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/None.h" 16 #include "llvm/ADT/Optional.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallString.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/ADT/TinyPtrVector.h" 21 #include "llvm/ADT/Triple.h" 22 #include "llvm/ADT/Twine.h" 23 #include "llvm/BinaryFormat/COFF.h" 24 #include "llvm/BinaryFormat/Dwarf.h" 25 #include "llvm/CodeGen/AsmPrinter.h" 26 #include "llvm/CodeGen/LexicalScopes.h" 27 #include "llvm/CodeGen/MachineFrameInfo.h" 28 #include "llvm/CodeGen/MachineFunction.h" 29 #include "llvm/CodeGen/MachineInstr.h" 30 #include "llvm/CodeGen/MachineModuleInfo.h" 31 #include "llvm/CodeGen/TargetFrameLowering.h" 32 #include "llvm/CodeGen/TargetRegisterInfo.h" 33 #include "llvm/CodeGen/TargetSubtargetInfo.h" 34 #include "llvm/Config/llvm-config.h" 35 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 36 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 37 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 38 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 39 #include "llvm/DebugInfo/CodeView/EnumTables.h" 40 #include "llvm/DebugInfo/CodeView/Line.h" 41 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 42 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 43 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 44 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 45 #include "llvm/IR/Constants.h" 46 #include "llvm/IR/DataLayout.h" 47 #include "llvm/IR/DebugInfoMetadata.h" 48 #include "llvm/IR/Function.h" 49 #include "llvm/IR/GlobalValue.h" 50 #include "llvm/IR/GlobalVariable.h" 51 #include "llvm/IR/Metadata.h" 52 #include "llvm/IR/Module.h" 53 #include "llvm/MC/MCAsmInfo.h" 54 #include "llvm/MC/MCContext.h" 55 #include "llvm/MC/MCSectionCOFF.h" 56 #include "llvm/MC/MCStreamer.h" 57 #include "llvm/MC/MCSymbol.h" 58 #include "llvm/Support/BinaryStreamWriter.h" 59 #include "llvm/Support/Casting.h" 60 #include "llvm/Support/Endian.h" 61 #include "llvm/Support/Error.h" 62 #include "llvm/Support/ErrorHandling.h" 63 #include "llvm/Support/FormatVariadic.h" 64 #include "llvm/Support/Path.h" 65 #include "llvm/Support/Program.h" 66 #include "llvm/Support/SMLoc.h" 67 #include "llvm/Support/ScopedPrinter.h" 68 #include "llvm/Target/TargetLoweringObjectFile.h" 69 #include "llvm/Target/TargetMachine.h" 70 #include <algorithm> 71 #include <cassert> 72 #include <cctype> 73 #include <cstddef> 74 #include <iterator> 75 #include <limits> 76 77 using namespace llvm; 78 using namespace llvm::codeview; 79 80 namespace { 81 class CVMCAdapter : public CodeViewRecordStreamer { 82 public: 83 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 84 : OS(&OS), TypeTable(TypeTable) {} 85 86 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 87 88 void emitIntValue(uint64_t Value, unsigned Size) override { 89 OS->emitIntValueInHex(Value, Size); 90 } 91 92 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 93 94 void AddComment(const Twine &T) override { OS->AddComment(T); } 95 96 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 97 98 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 99 100 std::string getTypeName(TypeIndex TI) override { 101 std::string TypeName; 102 if (!TI.isNoneType()) { 103 if (TI.isSimple()) 104 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 105 else 106 TypeName = std::string(TypeTable.getTypeName(TI)); 107 } 108 return TypeName; 109 } 110 111 private: 112 MCStreamer *OS = nullptr; 113 TypeCollection &TypeTable; 114 }; 115 } // namespace 116 117 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 118 switch (Type) { 119 case Triple::ArchType::x86: 120 return CPUType::Pentium3; 121 case Triple::ArchType::x86_64: 122 return CPUType::X64; 123 case Triple::ArchType::thumb: 124 // LLVM currently doesn't support Windows CE and so thumb 125 // here is indiscriminately mapped to ARMNT specifically. 126 return CPUType::ARMNT; 127 case Triple::ArchType::aarch64: 128 return CPUType::ARM64; 129 default: 130 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 131 } 132 } 133 134 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 135 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 136 137 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 138 std::string &Filepath = FileToFilepathMap[File]; 139 if (!Filepath.empty()) 140 return Filepath; 141 142 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 143 144 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 145 // it textually because one of the path components could be a symlink. 146 if (Dir.startswith("/") || Filename.startswith("/")) { 147 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 148 return Filename; 149 Filepath = std::string(Dir); 150 if (Dir.back() != '/') 151 Filepath += '/'; 152 Filepath += Filename; 153 return Filepath; 154 } 155 156 // Clang emits directory and relative filename info into the IR, but CodeView 157 // operates on full paths. We could change Clang to emit full paths too, but 158 // that would increase the IR size and probably not needed for other users. 159 // For now, just concatenate and canonicalize the path here. 160 if (Filename.find(':') == 1) 161 Filepath = std::string(Filename); 162 else 163 Filepath = (Dir + "\\" + Filename).str(); 164 165 // Canonicalize the path. We have to do it textually because we may no longer 166 // have access the file in the filesystem. 167 // First, replace all slashes with backslashes. 168 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 169 170 // Remove all "\.\" with "\". 171 size_t Cursor = 0; 172 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 173 Filepath.erase(Cursor, 2); 174 175 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 176 // path should be well-formatted, e.g. start with a drive letter, etc. 177 Cursor = 0; 178 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 179 // Something's wrong if the path starts with "\..\", abort. 180 if (Cursor == 0) 181 break; 182 183 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 184 if (PrevSlash == std::string::npos) 185 // Something's wrong, abort. 186 break; 187 188 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 189 // The next ".." might be following the one we've just erased. 190 Cursor = PrevSlash; 191 } 192 193 // Remove all duplicate backslashes. 194 Cursor = 0; 195 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 196 Filepath.erase(Cursor, 1); 197 198 return Filepath; 199 } 200 201 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 202 StringRef FullPath = getFullFilepath(F); 203 unsigned NextId = FileIdMap.size() + 1; 204 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 205 if (Insertion.second) { 206 // We have to compute the full filepath and emit a .cv_file directive. 207 ArrayRef<uint8_t> ChecksumAsBytes; 208 FileChecksumKind CSKind = FileChecksumKind::None; 209 if (F->getChecksum()) { 210 std::string Checksum = fromHex(F->getChecksum()->Value); 211 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 212 memcpy(CKMem, Checksum.data(), Checksum.size()); 213 ChecksumAsBytes = ArrayRef<uint8_t>( 214 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 215 switch (F->getChecksum()->Kind) { 216 case DIFile::CSK_MD5: 217 CSKind = FileChecksumKind::MD5; 218 break; 219 case DIFile::CSK_SHA1: 220 CSKind = FileChecksumKind::SHA1; 221 break; 222 case DIFile::CSK_SHA256: 223 CSKind = FileChecksumKind::SHA256; 224 break; 225 } 226 } 227 bool Success = OS.emitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 228 static_cast<unsigned>(CSKind)); 229 (void)Success; 230 assert(Success && ".cv_file directive failed"); 231 } 232 return Insertion.first->second; 233 } 234 235 CodeViewDebug::InlineSite & 236 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 237 const DISubprogram *Inlinee) { 238 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 239 InlineSite *Site = &SiteInsertion.first->second; 240 if (SiteInsertion.second) { 241 unsigned ParentFuncId = CurFn->FuncId; 242 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 243 ParentFuncId = 244 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 245 .SiteFuncId; 246 247 Site->SiteFuncId = NextFuncId++; 248 OS.emitCVInlineSiteIdDirective( 249 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 250 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 251 Site->Inlinee = Inlinee; 252 InlinedSubprograms.insert(Inlinee); 253 getFuncIdForSubprogram(Inlinee); 254 } 255 return *Site; 256 } 257 258 static StringRef getPrettyScopeName(const DIScope *Scope) { 259 StringRef ScopeName = Scope->getName(); 260 if (!ScopeName.empty()) 261 return ScopeName; 262 263 switch (Scope->getTag()) { 264 case dwarf::DW_TAG_enumeration_type: 265 case dwarf::DW_TAG_class_type: 266 case dwarf::DW_TAG_structure_type: 267 case dwarf::DW_TAG_union_type: 268 return "<unnamed-tag>"; 269 case dwarf::DW_TAG_namespace: 270 return "`anonymous namespace'"; 271 default: 272 return StringRef(); 273 } 274 } 275 276 const DISubprogram *CodeViewDebug::collectParentScopeNames( 277 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 278 const DISubprogram *ClosestSubprogram = nullptr; 279 while (Scope != nullptr) { 280 if (ClosestSubprogram == nullptr) 281 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 282 283 // If a type appears in a scope chain, make sure it gets emitted. The 284 // frontend will be responsible for deciding if this should be a forward 285 // declaration or a complete type. 286 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 287 DeferredCompleteTypes.push_back(Ty); 288 289 StringRef ScopeName = getPrettyScopeName(Scope); 290 if (!ScopeName.empty()) 291 QualifiedNameComponents.push_back(ScopeName); 292 Scope = Scope->getScope(); 293 } 294 return ClosestSubprogram; 295 } 296 297 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 298 StringRef TypeName) { 299 std::string FullyQualifiedName; 300 for (StringRef QualifiedNameComponent : 301 llvm::reverse(QualifiedNameComponents)) { 302 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 303 FullyQualifiedName.append("::"); 304 } 305 FullyQualifiedName.append(std::string(TypeName)); 306 return FullyQualifiedName; 307 } 308 309 struct CodeViewDebug::TypeLoweringScope { 310 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 311 ~TypeLoweringScope() { 312 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 313 // inner TypeLoweringScopes don't attempt to emit deferred types. 314 if (CVD.TypeEmissionLevel == 1) 315 CVD.emitDeferredCompleteTypes(); 316 --CVD.TypeEmissionLevel; 317 } 318 CodeViewDebug &CVD; 319 }; 320 321 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 322 StringRef Name) { 323 // Ensure types in the scope chain are emitted as soon as possible. 324 // This can create otherwise a situation where S_UDTs are emitted while 325 // looping in emitDebugInfoForUDTs. 326 TypeLoweringScope S(*this); 327 SmallVector<StringRef, 5> QualifiedNameComponents; 328 collectParentScopeNames(Scope, QualifiedNameComponents); 329 return formatNestedName(QualifiedNameComponents, Name); 330 } 331 332 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 333 const DIScope *Scope = Ty->getScope(); 334 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 335 } 336 337 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 338 // No scope means global scope and that uses the zero index. 339 // 340 // We also use zero index when the scope is a DISubprogram 341 // to suppress the emission of LF_STRING_ID for the function, 342 // which can trigger a link-time error with the linker in 343 // VS2019 version 16.11.2 or newer. 344 // Note, however, skipping the debug info emission for the DISubprogram 345 // is a temporary fix. The root issue here is that we need to figure out 346 // the proper way to encode a function nested in another function 347 // (as introduced by the Fortran 'contains' keyword) in CodeView. 348 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 349 return TypeIndex(); 350 351 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 352 353 // Check if we've already translated this scope. 354 auto I = TypeIndices.find({Scope, nullptr}); 355 if (I != TypeIndices.end()) 356 return I->second; 357 358 // Build the fully qualified name of the scope. 359 std::string ScopeName = getFullyQualifiedName(Scope); 360 StringIdRecord SID(TypeIndex(), ScopeName); 361 auto TI = TypeTable.writeLeafType(SID); 362 return recordTypeIndexForDINode(Scope, TI); 363 } 364 365 static StringRef removeTemplateArgs(StringRef Name) { 366 // Remove template args from the display name. Assume that the template args 367 // are the last thing in the name. 368 if (Name.empty() || Name.back() != '>') 369 return Name; 370 371 int OpenBrackets = 0; 372 for (int i = Name.size() - 1; i >= 0; --i) { 373 if (Name[i] == '>') 374 ++OpenBrackets; 375 else if (Name[i] == '<') { 376 --OpenBrackets; 377 if (OpenBrackets == 0) 378 return Name.substr(0, i); 379 } 380 } 381 return Name; 382 } 383 384 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 385 assert(SP); 386 387 // Check if we've already translated this subprogram. 388 auto I = TypeIndices.find({SP, nullptr}); 389 if (I != TypeIndices.end()) 390 return I->second; 391 392 // The display name includes function template arguments. Drop them to match 393 // MSVC. We need to have the template arguments in the DISubprogram name 394 // because they are used in other symbol records, such as S_GPROC32_IDs. 395 StringRef DisplayName = removeTemplateArgs(SP->getName()); 396 397 const DIScope *Scope = SP->getScope(); 398 TypeIndex TI; 399 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 400 // If the scope is a DICompositeType, then this must be a method. Member 401 // function types take some special handling, and require access to the 402 // subprogram. 403 TypeIndex ClassType = getTypeIndex(Class); 404 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 405 DisplayName); 406 TI = TypeTable.writeLeafType(MFuncId); 407 } else { 408 // Otherwise, this must be a free function. 409 TypeIndex ParentScope = getScopeIndex(Scope); 410 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 411 TI = TypeTable.writeLeafType(FuncId); 412 } 413 414 return recordTypeIndexForDINode(SP, TI); 415 } 416 417 static bool isNonTrivial(const DICompositeType *DCTy) { 418 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 419 } 420 421 static FunctionOptions 422 getFunctionOptions(const DISubroutineType *Ty, 423 const DICompositeType *ClassTy = nullptr, 424 StringRef SPName = StringRef("")) { 425 FunctionOptions FO = FunctionOptions::None; 426 const DIType *ReturnTy = nullptr; 427 if (auto TypeArray = Ty->getTypeArray()) { 428 if (TypeArray.size()) 429 ReturnTy = TypeArray[0]; 430 } 431 432 // Add CxxReturnUdt option to functions that return nontrivial record types 433 // or methods that return record types. 434 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 435 if (isNonTrivial(ReturnDCTy) || ClassTy) 436 FO |= FunctionOptions::CxxReturnUdt; 437 438 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 439 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 440 FO |= FunctionOptions::Constructor; 441 442 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 443 444 } 445 return FO; 446 } 447 448 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 449 const DICompositeType *Class) { 450 // Always use the method declaration as the key for the function type. The 451 // method declaration contains the this adjustment. 