening here. |
| 1837 | if (lastGlobalOp != gv.getOperation()) { |
| 1838 | if (lastGlobalOp) |
| 1839 | gv->moveAfter(existingOp: lastGlobalOp); |
| 1840 | else |
| 1841 | gv->moveBefore(block: getModule().getBody(), iterator: getModule().getBody()->begin()); |
| 1842 | |
| 1843 | lastGlobalOp = gv.getOperation(); |
| 1844 | } |
| 1845 | |
| 1846 | emitCXXGlobalVarDeclInitFunc(vd, addr: gv, performInit: needsGlobalCtor); |
| 1847 | } |
| 1848 | } |
| 1849 | |
| 1850 | bool CIRGenModule::shouldEmitFunction(GlobalDecl gd) { |
| 1851 | if (getFunctionLinkage(gd) != |
| 1852 | cir::GlobalLinkageKind::AvailableExternallyLinkage) |
| 1853 | return true; |
| 1854 | |
| 1855 | const auto *fd = cast<FunctionDecl>(Val: gd.getDecl()); |
| 1856 | // Inline builtins must be emitted; the body is redirected to a `.inline` |
| 1857 | // symbol in CIRGenFunction::generateCode. |
| 1858 | if (fd->isInlineBuiltinDeclaration()) |
| 1859 | return true; |
| 1860 | |
| 1861 | if (codeGenOpts.OptimizationLevel == 0 && !fd->hasAttr<AlwaysInlineAttr>()) |
| 1862 | return false; |
| 1863 | |
| 1864 | // We don't import function bodies from other named module units since that |
| 1865 | // behavior may break ABI compatibility of the current unit. |
| 1866 | if (const Module *m = fd->getOwningModule(); |
| 1867 | m && m->getTopLevelModule()->isNamedModule() && |
| 1868 | getASTContext().getCurrentNamedModule() != m->getTopLevelModule()) { |
| 1869 | // There are practices to mark template member function as always-inline |
| 1870 | // and mark the template as extern explicit instantiation but not give |
| 1871 | // the definition for member function. So we have to emit the function |
| 1872 | // from explicitly instantiation with always-inline. |
| 1873 | // |
| 1874 | // See https://github.com/llvm/llvm-project/issues/86893 for details. |
| 1875 | // |
| 1876 | // TODO: Maybe it is better to give it a warning if we call a non-inline |
| 1877 | // function from other module units which is marked as always-inline. |
| 1878 | if (!fd->isTemplateInstantiation() || !fd->hasAttr<AlwaysInlineAttr>()) { |
| 1879 | return false; |
| 1880 | } |
| 1881 | } |
| 1882 | |
| 1883 | if (fd->hasAttr<NoInlineAttr>()) |
| 1884 | return false; |
| 1885 | |
| 1886 | // PR9614 / glibc btowc workaround: an available_externally function whose |
| 1887 | // body just calls itself (via asm label or __builtin_* lowering on the |
| 1888 | // same name) is not a valid stand-in for the real implementation. Drop |
| 1889 | // it from the IR so the optimizer doesn't reason about its body. |
| 1890 | return !getCXXABI().getMangleContext().isTriviallyRecursive(FD: fd); |
| 1891 | } |
| 1892 | |
| 1893 | void CIRGenModule::emitGlobalDefinition(clang::GlobalDecl gd, |
| 1894 | mlir::Operation *op) { |
| 1895 | const auto *decl = cast<ValueDecl>(Val: gd.getDecl()); |
| 1896 | if (const auto *fd = dyn_cast<FunctionDecl>(Val: decl)) { |
| 1897 | if (!shouldEmitFunction(gd)) |
| 1898 | return; |
| 1899 | |
| 1900 | if (const auto *method = dyn_cast<CXXMethodDecl>(Val: decl)) { |
| 1901 | // Make sure to emit the definition(s) before we emit the thunks. This is |
| 1902 | // necessary for the generation of certain thunks. |
| 1903 | if (isa<CXXConstructorDecl>(Val: method) || isa<CXXDestructorDecl>(Val: method)) |
| 1904 | abi->emitCXXStructor(gd); |
| 1905 | else if (fd->isMultiVersion()) |
| 1906 | errorNYI(method->getSourceRange(), "multiversion functions" ); |
| 1907 | else |
| 1908 | emitGlobalFunctionDefinition(gd, op); |
| 1909 | |
| 1910 | if (method->isVirtual()) |
| 1911 | getVTables().emitThunks(gd); |
| 1912 | |
| 1913 | return; |
| 1914 | } |
| 1915 | |
| 1916 | if (fd->isMultiVersion()) |
| 1917 | errorNYI(fd->getSourceRange(), "multiversion functions" ); |
| 1918 | emitGlobalFunctionDefinition(gd, op); |
| 1919 | return; |
| 1920 | } |
| 1921 | |
| 1922 | if (const auto *vd = dyn_cast<VarDecl>(Val: decl)) |
| 1923 | return emitGlobalVarDefinition(vd, isTentative: !vd->hasDefinition()); |
| 1924 | |
| 1925 | llvm_unreachable("Invalid argument to CIRGenModule::emitGlobalDefinition" ); |
| 1926 | } |
| 1927 | |
| 1928 | mlir::Attribute |
| 1929 | CIRGenModule::getConstantArrayFromStringLiteral(const StringLiteral *e) { |
| 1930 | assert(!e->getType()->isPointerType() && "Strings are always arrays" ); |
| 1931 | |
| 1932 | // Don't emit it as the address of the string, emit the string data itself |
| 1933 | // as an inline array. |
| 1934 | if (e->getCharByteWidth() == 1) { |
| 1935 | SmallString<64> str(e->getString()); |
| 1936 | |
| 1937 | // Resize the string to the right size, which is indicated by its type. |
| 1938 | const ConstantArrayType *cat = |
| 1939 | astContext.getAsConstantArrayType(T: e->getType()); |
| 1940 | uint64_t finalSize = cat->getZExtSize(); |
| 1941 | str.resize(N: finalSize); |
| 1942 | |
| 1943 | mlir::Type eltTy = convertType(type: cat->getElementType()); |
| 1944 | return builder.getString(str, eltTy, size: finalSize, /*ensureNullTerm=*/false); |
| 1945 | } |
| 1946 | |
| 1947 | auto arrayTy = mlir::cast<cir::ArrayType>(Val: convertType(type: e->getType())); |
| 1948 | |
| 1949 | auto arrayEltTy = mlir::cast<cir::IntType>(Val: arrayTy.getElementType()); |
| 1950 | |
| 1951 | uint64_t arraySize = arrayTy.getSize(); |
| 1952 | unsigned literalSize = e->getLength(); |
| 1953 | assert(arraySize > literalSize && |
| 1954 | "wide string literal array size must have room for null terminator?" ); |
| 1955 | |
| 1956 | // Check if the string is all null bytes before building the vector. |
| 1957 | // In most non-zero cases, this will break out on the first element. |
| 1958 | bool isAllZero = true; |
| 1959 | for (unsigned i = 0; i < literalSize; ++i) { |
| 1960 | if (e->getCodeUnit(I: i) != 0) { |
| 1961 | isAllZero = false; |
| 1962 | break; |
| 1963 | } |
| 1964 | } |
| 1965 | |
| 1966 | if (isAllZero) |
| 1967 | return cir::ZeroAttr::get(type: arrayTy); |
| 1968 | |
| 1969 | // Otherwise emit a constant array holding the characters. |
| 1970 | SmallVector<mlir::Attribute> elements; |
| 1971 | elements.reserve(N: arraySize); |
| 1972 | for (unsigned i = 0; i < literalSize; ++i) |
| 1973 | elements.push_back(Elt: cir::IntAttr::get( |
| 1974 | type: arrayEltTy, value: llvm::APInt(arrayEltTy.getWidth(), e->getCodeUnit(I: i)))); |
| 1975 | |
| 1976 | auto elementsAttr = mlir::ArrayAttr::get(context: &getMLIRContext(), value: elements); |
| 1977 | return builder.getConstArray(attrs: elementsAttr, arrayTy); |
| 1978 | } |
| 1979 | |
| 1980 | bool CIRGenModule::supportsCOMDAT() const { |
| 1981 | return getTriple().supportsCOMDAT(); |
| 1982 | } |
| 1983 | |
| 1984 | void CIRGenModule::maybeSetTrivialComdat(const Decl &d, mlir::Operation *op) { |
| 1985 | if (!CodeGenUtils::shouldBeInCOMDAT(Ctx: getASTContext(), D: d)) |
| 1986 | return; |
| 1987 | if (auto globalOp = dyn_cast_or_null<cir::GlobalOp>(Val: op)) { |
| 1988 | globalOp.setSelfComdat(); |
| 1989 | } else { |
| 1990 | auto funcOp = cast<cir::FuncOp>(Val: op); |
| 1991 | funcOp.setSelfComdat(); |
| 1992 | } |
| 1993 | } |
| 1994 | |
| 1995 | void CIRGenModule::updateCompletedType(const TagDecl *td) { |
| 1996 | // Make sure that this type is translated. |
| 1997 | genTypes.updateCompletedType(td); |
| 1998 | } |
| 1999 | |
| 2000 | void CIRGenModule::addReplacement(StringRef name, mlir::Operation *op) { |
| 2001 | replacements[name] = op; |
| 2002 | } |
| 2003 | |
| 2004 | #ifndef NDEBUG |
| 2005 | static bool verifyPointerTypeArgs(llvm::ArrayRef<mlir::Operation *> users, |
| 2006 | cir::FuncOp newF) { |
| 2007 | for (mlir::Operation *user : users) { |
| 2008 | auto call = mlir::dyn_cast<cir::CallOp>(user); |
| 2009 | if (!call) |
| 2010 | continue; |
| 2011 | |
| 2012 | for (auto [argOp, fnArgType] : |
| 2013 | llvm::zip(call.getArgs(), newF.getFunctionType().getInputs())) { |
| 2014 | if (argOp.getType() != fnArgType) |
| 2015 | return false; |
| 2016 | } |
| 2017 | } |
| 2018 | |
| 2019 | return true; |
| 2020 | } |
| 2021 | #endif // NDEBUG |
| 2022 | |
| 2023 | void CIRGenModule::applyReplacements() { |
| 2024 | if (replacements.empty()) |
| 2025 | return; |
| 2026 | |
| 2027 | // Collect the users of the functions being replaced in one walk. Skip CIR |
| 2028 | // types. None can hold a symbol reference (see CIRTypes.td), and walking |
| 2029 | // them descends into every nested struct of every signature. |
| 2030 | llvm::DenseMap<mlir::StringAttr, llvm::SetVector<mlir::Operation *>> users; |
| 2031 | for (auto &i : replacements) |
| 2032 | users.try_emplace(Key: mlir::StringAttr::get(context: &getMLIRContext(), bytes: i.first)); |
| 2033 | |
| 2034 | mlir::ModuleOp module = getModule(); |
| 2035 | module->walk<mlir::WalkOrder::PreOrder>(callback: [&](mlir::Operation *op) { |
| 2036 | auto collect = [&](mlir::Attribute attr) { |
| 2037 | attr.walk<mlir::WalkOrder::PreOrder>( |
| 2038 | walkFns: [&](mlir::SymbolRefAttr ref) { |
| 2039 | auto it = users.find(Val: ref.getRootReference()); |
| 2040 | if (it != users.end()) |
| 2041 | it->second.insert(X: op); |
| 2042 | return mlir::WalkResult::skip(); |
| 2043 | }, |
| 2044 | walkFns: [&](mlir::Type type) { |
| 2045 | return mlir::isa<cir::CIRDialect>(Val: type.getDialect()) |
| 2046 | ? mlir::WalkResult::skip() |
| 2047 | : mlir::WalkResult::advance(); |
| 2048 | }); |
| 2049 | }; |
| 2050 | collect(op->getRawDictionaryAttrs()); |
| 2051 | op->getName().walkInherentAttrs( |
| 2052 | op, visitor: [&](llvm::StringRef, mlir::Attribute &attr) { collect(attr); }); |
| 2053 | // Like mlir::SymbolUserMap, don't look into nested symbol tables. |
| 2054 | if (op != module.getOperation() && |
| 2055 | op->hasTrait<mlir::OpTrait::SymbolTable>()) |
| 2056 | return mlir::WalkResult::skip(); |
| 2057 | return mlir::WalkResult::advance(); |
| 2058 | }); |
| 2059 | |
| 2060 | for (auto &i : replacements) { |
| 2061 | StringRef mangledName = i.first; |
| 2062 | mlir::Operation *replacement = i.second; |
| 2063 | mlir::Operation *entry = getGlobalValue(name: mangledName); |
| 2064 | if (!entry) |
| 2065 | continue; |
| 2066 | assert(isa<cir::FuncOp>(entry) && "expected function" ); |
| 2067 | auto oldF = cast<cir::FuncOp>(Val: entry); |
| 2068 | auto newF = dyn_cast<cir::FuncOp>(Val: replacement); |
| 2069 | if (!newF) { |
| 2070 | // In classic codegen, this can be a global alias, a bitcast, or a GEP. |
| 2071 | errorNYI(loc: replacement->getLoc(), feature: "replacement is not a function" ); |
| 2072 | continue; |
| 2073 | } |
| 2074 | |
| 2075 | auto oldIt = users.find(Val: oldF.getSymNameAttr()); |
| 2076 | assert(oldIt != users.end() && "replaced function was not collected" ); |
| 2077 | llvm::ArrayRef<mlir::Operation *> oldUsers = oldIt->second.getArrayRef(); |
| 2078 | assert(verifyPointerTypeArgs(oldUsers, newF) && |
| 2079 | "call argument types do not match replacement function" ); |
| 2080 | |
| 2081 | // Replace old with new, but keep the old order. |
| 2082 | for (mlir::Operation *user : oldUsers) |
| 2083 | (void)mlir::SymbolTable::replaceAllSymbolUses(oldSymbol: oldF, newSymbolName: newF.getSymNameAttr(), |
| 2084 | from: user); |
| 2085 | // These users now name newF. Keep them in case newF is replaced later. |
| 2086 | auto newIt = users.find(Val: newF.getSymNameAttr()); |
| 2087 | if (newIt != users.end()) |
| 2088 | newIt->second.set_union(oldIt->second); |
| 2089 | |
| 2090 | newF->moveBefore(existingOp: oldF); |
| 2091 | eraseGlobalSymbol(op: oldF); |
| 2092 | oldF->erase(); |
| 2093 | } |
| 2094 | } |
| 2095 | |
| 2096 | cir::GlobalOp CIRGenModule::createOrReplaceCXXRuntimeVariable( |
| 2097 | mlir::Location loc, StringRef name, mlir::Type ty, |
| 2098 | cir::GlobalLinkageKind linkage, clang::CharUnits alignment) { |
| 2099 | auto gv = mlir::dyn_cast_or_null<cir::GlobalOp>(Val: getGlobalValue(name)); |
| 2100 | |
| 2101 | if (gv) { |
| 2102 | // Check if the variable has the right type. |
| 2103 | if (gv.getSymType() == ty) |
| 2104 | return gv; |
| 2105 | |
| 2106 | // Because of C++ name mangling, the only way we can end up with an already |
| 2107 | // existing global with the same name is if it has been declared extern |
| 2108 | // "C". |
| 2109 | assert(gv.isDeclaration() && "Declaration has wrong type!" ); |
| 2110 | |
| 2111 | errorNYI(loc, feature: "createOrReplaceCXXRuntimeVariable: declaration exists with " |
| 2112 | "wrong type" ); |
| 2113 | return gv; |
| 2114 | } |
| 2115 | |
| 2116 | // Create a new variable. |
| 2117 | gv = createGlobalOp(loc, name, t: ty, /*isConstant=*/true); |
| 2118 | |
| 2119 | // Set up extra information and add to the module |
| 2120 | gv.setLinkageAttr( |
| 2121 | cir::GlobalLinkageKindAttr::get(context: &getMLIRContext(), value: linkage)); |
| 2122 | mlir::SymbolTable::setSymbolVisibility(symbol: gv, |
| 2123 | vis: CIRGenModule::getMLIRVisibility(op: gv)); |
| 2124 | |
| 2125 | if (supportsCOMDAT() && cir::isWeakForLinker(linkage) && |
| 2126 | !gv.hasAvailableExternallyLinkage()) { |
| 2127 | gv.setSelfComdat(); |
| 2128 | } |
| 2129 | |
| 2130 | gv.setAlignmentAttr(getSize(size: alignment)); |
| 2131 | setDSOLocal(static_cast<mlir::Operation *>(gv)); |
| 2132 | return gv; |
| 2133 | } |
| 2134 | |
| 2135 | cir::GlobalLinkageKind |
| 2136 | CIRGenModule::getCIRLinkageForDeclarator(const DeclaratorDecl *dd, |
| 2137 | GVALinkage linkage) { |
| 2138 | if (linkage == GVA_Internal) |
| 2139 | return cir::GlobalLinkageKind::InternalLinkage; |
| 2140 | |
| 2141 | if (dd->hasAttr<WeakAttr>()) |
| 2142 | return cir::GlobalLinkageKind::WeakAnyLinkage; |
| 2143 | |
| 2144 | if (const auto *fd = dd->getAsFunction()) |
| 2145 | if (fd->isMultiVersion() && linkage == GVA_AvailableExternally) |
| 2146 | return cir::GlobalLinkageKind::LinkOnceAnyLinkage; |
| 2147 | |
| 2148 | // We are guaranteed to have a strong definition somewhere else, |
| 2149 | // so we can use available_externally linkage. |
| 2150 | if (linkage == GVA_AvailableExternally) |
| 2151 | return cir::GlobalLinkageKind::AvailableExternallyLinkage; |
| 2152 | |
| 2153 | // Note that Apple's kernel linker doesn't support symbol |
| 2154 | // coalescing, so we need to avoid linkonce and weak linkages there. |
| 2155 | // Normally, this means we just map to internal, but for explicit |
| 2156 | // instantiations we'll map to external. |
| 2157 | |
| 2158 | // In C++, the compiler has to emit a definition in every translation unit |
| 2159 | // that references the function. We should use linkonce_odr because |
| 2160 | // a) if all references in this translation unit are optimized away, we |
| 2161 | // don't need to codegen it. b) if the function persists, it needs to be |
| 2162 | // merged with other definitions. c) C++ has the ODR, so we know the |
| 2163 | // definition is dependable. |
| 2164 | if (linkage == GVA_DiscardableODR) |
| 2165 | return !astContext.getLangOpts().AppleKext |
| 2166 | ? cir::GlobalLinkageKind::LinkOnceODRLinkage |
| 2167 | : cir::GlobalLinkageKind::InternalLinkage; |
| 2168 | |
| 2169 | // An explicit instantiation of a template has weak linkage, since |
| 2170 | // explicit instantiations can occur in multiple translation units |
| 2171 | // and must all be equivalent. However, we are not allowed to |
| 2172 | // throw away these explicit instantiations. |
| 2173 | // |
| 2174 | // CUDA/HIP: For -fno-gpu-rdc case, device code is limited to one TU, |
| 2175 | // so say that CUDA templates are either external (for kernels) or internal. |
| 2176 | // This lets llvm perform aggressive inter-procedural optimizations. For |
| 2177 | // -fgpu-rdc case, device function calls across multiple TU's are allowed, |
| 2178 | // therefore we need to follow the normal linkage paradigm. |
| 2179 | if (linkage == GVA_StrongODR) { |
| 2180 | if (getLangOpts().AppleKext) |
| 2181 | return cir::GlobalLinkageKind::ExternalLinkage; |
| 2182 | if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice && |
| 2183 | !getLangOpts().GPURelocatableDeviceCode) |
| 2184 | return dd->hasAttr<CUDAGlobalAttr>() |
| 2185 | ? cir::GlobalLinkageKind::ExternalLinkage |
| 2186 | : cir::GlobalLinkageKind::InternalLinkage; |
| 2187 | return cir::GlobalLinkageKind::WeakODRLinkage; |
| 2188 | } |
| 2189 | |
| 2190 | // C++ doesn't have tentative definitions and thus cannot have common |
| 2191 | // linkage. |
| 2192 | if (!getLangOpts().CPlusPlus && isa<VarDecl>(Val: dd) && |
| 2193 | !CodeGenUtils::isVarDeclStrongDefinition(Ctx: astContext, D: cast<VarDecl>(Val: dd), |
| 2194 | NoCommon: getCodeGenOpts().NoCommon)) |
| 2195 | return cir::GlobalLinkageKind::CommonLinkage; |
| 2196 | |
| 2197 | // selectany symbols are externally visible, so use weak instead of |
| 2198 | // linkonce. MSVC optimizes away references to const selectany globals, so |
| 2199 | // all definitions should be the same and ODR linkage should be used. |
| 2200 | // http://msdn.microsoft.com/en-us/library/5tkz6s71.aspx |
| 2201 | if (dd->hasAttr<SelectAnyAttr>()) |
| 2202 | return cir::GlobalLinkageKind::WeakODRLinkage; |
| 2203 | |
| 2204 | // Otherwise, we have strong external linkage. |
| 2205 | assert(linkage == GVA_StrongExternal); |
| 2206 | return cir::GlobalLinkageKind::ExternalLinkage; |
| 2207 | } |
| 2208 | |
| 2209 | /// This function is called when we implement a function with no prototype, e.g. |
| 2210 | /// "int foo() {}". If there are existing call uses of the old function in the |
| 2211 | /// module, this adjusts them to call the new function directly. |
| 2212 | /// |
| 2213 | /// This is not just a cleanup: the always_inline pass requires direct calls to |
| 2214 | /// functions to be able to inline them. If there is a bitcast in the way, it |
| 2215 | /// won't inline them. Instcombine normally deletes these calls, but it isn't |
| 2216 | /// run at -O0. |
| 2217 | void CIRGenModule::replaceUsesOfNonProtoTypeWithRealFunction( |
