1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/None.h" 17 #include "llvm/ADT/Optional.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallString.h" 20 #include "llvm/ADT/StringRef.h" 21 #include "llvm/ADT/TinyPtrVector.h" 22 #include "llvm/ADT/Triple.h" 23 #include "llvm/ADT/Twine.h" 24 #include "llvm/BinaryFormat/COFF.h" 25 #include "llvm/BinaryFormat/Dwarf.h" 26 #include "llvm/CodeGen/AsmPrinter.h" 27 #include "llvm/CodeGen/LexicalScopes.h" 28 #include "llvm/CodeGen/MachineFrameInfo.h" 29 #include "llvm/CodeGen/MachineFunction.h" 30 #include "llvm/CodeGen/MachineInstr.h" 31 #include "llvm/CodeGen/MachineModuleInfo.h" 32 #include "llvm/CodeGen/MachineOperand.h" 33 #include "llvm/CodeGen/TargetFrameLowering.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/TargetSubtargetInfo.h" 36 #include "llvm/Config/llvm-config.h" 37 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 38 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 39 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 40 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 41 #include "llvm/DebugInfo/CodeView/EnumTables.h" 42 #include "llvm/DebugInfo/CodeView/Line.h" 43 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 44 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 45 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 46 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 47 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 48 #include "llvm/IR/Constants.h" 49 #include "llvm/IR/DataLayout.h" 50 #include "llvm/IR/DebugInfoMetadata.h" 51 #include "llvm/IR/Function.h" 52 #include "llvm/IR/GlobalValue.h" 53 #include "llvm/IR/GlobalVariable.h" 54 #include "llvm/IR/Metadata.h" 55 #include "llvm/IR/Module.h" 56 #include "llvm/MC/MCAsmInfo.h" 57 #include "llvm/MC/MCContext.h" 58 #include "llvm/MC/MCSectionCOFF.h" 59 #include "llvm/MC/MCStreamer.h" 60 #include "llvm/MC/MCSymbol.h" 61 #include "llvm/Support/BinaryByteStream.h" 62 #include "llvm/Support/BinaryStreamReader.h" 63 #include "llvm/Support/BinaryStreamWriter.h" 64 #include "llvm/Support/Casting.h" 65 #include "llvm/Support/CommandLine.h" 66 #include "llvm/Support/Endian.h" 67 #include "llvm/Support/Error.h" 68 #include "llvm/Support/ErrorHandling.h" 69 #include "llvm/Support/FormatVariadic.h" 70 #include "llvm/Support/Path.h" 71 #include "llvm/Support/SMLoc.h" 72 #include "llvm/Support/ScopedPrinter.h" 73 #include "llvm/Target/TargetLoweringObjectFile.h" 74 #include "llvm/Target/TargetMachine.h" 75 #include <algorithm> 76 #include <cassert> 77 #include <cctype> 78 #include <cstddef> 79 #include <iterator> 80 #include <limits> 81 82 using namespace llvm; 83 using namespace llvm::codeview; 84 85 namespace { 86 class CVMCAdapter : public CodeViewRecordStreamer { 87 public: 88 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 89 : OS(&OS), TypeTable(TypeTable) {} 90 91 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 92 93 void emitIntValue(uint64_t Value, unsigned Size) override { 94 OS->emitIntValueInHex(Value, Size); 95 } 96 97 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 98 99 void AddComment(const Twine &T) override { OS->AddComment(T); } 100 101 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 102 103 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 104 105 std::string getTypeName(TypeIndex TI) override { 106 std::string TypeName; 107 if (!TI.isNoneType()) { 108 if (TI.isSimple()) 109 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 110 else 111 TypeName = std::string(TypeTable.getTypeName(TI)); 112 } 113 return TypeName; 114 } 115 116 private: 117 MCStreamer *OS = nullptr; 118 TypeCollection &TypeTable; 119 }; 120 } // namespace 121 122 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 123 switch (Type) { 124 case Triple::ArchType::x86: 125 return CPUType::Pentium3; 126 case Triple::ArchType::x86_64: 127 return CPUType::X64; 128 case Triple::ArchType::thumb: 129 // LLVM currently doesn't support Windows CE and so thumb 130 // here is indiscriminately mapped to ARMNT specifically. 131 return CPUType::ARMNT; 132 case Triple::ArchType::aarch64: 133 return CPUType::ARM64; 134 default: 135 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 136 } 137 } 138 139 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 140 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 141 142 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 143 std::string &Filepath = FileToFilepathMap[File]; 144 if (!Filepath.empty()) 145 return Filepath; 146 147 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 148 149 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 150 // it textually because one of the path components could be a symlink. 151 if (Dir.startswith("/") || Filename.startswith("/")) { 152 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 153 return Filename; 154 Filepath = std::string(Dir); 155 if (Dir.back() != '/') 156 Filepath += '/'; 157 Filepath += Filename; 158 return Filepath; 159 } 160 161 // Clang emits directory and relative filename info into the IR, but CodeView 162 // operates on full paths. We could change Clang to emit full paths too, but 163 // that would increase the IR size and probably not needed for other users. 164 // For now, just concatenate and canonicalize the path here. 165 if (Filename.find(':') == 1) 166 Filepath = std::string(Filename); 167 else 168 Filepath = (Dir + "\\" + Filename).str(); 169 170 // Canonicalize the path. We have to do it textually because we may no longer 171 // have access the file in the filesystem. 172 // First, replace all slashes with backslashes. 173 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 174 175 // Remove all "\.\" with "\". 176 size_t Cursor = 0; 177 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 178 Filepath.erase(Cursor, 2); 179 180 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 181 // path should be well-formatted, e.g. start with a drive letter, etc. 182 Cursor = 0; 183 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 184 // Something's wrong if the path starts with "\..\", abort. 185 if (Cursor == 0) 186 break; 187 188 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 189 if (PrevSlash == std::string::npos) 190 // Something's wrong, abort. 191 break; 192 193 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 194 // The next ".." might be following the one we've just erased. 195 Cursor = PrevSlash; 196 } 197 198 // Remove all duplicate backslashes. 199 Cursor = 0; 200 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 201 Filepath.erase(Cursor, 1); 202 203 return Filepath; 204 } 205 206 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 207 StringRef FullPath = getFullFilepath(F); 208 unsigned NextId = FileIdMap.size() + 1; 209 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 210 if (Insertion.second) { 211 // We have to compute the full filepath and emit a .cv_file directive. 212 ArrayRef<uint8_t> ChecksumAsBytes; 213 FileChecksumKind CSKind = FileChecksumKind::None; 214 if (F->getChecksum()) { 215 std::string Checksum = fromHex(F->getChecksum()->Value); 216 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 217 memcpy(CKMem, Checksum.data(), Checksum.size()); 218 ChecksumAsBytes = ArrayRef<uint8_t>( 219 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 220 switch (F->getChecksum()->Kind) { 221 case DIFile::CSK_MD5: 222 CSKind = FileChecksumKind::MD5; 223 break; 224 case DIFile::CSK_SHA1: 225 CSKind = FileChecksumKind::SHA1; 226 break; 227 case DIFile::CSK_SHA256: 228 CSKind = FileChecksumKind::SHA256; 229 break; 230 } 231 } 232 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 233 static_cast<unsigned>(CSKind)); 234 (void)Success; 235 assert(Success && ".cv_file directive failed"); 236 } 237 return Insertion.first->second; 238 } 239 240 CodeViewDebug::InlineSite & 241 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 242 const DISubprogram *Inlinee) { 243 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 244 InlineSite *Site = &SiteInsertion.first->second; 245 if (SiteInsertion.second) { 246 unsigned ParentFuncId = CurFn->FuncId; 247 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 248 ParentFuncId = 249 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 250 .SiteFuncId; 251 252 Site->SiteFuncId = NextFuncId++; 253 OS.EmitCVInlineSiteIdDirective( 254 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 255 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 256 Site->Inlinee = Inlinee; 257 InlinedSubprograms.insert(Inlinee); 258 getFuncIdForSubprogram(Inlinee); 259 } 260 return *Site; 261 } 262 263 static StringRef getPrettyScopeName(const DIScope *Scope) { 264 StringRef ScopeName = Scope->getName(); 265 if (!ScopeName.empty()) 266 return ScopeName; 267 268 switch (Scope->getTag()) { 269 case dwarf::DW_TAG_enumeration_type: 270 case dwarf::DW_TAG_class_type: 271 case dwarf::DW_TAG_structure_type: 272 case dwarf::DW_TAG_union_type: 273 return "<unnamed-tag>"; 274 case dwarf::DW_TAG_namespace: 275 return "`anonymous namespace'"; 276 default: 277 return StringRef(); 278 } 279 } 280 281 const DISubprogram *CodeViewDebug::collectParentScopeNames( 282 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 283 const DISubprogram *ClosestSubprogram = nullptr; 284 while (Scope != nullptr) { 285 if (ClosestSubprogram == nullptr) 286 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 287 288 // If a type appears in a scope chain, make sure it gets emitted. The 289 // frontend will be responsible for deciding if this should be a forward 290 // declaration or a complete type. 291 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 292 DeferredCompleteTypes.push_back(Ty); 293 294 StringRef ScopeName = getPrettyScopeName(Scope); 295 if (!ScopeName.empty()) 296 QualifiedNameComponents.push_back(ScopeName); 297 Scope = Scope->getScope(); 298 } 299 return ClosestSubprogram; 300 } 301 302 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 303 StringRef TypeName) { 304 std::string FullyQualifiedName; 305 for (StringRef QualifiedNameComponent : 306 llvm::reverse(QualifiedNameComponents)) { 307 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 308 FullyQualifiedName.append("::"); 309 } 310 FullyQualifiedName.append(std::string(TypeName)); 311 return FullyQualifiedName; 312 } 313 314 struct CodeViewDebug::TypeLoweringScope { 315 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 316 ~TypeLoweringScope() { 317 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 318 // inner TypeLoweringScopes don't attempt to emit deferred types. 319 if (CVD.TypeEmissionLevel == 1) 320 CVD.emitDeferredCompleteTypes(); 321 --CVD.TypeEmissionLevel; 322 } 323 CodeViewDebug &CVD; 324 }; 325 326 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 327 StringRef Name) { 328 // Ensure types in the scope chain are emitted as soon as possible. 329 // This can create otherwise a situation where S_UDTs are emitted while 330 // looping in emitDebugInfoForUDTs. 331 TypeLoweringScope S(*this); 332 SmallVector<StringRef, 5> QualifiedNameComponents; 333 collectParentScopeNames(Scope, QualifiedNameComponents); 334 return formatNestedName(QualifiedNameComponents, Name); 335 } 336 337 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 338 const DIScope *Scope = Ty->getScope(); 339 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 340 } 341 342 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 343 // No scope means global scope and that uses the zero index. 344 // 345 // We also use zero index when the scope is a DISubprogram 346 // to suppress the emission of LF_STRING_ID for the function, 347 // which can trigger a link-time error with the linker in 348 // VS2019 version 16.11.2 or newer. 349 // Note, however, skipping the debug info emission for the DISubprogram 350 // is a temporary fix. The root issue here is that we need to figure out 351 // the proper way to encode a function nested in another function 352 // (as introduced by the Fortran 'contains' keyword) in CodeView. 353 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 354 return TypeIndex(); 355 356 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 357 358 // Check if we've already translated this scope. 359 auto I = TypeIndices.find({Scope, nullptr}); 360 if (I != TypeIndices.end()) 361 return I->second; 362 363 // Build the fully qualified name of the scope. 364 std::string ScopeName = getFullyQualifiedName(Scope); 365 StringIdRecord SID(TypeIndex(), ScopeName); 366 auto TI = TypeTable.writeLeafType(SID); 367 return recordTypeIndexForDINode(Scope, TI); 368 } 369 370 static StringRef removeTemplateArgs(StringRef Name) { 371 // Remove template args from the display name. Assume that the template args 372 // are the last thing in the name. 373 if (Name.empty() || Name.back() != '>') 374 return Name; 375 376 int OpenBrackets = 0; 377 for (int i = Name.size() - 1; i >= 0; --i) { 378 if (Name[i] == '>') 379 ++OpenBrackets; 380 else if (Name[i] == '<') { 381 --OpenBrackets; 382 if (OpenBrackets == 0) 383 return Name.substr(0, i); 384 } 385 } 386 return Name; 387 } 388 389 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 390 assert(SP); 391 392 // Check if we've already translated this subprogram. 393 auto I = TypeIndices.find({SP, nullptr}); 394 if (I != TypeIndices.end()) 395 return I->second; 396 397 // The display name includes function template arguments. Drop them to match 398 // MSVC. We need to have the template arguments in the DISubprogram name 399 // because they are used in other symbol records, such as S_GPROC32_IDs. 400 StringRef DisplayName = removeTemplateArgs(SP->getName()); 401 402 const DIScope *Scope = SP->getScope(); 403 TypeIndex TI; 404 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 405 // If the scope is a DICompositeType, then this must be a method. Member 406 // function types take some special handling, and require access to the 407 // subprogram. 408 TypeIndex ClassType = getTypeIndex(Class); 409 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 410 DisplayName); 411 TI = TypeTable.writeLeafType(MFuncId); 412 } else { 413 // Otherwise, this must be a free function. 414 TypeIndex ParentScope = getScopeIndex(Scope); 415 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 416 TI = TypeTable.writeLeafType(FuncId); 417 } 418 419 return recordTypeIndexForDINode(SP, TI); 420 } 421 422 static bool isNonTrivial(const DICompositeType *DCTy) { 423 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 424 } 425 426 static FunctionOptions 427 getFunctionOptions(const DISubroutineType *Ty, 428 const DICompositeType *ClassTy = nullptr, 429 StringRef SPName = StringRef("")) { 430 FunctionOptions FO = FunctionOptions::None; 431 const DIType *ReturnTy = nullptr; 432 if (auto TypeArray = Ty->getTypeArray()) { 433 if (TypeArray.size()) 434 ReturnTy = TypeArray[0]; 435 } 436 437 // Add CxxReturnUdt option to functions that return nontrivial record types 438 // or methods that return record types. 