1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "llvm/ADT/APSInt.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallBitVector.h" 17 #include "llvm/ADT/SmallString.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/ADT/TinyPtrVector.h" 20 #include "llvm/ADT/Twine.h" 21 #include "llvm/BinaryFormat/COFF.h" 22 #include "llvm/BinaryFormat/Dwarf.h" 23 #include "llvm/CodeGen/AsmPrinter.h" 24 #include "llvm/CodeGen/LexicalScopes.h" 25 #include "llvm/CodeGen/MachineFrameInfo.h" 26 #include "llvm/CodeGen/MachineFunction.h" 27 #include "llvm/CodeGen/MachineInstr.h" 28 #include "llvm/CodeGen/MachineModuleInfo.h" 29 #include "llvm/CodeGen/TargetFrameLowering.h" 30 #include "llvm/CodeGen/TargetLowering.h" 31 #include "llvm/CodeGen/TargetRegisterInfo.h" 32 #include "llvm/CodeGen/TargetSubtargetInfo.h" 33 #include "llvm/Config/llvm-config.h" 34 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 35 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 36 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 37 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 38 #include "llvm/DebugInfo/CodeView/EnumTables.h" 39 #include "llvm/DebugInfo/CodeView/Line.h" 40 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 41 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 42 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 43 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 44 #include "llvm/IR/Constants.h" 45 #include "llvm/IR/DataLayout.h" 46 #include "llvm/IR/DebugInfoMetadata.h" 47 #include "llvm/IR/Function.h" 48 #include "llvm/IR/GlobalValue.h" 49 #include "llvm/IR/GlobalVariable.h" 50 #include "llvm/IR/Metadata.h" 51 #include "llvm/IR/Module.h" 52 #include "llvm/MC/MCAsmInfo.h" 53 #include "llvm/MC/MCContext.h" 54 #include "llvm/MC/MCSectionCOFF.h" 55 #include "llvm/MC/MCStreamer.h" 56 #include "llvm/MC/MCSymbol.h" 57 #include "llvm/Support/BinaryStreamWriter.h" 58 #include "llvm/Support/Casting.h" 59 #include "llvm/Support/Error.h" 60 #include "llvm/Support/ErrorHandling.h" 61 #include "llvm/Support/FormatVariadic.h" 62 #include "llvm/Support/Path.h" 63 #include "llvm/Support/Program.h" 64 #include "llvm/Support/SMLoc.h" 65 #include "llvm/Support/ScopedPrinter.h" 66 #include "llvm/Target/TargetLoweringObjectFile.h" 67 #include "llvm/Target/TargetMachine.h" 68 #include "llvm/TargetParser/Triple.h" 69 #include <algorithm> 70 #include <cassert> 71 #include <cctype> 72 #include <cstddef> 73 #include <iterator> 74 #include <limits> 75 76 using namespace llvm; 77 using namespace llvm::codeview; 78 79 namespace { 80 class CVMCAdapter : public CodeViewRecordStreamer { 81 public: 82 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 83 : OS(&OS), TypeTable(TypeTable) {} 84 85 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 86 87 void emitIntValue(uint64_t Value, unsigned Size) override { 88 OS->emitIntValueInHex(Value, Size); 89 } 90 91 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 92 93 void AddComment(const Twine &T) override { OS->AddComment(T); } 94 95 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 96 97 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 98 99 std::string getTypeName(TypeIndex TI) override { 100 std::string TypeName; 101 if (!TI.isNoneType()) { 102 if (TI.isSimple()) 103 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 104 else 105 TypeName = std::string(TypeTable.getTypeName(TI)); 106 } 107 return TypeName; 108 } 109 110 private: 111 MCStreamer *OS = nullptr; 112 TypeCollection &TypeTable; 113 }; 114 } // namespace 115 116 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 117 switch (Type) { 118 case Triple::ArchType::x86: 119 return CPUType::Pentium3; 120 case Triple::ArchType::x86_64: 121 return CPUType::X64; 122 case Triple::ArchType::thumb: 123 // LLVM currently doesn't support Windows CE and so thumb 124 // here is indiscriminately mapped to ARMNT specifically. 125 return CPUType::ARMNT; 126 case Triple::ArchType::aarch64: 127 return CPUType::ARM64; 128 default: 129 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 130 } 131 } 132 133 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 134 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 135 136 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 137 std::string &Filepath = FileToFilepathMap[File]; 138 if (!Filepath.empty()) 139 return Filepath; 140 141 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 142 143 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 144 // it textually because one of the path components could be a symlink. 145 if (Dir.starts_with("/") || Filename.starts_with("/")) { 146 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 147 return Filename; 148 Filepath = std::string(Dir); 149 if (Dir.back() != '/') 150 Filepath += '/'; 151 Filepath += Filename; 152 return Filepath; 153 } 154 155 // Clang emits directory and relative filename info into the IR, but CodeView 156 // operates on full paths. We could change Clang to emit full paths too, but 157 // that would increase the IR size and probably not needed for other users. 158 // For now, just concatenate and canonicalize the path here. 159 if (Filename.find(':') == 1) 160 Filepath = std::string(Filename); 161 else 162 Filepath = (Dir + "\\" + Filename).str(); 163 164 // Canonicalize the path. We have to do it textually because we may no longer 165 // have access the file in the filesystem. 166 // First, replace all slashes with backslashes. 167 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 168 169 // Remove all "\.\" with "\". 170 size_t Cursor = 0; 171 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 172 Filepath.erase(Cursor, 2); 173 174 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 175 // path should be well-formatted, e.g. start with a drive letter, etc. 176 Cursor = 0; 177 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 178 // Something's wrong if the path starts with "\..\", abort. 179 if (Cursor == 0) 180 break; 181 182 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 183 if (PrevSlash == std::string::npos) 184 // Something's wrong, abort. 185 break; 186 187 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 188 // The next ".." might be following the one we've just erased. 189 Cursor = PrevSlash; 190 } 191 192 // Remove all duplicate backslashes. 193 Cursor = 0; 194 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 195 Filepath.erase(Cursor, 1); 196 197 return Filepath; 198 } 199 200 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 201 StringRef FullPath = getFullFilepath(F); 202 unsigned NextId = FileIdMap.size() + 1; 203 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 204 if (Insertion.second) { 205 // We have to compute the full filepath and emit a .cv_file directive. 206 ArrayRef<uint8_t> ChecksumAsBytes; 207 FileChecksumKind CSKind = FileChecksumKind::None; 208 if (F->getChecksum()) { 209 std::string Checksum = fromHex(F->getChecksum()->Value); 210 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 211 memcpy(CKMem, Checksum.data(), Checksum.size()); 212 ChecksumAsBytes = ArrayRef<uint8_t>( 213 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 214 switch (F->getChecksum()->Kind) { 215 case DIFile::CSK_MD5: 216 CSKind = FileChecksumKind::MD5; 217 break; 218 case DIFile::CSK_SHA1: 219 CSKind = FileChecksumKind::SHA1; 220 break; 221 case DIFile::CSK_SHA256: 222 CSKind = FileChecksumKind::SHA256; 223 break; 224 } 225 } 226 bool Success = OS.emitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 227 static_cast<unsigned>(CSKind)); 228 (void)Success; 229 assert(Success && ".cv_file directive failed"); 230 } 231 return Insertion.first->second; 232 } 233 234 CodeViewDebug::InlineSite & 235 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 236 const DISubprogram *Inlinee) { 237 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 238 InlineSite *Site = &SiteInsertion.first->second; 239 if (SiteInsertion.second) { 240 unsigned ParentFuncId = CurFn->FuncId; 241 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 242 ParentFuncId = 243 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 244 .SiteFuncId; 245 246 Site->SiteFuncId = NextFuncId++; 247 OS.emitCVInlineSiteIdDirective( 248 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 249 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 250 Site->Inlinee = Inlinee; 251 InlinedSubprograms.insert(Inlinee); 252 auto InlineeIdx = getFuncIdForSubprogram(Inlinee); 253 254 if (InlinedAt->getInlinedAt() == nullptr) 255 CurFn->Inlinees.insert(InlineeIdx); 256 } 257 return *Site; 258 } 259 260 static StringRef getPrettyScopeName(const DIScope *Scope) { 261 StringRef ScopeName = Scope->getName(); 262 if (!ScopeName.empty()) 263 return ScopeName; 264 265 switch (Scope->getTag()) { 266 case dwarf::DW_TAG_enumeration_type: 267 case dwarf::DW_TAG_class_type: 268 case dwarf::DW_TAG_structure_type: 269 case dwarf::DW_TAG_union_type: 270 return "<unnamed-tag>"; 271 case dwarf::DW_TAG_namespace: 272 return "`anonymous namespace'"; 273 default: 274 return StringRef(); 275 } 276 } 277 278 const DISubprogram *CodeViewDebug::collectParentScopeNames( 279 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 280 const DISubprogram *ClosestSubprogram = nullptr; 281 while (Scope != nullptr) { 282 if (ClosestSubprogram == nullptr) 283 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 284 285 // If a type appears in a scope chain, make sure it gets emitted. The 286 // frontend will be responsible for deciding if this should be a forward 287 // declaration or a complete type. 288 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 289 DeferredCompleteTypes.push_back(Ty); 290 291 StringRef ScopeName = getPrettyScopeName(Scope); 292 if (!ScopeName.empty()) 293 QualifiedNameComponents.push_back(ScopeName); 294 Scope = Scope->getScope(); 295 } 296 return ClosestSubprogram; 297 } 298 299 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 300 StringRef TypeName) { 301 std::string FullyQualifiedName; 302 for (StringRef QualifiedNameComponent : 303 llvm::reverse(QualifiedNameComponents)) { 304 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 305 FullyQualifiedName.append("::"); 306 } 307 FullyQualifiedName.append(std::string(TypeName)); 308 return FullyQualifiedName; 309 } 310 311 struct CodeViewDebug::TypeLoweringScope { 312 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 313 ~TypeLoweringScope() { 314 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 315 // inner TypeLoweringScopes don't attempt to emit deferred types. 316 if (CVD.TypeEmissionLevel == 1) 317 CVD.emitDeferredCompleteTypes(); 318 --CVD.TypeEmissionLevel; 319 } 320 CodeViewDebug &CVD; 321 }; 322 323 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 324 StringRef Name) { 325 // Ensure types in the scope chain are emitted as soon as possible. 326 // This can create otherwise a situation where S_UDTs are emitted while 327 // looping in emitDebugInfoForUDTs. 328 TypeLoweringScope S(*this); 329 SmallVector<StringRef, 5> QualifiedNameComponents; 330 collectParentScopeNames(Scope, QualifiedNameComponents); 331 return formatNestedName(QualifiedNameComponents, Name); 332 } 333 334 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 335 const DIScope *Scope = Ty->getScope(); 336 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 337 } 338 339 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 340 // No scope means global scope and that uses the zero index. 341 // 342 // We also use zero index when the scope is a DISubprogram 343 // to suppress the emission of LF_STRING_ID for the function, 344 // which can trigger a link-time error with the linker in 345 // VS2019 version 16.11.2 or newer. 346 // Note, however, skipping the debug info emission for the DISubprogram 347 // is a temporary fix. The root issue here is that we need to figure out 348 // the proper way to encode a function nested in another function 349 // (as introduced by the Fortran 'contains' keyword) in CodeView. 350 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 351 return TypeIndex(); 352 353 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 354 355 // Check if we've already translated this scope. 356 auto I = TypeIndices.find({Scope, nullptr}); 357 if (I != TypeIndices.end()) 358 return I->second; 359 360 // Build the fully qualified name of the scope. 361 std::string ScopeName = getFullyQualifiedName(Scope); 362 StringIdRecord SID(TypeIndex(), ScopeName); 363 auto TI = TypeTable.writeLeafType(SID); 364 return recordTypeIndexForDINode(Scope, TI); 365 } 366 367 static StringRef removeTemplateArgs(StringRef Name) { 368 // Remove template args from the display name. Assume that the template args 369 // are the last thing in the name. 370 if (Name.empty() || Name.back() != '>') 371 return Name; 372 373 int OpenBrackets = 0; 374 for (int i = Name.size() - 1; i >= 0; --i) { 375 if (Name[i] == '>') 376 ++OpenBrackets; 377 else if (Name[i] == '<') { 378 --OpenBrackets; 379 if (OpenBrackets == 0) 380 return Name.substr(0, i); 381 } 382 } 383 return Name; 384 } 385 386 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 387 assert(SP); 388 389 // Check if we've already translated this subprogram. 390 auto I = TypeIndices.find({SP, nullptr}); 391 if (I != TypeIndices.end()) 392 return I->second; 393 394 // The display name includes function template arguments. Drop them to match 395 // MSVC. We need to have the template arguments in the DISubprogram name 396 // because they are used in other symbol records, such as S_GPROC32_IDs. 397 StringRef DisplayName = removeTemplateArgs(SP->getName()); 398 399 const DIScope *Scope = SP->getScope(); 400 TypeIndex TI; 401 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 402 // If the scope is a DICompositeType, then this must be a method. Member 403 // function types take some special handling, and require access to the 404 // subprogram. 405 TypeIndex ClassType = getTypeIndex(Class); 406 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 407 DisplayName); 408 TI = TypeTable.writeLeafType(MFuncId); 409 } else { 410 // Otherwise, this must be a free function. 411 TypeIndex ParentScope = getScopeIndex(Scope); 412 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 413 TI = TypeTable.writeLeafType(FuncId); 414 } 415 416 return recordTypeIndexForDINode(SP, TI); 417 } 418 419 static bool isNonTrivial(const DICompositeType *DCTy) { 420 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 421 } 422 423 static FunctionOptions 424 getFunctionOptions(const DISubroutineType *Ty, 425 const DICompositeType *ClassTy = nullptr, 426 StringRef SPName = StringRef("")) { 427 FunctionOptions FO = FunctionOptions::None; 428 const DIType *ReturnTy = nullptr; 429 if (auto TypeArray = Ty->getTypeArray()) { 430 if (TypeArray.size()) 431 ReturnTy = TypeArray[0]; 432 } 433 434 // Add CxxReturnUdt option to functions that return nontrivial record types 435 // or methods that return record types. 436 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 437 if (isNonTrivial(ReturnDCTy) || ClassTy) 438 FO |= FunctionOptions::CxxReturnUdt; 439 440 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 441 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 442 FO |= FunctionOptions::Constructor; 443 444 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 445 446 } 447 return FO; 448 } 449 450 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 451 const DICompositeType *Class) { 452 // Always use the method declaration as the key for the function type. The 453 // method declaration contains the this adjustment. 454 if (SP->getDeclaration()) 455 SP = SP->getDeclaration(); 456 assert(!SP->getDeclaration() && "should use declaration as key"); 457 458 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 459 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 460 auto I = TypeIndices.find({SP, Class}); 461 if (I != TypeIndices.end()) 462 return I->second; 463 464 // Make sure complete type info for the class is emitted *after* the member 465 // function type, as the complete class type is likely to reference this 466 // member function type. 467 TypeLoweringScope S(*this); 468 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 469 470 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 471 TypeIndex TI = lowerTypeMemberFunction( 472 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 473 return recordTypeIndexForDINode(SP, TI, Class); 474 } 475 476 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 477 TypeIndex TI, 478 const DIType *ClassTy) { 479 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 480 (void)InsertResult; 481 assert(InsertResult.second && "DINode was already assigned a type index"); 482 return TI; 483 } 484 485 unsigned CodeViewDebug::getPointerSizeInBytes() { 486 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 487 } 488 489 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 490 const LexicalScope *LS) { 491 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 492 // This variable was inlined. Associate it with the InlineSite. 