| 2218 | mlir::Operation *old, cir::FuncOp newFn) { |
| 2219 | // If we're redefining a global as a function, don't transform it. |
| 2220 | auto oldFn = mlir::dyn_cast<cir::FuncOp>(Val: old); |
| 2221 | if (!oldFn) |
| 2222 | return; |
| 2223 | |
| 2224 | // TODO(cir): this RAUW ignores the features below. |
| 2225 | assert(!cir::MissingFeatures::opFuncExceptions()); |
| 2226 | assert(!cir::MissingFeatures::opFuncParameterAttributes()); |
| 2227 | assert(!cir::MissingFeatures::opFuncOperandBundles()); |
| 2228 | // The caller sets function attributes from the replacement's declaration. |
| 2229 | // Copying the old attributes could retain incompatible assumptions from a |
| 2230 | // different declaration that happens to share the same symbol name. |
| 2231 | |
| 2232 | // Mark new function as originated from a no-proto declaration. |
| 2233 | newFn.setNoProto(oldFn.getNoProto()); |
| 2234 | |
| 2235 | // Iterate through all calls of the no-proto function. |
| 2236 | std::optional<mlir::SymbolTable::UseRange> symUses = |
| 2237 | oldFn.getSymbolUses(from: oldFn->getParentOp()); |
| 2238 | |
| 2239 | if (!symUses) |
| 2240 | return; |
| 2241 | |
| 2242 | for (const mlir::SymbolTable::SymbolUse &use : symUses.value()) { |
| 2243 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 2244 | |
| 2245 | if (auto noProtoCallOp = mlir::dyn_cast<cir::CallOp>(Val: use.getUser())) { |
| 2246 | builder.setInsertionPoint(noProtoCallOp); |
| 2247 | |
| 2248 | // Patch call type with the real function type. |
| 2249 | cir::FuncType newFnType = newFn.getFunctionType(); |
| 2250 | mlir::OperandRange callOperands = noProtoCallOp.getOperands(); |
| 2251 | bool returnTypeMatches = |
| 2252 | newFnType.hasVoidReturn() |
| 2253 | ? noProtoCallOp.getNumResults() == 0 |
| 2254 | : noProtoCallOp.getNumResults() == 1 && |
| 2255 | noProtoCallOp.getResultTypes().front() == |
| 2256 | newFnType.getReturnType(); |
| 2257 | bool typesMatch = !newFn.getNoProto() && returnTypeMatches && |
| 2258 | callOperands.size() == newFnType.getNumInputs(); |
| 2259 | for (unsigned i = 0, e = newFnType.getNumInputs(); typesMatch && i != e; |
| 2260 | ++i) { |
| 2261 | if (callOperands[i].getType() != newFnType.getInput(i)) |
| 2262 | typesMatch = false; |
| 2263 | } |
| 2264 | |
| 2265 | cir::CallOp realCallOp; |
| 2266 | if (typesMatch) { |
| 2267 | // Patch call type with the real function type. |
| 2268 | realCallOp = |
| 2269 | builder.createCallOp(loc: noProtoCallOp.getLoc(), callee: newFn, operands: callOperands); |
| 2270 | } else { |
| 2271 | // Build an indirect call whose function-pointer signature matches |
| 2272 | // the existing call site. A prototyped declaration keeps its own |
| 2273 | // type, ellipsis included, so arguments passed through the ellipsis |
| 2274 | // stay variadic. A direct call to an unprototyped declaration, such |
| 2275 | // as a library call CIRGen emitted by name, keeps its own operand |
| 2276 | // types and stays non-variadic. |
| 2277 | cir::FuncType origFnType = oldFn.getFunctionType(); |
| 2278 | cir::FuncType callFnType = |
| 2279 | oldFn.getNoProto() |
| 2280 | ? cir::FuncType::get(inputs: llvm::to_vector(Range: callOperands.getTypes()), |
| 2281 | returnType: origFnType.getReturnType(), |
| 2282 | /*isVarArg=*/false) |
| 2283 | : origFnType; |
| 2284 | mlir::Value addr = cir::GetGlobalOp::create( |
| 2285 | builder, location: noProtoCallOp.getLoc(), addr: cir::PointerType::get(pointee: newFnType), |
| 2286 | name: newFn.getSymName()); |
| 2287 | mlir::Value casted = |
| 2288 | builder.createBitcast(src: addr, newTy: cir::PointerType::get(pointee: callFnType)); |
| 2289 | realCallOp = builder.createIndirectCallOp( |
| 2290 | loc: noProtoCallOp.getLoc(), indirectTarget: casted, funcType: callFnType, operands: callOperands); |
| 2291 | } |
| 2292 | // The argument and result types are unchanged, so preserve their |
| 2293 | // attributes and the call semantics. The new callee is already set. |
| 2294 | for (mlir::NamedAttribute attr : noProtoCallOp->getAttrs()) |
| 2295 | if (attr.getName() != cir::CIRDialect::getCalleeAttrName()) |
| 2296 | realCallOp->setAttr(name: attr.getName(), value: attr.getValue()); |
| 2297 | |
| 2298 | // Replace old no proto call with fixed call. |
| 2299 | noProtoCallOp.replaceAllUsesWith(values&: realCallOp); |
| 2300 | noProtoCallOp.erase(); |
| 2301 | } else if (auto getGlobalOp = |
| 2302 | mlir::dyn_cast<cir::GetGlobalOp>(Val: use.getUser())) { |
| 2303 | // The GetGlobal was emitted with the no-proto FuncType. Uses of this |
| 2304 | // operation (cir.store, cir.cast) were built for that pointer type. When |
| 2305 | // we re-type the result to the real FuncType, we need to add a bit the |
| 2306 | // old pointer type so those uses are still valid. This can lead to |
| 2307 | // some redundant bitcast chains, but those will be cleaned up by the |
| 2308 | // canonicalizer. |
| 2309 | mlir::Value res = getGlobalOp.getAddr(); |
| 2310 | const mlir::Type oldResTy = res.getType(); |
| 2311 | const auto newPtrTy = cir::PointerType::get( |
| 2312 | pointee: newFn.getFunctionType(), |
| 2313 | addrSpace: mlir::cast<cir::PointerType>(Val: oldResTy).getAddrSpace()); |
| 2314 | if (oldResTy != newPtrTy) { |
| 2315 | res.setType(newPtrTy); |
| 2316 | builder.setInsertionPointAfter(getGlobalOp.getOperation()); |
| 2317 | mlir::Value castRes = |
| 2318 | cir::CastOp::create(builder, location: getGlobalOp.getLoc(), result: oldResTy, |
| 2319 | kind: cir::CastKind::bitcast, src: res); |
| 2320 | res.replaceAllUsesExcept(newValue: castRes, exceptedUser: castRes.getDefiningOp()); |
| 2321 | } |
| 2322 | } else if (mlir::isa<cir::GlobalOp>(Val: use.getUser())) { |
| 2323 | // Function addresses in global initializers use GlobalViewAttrs typed to |
| 2324 | // the initializer context (e.g. struct field type), not the FuncOp type, |
| 2325 | // so no update is required when the no-proto FuncOp is replaced. |
| 2326 | } else { |
| 2327 | llvm_unreachable( |
| 2328 | "replaceUsesOfNonProtoTypeWithRealFunction: unexpected use type" ); |
| 2329 | } |
| 2330 | } |
| 2331 | } |
| 2332 | |
| 2333 | cir::GlobalLinkageKind |
| 2334 | CIRGenModule::getCIRLinkageVarDefinition(const VarDecl *vd) { |
| 2335 | GVALinkage linkage = astContext.GetGVALinkageForVariable(VD: vd); |
| 2336 | return getCIRLinkageForDeclarator(dd: vd, linkage); |
| 2337 | } |
| 2338 | |
| 2339 | cir::GlobalLinkageKind CIRGenModule::getFunctionLinkage(GlobalDecl gd) { |
| 2340 | const auto *d = cast<FunctionDecl>(Val: gd.getDecl()); |
| 2341 | |
| 2342 | GVALinkage linkage = astContext.GetGVALinkageForFunction(FD: d); |
| 2343 | |
| 2344 | if (const auto *dtor = dyn_cast<CXXDestructorDecl>(Val: d)) |
| 2345 | return getCXXABI().getCXXDestructorLinkage(linkage, dtor, dt: gd.getDtorType()); |
| 2346 | |
| 2347 | return getCIRLinkageForDeclarator(dd: d, linkage); |
| 2348 | } |
| 2349 | |
| 2350 | static cir::GlobalOp |
| 2351 | generateStringLiteral(mlir::Location loc, mlir::TypedAttr c, |
| 2352 | cir::GlobalLinkageKind lt, CIRGenModule &cgm, |
| 2353 | StringRef globalName, CharUnits alignment) { |
| 2354 | LangAS as = CodeGenUtils::getGlobalConstantAddressSpace(LangOpts: cgm.getLangOpts(), |
| 2355 | Target: cgm.getTarget()); |
| 2356 | mlir::ptr::MemorySpaceAttrInterface addrSpace = |
| 2357 | cgm.getGlobalCIRAddressSpace(as); |
| 2358 | |
| 2359 | // Create a global variable for this string |
| 2360 | // FIXME(cir): check for insertion point in module level. |
| 2361 | cir::GlobalOp gv = |
| 2362 | cgm.createGlobalOp(loc, name: globalName, t: c.getType(), |
| 2363 | isConstant: !cgm.getLangOpts().WritableStrings, addrSpace); |
| 2364 | |
| 2365 | // Set up extra information and add to the module |
| 2366 | gv.setAlignmentAttr(cgm.getSize(size: alignment)); |
| 2367 | gv.setLinkageAttr( |
| 2368 | cir::GlobalLinkageKindAttr::get(context: cgm.getBuilder().getContext(), value: lt)); |
| 2369 | assert(!cir::MissingFeatures::opGlobalThreadLocal()); |
| 2370 | gv.setUnnamedAddr(true); |
| 2371 | CIRGenModule::setInitializer(op&: gv, value: c); |
| 2372 | if (gv.isWeakForLinker()) { |
| 2373 | assert(cgm.supportsCOMDAT() && "Only COFF uses weak string literals" ); |
| 2374 | gv.setSelfComdat(); |
| 2375 | } |
| 2376 | cgm.setDSOLocal(static_cast<mlir::Operation *>(gv)); |
| 2377 | return gv; |
| 2378 | } |
| 2379 | |
| 2380 | // LLVM IR automatically uniques names when new llvm::GlobalVariables are |
| 2381 | // created. This is handy, for example, when creating globals for string |
| 2382 | // literals. Since we don't do that when creating cir::GlobalOp's, we need |
| 2383 | // a mechanism to generate a unique name in advance. |
| 2384 | // |
| 2385 | // For now, this mechanism is only used in cases where we know that the |
| 2386 | // name is compiler-generated, so we don't use the MLIR symbol table for |
| 2387 | // the lookup. |
| 2388 | std::string CIRGenModule::getUniqueGlobalName(const std::string &baseName) { |
| 2389 | // If this is the first time we've generated a name for this basename, use |
| 2390 | // it as is and start a counter for this base name. |
| 2391 | auto it = cgGlobalNames.find(Key: baseName); |
| 2392 | if (it == cgGlobalNames.end()) { |
| 2393 | cgGlobalNames[baseName] = 1; |
| 2394 | return baseName; |
| 2395 | } |
| 2396 | |
| 2397 | std::string result = |
| 2398 | baseName + "." + std::to_string(val: cgGlobalNames[baseName]++); |
| 2399 | // There should not be any symbol with this name in the module. |
| 2400 | assert(!getGlobalValue(result)); |
| 2401 | return result; |
| 2402 | } |
| 2403 | |
| 2404 | /// Return a pointer to a constant array for the given string literal. |
| 2405 | cir::GlobalOp CIRGenModule::getGlobalForStringLiteral(const StringLiteral *s, |
| 2406 | StringRef name) { |
| 2407 | CharUnits alignment = |
| 2408 | astContext.getAlignOfGlobalVarInChars(T: s->getType(), /*VD=*/nullptr); |
| 2409 | |
| 2410 | mlir::Attribute c = getConstantArrayFromStringLiteral(e: s); |
| 2411 | |
| 2412 | cir::GlobalOp gv; |
| 2413 | if (!getLangOpts().WritableStrings && constantStringMap.count(Val: c)) { |
| 2414 | gv = constantStringMap[c]; |
| 2415 | // The bigger alignment always wins. |
| 2416 | if (!gv.getAlignment() || |
| 2417 | uint64_t(alignment.getQuantity()) > *gv.getAlignment()) |
| 2418 | gv.setAlignmentAttr(getSize(size: alignment)); |
| 2419 | } else { |
| 2420 | // Mangle the string literal if that's how the ABI merges duplicate strings. |
| 2421 | // Don't do it if they are writable, since we don't want writes in one TU to |
| 2422 | // affect strings in another. |
| 2423 | if (getCXXABI().getMangleContext().shouldMangleStringLiteral(SL: s) && |
| 2424 | !getLangOpts().WritableStrings) { |
| 2425 | errorNYI(s->getSourceRange(), |
| 2426 | "getGlobalForStringLiteral: mangle string literals" ); |
| 2427 | } |
| 2428 | |
| 2429 | // Unlike LLVM IR, CIR doesn't automatically unique names for globals, so |
| 2430 | // we need to do that explicitly. |
| 2431 | std::string uniqueName = getUniqueGlobalName(baseName: name.str()); |
| 2432 | // Synthetic string literals (e.g., from SourceLocExpr) may not have valid |
| 2433 | // source locations. Use unknown location in those cases. |
| 2434 | mlir::Location loc = s->getBeginLoc().isValid() |
| 2435 | ? getLoc(cRange: s->getSourceRange()) |
| 2436 | : builder.getUnknownLoc(); |
| 2437 | auto typedC = llvm::cast<mlir::TypedAttr>(Val&: c); |
| 2438 | gv = generateStringLiteral(loc, c: typedC, |
| 2439 | lt: cir::GlobalLinkageKind::PrivateLinkage, cgm&: *this, |
| 2440 | globalName: uniqueName, alignment); |
| 2441 | setDSOLocal(static_cast<mlir::Operation *>(gv)); |
| 2442 | constantStringMap[c] = gv; |
| 2443 | |
| 2444 | assert(!cir::MissingFeatures::sanitizers()); |
| 2445 | } |
| 2446 | return gv; |
| 2447 | } |
| 2448 | |
| 2449 | /// Return a pointer to a constant array for the given string literal. |
| 2450 | cir::GlobalViewAttr |
| 2451 | CIRGenModule::getAddrOfConstantStringFromLiteral(const StringLiteral *s, |
| 2452 | StringRef name) { |
| 2453 | cir::GlobalOp gv = getGlobalForStringLiteral(s, name); |
| 2454 | auto arrayTy = mlir::dyn_cast<cir::ArrayType>(Val: gv.getSymType()); |
| 2455 | assert(arrayTy && "String literal must be array" ); |
| 2456 | cir::PointerType ptrTy = getBuilder().getPointerTo( |
| 2457 | ty: arrayTy.getElementType(), |
| 2458 | as: getTypes().getPointerAddressSpace(pointeeTy: s->getType())); |
| 2459 | |
| 2460 | return builder.getGlobalViewAttr(type: ptrTy, globalOp: gv); |
| 2461 | } |
| 2462 | |
| 2463 | cir::GlobalOp CIRGenModule::createGlobalString(mlir::Location loc, |
| 2464 | StringRef str, StringRef name) { |
| 2465 | mlir::Attribute init = builder.getString(str, eltTy: convertType(type: astContext.CharTy), |
| 2466 | /*size=*/std::nullopt); |
| 2467 | |
| 2468 | // Unlike LLVM IR, CIR doesn't automatically unique names for globals, so |
| 2469 | // we need to do that explicitly. |
| 2470 | return generateStringLiteral(loc, c: mlir::cast<mlir::TypedAttr>(Val&: init), |
| 2471 | lt: cir::GlobalLinkageKind::PrivateLinkage, cgm&: *this, |
| 2472 | globalName: getUniqueGlobalName(baseName: name.str()), |
| 2473 | alignment: CharUnits::One()); |
| 2474 | } |
| 2475 | |
| 2476 | mlir::ptr::MemorySpaceAttrInterface |
| 2477 | CIRGenModule::getGlobalCIRAddressSpace(LangAS as) { |
| 2478 | return getTargetCIRGenInfo().getGlobalCIRAddressSpace(as); |
| 2479 | } |
| 2480 | |
| 2481 | // TODO(cir): this could be a common AST helper for both CIR and LLVM codegen. |
| 2482 | LangAS CIRGenModule::getLangTempAllocaAddressSpace() const { |
| 2483 | if (getLangOpts().OpenCL) |
| 2484 | return LangAS::opencl_private; |
| 2485 | |
| 2486 | // For temporaries inside functions, CUDA treats them as normal variables. |
| 2487 | // LangAS::cuda_device, on the other hand, is reserved for those variables |
| 2488 | // explicitly marked with __device__. |
| 2489 | if (getLangOpts().CUDAIsDevice) |
| 2490 | return LangAS::Default; |
| 2491 | |
| 2492 | if (getLangOpts().OpenMP && getLangOpts().OpenMPIsTargetDevice) |
| 2493 | assert(!cir::MissingFeatures::openMP()); |
| 2494 | |
| 2495 | if (getLangOpts().SYCLIsDevice) |
| 2496 | return LangAS::Default; |
| 2497 | |
| 2498 | return LangAS::Default; |
| 2499 | } |
| 2500 | |
| 2501 | void CIRGenModule::emitExplicitCastExprType(const ExplicitCastExpr *e, |
| 2502 | CIRGenFunction *cgf) { |
| 2503 | if (cgf && e->getType()->isVariablyModifiedType()) |
| 2504 | cgf->emitVariablyModifiedType(ty: e->getType()); |
| 2505 | |
| 2506 | assert(!cir::MissingFeatures::generateDebugInfo() && |
| 2507 | "emitExplicitCastExprType" ); |
| 2508 | } |
| 2509 | |
| 2510 | mlir::TypedAttr CIRGenModule::emitNullMemberAttr(QualType destTy, |
| 2511 | const MemberPointerType *mpt) { |
| 2512 | if (mpt->isMemberFunctionPointerType()) { |
| 2513 | auto ty = mlir::cast<cir::MethodType>(Val: convertType(type: destTy)); |
| 2514 | return builder.getNullMethodAttr(ty); |
| 2515 | } |
| 2516 | |
| 2517 | auto ty = mlir::cast<cir::DataMemberType>(Val: convertType(type: destTy)); |
| 2518 | return builder.getNullDataMemberAttr(ty); |
| 2519 | } |
| 2520 | |
| 2521 | mlir::Value CIRGenModule::emitMemberPointerConstant(const UnaryOperator *e) { |
| 2522 | assert(!cir::MissingFeatures::cxxABI()); |
| 2523 | |
| 2524 | mlir::Location loc = getLoc(cRange: e->getSourceRange()); |
| 2525 | |
| 2526 | const ValueDecl *decl = cast<DeclRefExpr>(Val: e->getSubExpr())->getDecl(); |
| 2527 | |
| 2528 | // A member function pointer. |
| 2529 | if (const auto *methodDecl = dyn_cast<CXXMethodDecl>(Val: decl)) { |
| 2530 | auto ty = mlir::cast<cir::MethodType>(Val: convertType(type: e->getType())); |
| 2531 | if (methodDecl->isVirtual()) |
| 2532 | return cir::ConstantOp::create( |
| 2533 | builder, location: loc, value: getCXXABI().buildVirtualMethodAttr(methodTy: ty, md: methodDecl)); |
| 2534 | |
| 2535 | const CIRGenFunctionInfo &fi = |
| 2536 | getTypes().arrangeCXXMethodDeclaration(md: methodDecl); |
| 2537 | cir::FuncType funcTy = getTypes().getFunctionType(info: fi); |
| 2538 | cir::FuncOp methodFuncOp = getAddrOfFunction(gd: methodDecl, funcType: funcTy); |
| 2539 | return cir::ConstantOp::create(builder, location: loc, |
| 2540 | value: builder.getMethodAttr(ty, methodFuncOp)); |
| 2541 | } |
| 2542 | |
| 2543 | // Otherwise, a member data pointer. |
| 2544 | auto ty = mlir::cast<cir::DataMemberType>(Val: convertType(type: e->getType())); |
| 2545 | const auto *mpt = e->getType()->castAs<MemberPointerType>(); |
| 2546 | const auto *destClass = mpt->getMostRecentCXXRecordDecl(); |
| 2547 | |
| 2548 | // Empty [[no_unique_address]] fields have no CIR field index; represent the |
| 2549 | // pointer-to-data-member by its concrete byte offset within the class. |
| 2550 | if (const auto *fieldDecl = dyn_cast<FieldDecl>(Val: decl); |
| 2551 | fieldDecl && isEmptyFieldForMemberPointer(field: fieldDecl)) { |
| 2552 | // This function should ONLY be accessed in reference to itself, I don't see |
| 2553 | // any cases/couldn't find any cases where anything else could get here, and |
| 2554 | // classic-codegen does the same. |
| 2555 | assert(fieldDecl->getParent() == destClass && |
| 2556 | "scalar member pointer should be relative to the declaring class" ); |
| 2557 | uint64_t offset = |
| 2558 | astContext.toCharUnitsFromBits(BitSize: astContext.getFieldOffset(FD: fieldDecl)) |
| 2559 | .getQuantity(); |
| 2560 | return cir::ConstantOp::create(builder, location: loc, |
| 2561 | value: cir::DataMemberOffsetAttr::get(type: ty, offset)); |
| 2562 | } |
| 2563 | |
| 2564 | std::optional<llvm::SmallVector<int32_t>> path = |
| 2565 | buildMemberPath(destClass, decl); |
| 2566 | if (!path) |
| 2567 | return {}; |
| 2568 | return cir::ConstantOp::create(builder, location: loc, |