439 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 440 if (isNonTrivial(ReturnDCTy) || ClassTy) 441 FO |= FunctionOptions::CxxReturnUdt; 442 443 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 444 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 445 FO |= FunctionOptions::Constructor; 446 447 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 448 449 } 450 return FO; 451 } 452 453 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 454 const DICompositeType *Class) { 455 // Always use the method declaration as the key for the function type. The 456 // method declaration contains the this adjustment. 457 if (SP->getDeclaration()) 458 SP = SP->getDeclaration(); 459 assert(!SP->getDeclaration() && "should use declaration as key"); 460 461 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 462 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 463 auto I = TypeIndices.find({SP, Class}); 464 if (I != TypeIndices.end()) 465 return I->second; 466 467 // Make sure complete type info for the class is emitted *after* the member 468 // function type, as the complete class type is likely to reference this 469 // member function type. 470 TypeLoweringScope S(*this); 471 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 472 473 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 474 TypeIndex TI = lowerTypeMemberFunction( 475 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 476 return recordTypeIndexForDINode(SP, TI, Class); 477 } 478 479 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 480 TypeIndex TI, 481 const DIType *ClassTy) { 482 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 483 (void)InsertResult; 484 assert(InsertResult.second && "DINode was already assigned a type index"); 485 return TI; 486 } 487 488 unsigned CodeViewDebug::getPointerSizeInBytes() { 489 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 490 } 491 492 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 493 const LexicalScope *LS) { 494 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 495 // This variable was inlined. Associate it with the InlineSite. 496 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 497 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 498 Site.InlinedLocals.emplace_back(Var); 499 } else { 500 // This variable goes into the corresponding lexical scope. 501 ScopeVariables[LS].emplace_back(Var); 502 } 503 } 504 505 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 506 const DILocation *Loc) { 507 if (!llvm::is_contained(Locs, Loc)) 508 Locs.push_back(Loc); 509 } 510 511 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 512 const MachineFunction *MF) { 513 // Skip this instruction if it has the same location as the previous one. 514 if (!DL || DL == PrevInstLoc) 515 return; 516 517 const DIScope *Scope = DL.get()->getScope(); 518 if (!Scope) 519 return; 520 521 // Skip this line if it is longer than the maximum we can record. 522 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 523 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 524 LI.isNeverStepInto()) 525 return; 526 527 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 528 if (CI.getStartColumn() != DL.getCol()) 529 return; 530 531 if (!CurFn->HaveLineInfo) 532 CurFn->HaveLineInfo = true; 533 unsigned FileId = 0; 534 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 535 FileId = CurFn->LastFileId; 536 else 537 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 538 PrevInstLoc = DL; 539 540 unsigned FuncId = CurFn->FuncId; 541 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 542 const DILocation *Loc = DL.get(); 543 544 // If this location was actually inlined from somewhere else, give it the ID 545 // of the inline call site. 546 FuncId = 547 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 548 549 // Ensure we have links in the tree of inline call sites. 550 bool FirstLoc = true; 551 while ((SiteLoc = Loc->getInlinedAt())) { 552 InlineSite &Site = 553 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 554 if (!FirstLoc) 555 addLocIfNotPresent(Site.ChildSites, Loc); 556 FirstLoc = false; 557 Loc = SiteLoc; 558 } 559 addLocIfNotPresent(CurFn->ChildSites, Loc); 560 } 561 562 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 563 /*PrologueEnd=*/false, /*IsStmt=*/false, 564 DL->getFilename(), SMLoc()); 565 } 566 567 void CodeViewDebug::emitCodeViewMagicVersion() { 568 OS.emitValueToAlignment(4); 569 OS.AddComment("Debug section magic"); 570 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 571 } 572 573 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 574 switch (DWLang) { 575 case dwarf::DW_LANG_C: 576 case dwarf::DW_LANG_C89: 577 case dwarf::DW_LANG_C99: 578 case dwarf::DW_LANG_C11: 579 case dwarf::DW_LANG_ObjC: 580 return SourceLanguage::C; 581 case dwarf::DW_LANG_C_plus_plus: 582 case dwarf::DW_LANG_C_plus_plus_03: 583 case dwarf::DW_LANG_C_plus_plus_11: 584 case dwarf::DW_LANG_C_plus_plus_14: 585 return SourceLanguage::Cpp; 586 case dwarf::DW_LANG_Fortran77: 587 case dwarf::DW_LANG_Fortran90: 588 case dwarf::DW_LANG_Fortran95: 589 case dwarf::DW_LANG_Fortran03: 590 case dwarf::DW_LANG_Fortran08: 591 return SourceLanguage::Fortran; 592 case dwarf::DW_LANG_Pascal83: 593 return SourceLanguage::Pascal; 594 case dwarf::DW_LANG_Cobol74: 595 case dwarf::DW_LANG_Cobol85: 596 return SourceLanguage::Cobol; 597 case dwarf::DW_LANG_Java: 598 return SourceLanguage::Java; 599 case dwarf::DW_LANG_D: 600 return SourceLanguage::D; 601 case dwarf::DW_LANG_Swift: 602 return SourceLanguage::Swift; 603 default: 604 // There's no CodeView representation for this language, and CV doesn't 605 // have an "unknown" option for the language field, so we'll use MASM, 606 // as it's very low level. 607 return SourceLanguage::Masm; 608 } 609 } 610 611 void CodeViewDebug::beginModule(Module *M) { 612 // If module doesn't have named metadata anchors or COFF debug section 613 // is not available, skip any debug info related stuff. 614 NamedMDNode *CUs = M->getNamedMetadata("llvm.dbg.cu"); 615 if (!CUs || !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 616 Asm = nullptr; 617 return; 618 } 619 // Tell MMI that we have and need debug info. 620 MMI->setDebugInfoAvailability(true); 621 622 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 623 624 // Get the current source language. 625 const MDNode *Node = *CUs->operands().begin(); 626 const auto *CU = cast<DICompileUnit>(Node); 627 628 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 629 630 collectGlobalVariableInfo(); 631 632 // Check if we should emit type record hashes. 633 ConstantInt *GH = 634 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 635 EmitDebugGlobalHashes = GH && !GH->isZero(); 636 } 637 638 void CodeViewDebug::endModule() { 639 if (!Asm || !MMI->hasDebugInfo()) 640 return; 641 642 // The COFF .debug$S section consists of several subsections, each starting 643 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 644 // of the payload followed by the payload itself. The subsections are 4-byte 645 // aligned. 646 647 // Use the generic .debug$S section, and make a subsection for all the inlined 648 // subprograms. 649 switchToDebugSectionForSymbol(nullptr); 650 651 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 652 emitObjName(); 653 emitCompilerInformation(); 654 endCVSubsection(CompilerInfo); 655 656 emitInlineeLinesSubsection(); 657 658 // Emit per-function debug information. 659 for (auto &P : FnDebugInfo) 660 if (!P.first->isDeclarationForLinker()) 661 emitDebugInfoForFunction(P.first, *P.second); 662 663 // Get types used by globals without emitting anything. 664 // This is meant to collect all static const data members so they can be 665 // emitted as globals. 666 collectDebugInfoForGlobals(); 667 668 // Emit retained types. 669 emitDebugInfoForRetainedTypes(); 670 671 // Emit global variable debug information. 672 setCurrentSubprogram(nullptr); 673 emitDebugInfoForGlobals(); 674 675 // Switch back to the generic .debug$S section after potentially processing 676 // comdat symbol sections. 677 switchToDebugSectionForSymbol(nullptr); 678 679 // Emit UDT records for any types used by global variables. 680 if (!GlobalUDTs.empty()) { 681 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 682 emitDebugInfoForUDTs(GlobalUDTs); 683 endCVSubsection(SymbolsEnd); 684 } 685 686 // This subsection holds a file index to offset in string table table. 687 OS.AddComment("File index to string table offset subsection"); 688 OS.emitCVFileChecksumsDirective(); 689 690 // This subsection holds the string table. 691 OS.AddComment("String table"); 692 OS.emitCVStringTableDirective(); 693 694 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 695 // subsection in the generic .debug$S section at the end. There is no 696 // particular reason for this ordering other than to match MSVC. 697 emitBuildInfo(); 698 699 // Emit type information and hashes last, so that any types we translate while 700 // emitting function info are included. 701 emitTypeInformation(); 702 703 if (EmitDebugGlobalHashes) 704 emitTypeGlobalHashes(); 705 706 clear(); 707 } 708 709 static void 710 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 711 unsigned MaxFixedRecordLength = 0xF00) { 712 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 713 // after a fixed length portion of the record. The fixed length portion should 714 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 715 // overall record size is less than the maximum allowed. 716 SmallString<32> NullTerminatedString( 717 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 718 NullTerminatedString.push_back('\0'); 719 OS.emitBytes(NullTerminatedString); 720 } 721 722 void CodeViewDebug::emitTypeInformation() { 723 if (TypeTable.empty()) 724 return; 725 726 // Start the .debug$T or .debug$P section with 0x4. 727 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 728 emitCodeViewMagicVersion(); 729 730 TypeTableCollection Table(TypeTable.records()); 731 TypeVisitorCallbackPipeline Pipeline; 732 733 // To emit type record using Codeview MCStreamer adapter 734 CVMCAdapter CVMCOS(OS, Table); 735 TypeRecordMapping typeMapping(CVMCOS); 736 Pipeline.addCallbackToPipeline(typeMapping); 737 738 Optional<TypeIndex> B = Table.getFirst(); 739 while (B) { 740 // This will fail if the record data is invalid. 741 CVType Record = Table.getType(*B); 742 743 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 744 745 if (E) { 746 logAllUnhandledErrors(std::move(E), errs(), "error: "); 747 llvm_unreachable("produced malformed type record"); 748 } 749 750 B = Table.getNext(*B); 751 } 752 } 753 754 void CodeViewDebug::emitTypeGlobalHashes() { 755 if (TypeTable.empty()) 756 return; 757 758 // Start the .debug$H section with the version and hash algorithm, currently 759 // hardcoded to version 0, SHA1. 760 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 761 762 OS.emitValueToAlignment(4); 763 OS.AddComment("Magic"); 764 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 765 OS.AddComment("Section Version"); 766 OS.emitInt16(0); 767 OS.AddComment("Hash Algorithm"); 768 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 769 770 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 771 for (const auto &GHR : TypeTable.hashes()) { 772 if (OS.isVerboseAsm()) { 773 // Emit an EOL-comment describing which TypeIndex this hash corresponds 774 // to, as well as the stringified SHA1 hash. 775 SmallString<32> Comment; 776 raw_svector_ostream CommentOS(Comment); 777 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 778 OS.AddComment(Comment); 779 ++TI; 780 } 781 assert(GHR.Hash.size() == 8); 782 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 783 GHR.Hash.size()); 784 OS.emitBinaryData(S); 785 } 786 } 787 788 void CodeViewDebug::emitObjName() { 789 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 790 791 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 792 llvm::SmallString<256> PathStore(PathRef); 793 794 if (PathRef.empty() || PathRef == "-") { 795 // Don't emit the filename if we're writing to stdout or to /dev/null. 796 PathRef = {}; 797 } else { 798 llvm::sys::path::remove_dots(PathStore, /*remove_dot_dot=*/true); 799 PathRef = PathStore; 800 } 801 802 OS.AddComment("Signature"); 803 OS.emitIntValue(0, 4); 804 805 OS.AddComment("Object name"); 806 emitNullTerminatedSymbolName(OS, PathRef); 807 808 endSymbolRecord(CompilerEnd); 809 } 810 811 namespace { 812 struct Version { 813 int Part[4]; 814 }; 815 } // end anonymous namespace 816 817 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 818 // the version number. 819 static Version parseVersion(StringRef Name) { 820 Version V = {{0}}; 821 int N = 0; 822 for (const char C : Name) { 823 if (isdigit(C)) { 824 V.Part[N] *= 10; 825 V.Part[N] += C - '0'; 826 } else if (C == '.') { 827 ++N; 828 if (N >= 4) 829 return V; 830 } else if (N > 0) 831 return V; 832 } 833 return V; 834 } 835 836 void CodeViewDebug::emitCompilerInformation() { 837 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 838 uint32_t Flags = 0; 839 840 // The low byte of the flags indicates the source language. 841 Flags = CurrentSourceLanguage; 842 // TODO: Figure out which other flags need to be set. 843 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 844 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 845 } 846 847 OS.AddComment("Flags and language"); 848 OS.emitInt32(Flags); 849 850 OS.AddComment("CPUType"); 851 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 852 853 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 854 const MDNode *Node = *CUs->operands().begin(); 855 const auto *CU = cast<DICompileUnit>(Node); 856 857 StringRef CompilerVersion = CU->getProducer(); 858 Version FrontVer = parseVersion(CompilerVersion); 859 OS.AddComment("Frontend version"); 860 for (int N : FrontVer.Part) 861 OS.emitInt16(N); 862 863 // Some Microsoft tools, like Binscope, expect a backend version number of at 864 // least 8.something, so we'll coerce the LLVM version into a form that 865 // guarantees it'll be big enough without really lying about the version. 866 int Major = 1000 * LLVM_VERSION_MAJOR + 867 10 * LLVM_VERSION_MINOR + 868 LLVM_VERSION_PATCH; 869 // Clamp it for builds that use unusually large version numbers. 870 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 871 Version BackVer = {{ Major, 0, 0, 0 }}; 872 OS.AddComment("Backend version"); 873 for (int N : BackVer.Part) 874 OS.emitInt16(N); 875 876 OS.AddComment("Null-terminated compiler version string"); 877 emitNullTerminatedSymbolName(OS, CompilerVersion); 878 879 endSymbolRecord(CompilerEnd); 880 } 881 882 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 883 StringRef S) { 884 StringIdRecord SIR(TypeIndex(0x0), S); 885 return TypeTable.writeLeafType(SIR); 886 } 887 888 void CodeViewDebug::emitBuildInfo() { 889 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 890 // build info. The known prefix is: 891 // - Absolute path of current directory 892 // - Compiler path 893 // - Main source file path, relative to CWD or absolute 894 // - Type server PDB file 895 // - Canonical compiler command line 896 // If frontend and backend compilation are separated (think llc or LTO), it's 897 // not clear if the compiler path should refer to the executable for the 898 // frontend or the backend. Leave it blank for now. 899 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 900 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 901 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 902 const auto *CU = cast<DICompileUnit>(Node); 903 const DIFile *MainSourceFile = CU->getFile(); 904 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 905 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 906 BuildInfoArgs[BuildInfoRecord::SourceFile] = 907 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 908 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 909 // we implement /Zi type servers. 