493 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 494 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 495 Site.InlinedLocals.emplace_back(std::move(Var)); 496 } else { 497 // This variable goes into the corresponding lexical scope. 498 ScopeVariables[LS].emplace_back(std::move(Var)); 499 } 500 } 501 502 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 503 const DILocation *Loc) { 504 if (!llvm::is_contained(Locs, Loc)) 505 Locs.push_back(Loc); 506 } 507 508 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 509 const MachineFunction *MF) { 510 // Skip this instruction if it has the same location as the previous one. 511 if (!DL || DL == PrevInstLoc) 512 return; 513 514 const DIScope *Scope = DL->getScope(); 515 if (!Scope) 516 return; 517 518 // Skip this line if it is longer than the maximum we can record. 519 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 520 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 521 LI.isNeverStepInto()) 522 return; 523 524 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 525 if (CI.getStartColumn() != DL.getCol()) 526 return; 527 528 if (!CurFn->HaveLineInfo) 529 CurFn->HaveLineInfo = true; 530 unsigned FileId = 0; 531 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 532 FileId = CurFn->LastFileId; 533 else 534 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 535 PrevInstLoc = DL; 536 537 unsigned FuncId = CurFn->FuncId; 538 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 539 const DILocation *Loc = DL.get(); 540 541 // If this location was actually inlined from somewhere else, give it the ID 542 // of the inline call site. 543 FuncId = 544 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 545 546 // Ensure we have links in the tree of inline call sites. 547 bool FirstLoc = true; 548 while ((SiteLoc = Loc->getInlinedAt())) { 549 InlineSite &Site = 550 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 551 if (!FirstLoc) 552 addLocIfNotPresent(Site.ChildSites, Loc); 553 FirstLoc = false; 554 Loc = SiteLoc; 555 } 556 addLocIfNotPresent(CurFn->ChildSites, Loc); 557 } 558 559 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 560 /*PrologueEnd=*/false, /*IsStmt=*/false, 561 DL->getFilename(), SMLoc()); 562 } 563 564 void CodeViewDebug::emitCodeViewMagicVersion() { 565 OS.emitValueToAlignment(Align(4)); 566 OS.AddComment("Debug section magic"); 567 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 568 } 569 570 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 571 switch (DWLang) { 572 case dwarf::DW_LANG_C: 573 case dwarf::DW_LANG_C89: 574 case dwarf::DW_LANG_C99: 575 case dwarf::DW_LANG_C11: 576 return SourceLanguage::C; 577 case dwarf::DW_LANG_C_plus_plus: 578 case dwarf::DW_LANG_C_plus_plus_03: 579 case dwarf::DW_LANG_C_plus_plus_11: 580 case dwarf::DW_LANG_C_plus_plus_14: 581 return SourceLanguage::Cpp; 582 case dwarf::DW_LANG_Fortran77: 583 case dwarf::DW_LANG_Fortran90: 584 case dwarf::DW_LANG_Fortran95: 585 case dwarf::DW_LANG_Fortran03: 586 case dwarf::DW_LANG_Fortran08: 587 return SourceLanguage::Fortran; 588 case dwarf::DW_LANG_Pascal83: 589 return SourceLanguage::Pascal; 590 case dwarf::DW_LANG_Cobol74: 591 case dwarf::DW_LANG_Cobol85: 592 return SourceLanguage::Cobol; 593 case dwarf::DW_LANG_Java: 594 return SourceLanguage::Java; 595 case dwarf::DW_LANG_D: 596 return SourceLanguage::D; 597 case dwarf::DW_LANG_Swift: 598 return SourceLanguage::Swift; 599 case dwarf::DW_LANG_Rust: 600 return SourceLanguage::Rust; 601 case dwarf::DW_LANG_ObjC: 602 return SourceLanguage::ObjC; 603 case dwarf::DW_LANG_ObjC_plus_plus: 604 return SourceLanguage::ObjCpp; 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 if (!MMI->hasDebugInfo() || 617 !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 618 Asm = nullptr; 619 return; 620 } 621 622 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 623 624 // Get the current source language. 625 const MDNode *Node = *M->debug_compile_units_begin(); 626 const auto *CU = cast<DICompileUnit>(Node); 627 628 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 629 630 collectGlobalVariableInfo(); 631 632 // Check if we should emit type record hashes. 633 ConstantInt *GH = 634 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 635 EmitDebugGlobalHashes = GH && !GH->isZero(); 636 } 637 638 void CodeViewDebug::endModule() { 639 if (!Asm || !MMI->hasDebugInfo()) 640 return; 641 642 // The COFF .debug$S section consists of several subsections, each starting 643 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 644 // of the payload followed by the payload itself. The subsections are 4-byte 645 // aligned. 646 647 // Use the generic .debug$S section, and make a subsection for all the inlined 648 // subprograms. 649 switchToDebugSectionForSymbol(nullptr); 650 651 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 652 emitObjName(); 653 emitCompilerInformation(); 654 endCVSubsection(CompilerInfo); 655 656 emitInlineeLinesSubsection(); 657 658 // Emit per-function debug information. 659 for (auto &P : FnDebugInfo) 660 if (!P.first->isDeclarationForLinker()) 661 emitDebugInfoForFunction(P.first, *P.second); 662 663 // Get types used by globals without emitting anything. 664 // This is meant to collect all static const data members so they can be 665 // emitted as globals. 666 collectDebugInfoForGlobals(); 667 668 // Emit retained types. 669 emitDebugInfoForRetainedTypes(); 670 671 // Emit global variable debug information. 672 setCurrentSubprogram(nullptr); 673 emitDebugInfoForGlobals(); 674 675 // Switch back to the generic .debug$S section after potentially processing 676 // comdat symbol sections. 677 switchToDebugSectionForSymbol(nullptr); 678 679 // Emit UDT records for any types used by global variables. 680 if (!GlobalUDTs.empty()) { 681 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 682 emitDebugInfoForUDTs(GlobalUDTs); 683 endCVSubsection(SymbolsEnd); 684 } 685 686 // This subsection holds a file index to offset in string table table. 687 OS.AddComment("File index to string table offset subsection"); 688 OS.emitCVFileChecksumsDirective(); 689 690 // This subsection holds the string table. 691 OS.AddComment("String table"); 692 OS.emitCVStringTableDirective(); 693 694 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 695 // subsection in the generic .debug$S section at the end. There is no 696 // particular reason for this ordering other than to match MSVC. 697 emitBuildInfo(); 698 699 // Emit type information and hashes last, so that any types we translate while 700 // emitting function info are included. 701 emitTypeInformation(); 702 703 if (EmitDebugGlobalHashes) 704 emitTypeGlobalHashes(); 705 706 clear(); 707 } 708 709 static void 710 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 711 unsigned MaxFixedRecordLength = 0xF00) { 712 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 713 // after a fixed length portion of the record. The fixed length portion should 714 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 715 // overall record size is less than the maximum allowed. 716 SmallString<32> NullTerminatedString( 717 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 718 NullTerminatedString.push_back('\0'); 719 OS.emitBytes(NullTerminatedString); 720 } 721 722 void CodeViewDebug::emitTypeInformation() { 723 if (TypeTable.empty()) 724 return; 725 726 // Start the .debug$T or .debug$P section with 0x4. 727 OS.switchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 728 emitCodeViewMagicVersion(); 729 730 TypeTableCollection Table(TypeTable.records()); 731 TypeVisitorCallbackPipeline Pipeline; 732 733 // To emit type record using Codeview MCStreamer adapter 734 CVMCAdapter CVMCOS(OS, Table); 735 TypeRecordMapping typeMapping(CVMCOS); 736 Pipeline.addCallbackToPipeline(typeMapping); 737 738 std::optional<TypeIndex> B = Table.getFirst(); 739 while (B) { 740 // This will fail if the record data is invalid. 741 CVType Record = Table.getType(*B); 742 743 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 744 745 if (E) { 746 logAllUnhandledErrors(std::move(E), errs(), "error: "); 747 llvm_unreachable("produced malformed type record"); 748 } 749 750 B = Table.getNext(*B); 751 } 752 } 753 754 void CodeViewDebug::emitTypeGlobalHashes() { 755 if (TypeTable.empty()) 756 return; 757 758 // Start the .debug$H section with the version and hash algorithm, currently 759 // hardcoded to version 0, SHA1. 760 OS.switchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 761 762 OS.emitValueToAlignment(Align(4)); 763 OS.AddComment("Magic"); 764 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 765 OS.AddComment("Section Version"); 766 OS.emitInt16(0); 767 OS.AddComment("Hash Algorithm"); 768 OS.emitInt16(uint16_t(GlobalTypeHashAlg::BLAKE3)); 769 770 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 771 for (const auto &GHR : TypeTable.hashes()) { 772 if (OS.isVerboseAsm()) { 773 // Emit an EOL-comment describing which TypeIndex this hash corresponds 774 // to, as well as the stringified SHA1 hash. 775 SmallString<32> Comment; 776 raw_svector_ostream CommentOS(Comment); 777 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 778 OS.AddComment(Comment); 779 ++TI; 780 } 781 assert(GHR.Hash.size() == 8); 782 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 783 GHR.Hash.size()); 784 OS.emitBinaryData(S); 785 } 786 } 787 788 void CodeViewDebug::emitObjName() { 789 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 790 791 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 792 llvm::SmallString<256> PathStore(PathRef); 793 794 if (PathRef.empty() || PathRef == "-") { 795 // Don't emit the filename if we're writing to stdout or to /dev/null. 796 PathRef = {}; 797 } else { 798 PathRef = PathStore; 799 } 800 801 OS.AddComment("Signature"); 802 OS.emitIntValue(0, 4); 803 804 OS.AddComment("Object name"); 805 emitNullTerminatedSymbolName(OS, PathRef); 806 807 endSymbolRecord(CompilerEnd); 808 } 809 810 namespace { 811 struct Version { 812 int Part[4]; 813 }; 814 } // end anonymous namespace 815 816 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 817 // the version number. 818 static Version parseVersion(StringRef Name) { 819 Version V = {{0}}; 820 int N = 0; 821 for (const char C : Name) { 822 if (isdigit(C)) { 823 V.Part[N] *= 10; 824 V.Part[N] += C - '0'; 825 V.Part[N] = 826 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 827 } else if (C == '.') { 828 ++N; 829 if (N >= 4) 830 return V; 831 } else if (N > 0) 832 return V; 833 } 834 return V; 835 } 836 837 void CodeViewDebug::emitCompilerInformation() { 838 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 839 uint32_t Flags = 0; 840 841 // The low byte of the flags indicates the source language. 842 Flags = CurrentSourceLanguage; 843 // TODO: Figure out which other flags need to be set. 844 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 845 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 846 } 847 using ArchType = llvm::Triple::ArchType; 848 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 849 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 850 Arch == ArchType::aarch64) { 851 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 852 } 853 854 OS.AddComment("Flags and language"); 855 OS.emitInt32(Flags); 856 857 OS.AddComment("CPUType"); 858 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 859 860 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 861 const MDNode *Node = *CUs->operands().begin(); 862 const auto *CU = cast<DICompileUnit>(Node); 863 864 StringRef CompilerVersion = CU->getProducer(); 865 Version FrontVer = parseVersion(CompilerVersion); 866 OS.AddComment("Frontend version"); 867 for (int N : FrontVer.Part) { 868 OS.emitInt16(N); 869 } 870 871 // Some Microsoft tools, like Binscope, expect a backend version number of at 872 // least 8.something, so we'll coerce the LLVM version into a form that 873 // guarantees it'll be big enough without really lying about the version. 874 int Major = 1000 * LLVM_VERSION_MAJOR + 875 10 * LLVM_VERSION_MINOR + 876 LLVM_VERSION_PATCH; 877 // Clamp it for builds that use unusually large version numbers. 878 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 879 Version BackVer = {{ Major, 0, 0, 0 }}; 880 OS.AddComment("Backend version"); 881 for (int N : BackVer.Part) 882 OS.emitInt16(N); 883 884 OS.AddComment("Null-terminated compiler version string"); 885 emitNullTerminatedSymbolName(OS, CompilerVersion); 886 887 endSymbolRecord(CompilerEnd); 888 } 889 890 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 891 StringRef S) { 892 StringIdRecord SIR(TypeIndex(0x0), S); 893 return TypeTable.writeLeafType(SIR); 894 } 895 896 static std::string flattenCommandLine(ArrayRef<std::string> Args, 897 StringRef MainFilename) { 898 std::string FlatCmdLine; 899 raw_string_ostream OS(FlatCmdLine); 900 bool PrintedOneArg = false; 901 if (!StringRef(Args[0]).contains("-cc1")) { 902 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 903 PrintedOneArg = true; 904 } 905 for (unsigned i = 0; i < Args.size(); i++) { 906 StringRef Arg = Args[i]; 907 if (Arg.empty()) 908 continue; 909 if (Arg == "-main-file-name" || Arg == "-o") { 910 i++; // Skip this argument and next one. 911 continue; 912 } 913 if (Arg.starts_with("-object-file-name") || Arg == MainFilename) 914 continue; 915 // Skip fmessage-length for reproduciability. 916 if (Arg.starts_with("-fmessage-length")) 917 continue; 918 if (PrintedOneArg) 919 OS << " "; 920 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 921 PrintedOneArg = true; 922 } 923 OS.flush(); 924 return FlatCmdLine; 925 } 926 927 void CodeViewDebug::emitBuildInfo() { 928 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 929 // build info. The known prefix is: 930 // - Absolute path of current directory 931 // - Compiler path 932 // - Main source file path, relative to CWD or absolute 933 // - Type server PDB file 934 // - Canonical compiler command line 935 // If frontend and backend compilation are separated (think llc or LTO), it's 936 // not clear if the compiler path should refer to the executable for the 937 // frontend or the backend. Leave it blank for now. 938 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 939 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 940 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 941 const auto *CU = cast<DICompileUnit>(Node); 942 const DIFile *MainSourceFile = CU->getFile(); 943 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 944 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 945 BuildInfoArgs[BuildInfoRecord::SourceFile] = 946 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 947 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 948 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 949 getStringIdTypeIdx(TypeTable, ""); 950 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 951 BuildInfoArgs[BuildInfoRecord::BuildTool] = 952 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 953 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 954 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 955 MainSourceFile->getFilename())); 956 } 957 BuildInfoRecord BIR(BuildInfoArgs); 958 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 959 960 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 961 // from the module symbols into the type stream. 962 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 963 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 964 OS.AddComment("LF_BUILDINFO index"); 965 OS.emitInt32(BuildInfoIndex.getIndex()); 966 endSymbolRecord(BIEnd); 967 endCVSubsection(BISubsecEnd); 968 } 969 970 void CodeViewDebug::emitInlineeLinesSubsection() { 971 if (InlinedSubprograms.empty()) 972 return; 973 974 OS.AddComment("Inlinee lines subsection"); 975 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 976 977 // We emit the checksum info for files. This is used by debuggers to 978 // determine if a pdb matches the source before loading it. Visual Studio, 979 // for instance, will display a warning that the breakpoints are not valid if 980 // the pdb does not match the source. 