| 2569 | value: builder.getDataMemberAttr(ty, path: *path)); |
| 2570 | } |
| 2571 | |
| 2572 | std::optional<llvm::SmallVector<int32_t>> |
| 2573 | CIRGenModule::buildMemberPath(const CXXRecordDecl *destClass, |
| 2574 | const ValueDecl *decl) { |
| 2575 | llvm::SmallVector<int32_t> path; |
| 2576 | |
| 2577 | // Members of an anonymous struct/union have an IndirectFieldDecl, which |
| 2578 | // contains the whole chain of how to get to it, so to get the 'path', we dig |
| 2579 | // through those rather than searching. |
| 2580 | if (const auto *indirectField = dyn_cast<IndirectFieldDecl>(Val: decl)) { |
| 2581 | const CXXRecordDecl *currentClass = destClass; |
| 2582 | for (const NamedDecl *nd : indirectField->chain()) { |
| 2583 | const auto *field = cast<FieldDecl>(Val: nd); |
| 2584 | if (!findFieldMemberPath(currentClass, field, path)) |
| 2585 | return std::nullopt; |
| 2586 | currentClass = field->getType()->getAsCXXRecordDecl(); |
| 2587 | } |
| 2588 | return path; |
| 2589 | } |
| 2590 | |
| 2591 | if (!findFieldMemberPath(currentClass: destClass, field: cast<FieldDecl>(Val: decl), path)) |
| 2592 | return std::nullopt; |
| 2593 | return path; |
| 2594 | } |
| 2595 | |
| 2596 | bool CIRGenModule::findFieldMemberPath(const CXXRecordDecl *currentClass, |
| 2597 | const FieldDecl *field, |
| 2598 | llvm::SmallVectorImpl<int32_t> &path) { |
| 2599 | const CIRGenRecordLayout &layout = |
| 2600 | getTypes().getCIRGenRecordLayout(rd: currentClass); |
| 2601 | |
| 2602 | // The field is declared directly in this class. |
| 2603 | if (astContext.isSameEntity(X: field->getParent()->getMostRecentDecl(), |
| 2604 | Y: currentClass->getMostRecentDecl())) { |
| 2605 | int32_t fieldIdx; |
| 2606 | if (currentClass->isUnion()) { |
| 2607 | // For unions, getCIRFieldNo always returns 0 for every union member (all |
| 2608 | // members share offset 0 in the CIR record). Use the declaration-order |
| 2609 | // index to distinguish members with the same type at the same offset. |
| 2610 | if (!layout.isZeroInitializable()) { |
| 2611 | errorNYI(field->getLocation(), |
| 2612 | "data member pointer for non-zero-initializable union" ); |
| 2613 | return false; |
| 2614 | } |
| 2615 | fieldIdx = static_cast<int32_t>(field->getFieldIndex()); |
| 2616 | } else { |
| 2617 | fieldIdx = static_cast<int32_t>(layout.getCIRFieldNo(fd: field)); |
| 2618 | } |
| 2619 | path.push_back(Elt: fieldIdx); |
| 2620 | return true; |
| 2621 | } |
| 2622 | |
| 2623 | // Otherwise search the base subobjects. A virtual base only blocks lowering |
| 2624 | // when the field actually lives within it; a virtual base elsewhere in the |
| 2625 | // hierarchy must not stop us from reaching a member through a non-virtual |
| 2626 | // path. |
| 2627 | for (const CXXBaseSpecifier &base : currentClass->bases()) { |
| 2628 | const auto *baseDecl = |
| 2629 | cast<CXXRecordDecl>(Val: base.getType()->getAsRecordDecl()); |
| 2630 | |
| 2631 | if (base.isVirtual()) { |
| 2632 | // A pointer to a data member that traverses a virtual base is ill-formed, |
| 2633 | // so this guard only fires defensively if the member is reached through |
| 2634 | // the virtual base. An unrelated virtual base is skipped so it does not |
| 2635 | // block members reached through a non-virtual path. |
| 2636 | llvm::SmallVector<int32_t> discardedPath; |
| 2637 | if (findFieldMemberPath(currentClass: baseDecl, field, path&: discardedPath)) { |
| 2638 | errorNYI(field->getLocation(), |
| 2639 | "data member pointer through virtual base" ); |
| 2640 | return false; |
| 2641 | } |
| 2642 | continue; |
| 2643 | } |
| 2644 | |
| 2645 | // If a base class doesn't participate in layout, the field cannot be in it, |
| 2646 | // skip it. |
| 2647 | if (!layout.hasNonVirtualBaseCIRField(rd: baseDecl)) |
| 2648 | continue; |
| 2649 | |
| 2650 | auto baseFieldIdx = |
| 2651 | static_cast<int32_t>(layout.getNonVirtualBaseCIRFieldNo(rd: baseDecl)); |
| 2652 | path.push_back(Elt: baseFieldIdx); |
| 2653 | if (findFieldMemberPath(currentClass: baseDecl, field, path)) |
| 2654 | return true; |
| 2655 | path.pop_back(); |
| 2656 | } |
| 2657 | return false; |
| 2658 | } |
| 2659 | |
| 2660 | bool CIRGenModule::isEmptyFieldForMemberPointer(const FieldDecl *field) { |
| 2661 | if (!field->isPotentiallyOverlapping() || |
| 2662 | !CodeGenUtils::isEmptyFieldForLayout(Ctx: astContext, FD: field)) |
| 2663 | return false; |
| 2664 | |
| 2665 | // Unions always have a field even if they are empty. |
| 2666 | const RecordDecl *rec = field->getParent(); |
| 2667 | if (rec->isUnion()) |
| 2668 | return true; |
| 2669 | |
| 2670 | // Otherwise, count on whether accumulateFields gave this a member. |
| 2671 | return !getTypes().getCIRGenRecordLayout(rd: rec).hasCIRField(fd: field); |
| 2672 | } |
| 2673 | |
| 2674 | void CIRGenModule::emitDeclContext(const DeclContext *dc) { |
| 2675 | for (Decl *decl : dc->decls()) { |
| 2676 | // Unlike other DeclContexts, the contents of an ObjCImplDecl at TU scope |
| 2677 | // are themselves considered "top-level", so EmitTopLevelDecl on an |
| 2678 | // ObjCImplDecl does not recursively visit them. We need to do that in |
| 2679 | // case they're nested inside another construct (LinkageSpecDecl / |
| 2680 | // ExportDecl) that does stop them from being considered "top-level". |
| 2681 | if (auto *oid = dyn_cast<ObjCImplDecl>(Val: decl)) |
| 2682 | errorNYI(oid->getSourceRange(), "emitDeclConext: ObjCImplDecl" ); |
| 2683 | |
| 2684 | emitTopLevelDecl(decl); |
| 2685 | } |
| 2686 | } |
| 2687 | |
| 2688 | // Emit code for a single top level declaration. |
| 2689 | void CIRGenModule::emitTopLevelDecl(Decl *decl) { |
| 2690 | |
| 2691 | // Ignore dependent declarations. |
| 2692 | if (decl->isTemplated()) |
| 2693 | return; |
| 2694 | |
| 2695 | switch (decl->getKind()) { |
| 2696 | default: |
| 2697 | errorNYI(loc: decl->getBeginLoc(), feature: "declaration of kind" , |
| 2698 | name: decl->getDeclKindName()); |
| 2699 | break; |
| 2700 | |
| 2701 | case Decl::CXXConversion: |
| 2702 | case Decl::CXXMethod: |
| 2703 | case Decl::Function: { |
| 2704 | auto *fd = cast<FunctionDecl>(Val: decl); |
| 2705 | // Consteval functions shouldn't be emitted. |
| 2706 | if (!fd->isConsteval()) |
| 2707 | emitGlobal(gd: fd); |
| 2708 | break; |
| 2709 | } |
| 2710 | |
| 2711 | case Decl::Import: { |
| 2712 | auto *importDecl = cast<ImportDecl>(Val: decl); |
| 2713 | clang::Module *mod = importDecl->getImportedModule(); |
| 2714 | |
| 2715 | // Nothing to record or to emit without a module. |
| 2716 | if (!mod) |
| 2717 | break; |
| 2718 | |
| 2719 | // If we've already imported this module, we're done. |
| 2720 | if (!importedModules.insert(X: mod)) |
| 2721 | break; |
| 2722 | moduleImportLocs[mod] = importDecl->getLocation(); |
| 2723 | |
| 2724 | assert(!cir::MissingFeatures::generateDebugInfo()); |
| 2725 | |
| 2726 | // A C++20 named module has its own initializer function, so its |
| 2727 | // initializers are not emitted here. The global init function of this |
| 2728 | // translation unit calls that initializer, which in turn covers the |
| 2729 | // modules it imports itself; release() records the modules to call. |
| 2730 | if (cxx20ModuleInits && mod->isNamedModule()) |
| 2731 | break; |
| 2732 | |
| 2733 | // For clang C++ module map modules the initializers for sub-modules are |
| 2734 | // emitted here. |
| 2735 | |
| 2736 | // Find all of the submodules and emit the module initializers. |
| 2737 | llvm::SmallPtrSet<clang::Module *, 16> visited; |
| 2738 | SmallVector<clang::Module *, 16> stack; |
| 2739 | visited.insert(Ptr: mod); |
| 2740 | stack.push_back(Elt: mod); |
| 2741 | |
| 2742 | while (!stack.empty()) { |
| 2743 | clang::Module *cur = stack.pop_back_val(); |
| 2744 | if (!emittedModuleInitializers.insert(Ptr: cur).second) |
| 2745 | continue; |
| 2746 | |
| 2747 | for (Decl *initializer : astContext.getModuleInitializers(M: cur)) |
| 2748 | emitTopLevelDecl(decl: initializer); |
| 2749 | |
| 2750 | // Visit the submodules of this module. |
| 2751 | for (clang::Module *submodule : cur->submodules()) { |
| 2752 | // Skip explicit children; they need to be explicitly imported to emit |
| 2753 | // the initializers. |
| 2754 | if (submodule->IsExplicit) |
| 2755 | continue; |
| 2756 | |
| 2757 | if (visited.insert(Ptr: submodule).second) |
| 2758 | stack.push_back(Elt: submodule); |
| 2759 | } |
| 2760 | } |
| 2761 | break; |
| 2762 | } |
| 2763 | |
| 2764 | case Decl::Export: |
| 2765 | emitDeclContext(dc: cast<ExportDecl>(Val: decl)); |
| 2766 | break; |
| 2767 | |
| 2768 | case Decl::Var: |
| 2769 | case Decl::Decomposition: |
| 2770 | case Decl::VarTemplateSpecialization: { |
| 2771 | emitGlobal(gd: cast<VarDecl>(Val: decl)); |
| 2772 | if (auto *decomp = dyn_cast<DecompositionDecl>(Val: decl)) |
| 2773 | for (auto *binding : decomp->flat_bindings()) |
| 2774 | if (auto *holdingVar = binding->getHoldingVar()) |
| 2775 | emitGlobal(gd: holdingVar); |
| 2776 | break; |
| 2777 | } |
| 2778 | case Decl::OpenACCRoutine: |
| 2779 | emitGlobalOpenACCRoutineDecl(cd: cast<OpenACCRoutineDecl>(Val: decl)); |
| 2780 | break; |
| 2781 | case Decl::OpenACCDeclare: |
| 2782 | emitGlobalOpenACCDeclareDecl(cd: cast<OpenACCDeclareDecl>(Val: decl)); |
| 2783 | break; |
| 2784 | case Decl::OMPThreadPrivate: |
| 2785 | emitOMPThreadPrivateDecl(d: cast<OMPThreadPrivateDecl>(Val: decl)); |
| 2786 | break; |
| 2787 | case Decl::OMPGroupPrivate: |
| 2788 | emitOMPGroupPrivateDecl(d: cast<OMPGroupPrivateDecl>(Val: decl)); |
| 2789 | break; |
| 2790 | case Decl::OMPAllocate: |
| 2791 | emitOMPAllocateDecl(d: cast<OMPAllocateDecl>(Val: decl)); |
| 2792 | break; |
| 2793 | case Decl::OMPCapturedExpr: |
| 2794 | emitOMPCapturedExpr(d: cast<OMPCapturedExprDecl>(Val: decl)); |
| 2795 | break; |
| 2796 | case Decl::OMPDeclareReduction: |
| 2797 | emitOMPDeclareReduction(d: cast<OMPDeclareReductionDecl>(Val: decl)); |
| 2798 | break; |
| 2799 | case Decl::OMPDeclareMapper: |
| 2800 | emitOMPDeclareMapper(d: cast<OMPDeclareMapperDecl>(Val: decl)); |
| 2801 | break; |
| 2802 | case Decl::OMPRequires: |
| 2803 | emitOMPRequiresDecl(d: cast<OMPRequiresDecl>(Val: decl)); |
| 2804 | break; |
| 2805 | case Decl::Enum: |
| 2806 | case Decl::Using: // using X; [C++] |
| 2807 | case Decl::UsingDirective: // using namespace X; [C++] |
| 2808 | case Decl::UsingEnum: // using enum X; [C++] |
| 2809 | case Decl::NamespaceAlias: |
| 2810 | case Decl::Typedef: |
| 2811 | case Decl::TypeAlias: // using foo = bar; [C++11] |
| 2812 | case Decl::Record: |
| 2813 | assert(!cir::MissingFeatures::generateDebugInfo()); |
| 2814 | break; |
| 2815 | |
| 2816 | // Indirect fields from global anonymous structs and unions can be |
| 2817 | // ignored; only the actual variable requires IR gen support. |
| 2818 | case Decl::IndirectField: |
| 2819 | break; |
| 2820 | |
| 2821 | // No code generation needed. |
| 2822 | case Decl::ClassTemplate: |
| 2823 | case Decl::Concept: |
| 2824 | case Decl::CXXDeductionGuide: |
| 2825 | case Decl::Empty: |
| 2826 | case Decl::ExplicitInstantiation: |
| 2827 | case Decl::FunctionTemplate: |
| 2828 | case Decl::StaticAssert: |
| 2829 | case Decl::TypeAliasTemplate: |
| 2830 | case Decl::UsingShadow: |
| 2831 | case Decl::VarTemplate: |
| 2832 | case Decl::VarTemplatePartialSpecialization: |
| 2833 | break; |
| 2834 | |
| 2835 | case Decl::CXXConstructor: |
| 2836 | getCXXABI().emitCXXConstructors(d: cast<CXXConstructorDecl>(Val: decl)); |
| 2837 | break; |
| 2838 | case Decl::CXXDestructor: |
| 2839 | getCXXABI().emitCXXDestructors(d: cast<CXXDestructorDecl>(Val: decl)); |
| 2840 | break; |
| 2841 | |
| 2842 | // C++ Decls |
| 2843 | case Decl::LinkageSpec: |
| 2844 | case Decl::Namespace: |
| 2845 | emitDeclContext(dc: Decl::castToDeclContext(decl)); |
| 2846 | break; |
| 2847 | |
| 2848 | case Decl::ClassTemplateSpecialization: |
| 2849 | case Decl::CXXRecord: { |
| 2850 | CXXRecordDecl *crd = cast<CXXRecordDecl>(Val: decl); |
| 2851 | assert(!cir::MissingFeatures::generateDebugInfo()); |
| 2852 | for (auto *childDecl : crd->decls()) |
| 2853 | if (isa<VarDecl, CXXRecordDecl, EnumDecl, OpenACCDeclareDecl>(Val: childDecl)) |
| 2854 | emitTopLevelDecl(decl: childDecl); |
| 2855 | break; |
| 2856 | } |
| 2857 | |
| 2858 | case Decl::FileScopeAsm: |
| 2859 | // File-scope asm is ignored during device-side CUDA compilation. |
| 2860 | if (langOpts.CUDA && langOpts.CUDAIsDevice) |
| 2861 | break; |
| 2862 | // File-scope asm is ignored during device-side OpenMP compilation. |
| 2863 | if (langOpts.OpenMPIsTargetDevice) |
| 2864 | break; |
| 2865 | // File-scope asm is ignored during device-side SYCL compilation. |
| 2866 | if (langOpts.SYCLIsDevice) |
| 2867 | break; |
| 2868 | auto *file_asm = cast<FileScopeAsmDecl>(Val: decl); |
| 2869 | std::string line = file_asm->getAsmString(); |
| 2870 | globalScopeAsm.push_back(Elt: builder.getStringAttr(bytes: line)); |
| 2871 | break; |
| 2872 | } |
| 2873 | } |
| 2874 | |
| 2875 | void CIRGenModule::setInitializer(cir::GlobalOp &op, mlir::Attribute value) { |
| 2876 | // Recompute visibility when updating initializer. |
| 2877 | op.setInitialValueAttr(value); |
| 2878 | assert(!cir::MissingFeatures::opGlobalVisibility()); |
| 2879 | } |
| 2880 | |
| 2881 | std::pair<cir::FuncType, cir::FuncOp> CIRGenModule::getAddrAndTypeOfCXXStructor( |
| 2882 | GlobalDecl gd, const CIRGenFunctionInfo *fnInfo, cir::FuncType fnType, |
| 2883 | bool dontDefer, ForDefinition_t isForDefinition) { |
| 2884 | auto *md = cast<CXXMethodDecl>(Val: gd.getDecl()); |
| 2885 | |
| 2886 | if (isa<CXXDestructorDecl>(Val: md)) { |
| 2887 | // Always alias equivalent complete destructors to base destructors in the |
| 2888 | // MS ABI. |
| 2889 | if (getTarget().getCXXABI().isMicrosoft() && |
| 2890 | gd.getDtorType() == Dtor_Complete && |
| 2891 | md->getParent()->getNumVBases() == 0) |
| 2892 | errorNYI(md->getSourceRange(), |
| 2893 | "getAddrAndTypeOfCXXStructor: MS ABI complete destructor" ); |
| 2894 | } |
| 2895 | |
| 2896 | if (!fnType) { |
| 2897 | if (!fnInfo) |
| 2898 | fnInfo = &getTypes().arrangeCXXStructorDeclaration(gd); |
| 2899 | fnType = getTypes().getFunctionType(info: *fnInfo); |
| 2900 | } |
| 2901 | |
| 2902 | auto fn = getOrCreateCIRFunction(mangledName: getMangledName(gd), funcType: fnType, gd, |
| 2903 | /*ForVtable=*/forVTable: false, dontDefer, |
| 2904 | /*IsThunk=*/isThunk: false, isForDefinition); |
| 2905 | |
| 2906 | return {fnType, fn}; |
| 2907 | } |
| 2908 | |
| 2909 | cir::FuncOp CIRGenModule::getAddrOfFunction(clang::GlobalDecl gd, |
| 2910 | mlir::Type funcType, bool forVTable, |
| 2911 | bool dontDefer, |
| 2912 | ForDefinition_t isForDefinition) { |
| 2913 | assert(!cast<FunctionDecl>(gd.getDecl())->isConsteval() && |
| 2914 | "consteval function should never be emitted" ); |
| 2915 | |
| 2916 | if (!funcType) { |
| 2917 | const auto *fd = cast<FunctionDecl>(Val: gd.getDecl()); |
| 2918 | funcType = convertType(type: fd->getType()); |
| 2919 | } |
| 2920 | |
| 2921 | // Devirtualized destructor calls may come through here instead of via |
| 2922 | // getAddrOfCXXStructor. Make sure we use the MS ABI base destructor instead |
| 2923 | // of the complete destructor when necessary. |
| 2924 | if (const auto *dd = dyn_cast<CXXDestructorDecl>(Val: gd.getDecl())) { |
| 2925 | if (getTarget().getCXXABI().isMicrosoft() && |
| 2926 | gd.getDtorType() == Dtor_Complete && |
| 2927 | dd->getParent()->getNumVBases() == 0) |
| 2928 | errorNYI(dd->getSourceRange(), |
| 2929 | "getAddrOfFunction: MS ABI complete destructor" ); |
| 2930 | } |
| 2931 | |
| 2932 | StringRef mangledName = getMangledName(gd); |
| 2933 | cir::FuncOp func = |
| 2934 | getOrCreateCIRFunction(mangledName, funcType, gd, forVTable, dontDefer, |
| 2935 | /*isThunk=*/false, isForDefinition); |
| 2936 | // Returns kernel handle for HIP kernel stub function. |
| 2937 | if (langOpts.CUDA && !langOpts.CUDAIsDevice && |
| 2938 | cast<FunctionDecl>(Val: gd.getDecl())->hasAttr<CUDAGlobalAttr>()) { |
| 2939 | mlir::Operation *handle = getCUDARuntime().getKernelHandle(fn: func, gd); |
| 2940 | |
| 2941 | // For HIP the kernel handle is a GlobalOp, which cannot be cast to |
| 2942 | // FuncOp. Return the stub directly in that case. |
| 2943 | bool isHIPHandle = mlir::isa<cir::GlobalOp>(Val: *handle); |
| 2944 | if (isForDefinition || isHIPHandle) |
| 2945 | return func; |
| 2946 | return mlir::dyn_cast<cir::FuncOp>(Val&: *handle); |
| 2947 | } |
| 2948 | return func; |
| 2949 | } |
| 2950 | |
| 2951 | static std::string getMangledNameImpl(CIRGenModule &cgm, GlobalDecl gd, |
| 2952 | const NamedDecl *nd) { |
| 2953 | SmallString<256> buffer; |
| 2954 | |
| 2955 | llvm::raw_svector_ostream out(buffer); |
| 2956 | MangleContext &mc = cgm.getCXXABI().getMangleContext(); |
| 2957 | |
| 2958 | assert(!cir::MissingFeatures::moduleNameHash()); |
| 2959 | |
| 2960 | if (mc.shouldMangleDeclName(D: nd)) { |
| 2961 | mc.mangleName(GD: gd.getWithDecl(D: nd), out); |
| 2962 | } else { |
| 2963 | IdentifierInfo *ii = nd->getIdentifier(); |
| 2964 | assert(ii && "Attempt to mangle unnamed decl." ); |
| 2965 | |
| 2966 | const auto *fd = dyn_cast<FunctionDecl>(Val: nd); |
| 2967 | if (fd && |
| 2968 | fd->getType()->castAs<FunctionType>()->getCallConv() == CC_X86RegCall) { |
| 2969 | cgm.errorNYI(nd->getSourceRange(), "getMangledName: X86RegCall" ); |
| 2970 | } else if (fd && fd->hasAttr<CUDAGlobalAttr>() && |
| 2971 | gd.getKernelReferenceKind() == KernelReferenceKind::Stub) { |
| 2972 | out << "__device_stub__" << ii->getName(); |
| 2973 | } else if (fd && |
| 2974 | DeviceKernelAttr::isOpenCLSpelling( |
| 2975 | A: fd->getAttr<DeviceKernelAttr>()) && |
| 2976 | gd.getKernelReferenceKind() == KernelReferenceKind::Stub) { |
| 2977 | cgm.errorNYI(nd->getSourceRange(), "getMangledName: OpenCL Stub" ); |
