| 1 | //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file contains support for writing Microsoft CodeView debug info. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "CodeViewDebug.h" |
| 14 | #include "llvm/ADT/APSInt.h" |
| 15 | #include "llvm/ADT/STLExtras.h" |
| 16 | #include "llvm/ADT/SmallBitVector.h" |
| 17 | #include "llvm/ADT/SmallString.h" |
| 18 | #include "llvm/ADT/StringRef.h" |
| 19 | #include "llvm/ADT/TinyPtrVector.h" |
| 20 | #include "llvm/ADT/Twine.h" |
| 21 | #include "llvm/BinaryFormat/COFF.h" |
| 22 | #include "llvm/BinaryFormat/Dwarf.h" |
| 23 | #include "llvm/CodeGen/AsmPrinter.h" |
| 24 | #include "llvm/CodeGen/LexicalScopes.h" |
| 25 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 26 | #include "llvm/CodeGen/MachineFunction.h" |
| 27 | #include "llvm/CodeGen/MachineInstr.h" |
| 28 | #include "llvm/CodeGen/MachineModuleInfo.h" |
| 29 | #include "llvm/CodeGen/TargetFrameLowering.h" |
| 30 | #include "llvm/CodeGen/TargetLowering.h" |
| 31 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 32 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 33 | #include "llvm/Config/llvm-config.h" |
| 34 | #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" |
| 35 | #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" |
| 36 | #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" |
| 37 | #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" |
| 38 | #include "llvm/DebugInfo/CodeView/EnumTables.h" |
| 39 | #include "llvm/DebugInfo/CodeView/Line.h" |
| 40 | #include "llvm/DebugInfo/CodeView/SymbolRecord.h" |
| 41 | #include "llvm/DebugInfo/CodeView/TypeRecord.h" |
| 42 | #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" |
| 43 | #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" |
| 44 | #include "llvm/IR/Constants.h" |
| 45 | #include "llvm/IR/DataLayout.h" |
| 46 | #include "llvm/IR/DebugInfoMetadata.h" |
| 47 | #include "llvm/IR/Function.h" |
| 48 | #include "llvm/IR/GlobalValue.h" |
| 49 | #include "llvm/IR/GlobalVariable.h" |
| 50 | #include "llvm/IR/Metadata.h" |
| 51 | #include "llvm/IR/Module.h" |
| 52 | #include "llvm/MC/MCAsmInfo.h" |
| 53 | #include "llvm/MC/MCContext.h" |
| 54 | #include "llvm/MC/MCSectionCOFF.h" |
| 55 | #include "llvm/MC/MCStreamer.h" |
| 56 | #include "llvm/MC/MCSymbol.h" |
| 57 | #include "llvm/Support/BinaryStreamWriter.h" |
| 58 | #include "llvm/Support/Casting.h" |
| 59 | #include "llvm/Support/Error.h" |
| 60 | #include "llvm/Support/ErrorHandling.h" |
| 61 | #include "llvm/Support/FormatVariadic.h" |
| 62 | #include "llvm/Support/Path.h" |
| 63 | #include "llvm/Support/SMLoc.h" |
| 64 | #include "llvm/Support/ScopedPrinter.h" |
| 65 | #include "llvm/Target/TargetLoweringObjectFile.h" |
| 66 | #include "llvm/Target/TargetMachine.h" |
| 67 | #include "llvm/TargetParser/Triple.h" |
| 68 | #include <algorithm> |
| 69 | #include <cassert> |
| 70 | #include <cctype> |
| 71 | #include <cstddef> |
| 72 | #include <limits> |
| 73 | |
| 74 | using namespace llvm; |
| 75 | using namespace llvm::codeview; |
| 76 | |
| 77 | namespace { |
| 78 | class CVMCAdapter : public CodeViewRecordStreamer { |
| 79 | public: |
| 80 | CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) |
| 81 | : OS(&OS), TypeTable(TypeTable) {} |
| 82 | |
| 83 | void emitBytes(StringRef Data) override { OS->emitBytes(Data); } |
| 84 | |
| 85 | void emitIntValue(uint64_t Value, unsigned Size) override { |
| 86 | OS->emitIntValueInHex(Value, Size); |
| 87 | } |
| 88 | |
| 89 | void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } |
| 90 | |
| 91 | void (const Twine &T) override { OS->AddComment(T); } |
| 92 | |
| 93 | void (const Twine &T) override { OS->emitRawComment(T); } |
| 94 | |
| 95 | bool isVerboseAsm() override { return OS->isVerboseAsm(); } |
| 96 | |
| 97 | std::string getTypeName(TypeIndex TI) override { |
| 98 | std::string TypeName; |
| 99 | if (!TI.isNoneType()) { |
| 100 | if (TI.isSimple()) |
| 101 | TypeName = std::string(TypeIndex::simpleTypeName(TI)); |
| 102 | else |
| 103 | TypeName = std::string(TypeTable.getTypeName(Index: TI)); |
| 104 | } |
| 105 | return TypeName; |
| 106 | } |
| 107 | |
| 108 | private: |
| 109 | MCStreamer *OS = nullptr; |
| 110 | TypeCollection &TypeTable; |
| 111 | }; |
| 112 | } // namespace |
| 113 | |
| 114 | static CPUType mapArchToCVCPUType(Triple::ArchType Type) { |
| 115 | switch (Type) { |
| 116 | case Triple::ArchType::x86: |
| 117 | return CPUType::Pentium3; |
| 118 | case Triple::ArchType::x86_64: |
| 119 | return CPUType::X64; |
| 120 | case Triple::ArchType::thumb: |
| 121 | // LLVM currently doesn't support Windows CE and so thumb |
| 122 | // here is indiscriminately mapped to ARMNT specifically. |
| 123 | return CPUType::ARMNT; |
| 124 | case Triple::ArchType::aarch64: |
| 125 | return CPUType::ARM64; |
| 126 | case Triple::ArchType::mipsel: |
| 127 | return CPUType::MIPS; |
| 128 | case Triple::ArchType::UnknownArch: |
| 129 | return CPUType::Unknown; |
| 130 | default: |
| 131 | report_fatal_error(reason: "target architecture doesn't map to a CodeView CPUType" ); |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | CodeViewDebug::CodeViewDebug(AsmPrinter *AP) |
| 136 | : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} |
| 137 | |
| 138 | StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { |
| 139 | std::string &Filepath = FileToFilepathMap[File]; |
| 140 | if (!Filepath.empty()) |
| 141 | return Filepath; |
| 142 | |
| 143 | StringRef Dir = File->getDirectory(), Filename = File->getFilename(); |
| 144 | |
| 145 | // If this is a Unix-style path, just use it as is. Don't try to canonicalize |
| 146 | // it textually because one of the path components could be a symlink. |
| 147 | if (Dir.starts_with(Prefix: "/" ) || Filename.starts_with(Prefix: "/" )) { |
| 148 | if (llvm::sys::path::is_absolute(path: Filename, style: llvm::sys::path::Style::posix)) |
| 149 | return Filename; |
| 150 | Filepath = std::string(Dir); |
| 151 | if (Dir.back() != '/') |
| 152 | Filepath += '/'; |
| 153 | Filepath += Filename; |
| 154 | return Filepath; |
| 155 | } |
| 156 | |
| 157 | // Clang emits directory and relative filename info into the IR, but CodeView |
| 158 | // operates on full paths. We could change Clang to emit full paths too, but |
| 159 | // that would increase the IR size and probably not needed for other users. |
| 160 | // For now, just concatenate and canonicalize the path here. |
| 161 | if (Filename.find(C: ':') == 1) |
| 162 | Filepath = std::string(Filename); |
| 163 | else |
| 164 | Filepath = (Dir + "\\" + Filename).str(); |
| 165 | |
| 166 | // Canonicalize the path. We have to do it textually because we may no longer |
| 167 | // have access the file in the filesystem. |
| 168 | // First, replace all slashes with backslashes. |
| 169 | llvm::replace(Range&: Filepath, OldValue: '/', NewValue: '\\'); |
| 170 | |
| 171 | // Remove all "\.\" with "\". |
| 172 | size_t Cursor = 0; |
| 173 | while ((Cursor = Filepath.find(s: "\\.\\" , pos: Cursor)) != std::string::npos) |
| 174 | Filepath.erase(pos: Cursor, n: 2); |
| 175 | |
| 176 | // Replace all "\XXX\..\" with "\". Don't try too hard though as the original |
| 177 | // path should be well-formatted, e.g. start with a drive letter, etc. |
| 178 | Cursor = 0; |
| 179 | while ((Cursor = Filepath.find(s: "\\..\\" , pos: Cursor)) != std::string::npos) { |
| 180 | // Something's wrong if the path starts with "\..\", abort. |
| 181 | if (Cursor == 0) |
| 182 | break; |
| 183 | |
| 184 | size_t PrevSlash = Filepath.rfind(c: '\\', pos: Cursor - 1); |
| 185 | if (PrevSlash == std::string::npos) |
| 186 | // Something's wrong, abort. |
| 187 | break; |
| 188 | |
| 189 | Filepath.erase(pos: PrevSlash, n: Cursor + 3 - PrevSlash); |
| 190 | // The next ".." might be following the one we've just erased. |
| 191 | Cursor = PrevSlash; |
| 192 | } |
| 193 | |
| 194 | // Remove all duplicate backslashes. |
| 195 | Cursor = 0; |
| 196 | while ((Cursor = Filepath.find(s: "\\\\" , pos: Cursor)) != std::string::npos) |
| 197 | Filepath.erase(pos: Cursor, n: 1); |
| 198 | |
| 199 | return Filepath; |
| 200 | } |
| 201 | |
| 202 | unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { |
| 203 | StringRef FullPath = getFullFilepath(File: F); |
| 204 | unsigned NextId = FileIdMap.size() + 1; |
| 205 | auto Insertion = FileIdMap.insert(KV: std::make_pair(x&: FullPath, y&: NextId)); |
| 206 | if (Insertion.second) { |
| 207 | // We have to compute the full filepath and emit a .cv_file directive. |
| 208 | ArrayRef<uint8_t> ChecksumAsBytes; |
| 209 | FileChecksumKind CSKind = FileChecksumKind::None; |
| 210 | if (F->getChecksum()) { |
| 211 | std::string Checksum = fromHex(Input: F->getChecksum()->Value); |
| 212 | void *CKMem = OS.getContext().allocate(Size: Checksum.size(), Align: 1); |
| 213 | memcpy(dest: CKMem, src: Checksum.data(), n: Checksum.size()); |
| 214 | ChecksumAsBytes = ArrayRef<uint8_t>( |
| 215 | reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); |
| 216 | switch (F->getChecksum()->Kind) { |
| 217 | case DIFile::CSK_MD5: |
| 218 | CSKind = FileChecksumKind::MD5; |
| 219 | break; |
| 220 | case DIFile::CSK_SHA1: |
| 221 | CSKind = FileChecksumKind::SHA1; |
| 222 | break; |
| 223 | case DIFile::CSK_SHA256: |
| 224 | CSKind = FileChecksumKind::SHA256; |
| 225 | break; |
| 226 | } |
| 227 | } |
| 228 | bool Success = OS.emitCVFileDirective(FileNo: NextId, Filename: FullPath, Checksum: ChecksumAsBytes, |
| 229 | ChecksumKind: static_cast<unsigned>(CSKind)); |
| 230 | (void)Success; |
| 231 | assert(Success && ".cv_file directive failed" ); |
| 232 | } |
| 233 | return Insertion.first->second; |
| 234 | } |
| 235 | |
| 236 | CodeViewDebug::InlineSite & |
| 237 | CodeViewDebug::getInlineSite(const DILocation *InlinedAt, |
| 238 | const DISubprogram *Inlinee) { |
| 239 | auto SiteInsertion = CurFn->InlineSites.try_emplace(k: InlinedAt); |
| 240 | InlineSite *Site = &SiteInsertion.first->second; |
| 241 | if (SiteInsertion.second) { |
| 242 | unsigned ParentFuncId = CurFn->FuncId; |
| 243 | if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) |
| 244 | ParentFuncId = |
| 245 | getInlineSite(InlinedAt: OuterIA, Inlinee: InlinedAt->getScope()->getSubprogram()) |
| 246 | .SiteFuncId; |
| 247 | |
| 248 | Site->SiteFuncId = NextFuncId++; |
| 249 | OS.emitCVInlineSiteIdDirective( |
| 250 | FunctionId: Site->SiteFuncId, IAFunc: ParentFuncId, IAFile: maybeRecordFile(F: InlinedAt->getFile()), |
| 251 | IALine: InlinedAt->getLine(), IACol: InlinedAt->getColumn(), Loc: SMLoc()); |
| 252 | Site->Inlinee = Inlinee; |
| 253 | InlinedSubprograms.insert(X: Inlinee); |
| 254 | auto InlineeIdx = getFuncIdForSubprogram(SP: Inlinee); |
| 255 | |
| 256 | if (InlinedAt->getInlinedAt() == nullptr) |
| 257 | CurFn->Inlinees.insert(V: InlineeIdx); |
| 258 | } |
| 259 | return *Site; |
| 260 | } |
| 261 | |
| 262 | static StringRef getPrettyScopeName(const DIScope *Scope) { |
| 263 | StringRef ScopeName = Scope->getName(); |
| 264 | if (!ScopeName.empty()) |
| 265 | return ScopeName; |
| 266 | |
| 267 | switch (Scope->getTag()) { |
| 268 | case dwarf::DW_TAG_enumeration_type: |
| 269 | case dwarf::DW_TAG_class_type: |
| 270 | case dwarf::DW_TAG_structure_type: |
| 271 | case dwarf::DW_TAG_union_type: |
| 272 | return "<unnamed-tag>" ; |
| 273 | case dwarf::DW_TAG_namespace: |
| 274 | return "`anonymous namespace'" ; |
| 275 | default: |
| 276 | return StringRef(); |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | const DISubprogram *CodeViewDebug::collectParentScopeNames( |
| 281 | const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { |
| 282 | const DISubprogram *ClosestSubprogram = nullptr; |
| 283 | while (Scope != nullptr) { |
| 284 | if (ClosestSubprogram == nullptr) |
| 285 | ClosestSubprogram = dyn_cast<DISubprogram>(Val: Scope); |
| 286 | |
| 287 | // If a type appears in a scope chain, make sure it gets emitted. The |
| 288 | // frontend will be responsible for deciding if this should be a forward |
| 289 | // declaration or a complete type. |
| 290 | if (const auto *Ty = dyn_cast<DICompositeType>(Val: Scope)) |
| 291 | DeferredCompleteTypes.push_back(Elt: Ty); |
| 292 | |
| 293 | StringRef ScopeName = getPrettyScopeName(Scope); |
| 294 | if (!ScopeName.empty()) |
| 295 | QualifiedNameComponents.push_back(Elt: ScopeName); |
| 296 | Scope = Scope->getScope(); |
| 297 | } |
| 298 | return ClosestSubprogram; |
| 299 | } |
| 300 | |
| 301 | static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, |
| 302 | StringRef TypeName) { |
| 303 | std::string FullyQualifiedName; |
| 304 | for (StringRef QualifiedNameComponent : |
| 305 | llvm::reverse(C&: QualifiedNameComponents)) { |
| 306 | FullyQualifiedName.append(str: std::string(QualifiedNameComponent)); |
| 307 | FullyQualifiedName.append(s: "::" ); |
| 308 | } |
| 309 | FullyQualifiedName.append(str: std::string(TypeName)); |
| 310 | return FullyQualifiedName; |
| 311 | } |
| 312 | |
| 313 | struct CodeViewDebug::TypeLoweringScope { |
| 314 | TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } |
| 315 | ~TypeLoweringScope() { |
| 316 | // Don't decrement TypeEmissionLevel until after emitting deferred types, so |
| 317 | // inner TypeLoweringScopes don't attempt to emit deferred types. |
| 318 | if (CVD.TypeEmissionLevel == 1) |
| 319 | CVD.emitDeferredCompleteTypes(); |
| 320 | --CVD.TypeEmissionLevel; |
| 321 | } |
| 322 | CodeViewDebug &CVD; |
| 323 | }; |
| 324 | |
| 325 | std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, |
| 326 | StringRef Name) { |
| 327 | // Ensure types in the scope chain are emitted as soon as possible. |
| 328 | // This can create otherwise a situation where S_UDTs are emitted while |
| 329 | // looping in emitDebugInfoForUDTs. |
| 330 | TypeLoweringScope S(*this); |
| 331 | SmallVector<StringRef, 5> QualifiedNameComponents; |
| 332 | collectParentScopeNames(Scope, QualifiedNameComponents); |
| 333 | return formatNestedName(QualifiedNameComponents, TypeName: Name); |
| 334 | } |
| 335 | |
| 336 | std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { |
| 337 | const DIScope *Scope = Ty->getScope(); |
| 338 | return getFullyQualifiedName(Scope, Name: getPrettyScopeName(Scope: Ty)); |
| 339 | } |
| 340 | |
| 341 | TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { |
| 342 | // No scope means global scope and that uses the zero index. |
| 343 | // |
| 344 | // We also use zero index when the scope is a DISubprogram |
| 345 | // to suppress the emission of LF_STRING_ID for the function, |
| 346 | // which can trigger a link-time error with the linker in |
| 347 | // VS2019 version 16.11.2 or newer. |
| 348 | // Note, however, skipping the debug info emission for the DISubprogram |
| 349 | // is a temporary fix. The root issue here is that we need to figure out |
| 350 | // the proper way to encode a function nested in another function |
| 351 | // (as introduced by the Fortran 'contains' keyword) in CodeView. |
| 352 | if (!Scope || isa<DIFile>(Val: Scope) || isa<DISubprogram>(Val: Scope)) |
| 353 | return TypeIndex(); |
| 354 | |
| 355 | assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type" ); |
| 356 | |
| 357 | // Check if we've already translated this scope. |
| 358 | auto I = TypeIndices.find(Val: {Scope, nullptr}); |
| 359 | if (I != TypeIndices.end()) |
| 360 | return I->second; |
| 361 | |
| 362 | // Build the fully qualified name of the scope. |
| 363 | std::string ScopeName = getFullyQualifiedName(Ty: Scope); |
| 364 | StringIdRecord SID(TypeIndex(), ScopeName); |
| 365 | auto TI = TypeTable.writeLeafType(Record&: SID); |
| 366 | return recordTypeIndexForDINode(Node: Scope, TI); |
| 367 | } |
| 368 | |
| 369 | static StringRef removeTemplateArgs(StringRef Name) { |
| 370 | // Remove template args from the display name. Assume that the template args |
| 371 | // are the last thing in the name. |
| 372 | if (Name.empty() || Name.back() != '>') |
| 373 | return Name; |
| 374 | |
| 375 | int OpenBrackets = 0; |
| 376 | for (int i = Name.size() - 1; i >= 0; --i) { |
| 377 | if (Name[i] == '>') |
| 378 | ++OpenBrackets; |
| 379 | else if (Name[i] == '<') { |
| 380 | --OpenBrackets; |
| 381 | if (OpenBrackets == 0) |
| 382 | return Name.substr(Start: 0, N: i); |
| 383 | } |
| 384 | } |
| 385 | return Name; |
| 386 | } |
| 387 | |
| 388 | TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { |
| 389 | assert(SP); |
| 390 | |
| 391 | // Check if we've already translated this subprogram. |
| 392 | auto I = TypeIndices.find(Val: {SP, nullptr}); |
| 393 | if (I != TypeIndices.end()) |
| 394 | return I->second; |
| 395 | |
| 396 | // The display name includes function template arguments. Drop them to match |
| 397 | // MSVC. We need to have the template arguments in the DISubprogram name |
| 398 | // because they are used in other symbol records, such as S_GPROC32_IDs. |
| 399 | StringRef DisplayName = removeTemplateArgs(Name: SP->getName()); |
| 400 | |
| 401 | const DIScope *Scope = SP->getScope(); |
| 402 | TypeIndex TI; |
| 403 | if (const auto *Class = dyn_cast_or_null<DICompositeType>(Val: Scope)) { |
| 404 | // If the scope is a DICompositeType, then this must be a method. Member |
| 405 | // function types take some special handling, and require access to the |
| 406 | // subprogram. |
| 407 | TypeIndex ClassType = getTypeIndex(Ty: Class); |
| 408 | MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), |
| 409 | DisplayName); |
| 410 | TI = TypeTable.writeLeafType(Record&: MFuncId); |
| 411 | } else { |
| 412 | // Otherwise, this must be a free function. |
| 413 | TypeIndex ParentScope = getScopeIndex(Scope); |
| 414 | FuncIdRecord FuncId(ParentScope, getTypeIndex(Ty: SP->getType()), DisplayName); |
| 415 | TI = TypeTable.writeLeafType(Record&: FuncId); |
| 416 | } |
| 417 | |
| 418 | return recordTypeIndexForDINode(Node: SP, TI); |
| 419 | } |
| 420 | |
| 421 | static bool isNonTrivial(const DICompositeType *DCTy) { |
| 422 | return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); |
| 423 | } |
| 424 | |
| 425 | static FunctionOptions |
| 426 | getFunctionOptions(const DISubroutineType *Ty, |
| 427 | const DICompositeType *ClassTy = nullptr, |
| 428 | StringRef SPName = StringRef("" )) { |
| 429 | FunctionOptions FO = FunctionOptions::None; |
| 430 | const DIType *ReturnTy = nullptr; |
| 431 | if (auto TypeArray = Ty->getTypeArray()) { |
| 432 | if (TypeArray.size()) |
| 433 | ReturnTy = TypeArray[0]; |
| 434 | } |
| 435 | |
| 436 | // Add CxxReturnUdt option to functions that return nontrivial record types |
| 437 | // or methods that return record types. |
| 438 | if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(Val: ReturnTy)) |
| 439 | if (isNonTrivial(DCTy: ReturnDCTy) || ClassTy) |
| 440 | FO |= FunctionOptions::CxxReturnUdt; |
| 441 | |
| 442 | // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. |
| 443 | if (ClassTy && isNonTrivial(DCTy: ClassTy) && SPName == ClassTy->getName()) { |
| 444 | FO |= FunctionOptions::Constructor; |
| 445 | |
| 446 | // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. |
| 447 | |
| 448 | } |
| 449 | return FO; |
| 450 | } |
| 451 | |
| 452 | TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, |
| 453 | const DICompositeType *Class) { |
| 454 | // Always use the method declaration as the key for the function type. The |
| 455 | // method declaration contains the this adjustment. |
| 456 | if (SP->getDeclaration()) |
| 457 | SP = SP->getDeclaration(); |
| 458 | assert(!SP->getDeclaration() && "should use declaration as key" ); |
| 459 | |
| 460 | // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide |
| 461 | // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. |
| 462 | auto I = TypeIndices.find(Val: {SP, Class}); |
| 463 | if (I != TypeIndices.end()) |
| 464 | return I->second; |
| 465 | |
| 466 | // Make sure complete type info for the class is emitted *after* the member |
| 467 | // function type, as the complete class type is likely to reference this |
| 468 | // member function type. |
| 469 | TypeLoweringScope S(*this); |
| 470 | const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; |
| 471 | |
| 472 | FunctionOptions FO = getFunctionOptions(Ty: SP->getType(), ClassTy: Class, SPName: SP->getName()); |
| 473 | TypeIndex TI = lowerTypeMemberFunction( |
| 474 | Ty: SP->getType(), ClassTy: Class, ThisAdjustment: SP->getThisAdjustment(), IsStaticMethod, FO); |
| 475 | return recordTypeIndexForDINode(Node: SP, TI, ClassTy: Class); |
| 476 | } |
| 477 | |
| 478 | TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, |
