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 "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/None.h" 17 #include "llvm/ADT/Optional.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallString.h" 20 #include "llvm/ADT/StringRef.h" 21 #include "llvm/ADT/TinyPtrVector.h" 22 #include "llvm/ADT/Triple.h" 23 #include "llvm/ADT/Twine.h" 24 #include "llvm/BinaryFormat/COFF.h" 25 #include "llvm/BinaryFormat/Dwarf.h" 26 #include "llvm/CodeGen/AsmPrinter.h" 27 #include "llvm/CodeGen/LexicalScopes.h" 28 #include "llvm/CodeGen/MachineFrameInfo.h" 29 #include "llvm/CodeGen/MachineFunction.h" 30 #include "llvm/CodeGen/MachineInstr.h" 31 #include "llvm/CodeGen/MachineModuleInfo.h" 32 #include "llvm/CodeGen/MachineOperand.h" 33 #include "llvm/CodeGen/TargetFrameLowering.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/TargetSubtargetInfo.h" 36 #include "llvm/Config/llvm-config.h" 37 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 38 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 39 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 40 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 41 #include "llvm/DebugInfo/CodeView/EnumTables.h" 42 #include "llvm/DebugInfo/CodeView/Line.h" 43 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 44 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 45 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 46 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 47 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 48 #include "llvm/IR/Constants.h" 49 #include "llvm/IR/DataLayout.h" 50 #include "llvm/IR/DebugInfoMetadata.h" 51 #include "llvm/IR/Function.h" 52 #include "llvm/IR/GlobalValue.h" 53 #include "llvm/IR/GlobalVariable.h" 54 #include "llvm/IR/Metadata.h" 55 #include "llvm/IR/Module.h" 56 #include "llvm/MC/MCAsmInfo.h" 57 #include "llvm/MC/MCContext.h" 58 #include "llvm/MC/MCSectionCOFF.h" 59 #include "llvm/MC/MCStreamer.h" 60 #include "llvm/MC/MCSymbol.h" 61 #include "llvm/Support/BinaryByteStream.h" 62 #include "llvm/Support/BinaryStreamReader.h" 63 #include "llvm/Support/BinaryStreamWriter.h" 64 #include "llvm/Support/Casting.h" 65 #include "llvm/Support/CommandLine.h" 66 #include "llvm/Support/Endian.h" 67 #include "llvm/Support/Error.h" 68 #include "llvm/Support/ErrorHandling.h" 69 #include "llvm/Support/FormatVariadic.h" 70 #include "llvm/Support/Path.h" 71 #include "llvm/Support/SMLoc.h" 72 #include "llvm/Support/ScopedPrinter.h" 73 #include "llvm/Target/TargetLoweringObjectFile.h" 74 #include "llvm/Target/TargetMachine.h" 75 #include <algorithm> 76 #include <cassert> 77 #include <cctype> 78 #include <cstddef> 79 #include <iterator> 80 #include <limits> 81 82 using namespace llvm; 83 using namespace llvm::codeview; 84 85 namespace { 86 class CVMCAdapter : public CodeViewRecordStreamer { 87 public: 88 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 89 : OS(&OS), TypeTable(TypeTable) {} 90 91 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 92 93 void emitIntValue(uint64_t Value, unsigned Size) override { 94 OS->emitIntValueInHex(Value, Size); 95 } 96 97 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 98 99 void AddComment(const Twine &T) override { OS->AddComment(T); } 100 101 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 102 103 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 104 105 std::string getTypeName(TypeIndex TI) override { 106 std::string TypeName; 107 if (!TI.isNoneType()) { 108 if (TI.isSimple()) 109 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 110 else 111 TypeName = std::string(TypeTable.getTypeName(TI)); 112 } 113 return TypeName; 114 } 115 116 private: 117 MCStreamer *OS = nullptr; 118 TypeCollection &TypeTable; 119 }; 120 } // namespace 121 122 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 123 switch (Type) { 124 case Triple::ArchType::x86: 125 return CPUType::Pentium3; 126 case Triple::ArchType::x86_64: 127 return CPUType::X64; 128 case Triple::ArchType::thumb: 129 // LLVM currently doesn't support Windows CE and so thumb 130 // here is indiscriminately mapped to ARMNT specifically. 131 return CPUType::ARMNT; 132 case Triple::ArchType::aarch64: 133 return CPUType::ARM64; 134 default: 135 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 136 } 137 } 138 139 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 140 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 141 142 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 143 std::string &Filepath = FileToFilepathMap[File]; 144 if (!Filepath.empty()) 145 return Filepath; 146 147 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 148 149 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 150 // it textually because one of the path components could be a symlink. 151 if (Dir.startswith("/") || Filename.startswith("/")) { 152 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 153 return Filename; 154 Filepath = std::string(Dir); 155 if (Dir.back() != '/') 156 Filepath += '/'; 157 Filepath += Filename; 158 return Filepath; 159 } 160 161 // Clang emits directory and relative filename info into the IR, but CodeView 162 // operates on full paths. We could change Clang to emit full paths too, but 163 // that would increase the IR size and probably not needed for other users. 164 // For now, just concatenate and canonicalize the path here. 165 if (Filename.find(':') == 1) 166 Filepath = std::string(Filename); 167 else 168 Filepath = (Dir + "\\" + Filename).str(); 169 170 // Canonicalize the path. We have to do it textually because we may no longer 171 // have access the file in the filesystem. 172 // First, replace all slashes with backslashes. 173 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 174 175 // Remove all "\.\" with "\". 176 size_t Cursor = 0; 177 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 178 Filepath.erase(Cursor, 2); 179 180 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 181 // path should be well-formatted, e.g. start with a drive letter, etc. 182 Cursor = 0; 183 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 184 // Something's wrong if the path starts with "\..\", abort. 185 if (Cursor == 0) 186 break; 187 188 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 189 if (PrevSlash == std::string::npos) 190 // Something's wrong, abort. 191 break; 192 193 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 194 // The next ".." might be following the one we've just erased. 195 Cursor = PrevSlash; 196 } 197 198 // Remove all duplicate backslashes. 199 Cursor = 0; 200 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 201 Filepath.erase(Cursor, 1); 202 203 return Filepath; 204 } 205 206 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 207 StringRef FullPath = getFullFilepath(F); 208 unsigned NextId = FileIdMap.size() + 1; 209 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 210 if (Insertion.second) { 211 // We have to compute the full filepath and emit a .cv_file directive. 212 ArrayRef<uint8_t> ChecksumAsBytes; 213 FileChecksumKind CSKind = FileChecksumKind::None; 214 if (F->getChecksum()) { 215 std::string Checksum = fromHex(F->getChecksum()->Value); 216 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 217 memcpy(CKMem, Checksum.data(), Checksum.size()); 218 ChecksumAsBytes = ArrayRef<uint8_t>( 219 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 220 switch (F->getChecksum()->Kind) { 221 case DIFile::CSK_MD5: 222 CSKind = FileChecksumKind::MD5; 223 break; 224 case DIFile::CSK_SHA1: 225 CSKind = FileChecksumKind::SHA1; 226 break; 227 case DIFile::CSK_SHA256: 228 CSKind = FileChecksumKind::SHA256; 229 break; 230 } 231 } 232 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 233 static_cast<unsigned>(CSKind)); 234 (void)Success; 235 assert(Success && ".cv_file directive failed"); 236 } 237 return Insertion.first->second; 238 } 239 240 CodeViewDebug::InlineSite & 241 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 242 const DISubprogram *Inlinee) { 243 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 244 InlineSite *Site = &SiteInsertion.first->second; 245 if (SiteInsertion.second) { 246 unsigned ParentFuncId = CurFn->FuncId; 247 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 248 ParentFuncId = 249 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 250 .SiteFuncId; 251 252 Site->SiteFuncId = NextFuncId++; 253 OS.EmitCVInlineSiteIdDirective( 254 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 255 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 256 Site->Inlinee = Inlinee; 257 InlinedSubprograms.insert(Inlinee); 258 getFuncIdForSubprogram(Inlinee); 259 } 260 return *Site; 261 } 262 263 static StringRef getPrettyScopeName(const DIScope *Scope) { 264 StringRef ScopeName = Scope->getName(); 265 if (!ScopeName.empty()) 266 return ScopeName; 267 268 switch (Scope->getTag()) { 269 case dwarf::DW_TAG_enumeration_type: 270 case dwarf::DW_TAG_class_type: 271 case dwarf::DW_TAG_structure_type: 272 case dwarf::DW_TAG_union_type: 273 return "<unnamed-tag>"; 274 case dwarf::DW_TAG_namespace: 275 return "`anonymous namespace'"; 276 default: 277 return StringRef(); 278 } 279 } 280 281 const DISubprogram *CodeViewDebug::collectParentScopeNames( 282 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 283 const DISubprogram *ClosestSubprogram = nullptr; 284 while (Scope != nullptr) { 285 if (ClosestSubprogram == nullptr) 286 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 287 288 // If a type appears in a scope chain, make sure it gets emitted. The 289 // frontend will be responsible for deciding if this should be a forward 290 // declaration or a complete type. 291 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 292 DeferredCompleteTypes.push_back(Ty); 293 294 StringRef ScopeName = getPrettyScopeName(Scope); 295 if (!ScopeName.empty()) 296 QualifiedNameComponents.push_back(ScopeName); 297 Scope = Scope->getScope(); 298 } 299 return ClosestSubprogram; 300 } 301 302 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 303 StringRef TypeName) { 304 std::string FullyQualifiedName; 305 for (StringRef QualifiedNameComponent : 306 llvm::reverse(QualifiedNameComponents)) { 307 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 308 FullyQualifiedName.append("::"); 309 } 310 FullyQualifiedName.append(std::string(TypeName)); 311 return FullyQualifiedName; 312 } 313 314 struct CodeViewDebug::TypeLoweringScope { 315 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 316 ~TypeLoweringScope() { 317 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 318 // inner TypeLoweringScopes don't attempt to emit deferred types. 319 if (CVD.TypeEmissionLevel == 1) 320 CVD.emitDeferredCompleteTypes(); 321 --CVD.TypeEmissionLevel; 322 } 323 CodeViewDebug &CVD; 324 }; 325 326 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 327 StringRef Name) { 328 // Ensure types in the scope chain are emitted as soon as possible. 329 // This can create otherwise a situation where S_UDTs are emitted while 330 // looping in emitDebugInfoForUDTs. 331 TypeLoweringScope S(*this); 332 SmallVector<StringRef, 5> QualifiedNameComponents; 333 collectParentScopeNames(Scope, QualifiedNameComponents); 334 return formatNestedName(QualifiedNameComponents, Name); 335 } 336 337 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 338 const DIScope *Scope = Ty->getScope(); 339 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 340 } 341 342 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 343 // No scope means global scope and that uses the zero index. 344 // 345 // We also use zero index when the scope is a DISubprogram 346 // to suppress the emission of LF_STRING_ID for the function, 347 // which can trigger a link-time error with the linker in 348 // VS2019 version 16.11.2 or newer. 349 // Note, however, skipping the debug info emission for the DISubprogram 350 // is a temporary fix. The root issue here is that we need to figure out 351 // the proper way to encode a function nested in another function 352 // (as introduced by the Fortran 'contains' keyword) in CodeView. 353 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 354 return TypeIndex(); 355 356 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 357 358 // Check if we've already translated this scope. 359 auto I = TypeIndices.find({Scope, nullptr}); 360 if (I != TypeIndices.end()) 361 return I->second; 362 363 // Build the fully qualified name of the scope. 364 std::string ScopeName = getFullyQualifiedName(Scope); 365 StringIdRecord SID(TypeIndex(), ScopeName); 366 auto TI = TypeTable.writeLeafType(SID); 367 return recordTypeIndexForDINode(Scope, TI); 368 } 369 370 static StringRef removeTemplateArgs(StringRef Name) { 371 // Remove template args from the display name. Assume that the template args 372 // are the last thing in the name. 373 if (Name.empty() || Name.back() != '>') 374 return Name; 375 376 int OpenBrackets = 0; 377 for (int i = Name.size() - 1; i >= 0; --i) { 378 if (Name[i] == '>') 379 ++OpenBrackets; 380 else if (Name[i] == '<') { 381 --OpenBrackets; 382 if (OpenBrackets == 0) 383 return Name.substr(0, i); 384 } 385 } 386 return Name; 387 } 388 389 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 390 assert(SP); 391 392 // Check if we've already translated this subprogram. 393 auto I = TypeIndices.find({SP, nullptr}); 394 if (I != TypeIndices.end()) 395 return I->second; 396 397 // The display name includes function template arguments. Drop them to match 398 // MSVC. We need to have the template arguments in the DISubprogram name 399 // because they are used in other symbol records, such as S_GPROC32_IDs. 400 StringRef DisplayName = removeTemplateArgs(SP->getName()); 401 402 const DIScope *Scope = SP->getScope(); 403 TypeIndex TI; 404 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 405 // If the scope is a DICompositeType, then this must be a method. Member 406 // function types take some special handling, and require access to the 407 // subprogram. 408 TypeIndex ClassType = getTypeIndex(Class); 409 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 410 DisplayName); 411 TI = TypeTable.writeLeafType(MFuncId); 412 } else { 413 // Otherwise, this must be a free function. 414 TypeIndex ParentScope = getScopeIndex(Scope); 415 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 416 TI = TypeTable.writeLeafType(FuncId); 417 } 418 419 return recordTypeIndexForDINode(SP, TI); 420 } 421 422 static bool isNonTrivial(const DICompositeType *DCTy) { 423 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 424 } 425 426 static FunctionOptions 427 getFunctionOptions(const DISubroutineType *Ty, 428 const DICompositeType *ClassTy = nullptr, 429 StringRef SPName = StringRef("")) { 430 FunctionOptions FO = FunctionOptions::None; 431 const DIType *ReturnTy = nullptr; 432 if (auto TypeArray = Ty->getTypeArray()) { 433 if (TypeArray.size()) 434 ReturnTy = TypeArray[0]; 435 } 436 437 // Add CxxReturnUdt option to functions that return nontrivial record types 438 // or methods that return record types. 