910 BuildInfoRecord BIR(BuildInfoArgs); 911 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 912 913 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 914 // from the module symbols into the type stream. 915 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 916 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 917 OS.AddComment("LF_BUILDINFO index"); 918 OS.emitInt32(BuildInfoIndex.getIndex()); 919 endSymbolRecord(BIEnd); 920 endCVSubsection(BISubsecEnd); 921 } 922 923 void CodeViewDebug::emitInlineeLinesSubsection() { 924 if (InlinedSubprograms.empty()) 925 return; 926 927 OS.AddComment("Inlinee lines subsection"); 928 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 929 930 // We emit the checksum info for files. This is used by debuggers to 931 // determine if a pdb matches the source before loading it. Visual Studio, 932 // for instance, will display a warning that the breakpoints are not valid if 933 // the pdb does not match the source. 934 OS.AddComment("Inlinee lines signature"); 935 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 936 937 for (const DISubprogram *SP : InlinedSubprograms) { 938 assert(TypeIndices.count({SP, nullptr})); 939 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 940 941 OS.AddBlankLine(); 942 unsigned FileId = maybeRecordFile(SP->getFile()); 943 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 944 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 945 OS.AddBlankLine(); 946 OS.AddComment("Type index of inlined function"); 947 OS.emitInt32(InlineeIdx.getIndex()); 948 OS.AddComment("Offset into filechecksum table"); 949 OS.emitCVFileChecksumOffsetDirective(FileId); 950 OS.AddComment("Starting line number"); 951 OS.emitInt32(SP->getLine()); 952 } 953 954 endCVSubsection(InlineEnd); 955 } 956 957 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 958 const DILocation *InlinedAt, 959 const InlineSite &Site) { 960 assert(TypeIndices.count({Site.Inlinee, nullptr})); 961 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 962 963 // SymbolRecord 964 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 965 966 OS.AddComment("PtrParent"); 967 OS.emitInt32(0); 968 OS.AddComment("PtrEnd"); 969 OS.emitInt32(0); 970 OS.AddComment("Inlinee type index"); 971 OS.emitInt32(InlineeIdx.getIndex()); 972 973 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 974 unsigned StartLineNum = Site.Inlinee->getLine(); 975 976 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 977 FI.Begin, FI.End); 978 979 endSymbolRecord(InlineEnd); 980 981 emitLocalVariableList(FI, Site.InlinedLocals); 982 983 // Recurse on child inlined call sites before closing the scope. 984 for (const DILocation *ChildSite : Site.ChildSites) { 985 auto I = FI.InlineSites.find(ChildSite); 986 assert(I != FI.InlineSites.end() && 987 "child site not in function inline site map"); 988 emitInlinedCallSite(FI, ChildSite, I->second); 989 } 990 991 // Close the scope. 992 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 993 } 994 995 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 996 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 997 // comdat key. A section may be comdat because of -ffunction-sections or 998 // because it is comdat in the IR. 999 MCSectionCOFF *GVSec = 1000 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1001 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1002 1003 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1004 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1005 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1006 1007 OS.SwitchSection(DebugSec); 1008 1009 // Emit the magic version number if this is the first time we've switched to 1010 // this section. 1011 if (ComdatDebugSections.insert(DebugSec).second) 1012 emitCodeViewMagicVersion(); 1013 } 1014 1015 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1016 // The only supported thunk ordinal is currently the standard type. 1017 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1018 FunctionInfo &FI, 1019 const MCSymbol *Fn) { 1020 std::string FuncName = 1021 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1022 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1023 1024 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1025 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1026 1027 // Emit S_THUNK32 1028 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1029 OS.AddComment("PtrParent"); 1030 OS.emitInt32(0); 1031 OS.AddComment("PtrEnd"); 1032 OS.emitInt32(0); 1033 OS.AddComment("PtrNext"); 1034 OS.emitInt32(0); 1035 OS.AddComment("Thunk section relative address"); 1036 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1037 OS.AddComment("Thunk section index"); 1038 OS.EmitCOFFSectionIndex(Fn); 1039 OS.AddComment("Code size"); 1040 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1041 OS.AddComment("Ordinal"); 1042 OS.emitInt8(unsigned(ordinal)); 1043 OS.AddComment("Function name"); 1044 emitNullTerminatedSymbolName(OS, FuncName); 1045 // Additional fields specific to the thunk ordinal would go here. 1046 endSymbolRecord(ThunkRecordEnd); 1047 1048 // Local variables/inlined routines are purposely omitted here. The point of 1049 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1050 1051 // Emit S_PROC_ID_END 1052 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1053 1054 endCVSubsection(SymbolsEnd); 1055 } 1056 1057 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1058 FunctionInfo &FI) { 1059 // For each function there is a separate subsection which holds the PC to 1060 // file:line table. 1061 const MCSymbol *Fn = Asm->getSymbol(GV); 1062 assert(Fn); 1063 1064 // Switch to the to a comdat section, if appropriate. 1065 switchToDebugSectionForSymbol(Fn); 1066 1067 std::string FuncName; 1068 auto *SP = GV->getSubprogram(); 1069 assert(SP); 1070 setCurrentSubprogram(SP); 1071 1072 if (SP->isThunk()) { 1073 emitDebugInfoForThunk(GV, FI, Fn); 1074 return; 1075 } 1076 1077 // If we have a display name, build the fully qualified name by walking the 1078 // chain of scopes. 1079 if (!SP->getName().empty()) 1080 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1081 1082 // If our DISubprogram name is empty, use the mangled name. 1083 if (FuncName.empty()) 1084 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1085 1086 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1087 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1088 OS.EmitCVFPOData(Fn); 1089 1090 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1091 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1092 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1093 { 1094 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1095 : SymbolKind::S_GPROC32_ID; 1096 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1097 1098 // These fields are filled in by tools like CVPACK which run after the fact. 1099 OS.AddComment("PtrParent"); 1100 OS.emitInt32(0); 1101 OS.AddComment("PtrEnd"); 1102 OS.emitInt32(0); 1103 OS.AddComment("PtrNext"); 1104 OS.emitInt32(0); 1105 // This is the important bit that tells the debugger where the function 1106 // code is located and what's its size: 1107 OS.AddComment("Code size"); 1108 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1109 OS.AddComment("Offset after prologue"); 1110 OS.emitInt32(0); 1111 OS.AddComment("Offset before epilogue"); 1112 OS.emitInt32(0); 1113 OS.AddComment("Function type index"); 1114 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1115 OS.AddComment("Function section relative address"); 1116 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1117 OS.AddComment("Function section index"); 1118 OS.EmitCOFFSectionIndex(Fn); 1119 OS.AddComment("Flags"); 1120 OS.emitInt8(0); 1121 // Emit the function display name as a null-terminated string. 1122 OS.AddComment("Function name"); 1123 // Truncate the name so we won't overflow the record length field. 1124 emitNullTerminatedSymbolName(OS, FuncName); 1125 endSymbolRecord(ProcRecordEnd); 1126 1127 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1128 // Subtract out the CSR size since MSVC excludes that and we include it. 1129 OS.AddComment("FrameSize"); 1130 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1131 OS.AddComment("Padding"); 1132 OS.emitInt32(0); 1133 OS.AddComment("Offset of padding"); 1134 OS.emitInt32(0); 1135 OS.AddComment("Bytes of callee saved registers"); 1136 OS.emitInt32(FI.CSRSize); 1137 OS.AddComment("Exception handler offset"); 1138 OS.emitInt32(0); 1139 OS.AddComment("Exception handler section"); 1140 OS.emitInt16(0); 1141 OS.AddComment("Flags (defines frame register)"); 1142 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1143 endSymbolRecord(FrameProcEnd); 1144 1145 emitLocalVariableList(FI, FI.Locals); 1146 emitGlobalVariableList(FI.Globals); 1147 emitLexicalBlockList(FI.ChildBlocks, FI); 1148 1149 // Emit inlined call site information. Only emit functions inlined directly 1150 // into the parent function. We'll emit the other sites recursively as part 1151 // of their parent inline site. 1152 for (const DILocation *InlinedAt : FI.ChildSites) { 1153 auto I = FI.InlineSites.find(InlinedAt); 1154 assert(I != FI.InlineSites.end() && 1155 "child site not in function inline site map"); 1156 emitInlinedCallSite(FI, InlinedAt, I->second); 1157 } 1158 1159 for (auto Annot : FI.Annotations) { 1160 MCSymbol *Label = Annot.first; 1161 MDTuple *Strs = cast<MDTuple>(Annot.second); 1162 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1163 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1164 // FIXME: Make sure we don't overflow the max record size. 1165 OS.EmitCOFFSectionIndex(Label); 1166 OS.emitInt16(Strs->getNumOperands()); 1167 for (Metadata *MD : Strs->operands()) { 1168 // MDStrings are null terminated, so we can do EmitBytes and get the 1169 // nice .asciz directive. 1170 StringRef Str = cast<MDString>(MD)->getString(); 1171 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1172 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1173 } 1174 endSymbolRecord(AnnotEnd); 1175 } 1176 1177 for (auto HeapAllocSite : FI.HeapAllocSites) { 1178 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1179 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1180 const DIType *DITy = std::get<2>(HeapAllocSite); 1181 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1182 OS.AddComment("Call site offset"); 1183 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1184 OS.AddComment("Call site section index"); 1185 OS.EmitCOFFSectionIndex(BeginLabel); 1186 OS.AddComment("Call instruction length"); 1187 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1188 OS.AddComment("Type index"); 1189 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1190 endSymbolRecord(HeapAllocEnd); 1191 } 1192 1193 if (SP != nullptr) 1194 emitDebugInfoForUDTs(LocalUDTs); 1195 1196 // We're done with this function. 1197 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1198 } 1199 endCVSubsection(SymbolsEnd); 1200 1201 // We have an assembler directive that takes care of the whole line table. 1202 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1203 } 1204 1205 CodeViewDebug::LocalVarDefRange 1206 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1207 LocalVarDefRange DR; 1208 DR.InMemory = -1; 1209 DR.DataOffset = Offset; 1210 assert(DR.DataOffset == Offset && "truncation"); 1211 DR.IsSubfield = 0; 1212 DR.StructOffset = 0; 1213 DR.CVRegister = CVRegister; 1214 return DR; 1215 } 1216 1217 void CodeViewDebug::collectVariableInfoFromMFTable( 1218 DenseSet<InlinedEntity> &Processed) { 1219 const MachineFunction &MF = *Asm->MF; 1220 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1221 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1222 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1223 1224 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1225 if (!VI.Var) 1226 continue; 1227 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1228 "Expected inlined-at fields to agree"); 1229 1230 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1231 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1232 1233 // If variable scope is not found then skip this variable. 1234 if (!Scope) 1235 continue; 1236 1237 // If the variable has an attached offset expression, extract it. 1238 // FIXME: Try to handle DW_OP_deref as well. 1239 int64_t ExprOffset = 0; 1240 bool Deref = false; 1241 if (VI.Expr) { 1242 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1243 if (VI.Expr->getNumElements() == 1 && 1244 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1245 Deref = true; 1246 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1247 continue; 1248 } 1249 1250 // Get the frame register used and the offset. 1251 Register FrameReg; 1252 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1253 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1254 1255 assert(!FrameOffset.getScalable() && 1256 "Frame offsets with a scalable component are not supported"); 1257 1258 // Calculate the label ranges. 1259 LocalVarDefRange DefRange = 1260 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1261 1262 for (const InsnRange &Range : Scope->getRanges()) { 1263 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1264 const MCSymbol *End = getLabelAfterInsn(Range.second); 1265 End = End ? End : Asm->getFunctionEnd(); 1266 DefRange.Ranges.emplace_back(Begin, End); 1267 } 1268 1269 LocalVariable Var; 1270 Var.DIVar = VI.Var; 1271 Var.DefRanges.emplace_back(std::move(DefRange)); 1272 if (Deref) 1273 Var.UseReferenceType = true; 1274 1275 recordLocalVariable(std::move(Var), Scope); 1276 } 1277 } 1278 1279 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1280 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1281 } 1282 1283 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1284 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1285 } 1286 1287 void CodeViewDebug::calculateRanges( 1288 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1289 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1290 1291 // Calculate the definition ranges. 1292 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1293 const auto &Entry = *I; 1294 if (!Entry.isDbgValue()) 1295 continue; 1296 const MachineInstr *DVInst = Entry.getInstr(); 1297 assert(DVInst->isDebugValue() && "Invalid History entry"); 1298 // FIXME: Find a way to represent constant variables, since they are 1299 // relatively common. 