| 479 | TypeIndex TI, |
| 480 | const DIType *ClassTy) { |
| 481 | auto InsertResult = TypeIndices.insert(KV: {{Node, ClassTy}, TI}); |
| 482 | (void)InsertResult; |
| 483 | assert(InsertResult.second && "DINode was already assigned a type index" ); |
| 484 | return TI; |
| 485 | } |
| 486 | |
| 487 | unsigned CodeViewDebug::getPointerSizeInBytes() { |
| 488 | return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; |
| 489 | } |
| 490 | |
| 491 | void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, |
| 492 | const LexicalScope *LS) { |
| 493 | if (const DILocation *InlinedAt = LS->getInlinedAt()) { |
| 494 | // This variable was inlined. Associate it with the InlineSite. |
| 495 | const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); |
| 496 | InlineSite &Site = getInlineSite(InlinedAt, Inlinee); |
| 497 | Site.InlinedLocals.emplace_back(Args: std::move(Var)); |
| 498 | } else { |
| 499 | // This variable goes into the corresponding lexical scope. |
| 500 | ScopeVariables[LS].emplace_back(Args: std::move(Var)); |
| 501 | } |
| 502 | } |
| 503 | |
| 504 | static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, |
| 505 | const DILocation *Loc) { |
| 506 | if (!llvm::is_contained(Range&: Locs, Element: Loc)) |
| 507 | Locs.push_back(Elt: Loc); |
| 508 | } |
| 509 | |
| 510 | void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, |
| 511 | const MachineFunction *MF) { |
| 512 | // Skip this instruction if it has the same location as the previous one. |
| 513 | if (!DL || DL == PrevInstLoc) |
| 514 | return; |
| 515 | |
| 516 | const DIScope *Scope = DL->getScope(); |
| 517 | if (!Scope) |
| 518 | return; |
| 519 | |
| 520 | // Skip this line if it is longer than the maximum we can record. |
| 521 | LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); |
| 522 | if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || |
| 523 | LI.isNeverStepInto()) |
| 524 | return; |
| 525 | |
| 526 | ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); |
| 527 | if (CI.getStartColumn() != DL.getCol()) |
| 528 | return; |
| 529 | |
| 530 | if (!CurFn->HaveLineInfo) |
| 531 | CurFn->HaveLineInfo = true; |
| 532 | unsigned FileId = 0; |
| 533 | if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) |
| 534 | FileId = CurFn->LastFileId; |
| 535 | else |
| 536 | FileId = CurFn->LastFileId = maybeRecordFile(F: DL->getFile()); |
| 537 | PrevInstLoc = DL; |
| 538 | |
| 539 | unsigned FuncId = CurFn->FuncId; |
| 540 | if (const DILocation *SiteLoc = DL->getInlinedAt()) { |
| 541 | const DILocation *Loc = DL.get(); |
| 542 | |
| 543 | // If this location was actually inlined from somewhere else, give it the ID |
| 544 | // of the inline call site. |
| 545 | FuncId = |
| 546 | getInlineSite(InlinedAt: SiteLoc, Inlinee: Loc->getScope()->getSubprogram()).SiteFuncId; |
| 547 | |
| 548 | // Ensure we have links in the tree of inline call sites. |
| 549 | bool FirstLoc = true; |
| 550 | while ((SiteLoc = Loc->getInlinedAt())) { |
| 551 | InlineSite &Site = |
| 552 | getInlineSite(InlinedAt: SiteLoc, Inlinee: Loc->getScope()->getSubprogram()); |
| 553 | if (!FirstLoc) |
| 554 | addLocIfNotPresent(Locs&: Site.ChildSites, Loc); |
| 555 | FirstLoc = false; |
| 556 | Loc = SiteLoc; |
| 557 | } |
| 558 | addLocIfNotPresent(Locs&: CurFn->ChildSites, Loc); |
| 559 | } |
| 560 | |
| 561 | OS.emitCVLocDirective(FunctionId: FuncId, FileNo: FileId, Line: DL.getLine(), Column: DL.getCol(), |
| 562 | /*PrologueEnd=*/false, /*IsStmt=*/false, |
| 563 | FileName: DL->getFilename(), Loc: SMLoc()); |
| 564 | } |
| 565 | |
| 566 | void CodeViewDebug::emitCodeViewMagicVersion() { |
| 567 | OS.emitValueToAlignment(Alignment: Align(4)); |
| 568 | OS.AddComment(T: "Debug section magic" ); |
| 569 | OS.emitInt32(Value: COFF::DEBUG_SECTION_MAGIC); |
| 570 | } |
| 571 | |
| 572 | static SourceLanguage |
| 573 | MapDWARFLanguageToCVLang(dwarf::SourceLanguageName DWLName) { |
| 574 | switch (DWLName) { |
| 575 | case dwarf::DW_LNAME_C: |
| 576 | return SourceLanguage::C; |
| 577 | case dwarf::DW_LNAME_C_plus_plus: |
| 578 | return SourceLanguage::Cpp; |
| 579 | case dwarf::DW_LNAME_Fortran: |
| 580 | return SourceLanguage::Fortran; |
| 581 | case dwarf::DW_LNAME_Pascal: |
| 582 | return SourceLanguage::Pascal; |
| 583 | case dwarf::DW_LNAME_Cobol: |
| 584 | return SourceLanguage::Cobol; |
| 585 | case dwarf::DW_LNAME_Java: |
| 586 | return SourceLanguage::Java; |
| 587 | case dwarf::DW_LNAME_D: |
| 588 | return SourceLanguage::D; |
| 589 | case dwarf::DW_LNAME_Swift: |
| 590 | return SourceLanguage::Swift; |
| 591 | case dwarf::DW_LNAME_Rust: |
| 592 | return SourceLanguage::Rust; |
| 593 | case dwarf::DW_LNAME_ObjC: |
| 594 | return SourceLanguage::ObjC; |
| 595 | case dwarf::DW_LNAME_ObjC_plus_plus: |
| 596 | return SourceLanguage::ObjCpp; |
| 597 | default: |
| 598 | // There's no CodeView representation for this language, and CV doesn't |
| 599 | // have an "unknown" option for the language field, so we'll use MASM, |
| 600 | // as it's very low level. |
| 601 | return SourceLanguage::Masm; |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | static SourceLanguage MapDWARFLanguageToCVLang(dwarf::SourceLanguage DWLang) { |
| 606 | auto MaybeLName = dwarf::toDW_LNAME(language: DWLang); |
| 607 | if (!MaybeLName) |
| 608 | return MapDWARFLanguageToCVLang(DWLName: static_cast<dwarf::SourceLanguageName>(0)); |
| 609 | |
| 610 | return MapDWARFLanguageToCVLang(DWLName: MaybeLName->first); |
| 611 | } |
| 612 | |
| 613 | void CodeViewDebug::beginModule(Module *M) { |
| 614 | // If COFF debug section is not available, skip any debug info related stuff. |
| 615 | if (!Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { |
| 616 | Asm = nullptr; |
| 617 | return; |
| 618 | } |
| 619 | |
| 620 | CompilerInfoAsm = Asm; |
| 621 | TheCPU = mapArchToCVCPUType(Type: M->getTargetTriple().getArch()); |
| 622 | |
| 623 | // Get the current source language. |
| 624 | const MDNode *Node; |
| 625 | if (Asm->hasDebugInfo()) { |
| 626 | Node = *M->debug_compile_units_begin(); |
| 627 | } else { |
| 628 | // When emitting only compiler information, we may have only NoDebug CUs, |
| 629 | // which would be skipped by debug_compile_units_begin. |
| 630 | NamedMDNode *CUs = MMI->getModule()->getNamedMetadata(Name: "llvm.dbg.cu" ); |
| 631 | if (CUs->operands().empty()) { |
| 632 | Asm = nullptr; |
| 633 | return; |
| 634 | } |
| 635 | Node = *CUs->operands().begin(); |
| 636 | } |
| 637 | |
| 638 | TheCU = cast<DICompileUnit>(Val: Node); |
| 639 | DISourceLanguageName Lang = TheCU->getSourceLanguage(); |
| 640 | CurrentSourceLanguage = |
| 641 | Lang.hasVersionedName() |
| 642 | ? MapDWARFLanguageToCVLang( |
| 643 | DWLName: static_cast<dwarf::SourceLanguageName>(Lang.getName())) |
| 644 | : MapDWARFLanguageToCVLang( |
| 645 | DWLang: static_cast<dwarf::SourceLanguage>(Lang.getName())); |
| 646 | if (!M->getCodeViewFlag() || |
| 647 | TheCU->getEmissionKind() == DICompileUnit::NoDebug) { |
| 648 | Asm = nullptr; |
| 649 | return; |
| 650 | } |
| 651 | |
| 652 | collectGlobalVariableInfo(); |
| 653 | |
| 654 | // Check if we should emit type record hashes. |
| 655 | ConstantInt *GH = |
| 656 | mdconst::extract_or_null<ConstantInt>(MD: M->getModuleFlag(Key: "CodeViewGHash" )); |
| 657 | EmitDebugGlobalHashes = GH && !GH->isZero(); |
| 658 | } |
| 659 | |
| 660 | void CodeViewDebug::endModule() { |
| 661 | if (!CompilerInfoAsm) |
| 662 | return; |
| 663 | |
| 664 | // The COFF .debug$S section consists of several subsections, each starting |
| 665 | // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length |
| 666 | // of the payload followed by the payload itself. The subsections are 4-byte |
| 667 | // aligned. |
| 668 | |
| 669 | // Use the generic .debug$S section, and make a subsection for all the inlined |
| 670 | // subprograms. |
| 671 | switchToDebugSectionForSymbol(GVSym: nullptr); |
| 672 | |
| 673 | MCSymbol *CompilerInfo = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 674 | emitObjName(); |
| 675 | emitCompilerInformation(); |
| 676 | endCVSubsection(EndLabel: CompilerInfo); |
| 677 | if (!Asm) |
| 678 | return; |
| 679 | |
| 680 | emitSecureHotPatchInformation(); |
| 681 | |
| 682 | emitInlineeLinesSubsection(); |
| 683 | |
| 684 | // Emit per-function debug information. |
| 685 | for (auto &P : FnDebugInfo) |
| 686 | if (!P.first->isDeclarationForLinker()) |
| 687 | emitDebugInfoForFunction(GV: P.first, FI&: *P.second); |
| 688 | |
| 689 | // Get types used by globals without emitting anything. |
| 690 | // This is meant to collect all static const data members so they can be |
| 691 | // emitted as globals. |
| 692 | collectDebugInfoForGlobals(); |
| 693 | |
| 694 | // Emit retained types. |
| 695 | emitDebugInfoForRetainedTypes(); |
| 696 | |
| 697 | // Emit global variable debug information. |
| 698 | setCurrentSubprogram(nullptr); |
| 699 | emitDebugInfoForGlobals(); |
| 700 | |
| 701 | // Switch back to the generic .debug$S section after potentially processing |
| 702 | // comdat symbol sections. |
| 703 | switchToDebugSectionForSymbol(GVSym: nullptr); |
| 704 | |
| 705 | // Emit UDT records for any types used by global variables. |
| 706 | if (!GlobalUDTs.empty()) { |
| 707 | MCSymbol *SymbolsEnd = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 708 | emitDebugInfoForUDTs(UDTs: GlobalUDTs); |
| 709 | endCVSubsection(EndLabel: SymbolsEnd); |
| 710 | } |
| 711 | |
| 712 | // This subsection holds a file index to offset in string table table. |
| 713 | OS.AddComment(T: "File index to string table offset subsection" ); |
| 714 | OS.emitCVFileChecksumsDirective(); |
| 715 | |
| 716 | // This subsection holds the string table. |
| 717 | OS.AddComment(T: "String table" ); |
| 718 | OS.emitCVStringTableDirective(); |
| 719 | |
| 720 | // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol |
| 721 | // subsection in the generic .debug$S section at the end. There is no |
| 722 | // particular reason for this ordering other than to match MSVC. |
| 723 | emitBuildInfo(); |
| 724 | |
| 725 | // Emit type information and hashes last, so that any types we translate while |
| 726 | // emitting function info are included. |
| 727 | emitTypeInformation(); |
| 728 | |
| 729 | if (EmitDebugGlobalHashes) |
| 730 | emitTypeGlobalHashes(); |
| 731 | |
| 732 | clear(); |
| 733 | } |
| 734 | |
| 735 | static void |
| 736 | emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, |
| 737 | unsigned MaxFixedRecordLength = 0xF00) { |
| 738 | // The maximum CV record length is 0xFF00. Most of the strings we emit appear |
| 739 | // after a fixed length portion of the record. The fixed length portion should |
| 740 | // always be less than 0xF00 (3840) bytes, so truncate the string so that the |
| 741 | // overall record size is less than the maximum allowed. |
| 742 | SmallString<32> NullTerminatedString( |
| 743 | S.take_front(N: MaxRecordLength - MaxFixedRecordLength - 1)); |
| 744 | NullTerminatedString.push_back(Elt: '\0'); |
| 745 | OS.emitBytes(Data: NullTerminatedString); |
| 746 | } |
| 747 | |
| 748 | void CodeViewDebug::emitTypeInformation() { |
| 749 | if (TypeTable.empty()) |
| 750 | return; |
| 751 | |
| 752 | // Start the .debug$T or .debug$P section with 0x4. |
| 753 | OS.switchSection(Section: Asm->getObjFileLowering().getCOFFDebugTypesSection()); |
| 754 | emitCodeViewMagicVersion(); |
| 755 | |
| 756 | TypeTableCollection Table(TypeTable.records()); |
| 757 | TypeVisitorCallbackPipeline Pipeline; |
| 758 | |
| 759 | // To emit type record using Codeview MCStreamer adapter |
| 760 | CVMCAdapter CVMCOS(OS, Table); |
| 761 | TypeRecordMapping typeMapping(CVMCOS); |
| 762 | Pipeline.addCallbackToPipeline(Callbacks&: typeMapping); |
| 763 | |
| 764 | std::optional<TypeIndex> B = Table.getFirst(); |
| 765 | while (B) { |
| 766 | // This will fail if the record data is invalid. |
| 767 | CVType Record = Table.getType(Index: *B); |
| 768 | |
| 769 | Error E = codeview::visitTypeRecord(Record, Index: *B, Callbacks&: Pipeline); |
| 770 | |
| 771 | if (E) { |
| 772 | logAllUnhandledErrors(E: std::move(E), OS&: errs(), ErrorBanner: "error: " ); |
| 773 | llvm_unreachable("produced malformed type record" ); |
| 774 | } |
| 775 | |
| 776 | B = Table.getNext(Prev: *B); |
| 777 | } |
| 778 | } |
| 779 | |
| 780 | void CodeViewDebug::emitTypeGlobalHashes() { |
| 781 | if (TypeTable.empty()) |
| 782 | return; |
| 783 | |
| 784 | // Start the .debug$H section with the version and hash algorithm, currently |
| 785 | // hardcoded to version 0, SHA1. |
| 786 | OS.switchSection(Section: Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); |
| 787 | |
| 788 | OS.emitValueToAlignment(Alignment: Align(4)); |
| 789 | OS.AddComment(T: "Magic" ); |
| 790 | OS.emitInt32(Value: COFF::DEBUG_HASHES_SECTION_MAGIC); |
| 791 | OS.AddComment(T: "Section Version" ); |
| 792 | OS.emitInt16(Value: 0); |
| 793 | OS.AddComment(T: "Hash Algorithm" ); |
| 794 | OS.emitInt16(Value: uint16_t(GlobalTypeHashAlg::BLAKE3)); |
| 795 | |
| 796 | TypeIndex TI(TypeIndex::FirstNonSimpleIndex); |
| 797 | for (const auto &GHR : TypeTable.hashes()) { |
| 798 | if (OS.isVerboseAsm()) { |
| 799 | // Emit an EOL-comment describing which TypeIndex this hash corresponds |
| 800 | // to, as well as the stringified SHA1 hash. |
| 801 | SmallString<32> ; |
| 802 | raw_svector_ostream (Comment); |
| 803 | CommentOS << formatv(Fmt: "{0:X+} [{1}]" , Vals: TI.getIndex(), Vals: GHR); |
| 804 | OS.AddComment(T: Comment); |
| 805 | ++TI; |
| 806 | } |
| 807 | assert(GHR.Hash.size() == 8); |
| 808 | StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), |
| 809 | GHR.Hash.size()); |
| 810 | OS.emitBinaryData(Data: S); |
| 811 | } |
| 812 | } |
| 813 | |
| 814 | void CodeViewDebug::emitObjName() { |
| 815 | MCSymbol *CompilerEnd = beginSymbolRecord(Kind: SymbolKind::S_OBJNAME); |
| 816 | |
| 817 | StringRef PathRef(CompilerInfoAsm->TM.Options.ObjectFilenameForDebug); |
| 818 | llvm::SmallString<256> PathStore(PathRef); |
| 819 | |
| 820 | if (PathRef.empty() || PathRef == "-" ) { |
| 821 | // Don't emit the filename if we're writing to stdout or to /dev/null. |
| 822 | PathRef = {}; |
| 823 | } else { |
| 824 | PathRef = PathStore; |
| 825 | } |
| 826 | |
| 827 | OS.AddComment(T: "Signature" ); |
| 828 | OS.emitIntValue(Value: 0, Size: 4); |
| 829 | |
| 830 | OS.AddComment(T: "Object name" ); |
| 831 | emitNullTerminatedSymbolName(OS, S: PathRef); |
| 832 | |
| 833 | endSymbolRecord(SymEnd: CompilerEnd); |
| 834 | } |
| 835 | |
| 836 | void CodeViewDebug::emitSecureHotPatchInformation() { |
| 837 | MCSymbol *hotPatchInfo = nullptr; |
| 838 | |
| 839 | for (const auto &F : MMI->getModule()->functions()) { |
| 840 | if (!F.isDeclarationForLinker() && |
| 841 | F.hasFnAttribute(Kind: "marked_for_windows_hot_patching" )) { |
| 842 | if (hotPatchInfo == nullptr) |
| 843 | hotPatchInfo = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 844 | MCSymbol *HotPatchEnd = beginSymbolRecord(Kind: SymbolKind::S_HOTPATCHFUNC); |
| 845 | auto *SP = F.getSubprogram(); |
| 846 | OS.AddComment(T: "Function" ); |
| 847 | OS.emitInt32(Value: getFuncIdForSubprogram(SP).getIndex()); |
| 848 | OS.AddComment(T: "Name" ); |
| 849 | emitNullTerminatedSymbolName(OS, S: F.getName()); |
| 850 | endSymbolRecord(SymEnd: HotPatchEnd); |
| 851 | } |
| 852 | } |
| 853 | |
| 854 | if (hotPatchInfo != nullptr) |
| 855 | endCVSubsection(EndLabel: hotPatchInfo); |
| 856 | } |
| 857 | |
| 858 | namespace { |
| 859 | struct Version { |
| 860 | int Part[4]; |
| 861 | }; |
| 862 | } // end anonymous namespace |
| 863 | |
| 864 | // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out |
| 865 | // the version number. |
| 866 | static Version parseVersion(StringRef Name) { |
| 867 | Version V = {.Part: {0}}; |
| 868 | int N = 0; |
| 869 | for (const char C : Name) { |
| 870 | if (isdigit(C)) { |
| 871 | V.Part[N] *= 10; |
| 872 | V.Part[N] += C - '0'; |
| 873 | V.Part[N] = |
| 874 | std::min<int>(a: V.Part[N], b: std::numeric_limits<uint16_t>::max()); |
| 875 | } else if (C == '.') { |
| 876 | ++N; |
| 877 | if (N >= 4) |
| 878 | return V; |
| 879 | } else if (N > 0) |
| 880 | return V; |
| 881 | } |
| 882 | return V; |
| 883 | } |
| 884 | |
| 885 | void CodeViewDebug::emitCompilerInformation() { |
| 886 | MCSymbol *CompilerEnd = beginSymbolRecord(Kind: SymbolKind::S_COMPILE3); |
| 887 | uint32_t Flags = 0; |
| 888 | |
| 889 | // The low byte of the flags indicates the source language. |
| 890 | Flags = CurrentSourceLanguage; |
| 891 | // TODO: Figure out which other flags need to be set. |
| 892 | if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { |
| 893 | Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); |
| 894 | } |
| 895 | using ArchType = llvm::Triple::ArchType; |
| 896 | ArchType Arch = MMI->getModule()->getTargetTriple().getArch(); |
| 897 | if (CompilerInfoAsm->TM.Options.Hotpatch || Arch == ArchType::thumb || |
| 898 | Arch == ArchType::aarch64) { |
| 899 | Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); |
| 900 | } |
| 901 | |
| 902 | OS.AddComment(T: "Flags and language" ); |
| 903 | OS.emitInt32(Value: Flags); |
| 904 | |
| 905 | OS.AddComment(T: "CPUType" ); |
| 906 | OS.emitInt16(Value: static_cast<uint64_t>(TheCPU)); |
| 907 | |
| 908 | StringRef CompilerVersion = "0" ; |
| 909 | if (TheCU) |
| 910 | CompilerVersion = TheCU->getProducer(); |
| 911 | |
| 912 | Version FrontVer = parseVersion(Name: CompilerVersion); |
| 913 | OS.AddComment(T: "Frontend version" ); |
| 914 | for (int N : FrontVer.Part) { |
| 915 | OS.emitInt16(Value: N); |
| 916 | } |
| 917 | |
| 918 | // Some Microsoft tools, like Binscope, expect a backend version number of at |
| 919 | // least 8.something, so we'll coerce the LLVM version into a form that |
| 920 | // guarantees it'll be big enough without really lying about the version. |
| 921 | int Major = 1000 * LLVM_VERSION_MAJOR + |
| 922 | 10 * LLVM_VERSION_MINOR + |
| 923 | LLVM_VERSION_PATCH; |
| 924 | // Clamp it for builds that use unusually large version numbers. |
| 925 | Major = std::min<int>(a: Major, b: std::numeric_limits<uint16_t>::max()); |
| 926 | Version BackVer = {.Part: { Major, 0, 0, 0 }}; |
| 927 | OS.AddComment(T: "Backend version" ); |
| 928 | for (int N : BackVer.Part) |
| 929 | OS.emitInt16(Value: N); |
| 930 | |
| 931 | OS.AddComment(T: "Null-terminated compiler version string" ); |
| 932 | emitNullTerminatedSymbolName(OS, S: CompilerVersion); |
| 933 | |
| 934 | endSymbolRecord(SymEnd: CompilerEnd); |
| 935 | } |
| 936 | |
| 937 | static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, |
| 938 | StringRef S) { |
| 939 | StringIdRecord SIR(TypeIndex(0x0), S); |
| 940 | return TypeTable.writeLeafType(Record&: SIR); |
| 941 | } |
| 942 | |
| 943 | void CodeViewDebug::emitBuildInfo() { |
| 944 | // First, make LF_BUILDINFO. It's a sequence of strings with various bits of |
| 945 | // build info. The known prefix is: |
| 946 | // - Absolute path of current directory |
| 947 | // - Compiler path |
| 948 | // - Main source file path, relative to CWD or absolute |
| 949 | // - Type server PDB file |
| 950 | // - Canonical compiler command line |
| 951 | // If frontend and backend compilation are separated (think llc or LTO), it's |
| 952 | // not clear if the compiler path should refer to the executable for the |
| 953 | // frontend or the backend. Leave it blank for now. |
| 954 | TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; |
| 955 | NamedMDNode *CUs = MMI->getModule()->getNamedMetadata(Name: "llvm.dbg.cu" ); |
| 956 | const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. |
| 957 | const auto *CU = cast<DICompileUnit>(Val: Node); |
| 958 | const DIFile *MainSourceFile = CU->getFile(); |
| 959 | BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = |
| 960 | getStringIdTypeIdx(TypeTable, S: MainSourceFile->getDirectory()); |
| 961 | BuildInfoArgs[BuildInfoRecord::SourceFile] = |
| 962 | getStringIdTypeIdx(TypeTable, S: MainSourceFile->getFilename()); |
| 963 | // FIXME: PDB is intentionally blank unless we implement /Zi type servers. |
| 964 | BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = |
| 965 | getStringIdTypeIdx(TypeTable, S: "" ); |
| 966 | BuildInfoArgs[BuildInfoRecord::BuildTool] = |
| 967 | getStringIdTypeIdx(TypeTable, S: Asm->TM.Options.MCOptions.Argv0); |
| 968 | BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( |
| 969 | TypeTable, S: Asm->TM.Options.MCOptions.CommandlineArgs); |
| 970 | |
| 971 | BuildInfoRecord BIR(BuildInfoArgs); |