| 2978 | } else { |
| 2979 | out << ii->getName(); |
| 2980 | } |
| 2981 | } |
| 2982 | |
| 2983 | // Check if the module name hash should be appended for internal linkage |
| 2984 | // symbols. This should come before multi-version target suffixes are |
| 2985 | // appendded. This is to keep the name and module hash suffix of the internal |
| 2986 | // linkage function together. The unique suffix should only be added when name |
| 2987 | // mangling is done to make sure that the final name can be properly |
| 2988 | // demangled. For example, for C functions without prototypes, name mangling |
| 2989 | // is not done and the unique suffix should not be appended then. |
| 2990 | assert(!cir::MissingFeatures::moduleNameHash()); |
| 2991 | |
| 2992 | if (const auto *fd = dyn_cast<FunctionDecl>(Val: nd)) { |
| 2993 | if (fd->isMultiVersion()) { |
| 2994 | cgm.errorNYI(nd->getSourceRange(), |
| 2995 | "getMangledName: multi-version functions" ); |
| 2996 | } |
| 2997 | } |
| 2998 | // Make unique name for device side static file-scope variable for HIP. |
| 2999 | if (cgm.getASTContext().shouldExternalize(D: nd) && |
| 3000 | cgm.getLangOpts().GPURelocatableDeviceCode) |
| 3001 | if (cgm.getLangOpts().CUDA) |
| 3002 | cgm.errorNYI(nd->getSourceRange(), |
| 3003 | "getMangledName: GPU relocatable device code" ); |
| 3004 | |
| 3005 | return std::string(out.str()); |
| 3006 | } |
| 3007 | |
| 3008 | static FunctionDecl * |
| 3009 | createOpenACCBindTempFunction(ASTContext &ctx, const IdentifierInfo *bindName, |
| 3010 | const FunctionDecl *protoFunc) { |
| 3011 | // If this is a C no-prototype function, we can take the 'easy' way out and |
| 3012 | // just create a function with no arguments/functions, etc. |
| 3013 | if (!protoFunc->hasPrototype()) |
| 3014 | return FunctionDecl::Create( |
| 3015 | C&: ctx, /*DC=*/ctx.getTranslationUnitDecl(), |
| 3016 | /*StartLoc=*/SourceLocation{}, /*NLoc=*/SourceLocation{}, N: bindName, |
| 3017 | T: protoFunc->getType(), /*TInfo=*/nullptr, SC: StorageClass::SC_None); |
| 3018 | |
| 3019 | QualType funcTy = protoFunc->getType(); |
| 3020 | auto *fpt = cast<FunctionProtoType>(Val: protoFunc->getType()); |
| 3021 | |
| 3022 | // If this is a member function, add an explicit 'this' to the function type. |
| 3023 | if (auto *methodDecl = dyn_cast<CXXMethodDecl>(Val: protoFunc); |
| 3024 | methodDecl && methodDecl->isImplicitObjectMemberFunction()) { |
| 3025 | llvm::SmallVector<QualType> paramTypes{fpt->getParamTypes()}; |
| 3026 | paramTypes.insert(I: paramTypes.begin(), Elt: methodDecl->getThisType()); |
| 3027 | |
| 3028 | funcTy = ctx.getFunctionType(ResultTy: fpt->getReturnType(), Args: paramTypes, |
| 3029 | EPI: fpt->getExtProtoInfo()); |
| 3030 | fpt = cast<FunctionProtoType>(Val&: funcTy); |
| 3031 | } |
| 3032 | |
| 3033 | auto *tempFunc = |
| 3034 | FunctionDecl::Create(C&: ctx, /*DC=*/ctx.getTranslationUnitDecl(), |
| 3035 | /*StartLoc=*/SourceLocation{}, |
| 3036 | /*NLoc=*/SourceLocation{}, N: bindName, T: funcTy, |
| 3037 | /*TInfo=*/nullptr, SC: StorageClass::SC_None); |
| 3038 | |
| 3039 | SmallVector<ParmVarDecl *> params; |
| 3040 | params.reserve(N: fpt->getNumParams()); |
| 3041 | |
| 3042 | // Add all of the parameters. |
| 3043 | for (unsigned i = 0, e = fpt->getNumParams(); i != e; ++i) { |
| 3044 | ParmVarDecl *parm = ParmVarDecl::Create( |
| 3045 | C&: ctx, DC: tempFunc, /*StartLoc=*/SourceLocation{}, |
| 3046 | /*IdLoc=*/SourceLocation{}, |
| 3047 | /*Id=*/nullptr, T: fpt->getParamType(i), /*TInfo=*/nullptr, |
| 3048 | S: StorageClass::SC_None, /*DefArg=*/nullptr); |
| 3049 | parm->setScopeInfo(scopeDepth: 0, parameterIndex: i); |
| 3050 | params.push_back(Elt: parm); |
| 3051 | } |
| 3052 | |
| 3053 | tempFunc->setParams(params); |
| 3054 | |
| 3055 | return tempFunc; |
| 3056 | } |
| 3057 | |
| 3058 | std::string |
| 3059 | CIRGenModule::getOpenACCBindMangledName(const IdentifierInfo *bindName, |
| 3060 | const FunctionDecl *attachedFunction) { |
| 3061 | FunctionDecl *tempFunc = createOpenACCBindTempFunction( |
| 3062 | ctx&: getASTContext(), bindName, protoFunc: attachedFunction); |
| 3063 | |
| 3064 | std::string ret = getMangledNameImpl(cgm&: *this, gd: GlobalDecl(tempFunc), nd: tempFunc); |
| 3065 | |
| 3066 | // This does nothing (it is a do-nothing function), since this is a |
| 3067 | // slab-allocator, but leave a call in to immediately destroy this in case we |
| 3068 | // ever come up with a way of getting allocations back. |
| 3069 | getASTContext().Deallocate(Ptr: tempFunc); |
| 3070 | return ret; |
| 3071 | } |
| 3072 | |
| 3073 | StringRef CIRGenModule::getMangledName(GlobalDecl gd) { |
| 3074 | GlobalDecl canonicalGd = gd.getCanonicalDecl(); |
| 3075 | |
| 3076 | // Some ABIs don't have constructor variants. Make sure that base and complete |
| 3077 | // constructors get mangled the same. |
| 3078 | if (const auto *cd = dyn_cast<CXXConstructorDecl>(Val: canonicalGd.getDecl())) { |
| 3079 | if (!getTarget().getCXXABI().hasConstructorVariants()) { |
| 3080 | errorNYI(cd->getSourceRange(), |
| 3081 | "getMangledName: C++ constructor without variants" ); |
| 3082 | return cast<NamedDecl>(Val: gd.getDecl())->getIdentifier()->getName(); |
| 3083 | } |
| 3084 | } |
| 3085 | |
| 3086 | // In CUDA/HIP device compilation with -fgpu-rdc, the mangled name of a |
| 3087 | // static device variable depends on whether the variable is referenced by |
| 3088 | // a host or device host function. Therefore the mangled name cannot be |
| 3089 | // cached. |
| 3090 | if (!langOpts.CUDAIsDevice || !astContext.mayExternalize(D: gd.getDecl())) { |
| 3091 | auto foundName = mangledDeclNames.find(Key: canonicalGd); |
| 3092 | if (foundName != mangledDeclNames.end()) |
| 3093 | return foundName->second; |
| 3094 | } |
| 3095 | |
| 3096 | // Keep the first result in the case of a mangling collision. |
| 3097 | const auto *nd = cast<NamedDecl>(Val: gd.getDecl()); |
| 3098 | std::string mangledName = getMangledNameImpl(cgm&: *this, gd, nd); |
| 3099 | |
| 3100 | auto result = manglings.insert(KV: std::make_pair(x&: mangledName, y&: gd)); |
| 3101 | return mangledDeclNames[canonicalGd] = result.first->first(); |
| 3102 | } |
| 3103 | |
| 3104 | void CIRGenModule::emitTentativeDefinition(const VarDecl *d) { |
| 3105 | assert(!d->getInit() && "Cannot emit definite definitions here!" ); |
| 3106 | |
| 3107 | StringRef mangledName = getMangledName(gd: d); |
| 3108 | mlir::Operation *gv = getGlobalValue(name: mangledName); |
| 3109 | |
| 3110 | // If we already have a definition, not declaration, with the same mangled |
| 3111 | // name, emitting of declaration is not required (and would actually overwrite |
| 3112 | // the emitted definition). |
| 3113 | if (gv) { |
| 3114 | auto gvOp = dyn_cast<cir::GlobalOp>(Val: gv); |
| 3115 | if (!gvOp) { |
| 3116 | errorNYI(d->getSourceRange(), |
| 3117 | "emitTentativeDefinition: name already used by a non-global" ); |
| 3118 | return; |
| 3119 | } |
| 3120 | if (!gvOp.isDeclaration()) |
| 3121 | return; |
| 3122 | } |
| 3123 | |
| 3124 | // If we have not seen a reference to this variable yet, place it into the |
| 3125 | // deferred declarations table to be emitted if needed later. |
| 3126 | if (!mustBeEmitted(d) && !gv) { |
| 3127 | deferredDecls[mangledName] = d; |
| 3128 | return; |
| 3129 | } |
| 3130 | |
| 3131 | // The tentative definition is the only definition. |
| 3132 | emitGlobalVarDefinition(vd: d); |
| 3133 | } |
| 3134 | |
| 3135 | bool CIRGenModule::mustBeEmitted(const ValueDecl *global) { |
| 3136 | // Never defer when EmitAllDecls is specified. |
| 3137 | if (langOpts.EmitAllDecls) |
| 3138 | return true; |
| 3139 | |
| 3140 | const auto *vd = dyn_cast<VarDecl>(Val: global); |
| 3141 | if (vd && |
| 3142 | ((codeGenOpts.KeepPersistentStorageVariables && |
| 3143 | (vd->getStorageDuration() == SD_Static || |
| 3144 | vd->getStorageDuration() == SD_Thread)) || |
| 3145 | (codeGenOpts.KeepStaticConsts && vd->getStorageDuration() == SD_Static && |
| 3146 | vd->getType().isConstQualified()))) |
| 3147 | return true; |
| 3148 | |
| 3149 | return getASTContext().DeclMustBeEmitted(D: global); |
| 3150 | } |
| 3151 | |
| 3152 | bool CIRGenModule::mayBeEmittedEagerly(const ValueDecl *global) { |
| 3153 | // In OpenMP 5.0 variables and function may be marked as |
| 3154 | // device_type(host/nohost) and we should not emit them eagerly unless we sure |
| 3155 | // that they must be emitted on the host/device. To be sure we need to have |
| 3156 | // seen a declare target with an explicit mentioning of the function, we know |
| 3157 | // we have if the level of the declare target attribute is -1. Note that we |
| 3158 | // check somewhere else if we should emit this at all. |
| 3159 | if (langOpts.OpenMP >= 50 && !langOpts.OpenMPSimd) { |
| 3160 | std::optional<OMPDeclareTargetDeclAttr *> activeAttr = |
| 3161 | OMPDeclareTargetDeclAttr::getActiveAttr(VD: global); |
| 3162 | if (!activeAttr || (*activeAttr)->getLevel() != (unsigned)-1) |
| 3163 | return false; |
| 3164 | } |
| 3165 | |
| 3166 | const auto *fd = dyn_cast<FunctionDecl>(Val: global); |
| 3167 | if (fd) { |
| 3168 | // Implicit template instantiations may change linkage if they are later |
| 3169 | // explicitly instantiated, so they should not be emitted eagerly. |
| 3170 | if (fd->getTemplateSpecializationKind() == TSK_ImplicitInstantiation) |
| 3171 | return false; |
| 3172 | // Defer until all versions have been semantically checked. |
| 3173 | if (fd->hasAttr<TargetVersionAttr>() && !fd->isMultiVersion()) |
| 3174 | return false; |
| 3175 | // Defer emission of SYCL kernel entry point functions during device |
| 3176 | // compilation. |
| 3177 | if (langOpts.SYCLIsDevice && fd->hasAttr<SYCLKernelEntryPointAttr>()) |
| 3178 | return false; |
| 3179 | } |
| 3180 | const auto *vd = dyn_cast<VarDecl>(Val: global); |
| 3181 | if (vd) |
| 3182 | if (astContext.getInlineVariableDefinitionKind(VD: vd) == |
| 3183 | ASTContext::InlineVariableDefinitionKind::WeakUnknown) |
| 3184 | // A definition of an inline constexpr static data member may change |
| 3185 | // linkage later if it's redeclared outside the class. |
| 3186 | return false; |
| 3187 | |
| 3188 | // If OpenMP is enabled and threadprivates must be generated like TLS, delay |
| 3189 | // codegen for global variables, because they may be marked as threadprivate. |
| 3190 | if (langOpts.OpenMP && langOpts.OpenMPUseTLS && |
| 3191 | astContext.getTargetInfo().isTLSSupported() && isa<VarDecl>(Val: global) && |
| 3192 | !global->getType().isConstantStorage(Ctx: astContext, ExcludeCtor: false, ExcludeDtor: false) && |
| 3193 | !OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD: global)) |
| 3194 | return false; |
| 3195 | |
| 3196 | assert((fd || vd) && |
| 3197 | "Only FunctionDecl and VarDecl should hit this path so far." ); |
| 3198 | return true; |
| 3199 | } |
| 3200 | |
| 3201 | static bool shouldAssumeDSOLocal(const CIRGenModule &cgm, |
| 3202 | cir::CIRGlobalValueInterface gv) { |
| 3203 | if (gv.hasLocalLinkage()) |
| 3204 | return true; |
| 3205 | |
| 3206 | if (!gv.hasDefaultVisibility() && !gv.hasExternalWeakLinkage()) |
| 3207 | return true; |
| 3208 | |
| 3209 | // DLLImport explicitly marks the GV as external. |
| 3210 | // so it shouldn't be dso_local |
| 3211 | // But we don't have the info set now |
| 3212 | assert(!cir::MissingFeatures::opGlobalDLLImportExport()); |
| 3213 | |
| 3214 | const llvm::Triple &tt = cgm.getTriple(); |
| 3215 | const CodeGenOptions &cgOpts = cgm.getCodeGenOpts(); |
| 3216 | if (tt.isOSCygMing()) { |
| 3217 | // In MinGW and Cygwin, variables without DLLImport can still be |
| 3218 | // automatically imported from a DLL by the linker; don't mark variables |
| 3219 | // that potentially could come from another DLL as DSO local. |
| 3220 | |
| 3221 | // With EmulatedTLS, TLS variables can be autoimported from other DLLs |
| 3222 | // (and this actually happens in the public interface of libstdc++), so |
| 3223 | // such variables can't be marked as DSO local. (Native TLS variables |
| 3224 | // can't be dllimported at all, though.) |
| 3225 | cgm.errorNYI(feature: "shouldAssumeDSOLocal: MinGW" ); |
| 3226 | } |
| 3227 | |
| 3228 | // On COFF, don't mark 'extern_weak' symbols as DSO local. If these symbols |
| 3229 | // remain unresolved in the link, they can be resolved to zero, which is |
| 3230 | // outside the current DSO. |
| 3231 | if (tt.isOSBinFormatCOFF() && gv.hasExternalWeakLinkage()) |
| 3232 | return false; |
| 3233 | |
| 3234 | // Every other GV is local on COFF. |
| 3235 | // Make an exception for windows OS in the triple: Some firmware builds use |
| 3236 | // *-win32-macho triples. This (accidentally?) produced windows relocations |
| 3237 | // without GOT tables in older clang versions; Keep this behaviour. |
| 3238 | // FIXME: even thread local variables? |
| 3239 | if (tt.isOSBinFormatCOFF() || (tt.isOSWindows() && tt.isOSBinFormatMachO())) |
| 3240 | return true; |
| 3241 | |
| 3242 | // Only handle COFF and ELF for now. |
| 3243 | if (!tt.isOSBinFormatELF()) |
| 3244 | return false; |
| 3245 | |
| 3246 | llvm::Reloc::Model rm = cgOpts.RelocationModel; |
| 3247 | const LangOptions &lOpts = cgm.getLangOpts(); |
| 3248 | if (rm != llvm::Reloc::Static && !lOpts.PIE) { |
| 3249 | // On ELF, if -fno-semantic-interposition is specified and the target |
| 3250 | // supports local aliases, there will be neither CC1 |
| 3251 | // -fsemantic-interposition nor -fhalf-no-semantic-interposition. Set |
| 3252 | // dso_local on the function if using a local alias is preferable (can avoid |
| 3253 | // PLT indirection). |
| 3254 | auto func = dyn_cast<cir::FuncOp>(Val: gv.getOperation()); |
| 3255 | if (!func || func.getAliasee() || !gv.canBenefitFromLocalAlias()) |
| 3256 | return false; |
| 3257 | return !(lOpts.SemanticInterposition || lOpts.HalfNoSemanticInterposition); |
| 3258 | } |
| 3259 | |
| 3260 | // A definition cannot be preempted from an executable. |
| 3261 | if (!gv.isDeclarationForLinker()) |
| 3262 | return true; |
| 3263 | |
| 3264 | // Most PIC code sequences that assume that a symbol is local cannot produce a |
| 3265 | // 0 if it turns out the symbol is undefined. While this is ABI and relocation |
| 3266 | // depended, it seems worth it to handle it here. |
| 3267 | if (rm == llvm::Reloc::PIC_ && gv.hasExternalWeakLinkage()) |
| 3268 | return false; |
| 3269 | |
| 3270 | // PowerPC64 prefers TOC indirection to avoid copy relocations. |
| 3271 | if (tt.isPPC64()) |
| 3272 | return false; |
| 3273 | |
| 3274 | if (cgOpts.DirectAccessExternalData) { |
| 3275 | // If -fdirect-access-external-data (default for -fno-pic), set dso_local |
| 3276 | // for non-thread-local variables. If the symbol is not defined in the |
| 3277 | // executable, a copy relocation will be needed at link time. dso_local is |
| 3278 | // excluded for thread-local variables because they generally don't support |
| 3279 | // copy relocations. |
| 3280 | if (auto globalOp = dyn_cast<cir::GlobalOp>(Val: gv.getOperation())) { |
| 3281 | // Assume variables are not thread-local until that support is added. |
| 3282 | assert(!cir::MissingFeatures::opGlobalThreadLocal()); |
| 3283 | return true; |
| 3284 | } |
| 3285 | |
| 3286 | // -fno-pic sets dso_local on a function declaration to allow direct |
| 3287 | // accesses when taking its address (similar to a data symbol). If the |
| 3288 | // function is not defined in the executable, a canonical PLT entry will be |
| 3289 | // needed at link time. -fno-direct-access-external-data can avoid the |
| 3290 | // canonical PLT entry. We don't generalize this condition to -fpie/-fpic as |
| 3291 | // it could just cause trouble without providing perceptible benefits. |
| 3292 | if (isa<cir::FuncOp>(Val: gv) && !cgOpts.NoPLT && rm == llvm::Reloc::Static) |
| 3293 | return true; |
| 3294 | } |
| 3295 | |
| 3296 | // If we can use copy relocations we can assume it is local. |
| 3297 | |
| 3298 | // Otherwise don't assume it is local. |
| 3299 | |
| 3300 | return false; |
| 3301 | } |
| 3302 | |
| 3303 | void CIRGenModule::setGlobalVisibility(cir::CIRGlobalValueInterface gv, |
| 3304 | const NamedDecl *d) const { |
| 3305 | // Internal definitions always have default visibility. |
| 3306 | if (gv.hasLocalLinkage()) { |
| 3307 | gv.setGlobalVisibility(cir::VisibilityKind::Default); |
| 3308 | return; |
| 3309 | } |
| 3310 | if (!d) |
| 3311 | return; |
| 3312 | |
| 3313 | // Set visibility for definitions, and for declarations if requested globally |
| 3314 | // or set explicitly. |
| 3315 | LinkageInfo lv = d->getLinkageAndVisibility(); |
| 3316 | |
| 3317 | // OpenMP declare target variables must be visible to the host so they can |
| 3318 | // be registered. We require protected visibility unless the variable has |
| 3319 | // the DT_nohost modifier and does not need to be registered. |
| 3320 | if (getASTContext().getLangOpts().OpenMP && |
| 3321 | getASTContext().getLangOpts().OpenMPIsTargetDevice && isa<VarDecl>(Val: d) && |
| 3322 | d->hasAttr<OMPDeclareTargetDeclAttr>() && |
| 3323 | d->getAttr<OMPDeclareTargetDeclAttr>()->getDevType() != |
| 3324 | OMPDeclareTargetDeclAttr::DT_NoHost && |
| 3325 | lv.getVisibility() == HiddenVisibility) { |
| 3326 | llvm_unreachable("setGlobalVisibility: OpenMP is NYI" ); |
| 3327 | return; |
| 3328 | } |
| 3329 | |
| 3330 | // CUDA/HIP device kernels and global variables must be visible to the host |
| 3331 | // so they can be registered / initialized. We require protected visibility |
| 3332 | // unless the user explicitly requested hidden via an attribute. |
| 3333 | if (getASTContext().getLangOpts().CUDAIsDevice && |
| 3334 | lv.getVisibility() == HiddenVisibility && !lv.isVisibilityExplicit() && |