439 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 440 if (isNonTrivial(ReturnDCTy) || ClassTy) 441 FO |= FunctionOptions::CxxReturnUdt; 442 443 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 444 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 445 FO |= FunctionOptions::Constructor; 446 447 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 448 449 } 450 return FO; 451 } 452 453 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 454 const DICompositeType *Class) { 455 // Always use the method declaration as the key for the function type. The 456 // method declaration contains the this adjustment. 457 if (SP->getDeclaration()) 458 SP = SP->getDeclaration(); 459 assert(!SP->getDeclaration() && "should use declaration as key"); 460 461 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 462 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 463 auto I = TypeIndices.find({SP, Class}); 464 if (I != TypeIndices.end()) 465 return I->second; 466 467 // Make sure complete type info for the class is emitted *after* the member 468 // function type, as the complete class type is likely to reference this 469 // member function type. 470 TypeLoweringScope S(*this); 471 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 472 473 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 474 TypeIndex TI = lowerTypeMemberFunction( 475 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 476 return recordTypeIndexForDINode(SP, TI, Class); 477 } 478 479 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 480 TypeIndex TI, 481 const DIType *ClassTy) { 482 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 483 (void)InsertResult; 484 assert(InsertResult.second && "DINode was already assigned a type index"); 485 return TI; 486 } 487 488 unsigned CodeViewDebug::getPointerSizeInBytes() { 489 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 490 } 491 492 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 493 const LexicalScope *LS) { 494 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 495 // This variable was inlined. Associate it with the InlineSite. 496 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 497 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 498 Site.InlinedLocals.emplace_back(Var); 499 } else { 500 // This variable goes into the corresponding lexical scope. 501 ScopeVariables[LS].emplace_back(Var); 502 } 503 } 504 505 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 506 const DILocation *Loc) { 507 if (!llvm::is_contained(Locs, Loc)) 508 Locs.push_back(Loc); 509 } 510 511 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 512 const MachineFunction *MF) { 513 // Skip this instruction if it has the same location as the previous one. 514 if (!DL || DL == PrevInstLoc) 515 return; 516 517 const DIScope *Scope = DL.get()->getScope(); 518 if (!Scope) 519 return; 520 521 // Skip this line if it is longer than the maximum we can record. 522 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 523 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 524 LI.isNeverStepInto()) 525 return; 526 527 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 528 if (CI.getStartColumn() != DL.getCol()) 529 return; 530 531 if (!CurFn->HaveLineInfo) 532 CurFn->HaveLineInfo = true; 533 unsigned FileId = 0; 534 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 535 FileId = CurFn->LastFileId; 536 else 537 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 538 PrevInstLoc = DL; 539 540 unsigned FuncId = CurFn->FuncId; 541 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 542 const DILocation *Loc = DL.get(); 543 544 // If this location was actually inlined from somewhere else, give it the ID 545 // of the inline call site. 546 FuncId = 547 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 548 549 // Ensure we have links in the tree of inline call sites. 550 bool FirstLoc = true; 551 while ((SiteLoc = Loc->getInlinedAt())) { 552 InlineSite &Site = 553 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 554 if (!FirstLoc) 555 addLocIfNotPresent(Site.ChildSites, Loc); 556 FirstLoc = false; 557 Loc = SiteLoc; 558 } 559 addLocIfNotPresent(CurFn->ChildSites, Loc); 560 } 561 562 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 563 /*PrologueEnd=*/false, /*IsStmt=*/false, 564 DL->getFilename(), SMLoc()); 565 } 566 567 void CodeViewDebug::emitCodeViewMagicVersion() { 568 OS.emitValueToAlignment(4); 569 OS.AddComment("Debug section magic"); 570 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 571 } 572 573 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 574 switch (DWLang) { 575 case dwarf::DW_LANG_C: 576 case dwarf::DW_LANG_C89: 577 case dwarf::DW_LANG_C99: 578 case dwarf::DW_LANG_C11: 579 case dwarf::DW_LANG_ObjC: 580 return SourceLanguage::C; 581 case dwarf::DW_LANG_C_plus_plus: 582 case dwarf::DW_LANG_C_plus_plus_03: 583 case dwarf::DW_LANG_C_plus_plus_11: 584 case dwarf::DW_LANG_C_plus_plus_14: 585 return SourceLanguage::Cpp; 586 case dwarf::DW_LANG_Fortran77: 587 case dwarf::DW_LANG_Fortran90: 588 case dwarf::DW_LANG_Fortran95: 589 case dwarf::DW_LANG_Fortran03: 590 case dwarf::DW_LANG_Fortran08: 591 return SourceLanguage::Fortran; 592 case dwarf::DW_LANG_Pascal83: 593 return SourceLanguage::Pascal; 594 case dwarf::DW_LANG_Cobol74: 595 case dwarf::DW_LANG_Cobol85: 596 return SourceLanguage::Cobol; 597 case dwarf::DW_LANG_Java: 598 return SourceLanguage::Java; 599 case dwarf::DW_LANG_D: 600 return SourceLanguage::D; 601 case dwarf::DW_LANG_Swift: 602 return SourceLanguage::Swift; 603 case dwarf::DW_LANG_Rust: 604 return SourceLanguage::Rust; 605 default: 606 // There's no CodeView representation for this language, and CV doesn't 607 // have an "unknown" option for the language field, so we'll use MASM, 608 // as it's very low level. 609 return SourceLanguage::Masm; 610 } 611 } 612 613 void CodeViewDebug::beginModule(Module *M) { 614 // If module doesn't have named metadata anchors or COFF debug section 615 // is not available, skip any debug info related stuff. 616 NamedMDNode *CUs = M->getNamedMetadata("llvm.dbg.cu"); 617 if (!CUs || !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 618 Asm = nullptr; 619 return; 620 } 621 // Tell MMI that we have and need debug info. 622 MMI->setDebugInfoAvailability(true); 623 624 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 625 626 // Get the current source language. 627 const MDNode *Node = *CUs->operands().begin(); 628 const auto *CU = cast<DICompileUnit>(Node); 629 630 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 631 632 collectGlobalVariableInfo(); 633 634 // Check if we should emit type record hashes. 635 ConstantInt *GH = 636 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 637 EmitDebugGlobalHashes = GH && !GH->isZero(); 638 } 639 640 void CodeViewDebug::endModule() { 641 if (!Asm || !MMI->hasDebugInfo()) 642 return; 643 644 // The COFF .debug$S section consists of several subsections, each starting 645 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 646 // of the payload followed by the payload itself. The subsections are 4-byte 647 // aligned. 648 649 // Use the generic .debug$S section, and make a subsection for all the inlined 650 // subprograms. 651 switchToDebugSectionForSymbol(nullptr); 652 653 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 654 emitObjName(); 655 emitCompilerInformation(); 656 endCVSubsection(CompilerInfo); 657 658 emitInlineeLinesSubsection(); 659 660 // Emit per-function debug information. 661 for (auto &P : FnDebugInfo) 662 if (!P.first->isDeclarationForLinker()) 663 emitDebugInfoForFunction(P.first, *P.second); 664 665 // Get types used by globals without emitting anything. 666 // This is meant to collect all static const data members so they can be 667 // emitted as globals. 668 collectDebugInfoForGlobals(); 669 670 // Emit retained types. 671 emitDebugInfoForRetainedTypes(); 672 673 // Emit global variable debug information. 674 setCurrentSubprogram(nullptr); 675 emitDebugInfoForGlobals(); 676 677 // Switch back to the generic .debug$S section after potentially processing 678 // comdat symbol sections. 679 switchToDebugSectionForSymbol(nullptr); 680 681 // Emit UDT records for any types used by global variables. 682 if (!GlobalUDTs.empty()) { 683 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 684 emitDebugInfoForUDTs(GlobalUDTs); 685 endCVSubsection(SymbolsEnd); 686 } 687 688 // This subsection holds a file index to offset in string table table. 689 OS.AddComment("File index to string table offset subsection"); 690 OS.emitCVFileChecksumsDirective(); 691 692 // This subsection holds the string table. 693 OS.AddComment("String table"); 694 OS.emitCVStringTableDirective(); 695 696 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 697 // subsection in the generic .debug$S section at the end. There is no 698 // particular reason for this ordering other than to match MSVC. 699 emitBuildInfo(); 700 701 // Emit type information and hashes last, so that any types we translate while 702 // emitting function info are included. 703 emitTypeInformation(); 704 705 if (EmitDebugGlobalHashes) 706 emitTypeGlobalHashes(); 707 708 clear(); 709 } 710 711 static void 712 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 713 unsigned MaxFixedRecordLength = 0xF00) { 714 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 715 // after a fixed length portion of the record. The fixed length portion should 716 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 717 // overall record size is less than the maximum allowed. 718 SmallString<32> NullTerminatedString( 719 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 720 NullTerminatedString.push_back('\0'); 721 OS.emitBytes(NullTerminatedString); 722 } 723 724 void CodeViewDebug::emitTypeInformation() { 725 if (TypeTable.empty()) 726 return; 727 728 // Start the .debug$T or .debug$P section with 0x4. 729 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 730 emitCodeViewMagicVersion(); 731 732 TypeTableCollection Table(TypeTable.records()); 733 TypeVisitorCallbackPipeline Pipeline; 734 735 // To emit type record using Codeview MCStreamer adapter 736 CVMCAdapter CVMCOS(OS, Table); 737 TypeRecordMapping typeMapping(CVMCOS); 738 Pipeline.addCallbackToPipeline(typeMapping); 739 740 Optional<TypeIndex> B = Table.getFirst(); 741 while (B) { 742 // This will fail if the record data is invalid. 743 CVType Record = Table.getType(*B); 744 745 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 746 747 if (E) { 748 logAllUnhandledErrors(std::move(E), errs(), "error: "); 749 llvm_unreachable("produced malformed type record"); 750 } 751 752 B = Table.getNext(*B); 753 } 754 } 755 756 void CodeViewDebug::emitTypeGlobalHashes() { 757 if (TypeTable.empty()) 758 return; 759 760 // Start the .debug$H section with the version and hash algorithm, currently 761 // hardcoded to version 0, SHA1. 762 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 763 764 OS.emitValueToAlignment(4); 765 OS.AddComment("Magic"); 766 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 767 OS.AddComment("Section Version"); 768 OS.emitInt16(0); 769 OS.AddComment("Hash Algorithm"); 770 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 771 772 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 773 for (const auto &GHR : TypeTable.hashes()) { 774 if (OS.isVerboseAsm()) { 775 // Emit an EOL-comment describing which TypeIndex this hash corresponds 776 // to, as well as the stringified SHA1 hash. 777 SmallString<32> Comment; 778 raw_svector_ostream CommentOS(Comment); 779 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 780 OS.AddComment(Comment); 781 ++TI; 782 } 783 assert(GHR.Hash.size() == 8); 784 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 785 GHR.Hash.size()); 786 OS.emitBinaryData(S); 787 } 788 } 789 790 void CodeViewDebug::emitObjName() { 791 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 792 793 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 794 llvm::SmallString<256> PathStore(PathRef); 795 796 if (PathRef.empty() || PathRef == "-") { 797 // Don't emit the filename if we're writing to stdout or to /dev/null. 798 PathRef = {}; 799 } else { 800 llvm::sys::path::remove_dots(PathStore, /*remove_dot_dot=*/true); 801 PathRef = PathStore; 802 } 803 804 OS.AddComment("Signature"); 805 OS.emitIntValue(0, 4); 806 807 OS.AddComment("Object name"); 808 emitNullTerminatedSymbolName(OS, PathRef); 809 810 endSymbolRecord(CompilerEnd); 811 } 812 813 namespace { 814 struct Version { 815 int Part[4]; 816 }; 817 } // end anonymous namespace 818 819 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 820 // the version number. 821 static Version parseVersion(StringRef Name) { 822 Version V = {{0}}; 823 int N = 0; 824 for (const char C : Name) { 825 if (isdigit(C)) { 826 V.Part[N] *= 10; 827 V.Part[N] += C - '0'; 828 } else if (C == '.') { 829 ++N; 830 if (N >= 4) 831 return V; 832 } else if (N > 0) 833 return V; 834 } 835 return V; 836 } 837 838 void CodeViewDebug::emitCompilerInformation() { 839 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 840 uint32_t Flags = 0; 841 842 // The low byte of the flags indicates the source language. 843 Flags = CurrentSourceLanguage; 844 // TODO: Figure out which other flags need to be set. 845 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 846 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 847 } 848 849 OS.AddComment("Flags and language"); 850 OS.emitInt32(Flags); 851 852 OS.AddComment("CPUType"); 853 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 854 855 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 856 const MDNode *Node = *CUs->operands().begin(); 857 const auto *CU = cast<DICompileUnit>(Node); 858 859 StringRef CompilerVersion = CU->getProducer(); 860 Version FrontVer = parseVersion(CompilerVersion); 861 OS.AddComment("Frontend version"); 862 for (int N : FrontVer.Part) { 863 N = std::min<int>(N, std::numeric_limits<uint16_t>::max()); 864 OS.emitInt16(N); 865 } 866 867 // Some Microsoft tools, like Binscope, expect a backend version number of at 868 // least 8.something, so we'll coerce the LLVM version into a form that 869 // guarantees it'll be big enough without really lying about the version. 870 int Major = 1000 * LLVM_VERSION_MAJOR + 871 10 * LLVM_VERSION_MINOR + 872 LLVM_VERSION_PATCH; 873 // Clamp it for builds that use unusually large version numbers. 874 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 875 Version BackVer = {{ Major, 0, 0, 0 }}; 876 OS.AddComment("Backend version"); 877 for (int N : BackVer.Part) 878 OS.emitInt16(N); 879 880 OS.AddComment("Null-terminated compiler version string"); 881 emitNullTerminatedSymbolName(OS, CompilerVersion); 882 883 endSymbolRecord(CompilerEnd); 884 } 885 886 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 887 StringRef S) { 888 StringIdRecord SIR(TypeIndex(0x0), S); 889 return TypeTable.writeLeafType(SIR); 890 } 891 892 void CodeViewDebug::emitBuildInfo() { 893 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 894 // build info. The known prefix is: 895 // - Absolute path of current directory 896 // - Compiler path 897 // - Main source file path, relative to CWD or absolute 898 // - Type server PDB file 899 // - Canonical compiler command line 900 // If frontend and backend compilation are separated (think llc or LTO), it's 901 // not clear if the compiler path should refer to the executable for the 902 // frontend or the backend. Leave it blank for now. 903 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 904 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 905 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 906 const auto *CU = cast<DICompileUnit>(Node); 907 const DIFile *MainSourceFile = CU->getFile(); 908 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 909 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 910 BuildInfoArgs[BuildInfoRecord::SourceFile] = 911 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 912 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 913 // we implement /Zi type servers. 