| 972 | TypeIndex BuildInfoIndex = TypeTable.writeLeafType(Record&: BIR); |
| 973 | |
| 974 | // Make a new .debug$S subsection for the S_BUILDINFO record, which points |
| 975 | // from the module symbols into the type stream. |
| 976 | MCSymbol *BISubsecEnd = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 977 | MCSymbol *BIEnd = beginSymbolRecord(Kind: SymbolKind::S_BUILDINFO); |
| 978 | OS.AddComment(T: "LF_BUILDINFO index" ); |
| 979 | OS.emitInt32(Value: BuildInfoIndex.getIndex()); |
| 980 | endSymbolRecord(SymEnd: BIEnd); |
| 981 | endCVSubsection(EndLabel: BISubsecEnd); |
| 982 | } |
| 983 | |
| 984 | void CodeViewDebug::emitInlineeLinesSubsection() { |
| 985 | if (InlinedSubprograms.empty()) |
| 986 | return; |
| 987 | |
| 988 | OS.AddComment(T: "Inlinee lines subsection" ); |
| 989 | MCSymbol *InlineEnd = beginCVSubsection(Kind: DebugSubsectionKind::InlineeLines); |
| 990 | |
| 991 | // We emit the checksum info for files. This is used by debuggers to |
| 992 | // determine if a pdb matches the source before loading it. Visual Studio, |
| 993 | // for instance, will display a warning that the breakpoints are not valid if |
| 994 | // the pdb does not match the source. |
| 995 | OS.AddComment(T: "Inlinee lines signature" ); |
| 996 | OS.emitInt32(Value: unsigned(InlineeLinesSignature::Normal)); |
| 997 | |
| 998 | for (const DISubprogram *SP : InlinedSubprograms) { |
| 999 | assert(TypeIndices.count({SP, nullptr})); |
| 1000 | TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; |
| 1001 | |
| 1002 | OS.addBlankLine(); |
| 1003 | unsigned FileId = maybeRecordFile(F: SP->getFile()); |
| 1004 | OS.AddComment(T: "Inlined function " + SP->getName() + " starts at " + |
| 1005 | SP->getFilename() + Twine(':') + Twine(SP->getLine())); |
| 1006 | OS.addBlankLine(); |
| 1007 | OS.AddComment(T: "Type index of inlined function" ); |
| 1008 | OS.emitInt32(Value: InlineeIdx.getIndex()); |
| 1009 | OS.AddComment(T: "Offset into filechecksum table" ); |
| 1010 | OS.emitCVFileChecksumOffsetDirective(FileNo: FileId); |
| 1011 | OS.AddComment(T: "Starting line number" ); |
| 1012 | OS.emitInt32(Value: SP->getLine()); |
| 1013 | } |
| 1014 | |
| 1015 | endCVSubsection(EndLabel: InlineEnd); |
| 1016 | } |
| 1017 | |
| 1018 | void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, |
| 1019 | const DILocation *InlinedAt, |
| 1020 | const InlineSite &Site) { |
| 1021 | assert(TypeIndices.count({Site.Inlinee, nullptr})); |
| 1022 | TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; |
| 1023 | |
| 1024 | // SymbolRecord |
| 1025 | MCSymbol *InlineEnd = beginSymbolRecord(Kind: SymbolKind::S_INLINESITE); |
| 1026 | |
| 1027 | OS.AddComment(T: "PtrParent" ); |
| 1028 | OS.emitInt32(Value: 0); |
| 1029 | OS.AddComment(T: "PtrEnd" ); |
| 1030 | OS.emitInt32(Value: 0); |
| 1031 | OS.AddComment(T: "Inlinee type index" ); |
| 1032 | OS.emitInt32(Value: InlineeIdx.getIndex()); |
| 1033 | |
| 1034 | unsigned FileId = maybeRecordFile(F: Site.Inlinee->getFile()); |
| 1035 | unsigned StartLineNum = Site.Inlinee->getLine(); |
| 1036 | |
| 1037 | OS.emitCVInlineLinetableDirective(PrimaryFunctionId: Site.SiteFuncId, SourceFileId: FileId, SourceLineNum: StartLineNum, |
| 1038 | FnStartSym: FI.Begin, FnEndSym: FI.End); |
| 1039 | |
| 1040 | endSymbolRecord(SymEnd: InlineEnd); |
| 1041 | |
| 1042 | emitLocalVariableList(FI, Locals: Site.InlinedLocals); |
| 1043 | |
| 1044 | // Recurse on child inlined call sites before closing the scope. |
| 1045 | for (const DILocation *ChildSite : Site.ChildSites) { |
| 1046 | auto I = FI.InlineSites.find(x: ChildSite); |
| 1047 | assert(I != FI.InlineSites.end() && |
| 1048 | "child site not in function inline site map" ); |
| 1049 | emitInlinedCallSite(FI, InlinedAt: ChildSite, Site: I->second); |
| 1050 | } |
| 1051 | |
| 1052 | // Close the scope. |
| 1053 | emitEndSymbolRecord(EndKind: SymbolKind::S_INLINESITE_END); |
| 1054 | } |
| 1055 | |
| 1056 | void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { |
| 1057 | // If we have a symbol, it may be in a section that is COMDAT. If so, find the |
| 1058 | // comdat key. A section may be comdat because of -ffunction-sections or |
| 1059 | // because it is comdat in the IR. |
| 1060 | MCSectionCOFF *GVSec = |
| 1061 | GVSym ? static_cast<MCSectionCOFF *>(&GVSym->getSection()) : nullptr; |
| 1062 | const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; |
| 1063 | |
| 1064 | auto *DebugSec = static_cast<MCSectionCOFF *>( |
| 1065 | CompilerInfoAsm->getObjFileLowering().getCOFFDebugSymbolsSection()); |
| 1066 | DebugSec = OS.getContext().getAssociativeCOFFSection(Sec: DebugSec, KeySym); |
| 1067 | |
| 1068 | OS.switchSection(Section: DebugSec); |
| 1069 | |
| 1070 | // Emit the magic version number if this is the first time we've switched to |
| 1071 | // this section. |
| 1072 | if (ComdatDebugSections.insert(V: DebugSec).second) |
| 1073 | emitCodeViewMagicVersion(); |
| 1074 | } |
| 1075 | |
| 1076 | // Emit an S_THUNK32/S_END symbol pair for a thunk routine. |
| 1077 | // The only supported thunk ordinal is currently the standard type. |
| 1078 | void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, |
| 1079 | FunctionInfo &FI, |
| 1080 | const MCSymbol *Fn) { |
| 1081 | std::string FuncName = |
| 1082 | std::string(GlobalValue::dropLLVMManglingEscape(Name: GV->getName())); |
| 1083 | const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. |
| 1084 | |
| 1085 | OS.AddComment(T: "Symbol subsection for " + Twine(FuncName)); |
| 1086 | MCSymbol *SymbolsEnd = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 1087 | |
| 1088 | // Emit S_THUNK32 |
| 1089 | MCSymbol *ThunkRecordEnd = beginSymbolRecord(Kind: SymbolKind::S_THUNK32); |
| 1090 | OS.AddComment(T: "PtrParent" ); |
| 1091 | OS.emitInt32(Value: 0); |
| 1092 | OS.AddComment(T: "PtrEnd" ); |
| 1093 | OS.emitInt32(Value: 0); |
| 1094 | OS.AddComment(T: "PtrNext" ); |
| 1095 | OS.emitInt32(Value: 0); |
| 1096 | OS.AddComment(T: "Thunk section relative address" ); |
| 1097 | OS.emitCOFFSecRel32(Symbol: Fn, /*Offset=*/0); |
| 1098 | OS.AddComment(T: "Thunk section index" ); |
| 1099 | OS.emitCOFFSectionIndex(Symbol: Fn); |
| 1100 | OS.AddComment(T: "Code size" ); |
| 1101 | OS.emitAbsoluteSymbolDiff(Hi: FI.End, Lo: Fn, Size: 2); |
| 1102 | OS.AddComment(T: "Ordinal" ); |
| 1103 | OS.emitInt8(Value: unsigned(ordinal)); |
| 1104 | OS.AddComment(T: "Function name" ); |
| 1105 | emitNullTerminatedSymbolName(OS, S: FuncName); |
| 1106 | // Additional fields specific to the thunk ordinal would go here. |
| 1107 | endSymbolRecord(SymEnd: ThunkRecordEnd); |
| 1108 | |
| 1109 | // Local variables/inlined routines are purposely omitted here. The point of |
| 1110 | // marking this as a thunk is so Visual Studio will NOT stop in this routine. |
| 1111 | |
| 1112 | // Emit S_PROC_ID_END |
| 1113 | emitEndSymbolRecord(EndKind: SymbolKind::S_PROC_ID_END); |
| 1114 | |
| 1115 | endCVSubsection(EndLabel: SymbolsEnd); |
| 1116 | } |
| 1117 | |
| 1118 | void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, |
| 1119 | FunctionInfo &FI) { |
| 1120 | // For each function there is a separate subsection which holds the PC to |
| 1121 | // file:line table. |
| 1122 | const MCSymbol *Fn = Asm->getSymbol(GV); |
| 1123 | assert(Fn); |
| 1124 | |
| 1125 | // Switch to the to a comdat section, if appropriate. |
| 1126 | switchToDebugSectionForSymbol(GVSym: Fn); |
| 1127 | |
| 1128 | std::string FuncName; |
| 1129 | auto *SP = GV->getSubprogram(); |
| 1130 | assert(SP); |
| 1131 | setCurrentSubprogram(SP); |
| 1132 | |
| 1133 | if (SP->isThunk()) { |
| 1134 | emitDebugInfoForThunk(GV, FI, Fn); |
| 1135 | return; |
| 1136 | } |
| 1137 | |
| 1138 | // If we have a display name, build the fully qualified name by walking the |
| 1139 | // chain of scopes. |
| 1140 | if (!SP->getName().empty()) |
| 1141 | FuncName = getFullyQualifiedName(Scope: SP->getScope(), Name: SP->getName()); |
| 1142 | |
| 1143 | // If our DISubprogram name is empty, use the mangled name. |
| 1144 | if (FuncName.empty()) |
| 1145 | FuncName = std::string(GlobalValue::dropLLVMManglingEscape(Name: GV->getName())); |
| 1146 | |
| 1147 | // Emit FPO data, but only on 32-bit x86. No other platforms use it. |
| 1148 | if (MMI->getModule()->getTargetTriple().getArch() == Triple::x86) |
| 1149 | OS.emitCVFPOData(ProcSym: Fn); |
| 1150 | |
| 1151 | // Emit a symbol subsection, required by VS2012+ to find function boundaries. |
| 1152 | OS.AddComment(T: "Symbol subsection for " + Twine(FuncName)); |
| 1153 | MCSymbol *SymbolsEnd = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 1154 | { |
| 1155 | SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID |
| 1156 | : SymbolKind::S_GPROC32_ID; |
| 1157 | MCSymbol *ProcRecordEnd = beginSymbolRecord(Kind: ProcKind); |
| 1158 | |
| 1159 | // These fields are filled in by tools like CVPACK which run after the fact. |
| 1160 | OS.AddComment(T: "PtrParent" ); |
| 1161 | OS.emitInt32(Value: 0); |
| 1162 | OS.AddComment(T: "PtrEnd" ); |
| 1163 | OS.emitInt32(Value: 0); |
| 1164 | OS.AddComment(T: "PtrNext" ); |
| 1165 | OS.emitInt32(Value: 0); |
| 1166 | // This is the important bit that tells the debugger where the function |
| 1167 | // code is located and what's its size: |
| 1168 | OS.AddComment(T: "Code size" ); |
| 1169 | OS.emitAbsoluteSymbolDiff(Hi: FI.End, Lo: Fn, Size: 4); |
| 1170 | OS.AddComment(T: "Offset after prologue" ); |
| 1171 | OS.emitInt32(Value: 0); |
| 1172 | OS.AddComment(T: "Offset before epilogue" ); |
| 1173 | OS.emitInt32(Value: 0); |
| 1174 | OS.AddComment(T: "Function type index" ); |
| 1175 | OS.emitInt32(Value: getFuncIdForSubprogram(SP: GV->getSubprogram()).getIndex()); |
| 1176 | OS.AddComment(T: "Function section relative address" ); |
| 1177 | OS.emitCOFFSecRel32(Symbol: Fn, /*Offset=*/0); |
| 1178 | OS.AddComment(T: "Function section index" ); |
| 1179 | OS.emitCOFFSectionIndex(Symbol: Fn); |
| 1180 | OS.AddComment(T: "Flags" ); |
| 1181 | ProcSymFlags ProcFlags = ProcSymFlags::HasOptimizedDebugInfo; |
| 1182 | if (FI.HasFramePointer) |
| 1183 | ProcFlags |= ProcSymFlags::HasFP; |
| 1184 | if (GV->hasFnAttribute(Kind: Attribute::NoReturn)) |
| 1185 | ProcFlags |= ProcSymFlags::IsNoReturn; |
| 1186 | if (GV->hasFnAttribute(Kind: Attribute::NoInline)) |
| 1187 | ProcFlags |= ProcSymFlags::IsNoInline; |
| 1188 | OS.emitInt8(Value: static_cast<uint8_t>(ProcFlags)); |
| 1189 | // Emit the function display name as a null-terminated string. |
| 1190 | OS.AddComment(T: "Function name" ); |
| 1191 | // Truncate the name so we won't overflow the record length field. |
| 1192 | emitNullTerminatedSymbolName(OS, S: FuncName); |
| 1193 | endSymbolRecord(SymEnd: ProcRecordEnd); |
| 1194 | |
| 1195 | MCSymbol *FrameProcEnd = beginSymbolRecord(Kind: SymbolKind::S_FRAMEPROC); |
| 1196 | // Subtract out the CSR size since MSVC excludes that and we include it. |
| 1197 | OS.AddComment(T: "FrameSize" ); |
| 1198 | OS.emitInt32(Value: FI.FrameSize - FI.CSRSize); |
| 1199 | OS.AddComment(T: "Padding" ); |
| 1200 | OS.emitInt32(Value: 0); |
| 1201 | OS.AddComment(T: "Offset of padding" ); |
| 1202 | OS.emitInt32(Value: 0); |
| 1203 | OS.AddComment(T: "Bytes of callee saved registers" ); |
| 1204 | OS.emitInt32(Value: FI.CSRSize); |
| 1205 | OS.AddComment(T: "Exception handler offset" ); |
| 1206 | OS.emitInt32(Value: 0); |
| 1207 | OS.AddComment(T: "Exception handler section" ); |
| 1208 | OS.emitInt16(Value: 0); |
| 1209 | OS.AddComment(T: "Flags (defines frame register)" ); |
| 1210 | OS.emitInt32(Value: uint32_t(FI.FrameProcOpts)); |
| 1211 | endSymbolRecord(SymEnd: FrameProcEnd); |
| 1212 | |
| 1213 | emitInlinees(Inlinees: FI.Inlinees); |
| 1214 | emitLocalVariableList(FI, Locals: FI.Locals); |
| 1215 | emitGlobalVariableList(Globals: FI.Globals); |
| 1216 | emitLexicalBlockList(Blocks: FI.ChildBlocks, FI); |
| 1217 | |
| 1218 | // Emit inlined call site information. Only emit functions inlined directly |
| 1219 | // into the parent function. We'll emit the other sites recursively as part |
| 1220 | // of their parent inline site. |
| 1221 | for (const DILocation *InlinedAt : FI.ChildSites) { |
| 1222 | auto I = FI.InlineSites.find(x: InlinedAt); |
| 1223 | assert(I != FI.InlineSites.end() && |
| 1224 | "child site not in function inline site map" ); |
| 1225 | emitInlinedCallSite(FI, InlinedAt, Site: I->second); |
| 1226 | } |
| 1227 | |
| 1228 | for (auto Annot : FI.Annotations) { |
| 1229 | MCSymbol *Label = Annot.first; |
| 1230 | MDTuple *Strs = cast<MDTuple>(Val: Annot.second); |
| 1231 | MCSymbol *AnnotEnd = beginSymbolRecord(Kind: SymbolKind::S_ANNOTATION); |
| 1232 | OS.emitCOFFSecRel32(Symbol: Label, /*Offset=*/0); |
| 1233 | // FIXME: Make sure we don't overflow the max record size. |
| 1234 | OS.emitCOFFSectionIndex(Symbol: Label); |
| 1235 | OS.emitInt16(Value: Strs->getNumOperands()); |
| 1236 | for (Metadata *MD : Strs->operands()) { |
| 1237 | // MDStrings are null terminated, so we can do EmitBytes and get the |
| 1238 | // nice .asciz directive. |
| 1239 | StringRef Str = cast<MDString>(Val: MD)->getString(); |
| 1240 | assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString" ); |
| 1241 | OS.emitBytes(Data: StringRef(Str.data(), Str.size() + 1)); |
| 1242 | } |
| 1243 | endSymbolRecord(SymEnd: AnnotEnd); |
| 1244 | } |
| 1245 | |
| 1246 | for (auto HeapAllocSite : FI.HeapAllocSites) { |
| 1247 | const MCSymbol *BeginLabel = std::get<0>(t&: HeapAllocSite); |
| 1248 | const MCSymbol *EndLabel = std::get<1>(t&: HeapAllocSite); |
| 1249 | const DIType *DITy = std::get<2>(t&: HeapAllocSite); |
| 1250 | MCSymbol *HeapAllocEnd = beginSymbolRecord(Kind: SymbolKind::S_HEAPALLOCSITE); |
| 1251 | OS.AddComment(T: "Call site offset" ); |
| 1252 | OS.emitCOFFSecRel32(Symbol: BeginLabel, /*Offset=*/0); |
| 1253 | OS.AddComment(T: "Call site section index" ); |
| 1254 | OS.emitCOFFSectionIndex(Symbol: BeginLabel); |
| 1255 | OS.AddComment(T: "Call instruction length" ); |
| 1256 | OS.emitAbsoluteSymbolDiff(Hi: EndLabel, Lo: BeginLabel, Size: 2); |
| 1257 | OS.AddComment(T: "Type index" ); |
| 1258 | OS.emitInt32(Value: getCompleteTypeIndex(Ty: DITy).getIndex()); |
| 1259 | endSymbolRecord(SymEnd: HeapAllocEnd); |
| 1260 | } |
| 1261 | |
| 1262 | if (SP != nullptr) |
| 1263 | emitDebugInfoForUDTs(UDTs: LocalUDTs); |
| 1264 | |
| 1265 | emitDebugInfoForJumpTables(FI); |
| 1266 | |
| 1267 | // We're done with this function. |
| 1268 | emitEndSymbolRecord(EndKind: SymbolKind::S_PROC_ID_END); |
| 1269 | } |
| 1270 | endCVSubsection(EndLabel: SymbolsEnd); |
| 1271 | |
| 1272 | // We have an assembler directive that takes care of the whole line table. |
| 1273 | OS.emitCVLinetableDirective(FunctionId: FI.FuncId, FnStart: Fn, FnEnd: FI.End); |
| 1274 | } |
| 1275 | |
| 1276 | CodeViewDebug::LocalVarDef |
| 1277 | CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { |
| 1278 | LocalVarDef DR; |
| 1279 | DR.InMemory = -1; |
| 1280 | DR.DataOffset = Offset; |
| 1281 | assert(DR.DataOffset == Offset && "truncation" ); |
| 1282 | DR.IsSubfield = 0; |
| 1283 | DR.StructOffset = 0; |
| 1284 | DR.CVRegister = CVRegister; |
| 1285 | return DR; |
| 1286 | } |
| 1287 | |
| 1288 | void CodeViewDebug::collectVariableInfoFromMFTable( |
| 1289 | DenseSet<InlinedEntity> &Processed) { |
| 1290 | const MachineFunction &MF = *Asm->MF; |
| 1291 | const TargetSubtargetInfo &TSI = MF.getSubtarget(); |
| 1292 | const TargetFrameLowering *TFI = TSI.getFrameLowering(); |
| 1293 | const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); |
| 1294 | |
| 1295 | for (const MachineFunction::VariableDbgInfo &VI : |
| 1296 | MF.getInStackSlotVariableDbgInfo()) { |
| 1297 | if (!VI.Var) |
| 1298 | continue; |
| 1299 | assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && |
| 1300 | "Expected inlined-at fields to agree" ); |
| 1301 | |
| 1302 | Processed.insert(V: InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); |
| 1303 | LexicalScope *Scope = LScopes.findLexicalScope(DL: VI.Loc); |
| 1304 | |
| 1305 | // If variable scope is not found then skip this variable. |
| 1306 | if (!Scope) |
| 1307 | continue; |
| 1308 | |
| 1309 | // If the variable has an attached offset expression, extract it. |
| 1310 | // FIXME: Try to handle DW_OP_deref as well. |
| 1311 | int64_t ExprOffset = 0; |
| 1312 | bool Deref = false; |
| 1313 | if (VI.Expr) { |
| 1314 | // If there is one DW_OP_deref element, use offset of 0 and keep going. |
| 1315 | if (VI.Expr->getNumElements() == 1 && |
| 1316 | VI.Expr->getElement(I: 0) == llvm::dwarf::DW_OP_deref) |
| 1317 | Deref = true; |
| 1318 | else if (!VI.Expr->extractIfOffset(Offset&: ExprOffset)) |
| 1319 | continue; |
| 1320 | } |
| 1321 | |
| 1322 | // Get the frame register used and the offset. |
| 1323 | Register FrameReg; |
| 1324 | StackOffset FrameOffset = |
| 1325 | TFI->getFrameIndexReference(MF: *Asm->MF, FI: VI.getStackSlot(), FrameReg); |
| 1326 | uint16_t CVReg = TRI->getCodeViewRegNum(Reg: FrameReg); |
| 1327 | |
| 1328 | if (FrameOffset.getScalable()) { |
| 1329 | // No encoding currently exists for scalable offsets; bail out. |
| 1330 | continue; |
| 1331 | } |
| 1332 | |
| 1333 | // Calculate the label ranges. |
| 1334 | LocalVarDef DefRange = |
| 1335 | createDefRangeMem(CVRegister: CVReg, Offset: FrameOffset.getFixed() + ExprOffset); |
| 1336 | |
| 1337 | LocalVariable Var; |
| 1338 | Var.DIVar = VI.Var; |
| 1339 | |
| 1340 | for (const InsnRange &Range : Scope->getRanges()) { |
| 1341 | const MCSymbol *Begin = getLabelBeforeInsn(MI: Range.first); |
| 1342 | const MCSymbol *End = getLabelAfterInsn(MI: Range.second); |
| 1343 | End = End ? End : Asm->getFunctionEnd(); |
| 1344 | Var.DefRanges[DefRange].emplace_back(Args&: Begin, Args&: End); |
| 1345 | } |
| 1346 | |
| 1347 | if (Deref) |
| 1348 | Var.UseReferenceType = true; |
| 1349 | |
| 1350 | recordLocalVariable(Var: std::move(Var), LS: Scope); |
| 1351 | } |
| 1352 | } |
| 1353 | |
| 1354 | static bool canUseReferenceType(const DbgVariableLocation &Loc) { |
| 1355 | return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; |
| 1356 | } |
| 1357 | |
| 1358 | static bool needsReferenceType(const DbgVariableLocation &Loc) { |
| 1359 | return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; |
| 1360 | } |
| 1361 | |
| 1362 | void CodeViewDebug::calculateRanges( |
| 1363 | LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { |
| 1364 | const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); |
| 1365 | |
| 1366 | // Calculate the definition ranges. |
| 1367 | for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { |
| 1368 | const auto &Entry = *I; |
| 1369 | if (!Entry.isDbgValue()) |
| 1370 | continue; |
| 1371 | const MachineInstr *DVInst = Entry.getInstr(); |
| 1372 | assert(DVInst->isDebugValue() && "Invalid History entry" ); |
| 1373 | // FIXME: Find a way to represent constant variables, since they are |
| 1374 | // relatively common. |
| 1375 | std::optional<DbgVariableLocation> Location = |
| 1376 | DbgVariableLocation::extractFromMachineInstruction(Instruction: *DVInst); |
| 1377 | if (!Location) |
| 1378 | { |
| 1379 | // When we don't have a location this is usually because LLVM has |
| 1380 | // transformed it into a constant and we only have an llvm.dbg.value. We |
| 1381 | // can't represent these well in CodeView since S_LOCAL only works on |
| 1382 | // registers and memory locations. Instead, we will pretend this to be a |
| 1383 | // constant value to at least have it show up in the debugger. |
| 1384 | auto Op = DVInst->getDebugOperand(Index: 0); |
| 1385 | if (Op.isImm()) |
| 1386 | Var.ConstantValue = APSInt(APInt(64, Op.getImm()), false); |
| 1387 | continue; |
| 1388 | } |
| 1389 | |
| 1390 | // CodeView can only express variables in register and variables in memory |
| 1391 | // at a constant offset from a register. However, for variables passed |
| 1392 | // indirectly by pointer, it is common for that pointer to be spilled to a |
| 1393 | // stack location. For the special case of one offseted load followed by a |
| 1394 | // zero offset load (a pointer spilled to the stack), we change the type of |
| 1395 | // the local variable from a value type to a reference type. This tricks the |
| 1396 | // debugger into doing the load for us. |
| 1397 | if (Var.UseReferenceType) { |
| 1398 | // We're using a reference type. Drop the last zero offset load. |
| 1399 | if (canUseReferenceType(Loc: *Location)) |
| 1400 | Location->LoadChain.pop_back(); |
| 1401 | else |
| 1402 | continue; |