1300 Optional<DbgVariableLocation> Location = 1301 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1302 if (!Location) 1303 continue; 1304 1305 // CodeView can only express variables in register and variables in memory 1306 // at a constant offset from a register. However, for variables passed 1307 // indirectly by pointer, it is common for that pointer to be spilled to a 1308 // stack location. For the special case of one offseted load followed by a 1309 // zero offset load (a pointer spilled to the stack), we change the type of 1310 // the local variable from a value type to a reference type. This tricks the 1311 // debugger into doing the load for us. 1312 if (Var.UseReferenceType) { 1313 // We're using a reference type. Drop the last zero offset load. 1314 if (canUseReferenceType(*Location)) 1315 Location->LoadChain.pop_back(); 1316 else 1317 continue; 1318 } else if (needsReferenceType(*Location)) { 1319 // This location can't be expressed without switching to a reference type. 1320 // Start over using that. 1321 Var.UseReferenceType = true; 1322 Var.DefRanges.clear(); 1323 calculateRanges(Var, Entries); 1324 return; 1325 } 1326 1327 // We can only handle a register or an offseted load of a register. 1328 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1329 continue; 1330 { 1331 LocalVarDefRange DR; 1332 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1333 DR.InMemory = !Location->LoadChain.empty(); 1334 DR.DataOffset = 1335 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1336 if (Location->FragmentInfo) { 1337 DR.IsSubfield = true; 1338 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1339 } else { 1340 DR.IsSubfield = false; 1341 DR.StructOffset = 0; 1342 } 1343 1344 if (Var.DefRanges.empty() || 1345 Var.DefRanges.back().isDifferentLocation(DR)) { 1346 Var.DefRanges.emplace_back(std::move(DR)); 1347 } 1348 } 1349 1350 // Compute the label range. 1351 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1352 const MCSymbol *End; 1353 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1354 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1355 End = EndingEntry.isDbgValue() 1356 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1357 : getLabelAfterInsn(EndingEntry.getInstr()); 1358 } else 1359 End = Asm->getFunctionEnd(); 1360 1361 // If the last range end is our begin, just extend the last range. 1362 // Otherwise make a new range. 1363 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1364 Var.DefRanges.back().Ranges; 1365 if (!R.empty() && R.back().second == Begin) 1366 R.back().second = End; 1367 else 1368 R.emplace_back(Begin, End); 1369 1370 // FIXME: Do more range combining. 1371 } 1372 } 1373 1374 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1375 DenseSet<InlinedEntity> Processed; 1376 // Grab the variable info that was squirreled away in the MMI side-table. 1377 collectVariableInfoFromMFTable(Processed); 1378 1379 for (const auto &I : DbgValues) { 1380 InlinedEntity IV = I.first; 1381 if (Processed.count(IV)) 1382 continue; 1383 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1384 const DILocation *InlinedAt = IV.second; 1385 1386 // Instruction ranges, specifying where IV is accessible. 1387 const auto &Entries = I.second; 1388 1389 LexicalScope *Scope = nullptr; 1390 if (InlinedAt) 1391 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1392 else 1393 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1394 // If variable scope is not found then skip this variable. 1395 if (!Scope) 1396 continue; 1397 1398 LocalVariable Var; 1399 Var.DIVar = DIVar; 1400 1401 calculateRanges(Var, Entries); 1402 recordLocalVariable(std::move(Var), Scope); 1403 } 1404 } 1405 1406 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1407 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1408 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1409 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1410 const Function &GV = MF->getFunction(); 1411 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1412 assert(Insertion.second && "function already has info"); 1413 CurFn = Insertion.first->second.get(); 1414 CurFn->FuncId = NextFuncId++; 1415 CurFn->Begin = Asm->getFunctionBegin(); 1416 1417 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1418 // callee-saved registers were used. For targets that don't use a PUSH 1419 // instruction (AArch64), this will be zero. 1420 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1421 CurFn->FrameSize = MFI.getStackSize(); 1422 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1423 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1424 1425 // For this function S_FRAMEPROC record, figure out which codeview register 1426 // will be the frame pointer. 1427 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1428 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1429 if (CurFn->FrameSize > 0) { 1430 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1431 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1432 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1433 } else { 1434 // If there is an FP, parameters are always relative to it. 1435 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1436 if (CurFn->HasStackRealignment) { 1437 // If the stack needs realignment, locals are relative to SP or VFRAME. 1438 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1439 } else { 1440 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1441 // other stack adjustments. 1442 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1443 } 1444 } 1445 } 1446 1447 // Compute other frame procedure options. 1448 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1449 if (MFI.hasVarSizedObjects()) 1450 FPO |= FrameProcedureOptions::HasAlloca; 1451 if (MF->exposesReturnsTwice()) 1452 FPO |= FrameProcedureOptions::HasSetJmp; 1453 // FIXME: Set HasLongJmp if we ever track that info. 1454 if (MF->hasInlineAsm()) 1455 FPO |= FrameProcedureOptions::HasInlineAssembly; 1456 if (GV.hasPersonalityFn()) { 1457 if (isAsynchronousEHPersonality( 1458 classifyEHPersonality(GV.getPersonalityFn()))) 1459 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1460 else 1461 FPO |= FrameProcedureOptions::HasExceptionHandling; 1462 } 1463 if (GV.hasFnAttribute(Attribute::InlineHint)) 1464 FPO |= FrameProcedureOptions::MarkedInline; 1465 if (GV.hasFnAttribute(Attribute::Naked)) 1466 FPO |= FrameProcedureOptions::Naked; 1467 if (MFI.hasStackProtectorIndex()) 1468 FPO |= FrameProcedureOptions::SecurityChecks; 1469 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1470 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1471 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1472 !GV.hasOptSize() && !GV.hasOptNone()) 1473 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1474 if (GV.hasProfileData()) { 1475 FPO |= FrameProcedureOptions::ValidProfileCounts; 1476 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1477 } 1478 // FIXME: Set GuardCfg when it is implemented. 1479 CurFn->FrameProcOpts = FPO; 1480 1481 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1482 1483 // Find the end of the function prolog. First known non-DBG_VALUE and 1484 // non-frame setup location marks the beginning of the function body. 1485 // FIXME: is there a simpler a way to do this? Can we just search 1486 // for the first instruction of the function, not the last of the prolog? 1487 DebugLoc PrologEndLoc; 1488 bool EmptyPrologue = true; 1489 for (const auto &MBB : *MF) { 1490 for (const auto &MI : MBB) { 1491 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1492 MI.getDebugLoc()) { 1493 PrologEndLoc = MI.getDebugLoc(); 1494 break; 1495 } else if (!MI.isMetaInstruction()) { 1496 EmptyPrologue = false; 1497 } 1498 } 1499 } 1500 1501 // Record beginning of function if we have a non-empty prologue. 1502 if (PrologEndLoc && !EmptyPrologue) { 1503 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1504 maybeRecordLocation(FnStartDL, MF); 1505 } 1506 1507 // Find heap alloc sites and emit labels around them. 1508 for (const auto &MBB : *MF) { 1509 for (const auto &MI : MBB) { 1510 if (MI.getHeapAllocMarker()) { 1511 requestLabelBeforeInsn(&MI); 1512 requestLabelAfterInsn(&MI); 1513 } 1514 } 1515 } 1516 } 1517 1518 static bool shouldEmitUdt(const DIType *T) { 1519 if (!T) 1520 return false; 1521 1522 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1523 if (T->getTag() == dwarf::DW_TAG_typedef) { 1524 if (DIScope *Scope = T->getScope()) { 1525 switch (Scope->getTag()) { 1526 case dwarf::DW_TAG_structure_type: 1527 case dwarf::DW_TAG_class_type: 1528 case dwarf::DW_TAG_union_type: 1529 return false; 1530 default: 1531 // do nothing. 1532 ; 1533 } 1534 } 1535 } 1536 1537 while (true) { 1538 if (!T || T->isForwardDecl()) 1539 return false; 1540 1541 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1542 if (!DT) 1543 return true; 1544 T = DT->getBaseType(); 1545 } 1546 return true; 1547 } 1548 1549 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1550 // Don't record empty UDTs. 1551 if (Ty->getName().empty()) 1552 return; 1553 if (!shouldEmitUdt(Ty)) 1554 return; 1555 1556 SmallVector<StringRef, 5> ParentScopeNames; 1557 const DISubprogram *ClosestSubprogram = 1558 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1559 1560 std::string FullyQualifiedName = 1561 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1562 1563 if (ClosestSubprogram == nullptr) { 1564 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1565 } else if (ClosestSubprogram == CurrentSubprogram) { 1566 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1567 } 1568 1569 // TODO: What if the ClosestSubprogram is neither null or the current 1570 // subprogram? Currently, the UDT just gets dropped on the floor. 1571 // 1572 // The current behavior is not desirable. To get maximal fidelity, we would 1573 // need to perform all type translation before beginning emission of .debug$S 1574 // and then make LocalUDTs a member of FunctionInfo 1575 } 1576 1577 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1578 // Generic dispatch for lowering an unknown type. 1579 switch (Ty->getTag()) { 1580 case dwarf::DW_TAG_array_type: 1581 return lowerTypeArray(cast<DICompositeType>(Ty)); 1582 case dwarf::DW_TAG_typedef: 1583 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1584 case dwarf::DW_TAG_base_type: 1585 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1586 case dwarf::DW_TAG_pointer_type: 1587 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1588 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1589 LLVM_FALLTHROUGH; 1590 case dwarf::DW_TAG_reference_type: 1591 case dwarf::DW_TAG_rvalue_reference_type: 1592 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1593 case dwarf::DW_TAG_ptr_to_member_type: 1594 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1595 case dwarf::DW_TAG_restrict_type: 1596 case dwarf::DW_TAG_const_type: 1597 case dwarf::DW_TAG_volatile_type: 1598 // TODO: add support for DW_TAG_atomic_type here 1599 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1600 case dwarf::DW_TAG_subroutine_type: 1601 if (ClassTy) { 1602 // The member function type of a member function pointer has no 1603 // ThisAdjustment. 1604 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1605 /*ThisAdjustment=*/0, 1606 /*IsStaticMethod=*/false); 1607 } 1608 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1609 case dwarf::DW_TAG_enumeration_type: 1610 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1611 case dwarf::DW_TAG_class_type: 1612 case dwarf::DW_TAG_structure_type: 1613 return lowerTypeClass(cast<DICompositeType>(Ty)); 1614 case dwarf::DW_TAG_union_type: 1615 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1616 case dwarf::DW_TAG_string_type: 1617 return lowerTypeString(cast<DIStringType>(Ty)); 1618 case dwarf::DW_TAG_unspecified_type: 1619 if (Ty->getName() == "decltype(nullptr)") 1620 return TypeIndex::NullptrT(); 1621 return TypeIndex::None(); 1622 default: 1623 // Use the null type index. 1624 return TypeIndex(); 1625 } 1626 } 1627 1628 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1629 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1630 StringRef TypeName = Ty->getName(); 1631 1632 addToUDTs(Ty); 1633 1634 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1635 TypeName == "HRESULT") 1636 return TypeIndex(SimpleTypeKind::HResult); 1637 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1638 TypeName == "wchar_t") 1639 return TypeIndex(SimpleTypeKind::WideCharacter); 1640 1641 return UnderlyingTypeIndex; 1642 } 1643 1644 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1645 const DIType *ElementType = Ty->getBaseType(); 1646 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1647 // IndexType is size_t, which depends on the bitness of the target. 1648 TypeIndex IndexType = getPointerSizeInBytes() == 8 1649 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1650 : TypeIndex(SimpleTypeKind::UInt32Long); 1651 1652 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1653 1654 // Add subranges to array type. 1655 DINodeArray Elements = Ty->getElements(); 1656 for (int i = Elements.size() - 1; i >= 0; --i) { 1657 const DINode *Element = Elements[i]; 1658 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1659 1660 const DISubrange *Subrange = cast<DISubrange>(Element); 1661 int64_t Count = -1; 1662 1663 // If Subrange has a Count field, use it. 1664 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1665 // where lowerbound is from the LowerBound field of the Subrange, 1666 // or the language default lowerbound if that field is unspecified. 1667 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1668 Count = CI->getSExtValue(); 1669 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1670 // Fortran uses 1 as the default lowerbound; other languages use 0. 1671 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1672 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1673 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1674 Count = UI->getSExtValue() - Lowerbound + 1; 1675 } 1676 1677 // Forward declarations of arrays without a size and VLAs use a count of -1. 1678 // Emit a count of zero in these cases to match what MSVC does for arrays 1679 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1680 // should do for them even if we could distinguish them. 1681 if (Count == -1) 1682 Count = 0; 1683 1684 // Update the element size and element type index for subsequent subranges. 1685 ElementSize *= Count; 1686 1687 // If this is the outermost array, use the size from the array. It will be 1688 // more accurate if we had a VLA or an incomplete element type size. 1689 uint64_t ArraySize = 1690 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1691 1692 StringRef Name = (i == 0) ? Ty->getName() : ""; 1693 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1694 ElementTypeIndex = TypeTable.writeLeafType(AR); 1695 } 1696 1697 return ElementTypeIndex; 1698 } 1699 1700 // This function lowers a Fortran character type (DIStringType). 1701 // Note that it handles only the character*n variant (using SizeInBits 1702 // field in DIString to describe the type size) at the moment. 1703 // Other variants (leveraging the StringLength and StringLengthExp 1704 // fields in DIStringType) remain TBD. 1705 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1706 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1707 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1708 StringRef Name = Ty->getName(); 1709 // IndexType is size_t, which depends on the bitness of the target. 1710 TypeIndex IndexType = getPointerSizeInBytes() == 8 1711 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1712 : TypeIndex(SimpleTypeKind::UInt32Long); 1713 1714 // Create a type of character array of ArraySize. 