914 BuildInfoRecord BIR(BuildInfoArgs); 915 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 916 917 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 918 // from the module symbols into the type stream. 919 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 920 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 921 OS.AddComment("LF_BUILDINFO index"); 922 OS.emitInt32(BuildInfoIndex.getIndex()); 923 endSymbolRecord(BIEnd); 924 endCVSubsection(BISubsecEnd); 925 } 926 927 void CodeViewDebug::emitInlineeLinesSubsection() { 928 if (InlinedSubprograms.empty()) 929 return; 930 931 OS.AddComment("Inlinee lines subsection"); 932 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 933 934 // We emit the checksum info for files. This is used by debuggers to 935 // determine if a pdb matches the source before loading it. Visual Studio, 936 // for instance, will display a warning that the breakpoints are not valid if 937 // the pdb does not match the source. 938 OS.AddComment("Inlinee lines signature"); 939 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 940 941 for (const DISubprogram *SP : InlinedSubprograms) { 942 assert(TypeIndices.count({SP, nullptr})); 943 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 944 945 OS.AddBlankLine(); 946 unsigned FileId = maybeRecordFile(SP->getFile()); 947 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 948 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 949 OS.AddBlankLine(); 950 OS.AddComment("Type index of inlined function"); 951 OS.emitInt32(InlineeIdx.getIndex()); 952 OS.AddComment("Offset into filechecksum table"); 953 OS.emitCVFileChecksumOffsetDirective(FileId); 954 OS.AddComment("Starting line number"); 955 OS.emitInt32(SP->getLine()); 956 } 957 958 endCVSubsection(InlineEnd); 959 } 960 961 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 962 const DILocation *InlinedAt, 963 const InlineSite &Site) { 964 assert(TypeIndices.count({Site.Inlinee, nullptr})); 965 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 966 967 // SymbolRecord 968 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 969 970 OS.AddComment("PtrParent"); 971 OS.emitInt32(0); 972 OS.AddComment("PtrEnd"); 973 OS.emitInt32(0); 974 OS.AddComment("Inlinee type index"); 975 OS.emitInt32(InlineeIdx.getIndex()); 976 977 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 978 unsigned StartLineNum = Site.Inlinee->getLine(); 979 980 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 981 FI.Begin, FI.End); 982 983 endSymbolRecord(InlineEnd); 984 985 emitLocalVariableList(FI, Site.InlinedLocals); 986 987 // Recurse on child inlined call sites before closing the scope. 988 for (const DILocation *ChildSite : Site.ChildSites) { 989 auto I = FI.InlineSites.find(ChildSite); 990 assert(I != FI.InlineSites.end() && 991 "child site not in function inline site map"); 992 emitInlinedCallSite(FI, ChildSite, I->second); 993 } 994 995 // Close the scope. 996 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 997 } 998 999 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1000 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1001 // comdat key. A section may be comdat because of -ffunction-sections or 1002 // because it is comdat in the IR. 1003 MCSectionCOFF *GVSec = 1004 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1005 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1006 1007 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1008 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1009 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1010 1011 OS.SwitchSection(DebugSec); 1012 1013 // Emit the magic version number if this is the first time we've switched to 1014 // this section. 1015 if (ComdatDebugSections.insert(DebugSec).second) 1016 emitCodeViewMagicVersion(); 1017 } 1018 1019 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1020 // The only supported thunk ordinal is currently the standard type. 1021 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1022 FunctionInfo &FI, 1023 const MCSymbol *Fn) { 1024 std::string FuncName = 1025 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1026 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1027 1028 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1029 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1030 1031 // Emit S_THUNK32 1032 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1033 OS.AddComment("PtrParent"); 1034 OS.emitInt32(0); 1035 OS.AddComment("PtrEnd"); 1036 OS.emitInt32(0); 1037 OS.AddComment("PtrNext"); 1038 OS.emitInt32(0); 1039 OS.AddComment("Thunk section relative address"); 1040 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1041 OS.AddComment("Thunk section index"); 1042 OS.EmitCOFFSectionIndex(Fn); 1043 OS.AddComment("Code size"); 1044 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1045 OS.AddComment("Ordinal"); 1046 OS.emitInt8(unsigned(ordinal)); 1047 OS.AddComment("Function name"); 1048 emitNullTerminatedSymbolName(OS, FuncName); 1049 // Additional fields specific to the thunk ordinal would go here. 1050 endSymbolRecord(ThunkRecordEnd); 1051 1052 // Local variables/inlined routines are purposely omitted here. The point of 1053 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1054 1055 // Emit S_PROC_ID_END 1056 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1057 1058 endCVSubsection(SymbolsEnd); 1059 } 1060 1061 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1062 FunctionInfo &FI) { 1063 // For each function there is a separate subsection which holds the PC to 1064 // file:line table. 1065 const MCSymbol *Fn = Asm->getSymbol(GV); 1066 assert(Fn); 1067 1068 // Switch to the to a comdat section, if appropriate. 1069 switchToDebugSectionForSymbol(Fn); 1070 1071 std::string FuncName; 1072 auto *SP = GV->getSubprogram(); 1073 assert(SP); 1074 setCurrentSubprogram(SP); 1075 1076 if (SP->isThunk()) { 1077 emitDebugInfoForThunk(GV, FI, Fn); 1078 return; 1079 } 1080 1081 // If we have a display name, build the fully qualified name by walking the 1082 // chain of scopes. 1083 if (!SP->getName().empty()) 1084 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1085 1086 // If our DISubprogram name is empty, use the mangled name. 1087 if (FuncName.empty()) 1088 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1089 1090 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1091 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1092 OS.EmitCVFPOData(Fn); 1093 1094 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1095 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1096 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1097 { 1098 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1099 : SymbolKind::S_GPROC32_ID; 1100 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1101 1102 // These fields are filled in by tools like CVPACK which run after the fact. 1103 OS.AddComment("PtrParent"); 1104 OS.emitInt32(0); 1105 OS.AddComment("PtrEnd"); 1106 OS.emitInt32(0); 1107 OS.AddComment("PtrNext"); 1108 OS.emitInt32(0); 1109 // This is the important bit that tells the debugger where the function 1110 // code is located and what's its size: 1111 OS.AddComment("Code size"); 1112 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1113 OS.AddComment("Offset after prologue"); 1114 OS.emitInt32(0); 1115 OS.AddComment("Offset before epilogue"); 1116 OS.emitInt32(0); 1117 OS.AddComment("Function type index"); 1118 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1119 OS.AddComment("Function section relative address"); 1120 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1121 OS.AddComment("Function section index"); 1122 OS.EmitCOFFSectionIndex(Fn); 1123 OS.AddComment("Flags"); 1124 OS.emitInt8(0); 1125 // Emit the function display name as a null-terminated string. 1126 OS.AddComment("Function name"); 1127 // Truncate the name so we won't overflow the record length field. 1128 emitNullTerminatedSymbolName(OS, FuncName); 1129 endSymbolRecord(ProcRecordEnd); 1130 1131 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1132 // Subtract out the CSR size since MSVC excludes that and we include it. 1133 OS.AddComment("FrameSize"); 1134 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1135 OS.AddComment("Padding"); 1136 OS.emitInt32(0); 1137 OS.AddComment("Offset of padding"); 1138 OS.emitInt32(0); 1139 OS.AddComment("Bytes of callee saved registers"); 1140 OS.emitInt32(FI.CSRSize); 1141 OS.AddComment("Exception handler offset"); 1142 OS.emitInt32(0); 1143 OS.AddComment("Exception handler section"); 1144 OS.emitInt16(0); 1145 OS.AddComment("Flags (defines frame register)"); 1146 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1147 endSymbolRecord(FrameProcEnd); 1148 1149 emitLocalVariableList(FI, FI.Locals); 1150 emitGlobalVariableList(FI.Globals); 1151 emitLexicalBlockList(FI.ChildBlocks, FI); 1152 1153 // Emit inlined call site information. Only emit functions inlined directly 1154 // into the parent function. We'll emit the other sites recursively as part 1155 // of their parent inline site. 1156 for (const DILocation *InlinedAt : FI.ChildSites) { 1157 auto I = FI.InlineSites.find(InlinedAt); 1158 assert(I != FI.InlineSites.end() && 1159 "child site not in function inline site map"); 1160 emitInlinedCallSite(FI, InlinedAt, I->second); 1161 } 1162 1163 for (auto Annot : FI.Annotations) { 1164 MCSymbol *Label = Annot.first; 1165 MDTuple *Strs = cast<MDTuple>(Annot.second); 1166 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1167 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1168 // FIXME: Make sure we don't overflow the max record size. 1169 OS.EmitCOFFSectionIndex(Label); 1170 OS.emitInt16(Strs->getNumOperands()); 1171 for (Metadata *MD : Strs->operands()) { 1172 // MDStrings are null terminated, so we can do EmitBytes and get the 1173 // nice .asciz directive. 1174 StringRef Str = cast<MDString>(MD)->getString(); 1175 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1176 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1177 } 1178 endSymbolRecord(AnnotEnd); 1179 } 1180 1181 for (auto HeapAllocSite : FI.HeapAllocSites) { 1182 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1183 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1184 const DIType *DITy = std::get<2>(HeapAllocSite); 1185 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1186 OS.AddComment("Call site offset"); 1187 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1188 OS.AddComment("Call site section index"); 1189 OS.EmitCOFFSectionIndex(BeginLabel); 1190 OS.AddComment("Call instruction length"); 1191 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1192 OS.AddComment("Type index"); 1193 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1194 endSymbolRecord(HeapAllocEnd); 1195 } 1196 1197 if (SP != nullptr) 1198 emitDebugInfoForUDTs(LocalUDTs); 1199 1200 // We're done with this function. 1201 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1202 } 1203 endCVSubsection(SymbolsEnd); 1204 1205 // We have an assembler directive that takes care of the whole line table. 1206 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1207 } 1208 1209 CodeViewDebug::LocalVarDefRange 1210 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1211 LocalVarDefRange DR; 1212 DR.InMemory = -1; 1213 DR.DataOffset = Offset; 1214 assert(DR.DataOffset == Offset && "truncation"); 1215 DR.IsSubfield = 0; 1216 DR.StructOffset = 0; 1217 DR.CVRegister = CVRegister; 1218 return DR; 1219 } 1220 1221 void CodeViewDebug::collectVariableInfoFromMFTable( 1222 DenseSet<InlinedEntity> &Processed) { 1223 const MachineFunction &MF = *Asm->MF; 1224 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1225 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1226 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1227 1228 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1229 if (!VI.Var) 1230 continue; 1231 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1232 "Expected inlined-at fields to agree"); 1233 1234 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1235 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1236 1237 // If variable scope is not found then skip this variable. 1238 if (!Scope) 1239 continue; 1240 1241 // If the variable has an attached offset expression, extract it. 1242 // FIXME: Try to handle DW_OP_deref as well. 1243 int64_t ExprOffset = 0; 1244 bool Deref = false; 1245 if (VI.Expr) { 1246 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1247 if (VI.Expr->getNumElements() == 1 && 1248 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1249 Deref = true; 1250 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1251 continue; 1252 } 1253 1254 // Get the frame register used and the offset. 1255 Register FrameReg; 1256 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1257 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1258 1259 assert(!FrameOffset.getScalable() && 1260 "Frame offsets with a scalable component are not supported"); 1261 1262 // Calculate the label ranges. 1263 LocalVarDefRange DefRange = 1264 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1265 1266 for (const InsnRange &Range : Scope->getRanges()) { 1267 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1268 const MCSymbol *End = getLabelAfterInsn(Range.second); 1269 End = End ? End : Asm->getFunctionEnd(); 1270 DefRange.Ranges.emplace_back(Begin, End); 1271 } 1272 1273 LocalVariable Var; 1274 Var.DIVar = VI.Var; 1275 Var.DefRanges.emplace_back(std::move(DefRange)); 1276 if (Deref) 1277 Var.UseReferenceType = true; 1278 1279 recordLocalVariable(std::move(Var), Scope); 1280 } 1281 } 1282 1283 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1284 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1285 } 1286 1287 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1288 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1289 } 1290 1291 void CodeViewDebug::calculateRanges( 1292 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1293 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1294 1295 // Calculate the definition ranges. 1296 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1297 const auto &Entry = *I; 1298 if (!Entry.isDbgValue()) 1299 continue; 1300 const MachineInstr *DVInst = Entry.getInstr(); 1301 assert(DVInst->isDebugValue() && "Invalid History entry"); 1302 // FIXME: Find a way to represent constant variables, since they are 1303 // relatively common. 