| 1403 | } else if (needsReferenceType(Loc: *Location)) { |
| 1404 | // This location can't be expressed without switching to a reference type. |
| 1405 | // Start over using that. |
| 1406 | Var.UseReferenceType = true; |
| 1407 | Var.DefRanges.clear(); |
| 1408 | calculateRanges(Var, Entries); |
| 1409 | return; |
| 1410 | } |
| 1411 | |
| 1412 | // We can only handle a register or an offseted load of a register. |
| 1413 | if (!Location->Register || Location->LoadChain.size() > 1) |
| 1414 | continue; |
| 1415 | |
| 1416 | // Codeview can only express byte-aligned offsets, ensure that we have a |
| 1417 | // byte-boundaried location. |
| 1418 | if (Location->FragmentInfo) |
| 1419 | if (Location->FragmentInfo->OffsetInBits % 8) |
| 1420 | continue; |
| 1421 | |
| 1422 | if (TRI->isIgnoredCVReg(LLVMReg: Location->Register)) { |
| 1423 | // No encoding currently exists for this register; bail out. |
| 1424 | continue; |
| 1425 | } |
| 1426 | |
| 1427 | LocalVarDef DR; |
| 1428 | DR.CVRegister = TRI->getCodeViewRegNum(Reg: Location->Register); |
| 1429 | DR.InMemory = !Location->LoadChain.empty(); |
| 1430 | DR.DataOffset = |
| 1431 | !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; |
| 1432 | if (Location->FragmentInfo) { |
| 1433 | DR.IsSubfield = true; |
| 1434 | DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; |
| 1435 | } else { |
| 1436 | DR.IsSubfield = false; |
| 1437 | DR.StructOffset = 0; |
| 1438 | } |
| 1439 | |
| 1440 | // Compute the label range. |
| 1441 | const MCSymbol *Begin = getLabelBeforeInsn(MI: Entry.getInstr()); |
| 1442 | const MCSymbol *End; |
| 1443 | if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { |
| 1444 | auto &EndingEntry = Entries[Entry.getEndIndex()]; |
| 1445 | End = EndingEntry.isDbgValue() |
| 1446 | ? getLabelBeforeInsn(MI: EndingEntry.getInstr()) |
| 1447 | : getLabelAfterInsn(MI: EndingEntry.getInstr()); |
| 1448 | } else |
| 1449 | End = Asm->getFunctionEnd(); |
| 1450 | |
| 1451 | // If the last range end is our begin, just extend the last range. |
| 1452 | // Otherwise make a new range. |
| 1453 | SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = |
| 1454 | Var.DefRanges[DR]; |
| 1455 | if (!R.empty() && R.back().second == Begin) |
| 1456 | R.back().second = End; |
| 1457 | else |
| 1458 | R.emplace_back(Args&: Begin, Args&: End); |
| 1459 | |
| 1460 | // FIXME: Do more range combining. |
| 1461 | } |
| 1462 | } |
| 1463 | |
| 1464 | void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { |
| 1465 | DenseSet<InlinedEntity> Processed; |
| 1466 | // Grab the variable info that was squirreled away in the MMI side-table. |
| 1467 | collectVariableInfoFromMFTable(Processed); |
| 1468 | |
| 1469 | for (const auto &I : DbgValues) { |
| 1470 | InlinedEntity IV = I.first; |
| 1471 | if (Processed.count(V: IV)) |
| 1472 | continue; |
| 1473 | const DILocalVariable *DIVar = cast<DILocalVariable>(Val: IV.first); |
| 1474 | const DILocation *InlinedAt = IV.second; |
| 1475 | |
| 1476 | // Instruction ranges, specifying where IV is accessible. |
| 1477 | const auto &Entries = I.second; |
| 1478 | |
| 1479 | LexicalScope *Scope = nullptr; |
| 1480 | if (InlinedAt) |
| 1481 | Scope = LScopes.findInlinedScope(N: DIVar->getScope(), IA: InlinedAt); |
| 1482 | else |
| 1483 | Scope = LScopes.findLexicalScope(N: DIVar->getScope()); |
| 1484 | // If variable scope is not found then skip this variable. |
| 1485 | if (!Scope) |
| 1486 | continue; |
| 1487 | |
| 1488 | LocalVariable Var; |
| 1489 | Var.DIVar = DIVar; |
| 1490 | |
| 1491 | calculateRanges(Var, Entries); |
| 1492 | recordLocalVariable(Var: std::move(Var), LS: Scope); |
| 1493 | } |
| 1494 | } |
| 1495 | |
| 1496 | void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { |
| 1497 | const TargetSubtargetInfo &TSI = MF->getSubtarget(); |
| 1498 | const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); |
| 1499 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
| 1500 | const Function &GV = MF->getFunction(); |
| 1501 | auto Insertion = FnDebugInfo.insert(KV: {&GV, std::make_unique<FunctionInfo>()}); |
| 1502 | assert(Insertion.second && "function already has info" ); |
| 1503 | CurFn = Insertion.first->second.get(); |
| 1504 | CurFn->FuncId = NextFuncId++; |
| 1505 | CurFn->Begin = Asm->getFunctionBegin(); |
| 1506 | |
| 1507 | // The S_FRAMEPROC record reports the stack size, and how many bytes of |
| 1508 | // callee-saved registers were used. For targets that don't use a PUSH |
| 1509 | // instruction (AArch64), this will be zero. |
| 1510 | CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); |
| 1511 | CurFn->FrameSize = MFI.getStackSize(); |
| 1512 | CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); |
| 1513 | CurFn->HasStackRealignment = TRI->hasStackRealignment(MF: *MF); |
| 1514 | |
| 1515 | // For this function S_FRAMEPROC record, figure out which codeview register |
| 1516 | // will be the frame pointer. |
| 1517 | CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. |
| 1518 | CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. |
| 1519 | if (CurFn->FrameSize > 0) { |
| 1520 | if (!TSI.getFrameLowering()->hasFP(MF: *MF)) { |
| 1521 | CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; |
| 1522 | CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; |
| 1523 | } else { |
| 1524 | CurFn->HasFramePointer = true; |
| 1525 | // If there is an FP, parameters are always relative to it. |
| 1526 | CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; |
| 1527 | if (CurFn->HasStackRealignment) { |
| 1528 | // If the stack needs realignment, locals are relative to SP or VFRAME. |
| 1529 | CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; |
| 1530 | } else { |
| 1531 | // Otherwise, locals are relative to EBP, and we probably have VLAs or |
| 1532 | // other stack adjustments. |
| 1533 | CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; |
| 1534 | } |
| 1535 | } |
| 1536 | } |
| 1537 | |
| 1538 | // Compute other frame procedure options. |
| 1539 | FrameProcedureOptions FPO = FrameProcedureOptions::None; |
| 1540 | if (MFI.hasVarSizedObjects()) |
| 1541 | FPO |= FrameProcedureOptions::HasAlloca; |
| 1542 | if (MF->exposesReturnsTwice()) |
| 1543 | FPO |= FrameProcedureOptions::HasSetJmp; |
| 1544 | // FIXME: Set HasLongJmp if we ever track that info. |
| 1545 | if (MF->hasInlineAsm()) |
| 1546 | FPO |= FrameProcedureOptions::HasInlineAssembly; |
| 1547 | if (GV.hasPersonalityFn()) { |
| 1548 | if (isAsynchronousEHPersonality( |
| 1549 | Pers: classifyEHPersonality(Pers: GV.getPersonalityFn()))) |
| 1550 | FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; |
| 1551 | else |
| 1552 | FPO |= FrameProcedureOptions::HasExceptionHandling; |
| 1553 | } |
| 1554 | if (GV.hasFnAttribute(Kind: Attribute::InlineHint)) |
| 1555 | FPO |= FrameProcedureOptions::MarkedInline; |
| 1556 | if (GV.hasFnAttribute(Kind: Attribute::Naked)) |
| 1557 | FPO |= FrameProcedureOptions::Naked; |
| 1558 | if (MFI.hasStackProtectorIndex()) { |
| 1559 | FPO |= FrameProcedureOptions::SecurityChecks; |
| 1560 | if (GV.hasFnAttribute(Kind: Attribute::StackProtectStrong) || |
| 1561 | GV.hasFnAttribute(Kind: Attribute::StackProtectReq)) { |
| 1562 | FPO |= FrameProcedureOptions::StrictSecurityChecks; |
| 1563 | } |
| 1564 | } else if (!GV.hasStackProtectorFnAttr()) { |
| 1565 | // __declspec(safebuffers) disables stack guards. |
| 1566 | FPO |= FrameProcedureOptions::SafeBuffers; |
| 1567 | } |
| 1568 | FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); |
| 1569 | FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); |
| 1570 | if (Asm->TM.getOptLevel() != CodeGenOptLevel::None && !GV.hasOptSize() && |
| 1571 | !GV.hasOptNone()) |
| 1572 | FPO |= FrameProcedureOptions::OptimizedForSpeed; |
| 1573 | if (GV.hasProfileData()) { |
| 1574 | FPO |= FrameProcedureOptions::ValidProfileCounts; |
| 1575 | FPO |= FrameProcedureOptions::ProfileGuidedOptimization; |
| 1576 | } |
| 1577 | // FIXME: Set GuardCfg when it is implemented. |
| 1578 | CurFn->FrameProcOpts = FPO; |
| 1579 | |
| 1580 | OS.emitCVFuncIdDirective(FunctionId: CurFn->FuncId); |
| 1581 | |
| 1582 | // Find the end of the function prolog. First known non-DBG_VALUE and |
| 1583 | // non-frame setup location marks the beginning of the function body. |
| 1584 | // FIXME: is there a simpler a way to do this? Can we just search |
| 1585 | // for the first instruction of the function, not the last of the prolog? |
| 1586 | DebugLoc PrologEndLoc; |
| 1587 | bool EmptyPrologue = true; |
| 1588 | for (const auto &MBB : *MF) { |
| 1589 | for (const auto &MI : MBB) { |
| 1590 | if (!MI.isMetaInstruction() && !MI.getFlag(Flag: MachineInstr::FrameSetup) && |
| 1591 | MI.getDebugLoc()) { |
| 1592 | PrologEndLoc = MI.getDebugLoc(); |
| 1593 | break; |
| 1594 | } else if (!MI.isMetaInstruction()) { |
| 1595 | EmptyPrologue = false; |
| 1596 | } |
| 1597 | } |
| 1598 | } |
| 1599 | |
| 1600 | // Record beginning of function if we have a non-empty prologue. |
| 1601 | if (PrologEndLoc && !EmptyPrologue) { |
| 1602 | DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); |
| 1603 | maybeRecordLocation(DL: FnStartDL, MF); |
| 1604 | } |
| 1605 | |
| 1606 | // Find heap alloc sites and emit labels around them. |
| 1607 | for (const auto &MBB : *MF) { |
| 1608 | for (const auto &MI : MBB) { |
| 1609 | if (MI.getHeapAllocMarker()) { |
| 1610 | requestLabelBeforeInsn(MI: &MI); |
| 1611 | requestLabelAfterInsn(MI: &MI); |
| 1612 | } |
| 1613 | } |
| 1614 | } |
| 1615 | |
| 1616 | // Mark branches that may potentially be using jump tables with labels. |
| 1617 | bool isThumb = MMI->getModule()->getTargetTriple().getArch() == |
| 1618 | llvm::Triple::ArchType::thumb; |
| 1619 | discoverJumpTableBranches(MF, isThumb); |
| 1620 | } |
| 1621 | |
| 1622 | static bool shouldEmitUdt(const DIType *T) { |
| 1623 | if (!T) |
| 1624 | return false; |
| 1625 | |
| 1626 | // MSVC does not emit UDTs for typedefs that are scoped to classes. |
| 1627 | if (T->getTag() == dwarf::DW_TAG_typedef) { |
| 1628 | if (DIScope *Scope = T->getScope()) { |
| 1629 | switch (Scope->getTag()) { |
| 1630 | case dwarf::DW_TAG_structure_type: |
| 1631 | case dwarf::DW_TAG_class_type: |
| 1632 | case dwarf::DW_TAG_union_type: |
| 1633 | return false; |
| 1634 | default: |
| 1635 | // do nothing. |
| 1636 | ; |
| 1637 | } |
| 1638 | } |
| 1639 | } |
| 1640 | |
| 1641 | while (true) { |
| 1642 | if (!T || T->isForwardDecl()) |
| 1643 | return false; |
| 1644 | |
| 1645 | const DIDerivedType *DT = dyn_cast<DIDerivedType>(Val: T); |
| 1646 | if (!DT) |
| 1647 | return true; |
| 1648 | T = DT->getBaseType(); |
| 1649 | } |
| 1650 | return true; |
| 1651 | } |
| 1652 | |
| 1653 | void CodeViewDebug::addToUDTs(const DIType *Ty) { |
| 1654 | // Don't record empty UDTs. |
| 1655 | if (Ty->getName().empty()) |
| 1656 | return; |
| 1657 | if (!shouldEmitUdt(T: Ty)) |
| 1658 | return; |
| 1659 | |
| 1660 | SmallVector<StringRef, 5> ParentScopeNames; |
| 1661 | const DISubprogram *ClosestSubprogram = |
| 1662 | collectParentScopeNames(Scope: Ty->getScope(), QualifiedNameComponents&: ParentScopeNames); |
| 1663 | |
| 1664 | std::string FullyQualifiedName = |
| 1665 | formatNestedName(QualifiedNameComponents: ParentScopeNames, TypeName: getPrettyScopeName(Scope: Ty)); |
| 1666 | |
| 1667 | if (ClosestSubprogram == nullptr) { |
| 1668 | GlobalUDTs.emplace_back(args: std::move(FullyQualifiedName), args&: Ty); |
| 1669 | } else if (ClosestSubprogram == CurrentSubprogram) { |
| 1670 | LocalUDTs.emplace_back(args: std::move(FullyQualifiedName), args&: Ty); |
| 1671 | } |
| 1672 | |
| 1673 | // TODO: What if the ClosestSubprogram is neither null or the current |
| 1674 | // subprogram? Currently, the UDT just gets dropped on the floor. |
| 1675 | // |
| 1676 | // The current behavior is not desirable. To get maximal fidelity, we would |
| 1677 | // need to perform all type translation before beginning emission of .debug$S |
| 1678 | // and then make LocalUDTs a member of FunctionInfo |
| 1679 | } |
| 1680 | |
| 1681 | TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { |
| 1682 | // Generic dispatch for lowering an unknown type. |
| 1683 | switch (Ty->getTag()) { |
| 1684 | case dwarf::DW_TAG_array_type: |
| 1685 | return lowerTypeArray(Ty: cast<DICompositeType>(Val: Ty)); |
| 1686 | case dwarf::DW_TAG_typedef: |
| 1687 | return lowerTypeAlias(Ty: cast<DIDerivedType>(Val: Ty)); |
| 1688 | case dwarf::DW_TAG_base_type: |
| 1689 | return lowerTypeBasic(Ty: cast<DIBasicType>(Val: Ty)); |
| 1690 | case dwarf::DW_TAG_pointer_type: |
| 1691 | if (cast<DIDerivedType>(Val: Ty)->getName() == "__vtbl_ptr_type" ) |
| 1692 | return lowerTypeVFTableShape(Ty: cast<DIDerivedType>(Val: Ty)); |
| 1693 | [[fallthrough]]; |
| 1694 | case dwarf::DW_TAG_reference_type: |
| 1695 | case dwarf::DW_TAG_rvalue_reference_type: |
| 1696 | return lowerTypePointer(Ty: cast<DIDerivedType>(Val: Ty)); |
| 1697 | case dwarf::DW_TAG_ptr_to_member_type: |
| 1698 | return lowerTypeMemberPointer(Ty: cast<DIDerivedType>(Val: Ty)); |
| 1699 | case dwarf::DW_TAG_restrict_type: |
| 1700 | case dwarf::DW_TAG_const_type: |
| 1701 | case dwarf::DW_TAG_volatile_type: |
| 1702 | // TODO: add support for DW_TAG_atomic_type here |
| 1703 | return lowerTypeModifier(Ty: cast<DIDerivedType>(Val: Ty)); |
| 1704 | case dwarf::DW_TAG_subroutine_type: |
| 1705 | if (ClassTy) { |
| 1706 | // The member function type of a member function pointer has no |
| 1707 | // ThisAdjustment. |
| 1708 | return lowerTypeMemberFunction(Ty: cast<DISubroutineType>(Val: Ty), ClassTy, |
| 1709 | /*ThisAdjustment=*/0, |
| 1710 | /*IsStaticMethod=*/false); |
| 1711 | } |
| 1712 | return lowerTypeFunction(Ty: cast<DISubroutineType>(Val: Ty)); |
| 1713 | case dwarf::DW_TAG_enumeration_type: |
| 1714 | return lowerTypeEnum(Ty: cast<DICompositeType>(Val: Ty)); |
| 1715 | case dwarf::DW_TAG_class_type: |
| 1716 | case dwarf::DW_TAG_structure_type: |
| 1717 | return lowerTypeClass(Ty: cast<DICompositeType>(Val: Ty)); |
| 1718 | case dwarf::DW_TAG_union_type: |
| 1719 | return lowerTypeUnion(Ty: cast<DICompositeType>(Val: Ty)); |
| 1720 | case dwarf::DW_TAG_string_type: |
| 1721 | return lowerTypeString(Ty: cast<DIStringType>(Val: Ty)); |
| 1722 | case dwarf::DW_TAG_unspecified_type: |
| 1723 | if (Ty->getName() == "decltype(nullptr)" ) |
| 1724 | return TypeIndex::NullptrT(); |
| 1725 | return TypeIndex::None(); |
| 1726 | default: |
| 1727 | // Use the null type index. |
| 1728 | return TypeIndex(); |
| 1729 | } |
| 1730 | } |
| 1731 | |
| 1732 | TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { |
| 1733 | TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty: Ty->getBaseType()); |
| 1734 | StringRef TypeName = Ty->getName(); |
| 1735 | |
| 1736 | addToUDTs(Ty); |
| 1737 | |
| 1738 | if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && |
| 1739 | TypeName == "HRESULT" ) |
| 1740 | return TypeIndex(SimpleTypeKind::HResult); |
| 1741 | if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && |
| 1742 | TypeName == "wchar_t" ) |
| 1743 | return TypeIndex(SimpleTypeKind::WideCharacter); |
| 1744 | |
| 1745 | return UnderlyingTypeIndex; |
| 1746 | } |
| 1747 | |
| 1748 | TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { |
| 1749 | const DIType *ElementType = Ty->getBaseType(); |
| 1750 | TypeIndex ElementTypeIndex = getTypeIndex(Ty: ElementType); |
| 1751 | // IndexType is size_t, which depends on the bitness of the target. |
| 1752 | TypeIndex IndexType = getPointerSizeInBytes() == 8 |
| 1753 | ? TypeIndex(SimpleTypeKind::UInt64Quad) |
| 1754 | : TypeIndex(SimpleTypeKind::UInt32Long); |
| 1755 | |
| 1756 | uint64_t ElementSize = getBaseTypeSize(Ty: ElementType) / 8; |
| 1757 | |
| 1758 | // Add subranges to array type. |
| 1759 | DINodeArray Elements = Ty->getElements(); |
| 1760 | for (int i = Elements.size() - 1; i >= 0; --i) { |
| 1761 | const DINode *Element = Elements[i]; |
| 1762 | assert(Element->getTag() == dwarf::DW_TAG_subrange_type); |
| 1763 | |
| 1764 | const DISubrange *Subrange = cast<DISubrange>(Val: Element); |
| 1765 | int64_t Count = -1; |
| 1766 | |
| 1767 | // If Subrange has a Count field, use it. |
| 1768 | // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), |
| 1769 | // where lowerbound is from the LowerBound field of the Subrange, |
| 1770 | // or the language default lowerbound if that field is unspecified. |
| 1771 | if (auto *CI = dyn_cast_if_present<ConstantInt *>(Val: Subrange->getCount())) |
| 1772 | Count = CI->getSExtValue(); |
| 1773 | else if (auto *UI = dyn_cast_if_present<ConstantInt *>( |
| 1774 | Val: Subrange->getUpperBound())) { |
| 1775 | // Fortran uses 1 as the default lowerbound; other languages use 0. |
| 1776 | int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; |
| 1777 | auto *LI = dyn_cast_if_present<ConstantInt *>(Val: Subrange->getLowerBound()); |
| 1778 | Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; |
| 1779 | Count = UI->getSExtValue() - Lowerbound + 1; |
| 1780 | } |
| 1781 | |
| 1782 | // Forward declarations of arrays without a size and VLAs use a count of -1. |
| 1783 | // Emit a count of zero in these cases to match what MSVC does for arrays |
| 1784 | // without a size. MSVC doesn't support VLAs, so it's not clear what we |
| 1785 | // should do for them even if we could distinguish them. |
| 1786 | if (Count == -1) |
| 1787 | Count = 0; |
| 1788 | |
| 1789 | // Update the element size and element type index for subsequent subranges. |
| 1790 | ElementSize *= Count; |
| 1791 | |
| 1792 | // If this is the outermost array, use the size from the array. It will be |
| 1793 | // more accurate if we had a VLA or an incomplete element type size. |
| 1794 | uint64_t ArraySize = |
| 1795 | (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; |
| 1796 | |
| 1797 | StringRef Name = (i == 0) ? Ty->getName() : "" ; |
| 1798 | ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); |
| 1799 | ElementTypeIndex = TypeTable.writeLeafType(Record&: AR); |
| 1800 | } |
| 1801 | |
| 1802 | return ElementTypeIndex; |
| 1803 | } |
| 1804 | |
| 1805 | // This function lowers a Fortran character type (DIStringType). |
| 1806 | // Note that it handles only the character*n variant (using SizeInBits |
| 1807 | // field in DIString to describe the type size) at the moment. |
| 1808 | // Other variants (leveraging the StringLength and StringLengthExp |
| 1809 | // fields in DIStringType) remain TBD. |
| 1810 | TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { |
| 1811 | TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); |
| 1812 | uint64_t ArraySize = Ty->getSizeInBits() >> 3; |
| 1813 | StringRef Name = Ty->getName(); |
| 1814 | // IndexType is size_t, which depends on the bitness of the target. |
| 1815 | TypeIndex IndexType = getPointerSizeInBytes() == 8 |
| 1816 | ? TypeIndex(SimpleTypeKind::UInt64Quad) |
| 1817 | : TypeIndex(SimpleTypeKind::UInt32Long); |
| 1818 | |
| 1819 | // Create a type of character array of ArraySize. |
| 1820 | ArrayRecord AR(CharType, IndexType, ArraySize, Name); |
| 1821 | |
| 1822 | return TypeTable.writeLeafType(Record&: AR); |
| 1823 | } |
| 1824 | |
| 1825 | TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { |
| 1826 | TypeIndex Index; |
| 1827 | dwarf::TypeKind Kind; |
| 1828 | uint32_t ByteSize; |
| 1829 | |
| 1830 | Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); |
| 1831 | ByteSize = Ty->getSizeInBits() / 8; |
| 1832 | |
| 1833 | SimpleTypeKind STK = SimpleTypeKind::None; |
| 1834 | switch (Kind) { |
| 1835 | case dwarf::DW_ATE_address: |
| 1836 | // FIXME: Translate |
| 1837 | break; |
| 1838 | case dwarf::DW_ATE_boolean: |
| 1839 | switch (ByteSize) { |
| 1840 | case 1: STK = SimpleTypeKind::Boolean8; break; |
| 1841 | case 2: STK = SimpleTypeKind::Boolean16; break; |
| 1842 | case 4: STK = SimpleTypeKind::Boolean32; break; |
| 1843 | case 8: STK = SimpleTypeKind::Boolean64; break; |
| 1844 | case 16: STK = SimpleTypeKind::Boolean128; break; |
| 1845 | } |
| 1846 | break; |
| 1847 | case dwarf::DW_ATE_complex_float: |
| 1848 | // The CodeView size for a complex represents the size of |
| 1849 | // an individual component. |
| 1850 | switch (ByteSize) { |
| 1851 | case 4: STK = SimpleTypeKind::Complex16; break; |
| 1852 | case 8: STK = SimpleTypeKind::Complex32; break; |
| 1853 | case 16: STK = SimpleTypeKind::Complex64; break; |
| 1854 | case 20: STK = SimpleTypeKind::Complex80; break; |
| 1855 | case 32: STK = SimpleTypeKind::Complex128; break; |
| 1856 | } |
| 1857 | break; |