1715 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1716 1717 return TypeTable.writeLeafType(AR); 1718 } 1719 1720 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1721 TypeIndex Index; 1722 dwarf::TypeKind Kind; 1723 uint32_t ByteSize; 1724 1725 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1726 ByteSize = Ty->getSizeInBits() / 8; 1727 1728 SimpleTypeKind STK = SimpleTypeKind::None; 1729 switch (Kind) { 1730 case dwarf::DW_ATE_address: 1731 // FIXME: Translate 1732 break; 1733 case dwarf::DW_ATE_boolean: 1734 switch (ByteSize) { 1735 case 1: STK = SimpleTypeKind::Boolean8; break; 1736 case 2: STK = SimpleTypeKind::Boolean16; break; 1737 case 4: STK = SimpleTypeKind::Boolean32; break; 1738 case 8: STK = SimpleTypeKind::Boolean64; break; 1739 case 16: STK = SimpleTypeKind::Boolean128; break; 1740 } 1741 break; 1742 case dwarf::DW_ATE_complex_float: 1743 switch (ByteSize) { 1744 case 2: STK = SimpleTypeKind::Complex16; break; 1745 case 4: STK = SimpleTypeKind::Complex32; break; 1746 case 8: STK = SimpleTypeKind::Complex64; break; 1747 case 10: STK = SimpleTypeKind::Complex80; break; 1748 case 16: STK = SimpleTypeKind::Complex128; break; 1749 } 1750 break; 1751 case dwarf::DW_ATE_float: 1752 switch (ByteSize) { 1753 case 2: STK = SimpleTypeKind::Float16; break; 1754 case 4: STK = SimpleTypeKind::Float32; break; 1755 case 6: STK = SimpleTypeKind::Float48; break; 1756 case 8: STK = SimpleTypeKind::Float64; break; 1757 case 10: STK = SimpleTypeKind::Float80; break; 1758 case 16: STK = SimpleTypeKind::Float128; break; 1759 } 1760 break; 1761 case dwarf::DW_ATE_signed: 1762 switch (ByteSize) { 1763 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1764 case 2: STK = SimpleTypeKind::Int16Short; break; 1765 case 4: STK = SimpleTypeKind::Int32; break; 1766 case 8: STK = SimpleTypeKind::Int64Quad; break; 1767 case 16: STK = SimpleTypeKind::Int128Oct; break; 1768 } 1769 break; 1770 case dwarf::DW_ATE_unsigned: 1771 switch (ByteSize) { 1772 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1773 case 2: STK = SimpleTypeKind::UInt16Short; break; 1774 case 4: STK = SimpleTypeKind::UInt32; break; 1775 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1776 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1777 } 1778 break; 1779 case dwarf::DW_ATE_UTF: 1780 switch (ByteSize) { 1781 case 2: STK = SimpleTypeKind::Character16; break; 1782 case 4: STK = SimpleTypeKind::Character32; break; 1783 } 1784 break; 1785 case dwarf::DW_ATE_signed_char: 1786 if (ByteSize == 1) 1787 STK = SimpleTypeKind::SignedCharacter; 1788 break; 1789 case dwarf::DW_ATE_unsigned_char: 1790 if (ByteSize == 1) 1791 STK = SimpleTypeKind::UnsignedCharacter; 1792 break; 1793 default: 1794 break; 1795 } 1796 1797 // Apply some fixups based on the source-level type name. 1798 // Include some amount of canonicalization from an old naming scheme Clang 1799 // used to use for integer types (in an outdated effort to be compatible with 1800 // GCC's debug info/GDB's behavior, which has since been addressed). 1801 if (STK == SimpleTypeKind::Int32 && 1802 (Ty->getName() == "long int" || Ty->getName() == "long")) 1803 STK = SimpleTypeKind::Int32Long; 1804 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1805 Ty->getName() == "unsigned long")) 1806 STK = SimpleTypeKind::UInt32Long; 1807 if (STK == SimpleTypeKind::UInt16Short && 1808 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1809 STK = SimpleTypeKind::WideCharacter; 1810 if ((STK == SimpleTypeKind::SignedCharacter || 1811 STK == SimpleTypeKind::UnsignedCharacter) && 1812 Ty->getName() == "char") 1813 STK = SimpleTypeKind::NarrowCharacter; 1814 1815 return TypeIndex(STK); 1816 } 1817 1818 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1819 PointerOptions PO) { 1820 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1821 1822 // Pointers to simple types without any options can use SimpleTypeMode, rather 1823 // than having a dedicated pointer type record. 1824 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1825 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1826 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1827 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1828 ? SimpleTypeMode::NearPointer64 1829 : SimpleTypeMode::NearPointer32; 1830 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1831 } 1832 1833 PointerKind PK = 1834 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1835 PointerMode PM = PointerMode::Pointer; 1836 switch (Ty->getTag()) { 1837 default: llvm_unreachable("not a pointer tag type"); 1838 case dwarf::DW_TAG_pointer_type: 1839 PM = PointerMode::Pointer; 1840 break; 1841 case dwarf::DW_TAG_reference_type: 1842 PM = PointerMode::LValueReference; 1843 break; 1844 case dwarf::DW_TAG_rvalue_reference_type: 1845 PM = PointerMode::RValueReference; 1846 break; 1847 } 1848 1849 if (Ty->isObjectPointer()) 1850 PO |= PointerOptions::Const; 1851 1852 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1853 return TypeTable.writeLeafType(PR); 1854 } 1855 1856 static PointerToMemberRepresentation 1857 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1858 // SizeInBytes being zero generally implies that the member pointer type was 1859 // incomplete, which can happen if it is part of a function prototype. In this 1860 // case, use the unknown model instead of the general model. 1861 if (IsPMF) { 1862 switch (Flags & DINode::FlagPtrToMemberRep) { 1863 case 0: 1864 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1865 : PointerToMemberRepresentation::GeneralFunction; 1866 case DINode::FlagSingleInheritance: 1867 return PointerToMemberRepresentation::SingleInheritanceFunction; 1868 case DINode::FlagMultipleInheritance: 1869 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1870 case DINode::FlagVirtualInheritance: 1871 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1872 } 1873 } else { 1874 switch (Flags & DINode::FlagPtrToMemberRep) { 1875 case 0: 1876 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1877 : PointerToMemberRepresentation::GeneralData; 1878 case DINode::FlagSingleInheritance: 1879 return PointerToMemberRepresentation::SingleInheritanceData; 1880 case DINode::FlagMultipleInheritance: 1881 return PointerToMemberRepresentation::MultipleInheritanceData; 1882 case DINode::FlagVirtualInheritance: 1883 return PointerToMemberRepresentation::VirtualInheritanceData; 1884 } 1885 } 1886 llvm_unreachable("invalid ptr to member representation"); 1887 } 1888 1889 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1890 PointerOptions PO) { 1891 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1892 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1893 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1894 TypeIndex PointeeTI = 1895 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1896 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1897 : PointerKind::Near32; 1898 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1899 : PointerMode::PointerToDataMember; 1900 1901 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1902 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1903 MemberPointerInfo MPI( 1904 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1905 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1906 return TypeTable.writeLeafType(PR); 1907 } 1908 1909 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1910 /// have a translation, use the NearC convention. 1911 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1912 switch (DwarfCC) { 1913 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1914 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1915 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1916 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1917 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1918 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1919 } 1920 return CallingConvention::NearC; 1921 } 1922 1923 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1924 ModifierOptions Mods = ModifierOptions::None; 1925 PointerOptions PO = PointerOptions::None; 1926 bool IsModifier = true; 1927 const DIType *BaseTy = Ty; 1928 while (IsModifier && BaseTy) { 1929 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1930 switch (BaseTy->getTag()) { 1931 case dwarf::DW_TAG_const_type: 1932 Mods |= ModifierOptions::Const; 1933 PO |= PointerOptions::Const; 1934 break; 1935 case dwarf::DW_TAG_volatile_type: 1936 Mods |= ModifierOptions::Volatile; 1937 PO |= PointerOptions::Volatile; 1938 break; 1939 case dwarf::DW_TAG_restrict_type: 1940 // Only pointer types be marked with __restrict. There is no known flag 1941 // for __restrict in LF_MODIFIER records. 1942 PO |= PointerOptions::Restrict; 1943 break; 1944 default: 1945 IsModifier = false; 1946 break; 1947 } 1948 if (IsModifier) 1949 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1950 } 1951 1952 // Check if the inner type will use an LF_POINTER record. If so, the 1953 // qualifiers will go in the LF_POINTER record. This comes up for types like 1954 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1955 // char *'. 1956 if (BaseTy) { 1957 switch (BaseTy->getTag()) { 1958 case dwarf::DW_TAG_pointer_type: 1959 case dwarf::DW_TAG_reference_type: 1960 case dwarf::DW_TAG_rvalue_reference_type: 1961 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1962 case dwarf::DW_TAG_ptr_to_member_type: 1963 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1964 default: 1965 break; 1966 } 1967 } 1968 1969 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1970 1971 // Return the base type index if there aren't any modifiers. For example, the 1972 // metadata could contain restrict wrappers around non-pointer types. 1973 if (Mods == ModifierOptions::None) 1974 return ModifiedTI; 1975 1976 ModifierRecord MR(ModifiedTI, Mods); 1977 return TypeTable.writeLeafType(MR); 1978 } 1979 1980 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1981 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1982 for (const DIType *ArgType : Ty->getTypeArray()) 1983 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1984 1985 // MSVC uses type none for variadic argument. 1986 if (ReturnAndArgTypeIndices.size() > 1 && 1987 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1988 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1989 } 1990 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1991 ArrayRef<TypeIndex> ArgTypeIndices = None; 1992 if (!ReturnAndArgTypeIndices.empty()) { 1993 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1994 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1995 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1996 } 1997 1998 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1999 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2000 2001 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2002 2003 FunctionOptions FO = getFunctionOptions(Ty); 2004 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2005 ArgListIndex); 2006 return TypeTable.writeLeafType(Procedure); 2007 } 2008 2009 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2010 const DIType *ClassTy, 2011 int ThisAdjustment, 2012 bool IsStaticMethod, 2013 FunctionOptions FO) { 2014 // Lower the containing class type. 2015 TypeIndex ClassType = getTypeIndex(ClassTy); 2016 2017 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2018 2019 unsigned Index = 0; 2020 SmallVector<TypeIndex, 8> ArgTypeIndices; 2021 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2022 if (ReturnAndArgs.size() > Index) { 2023 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2024 } 2025 2026 // If the first argument is a pointer type and this isn't a static method, 2027 // treat it as the special 'this' parameter, which is encoded separately from 2028 // the arguments. 2029 TypeIndex ThisTypeIndex; 2030 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2031 if (const DIDerivedType *PtrTy = 2032 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2033 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2034 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2035 Index++; 2036 } 2037 } 2038 } 2039 2040 while (Index < ReturnAndArgs.size()) 2041 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2042 2043 // MSVC uses type none for variadic argument. 2044 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2045 ArgTypeIndices.back() = TypeIndex::None(); 2046 2047 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2048 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2049 2050 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2051 2052 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2053 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2054 return TypeTable.writeLeafType(MFR); 2055 } 2056 2057 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2058 unsigned VSlotCount = 2059 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2060 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2061 2062 VFTableShapeRecord VFTSR(Slots); 2063 return TypeTable.writeLeafType(VFTSR); 2064 } 2065 2066 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2067 switch (Flags & DINode::FlagAccessibility) { 2068 case DINode::FlagPrivate: return MemberAccess::Private; 2069 case DINode::FlagPublic: return MemberAccess::Public; 2070 case DINode::FlagProtected: return MemberAccess::Protected; 2071 case 0: 2072 // If there was no explicit access control, provide the default for the tag. 2073 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2074 : MemberAccess::Public; 2075 } 2076 llvm_unreachable("access flags are exclusive"); 2077 } 2078 2079 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2080 if (SP->isArtificial()) 2081 return MethodOptions::CompilerGenerated; 2082 2083 // FIXME: Handle other MethodOptions. 2084 2085 return MethodOptions::None; 2086 } 2087 2088 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2089 bool Introduced) { 2090 if (SP->getFlags() & DINode::FlagStaticMember) 2091 return MethodKind::Static; 2092 2093 switch (SP->getVirtuality()) { 2094 case dwarf::DW_VIRTUALITY_none: 2095 break; 2096 case dwarf::DW_VIRTUALITY_virtual: 2097 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2098 case dwarf::DW_VIRTUALITY_pure_virtual: 2099 return Introduced ? MethodKind::PureIntroducingVirtual 2100 : MethodKind::PureVirtual; 2101 default: 2102 llvm_unreachable("unhandled virtuality case"); 2103 } 2104 2105 return MethodKind::Vanilla; 2106 } 2107 2108 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2109 switch (Ty->getTag()) { 2110 case dwarf::DW_TAG_class_type: 2111 return TypeRecordKind::Class; 2112 case dwarf::DW_TAG_structure_type: 2113 return TypeRecordKind::Struct; 2114 default: 2115 llvm_unreachable("unexpected tag"); 2116 } 2117 } 2118 2119 /// Return ClassOptions that should be present on both the forward declaration 2120 /// and the defintion of a tag type. 2121 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2122 ClassOptions CO = ClassOptions::None; 2123 2124 // MSVC always sets this flag, even for local types. Clang doesn't always 2125 // appear to give every type a linkage name, which may be problematic for us. 2126 // FIXME: Investigate the consequences of not following them here. 2127 if (!Ty->getIdentifier().empty()) 2128 CO |= ClassOptions::HasUniqueName; 2129 2130 // Put the Nested flag on a type if it appears immediately inside a tag type. 2131 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2132 // here. That flag is only set on definitions, and not forward declarations. 