| 3335 | !d->hasAttr<OMPDeclareTargetDeclAttr>()) { |
| 3336 | bool needsProtected = false; |
| 3337 | if (isa<FunctionDecl>(Val: d)) { |
| 3338 | needsProtected = |
| 3339 | d->hasAttr<CUDAGlobalAttr>() || d->hasAttr<DeviceKernelAttr>(); |
| 3340 | } else if (const auto *vd = dyn_cast<VarDecl>(Val: d)) { |
| 3341 | needsProtected = vd->hasAttr<CUDADeviceAttr>() || |
| 3342 | vd->hasAttr<CUDAConstantAttr>() || |
| 3343 | vd->getType()->isCUDADeviceBuiltinSurfaceType() || |
| 3344 | vd->getType()->isCUDADeviceBuiltinTextureType(); |
| 3345 | } |
| 3346 | if (needsProtected) { |
| 3347 | gv.setGlobalVisibility(cir::VisibilityKind::Protected); |
| 3348 | return; |
| 3349 | } |
| 3350 | } |
| 3351 | |
| 3352 | if (getASTContext().getLangOpts().HLSL && !d->isInExportDeclContext()) { |
| 3353 | gv.setGlobalVisibility(cir::VisibilityKind::Hidden); |
| 3354 | return; |
| 3355 | } |
| 3356 | |
| 3357 | assert(!cir::MissingFeatures::opGlobalDLLImportExport()); |
| 3358 | |
| 3359 | if (lv.isVisibilityExplicit() || getLangOpts().SetVisibilityForExternDecls || |
| 3360 | !gv.isDeclarationForLinker()) |
| 3361 | gv.setGlobalVisibility(getCIRVisibilityKind(v: lv.getVisibility())); |
| 3362 | } |
| 3363 | |
| 3364 | void CIRGenModule::setDSOLocal(cir::CIRGlobalValueInterface gv) const { |
| 3365 | gv.setDSOLocal(shouldAssumeDSOLocal(cgm: *this, gv)); |
| 3366 | } |
| 3367 | |
| 3368 | void CIRGenModule::setDSOLocal(mlir::Operation *op) const { |
| 3369 | if (auto globalValue = dyn_cast<cir::CIRGlobalValueInterface>(Val: op)) |
| 3370 | setDSOLocal(globalValue); |
| 3371 | } |
| 3372 | |
| 3373 | void CIRGenModule::setGVProperties(mlir::Operation *op, |
| 3374 | const NamedDecl *d) const { |
| 3375 | assert(!cir::MissingFeatures::opGlobalDLLImportExport()); |
| 3376 | setGVPropertiesAux(op, d); |
| 3377 | } |
| 3378 | |
| 3379 | void CIRGenModule::setGVPropertiesAux(mlir::Operation *op, |
| 3380 | const NamedDecl *d) const { |
| 3381 | setGlobalVisibility(gv: cast<cir::CIRGlobalValueInterface>(Val: op), d); |
| 3382 | setDSOLocal(op); |
| 3383 | assert(!cir::MissingFeatures::opGlobalPartition()); |
| 3384 | } |
| 3385 | |
| 3386 | bool CIRGenModule::lookupRepresentativeDecl(StringRef mangledName, |
| 3387 | GlobalDecl &result) const { |
| 3388 | auto res = manglings.find(Key: mangledName); |
| 3389 | if (res == manglings.end()) |
| 3390 | return false; |
| 3391 | result = res->getValue(); |
| 3392 | return true; |
| 3393 | } |
| 3394 | |
| 3395 | static cir::TLSModel getCIRTLSModel(StringRef S) { |
| 3396 | return llvm::StringSwitch<cir::TLSModel>(S) |
| 3397 | .Case(S: "global-dynamic" , Value: cir::TLSModel::GeneralDynamic) |
| 3398 | .Case(S: "local-dynamic" , Value: cir::TLSModel::LocalDynamic) |
| 3399 | .Case(S: "initial-exec" , Value: cir::TLSModel::InitialExec) |
| 3400 | .Case(S: "local-exec" , Value: cir::TLSModel::LocalExec); |
| 3401 | } |
| 3402 | |
| 3403 | cir::TLSModel CIRGenModule::getDefaultCIRTLSModel() const { |
| 3404 | switch (getCodeGenOpts().getDefaultTLSModel()) { |
| 3405 | case CodeGenOptions::GeneralDynamicTLSModel: |
| 3406 | return cir::TLSModel::GeneralDynamic; |
| 3407 | case CodeGenOptions::LocalDynamicTLSModel: |
| 3408 | return cir::TLSModel::LocalDynamic; |
| 3409 | case CodeGenOptions::InitialExecTLSModel: |
| 3410 | return cir::TLSModel::InitialExec; |
| 3411 | case CodeGenOptions::LocalExecTLSModel: |
| 3412 | return cir::TLSModel::LocalExec; |
| 3413 | } |
| 3414 | llvm_unreachable("Invalid TLS model!" ); |
| 3415 | } |
| 3416 | |
| 3417 | void CIRGenModule::setTLSMode(mlir::Operation *op, const VarDecl &d, |
| 3418 | bool isExtendingDecl) { |
| 3419 | assert(d.getTLSKind() && "setting TLS mode on non-TLS var!" ); |
| 3420 | |
| 3421 | cir::TLSModel tlm = getDefaultCIRTLSModel(); |
| 3422 | |
| 3423 | // Override the TLS model if it is explicitly specified. |
| 3424 | if (const auto *attr = d.getAttr<TLSModelAttr>()) |
| 3425 | tlm = getCIRTLSModel(S: attr->getModel()); |
| 3426 | |
| 3427 | auto global = cast<cir::GlobalOp>(Val: op); |
| 3428 | global.setTlsModel(tlm); |
| 3429 | |
| 3430 | // For namespace-scope dyanmic TLS we need to set the wrapper, int, or guard |
| 3431 | // info. |
| 3432 | if (d.isStaticLocal()) |
| 3433 | return; |
| 3434 | |
| 3435 | // If this function was called to set the TLS mode for a temporary whose |
| 3436 | // lifetime is extended by the variable declared by `d`, don't emit the |
| 3437 | // wrapper, init, and guard info. |
| 3438 | if (isExtendingDecl) |
| 3439 | return; |
| 3440 | |
| 3441 | setGlobalTlsReferences(vd: d, globalOp: global); |
| 3442 | } |
| 3443 | |
| 3444 | void CIRGenModule::setCIRFunctionAttributes(GlobalDecl globalDecl, |
| 3445 | const CIRGenFunctionInfo &info, |
| 3446 | cir::FuncOp func, bool isThunk) { |
| 3447 | // TODO(cir): More logic of constructAttributeList is needed. |
| 3448 | cir::CallingConv callingConv; |
| 3449 | |
| 3450 | // TODO(cir): The current list should be initialized with the extra function |
| 3451 | // attributes, but we don't have those yet. For now, the PAL is initialized |
| 3452 | // with nothing. |
| 3453 | assert(!cir::MissingFeatures::opFuncExtraAttrs()); |
| 3454 | // Initialize PAL with existing attributes to merge attributes. |
| 3455 | mlir::NamedAttrList pal{}; |
| 3456 | std::vector<mlir::NamedAttrList> argAttrs(info.arguments().size()); |
| 3457 | mlir::NamedAttrList retAttrs{}; |
| 3458 | constructAttributeList(name: func.getName(), info, calleeInfo: globalDecl, attrs&: pal, argAttrs, |
| 3459 | retAttrs, callingConv, |
| 3460 | /*attrOnCallSite=*/false, isThunk); |
| 3461 | |
| 3462 | for (mlir::NamedAttribute attr : pal) |
| 3463 | func->setAttr(name: attr.getName(), value: attr.getValue()); |
| 3464 | |
| 3465 | llvm::for_each(Range: llvm::enumerate(First&: argAttrs), F: [func](auto idx_arg_pair) { |
| 3466 | mlir::function_interface_impl::setArgAttrs(func, idx_arg_pair.index(), |
| 3467 | idx_arg_pair.value()); |
| 3468 | }); |
| 3469 | if (!retAttrs.empty()) |
| 3470 | mlir::function_interface_impl::setResultAttrs(op: func, index: 0, attributes: retAttrs); |
| 3471 | |
| 3472 | // TODO(cir): Check X86_VectorCall incompatibility wiht WinARM64EC |
| 3473 | |
| 3474 | func.setCallingConv(callingConv); |
| 3475 | } |
| 3476 | |
| 3477 | void CIRGenModule::setFunctionAttributes(GlobalDecl globalDecl, |
| 3478 | cir::FuncOp func, |
| 3479 | bool isIncompleteFunction, |
| 3480 | bool isThunk) { |
| 3481 | // NOTE(cir): Original CodeGen checks if this is an intrinsic. In CIR we |
| 3482 | // represent them in dedicated ops. The correct attributes are ensured during |
| 3483 | // translation to LLVM. Thus, we don't need to check for them here. |
| 3484 | |
| 3485 | const auto *funcDecl = cast<FunctionDecl>(Val: globalDecl.getDecl()); |
| 3486 | |
| 3487 | if (!isIncompleteFunction) |
| 3488 | setCIRFunctionAttributes(globalDecl, |
| 3489 | info: getTypes().arrangeGlobalDeclaration(gd: globalDecl), |
| 3490 | func, isThunk); |
| 3491 | |
| 3492 | // Add the Returned attribute for "this", except for iOS 5 and earlier |
| 3493 | // where substantial code, including the libstdc++ dylib, was compiled with |
| 3494 | // GCC and does not actually return "this". |
| 3495 | if (!isThunk && getCXXABI().hasThisReturn(gd: globalDecl) && |
| 3496 | !(getTriple().isiOS() && getTriple().isOSVersionLT(Major: 6))) |
| 3497 | func.setArgAttr(index: 0, name: mlir::LLVM::LLVMDialect::getReturnedAttrName(), |
| 3498 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3499 | |
| 3500 | if (!isIncompleteFunction && func.isDeclaration()) |
| 3501 | getTargetCIRGenInfo().setTargetAttributes(decl: funcDecl, global: func, module&: *this); |
| 3502 | |
| 3503 | // Diagnose calls to this function at the backend level, mirroring |
| 3504 | // CodeGenModule::SetFunctionAttributes's "dontcall-error"/"dontcall-warn". |
| 3505 | if (const auto *errorAttr = funcDecl->getAttr<ErrorAttr>()) { |
| 3506 | if (errorAttr->isError()) |
| 3507 | func->setAttr(name: cir::CIRDialect::getDontCallErrorAttrName(), |
| 3508 | value: mlir::StringAttr::get(context: &getMLIRContext(), |
| 3509 | bytes: errorAttr->getUserDiagnostic())); |
| 3510 | else if (errorAttr->isWarning()) |
| 3511 | func->setAttr(name: cir::CIRDialect::getDontCallWarnAttrName(), |
| 3512 | value: mlir::StringAttr::get(context: &getMLIRContext(), |
| 3513 | bytes: errorAttr->getUserDiagnostic())); |
| 3514 | } |
| 3515 | |
| 3516 | // Mirrors setLinkageForGV in CodeGenModule::SetFunctionAttributes. |
| 3517 | setLinkageForFunction(cgm&: *this, func, nd: funcDecl); |
| 3518 | |
| 3519 | // If we plan on emitting this inline builtin, we can't treat it as a builtin. |
| 3520 | if (funcDecl->isInlineBuiltinDeclaration()) { |
| 3521 | const FunctionDecl *fdBody; |
| 3522 | bool hasBody = funcDecl->hasBody(Definition&: fdBody); |
| 3523 | (void)hasBody; |
| 3524 | assert(hasBody && "Inline builtin declarations should always have an " |
| 3525 | "available body!" ); |
| 3526 | assert(!cir::MissingFeatures::attributeNoBuiltin()); |
| 3527 | } |
| 3528 | |
| 3529 | if (funcDecl->isReplaceableGlobalAllocationFunction()) { |
| 3530 | // A replaceable global allocation function does not act like a builtin by |
| 3531 | // default, only if it is invoked by a new-expression or delete-expression. |
| 3532 | func->setAttr(name: cir::CIRDialect::getNoBuiltinAttrName(), |
| 3533 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3534 | } |
| 3535 | |
| 3536 | if (isa<CXXConstructorDecl>(Val: funcDecl) || isa<CXXDestructorDecl>(Val: funcDecl)) |
| 3537 | func.setUnnamedAddr(true); |
| 3538 | else if (const auto *md = dyn_cast<CXXMethodDecl>(Val: funcDecl)) |
| 3539 | if (md->isVirtual()) |
| 3540 | func.setUnnamedAddr(true); |
| 3541 | } |
| 3542 | |
| 3543 | void CIRGenModule::setCIRFunctionAttributesForDefinition( |
| 3544 | const clang::Decl *decl, cir::FuncOp f) { |
| 3545 | |
| 3546 | if ((!decl || !decl->hasAttr<NoUwtableAttr>()) && codeGenOpts.UnwindTables) |
| 3547 | f.setUwtable(static_cast<cir::UnwindTableKind>(codeGenOpts.UnwindTables)); |
| 3548 | |
| 3549 | assert(!cir::MissingFeatures::stackProtector()); |
| 3550 | |
| 3551 | if (!CodeGenUtils::hasUnwindExceptions(LangOpts: langOpts)) |
| 3552 | f->setAttr(name: cir::CIRDialect::getNoThrowAttrName(), |
| 3553 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3554 | |
| 3555 | std::optional<cir::InlineKind> existingInlineKind = f.getInlineKind(); |
| 3556 | bool isNoInline = |
| 3557 | existingInlineKind && *existingInlineKind == cir::InlineKind::NoInline; |
| 3558 | bool isAlwaysInline = existingInlineKind && |
| 3559 | *existingInlineKind == cir::InlineKind::AlwaysInline; |
| 3560 | if (!decl) { |
| 3561 | assert(!cir::MissingFeatures::hlsl()); |
| 3562 | |
| 3563 | if (!isAlwaysInline && |
| 3564 | codeGenOpts.getInlining() == CodeGenOptions::OnlyAlwaysInlining) { |
| 3565 | // If inlining is disabled and we don't have a declaration to control |
| 3566 | // inlining, mark the function as 'noinline' unless it is explicitly |
| 3567 | // marked as 'alwaysinline'. |
| 3568 | f.setInlineKind(cir::InlineKind::NoInline); |
| 3569 | } |
| 3570 | |
| 3571 | return; |
| 3572 | } |
| 3573 | |
| 3574 | assert(!cir::MissingFeatures::opFuncArmStreamingAttr()); |
| 3575 | assert(!cir::MissingFeatures::opFuncArmNewAttr()); |
| 3576 | assert(!cir::MissingFeatures::hlsl()); |
| 3577 | |
| 3578 | // Track whether we need to add the optnone attribute, starting with the |
| 3579 | // default for this optimization level. |
| 3580 | bool shouldAddOptNone = |
| 3581 | !codeGenOpts.DisableO0ImplyOptNone && codeGenOpts.OptimizationLevel == 0; |
| 3582 | // optnone is incompatible with minsize and alwaysinline. |
| 3583 | shouldAddOptNone &= !decl->hasAttr<MinSizeAttr>(); |
| 3584 | shouldAddOptNone &= !decl->hasAttr<AlwaysInlineAttr>(); |
| 3585 | |
| 3586 | // Handle inline attributes |
| 3587 | if ((shouldAddOptNone || decl->hasAttr<OptimizeNoneAttr>()) && |
| 3588 | !isAlwaysInline) { |
| 3589 | // Add optnone, but do so only if the function isn't always_inline. |
| 3590 | f->setAttr(name: cir::CIRDialect::getOptNoneAttrName(), |
| 3591 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3592 | |
| 3593 | // OptimizeNone implies noinline; we should not be inlining such functions. |
| 3594 | f.setInlineKind(cir::InlineKind::NoInline); |
| 3595 | |
| 3596 | // We still need to handle naked functions even though optnone subsumes |
| 3597 | // much of their semantics. |
| 3598 | if (decl->hasAttr<NakedAttr>()) |
| 3599 | f->setAttr(name: cir::CIRDialect::getNakedAttrName(), |
| 3600 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3601 | |
| 3602 | // OptimizeNone wins over OptimizeForSize and MinSize. |
| 3603 | f->removeAttr(name: cir::CIRDialect::getOptimizeForSizeAttrName()); |
| 3604 | f->removeAttr(name: cir::CIRDialect::getMinSizeAttrName()); |
| 3605 | } else if (decl->hasAttr<NakedAttr>()) { |
| 3606 | // Naked implies noinline: we should not be inlining such functions. |
| 3607 | f->setAttr(name: cir::CIRDialect::getNakedAttrName(), |
| 3608 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3609 | f.setInlineKind(cir::InlineKind::NoInline); |
| 3610 | } else if (decl->hasAttr<NoDuplicateAttr>()) { |
| 3611 | f->setAttr(name: cir::CIRDialect::getNoDuplicatesAttrName(), |
| 3612 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3613 | } else if (decl->hasAttr<NoInlineAttr>() && !isAlwaysInline) { |
| 3614 | // Add noinline if the function isn't always_inline. |
| 3615 | f.setInlineKind(cir::InlineKind::NoInline); |
| 3616 | } else if (decl->hasAttr<AlwaysInlineAttr>() && !isNoInline) { |
| 3617 | // Don't override AlwaysInline with NoInline, or vice versa, since we can't |
| 3618 | // specify both in IR. |
| 3619 | f.setInlineKind(cir::InlineKind::AlwaysInline); |
| 3620 | } else if (codeGenOpts.getInlining() == CodeGenOptions::OnlyAlwaysInlining) { |
| 3621 | // If inlining is disabled, force everything that isn't always_inline |
| 3622 | // to carry an explicit noinline attribute. |
| 3623 | if (!isAlwaysInline) |
| 3624 | f.setInlineKind(cir::InlineKind::NoInline); |
| 3625 | } else { |
| 3626 | // Otherwise, propagate the inline hint attribute and potentially use its |
| 3627 | // absence to mark things as noinline. |
| 3628 | // Search function and template pattern redeclarations for inline. |
| 3629 | if (auto *fd = dyn_cast<FunctionDecl>(Val: decl)) { |
| 3630 | // TODO: Share this checkForInline implementation with classic codegen. |
| 3631 | // This logic is likely to change over time, so sharing would help ensure |
| 3632 | // consistency. |
| 3633 | auto checkForInline = [](const FunctionDecl *decl) { |
| 3634 | auto checkRedeclForInline = [](const FunctionDecl *redecl) { |
| 3635 | return redecl->isInlineSpecified(); |
| 3636 | }; |
| 3637 | if (any_of(Range: decl->redecls(), P: checkRedeclForInline)) |
| 3638 | return true; |
| 3639 | const FunctionDecl *pattern = decl->getTemplateInstantiationPattern(); |
| 3640 | if (!pattern) |
| 3641 | return false; |
| 3642 | return any_of(Range: pattern->redecls(), P: checkRedeclForInline); |
| 3643 | }; |
| 3644 | if (checkForInline(fd)) { |
| 3645 | f.setInlineKind(cir::InlineKind::InlineHint); |
| 3646 | } else if (codeGenOpts.getInlining() == |
| 3647 | CodeGenOptions::OnlyHintInlining && |
| 3648 | !fd->isInlined() && !isAlwaysInline) { |
| 3649 | f.setInlineKind(cir::InlineKind::NoInline); |
| 3650 | } |
| 3651 | } |
| 3652 | } |
| 3653 | |
| 3654 | // Add other optimization related attributes if we are optimizing this |
| 3655 | // function. |
| 3656 | if (!decl->hasAttr<OptimizeNoneAttr>()) { |
| 3657 | if (decl->hasAttr<ColdAttr>()) { |
| 3658 | if (!shouldAddOptNone) |
| 3659 | f->setAttr(name: cir::CIRDialect::getOptimizeForSizeAttrName(), |
| 3660 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3661 | |
| 3662 | f->setAttr(name: cir::CIRDialect::getColdAttrName(), |
| 3663 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3664 | } |
| 3665 | |
| 3666 | if (decl->hasAttr<HotAttr>()) |
| 3667 | f->setAttr(name: cir::CIRDialect::getHotAttrName(), |
| 3668 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3669 | |
| 3670 | if (decl->hasAttr<MinSizeAttr>()) |
| 3671 | f->setAttr(name: cir::CIRDialect::getMinSizeAttrName(), |
| 3672 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 3673 | } |
| 3674 | |
| 3675 | std::optional<uint64_t> explicitAlignment; |
| 3676 | if (unsigned alignment = |
| 3677 | decl->getMaxAlignment() / getASTContext().getCharWidth()) |
| 3678 | explicitAlignment = alignment; |
| 3679 | else if (langOpts.FunctionAlignment) |
| 3680 | explicitAlignment = 1ull << langOpts.FunctionAlignment; |
| 3681 | |
| 3682 | if (explicitAlignment) { |
| 3683 | f.setAlignment(*explicitAlignment); |
| 3684 | f.setPreferredAlignment(*explicitAlignment); |
| 3685 | } else if (langOpts.PreferredFunctionAlignment) { |
| 3686 | f.setPreferredAlignment(langOpts.PreferredFunctionAlignment); |
| 3687 | } |
| 3688 | |
| 3689 | // Some C++ ABIs require 2-byte alignment for member functions, in order to |
| 3690 | // reserve a bit for differentiating between virtual and non-virtual member |
| 3691 | // functions. If the current target's C++ ABI requires this and this is a |
| 3692 | // member function, set its alignment accordingly. |
| 3693 | if (getTarget().getCXXABI().areMemberFunctionsAligned()) { |
| 3694 | if (isa<CXXMethodDecl>(Val: decl) && f.getAlignment().value_or(u: 1) < 2) |
| 3695 | f.setAlignment(2); |
| 3696 | } |
| 3697 | |
| 3698 | // Attach "sycl-module-id" to sycl_external function definitions to mark |
| 3699 | // them as entry points for per-translation-unit device-code splitting. |
| 3700 | if (getLangOpts().SYCLIsDevice && decl->hasAttr<SYCLExternalAttr>()) |
| 3701 | addSYCLModuleIdAttr(fn: f); |
| 3702 | } |