981 OS.AddComment("Inlinee lines signature"); 982 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 983 984 for (const DISubprogram *SP : InlinedSubprograms) { 985 assert(TypeIndices.count({SP, nullptr})); 986 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 987 988 OS.addBlankLine(); 989 unsigned FileId = maybeRecordFile(SP->getFile()); 990 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 991 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 992 OS.addBlankLine(); 993 OS.AddComment("Type index of inlined function"); 994 OS.emitInt32(InlineeIdx.getIndex()); 995 OS.AddComment("Offset into filechecksum table"); 996 OS.emitCVFileChecksumOffsetDirective(FileId); 997 OS.AddComment("Starting line number"); 998 OS.emitInt32(SP->getLine()); 999 } 1000 1001 endCVSubsection(InlineEnd); 1002 } 1003 1004 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 1005 const DILocation *InlinedAt, 1006 const InlineSite &Site) { 1007 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1008 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1009 1010 // SymbolRecord 1011 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1012 1013 OS.AddComment("PtrParent"); 1014 OS.emitInt32(0); 1015 OS.AddComment("PtrEnd"); 1016 OS.emitInt32(0); 1017 OS.AddComment("Inlinee type index"); 1018 OS.emitInt32(InlineeIdx.getIndex()); 1019 1020 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1021 unsigned StartLineNum = Site.Inlinee->getLine(); 1022 1023 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1024 FI.Begin, FI.End); 1025 1026 endSymbolRecord(InlineEnd); 1027 1028 emitLocalVariableList(FI, Site.InlinedLocals); 1029 1030 // Recurse on child inlined call sites before closing the scope. 1031 for (const DILocation *ChildSite : Site.ChildSites) { 1032 auto I = FI.InlineSites.find(ChildSite); 1033 assert(I != FI.InlineSites.end() && 1034 "child site not in function inline site map"); 1035 emitInlinedCallSite(FI, ChildSite, I->second); 1036 } 1037 1038 // Close the scope. 1039 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1040 } 1041 1042 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1043 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1044 // comdat key. A section may be comdat because of -ffunction-sections or 1045 // because it is comdat in the IR. 1046 MCSectionCOFF *GVSec = 1047 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1048 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1049 1050 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1051 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1052 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1053 1054 OS.switchSection(DebugSec); 1055 1056 // Emit the magic version number if this is the first time we've switched to 1057 // this section. 1058 if (ComdatDebugSections.insert(DebugSec).second) 1059 emitCodeViewMagicVersion(); 1060 } 1061 1062 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1063 // The only supported thunk ordinal is currently the standard type. 1064 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1065 FunctionInfo &FI, 1066 const MCSymbol *Fn) { 1067 std::string FuncName = 1068 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1069 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1070 1071 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1072 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1073 1074 // Emit S_THUNK32 1075 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1076 OS.AddComment("PtrParent"); 1077 OS.emitInt32(0); 1078 OS.AddComment("PtrEnd"); 1079 OS.emitInt32(0); 1080 OS.AddComment("PtrNext"); 1081 OS.emitInt32(0); 1082 OS.AddComment("Thunk section relative address"); 1083 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1084 OS.AddComment("Thunk section index"); 1085 OS.emitCOFFSectionIndex(Fn); 1086 OS.AddComment("Code size"); 1087 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1088 OS.AddComment("Ordinal"); 1089 OS.emitInt8(unsigned(ordinal)); 1090 OS.AddComment("Function name"); 1091 emitNullTerminatedSymbolName(OS, FuncName); 1092 // Additional fields specific to the thunk ordinal would go here. 1093 endSymbolRecord(ThunkRecordEnd); 1094 1095 // Local variables/inlined routines are purposely omitted here. The point of 1096 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1097 1098 // Emit S_PROC_ID_END 1099 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1100 1101 endCVSubsection(SymbolsEnd); 1102 } 1103 1104 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1105 FunctionInfo &FI) { 1106 // For each function there is a separate subsection which holds the PC to 1107 // file:line table. 1108 const MCSymbol *Fn = Asm->getSymbol(GV); 1109 assert(Fn); 1110 1111 // Switch to the to a comdat section, if appropriate. 1112 switchToDebugSectionForSymbol(Fn); 1113 1114 std::string FuncName; 1115 auto *SP = GV->getSubprogram(); 1116 assert(SP); 1117 setCurrentSubprogram(SP); 1118 1119 if (SP->isThunk()) { 1120 emitDebugInfoForThunk(GV, FI, Fn); 1121 return; 1122 } 1123 1124 // If we have a display name, build the fully qualified name by walking the 1125 // chain of scopes. 1126 if (!SP->getName().empty()) 1127 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1128 1129 // If our DISubprogram name is empty, use the mangled name. 1130 if (FuncName.empty()) 1131 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1132 1133 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1134 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1135 OS.emitCVFPOData(Fn); 1136 1137 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1138 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1139 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1140 { 1141 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1142 : SymbolKind::S_GPROC32_ID; 1143 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1144 1145 // These fields are filled in by tools like CVPACK which run after the fact. 1146 OS.AddComment("PtrParent"); 1147 OS.emitInt32(0); 1148 OS.AddComment("PtrEnd"); 1149 OS.emitInt32(0); 1150 OS.AddComment("PtrNext"); 1151 OS.emitInt32(0); 1152 // This is the important bit that tells the debugger where the function 1153 // code is located and what's its size: 1154 OS.AddComment("Code size"); 1155 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1156 OS.AddComment("Offset after prologue"); 1157 OS.emitInt32(0); 1158 OS.AddComment("Offset before epilogue"); 1159 OS.emitInt32(0); 1160 OS.AddComment("Function type index"); 1161 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1162 OS.AddComment("Function section relative address"); 1163 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1164 OS.AddComment("Function section index"); 1165 OS.emitCOFFSectionIndex(Fn); 1166 OS.AddComment("Flags"); 1167 ProcSymFlags ProcFlags = ProcSymFlags::HasOptimizedDebugInfo; 1168 if (FI.HasFramePointer) 1169 ProcFlags |= ProcSymFlags::HasFP; 1170 if (GV->hasFnAttribute(Attribute::NoReturn)) 1171 ProcFlags |= ProcSymFlags::IsNoReturn; 1172 if (GV->hasFnAttribute(Attribute::NoInline)) 1173 ProcFlags |= ProcSymFlags::IsNoInline; 1174 OS.emitInt8(static_cast<uint8_t>(ProcFlags)); 1175 // Emit the function display name as a null-terminated string. 1176 OS.AddComment("Function name"); 1177 // Truncate the name so we won't overflow the record length field. 1178 emitNullTerminatedSymbolName(OS, FuncName); 1179 endSymbolRecord(ProcRecordEnd); 1180 1181 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1182 // Subtract out the CSR size since MSVC excludes that and we include it. 1183 OS.AddComment("FrameSize"); 1184 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1185 OS.AddComment("Padding"); 1186 OS.emitInt32(0); 1187 OS.AddComment("Offset of padding"); 1188 OS.emitInt32(0); 1189 OS.AddComment("Bytes of callee saved registers"); 1190 OS.emitInt32(FI.CSRSize); 1191 OS.AddComment("Exception handler offset"); 1192 OS.emitInt32(0); 1193 OS.AddComment("Exception handler section"); 1194 OS.emitInt16(0); 1195 OS.AddComment("Flags (defines frame register)"); 1196 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1197 endSymbolRecord(FrameProcEnd); 1198 1199 emitInlinees(FI.Inlinees); 1200 emitLocalVariableList(FI, FI.Locals); 1201 emitGlobalVariableList(FI.Globals); 1202 emitLexicalBlockList(FI.ChildBlocks, FI); 1203 1204 // Emit inlined call site information. Only emit functions inlined directly 1205 // into the parent function. We'll emit the other sites recursively as part 1206 // of their parent inline site. 1207 for (const DILocation *InlinedAt : FI.ChildSites) { 1208 auto I = FI.InlineSites.find(InlinedAt); 1209 assert(I != FI.InlineSites.end() && 1210 "child site not in function inline site map"); 1211 emitInlinedCallSite(FI, InlinedAt, I->second); 1212 } 1213 1214 for (auto Annot : FI.Annotations) { 1215 MCSymbol *Label = Annot.first; 1216 MDTuple *Strs = cast<MDTuple>(Annot.second); 1217 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1218 OS.emitCOFFSecRel32(Label, /*Offset=*/0); 1219 // FIXME: Make sure we don't overflow the max record size. 1220 OS.emitCOFFSectionIndex(Label); 1221 OS.emitInt16(Strs->getNumOperands()); 1222 for (Metadata *MD : Strs->operands()) { 1223 // MDStrings are null terminated, so we can do EmitBytes and get the 1224 // nice .asciz directive. 1225 StringRef Str = cast<MDString>(MD)->getString(); 1226 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1227 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1228 } 1229 endSymbolRecord(AnnotEnd); 1230 } 1231 1232 for (auto HeapAllocSite : FI.HeapAllocSites) { 1233 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1234 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1235 const DIType *DITy = std::get<2>(HeapAllocSite); 1236 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1237 OS.AddComment("Call site offset"); 1238 OS.emitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1239 OS.AddComment("Call site section index"); 1240 OS.emitCOFFSectionIndex(BeginLabel); 1241 OS.AddComment("Call instruction length"); 1242 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1243 OS.AddComment("Type index"); 1244 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1245 endSymbolRecord(HeapAllocEnd); 1246 } 1247 1248 if (SP != nullptr) 1249 emitDebugInfoForUDTs(LocalUDTs); 1250 1251 emitDebugInfoForJumpTables(FI); 1252 1253 // We're done with this function. 1254 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1255 } 1256 endCVSubsection(SymbolsEnd); 1257 1258 // We have an assembler directive that takes care of the whole line table. 1259 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1260 } 1261 1262 CodeViewDebug::LocalVarDef 1263 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1264 LocalVarDef DR; 1265 DR.InMemory = -1; 1266 DR.DataOffset = Offset; 1267 assert(DR.DataOffset == Offset && "truncation"); 1268 DR.IsSubfield = 0; 1269 DR.StructOffset = 0; 1270 DR.CVRegister = CVRegister; 1271 return DR; 1272 } 1273 1274 void CodeViewDebug::collectVariableInfoFromMFTable( 1275 DenseSet<InlinedEntity> &Processed) { 1276 const MachineFunction &MF = *Asm->MF; 1277 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1278 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1279 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1280 1281 for (const MachineFunction::VariableDbgInfo &VI : 1282 MF.getInStackSlotVariableDbgInfo()) { 1283 if (!VI.Var) 1284 continue; 1285 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1286 "Expected inlined-at fields to agree"); 1287 1288 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1289 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1290 1291 // If variable scope is not found then skip this variable. 1292 if (!Scope) 1293 continue; 1294 1295 // If the variable has an attached offset expression, extract it. 1296 // FIXME: Try to handle DW_OP_deref as well. 1297 int64_t ExprOffset = 0; 1298 bool Deref = false; 1299 if (VI.Expr) { 1300 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1301 if (VI.Expr->getNumElements() == 1 && 1302 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1303 Deref = true; 1304 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1305 continue; 1306 } 1307 1308 // Get the frame register used and the offset. 1309 Register FrameReg; 1310 StackOffset FrameOffset = 1311 TFI->getFrameIndexReference(*Asm->MF, VI.getStackSlot(), FrameReg); 1312 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1313 1314 assert(!FrameOffset.getScalable() && 1315 "Frame offsets with a scalable component are not supported"); 1316 1317 // Calculate the label ranges. 1318 LocalVarDef DefRange = 1319 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1320 1321 LocalVariable Var; 1322 Var.DIVar = VI.Var; 1323 1324 for (const InsnRange &Range : Scope->getRanges()) { 1325 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1326 const MCSymbol *End = getLabelAfterInsn(Range.second); 1327 End = End ? End : Asm->getFunctionEnd(); 1328 Var.DefRanges[DefRange].emplace_back(Begin, End); 1329 } 1330 1331 if (Deref) 1332 Var.UseReferenceType = true; 1333 1334 recordLocalVariable(std::move(Var), Scope); 1335 } 1336 } 1337 1338 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1339 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1340 } 1341 1342 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1343 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1344 } 1345 1346 void CodeViewDebug::calculateRanges( 1347 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1348 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1349 1350 // Calculate the definition ranges. 1351 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1352 const auto &Entry = *I; 1353 if (!Entry.isDbgValue()) 1354 continue; 1355 const MachineInstr *DVInst = Entry.getInstr(); 1356 assert(DVInst->isDebugValue() && "Invalid History entry"); 1357 // FIXME: Find a way to represent constant variables, since they are 1358 // relatively common. 1359 std::optional<DbgVariableLocation> Location = 1360 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1361 if (!Location) 1362 { 1363 // When we don't have a location this is usually because LLVM has 1364 // transformed it into a constant and we only have an llvm.dbg.value. We 1365 // can't represent these well in CodeView since S_LOCAL only works on 1366 // registers and memory locations. Instead, we will pretend this to be a 1367 // constant value to at least have it show up in the debugger. 1368 auto Op = DVInst->getDebugOperand(0); 1369 if (Op.isImm()) 1370 Var.ConstantValue = APSInt(APInt(64, Op.getImm()), false); 1371 continue; 1372 } 1373 1374 // CodeView can only express variables in register and variables in memory 1375 // at a constant offset from a register. However, for variables passed 1376 // indirectly by pointer, it is common for that pointer to be spilled to a 1377 // stack location. For the special case of one offseted load followed by a 1378 // zero offset load (a pointer spilled to the stack), we change the type of 1379 // the local variable from a value type to a reference type. This tricks the 1380 // debugger into doing the load for us. 1381 if (Var.UseReferenceType) { 1382 // We're using a reference type. Drop the last zero offset load. 1383 if (canUseReferenceType(*Location)) 1384 Location->LoadChain.pop_back(); 1385 else 1386 continue; 1387 } else if (needsReferenceType(*Location)) { 1388 // This location can't be expressed without switching to a reference type. 1389 // Start over using that. 1390 Var.UseReferenceType = true; 1391 Var.DefRanges.clear(); 1392 calculateRanges(Var, Entries); 1393 return; 1394 } 1395 1396 // We can only handle a register or an offseted load of a register. 1397 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1398 continue; 1399 1400 // Codeview can only express byte-aligned offsets, ensure that we have a 1401 // byte-boundaried location. 1402 if (Location->FragmentInfo) 1403 if (Location->FragmentInfo->OffsetInBits % 8) 1404 continue; 1405 1406 LocalVarDef DR; 1407 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1408 DR.InMemory = !Location->LoadChain.empty(); 1409 DR.DataOffset = 1410 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1411 if (Location->FragmentInfo) { 1412 DR.IsSubfield = true; 1413 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1414 } else { 1415 DR.IsSubfield = false; 1416 DR.StructOffset = 0; 1417 } 1418 1419 // Compute the label range. 