| 1858 | case dwarf::DW_ATE_float: |
| 1859 | switch (ByteSize) { |
| 1860 | case 2: STK = SimpleTypeKind::Float16; break; |
| 1861 | case 4: STK = SimpleTypeKind::Float32; break; |
| 1862 | case 6: STK = SimpleTypeKind::Float48; break; |
| 1863 | case 8: STK = SimpleTypeKind::Float64; break; |
| 1864 | case 10: STK = SimpleTypeKind::Float80; break; |
| 1865 | case 16: STK = SimpleTypeKind::Float128; break; |
| 1866 | } |
| 1867 | break; |
| 1868 | case dwarf::DW_ATE_signed: |
| 1869 | switch (ByteSize) { |
| 1870 | case 1: STK = SimpleTypeKind::SignedCharacter; break; |
| 1871 | case 2: STK = SimpleTypeKind::Int16Short; break; |
| 1872 | case 4: STK = SimpleTypeKind::Int32; break; |
| 1873 | case 8: STK = SimpleTypeKind::Int64Quad; break; |
| 1874 | case 16: STK = SimpleTypeKind::Int128Oct; break; |
| 1875 | } |
| 1876 | break; |
| 1877 | case dwarf::DW_ATE_unsigned: |
| 1878 | switch (ByteSize) { |
| 1879 | case 1: STK = SimpleTypeKind::UnsignedCharacter; break; |
| 1880 | case 2: STK = SimpleTypeKind::UInt16Short; break; |
| 1881 | case 4: STK = SimpleTypeKind::UInt32; break; |
| 1882 | case 8: STK = SimpleTypeKind::UInt64Quad; break; |
| 1883 | case 16: STK = SimpleTypeKind::UInt128Oct; break; |
| 1884 | } |
| 1885 | break; |
| 1886 | case dwarf::DW_ATE_UTF: |
| 1887 | switch (ByteSize) { |
| 1888 | case 1: STK = SimpleTypeKind::Character8; break; |
| 1889 | case 2: STK = SimpleTypeKind::Character16; break; |
| 1890 | case 4: STK = SimpleTypeKind::Character32; break; |
| 1891 | } |
| 1892 | break; |
| 1893 | case dwarf::DW_ATE_signed_char: |
| 1894 | if (ByteSize == 1) |
| 1895 | STK = SimpleTypeKind::SignedCharacter; |
| 1896 | break; |
| 1897 | case dwarf::DW_ATE_unsigned_char: |
| 1898 | if (ByteSize == 1) |
| 1899 | STK = SimpleTypeKind::UnsignedCharacter; |
| 1900 | break; |
| 1901 | default: |
| 1902 | break; |
| 1903 | } |
| 1904 | |
| 1905 | // Apply some fixups based on the source-level type name. |
| 1906 | // Include some amount of canonicalization from an old naming scheme Clang |
| 1907 | // used to use for integer types (in an outdated effort to be compatible with |
| 1908 | // GCC's debug info/GDB's behavior, which has since been addressed). |
| 1909 | if (STK == SimpleTypeKind::Int32 && |
| 1910 | (Ty->getName() == "long int" || Ty->getName() == "long" )) |
| 1911 | STK = SimpleTypeKind::Int32Long; |
| 1912 | if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || |
| 1913 | Ty->getName() == "unsigned long" )) |
| 1914 | STK = SimpleTypeKind::UInt32Long; |
| 1915 | if (STK == SimpleTypeKind::UInt16Short && |
| 1916 | (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t" )) |
| 1917 | STK = SimpleTypeKind::WideCharacter; |
| 1918 | if ((STK == SimpleTypeKind::SignedCharacter || |
| 1919 | STK == SimpleTypeKind::UnsignedCharacter) && |
| 1920 | Ty->getName() == "char" ) |
| 1921 | STK = SimpleTypeKind::NarrowCharacter; |
| 1922 | |
| 1923 | return TypeIndex(STK); |
| 1924 | } |
| 1925 | |
| 1926 | TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, |
| 1927 | PointerOptions PO) { |
| 1928 | TypeIndex PointeeTI = getTypeIndex(Ty: Ty->getBaseType()); |
| 1929 | |
| 1930 | // Pointers to simple types without any options can use SimpleTypeMode, rather |
| 1931 | // than having a dedicated pointer type record. |
| 1932 | if (PointeeTI.isSimple() && PO == PointerOptions::None && |
| 1933 | PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && |
| 1934 | Ty->getTag() == dwarf::DW_TAG_pointer_type) { |
| 1935 | SimpleTypeMode Mode = Ty->getSizeInBits() == 64 |
| 1936 | ? SimpleTypeMode::NearPointer64 |
| 1937 | : SimpleTypeMode::NearPointer32; |
| 1938 | return TypeIndex(PointeeTI.getSimpleKind(), Mode); |
| 1939 | } |
| 1940 | |
| 1941 | PointerKind PK = |
| 1942 | Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; |
| 1943 | PointerMode PM = PointerMode::Pointer; |
| 1944 | switch (Ty->getTag()) { |
| 1945 | default: llvm_unreachable("not a pointer tag type" ); |
| 1946 | case dwarf::DW_TAG_pointer_type: |
| 1947 | PM = PointerMode::Pointer; |
| 1948 | break; |
| 1949 | case dwarf::DW_TAG_reference_type: |
| 1950 | PM = PointerMode::LValueReference; |
| 1951 | break; |
| 1952 | case dwarf::DW_TAG_rvalue_reference_type: |
| 1953 | PM = PointerMode::RValueReference; |
| 1954 | break; |
| 1955 | } |
| 1956 | |
| 1957 | if (Ty->isObjectPointer()) |
| 1958 | PO |= PointerOptions::Const; |
| 1959 | |
| 1960 | PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); |
| 1961 | return TypeTable.writeLeafType(Record&: PR); |
| 1962 | } |
| 1963 | |
| 1964 | static PointerToMemberRepresentation |
| 1965 | translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { |
| 1966 | // SizeInBytes being zero generally implies that the member pointer type was |
| 1967 | // incomplete, which can happen if it is part of a function prototype. In this |
| 1968 | // case, use the unknown model instead of the general model. |
| 1969 | if (IsPMF) { |
| 1970 | switch (Flags & DINode::FlagPtrToMemberRep) { |
| 1971 | case 0: |
| 1972 | return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown |
| 1973 | : PointerToMemberRepresentation::GeneralFunction; |
| 1974 | case DINode::FlagSingleInheritance: |
| 1975 | return PointerToMemberRepresentation::SingleInheritanceFunction; |
| 1976 | case DINode::FlagMultipleInheritance: |
| 1977 | return PointerToMemberRepresentation::MultipleInheritanceFunction; |
| 1978 | case DINode::FlagVirtualInheritance: |
| 1979 | return PointerToMemberRepresentation::VirtualInheritanceFunction; |
| 1980 | } |
| 1981 | } else { |
| 1982 | switch (Flags & DINode::FlagPtrToMemberRep) { |
| 1983 | case 0: |
| 1984 | return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown |
| 1985 | : PointerToMemberRepresentation::GeneralData; |
| 1986 | case DINode::FlagSingleInheritance: |
| 1987 | return PointerToMemberRepresentation::SingleInheritanceData; |
| 1988 | case DINode::FlagMultipleInheritance: |
| 1989 | return PointerToMemberRepresentation::MultipleInheritanceData; |
| 1990 | case DINode::FlagVirtualInheritance: |
| 1991 | return PointerToMemberRepresentation::VirtualInheritanceData; |
| 1992 | } |
| 1993 | } |
| 1994 | llvm_unreachable("invalid ptr to member representation" ); |
| 1995 | } |
| 1996 | |
| 1997 | TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, |
| 1998 | PointerOptions PO) { |
| 1999 | assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); |
| 2000 | bool IsPMF = isa<DISubroutineType>(Val: Ty->getBaseType()); |
| 2001 | TypeIndex ClassTI = getTypeIndex(Ty: Ty->getClassType()); |
| 2002 | TypeIndex PointeeTI = |
| 2003 | getTypeIndex(Ty: Ty->getBaseType(), ClassTy: IsPMF ? Ty->getClassType() : nullptr); |
| 2004 | PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 |
| 2005 | : PointerKind::Near32; |
| 2006 | PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction |
| 2007 | : PointerMode::PointerToDataMember; |
| 2008 | |
| 2009 | assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big" ); |
| 2010 | uint8_t SizeInBytes = Ty->getSizeInBits() / 8; |
| 2011 | MemberPointerInfo MPI( |
| 2012 | ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Flags: Ty->getFlags())); |
| 2013 | PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); |
| 2014 | return TypeTable.writeLeafType(Record&: PR); |
| 2015 | } |
| 2016 | |
| 2017 | /// Given a DWARF calling convention, get the CodeView equivalent. If we don't |
| 2018 | /// have a translation, use the NearC convention. |
| 2019 | static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { |
| 2020 | switch (DwarfCC) { |
| 2021 | case dwarf::DW_CC_normal: return CallingConvention::NearC; |
| 2022 | case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; |
| 2023 | case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; |
| 2024 | case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; |
| 2025 | case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; |
| 2026 | case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; |
| 2027 | } |
| 2028 | return CallingConvention::NearC; |
| 2029 | } |
| 2030 | |
| 2031 | TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { |
| 2032 | ModifierOptions Mods = ModifierOptions::None; |
| 2033 | PointerOptions PO = PointerOptions::None; |
| 2034 | bool IsModifier = true; |
| 2035 | const DIType *BaseTy = Ty; |
| 2036 | while (IsModifier && BaseTy) { |
| 2037 | // FIXME: Need to add DWARF tags for __unaligned and _Atomic |
| 2038 | switch (BaseTy->getTag()) { |
| 2039 | case dwarf::DW_TAG_const_type: |
| 2040 | Mods |= ModifierOptions::Const; |
| 2041 | PO |= PointerOptions::Const; |
| 2042 | break; |
| 2043 | case dwarf::DW_TAG_volatile_type: |
| 2044 | Mods |= ModifierOptions::Volatile; |
| 2045 | PO |= PointerOptions::Volatile; |
| 2046 | break; |
| 2047 | case dwarf::DW_TAG_restrict_type: |
| 2048 | // Only pointer types be marked with __restrict. There is no known flag |
| 2049 | // for __restrict in LF_MODIFIER records. |
| 2050 | PO |= PointerOptions::Restrict; |
| 2051 | break; |
| 2052 | default: |
| 2053 | IsModifier = false; |
| 2054 | break; |
| 2055 | } |
| 2056 | if (IsModifier) |
| 2057 | BaseTy = cast<DIDerivedType>(Val: BaseTy)->getBaseType(); |
| 2058 | } |
| 2059 | |
| 2060 | // Check if the inner type will use an LF_POINTER record. If so, the |
| 2061 | // qualifiers will go in the LF_POINTER record. This comes up for types like |
| 2062 | // 'int *const' and 'int *__restrict', not the more common cases like 'const |
| 2063 | // char *'. |
| 2064 | if (BaseTy) { |
| 2065 | switch (BaseTy->getTag()) { |
| 2066 | case dwarf::DW_TAG_pointer_type: |
| 2067 | case dwarf::DW_TAG_reference_type: |
| 2068 | case dwarf::DW_TAG_rvalue_reference_type: |
| 2069 | return lowerTypePointer(Ty: cast<DIDerivedType>(Val: BaseTy), PO); |
| 2070 | case dwarf::DW_TAG_ptr_to_member_type: |
| 2071 | return lowerTypeMemberPointer(Ty: cast<DIDerivedType>(Val: BaseTy), PO); |
| 2072 | default: |
| 2073 | break; |
| 2074 | } |
| 2075 | } |
| 2076 | |
| 2077 | TypeIndex ModifiedTI = getTypeIndex(Ty: BaseTy); |
| 2078 | |
| 2079 | // Return the base type index if there aren't any modifiers. For example, the |
| 2080 | // metadata could contain restrict wrappers around non-pointer types. |
| 2081 | if (Mods == ModifierOptions::None) |
| 2082 | return ModifiedTI; |
| 2083 | |
| 2084 | ModifierRecord MR(ModifiedTI, Mods); |
| 2085 | return TypeTable.writeLeafType(Record&: MR); |
| 2086 | } |
| 2087 | |
| 2088 | TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { |
| 2089 | SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; |
| 2090 | for (const DIType *ArgType : Ty->getTypeArray()) |
| 2091 | ReturnAndArgTypeIndices.push_back(Elt: getTypeIndex(Ty: ArgType)); |
| 2092 | |
| 2093 | // MSVC uses type none for variadic argument. |
| 2094 | if (ReturnAndArgTypeIndices.size() > 1 && |
| 2095 | ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { |
| 2096 | ReturnAndArgTypeIndices.back() = TypeIndex::None(); |
| 2097 | } |
| 2098 | TypeIndex ReturnTypeIndex = TypeIndex::Void(); |
| 2099 | ArrayRef<TypeIndex> ArgTypeIndices = {}; |
| 2100 | if (!ReturnAndArgTypeIndices.empty()) { |
| 2101 | auto ReturnAndArgTypesRef = ArrayRef(ReturnAndArgTypeIndices); |
| 2102 | ReturnTypeIndex = ReturnAndArgTypesRef.consume_front(); |
| 2103 | ArgTypeIndices = ReturnAndArgTypesRef; |
| 2104 | } |
| 2105 | |
| 2106 | ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); |
| 2107 | TypeIndex ArgListIndex = TypeTable.writeLeafType(Record&: ArgListRec); |
| 2108 | |
| 2109 | CallingConvention CC = dwarfCCToCodeView(DwarfCC: Ty->getCC()); |
| 2110 | |
| 2111 | FunctionOptions FO = getFunctionOptions(Ty); |
| 2112 | ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), |
| 2113 | ArgListIndex); |
| 2114 | return TypeTable.writeLeafType(Record&: Procedure); |
| 2115 | } |
| 2116 | |
| 2117 | TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, |
| 2118 | const DIType *ClassTy, |
| 2119 | int ThisAdjustment, |
| 2120 | bool IsStaticMethod, |
| 2121 | FunctionOptions FO) { |
| 2122 | // Lower the containing class type. |
| 2123 | TypeIndex ClassType = getTypeIndex(Ty: ClassTy); |
| 2124 | |
| 2125 | DITypeArray ReturnAndArgs = Ty->getTypeArray(); |
| 2126 | |
| 2127 | unsigned Index = 0; |
| 2128 | SmallVector<TypeIndex, 8> ArgTypeIndices; |
| 2129 | TypeIndex ReturnTypeIndex = TypeIndex::Void(); |
| 2130 | if (ReturnAndArgs.size() > Index) { |
| 2131 | ReturnTypeIndex = getTypeIndex(Ty: ReturnAndArgs[Index++]); |
| 2132 | } |
| 2133 | |
| 2134 | // If the first argument is a pointer type and this isn't a static method, |
| 2135 | // treat it as the special 'this' parameter, which is encoded separately from |
| 2136 | // the arguments. |
| 2137 | TypeIndex ThisTypeIndex; |
| 2138 | if (!IsStaticMethod && ReturnAndArgs.size() > Index) { |
| 2139 | if (const DIDerivedType *PtrTy = |
| 2140 | dyn_cast_or_null<DIDerivedType>(Val: ReturnAndArgs[Index])) { |
| 2141 | if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { |
| 2142 | ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, SubroutineTy: Ty); |
| 2143 | Index++; |
| 2144 | } |
| 2145 | } |
| 2146 | } |
| 2147 | |
| 2148 | while (Index < ReturnAndArgs.size()) |
| 2149 | ArgTypeIndices.push_back(Elt: getTypeIndex(Ty: ReturnAndArgs[Index++])); |
| 2150 | |
| 2151 | // MSVC uses type none for variadic argument. |
| 2152 | if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) |
| 2153 | ArgTypeIndices.back() = TypeIndex::None(); |
| 2154 | |
| 2155 | ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); |
| 2156 | TypeIndex ArgListIndex = TypeTable.writeLeafType(Record&: ArgListRec); |
| 2157 | |
| 2158 | CallingConvention CC = dwarfCCToCodeView(DwarfCC: Ty->getCC()); |
| 2159 | |
| 2160 | MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, |
| 2161 | ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); |
| 2162 | return TypeTable.writeLeafType(Record&: MFR); |
| 2163 | } |
| 2164 | |
| 2165 | TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { |
| 2166 | unsigned VSlotCount = |
| 2167 | Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); |
| 2168 | SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); |
| 2169 | |
| 2170 | VFTableShapeRecord VFTSR(Slots); |
| 2171 | return TypeTable.writeLeafType(Record&: VFTSR); |
| 2172 | } |
| 2173 | |
| 2174 | static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { |
| 2175 | switch (Flags & DINode::FlagAccessibility) { |
| 2176 | case DINode::FlagPrivate: return MemberAccess::Private; |
| 2177 | case DINode::FlagPublic: return MemberAccess::Public; |
| 2178 | case DINode::FlagProtected: return MemberAccess::Protected; |
| 2179 | case 0: |
| 2180 | // If there was no explicit access control, provide the default for the tag. |
| 2181 | return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private |
| 2182 | : MemberAccess::Public; |
| 2183 | } |
| 2184 | llvm_unreachable("access flags are exclusive" ); |
| 2185 | } |
| 2186 | |
| 2187 | static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { |
| 2188 | if (SP->isArtificial()) |
| 2189 | return MethodOptions::CompilerGenerated; |
| 2190 | |
| 2191 | // FIXME: Handle other MethodOptions. |
| 2192 | |
| 2193 | return MethodOptions::None; |
| 2194 | } |
| 2195 | |
| 2196 | static MethodKind translateMethodKindFlags(const DISubprogram *SP, |
| 2197 | bool Introduced) { |
| 2198 | if (SP->getFlags() & DINode::FlagStaticMember) |
| 2199 | return MethodKind::Static; |
| 2200 | |
| 2201 | switch (SP->getVirtuality()) { |
| 2202 | case dwarf::DW_VIRTUALITY_none: |
| 2203 | break; |
| 2204 | case dwarf::DW_VIRTUALITY_virtual: |
| 2205 | return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; |
| 2206 | case dwarf::DW_VIRTUALITY_pure_virtual: |
| 2207 | return Introduced ? MethodKind::PureIntroducingVirtual |
| 2208 | : MethodKind::PureVirtual; |
| 2209 | default: |
| 2210 | llvm_unreachable("unhandled virtuality case" ); |
| 2211 | } |
| 2212 | |
| 2213 | return MethodKind::Vanilla; |
| 2214 | } |
| 2215 | |
| 2216 | static TypeRecordKind getRecordKind(const DICompositeType *Ty) { |
| 2217 | switch (Ty->getTag()) { |
| 2218 | case dwarf::DW_TAG_class_type: |
| 2219 | return TypeRecordKind::Class; |
| 2220 | case dwarf::DW_TAG_structure_type: |
| 2221 | return TypeRecordKind::Struct; |
| 2222 | default: |
| 2223 | llvm_unreachable("unexpected tag" ); |
| 2224 | } |
| 2225 | } |
| 2226 | |
| 2227 | /// Return ClassOptions that should be present on both the forward declaration |
| 2228 | /// and the defintion of a tag type. |
| 2229 | static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { |
| 2230 | ClassOptions CO = ClassOptions::None; |
| 2231 | |
| 2232 | // MSVC always sets this flag, even for local types. Clang doesn't always |
| 2233 | // appear to give every type a linkage name, which may be problematic for us. |
| 2234 | // FIXME: Investigate the consequences of not following them here. |
| 2235 | if (!Ty->getIdentifier().empty()) |
| 2236 | CO |= ClassOptions::HasUniqueName; |
| 2237 | |
| 2238 | // Put the Nested flag on a type if it appears immediately inside a tag type. |
| 2239 | // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass |
| 2240 | // here. That flag is only set on definitions, and not forward declarations. |
| 2241 | const DIScope *ImmediateScope = Ty->getScope(); |
| 2242 | if (ImmediateScope && isa<DICompositeType>(Val: ImmediateScope)) |
| 2243 | CO |= ClassOptions::Nested; |
| 2244 | |
| 2245 | // Put the Scoped flag on function-local types. MSVC puts this flag for enum |
| 2246 | // type only when it has an immediate function scope. Clang never puts enums |
| 2247 | // inside DILexicalBlock scopes. Enum types, as generated by clang, are |
| 2248 | // always in function, class, or file scopes. |
| 2249 | if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { |
| 2250 | if (ImmediateScope && isa<DISubprogram>(Val: ImmediateScope)) |
| 2251 | CO |= ClassOptions::Scoped; |
| 2252 | } else { |
| 2253 | for (const DIScope *Scope = ImmediateScope; Scope != nullptr; |
| 2254 | Scope = Scope->getScope()) { |
| 2255 | if (isa<DISubprogram>(Val: Scope)) { |
| 2256 | CO |= ClassOptions::Scoped; |
| 2257 | break; |
| 2258 | } |
| 2259 | } |
| 2260 | } |
| 2261 | |
| 2262 | return CO; |
| 2263 | } |
| 2264 | |
| 2265 | void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { |
| 2266 | switch (Ty->getTag()) { |
| 2267 | case dwarf::DW_TAG_class_type: |
| 2268 | case dwarf::DW_TAG_structure_type: |
| 2269 | case dwarf::DW_TAG_union_type: |
| 2270 | case dwarf::DW_TAG_enumeration_type: |
| 2271 | break; |
| 2272 | default: |
| 2273 | return; |
| 2274 | } |
| 2275 | |
| 2276 | if (const auto *File = Ty->getFile()) { |
| 2277 | StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); |
| 2278 | TypeIndex SIDI = TypeTable.writeLeafType(Record&: SIDR); |
| 2279 | |
| 2280 | UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); |
| 2281 | TypeTable.writeLeafType(Record&: USLR); |
| 2282 | } |
| 2283 | } |
| 2284 | |
| 2285 | TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { |
| 2286 | ClassOptions CO = getCommonClassOptions(Ty); |
| 2287 | TypeIndex FTI; |
| 2288 | unsigned EnumeratorCount = 0; |
| 2289 | |
| 2290 | if (Ty->isForwardDecl()) { |
| 2291 | CO |= ClassOptions::ForwardReference; |
| 2292 | } else { |
| 2293 | ContinuationRecordBuilder ContinuationBuilder; |
| 2294 | ContinuationBuilder.begin(RecordKind: ContinuationRecordKind::FieldList); |
| 2295 | for (const DINode *Element : Ty->getElements()) { |
| 2296 | // We assume that the frontend provides all members in source declaration |
| 2297 | // order, which is what MSVC does. |
| 2298 | if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Val: Element)) { |
| 2299 | // FIXME: Is it correct to always emit these as unsigned here? |
| 2300 | EnumeratorRecord ER(MemberAccess::Public, |
| 2301 | APSInt(Enumerator->getValue(), true), |
| 2302 | Enumerator->getName()); |
| 2303 | ContinuationBuilder.writeMemberType(Record&: ER); |
| 2304 | EnumeratorCount++; |
| 2305 | } |
| 2306 | } |
| 2307 | FTI = TypeTable.insertRecord(Builder&: ContinuationBuilder); |
| 2308 | } |
| 2309 | |
| 2310 | std::string FullName = getFullyQualifiedName(Ty); |
| 2311 | |
| 2312 | EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), |
| 2313 | getTypeIndex(Ty: Ty->getBaseType())); |
| 2314 | TypeIndex EnumTI = TypeTable.writeLeafType(Record&: ER); |
| 2315 | |
| 2316 | addUDTSrcLine(Ty, TI: EnumTI); |
| 2317 | |
| 2318 | return EnumTI; |
| 2319 | } |
| 2320 | |