452 if (SP->getDeclaration()) 453 SP = SP->getDeclaration(); 454 assert(!SP->getDeclaration() && "should use declaration as key"); 455 456 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 457 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 458 auto I = TypeIndices.find({SP, Class}); 459 if (I != TypeIndices.end()) 460 return I->second; 461 462 // Make sure complete type info for the class is emitted *after* the member 463 // function type, as the complete class type is likely to reference this 464 // member function type. 465 TypeLoweringScope S(*this); 466 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 467 468 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 469 TypeIndex TI = lowerTypeMemberFunction( 470 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 471 return recordTypeIndexForDINode(SP, TI, Class); 472 } 473 474 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 475 TypeIndex TI, 476 const DIType *ClassTy) { 477 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 478 (void)InsertResult; 479 assert(InsertResult.second && "DINode was already assigned a type index"); 480 return TI; 481 } 482 483 unsigned CodeViewDebug::getPointerSizeInBytes() { 484 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 485 } 486 487 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 488 const LexicalScope *LS) { 489 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 490 // This variable was inlined. Associate it with the InlineSite. 491 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 492 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 493 Site.InlinedLocals.emplace_back(Var); 494 } else { 495 // This variable goes into the corresponding lexical scope. 496 ScopeVariables[LS].emplace_back(Var); 497 } 498 } 499 500 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 501 const DILocation *Loc) { 502 if (!llvm::is_contained(Locs, Loc)) 503 Locs.push_back(Loc); 504 } 505 506 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 507 const MachineFunction *MF) { 508 // Skip this instruction if it has the same location as the previous one. 509 if (!DL || DL == PrevInstLoc) 510 return; 511 512 const DIScope *Scope = DL->getScope(); 513 if (!Scope) 514 return; 515 516 // Skip this line if it is longer than the maximum we can record. 517 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 518 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 519 LI.isNeverStepInto()) 520 return; 521 522 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 523 if (CI.getStartColumn() != DL.getCol()) 524 return; 525 526 if (!CurFn->HaveLineInfo) 527 CurFn->HaveLineInfo = true; 528 unsigned FileId = 0; 529 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 530 FileId = CurFn->LastFileId; 531 else 532 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 533 PrevInstLoc = DL; 534 535 unsigned FuncId = CurFn->FuncId; 536 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 537 const DILocation *Loc = DL.get(); 538 539 // If this location was actually inlined from somewhere else, give it the ID 540 // of the inline call site. 541 FuncId = 542 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 543 544 // Ensure we have links in the tree of inline call sites. 545 bool FirstLoc = true; 546 while ((SiteLoc = Loc->getInlinedAt())) { 547 InlineSite &Site = 548 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 549 if (!FirstLoc) 550 addLocIfNotPresent(Site.ChildSites, Loc); 551 FirstLoc = false; 552 Loc = SiteLoc; 553 } 554 addLocIfNotPresent(CurFn->ChildSites, Loc); 555 } 556 557 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 558 /*PrologueEnd=*/false, /*IsStmt=*/false, 559 DL->getFilename(), SMLoc()); 560 } 561 562 void CodeViewDebug::emitCodeViewMagicVersion() { 563 OS.emitValueToAlignment(4); 564 OS.AddComment("Debug section magic"); 565 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 566 } 567 568 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 569 switch (DWLang) { 570 case dwarf::DW_LANG_C: 571 case dwarf::DW_LANG_C89: 572 case dwarf::DW_LANG_C99: 573 case dwarf::DW_LANG_C11: 574 case dwarf::DW_LANG_ObjC: 575 return SourceLanguage::C; 576 case dwarf::DW_LANG_C_plus_plus: 577 case dwarf::DW_LANG_C_plus_plus_03: 578 case dwarf::DW_LANG_C_plus_plus_11: 579 case dwarf::DW_LANG_C_plus_plus_14: 580 return SourceLanguage::Cpp; 581 case dwarf::DW_LANG_Fortran77: 582 case dwarf::DW_LANG_Fortran90: 583 case dwarf::DW_LANG_Fortran95: 584 case dwarf::DW_LANG_Fortran03: 585 case dwarf::DW_LANG_Fortran08: 586 return SourceLanguage::Fortran; 587 case dwarf::DW_LANG_Pascal83: 588 return SourceLanguage::Pascal; 589 case dwarf::DW_LANG_Cobol74: 590 case dwarf::DW_LANG_Cobol85: 591 return SourceLanguage::Cobol; 592 case dwarf::DW_LANG_Java: 593 return SourceLanguage::Java; 594 case dwarf::DW_LANG_D: 595 return SourceLanguage::D; 596 case dwarf::DW_LANG_Swift: 597 return SourceLanguage::Swift; 598 case dwarf::DW_LANG_Rust: 599 return SourceLanguage::Rust; 600 default: 601 // There's no CodeView representation for this language, and CV doesn't 602 // have an "unknown" option for the language field, so we'll use MASM, 603 // as it's very low level. 604 return SourceLanguage::Masm; 605 } 606 } 607 608 void CodeViewDebug::beginModule(Module *M) { 609 // If module doesn't have named metadata anchors or COFF debug section 610 // is not available, skip any debug info related stuff. 611 if (!MMI->hasDebugInfo() || 612 !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 613 Asm = nullptr; 614 return; 615 } 616 617 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 618 619 // Get the current source language. 620 const MDNode *Node = *M->debug_compile_units_begin(); 621 const auto *CU = cast<DICompileUnit>(Node); 622 623 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 624 625 collectGlobalVariableInfo(); 626 627 // Check if we should emit type record hashes. 628 ConstantInt *GH = 629 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 630 EmitDebugGlobalHashes = GH && !GH->isZero(); 631 } 632 633 void CodeViewDebug::endModule() { 634 if (!Asm || !MMI->hasDebugInfo()) 635 return; 636 637 // The COFF .debug$S section consists of several subsections, each starting 638 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 639 // of the payload followed by the payload itself. The subsections are 4-byte 640 // aligned. 641 642 // Use the generic .debug$S section, and make a subsection for all the inlined 643 // subprograms. 644 switchToDebugSectionForSymbol(nullptr); 645 646 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 647 emitObjName(); 648 emitCompilerInformation(); 649 endCVSubsection(CompilerInfo); 650 651 emitInlineeLinesSubsection(); 652 653 // Emit per-function debug information. 654 for (auto &P : FnDebugInfo) 655 if (!P.first->isDeclarationForLinker()) 656 emitDebugInfoForFunction(P.first, *P.second); 657 658 // Get types used by globals without emitting anything. 659 // This is meant to collect all static const data members so they can be 660 // emitted as globals. 661 collectDebugInfoForGlobals(); 662 663 // Emit retained types. 664 emitDebugInfoForRetainedTypes(); 665 666 // Emit global variable debug information. 667 setCurrentSubprogram(nullptr); 668 emitDebugInfoForGlobals(); 669 670 // Switch back to the generic .debug$S section after potentially processing 671 // comdat symbol sections. 672 switchToDebugSectionForSymbol(nullptr); 673 674 // Emit UDT records for any types used by global variables. 675 if (!GlobalUDTs.empty()) { 676 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 677 emitDebugInfoForUDTs(GlobalUDTs); 678 endCVSubsection(SymbolsEnd); 679 } 680 681 // This subsection holds a file index to offset in string table table. 682 OS.AddComment("File index to string table offset subsection"); 683 OS.emitCVFileChecksumsDirective(); 684 685 // This subsection holds the string table. 686 OS.AddComment("String table"); 687 OS.emitCVStringTableDirective(); 688 689 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 690 // subsection in the generic .debug$S section at the end. There is no 691 // particular reason for this ordering other than to match MSVC. 692 emitBuildInfo(); 693 694 // Emit type information and hashes last, so that any types we translate while 695 // emitting function info are included. 696 emitTypeInformation(); 697 698 if (EmitDebugGlobalHashes) 699 emitTypeGlobalHashes(); 700 701 clear(); 702 } 703 704 static void 705 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 706 unsigned MaxFixedRecordLength = 0xF00) { 707 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 708 // after a fixed length portion of the record. The fixed length portion should 709 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 710 // overall record size is less than the maximum allowed. 711 SmallString<32> NullTerminatedString( 712 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 713 NullTerminatedString.push_back('\0'); 714 OS.emitBytes(NullTerminatedString); 715 } 716 717 void CodeViewDebug::emitTypeInformation() { 718 if (TypeTable.empty()) 719 return; 720 721 // Start the .debug$T or .debug$P section with 0x4. 722 OS.switchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 723 emitCodeViewMagicVersion(); 724 725 TypeTableCollection Table(TypeTable.records()); 726 TypeVisitorCallbackPipeline Pipeline; 727 728 // To emit type record using Codeview MCStreamer adapter 729 CVMCAdapter CVMCOS(OS, Table); 730 TypeRecordMapping typeMapping(CVMCOS); 731 Pipeline.addCallbackToPipeline(typeMapping); 732 733 Optional<TypeIndex> B = Table.getFirst(); 734 while (B) { 735 // This will fail if the record data is invalid. 736 CVType Record = Table.getType(*B); 737 738 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 739 740 if (E) { 741 logAllUnhandledErrors(std::move(E), errs(), "error: "); 742 llvm_unreachable("produced malformed type record"); 743 } 744 745 B = Table.getNext(*B); 746 } 747 } 748 749 void CodeViewDebug::emitTypeGlobalHashes() { 750 if (TypeTable.empty()) 751 return; 752 753 // Start the .debug$H section with the version and hash algorithm, currently 754 // hardcoded to version 0, SHA1. 755 OS.switchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 756 757 OS.emitValueToAlignment(4); 758 OS.AddComment("Magic"); 759 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 760 OS.AddComment("Section Version"); 761 OS.emitInt16(0); 762 OS.AddComment("Hash Algorithm"); 763 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 764 765 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 766 for (const auto &GHR : TypeTable.hashes()) { 767 if (OS.isVerboseAsm()) { 768 // Emit an EOL-comment describing which TypeIndex this hash corresponds 769 // to, as well as the stringified SHA1 hash. 770 SmallString<32> Comment; 771 raw_svector_ostream CommentOS(Comment); 772 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 773 OS.AddComment(Comment); 774 ++TI; 775 } 776 assert(GHR.Hash.size() == 8); 777 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 778 GHR.Hash.size()); 779 OS.emitBinaryData(S); 780 } 781 } 782 783 void CodeViewDebug::emitObjName() { 784 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 785 786 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 787 llvm::SmallString<256> PathStore(PathRef); 788 789 if (PathRef.empty() || PathRef == "-") { 790 // Don't emit the filename if we're writing to stdout or to /dev/null. 791 PathRef = {}; 792 } else { 793 llvm::sys::path::remove_dots(PathStore, /*remove_dot_dot=*/true); 794 PathRef = PathStore; 795 } 796 797 OS.AddComment("Signature"); 798 OS.emitIntValue(0, 4); 799 800 OS.AddComment("Object name"); 801 emitNullTerminatedSymbolName(OS, PathRef); 802 803 endSymbolRecord(CompilerEnd); 804 } 805 806 namespace { 807 struct Version { 808 int Part[4]; 809 }; 810 } // end anonymous namespace 811 812 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 813 // the version number. 814 static Version parseVersion(StringRef Name) { 815 Version V = {{0}}; 816 int N = 0; 817 for (const char C : Name) { 818 if (isdigit(C)) { 819 V.Part[N] *= 10; 820 V.Part[N] += C - '0'; 821 V.Part[N] = 822 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 823 } else if (C == '.') { 824 ++N; 825 if (N >= 4) 826 return V; 827 } else if (N > 0) 828 return V; 829 } 830 return V; 831 } 832 833 void CodeViewDebug::emitCompilerInformation() { 834 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 835 uint32_t Flags = 0; 836 837 // The low byte of the flags indicates the source language. 838 Flags = CurrentSourceLanguage; 839 // TODO: Figure out which other flags need to be set. 840 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 841 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 842 } 843 using ArchType = llvm::Triple::ArchType; 844 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 845 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 846 Arch == ArchType::aarch64) { 847 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 848 } 849 850 OS.AddComment("Flags and language"); 851 OS.emitInt32(Flags); 852 853 OS.AddComment("CPUType"); 854 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 855 856 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 857 const MDNode *Node = *CUs->operands().begin(); 858 const auto *CU = cast<DICompileUnit>(Node); 859 860 StringRef CompilerVersion = CU->getProducer(); 861 Version FrontVer = parseVersion(CompilerVersion); 862 OS.AddComment("Frontend version"); 863 for (int N : FrontVer.Part) { 864 OS.emitInt16(N); 865 } 866 867 // Some Microsoft tools, like Binscope, expect a backend version number of at 868 // least 8.something, so we'll coerce the LLVM version into a form that 869 // guarantees it'll be big enough without really lying about the version. 870 int Major = 1000 * LLVM_VERSION_MAJOR + 871 10 * LLVM_VERSION_MINOR + 872 LLVM_VERSION_PATCH; 873 // Clamp it for builds that use unusually large version numbers. 874 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 875 Version BackVer = {{ Major, 0, 0, 0 }}; 876 OS.AddComment("Backend version"); 877 for (int N : BackVer.Part) 878 OS.emitInt16(N); 879 880 OS.AddComment("Null-terminated compiler version string"); 881 emitNullTerminatedSymbolName(OS, CompilerVersion); 882 883 endSymbolRecord(CompilerEnd); 884 } 885 886 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 887 StringRef S) { 888 StringIdRecord SIR(TypeIndex(0x0), S); 889 return TypeTable.writeLeafType(SIR); 890 } 891 892 static std::string flattenCommandLine(ArrayRef<std::string> Args, 893 StringRef MainFilename) { 894 std::string FlatCmdLine; 895 raw_string_ostream OS(FlatCmdLine); 896 bool PrintedOneArg = false; 897 if (!StringRef(Args[0]).contains("-cc1")) { 898 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 899 PrintedOneArg = true; 900 } 901 for (unsigned i = 0; i < Args.size(); i++) { 902 StringRef Arg = Args[i]; 903 if (Arg.empty()) 904 continue; 905 if (Arg == "-main-file-name" || Arg == "-o") { 906 i++; // Skip this argument and next one. 907 continue; 908 } 909 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 910 continue; 911 if (PrintedOneArg) 912 OS << " "; 913 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 914 PrintedOneArg = true; 915 } 916 OS.flush(); 917 return FlatCmdLine; 918 } 919 920 void CodeViewDebug::emitBuildInfo() { 921 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 922 // build info. The known prefix is: 923 // - Absolute path of current directory 924 // - Compiler path 925 // - Main source file path, relative to CWD or absolute 926 // - Type server PDB file 927 // - Canonical compiler command line 928 // If frontend and backend compilation are separated (think llc or LTO), it's 929 // not clear if the compiler path should refer to the executable for the 930 // frontend or the backend. Leave it blank for now. 931 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 932 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 933 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 934 const auto *CU = cast<DICompileUnit>(Node); 935 const DIFile *MainSourceFile = CU->getFile(); 936 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 937 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 938 BuildInfoArgs[BuildInfoRecord::SourceFile] = 939 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 940 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 941 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 942 getStringIdTypeIdx(TypeTable, ""); 943 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 944 BuildInfoArgs[BuildInfoRecord::BuildTool] = 945 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 946 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 947 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 948 MainSourceFile->getFilename())); 949 } 950 BuildInfoRecord BIR(BuildInfoArgs); 951 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 952 953 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 954 // from the module symbols into the type stream. 