2133 const DIScope *ImmediateScope = Ty->getScope(); 2134 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2135 CO |= ClassOptions::Nested; 2136 2137 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2138 // type only when it has an immediate function scope. Clang never puts enums 2139 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2140 // always in function, class, or file scopes. 2141 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2142 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2143 CO |= ClassOptions::Scoped; 2144 } else { 2145 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2146 Scope = Scope->getScope()) { 2147 if (isa<DISubprogram>(Scope)) { 2148 CO |= ClassOptions::Scoped; 2149 break; 2150 } 2151 } 2152 } 2153 2154 return CO; 2155 } 2156 2157 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2158 switch (Ty->getTag()) { 2159 case dwarf::DW_TAG_class_type: 2160 case dwarf::DW_TAG_structure_type: 2161 case dwarf::DW_TAG_union_type: 2162 case dwarf::DW_TAG_enumeration_type: 2163 break; 2164 default: 2165 return; 2166 } 2167 2168 if (const auto *File = Ty->getFile()) { 2169 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2170 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2171 2172 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2173 TypeTable.writeLeafType(USLR); 2174 } 2175 } 2176 2177 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2178 ClassOptions CO = getCommonClassOptions(Ty); 2179 TypeIndex FTI; 2180 unsigned EnumeratorCount = 0; 2181 2182 if (Ty->isForwardDecl()) { 2183 CO |= ClassOptions::ForwardReference; 2184 } else { 2185 ContinuationRecordBuilder ContinuationBuilder; 2186 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2187 for (const DINode *Element : Ty->getElements()) { 2188 // We assume that the frontend provides all members in source declaration 2189 // order, which is what MSVC does. 2190 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2191 // FIXME: Is it correct to always emit these as unsigned here? 2192 EnumeratorRecord ER(MemberAccess::Public, 2193 APSInt(Enumerator->getValue(), true), 2194 Enumerator->getName()); 2195 ContinuationBuilder.writeMemberType(ER); 2196 EnumeratorCount++; 2197 } 2198 } 2199 FTI = TypeTable.insertRecord(ContinuationBuilder); 2200 } 2201 2202 std::string FullName = getFullyQualifiedName(Ty); 2203 2204 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2205 getTypeIndex(Ty->getBaseType())); 2206 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2207 2208 addUDTSrcLine(Ty, EnumTI); 2209 2210 return EnumTI; 2211 } 2212 2213 //===----------------------------------------------------------------------===// 2214 // ClassInfo 2215 //===----------------------------------------------------------------------===// 2216 2217 struct llvm::ClassInfo { 2218 struct MemberInfo { 2219 const DIDerivedType *MemberTypeNode; 2220 uint64_t BaseOffset; 2221 }; 2222 // [MemberInfo] 2223 using MemberList = std::vector<MemberInfo>; 2224 2225 using MethodsList = TinyPtrVector<const DISubprogram *>; 2226 // MethodName -> MethodsList 2227 using MethodsMap = MapVector<MDString *, MethodsList>; 2228 2229 /// Base classes. 2230 std::vector<const DIDerivedType *> Inheritance; 2231 2232 /// Direct members. 2233 MemberList Members; 2234 // Direct overloaded methods gathered by name. 2235 MethodsMap Methods; 2236 2237 TypeIndex VShapeTI; 2238 2239 std::vector<const DIType *> NestedTypes; 2240 }; 2241 2242 void CodeViewDebug::clear() { 2243 assert(CurFn == nullptr); 2244 FileIdMap.clear(); 2245 FnDebugInfo.clear(); 2246 FileToFilepathMap.clear(); 2247 LocalUDTs.clear(); 2248 GlobalUDTs.clear(); 2249 TypeIndices.clear(); 2250 CompleteTypeIndices.clear(); 2251 ScopeGlobals.clear(); 2252 CVGlobalVariableOffsets.clear(); 2253 } 2254 2255 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2256 const DIDerivedType *DDTy) { 2257 if (!DDTy->getName().empty()) { 2258 Info.Members.push_back({DDTy, 0}); 2259 2260 // Collect static const data members with values. 2261 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2262 DINode::FlagStaticMember) { 2263 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2264 isa<ConstantFP>(DDTy->getConstant()))) 2265 StaticConstMembers.push_back(DDTy); 2266 } 2267 2268 return; 2269 } 2270 2271 // An unnamed member may represent a nested struct or union. Attempt to 2272 // interpret the unnamed member as a DICompositeType possibly wrapped in 2273 // qualifier types. Add all the indirect fields to the current record if that 2274 // succeeds, and drop the member if that fails. 2275 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2276 uint64_t Offset = DDTy->getOffsetInBits(); 2277 const DIType *Ty = DDTy->getBaseType(); 2278 bool FullyResolved = false; 2279 while (!FullyResolved) { 2280 switch (Ty->getTag()) { 2281 case dwarf::DW_TAG_const_type: 2282 case dwarf::DW_TAG_volatile_type: 2283 // FIXME: we should apply the qualifier types to the indirect fields 2284 // rather than dropping them. 2285 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2286 break; 2287 default: 2288 FullyResolved = true; 2289 break; 2290 } 2291 } 2292 2293 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2294 if (!DCTy) 2295 return; 2296 2297 ClassInfo NestedInfo = collectClassInfo(DCTy); 2298 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2299 Info.Members.push_back( 2300 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2301 } 2302 2303 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2304 ClassInfo Info; 2305 // Add elements to structure type. 2306 DINodeArray Elements = Ty->getElements(); 2307 for (auto *Element : Elements) { 2308 // We assume that the frontend provides all members in source declaration 2309 // order, which is what MSVC does. 2310 if (!Element) 2311 continue; 2312 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2313 Info.Methods[SP->getRawName()].push_back(SP); 2314 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2315 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2316 collectMemberInfo(Info, DDTy); 2317 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2318 Info.Inheritance.push_back(DDTy); 2319 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2320 DDTy->getName() == "__vtbl_ptr_type") { 2321 Info.VShapeTI = getTypeIndex(DDTy); 2322 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2323 Info.NestedTypes.push_back(DDTy); 2324 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2325 // Ignore friend members. It appears that MSVC emitted info about 2326 // friends in the past, but modern versions do not. 2327 } 2328 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2329 Info.NestedTypes.push_back(Composite); 2330 } 2331 // Skip other unrecognized kinds of elements. 2332 } 2333 return Info; 2334 } 2335 2336 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2337 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2338 // if a complete type should be emitted instead of a forward reference. 2339 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2340 !Ty->isForwardDecl(); 2341 } 2342 2343 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2344 // Emit the complete type for unnamed structs. C++ classes with methods 2345 // which have a circular reference back to the class type are expected to 2346 // be named by the front-end and should not be "unnamed". C unnamed 2347 // structs should not have circular references. 2348 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2349 // If this unnamed complete type is already in the process of being defined 2350 // then the description of the type is malformed and cannot be emitted 2351 // into CodeView correctly so report a fatal error. 2352 auto I = CompleteTypeIndices.find(Ty); 2353 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2354 report_fatal_error("cannot debug circular reference to unnamed type"); 2355 return getCompleteTypeIndex(Ty); 2356 } 2357 2358 // First, construct the forward decl. Don't look into Ty to compute the 2359 // forward decl options, since it might not be available in all TUs. 2360 TypeRecordKind Kind = getRecordKind(Ty); 2361 ClassOptions CO = 2362 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2363 std::string FullName = getFullyQualifiedName(Ty); 2364 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2365 FullName, Ty->getIdentifier()); 2366 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2367 if (!Ty->isForwardDecl()) 2368 DeferredCompleteTypes.push_back(Ty); 2369 return FwdDeclTI; 2370 } 2371 2372 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2373 // Construct the field list and complete type record. 2374 TypeRecordKind Kind = getRecordKind(Ty); 2375 ClassOptions CO = getCommonClassOptions(Ty); 2376 TypeIndex FieldTI; 2377 TypeIndex VShapeTI; 2378 unsigned FieldCount; 2379 bool ContainsNestedClass; 2380 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2381 lowerRecordFieldList(Ty); 2382 2383 if (ContainsNestedClass) 2384 CO |= ClassOptions::ContainsNestedClass; 2385 2386 // MSVC appears to set this flag by searching any destructor or method with 2387 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2388 // the members, however special member functions are not yet emitted into 2389 // debug information. For now checking a class's non-triviality seems enough. 2390 // FIXME: not true for a nested unnamed struct. 2391 if (isNonTrivial(Ty)) 2392 CO |= ClassOptions::HasConstructorOrDestructor; 2393 2394 std::string FullName = getFullyQualifiedName(Ty); 2395 2396 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2397 2398 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2399 SizeInBytes, FullName, Ty->getIdentifier()); 2400 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2401 2402 addUDTSrcLine(Ty, ClassTI); 2403 2404 addToUDTs(Ty); 2405 2406 return ClassTI; 2407 } 2408 2409 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2410 // Emit the complete type for unnamed unions. 2411 if (shouldAlwaysEmitCompleteClassType(Ty)) 2412 return getCompleteTypeIndex(Ty); 2413 2414 ClassOptions CO = 2415 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2416 std::string FullName = getFullyQualifiedName(Ty); 2417 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2418 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2419 if (!Ty->isForwardDecl()) 2420 DeferredCompleteTypes.push_back(Ty); 2421 return FwdDeclTI; 2422 } 2423 2424 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2425 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2426 TypeIndex FieldTI; 2427 unsigned FieldCount; 2428 bool ContainsNestedClass; 2429 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2430 lowerRecordFieldList(Ty); 2431 2432 if (ContainsNestedClass) 2433 CO |= ClassOptions::ContainsNestedClass; 2434 2435 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2436 std::string FullName = getFullyQualifiedName(Ty); 2437 2438 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2439 Ty->getIdentifier()); 2440 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2441 2442 addUDTSrcLine(Ty, UnionTI); 2443 2444 addToUDTs(Ty); 2445 2446 return UnionTI; 2447 } 2448 2449 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2450 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2451 // Manually count members. MSVC appears to count everything that generates a 2452 // field list record. Each individual overload in a method overload group 2453 // contributes to this count, even though the overload group is a single field 2454 // list record. 2455 unsigned MemberCount = 0; 2456 ClassInfo Info = collectClassInfo(Ty); 2457 ContinuationRecordBuilder ContinuationBuilder; 2458 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2459 2460 // Create base classes. 2461 for (const DIDerivedType *I : Info.Inheritance) { 2462 if (I->getFlags() & DINode::FlagVirtual) { 2463 // Virtual base. 2464 unsigned VBPtrOffset = I->getVBPtrOffset(); 2465 // FIXME: Despite the accessor name, the offset is really in bytes. 2466 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2467 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2468 ? TypeRecordKind::IndirectVirtualBaseClass 2469 : TypeRecordKind::VirtualBaseClass; 2470 VirtualBaseClassRecord VBCR( 2471 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2472 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2473 VBTableIndex); 2474 2475 ContinuationBuilder.writeMemberType(VBCR); 2476 MemberCount++; 2477 } else { 2478 assert(I->getOffsetInBits() % 8 == 0 && 2479 "bases must be on byte boundaries"); 2480 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2481 getTypeIndex(I->getBaseType()), 2482 I->getOffsetInBits() / 8); 2483 ContinuationBuilder.writeMemberType(BCR); 2484 MemberCount++; 2485 } 2486 } 2487 2488 // Create members. 2489 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2490 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2491 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2492 StringRef MemberName = Member->getName(); 2493 MemberAccess Access = 2494 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2495 2496 if (Member->isStaticMember()) { 2497 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2498 ContinuationBuilder.writeMemberType(SDMR); 2499 MemberCount++; 2500 continue; 2501 } 2502 2503 // Virtual function pointer member. 2504 if ((Member->getFlags() & DINode::FlagArtificial) && 2505 Member->getName().startswith("_vptr$")) { 2506 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2507 ContinuationBuilder.writeMemberType(VFPR); 2508 MemberCount++; 2509 continue; 2510 } 2511 2512 // Data member. 2513 uint64_t MemberOffsetInBits = 2514 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2515 if (Member->isBitField()) { 2516 uint64_t StartBitOffset = MemberOffsetInBits; 2517 if (const auto *CI = 2518 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2519 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2520 } 2521 StartBitOffset -= MemberOffsetInBits; 2522 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2523 StartBitOffset); 2524 MemberBaseType = TypeTable.writeLeafType(BFR); 2525 } 2526 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2527 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2528 MemberName); 2529 ContinuationBuilder.writeMemberType(DMR); 2530 MemberCount++; 2531 } 2532 2533 // Create methods 2534 for (auto &MethodItr : Info.Methods) { 2535 StringRef Name = MethodItr.first->getString(); 2536 2537 std::vector<OneMethodRecord> Methods; 2538 for (const DISubprogram *SP : MethodItr.second) { 2539 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2540 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2541 2542 unsigned VFTableOffset = -1; 2543 if (Introduced) 2544 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2545 2546 Methods.push_back(OneMethodRecord( 2547 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2548 translateMethodKindFlags(SP, Introduced), 2549 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2550 MemberCount++; 2551 } 2552 assert(!Methods.empty() && "Empty methods map entry"); 2553 if (Methods.size() == 1) 2554 ContinuationBuilder.writeMemberType(Methods[0]); 2555 else { 2556 // FIXME: Make this use its own ContinuationBuilder so that 2557 // MethodOverloadList can be split correctly. 