1420 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1421 const MCSymbol *End; 1422 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1423 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1424 End = EndingEntry.isDbgValue() 1425 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1426 : getLabelAfterInsn(EndingEntry.getInstr()); 1427 } else 1428 End = Asm->getFunctionEnd(); 1429 1430 // If the last range end is our begin, just extend the last range. 1431 // Otherwise make a new range. 1432 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1433 Var.DefRanges[DR]; 1434 if (!R.empty() && R.back().second == Begin) 1435 R.back().second = End; 1436 else 1437 R.emplace_back(Begin, End); 1438 1439 // FIXME: Do more range combining. 1440 } 1441 } 1442 1443 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1444 DenseSet<InlinedEntity> Processed; 1445 // Grab the variable info that was squirreled away in the MMI side-table. 1446 collectVariableInfoFromMFTable(Processed); 1447 1448 for (const auto &I : DbgValues) { 1449 InlinedEntity IV = I.first; 1450 if (Processed.count(IV)) 1451 continue; 1452 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1453 const DILocation *InlinedAt = IV.second; 1454 1455 // Instruction ranges, specifying where IV is accessible. 1456 const auto &Entries = I.second; 1457 1458 LexicalScope *Scope = nullptr; 1459 if (InlinedAt) 1460 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1461 else 1462 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1463 // If variable scope is not found then skip this variable. 1464 if (!Scope) 1465 continue; 1466 1467 LocalVariable Var; 1468 Var.DIVar = DIVar; 1469 1470 calculateRanges(Var, Entries); 1471 recordLocalVariable(std::move(Var), Scope); 1472 } 1473 } 1474 1475 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1476 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1477 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1478 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1479 const Function &GV = MF->getFunction(); 1480 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1481 assert(Insertion.second && "function already has info"); 1482 CurFn = Insertion.first->second.get(); 1483 CurFn->FuncId = NextFuncId++; 1484 CurFn->Begin = Asm->getFunctionBegin(); 1485 1486 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1487 // callee-saved registers were used. For targets that don't use a PUSH 1488 // instruction (AArch64), this will be zero. 1489 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1490 CurFn->FrameSize = MFI.getStackSize(); 1491 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1492 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1493 1494 // For this function S_FRAMEPROC record, figure out which codeview register 1495 // will be the frame pointer. 1496 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1497 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1498 if (CurFn->FrameSize > 0) { 1499 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1500 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1501 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1502 } else { 1503 CurFn->HasFramePointer = true; 1504 // If there is an FP, parameters are always relative to it. 1505 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1506 if (CurFn->HasStackRealignment) { 1507 // If the stack needs realignment, locals are relative to SP or VFRAME. 1508 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1509 } else { 1510 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1511 // other stack adjustments. 1512 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1513 } 1514 } 1515 } 1516 1517 // Compute other frame procedure options. 1518 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1519 if (MFI.hasVarSizedObjects()) 1520 FPO |= FrameProcedureOptions::HasAlloca; 1521 if (MF->exposesReturnsTwice()) 1522 FPO |= FrameProcedureOptions::HasSetJmp; 1523 // FIXME: Set HasLongJmp if we ever track that info. 1524 if (MF->hasInlineAsm()) 1525 FPO |= FrameProcedureOptions::HasInlineAssembly; 1526 if (GV.hasPersonalityFn()) { 1527 if (isAsynchronousEHPersonality( 1528 classifyEHPersonality(GV.getPersonalityFn()))) 1529 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1530 else 1531 FPO |= FrameProcedureOptions::HasExceptionHandling; 1532 } 1533 if (GV.hasFnAttribute(Attribute::InlineHint)) 1534 FPO |= FrameProcedureOptions::MarkedInline; 1535 if (GV.hasFnAttribute(Attribute::Naked)) 1536 FPO |= FrameProcedureOptions::Naked; 1537 if (MFI.hasStackProtectorIndex()) { 1538 FPO |= FrameProcedureOptions::SecurityChecks; 1539 if (GV.hasFnAttribute(Attribute::StackProtectStrong) || 1540 GV.hasFnAttribute(Attribute::StackProtectReq)) { 1541 FPO |= FrameProcedureOptions::StrictSecurityChecks; 1542 } 1543 } else if (!GV.hasStackProtectorFnAttr()) { 1544 // __declspec(safebuffers) disables stack guards. 1545 FPO |= FrameProcedureOptions::SafeBuffers; 1546 } 1547 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1548 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1549 if (Asm->TM.getOptLevel() != CodeGenOptLevel::None && !GV.hasOptSize() && 1550 !GV.hasOptNone()) 1551 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1552 if (GV.hasProfileData()) { 1553 FPO |= FrameProcedureOptions::ValidProfileCounts; 1554 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1555 } 1556 // FIXME: Set GuardCfg when it is implemented. 1557 CurFn->FrameProcOpts = FPO; 1558 1559 OS.emitCVFuncIdDirective(CurFn->FuncId); 1560 1561 // Find the end of the function prolog. First known non-DBG_VALUE and 1562 // non-frame setup location marks the beginning of the function body. 1563 // FIXME: is there a simpler a way to do this? Can we just search 1564 // for the first instruction of the function, not the last of the prolog? 1565 DebugLoc PrologEndLoc; 1566 bool EmptyPrologue = true; 1567 for (const auto &MBB : *MF) { 1568 for (const auto &MI : MBB) { 1569 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1570 MI.getDebugLoc()) { 1571 PrologEndLoc = MI.getDebugLoc(); 1572 break; 1573 } else if (!MI.isMetaInstruction()) { 1574 EmptyPrologue = false; 1575 } 1576 } 1577 } 1578 1579 // Record beginning of function if we have a non-empty prologue. 1580 if (PrologEndLoc && !EmptyPrologue) { 1581 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1582 maybeRecordLocation(FnStartDL, MF); 1583 } 1584 1585 // Find heap alloc sites and emit labels around them. 1586 for (const auto &MBB : *MF) { 1587 for (const auto &MI : MBB) { 1588 if (MI.getHeapAllocMarker()) { 1589 requestLabelBeforeInsn(&MI); 1590 requestLabelAfterInsn(&MI); 1591 } 1592 } 1593 } 1594 1595 // Mark branches that may potentially be using jump tables with labels. 1596 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 1597 llvm::Triple::ArchType::thumb; 1598 discoverJumpTableBranches(MF, isThumb); 1599 } 1600 1601 static bool shouldEmitUdt(const DIType *T) { 1602 if (!T) 1603 return false; 1604 1605 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1606 if (T->getTag() == dwarf::DW_TAG_typedef) { 1607 if (DIScope *Scope = T->getScope()) { 1608 switch (Scope->getTag()) { 1609 case dwarf::DW_TAG_structure_type: 1610 case dwarf::DW_TAG_class_type: 1611 case dwarf::DW_TAG_union_type: 1612 return false; 1613 default: 1614 // do nothing. 1615 ; 1616 } 1617 } 1618 } 1619 1620 while (true) { 1621 if (!T || T->isForwardDecl()) 1622 return false; 1623 1624 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1625 if (!DT) 1626 return true; 1627 T = DT->getBaseType(); 1628 } 1629 return true; 1630 } 1631 1632 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1633 // Don't record empty UDTs. 1634 if (Ty->getName().empty()) 1635 return; 1636 if (!shouldEmitUdt(Ty)) 1637 return; 1638 1639 SmallVector<StringRef, 5> ParentScopeNames; 1640 const DISubprogram *ClosestSubprogram = 1641 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1642 1643 std::string FullyQualifiedName = 1644 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1645 1646 if (ClosestSubprogram == nullptr) { 1647 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1648 } else if (ClosestSubprogram == CurrentSubprogram) { 1649 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1650 } 1651 1652 // TODO: What if the ClosestSubprogram is neither null or the current 1653 // subprogram? Currently, the UDT just gets dropped on the floor. 1654 // 1655 // The current behavior is not desirable. To get maximal fidelity, we would 1656 // need to perform all type translation before beginning emission of .debug$S 1657 // and then make LocalUDTs a member of FunctionInfo 1658 } 1659 1660 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1661 // Generic dispatch for lowering an unknown type. 1662 switch (Ty->getTag()) { 1663 case dwarf::DW_TAG_array_type: 1664 return lowerTypeArray(cast<DICompositeType>(Ty)); 1665 case dwarf::DW_TAG_typedef: 1666 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1667 case dwarf::DW_TAG_base_type: 1668 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1669 case dwarf::DW_TAG_pointer_type: 1670 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1671 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1672 [[fallthrough]]; 1673 case dwarf::DW_TAG_reference_type: 1674 case dwarf::DW_TAG_rvalue_reference_type: 1675 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1676 case dwarf::DW_TAG_ptr_to_member_type: 1677 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1678 case dwarf::DW_TAG_restrict_type: 1679 case dwarf::DW_TAG_const_type: 1680 case dwarf::DW_TAG_volatile_type: 1681 // TODO: add support for DW_TAG_atomic_type here 1682 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1683 case dwarf::DW_TAG_subroutine_type: 1684 if (ClassTy) { 1685 // The member function type of a member function pointer has no 1686 // ThisAdjustment. 1687 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1688 /*ThisAdjustment=*/0, 1689 /*IsStaticMethod=*/false); 1690 } 1691 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1692 case dwarf::DW_TAG_enumeration_type: 1693 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1694 case dwarf::DW_TAG_class_type: 1695 case dwarf::DW_TAG_structure_type: 1696 return lowerTypeClass(cast<DICompositeType>(Ty)); 1697 case dwarf::DW_TAG_union_type: 1698 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1699 case dwarf::DW_TAG_string_type: 1700 return lowerTypeString(cast<DIStringType>(Ty)); 1701 case dwarf::DW_TAG_unspecified_type: 1702 if (Ty->getName() == "decltype(nullptr)") 1703 return TypeIndex::NullptrT(); 1704 return TypeIndex::None(); 1705 default: 1706 // Use the null type index. 1707 return TypeIndex(); 1708 } 1709 } 1710 1711 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1712 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1713 StringRef TypeName = Ty->getName(); 1714 1715 addToUDTs(Ty); 1716 1717 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1718 TypeName == "HRESULT") 1719 return TypeIndex(SimpleTypeKind::HResult); 1720 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1721 TypeName == "wchar_t") 1722 return TypeIndex(SimpleTypeKind::WideCharacter); 1723 1724 return UnderlyingTypeIndex; 1725 } 1726 1727 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1728 const DIType *ElementType = Ty->getBaseType(); 1729 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1730 // IndexType is size_t, which depends on the bitness of the target. 1731 TypeIndex IndexType = getPointerSizeInBytes() == 8 1732 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1733 : TypeIndex(SimpleTypeKind::UInt32Long); 1734 1735 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1736 1737 // Add subranges to array type. 1738 DINodeArray Elements = Ty->getElements(); 1739 for (int i = Elements.size() - 1; i >= 0; --i) { 1740 const DINode *Element = Elements[i]; 1741 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1742 1743 const DISubrange *Subrange = cast<DISubrange>(Element); 1744 int64_t Count = -1; 1745 1746 // If Subrange has a Count field, use it. 1747 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1748 // where lowerbound is from the LowerBound field of the Subrange, 1749 // or the language default lowerbound if that field is unspecified. 1750 if (auto *CI = dyn_cast_if_present<ConstantInt *>(Subrange->getCount())) 1751 Count = CI->getSExtValue(); 1752 else if (auto *UI = dyn_cast_if_present<ConstantInt *>( 1753 Subrange->getUpperBound())) { 1754 // Fortran uses 1 as the default lowerbound; other languages use 0. 1755 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1756 auto *LI = dyn_cast_if_present<ConstantInt *>(Subrange->getLowerBound()); 1757 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1758 Count = UI->getSExtValue() - Lowerbound + 1; 1759 } 1760 1761 // Forward declarations of arrays without a size and VLAs use a count of -1. 1762 // Emit a count of zero in these cases to match what MSVC does for arrays 1763 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1764 // should do for them even if we could distinguish them. 1765 if (Count == -1) 1766 Count = 0; 1767 1768 // Update the element size and element type index for subsequent subranges. 1769 ElementSize *= Count; 1770 1771 // If this is the outermost array, use the size from the array. It will be 1772 // more accurate if we had a VLA or an incomplete element type size. 1773 uint64_t ArraySize = 1774 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1775 1776 StringRef Name = (i == 0) ? Ty->getName() : ""; 1777 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1778 ElementTypeIndex = TypeTable.writeLeafType(AR); 1779 } 1780 1781 return ElementTypeIndex; 1782 } 1783 1784 // This function lowers a Fortran character type (DIStringType). 1785 // Note that it handles only the character*n variant (using SizeInBits 1786 // field in DIString to describe the type size) at the moment. 1787 // Other variants (leveraging the StringLength and StringLengthExp 1788 // fields in DIStringType) remain TBD. 1789 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1790 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1791 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1792 StringRef Name = Ty->getName(); 1793 // IndexType is size_t, which depends on the bitness of the target. 1794 TypeIndex IndexType = getPointerSizeInBytes() == 8 1795 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1796 : TypeIndex(SimpleTypeKind::UInt32Long); 1797 1798 // Create a type of character array of ArraySize. 1799 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1800 1801 return TypeTable.writeLeafType(AR); 1802 } 1803 1804 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1805 TypeIndex Index; 1806 dwarf::TypeKind Kind; 1807 uint32_t ByteSize; 1808 1809 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1810 ByteSize = Ty->getSizeInBits() / 8; 1811 1812 SimpleTypeKind STK = SimpleTypeKind::None; 1813 switch (Kind) { 1814 case dwarf::DW_ATE_address: 1815 // FIXME: Translate 1816 break; 1817 case dwarf::DW_ATE_boolean: 1818 switch (ByteSize) { 1819 case 1: STK = SimpleTypeKind::Boolean8; break; 1820 case 2: STK = SimpleTypeKind::Boolean16; break; 1821 case 4: STK = SimpleTypeKind::Boolean32; break; 1822 case 8: STK = SimpleTypeKind::Boolean64; break; 1823 case 16: STK = SimpleTypeKind::Boolean128; break; 1824 } 1825 break; 1826 case dwarf::DW_ATE_complex_float: 1827 // The CodeView size for a complex represents the size of 1828 // an individual component. 