| 3703 | |
| 3704 | // Maps an AST address space to the OpenCL logical address space kind recorded |
| 3705 | // in kernel argument metadata. This mapping is independent of the target |
| 3706 | // address space map, allowing consumers to distinguish OpenCL logical address |
| 3707 | // spaces even when the target maps them to the same address space. |
| 3708 | static cir::LangAddressSpace |
| 3709 | getOpenCLKernelArgAddressSpace(LangAS addressSpace) { |
| 3710 | switch (addressSpace) { |
| 3711 | case LangAS::opencl_global: |
| 3712 | return cir::LangAddressSpace::OffloadGlobal; |
| 3713 | case LangAS::opencl_constant: |
| 3714 | return cir::LangAddressSpace::OffloadConstant; |
| 3715 | case LangAS::opencl_local: |
| 3716 | return cir::LangAddressSpace::OffloadLocal; |
| 3717 | case LangAS::opencl_generic: |
| 3718 | return cir::LangAddressSpace::OffloadGeneric; |
| 3719 | case LangAS::opencl_global_device: |
| 3720 | return cir::LangAddressSpace::OffloadGlobalDevice; |
| 3721 | case LangAS::opencl_global_host: |
| 3722 | return cir::LangAddressSpace::OffloadGlobalHost; |
| 3723 | default: |
| 3724 | // All other AST address spaces, including target-specific ones, use the |
| 3725 | // OpenCL metadata default, which lowers to SPIR address space ID 0. |
| 3726 | return cir::LangAddressSpace::Default; |
| 3727 | } |
| 3728 | } |
| 3729 | |
| 3730 | void CIRGenModule::emitOpenCLKernelArgMetadata(cir::FuncOp func, |
| 3731 | const clang::FunctionDecl *fd) { |
| 3732 | assert(fd && "expected a kernel function declaration" ); |
| 3733 | const PrintingPolicy &policy = getASTContext().getPrintingPolicy(); |
| 3734 | |
| 3735 | // Create arrays that represent the kernel argument metadata. Each array has |
| 3736 | // one value per kernel argument, in source order. |
| 3737 | SmallVector<mlir::Attribute> addressQuals; |
| 3738 | SmallVector<mlir::Attribute> accessQuals; |
| 3739 | SmallVector<mlir::Attribute> argTypeNames; |
| 3740 | SmallVector<mlir::Attribute> argBaseTypeNames; |
| 3741 | SmallVector<mlir::Attribute> argTypeQuals; |
| 3742 | SmallVector<mlir::Attribute> argNames; |
| 3743 | |
| 3744 | for (const ParmVarDecl *param : fd->parameters()) { |
| 3745 | argNames.push_back(Elt: builder.getStringAttr(bytes: param->getName())); |
| 3746 | |
| 3747 | QualType type = param->getType(); |
| 3748 | std::string typeQuals; |
| 3749 | |
| 3750 | if (type->isImageType() || type->isPipeType()) { |
| 3751 | errorNYI(param->getSourceRange(), |
| 3752 | "OpenCL kernel argument metadata for image and pipe types" ); |
| 3753 | return; |
| 3754 | } |
| 3755 | |
| 3756 | accessQuals.push_back(Elt: builder.getStringAttr(bytes: "none" )); |
| 3757 | |
| 3758 | auto getTypeSpelling = [&](QualType paramType) { |
| 3759 | std::string typeName = paramType.getUnqualifiedType().getAsString(Policy: policy); |
| 3760 | |
| 3761 | if (paramType.isCanonical()) { |
| 3762 | StringRef typeNameRef = typeName; |
| 3763 | if (typeNameRef.consume_front(Prefix: "unsigned " )) |
| 3764 | return std::string("u" ) + typeNameRef.str(); |
| 3765 | if (typeNameRef.consume_front(Prefix: "signed " )) |
| 3766 | return typeNameRef.str(); |
| 3767 | } |
| 3768 | |
| 3769 | return typeName; |
| 3770 | }; |
| 3771 | |
| 3772 | // Type metadata preserves source spelling, while base type metadata uses |
| 3773 | // canonical spelling without typedefs. |
| 3774 | if (type->isPointerType()) { |
| 3775 | QualType pointeeType = type->getPointeeType(); |
| 3776 | addressQuals.push_back(Elt: cir::LangAddressSpaceAttr::get( |
| 3777 | context: &getMLIRContext(), |
| 3778 | value: getOpenCLKernelArgAddressSpace(addressSpace: pointeeType.getAddressSpace()))); |
| 3779 | |
| 3780 | argTypeNames.push_back( |
| 3781 | Elt: builder.getStringAttr(bytes: getTypeSpelling(pointeeType) + "*" )); |
| 3782 | argBaseTypeNames.push_back(Elt: builder.getStringAttr( |
| 3783 | bytes: getTypeSpelling(pointeeType.getCanonicalType()) + "*" )); |
| 3784 | |
| 3785 | if (type.isRestrictQualified()) |
| 3786 | typeQuals = "restrict" ; |
| 3787 | if (pointeeType.isConstQualified() || |
| 3788 | pointeeType.getAddressSpace() == LangAS::opencl_constant) |
| 3789 | typeQuals += typeQuals.empty() ? "const" : " const" ; |
| 3790 | if (pointeeType.isVolatileQualified()) |
| 3791 | typeQuals += typeQuals.empty() ? "volatile" : " volatile" ; |
| 3792 | } else { |
| 3793 | addressQuals.push_back(Elt: cir::LangAddressSpaceAttr::get( |
| 3794 | context: &getMLIRContext(), value: cir::LangAddressSpace::Default)); |
| 3795 | |
| 3796 | argTypeNames.push_back(Elt: builder.getStringAttr(bytes: getTypeSpelling(type))); |
| 3797 | argBaseTypeNames.push_back( |
| 3798 | Elt: builder.getStringAttr(bytes: getTypeSpelling(type.getCanonicalType()))); |
| 3799 | } |
| 3800 | |
| 3801 | argTypeQuals.push_back(Elt: builder.getStringAttr(bytes: typeQuals)); |
| 3802 | } |
| 3803 | |
| 3804 | mlir::ArrayAttr names; |
| 3805 | if (getCodeGenOpts().EmitOpenCLArgMetadata) |
| 3806 | names = builder.getArrayAttr(value: argNames); |
| 3807 | |
| 3808 | mlir::Attribute metadata = cir::OpenCLKernelArgMetadataAttr::get( |
| 3809 | context: func.getContext(), addr_space: builder.getArrayAttr(value: addressQuals), |
| 3810 | access_qual: builder.getArrayAttr(value: accessQuals), type: builder.getArrayAttr(value: argTypeNames), |
| 3811 | base_type: builder.getArrayAttr(value: argBaseTypeNames), |
| 3812 | type_qual: builder.getArrayAttr(value: argTypeQuals), name: names); |
| 3813 | func->setAttr(name: cir::CIRDialect::getOpenCLKernelArgMetadataAttrName(), |
| 3814 | value: metadata); |
| 3815 | } |
| 3816 | |
| 3817 | cir::FuncOp CIRGenModule::getOrCreateCIRFunction( |
| 3818 | StringRef mangledName, mlir::Type funcType, GlobalDecl gd, bool forVTable, |
| 3819 | bool dontDefer, bool isThunk, ForDefinition_t isForDefinition, |
| 3820 | mlir::NamedAttrList ) { |
| 3821 | const Decl *d = gd.getDecl(); |
| 3822 | |
| 3823 | if (const auto *fd = cast_or_null<FunctionDecl>(Val: d)) { |
| 3824 | // For the device, mark the function as one that should be emitted. |
| 3825 | if (getLangOpts().OpenMPIsTargetDevice && openMPRuntime && |
| 3826 | !getOpenMPRuntime().markAsGlobalTarget(gd) && fd->isDefined() && |
| 3827 | !dontDefer && !isForDefinition) { |
| 3828 | if (const FunctionDecl *fdDef = fd->getDefinition()) { |
| 3829 | GlobalDecl gdDef; |
| 3830 | if (const auto *cd = dyn_cast<CXXConstructorDecl>(Val: fdDef)) |
| 3831 | gdDef = GlobalDecl(cd, gd.getCtorType()); |
| 3832 | else if (const auto *dd = dyn_cast<CXXDestructorDecl>(Val: fdDef)) |
| 3833 | gdDef = GlobalDecl(dd, gd.getDtorType()); |
| 3834 | else |
| 3835 | gdDef = GlobalDecl(fdDef); |
| 3836 | emitGlobal(gd: gdDef); |
| 3837 | } |
| 3838 | } |
| 3839 | |
| 3840 | // Any attempts to use a MultiVersion function should result in retrieving |
| 3841 | // the iFunc instead. Name mangling will handle the rest of the changes. |
| 3842 | if (fd->isMultiVersion()) |
| 3843 | errorNYI(fd->getSourceRange(), "getOrCreateCIRFunction: multi-version" ); |
| 3844 | } |
| 3845 | |
| 3846 | // Lookup the entry, lazily creating it if necessary. |
| 3847 | mlir::Operation *entry = getGlobalValue(name: mangledName); |
| 3848 | if (entry) { |
| 3849 | assert(mlir::isa<cir::FuncOp>(entry)); |
| 3850 | |
| 3851 | // A weakref's target is only weak until something else names it. |
| 3852 | if (weakRefReferences.erase(Ptr: entry)) { |
| 3853 | const auto *fd = cast_or_null<FunctionDecl>(Val: d); |
| 3854 | if (fd && !fd->hasAttr<WeakAttr>()) |
| 3855 | cast<cir::FuncOp>(Val: entry).setLinkage( |
| 3856 | cir::GlobalLinkageKind::ExternalLinkage); |
| 3857 | } |
| 3858 | |
| 3859 | // Handle dropped DLL attributes. |
| 3860 | if (d && !d->hasAttr<DLLImportAttr>() && !d->hasAttr<DLLExportAttr>()) { |
| 3861 | assert(!cir::MissingFeatures::setDLLStorageClass()); |
| 3862 | setDSOLocal(entry); |
| 3863 | } |
| 3864 | |
| 3865 | // If there are two attempts to define the same mangled name, issue an |
| 3866 | // error. |
| 3867 | auto fn = cast<cir::FuncOp>(Val: entry); |
| 3868 | if (isForDefinition && fn && !fn.isDeclaration()) { |
| 3869 | GlobalDecl otherGd; |
| 3870 | // Check that GD is not yet in DiagnosedConflictingDefinitions is required |
| 3871 | // to make sure that we issue an error only once. |
| 3872 | if (lookupRepresentativeDecl(mangledName, result&: otherGd) && |
| 3873 | (gd.getCanonicalDecl().getDecl() != |
| 3874 | otherGd.getCanonicalDecl().getDecl()) && |
| 3875 | diagnosedConflictingDefinitions.insert(V: gd).second) { |
| 3876 | getDiags().Report(Loc: d->getLocation(), DiagID: diag::err_duplicate_mangled_name) |
| 3877 | << mangledName; |
| 3878 | getDiags().Report(Loc: otherGd.getDecl()->getLocation(), |
| 3879 | DiagID: diag::note_previous_definition); |
| 3880 | } |
| 3881 | } |
| 3882 | |
| 3883 | if (fn && fn.getFunctionType() == funcType) { |
| 3884 | return fn; |
| 3885 | } |
| 3886 | |
| 3887 | if (!isForDefinition) { |
| 3888 | return fn; |
| 3889 | } |
| 3890 | |
| 3891 | // TODO(cir): classic codegen checks here if this is a llvm::GlobalAlias. |
| 3892 | // How will we support this? |
| 3893 | } |
| 3894 | |
| 3895 | auto *funcDecl = llvm::cast_or_null<FunctionDecl>(Val: gd.getDecl()); |
| 3896 | bool invalidLoc = !funcDecl || |
| 3897 | funcDecl->getSourceRange().getBegin().isInvalid() || |
| 3898 | funcDecl->getSourceRange().getEnd().isInvalid(); |
| 3899 | cir::FuncOp funcOp = createCIRFunction( |
| 3900 | loc: invalidLoc ? theModule->getLoc() : getLoc(cRange: funcDecl->getSourceRange()), |
| 3901 | name: mangledName, funcType: mlir::cast<cir::FuncType>(Val&: funcType), funcDecl); |
| 3902 | |
| 3903 | if (funcDecl && funcDecl->hasAttr<AnnotateAttr>()) |
| 3904 | deferredAnnotations[mangledName] = funcDecl; |
| 3905 | |
| 3906 | // If we already created a function with the same mangled name (but different |
| 3907 | // type) before, take its name and add it to the list of functions to be |
| 3908 | // replaced with F at the end of CodeGen. |
| 3909 | // |
| 3910 | // This happens if there is a prototype for a function (e.g. "int f()") and |
| 3911 | // then a definition of a different type (e.g. "int f(int x)"). |
| 3912 | if (entry) { |
| 3913 | |
| 3914 | // Fetch a generic symbol-defining operation and its uses. |
| 3915 | auto symbolOp = mlir::cast<mlir::SymbolOpInterface>(Val: entry); |
| 3916 | |
| 3917 | // This might be an implementation of a function without a prototype, in |
| 3918 | // which case, try to do special replacement of calls which match the new |
| 3919 | // prototype. The really key thing here is that we also potentially drop |
| 3920 | // arguments from the call site so as to make a direct call, which makes the |
| 3921 | // inliner happier and suppresses a number of optimizer warnings (!) about |
| 3922 | // dropping arguments. |
| 3923 | if (symbolOp.getSymbolUses(from: symbolOp->getParentOp())) |
| 3924 | replaceUsesOfNonProtoTypeWithRealFunction(old: entry, newFn: funcOp); |
| 3925 | |
| 3926 | // Obliterate no-proto declaration. |
| 3927 | eraseGlobalSymbol(op: entry); |
| 3928 | entry->erase(); |
| 3929 | } |
| 3930 | |
| 3931 | if (d) |
| 3932 | setFunctionAttributes(globalDecl: gd, func: funcOp, /*isIncompleteFunction=*/false, isThunk); |
| 3933 | if (!extraAttrs.empty()) |
| 3934 | for (mlir::NamedAttribute attr : extraAttrs) |
| 3935 | if (!funcOp->hasDiscardableAttr(name: attr.getName())) |
| 3936 | funcOp->setDiscardableAttr(name: attr.getName(), value: attr.getValue()); |
| 3937 | |
| 3938 | // 'dontDefer' actually means don't move this to the deferredDeclsToEmit list. |
| 3939 | if (dontDefer) { |
| 3940 | // TODO(cir): This assertion will need an additional condition when we |
| 3941 | // support incomplete functions. |
| 3942 | assert(funcOp.getFunctionType() == funcType); |
| 3943 | return funcOp; |
| 3944 | } |
| 3945 | |
| 3946 | // All MSVC dtors other than the base dtor are linkonce_odr and delegate to |
| 3947 | // each other bottoming out wiht the base dtor. Therefore we emit non-base |
| 3948 | // dtors on usage, even if there is no dtor definition in the TU. |
| 3949 | if (isa_and_nonnull<CXXDestructorDecl>(Val: d) && |
| 3950 | getCXXABI().useThunkForDtorVariant(dtor: cast<CXXDestructorDecl>(Val: d), |
| 3951 | dt: gd.getDtorType())) |
| 3952 | errorNYI(d->getSourceRange(), "getOrCreateCIRFunction: dtor" ); |
| 3953 | |
| 3954 | // This is the first use or definition of a mangled name. If there is a |
| 3955 | // deferred decl with this name, remember that we need to emit it at the end |
| 3956 | // of the file. |
| 3957 | auto ddi = deferredDecls.find(x: mangledName); |
| 3958 | if (ddi != deferredDecls.end()) { |
| 3959 | // Move the potentially referenced deferred decl to the |
| 3960 | // DeferredDeclsToEmit list, and remove it from DeferredDecls (since we |
| 3961 | // don't need it anymore). |
| 3962 | addDeferredDeclToEmit(GD: ddi->second); |
| 3963 | deferredDecls.erase(position: ddi); |
| 3964 | |
| 3965 | // Otherwise, there are cases we have to worry about where we're using a |
| 3966 | // declaration for which we must emit a definition but where we might not |
| 3967 | // find a top-level definition. |
| 3968 | // - member functions defined inline in their classes |
| 3969 | // - friend functions defined inline in some class |
| 3970 | // - special member functions with implicit definitions |
| 3971 | // If we ever change our AST traversal to walk into class methods, this |
| 3972 | // will be unnecessary. |
| 3973 | // |
| 3974 | // We also don't emit a definition for a function if it's going to be an |
| 3975 | // entry in a vtable, unless it's already marked as used. |
| 3976 | } else if (getLangOpts().CPlusPlus && d) { |
| 3977 | // Look for a declaration that's lexically in a record. |
| 3978 | for (const auto *fd = cast<FunctionDecl>(Val: d)->getMostRecentDecl(); fd; |
| 3979 | fd = fd->getPreviousDecl()) { |
| 3980 | if (isa<CXXRecordDecl>(Val: fd->getLexicalDeclContext())) { |
| 3981 | if (fd->doesThisDeclarationHaveABody()) { |
| 3982 | addDeferredDeclToEmit(GD: gd.getWithDecl(D: fd)); |
| 3983 | break; |
| 3984 | } |
| 3985 | } |
| 3986 | } |
| 3987 | } |
| 3988 | |
| 3989 | return funcOp; |
| 3990 | } |
| 3991 | |
| 3992 | cir::FuncOp |
| 3993 | CIRGenModule::createCIRFunction(mlir::Location loc, StringRef name, |
| 3994 | cir::FuncType funcType, |
| 3995 | const clang::FunctionDecl *funcDecl) { |
| 3996 | cir::FuncOp func; |
| 3997 | { |
| 3998 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 3999 | |
| 4000 | // Functions always belong at module scope, but the ambient insertion |
| 4001 | // point may be inside another op's region, e.g. a thunk body or a |
| 4002 | // global's ctor region, so it cannot be used here. |
| 4003 | builder.setInsertionPointToEnd(getModule().getBody()); |
| 4004 | |
| 4005 | func = cir::FuncOp::create(builder, location: loc, sym_name: name, type: funcType); |
| 4006 | |
| 4007 | symbolLookupCache[func.getSymNameAttr()] = func; |
| 4008 | |
| 4009 | assert(!cir::MissingFeatures::opFuncAstDeclAttr()); |
| 4010 | |
| 4011 | if (funcDecl && !funcDecl->hasPrototype()) |
| 4012 | func.setNoProto(true); |
| 4013 | |
| 4014 | assert(func.isDeclaration() && "expected empty body" ); |
| 4015 | |
| 4016 | // A declaration gets private visibility by default, but external linkage |
| 4017 | // as the default linkage. |
| 4018 | func.setLinkageAttr(cir::GlobalLinkageKindAttr::get( |
| 4019 | context: &getMLIRContext(), value: cir::GlobalLinkageKind::ExternalLinkage)); |
| 4020 | mlir::SymbolTable::setSymbolVisibility( |
| 4021 | symbol: func, vis: mlir::SymbolTable::Visibility::Private); |
| 4022 | |
| 4023 | assert(!cir::MissingFeatures::opFuncExtraAttrs()); |
| 4024 | |
| 4025 | // Record the func_info tag, a C++ special member form or a known standard |
| 4026 | // library entity. |
| 4027 | setFuncInfoAttr(funcOp: func, funcDecl); |
| 4028 | |
| 4029 | if (funcDecl && this->getLangOpts().OpenACC) { |
| 4030 | // We only have to handle this attribute, since OpenACCAnnotAttrs are |
| 4031 | // handled via the end-of-TU work. |
| 4032 | for (const auto *attr : |
| 4033 | funcDecl->specific_attrs<OpenACCRoutineDeclAttr>()) |
| 4034 | emitOpenACCRoutineDecl(funcDecl, func, pragmaLoc: attr->getLocation(), |
| 4035 | clauses: attr->Clauses); |
| 4036 | } |
| 4037 | } |
| 4038 | return func; |
| 4039 | } |
| 4040 | |
| 4041 | cir::FuncOp |
| 4042 | CIRGenModule::createCIRBuiltinFunction(mlir::Location loc, StringRef name, |
| 4043 | cir::FuncType ty, |
| 4044 | const clang::FunctionDecl *fd) { |
| 4045 | cir::FuncOp fnOp = createCIRFunction(loc, name, funcType: ty, funcDecl: fd); |
| 4046 | fnOp.setBuiltin(true); |
| 4047 | return fnOp; |
| 4048 | } |
| 4049 | |
| 4050 | static cir::CtorKind getCtorKindFromDecl(const CXXConstructorDecl *ctor) { |
| 4051 | if (ctor->isDefaultConstructor()) |
| 4052 | return cir::CtorKind::Default; |
| 4053 | if (ctor->isCopyConstructor()) |
| 4054 | return cir::CtorKind::Copy; |
| 4055 | if (ctor->isMoveConstructor()) |
| 4056 | return cir::CtorKind::Move; |
| 4057 | return cir::CtorKind::Custom; |
| 4058 | } |
| 4059 | |
| 4060 | static cir::AssignKind getAssignKindFromDecl(const CXXMethodDecl *method) { |
| 4061 | if (method->isCopyAssignmentOperator()) |
| 4062 | return cir::AssignKind::Copy; |
| 4063 | if (method->isMoveAssignmentOperator()) |
| 4064 | return cir::AssignKind::Move; |
| 4065 | llvm_unreachable("not a copy or move assignment operator" ); |
| 4066 | } |
| 4067 | |
| 4068 | void CIRGenModule::setFuncInfoAttr(cir::FuncOp funcOp, |
| 4069 | const clang::FunctionDecl *funcDecl) { |
| 4070 | if (!funcDecl) |
| 4071 | return; |
| 4072 | |
| 4073 | if (const auto *dtor = dyn_cast<CXXDestructorDecl>(Val: funcDecl)) { |
| 4074 | auto cxxDtor = cir::CXXDtorAttr::get( |