2558 MethodOverloadListRecord MOLR(Methods); 2559 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2560 2561 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2562 ContinuationBuilder.writeMemberType(OMR); 2563 } 2564 } 2565 2566 // Create nested classes. 2567 for (const DIType *Nested : Info.NestedTypes) { 2568 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2569 ContinuationBuilder.writeMemberType(R); 2570 MemberCount++; 2571 } 2572 2573 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2574 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2575 !Info.NestedTypes.empty()); 2576 } 2577 2578 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2579 if (!VBPType.getIndex()) { 2580 // Make a 'const int *' type. 2581 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2582 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2583 2584 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2585 : PointerKind::Near32; 2586 PointerMode PM = PointerMode::Pointer; 2587 PointerOptions PO = PointerOptions::None; 2588 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2589 VBPType = TypeTable.writeLeafType(PR); 2590 } 2591 2592 return VBPType; 2593 } 2594 2595 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2596 // The null DIType is the void type. Don't try to hash it. 2597 if (!Ty) 2598 return TypeIndex::Void(); 2599 2600 // Check if we've already translated this type. Don't try to do a 2601 // get-or-create style insertion that caches the hash lookup across the 2602 // lowerType call. It will update the TypeIndices map. 2603 auto I = TypeIndices.find({Ty, ClassTy}); 2604 if (I != TypeIndices.end()) 2605 return I->second; 2606 2607 TypeLoweringScope S(*this); 2608 TypeIndex TI = lowerType(Ty, ClassTy); 2609 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2610 } 2611 2612 codeview::TypeIndex 2613 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2614 const DISubroutineType *SubroutineTy) { 2615 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2616 "this type must be a pointer type"); 2617 2618 PointerOptions Options = PointerOptions::None; 2619 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2620 Options = PointerOptions::LValueRefThisPointer; 2621 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2622 Options = PointerOptions::RValueRefThisPointer; 2623 2624 // Check if we've already translated this type. If there is no ref qualifier 2625 // on the function then we look up this pointer type with no associated class 2626 // so that the TypeIndex for the this pointer can be shared with the type 2627 // index for other pointers to this class type. If there is a ref qualifier 2628 // then we lookup the pointer using the subroutine as the parent type. 2629 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2630 if (I != TypeIndices.end()) 2631 return I->second; 2632 2633 TypeLoweringScope S(*this); 2634 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2635 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2636 } 2637 2638 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2639 PointerRecord PR(getTypeIndex(Ty), 2640 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2641 : PointerKind::Near32, 2642 PointerMode::LValueReference, PointerOptions::None, 2643 Ty->getSizeInBits() / 8); 2644 return TypeTable.writeLeafType(PR); 2645 } 2646 2647 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2648 // The null DIType is the void type. Don't try to hash it. 2649 if (!Ty) 2650 return TypeIndex::Void(); 2651 2652 // Look through typedefs when getting the complete type index. Call 2653 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2654 // emitted only once. 2655 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2656 (void)getTypeIndex(Ty); 2657 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2658 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2659 2660 // If this is a non-record type, the complete type index is the same as the 2661 // normal type index. Just call getTypeIndex. 2662 switch (Ty->getTag()) { 2663 case dwarf::DW_TAG_class_type: 2664 case dwarf::DW_TAG_structure_type: 2665 case dwarf::DW_TAG_union_type: 2666 break; 2667 default: 2668 return getTypeIndex(Ty); 2669 } 2670 2671 const auto *CTy = cast<DICompositeType>(Ty); 2672 2673 TypeLoweringScope S(*this); 2674 2675 // Make sure the forward declaration is emitted first. It's unclear if this 2676 // is necessary, but MSVC does it, and we should follow suit until we can show 2677 // otherwise. 2678 // We only emit a forward declaration for named types. 2679 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2680 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2681 2682 // Just use the forward decl if we don't have complete type info. This 2683 // might happen if the frontend is using modules and expects the complete 2684 // definition to be emitted elsewhere. 2685 if (CTy->isForwardDecl()) 2686 return FwdDeclTI; 2687 } 2688 2689 // Check if we've already translated the complete record type. 2690 // Insert the type with a null TypeIndex to signify that the type is currently 2691 // being lowered. 2692 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2693 if (!InsertResult.second) 2694 return InsertResult.first->second; 2695 2696 TypeIndex TI; 2697 switch (CTy->getTag()) { 2698 case dwarf::DW_TAG_class_type: 2699 case dwarf::DW_TAG_structure_type: 2700 TI = lowerCompleteTypeClass(CTy); 2701 break; 2702 case dwarf::DW_TAG_union_type: 2703 TI = lowerCompleteTypeUnion(CTy); 2704 break; 2705 default: 2706 llvm_unreachable("not a record"); 2707 } 2708 2709 // Update the type index associated with this CompositeType. This cannot 2710 // use the 'InsertResult' iterator above because it is potentially 2711 // invalidated by map insertions which can occur while lowering the class 2712 // type above. 2713 CompleteTypeIndices[CTy] = TI; 2714 return TI; 2715 } 2716 2717 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2718 /// and do this until fixpoint, as each complete record type typically 2719 /// references 2720 /// many other record types. 2721 void CodeViewDebug::emitDeferredCompleteTypes() { 2722 SmallVector<const DICompositeType *, 4> TypesToEmit; 2723 while (!DeferredCompleteTypes.empty()) { 2724 std::swap(DeferredCompleteTypes, TypesToEmit); 2725 for (const DICompositeType *RecordTy : TypesToEmit) 2726 getCompleteTypeIndex(RecordTy); 2727 TypesToEmit.clear(); 2728 } 2729 } 2730 2731 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2732 ArrayRef<LocalVariable> Locals) { 2733 // Get the sorted list of parameters and emit them first. 2734 SmallVector<const LocalVariable *, 6> Params; 2735 for (const LocalVariable &L : Locals) 2736 if (L.DIVar->isParameter()) 2737 Params.push_back(&L); 2738 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2739 return L->DIVar->getArg() < R->DIVar->getArg(); 2740 }); 2741 for (const LocalVariable *L : Params) 2742 emitLocalVariable(FI, *L); 2743 2744 // Next emit all non-parameters in the order that we found them. 2745 for (const LocalVariable &L : Locals) 2746 if (!L.DIVar->isParameter()) 2747 emitLocalVariable(FI, L); 2748 } 2749 2750 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2751 const LocalVariable &Var) { 2752 // LocalSym record, see SymbolRecord.h for more info. 2753 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2754 2755 LocalSymFlags Flags = LocalSymFlags::None; 2756 if (Var.DIVar->isParameter()) 2757 Flags |= LocalSymFlags::IsParameter; 2758 if (Var.DefRanges.empty()) 2759 Flags |= LocalSymFlags::IsOptimizedOut; 2760 2761 OS.AddComment("TypeIndex"); 2762 TypeIndex TI = Var.UseReferenceType 2763 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2764 : getCompleteTypeIndex(Var.DIVar->getType()); 2765 OS.emitInt32(TI.getIndex()); 2766 OS.AddComment("Flags"); 2767 OS.emitInt16(static_cast<uint16_t>(Flags)); 2768 // Truncate the name so we won't overflow the record length field. 2769 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2770 endSymbolRecord(LocalEnd); 2771 2772 // Calculate the on disk prefix of the appropriate def range record. The 2773 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2774 // should be big enough to hold all forms without memory allocation. 2775 SmallString<20> BytePrefix; 2776 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2777 BytePrefix.clear(); 2778 if (DefRange.InMemory) { 2779 int Offset = DefRange.DataOffset; 2780 unsigned Reg = DefRange.CVRegister; 2781 2782 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2783 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2784 // instead. In frames without stack realignment, $T0 will be the CFA. 2785 if (RegisterId(Reg) == RegisterId::ESP) { 2786 Reg = unsigned(RegisterId::VFRAME); 2787 Offset += FI.OffsetAdjustment; 2788 } 2789 2790 // If we can use the chosen frame pointer for the frame and this isn't a 2791 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2792 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2793 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2794 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2795 (bool(Flags & LocalSymFlags::IsParameter) 2796 ? (EncFP == FI.EncodedParamFramePtrReg) 2797 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2798 DefRangeFramePointerRelHeader DRHdr; 2799 DRHdr.Offset = Offset; 2800 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2801 } else { 2802 uint16_t RegRelFlags = 0; 2803 if (DefRange.IsSubfield) { 2804 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2805 (DefRange.StructOffset 2806 << DefRangeRegisterRelSym::OffsetInParentShift); 2807 } 2808 DefRangeRegisterRelHeader DRHdr; 2809 DRHdr.Register = Reg; 2810 DRHdr.Flags = RegRelFlags; 2811 DRHdr.BasePointerOffset = Offset; 2812 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2813 } 2814 } else { 2815 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2816 if (DefRange.IsSubfield) { 2817 DefRangeSubfieldRegisterHeader DRHdr; 2818 DRHdr.Register = DefRange.CVRegister; 2819 DRHdr.MayHaveNoName = 0; 2820 DRHdr.OffsetInParent = DefRange.StructOffset; 2821 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2822 } else { 2823 DefRangeRegisterHeader DRHdr; 2824 DRHdr.Register = DefRange.CVRegister; 2825 DRHdr.MayHaveNoName = 0; 2826 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2827 } 2828 } 2829 } 2830 } 2831 2832 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2833 const FunctionInfo& FI) { 2834 for (LexicalBlock *Block : Blocks) 2835 emitLexicalBlock(*Block, FI); 2836 } 2837 2838 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2839 /// lexical block scope. 2840 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2841 const FunctionInfo& FI) { 2842 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2843 OS.AddComment("PtrParent"); 2844 OS.emitInt32(0); // PtrParent 2845 OS.AddComment("PtrEnd"); 2846 OS.emitInt32(0); // PtrEnd 2847 OS.AddComment("Code size"); 2848 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2849 OS.AddComment("Function section relative address"); 2850 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2851 OS.AddComment("Function section index"); 2852 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2853 OS.AddComment("Lexical block name"); 2854 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2855 endSymbolRecord(RecordEnd); 2856 2857 // Emit variables local to this lexical block. 2858 emitLocalVariableList(FI, Block.Locals); 2859 emitGlobalVariableList(Block.Globals); 2860 2861 // Emit lexical blocks contained within this block. 2862 emitLexicalBlockList(Block.Children, FI); 2863 2864 // Close the lexical block scope. 2865 emitEndSymbolRecord(SymbolKind::S_END); 2866 } 2867 2868 /// Convenience routine for collecting lexical block information for a list 2869 /// of lexical scopes. 2870 void CodeViewDebug::collectLexicalBlockInfo( 2871 SmallVectorImpl<LexicalScope *> &Scopes, 2872 SmallVectorImpl<LexicalBlock *> &Blocks, 2873 SmallVectorImpl<LocalVariable> &Locals, 2874 SmallVectorImpl<CVGlobalVariable> &Globals) { 2875 for (LexicalScope *Scope : Scopes) 2876 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2877 } 2878 2879 /// Populate the lexical blocks and local variable lists of the parent with 2880 /// information about the specified lexical scope. 2881 void CodeViewDebug::collectLexicalBlockInfo( 2882 LexicalScope &Scope, 2883 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2884 SmallVectorImpl<LocalVariable> &ParentLocals, 2885 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2886 if (Scope.isAbstractScope()) 2887 return; 2888 2889 // Gather information about the lexical scope including local variables, 2890 // global variables, and address ranges. 2891 bool IgnoreScope = false; 2892 auto LI = ScopeVariables.find(&Scope); 2893 SmallVectorImpl<LocalVariable> *Locals = 2894 LI != ScopeVariables.end() ? &LI->second : nullptr; 2895 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2896 SmallVectorImpl<CVGlobalVariable> *Globals = 2897 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2898 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2899 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2900 2901 // Ignore lexical scopes which do not contain variables. 2902 if (!Locals && !Globals) 2903 IgnoreScope = true; 2904 2905 // Ignore lexical scopes which are not lexical blocks. 2906 if (!DILB) 2907 IgnoreScope = true; 2908 2909 // Ignore scopes which have too many address ranges to represent in the 2910 // current CodeView format or do not have a valid address range. 2911 // 2912 // For lexical scopes with multiple address ranges you may be tempted to 2913 // construct a single range covering every instruction where the block is 2914 // live and everything in between. Unfortunately, Visual Studio only 2915 // displays variables from the first matching lexical block scope. If the 2916 // first lexical block contains exception handling code or cold code which 2917 // is moved to the bottom of the routine creating a single range covering 2918 // nearly the entire routine, then it will hide all other lexical blocks 2919 // and the variables they contain. 2920 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2921 IgnoreScope = true; 2922 2923 if (IgnoreScope) { 2924 // This scope can be safely ignored and eliminating it will reduce the 2925 // size of the debug information. Be sure to collect any variable and scope 2926 // information from the this scope or any of its children and collapse them 2927 // into the parent scope. 2928 if (Locals) 2929 ParentLocals.append(Locals->begin(), Locals->end()); 2930 if (Globals) 2931 ParentGlobals.append(Globals->begin(), Globals->end()); 2932 collectLexicalBlockInfo(Scope.getChildren(), 2933 ParentBlocks, 2934 ParentLocals, 2935 ParentGlobals); 2936 return; 2937 } 2938 2939 // Create a new CodeView lexical block for this lexical scope. If we've 2940 // seen this DILexicalBlock before then the scope tree is malformed and 2941 // we can handle this gracefully by not processing it a second time. 2942 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2943 if (!BlockInsertion.second) 2944 return; 2945 2946 // Create a lexical block containing the variables and collect the the 2947 // lexical block information for the children. 