1829 switch (ByteSize) { 1830 case 4: STK = SimpleTypeKind::Complex16; break; 1831 case 8: STK = SimpleTypeKind::Complex32; break; 1832 case 16: STK = SimpleTypeKind::Complex64; break; 1833 case 20: STK = SimpleTypeKind::Complex80; break; 1834 case 32: STK = SimpleTypeKind::Complex128; break; 1835 } 1836 break; 1837 case dwarf::DW_ATE_float: 1838 switch (ByteSize) { 1839 case 2: STK = SimpleTypeKind::Float16; break; 1840 case 4: STK = SimpleTypeKind::Float32; break; 1841 case 6: STK = SimpleTypeKind::Float48; break; 1842 case 8: STK = SimpleTypeKind::Float64; break; 1843 case 10: STK = SimpleTypeKind::Float80; break; 1844 case 16: STK = SimpleTypeKind::Float128; break; 1845 } 1846 break; 1847 case dwarf::DW_ATE_signed: 1848 switch (ByteSize) { 1849 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1850 case 2: STK = SimpleTypeKind::Int16Short; break; 1851 case 4: STK = SimpleTypeKind::Int32; break; 1852 case 8: STK = SimpleTypeKind::Int64Quad; break; 1853 case 16: STK = SimpleTypeKind::Int128Oct; break; 1854 } 1855 break; 1856 case dwarf::DW_ATE_unsigned: 1857 switch (ByteSize) { 1858 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1859 case 2: STK = SimpleTypeKind::UInt16Short; break; 1860 case 4: STK = SimpleTypeKind::UInt32; break; 1861 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1862 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1863 } 1864 break; 1865 case dwarf::DW_ATE_UTF: 1866 switch (ByteSize) { 1867 case 1: STK = SimpleTypeKind::Character8; break; 1868 case 2: STK = SimpleTypeKind::Character16; break; 1869 case 4: STK = SimpleTypeKind::Character32; break; 1870 } 1871 break; 1872 case dwarf::DW_ATE_signed_char: 1873 if (ByteSize == 1) 1874 STK = SimpleTypeKind::SignedCharacter; 1875 break; 1876 case dwarf::DW_ATE_unsigned_char: 1877 if (ByteSize == 1) 1878 STK = SimpleTypeKind::UnsignedCharacter; 1879 break; 1880 default: 1881 break; 1882 } 1883 1884 // Apply some fixups based on the source-level type name. 1885 // Include some amount of canonicalization from an old naming scheme Clang 1886 // used to use for integer types (in an outdated effort to be compatible with 1887 // GCC's debug info/GDB's behavior, which has since been addressed). 1888 if (STK == SimpleTypeKind::Int32 && 1889 (Ty->getName() == "long int" || Ty->getName() == "long")) 1890 STK = SimpleTypeKind::Int32Long; 1891 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1892 Ty->getName() == "unsigned long")) 1893 STK = SimpleTypeKind::UInt32Long; 1894 if (STK == SimpleTypeKind::UInt16Short && 1895 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1896 STK = SimpleTypeKind::WideCharacter; 1897 if ((STK == SimpleTypeKind::SignedCharacter || 1898 STK == SimpleTypeKind::UnsignedCharacter) && 1899 Ty->getName() == "char") 1900 STK = SimpleTypeKind::NarrowCharacter; 1901 1902 return TypeIndex(STK); 1903 } 1904 1905 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1906 PointerOptions PO) { 1907 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1908 1909 // Pointers to simple types without any options can use SimpleTypeMode, rather 1910 // than having a dedicated pointer type record. 1911 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1912 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1913 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1914 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1915 ? SimpleTypeMode::NearPointer64 1916 : SimpleTypeMode::NearPointer32; 1917 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1918 } 1919 1920 PointerKind PK = 1921 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1922 PointerMode PM = PointerMode::Pointer; 1923 switch (Ty->getTag()) { 1924 default: llvm_unreachable("not a pointer tag type"); 1925 case dwarf::DW_TAG_pointer_type: 1926 PM = PointerMode::Pointer; 1927 break; 1928 case dwarf::DW_TAG_reference_type: 1929 PM = PointerMode::LValueReference; 1930 break; 1931 case dwarf::DW_TAG_rvalue_reference_type: 1932 PM = PointerMode::RValueReference; 1933 break; 1934 } 1935 1936 if (Ty->isObjectPointer()) 1937 PO |= PointerOptions::Const; 1938 1939 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1940 return TypeTable.writeLeafType(PR); 1941 } 1942 1943 static PointerToMemberRepresentation 1944 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1945 // SizeInBytes being zero generally implies that the member pointer type was 1946 // incomplete, which can happen if it is part of a function prototype. In this 1947 // case, use the unknown model instead of the general model. 1948 if (IsPMF) { 1949 switch (Flags & DINode::FlagPtrToMemberRep) { 1950 case 0: 1951 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1952 : PointerToMemberRepresentation::GeneralFunction; 1953 case DINode::FlagSingleInheritance: 1954 return PointerToMemberRepresentation::SingleInheritanceFunction; 1955 case DINode::FlagMultipleInheritance: 1956 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1957 case DINode::FlagVirtualInheritance: 1958 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1959 } 1960 } else { 1961 switch (Flags & DINode::FlagPtrToMemberRep) { 1962 case 0: 1963 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1964 : PointerToMemberRepresentation::GeneralData; 1965 case DINode::FlagSingleInheritance: 1966 return PointerToMemberRepresentation::SingleInheritanceData; 1967 case DINode::FlagMultipleInheritance: 1968 return PointerToMemberRepresentation::MultipleInheritanceData; 1969 case DINode::FlagVirtualInheritance: 1970 return PointerToMemberRepresentation::VirtualInheritanceData; 1971 } 1972 } 1973 llvm_unreachable("invalid ptr to member representation"); 1974 } 1975 1976 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1977 PointerOptions PO) { 1978 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1979 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1980 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1981 TypeIndex PointeeTI = 1982 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1983 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1984 : PointerKind::Near32; 1985 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1986 : PointerMode::PointerToDataMember; 1987 1988 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1989 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1990 MemberPointerInfo MPI( 1991 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1992 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1993 return TypeTable.writeLeafType(PR); 1994 } 1995 1996 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1997 /// have a translation, use the NearC convention. 1998 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1999 switch (DwarfCC) { 2000 case dwarf::DW_CC_normal: return CallingConvention::NearC; 2001 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 2002 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 2003 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 2004 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 2005 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 2006 } 2007 return CallingConvention::NearC; 2008 } 2009 2010 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 2011 ModifierOptions Mods = ModifierOptions::None; 2012 PointerOptions PO = PointerOptions::None; 2013 bool IsModifier = true; 2014 const DIType *BaseTy = Ty; 2015 while (IsModifier && BaseTy) { 2016 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 2017 switch (BaseTy->getTag()) { 2018 case dwarf::DW_TAG_const_type: 2019 Mods |= ModifierOptions::Const; 2020 PO |= PointerOptions::Const; 2021 break; 2022 case dwarf::DW_TAG_volatile_type: 2023 Mods |= ModifierOptions::Volatile; 2024 PO |= PointerOptions::Volatile; 2025 break; 2026 case dwarf::DW_TAG_restrict_type: 2027 // Only pointer types be marked with __restrict. There is no known flag 2028 // for __restrict in LF_MODIFIER records. 2029 PO |= PointerOptions::Restrict; 2030 break; 2031 default: 2032 IsModifier = false; 2033 break; 2034 } 2035 if (IsModifier) 2036 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 2037 } 2038 2039 // Check if the inner type will use an LF_POINTER record. If so, the 2040 // qualifiers will go in the LF_POINTER record. This comes up for types like 2041 // 'int *const' and 'int *__restrict', not the more common cases like 'const 2042 // char *'. 2043 if (BaseTy) { 2044 switch (BaseTy->getTag()) { 2045 case dwarf::DW_TAG_pointer_type: 2046 case dwarf::DW_TAG_reference_type: 2047 case dwarf::DW_TAG_rvalue_reference_type: 2048 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2049 case dwarf::DW_TAG_ptr_to_member_type: 2050 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2051 default: 2052 break; 2053 } 2054 } 2055 2056 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2057 2058 // Return the base type index if there aren't any modifiers. For example, the 2059 // metadata could contain restrict wrappers around non-pointer types. 2060 if (Mods == ModifierOptions::None) 2061 return ModifiedTI; 2062 2063 ModifierRecord MR(ModifiedTI, Mods); 2064 return TypeTable.writeLeafType(MR); 2065 } 2066 2067 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2068 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2069 for (const DIType *ArgType : Ty->getTypeArray()) 2070 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2071 2072 // MSVC uses type none for variadic argument. 2073 if (ReturnAndArgTypeIndices.size() > 1 && 2074 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2075 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2076 } 2077 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2078 ArrayRef<TypeIndex> ArgTypeIndices = std::nullopt; 2079 if (!ReturnAndArgTypeIndices.empty()) { 2080 auto ReturnAndArgTypesRef = ArrayRef(ReturnAndArgTypeIndices); 2081 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2082 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2083 } 2084 2085 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2086 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2087 2088 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2089 2090 FunctionOptions FO = getFunctionOptions(Ty); 2091 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2092 ArgListIndex); 2093 return TypeTable.writeLeafType(Procedure); 2094 } 2095 2096 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2097 const DIType *ClassTy, 2098 int ThisAdjustment, 2099 bool IsStaticMethod, 2100 FunctionOptions FO) { 2101 // Lower the containing class type. 2102 TypeIndex ClassType = getTypeIndex(ClassTy); 2103 2104 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2105 2106 unsigned Index = 0; 2107 SmallVector<TypeIndex, 8> ArgTypeIndices; 2108 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2109 if (ReturnAndArgs.size() > Index) { 2110 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2111 } 2112 2113 // If the first argument is a pointer type and this isn't a static method, 2114 // treat it as the special 'this' parameter, which is encoded separately from 2115 // the arguments. 2116 TypeIndex ThisTypeIndex; 2117 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2118 if (const DIDerivedType *PtrTy = 2119 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2120 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2121 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2122 Index++; 2123 } 2124 } 2125 } 2126 2127 while (Index < ReturnAndArgs.size()) 2128 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2129 2130 // MSVC uses type none for variadic argument. 2131 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2132 ArgTypeIndices.back() = TypeIndex::None(); 2133 2134 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2135 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2136 2137 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2138 2139 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2140 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2141 return TypeTable.writeLeafType(MFR); 2142 } 2143 2144 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2145 unsigned VSlotCount = 2146 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2147 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2148 2149 VFTableShapeRecord VFTSR(Slots); 2150 return TypeTable.writeLeafType(VFTSR); 2151 } 2152 2153 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2154 switch (Flags & DINode::FlagAccessibility) { 2155 case DINode::FlagPrivate: return MemberAccess::Private; 2156 case DINode::FlagPublic: return MemberAccess::Public; 2157 case DINode::FlagProtected: return MemberAccess::Protected; 2158 case 0: 2159 // If there was no explicit access control, provide the default for the tag. 2160 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2161 : MemberAccess::Public; 2162 } 2163 llvm_unreachable("access flags are exclusive"); 2164 } 2165 2166 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2167 if (SP->isArtificial()) 2168 return MethodOptions::CompilerGenerated; 2169 2170 // FIXME: Handle other MethodOptions. 2171 2172 return MethodOptions::None; 2173 } 2174 2175 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2176 bool Introduced) { 2177 if (SP->getFlags() & DINode::FlagStaticMember) 2178 return MethodKind::Static; 2179 2180 switch (SP->getVirtuality()) { 2181 case dwarf::DW_VIRTUALITY_none: 2182 break; 2183 case dwarf::DW_VIRTUALITY_virtual: 2184 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2185 case dwarf::DW_VIRTUALITY_pure_virtual: 2186 return Introduced ? MethodKind::PureIntroducingVirtual 2187 : MethodKind::PureVirtual; 2188 default: 2189 llvm_unreachable("unhandled virtuality case"); 2190 } 2191 2192 return MethodKind::Vanilla; 2193 } 2194 2195 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2196 switch (Ty->getTag()) { 2197 case dwarf::DW_TAG_class_type: 2198 return TypeRecordKind::Class; 2199 case dwarf::DW_TAG_structure_type: 2200 return TypeRecordKind::Struct; 2201 default: 2202 llvm_unreachable("unexpected tag"); 2203 } 2204 } 2205 2206 /// Return ClassOptions that should be present on both the forward declaration 2207 /// and the defintion of a tag type. 2208 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2209 ClassOptions CO = ClassOptions::None; 2210 2211 // MSVC always sets this flag, even for local types. Clang doesn't always 2212 // appear to give every type a linkage name, which may be problematic for us. 2213 // FIXME: Investigate the consequences of not following them here. 2214 if (!Ty->getIdentifier().empty()) 2215 CO |= ClassOptions::HasUniqueName; 2216 2217 // Put the Nested flag on a type if it appears immediately inside a tag type. 2218 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2219 // here. That flag is only set on definitions, and not forward declarations. 2220 const DIScope *ImmediateScope = Ty->getScope(); 2221 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2222 CO |= ClassOptions::Nested; 2223 2224 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2225 // type only when it has an immediate function scope. Clang never puts enums 2226 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2227 // always in function, class, or file scopes. 2228 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2229 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2230 CO |= ClassOptions::Scoped; 2231 } else { 2232 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2233 Scope = Scope->getScope()) { 2234 if (isa<DISubprogram>(Scope)) { 2235 CO |= ClassOptions::Scoped; 2236 break; 2237 } 2238 } 2239 } 2240 2241 return CO; 2242 } 2243 2244 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2245 switch (Ty->getTag()) { 2246 case dwarf::DW_TAG_class_type: 2247 case dwarf::DW_TAG_structure_type: 2248 case dwarf::DW_TAG_union_type: 2249 case dwarf::DW_TAG_enumeration_type: 2250 break; 2251 default: 2252 return; 2253 } 2254 2255 if (const auto *File = Ty->getFile()) { 2256 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2257 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2258 2259 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2260 TypeTable.writeLeafType(USLR); 2261 } 2262 } 2263 2264 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2265 ClassOptions CO = getCommonClassOptions(Ty); 2266 TypeIndex FTI; 2267 unsigned EnumeratorCount = 0; 2268 2269 if (Ty->isForwardDecl()) { 2270 CO |= ClassOptions::ForwardReference; 2271 } else { 2272 ContinuationRecordBuilder ContinuationBuilder; 2273 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2274 for (const DINode *Element : Ty->getElements()) { 2275 // We assume that the frontend provides all members in source declaration 2276 // order, which is what MSVC does. 2277 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2278 // FIXME: Is it correct to always emit these as unsigned here? 2279 EnumeratorRecord ER(MemberAccess::Public, 2280 APSInt(Enumerator->getValue(), true), 2281 Enumerator->getName()); 2282 ContinuationBuilder.writeMemberType(ER); 2283 EnumeratorCount++; 2284 } 2285 } 2286 FTI = TypeTable.insertRecord(ContinuationBuilder); 2287 } 2288 2289 std::string FullName = getFullyQualifiedName(Ty); 2290 2291 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2292 getTypeIndex(Ty->getBaseType())); 2293 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2294 2295 addUDTSrcLine(Ty, EnumTI); 2296 2297 return EnumTI; 2298 } 2299 2300 //===----------------------------------------------------------------------===// 2301 // ClassInfo 2302 //===----------------------------------------------------------------------===// 2303 2304 struct llvm::ClassInfo { 2305 struct MemberInfo { 2306 const DIDerivedType *MemberTypeNode; 2307 uint64_t BaseOffset; 2308 }; 2309 // [MemberInfo] 2310 using MemberList = std::vector<MemberInfo>; 2311 2312 using MethodsList = TinyPtrVector<const DISubprogram *>; 2313 // MethodName -> MethodsList 2314 using MethodsMap = MapVector<MDString *, MethodsList>; 2315 2316 /// Base classes. 