955 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 956 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 957 OS.AddComment("LF_BUILDINFO index"); 958 OS.emitInt32(BuildInfoIndex.getIndex()); 959 endSymbolRecord(BIEnd); 960 endCVSubsection(BISubsecEnd); 961 } 962 963 void CodeViewDebug::emitInlineeLinesSubsection() { 964 if (InlinedSubprograms.empty()) 965 return; 966 967 OS.AddComment("Inlinee lines subsection"); 968 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 969 970 // We emit the checksum info for files. This is used by debuggers to 971 // determine if a pdb matches the source before loading it. Visual Studio, 972 // for instance, will display a warning that the breakpoints are not valid if 973 // the pdb does not match the source. 974 OS.AddComment("Inlinee lines signature"); 975 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 976 977 for (const DISubprogram *SP : InlinedSubprograms) { 978 assert(TypeIndices.count({SP, nullptr})); 979 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 980 981 OS.addBlankLine(); 982 unsigned FileId = maybeRecordFile(SP->getFile()); 983 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 984 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 985 OS.addBlankLine(); 986 OS.AddComment("Type index of inlined function"); 987 OS.emitInt32(InlineeIdx.getIndex()); 988 OS.AddComment("Offset into filechecksum table"); 989 OS.emitCVFileChecksumOffsetDirective(FileId); 990 OS.AddComment("Starting line number"); 991 OS.emitInt32(SP->getLine()); 992 } 993 994 endCVSubsection(InlineEnd); 995 } 996 997 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 998 const DILocation *InlinedAt, 999 const InlineSite &Site) { 1000 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1001 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1002 1003 // SymbolRecord 1004 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1005 1006 OS.AddComment("PtrParent"); 1007 OS.emitInt32(0); 1008 OS.AddComment("PtrEnd"); 1009 OS.emitInt32(0); 1010 OS.AddComment("Inlinee type index"); 1011 OS.emitInt32(InlineeIdx.getIndex()); 1012 1013 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1014 unsigned StartLineNum = Site.Inlinee->getLine(); 1015 1016 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1017 FI.Begin, FI.End); 1018 1019 endSymbolRecord(InlineEnd); 1020 1021 emitLocalVariableList(FI, Site.InlinedLocals); 1022 1023 // Recurse on child inlined call sites before closing the scope. 1024 for (const DILocation *ChildSite : Site.ChildSites) { 1025 auto I = FI.InlineSites.find(ChildSite); 1026 assert(I != FI.InlineSites.end() && 1027 "child site not in function inline site map"); 1028 emitInlinedCallSite(FI, ChildSite, I->second); 1029 } 1030 1031 // Close the scope. 1032 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1033 } 1034 1035 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1036 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1037 // comdat key. A section may be comdat because of -ffunction-sections or 1038 // because it is comdat in the IR. 1039 MCSectionCOFF *GVSec = 1040 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1041 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1042 1043 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1044 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1045 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1046 1047 OS.switchSection(DebugSec); 1048 1049 // Emit the magic version number if this is the first time we've switched to 1050 // this section. 1051 if (ComdatDebugSections.insert(DebugSec).second) 1052 emitCodeViewMagicVersion(); 1053 } 1054 1055 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1056 // The only supported thunk ordinal is currently the standard type. 1057 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1058 FunctionInfo &FI, 1059 const MCSymbol *Fn) { 1060 std::string FuncName = 1061 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1062 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1063 1064 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1065 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1066 1067 // Emit S_THUNK32 1068 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1069 OS.AddComment("PtrParent"); 1070 OS.emitInt32(0); 1071 OS.AddComment("PtrEnd"); 1072 OS.emitInt32(0); 1073 OS.AddComment("PtrNext"); 1074 OS.emitInt32(0); 1075 OS.AddComment("Thunk section relative address"); 1076 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1077 OS.AddComment("Thunk section index"); 1078 OS.emitCOFFSectionIndex(Fn); 1079 OS.AddComment("Code size"); 1080 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1081 OS.AddComment("Ordinal"); 1082 OS.emitInt8(unsigned(ordinal)); 1083 OS.AddComment("Function name"); 1084 emitNullTerminatedSymbolName(OS, FuncName); 1085 // Additional fields specific to the thunk ordinal would go here. 1086 endSymbolRecord(ThunkRecordEnd); 1087 1088 // Local variables/inlined routines are purposely omitted here. The point of 1089 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1090 1091 // Emit S_PROC_ID_END 1092 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1093 1094 endCVSubsection(SymbolsEnd); 1095 } 1096 1097 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1098 FunctionInfo &FI) { 1099 // For each function there is a separate subsection which holds the PC to 1100 // file:line table. 1101 const MCSymbol *Fn = Asm->getSymbol(GV); 1102 assert(Fn); 1103 1104 // Switch to the to a comdat section, if appropriate. 1105 switchToDebugSectionForSymbol(Fn); 1106 1107 std::string FuncName; 1108 auto *SP = GV->getSubprogram(); 1109 assert(SP); 1110 setCurrentSubprogram(SP); 1111 1112 if (SP->isThunk()) { 1113 emitDebugInfoForThunk(GV, FI, Fn); 1114 return; 1115 } 1116 1117 // If we have a display name, build the fully qualified name by walking the 1118 // chain of scopes. 1119 if (!SP->getName().empty()) 1120 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1121 1122 // If our DISubprogram name is empty, use the mangled name. 1123 if (FuncName.empty()) 1124 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1125 1126 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1127 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1128 OS.emitCVFPOData(Fn); 1129 1130 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1131 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1132 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1133 { 1134 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1135 : SymbolKind::S_GPROC32_ID; 1136 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1137 1138 // These fields are filled in by tools like CVPACK which run after the fact. 1139 OS.AddComment("PtrParent"); 1140 OS.emitInt32(0); 1141 OS.AddComment("PtrEnd"); 1142 OS.emitInt32(0); 1143 OS.AddComment("PtrNext"); 1144 OS.emitInt32(0); 1145 // This is the important bit that tells the debugger where the function 1146 // code is located and what's its size: 1147 OS.AddComment("Code size"); 1148 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1149 OS.AddComment("Offset after prologue"); 1150 OS.emitInt32(0); 1151 OS.AddComment("Offset before epilogue"); 1152 OS.emitInt32(0); 1153 OS.AddComment("Function type index"); 1154 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1155 OS.AddComment("Function section relative address"); 1156 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1157 OS.AddComment("Function section index"); 1158 OS.emitCOFFSectionIndex(Fn); 1159 OS.AddComment("Flags"); 1160 OS.emitInt8(0); 1161 // Emit the function display name as a null-terminated string. 1162 OS.AddComment("Function name"); 1163 // Truncate the name so we won't overflow the record length field. 1164 emitNullTerminatedSymbolName(OS, FuncName); 1165 endSymbolRecord(ProcRecordEnd); 1166 1167 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1168 // Subtract out the CSR size since MSVC excludes that and we include it. 1169 OS.AddComment("FrameSize"); 1170 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1171 OS.AddComment("Padding"); 1172 OS.emitInt32(0); 1173 OS.AddComment("Offset of padding"); 1174 OS.emitInt32(0); 1175 OS.AddComment("Bytes of callee saved registers"); 1176 OS.emitInt32(FI.CSRSize); 1177 OS.AddComment("Exception handler offset"); 1178 OS.emitInt32(0); 1179 OS.AddComment("Exception handler section"); 1180 OS.emitInt16(0); 1181 OS.AddComment("Flags (defines frame register)"); 1182 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1183 endSymbolRecord(FrameProcEnd); 1184 1185 emitLocalVariableList(FI, FI.Locals); 1186 emitGlobalVariableList(FI.Globals); 1187 emitLexicalBlockList(FI.ChildBlocks, FI); 1188 1189 // Emit inlined call site information. Only emit functions inlined directly 1190 // into the parent function. We'll emit the other sites recursively as part 1191 // of their parent inline site. 1192 for (const DILocation *InlinedAt : FI.ChildSites) { 1193 auto I = FI.InlineSites.find(InlinedAt); 1194 assert(I != FI.InlineSites.end() && 1195 "child site not in function inline site map"); 1196 emitInlinedCallSite(FI, InlinedAt, I->second); 1197 } 1198 1199 for (auto Annot : FI.Annotations) { 1200 MCSymbol *Label = Annot.first; 1201 MDTuple *Strs = cast<MDTuple>(Annot.second); 1202 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1203 OS.emitCOFFSecRel32(Label, /*Offset=*/0); 1204 // FIXME: Make sure we don't overflow the max record size. 1205 OS.emitCOFFSectionIndex(Label); 1206 OS.emitInt16(Strs->getNumOperands()); 1207 for (Metadata *MD : Strs->operands()) { 1208 // MDStrings are null terminated, so we can do EmitBytes and get the 1209 // nice .asciz directive. 1210 StringRef Str = cast<MDString>(MD)->getString(); 1211 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1212 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1213 } 1214 endSymbolRecord(AnnotEnd); 1215 } 1216 1217 for (auto HeapAllocSite : FI.HeapAllocSites) { 1218 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1219 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1220 const DIType *DITy = std::get<2>(HeapAllocSite); 1221 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1222 OS.AddComment("Call site offset"); 1223 OS.emitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1224 OS.AddComment("Call site section index"); 1225 OS.emitCOFFSectionIndex(BeginLabel); 1226 OS.AddComment("Call instruction length"); 1227 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1228 OS.AddComment("Type index"); 1229 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1230 endSymbolRecord(HeapAllocEnd); 1231 } 1232 1233 if (SP != nullptr) 1234 emitDebugInfoForUDTs(LocalUDTs); 1235 1236 // We're done with this function. 1237 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1238 } 1239 endCVSubsection(SymbolsEnd); 1240 1241 // We have an assembler directive that takes care of the whole line table. 1242 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1243 } 1244 1245 CodeViewDebug::LocalVarDef 1246 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1247 LocalVarDef DR; 1248 DR.InMemory = -1; 1249 DR.DataOffset = Offset; 1250 assert(DR.DataOffset == Offset && "truncation"); 1251 DR.IsSubfield = 0; 1252 DR.StructOffset = 0; 1253 DR.CVRegister = CVRegister; 1254 return DR; 1255 } 1256 1257 void CodeViewDebug::collectVariableInfoFromMFTable( 1258 DenseSet<InlinedEntity> &Processed) { 1259 const MachineFunction &MF = *Asm->MF; 1260 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1261 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1262 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1263 1264 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1265 if (!VI.Var) 1266 continue; 1267 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1268 "Expected inlined-at fields to agree"); 1269 1270 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1271 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1272 1273 // If variable scope is not found then skip this variable. 1274 if (!Scope) 1275 continue; 1276 1277 // If the variable has an attached offset expression, extract it. 1278 // FIXME: Try to handle DW_OP_deref as well. 1279 int64_t ExprOffset = 0; 1280 bool Deref = false; 1281 if (VI.Expr) { 1282 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1283 if (VI.Expr->getNumElements() == 1 && 1284 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1285 Deref = true; 1286 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1287 continue; 1288 } 1289 1290 // Get the frame register used and the offset. 1291 Register FrameReg; 1292 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1293 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1294 1295 assert(!FrameOffset.getScalable() && 1296 "Frame offsets with a scalable component are not supported"); 1297 1298 // Calculate the label ranges. 1299 LocalVarDef DefRange = 1300 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1301 1302 LocalVariable Var; 1303 Var.DIVar = VI.Var; 1304 1305 for (const InsnRange &Range : Scope->getRanges()) { 1306 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1307 const MCSymbol *End = getLabelAfterInsn(Range.second); 1308 End = End ? End : Asm->getFunctionEnd(); 1309 Var.DefRanges[DefRange].emplace_back(Begin, End); 1310 } 1311 1312 if (Deref) 1313 Var.UseReferenceType = true; 1314 1315 recordLocalVariable(std::move(Var), Scope); 1316 } 1317 } 1318 1319 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1320 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1321 } 1322 1323 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1324 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1325 } 1326 1327 void CodeViewDebug::calculateRanges( 1328 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1329 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1330 1331 // Calculate the definition ranges. 