1304 Optional<DbgVariableLocation> Location = 1305 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1306 if (!Location) 1307 continue; 1308 1309 // CodeView can only express variables in register and variables in memory 1310 // at a constant offset from a register. However, for variables passed 1311 // indirectly by pointer, it is common for that pointer to be spilled to a 1312 // stack location. For the special case of one offseted load followed by a 1313 // zero offset load (a pointer spilled to the stack), we change the type of 1314 // the local variable from a value type to a reference type. This tricks the 1315 // debugger into doing the load for us. 1316 if (Var.UseReferenceType) { 1317 // We're using a reference type. Drop the last zero offset load. 1318 if (canUseReferenceType(*Location)) 1319 Location->LoadChain.pop_back(); 1320 else 1321 continue; 1322 } else if (needsReferenceType(*Location)) { 1323 // This location can't be expressed without switching to a reference type. 1324 // Start over using that. 1325 Var.UseReferenceType = true; 1326 Var.DefRanges.clear(); 1327 calculateRanges(Var, Entries); 1328 return; 1329 } 1330 1331 // We can only handle a register or an offseted load of a register. 1332 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1333 continue; 1334 { 1335 LocalVarDefRange DR; 1336 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1337 DR.InMemory = !Location->LoadChain.empty(); 1338 DR.DataOffset = 1339 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1340 if (Location->FragmentInfo) { 1341 DR.IsSubfield = true; 1342 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1343 } else { 1344 DR.IsSubfield = false; 1345 DR.StructOffset = 0; 1346 } 1347 1348 if (Var.DefRanges.empty() || 1349 Var.DefRanges.back().isDifferentLocation(DR)) { 1350 Var.DefRanges.emplace_back(std::move(DR)); 1351 } 1352 } 1353 1354 // Compute the label range. 1355 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1356 const MCSymbol *End; 1357 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1358 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1359 End = EndingEntry.isDbgValue() 1360 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1361 : getLabelAfterInsn(EndingEntry.getInstr()); 1362 } else 1363 End = Asm->getFunctionEnd(); 1364 1365 // If the last range end is our begin, just extend the last range. 1366 // Otherwise make a new range. 1367 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1368 Var.DefRanges.back().Ranges; 1369 if (!R.empty() && R.back().second == Begin) 1370 R.back().second = End; 1371 else 1372 R.emplace_back(Begin, End); 1373 1374 // FIXME: Do more range combining. 1375 } 1376 } 1377 1378 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1379 DenseSet<InlinedEntity> Processed; 1380 // Grab the variable info that was squirreled away in the MMI side-table. 1381 collectVariableInfoFromMFTable(Processed); 1382 1383 for (const auto &I : DbgValues) { 1384 InlinedEntity IV = I.first; 1385 if (Processed.count(IV)) 1386 continue; 1387 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1388 const DILocation *InlinedAt = IV.second; 1389 1390 // Instruction ranges, specifying where IV is accessible. 1391 const auto &Entries = I.second; 1392 1393 LexicalScope *Scope = nullptr; 1394 if (InlinedAt) 1395 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1396 else 1397 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1398 // If variable scope is not found then skip this variable. 1399 if (!Scope) 1400 continue; 1401 1402 LocalVariable Var; 1403 Var.DIVar = DIVar; 1404 1405 calculateRanges(Var, Entries); 1406 recordLocalVariable(std::move(Var), Scope); 1407 } 1408 } 1409 1410 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1411 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1412 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1413 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1414 const Function &GV = MF->getFunction(); 1415 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1416 assert(Insertion.second && "function already has info"); 1417 CurFn = Insertion.first->second.get(); 1418 CurFn->FuncId = NextFuncId++; 1419 CurFn->Begin = Asm->getFunctionBegin(); 1420 1421 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1422 // callee-saved registers were used. For targets that don't use a PUSH 1423 // instruction (AArch64), this will be zero. 1424 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1425 CurFn->FrameSize = MFI.getStackSize(); 1426 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1427 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1428 1429 // For this function S_FRAMEPROC record, figure out which codeview register 1430 // will be the frame pointer. 1431 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1432 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1433 if (CurFn->FrameSize > 0) { 1434 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1435 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1436 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1437 } else { 1438 // If there is an FP, parameters are always relative to it. 1439 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1440 if (CurFn->HasStackRealignment) { 1441 // If the stack needs realignment, locals are relative to SP or VFRAME. 1442 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1443 } else { 1444 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1445 // other stack adjustments. 1446 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1447 } 1448 } 1449 } 1450 1451 // Compute other frame procedure options. 1452 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1453 if (MFI.hasVarSizedObjects()) 1454 FPO |= FrameProcedureOptions::HasAlloca; 1455 if (MF->exposesReturnsTwice()) 1456 FPO |= FrameProcedureOptions::HasSetJmp; 1457 // FIXME: Set HasLongJmp if we ever track that info. 1458 if (MF->hasInlineAsm()) 1459 FPO |= FrameProcedureOptions::HasInlineAssembly; 1460 if (GV.hasPersonalityFn()) { 1461 if (isAsynchronousEHPersonality( 1462 classifyEHPersonality(GV.getPersonalityFn()))) 1463 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1464 else 1465 FPO |= FrameProcedureOptions::HasExceptionHandling; 1466 } 1467 if (GV.hasFnAttribute(Attribute::InlineHint)) 1468 FPO |= FrameProcedureOptions::MarkedInline; 1469 if (GV.hasFnAttribute(Attribute::Naked)) 1470 FPO |= FrameProcedureOptions::Naked; 1471 if (MFI.hasStackProtectorIndex()) 1472 FPO |= FrameProcedureOptions::SecurityChecks; 1473 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1474 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1475 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1476 !GV.hasOptSize() && !GV.hasOptNone()) 1477 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1478 if (GV.hasProfileData()) { 1479 FPO |= FrameProcedureOptions::ValidProfileCounts; 1480 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1481 } 1482 // FIXME: Set GuardCfg when it is implemented. 1483 CurFn->FrameProcOpts = FPO; 1484 1485 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1486 1487 // Find the end of the function prolog. First known non-DBG_VALUE and 1488 // non-frame setup location marks the beginning of the function body. 1489 // FIXME: is there a simpler a way to do this? Can we just search 1490 // for the first instruction of the function, not the last of the prolog? 1491 DebugLoc PrologEndLoc; 1492 bool EmptyPrologue = true; 1493 for (const auto &MBB : *MF) { 1494 for (const auto &MI : MBB) { 1495 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1496 MI.getDebugLoc()) { 1497 PrologEndLoc = MI.getDebugLoc(); 1498 break; 1499 } else if (!MI.isMetaInstruction()) { 1500 EmptyPrologue = false; 1501 } 1502 } 1503 } 1504 1505 // Record beginning of function if we have a non-empty prologue. 1506 if (PrologEndLoc && !EmptyPrologue) { 1507 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1508 maybeRecordLocation(FnStartDL, MF); 1509 } 1510 1511 // Find heap alloc sites and emit labels around them. 1512 for (const auto &MBB : *MF) { 1513 for (const auto &MI : MBB) { 1514 if (MI.getHeapAllocMarker()) { 1515 requestLabelBeforeInsn(&MI); 1516 requestLabelAfterInsn(&MI); 1517 } 1518 } 1519 } 1520 } 1521 1522 static bool shouldEmitUdt(const DIType *T) { 1523 if (!T) 1524 return false; 1525 1526 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1527 if (T->getTag() == dwarf::DW_TAG_typedef) { 1528 if (DIScope *Scope = T->getScope()) { 1529 switch (Scope->getTag()) { 1530 case dwarf::DW_TAG_structure_type: 1531 case dwarf::DW_TAG_class_type: 1532 case dwarf::DW_TAG_union_type: 1533 return false; 1534 default: 1535 // do nothing. 1536 ; 1537 } 1538 } 1539 } 1540 1541 while (true) { 1542 if (!T || T->isForwardDecl()) 1543 return false; 1544 1545 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1546 if (!DT) 1547 return true; 1548 T = DT->getBaseType(); 1549 } 1550 return true; 1551 } 1552 1553 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1554 // Don't record empty UDTs. 1555 if (Ty->getName().empty()) 1556 return; 1557 if (!shouldEmitUdt(Ty)) 1558 return; 1559 1560 SmallVector<StringRef, 5> ParentScopeNames; 1561 const DISubprogram *ClosestSubprogram = 1562 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1563 1564 std::string FullyQualifiedName = 1565 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1566 1567 if (ClosestSubprogram == nullptr) { 1568 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1569 } else if (ClosestSubprogram == CurrentSubprogram) { 1570 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1571 } 1572 1573 // TODO: What if the ClosestSubprogram is neither null or the current 1574 // subprogram? Currently, the UDT just gets dropped on the floor. 1575 // 1576 // The current behavior is not desirable. To get maximal fidelity, we would 1577 // need to perform all type translation before beginning emission of .debug$S 1578 // and then make LocalUDTs a member of FunctionInfo 1579 } 1580 1581 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1582 // Generic dispatch for lowering an unknown type. 1583 switch (Ty->getTag()) { 1584 case dwarf::DW_TAG_array_type: 1585 return lowerTypeArray(cast<DICompositeType>(Ty)); 1586 case dwarf::DW_TAG_typedef: 1587 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1588 case dwarf::DW_TAG_base_type: 1589 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1590 case dwarf::DW_TAG_pointer_type: 1591 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1592 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1593 LLVM_FALLTHROUGH; 1594 case dwarf::DW_TAG_reference_type: 1595 case dwarf::DW_TAG_rvalue_reference_type: 1596 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1597 case dwarf::DW_TAG_ptr_to_member_type: 1598 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1599 case dwarf::DW_TAG_restrict_type: 1600 case dwarf::DW_TAG_const_type: 1601 case dwarf::DW_TAG_volatile_type: 1602 // TODO: add support for DW_TAG_atomic_type here 1603 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1604 case dwarf::DW_TAG_subroutine_type: 1605 if (ClassTy) { 1606 // The member function type of a member function pointer has no 1607 // ThisAdjustment. 1608 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1609 /*ThisAdjustment=*/0, 1610 /*IsStaticMethod=*/false); 1611 } 1612 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1613 case dwarf::DW_TAG_enumeration_type: 1614 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1615 case dwarf::DW_TAG_class_type: 1616 case dwarf::DW_TAG_structure_type: 1617 return lowerTypeClass(cast<DICompositeType>(Ty)); 1618 case dwarf::DW_TAG_union_type: 1619 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1620 case dwarf::DW_TAG_string_type: 1621 return lowerTypeString(cast<DIStringType>(Ty)); 1622 case dwarf::DW_TAG_unspecified_type: 1623 if (Ty->getName() == "decltype(nullptr)") 1624 return TypeIndex::NullptrT(); 1625 return TypeIndex::None(); 1626 default: 1627 // Use the null type index. 1628 return TypeIndex(); 1629 } 1630 } 1631 1632 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1633 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1634 StringRef TypeName = Ty->getName(); 1635 1636 addToUDTs(Ty); 1637 1638 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1639 TypeName == "HRESULT") 1640 return TypeIndex(SimpleTypeKind::HResult); 1641 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1642 TypeName == "wchar_t") 1643 return TypeIndex(SimpleTypeKind::WideCharacter); 1644 1645 return UnderlyingTypeIndex; 1646 } 1647 1648 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1649 const DIType *ElementType = Ty->getBaseType(); 1650 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1651 // IndexType is size_t, which depends on the bitness of the target. 1652 TypeIndex IndexType = getPointerSizeInBytes() == 8 1653 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1654 : TypeIndex(SimpleTypeKind::UInt32Long); 1655 1656 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1657 1658 // Add subranges to array type. 1659 DINodeArray Elements = Ty->getElements(); 1660 for (int i = Elements.size() - 1; i >= 0; --i) { 1661 const DINode *Element = Elements[i]; 1662 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1663 1664 const DISubrange *Subrange = cast<DISubrange>(Element); 1665 int64_t Count = -1; 1666 1667 // If Subrange has a Count field, use it. 1668 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1669 // where lowerbound is from the LowerBound field of the Subrange, 1670 // or the language default lowerbound if that field is unspecified. 1671 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1672 Count = CI->getSExtValue(); 1673 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1674 // Fortran uses 1 as the default lowerbound; other languages use 0. 1675 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1676 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1677 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1678 Count = UI->getSExtValue() - Lowerbound + 1; 1679 } 1680 1681 // Forward declarations of arrays without a size and VLAs use a count of -1. 1682 // Emit a count of zero in these cases to match what MSVC does for arrays 1683 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1684 // should do for them even if we could distinguish them. 1685 if (Count == -1) 1686 Count = 0; 1687 1688 // Update the element size and element type index for subsequent subranges. 1689 ElementSize *= Count; 1690 1691 // If this is the outermost array, use the size from the array. It will be 1692 // more accurate if we had a VLA or an incomplete element type size. 1693 uint64_t ArraySize = 1694 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1695 1696 StringRef Name = (i == 0) ? Ty->getName() : ""; 1697 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1698 ElementTypeIndex = TypeTable.writeLeafType(AR); 1699 } 1700 1701 return ElementTypeIndex; 1702 } 1703 1704 // This function lowers a Fortran character type (DIStringType). 1705 // Note that it handles only the character*n variant (using SizeInBits 1706 // field in DIString to describe the type size) at the moment. 1707 // Other variants (leveraging the StringLength and StringLengthExp 1708 // fields in DIStringType) remain TBD. 1709 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1710 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1711 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1712 StringRef Name = Ty->getName(); 1713 // IndexType is size_t, which depends on the bitness of the target. 