| 4075 | type: convertType(type: getASTContext().getCanonicalTagType(TD: dtor->getParent())), |
| 4076 | isTrivial: dtor->isTrivial()); |
| 4077 | funcOp.setFuncInfoAttr(cxxDtor); |
| 4078 | return; |
| 4079 | } |
| 4080 | |
| 4081 | if (const auto *ctor = dyn_cast<CXXConstructorDecl>(Val: funcDecl)) { |
| 4082 | cir::CtorKind kind = getCtorKindFromDecl(ctor); |
| 4083 | auto cxxCtor = cir::CXXCtorAttr::get( |
| 4084 | type: convertType(type: getASTContext().getCanonicalTagType(TD: ctor->getParent())), |
| 4085 | ctorKind: kind, isTrivial: ctor->isTrivial()); |
| 4086 | funcOp.setFuncInfoAttr(cxxCtor); |
| 4087 | return; |
| 4088 | } |
| 4089 | |
| 4090 | const auto *method = dyn_cast<CXXMethodDecl>(Val: funcDecl); |
| 4091 | if (method && (method->isCopyAssignmentOperator() || |
| 4092 | method->isMoveAssignmentOperator())) { |
| 4093 | cir::AssignKind assignKind = getAssignKindFromDecl(method); |
| 4094 | auto cxxAssign = cir::CXXAssignAttr::get( |
| 4095 | type: convertType(type: getASTContext().getCanonicalTagType(TD: method->getParent())), |
| 4096 | assignKind, isTrivial: method->isTrivial()); |
| 4097 | funcOp.setFuncInfoAttr(cxxAssign); |
| 4098 | return; |
| 4099 | } |
| 4100 | |
| 4101 | // Otherwise tag a function that matches a known standard library entity. A |
| 4102 | // known entity is named by a plain identifier in std. For a member the |
| 4103 | // record decides std membership. Inline namespaces, like the versioning |
| 4104 | // namespace of libc++, count as part of std. |
| 4105 | if (!funcDecl->getIdentifier()) |
| 4106 | return; |
| 4107 | bool inStdNamespace = method ? method->getParent()->isInStdNamespace() |
| 4108 | : funcDecl->isInStdNamespace(); |
| 4109 | if (!inStdNamespace) |
| 4110 | return; |
| 4111 | |
| 4112 | // The names and the tags come from CIRStdOps.td, and the recognizer checks |
| 4113 | // the shape of each call. Only free functions name a known entity today, so |
| 4114 | // a member like char_traits::find never shares the tag of the free std::find. |
| 4115 | std::optional<cir::KnownFuncKind> kind; |
| 4116 | if (!method) { |
| 4117 | kind = llvm::StringSwitch<std::optional<cir::KnownFuncKind>>( |
| 4118 | funcDecl->getName()) |
| 4119 | .Case(S: cir::StdFindOp::getFunctionName(), |
| 4120 | Value: cir::StdFindOp::getFuncKind()) |
| 4121 | .Default(Value: std::nullopt); |
| 4122 | } |
| 4123 | if (!kind) |
| 4124 | return; |
| 4125 | |
| 4126 | funcOp.setFuncInfoAttr(cir::FuncIdentityAttr::get(context: &getMLIRContext(), kind: *kind)); |
| 4127 | } |
| 4128 | |
| 4129 | static void setWindowsItaniumDLLImport(CIRGenModule &cgm, bool isLocal, |
| 4130 | cir::FuncOp funcOp, StringRef name) { |
| 4131 | // In Windows Itanium environments, try to mark runtime functions |
| 4132 | // dllimport. For Mingw and MSVC, don't. We don't really know if the user |
| 4133 | // will link their standard library statically or dynamically. Marking |
| 4134 | // functions imported when they are not imported can cause linker errors |
| 4135 | // and warnings. |
| 4136 | if (!isLocal && cgm.getTarget().getTriple().isWindowsItaniumEnvironment() && |
| 4137 | !cgm.getCodeGenOpts().LTOVisibilityPublicStd) { |
| 4138 | assert(!cir::MissingFeatures::getRuntimeFunctionDecl()); |
| 4139 | assert(!cir::MissingFeatures::setDLLStorageClass()); |
| 4140 | assert(!cir::MissingFeatures::opGlobalDLLImportExport()); |
| 4141 | } |
| 4142 | } |
| 4143 | |
| 4144 | cir::FuncOp CIRGenModule::createRuntimeFunction(cir::FuncType ty, |
| 4145 | StringRef name, |
| 4146 | mlir::NamedAttrList , |
| 4147 | bool isLocal, |
| 4148 | bool assumeConvergent) { |
| 4149 | if (assumeConvergent) |
| 4150 | errorNYI(feature: "createRuntimeFunction: assumeConvergent" ); |
| 4151 | |
| 4152 | extraAttrs.set(name: cir::CIRDialect::getRuntimeFuncAttrName(), |
| 4153 | value: mlir::UnitAttr::get(context: &getMLIRContext())); |
| 4154 | cir::FuncOp entry = getOrCreateCIRFunction(mangledName: name, funcType: ty, gd: GlobalDecl(), |
| 4155 | /*forVtable=*/forVTable: false, extraAttrs); |
| 4156 | |
| 4157 | if (entry) { |
| 4158 | // TODO(cir): set the attributes of the function. |
| 4159 | assert(!cir::MissingFeatures::setLLVMFunctionFEnvAttributes()); |
| 4160 | assert(!cir::MissingFeatures::opFuncCallingConv()); |
| 4161 | setWindowsItaniumDLLImport(cgm&: *this, isLocal, funcOp: entry, name); |
| 4162 | entry.setDSOLocal(true); |
| 4163 | } |
| 4164 | |
| 4165 | return entry; |
| 4166 | } |
| 4167 | |
| 4168 | mlir::SymbolTable::Visibility |
| 4169 | CIRGenModule::getMLIRVisibility(cir::GlobalOp op) { |
| 4170 | // MLIR doesn't accept public symbols declarations (only |
| 4171 | // definitions). |
| 4172 | if (op.isDeclaration()) |
| 4173 | return mlir::SymbolTable::Visibility::Private; |
| 4174 | return getMLIRVisibilityFromCIRLinkage(GLK: op.getLinkage()); |
| 4175 | } |
| 4176 | |
| 4177 | mlir::SymbolTable::Visibility |
| 4178 | CIRGenModule::getMLIRVisibilityFromCIRLinkage(cir::GlobalLinkageKind glk) { |
| 4179 | switch (glk) { |
| 4180 | case cir::GlobalLinkageKind::InternalLinkage: |
| 4181 | case cir::GlobalLinkageKind::PrivateLinkage: |
| 4182 | return mlir::SymbolTable::Visibility::Private; |
| 4183 | case cir::GlobalLinkageKind::ExternalLinkage: |
| 4184 | case cir::GlobalLinkageKind::ExternalWeakLinkage: |
| 4185 | case cir::GlobalLinkageKind::LinkOnceODRLinkage: |
| 4186 | case cir::GlobalLinkageKind::AvailableExternallyLinkage: |
| 4187 | case cir::GlobalLinkageKind::CommonLinkage: |
| 4188 | case cir::GlobalLinkageKind::WeakAnyLinkage: |
| 4189 | case cir::GlobalLinkageKind::WeakODRLinkage: |
| 4190 | return mlir::SymbolTable::Visibility::Public; |
| 4191 | default: { |
| 4192 | llvm::errs() << "visibility not implemented for '" |
| 4193 | << stringifyGlobalLinkageKind(glk) << "'\n" ; |
| 4194 | assert(0 && "not implemented" ); |
| 4195 | } |
| 4196 | } |
| 4197 | llvm_unreachable("linkage should be handled above!" ); |
| 4198 | } |
| 4199 | |
| 4200 | std::string CIRGenModule::getModuleInitializerName(clang::Module *mod) { |
| 4201 | std::string name; |
| 4202 | llvm::raw_string_ostream out(name); |
| 4203 | cast<clang::ItaniumMangleContext>(Val&: getCXXABI().getMangleContext()) |
| 4204 | .mangleModuleInitializer(Module: mod, out); |
| 4205 | return name; |
| 4206 | } |
| 4207 | |
| 4208 | void CIRGenModule::release() { |
| 4209 | emitDeferred(); |
| 4210 | emitVTablesOpportunistically(); |
| 4211 | applyReplacements(); |
| 4212 | |
| 4213 | getModule()->setAttr(name: cir::CIRDialect::getModuleLevelAsmAttrName(), |
| 4214 | value: builder.getArrayAttr(value: globalScopeAsm)); |
| 4215 | |
| 4216 | emitGlobalAnnotations(); |
| 4217 | |
| 4218 | if (std::optional<cir::FramePointerKind> kind = |
| 4219 | getFramePointerKind(kind: codeGenOpts.getFramePointer())) |
| 4220 | getModule()->setAttr( |
| 4221 | name: cir::CIRDialect::getModuleFramePointerAttrName(), |
| 4222 | value: cir::FramePointerKindAttr::get(context: &getMLIRContext(), value: *kind)); |
| 4223 | |
| 4224 | if (!recordLayoutEntries.empty()) |
| 4225 | getModule()->setAttr( |
| 4226 | name: cir::CIRDialect::getRecordLayoutsAttrName(), |
| 4227 | value: mlir::DictionaryAttr::get(context: &getMLIRContext(), value: recordLayoutEntries)); |
| 4228 | |
| 4229 | if (getTriple().isAMDGPU() || |
| 4230 | (getTriple().isSPIRV() && getTriple().getVendor() == llvm::Triple::AMD)) |
| 4231 | emitAMDGPUMetadata(); |
| 4232 | |
| 4233 | if (getLangOpts().HIP) { |
| 4234 | // Emit a unique ID so that host and device binaries from the same |
| 4235 | // compilation unit can be associated. |
| 4236 | std::string cuidName = |
| 4237 | ("__hip_cuid_" + getASTContext().getCUIDHash()).str(); |
| 4238 | auto int8Ty = cir::IntType::get(context: &getMLIRContext(), width: 8, /*isSigned=*/is_signed: false); |
| 4239 | auto loc = builder.getUnknownLoc(); |
| 4240 | mlir::ptr::MemorySpaceAttrInterface addrSpace = |
| 4241 | getGlobalCIRAddressSpace(as: getGlobalVarAddressSpace(d: nullptr)); |
| 4242 | |
| 4243 | auto gv = createGlobalOp(loc, name: cuidName, t: int8Ty, |
| 4244 | /*isConstant=*/false, addrSpace); |
| 4245 | gv.setLinkage(cir::GlobalLinkageKind::ExternalLinkage); |
| 4246 | // Initialize with zero |
| 4247 | auto zeroAttr = cir::IntAttr::get(type: int8Ty, value: 0); |
| 4248 | gv.setInitialValueAttr(zeroAttr); |
| 4249 | // External linkage requires public visibility |
| 4250 | mlir::SymbolTable::setSymbolVisibility( |
| 4251 | symbol: gv, vis: mlir::SymbolTable::Visibility::Public); |
| 4252 | |
| 4253 | addCompilerUsedGlobal(gv); |
| 4254 | } |
| 4255 | |
| 4256 | if (astContext.getLangOpts().CUDA && cudaRuntime) |
| 4257 | getCUDARuntime().finalizeModule(); |
| 4258 | |
| 4259 | emitLLVMUsed(); |
| 4260 | |
| 4261 | // Precompute the mangled C++20 named-module initializer function name and |
| 4262 | // stash it on the ModuleOp so LoweringPrepare (which runs without a live |
| 4263 | // ASTContext) can read it back as an attribute. This attribute is the only |
| 4264 | // channel through which the named-module initializer reaches lowering: its |
| 4265 | // presence tells LoweringPrepare both what to call the global-init function |
| 4266 | // and that the function needs external linkage, and its absence selects the |
| 4267 | // `_GLOBAL__sub_I_` form. Lowering therefore never has to rediscover the |
| 4268 | // module from the AST. |
| 4269 | if (cxx20ModuleInits) { |
| 4270 | clang::Module *primary = astContext.getCurrentNamedModule(); |
| 4271 | bool isInterfaceUnit = primary && !primary->isModuleImplementation(); |
| 4272 | if (isInterfaceUnit) |
| 4273 | getModule()->setAttr( |
| 4274 | name: cir::CIRDialect::getCXXModuleInitFnNameAttrName(), |
| 4275 | value: builder.getStringAttr(bytes: getModuleInitializerName(mod: primary))); |
| 4276 | |
| 4277 | // The imported modules whose initializers the global init function of |
| 4278 | // this translation unit calls first, selected as in classic CodeGen. |
| 4279 | llvm::SmallVector<mlir::Attribute> importedInits; |
| 4280 | for (clang::Module *mod : CodeGenUtils::importedModulesToInitialize( |
| 4281 | Primary: primary, Imported: importedModules.getArrayRef())) { |
| 4282 | // A module that only the interface unit's export or import lists name |
| 4283 | // takes the location of the module declaration. |
| 4284 | auto importLoc = moduleImportLocs.find(Val: mod); |
| 4285 | mlir::Location loc = |
| 4286 | getLoc(cLoc: importLoc != moduleImportLocs.end() ? importLoc->second |
| 4287 | : primary->DefinitionLoc); |
| 4288 | importedInits.push_back(Elt: cir::CXXModuleInitAttr::get( |
| 4289 | context: &getMLIRContext(), |
| 4290 | name: builder.getStringAttr(bytes: getModuleInitializerName(mod)), loc)); |
| 4291 | } |
| 4292 | if (!importedInits.empty()) |
| 4293 | getModule()->setAttr(name: cir::CIRDialect::getCXXModuleImportedInitsAttrName(), |
| 4294 | value: builder.getArrayAttr(value: importedInits)); |
| 4295 | } |
| 4296 | |
| 4297 | // Autolink metadata for the imported modules is not emitted yet. |
| 4298 | assert(!cir::MissingFeatures::emitModuleLinkOptions()); |
| 4299 | |
| 4300 | // Classic codegen calls `checkAliases` here to validate any alias |
| 4301 | // definitions emitted during codegen. |
| 4302 | assert(!cir::MissingFeatures::checkAliases()); |
| 4303 | |
| 4304 | // There's a lot of code that is not implemented yet. |
| 4305 | assert(!cir::MissingFeatures::cgmRelease()); |
| 4306 | } |
| 4307 | |
| 4308 | void CIRGenModule::emitAliasDefinition(GlobalDecl gd) { |
| 4309 | const auto *d = cast<ValueDecl>(Val: gd.getDecl()); |
| 4310 | const AliasAttr *aa = d->getAttr<AliasAttr>(); |
| 4311 | assert(aa && "Not an alias?" ); |
| 4312 | |
| 4313 | StringRef mangledName = getMangledName(gd); |
| 4314 | |
| 4315 | if (aa->getAliasee() == mangledName) { |
| 4316 | diags.Report(Loc: aa->getLocation(), DiagID: diag::err_cyclic_alias) << 0; |
| 4317 | return; |
| 4318 | } |
| 4319 | |
| 4320 | // If there is a definition in the module, then it wins over the alias. |
| 4321 | // This is dubious, but allow it to be safe. Just ignore the alias. |
| 4322 | mlir::Operation *entry = getGlobalValue(name: mangledName); |
| 4323 | if (entry) { |
| 4324 | auto entryGV = mlir::dyn_cast<cir::CIRGlobalValueInterface>(Val: entry); |
| 4325 | if (entryGV && entryGV.isDefinition()) |
| 4326 | return; |
| 4327 | } |
| 4328 | |
| 4329 | // Classic codegen pushes the alias onto an `Aliases` list at this point so |
| 4330 | // that `checkAliases` can later validate the alias and recover on error. |
| 4331 | assert(!cir::MissingFeatures::checkAliases()); |
| 4332 | |
| 4333 | mlir::Location loc = getLoc(cRange: d->getSourceRange()); |
| 4334 | bool isFunction = isa<FunctionDecl>(Val: d); |
| 4335 | |
| 4336 | // Get the linkage and the type of the alias. |
| 4337 | mlir::Type declTy; |
| 4338 | cir::GlobalLinkageKind linkage; |
| 4339 | if (isFunction) { |
| 4340 | declTy = getTypes().getFunctionType(gd); |
| 4341 | linkage = getFunctionLinkage(gd); |
| 4342 | } else { |
| 4343 | declTy = getTypes().convertTypeForMem(d->getType()); |
| 4344 | const auto *vd = cast<VarDecl>(Val: d); |
| 4345 | linkage = getCIRLinkageVarDefinition(vd); |
| 4346 | } |
| 4347 | // |
| 4348 | // Create the alias op. |
| 4349 | // TODO(cir): Make GlobalAlias a separate op. |
| 4350 | cir::CIRGlobalValueInterface alias = |
| 4351 | isFunction ? mlir::cast<cir::CIRGlobalValueInterface>( |
| 4352 | Val: createCIRFunction(loc, name: mangledName, |
| 4353 | funcType: mlir::cast<cir::FuncType>(Val&: declTy), |
| 4354 | funcDecl: cast<FunctionDecl>(Val: d)) |
| 4355 | .getOperation()) |
| 4356 | : mlir::cast<cir::CIRGlobalValueInterface>( |
| 4357 | Val: createGlobalOp(loc, name: mangledName, t: declTy).getOperation()); |
| 4358 | |
| 4359 | // Create the alias op. If there is an existing declaration with the same |
| 4360 | // name, erase it: any references to it via flat symbol reference will |
| 4361 | // automatically resolve to the new alias. |
| 4362 | // However, function aliases actually change its type, so we have to replace |
| 4363 | // uses of it. |
| 4364 | if (entry) { |
| 4365 | if (isFunction) |
| 4366 | replaceUsesOfNonProtoTypeWithRealFunction( |
| 4367 | old: entry, newFn: mlir::cast<cir::FuncOp>(Val: alias.getOperation())); |
| 4368 | eraseGlobalSymbol(op: entry); |
| 4369 | entry->erase(); |
| 4370 | } |
| 4371 | |
| 4372 | // Aliases that target weak symbols must themselves be marked weak. |
| 4373 | if (d->hasAttr<WeakAttr>() || d->hasAttr<WeakRefAttr>() || |
| 4374 | d->isWeakImported()) |
| 4375 | linkage = cir::GlobalLinkageKind::WeakAnyLinkage; |
| 4376 | |
| 4377 | // Aliases are always definitions, so the MLIR visibility should match the |
| 4378 | // linkage rather than defaulting to private. |
| 4379 | mlir::SymbolTable::Visibility visibility = |
| 4380 | getMLIRVisibilityFromCIRLinkage(glk: linkage); |
| 4381 | |
| 4382 | alias.setAliasee(aa->getAliasee()); |
| 4383 | alias.setLinkage(linkage); |
| 4384 | mlir::SymbolTable::setSymbolVisibility(symbol: alias, vis: visibility); |
| 4385 | assert(!cir::MissingFeatures::opGlobalThreadLocal()); |
| 4386 | setCommonAttributes(gd, gv: alias); |
| 4387 | assert(!cir::MissingFeatures::generateDebugInfo()); |
| 4388 | } |
| 4389 | |
| 4390 | cir::FuncOp CIRGenModule::emitAliasForGlobal(StringRef mangledName, |
| 4391 | mlir::Operation *op, |
| 4392 | GlobalDecl aliasGD, |
| 4393 | cir::FuncOp aliasee, |
| 4394 | cir::GlobalLinkageKind linkage) { |
| 4395 | |
| 4396 | auto *aliasFD = dyn_cast<FunctionDecl>(Val: aliasGD.getDecl()); |
| 4397 | assert(aliasFD && "expected FunctionDecl" ); |
| 4398 | |
| 4399 | // The aliasee function type is different from the alias one, this difference |
| 4400 | // is specific to CIR because in LLVM the ptr types are already erased at this |
| 4401 | // point. |
| 4402 | const CIRGenFunctionInfo &fnInfo = |
| 4403 | getTypes().arrangeCXXStructorDeclaration(gd: aliasGD); |
| 4404 | cir::FuncType fnType = getTypes().getFunctionType(info: fnInfo); |
| 4405 | |
| 4406 | cir::FuncOp alias = |
| 4407 | createCIRFunction(loc: getLoc(cRange: aliasGD.getDecl()->getSourceRange()), |
| 4408 | name: mangledName, funcType: fnType, funcDecl: aliasFD); |
| 4409 | alias.setAliasee(aliasee.getName()); |
| 4410 | alias.setLinkage(linkage); |
| 4411 | // Declarations cannot have public MLIR visibility, just mark them private |
| 4412 | // but this really should have no meaning since CIR should not be using |
| 4413 | // this information to derive linkage information. |
| 4414 | mlir::SymbolTable::setSymbolVisibility( |
| 4415 | symbol: alias, vis: mlir::SymbolTable::Visibility::Private); |
| 4416 | |
| 4417 | if (op) { |
| 4418 | // Any existing users of the existing function declaration will be |
| 4419 | // referencing the function by flat symbol reference (i.e. the name), so |
| 4420 | // those uses will automatically resolve to the alias now that we've |
| 4421 | // replaced the function declaration. We can safely erase the existing |
| 4422 | // function declaration. |
| 4423 | assert(cast<cir::FuncOp>(op).getFunctionType() == alias.getFunctionType() && |
| 4424 | "declaration exists with different type" ); |
| 4425 | eraseGlobalSymbol(op); |
| 4426 | op->erase(); |
| 4427 | } else { |
| 4428 | // Name already set by createCIRFunction |
| 4429 | } |
| 4430 | |
| 4431 | // Finally, set up the alias with its proper name and attributes. |
| 4432 | setCommonAttributes(gd: aliasGD, gv: alias); |
| 4433 | return alias; |
| 4434 | } |
| 4435 | |
| 4436 | mlir::Type CIRGenModule::convertType(QualType type) { |
| 4437 | return genTypes.convertType(type); |
| 4438 | } |
| 4439 | |
| 4440 | bool CIRGenModule::verifyModule() const { |
| 4441 | // Verify the module after we have finished constructing it, this will |