1332 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1333 const auto &Entry = *I; 1334 if (!Entry.isDbgValue()) 1335 continue; 1336 const MachineInstr *DVInst = Entry.getInstr(); 1337 assert(DVInst->isDebugValue() && "Invalid History entry"); 1338 // FIXME: Find a way to represent constant variables, since they are 1339 // relatively common. 1340 Optional<DbgVariableLocation> Location = 1341 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1342 if (!Location) 1343 continue; 1344 1345 // CodeView can only express variables in register and variables in memory 1346 // at a constant offset from a register. However, for variables passed 1347 // indirectly by pointer, it is common for that pointer to be spilled to a 1348 // stack location. For the special case of one offseted load followed by a 1349 // zero offset load (a pointer spilled to the stack), we change the type of 1350 // the local variable from a value type to a reference type. This tricks the 1351 // debugger into doing the load for us. 1352 if (Var.UseReferenceType) { 1353 // We're using a reference type. Drop the last zero offset load. 1354 if (canUseReferenceType(*Location)) 1355 Location->LoadChain.pop_back(); 1356 else 1357 continue; 1358 } else if (needsReferenceType(*Location)) { 1359 // This location can't be expressed without switching to a reference type. 1360 // Start over using that. 1361 Var.UseReferenceType = true; 1362 Var.DefRanges.clear(); 1363 calculateRanges(Var, Entries); 1364 return; 1365 } 1366 1367 // We can only handle a register or an offseted load of a register. 1368 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1369 continue; 1370 1371 LocalVarDef DR; 1372 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1373 DR.InMemory = !Location->LoadChain.empty(); 1374 DR.DataOffset = 1375 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1376 if (Location->FragmentInfo) { 1377 DR.IsSubfield = true; 1378 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1379 } else { 1380 DR.IsSubfield = false; 1381 DR.StructOffset = 0; 1382 } 1383 1384 // Compute the label range. 1385 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1386 const MCSymbol *End; 1387 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1388 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1389 End = EndingEntry.isDbgValue() 1390 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1391 : getLabelAfterInsn(EndingEntry.getInstr()); 1392 } else 1393 End = Asm->getFunctionEnd(); 1394 1395 // If the last range end is our begin, just extend the last range. 1396 // Otherwise make a new range. 1397 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1398 Var.DefRanges[DR]; 1399 if (!R.empty() && R.back().second == Begin) 1400 R.back().second = End; 1401 else 1402 R.emplace_back(Begin, End); 1403 1404 // FIXME: Do more range combining. 1405 } 1406 } 1407 1408 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1409 DenseSet<InlinedEntity> Processed; 1410 // Grab the variable info that was squirreled away in the MMI side-table. 1411 collectVariableInfoFromMFTable(Processed); 1412 1413 for (const auto &I : DbgValues) { 1414 InlinedEntity IV = I.first; 1415 if (Processed.count(IV)) 1416 continue; 1417 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1418 const DILocation *InlinedAt = IV.second; 1419 1420 // Instruction ranges, specifying where IV is accessible. 1421 const auto &Entries = I.second; 1422 1423 LexicalScope *Scope = nullptr; 1424 if (InlinedAt) 1425 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1426 else 1427 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1428 // If variable scope is not found then skip this variable. 1429 if (!Scope) 1430 continue; 1431 1432 LocalVariable Var; 1433 Var.DIVar = DIVar; 1434 1435 calculateRanges(Var, Entries); 1436 recordLocalVariable(std::move(Var), Scope); 1437 } 1438 } 1439 1440 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1441 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1442 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1443 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1444 const Function &GV = MF->getFunction(); 1445 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1446 assert(Insertion.second && "function already has info"); 1447 CurFn = Insertion.first->second.get(); 1448 CurFn->FuncId = NextFuncId++; 1449 CurFn->Begin = Asm->getFunctionBegin(); 1450 1451 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1452 // callee-saved registers were used. For targets that don't use a PUSH 1453 // instruction (AArch64), this will be zero. 1454 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1455 CurFn->FrameSize = MFI.getStackSize(); 1456 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1457 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1458 1459 // For this function S_FRAMEPROC record, figure out which codeview register 1460 // will be the frame pointer. 1461 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1462 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1463 if (CurFn->FrameSize > 0) { 1464 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1465 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1466 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1467 } else { 1468 // If there is an FP, parameters are always relative to it. 1469 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1470 if (CurFn->HasStackRealignment) { 1471 // If the stack needs realignment, locals are relative to SP or VFRAME. 1472 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1473 } else { 1474 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1475 // other stack adjustments. 1476 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1477 } 1478 } 1479 } 1480 1481 // Compute other frame procedure options. 1482 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1483 if (MFI.hasVarSizedObjects()) 1484 FPO |= FrameProcedureOptions::HasAlloca; 1485 if (MF->exposesReturnsTwice()) 1486 FPO |= FrameProcedureOptions::HasSetJmp; 1487 // FIXME: Set HasLongJmp if we ever track that info. 1488 if (MF->hasInlineAsm()) 1489 FPO |= FrameProcedureOptions::HasInlineAssembly; 1490 if (GV.hasPersonalityFn()) { 1491 if (isAsynchronousEHPersonality( 1492 classifyEHPersonality(GV.getPersonalityFn()))) 1493 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1494 else 1495 FPO |= FrameProcedureOptions::HasExceptionHandling; 1496 } 1497 if (GV.hasFnAttribute(Attribute::InlineHint)) 1498 FPO |= FrameProcedureOptions::MarkedInline; 1499 if (GV.hasFnAttribute(Attribute::Naked)) 1500 FPO |= FrameProcedureOptions::Naked; 1501 if (MFI.hasStackProtectorIndex()) 1502 FPO |= FrameProcedureOptions::SecurityChecks; 1503 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1504 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1505 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1506 !GV.hasOptSize() && !GV.hasOptNone()) 1507 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1508 if (GV.hasProfileData()) { 1509 FPO |= FrameProcedureOptions::ValidProfileCounts; 1510 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1511 } 1512 // FIXME: Set GuardCfg when it is implemented. 1513 CurFn->FrameProcOpts = FPO; 1514 1515 OS.emitCVFuncIdDirective(CurFn->FuncId); 1516 1517 // Find the end of the function prolog. First known non-DBG_VALUE and 1518 // non-frame setup location marks the beginning of the function body. 1519 // FIXME: is there a simpler a way to do this? Can we just search 1520 // for the first instruction of the function, not the last of the prolog? 1521 DebugLoc PrologEndLoc; 1522 bool EmptyPrologue = true; 1523 for (const auto &MBB : *MF) { 1524 for (const auto &MI : MBB) { 1525 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1526 MI.getDebugLoc()) { 1527 PrologEndLoc = MI.getDebugLoc(); 1528 break; 1529 } else if (!MI.isMetaInstruction()) { 1530 EmptyPrologue = false; 1531 } 1532 } 1533 } 1534 1535 // Record beginning of function if we have a non-empty prologue. 1536 if (PrologEndLoc && !EmptyPrologue) { 1537 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1538 maybeRecordLocation(FnStartDL, MF); 1539 } 1540 1541 // Find heap alloc sites and emit labels around them. 1542 for (const auto &MBB : *MF) { 1543 for (const auto &MI : MBB) { 1544 if (MI.getHeapAllocMarker()) { 1545 requestLabelBeforeInsn(&MI); 1546 requestLabelAfterInsn(&MI); 1547 } 1548 } 1549 } 1550 } 1551 1552 static bool shouldEmitUdt(const DIType *T) { 1553 if (!T) 1554 return false; 1555 1556 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1557 if (T->getTag() == dwarf::DW_TAG_typedef) { 1558 if (DIScope *Scope = T->getScope()) { 1559 switch (Scope->getTag()) { 1560 case dwarf::DW_TAG_structure_type: 1561 case dwarf::DW_TAG_class_type: 1562 case dwarf::DW_TAG_union_type: 1563 return false; 1564 default: 1565 // do nothing. 1566 ; 1567 } 1568 } 1569 } 1570 1571 while (true) { 1572 if (!T || T->isForwardDecl()) 1573 return false; 1574 1575 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1576 if (!DT) 1577 return true; 1578 T = DT->getBaseType(); 1579 } 1580 return true; 1581 } 1582 1583 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1584 // Don't record empty UDTs. 1585 if (Ty->getName().empty()) 1586 return; 1587 if (!shouldEmitUdt(Ty)) 1588 return; 1589 1590 SmallVector<StringRef, 5> ParentScopeNames; 1591 const DISubprogram *ClosestSubprogram = 1592 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1593 1594 std::string FullyQualifiedName = 1595 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1596 1597 if (ClosestSubprogram == nullptr) { 1598 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1599 } else if (ClosestSubprogram == CurrentSubprogram) { 1600 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1601 } 1602 1603 // TODO: What if the ClosestSubprogram is neither null or the current 1604 // subprogram? Currently, the UDT just gets dropped on the floor. 1605 // 1606 // The current behavior is not desirable. To get maximal fidelity, we would 1607 // need to perform all type translation before beginning emission of .debug$S 1608 // and then make LocalUDTs a member of FunctionInfo 1609 } 1610 1611 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1612 // Generic dispatch for lowering an unknown type. 1613 switch (Ty->getTag()) { 1614 case dwarf::DW_TAG_array_type: 1615 return lowerTypeArray(cast<DICompositeType>(Ty)); 1616 case dwarf::DW_TAG_typedef: 1617 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1618 case dwarf::DW_TAG_base_type: 1619 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1620 case dwarf::DW_TAG_pointer_type: 1621 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1622 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1623 LLVM_FALLTHROUGH; 1624 case dwarf::DW_TAG_reference_type: 1625 case dwarf::DW_TAG_rvalue_reference_type: 1626 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1627 case dwarf::DW_TAG_ptr_to_member_type: 1628 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1629 case dwarf::DW_TAG_restrict_type: 1630 case dwarf::DW_TAG_const_type: 1631 case dwarf::DW_TAG_volatile_type: 1632 // TODO: add support for DW_TAG_atomic_type here 1633 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1634 case dwarf::DW_TAG_subroutine_type: 1635 if (ClassTy) { 1636 // The member function type of a member function pointer has no 1637 // ThisAdjustment. 1638 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1639 /*ThisAdjustment=*/0, 1640 /*IsStaticMethod=*/false); 1641 } 1642 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1643 case dwarf::DW_TAG_enumeration_type: 1644 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1645 case dwarf::DW_TAG_class_type: 1646 case dwarf::DW_TAG_structure_type: 1647 return lowerTypeClass(cast<DICompositeType>(Ty)); 1648 case dwarf::DW_TAG_union_type: 1649 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1650 case dwarf::DW_TAG_string_type: 1651 return lowerTypeString(cast<DIStringType>(Ty)); 1652 case dwarf::DW_TAG_unspecified_type: 1653 if (Ty->getName() == "decltype(nullptr)") 1654 return TypeIndex::NullptrT(); 1655 return TypeIndex::None(); 1656 default: 1657 // Use the null type index. 1658 return TypeIndex(); 1659 } 1660 } 1661 1662 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1663 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1664 StringRef TypeName = Ty->getName(); 1665 1666 addToUDTs(Ty); 1667 1668 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1669 TypeName == "HRESULT") 1670 return TypeIndex(SimpleTypeKind::HResult); 1671 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1672 TypeName == "wchar_t") 1673 return TypeIndex(SimpleTypeKind::WideCharacter); 1674 1675 return UnderlyingTypeIndex; 1676 } 1677 1678 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1679 const DIType *ElementType = Ty->getBaseType(); 1680 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1681 // IndexType is size_t, which depends on the bitness of the target. 1682 TypeIndex IndexType = getPointerSizeInBytes() == 8 1683 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1684 : TypeIndex(SimpleTypeKind::UInt32Long); 1685 1686 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1687 1688 // Add subranges to array type. 1689 DINodeArray Elements = Ty->getElements(); 1690 for (int i = Elements.size() - 1; i >= 0; --i) { 1691 const DINode *Element = Elements[i]; 1692 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1693 1694 const DISubrange *Subrange = cast<DISubrange>(Element); 1695 int64_t Count = -1; 1696 1697 // If Subrange has a Count field, use it. 1698 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1699 // where lowerbound is from the LowerBound field of the Subrange, 1700 // or the language default lowerbound if that field is unspecified. 1701 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1702 Count = CI->getSExtValue(); 1703 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1704 // Fortran uses 1 as the default lowerbound; other languages use 0. 1705 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1706 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1707 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1708 Count = UI->getSExtValue() - Lowerbound + 1; 1709 } 1710 1711 // Forward declarations of arrays without a size and VLAs use a count of -1. 1712 // Emit a count of zero in these cases to match what MSVC does for arrays 1713 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1714 // should do for them even if we could distinguish them. 1715 if (Count == -1) 1716 Count = 0; 1717 1718 // Update the element size and element type index for subsequent subranges. 1719 ElementSize *= Count; 1720 1721 // If this is the outermost array, use the size from the array. It will be 1722 // more accurate if we had a VLA or an incomplete element type size. 1723 uint64_t ArraySize = 1724 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1725 1726 StringRef Name = (i == 0) ? Ty->getName() : ""; 1727 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1728 ElementTypeIndex = TypeTable.writeLeafType(AR); 1729 } 1730 1731 return ElementTypeIndex; 1732 } 1733 1734 // This function lowers a Fortran character type (DIStringType). 1735 // Note that it handles only the character*n variant (using SizeInBits 1736 // field in DIString to describe the type size) at the moment. 1737 // Other variants (leveraging the StringLength and StringLengthExp 1738 // fields in DIStringType) remain TBD. 1739 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1740 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1741 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1742 StringRef Name = Ty->getName(); 1743 // IndexType is size_t, which depends on the bitness of the target. 