1714 TypeIndex IndexType = getPointerSizeInBytes() == 8 1715 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1716 : TypeIndex(SimpleTypeKind::UInt32Long); 1717 1718 // Create a type of character array of ArraySize. 1719 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1720 1721 return TypeTable.writeLeafType(AR); 1722 } 1723 1724 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1725 TypeIndex Index; 1726 dwarf::TypeKind Kind; 1727 uint32_t ByteSize; 1728 1729 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1730 ByteSize = Ty->getSizeInBits() / 8; 1731 1732 SimpleTypeKind STK = SimpleTypeKind::None; 1733 switch (Kind) { 1734 case dwarf::DW_ATE_address: 1735 // FIXME: Translate 1736 break; 1737 case dwarf::DW_ATE_boolean: 1738 switch (ByteSize) { 1739 case 1: STK = SimpleTypeKind::Boolean8; break; 1740 case 2: STK = SimpleTypeKind::Boolean16; break; 1741 case 4: STK = SimpleTypeKind::Boolean32; break; 1742 case 8: STK = SimpleTypeKind::Boolean64; break; 1743 case 16: STK = SimpleTypeKind::Boolean128; break; 1744 } 1745 break; 1746 case dwarf::DW_ATE_complex_float: 1747 switch (ByteSize) { 1748 case 2: STK = SimpleTypeKind::Complex16; break; 1749 case 4: STK = SimpleTypeKind::Complex32; break; 1750 case 8: STK = SimpleTypeKind::Complex64; break; 1751 case 10: STK = SimpleTypeKind::Complex80; break; 1752 case 16: STK = SimpleTypeKind::Complex128; break; 1753 } 1754 break; 1755 case dwarf::DW_ATE_float: 1756 switch (ByteSize) { 1757 case 2: STK = SimpleTypeKind::Float16; break; 1758 case 4: STK = SimpleTypeKind::Float32; break; 1759 case 6: STK = SimpleTypeKind::Float48; break; 1760 case 8: STK = SimpleTypeKind::Float64; break; 1761 case 10: STK = SimpleTypeKind::Float80; break; 1762 case 16: STK = SimpleTypeKind::Float128; break; 1763 } 1764 break; 1765 case dwarf::DW_ATE_signed: 1766 switch (ByteSize) { 1767 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1768 case 2: STK = SimpleTypeKind::Int16Short; break; 1769 case 4: STK = SimpleTypeKind::Int32; break; 1770 case 8: STK = SimpleTypeKind::Int64Quad; break; 1771 case 16: STK = SimpleTypeKind::Int128Oct; break; 1772 } 1773 break; 1774 case dwarf::DW_ATE_unsigned: 1775 switch (ByteSize) { 1776 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1777 case 2: STK = SimpleTypeKind::UInt16Short; break; 1778 case 4: STK = SimpleTypeKind::UInt32; break; 1779 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1780 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1781 } 1782 break; 1783 case dwarf::DW_ATE_UTF: 1784 switch (ByteSize) { 1785 case 2: STK = SimpleTypeKind::Character16; break; 1786 case 4: STK = SimpleTypeKind::Character32; break; 1787 } 1788 break; 1789 case dwarf::DW_ATE_signed_char: 1790 if (ByteSize == 1) 1791 STK = SimpleTypeKind::SignedCharacter; 1792 break; 1793 case dwarf::DW_ATE_unsigned_char: 1794 if (ByteSize == 1) 1795 STK = SimpleTypeKind::UnsignedCharacter; 1796 break; 1797 default: 1798 break; 1799 } 1800 1801 // Apply some fixups based on the source-level type name. 1802 // Include some amount of canonicalization from an old naming scheme Clang 1803 // used to use for integer types (in an outdated effort to be compatible with 1804 // GCC's debug info/GDB's behavior, which has since been addressed). 1805 if (STK == SimpleTypeKind::Int32 && 1806 (Ty->getName() == "long int" || Ty->getName() == "long")) 1807 STK = SimpleTypeKind::Int32Long; 1808 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1809 Ty->getName() == "unsigned long")) 1810 STK = SimpleTypeKind::UInt32Long; 1811 if (STK == SimpleTypeKind::UInt16Short && 1812 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1813 STK = SimpleTypeKind::WideCharacter; 1814 if ((STK == SimpleTypeKind::SignedCharacter || 1815 STK == SimpleTypeKind::UnsignedCharacter) && 1816 Ty->getName() == "char") 1817 STK = SimpleTypeKind::NarrowCharacter; 1818 1819 return TypeIndex(STK); 1820 } 1821 1822 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1823 PointerOptions PO) { 1824 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1825 1826 // Pointers to simple types without any options can use SimpleTypeMode, rather 1827 // than having a dedicated pointer type record. 1828 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1829 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1830 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1831 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1832 ? SimpleTypeMode::NearPointer64 1833 : SimpleTypeMode::NearPointer32; 1834 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1835 } 1836 1837 PointerKind PK = 1838 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1839 PointerMode PM = PointerMode::Pointer; 1840 switch (Ty->getTag()) { 1841 default: llvm_unreachable("not a pointer tag type"); 1842 case dwarf::DW_TAG_pointer_type: 1843 PM = PointerMode::Pointer; 1844 break; 1845 case dwarf::DW_TAG_reference_type: 1846 PM = PointerMode::LValueReference; 1847 break; 1848 case dwarf::DW_TAG_rvalue_reference_type: 1849 PM = PointerMode::RValueReference; 1850 break; 1851 } 1852 1853 if (Ty->isObjectPointer()) 1854 PO |= PointerOptions::Const; 1855 1856 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1857 return TypeTable.writeLeafType(PR); 1858 } 1859 1860 static PointerToMemberRepresentation 1861 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1862 // SizeInBytes being zero generally implies that the member pointer type was 1863 // incomplete, which can happen if it is part of a function prototype. In this 1864 // case, use the unknown model instead of the general model. 1865 if (IsPMF) { 1866 switch (Flags & DINode::FlagPtrToMemberRep) { 1867 case 0: 1868 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1869 : PointerToMemberRepresentation::GeneralFunction; 1870 case DINode::FlagSingleInheritance: 1871 return PointerToMemberRepresentation::SingleInheritanceFunction; 1872 case DINode::FlagMultipleInheritance: 1873 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1874 case DINode::FlagVirtualInheritance: 1875 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1876 } 1877 } else { 1878 switch (Flags & DINode::FlagPtrToMemberRep) { 1879 case 0: 1880 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1881 : PointerToMemberRepresentation::GeneralData; 1882 case DINode::FlagSingleInheritance: 1883 return PointerToMemberRepresentation::SingleInheritanceData; 1884 case DINode::FlagMultipleInheritance: 1885 return PointerToMemberRepresentation::MultipleInheritanceData; 1886 case DINode::FlagVirtualInheritance: 1887 return PointerToMemberRepresentation::VirtualInheritanceData; 1888 } 1889 } 1890 llvm_unreachable("invalid ptr to member representation"); 1891 } 1892 1893 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1894 PointerOptions PO) { 1895 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1896 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1897 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1898 TypeIndex PointeeTI = 1899 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1900 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1901 : PointerKind::Near32; 1902 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1903 : PointerMode::PointerToDataMember; 1904 1905 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1906 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1907 MemberPointerInfo MPI( 1908 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1909 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1910 return TypeTable.writeLeafType(PR); 1911 } 1912 1913 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1914 /// have a translation, use the NearC convention. 1915 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1916 switch (DwarfCC) { 1917 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1918 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1919 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1920 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1921 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1922 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1923 } 1924 return CallingConvention::NearC; 1925 } 1926 1927 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1928 ModifierOptions Mods = ModifierOptions::None; 1929 PointerOptions PO = PointerOptions::None; 1930 bool IsModifier = true; 1931 const DIType *BaseTy = Ty; 1932 while (IsModifier && BaseTy) { 1933 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1934 switch (BaseTy->getTag()) { 1935 case dwarf::DW_TAG_const_type: 1936 Mods |= ModifierOptions::Const; 1937 PO |= PointerOptions::Const; 1938 break; 1939 case dwarf::DW_TAG_volatile_type: 1940 Mods |= ModifierOptions::Volatile; 1941 PO |= PointerOptions::Volatile; 1942 break; 1943 case dwarf::DW_TAG_restrict_type: 1944 // Only pointer types be marked with __restrict. There is no known flag 1945 // for __restrict in LF_MODIFIER records. 1946 PO |= PointerOptions::Restrict; 1947 break; 1948 default: 1949 IsModifier = false; 1950 break; 1951 } 1952 if (IsModifier) 1953 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1954 } 1955 1956 // Check if the inner type will use an LF_POINTER record. If so, the 1957 // qualifiers will go in the LF_POINTER record. This comes up for types like 1958 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1959 // char *'. 1960 if (BaseTy) { 1961 switch (BaseTy->getTag()) { 1962 case dwarf::DW_TAG_pointer_type: 1963 case dwarf::DW_TAG_reference_type: 1964 case dwarf::DW_TAG_rvalue_reference_type: 1965 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1966 case dwarf::DW_TAG_ptr_to_member_type: 1967 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1968 default: 1969 break; 1970 } 1971 } 1972 1973 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1974 1975 // Return the base type index if there aren't any modifiers. For example, the 1976 // metadata could contain restrict wrappers around non-pointer types. 1977 if (Mods == ModifierOptions::None) 1978 return ModifiedTI; 1979 1980 ModifierRecord MR(ModifiedTI, Mods); 1981 return TypeTable.writeLeafType(MR); 1982 } 1983 1984 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1985 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1986 for (const DIType *ArgType : Ty->getTypeArray()) 1987 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1988 1989 // MSVC uses type none for variadic argument. 1990 if (ReturnAndArgTypeIndices.size() > 1 && 1991 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1992 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1993 } 1994 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1995 ArrayRef<TypeIndex> ArgTypeIndices = None; 1996 if (!ReturnAndArgTypeIndices.empty()) { 1997 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1998 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1999 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2000 } 2001 2002 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2003 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2004 2005 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2006 2007 FunctionOptions FO = getFunctionOptions(Ty); 2008 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2009 ArgListIndex); 2010 return TypeTable.writeLeafType(Procedure); 2011 } 2012 2013 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2014 const DIType *ClassTy, 2015 int ThisAdjustment, 2016 bool IsStaticMethod, 2017 FunctionOptions FO) { 2018 // Lower the containing class type. 2019 TypeIndex ClassType = getTypeIndex(ClassTy); 2020 2021 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2022 2023 unsigned Index = 0; 2024 SmallVector<TypeIndex, 8> ArgTypeIndices; 2025 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2026 if (ReturnAndArgs.size() > Index) { 2027 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2028 } 2029 2030 // If the first argument is a pointer type and this isn't a static method, 2031 // treat it as the special 'this' parameter, which is encoded separately from 2032 // the arguments. 2033 TypeIndex ThisTypeIndex; 2034 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2035 if (const DIDerivedType *PtrTy = 2036 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2037 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2038 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2039 Index++; 2040 } 2041 } 2042 } 2043 2044 while (Index < ReturnAndArgs.size()) 2045 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2046 2047 // MSVC uses type none for variadic argument. 2048 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2049 ArgTypeIndices.back() = TypeIndex::None(); 2050 2051 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2052 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2053 2054 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2055 2056 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2057 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2058 return TypeTable.writeLeafType(MFR); 2059 } 2060 2061 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2062 unsigned VSlotCount = 2063 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2064 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2065 2066 VFTableShapeRecord VFTSR(Slots); 2067 return TypeTable.writeLeafType(VFTSR); 2068 } 2069 2070 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2071 switch (Flags & DINode::FlagAccessibility) { 2072 case DINode::FlagPrivate: return MemberAccess::Private; 2073 case DINode::FlagPublic: return MemberAccess::Public; 2074 case DINode::FlagProtected: return MemberAccess::Protected; 2075 case 0: 2076 // If there was no explicit access control, provide the default for the tag. 2077 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2078 : MemberAccess::Public; 2079 } 2080 llvm_unreachable("access flags are exclusive"); 2081 } 2082 2083 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2084 if (SP->isArtificial()) 2085 return MethodOptions::CompilerGenerated; 2086 2087 // FIXME: Handle other MethodOptions. 2088 2089 return MethodOptions::None; 2090 } 2091 2092 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2093 bool Introduced) { 2094 if (SP->getFlags() & DINode::FlagStaticMember) 2095 return MethodKind::Static; 2096 2097 switch (SP->getVirtuality()) { 2098 case dwarf::DW_VIRTUALITY_none: 2099 break; 2100 case dwarf::DW_VIRTUALITY_virtual: 2101 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2102 case dwarf::DW_VIRTUALITY_pure_virtual: 2103 return Introduced ? MethodKind::PureIntroducingVirtual 2104 : MethodKind::PureVirtual; 2105 default: 2106 llvm_unreachable("unhandled virtuality case"); 2107 } 2108 2109 return MethodKind::Vanilla; 2110 } 2111 2112 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2113 switch (Ty->getTag()) { 2114 case dwarf::DW_TAG_class_type: 2115 return TypeRecordKind::Class; 2116 case dwarf::DW_TAG_structure_type: 2117 return TypeRecordKind::Struct; 2118 default: 2119 llvm_unreachable("unexpected tag"); 2120 } 2121 } 2122 2123 /// Return ClassOptions that should be present on both the forward declaration 2124 /// and the defintion of a tag type. 2125 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2126 ClassOptions CO = ClassOptions::None; 2127 2128 // MSVC always sets this flag, even for local types. Clang doesn't always 2129 // appear to give every type a linkage name, which may be problematic for us. 2130 // FIXME: Investigate the consequences of not following them here. 2131 if (!Ty->getIdentifier().empty()) 2132 CO |= ClassOptions::HasUniqueName; 2133 2134 // Put the Nested flag on a type if it appears immediately inside a tag type. 2135 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2136 // here. That flag is only set on definitions, and not forward declarations. 