| 4442 | // check the structural properties of the IR and invoke any specific |
| 4443 | // verifiers we have on the CIR operations. |
| 4444 | return mlir::verify(op: getModule()).succeeded(); |
| 4445 | } |
| 4446 | |
| 4447 | mlir::Attribute CIRGenModule::getAddrOfRTTIDescriptor(mlir::Location loc, |
| 4448 | QualType ty, bool forEh) { |
| 4449 | // Return a bogus pointer if RTTI is disabled, unless it's for EH. |
| 4450 | // FIXME: should we even be calling this method if RTTI is disabled |
| 4451 | // and it's not for EH? |
| 4452 | if (!shouldEmitRTTI(forEH: forEh)) |
| 4453 | return builder.getConstNullPtrAttr(t: builder.getUInt8PtrTy()); |
| 4454 | |
| 4455 | if (forEh && ty->isObjCObjectPointerType() && |
| 4456 | langOpts.ObjCRuntime.isGNUFamily()) { |
| 4457 | errorNYI(loc, feature: "getAddrOfRTTIDescriptor: Objc PtrType & Objc RT GUN" ); |
| 4458 | return {}; |
| 4459 | } |
| 4460 | |
| 4461 | return getCXXABI().getAddrOfRTTIDescriptor(loc, ty); |
| 4462 | } |
| 4463 | |
| 4464 | // TODO(cir): this can be shared with LLVM codegen. |
| 4465 | CharUnits CIRGenModule::computeNonVirtualBaseClassOffset( |
| 4466 | const CXXRecordDecl *derivedClass, |
| 4467 | llvm::iterator_range<CastExpr::path_const_iterator> path) { |
| 4468 | CharUnits offset = CharUnits::Zero(); |
| 4469 | |
| 4470 | const ASTContext &astContext = getASTContext(); |
| 4471 | const CXXRecordDecl *rd = derivedClass; |
| 4472 | |
| 4473 | for (const CXXBaseSpecifier *base : path) { |
| 4474 | assert(!base->isVirtual() && "Should not see virtual bases here!" ); |
| 4475 | |
| 4476 | // Get the layout. |
| 4477 | const ASTRecordLayout &layout = astContext.getASTRecordLayout(D: rd); |
| 4478 | |
| 4479 | const auto *baseDecl = base->getType()->castAsCXXRecordDecl(); |
| 4480 | |
| 4481 | // Add the offset. |
| 4482 | offset += layout.getBaseClassOffset(Base: baseDecl); |
| 4483 | |
| 4484 | rd = baseDecl; |
| 4485 | } |
| 4486 | |
| 4487 | return offset; |
| 4488 | } |
| 4489 | |
| 4490 | DiagnosticBuilder CIRGenModule::errorNYI(SourceLocation loc, |
| 4491 | llvm::StringRef feature) { |
| 4492 | unsigned diagID = diags.getCustomDiagID( |
| 4493 | L: DiagnosticsEngine::Error, FormatString: "ClangIR code gen Not Yet Implemented: %0" ); |
| 4494 | return diags.Report(Loc: loc, DiagID: diagID) << feature; |
| 4495 | } |
| 4496 | |
| 4497 | DiagnosticBuilder CIRGenModule::errorNYI(SourceRange loc, |
| 4498 | llvm::StringRef feature) { |
| 4499 | return errorNYI(loc: loc.getBegin(), feature) << loc; |
| 4500 | } |
| 4501 | |
| 4502 | void CIRGenModule::error(SourceLocation loc, StringRef error) { |
| 4503 | unsigned diagID = getDiags().getCustomDiagID(L: DiagnosticsEngine::Error, FormatString: "%0" ); |
| 4504 | getDiags().Report(Loc: astContext.getFullLoc(Loc: loc), DiagID: diagID) << error; |
| 4505 | } |
| 4506 | |
| 4507 | /// Print out an error that codegen doesn't support the specified stmt yet. |
| 4508 | void CIRGenModule::errorUnsupported(const Stmt *s, llvm::StringRef type) { |
| 4509 | unsigned diagId = diags.getCustomDiagID(L: DiagnosticsEngine::Error, |
| 4510 | FormatString: "cannot compile this %0 yet" ); |
| 4511 | diags.Report(Loc: astContext.getFullLoc(Loc: s->getBeginLoc()), DiagID: diagId) |
| 4512 | << type << s->getSourceRange(); |
| 4513 | } |
| 4514 | |
| 4515 | /// Print out an error that codegen doesn't support the specified decl yet. |
| 4516 | void CIRGenModule::errorUnsupported(const Decl *d, llvm::StringRef type) { |
| 4517 | unsigned diagId = diags.getCustomDiagID(L: DiagnosticsEngine::Error, |
| 4518 | FormatString: "cannot compile this %0 yet" ); |
| 4519 | diags.Report(Loc: astContext.getFullLoc(Loc: d->getLocation()), DiagID: diagId) << type; |
| 4520 | } |
| 4521 | |
| 4522 | mlir::Operation * |
| 4523 | CIRGenModule::getAddrOfGlobalTemporary(const MaterializeTemporaryExpr *mte, |
| 4524 | const Expr *init) { |
| 4525 | assert((mte->getStorageDuration() == SD_Static || |
| 4526 | mte->getStorageDuration() == SD_Thread) && |
| 4527 | "not a global temporary" ); |
| 4528 | const auto *varDecl = cast<VarDecl>(Val: mte->getExtendingDecl()); |
| 4529 | |
| 4530 | // Use the MaterializeTemporaryExpr's type if it has the same unqualified |
| 4531 | // base type as Init. This preserves cv-qualifiers (e.g. const from a |
| 4532 | // constexpr or const-ref binding) that skipRValueSubobjectAdjustments may |
| 4533 | // have dropped via NoOp casts, while correctly falling back to Init's type |
| 4534 | // when a real subobject adjustment changed the type (e.g. member access or |
| 4535 | // base-class cast in C++98), where E->getType() reflects the reference type, |
| 4536 | // not the actual storage type. |
| 4537 | QualType materializedType = init->getType(); |
| 4538 | if (getASTContext().hasSameUnqualifiedType(T1: mte->getType(), T2: materializedType)) |
| 4539 | materializedType = mte->getType(); |
| 4540 | |
| 4541 | CharUnits align = getASTContext().getTypeAlignInChars(T: materializedType); |
| 4542 | mlir::Location loc = getLoc(cRange: mte->getSourceRange()); |
| 4543 | |
| 4544 | // FIXME: If an externally-visible declaration extends multiple temporaries, |
| 4545 | // we need to give each temporary the same name in every translation unit (and |
| 4546 | // we also need to make the temporaries externally-visible). |
| 4547 | llvm::SmallString<256> name; |
| 4548 | llvm::raw_svector_ostream out(name); |
| 4549 | getCXXABI().getMangleContext().mangleReferenceTemporary( |
| 4550 | D: varDecl, ManglingNumber: mte->getManglingNumber(), out); |
| 4551 | |
| 4552 | auto insertResult = materializedGlobalTemporaryMap.insert(KV: {mte, nullptr}); |
| 4553 | if (!insertResult.second) { |
| 4554 | mlir::Type type = getTypes().convertTypeForMem(materializedType); |
| 4555 | // We've seen this before: either we already created it or we're in the |
| 4556 | // process of doing so. |
| 4557 | if (!insertResult.first->second) { |
| 4558 | // We recursively re-entered this function, probably during emission of |
| 4559 | // the initializer. Create a placeholder. |
| 4560 | insertResult.first->second = |
| 4561 | createGlobalOp(loc, name, t: type, /*isConstant=*/false); |
| 4562 | } |
| 4563 | return insertResult.first->second; |
| 4564 | } |
| 4565 | |
| 4566 | APValue *value = nullptr; |
| 4567 | if (mte->getStorageDuration() == SD_Static && varDecl->evaluateValue()) { |
| 4568 | // If the initializer of the extending declaration is a constant |
| 4569 | // initializer, we should have a cached constant initializer for this |
| 4570 | // temporay. Note taht this m ight have a different value from the value |
| 4571 | // computed by evaluating the initializer if the surrounding constant |
| 4572 | // expression modifies the temporary. |
| 4573 | value = mte->getOrCreateValue(/*MayCreate=*/false); |
| 4574 | } |
| 4575 | |
| 4576 | // Try evaluating it now, it might have a constant initializer |
| 4577 | Expr::EvalResult evalResult; |
| 4578 | if (!value && init->EvaluateAsRValue(Result&: evalResult, Ctx: getASTContext()) && |
| 4579 | !evalResult.hasSideEffects()) |
| 4580 | value = &evalResult.Val; |
| 4581 | |
| 4582 | assert(!cir::MissingFeatures::addressSpace()); |
| 4583 | |
| 4584 | std::optional<ConstantEmitter> emitter; |
| 4585 | mlir::Attribute initialValue = nullptr; |
| 4586 | bool isConstant = false; |
| 4587 | mlir::Type type; |
| 4588 | |
| 4589 | if (value) { |
| 4590 | emitter.emplace(args&: *this); |
| 4591 | initialValue = emitter->emitForInitializer(value: *value, destType: materializedType); |
| 4592 | |
| 4593 | isConstant = materializedType.isConstantStorage( |
| 4594 | Ctx: getASTContext(), /*ExcludeCtor=*/value, /*ExcludeDtor=*/false); |
| 4595 | |
| 4596 | type = mlir::cast<mlir::TypedAttr>(Val&: initialValue).getType(); |
| 4597 | } else { |
| 4598 | // No initializer, the initialization will be provided when we initialize |
| 4599 | // the declaration which performed lifetime extension. |
| 4600 | type = getTypes().convertTypeForMem(materializedType); |
| 4601 | } |
| 4602 | |
| 4603 | // Create a global variable for this lifetime-extended temporary. |
| 4604 | cir::GlobalLinkageKind linkage = getCIRLinkageVarDefinition(vd: varDecl); |
| 4605 | if (linkage == cir::GlobalLinkageKind::ExternalLinkage) { |
| 4606 | const VarDecl *initVD; |
| 4607 | if (varDecl->isStaticDataMember() && varDecl->getAnyInitializer(D&: initVD) && |
| 4608 | isa<CXXRecordDecl>(Val: initVD->getLexicalDeclContext())) { |
| 4609 | // Temporaries defined inside a class get linkonce_odr linkage because the |
| 4610 | // calss can be defined in multiple translation units. |
| 4611 | errorNYI(loc: mte->getSourceRange(), feature: "static data member initialization" ); |
| 4612 | } else { |
| 4613 | // There is no need for this temporary to have external linkage if the |
| 4614 | // VarDecl has external linkage. |
| 4615 | linkage = cir::GlobalLinkageKind::InternalLinkage; |
| 4616 | } |
| 4617 | } |
| 4618 | cir::GlobalOp gv = createGlobalOp(loc, name, t: type, isConstant); |
| 4619 | gv.setInitialValueAttr(initialValue); |
| 4620 | gv.setLinkage(linkage); |
| 4621 | gv.setVisibility(getMLIRVisibilityFromCIRLinkage(glk: linkage)); |
| 4622 | |
| 4623 | if (emitter) |
| 4624 | emitter->finalize(gv); |
| 4625 | // Don't assign dllimport or dllexport to local linkage globals |
| 4626 | if (!gv.hasLocalLinkage()) { |
| 4627 | setGVProperties(op: gv, d: varDecl); |
| 4628 | assert(!cir::MissingFeatures::setDLLStorageClass()); |
| 4629 | } |
| 4630 | |
| 4631 | gv.setAlignment(align.getAsAlign().value()); |
| 4632 | if (supportsCOMDAT() && gv.isWeakForLinker()) |
| 4633 | gv.setSelfComdat(); |
| 4634 | if (varDecl->getTLSKind()) |
| 4635 | setTLSMode(op: gv, d: *varDecl, /*isExtendingDecl=*/true); |
| 4636 | mlir::Operation *cv = gv; |
| 4637 | |
| 4638 | assert(!cir::MissingFeatures::addressSpace()); |
| 4639 | |
| 4640 | // Update the map with the new temporary. If we created a placeholder above, |
| 4641 | // erase it as well, the name will have been the same, so our symbol |
| 4642 | // references would have been correct. We still do a 'replaceAllUsesWith' in |
| 4643 | // case some sort of expression formed a reference to the placeholder |
| 4644 | // temporary. |
| 4645 | mlir::Operation *&entry = materializedGlobalTemporaryMap[mte]; |
| 4646 | if (entry) { |
| 4647 | entry->replaceAllUsesWith(values&: cv); |
| 4648 | eraseGlobalSymbol(op: entry); |
| 4649 | entry->erase(); |
| 4650 | } |
| 4651 | entry = cv; |
| 4652 | |
| 4653 | return cv; |
| 4654 | } |
| 4655 | |
| 4656 | cir::GlobalOp CIRGenModule::getAddrOfUnnamedGlobalConstantDecl( |
| 4657 | const UnnamedGlobalConstantDecl *gcd) { |
| 4658 | unsigned numEntries = unnamedGlobalConstantDeclMap.size(); |
| 4659 | cir::GlobalOp *globalOpEntry = &unnamedGlobalConstantDeclMap[gcd]; |
| 4660 | |
| 4661 | if (*globalOpEntry) |
| 4662 | return *globalOpEntry; |
| 4663 | |
| 4664 | ConstantEmitter emitter(*this); |
| 4665 | |
| 4666 | const APValue &value = gcd->getValue(); |
| 4667 | assert(!value.isAbsent()); |
| 4668 | assert(!cir::MissingFeatures::addressSpace() && |
| 4669 | "emitForInitializer should take gcd->getType().getAddressSpace()" ); |
| 4670 | mlir::Attribute init = emitter.emitForInitializer(value, destType: gcd->getType()); |
| 4671 | auto typedInit = dyn_cast<mlir::TypedAttr>(Val&: init); |
| 4672 | |
| 4673 | if (!typedInit) |
| 4674 | errorNYI(loc: gcd->getSourceRange(), |
| 4675 | feature: "getAddrOfUnnamedGlobalConstantDecl: non-typed initializer" ); |
| 4676 | |
| 4677 | assert(!cir::MissingFeatures::addressSpace()); |
| 4678 | |
| 4679 | // Classic codegen always creates these with .constant, then counts on the |
| 4680 | // auto-addition of '.#'. CIR global doesn't have this, so we'll just auto-add |
| 4681 | // one if this isn't the first. We could probably choose a better name than |
| 4682 | // .constant to be unique for this type of decl, but this is consistent with |
| 4683 | // classic codegen. |
| 4684 | std::string name = numEntries == 0 |
| 4685 | ? ".constant" |
| 4686 | : (Twine(".constant." ) + Twine(numEntries)).str(); |
| 4687 | auto globalOp = createGlobalOp(loc: builder.getUnknownLoc(), name, |
| 4688 | t: typedInit.getType(), /*is_constant=*/isConstant: true); |
| 4689 | globalOp.setLinkage(cir::GlobalLinkageKind::PrivateLinkage); |
| 4690 | |
| 4691 | CharUnits alignment = getASTContext().getTypeAlignInChars(T: gcd->getType()); |
| 4692 | globalOp.setAlignment(alignment.getAsAlign().value()); |
| 4693 | globalOp.setUnnamedAddr(true); |
| 4694 | CIRGenModule::setInitializer(op&: globalOp, value: init); |
| 4695 | |
| 4696 | emitter.finalize(gv: globalOp); |
| 4697 | *globalOpEntry = globalOp; |
| 4698 | return globalOp; |
| 4699 | } |
| 4700 | |
| 4701 | cir::GlobalOp |
| 4702 | CIRGenModule::getAddrOfTemplateParamObject(const TemplateParamObjectDecl *tpo) { |
| 4703 | StringRef name = getMangledName(gd: tpo); |
| 4704 | CharUnits alignment = getNaturalTypeAlignment(t: tpo->getType()); |
| 4705 | |
| 4706 | if (auto globalOp = |
| 4707 | mlir::dyn_cast_or_null<cir::GlobalOp>(Val: getGlobalValue(name))) |
| 4708 | return globalOp; |
| 4709 | |
| 4710 | ConstantEmitter emitter(*this); |
| 4711 | assert(!cir::MissingFeatures::addressSpace() && |
| 4712 | "emitForInitializer should take tpo->getType().getAddressSpace()" ); |
| 4713 | mlir::Attribute init = |
| 4714 | emitter.emitForInitializer(value: tpo->getValue(), destType: tpo->getType()); |
| 4715 | |
| 4716 | if (!init) { |
| 4717 | errorUnsupported(d: tpo, type: "template parameter object" ); |
| 4718 | return {}; |
| 4719 | } |
| 4720 | |
| 4721 | mlir::TypedAttr typedInit = cast<mlir::TypedAttr>(Val&: init); |
| 4722 | |
| 4723 | cir::GlobalLinkageKind linkage = |
| 4724 | isExternallyVisible(L: tpo->getLinkageAndVisibility().getLinkage()) |
| 4725 | ? cir::GlobalLinkageKind::LinkOnceODRLinkage |
| 4726 | : cir::GlobalLinkageKind::InternalLinkage; |
| 4727 | |
| 4728 | assert(!cir::MissingFeatures::addressSpace()); |
| 4729 | auto globalOp = createGlobalOp(loc: builder.getUnknownLoc(), name, |
| 4730 | t: typedInit.getType(), /*is_constant=*/isConstant: true); |
| 4731 | globalOp.setLinkage(linkage); |
| 4732 | globalOp.setAlignment(alignment.getAsAlign().value()); |
| 4733 | if (supportsCOMDAT() && linkage == cir::GlobalLinkageKind::LinkOnceODRLinkage) |
| 4734 | globalOp.setSelfComdat(); |
| 4735 | |
| 4736 | CIRGenModule::setInitializer(op&: globalOp, value: init); |
| 4737 | emitter.finalize(gv: globalOp); |
| 4738 | |
| 4739 | insertGlobalSymbol(op: globalOp); |
| 4740 | |
| 4741 | return globalOp; |
| 4742 | } |
| 4743 | |
| 4744 | //===----------------------------------------------------------------------===// |
| 4745 | // Annotations |
| 4746 | //===----------------------------------------------------------------------===// |
| 4747 | |
| 4748 | mlir::ArrayAttr |
| 4749 | CIRGenModule::getOrCreateAnnotationArgs(const AnnotateAttr *attr) { |
| 4750 | ArrayRef<Expr *> exprs = {attr->args_begin(), attr->args_size()}; |
| 4751 | // Return a null attr for no-args annotations so OptionalParameter omits |
| 4752 | // the args portion entirely from the printed IR. |
| 4753 | if (exprs.empty()) |
| 4754 | return {}; |
| 4755 | |
| 4756 | llvm::FoldingSetNodeID id; |
| 4757 | for (Expr *e : exprs) |
| 4758 | id.Add(x: cast<clang::ConstantExpr>(Val: e)->getAPValueResult()); |
| 4759 | |
| 4760 | mlir::ArrayAttr &lookup = annotationArgs[id.computeHash()]; |
| 4761 | if (lookup) |
| 4762 | return lookup; |
| 4763 | |
| 4764 | llvm::SmallVector<mlir::Attribute> args; |
| 4765 | args.reserve(N: exprs.size()); |
| 4766 | for (Expr *e : exprs) { |
| 4767 | if (auto *strE = dyn_cast<clang::StringLiteral>(Val: e->IgnoreParenCasts())) { |
| 4768 | args.push_back(Elt: builder.getStringAttr(bytes: strE->getString())); |
| 4769 | } else if (auto *intE = |
| 4770 | dyn_cast<clang::IntegerLiteral>(Val: e->IgnoreParenCasts())) { |
| 4771 | auto intTy = builder.getIntegerType(width: intE->getValue().getBitWidth()); |
| 4772 | args.push_back(Elt: builder.getIntegerAttr(type: intTy, value: intE->getValue())); |
| 4773 | } else { |
| 4774 | errorNYI(loc: e->getExprLoc(), feature: "annotation argument expression" ); |
| 4775 | } |
| 4776 | } |
| 4777 | |
| 4778 | return lookup = builder.getArrayAttr(value: args); |
| 4779 | } |
| 4780 | |
| 4781 | cir::AnnotationAttr CIRGenModule::emitAnnotateAttr(const AnnotateAttr *aa) { |
| 4782 | mlir::StringAttr annoGV = builder.getStringAttr(bytes: aa->getAnnotation()); |
| 4783 | mlir::ArrayAttr args = getOrCreateAnnotationArgs(attr: aa); |
| 4784 | return cir::AnnotationAttr::get(context: &getMLIRContext(), name: annoGV, args); |
| 4785 | } |
| 4786 | |
| 4787 | void CIRGenModule::addGlobalAnnotations(const ValueDecl *d, |
| 4788 | mlir::Operation *gv) { |
| 4789 | assert(d->hasAttr<AnnotateAttr>() && "no annotate attribute" ); |
| 4790 | assert((isa<cir::GlobalOp>(gv) || isa<cir::FuncOp>(gv)) && |
| 4791 | "annotation only on globals" ); |
| 4792 | llvm::SmallVector<mlir::Attribute> annotations; |
| 4793 | for (const auto *i : d->specific_attrs<AnnotateAttr>()) |
| 4794 | annotations.push_back(Elt: emitAnnotateAttr(aa: i)); |
| 4795 | if (auto global = dyn_cast<cir::GlobalOp>(Val: gv)) |
| 4796 | global.setAnnotationsAttr(builder.getArrayAttr(value: annotations)); |
| 4797 | else if (auto func = dyn_cast<cir::FuncOp>(Val: gv)) |
| 4798 | func.setAnnotationsAttr(builder.getArrayAttr(value: annotations)); |
| 4799 | } |
| 4800 | |
| 4801 | void CIRGenModule::emitGlobalAnnotations() { |
| 4802 | for (const auto &[mangledName, vd] : deferredAnnotations) { |
| 4803 | mlir::Operation *gv = getGlobalValue(name: mangledName); |
| 4804 | if (gv) |
| 4805 | addGlobalAnnotations(d: vd, gv); |
| 4806 | } |
| 4807 | deferredAnnotations.clear(); |
| 4808 | } |
| 4809 | |