1744 TypeIndex IndexType = getPointerSizeInBytes() == 8 1745 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1746 : TypeIndex(SimpleTypeKind::UInt32Long); 1747 1748 // Create a type of character array of ArraySize. 1749 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1750 1751 return TypeTable.writeLeafType(AR); 1752 } 1753 1754 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1755 TypeIndex Index; 1756 dwarf::TypeKind Kind; 1757 uint32_t ByteSize; 1758 1759 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1760 ByteSize = Ty->getSizeInBits() / 8; 1761 1762 SimpleTypeKind STK = SimpleTypeKind::None; 1763 switch (Kind) { 1764 case dwarf::DW_ATE_address: 1765 // FIXME: Translate 1766 break; 1767 case dwarf::DW_ATE_boolean: 1768 switch (ByteSize) { 1769 case 1: STK = SimpleTypeKind::Boolean8; break; 1770 case 2: STK = SimpleTypeKind::Boolean16; break; 1771 case 4: STK = SimpleTypeKind::Boolean32; break; 1772 case 8: STK = SimpleTypeKind::Boolean64; break; 1773 case 16: STK = SimpleTypeKind::Boolean128; break; 1774 } 1775 break; 1776 case dwarf::DW_ATE_complex_float: 1777 switch (ByteSize) { 1778 case 2: STK = SimpleTypeKind::Complex16; break; 1779 case 4: STK = SimpleTypeKind::Complex32; break; 1780 case 8: STK = SimpleTypeKind::Complex64; break; 1781 case 10: STK = SimpleTypeKind::Complex80; break; 1782 case 16: STK = SimpleTypeKind::Complex128; break; 1783 } 1784 break; 1785 case dwarf::DW_ATE_float: 1786 switch (ByteSize) { 1787 case 2: STK = SimpleTypeKind::Float16; break; 1788 case 4: STK = SimpleTypeKind::Float32; break; 1789 case 6: STK = SimpleTypeKind::Float48; break; 1790 case 8: STK = SimpleTypeKind::Float64; break; 1791 case 10: STK = SimpleTypeKind::Float80; break; 1792 case 16: STK = SimpleTypeKind::Float128; break; 1793 } 1794 break; 1795 case dwarf::DW_ATE_signed: 1796 switch (ByteSize) { 1797 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1798 case 2: STK = SimpleTypeKind::Int16Short; break; 1799 case 4: STK = SimpleTypeKind::Int32; break; 1800 case 8: STK = SimpleTypeKind::Int64Quad; break; 1801 case 16: STK = SimpleTypeKind::Int128Oct; break; 1802 } 1803 break; 1804 case dwarf::DW_ATE_unsigned: 1805 switch (ByteSize) { 1806 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1807 case 2: STK = SimpleTypeKind::UInt16Short; break; 1808 case 4: STK = SimpleTypeKind::UInt32; break; 1809 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1810 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1811 } 1812 break; 1813 case dwarf::DW_ATE_UTF: 1814 switch (ByteSize) { 1815 case 1: STK = SimpleTypeKind::Character8; break; 1816 case 2: STK = SimpleTypeKind::Character16; break; 1817 case 4: STK = SimpleTypeKind::Character32; break; 1818 } 1819 break; 1820 case dwarf::DW_ATE_signed_char: 1821 if (ByteSize == 1) 1822 STK = SimpleTypeKind::SignedCharacter; 1823 break; 1824 case dwarf::DW_ATE_unsigned_char: 1825 if (ByteSize == 1) 1826 STK = SimpleTypeKind::UnsignedCharacter; 1827 break; 1828 default: 1829 break; 1830 } 1831 1832 // Apply some fixups based on the source-level type name. 1833 // Include some amount of canonicalization from an old naming scheme Clang 1834 // used to use for integer types (in an outdated effort to be compatible with 1835 // GCC's debug info/GDB's behavior, which has since been addressed). 1836 if (STK == SimpleTypeKind::Int32 && 1837 (Ty->getName() == "long int" || Ty->getName() == "long")) 1838 STK = SimpleTypeKind::Int32Long; 1839 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1840 Ty->getName() == "unsigned long")) 1841 STK = SimpleTypeKind::UInt32Long; 1842 if (STK == SimpleTypeKind::UInt16Short && 1843 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1844 STK = SimpleTypeKind::WideCharacter; 1845 if ((STK == SimpleTypeKind::SignedCharacter || 1846 STK == SimpleTypeKind::UnsignedCharacter) && 1847 Ty->getName() == "char") 1848 STK = SimpleTypeKind::NarrowCharacter; 1849 1850 return TypeIndex(STK); 1851 } 1852 1853 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1854 PointerOptions PO) { 1855 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1856 1857 // Pointers to simple types without any options can use SimpleTypeMode, rather 1858 // than having a dedicated pointer type record. 1859 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1860 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1861 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1862 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1863 ? SimpleTypeMode::NearPointer64 1864 : SimpleTypeMode::NearPointer32; 1865 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1866 } 1867 1868 PointerKind PK = 1869 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1870 PointerMode PM = PointerMode::Pointer; 1871 switch (Ty->getTag()) { 1872 default: llvm_unreachable("not a pointer tag type"); 1873 case dwarf::DW_TAG_pointer_type: 1874 PM = PointerMode::Pointer; 1875 break; 1876 case dwarf::DW_TAG_reference_type: 1877 PM = PointerMode::LValueReference; 1878 break; 1879 case dwarf::DW_TAG_rvalue_reference_type: 1880 PM = PointerMode::RValueReference; 1881 break; 1882 } 1883 1884 if (Ty->isObjectPointer()) 1885 PO |= PointerOptions::Const; 1886 1887 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1888 return TypeTable.writeLeafType(PR); 1889 } 1890 1891 static PointerToMemberRepresentation 1892 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1893 // SizeInBytes being zero generally implies that the member pointer type was 1894 // incomplete, which can happen if it is part of a function prototype. In this 1895 // case, use the unknown model instead of the general model. 1896 if (IsPMF) { 1897 switch (Flags & DINode::FlagPtrToMemberRep) { 1898 case 0: 1899 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1900 : PointerToMemberRepresentation::GeneralFunction; 1901 case DINode::FlagSingleInheritance: 1902 return PointerToMemberRepresentation::SingleInheritanceFunction; 1903 case DINode::FlagMultipleInheritance: 1904 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1905 case DINode::FlagVirtualInheritance: 1906 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1907 } 1908 } else { 1909 switch (Flags & DINode::FlagPtrToMemberRep) { 1910 case 0: 1911 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1912 : PointerToMemberRepresentation::GeneralData; 1913 case DINode::FlagSingleInheritance: 1914 return PointerToMemberRepresentation::SingleInheritanceData; 1915 case DINode::FlagMultipleInheritance: 1916 return PointerToMemberRepresentation::MultipleInheritanceData; 1917 case DINode::FlagVirtualInheritance: 1918 return PointerToMemberRepresentation::VirtualInheritanceData; 1919 } 1920 } 1921 llvm_unreachable("invalid ptr to member representation"); 1922 } 1923 1924 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1925 PointerOptions PO) { 1926 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1927 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1928 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1929 TypeIndex PointeeTI = 1930 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1931 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1932 : PointerKind::Near32; 1933 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1934 : PointerMode::PointerToDataMember; 1935 1936 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1937 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1938 MemberPointerInfo MPI( 1939 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1940 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1941 return TypeTable.writeLeafType(PR); 1942 } 1943 1944 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1945 /// have a translation, use the NearC convention. 1946 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1947 switch (DwarfCC) { 1948 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1949 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1950 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1951 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1952 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1953 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1954 } 1955 return CallingConvention::NearC; 1956 } 1957 1958 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1959 ModifierOptions Mods = ModifierOptions::None; 1960 PointerOptions PO = PointerOptions::None; 1961 bool IsModifier = true; 1962 const DIType *BaseTy = Ty; 1963 while (IsModifier && BaseTy) { 1964 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1965 switch (BaseTy->getTag()) { 1966 case dwarf::DW_TAG_const_type: 1967 Mods |= ModifierOptions::Const; 1968 PO |= PointerOptions::Const; 1969 break; 1970 case dwarf::DW_TAG_volatile_type: 1971 Mods |= ModifierOptions::Volatile; 1972 PO |= PointerOptions::Volatile; 1973 break; 1974 case dwarf::DW_TAG_restrict_type: 1975 // Only pointer types be marked with __restrict. There is no known flag 1976 // for __restrict in LF_MODIFIER records. 1977 PO |= PointerOptions::Restrict; 1978 break; 1979 default: 1980 IsModifier = false; 1981 break; 1982 } 1983 if (IsModifier) 1984 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1985 } 1986 1987 // Check if the inner type will use an LF_POINTER record. If so, the 1988 // qualifiers will go in the LF_POINTER record. This comes up for types like 1989 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1990 // char *'. 1991 if (BaseTy) { 1992 switch (BaseTy->getTag()) { 1993 case dwarf::DW_TAG_pointer_type: 1994 case dwarf::DW_TAG_reference_type: 1995 case dwarf::DW_TAG_rvalue_reference_type: 1996 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1997 case dwarf::DW_TAG_ptr_to_member_type: 1998 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1999 default: 2000 break; 2001 } 2002 } 2003 2004 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2005 2006 // Return the base type index if there aren't any modifiers. For example, the 2007 // metadata could contain restrict wrappers around non-pointer types. 2008 if (Mods == ModifierOptions::None) 2009 return ModifiedTI; 2010 2011 ModifierRecord MR(ModifiedTI, Mods); 2012 return TypeTable.writeLeafType(MR); 2013 } 2014 2015 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2016 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2017 for (const DIType *ArgType : Ty->getTypeArray()) 2018 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2019 2020 // MSVC uses type none for variadic argument. 2021 if (ReturnAndArgTypeIndices.size() > 1 && 2022 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2023 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2024 } 2025 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2026 ArrayRef<TypeIndex> ArgTypeIndices = None; 2027 if (!ReturnAndArgTypeIndices.empty()) { 2028 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 2029 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2030 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2031 } 2032 2033 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2034 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2035 2036 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2037 2038 FunctionOptions FO = getFunctionOptions(Ty); 2039 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2040 ArgListIndex); 2041 return TypeTable.writeLeafType(Procedure); 2042 } 2043 2044 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2045 const DIType *ClassTy, 2046 int ThisAdjustment, 2047 bool IsStaticMethod, 2048 FunctionOptions FO) { 2049 // Lower the containing class type. 2050 TypeIndex ClassType = getTypeIndex(ClassTy); 2051 2052 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2053 2054 unsigned Index = 0; 2055 SmallVector<TypeIndex, 8> ArgTypeIndices; 2056 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2057 if (ReturnAndArgs.size() > Index) { 2058 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2059 } 2060 2061 // If the first argument is a pointer type and this isn't a static method, 2062 // treat it as the special 'this' parameter, which is encoded separately from 2063 // the arguments. 2064 TypeIndex ThisTypeIndex; 2065 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2066 if (const DIDerivedType *PtrTy = 2067 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2068 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2069 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2070 Index++; 2071 } 2072 } 2073 } 2074 2075 while (Index < ReturnAndArgs.size()) 2076 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2077 2078 // MSVC uses type none for variadic argument. 2079 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2080 ArgTypeIndices.back() = TypeIndex::None(); 2081 2082 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2083 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2084 2085 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2086 2087 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2088 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2089 return TypeTable.writeLeafType(MFR); 2090 } 2091 2092 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2093 unsigned VSlotCount = 2094 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2095 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2096 2097 VFTableShapeRecord VFTSR(Slots); 2098 return TypeTable.writeLeafType(VFTSR); 2099 } 2100 2101 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2102 switch (Flags & DINode::FlagAccessibility) { 2103 case DINode::FlagPrivate: return MemberAccess::Private; 2104 case DINode::FlagPublic: return MemberAccess::Public; 2105 case DINode::FlagProtected: return MemberAccess::Protected; 2106 case 0: 2107 // If there was no explicit access control, provide the default for the tag. 2108 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2109 : MemberAccess::Public; 2110 } 2111 llvm_unreachable("access flags are exclusive"); 2112 } 2113 2114 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2115 if (SP->isArtificial()) 2116 return MethodOptions::CompilerGenerated; 2117 2118 // FIXME: Handle other MethodOptions. 2119 2120 return MethodOptions::None; 2121 } 2122 2123 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2124 bool Introduced) { 2125 if (SP->getFlags() & DINode::FlagStaticMember) 2126 return MethodKind::Static; 2127 2128 switch (SP->getVirtuality()) { 2129 case dwarf::DW_VIRTUALITY_none: 2130 break; 2131 case dwarf::DW_VIRTUALITY_virtual: 2132 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2133 case dwarf::DW_VIRTUALITY_pure_virtual: 2134 return Introduced ? MethodKind::PureIntroducingVirtual 2135 : MethodKind::PureVirtual; 2136 default: 2137 llvm_unreachable("unhandled virtuality case"); 2138 } 2139 2140 return MethodKind::Vanilla; 2141 } 2142 2143 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2144 switch (Ty->getTag()) { 2145 case dwarf::DW_TAG_class_type: 2146 return TypeRecordKind::Class; 2147 case dwarf::DW_TAG_structure_type: 2148 return TypeRecordKind::Struct; 2149 default: 2150 llvm_unreachable("unexpected tag"); 2151 } 2152 } 2153 2154 /// Return ClassOptions that should be present on both the forward declaration 2155 /// and the defintion of a tag type. 2156 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2157 ClassOptions CO = ClassOptions::None; 2158 2159 // MSVC always sets this flag, even for local types. Clang doesn't always 2160 // appear to give every type a linkage name, which may be problematic for us. 2161 // FIXME: Investigate the consequences of not following them here. 2162 if (!Ty->getIdentifier().empty()) 2163 CO |= ClassOptions::HasUniqueName; 2164 2165 // Put the Nested flag on a type if it appears immediately inside a tag type. 2166 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2167 // here. That flag is only set on definitions, and not forward declarations. 