2137 const DIScope *ImmediateScope = Ty->getScope(); 2138 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2139 CO |= ClassOptions::Nested; 2140 2141 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2142 // type only when it has an immediate function scope. Clang never puts enums 2143 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2144 // always in function, class, or file scopes. 2145 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2146 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2147 CO |= ClassOptions::Scoped; 2148 } else { 2149 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2150 Scope = Scope->getScope()) { 2151 if (isa<DISubprogram>(Scope)) { 2152 CO |= ClassOptions::Scoped; 2153 break; 2154 } 2155 } 2156 } 2157 2158 return CO; 2159 } 2160 2161 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2162 switch (Ty->getTag()) { 2163 case dwarf::DW_TAG_class_type: 2164 case dwarf::DW_TAG_structure_type: 2165 case dwarf::DW_TAG_union_type: 2166 case dwarf::DW_TAG_enumeration_type: 2167 break; 2168 default: 2169 return; 2170 } 2171 2172 if (const auto *File = Ty->getFile()) { 2173 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2174 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2175 2176 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2177 TypeTable.writeLeafType(USLR); 2178 } 2179 } 2180 2181 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2182 ClassOptions CO = getCommonClassOptions(Ty); 2183 TypeIndex FTI; 2184 unsigned EnumeratorCount = 0; 2185 2186 if (Ty->isForwardDecl()) { 2187 CO |= ClassOptions::ForwardReference; 2188 } else { 2189 ContinuationRecordBuilder ContinuationBuilder; 2190 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2191 for (const DINode *Element : Ty->getElements()) { 2192 // We assume that the frontend provides all members in source declaration 2193 // order, which is what MSVC does. 2194 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2195 // FIXME: Is it correct to always emit these as unsigned here? 2196 EnumeratorRecord ER(MemberAccess::Public, 2197 APSInt(Enumerator->getValue(), true), 2198 Enumerator->getName()); 2199 ContinuationBuilder.writeMemberType(ER); 2200 EnumeratorCount++; 2201 } 2202 } 2203 FTI = TypeTable.insertRecord(ContinuationBuilder); 2204 } 2205 2206 std::string FullName = getFullyQualifiedName(Ty); 2207 2208 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2209 getTypeIndex(Ty->getBaseType())); 2210 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2211 2212 addUDTSrcLine(Ty, EnumTI); 2213 2214 return EnumTI; 2215 } 2216 2217 //===----------------------------------------------------------------------===// 2218 // ClassInfo 2219 //===----------------------------------------------------------------------===// 2220 2221 struct llvm::ClassInfo { 2222 struct MemberInfo { 2223 const DIDerivedType *MemberTypeNode; 2224 uint64_t BaseOffset; 2225 }; 2226 // [MemberInfo] 2227 using MemberList = std::vector<MemberInfo>; 2228 2229 using MethodsList = TinyPtrVector<const DISubprogram *>; 2230 // MethodName -> MethodsList 2231 using MethodsMap = MapVector<MDString *, MethodsList>; 2232 2233 /// Base classes. 2234 std::vector<const DIDerivedType *> Inheritance; 2235 2236 /// Direct members. 2237 MemberList Members; 2238 // Direct overloaded methods gathered by name. 2239 MethodsMap Methods; 2240 2241 TypeIndex VShapeTI; 2242 2243 std::vector<const DIType *> NestedTypes; 2244 }; 2245 2246 void CodeViewDebug::clear() { 2247 assert(CurFn == nullptr); 2248 FileIdMap.clear(); 2249 FnDebugInfo.clear(); 2250 FileToFilepathMap.clear(); 2251 LocalUDTs.clear(); 2252 GlobalUDTs.clear(); 2253 TypeIndices.clear(); 2254 CompleteTypeIndices.clear(); 2255 ScopeGlobals.clear(); 2256 CVGlobalVariableOffsets.clear(); 2257 } 2258 2259 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2260 const DIDerivedType *DDTy) { 2261 if (!DDTy->getName().empty()) { 2262 Info.Members.push_back({DDTy, 0}); 2263 2264 // Collect static const data members with values. 2265 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2266 DINode::FlagStaticMember) { 2267 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2268 isa<ConstantFP>(DDTy->getConstant()))) 2269 StaticConstMembers.push_back(DDTy); 2270 } 2271 2272 return; 2273 } 2274 2275 // An unnamed member may represent a nested struct or union. Attempt to 2276 // interpret the unnamed member as a DICompositeType possibly wrapped in 2277 // qualifier types. Add all the indirect fields to the current record if that 2278 // succeeds, and drop the member if that fails. 2279 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2280 uint64_t Offset = DDTy->getOffsetInBits(); 2281 const DIType *Ty = DDTy->getBaseType(); 2282 bool FullyResolved = false; 2283 while (!FullyResolved) { 2284 switch (Ty->getTag()) { 2285 case dwarf::DW_TAG_const_type: 2286 case dwarf::DW_TAG_volatile_type: 2287 // FIXME: we should apply the qualifier types to the indirect fields 2288 // rather than dropping them. 2289 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2290 break; 2291 default: 2292 FullyResolved = true; 2293 break; 2294 } 2295 } 2296 2297 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2298 if (!DCTy) 2299 return; 2300 2301 ClassInfo NestedInfo = collectClassInfo(DCTy); 2302 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2303 Info.Members.push_back( 2304 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2305 } 2306 2307 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2308 ClassInfo Info; 2309 // Add elements to structure type. 2310 DINodeArray Elements = Ty->getElements(); 2311 for (auto *Element : Elements) { 2312 // We assume that the frontend provides all members in source declaration 2313 // order, which is what MSVC does. 2314 if (!Element) 2315 continue; 2316 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2317 Info.Methods[SP->getRawName()].push_back(SP); 2318 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2319 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2320 collectMemberInfo(Info, DDTy); 2321 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2322 Info.Inheritance.push_back(DDTy); 2323 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2324 DDTy->getName() == "__vtbl_ptr_type") { 2325 Info.VShapeTI = getTypeIndex(DDTy); 2326 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2327 Info.NestedTypes.push_back(DDTy); 2328 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2329 // Ignore friend members. It appears that MSVC emitted info about 2330 // friends in the past, but modern versions do not. 2331 } 2332 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2333 Info.NestedTypes.push_back(Composite); 2334 } 2335 // Skip other unrecognized kinds of elements. 2336 } 2337 return Info; 2338 } 2339 2340 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2341 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2342 // if a complete type should be emitted instead of a forward reference. 2343 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2344 !Ty->isForwardDecl(); 2345 } 2346 2347 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2348 // Emit the complete type for unnamed structs. C++ classes with methods 2349 // which have a circular reference back to the class type are expected to 2350 // be named by the front-end and should not be "unnamed". C unnamed 2351 // structs should not have circular references. 2352 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2353 // If this unnamed complete type is already in the process of being defined 2354 // then the description of the type is malformed and cannot be emitted 2355 // into CodeView correctly so report a fatal error. 2356 auto I = CompleteTypeIndices.find(Ty); 2357 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2358 report_fatal_error("cannot debug circular reference to unnamed type"); 2359 return getCompleteTypeIndex(Ty); 2360 } 2361 2362 // First, construct the forward decl. Don't look into Ty to compute the 2363 // forward decl options, since it might not be available in all TUs. 2364 TypeRecordKind Kind = getRecordKind(Ty); 2365 ClassOptions CO = 2366 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2367 std::string FullName = getFullyQualifiedName(Ty); 2368 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2369 FullName, Ty->getIdentifier()); 2370 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2371 if (!Ty->isForwardDecl()) 2372 DeferredCompleteTypes.push_back(Ty); 2373 return FwdDeclTI; 2374 } 2375 2376 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2377 // Construct the field list and complete type record. 2378 TypeRecordKind Kind = getRecordKind(Ty); 2379 ClassOptions CO = getCommonClassOptions(Ty); 2380 TypeIndex FieldTI; 2381 TypeIndex VShapeTI; 2382 unsigned FieldCount; 2383 bool ContainsNestedClass; 2384 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2385 lowerRecordFieldList(Ty); 2386 2387 if (ContainsNestedClass) 2388 CO |= ClassOptions::ContainsNestedClass; 2389 2390 // MSVC appears to set this flag by searching any destructor or method with 2391 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2392 // the members, however special member functions are not yet emitted into 2393 // debug information. For now checking a class's non-triviality seems enough. 2394 // FIXME: not true for a nested unnamed struct. 2395 if (isNonTrivial(Ty)) 2396 CO |= ClassOptions::HasConstructorOrDestructor; 2397 2398 std::string FullName = getFullyQualifiedName(Ty); 2399 2400 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2401 2402 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2403 SizeInBytes, FullName, Ty->getIdentifier()); 2404 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2405 2406 addUDTSrcLine(Ty, ClassTI); 2407 2408 addToUDTs(Ty); 2409 2410 return ClassTI; 2411 } 2412 2413 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2414 // Emit the complete type for unnamed unions. 2415 if (shouldAlwaysEmitCompleteClassType(Ty)) 2416 return getCompleteTypeIndex(Ty); 2417 2418 ClassOptions CO = 2419 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2420 std::string FullName = getFullyQualifiedName(Ty); 2421 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2422 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2423 if (!Ty->isForwardDecl()) 2424 DeferredCompleteTypes.push_back(Ty); 2425 return FwdDeclTI; 2426 } 2427 2428 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2429 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2430 TypeIndex FieldTI; 2431 unsigned FieldCount; 2432 bool ContainsNestedClass; 2433 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2434 lowerRecordFieldList(Ty); 2435 2436 if (ContainsNestedClass) 2437 CO |= ClassOptions::ContainsNestedClass; 2438 2439 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2440 std::string FullName = getFullyQualifiedName(Ty); 2441 2442 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2443 Ty->getIdentifier()); 2444 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2445 2446 addUDTSrcLine(Ty, UnionTI); 2447 2448 addToUDTs(Ty); 2449 2450 return UnionTI; 2451 } 2452 2453 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2454 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2455 // Manually count members. MSVC appears to count everything that generates a 2456 // field list record. Each individual overload in a method overload group 2457 // contributes to this count, even though the overload group is a single field 2458 // list record. 2459 unsigned MemberCount = 0; 2460 ClassInfo Info = collectClassInfo(Ty); 2461 ContinuationRecordBuilder ContinuationBuilder; 2462 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2463 2464 // Create base classes. 2465 for (const DIDerivedType *I : Info.Inheritance) { 2466 if (I->getFlags() & DINode::FlagVirtual) { 2467 // Virtual base. 2468 unsigned VBPtrOffset = I->getVBPtrOffset(); 2469 // FIXME: Despite the accessor name, the offset is really in bytes. 2470 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2471 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2472 ? TypeRecordKind::IndirectVirtualBaseClass 2473 : TypeRecordKind::VirtualBaseClass; 2474 VirtualBaseClassRecord VBCR( 2475 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2476 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2477 VBTableIndex); 2478 2479 ContinuationBuilder.writeMemberType(VBCR); 2480 MemberCount++; 2481 } else { 2482 assert(I->getOffsetInBits() % 8 == 0 && 2483 "bases must be on byte boundaries"); 2484 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2485 getTypeIndex(I->getBaseType()), 2486 I->getOffsetInBits() / 8); 2487 ContinuationBuilder.writeMemberType(BCR); 2488 MemberCount++; 2489 } 2490 } 2491 2492 // Create members. 2493 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2494 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2495 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2496 StringRef MemberName = Member->getName(); 2497 MemberAccess Access = 2498 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2499 2500 if (Member->isStaticMember()) { 2501 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2502 ContinuationBuilder.writeMemberType(SDMR); 2503 MemberCount++; 2504 continue; 2505 } 2506 2507 // Virtual function pointer member. 2508 if ((Member->getFlags() & DINode::FlagArtificial) && 2509 Member->getName().startswith("_vptr$")) { 2510 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2511 ContinuationBuilder.writeMemberType(VFPR); 2512 MemberCount++; 2513 continue; 2514 } 2515 2516 // Data member. 2517 uint64_t MemberOffsetInBits = 2518 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2519 if (Member->isBitField()) { 2520 uint64_t StartBitOffset = MemberOffsetInBits; 2521 if (const auto *CI = 2522 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2523 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2524 } 2525 StartBitOffset -= MemberOffsetInBits; 2526 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2527 StartBitOffset); 2528 MemberBaseType = TypeTable.writeLeafType(BFR); 2529 } 2530 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2531 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2532 MemberName); 2533 ContinuationBuilder.writeMemberType(DMR); 2534 MemberCount++; 2535 } 2536 2537 // Create methods 2538 for (auto &MethodItr : Info.Methods) { 2539 StringRef Name = MethodItr.first->getString(); 2540 2541 std::vector<OneMethodRecord> Methods; 2542 for (const DISubprogram *SP : MethodItr.second) { 2543 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2544 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2545 2546 unsigned VFTableOffset = -1; 2547 if (Introduced) 2548 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2549 2550 Methods.push_back(OneMethodRecord( 2551 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2552 translateMethodKindFlags(SP, Introduced), 2553 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2554 MemberCount++; 2555 } 2556 assert(!Methods.empty() && "Empty methods map entry"); 2557 if (Methods.size() == 1) 2558 ContinuationBuilder.writeMemberType(Methods[0]); 2559 else { 2560 // FIXME: Make this use its own ContinuationBuilder so that 2561 // MethodOverloadList can be split correctly. 