2317 std::vector<const DIDerivedType *> Inheritance; 2318 2319 /// Direct members. 2320 MemberList Members; 2321 // Direct overloaded methods gathered by name. 2322 MethodsMap Methods; 2323 2324 TypeIndex VShapeTI; 2325 2326 std::vector<const DIType *> NestedTypes; 2327 }; 2328 2329 void CodeViewDebug::clear() { 2330 assert(CurFn == nullptr); 2331 FileIdMap.clear(); 2332 FnDebugInfo.clear(); 2333 FileToFilepathMap.clear(); 2334 LocalUDTs.clear(); 2335 GlobalUDTs.clear(); 2336 TypeIndices.clear(); 2337 CompleteTypeIndices.clear(); 2338 ScopeGlobals.clear(); 2339 CVGlobalVariableOffsets.clear(); 2340 } 2341 2342 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2343 const DIDerivedType *DDTy) { 2344 if (!DDTy->getName().empty()) { 2345 Info.Members.push_back({DDTy, 0}); 2346 2347 // Collect static const data members with values. 2348 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2349 DINode::FlagStaticMember) { 2350 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2351 isa<ConstantFP>(DDTy->getConstant()))) 2352 StaticConstMembers.push_back(DDTy); 2353 } 2354 2355 return; 2356 } 2357 2358 // An unnamed member may represent a nested struct or union. Attempt to 2359 // interpret the unnamed member as a DICompositeType possibly wrapped in 2360 // qualifier types. Add all the indirect fields to the current record if that 2361 // succeeds, and drop the member if that fails. 2362 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2363 uint64_t Offset = DDTy->getOffsetInBits(); 2364 const DIType *Ty = DDTy->getBaseType(); 2365 bool FullyResolved = false; 2366 while (!FullyResolved) { 2367 switch (Ty->getTag()) { 2368 case dwarf::DW_TAG_const_type: 2369 case dwarf::DW_TAG_volatile_type: 2370 // FIXME: we should apply the qualifier types to the indirect fields 2371 // rather than dropping them. 2372 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2373 break; 2374 default: 2375 FullyResolved = true; 2376 break; 2377 } 2378 } 2379 2380 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2381 if (!DCTy) 2382 return; 2383 2384 ClassInfo NestedInfo = collectClassInfo(DCTy); 2385 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2386 Info.Members.push_back( 2387 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2388 } 2389 2390 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2391 ClassInfo Info; 2392 // Add elements to structure type. 2393 DINodeArray Elements = Ty->getElements(); 2394 for (auto *Element : Elements) { 2395 // We assume that the frontend provides all members in source declaration 2396 // order, which is what MSVC does. 2397 if (!Element) 2398 continue; 2399 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2400 Info.Methods[SP->getRawName()].push_back(SP); 2401 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2402 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2403 collectMemberInfo(Info, DDTy); 2404 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2405 Info.Inheritance.push_back(DDTy); 2406 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2407 DDTy->getName() == "__vtbl_ptr_type") { 2408 Info.VShapeTI = getTypeIndex(DDTy); 2409 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2410 Info.NestedTypes.push_back(DDTy); 2411 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2412 // Ignore friend members. It appears that MSVC emitted info about 2413 // friends in the past, but modern versions do not. 2414 } 2415 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2416 Info.NestedTypes.push_back(Composite); 2417 } 2418 // Skip other unrecognized kinds of elements. 2419 } 2420 return Info; 2421 } 2422 2423 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2424 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2425 // if a complete type should be emitted instead of a forward reference. 2426 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2427 !Ty->isForwardDecl(); 2428 } 2429 2430 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2431 // Emit the complete type for unnamed structs. C++ classes with methods 2432 // which have a circular reference back to the class type are expected to 2433 // be named by the front-end and should not be "unnamed". C unnamed 2434 // structs should not have circular references. 2435 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2436 // If this unnamed complete type is already in the process of being defined 2437 // then the description of the type is malformed and cannot be emitted 2438 // into CodeView correctly so report a fatal error. 2439 auto I = CompleteTypeIndices.find(Ty); 2440 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2441 report_fatal_error("cannot debug circular reference to unnamed type"); 2442 return getCompleteTypeIndex(Ty); 2443 } 2444 2445 // First, construct the forward decl. Don't look into Ty to compute the 2446 // forward decl options, since it might not be available in all TUs. 2447 TypeRecordKind Kind = getRecordKind(Ty); 2448 ClassOptions CO = 2449 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2450 std::string FullName = getFullyQualifiedName(Ty); 2451 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2452 FullName, Ty->getIdentifier()); 2453 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2454 if (!Ty->isForwardDecl()) 2455 DeferredCompleteTypes.push_back(Ty); 2456 return FwdDeclTI; 2457 } 2458 2459 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2460 // Construct the field list and complete type record. 2461 TypeRecordKind Kind = getRecordKind(Ty); 2462 ClassOptions CO = getCommonClassOptions(Ty); 2463 TypeIndex FieldTI; 2464 TypeIndex VShapeTI; 2465 unsigned FieldCount; 2466 bool ContainsNestedClass; 2467 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2468 lowerRecordFieldList(Ty); 2469 2470 if (ContainsNestedClass) 2471 CO |= ClassOptions::ContainsNestedClass; 2472 2473 // MSVC appears to set this flag by searching any destructor or method with 2474 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2475 // the members, however special member functions are not yet emitted into 2476 // debug information. For now checking a class's non-triviality seems enough. 2477 // FIXME: not true for a nested unnamed struct. 2478 if (isNonTrivial(Ty)) 2479 CO |= ClassOptions::HasConstructorOrDestructor; 2480 2481 std::string FullName = getFullyQualifiedName(Ty); 2482 2483 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2484 2485 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2486 SizeInBytes, FullName, Ty->getIdentifier()); 2487 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2488 2489 addUDTSrcLine(Ty, ClassTI); 2490 2491 addToUDTs(Ty); 2492 2493 return ClassTI; 2494 } 2495 2496 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2497 // Emit the complete type for unnamed unions. 2498 if (shouldAlwaysEmitCompleteClassType(Ty)) 2499 return getCompleteTypeIndex(Ty); 2500 2501 ClassOptions CO = 2502 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2503 std::string FullName = getFullyQualifiedName(Ty); 2504 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2505 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2506 if (!Ty->isForwardDecl()) 2507 DeferredCompleteTypes.push_back(Ty); 2508 return FwdDeclTI; 2509 } 2510 2511 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2512 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2513 TypeIndex FieldTI; 2514 unsigned FieldCount; 2515 bool ContainsNestedClass; 2516 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2517 lowerRecordFieldList(Ty); 2518 2519 if (ContainsNestedClass) 2520 CO |= ClassOptions::ContainsNestedClass; 2521 2522 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2523 std::string FullName = getFullyQualifiedName(Ty); 2524 2525 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2526 Ty->getIdentifier()); 2527 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2528 2529 addUDTSrcLine(Ty, UnionTI); 2530 2531 addToUDTs(Ty); 2532 2533 return UnionTI; 2534 } 2535 2536 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2537 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2538 // Manually count members. MSVC appears to count everything that generates a 2539 // field list record. Each individual overload in a method overload group 2540 // contributes to this count, even though the overload group is a single field 2541 // list record. 2542 unsigned MemberCount = 0; 2543 ClassInfo Info = collectClassInfo(Ty); 2544 ContinuationRecordBuilder ContinuationBuilder; 2545 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2546 2547 // Create base classes. 2548 for (const DIDerivedType *I : Info.Inheritance) { 2549 if (I->getFlags() & DINode::FlagVirtual) { 2550 // Virtual base. 2551 unsigned VBPtrOffset = I->getVBPtrOffset(); 2552 // FIXME: Despite the accessor name, the offset is really in bytes. 2553 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2554 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2555 ? TypeRecordKind::IndirectVirtualBaseClass 2556 : TypeRecordKind::VirtualBaseClass; 2557 VirtualBaseClassRecord VBCR( 2558 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2559 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2560 VBTableIndex); 2561 2562 ContinuationBuilder.writeMemberType(VBCR); 2563 MemberCount++; 2564 } else { 2565 assert(I->getOffsetInBits() % 8 == 0 && 2566 "bases must be on byte boundaries"); 2567 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2568 getTypeIndex(I->getBaseType()), 2569 I->getOffsetInBits() / 8); 2570 ContinuationBuilder.writeMemberType(BCR); 2571 MemberCount++; 2572 } 2573 } 2574 2575 // Create members. 2576 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2577 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2578 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2579 StringRef MemberName = Member->getName(); 2580 MemberAccess Access = 2581 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2582 2583 if (Member->isStaticMember()) { 2584 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2585 ContinuationBuilder.writeMemberType(SDMR); 2586 MemberCount++; 2587 continue; 2588 } 2589 2590 // Virtual function pointer member. 2591 if ((Member->getFlags() & DINode::FlagArtificial) && 2592 Member->getName().starts_with("_vptr$")) { 2593 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2594 ContinuationBuilder.writeMemberType(VFPR); 2595 MemberCount++; 2596 continue; 2597 } 2598 2599 // Data member. 2600 uint64_t MemberOffsetInBits = 2601 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2602 if (Member->isBitField()) { 2603 uint64_t StartBitOffset = MemberOffsetInBits; 2604 if (const auto *CI = 2605 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2606 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2607 } 2608 StartBitOffset -= MemberOffsetInBits; 2609 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2610 StartBitOffset); 2611 MemberBaseType = TypeTable.writeLeafType(BFR); 2612 } 2613 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2614 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2615 MemberName); 2616 ContinuationBuilder.writeMemberType(DMR); 2617 MemberCount++; 2618 } 2619 2620 // Create methods 2621 for (auto &MethodItr : Info.Methods) { 2622 StringRef Name = MethodItr.first->getString(); 2623 2624 std::vector<OneMethodRecord> Methods; 2625 for (const DISubprogram *SP : MethodItr.second) { 2626 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2627 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2628 2629 unsigned VFTableOffset = -1; 2630 if (Introduced) 2631 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2632 2633 Methods.push_back(OneMethodRecord( 2634 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2635 translateMethodKindFlags(SP, Introduced), 2636 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2637 MemberCount++; 2638 } 2639 assert(!Methods.empty() && "Empty methods map entry"); 2640 if (Methods.size() == 1) 2641 ContinuationBuilder.writeMemberType(Methods[0]); 2642 else { 2643 // FIXME: Make this use its own ContinuationBuilder so that 2644 // MethodOverloadList can be split correctly. 2645 MethodOverloadListRecord MOLR(Methods); 2646 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2647 2648 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2649 ContinuationBuilder.writeMemberType(OMR); 2650 } 2651 } 2652 2653 // Create nested classes. 2654 for (const DIType *Nested : Info.NestedTypes) { 2655 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2656 ContinuationBuilder.writeMemberType(R); 2657 MemberCount++; 2658 } 2659 2660 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2661 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2662 !Info.NestedTypes.empty()); 2663 } 2664 2665 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2666 if (!VBPType.getIndex()) { 2667 // Make a 'const int *' type. 2668 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2669 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2670 2671 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2672 : PointerKind::Near32; 2673 PointerMode PM = PointerMode::Pointer; 2674 PointerOptions PO = PointerOptions::None; 2675 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2676 VBPType = TypeTable.writeLeafType(PR); 2677 } 2678 2679 return VBPType; 2680 } 2681 2682 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2683 // The null DIType is the void type. Don't try to hash it. 2684 if (!Ty) 2685 return TypeIndex::Void(); 2686 2687 // Check if we've already translated this type. Don't try to do a 2688 // get-or-create style insertion that caches the hash lookup across the 2689 // lowerType call. It will update the TypeIndices map. 2690 auto I = TypeIndices.find({Ty, ClassTy}); 2691 if (I != TypeIndices.end()) 2692 return I->second; 2693 2694 TypeLoweringScope S(*this); 2695 TypeIndex TI = lowerType(Ty, ClassTy); 2696 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2697 } 2698 2699 codeview::TypeIndex 2700 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2701 const DISubroutineType *SubroutineTy) { 2702 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2703 "this type must be a pointer type"); 2704 2705 PointerOptions Options = PointerOptions::None; 2706 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2707 Options = PointerOptions::LValueRefThisPointer; 2708 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2709 Options = PointerOptions::RValueRefThisPointer; 2710 2711 // Check if we've already translated this type. If there is no ref qualifier 2712 // on the function then we look up this pointer type with no associated class 2713 // so that the TypeIndex for the this pointer can be shared with the type 2714 // index for other pointers to this class type. If there is a ref qualifier 2715 // then we lookup the pointer using the subroutine as the parent type. 2716 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2717 if (I != TypeIndices.end()) 2718 return I->second; 2719 2720 TypeLoweringScope S(*this); 2721 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2722 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2723 } 2724 2725 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2726 PointerRecord PR(getTypeIndex(Ty), 2727 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2728 : PointerKind::Near32, 2729 PointerMode::LValueReference, PointerOptions::None, 2730 Ty->getSizeInBits() / 8); 2731 return TypeTable.writeLeafType(PR); 2732 } 2733 2734 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2735 // The null DIType is the void type. Don't try to hash it. 2736 if (!Ty) 2737 return TypeIndex::Void(); 2738 2739 // Look through typedefs when getting the complete type index. Call 2740 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2741 // emitted only once. 2742 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2743 (void)getTypeIndex(Ty); 2744 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2745 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2746 2747 // If this is a non-record type, the complete type index is the same as the 2748 // normal type index. Just call getTypeIndex. 