2948 const InsnRange &Range = Ranges.front(); 2949 assert(Range.first && Range.second); 2950 LexicalBlock &Block = BlockInsertion.first->second; 2951 Block.Begin = getLabelBeforeInsn(Range.first); 2952 Block.End = getLabelAfterInsn(Range.second); 2953 assert(Block.Begin && "missing label for scope begin"); 2954 assert(Block.End && "missing label for scope end"); 2955 Block.Name = DILB->getName(); 2956 if (Locals) 2957 Block.Locals = std::move(*Locals); 2958 if (Globals) 2959 Block.Globals = std::move(*Globals); 2960 ParentBlocks.push_back(&Block); 2961 collectLexicalBlockInfo(Scope.getChildren(), 2962 Block.Children, 2963 Block.Locals, 2964 Block.Globals); 2965 } 2966 2967 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2968 const Function &GV = MF->getFunction(); 2969 assert(FnDebugInfo.count(&GV)); 2970 assert(CurFn == FnDebugInfo[&GV].get()); 2971 2972 collectVariableInfo(GV.getSubprogram()); 2973 2974 // Build the lexical block structure to emit for this routine. 2975 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2976 collectLexicalBlockInfo(*CFS, 2977 CurFn->ChildBlocks, 2978 CurFn->Locals, 2979 CurFn->Globals); 2980 2981 // Clear the scope and variable information from the map which will not be 2982 // valid after we have finished processing this routine. This also prepares 2983 // the map for the subsequent routine. 2984 ScopeVariables.clear(); 2985 2986 // Don't emit anything if we don't have any line tables. 2987 // Thunks are compiler-generated and probably won't have source correlation. 2988 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2989 FnDebugInfo.erase(&GV); 2990 CurFn = nullptr; 2991 return; 2992 } 2993 2994 // Find heap alloc sites and add to list. 2995 for (const auto &MBB : *MF) { 2996 for (const auto &MI : MBB) { 2997 if (MDNode *MD = MI.getHeapAllocMarker()) { 2998 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 2999 getLabelAfterInsn(&MI), 3000 dyn_cast<DIType>(MD))); 3001 } 3002 } 3003 } 3004 3005 CurFn->Annotations = MF->getCodeViewAnnotations(); 3006 3007 CurFn->End = Asm->getFunctionEnd(); 3008 3009 CurFn = nullptr; 3010 } 3011 3012 // Usable locations are valid with non-zero line numbers. A line number of zero 3013 // corresponds to optimized code that doesn't have a distinct source location. 3014 // In this case, we try to use the previous or next source location depending on 3015 // the context. 3016 static bool isUsableDebugLoc(DebugLoc DL) { 3017 return DL && DL.getLine() != 0; 3018 } 3019 3020 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3021 DebugHandlerBase::beginInstruction(MI); 3022 3023 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3024 if (!Asm || !CurFn || MI->isDebugInstr() || 3025 MI->getFlag(MachineInstr::FrameSetup)) 3026 return; 3027 3028 // If the first instruction of a new MBB has no location, find the first 3029 // instruction with a location and use that. 3030 DebugLoc DL = MI->getDebugLoc(); 3031 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3032 for (const auto &NextMI : *MI->getParent()) { 3033 if (NextMI.isDebugInstr()) 3034 continue; 3035 DL = NextMI.getDebugLoc(); 3036 if (isUsableDebugLoc(DL)) 3037 break; 3038 } 3039 // FIXME: Handle the case where the BB has no valid locations. This would 3040 // probably require doing a real dataflow analysis. 3041 } 3042 PrevInstBB = MI->getParent(); 3043 3044 // If we still don't have a debug location, don't record a location. 3045 if (!isUsableDebugLoc(DL)) 3046 return; 3047 3048 maybeRecordLocation(DL, Asm->MF); 3049 } 3050 3051 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3052 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3053 *EndLabel = MMI->getContext().createTempSymbol(); 3054 OS.emitInt32(unsigned(Kind)); 3055 OS.AddComment("Subsection size"); 3056 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3057 OS.emitLabel(BeginLabel); 3058 return EndLabel; 3059 } 3060 3061 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3062 OS.emitLabel(EndLabel); 3063 // Every subsection must be aligned to a 4-byte boundary. 3064 OS.emitValueToAlignment(4); 3065 } 3066 3067 static StringRef getSymbolName(SymbolKind SymKind) { 3068 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3069 if (EE.Value == SymKind) 3070 return EE.Name; 3071 return ""; 3072 } 3073 3074 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3075 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3076 *EndLabel = MMI->getContext().createTempSymbol(); 3077 OS.AddComment("Record length"); 3078 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3079 OS.emitLabel(BeginLabel); 3080 if (OS.isVerboseAsm()) 3081 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3082 OS.emitInt16(unsigned(SymKind)); 3083 return EndLabel; 3084 } 3085 3086 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3087 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3088 // an extra copy of every symbol record in LLD. This increases object file 3089 // size by less than 1% in the clang build, and is compatible with the Visual 3090 // C++ linker. 3091 OS.emitValueToAlignment(4); 3092 OS.emitLabel(SymEnd); 3093 } 3094 3095 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3096 OS.AddComment("Record length"); 3097 OS.emitInt16(2); 3098 if (OS.isVerboseAsm()) 3099 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3100 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3101 } 3102 3103 void CodeViewDebug::emitDebugInfoForUDTs( 3104 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3105 #ifndef NDEBUG 3106 size_t OriginalSize = UDTs.size(); 3107 #endif 3108 for (const auto &UDT : UDTs) { 3109 const DIType *T = UDT.second; 3110 assert(shouldEmitUdt(T)); 3111 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3112 OS.AddComment("Type"); 3113 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3114 assert(OriginalSize == UDTs.size() && 3115 "getCompleteTypeIndex found new UDTs!"); 3116 emitNullTerminatedSymbolName(OS, UDT.first); 3117 endSymbolRecord(UDTRecordEnd); 3118 } 3119 } 3120 3121 void CodeViewDebug::collectGlobalVariableInfo() { 3122 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3123 GlobalMap; 3124 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3125 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3126 GV.getDebugInfo(GVEs); 3127 for (const auto *GVE : GVEs) 3128 GlobalMap[GVE] = &GV; 3129 } 3130 3131 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3132 for (const MDNode *Node : CUs->operands()) { 3133 const auto *CU = cast<DICompileUnit>(Node); 3134 for (const auto *GVE : CU->getGlobalVariables()) { 3135 const DIGlobalVariable *DIGV = GVE->getVariable(); 3136 const DIExpression *DIE = GVE->getExpression(); 3137 3138 if ((DIE->getNumElements() == 2) && 3139 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3140 // Record the constant offset for the variable. 3141 // 3142 // A Fortran common block uses this idiom to encode the offset 3143 // of a variable from the common block's starting address. 3144 CVGlobalVariableOffsets.insert( 3145 std::make_pair(DIGV, DIE->getElement(1))); 3146 3147 // Emit constant global variables in a global symbol section. 3148 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3149 CVGlobalVariable CVGV = {DIGV, DIE}; 3150 GlobalVariables.emplace_back(std::move(CVGV)); 3151 } 3152 3153 const auto *GV = GlobalMap.lookup(GVE); 3154 if (!GV || GV->isDeclarationForLinker()) 3155 continue; 3156 3157 DIScope *Scope = DIGV->getScope(); 3158 SmallVector<CVGlobalVariable, 1> *VariableList; 3159 if (Scope && isa<DILocalScope>(Scope)) { 3160 // Locate a global variable list for this scope, creating one if 3161 // necessary. 3162 auto Insertion = ScopeGlobals.insert( 3163 {Scope, std::unique_ptr<GlobalVariableList>()}); 3164 if (Insertion.second) 3165 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3166 VariableList = Insertion.first->second.get(); 3167 } else if (GV->hasComdat()) 3168 // Emit this global variable into a COMDAT section. 3169 VariableList = &ComdatVariables; 3170 else 3171 // Emit this global variable in a single global symbol section. 3172 VariableList = &GlobalVariables; 3173 CVGlobalVariable CVGV = {DIGV, GV}; 3174 VariableList->emplace_back(std::move(CVGV)); 3175 } 3176 } 3177 } 3178 3179 void CodeViewDebug::collectDebugInfoForGlobals() { 3180 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3181 const DIGlobalVariable *DIGV = CVGV.DIGV; 3182 const DIScope *Scope = DIGV->getScope(); 3183 getCompleteTypeIndex(DIGV->getType()); 3184 getFullyQualifiedName(Scope, DIGV->getName()); 3185 } 3186 3187 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3188 const DIGlobalVariable *DIGV = CVGV.DIGV; 3189 const DIScope *Scope = DIGV->getScope(); 3190 getCompleteTypeIndex(DIGV->getType()); 3191 getFullyQualifiedName(Scope, DIGV->getName()); 3192 } 3193 } 3194 3195 void CodeViewDebug::emitDebugInfoForGlobals() { 3196 // First, emit all globals that are not in a comdat in a single symbol 3197 // substream. MSVC doesn't like it if the substream is empty, so only open 3198 // it if we have at least one global to emit. 3199 switchToDebugSectionForSymbol(nullptr); 3200 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3201 OS.AddComment("Symbol subsection for globals"); 3202 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3203 emitGlobalVariableList(GlobalVariables); 3204 emitStaticConstMemberList(); 3205 endCVSubsection(EndLabel); 3206 } 3207 3208 // Second, emit each global that is in a comdat into its own .debug$S 3209 // section along with its own symbol substream. 3210 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3211 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3212 MCSymbol *GVSym = Asm->getSymbol(GV); 3213 OS.AddComment("Symbol subsection for " + 3214 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3215 switchToDebugSectionForSymbol(GVSym); 3216 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3217 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3218 emitDebugInfoForGlobal(CVGV); 3219 endCVSubsection(EndLabel); 3220 } 3221 } 3222 3223 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3224 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3225 for (const MDNode *Node : CUs->operands()) { 3226 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3227 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3228 getTypeIndex(RT); 3229 // FIXME: Add to global/local DTU list. 3230 } 3231 } 3232 } 3233 } 3234 3235 // Emit each global variable in the specified array. 3236 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3237 for (const CVGlobalVariable &CVGV : Globals) { 3238 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3239 emitDebugInfoForGlobal(CVGV); 3240 } 3241 } 3242 3243 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3244 const std::string &QualifiedName) { 3245 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3246 OS.AddComment("Type"); 3247 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3248 3249 OS.AddComment("Value"); 3250 3251 // Encoded integers shouldn't need more than 10 bytes. 3252 uint8_t Data[10]; 3253 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3254 CodeViewRecordIO IO(Writer); 3255 cantFail(IO.mapEncodedInteger(Value)); 3256 StringRef SRef((char *)Data, Writer.getOffset()); 3257 OS.emitBinaryData(SRef); 3258 3259 OS.AddComment("Name"); 3260 emitNullTerminatedSymbolName(OS, QualifiedName); 3261 endSymbolRecord(SConstantEnd); 3262 } 3263 3264 void CodeViewDebug::emitStaticConstMemberList() { 3265 for (const DIDerivedType *DTy : StaticConstMembers) { 3266 const DIScope *Scope = DTy->getScope(); 3267 3268 APSInt Value; 3269 if (const ConstantInt *CI = 3270 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3271 Value = APSInt(CI->getValue(), 3272 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3273 else if (const ConstantFP *CFP = 3274 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3275 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3276 else 3277 llvm_unreachable("cannot emit a constant without a value"); 3278 3279 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3280 getFullyQualifiedName(Scope, DTy->getName())); 3281 } 3282 } 3283 3284 static bool isFloatDIType(const DIType *Ty) { 3285 if (isa<DICompositeType>(Ty)) 3286 return false; 3287 3288 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3289 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3290 if (T == dwarf::DW_TAG_pointer_type || 3291 T == dwarf::DW_TAG_ptr_to_member_type || 3292 T == dwarf::DW_TAG_reference_type || 3293 T == dwarf::DW_TAG_rvalue_reference_type) 3294 return false; 3295 assert(DTy->getBaseType() && "Expected valid base type"); 3296 return isFloatDIType(DTy->getBaseType()); 3297 } 3298 3299 auto *BTy = cast<DIBasicType>(Ty); 3300 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3301 } 3302 3303 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3304 const DIGlobalVariable *DIGV = CVGV.DIGV; 3305 3306 const DIScope *Scope = DIGV->getScope(); 3307 // For static data members, get the scope from the declaration. 3308 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3309 DIGV->getRawStaticDataMemberDeclaration())) 3310 Scope = MemberDecl->getScope(); 3311 // For Fortran, the scoping portion is elided in its name so that we can 3312 // reference the variable in the command line of the VS debugger. 3313 std::string QualifiedName = 3314 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3315 : getFullyQualifiedName(Scope, DIGV->getName()); 3316 3317 if (const GlobalVariable *GV = 3318 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3319 // DataSym record, see SymbolRecord.h for more info. Thread local data 3320 // happens to have the same format as global data. 3321 MCSymbol *GVSym = Asm->getSymbol(GV); 3322 SymbolKind DataSym = GV->isThreadLocal() 3323 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3324 : SymbolKind::S_GTHREAD32) 3325 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3326 : SymbolKind::S_GDATA32); 3327 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3328 OS.AddComment("Type"); 3329 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3330 OS.AddComment("DataOffset"); 3331 3332 uint64_t Offset = 0; 3333 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3334 // Use the offset seen while collecting info on globals. 3335 Offset = CVGlobalVariableOffsets[DIGV]; 3336 OS.EmitCOFFSecRel32(GVSym, Offset); 3337 3338 OS.AddComment("Segment"); 3339 OS.EmitCOFFSectionIndex(GVSym); 3340 OS.AddComment("Name"); 3341 const unsigned LengthOfDataRecord = 12; 3342 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3343 endSymbolRecord(DataEnd); 3344 } else { 3345 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3346 assert(DIE->isConstant() && 3347 "Global constant variables must contain a constant expression."); 3348 3349 // Use unsigned for floats. 3350 bool isUnsigned = isFloatDIType(DIGV->getType()) 3351 ? true 3352 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3353 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3354 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3355 } 3356 } 3357