2168 const DIScope *ImmediateScope = Ty->getScope(); 2169 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2170 CO |= ClassOptions::Nested; 2171 2172 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2173 // type only when it has an immediate function scope. Clang never puts enums 2174 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2175 // always in function, class, or file scopes. 2176 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2177 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2178 CO |= ClassOptions::Scoped; 2179 } else { 2180 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2181 Scope = Scope->getScope()) { 2182 if (isa<DISubprogram>(Scope)) { 2183 CO |= ClassOptions::Scoped; 2184 break; 2185 } 2186 } 2187 } 2188 2189 return CO; 2190 } 2191 2192 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2193 switch (Ty->getTag()) { 2194 case dwarf::DW_TAG_class_type: 2195 case dwarf::DW_TAG_structure_type: 2196 case dwarf::DW_TAG_union_type: 2197 case dwarf::DW_TAG_enumeration_type: 2198 break; 2199 default: 2200 return; 2201 } 2202 2203 if (const auto *File = Ty->getFile()) { 2204 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2205 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2206 2207 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2208 TypeTable.writeLeafType(USLR); 2209 } 2210 } 2211 2212 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2213 ClassOptions CO = getCommonClassOptions(Ty); 2214 TypeIndex FTI; 2215 unsigned EnumeratorCount = 0; 2216 2217 if (Ty->isForwardDecl()) { 2218 CO |= ClassOptions::ForwardReference; 2219 } else { 2220 ContinuationRecordBuilder ContinuationBuilder; 2221 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2222 for (const DINode *Element : Ty->getElements()) { 2223 // We assume that the frontend provides all members in source declaration 2224 // order, which is what MSVC does. 2225 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2226 // FIXME: Is it correct to always emit these as unsigned here? 2227 EnumeratorRecord ER(MemberAccess::Public, 2228 APSInt(Enumerator->getValue(), true), 2229 Enumerator->getName()); 2230 ContinuationBuilder.writeMemberType(ER); 2231 EnumeratorCount++; 2232 } 2233 } 2234 FTI = TypeTable.insertRecord(ContinuationBuilder); 2235 } 2236 2237 std::string FullName = getFullyQualifiedName(Ty); 2238 2239 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2240 getTypeIndex(Ty->getBaseType())); 2241 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2242 2243 addUDTSrcLine(Ty, EnumTI); 2244 2245 return EnumTI; 2246 } 2247 2248 //===----------------------------------------------------------------------===// 2249 // ClassInfo 2250 //===----------------------------------------------------------------------===// 2251 2252 struct llvm::ClassInfo { 2253 struct MemberInfo { 2254 const DIDerivedType *MemberTypeNode; 2255 uint64_t BaseOffset; 2256 }; 2257 // [MemberInfo] 2258 using MemberList = std::vector<MemberInfo>; 2259 2260 using MethodsList = TinyPtrVector<const DISubprogram *>; 2261 // MethodName -> MethodsList 2262 using MethodsMap = MapVector<MDString *, MethodsList>; 2263 2264 /// Base classes. 2265 std::vector<const DIDerivedType *> Inheritance; 2266 2267 /// Direct members. 2268 MemberList Members; 2269 // Direct overloaded methods gathered by name. 2270 MethodsMap Methods; 2271 2272 TypeIndex VShapeTI; 2273 2274 std::vector<const DIType *> NestedTypes; 2275 }; 2276 2277 void CodeViewDebug::clear() { 2278 assert(CurFn == nullptr); 2279 FileIdMap.clear(); 2280 FnDebugInfo.clear(); 2281 FileToFilepathMap.clear(); 2282 LocalUDTs.clear(); 2283 GlobalUDTs.clear(); 2284 TypeIndices.clear(); 2285 CompleteTypeIndices.clear(); 2286 ScopeGlobals.clear(); 2287 CVGlobalVariableOffsets.clear(); 2288 } 2289 2290 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2291 const DIDerivedType *DDTy) { 2292 if (!DDTy->getName().empty()) { 2293 Info.Members.push_back({DDTy, 0}); 2294 2295 // Collect static const data members with values. 2296 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2297 DINode::FlagStaticMember) { 2298 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2299 isa<ConstantFP>(DDTy->getConstant()))) 2300 StaticConstMembers.push_back(DDTy); 2301 } 2302 2303 return; 2304 } 2305 2306 // An unnamed member may represent a nested struct or union. Attempt to 2307 // interpret the unnamed member as a DICompositeType possibly wrapped in 2308 // qualifier types. Add all the indirect fields to the current record if that 2309 // succeeds, and drop the member if that fails. 2310 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2311 uint64_t Offset = DDTy->getOffsetInBits(); 2312 const DIType *Ty = DDTy->getBaseType(); 2313 bool FullyResolved = false; 2314 while (!FullyResolved) { 2315 switch (Ty->getTag()) { 2316 case dwarf::DW_TAG_const_type: 2317 case dwarf::DW_TAG_volatile_type: 2318 // FIXME: we should apply the qualifier types to the indirect fields 2319 // rather than dropping them. 2320 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2321 break; 2322 default: 2323 FullyResolved = true; 2324 break; 2325 } 2326 } 2327 2328 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2329 if (!DCTy) 2330 return; 2331 2332 ClassInfo NestedInfo = collectClassInfo(DCTy); 2333 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2334 Info.Members.push_back( 2335 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2336 } 2337 2338 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2339 ClassInfo Info; 2340 // Add elements to structure type. 2341 DINodeArray Elements = Ty->getElements(); 2342 for (auto *Element : Elements) { 2343 // We assume that the frontend provides all members in source declaration 2344 // order, which is what MSVC does. 2345 if (!Element) 2346 continue; 2347 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2348 Info.Methods[SP->getRawName()].push_back(SP); 2349 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2350 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2351 collectMemberInfo(Info, DDTy); 2352 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2353 Info.Inheritance.push_back(DDTy); 2354 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2355 DDTy->getName() == "__vtbl_ptr_type") { 2356 Info.VShapeTI = getTypeIndex(DDTy); 2357 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2358 Info.NestedTypes.push_back(DDTy); 2359 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2360 // Ignore friend members. It appears that MSVC emitted info about 2361 // friends in the past, but modern versions do not. 2362 } 2363 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2364 Info.NestedTypes.push_back(Composite); 2365 } 2366 // Skip other unrecognized kinds of elements. 2367 } 2368 return Info; 2369 } 2370 2371 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2372 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2373 // if a complete type should be emitted instead of a forward reference. 2374 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2375 !Ty->isForwardDecl(); 2376 } 2377 2378 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2379 // Emit the complete type for unnamed structs. C++ classes with methods 2380 // which have a circular reference back to the class type are expected to 2381 // be named by the front-end and should not be "unnamed". C unnamed 2382 // structs should not have circular references. 2383 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2384 // If this unnamed complete type is already in the process of being defined 2385 // then the description of the type is malformed and cannot be emitted 2386 // into CodeView correctly so report a fatal error. 2387 auto I = CompleteTypeIndices.find(Ty); 2388 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2389 report_fatal_error("cannot debug circular reference to unnamed type"); 2390 return getCompleteTypeIndex(Ty); 2391 } 2392 2393 // First, construct the forward decl. Don't look into Ty to compute the 2394 // forward decl options, since it might not be available in all TUs. 2395 TypeRecordKind Kind = getRecordKind(Ty); 2396 ClassOptions CO = 2397 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2398 std::string FullName = getFullyQualifiedName(Ty); 2399 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2400 FullName, Ty->getIdentifier()); 2401 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2402 if (!Ty->isForwardDecl()) 2403 DeferredCompleteTypes.push_back(Ty); 2404 return FwdDeclTI; 2405 } 2406 2407 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2408 // Construct the field list and complete type record. 2409 TypeRecordKind Kind = getRecordKind(Ty); 2410 ClassOptions CO = getCommonClassOptions(Ty); 2411 TypeIndex FieldTI; 2412 TypeIndex VShapeTI; 2413 unsigned FieldCount; 2414 bool ContainsNestedClass; 2415 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2416 lowerRecordFieldList(Ty); 2417 2418 if (ContainsNestedClass) 2419 CO |= ClassOptions::ContainsNestedClass; 2420 2421 // MSVC appears to set this flag by searching any destructor or method with 2422 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2423 // the members, however special member functions are not yet emitted into 2424 // debug information. For now checking a class's non-triviality seems enough. 2425 // FIXME: not true for a nested unnamed struct. 2426 if (isNonTrivial(Ty)) 2427 CO |= ClassOptions::HasConstructorOrDestructor; 2428 2429 std::string FullName = getFullyQualifiedName(Ty); 2430 2431 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2432 2433 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2434 SizeInBytes, FullName, Ty->getIdentifier()); 2435 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2436 2437 addUDTSrcLine(Ty, ClassTI); 2438 2439 addToUDTs(Ty); 2440 2441 return ClassTI; 2442 } 2443 2444 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2445 // Emit the complete type for unnamed unions. 2446 if (shouldAlwaysEmitCompleteClassType(Ty)) 2447 return getCompleteTypeIndex(Ty); 2448 2449 ClassOptions CO = 2450 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2451 std::string FullName = getFullyQualifiedName(Ty); 2452 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2453 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2454 if (!Ty->isForwardDecl()) 2455 DeferredCompleteTypes.push_back(Ty); 2456 return FwdDeclTI; 2457 } 2458 2459 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2460 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2461 TypeIndex FieldTI; 2462 unsigned FieldCount; 2463 bool ContainsNestedClass; 2464 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2465 lowerRecordFieldList(Ty); 2466 2467 if (ContainsNestedClass) 2468 CO |= ClassOptions::ContainsNestedClass; 2469 2470 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2471 std::string FullName = getFullyQualifiedName(Ty); 2472 2473 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2474 Ty->getIdentifier()); 2475 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2476 2477 addUDTSrcLine(Ty, UnionTI); 2478 2479 addToUDTs(Ty); 2480 2481 return UnionTI; 2482 } 2483 2484 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2485 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2486 // Manually count members. MSVC appears to count everything that generates a 2487 // field list record. Each individual overload in a method overload group 2488 // contributes to this count, even though the overload group is a single field 2489 // list record. 2490 unsigned MemberCount = 0; 2491 ClassInfo Info = collectClassInfo(Ty); 2492 ContinuationRecordBuilder ContinuationBuilder; 2493 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2494 2495 // Create base classes. 2496 for (const DIDerivedType *I : Info.Inheritance) { 2497 if (I->getFlags() & DINode::FlagVirtual) { 2498 // Virtual base. 2499 unsigned VBPtrOffset = I->getVBPtrOffset(); 2500 // FIXME: Despite the accessor name, the offset is really in bytes. 2501 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2502 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2503 ? TypeRecordKind::IndirectVirtualBaseClass 2504 : TypeRecordKind::VirtualBaseClass; 2505 VirtualBaseClassRecord VBCR( 2506 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2507 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2508 VBTableIndex); 2509 2510 ContinuationBuilder.writeMemberType(VBCR); 2511 MemberCount++; 2512 } else { 2513 assert(I->getOffsetInBits() % 8 == 0 && 2514 "bases must be on byte boundaries"); 2515 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2516 getTypeIndex(I->getBaseType()), 2517 I->getOffsetInBits() / 8); 2518 ContinuationBuilder.writeMemberType(BCR); 2519 MemberCount++; 2520 } 2521 } 2522 2523 // Create members. 2524 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2525 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2526 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2527 StringRef MemberName = Member->getName(); 2528 MemberAccess Access = 2529 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2530 2531 if (Member->isStaticMember()) { 2532 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2533 ContinuationBuilder.writeMemberType(SDMR); 2534 MemberCount++; 2535 continue; 2536 } 2537 2538 // Virtual function pointer member. 2539 if ((Member->getFlags() & DINode::FlagArtificial) && 2540 Member->getName().startswith("_vptr$")) { 2541 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2542 ContinuationBuilder.writeMemberType(VFPR); 2543 MemberCount++; 2544 continue; 2545 } 2546 2547 // Data member. 2548 uint64_t MemberOffsetInBits = 2549 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2550 if (Member->isBitField()) { 2551 uint64_t StartBitOffset = MemberOffsetInBits; 2552 if (const auto *CI = 2553 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2554 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2555 } 2556 StartBitOffset -= MemberOffsetInBits; 2557 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2558 StartBitOffset); 2559 MemberBaseType = TypeTable.writeLeafType(BFR); 2560 } 2561 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2562 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2563 MemberName); 2564 ContinuationBuilder.writeMemberType(DMR); 2565 MemberCount++; 2566 } 2567 2568 // Create methods 2569 for (auto &MethodItr : Info.Methods) { 2570 StringRef Name = MethodItr.first->getString(); 2571 2572 std::vector<OneMethodRecord> Methods; 2573 for (const DISubprogram *SP : MethodItr.second) { 2574 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2575 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2576 2577 unsigned VFTableOffset = -1; 2578 if (Introduced) 2579 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2580 2581 Methods.push_back(OneMethodRecord( 2582 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2583 translateMethodKindFlags(SP, Introduced), 2584 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2585 MemberCount++; 2586 } 2587 assert(!Methods.empty() && "Empty methods map entry"); 2588 if (Methods.size() == 1) 2589 ContinuationBuilder.writeMemberType(Methods[0]); 2590 else { 2591 // FIXME: Make this use its own ContinuationBuilder so that 2592 // MethodOverloadList can be split correctly. 