2749 switch (Ty->getTag()) { 2750 case dwarf::DW_TAG_class_type: 2751 case dwarf::DW_TAG_structure_type: 2752 case dwarf::DW_TAG_union_type: 2753 break; 2754 default: 2755 return getTypeIndex(Ty); 2756 } 2757 2758 const auto *CTy = cast<DICompositeType>(Ty); 2759 2760 TypeLoweringScope S(*this); 2761 2762 // Make sure the forward declaration is emitted first. It's unclear if this 2763 // is necessary, but MSVC does it, and we should follow suit until we can show 2764 // otherwise. 2765 // We only emit a forward declaration for named types. 2766 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2767 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2768 2769 // Just use the forward decl if we don't have complete type info. This 2770 // might happen if the frontend is using modules and expects the complete 2771 // definition to be emitted elsewhere. 2772 if (CTy->isForwardDecl()) 2773 return FwdDeclTI; 2774 } 2775 2776 // Check if we've already translated the complete record type. 2777 // Insert the type with a null TypeIndex to signify that the type is currently 2778 // being lowered. 2779 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2780 if (!InsertResult.second) 2781 return InsertResult.first->second; 2782 2783 TypeIndex TI; 2784 switch (CTy->getTag()) { 2785 case dwarf::DW_TAG_class_type: 2786 case dwarf::DW_TAG_structure_type: 2787 TI = lowerCompleteTypeClass(CTy); 2788 break; 2789 case dwarf::DW_TAG_union_type: 2790 TI = lowerCompleteTypeUnion(CTy); 2791 break; 2792 default: 2793 llvm_unreachable("not a record"); 2794 } 2795 2796 // Update the type index associated with this CompositeType. This cannot 2797 // use the 'InsertResult' iterator above because it is potentially 2798 // invalidated by map insertions which can occur while lowering the class 2799 // type above. 2800 CompleteTypeIndices[CTy] = TI; 2801 return TI; 2802 } 2803 2804 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2805 /// and do this until fixpoint, as each complete record type typically 2806 /// references 2807 /// many other record types. 2808 void CodeViewDebug::emitDeferredCompleteTypes() { 2809 SmallVector<const DICompositeType *, 4> TypesToEmit; 2810 while (!DeferredCompleteTypes.empty()) { 2811 std::swap(DeferredCompleteTypes, TypesToEmit); 2812 for (const DICompositeType *RecordTy : TypesToEmit) 2813 getCompleteTypeIndex(RecordTy); 2814 TypesToEmit.clear(); 2815 } 2816 } 2817 2818 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2819 ArrayRef<LocalVariable> Locals) { 2820 // Get the sorted list of parameters and emit them first. 2821 SmallVector<const LocalVariable *, 6> Params; 2822 for (const LocalVariable &L : Locals) 2823 if (L.DIVar->isParameter()) 2824 Params.push_back(&L); 2825 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2826 return L->DIVar->getArg() < R->DIVar->getArg(); 2827 }); 2828 for (const LocalVariable *L : Params) 2829 emitLocalVariable(FI, *L); 2830 2831 // Next emit all non-parameters in the order that we found them. 2832 for (const LocalVariable &L : Locals) { 2833 if (!L.DIVar->isParameter()) { 2834 if (L.ConstantValue) { 2835 // If ConstantValue is set we will emit it as a S_CONSTANT instead of a 2836 // S_LOCAL in order to be able to represent it at all. 2837 const DIType *Ty = L.DIVar->getType(); 2838 APSInt Val(*L.ConstantValue); 2839 emitConstantSymbolRecord(Ty, Val, std::string(L.DIVar->getName())); 2840 } else { 2841 emitLocalVariable(FI, L); 2842 } 2843 } 2844 } 2845 } 2846 2847 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2848 const LocalVariable &Var) { 2849 // LocalSym record, see SymbolRecord.h for more info. 2850 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2851 2852 LocalSymFlags Flags = LocalSymFlags::None; 2853 if (Var.DIVar->isParameter()) 2854 Flags |= LocalSymFlags::IsParameter; 2855 if (Var.DefRanges.empty()) 2856 Flags |= LocalSymFlags::IsOptimizedOut; 2857 2858 OS.AddComment("TypeIndex"); 2859 TypeIndex TI = Var.UseReferenceType 2860 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2861 : getCompleteTypeIndex(Var.DIVar->getType()); 2862 OS.emitInt32(TI.getIndex()); 2863 OS.AddComment("Flags"); 2864 OS.emitInt16(static_cast<uint16_t>(Flags)); 2865 // Truncate the name so we won't overflow the record length field. 2866 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2867 endSymbolRecord(LocalEnd); 2868 2869 // Calculate the on disk prefix of the appropriate def range record. The 2870 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2871 // should be big enough to hold all forms without memory allocation. 2872 SmallString<20> BytePrefix; 2873 for (const auto &Pair : Var.DefRanges) { 2874 LocalVarDef DefRange = Pair.first; 2875 const auto &Ranges = Pair.second; 2876 BytePrefix.clear(); 2877 if (DefRange.InMemory) { 2878 int Offset = DefRange.DataOffset; 2879 unsigned Reg = DefRange.CVRegister; 2880 2881 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2882 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2883 // instead. In frames without stack realignment, $T0 will be the CFA. 2884 if (RegisterId(Reg) == RegisterId::ESP) { 2885 Reg = unsigned(RegisterId::VFRAME); 2886 Offset += FI.OffsetAdjustment; 2887 } 2888 2889 // If we can use the chosen frame pointer for the frame and this isn't a 2890 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2891 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2892 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2893 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2894 (bool(Flags & LocalSymFlags::IsParameter) 2895 ? (EncFP == FI.EncodedParamFramePtrReg) 2896 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2897 DefRangeFramePointerRelHeader DRHdr; 2898 DRHdr.Offset = Offset; 2899 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2900 } else { 2901 uint16_t RegRelFlags = 0; 2902 if (DefRange.IsSubfield) { 2903 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2904 (DefRange.StructOffset 2905 << DefRangeRegisterRelSym::OffsetInParentShift); 2906 } 2907 DefRangeRegisterRelHeader DRHdr; 2908 DRHdr.Register = Reg; 2909 DRHdr.Flags = RegRelFlags; 2910 DRHdr.BasePointerOffset = Offset; 2911 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2912 } 2913 } else { 2914 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2915 if (DefRange.IsSubfield) { 2916 DefRangeSubfieldRegisterHeader DRHdr; 2917 DRHdr.Register = DefRange.CVRegister; 2918 DRHdr.MayHaveNoName = 0; 2919 DRHdr.OffsetInParent = DefRange.StructOffset; 2920 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2921 } else { 2922 DefRangeRegisterHeader DRHdr; 2923 DRHdr.Register = DefRange.CVRegister; 2924 DRHdr.MayHaveNoName = 0; 2925 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2926 } 2927 } 2928 } 2929 } 2930 2931 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2932 const FunctionInfo& FI) { 2933 for (LexicalBlock *Block : Blocks) 2934 emitLexicalBlock(*Block, FI); 2935 } 2936 2937 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2938 /// lexical block scope. 2939 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2940 const FunctionInfo& FI) { 2941 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2942 OS.AddComment("PtrParent"); 2943 OS.emitInt32(0); // PtrParent 2944 OS.AddComment("PtrEnd"); 2945 OS.emitInt32(0); // PtrEnd 2946 OS.AddComment("Code size"); 2947 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2948 OS.AddComment("Function section relative address"); 2949 OS.emitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2950 OS.AddComment("Function section index"); 2951 OS.emitCOFFSectionIndex(FI.Begin); // Func Symbol 2952 OS.AddComment("Lexical block name"); 2953 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2954 endSymbolRecord(RecordEnd); 2955 2956 // Emit variables local to this lexical block. 2957 emitLocalVariableList(FI, Block.Locals); 2958 emitGlobalVariableList(Block.Globals); 2959 2960 // Emit lexical blocks contained within this block. 2961 emitLexicalBlockList(Block.Children, FI); 2962 2963 // Close the lexical block scope. 2964 emitEndSymbolRecord(SymbolKind::S_END); 2965 } 2966 2967 /// Convenience routine for collecting lexical block information for a list 2968 /// of lexical scopes. 2969 void CodeViewDebug::collectLexicalBlockInfo( 2970 SmallVectorImpl<LexicalScope *> &Scopes, 2971 SmallVectorImpl<LexicalBlock *> &Blocks, 2972 SmallVectorImpl<LocalVariable> &Locals, 2973 SmallVectorImpl<CVGlobalVariable> &Globals) { 2974 for (LexicalScope *Scope : Scopes) 2975 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2976 } 2977 2978 /// Populate the lexical blocks and local variable lists of the parent with 2979 /// information about the specified lexical scope. 2980 void CodeViewDebug::collectLexicalBlockInfo( 2981 LexicalScope &Scope, 2982 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2983 SmallVectorImpl<LocalVariable> &ParentLocals, 2984 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2985 if (Scope.isAbstractScope()) 2986 return; 2987 2988 // Gather information about the lexical scope including local variables, 2989 // global variables, and address ranges. 2990 bool IgnoreScope = false; 2991 auto LI = ScopeVariables.find(&Scope); 2992 SmallVectorImpl<LocalVariable> *Locals = 2993 LI != ScopeVariables.end() ? &LI->second : nullptr; 2994 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2995 SmallVectorImpl<CVGlobalVariable> *Globals = 2996 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2997 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2998 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2999 3000 // Ignore lexical scopes which do not contain variables. 3001 if (!Locals && !Globals) 3002 IgnoreScope = true; 3003 3004 // Ignore lexical scopes which are not lexical blocks. 3005 if (!DILB) 3006 IgnoreScope = true; 3007 3008 // Ignore scopes which have too many address ranges to represent in the 3009 // current CodeView format or do not have a valid address range. 3010 // 3011 // For lexical scopes with multiple address ranges you may be tempted to 3012 // construct a single range covering every instruction where the block is 3013 // live and everything in between. Unfortunately, Visual Studio only 3014 // displays variables from the first matching lexical block scope. If the 3015 // first lexical block contains exception handling code or cold code which 3016 // is moved to the bottom of the routine creating a single range covering 3017 // nearly the entire routine, then it will hide all other lexical blocks 3018 // and the variables they contain. 3019 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 3020 IgnoreScope = true; 3021 3022 if (IgnoreScope) { 3023 // This scope can be safely ignored and eliminating it will reduce the 3024 // size of the debug information. Be sure to collect any variable and scope 3025 // information from the this scope or any of its children and collapse them 3026 // into the parent scope. 3027 if (Locals) 3028 ParentLocals.append(Locals->begin(), Locals->end()); 3029 if (Globals) 3030 ParentGlobals.append(Globals->begin(), Globals->end()); 3031 collectLexicalBlockInfo(Scope.getChildren(), 3032 ParentBlocks, 3033 ParentLocals, 3034 ParentGlobals); 3035 return; 3036 } 3037 3038 // Create a new CodeView lexical block for this lexical scope. If we've 3039 // seen this DILexicalBlock before then the scope tree is malformed and 3040 // we can handle this gracefully by not processing it a second time. 3041 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 3042 if (!BlockInsertion.second) 3043 return; 3044 3045 // Create a lexical block containing the variables and collect the 3046 // lexical block information for the children. 3047 const InsnRange &Range = Ranges.front(); 3048 assert(Range.first && Range.second); 3049 LexicalBlock &Block = BlockInsertion.first->second; 3050 Block.Begin = getLabelBeforeInsn(Range.first); 3051 Block.End = getLabelAfterInsn(Range.second); 3052 assert(Block.Begin && "missing label for scope begin"); 3053 assert(Block.End && "missing label for scope end"); 3054 Block.Name = DILB->getName(); 3055 if (Locals) 3056 Block.Locals = std::move(*Locals); 3057 if (Globals) 3058 Block.Globals = std::move(*Globals); 3059 ParentBlocks.push_back(&Block); 3060 collectLexicalBlockInfo(Scope.getChildren(), 3061 Block.Children, 3062 Block.Locals, 3063 Block.Globals); 3064 } 3065 3066 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3067 const Function &GV = MF->getFunction(); 3068 assert(FnDebugInfo.count(&GV)); 3069 assert(CurFn == FnDebugInfo[&GV].get()); 3070 3071 collectVariableInfo(GV.getSubprogram()); 3072 3073 // Build the lexical block structure to emit for this routine. 3074 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3075 collectLexicalBlockInfo(*CFS, 3076 CurFn->ChildBlocks, 3077 CurFn->Locals, 3078 CurFn->Globals); 3079 3080 // Clear the scope and variable information from the map which will not be 3081 // valid after we have finished processing this routine. This also prepares 3082 // the map for the subsequent routine. 3083 ScopeVariables.clear(); 3084 3085 // Don't emit anything if we don't have any line tables. 3086 // Thunks are compiler-generated and probably won't have source correlation. 3087 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3088 FnDebugInfo.erase(&GV); 3089 CurFn = nullptr; 3090 return; 3091 } 3092 3093 // Find heap alloc sites and add to list. 3094 for (const auto &MBB : *MF) { 3095 for (const auto &MI : MBB) { 3096 if (MDNode *MD = MI.getHeapAllocMarker()) { 3097 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3098 getLabelAfterInsn(&MI), 3099 dyn_cast<DIType>(MD))); 3100 } 3101 } 3102 } 3103 3104 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 3105 llvm::Triple::ArchType::thumb; 3106 collectDebugInfoForJumpTables(MF, isThumb); 3107 3108 CurFn->Annotations = MF->getCodeViewAnnotations(); 3109 3110 CurFn->End = Asm->getFunctionEnd(); 3111 3112 CurFn = nullptr; 3113 } 3114 3115 // Usable locations are valid with non-zero line numbers. A line number of zero 3116 // corresponds to optimized code that doesn't have a distinct source location. 3117 // In this case, we try to use the previous or next source location depending on 3118 // the context. 3119 static bool isUsableDebugLoc(DebugLoc DL) { 3120 return DL && DL.getLine() != 0; 3121 } 3122 3123 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3124 DebugHandlerBase::beginInstruction(MI); 3125 3126 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3127 if (!Asm || !CurFn || MI->isDebugInstr() || 3128 MI->getFlag(MachineInstr::FrameSetup)) 3129 return; 3130 3131 // If the first instruction of a new MBB has no location, find the first 3132 // instruction with a location and use that. 3133 DebugLoc DL = MI->getDebugLoc(); 3134 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3135 for (const auto &NextMI : *MI->getParent()) { 3136 if (NextMI.isDebugInstr()) 3137 continue; 3138 DL = NextMI.getDebugLoc(); 3139 if (isUsableDebugLoc(DL)) 3140 break; 3141 } 3142 // FIXME: Handle the case where the BB has no valid locations. This would 3143 // probably require doing a real dataflow analysis. 3144 } 3145 PrevInstBB = MI->getParent(); 3146 3147 // If we still don't have a debug location, don't record a location. 3148 if (!isUsableDebugLoc(DL)) 3149 return; 3150 3151 maybeRecordLocation(DL, Asm->MF); 3152 } 3153 3154 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3155 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3156 *EndLabel = MMI->getContext().createTempSymbol(); 3157 OS.emitInt32(unsigned(Kind)); 3158 OS.AddComment("Subsection size"); 3159 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3160 OS.emitLabel(BeginLabel); 3161 return EndLabel; 3162 } 3163 3164 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3165 OS.emitLabel(EndLabel); 3166 // Every subsection must be aligned to a 4-byte boundary. 3167 OS.emitValueToAlignment(Align(4)); 3168 } 3169 3170 static StringRef getSymbolName(SymbolKind SymKind) { 3171 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3172 if (EE.Value == SymKind) 3173 return EE.Name; 3174 return ""; 3175 } 3176 3177 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3178 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3179 *EndLabel = MMI->getContext().createTempSymbol(); 3180 OS.AddComment("Record length"); 3181 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3182 OS.emitLabel(BeginLabel); 3183 if (OS.isVerboseAsm()) 3184 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3185 OS.emitInt16(unsigned(SymKind)); 3186 return EndLabel; 3187 } 3188 3189 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3190 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3191 // an extra copy of every symbol record in LLD. This increases object file 3192 // size by less than 1% in the clang build, and is compatible with the Visual 3193 // C++ linker. 