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