2593 MethodOverloadListRecord MOLR(Methods); 2594 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2595 2596 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2597 ContinuationBuilder.writeMemberType(OMR); 2598 } 2599 } 2600 2601 // Create nested classes. 2602 for (const DIType *Nested : Info.NestedTypes) { 2603 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2604 ContinuationBuilder.writeMemberType(R); 2605 MemberCount++; 2606 } 2607 2608 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2609 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2610 !Info.NestedTypes.empty()); 2611 } 2612 2613 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2614 if (!VBPType.getIndex()) { 2615 // Make a 'const int *' type. 2616 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2617 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2618 2619 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2620 : PointerKind::Near32; 2621 PointerMode PM = PointerMode::Pointer; 2622 PointerOptions PO = PointerOptions::None; 2623 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2624 VBPType = TypeTable.writeLeafType(PR); 2625 } 2626 2627 return VBPType; 2628 } 2629 2630 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2631 // The null DIType is the void type. Don't try to hash it. 2632 if (!Ty) 2633 return TypeIndex::Void(); 2634 2635 // Check if we've already translated this type. Don't try to do a 2636 // get-or-create style insertion that caches the hash lookup across the 2637 // lowerType call. It will update the TypeIndices map. 2638 auto I = TypeIndices.find({Ty, ClassTy}); 2639 if (I != TypeIndices.end()) 2640 return I->second; 2641 2642 TypeLoweringScope S(*this); 2643 TypeIndex TI = lowerType(Ty, ClassTy); 2644 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2645 } 2646 2647 codeview::TypeIndex 2648 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2649 const DISubroutineType *SubroutineTy) { 2650 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2651 "this type must be a pointer type"); 2652 2653 PointerOptions Options = PointerOptions::None; 2654 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2655 Options = PointerOptions::LValueRefThisPointer; 2656 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2657 Options = PointerOptions::RValueRefThisPointer; 2658 2659 // Check if we've already translated this type. If there is no ref qualifier 2660 // on the function then we look up this pointer type with no associated class 2661 // so that the TypeIndex for the this pointer can be shared with the type 2662 // index for other pointers to this class type. If there is a ref qualifier 2663 // then we lookup the pointer using the subroutine as the parent type. 2664 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2665 if (I != TypeIndices.end()) 2666 return I->second; 2667 2668 TypeLoweringScope S(*this); 2669 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2670 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2671 } 2672 2673 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2674 PointerRecord PR(getTypeIndex(Ty), 2675 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2676 : PointerKind::Near32, 2677 PointerMode::LValueReference, PointerOptions::None, 2678 Ty->getSizeInBits() / 8); 2679 return TypeTable.writeLeafType(PR); 2680 } 2681 2682 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2683 // The null DIType is the void type. Don't try to hash it. 2684 if (!Ty) 2685 return TypeIndex::Void(); 2686 2687 // Look through typedefs when getting the complete type index. Call 2688 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2689 // emitted only once. 2690 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2691 (void)getTypeIndex(Ty); 2692 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2693 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2694 2695 // If this is a non-record type, the complete type index is the same as the 2696 // normal type index. Just call getTypeIndex. 2697 switch (Ty->getTag()) { 2698 case dwarf::DW_TAG_class_type: 2699 case dwarf::DW_TAG_structure_type: 2700 case dwarf::DW_TAG_union_type: 2701 break; 2702 default: 2703 return getTypeIndex(Ty); 2704 } 2705 2706 const auto *CTy = cast<DICompositeType>(Ty); 2707 2708 TypeLoweringScope S(*this); 2709 2710 // Make sure the forward declaration is emitted first. It's unclear if this 2711 // is necessary, but MSVC does it, and we should follow suit until we can show 2712 // otherwise. 2713 // We only emit a forward declaration for named types. 2714 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2715 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2716 2717 // Just use the forward decl if we don't have complete type info. This 2718 // might happen if the frontend is using modules and expects the complete 2719 // definition to be emitted elsewhere. 2720 if (CTy->isForwardDecl()) 2721 return FwdDeclTI; 2722 } 2723 2724 // Check if we've already translated the complete record type. 2725 // Insert the type with a null TypeIndex to signify that the type is currently 2726 // being lowered. 2727 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2728 if (!InsertResult.second) 2729 return InsertResult.first->second; 2730 2731 TypeIndex TI; 2732 switch (CTy->getTag()) { 2733 case dwarf::DW_TAG_class_type: 2734 case dwarf::DW_TAG_structure_type: 2735 TI = lowerCompleteTypeClass(CTy); 2736 break; 2737 case dwarf::DW_TAG_union_type: 2738 TI = lowerCompleteTypeUnion(CTy); 2739 break; 2740 default: 2741 llvm_unreachable("not a record"); 2742 } 2743 2744 // Update the type index associated with this CompositeType. This cannot 2745 // use the 'InsertResult' iterator above because it is potentially 2746 // invalidated by map insertions which can occur while lowering the class 2747 // type above. 2748 CompleteTypeIndices[CTy] = TI; 2749 return TI; 2750 } 2751 2752 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2753 /// and do this until fixpoint, as each complete record type typically 2754 /// references 2755 /// many other record types. 2756 void CodeViewDebug::emitDeferredCompleteTypes() { 2757 SmallVector<const DICompositeType *, 4> TypesToEmit; 2758 while (!DeferredCompleteTypes.empty()) { 2759 std::swap(DeferredCompleteTypes, TypesToEmit); 2760 for (const DICompositeType *RecordTy : TypesToEmit) 2761 getCompleteTypeIndex(RecordTy); 2762 TypesToEmit.clear(); 2763 } 2764 } 2765 2766 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2767 ArrayRef<LocalVariable> Locals) { 2768 // Get the sorted list of parameters and emit them first. 2769 SmallVector<const LocalVariable *, 6> Params; 2770 for (const LocalVariable &L : Locals) 2771 if (L.DIVar->isParameter()) 2772 Params.push_back(&L); 2773 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2774 return L->DIVar->getArg() < R->DIVar->getArg(); 2775 }); 2776 for (const LocalVariable *L : Params) 2777 emitLocalVariable(FI, *L); 2778 2779 // Next emit all non-parameters in the order that we found them. 2780 for (const LocalVariable &L : Locals) 2781 if (!L.DIVar->isParameter()) 2782 emitLocalVariable(FI, L); 2783 } 2784 2785 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2786 const LocalVariable &Var) { 2787 // LocalSym record, see SymbolRecord.h for more info. 2788 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2789 2790 LocalSymFlags Flags = LocalSymFlags::None; 2791 if (Var.DIVar->isParameter()) 2792 Flags |= LocalSymFlags::IsParameter; 2793 if (Var.DefRanges.empty()) 2794 Flags |= LocalSymFlags::IsOptimizedOut; 2795 2796 OS.AddComment("TypeIndex"); 2797 TypeIndex TI = Var.UseReferenceType 2798 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2799 : getCompleteTypeIndex(Var.DIVar->getType()); 2800 OS.emitInt32(TI.getIndex()); 2801 OS.AddComment("Flags"); 2802 OS.emitInt16(static_cast<uint16_t>(Flags)); 2803 // Truncate the name so we won't overflow the record length field. 2804 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2805 endSymbolRecord(LocalEnd); 2806 2807 // Calculate the on disk prefix of the appropriate def range record. The 2808 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2809 // should be big enough to hold all forms without memory allocation. 2810 SmallString<20> BytePrefix; 2811 for (const auto &Pair : Var.DefRanges) { 2812 LocalVarDef DefRange = Pair.first; 2813 const auto &Ranges = Pair.second; 2814 BytePrefix.clear(); 2815 if (DefRange.InMemory) { 2816 int Offset = DefRange.DataOffset; 2817 unsigned Reg = DefRange.CVRegister; 2818 2819 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2820 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2821 // instead. In frames without stack realignment, $T0 will be the CFA. 2822 if (RegisterId(Reg) == RegisterId::ESP) { 2823 Reg = unsigned(RegisterId::VFRAME); 2824 Offset += FI.OffsetAdjustment; 2825 } 2826 2827 // If we can use the chosen frame pointer for the frame and this isn't a 2828 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2829 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2830 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2831 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2832 (bool(Flags & LocalSymFlags::IsParameter) 2833 ? (EncFP == FI.EncodedParamFramePtrReg) 2834 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2835 DefRangeFramePointerRelHeader DRHdr; 2836 DRHdr.Offset = Offset; 2837 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2838 } else { 2839 uint16_t RegRelFlags = 0; 2840 if (DefRange.IsSubfield) { 2841 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2842 (DefRange.StructOffset 2843 << DefRangeRegisterRelSym::OffsetInParentShift); 2844 } 2845 DefRangeRegisterRelHeader DRHdr; 2846 DRHdr.Register = Reg; 2847 DRHdr.Flags = RegRelFlags; 2848 DRHdr.BasePointerOffset = Offset; 2849 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2850 } 2851 } else { 2852 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2853 if (DefRange.IsSubfield) { 2854 DefRangeSubfieldRegisterHeader DRHdr; 2855 DRHdr.Register = DefRange.CVRegister; 2856 DRHdr.MayHaveNoName = 0; 2857 DRHdr.OffsetInParent = DefRange.StructOffset; 2858 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2859 } else { 2860 DefRangeRegisterHeader DRHdr; 2861 DRHdr.Register = DefRange.CVRegister; 2862 DRHdr.MayHaveNoName = 0; 2863 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2864 } 2865 } 2866 } 2867 } 2868 2869 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2870 const FunctionInfo& FI) { 2871 for (LexicalBlock *Block : Blocks) 2872 emitLexicalBlock(*Block, FI); 2873 } 2874 2875 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2876 /// lexical block scope. 2877 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2878 const FunctionInfo& FI) { 2879 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2880 OS.AddComment("PtrParent"); 2881 OS.emitInt32(0); // PtrParent 2882 OS.AddComment("PtrEnd"); 2883 OS.emitInt32(0); // PtrEnd 2884 OS.AddComment("Code size"); 2885 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2886 OS.AddComment("Function section relative address"); 2887 OS.emitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2888 OS.AddComment("Function section index"); 2889 OS.emitCOFFSectionIndex(FI.Begin); // Func Symbol 2890 OS.AddComment("Lexical block name"); 2891 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2892 endSymbolRecord(RecordEnd); 2893 2894 // Emit variables local to this lexical block. 2895 emitLocalVariableList(FI, Block.Locals); 2896 emitGlobalVariableList(Block.Globals); 2897 2898 // Emit lexical blocks contained within this block. 2899 emitLexicalBlockList(Block.Children, FI); 2900 2901 // Close the lexical block scope. 2902 emitEndSymbolRecord(SymbolKind::S_END); 2903 } 2904 2905 /// Convenience routine for collecting lexical block information for a list 2906 /// of lexical scopes. 2907 void CodeViewDebug::collectLexicalBlockInfo( 2908 SmallVectorImpl<LexicalScope *> &Scopes, 2909 SmallVectorImpl<LexicalBlock *> &Blocks, 2910 SmallVectorImpl<LocalVariable> &Locals, 2911 SmallVectorImpl<CVGlobalVariable> &Globals) { 2912 for (LexicalScope *Scope : Scopes) 2913 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2914 } 2915 2916 /// Populate the lexical blocks and local variable lists of the parent with 2917 /// information about the specified lexical scope. 2918 void CodeViewDebug::collectLexicalBlockInfo( 2919 LexicalScope &Scope, 2920 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2921 SmallVectorImpl<LocalVariable> &ParentLocals, 2922 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2923 if (Scope.isAbstractScope()) 2924 return; 2925 2926 // Gather information about the lexical scope including local variables, 2927 // global variables, and address ranges. 2928 bool IgnoreScope = false; 2929 auto LI = ScopeVariables.find(&Scope); 2930 SmallVectorImpl<LocalVariable> *Locals = 2931 LI != ScopeVariables.end() ? &LI->second : nullptr; 2932 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2933 SmallVectorImpl<CVGlobalVariable> *Globals = 2934 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2935 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2936 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2937 2938 // Ignore lexical scopes which do not contain variables. 2939 if (!Locals && !Globals) 2940 IgnoreScope = true; 2941 2942 // Ignore lexical scopes which are not lexical blocks. 2943 if (!DILB) 2944 IgnoreScope = true; 2945 2946 // Ignore scopes which have too many address ranges to represent in the 2947 // current CodeView format or do not have a valid address range. 2948 // 2949 // For lexical scopes with multiple address ranges you may be tempted to 2950 // construct a single range covering every instruction where the block is 2951 // live and everything in between. Unfortunately, Visual Studio only 2952 // displays variables from the first matching lexical block scope. If the 2953 // first lexical block contains exception handling code or cold code which 2954 // is moved to the bottom of the routine creating a single range covering 2955 // nearly the entire routine, then it will hide all other lexical blocks 2956 // and the variables they contain. 2957 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2958 IgnoreScope = true; 2959 2960 if (IgnoreScope) { 2961 // This scope can be safely ignored and eliminating it will reduce the 2962 // size of the debug information. Be sure to collect any variable and scope 2963 // information from the this scope or any of its children and collapse them 2964 // into the parent scope. 2965 if (Locals) 2966 ParentLocals.append(Locals->begin(), Locals->end()); 2967 if (Globals) 2968 ParentGlobals.append(Globals->begin(), Globals->end()); 2969 collectLexicalBlockInfo(Scope.getChildren(), 2970 ParentBlocks, 2971 ParentLocals, 2972 ParentGlobals); 2973 return; 2974 } 2975 2976 // Create a new CodeView lexical block for this lexical scope. If we've 2977 // seen this DILexicalBlock before then the scope tree is malformed and 2978 // we can handle this gracefully by not processing it a second time. 2979 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2980 if (!BlockInsertion.second) 2981 return; 2982 2983 // Create a lexical block containing the variables and collect the the 2984 // lexical block information for the children. 