2562 MethodOverloadListRecord MOLR(Methods); 2563 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2564 2565 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2566 ContinuationBuilder.writeMemberType(OMR); 2567 } 2568 } 2569 2570 // Create nested classes. 2571 for (const DIType *Nested : Info.NestedTypes) { 2572 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2573 ContinuationBuilder.writeMemberType(R); 2574 MemberCount++; 2575 } 2576 2577 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2578 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2579 !Info.NestedTypes.empty()); 2580 } 2581 2582 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2583 if (!VBPType.getIndex()) { 2584 // Make a 'const int *' type. 2585 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2586 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2587 2588 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2589 : PointerKind::Near32; 2590 PointerMode PM = PointerMode::Pointer; 2591 PointerOptions PO = PointerOptions::None; 2592 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2593 VBPType = TypeTable.writeLeafType(PR); 2594 } 2595 2596 return VBPType; 2597 } 2598 2599 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2600 // The null DIType is the void type. Don't try to hash it. 2601 if (!Ty) 2602 return TypeIndex::Void(); 2603 2604 // Check if we've already translated this type. Don't try to do a 2605 // get-or-create style insertion that caches the hash lookup across the 2606 // lowerType call. It will update the TypeIndices map. 2607 auto I = TypeIndices.find({Ty, ClassTy}); 2608 if (I != TypeIndices.end()) 2609 return I->second; 2610 2611 TypeLoweringScope S(*this); 2612 TypeIndex TI = lowerType(Ty, ClassTy); 2613 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2614 } 2615 2616 codeview::TypeIndex 2617 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2618 const DISubroutineType *SubroutineTy) { 2619 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2620 "this type must be a pointer type"); 2621 2622 PointerOptions Options = PointerOptions::None; 2623 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2624 Options = PointerOptions::LValueRefThisPointer; 2625 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2626 Options = PointerOptions::RValueRefThisPointer; 2627 2628 // Check if we've already translated this type. If there is no ref qualifier 2629 // on the function then we look up this pointer type with no associated class 2630 // so that the TypeIndex for the this pointer can be shared with the type 2631 // index for other pointers to this class type. If there is a ref qualifier 2632 // then we lookup the pointer using the subroutine as the parent type. 2633 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2634 if (I != TypeIndices.end()) 2635 return I->second; 2636 2637 TypeLoweringScope S(*this); 2638 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2639 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2640 } 2641 2642 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2643 PointerRecord PR(getTypeIndex(Ty), 2644 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2645 : PointerKind::Near32, 2646 PointerMode::LValueReference, PointerOptions::None, 2647 Ty->getSizeInBits() / 8); 2648 return TypeTable.writeLeafType(PR); 2649 } 2650 2651 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2652 // The null DIType is the void type. Don't try to hash it. 2653 if (!Ty) 2654 return TypeIndex::Void(); 2655 2656 // Look through typedefs when getting the complete type index. Call 2657 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2658 // emitted only once. 2659 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2660 (void)getTypeIndex(Ty); 2661 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2662 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2663 2664 // If this is a non-record type, the complete type index is the same as the 2665 // normal type index. Just call getTypeIndex. 2666 switch (Ty->getTag()) { 2667 case dwarf::DW_TAG_class_type: 2668 case dwarf::DW_TAG_structure_type: 2669 case dwarf::DW_TAG_union_type: 2670 break; 2671 default: 2672 return getTypeIndex(Ty); 2673 } 2674 2675 const auto *CTy = cast<DICompositeType>(Ty); 2676 2677 TypeLoweringScope S(*this); 2678 2679 // Make sure the forward declaration is emitted first. It's unclear if this 2680 // is necessary, but MSVC does it, and we should follow suit until we can show 2681 // otherwise. 2682 // We only emit a forward declaration for named types. 2683 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2684 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2685 2686 // Just use the forward decl if we don't have complete type info. This 2687 // might happen if the frontend is using modules and expects the complete 2688 // definition to be emitted elsewhere. 2689 if (CTy->isForwardDecl()) 2690 return FwdDeclTI; 2691 } 2692 2693 // Check if we've already translated the complete record type. 2694 // Insert the type with a null TypeIndex to signify that the type is currently 2695 // being lowered. 2696 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2697 if (!InsertResult.second) 2698 return InsertResult.first->second; 2699 2700 TypeIndex TI; 2701 switch (CTy->getTag()) { 2702 case dwarf::DW_TAG_class_type: 2703 case dwarf::DW_TAG_structure_type: 2704 TI = lowerCompleteTypeClass(CTy); 2705 break; 2706 case dwarf::DW_TAG_union_type: 2707 TI = lowerCompleteTypeUnion(CTy); 2708 break; 2709 default: 2710 llvm_unreachable("not a record"); 2711 } 2712 2713 // Update the type index associated with this CompositeType. This cannot 2714 // use the 'InsertResult' iterator above because it is potentially 2715 // invalidated by map insertions which can occur while lowering the class 2716 // type above. 2717 CompleteTypeIndices[CTy] = TI; 2718 return TI; 2719 } 2720 2721 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2722 /// and do this until fixpoint, as each complete record type typically 2723 /// references 2724 /// many other record types. 2725 void CodeViewDebug::emitDeferredCompleteTypes() { 2726 SmallVector<const DICompositeType *, 4> TypesToEmit; 2727 while (!DeferredCompleteTypes.empty()) { 2728 std::swap(DeferredCompleteTypes, TypesToEmit); 2729 for (const DICompositeType *RecordTy : TypesToEmit) 2730 getCompleteTypeIndex(RecordTy); 2731 TypesToEmit.clear(); 2732 } 2733 } 2734 2735 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2736 ArrayRef<LocalVariable> Locals) { 2737 // Get the sorted list of parameters and emit them first. 2738 SmallVector<const LocalVariable *, 6> Params; 2739 for (const LocalVariable &L : Locals) 2740 if (L.DIVar->isParameter()) 2741 Params.push_back(&L); 2742 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2743 return L->DIVar->getArg() < R->DIVar->getArg(); 2744 }); 2745 for (const LocalVariable *L : Params) 2746 emitLocalVariable(FI, *L); 2747 2748 // Next emit all non-parameters in the order that we found them. 2749 for (const LocalVariable &L : Locals) 2750 if (!L.DIVar->isParameter()) 2751 emitLocalVariable(FI, L); 2752 } 2753 2754 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2755 const LocalVariable &Var) { 2756 // LocalSym record, see SymbolRecord.h for more info. 2757 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2758 2759 LocalSymFlags Flags = LocalSymFlags::None; 2760 if (Var.DIVar->isParameter()) 2761 Flags |= LocalSymFlags::IsParameter; 2762 if (Var.DefRanges.empty()) 2763 Flags |= LocalSymFlags::IsOptimizedOut; 2764 2765 OS.AddComment("TypeIndex"); 2766 TypeIndex TI = Var.UseReferenceType 2767 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2768 : getCompleteTypeIndex(Var.DIVar->getType()); 2769 OS.emitInt32(TI.getIndex()); 2770 OS.AddComment("Flags"); 2771 OS.emitInt16(static_cast<uint16_t>(Flags)); 2772 // Truncate the name so we won't overflow the record length field. 2773 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2774 endSymbolRecord(LocalEnd); 2775 2776 // Calculate the on disk prefix of the appropriate def range record. The 2777 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2778 // should be big enough to hold all forms without memory allocation. 2779 SmallString<20> BytePrefix; 2780 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2781 BytePrefix.clear(); 2782 if (DefRange.InMemory) { 2783 int Offset = DefRange.DataOffset; 2784 unsigned Reg = DefRange.CVRegister; 2785 2786 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2787 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2788 // instead. In frames without stack realignment, $T0 will be the CFA. 2789 if (RegisterId(Reg) == RegisterId::ESP) { 2790 Reg = unsigned(RegisterId::VFRAME); 2791 Offset += FI.OffsetAdjustment; 2792 } 2793 2794 // If we can use the chosen frame pointer for the frame and this isn't a 2795 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2796 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2797 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2798 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2799 (bool(Flags & LocalSymFlags::IsParameter) 2800 ? (EncFP == FI.EncodedParamFramePtrReg) 2801 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2802 DefRangeFramePointerRelHeader DRHdr; 2803 DRHdr.Offset = Offset; 2804 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2805 } else { 2806 uint16_t RegRelFlags = 0; 2807 if (DefRange.IsSubfield) { 2808 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2809 (DefRange.StructOffset 2810 << DefRangeRegisterRelSym::OffsetInParentShift); 2811 } 2812 DefRangeRegisterRelHeader DRHdr; 2813 DRHdr.Register = Reg; 2814 DRHdr.Flags = RegRelFlags; 2815 DRHdr.BasePointerOffset = Offset; 2816 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2817 } 2818 } else { 2819 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2820 if (DefRange.IsSubfield) { 2821 DefRangeSubfieldRegisterHeader DRHdr; 2822 DRHdr.Register = DefRange.CVRegister; 2823 DRHdr.MayHaveNoName = 0; 2824 DRHdr.OffsetInParent = DefRange.StructOffset; 2825 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2826 } else { 2827 DefRangeRegisterHeader DRHdr; 2828 DRHdr.Register = DefRange.CVRegister; 2829 DRHdr.MayHaveNoName = 0; 2830 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2831 } 2832 } 2833 } 2834 } 2835 2836 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2837 const FunctionInfo& FI) { 2838 for (LexicalBlock *Block : Blocks) 2839 emitLexicalBlock(*Block, FI); 2840 } 2841 2842 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2843 /// lexical block scope. 2844 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2845 const FunctionInfo& FI) { 2846 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2847 OS.AddComment("PtrParent"); 2848 OS.emitInt32(0); // PtrParent 2849 OS.AddComment("PtrEnd"); 2850 OS.emitInt32(0); // PtrEnd 2851 OS.AddComment("Code size"); 2852 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2853 OS.AddComment("Function section relative address"); 2854 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2855 OS.AddComment("Function section index"); 2856 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2857 OS.AddComment("Lexical block name"); 2858 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2859 endSymbolRecord(RecordEnd); 2860 2861 // Emit variables local to this lexical block. 2862 emitLocalVariableList(FI, Block.Locals); 2863 emitGlobalVariableList(Block.Globals); 2864 2865 // Emit lexical blocks contained within this block. 2866 emitLexicalBlockList(Block.Children, FI); 2867 2868 // Close the lexical block scope. 2869 emitEndSymbolRecord(SymbolKind::S_END); 2870 } 2871 2872 /// Convenience routine for collecting lexical block information for a list 2873 /// of lexical scopes. 2874 void CodeViewDebug::collectLexicalBlockInfo( 2875 SmallVectorImpl<LexicalScope *> &Scopes, 2876 SmallVectorImpl<LexicalBlock *> &Blocks, 2877 SmallVectorImpl<LocalVariable> &Locals, 2878 SmallVectorImpl<CVGlobalVariable> &Globals) { 2879 for (LexicalScope *Scope : Scopes) 2880 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2881 } 2882 2883 /// Populate the lexical blocks and local variable lists of the parent with 2884 /// information about the specified lexical scope. 2885 void CodeViewDebug::collectLexicalBlockInfo( 2886 LexicalScope &Scope, 2887 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2888 SmallVectorImpl<LocalVariable> &ParentLocals, 2889 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2890 if (Scope.isAbstractScope()) 2891 return; 2892 2893 // Gather information about the lexical scope including local variables, 2894 // global variables, and address ranges. 2895 bool IgnoreScope = false; 2896 auto LI = ScopeVariables.find(&Scope); 2897 SmallVectorImpl<LocalVariable> *Locals = 2898 LI != ScopeVariables.end() ? &LI->second : nullptr; 2899 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2900 SmallVectorImpl<CVGlobalVariable> *Globals = 2901 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2902 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2903 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2904 2905 // Ignore lexical scopes which do not contain variables. 2906 if (!Locals && !Globals) 2907 IgnoreScope = true; 2908 2909 // Ignore lexical scopes which are not lexical blocks. 2910 if (!DILB) 2911 IgnoreScope = true; 2912 2913 // Ignore scopes which have too many address ranges to represent in the 2914 // current CodeView format or do not have a valid address range. 2915 // 2916 // For lexical scopes with multiple address ranges you may be tempted to 2917 // construct a single range covering every instruction where the block is 2918 // live and everything in between. Unfortunately, Visual Studio only 2919 // displays variables from the first matching lexical block scope. If the 2920 // first lexical block contains exception handling code or cold code which 2921 // is moved to the bottom of the routine creating a single range covering 2922 // nearly the entire routine, then it will hide all other lexical blocks 2923 // and the variables they contain. 2924 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2925 IgnoreScope = true; 2926 2927 if (IgnoreScope) { 2928 // This scope can be safely ignored and eliminating it will reduce the 2929 // size of the debug information. Be sure to collect any variable and scope 2930 // information from the this scope or any of its children and collapse them 2931 // into the parent scope. 2932 if (Locals) 2933 ParentLocals.append(Locals->begin(), Locals->end()); 2934 if (Globals) 2935 ParentGlobals.append(Globals->begin(), Globals->end()); 2936 collectLexicalBlockInfo(Scope.getChildren(), 2937 ParentBlocks, 2938 ParentLocals, 2939 ParentGlobals); 2940 return; 2941 } 2942 2943 // Create a new CodeView lexical block for this lexical scope. If we've 2944 // seen this DILexicalBlock before then the scope tree is malformed and 2945 // we can handle this gracefully by not processing it a second time. 2946 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2947 if (!BlockInsertion.second) 2948 return; 2949 2950 // Create a lexical block containing the variables and collect the the 2951 // lexical block information for the children. 