3194 OS.emitValueToAlignment(Align(4)); 3195 OS.emitLabel(SymEnd); 3196 } 3197 3198 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3199 OS.AddComment("Record length"); 3200 OS.emitInt16(2); 3201 if (OS.isVerboseAsm()) 3202 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3203 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3204 } 3205 3206 void CodeViewDebug::emitDebugInfoForUDTs( 3207 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3208 #ifndef NDEBUG 3209 size_t OriginalSize = UDTs.size(); 3210 #endif 3211 for (const auto &UDT : UDTs) { 3212 const DIType *T = UDT.second; 3213 assert(shouldEmitUdt(T)); 3214 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3215 OS.AddComment("Type"); 3216 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3217 assert(OriginalSize == UDTs.size() && 3218 "getCompleteTypeIndex found new UDTs!"); 3219 emitNullTerminatedSymbolName(OS, UDT.first); 3220 endSymbolRecord(UDTRecordEnd); 3221 } 3222 } 3223 3224 void CodeViewDebug::collectGlobalVariableInfo() { 3225 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3226 GlobalMap; 3227 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3228 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3229 GV.getDebugInfo(GVEs); 3230 for (const auto *GVE : GVEs) 3231 GlobalMap[GVE] = &GV; 3232 } 3233 3234 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3235 for (const MDNode *Node : CUs->operands()) { 3236 const auto *CU = cast<DICompileUnit>(Node); 3237 for (const auto *GVE : CU->getGlobalVariables()) { 3238 const DIGlobalVariable *DIGV = GVE->getVariable(); 3239 const DIExpression *DIE = GVE->getExpression(); 3240 // Don't emit string literals in CodeView, as the only useful parts are 3241 // generally the filename and line number, which isn't possible to output 3242 // in CodeView. String literals should be the only unnamed GlobalVariable 3243 // with debug info. 3244 if (DIGV->getName().empty()) continue; 3245 3246 if ((DIE->getNumElements() == 2) && 3247 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3248 // Record the constant offset for the variable. 3249 // 3250 // A Fortran common block uses this idiom to encode the offset 3251 // of a variable from the common block's starting address. 3252 CVGlobalVariableOffsets.insert( 3253 std::make_pair(DIGV, DIE->getElement(1))); 3254 3255 // Emit constant global variables in a global symbol section. 3256 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3257 CVGlobalVariable CVGV = {DIGV, DIE}; 3258 GlobalVariables.emplace_back(std::move(CVGV)); 3259 } 3260 3261 const auto *GV = GlobalMap.lookup(GVE); 3262 if (!GV || GV->isDeclarationForLinker()) 3263 continue; 3264 3265 DIScope *Scope = DIGV->getScope(); 3266 SmallVector<CVGlobalVariable, 1> *VariableList; 3267 if (Scope && isa<DILocalScope>(Scope)) { 3268 // Locate a global variable list for this scope, creating one if 3269 // necessary. 3270 auto Insertion = ScopeGlobals.insert( 3271 {Scope, std::unique_ptr<GlobalVariableList>()}); 3272 if (Insertion.second) 3273 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3274 VariableList = Insertion.first->second.get(); 3275 } else if (GV->hasComdat()) 3276 // Emit this global variable into a COMDAT section. 3277 VariableList = &ComdatVariables; 3278 else 3279 // Emit this global variable in a single global symbol section. 3280 VariableList = &GlobalVariables; 3281 CVGlobalVariable CVGV = {DIGV, GV}; 3282 VariableList->emplace_back(std::move(CVGV)); 3283 } 3284 } 3285 } 3286 3287 void CodeViewDebug::collectDebugInfoForGlobals() { 3288 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3289 const DIGlobalVariable *DIGV = CVGV.DIGV; 3290 const DIScope *Scope = DIGV->getScope(); 3291 getCompleteTypeIndex(DIGV->getType()); 3292 getFullyQualifiedName(Scope, DIGV->getName()); 3293 } 3294 3295 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3296 const DIGlobalVariable *DIGV = CVGV.DIGV; 3297 const DIScope *Scope = DIGV->getScope(); 3298 getCompleteTypeIndex(DIGV->getType()); 3299 getFullyQualifiedName(Scope, DIGV->getName()); 3300 } 3301 } 3302 3303 void CodeViewDebug::emitDebugInfoForGlobals() { 3304 // First, emit all globals that are not in a comdat in a single symbol 3305 // substream. MSVC doesn't like it if the substream is empty, so only open 3306 // it if we have at least one global to emit. 3307 switchToDebugSectionForSymbol(nullptr); 3308 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3309 OS.AddComment("Symbol subsection for globals"); 3310 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3311 emitGlobalVariableList(GlobalVariables); 3312 emitStaticConstMemberList(); 3313 endCVSubsection(EndLabel); 3314 } 3315 3316 // Second, emit each global that is in a comdat into its own .debug$S 3317 // section along with its own symbol substream. 3318 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3319 const GlobalVariable *GV = cast<const GlobalVariable *>(CVGV.GVInfo); 3320 MCSymbol *GVSym = Asm->getSymbol(GV); 3321 OS.AddComment("Symbol subsection for " + 3322 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3323 switchToDebugSectionForSymbol(GVSym); 3324 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3325 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3326 emitDebugInfoForGlobal(CVGV); 3327 endCVSubsection(EndLabel); 3328 } 3329 } 3330 3331 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3332 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3333 for (const MDNode *Node : CUs->operands()) { 3334 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3335 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3336 getTypeIndex(RT); 3337 // FIXME: Add to global/local DTU list. 3338 } 3339 } 3340 } 3341 } 3342 3343 // Emit each global variable in the specified array. 3344 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3345 for (const CVGlobalVariable &CVGV : Globals) { 3346 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3347 emitDebugInfoForGlobal(CVGV); 3348 } 3349 } 3350 3351 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3352 const std::string &QualifiedName) { 3353 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3354 OS.AddComment("Type"); 3355 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3356 3357 OS.AddComment("Value"); 3358 3359 // Encoded integers shouldn't need more than 10 bytes. 3360 uint8_t Data[10]; 3361 BinaryStreamWriter Writer(Data, llvm::endianness::little); 3362 CodeViewRecordIO IO(Writer); 3363 cantFail(IO.mapEncodedInteger(Value)); 3364 StringRef SRef((char *)Data, Writer.getOffset()); 3365 OS.emitBinaryData(SRef); 3366 3367 OS.AddComment("Name"); 3368 emitNullTerminatedSymbolName(OS, QualifiedName); 3369 endSymbolRecord(SConstantEnd); 3370 } 3371 3372 void CodeViewDebug::emitStaticConstMemberList() { 3373 for (const DIDerivedType *DTy : StaticConstMembers) { 3374 const DIScope *Scope = DTy->getScope(); 3375 3376 APSInt Value; 3377 if (const ConstantInt *CI = 3378 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3379 Value = APSInt(CI->getValue(), 3380 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3381 else if (const ConstantFP *CFP = 3382 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3383 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3384 else 3385 llvm_unreachable("cannot emit a constant without a value"); 3386 3387 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3388 getFullyQualifiedName(Scope, DTy->getName())); 3389 } 3390 } 3391 3392 static bool isFloatDIType(const DIType *Ty) { 3393 if (isa<DICompositeType>(Ty)) 3394 return false; 3395 3396 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3397 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3398 if (T == dwarf::DW_TAG_pointer_type || 3399 T == dwarf::DW_TAG_ptr_to_member_type || 3400 T == dwarf::DW_TAG_reference_type || 3401 T == dwarf::DW_TAG_rvalue_reference_type) 3402 return false; 3403 assert(DTy->getBaseType() && "Expected valid base type"); 3404 return isFloatDIType(DTy->getBaseType()); 3405 } 3406 3407 auto *BTy = cast<DIBasicType>(Ty); 3408 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3409 } 3410 3411 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3412 const DIGlobalVariable *DIGV = CVGV.DIGV; 3413 3414 const DIScope *Scope = DIGV->getScope(); 3415 // For static data members, get the scope from the declaration. 3416 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3417 DIGV->getRawStaticDataMemberDeclaration())) 3418 Scope = MemberDecl->getScope(); 3419 // For static local variables and Fortran, the scoping portion is elided 3420 // in its name so that we can reference the variable in the command line 3421 // of the VS debugger. 3422 std::string QualifiedName = 3423 (moduleIsInFortran() || (Scope && isa<DILocalScope>(Scope))) 3424 ? std::string(DIGV->getName()) 3425 : getFullyQualifiedName(Scope, DIGV->getName()); 3426 3427 if (const GlobalVariable *GV = 3428 dyn_cast_if_present<const GlobalVariable *>(CVGV.GVInfo)) { 3429 // DataSym record, see SymbolRecord.h for more info. Thread local data 3430 // happens to have the same format as global data. 3431 MCSymbol *GVSym = Asm->getSymbol(GV); 3432 SymbolKind DataSym = GV->isThreadLocal() 3433 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3434 : SymbolKind::S_GTHREAD32) 3435 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3436 : SymbolKind::S_GDATA32); 3437 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3438 OS.AddComment("Type"); 3439 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3440 OS.AddComment("DataOffset"); 3441 3442 uint64_t Offset = 0; 3443 if (CVGlobalVariableOffsets.contains(DIGV)) 3444 // Use the offset seen while collecting info on globals. 3445 Offset = CVGlobalVariableOffsets[DIGV]; 3446 OS.emitCOFFSecRel32(GVSym, Offset); 3447 3448 OS.AddComment("Segment"); 3449 OS.emitCOFFSectionIndex(GVSym); 3450 OS.AddComment("Name"); 3451 const unsigned LengthOfDataRecord = 12; 3452 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3453 endSymbolRecord(DataEnd); 3454 } else { 3455 const DIExpression *DIE = cast<const DIExpression *>(CVGV.GVInfo); 3456 assert(DIE->isConstant() && 3457 "Global constant variables must contain a constant expression."); 3458 3459 // Use unsigned for floats. 3460 bool isUnsigned = isFloatDIType(DIGV->getType()) 3461 ? true 3462 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3463 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3464 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3465 } 3466 } 3467 3468 void forEachJumpTableBranch( 3469 const MachineFunction *MF, bool isThumb, 3470 const std::function<void(const MachineJumpTableInfo &, const MachineInstr &, 3471 int64_t)> &Callback) { 3472 auto JTI = MF->getJumpTableInfo(); 3473 if (JTI && !JTI->isEmpty()) { 3474 #ifndef NDEBUG 3475 auto UsedJTs = llvm::SmallBitVector(JTI->getJumpTables().size()); 3476 #endif 3477 for (const auto &MBB : *MF) { 3478 // Search for indirect branches... 3479 const auto LastMI = MBB.getFirstTerminator(); 3480 if (LastMI != MBB.end() && LastMI->isIndirectBranch()) { 3481 if (isThumb) { 3482 // ... that directly use jump table operands. 3483 // NOTE: ARM uses pattern matching to lower its BR_JT SDNode to 3484 // machine instructions, hence inserting a JUMP_TABLE_DEBUG_INFO node 3485 // interferes with this process *but* the resulting pseudo-instruction 3486 // uses a Jump Table operand, so extract the jump table index directly 3487 // from that. 3488 for (const auto &MO : LastMI->operands()) { 3489 if (MO.isJTI()) { 3490 unsigned Index = MO.getIndex(); 3491 #ifndef NDEBUG 3492 UsedJTs.set(Index); 3493 #endif 3494 Callback(*JTI, *LastMI, Index); 3495 break; 3496 } 3497 } 3498 } else { 3499 // ... that have jump table debug info. 3500 // NOTE: The debug info is inserted as a JUMP_TABLE_DEBUG_INFO node 3501 // when lowering the BR_JT SDNode to an indirect branch. 3502 for (auto I = MBB.instr_rbegin(), E = MBB.instr_rend(); I != E; ++I) { 3503 if (I->isJumpTableDebugInfo()) { 3504 unsigned Index = I->getOperand(0).getImm(); 3505 #ifndef NDEBUG 3506 UsedJTs.set(Index); 3507 #endif 3508 Callback(*JTI, *LastMI, Index); 3509 break; 3510 } 3511 } 3512 } 3513 } 3514 } 3515 #ifndef NDEBUG 3516 assert(UsedJTs.all() && 3517 "Some of jump tables were not used in a debug info instruction"); 3518 #endif 3519 } 3520 } 3521 3522 void CodeViewDebug::discoverJumpTableBranches(const MachineFunction *MF, 3523 bool isThumb) { 3524 forEachJumpTableBranch( 3525 MF, isThumb, 3526 [this](const MachineJumpTableInfo &, const MachineInstr &BranchMI, 3527 int64_t) { requestLabelBeforeInsn(&BranchMI); }); 3528 } 3529 3530 void CodeViewDebug::collectDebugInfoForJumpTables(const MachineFunction *MF, 3531 bool isThumb) { 3532 forEachJumpTableBranch( 3533 MF, isThumb, 3534 [this, MF](const MachineJumpTableInfo &JTI, const MachineInstr &BranchMI, 3535 int64_t JumpTableIndex) { 3536 // For label-difference jump tables, find the base expression. 3537 // Otherwise the jump table uses an absolute address (so no base 3538 // is required). 3539 const MCSymbol *Base; 3540 uint64_t BaseOffset = 0; 3541 const MCSymbol *Branch = getLabelBeforeInsn(&BranchMI); 3542 JumpTableEntrySize EntrySize; 3543 switch (JTI.getEntryKind()) { 3544 case MachineJumpTableInfo::EK_Custom32: 3545 case MachineJumpTableInfo::EK_GPRel32BlockAddress: 3546 case MachineJumpTableInfo::EK_GPRel64BlockAddress: 3547 llvm_unreachable( 3548 "EK_Custom32, EK_GPRel32BlockAddress, and " 3549 "EK_GPRel64BlockAddress should never be emitted for COFF"); 3550 case MachineJumpTableInfo::EK_BlockAddress: 3551 // Each entry is an absolute address. 3552 EntrySize = JumpTableEntrySize::Pointer; 3553 Base = nullptr; 3554 break; 3555 case MachineJumpTableInfo::EK_Inline: 3556 case MachineJumpTableInfo::EK_LabelDifference32: 3557 case MachineJumpTableInfo::EK_LabelDifference64: 3558 // Ask the AsmPrinter. 3559 std::tie(Base, BaseOffset, Branch, EntrySize) = 3560 Asm->getCodeViewJumpTableInfo(JumpTableIndex, &BranchMI, Branch); 3561 break; 3562 } 3563 3564 CurFn->JumpTables.push_back( 3565 {EntrySize, Base, BaseOffset, Branch, 3566 MF->getJTISymbol(JumpTableIndex, MMI->getContext()), 3567 JTI.getJumpTables()[JumpTableIndex].MBBs.size()}); 3568 }); 3569 } 3570 3571 void CodeViewDebug::emitDebugInfoForJumpTables(const FunctionInfo &FI) { 3572 for (auto JumpTable : FI.JumpTables) { 3573 MCSymbol *JumpTableEnd = beginSymbolRecord(SymbolKind::S_ARMSWITCHTABLE); 3574 if (JumpTable.Base) { 3575 OS.AddComment("Base offset"); 3576 OS.emitCOFFSecRel32(JumpTable.Base, JumpTable.BaseOffset); 3577 OS.AddComment("Base section index"); 3578 OS.emitCOFFSectionIndex(JumpTable.Base); 3579 } else { 3580 OS.AddComment("Base offset"); 3581 OS.emitInt32(0); 3582 OS.AddComment("Base section index"); 3583 OS.emitInt16(0); 3584 } 3585 OS.AddComment("Switch type"); 3586 OS.emitInt16(static_cast<uint16_t>(JumpTable.EntrySize)); 3587 OS.AddComment("Branch offset"); 3588 OS.emitCOFFSecRel32(JumpTable.Branch, /*Offset=*/0); 3589 OS.AddComment("Table offset"); 3590 OS.emitCOFFSecRel32(JumpTable.Table, /*Offset=*/0); 3591 OS.AddComment("Branch section index"); 3592 OS.emitCOFFSectionIndex(JumpTable.Branch); 3593 OS.AddComment("Table section index"); 3594 OS.emitCOFFSectionIndex(JumpTable.Table); 3595 OS.AddComment("Entries count"); 3596 OS.emitInt32(JumpTable.TableSize); 3597 endSymbolRecord(JumpTableEnd); 3598 } 3599 } 3600 3601 void CodeViewDebug::emitInlinees( 3602 const SmallSet<codeview::TypeIndex, 1> &Inlinees) { 3603 // Divide the list of inlinees into chunks such that each chunk fits within 3604 // one record. 3605 constexpr size_t ChunkSize = 3606 (MaxRecordLength - sizeof(SymbolKind) - sizeof(uint32_t)) / 3607 sizeof(uint32_t); 3608 3609 SmallVector<TypeIndex> SortedInlinees{Inlinees.begin(), Inlinees.end()}; 3610 llvm::sort(SortedInlinees); 3611 3612 size_t CurrentIndex = 0; 3613 while (CurrentIndex < SortedInlinees.size()) { 3614 auto Symbol = beginSymbolRecord(SymbolKind::S_INLINEES); 3615 auto CurrentChunkSize = 3616 std::min(ChunkSize, SortedInlinees.size() - CurrentIndex); 3617 OS.AddComment("Count"); 3618 OS.emitInt32(CurrentChunkSize); 3619 3620 const size_t CurrentChunkEnd = CurrentIndex + CurrentChunkSize; 3621 for (; CurrentIndex < CurrentChunkEnd; ++CurrentIndex) { 3622 OS.AddComment("Inlinee"); 3623 OS.emitInt32(SortedInlinees[CurrentIndex].getIndex()); 3624 } 3625 endSymbolRecord(Symbol); 3626 } 3627 } 3628