2985 const InsnRange &Range = Ranges.front(); 2986 assert(Range.first && Range.second); 2987 LexicalBlock &Block = BlockInsertion.first->second; 2988 Block.Begin = getLabelBeforeInsn(Range.first); 2989 Block.End = getLabelAfterInsn(Range.second); 2990 assert(Block.Begin && "missing label for scope begin"); 2991 assert(Block.End && "missing label for scope end"); 2992 Block.Name = DILB->getName(); 2993 if (Locals) 2994 Block.Locals = std::move(*Locals); 2995 if (Globals) 2996 Block.Globals = std::move(*Globals); 2997 ParentBlocks.push_back(&Block); 2998 collectLexicalBlockInfo(Scope.getChildren(), 2999 Block.Children, 3000 Block.Locals, 3001 Block.Globals); 3002 } 3003 3004 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3005 const Function &GV = MF->getFunction(); 3006 assert(FnDebugInfo.count(&GV)); 3007 assert(CurFn == FnDebugInfo[&GV].get()); 3008 3009 collectVariableInfo(GV.getSubprogram()); 3010 3011 // Build the lexical block structure to emit for this routine. 3012 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3013 collectLexicalBlockInfo(*CFS, 3014 CurFn->ChildBlocks, 3015 CurFn->Locals, 3016 CurFn->Globals); 3017 3018 // Clear the scope and variable information from the map which will not be 3019 // valid after we have finished processing this routine. This also prepares 3020 // the map for the subsequent routine. 3021 ScopeVariables.clear(); 3022 3023 // Don't emit anything if we don't have any line tables. 3024 // Thunks are compiler-generated and probably won't have source correlation. 3025 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3026 FnDebugInfo.erase(&GV); 3027 CurFn = nullptr; 3028 return; 3029 } 3030 3031 // Find heap alloc sites and add to list. 3032 for (const auto &MBB : *MF) { 3033 for (const auto &MI : MBB) { 3034 if (MDNode *MD = MI.getHeapAllocMarker()) { 3035 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3036 getLabelAfterInsn(&MI), 3037 dyn_cast<DIType>(MD))); 3038 } 3039 } 3040 } 3041 3042 CurFn->Annotations = MF->getCodeViewAnnotations(); 3043 3044 CurFn->End = Asm->getFunctionEnd(); 3045 3046 CurFn = nullptr; 3047 } 3048 3049 // Usable locations are valid with non-zero line numbers. A line number of zero 3050 // corresponds to optimized code that doesn't have a distinct source location. 3051 // In this case, we try to use the previous or next source location depending on 3052 // the context. 3053 static bool isUsableDebugLoc(DebugLoc DL) { 3054 return DL && DL.getLine() != 0; 3055 } 3056 3057 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3058 DebugHandlerBase::beginInstruction(MI); 3059 3060 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3061 if (!Asm || !CurFn || MI->isDebugInstr() || 3062 MI->getFlag(MachineInstr::FrameSetup)) 3063 return; 3064 3065 // If the first instruction of a new MBB has no location, find the first 3066 // instruction with a location and use that. 3067 DebugLoc DL = MI->getDebugLoc(); 3068 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3069 for (const auto &NextMI : *MI->getParent()) { 3070 if (NextMI.isDebugInstr()) 3071 continue; 3072 DL = NextMI.getDebugLoc(); 3073 if (isUsableDebugLoc(DL)) 3074 break; 3075 } 3076 // FIXME: Handle the case where the BB has no valid locations. This would 3077 // probably require doing a real dataflow analysis. 3078 } 3079 PrevInstBB = MI->getParent(); 3080 3081 // If we still don't have a debug location, don't record a location. 3082 if (!isUsableDebugLoc(DL)) 3083 return; 3084 3085 maybeRecordLocation(DL, Asm->MF); 3086 } 3087 3088 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3089 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3090 *EndLabel = MMI->getContext().createTempSymbol(); 3091 OS.emitInt32(unsigned(Kind)); 3092 OS.AddComment("Subsection size"); 3093 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3094 OS.emitLabel(BeginLabel); 3095 return EndLabel; 3096 } 3097 3098 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3099 OS.emitLabel(EndLabel); 3100 // Every subsection must be aligned to a 4-byte boundary. 3101 OS.emitValueToAlignment(4); 3102 } 3103 3104 static StringRef getSymbolName(SymbolKind SymKind) { 3105 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3106 if (EE.Value == SymKind) 3107 return EE.Name; 3108 return ""; 3109 } 3110 3111 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3112 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3113 *EndLabel = MMI->getContext().createTempSymbol(); 3114 OS.AddComment("Record length"); 3115 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3116 OS.emitLabel(BeginLabel); 3117 if (OS.isVerboseAsm()) 3118 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3119 OS.emitInt16(unsigned(SymKind)); 3120 return EndLabel; 3121 } 3122 3123 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3124 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3125 // an extra copy of every symbol record in LLD. This increases object file 3126 // size by less than 1% in the clang build, and is compatible with the Visual 3127 // C++ linker. 3128 OS.emitValueToAlignment(4); 3129 OS.emitLabel(SymEnd); 3130 } 3131 3132 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3133 OS.AddComment("Record length"); 3134 OS.emitInt16(2); 3135 if (OS.isVerboseAsm()) 3136 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3137 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3138 } 3139 3140 void CodeViewDebug::emitDebugInfoForUDTs( 3141 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3142 #ifndef NDEBUG 3143 size_t OriginalSize = UDTs.size(); 3144 #endif 3145 for (const auto &UDT : UDTs) { 3146 const DIType *T = UDT.second; 3147 assert(shouldEmitUdt(T)); 3148 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3149 OS.AddComment("Type"); 3150 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3151 assert(OriginalSize == UDTs.size() && 3152 "getCompleteTypeIndex found new UDTs!"); 3153 emitNullTerminatedSymbolName(OS, UDT.first); 3154 endSymbolRecord(UDTRecordEnd); 3155 } 3156 } 3157 3158 void CodeViewDebug::collectGlobalVariableInfo() { 3159 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3160 GlobalMap; 3161 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3162 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3163 GV.getDebugInfo(GVEs); 3164 for (const auto *GVE : GVEs) 3165 GlobalMap[GVE] = &GV; 3166 } 3167 3168 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3169 for (const MDNode *Node : CUs->operands()) { 3170 const auto *CU = cast<DICompileUnit>(Node); 3171 for (const auto *GVE : CU->getGlobalVariables()) { 3172 const DIGlobalVariable *DIGV = GVE->getVariable(); 3173 const DIExpression *DIE = GVE->getExpression(); 3174 // Don't emit string literals in CodeView, as the only useful parts are 3175 // generally the filename and line number, which isn't possible to output 3176 // in CodeView. String literals should be the only unnamed GlobalVariable 3177 // with debug info. 3178 if (DIGV->getName().empty()) continue; 3179 3180 if ((DIE->getNumElements() == 2) && 3181 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3182 // Record the constant offset for the variable. 3183 // 3184 // A Fortran common block uses this idiom to encode the offset 3185 // of a variable from the common block's starting address. 3186 CVGlobalVariableOffsets.insert( 3187 std::make_pair(DIGV, DIE->getElement(1))); 3188 3189 // Emit constant global variables in a global symbol section. 3190 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3191 CVGlobalVariable CVGV = {DIGV, DIE}; 3192 GlobalVariables.emplace_back(std::move(CVGV)); 3193 } 3194 3195 const auto *GV = GlobalMap.lookup(GVE); 3196 if (!GV || GV->isDeclarationForLinker()) 3197 continue; 3198 3199 DIScope *Scope = DIGV->getScope(); 3200 SmallVector<CVGlobalVariable, 1> *VariableList; 3201 if (Scope && isa<DILocalScope>(Scope)) { 3202 // Locate a global variable list for this scope, creating one if 3203 // necessary. 3204 auto Insertion = ScopeGlobals.insert( 3205 {Scope, std::unique_ptr<GlobalVariableList>()}); 3206 if (Insertion.second) 3207 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3208 VariableList = Insertion.first->second.get(); 3209 } else if (GV->hasComdat()) 3210 // Emit this global variable into a COMDAT section. 3211 VariableList = &ComdatVariables; 3212 else 3213 // Emit this global variable in a single global symbol section. 3214 VariableList = &GlobalVariables; 3215 CVGlobalVariable CVGV = {DIGV, GV}; 3216 VariableList->emplace_back(std::move(CVGV)); 3217 } 3218 } 3219 } 3220 3221 void CodeViewDebug::collectDebugInfoForGlobals() { 3222 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3223 const DIGlobalVariable *DIGV = CVGV.DIGV; 3224 const DIScope *Scope = DIGV->getScope(); 3225 getCompleteTypeIndex(DIGV->getType()); 3226 getFullyQualifiedName(Scope, DIGV->getName()); 3227 } 3228 3229 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3230 const DIGlobalVariable *DIGV = CVGV.DIGV; 3231 const DIScope *Scope = DIGV->getScope(); 3232 getCompleteTypeIndex(DIGV->getType()); 3233 getFullyQualifiedName(Scope, DIGV->getName()); 3234 } 3235 } 3236 3237 void CodeViewDebug::emitDebugInfoForGlobals() { 3238 // First, emit all globals that are not in a comdat in a single symbol 3239 // substream. MSVC doesn't like it if the substream is empty, so only open 3240 // it if we have at least one global to emit. 3241 switchToDebugSectionForSymbol(nullptr); 3242 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3243 OS.AddComment("Symbol subsection for globals"); 3244 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3245 emitGlobalVariableList(GlobalVariables); 3246 emitStaticConstMemberList(); 3247 endCVSubsection(EndLabel); 3248 } 3249 3250 // Second, emit each global that is in a comdat into its own .debug$S 3251 // section along with its own symbol substream. 3252 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3253 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3254 MCSymbol *GVSym = Asm->getSymbol(GV); 3255 OS.AddComment("Symbol subsection for " + 3256 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3257 switchToDebugSectionForSymbol(GVSym); 3258 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3259 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3260 emitDebugInfoForGlobal(CVGV); 3261 endCVSubsection(EndLabel); 3262 } 3263 } 3264 3265 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3266 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3267 for (const MDNode *Node : CUs->operands()) { 3268 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3269 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3270 getTypeIndex(RT); 3271 // FIXME: Add to global/local DTU list. 3272 } 3273 } 3274 } 3275 } 3276 3277 // Emit each global variable in the specified array. 3278 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3279 for (const CVGlobalVariable &CVGV : Globals) { 3280 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3281 emitDebugInfoForGlobal(CVGV); 3282 } 3283 } 3284 3285 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3286 const std::string &QualifiedName) { 3287 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3288 OS.AddComment("Type"); 3289 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3290 3291 OS.AddComment("Value"); 3292 3293 // Encoded integers shouldn't need more than 10 bytes. 3294 uint8_t Data[10]; 3295 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3296 CodeViewRecordIO IO(Writer); 3297 cantFail(IO.mapEncodedInteger(Value)); 3298 StringRef SRef((char *)Data, Writer.getOffset()); 3299 OS.emitBinaryData(SRef); 3300 3301 OS.AddComment("Name"); 3302 emitNullTerminatedSymbolName(OS, QualifiedName); 3303 endSymbolRecord(SConstantEnd); 3304 } 3305 3306 void CodeViewDebug::emitStaticConstMemberList() { 3307 for (const DIDerivedType *DTy : StaticConstMembers) { 3308 const DIScope *Scope = DTy->getScope(); 3309 3310 APSInt Value; 3311 if (const ConstantInt *CI = 3312 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3313 Value = APSInt(CI->getValue(), 3314 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3315 else if (const ConstantFP *CFP = 3316 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3317 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3318 else 3319 llvm_unreachable("cannot emit a constant without a value"); 3320 3321 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3322 getFullyQualifiedName(Scope, DTy->getName())); 3323 } 3324 } 3325 3326 static bool isFloatDIType(const DIType *Ty) { 3327 if (isa<DICompositeType>(Ty)) 3328 return false; 3329 3330 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3331 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3332 if (T == dwarf::DW_TAG_pointer_type || 3333 T == dwarf::DW_TAG_ptr_to_member_type || 3334 T == dwarf::DW_TAG_reference_type || 3335 T == dwarf::DW_TAG_rvalue_reference_type) 3336 return false; 3337 assert(DTy->getBaseType() && "Expected valid base type"); 3338 return isFloatDIType(DTy->getBaseType()); 3339 } 3340 3341 auto *BTy = cast<DIBasicType>(Ty); 3342 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3343 } 3344 3345 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3346 const DIGlobalVariable *DIGV = CVGV.DIGV; 3347 3348 const DIScope *Scope = DIGV->getScope(); 3349 // For static data members, get the scope from the declaration. 3350 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3351 DIGV->getRawStaticDataMemberDeclaration())) 3352 Scope = MemberDecl->getScope(); 3353 // For Fortran, the scoping portion is elided in its name so that we can 3354 // reference the variable in the command line of the VS debugger. 3355 std::string QualifiedName = 3356 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3357 : getFullyQualifiedName(Scope, DIGV->getName()); 3358 3359 if (const GlobalVariable *GV = 3360 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3361 // DataSym record, see SymbolRecord.h for more info. Thread local data 3362 // happens to have the same format as global data. 3363 MCSymbol *GVSym = Asm->getSymbol(GV); 3364 SymbolKind DataSym = GV->isThreadLocal() 3365 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3366 : SymbolKind::S_GTHREAD32) 3367 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3368 : SymbolKind::S_GDATA32); 3369 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3370 OS.AddComment("Type"); 3371 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3372 OS.AddComment("DataOffset"); 3373 3374 uint64_t Offset = 0; 3375 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3376 // Use the offset seen while collecting info on globals. 3377 Offset = CVGlobalVariableOffsets[DIGV]; 3378 OS.emitCOFFSecRel32(GVSym, Offset); 3379 3380 OS.AddComment("Segment"); 3381 OS.emitCOFFSectionIndex(GVSym); 3382 OS.AddComment("Name"); 3383 const unsigned LengthOfDataRecord = 12; 3384 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3385 endSymbolRecord(DataEnd); 3386 } else { 3387 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3388 assert(DIE->isConstant() && 3389 "Global constant variables must contain a constant expression."); 3390 3391 // Use unsigned for floats. 3392 bool isUnsigned = isFloatDIType(DIGV->getType()) 3393 ? true 3394 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3395 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3396 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3397 } 3398 } 3399