2952 const InsnRange &Range = Ranges.front(); 2953 assert(Range.first && Range.second); 2954 LexicalBlock &Block = BlockInsertion.first->second; 2955 Block.Begin = getLabelBeforeInsn(Range.first); 2956 Block.End = getLabelAfterInsn(Range.second); 2957 assert(Block.Begin && "missing label for scope begin"); 2958 assert(Block.End && "missing label for scope end"); 2959 Block.Name = DILB->getName(); 2960 if (Locals) 2961 Block.Locals = std::move(*Locals); 2962 if (Globals) 2963 Block.Globals = std::move(*Globals); 2964 ParentBlocks.push_back(&Block); 2965 collectLexicalBlockInfo(Scope.getChildren(), 2966 Block.Children, 2967 Block.Locals, 2968 Block.Globals); 2969 } 2970 2971 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2972 const Function &GV = MF->getFunction(); 2973 assert(FnDebugInfo.count(&GV)); 2974 assert(CurFn == FnDebugInfo[&GV].get()); 2975 2976 collectVariableInfo(GV.getSubprogram()); 2977 2978 // Build the lexical block structure to emit for this routine. 2979 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2980 collectLexicalBlockInfo(*CFS, 2981 CurFn->ChildBlocks, 2982 CurFn->Locals, 2983 CurFn->Globals); 2984 2985 // Clear the scope and variable information from the map which will not be 2986 // valid after we have finished processing this routine. This also prepares 2987 // the map for the subsequent routine. 2988 ScopeVariables.clear(); 2989 2990 // Don't emit anything if we don't have any line tables. 2991 // Thunks are compiler-generated and probably won't have source correlation. 2992 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2993 FnDebugInfo.erase(&GV); 2994 CurFn = nullptr; 2995 return; 2996 } 2997 2998 // Find heap alloc sites and add to list. 2999 for (const auto &MBB : *MF) { 3000 for (const auto &MI : MBB) { 3001 if (MDNode *MD = MI.getHeapAllocMarker()) { 3002 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3003 getLabelAfterInsn(&MI), 3004 dyn_cast<DIType>(MD))); 3005 } 3006 } 3007 } 3008 3009 CurFn->Annotations = MF->getCodeViewAnnotations(); 3010 3011 CurFn->End = Asm->getFunctionEnd(); 3012 3013 CurFn = nullptr; 3014 } 3015 3016 // Usable locations are valid with non-zero line numbers. A line number of zero 3017 // corresponds to optimized code that doesn't have a distinct source location. 3018 // In this case, we try to use the previous or next source location depending on 3019 // the context. 3020 static bool isUsableDebugLoc(DebugLoc DL) { 3021 return DL && DL.getLine() != 0; 3022 } 3023 3024 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3025 DebugHandlerBase::beginInstruction(MI); 3026 3027 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3028 if (!Asm || !CurFn || MI->isDebugInstr() || 3029 MI->getFlag(MachineInstr::FrameSetup)) 3030 return; 3031 3032 // If the first instruction of a new MBB has no location, find the first 3033 // instruction with a location and use that. 3034 DebugLoc DL = MI->getDebugLoc(); 3035 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3036 for (const auto &NextMI : *MI->getParent()) { 3037 if (NextMI.isDebugInstr()) 3038 continue; 3039 DL = NextMI.getDebugLoc(); 3040 if (isUsableDebugLoc(DL)) 3041 break; 3042 } 3043 // FIXME: Handle the case where the BB has no valid locations. This would 3044 // probably require doing a real dataflow analysis. 3045 } 3046 PrevInstBB = MI->getParent(); 3047 3048 // If we still don't have a debug location, don't record a location. 3049 if (!isUsableDebugLoc(DL)) 3050 return; 3051 3052 maybeRecordLocation(DL, Asm->MF); 3053 } 3054 3055 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3056 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3057 *EndLabel = MMI->getContext().createTempSymbol(); 3058 OS.emitInt32(unsigned(Kind)); 3059 OS.AddComment("Subsection size"); 3060 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3061 OS.emitLabel(BeginLabel); 3062 return EndLabel; 3063 } 3064 3065 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3066 OS.emitLabel(EndLabel); 3067 // Every subsection must be aligned to a 4-byte boundary. 3068 OS.emitValueToAlignment(4); 3069 } 3070 3071 static StringRef getSymbolName(SymbolKind SymKind) { 3072 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3073 if (EE.Value == SymKind) 3074 return EE.Name; 3075 return ""; 3076 } 3077 3078 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3079 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3080 *EndLabel = MMI->getContext().createTempSymbol(); 3081 OS.AddComment("Record length"); 3082 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3083 OS.emitLabel(BeginLabel); 3084 if (OS.isVerboseAsm()) 3085 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3086 OS.emitInt16(unsigned(SymKind)); 3087 return EndLabel; 3088 } 3089 3090 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3091 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3092 // an extra copy of every symbol record in LLD. This increases object file 3093 // size by less than 1% in the clang build, and is compatible with the Visual 3094 // C++ linker. 3095 OS.emitValueToAlignment(4); 3096 OS.emitLabel(SymEnd); 3097 } 3098 3099 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3100 OS.AddComment("Record length"); 3101 OS.emitInt16(2); 3102 if (OS.isVerboseAsm()) 3103 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3104 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3105 } 3106 3107 void CodeViewDebug::emitDebugInfoForUDTs( 3108 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3109 #ifndef NDEBUG 3110 size_t OriginalSize = UDTs.size(); 3111 #endif 3112 for (const auto &UDT : UDTs) { 3113 const DIType *T = UDT.second; 3114 assert(shouldEmitUdt(T)); 3115 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3116 OS.AddComment("Type"); 3117 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3118 assert(OriginalSize == UDTs.size() && 3119 "getCompleteTypeIndex found new UDTs!"); 3120 emitNullTerminatedSymbolName(OS, UDT.first); 3121 endSymbolRecord(UDTRecordEnd); 3122 } 3123 } 3124 3125 void CodeViewDebug::collectGlobalVariableInfo() { 3126 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3127 GlobalMap; 3128 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3129 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3130 GV.getDebugInfo(GVEs); 3131 for (const auto *GVE : GVEs) 3132 GlobalMap[GVE] = &GV; 3133 } 3134 3135 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3136 for (const MDNode *Node : CUs->operands()) { 3137 const auto *CU = cast<DICompileUnit>(Node); 3138 for (const auto *GVE : CU->getGlobalVariables()) { 3139 const DIGlobalVariable *DIGV = GVE->getVariable(); 3140 const DIExpression *DIE = GVE->getExpression(); 3141 3142 if ((DIE->getNumElements() == 2) && 3143 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3144 // Record the constant offset for the variable. 3145 // 3146 // A Fortran common block uses this idiom to encode the offset 3147 // of a variable from the common block's starting address. 3148 CVGlobalVariableOffsets.insert( 3149 std::make_pair(DIGV, DIE->getElement(1))); 3150 3151 // Emit constant global variables in a global symbol section. 3152 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3153 CVGlobalVariable CVGV = {DIGV, DIE}; 3154 GlobalVariables.emplace_back(std::move(CVGV)); 3155 } 3156 3157 const auto *GV = GlobalMap.lookup(GVE); 3158 if (!GV || GV->isDeclarationForLinker()) 3159 continue; 3160 3161 DIScope *Scope = DIGV->getScope(); 3162 SmallVector<CVGlobalVariable, 1> *VariableList; 3163 if (Scope && isa<DILocalScope>(Scope)) { 3164 // Locate a global variable list for this scope, creating one if 3165 // necessary. 3166 auto Insertion = ScopeGlobals.insert( 3167 {Scope, std::unique_ptr<GlobalVariableList>()}); 3168 if (Insertion.second) 3169 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3170 VariableList = Insertion.first->second.get(); 3171 } else if (GV->hasComdat()) 3172 // Emit this global variable into a COMDAT section. 3173 VariableList = &ComdatVariables; 3174 else 3175 // Emit this global variable in a single global symbol section. 3176 VariableList = &GlobalVariables; 3177 CVGlobalVariable CVGV = {DIGV, GV}; 3178 VariableList->emplace_back(std::move(CVGV)); 3179 } 3180 } 3181 } 3182 3183 void CodeViewDebug::collectDebugInfoForGlobals() { 3184 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3185 const DIGlobalVariable *DIGV = CVGV.DIGV; 3186 const DIScope *Scope = DIGV->getScope(); 3187 getCompleteTypeIndex(DIGV->getType()); 3188 getFullyQualifiedName(Scope, DIGV->getName()); 3189 } 3190 3191 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3192 const DIGlobalVariable *DIGV = CVGV.DIGV; 3193 const DIScope *Scope = DIGV->getScope(); 3194 getCompleteTypeIndex(DIGV->getType()); 3195 getFullyQualifiedName(Scope, DIGV->getName()); 3196 } 3197 } 3198 3199 void CodeViewDebug::emitDebugInfoForGlobals() { 3200 // First, emit all globals that are not in a comdat in a single symbol 3201 // substream. MSVC doesn't like it if the substream is empty, so only open 3202 // it if we have at least one global to emit. 3203 switchToDebugSectionForSymbol(nullptr); 3204 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3205 OS.AddComment("Symbol subsection for globals"); 3206 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3207 emitGlobalVariableList(GlobalVariables); 3208 emitStaticConstMemberList(); 3209 endCVSubsection(EndLabel); 3210 } 3211 3212 // Second, emit each global that is in a comdat into its own .debug$S 3213 // section along with its own symbol substream. 3214 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3215 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3216 MCSymbol *GVSym = Asm->getSymbol(GV); 3217 OS.AddComment("Symbol subsection for " + 3218 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3219 switchToDebugSectionForSymbol(GVSym); 3220 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3221 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3222 emitDebugInfoForGlobal(CVGV); 3223 endCVSubsection(EndLabel); 3224 } 3225 } 3226 3227 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3228 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3229 for (const MDNode *Node : CUs->operands()) { 3230 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3231 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3232 getTypeIndex(RT); 3233 // FIXME: Add to global/local DTU list. 3234 } 3235 } 3236 } 3237 } 3238 3239 // Emit each global variable in the specified array. 3240 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3241 for (const CVGlobalVariable &CVGV : Globals) { 3242 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3243 emitDebugInfoForGlobal(CVGV); 3244 } 3245 } 3246 3247 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3248 const std::string &QualifiedName) { 3249 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3250 OS.AddComment("Type"); 3251 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3252 3253 OS.AddComment("Value"); 3254 3255 // Encoded integers shouldn't need more than 10 bytes. 3256 uint8_t Data[10]; 3257 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3258 CodeViewRecordIO IO(Writer); 3259 cantFail(IO.mapEncodedInteger(Value)); 3260 StringRef SRef((char *)Data, Writer.getOffset()); 3261 OS.emitBinaryData(SRef); 3262 3263 OS.AddComment("Name"); 3264 emitNullTerminatedSymbolName(OS, QualifiedName); 3265 endSymbolRecord(SConstantEnd); 3266 } 3267 3268 void CodeViewDebug::emitStaticConstMemberList() { 3269 for (const DIDerivedType *DTy : StaticConstMembers) { 3270 const DIScope *Scope = DTy->getScope(); 3271 3272 APSInt Value; 3273 if (const ConstantInt *CI = 3274 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3275 Value = APSInt(CI->getValue(), 3276 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3277 else if (const ConstantFP *CFP = 3278 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3279 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3280 else 3281 llvm_unreachable("cannot emit a constant without a value"); 3282 3283 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3284 getFullyQualifiedName(Scope, DTy->getName())); 3285 } 3286 } 3287 3288 static bool isFloatDIType(const DIType *Ty) { 3289 if (isa<DICompositeType>(Ty)) 3290 return false; 3291 3292 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3293 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3294 if (T == dwarf::DW_TAG_pointer_type || 3295 T == dwarf::DW_TAG_ptr_to_member_type || 3296 T == dwarf::DW_TAG_reference_type || 3297 T == dwarf::DW_TAG_rvalue_reference_type) 3298 return false; 3299 assert(DTy->getBaseType() && "Expected valid base type"); 3300 return isFloatDIType(DTy->getBaseType()); 3301 } 3302 3303 auto *BTy = cast<DIBasicType>(Ty); 3304 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3305 } 3306 3307 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3308 const DIGlobalVariable *DIGV = CVGV.DIGV; 3309 3310 const DIScope *Scope = DIGV->getScope(); 3311 // For static data members, get the scope from the declaration. 3312 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3313 DIGV->getRawStaticDataMemberDeclaration())) 3314 Scope = MemberDecl->getScope(); 3315 // For Fortran, the scoping portion is elided in its name so that we can 3316 // reference the variable in the command line of the VS debugger. 3317 std::string QualifiedName = 3318 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3319 : getFullyQualifiedName(Scope, DIGV->getName()); 3320 3321 if (const GlobalVariable *GV = 3322 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3323 // DataSym record, see SymbolRecord.h for more info. Thread local data 3324 // happens to have the same format as global data. 3325 MCSymbol *GVSym = Asm->getSymbol(GV); 3326 SymbolKind DataSym = GV->isThreadLocal() 3327 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3328 : SymbolKind::S_GTHREAD32) 3329 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3330 : SymbolKind::S_GDATA32); 3331 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3332 OS.AddComment("Type"); 3333 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3334 OS.AddComment("DataOffset"); 3335 3336 uint64_t Offset = 0; 3337 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3338 // Use the offset seen while collecting info on globals. 3339 Offset = CVGlobalVariableOffsets[DIGV]; 3340 OS.EmitCOFFSecRel32(GVSym, Offset); 3341 3342 OS.AddComment("Segment"); 3343 OS.EmitCOFFSectionIndex(GVSym); 3344 OS.AddComment("Name"); 3345 const unsigned LengthOfDataRecord = 12; 3346 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3347 endSymbolRecord(DataEnd); 3348 } else { 3349 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3350 assert(DIE->isConstant() && 3351 "Global constant variables must contain a constant expression."); 3352 3353 // Use unsigned for floats. 3354 bool isUnsigned = isFloatDIType(DIGV->getType()) 3355 ? true 3356 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3357 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3358 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3359 } 3360 } 3361