| 2321 | //===----------------------------------------------------------------------===// |
| 2322 | // ClassInfo |
| 2323 | //===----------------------------------------------------------------------===// |
| 2324 | |
| 2325 | struct llvm::ClassInfo { |
| 2326 | struct MemberInfo { |
| 2327 | const DIDerivedType *MemberTypeNode; |
| 2328 | uint64_t BaseOffset; |
| 2329 | }; |
| 2330 | // [MemberInfo] |
| 2331 | using MemberList = std::vector<MemberInfo>; |
| 2332 | |
| 2333 | using MethodsList = TinyPtrVector<const DISubprogram *>; |
| 2334 | // MethodName -> MethodsList |
| 2335 | using MethodsMap = MapVector<MDString *, MethodsList>; |
| 2336 | |
| 2337 | /// Base classes. |
| 2338 | std::vector<const DIDerivedType *> Inheritance; |
| 2339 | |
| 2340 | /// Direct members. |
| 2341 | MemberList Members; |
| 2342 | // Direct overloaded methods gathered by name. |
| 2343 | MethodsMap Methods; |
| 2344 | |
| 2345 | TypeIndex VShapeTI; |
| 2346 | |
| 2347 | std::vector<const DIType *> NestedTypes; |
| 2348 | }; |
| 2349 | |
| 2350 | void CodeViewDebug::clear() { |
| 2351 | assert(CurFn == nullptr); |
| 2352 | FileIdMap.clear(); |
| 2353 | FnDebugInfo.clear(); |
| 2354 | FileToFilepathMap.clear(); |
| 2355 | LocalUDTs.clear(); |
| 2356 | GlobalUDTs.clear(); |
| 2357 | TypeIndices.clear(); |
| 2358 | CompleteTypeIndices.clear(); |
| 2359 | ScopeGlobals.clear(); |
| 2360 | CVGlobalVariableOffsets.clear(); |
| 2361 | } |
| 2362 | |
| 2363 | void CodeViewDebug::collectMemberInfo(ClassInfo &Info, |
| 2364 | const DIDerivedType *DDTy) { |
| 2365 | if (!DDTy->getName().empty()) { |
| 2366 | Info.Members.push_back(x: {.MemberTypeNode: DDTy, .BaseOffset: 0}); |
| 2367 | |
| 2368 | // Collect static const data members with values. |
| 2369 | if ((DDTy->getFlags() & DINode::FlagStaticMember) == |
| 2370 | DINode::FlagStaticMember) { |
| 2371 | if (DDTy->getConstant() && (isa<ConstantInt>(Val: DDTy->getConstant()) || |
| 2372 | isa<ConstantFP>(Val: DDTy->getConstant()))) |
| 2373 | StaticConstMembers.push_back(Elt: DDTy); |
| 2374 | } |
| 2375 | |
| 2376 | return; |
| 2377 | } |
| 2378 | |
| 2379 | // An unnamed member may represent a nested struct or union. Attempt to |
| 2380 | // interpret the unnamed member as a DICompositeType possibly wrapped in |
| 2381 | // qualifier types. Add all the indirect fields to the current record if that |
| 2382 | // succeeds, and drop the member if that fails. |
| 2383 | assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!" ); |
| 2384 | uint64_t Offset = DDTy->getOffsetInBits(); |
| 2385 | const DIType *Ty = DDTy->getBaseType(); |
| 2386 | bool FullyResolved = false; |
| 2387 | while (!FullyResolved) { |
| 2388 | switch (Ty->getTag()) { |
| 2389 | case dwarf::DW_TAG_const_type: |
| 2390 | case dwarf::DW_TAG_volatile_type: |
| 2391 | // FIXME: we should apply the qualifier types to the indirect fields |
| 2392 | // rather than dropping them. |
| 2393 | Ty = cast<DIDerivedType>(Val: Ty)->getBaseType(); |
| 2394 | break; |
| 2395 | default: |
| 2396 | FullyResolved = true; |
| 2397 | break; |
| 2398 | } |
| 2399 | } |
| 2400 | |
| 2401 | const DICompositeType *DCTy = dyn_cast<DICompositeType>(Val: Ty); |
| 2402 | if (!DCTy) |
| 2403 | return; |
| 2404 | |
| 2405 | ClassInfo NestedInfo = collectClassInfo(Ty: DCTy); |
| 2406 | for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) |
| 2407 | Info.Members.push_back( |
| 2408 | x: {.MemberTypeNode: IndirectField.MemberTypeNode, .BaseOffset: IndirectField.BaseOffset + Offset}); |
| 2409 | } |
| 2410 | |
| 2411 | ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { |
| 2412 | ClassInfo Info; |
| 2413 | // Add elements to structure type. |
| 2414 | DINodeArray Elements = Ty->getElements(); |
| 2415 | for (auto *Element : Elements) { |
| 2416 | // We assume that the frontend provides all members in source declaration |
| 2417 | // order, which is what MSVC does. |
| 2418 | if (!Element) |
| 2419 | continue; |
| 2420 | if (auto *SP = dyn_cast<DISubprogram>(Val: Element)) { |
| 2421 | Info.Methods[SP->getRawName()].push_back(NewVal: SP); |
| 2422 | } else if (auto *DDTy = dyn_cast<DIDerivedType>(Val: Element)) { |
| 2423 | if (DDTy->getTag() == dwarf::DW_TAG_member) { |
| 2424 | collectMemberInfo(Info, DDTy); |
| 2425 | } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { |
| 2426 | Info.Inheritance.push_back(x: DDTy); |
| 2427 | } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && |
| 2428 | DDTy->getName() == "__vtbl_ptr_type" ) { |
| 2429 | Info.VShapeTI = getTypeIndex(Ty: DDTy); |
| 2430 | } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { |
| 2431 | Info.NestedTypes.push_back(x: DDTy); |
| 2432 | } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { |
| 2433 | // Ignore friend members. It appears that MSVC emitted info about |
| 2434 | // friends in the past, but modern versions do not. |
| 2435 | } |
| 2436 | } else if (auto *Composite = dyn_cast<DICompositeType>(Val: Element)) { |
| 2437 | Info.NestedTypes.push_back(x: Composite); |
| 2438 | } |
| 2439 | // Skip other unrecognized kinds of elements. |
| 2440 | } |
| 2441 | return Info; |
| 2442 | } |
| 2443 | |
| 2444 | static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { |
| 2445 | // This routine is used by lowerTypeClass and lowerTypeUnion to determine |
| 2446 | // if a complete type should be emitted instead of a forward reference. |
| 2447 | return Ty->getName().empty() && Ty->getIdentifier().empty() && |
| 2448 | !Ty->isForwardDecl(); |
| 2449 | } |
| 2450 | |
| 2451 | TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { |
| 2452 | // Emit the complete type for unnamed structs. C++ classes with methods |
| 2453 | // which have a circular reference back to the class type are expected to |
| 2454 | // be named by the front-end and should not be "unnamed". C unnamed |
| 2455 | // structs should not have circular references. |
| 2456 | if (shouldAlwaysEmitCompleteClassType(Ty)) { |
| 2457 | // If this unnamed complete type is already in the process of being defined |
| 2458 | // then the description of the type is malformed and cannot be emitted |
| 2459 | // into CodeView correctly so report a fatal error. |
| 2460 | auto I = CompleteTypeIndices.find(Val: Ty); |
| 2461 | if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) |
| 2462 | report_fatal_error(reason: "cannot debug circular reference to unnamed type" ); |
| 2463 | return getCompleteTypeIndex(Ty); |
| 2464 | } |
| 2465 | |
| 2466 | // First, construct the forward decl. Don't look into Ty to compute the |
| 2467 | // forward decl options, since it might not be available in all TUs. |
| 2468 | TypeRecordKind Kind = getRecordKind(Ty); |
| 2469 | ClassOptions CO = |
| 2470 | ClassOptions::ForwardReference | getCommonClassOptions(Ty); |
| 2471 | std::string FullName = getFullyQualifiedName(Ty); |
| 2472 | ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, |
| 2473 | FullName, Ty->getIdentifier()); |
| 2474 | TypeIndex FwdDeclTI = TypeTable.writeLeafType(Record&: CR); |
| 2475 | if (!Ty->isForwardDecl()) |
| 2476 | DeferredCompleteTypes.push_back(Elt: Ty); |
| 2477 | return FwdDeclTI; |
| 2478 | } |
| 2479 | |
| 2480 | TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { |
| 2481 | // Construct the field list and complete type record. |
| 2482 | TypeRecordKind Kind = getRecordKind(Ty); |
| 2483 | ClassOptions CO = getCommonClassOptions(Ty); |
| 2484 | TypeIndex FieldTI; |
| 2485 | TypeIndex VShapeTI; |
| 2486 | unsigned FieldCount; |
| 2487 | bool ContainsNestedClass; |
| 2488 | std::tie(args&: FieldTI, args&: VShapeTI, args&: FieldCount, args&: ContainsNestedClass) = |
| 2489 | lowerRecordFieldList(Ty); |
| 2490 | |
| 2491 | if (ContainsNestedClass) |
| 2492 | CO |= ClassOptions::ContainsNestedClass; |
| 2493 | |
| 2494 | // MSVC appears to set this flag by searching any destructor or method with |
| 2495 | // FunctionOptions::Constructor among the emitted members. Clang AST has all |
| 2496 | // the members, however special member functions are not yet emitted into |
| 2497 | // debug information. For now checking a class's non-triviality seems enough. |
| 2498 | // FIXME: not true for a nested unnamed struct. |
| 2499 | if (isNonTrivial(DCTy: Ty)) |
| 2500 | CO |= ClassOptions::HasConstructorOrDestructor; |
| 2501 | |
| 2502 | std::string FullName = getFullyQualifiedName(Ty); |
| 2503 | |
| 2504 | uint64_t SizeInBytes = Ty->getSizeInBits() / 8; |
| 2505 | |
| 2506 | ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, |
| 2507 | SizeInBytes, FullName, Ty->getIdentifier()); |
| 2508 | TypeIndex ClassTI = TypeTable.writeLeafType(Record&: CR); |
| 2509 | |
| 2510 | addUDTSrcLine(Ty, TI: ClassTI); |
| 2511 | |
| 2512 | addToUDTs(Ty); |
| 2513 | |
| 2514 | return ClassTI; |
| 2515 | } |
| 2516 | |
| 2517 | TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { |
| 2518 | // Emit the complete type for unnamed unions. |
| 2519 | if (shouldAlwaysEmitCompleteClassType(Ty)) |
| 2520 | return getCompleteTypeIndex(Ty); |
| 2521 | |
| 2522 | ClassOptions CO = |
| 2523 | ClassOptions::ForwardReference | getCommonClassOptions(Ty); |
| 2524 | std::string FullName = getFullyQualifiedName(Ty); |
| 2525 | UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); |
| 2526 | TypeIndex FwdDeclTI = TypeTable.writeLeafType(Record&: UR); |
| 2527 | if (!Ty->isForwardDecl()) |
| 2528 | DeferredCompleteTypes.push_back(Elt: Ty); |
| 2529 | return FwdDeclTI; |
| 2530 | } |
| 2531 | |
| 2532 | TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { |
| 2533 | ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); |
| 2534 | TypeIndex FieldTI; |
| 2535 | unsigned FieldCount; |
| 2536 | bool ContainsNestedClass; |
| 2537 | std::tie(args&: FieldTI, args: std::ignore, args&: FieldCount, args&: ContainsNestedClass) = |
| 2538 | lowerRecordFieldList(Ty); |
| 2539 | |
| 2540 | if (ContainsNestedClass) |
| 2541 | CO |= ClassOptions::ContainsNestedClass; |
| 2542 | |
| 2543 | uint64_t SizeInBytes = Ty->getSizeInBits() / 8; |
| 2544 | std::string FullName = getFullyQualifiedName(Ty); |
| 2545 | |
| 2546 | UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, |
| 2547 | Ty->getIdentifier()); |
| 2548 | TypeIndex UnionTI = TypeTable.writeLeafType(Record&: UR); |
| 2549 | |
| 2550 | addUDTSrcLine(Ty, TI: UnionTI); |
| 2551 | |
| 2552 | addToUDTs(Ty); |
| 2553 | |
| 2554 | return UnionTI; |
| 2555 | } |
| 2556 | |
| 2557 | std::tuple<TypeIndex, TypeIndex, unsigned, bool> |
| 2558 | CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { |
| 2559 | // Manually count members. MSVC appears to count everything that generates a |
| 2560 | // field list record. Each individual overload in a method overload group |
| 2561 | // contributes to this count, even though the overload group is a single field |
| 2562 | // list record. |
| 2563 | unsigned MemberCount = 0; |
| 2564 | ClassInfo Info = collectClassInfo(Ty); |
| 2565 | ContinuationRecordBuilder ContinuationBuilder; |
| 2566 | ContinuationBuilder.begin(RecordKind: ContinuationRecordKind::FieldList); |
| 2567 | |
| 2568 | // Create base classes. |
| 2569 | for (const DIDerivedType *I : Info.Inheritance) { |
| 2570 | if (I->getFlags() & DINode::FlagVirtual) { |
| 2571 | // Virtual base. |
| 2572 | unsigned VBPtrOffset = I->getVBPtrOffset(); |
| 2573 | // FIXME: Despite the accessor name, the offset is really in bytes. |
| 2574 | unsigned VBTableIndex = I->getOffsetInBits() / 4; |
| 2575 | auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase |
| 2576 | ? TypeRecordKind::IndirectVirtualBaseClass |
| 2577 | : TypeRecordKind::VirtualBaseClass; |
| 2578 | VirtualBaseClassRecord VBCR( |
| 2579 | RecordKind, translateAccessFlags(RecordTag: Ty->getTag(), Flags: I->getFlags()), |
| 2580 | getTypeIndex(Ty: I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, |
| 2581 | VBTableIndex); |
| 2582 | |
| 2583 | ContinuationBuilder.writeMemberType(Record&: VBCR); |
| 2584 | MemberCount++; |
| 2585 | } else { |
| 2586 | assert(I->getOffsetInBits() % 8 == 0 && |
| 2587 | "bases must be on byte boundaries" ); |
| 2588 | BaseClassRecord BCR(translateAccessFlags(RecordTag: Ty->getTag(), Flags: I->getFlags()), |
| 2589 | getTypeIndex(Ty: I->getBaseType()), |
| 2590 | I->getOffsetInBits() / 8); |
| 2591 | ContinuationBuilder.writeMemberType(Record&: BCR); |
| 2592 | MemberCount++; |
| 2593 | } |
| 2594 | } |
| 2595 | |
| 2596 | // Create members. |
| 2597 | for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { |
| 2598 | const DIDerivedType *Member = MemberInfo.MemberTypeNode; |
| 2599 | TypeIndex MemberBaseType = getTypeIndex(Ty: Member->getBaseType()); |
| 2600 | StringRef MemberName = Member->getName(); |
| 2601 | MemberAccess Access = |
| 2602 | translateAccessFlags(RecordTag: Ty->getTag(), Flags: Member->getFlags()); |
| 2603 | |
| 2604 | if (Member->isStaticMember()) { |
| 2605 | StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); |
| 2606 | ContinuationBuilder.writeMemberType(Record&: SDMR); |
| 2607 | MemberCount++; |
| 2608 | continue; |
| 2609 | } |
| 2610 | |
| 2611 | // Virtual function pointer member. |
| 2612 | if ((Member->getFlags() & DINode::FlagArtificial) && |
| 2613 | Member->getName().starts_with(Prefix: "_vptr$" )) { |
| 2614 | VFPtrRecord VFPR(getTypeIndex(Ty: Member->getBaseType())); |
| 2615 | ContinuationBuilder.writeMemberType(Record&: VFPR); |
| 2616 | MemberCount++; |
| 2617 | continue; |
| 2618 | } |
| 2619 | |
| 2620 | // Data member. |
| 2621 | uint64_t MemberOffsetInBits = |
| 2622 | Member->getOffsetInBits() + MemberInfo.BaseOffset; |
| 2623 | if (Member->isBitField()) { |
| 2624 | uint64_t StartBitOffset = MemberOffsetInBits; |
| 2625 | if (const auto *CI = |
| 2626 | dyn_cast_or_null<ConstantInt>(Val: Member->getStorageOffsetInBits())) { |
| 2627 | MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; |
| 2628 | } |
| 2629 | StartBitOffset -= MemberOffsetInBits; |
| 2630 | BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), |
| 2631 | StartBitOffset); |
| 2632 | MemberBaseType = TypeTable.writeLeafType(Record&: BFR); |
| 2633 | } |
| 2634 | uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; |
| 2635 | DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, |
| 2636 | MemberName); |
| 2637 | ContinuationBuilder.writeMemberType(Record&: DMR); |
| 2638 | MemberCount++; |
| 2639 | } |
| 2640 | |
| 2641 | // Create methods |
| 2642 | for (auto &MethodItr : Info.Methods) { |
| 2643 | StringRef Name = MethodItr.first->getString(); |
| 2644 | |
| 2645 | std::vector<OneMethodRecord> Methods; |
| 2646 | for (const DISubprogram *SP : MethodItr.second) { |
| 2647 | TypeIndex MethodType = getMemberFunctionType(SP, Class: Ty); |
| 2648 | bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; |
| 2649 | |
| 2650 | unsigned VFTableOffset = -1; |
| 2651 | if (Introduced) |
| 2652 | VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); |
| 2653 | |
| 2654 | Methods.push_back(x: OneMethodRecord( |
| 2655 | MethodType, translateAccessFlags(RecordTag: Ty->getTag(), Flags: SP->getFlags()), |
| 2656 | translateMethodKindFlags(SP, Introduced), |
| 2657 | translateMethodOptionFlags(SP), VFTableOffset, Name)); |
| 2658 | MemberCount++; |
| 2659 | } |
| 2660 | assert(!Methods.empty() && "Empty methods map entry" ); |
| 2661 | if (Methods.size() == 1) |
| 2662 | ContinuationBuilder.writeMemberType(Record&: Methods[0]); |
| 2663 | else { |
| 2664 | // FIXME: Make this use its own ContinuationBuilder so that |
| 2665 | // MethodOverloadList can be split correctly. |
| 2666 | MethodOverloadListRecord MOLR(Methods); |
| 2667 | TypeIndex MethodList = TypeTable.writeLeafType(Record&: MOLR); |
| 2668 | |
| 2669 | OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); |
| 2670 | ContinuationBuilder.writeMemberType(Record&: OMR); |
| 2671 | } |
| 2672 | } |
| 2673 | |
| 2674 | // Create nested classes. |
| 2675 | for (const DIType *Nested : Info.NestedTypes) { |
| 2676 | NestedTypeRecord R(getTypeIndex(Ty: Nested), Nested->getName()); |
| 2677 | ContinuationBuilder.writeMemberType(Record&: R); |
| 2678 | MemberCount++; |
| 2679 | } |
| 2680 | |
| 2681 | TypeIndex FieldTI = TypeTable.insertRecord(Builder&: ContinuationBuilder); |
| 2682 | return std::make_tuple(args&: FieldTI, args&: Info.VShapeTI, args&: MemberCount, |
| 2683 | args: !Info.NestedTypes.empty()); |
| 2684 | } |
| 2685 | |
| 2686 | TypeIndex CodeViewDebug::getVBPTypeIndex() { |
| 2687 | if (!VBPType.getIndex()) { |
| 2688 | // Make a 'const int *' type. |
| 2689 | ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); |
| 2690 | TypeIndex ModifiedTI = TypeTable.writeLeafType(Record&: MR); |
| 2691 | |
| 2692 | PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 |
| 2693 | : PointerKind::Near32; |
| 2694 | PointerMode PM = PointerMode::Pointer; |
| 2695 | PointerOptions PO = PointerOptions::None; |
| 2696 | PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); |
| 2697 | VBPType = TypeTable.writeLeafType(Record&: PR); |
| 2698 | } |
| 2699 | |
| 2700 | return VBPType; |
| 2701 | } |
| 2702 | |
| 2703 | TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { |
| 2704 | // The null DIType is the void type. Don't try to hash it. |
| 2705 | if (!Ty) |
| 2706 | return TypeIndex::Void(); |
| 2707 | |
| 2708 | // Check if we've already translated this type. Don't try to do a |
| 2709 | // get-or-create style insertion that caches the hash lookup across the |
| 2710 | // lowerType call. It will update the TypeIndices map. |
| 2711 | auto I = TypeIndices.find(Val: {Ty, ClassTy}); |
| 2712 | if (I != TypeIndices.end()) |
| 2713 | return I->second; |
| 2714 | |
| 2715 | TypeLoweringScope S(*this); |
| 2716 | TypeIndex TI = lowerType(Ty, ClassTy); |
| 2717 | return recordTypeIndexForDINode(Node: Ty, TI, ClassTy); |
| 2718 | } |
| 2719 | |
| 2720 | codeview::TypeIndex |
| 2721 | CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, |
| 2722 | const DISubroutineType *SubroutineTy) { |
| 2723 | assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && |
| 2724 | "this type must be a pointer type" ); |
| 2725 | |
| 2726 | PointerOptions Options = PointerOptions::None; |
| 2727 | if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) |
| 2728 | Options = PointerOptions::LValueRefThisPointer; |
| 2729 | else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) |
| 2730 | Options = PointerOptions::RValueRefThisPointer; |
| 2731 | |
| 2732 | // Check if we've already translated this type. If there is no ref qualifier |
| 2733 | // on the function then we look up this pointer type with no associated class |
| 2734 | // so that the TypeIndex for the this pointer can be shared with the type |
| 2735 | // index for other pointers to this class type. If there is a ref qualifier |
| 2736 | // then we lookup the pointer using the subroutine as the parent type. |
| 2737 | auto I = TypeIndices.find(Val: {PtrTy, SubroutineTy}); |
| 2738 | if (I != TypeIndices.end()) |
| 2739 | return I->second; |
| 2740 | |
| 2741 | TypeLoweringScope S(*this); |
| 2742 | TypeIndex TI = lowerTypePointer(Ty: PtrTy, PO: Options); |
| 2743 | return recordTypeIndexForDINode(Node: PtrTy, TI, ClassTy: SubroutineTy); |
| 2744 | } |
| 2745 | |
| 2746 | TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { |
| 2747 | PointerRecord PR(getTypeIndex(Ty), |
| 2748 | getPointerSizeInBytes() == 8 ? PointerKind::Near64 |
| 2749 | : PointerKind::Near32, |
| 2750 | PointerMode::LValueReference, PointerOptions::None, |
| 2751 | Ty->getSizeInBits() / 8); |
| 2752 | return TypeTable.writeLeafType(Record&: PR); |
| 2753 | } |
| 2754 | |
| 2755 | TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { |
| 2756 | // The null DIType is the void type. Don't try to hash it. |
| 2757 | if (!Ty) |
| 2758 | return TypeIndex::Void(); |
| 2759 | |
| 2760 | // Look through typedefs when getting the complete type index. Call |
| 2761 | // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are |
| 2762 | // emitted only once. |
| 2763 | if (Ty->getTag() == dwarf::DW_TAG_typedef) |
| 2764 | (void)getTypeIndex(Ty); |
| 2765 | while (Ty->getTag() == dwarf::DW_TAG_typedef) |
| 2766 | Ty = cast<DIDerivedType>(Val: Ty)->getBaseType(); |
| 2767 | |
| 2768 | // If this is a non-record type, the complete type index is the same as the |
| 2769 | // normal type index. Just call getTypeIndex. |
| 2770 | switch (Ty->getTag()) { |
| 2771 | case dwarf::DW_TAG_class_type: |
| 2772 | case dwarf::DW_TAG_structure_type: |
| 2773 | case dwarf::DW_TAG_union_type: |
| 2774 | break; |
| 2775 | default: |
| 2776 | return getTypeIndex(Ty); |
| 2777 | } |
| 2778 | |
| 2779 | const auto *CTy = cast<DICompositeType>(Val: Ty); |
| 2780 | |
| 2781 | TypeLoweringScope S(*this); |
| 2782 | |
| 2783 | // Make sure the forward declaration is emitted first. It's unclear if this |
| 2784 | // is necessary, but MSVC does it, and we should follow suit until we can show |
| 2785 | // otherwise. |
| 2786 | // We only emit a forward declaration for named types. |
| 2787 | if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { |
| 2788 | TypeIndex FwdDeclTI = getTypeIndex(Ty: CTy); |
| 2789 | |
| 2790 | // Just use the forward decl if we don't have complete type info. This |
| 2791 | // might happen if the frontend is using modules and expects the complete |
| 2792 | // definition to be emitted elsewhere. |
| 2793 | if (CTy->isForwardDecl()) |
| 2794 | return FwdDeclTI; |
| 2795 | } |
| 2796 | |
| 2797 | // Check if we've already translated the complete record type. |
| 2798 | // Insert the type with a null TypeIndex to signify that the type is currently |
| 2799 | // being lowered. |
| 2800 | auto InsertResult = CompleteTypeIndices.try_emplace(Key: CTy); |
| 2801 | if (!InsertResult.second) |
| 2802 | return InsertResult.first->second; |
| 2803 | |
| 2804 | TypeIndex TI; |
| 2805 | switch (CTy->getTag()) { |
| 2806 | case dwarf::DW_TAG_class_type: |
| 2807 | case dwarf::DW_TAG_structure_type: |
| 2808 | TI = lowerCompleteTypeClass(Ty: CTy); |
| 2809 | break; |
| 2810 | case dwarf::DW_TAG_union_type: |
| 2811 | TI = lowerCompleteTypeUnion(Ty: CTy); |
| 2812 | break; |
| 2813 | default: |
| 2814 | llvm_unreachable("not a record" ); |
| 2815 | } |
| 2816 | |
| 2817 | // Update the type index associated with this CompositeType. This cannot |
| 2818 | // use the 'InsertResult' iterator above because it is potentially |
| 2819 | // invalidated by map insertions which can occur while lowering the class |
| 2820 | // type above. |
| 2821 | CompleteTypeIndices[CTy] = TI; |
| 2822 | return TI; |
| 2823 | } |
| 2824 | |
| 2825 | /// Emit all the deferred complete record types. Try to do this in FIFO order, |
| 2826 | /// and do this until fixpoint, as each complete record type typically |
| 2827 | /// references |
| 2828 | /// many other record types. |
| 2829 | void CodeViewDebug::emitDeferredCompleteTypes() { |
| 2830 | SmallVector<const DICompositeType *, 4> TypesToEmit; |
| 2831 | while (!DeferredCompleteTypes.empty()) { |
| 2832 | std::swap(LHS&: DeferredCompleteTypes, RHS&: TypesToEmit); |
| 2833 | for (const DICompositeType *RecordTy : TypesToEmit) |
| 2834 | getCompleteTypeIndex(Ty: RecordTy); |
| 2835 | TypesToEmit.clear(); |
| 2836 | } |
| 2837 | } |
| 2838 | |
| 2839 | void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, |
| 2840 | ArrayRef<LocalVariable> Locals) { |
| 2841 | // Get the sorted list of parameters and emit them first. |
| 2842 | SmallVector<const LocalVariable *, 6> Params; |
| 2843 | for (const LocalVariable &L : Locals) |
| 2844 | if (L.DIVar->isParameter()) |
| 2845 | Params.push_back(Elt: &L); |
| 2846 | llvm::sort(C&: Params, Comp: [](const LocalVariable *L, const LocalVariable *R) { |
| 2847 | return L->DIVar->getArg() < R->DIVar->getArg(); |
| 2848 | }); |
| 2849 | for (const LocalVariable *L : Params) |
| 2850 | emitLocalVariable(FI, Var: *L); |
| 2851 | |
| 2852 | // Next emit all non-parameters in the order that we found them. |
| 2853 | for (const LocalVariable &L : Locals) { |
| 2854 | if (!L.DIVar->isParameter()) { |
| 2855 | if (L.ConstantValue) { |
| 2856 | // If ConstantValue is set we will emit it as a S_CONSTANT instead of a |
| 2857 | // S_LOCAL in order to be able to represent it at all. |
| 2858 | const DIType *Ty = L.DIVar->getType(); |
| 2859 | APSInt Val(*L.ConstantValue); |
| 2860 | emitConstantSymbolRecord(DTy: Ty, Value&: Val, QualifiedName: std::string(L.DIVar->getName())); |
| 2861 | } else { |
| 2862 | emitLocalVariable(FI, Var: L); |
| 2863 | } |
| 2864 | } |
| 2865 | } |
| 2866 | } |
| 2867 | |
| 2868 | void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, |
| 2869 | const LocalVariable &Var) { |
| 2870 | // LocalSym record, see SymbolRecord.h for more info. |
| 2871 | MCSymbol *LocalEnd = beginSymbolRecord(Kind: SymbolKind::S_LOCAL); |
| 2872 | |
| 2873 | LocalSymFlags Flags = LocalSymFlags::None; |
| 2874 | if (Var.DIVar->isParameter()) |
| 2875 | Flags |= LocalSymFlags::IsParameter; |
| 2876 | if (Var.DefRanges.empty()) |
| 2877 | Flags |= LocalSymFlags::IsOptimizedOut; |
| 2878 | |
| 2879 | OS.AddComment(T: "TypeIndex" ); |
| 2880 | TypeIndex TI = Var.UseReferenceType |
| 2881 | ? getTypeIndexForReferenceTo(Ty: Var.DIVar->getType()) |
| 2882 | : getCompleteTypeIndex(Ty: Var.DIVar->getType()); |
| 2883 | OS.emitInt32(Value: TI.getIndex()); |
| 2884 | OS.AddComment(T: "Flags" ); |
| 2885 | OS.emitInt16(Value: static_cast<uint16_t>(Flags)); |
| 2886 | // Truncate the name so we won't overflow the record length field. |
| 2887 | emitNullTerminatedSymbolName(OS, S: Var.DIVar->getName()); |
| 2888 | endSymbolRecord(SymEnd: LocalEnd); |
| 2889 | |
| 2890 | // Calculate the on disk prefix of the appropriate def range record. The |
| 2891 | // records and on disk formats are described in SymbolRecords.h. BytePrefix |
| 2892 | // should be big enough to hold all forms without memory allocation. |
| 2893 | SmallString<20> BytePrefix; |
| 2894 | for (const auto &Pair : Var.DefRanges) { |
| 2895 | LocalVarDef DefRange = Pair.first; |
| 2896 | const auto &Ranges = Pair.second; |
| 2897 | BytePrefix.clear(); |
| 2898 | if (DefRange.InMemory) { |
| 2899 | int Offset = DefRange.DataOffset; |
| 2900 | unsigned Reg = DefRange.CVRegister; |
| 2901 | |
| 2902 | // 32-bit x86 call sequences often use PUSH instructions, which disrupt |
| 2903 | // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, |
| 2904 | // instead. In frames without stack realignment, $T0 will be the CFA. |
| 2905 | if (RegisterId(Reg) == RegisterId::ESP) { |
| 2906 | Reg = unsigned(RegisterId::VFRAME); |
| 2907 | Offset += FI.OffsetAdjustment; |
| 2908 | } |
| 2909 | |
| 2910 | // If we can use the chosen frame pointer for the frame and this isn't a |
| 2911 | // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. |
| 2912 | // Otherwise, use S_DEFRANGE_REGISTER_REL. |
| 2913 | EncodedFramePtrReg EncFP = encodeFramePtrReg(Reg: RegisterId(Reg), CPU: TheCPU); |
| 2914 | if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && |
| 2915 | (bool(Flags & LocalSymFlags::IsParameter) |
| 2916 | ? (EncFP == FI.EncodedParamFramePtrReg) |
| 2917 | : (EncFP == FI.EncodedLocalFramePtrReg))) { |
| 2918 | DefRangeFramePointerRelHeader DRHdr; |
| 2919 | DRHdr.Offset = Offset; |
| 2920 | OS.emitCVDefRangeDirective(Ranges, DRHdr); |
| 2921 | } else { |
| 2922 | uint16_t RegRelFlags = 0; |
| 2923 | if (DefRange.IsSubfield) { |
| 2924 | RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | |
| 2925 | (DefRange.StructOffset |
| 2926 | << DefRangeRegisterRelSym::OffsetInParentShift); |
| 2927 | } |
| 2928 | DefRangeRegisterRelHeader DRHdr; |
| 2929 | DRHdr.Register = Reg; |
| 2930 | DRHdr.Flags = RegRelFlags; |
| 2931 | DRHdr.BasePointerOffset = Offset; |
| 2932 | OS.emitCVDefRangeDirective(Ranges, DRHdr); |
| 2933 | } |
| 2934 | } else { |
| 2935 | assert(DefRange.DataOffset == 0 && "unexpected offset into register" ); |
| 2936 | if (DefRange.IsSubfield) { |
| 2937 | DefRangeSubfieldRegisterHeader DRHdr; |
| 2938 | DRHdr.Register = DefRange.CVRegister; |
| 2939 | DRHdr.MayHaveNoName = 0; |
| 2940 | DRHdr.OffsetInParent = DefRange.StructOffset; |
| 2941 | OS.emitCVDefRangeDirective(Ranges, DRHdr); |
| 2942 | } else { |
| 2943 | DefRangeRegisterHeader DRHdr; |
| 2944 | DRHdr.Register = DefRange.CVRegister; |
| 2945 | DRHdr.MayHaveNoName = 0; |
| 2946 | OS.emitCVDefRangeDirective(Ranges, DRHdr); |
| 2947 | } |
| 2948 | } |
| 2949 | } |
| 2950 | } |
| 2951 | |
| 2952 | void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, |
| 2953 | const FunctionInfo& FI) { |
| 2954 | for (LexicalBlock *Block : Blocks) |
| 2955 | emitLexicalBlock(Block: *Block, FI); |
| 2956 | } |
| 2957 | |
| 2958 | /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a |
| 2959 | /// lexical block scope. |
| 2960 | void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, |
| 2961 | const FunctionInfo& FI) { |
| 2962 | MCSymbol *RecordEnd = beginSymbolRecord(Kind: SymbolKind::S_BLOCK32); |
| 2963 | OS.AddComment(T: "PtrParent" ); |
| 2964 | OS.emitInt32(Value: 0); // PtrParent |
| 2965 | OS.AddComment(T: "PtrEnd" ); |
| 2966 | OS.emitInt32(Value: 0); // PtrEnd |
| 2967 | OS.AddComment(T: "Code size" ); |
| 2968 | OS.emitAbsoluteSymbolDiff(Hi: Block.End, Lo: Block.Begin, Size: 4); // Code Size |
| 2969 | OS.AddComment(T: "Function section relative address" ); |
| 2970 | OS.emitCOFFSecRel32(Symbol: Block.Begin, /*Offset=*/0); // Func Offset |
| 2971 | OS.AddComment(T: "Function section index" ); |
| 2972 | OS.emitCOFFSectionIndex(Symbol: FI.Begin); // Func Symbol |
| 2973 | OS.AddComment(T: "Lexical block name" ); |
| 2974 | emitNullTerminatedSymbolName(OS, S: Block.Name); // Name |
| 2975 | endSymbolRecord(SymEnd: RecordEnd); |
| 2976 | |
| 2977 | // Emit variables local to this lexical block. |
| 2978 | emitLocalVariableList(FI, Locals: Block.Locals); |
| 2979 | emitGlobalVariableList(Globals: Block.Globals); |
| 2980 | |
| 2981 | // Emit lexical blocks contained within this block. |
| 2982 | emitLexicalBlockList(Blocks: Block.Children, FI); |
| 2983 | |
| 2984 | // Close the lexical block scope. |
| 2985 | emitEndSymbolRecord(EndKind: SymbolKind::S_END); |
| 2986 | } |
| 2987 | |
| 2988 | /// Convenience routine for collecting lexical block information for a list |
| 2989 | /// of lexical scopes. |
| 2990 | void CodeViewDebug::collectLexicalBlockInfo( |
| 2991 | SmallVectorImpl<LexicalScope *> &Scopes, |
| 2992 | SmallVectorImpl<LexicalBlock *> &Blocks, |
| 2993 | SmallVectorImpl<LocalVariable> &Locals, |
| 2994 | SmallVectorImpl<CVGlobalVariable> &Globals) { |
| 2995 | for (LexicalScope *Scope : Scopes) |
| 2996 | collectLexicalBlockInfo(Scope&: *Scope, ParentBlocks&: Blocks, ParentLocals&: Locals, ParentGlobals&: Globals); |
| 2997 | } |
| 2998 | |
| 2999 | /// Populate the lexical blocks and local variable lists of the parent with |
| 3000 | /// information about the specified lexical scope. |
| 3001 | void CodeViewDebug::collectLexicalBlockInfo( |
| 3002 | LexicalScope &Scope, |
| 3003 | SmallVectorImpl<LexicalBlock *> &ParentBlocks, |
| 3004 | SmallVectorImpl<LocalVariable> &ParentLocals, |
| 3005 | SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { |
| 3006 | if (Scope.isAbstractScope()) |
| 3007 | return; |
| 3008 | |
| 3009 | // Gather information about the lexical scope including local variables, |
| 3010 | // global variables, and address ranges. |
| 3011 | bool IgnoreScope = false; |
| 3012 | auto LI = ScopeVariables.find(Val: &Scope); |
| 3013 | SmallVectorImpl<LocalVariable> *Locals = |
| 3014 | LI != ScopeVariables.end() ? &LI->second : nullptr; |
| 3015 | auto GI = ScopeGlobals.find(Val: Scope.getScopeNode()); |
| 3016 | SmallVectorImpl<CVGlobalVariable> *Globals = |
| 3017 | GI != ScopeGlobals.end() ? GI->second.get() : nullptr; |
| 3018 | const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Val: Scope.getScopeNode()); |
| 3019 | const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); |
| 3020 | |
| 3021 | // Ignore lexical scopes which do not contain variables. |
| 3022 | if (!Locals && !Globals) |
| 3023 | IgnoreScope = true; |
| 3024 | |
| 3025 | // Ignore lexical scopes which are not lexical blocks. |
| 3026 | if (!DILB) |
| 3027 | IgnoreScope = true; |
| 3028 | |
| 3029 | // Ignore scopes which have too many address ranges to represent in the |
| 3030 | // current CodeView format or do not have a valid address range. |
| 3031 | // |
| 3032 | // For lexical scopes with multiple address ranges you may be tempted to |
| 3033 | // construct a single range covering every instruction where the block is |
| 3034 | // live and everything in between. Unfortunately, Visual Studio only |
| 3035 | // displays variables from the first matching lexical block scope. If the |
| 3036 | // first lexical block contains exception handling code or cold code which |
| 3037 | // is moved to the bottom of the routine creating a single range covering |
| 3038 | // nearly the entire routine, then it will hide all other lexical blocks |
| 3039 | // and the variables they contain. |
| 3040 | if (Ranges.size() != 1 || !getLabelAfterInsn(MI: Ranges.front().second)) |
| 3041 | IgnoreScope = true; |
| 3042 | |
| 3043 | if (IgnoreScope) { |
| 3044 | // This scope can be safely ignored and eliminating it will reduce the |
| 3045 | // size of the debug information. Be sure to collect any variable and scope |
| 3046 | // information from the this scope or any of its children and collapse them |
| 3047 | // into the parent scope. |
| 3048 | if (Locals) |
| 3049 | ParentLocals.append(in_start: Locals->begin(), in_end: Locals->end()); |
| 3050 | if (Globals) |
| 3051 | ParentGlobals.append(in_start: Globals->begin(), in_end: Globals->end()); |
| 3052 | collectLexicalBlockInfo(Scopes&: Scope.getChildren(), |
| 3053 | Blocks&: ParentBlocks, |
| 3054 | Locals&: ParentLocals, |
| 3055 | Globals&: ParentGlobals); |
| 3056 | return; |
| 3057 | } |
| 3058 | |
| 3059 | // Create a new CodeView lexical block for this lexical scope. If we've |
| 3060 | // seen this DILexicalBlock before then the scope tree is malformed and |
| 3061 | // we can handle this gracefully by not processing it a second time. |
| 3062 | auto BlockInsertion = CurFn->LexicalBlocks.try_emplace(k: DILB); |
| 3063 | if (!BlockInsertion.second) |
| 3064 | return; |
| 3065 | |
| 3066 | // Create a lexical block containing the variables and collect the |
| 3067 | // lexical block information for the children. |
| 3068 | const InsnRange &Range = Ranges.front(); |
| 3069 | assert(Range.first && Range.second); |
| 3070 | LexicalBlock &Block = BlockInsertion.first->second; |
| 3071 | Block.Begin = getLabelBeforeInsn(MI: Range.first); |
| 3072 | Block.End = getLabelAfterInsn(MI: Range.second); |
| 3073 | assert(Block.Begin && "missing label for scope begin" ); |
| 3074 | assert(Block.End && "missing label for scope end" ); |
| 3075 | Block.Name = DILB->getName(); |
| 3076 | if (Locals) |
| 3077 | Block.Locals = std::move(*Locals); |
| 3078 | if (Globals) |
| 3079 | Block.Globals = std::move(*Globals); |
| 3080 | ParentBlocks.push_back(Elt: &Block); |
| 3081 | collectLexicalBlockInfo(Scopes&: Scope.getChildren(), |
| 3082 | Blocks&: Block.Children, |
| 3083 | Locals&: Block.Locals, |
| 3084 | Globals&: Block.Globals); |
| 3085 | } |
| 3086 | |
| 3087 | void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { |
| 3088 | const Function &GV = MF->getFunction(); |
| 3089 | assert(FnDebugInfo.count(&GV)); |
| 3090 | assert(CurFn == FnDebugInfo[&GV].get()); |
| 3091 | |
| 3092 | collectVariableInfo(SP: GV.getSubprogram()); |
| 3093 | |
| 3094 | // Build the lexical block structure to emit for this routine. |
| 3095 | if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) |
| 3096 | collectLexicalBlockInfo(Scope&: *CFS, |
| 3097 | ParentBlocks&: CurFn->ChildBlocks, |
| 3098 | ParentLocals&: CurFn->Locals, |
| 3099 | ParentGlobals&: CurFn->Globals); |
| 3100 | |
| 3101 | // Clear the scope and variable information from the map which will not be |
| 3102 | // valid after we have finished processing this routine. This also prepares |
| 3103 | // the map for the subsequent routine. |
| 3104 | ScopeVariables.clear(); |
| 3105 | |
| 3106 | // Don't emit anything if we don't have any line tables. |
| 3107 | // Thunks are compiler-generated and probably won't have source correlation. |
| 3108 | if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { |
| 3109 | FnDebugInfo.erase(Key: &GV); |
| 3110 | CurFn = nullptr; |
| 3111 | return; |
| 3112 | } |
| 3113 | |
| 3114 | // Find heap alloc sites and add to list. |
| 3115 | for (const auto &MBB : *MF) { |
| 3116 | for (const auto &MI : MBB) { |
| 3117 | if (MDNode *MD = MI.getHeapAllocMarker()) { |
| 3118 | CurFn->HeapAllocSites.push_back(x: std::make_tuple(args: getLabelBeforeInsn(MI: &MI), |
| 3119 | args: getLabelAfterInsn(MI: &MI), |
| 3120 | args: dyn_cast<DIType>(Val: MD))); |
| 3121 | } |
| 3122 | } |
| 3123 | } |
| 3124 | |
| 3125 | bool isThumb = MMI->getModule()->getTargetTriple().getArch() == |
| 3126 | llvm::Triple::ArchType::thumb; |
| 3127 | collectDebugInfoForJumpTables(MF, isThumb); |
| 3128 | |
| 3129 | CurFn->Annotations = MF->getCodeViewAnnotations(); |
| 3130 | |
| 3131 | CurFn->End = Asm->getFunctionEnd(); |
| 3132 | |
| 3133 | CurFn = nullptr; |
| 3134 | } |
| 3135 | |
| 3136 | // Usable locations are valid with non-zero line numbers. A line number of zero |
| 3137 | // corresponds to optimized code that doesn't have a distinct source location. |
| 3138 | // In this case, we try to use the previous or next source location depending on |
| 3139 | // the context. |
| 3140 | static bool isUsableDebugLoc(DebugLoc DL) { |
| 3141 | return DL && DL.getLine() != 0; |
| 3142 | } |
| 3143 | |
| 3144 | void CodeViewDebug::beginInstruction(const MachineInstr *MI) { |
| 3145 | DebugHandlerBase::beginInstruction(MI); |
| 3146 | |
| 3147 | // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. |
| 3148 | if (!Asm || !CurFn || MI->isDebugInstr() || |
| 3149 | MI->getFlag(Flag: MachineInstr::FrameSetup)) |
| 3150 | return; |
| 3151 | |
| 3152 | // If the first instruction of a new MBB has no location, find the first |
| 3153 | // instruction with a location and use that. |
| 3154 | DebugLoc DL = MI->getDebugLoc(); |
| 3155 | if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { |
| 3156 | for (const auto &NextMI : *MI->getParent()) { |
| 3157 | if (NextMI.isDebugInstr()) |
| 3158 | continue; |
| 3159 | DL = NextMI.getDebugLoc(); |
| 3160 | if (isUsableDebugLoc(DL)) |
| 3161 | break; |
| 3162 | } |
| 3163 | // FIXME: Handle the case where the BB has no valid locations. This would |
| 3164 | // probably require doing a real dataflow analysis. |
| 3165 | } |
| 3166 | PrevInstBB = MI->getParent(); |
| 3167 | |
| 3168 | // If we still don't have a debug location, don't record a location. |
| 3169 | if (!isUsableDebugLoc(DL)) |
| 3170 | return; |
| 3171 | |
| 3172 | maybeRecordLocation(DL, MF: Asm->MF); |
| 3173 | } |
| 3174 | |
| 3175 | MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { |
| 3176 | MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), |
| 3177 | *EndLabel = MMI->getContext().createTempSymbol(); |
| 3178 | OS.emitInt32(Value: unsigned(Kind)); |
| 3179 | OS.AddComment(T: "Subsection size" ); |
| 3180 | OS.emitAbsoluteSymbolDiff(Hi: EndLabel, Lo: BeginLabel, Size: 4); |
| 3181 | OS.emitLabel(Symbol: BeginLabel); |
| 3182 | return EndLabel; |
| 3183 | } |
| 3184 | |
| 3185 | void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { |
| 3186 | OS.emitLabel(Symbol: EndLabel); |
| 3187 | // Every subsection must be aligned to a 4-byte boundary. |
| 3188 | OS.emitValueToAlignment(Alignment: Align(4)); |
| 3189 | } |
| 3190 | |
| 3191 | static StringRef getSymbolName(SymbolKind SymKind) { |
| 3192 | for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) |
| 3193 | if (EE.Value == SymKind) |
| 3194 | return EE.Name; |
| 3195 | return "" ; |
| 3196 | } |
| 3197 | |
| 3198 | MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { |
| 3199 | MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), |
| 3200 | *EndLabel = MMI->getContext().createTempSymbol(); |
| 3201 | OS.AddComment(T: "Record length" ); |
| 3202 | OS.emitAbsoluteSymbolDiff(Hi: EndLabel, Lo: BeginLabel, Size: 2); |
| 3203 | OS.emitLabel(Symbol: BeginLabel); |
| 3204 | if (OS.isVerboseAsm()) |
| 3205 | OS.AddComment(T: "Record kind: " + getSymbolName(SymKind)); |
| 3206 | OS.emitInt16(Value: unsigned(SymKind)); |
| 3207 | return EndLabel; |
| 3208 | } |
| 3209 | |
| 3210 | void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { |
| 3211 | // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid |
| 3212 | // an extra copy of every symbol record in LLD. This increases object file |
| 3213 | // size by less than 1% in the clang build, and is compatible with the Visual |
| 3214 | // C++ linker. |
| 3215 | OS.emitValueToAlignment(Alignment: Align(4)); |
| 3216 | OS.emitLabel(Symbol: SymEnd); |
| 3217 | } |
| 3218 | |
| 3219 | void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { |
| 3220 | OS.AddComment(T: "Record length" ); |
| 3221 | OS.emitInt16(Value: 2); |
| 3222 | if (OS.isVerboseAsm()) |
| 3223 | OS.AddComment(T: "Record kind: " + getSymbolName(SymKind: EndKind)); |
| 3224 | OS.emitInt16(Value: uint16_t(EndKind)); // Record Kind |
| 3225 | } |
| 3226 | |
| 3227 | void CodeViewDebug::emitDebugInfoForUDTs( |
| 3228 | const std::vector<std::pair<std::string, const DIType *>> &UDTs) { |
| 3229 | #ifndef NDEBUG |
| 3230 | size_t OriginalSize = UDTs.size(); |
| 3231 | #endif |
| 3232 | for (const auto &UDT : UDTs) { |
| 3233 | const DIType *T = UDT.second; |
| 3234 | assert(shouldEmitUdt(T)); |
| 3235 | MCSymbol *UDTRecordEnd = beginSymbolRecord(SymKind: SymbolKind::S_UDT); |
| 3236 | OS.AddComment(T: "Type" ); |
| 3237 | OS.emitInt32(Value: getCompleteTypeIndex(Ty: T).getIndex()); |
| 3238 | assert(OriginalSize == UDTs.size() && |
| 3239 | "getCompleteTypeIndex found new UDTs!" ); |
| 3240 | emitNullTerminatedSymbolName(OS, S: UDT.first); |
| 3241 | endSymbolRecord(SymEnd: UDTRecordEnd); |
| 3242 | } |
| 3243 | } |
| 3244 | |
| 3245 | void CodeViewDebug::collectGlobalVariableInfo() { |
| 3246 | DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> |
| 3247 | GlobalMap; |
| 3248 | for (const GlobalVariable &GV : MMI->getModule()->globals()) { |
| 3249 | SmallVector<DIGlobalVariableExpression *, 1> GVEs; |
| 3250 | GV.getDebugInfo(GVs&: GVEs); |
| 3251 | for (const auto *GVE : GVEs) |
| 3252 | GlobalMap[GVE] = &GV; |
| 3253 | } |
| 3254 | |
| 3255 | NamedMDNode *CUs = MMI->getModule()->getNamedMetadata(Name: "llvm.dbg.cu" ); |
| 3256 | for (const MDNode *Node : CUs->operands()) { |
| 3257 | const auto *CU = cast<DICompileUnit>(Val: Node); |
| 3258 | for (const auto *GVE : CU->getGlobalVariables()) { |
| 3259 | const DIGlobalVariable *DIGV = GVE->getVariable(); |
| 3260 | const DIExpression *DIE = GVE->getExpression(); |
| 3261 | // Don't emit string literals in CodeView, as the only useful parts are |
| 3262 | // generally the filename and line number, which isn't possible to output |
| 3263 | // in CodeView. String literals should be the only unnamed GlobalVariable |
| 3264 | // with debug info. |
| 3265 | if (DIGV->getName().empty()) continue; |
| 3266 | |
| 3267 | if ((DIE->getNumElements() == 2) && |
| 3268 | (DIE->getElement(I: 0) == dwarf::DW_OP_plus_uconst)) |
| 3269 | // Record the constant offset for the variable. |
| 3270 | // |
| 3271 | // A Fortran common block uses this idiom to encode the offset |
| 3272 | // of a variable from the common block's starting address. |
| 3273 | CVGlobalVariableOffsets.insert( |
| 3274 | KV: std::make_pair(x&: DIGV, y: DIE->getElement(I: 1))); |
| 3275 | |
| 3276 | // Emit constant global variables in a global symbol section. |
| 3277 | if (GlobalMap.count(Val: GVE) == 0 && DIE->isConstant()) { |
| 3278 | CVGlobalVariable CVGV = {.DIGV: DIGV, .GVInfo: DIE}; |
| 3279 | GlobalVariables.emplace_back(Args: std::move(CVGV)); |
| 3280 | } |
| 3281 | |
| 3282 | const auto *GV = GlobalMap.lookup(Val: GVE); |
| 3283 | if (!GV || GV->isDeclarationForLinker()) |
| 3284 | continue; |
| 3285 | |
| 3286 | DIScope *Scope = DIGV->getScope(); |
| 3287 | SmallVector<CVGlobalVariable, 1> *VariableList; |
| 3288 | if (Scope && isa<DILocalScope>(Val: Scope)) { |
| 3289 | // Locate a global variable list for this scope, creating one if |
| 3290 | // necessary. |
| 3291 | auto Insertion = ScopeGlobals.insert( |
| 3292 | KV: {Scope, std::unique_ptr<GlobalVariableList>()}); |
| 3293 | if (Insertion.second) |
| 3294 | Insertion.first->second = std::make_unique<GlobalVariableList>(); |
| 3295 | VariableList = Insertion.first->second.get(); |
| 3296 | } else if (GV->hasComdat()) |
| 3297 | // Emit this global variable into a COMDAT section. |
| 3298 | VariableList = &ComdatVariables; |
| 3299 | else |
| 3300 | // Emit this global variable in a single global symbol section. |
| 3301 | VariableList = &GlobalVariables; |
| 3302 | CVGlobalVariable CVGV = {.DIGV: DIGV, .GVInfo: GV}; |
| 3303 | VariableList->emplace_back(Args: std::move(CVGV)); |
| 3304 | } |
| 3305 | } |
| 3306 | } |
| 3307 | |
| 3308 | void CodeViewDebug::collectDebugInfoForGlobals() { |
| 3309 | for (const CVGlobalVariable &CVGV : GlobalVariables) { |
| 3310 | const DIGlobalVariable *DIGV = CVGV.DIGV; |
| 3311 | const DIScope *Scope = DIGV->getScope(); |
| 3312 | getCompleteTypeIndex(Ty: DIGV->getType()); |
| 3313 | getFullyQualifiedName(Scope, Name: DIGV->getName()); |
| 3314 | } |
| 3315 | |
| 3316 | for (const CVGlobalVariable &CVGV : ComdatVariables) { |
| 3317 | const DIGlobalVariable *DIGV = CVGV.DIGV; |
| 3318 | const DIScope *Scope = DIGV->getScope(); |
| 3319 | getCompleteTypeIndex(Ty: DIGV->getType()); |
| 3320 | getFullyQualifiedName(Scope, Name: DIGV->getName()); |
| 3321 | } |
| 3322 | } |
| 3323 | |
| 3324 | void CodeViewDebug::emitDebugInfoForGlobals() { |
| 3325 | // First, emit all globals that are not in a comdat in a single symbol |
| 3326 | // substream. MSVC doesn't like it if the substream is empty, so only open |
| 3327 | // it if we have at least one global to emit. |
| 3328 | switchToDebugSectionForSymbol(GVSym: nullptr); |
| 3329 | if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { |
| 3330 | OS.AddComment(T: "Symbol subsection for globals" ); |
| 3331 | MCSymbol *EndLabel = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 3332 | emitGlobalVariableList(Globals: GlobalVariables); |
| 3333 | emitStaticConstMemberList(); |
| 3334 | endCVSubsection(EndLabel); |
| 3335 | } |
| 3336 | |
| 3337 | // Second, emit each global that is in a comdat into its own .debug$S |
| 3338 | // section along with its own symbol substream. |
| 3339 | for (const CVGlobalVariable &CVGV : ComdatVariables) { |
| 3340 | const GlobalVariable *GV = cast<const GlobalVariable *>(Val: CVGV.GVInfo); |
| 3341 | MCSymbol *GVSym = Asm->getSymbol(GV); |
| 3342 | OS.AddComment(T: "Symbol subsection for " + |
| 3343 | Twine(GlobalValue::dropLLVMManglingEscape(Name: GV->getName()))); |
| 3344 | switchToDebugSectionForSymbol(GVSym); |
| 3345 | MCSymbol *EndLabel = beginCVSubsection(Kind: DebugSubsectionKind::Symbols); |
| 3346 | // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. |
| 3347 | emitDebugInfoForGlobal(CVGV); |
| 3348 | endCVSubsection(EndLabel); |
| 3349 | } |
| 3350 | } |
| 3351 | |
| 3352 | void CodeViewDebug::emitDebugInfoForRetainedTypes() { |
| 3353 | NamedMDNode *CUs = MMI->getModule()->getNamedMetadata(Name: "llvm.dbg.cu" ); |
| 3354 | for (const MDNode *Node : CUs->operands()) { |
| 3355 | for (auto *Ty : cast<DICompileUnit>(Val: Node)->getRetainedTypes()) { |
| 3356 | if (DIType *RT = dyn_cast<DIType>(Val: Ty)) { |
| 3357 | getTypeIndex(Ty: RT); |
| 3358 | // FIXME: Add to global/local DTU list. |
| 3359 | } |
| 3360 | } |
| 3361 | } |
| 3362 | } |
| 3363 | |
| 3364 | // Emit each global variable in the specified array. |
| 3365 | void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { |
| 3366 | for (const CVGlobalVariable &CVGV : Globals) { |
| 3367 | // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. |
| 3368 | emitDebugInfoForGlobal(CVGV); |
| 3369 | } |
| 3370 | } |
| 3371 | |
| 3372 | void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, |
| 3373 | const std::string &QualifiedName) { |
| 3374 | MCSymbol *SConstantEnd = beginSymbolRecord(SymKind: SymbolKind::S_CONSTANT); |
| 3375 | OS.AddComment(T: "Type" ); |
| 3376 | OS.emitInt32(Value: getTypeIndex(Ty: DTy).getIndex()); |
| 3377 | |
| 3378 | OS.AddComment(T: "Value" ); |
| 3379 | |
| 3380 | // Encoded integers shouldn't need more than 10 bytes. |
| 3381 | uint8_t Data[10]; |
| 3382 | BinaryStreamWriter Writer(Data, llvm::endianness::little); |
| 3383 | CodeViewRecordIO IO(Writer); |
| 3384 | cantFail(Err: IO.mapEncodedInteger(Value)); |
| 3385 | StringRef SRef((char *)Data, Writer.getOffset()); |
| 3386 | OS.emitBinaryData(Data: SRef); |
| 3387 | |
| 3388 | OS.AddComment(T: "Name" ); |
| 3389 | emitNullTerminatedSymbolName(OS, S: QualifiedName); |
| 3390 | endSymbolRecord(SymEnd: SConstantEnd); |
| 3391 | } |
| 3392 | |
| 3393 | void CodeViewDebug::emitStaticConstMemberList() { |
| 3394 | for (const DIDerivedType *DTy : StaticConstMembers) { |
| 3395 | const DIScope *Scope = DTy->getScope(); |
| 3396 | |
| 3397 | APSInt Value; |
| 3398 | if (const ConstantInt *CI = |
| 3399 | dyn_cast_or_null<ConstantInt>(Val: DTy->getConstant())) |
| 3400 | Value = APSInt(CI->getValue(), |
| 3401 | DebugHandlerBase::isUnsignedDIType(Ty: DTy->getBaseType())); |
| 3402 | else if (const ConstantFP *CFP = |
| 3403 | dyn_cast_or_null<ConstantFP>(Val: DTy->getConstant())) |
| 3404 | Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); |
| 3405 | else |
| 3406 | llvm_unreachable("cannot emit a constant without a value" ); |
| 3407 | |
| 3408 | emitConstantSymbolRecord(DTy: DTy->getBaseType(), Value, |
| 3409 | QualifiedName: getFullyQualifiedName(Scope, Name: DTy->getName())); |
| 3410 | } |
| 3411 | } |
| 3412 | |
| 3413 | static bool isFloatDIType(const DIType *Ty) { |
| 3414 | if (isa<DICompositeType>(Val: Ty)) |
| 3415 | return false; |
| 3416 | |
| 3417 | if (auto *DTy = dyn_cast<DIDerivedType>(Val: Ty)) { |
| 3418 | dwarf::Tag T = (dwarf::Tag)Ty->getTag(); |
| 3419 | if (T == dwarf::DW_TAG_pointer_type || |
| 3420 | T == dwarf::DW_TAG_ptr_to_member_type || |
| 3421 | T == dwarf::DW_TAG_reference_type || |
| 3422 | T == dwarf::DW_TAG_rvalue_reference_type) |
| 3423 | return false; |
| 3424 | assert(DTy->getBaseType() && "Expected valid base type" ); |
| 3425 | return isFloatDIType(Ty: DTy->getBaseType()); |
| 3426 | } |
| 3427 | |
| 3428 | auto *BTy = cast<DIBasicType>(Val: Ty); |
| 3429 | return (BTy->getEncoding() == dwarf::DW_ATE_float); |
| 3430 | } |
| 3431 | |
| 3432 | void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { |
| 3433 | const DIGlobalVariable *DIGV = CVGV.DIGV; |
| 3434 | |
| 3435 | const DIScope *Scope = DIGV->getScope(); |
| 3436 | // For static data members, get the scope from the declaration. |
| 3437 | if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( |
| 3438 | Val: DIGV->getRawStaticDataMemberDeclaration())) |
| 3439 | Scope = MemberDecl->getScope(); |
| 3440 | // For static local variables and Fortran, the scoping portion is elided |
| 3441 | // in its name so that we can reference the variable in the command line |
| 3442 | // of the VS debugger. |
| 3443 | std::string QualifiedName = |
| 3444 | (moduleIsInFortran() || (Scope && isa<DILocalScope>(Val: Scope))) |
| 3445 | ? std::string(DIGV->getName()) |
| 3446 | : getFullyQualifiedName(Scope, Name: DIGV->getName()); |
| 3447 | |
| 3448 | if (const GlobalVariable *GV = |
| 3449 | dyn_cast_if_present<const GlobalVariable *>(Val: CVGV.GVInfo)) { |
| 3450 | // DataSym record, see SymbolRecord.h for more info. Thread local data |
| 3451 | // happens to have the same format as global data. |
| 3452 | MCSymbol *GVSym = Asm->getSymbol(GV); |
| 3453 | SymbolKind DataSym = GV->isThreadLocal() |
| 3454 | ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 |
| 3455 | : SymbolKind::S_GTHREAD32) |
| 3456 | : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 |
| 3457 | : SymbolKind::S_GDATA32); |
| 3458 | MCSymbol *DataEnd = beginSymbolRecord(SymKind: DataSym); |
| 3459 | OS.AddComment(T: "Type" ); |
| 3460 | OS.emitInt32(Value: getCompleteTypeIndex(Ty: DIGV->getType()).getIndex()); |
| 3461 | OS.AddComment(T: "DataOffset" ); |
| 3462 | |
| 3463 | // Use the offset seen while collecting info on globals. |
| 3464 | uint64_t Offset = CVGlobalVariableOffsets.lookup(Val: DIGV); |
| 3465 | OS.emitCOFFSecRel32(Symbol: GVSym, Offset); |
| 3466 | |
| 3467 | OS.AddComment(T: "Segment" ); |
| 3468 | OS.emitCOFFSectionIndex(Symbol: GVSym); |
| 3469 | OS.AddComment(T: "Name" ); |
| 3470 | const unsigned LengthOfDataRecord = 12; |
| 3471 | emitNullTerminatedSymbolName(OS, S: QualifiedName, MaxFixedRecordLength: LengthOfDataRecord); |
| 3472 | endSymbolRecord(SymEnd: DataEnd); |
| 3473 | } else { |
| 3474 | const DIExpression *DIE = cast<const DIExpression *>(Val: CVGV.GVInfo); |
| 3475 | assert(DIE->isConstant() && |
| 3476 | "Global constant variables must contain a constant expression." ); |
| 3477 | |
| 3478 | // Use unsigned for floats. |
| 3479 | bool isUnsigned = isFloatDIType(Ty: DIGV->getType()) |
| 3480 | ? true |
| 3481 | : DebugHandlerBase::isUnsignedDIType(Ty: DIGV->getType()); |
| 3482 | APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(I: 1)), isUnsigned); |
| 3483 | emitConstantSymbolRecord(DTy: DIGV->getType(), Value, QualifiedName); |
| 3484 | } |
| 3485 | } |
| 3486 | |
| 3487 | void forEachJumpTableBranch( |
| 3488 | const MachineFunction *MF, bool isThumb, |
| 3489 | const std::function<void(const MachineJumpTableInfo &, const MachineInstr &, |
| 3490 | int64_t)> &Callback) { |
| 3491 | auto JTI = MF->getJumpTableInfo(); |
| 3492 | if (JTI && !JTI->isEmpty()) { |
| 3493 | #ifndef NDEBUG |
| 3494 | auto UsedJTs = llvm::SmallBitVector(JTI->getJumpTables().size()); |
| 3495 | #endif |
| 3496 | for (const auto &MBB : *MF) { |
| 3497 | // Search for indirect branches... |
| 3498 | const auto LastMI = MBB.getFirstTerminator(); |
| 3499 | if (LastMI != MBB.end() && LastMI->isIndirectBranch()) { |
| 3500 | if (isThumb) { |
| 3501 | // ... that directly use jump table operands. |
| 3502 | // NOTE: ARM uses pattern matching to lower its BR_JT SDNode to |
| 3503 | // machine instructions, hence inserting a JUMP_TABLE_DEBUG_INFO node |
| 3504 | // interferes with this process *but* the resulting pseudo-instruction |
| 3505 | // uses a Jump Table operand, so extract the jump table index directly |
| 3506 | // from that. |
| 3507 | for (const auto &MO : LastMI->operands()) { |
| 3508 | if (MO.isJTI()) { |
| 3509 | unsigned Index = MO.getIndex(); |
| 3510 | #ifndef NDEBUG |
| 3511 | UsedJTs.set(Index); |
| 3512 | #endif |
| 3513 | Callback(*JTI, *LastMI, Index); |
| 3514 | break; |
| 3515 | } |
| 3516 | } |
| 3517 | } else { |
| 3518 | // ... that have jump table debug info. |
| 3519 | // NOTE: The debug info is inserted as a JUMP_TABLE_DEBUG_INFO node |
| 3520 | // when lowering the BR_JT SDNode to an indirect branch. |
| 3521 | for (auto I = MBB.instr_rbegin(), E = MBB.instr_rend(); I != E; ++I) { |
| 3522 | if (I->isJumpTableDebugInfo()) { |
| 3523 | unsigned Index = I->getOperand(i: 0).getImm(); |
| 3524 | #ifndef NDEBUG |
| 3525 | UsedJTs.set(Index); |
| 3526 | #endif |
| 3527 | Callback(*JTI, *LastMI, Index); |
| 3528 | break; |
| 3529 | } |
| 3530 | } |
| 3531 | } |
| 3532 | } |
| 3533 | } |
| 3534 | #ifndef NDEBUG |
| 3535 | assert(UsedJTs.all() && |
| 3536 | "Some of jump tables were not used in a debug info instruction" ); |
| 3537 | #endif |
| 3538 | } |
| 3539 | } |
| 3540 | |
| 3541 | void CodeViewDebug::discoverJumpTableBranches(const MachineFunction *MF, |
| 3542 | bool isThumb) { |
| 3543 | forEachJumpTableBranch( |
| 3544 | MF, isThumb, |
| 3545 | Callback: [this](const MachineJumpTableInfo &, const MachineInstr &BranchMI, |
| 3546 | int64_t) { requestLabelBeforeInsn(MI: &BranchMI); }); |
| 3547 | } |
| 3548 | |
| 3549 | void CodeViewDebug::collectDebugInfoForJumpTables(const MachineFunction *MF, |
| 3550 | bool isThumb) { |
| 3551 | forEachJumpTableBranch( |
| 3552 | MF, isThumb, |
| 3553 | Callback: [this, MF](const MachineJumpTableInfo &JTI, const MachineInstr &BranchMI, |
| 3554 | int64_t JumpTableIndex) { |
| 3555 | // For label-difference jump tables, find the base expression. |
| 3556 | // Otherwise the jump table uses an absolute address (so no base |
| 3557 | // is required). |
| 3558 | const MCSymbol *Base; |
| 3559 | uint64_t BaseOffset = 0; |
| 3560 | const MCSymbol *Branch = getLabelBeforeInsn(MI: &BranchMI); |
| 3561 | JumpTableEntrySize EntrySize; |
| 3562 | switch (JTI.getEntryKind()) { |
| 3563 | case MachineJumpTableInfo::EK_Custom32: |
| 3564 | case MachineJumpTableInfo::EK_GPRel32BlockAddress: |
| 3565 | case MachineJumpTableInfo::EK_GPRel64BlockAddress: |
| 3566 | llvm_unreachable( |
| 3567 | "EK_Custom32, EK_GPRel32BlockAddress, and " |
| 3568 | "EK_GPRel64BlockAddress should never be emitted for COFF" ); |
| 3569 | case MachineJumpTableInfo::EK_BlockAddress: |
| 3570 | // Each entry is an absolute address. |
| 3571 | EntrySize = JumpTableEntrySize::Pointer; |
| 3572 | Base = nullptr; |
| 3573 | break; |
| 3574 | case MachineJumpTableInfo::EK_Inline: |
| 3575 | case MachineJumpTableInfo::EK_LabelDifference32: |
| 3576 | case MachineJumpTableInfo::EK_LabelDifference64: |
| 3577 | // Ask the AsmPrinter. |
| 3578 | std::tie(args&: Base, args&: BaseOffset, args&: Branch, args&: EntrySize) = |
| 3579 | Asm->getCodeViewJumpTableInfo(JTI: JumpTableIndex, BranchInstr: &BranchMI, BranchLabel: Branch); |
| 3580 | break; |
| 3581 | } |
| 3582 | |
| 3583 | const MachineJumpTableEntry &JTE = JTI.getJumpTables()[JumpTableIndex]; |
| 3584 | JumpTableInfo CVJTI{.EntrySize: EntrySize, |
| 3585 | .Base: Base, |
| 3586 | .BaseOffset: BaseOffset, |
| 3587 | .Branch: Branch, |
| 3588 | .Table: MF->getJTISymbol(JTI: JumpTableIndex, Ctx&: MMI->getContext()), |
| 3589 | .TableSize: JTE.MBBs.size(), |
| 3590 | .Cases: {}}; |
| 3591 | for (const auto &MBB : JTE.MBBs) |
| 3592 | CVJTI.Cases.push_back(x: MBB->getSymbol()); |
| 3593 | CurFn->JumpTables.push_back(x: std::move(CVJTI)); |
| 3594 | }); |
| 3595 | } |
| 3596 | |
| 3597 | void CodeViewDebug::emitDebugInfoForJumpTables(const FunctionInfo &FI) { |
| 3598 | // Emit S_LABEL32 records for each jump target |
| 3599 | for (const auto &JumpTable : FI.JumpTables) { |
| 3600 | for (const auto &CaseSym : JumpTable.Cases) { |
| 3601 | MCSymbol *LabelEnd = beginSymbolRecord(SymKind: SymbolKind::S_LABEL32); |
| 3602 | OS.AddComment(T: "Offset and segment" ); |
| 3603 | OS.emitCOFFSecRel32(Symbol: CaseSym, Offset: 0); |
| 3604 | OS.AddComment(T: "Flags" ); |
| 3605 | OS.emitInt8(Value: 0); |
| 3606 | emitNullTerminatedSymbolName(OS, S: CaseSym->getName()); |
| 3607 | endSymbolRecord(SymEnd: LabelEnd); |
| 3608 | } |
| 3609 | } |
| 3610 | |
| 3611 | for (const auto &JumpTable : FI.JumpTables) { |
| 3612 | MCSymbol *JumpTableEnd = beginSymbolRecord(SymKind: SymbolKind::S_ARMSWITCHTABLE); |
| 3613 | if (JumpTable.Base) { |
| 3614 | OS.AddComment(T: "Base offset" ); |
| 3615 | OS.emitCOFFSecRel32(Symbol: JumpTable.Base, Offset: JumpTable.BaseOffset); |
| 3616 | OS.AddComment(T: "Base section index" ); |
| 3617 | OS.emitCOFFSectionIndex(Symbol: JumpTable.Base); |
| 3618 | } else { |
| 3619 | OS.AddComment(T: "Base offset" ); |
| 3620 | OS.emitInt32(Value: 0); |
| 3621 | OS.AddComment(T: "Base section index" ); |
| 3622 | OS.emitInt16(Value: 0); |
| 3623 | } |
| 3624 | OS.AddComment(T: "Switch type" ); |
| 3625 | OS.emitInt16(Value: static_cast<uint16_t>(JumpTable.EntrySize)); |
| 3626 | OS.AddComment(T: "Branch offset" ); |
| 3627 | OS.emitCOFFSecRel32(Symbol: JumpTable.Branch, /*Offset=*/0); |
| 3628 | OS.AddComment(T: "Table offset" ); |
| 3629 | OS.emitCOFFSecRel32(Symbol: JumpTable.Table, /*Offset=*/0); |
| 3630 | OS.AddComment(T: "Branch section index" ); |
| 3631 | OS.emitCOFFSectionIndex(Symbol: JumpTable.Branch); |
| 3632 | OS.AddComment(T: "Table section index" ); |
| 3633 | OS.emitCOFFSectionIndex(Symbol: JumpTable.Table); |
| 3634 | OS.AddComment(T: "Entries count" ); |
| 3635 | OS.emitInt32(Value: JumpTable.TableSize); |
| 3636 | endSymbolRecord(SymEnd: JumpTableEnd); |
| 3637 | } |
| 3638 | } |
| 3639 | |
| 3640 | void CodeViewDebug::emitInlinees( |
| 3641 | const SmallSet<codeview::TypeIndex, 1> &Inlinees) { |
| 3642 | // Divide the list of inlinees into chunks such that each chunk fits within |
| 3643 | // one record. |
| 3644 | constexpr size_t ChunkSize = |
| 3645 | (MaxRecordLength - sizeof(SymbolKind) - sizeof(uint32_t)) / |
| 3646 | sizeof(uint32_t); |
| 3647 | |
| 3648 | SmallVector<TypeIndex> SortedInlinees{Inlinees.begin(), Inlinees.end()}; |
| 3649 | llvm::sort(C&: SortedInlinees); |
| 3650 | |
| 3651 | size_t CurrentIndex = 0; |
| 3652 | while (CurrentIndex < SortedInlinees.size()) { |
| 3653 | auto Symbol = beginSymbolRecord(SymKind: SymbolKind::S_INLINEES); |
| 3654 | auto CurrentChunkSize = |
| 3655 | std::min(a: ChunkSize, b: SortedInlinees.size() - CurrentIndex); |
| 3656 | OS.AddComment(T: "Count" ); |
| 3657 | OS.emitInt32(Value: CurrentChunkSize); |
| 3658 | |
| 3659 | const size_t CurrentChunkEnd = CurrentIndex + CurrentChunkSize; |
| 3660 | for (; CurrentIndex < CurrentChunkEnd; ++CurrentIndex) { |
| 3661 | OS.AddComment(T: "Inlinee" ); |
| 3662 | OS.emitInt32(Value: SortedInlinees[CurrentIndex].getIndex()); |
| 3663 | } |
| 3664 | endSymbolRecord(SymEnd: Symbol); |
| 3665 | } |
| 3666 | } |
| 3667 | |