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 N = std::min<int>(N, std::numeric_limits<uint16_t>::max()); 862 OS.emitInt16(N); 863 } 864 865 // Some Microsoft tools, like Binscope, expect a backend version number of at 866 // least 8.something, so we'll coerce the LLVM version into a form that 867 // guarantees it'll be big enough without really lying about the version. 868 int Major = 1000 * LLVM_VERSION_MAJOR + 869 10 * LLVM_VERSION_MINOR + 870 LLVM_VERSION_PATCH; 871 // Clamp it for builds that use unusually large version numbers. 872 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 873 Version BackVer = {{ Major, 0, 0, 0 }}; 874 OS.AddComment("Backend version"); 875 for (int N : BackVer.Part) 876 OS.emitInt16(N); 877 878 OS.AddComment("Null-terminated compiler version string"); 879 emitNullTerminatedSymbolName(OS, CompilerVersion); 880 881 endSymbolRecord(CompilerEnd); 882 } 883 884 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 885 StringRef S) { 886 StringIdRecord SIR(TypeIndex(0x0), S); 887 return TypeTable.writeLeafType(SIR); 888 } 889 890 void CodeViewDebug::emitBuildInfo() { 891 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 892 // build info. The known prefix is: 893 // - Absolute path of current directory 894 // - Compiler path 895 // - Main source file path, relative to CWD or absolute 896 // - Type server PDB file 897 // - Canonical compiler command line 898 // If frontend and backend compilation are separated (think llc or LTO), it's 899 // not clear if the compiler path should refer to the executable for the 900 // frontend or the backend. Leave it blank for now. 901 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 902 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 903 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 904 const auto *CU = cast<DICompileUnit>(Node); 905 const DIFile *MainSourceFile = CU->getFile(); 906 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 907 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 908 BuildInfoArgs[BuildInfoRecord::SourceFile] = 909 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 910 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 911 // we implement /Zi type servers. 912 BuildInfoRecord BIR(BuildInfoArgs); 913 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 914 915 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 916 // from the module symbols into the type stream. 917 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 918 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 919 OS.AddComment("LF_BUILDINFO index"); 920 OS.emitInt32(BuildInfoIndex.getIndex()); 921 endSymbolRecord(BIEnd); 922 endCVSubsection(BISubsecEnd); 923 } 924 925 void CodeViewDebug::emitInlineeLinesSubsection() { 926 if (InlinedSubprograms.empty()) 927 return; 928 929 OS.AddComment("Inlinee lines subsection"); 930 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 931 932 // We emit the checksum info for files. This is used by debuggers to 933 // determine if a pdb matches the source before loading it. Visual Studio, 934 // for instance, will display a warning that the breakpoints are not valid if 935 // the pdb does not match the source. 936 OS.AddComment("Inlinee lines signature"); 937 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 938 939 for (const DISubprogram *SP : InlinedSubprograms) { 940 assert(TypeIndices.count({SP, nullptr})); 941 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 942 943 OS.AddBlankLine(); 944 unsigned FileId = maybeRecordFile(SP->getFile()); 945 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 946 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 947 OS.AddBlankLine(); 948 OS.AddComment("Type index of inlined function"); 949 OS.emitInt32(InlineeIdx.getIndex()); 950 OS.AddComment("Offset into filechecksum table"); 951 OS.emitCVFileChecksumOffsetDirective(FileId); 952 OS.AddComment("Starting line number"); 953 OS.emitInt32(SP->getLine()); 954 } 955 956 endCVSubsection(InlineEnd); 957 } 958 959 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 960 const DILocation *InlinedAt, 961 const InlineSite &Site) { 962 assert(TypeIndices.count({Site.Inlinee, nullptr})); 963 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 964 965 // SymbolRecord 966 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 967 968 OS.AddComment("PtrParent"); 969 OS.emitInt32(0); 970 OS.AddComment("PtrEnd"); 971 OS.emitInt32(0); 972 OS.AddComment("Inlinee type index"); 973 OS.emitInt32(InlineeIdx.getIndex()); 974 975 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 976 unsigned StartLineNum = Site.Inlinee->getLine(); 977 978 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 979 FI.Begin, FI.End); 980 981 endSymbolRecord(InlineEnd); 982 983 emitLocalVariableList(FI, Site.InlinedLocals); 984 985 // Recurse on child inlined call sites before closing the scope. 986 for (const DILocation *ChildSite : Site.ChildSites) { 987 auto I = FI.InlineSites.find(ChildSite); 988 assert(I != FI.InlineSites.end() && 989 "child site not in function inline site map"); 990 emitInlinedCallSite(FI, ChildSite, I->second); 991 } 992 993 // Close the scope. 994 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 995 } 996 997 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 998 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 999 // comdat key. A section may be comdat because of -ffunction-sections or 1000 // because it is comdat in the IR. 1001 MCSectionCOFF *GVSec = 1002 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1003 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1004 1005 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1006 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1007 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1008 1009 OS.SwitchSection(DebugSec); 1010 1011 // Emit the magic version number if this is the first time we've switched to 1012 // this section. 1013 if (ComdatDebugSections.insert(DebugSec).second) 1014 emitCodeViewMagicVersion(); 1015 } 1016 1017 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1018 // The only supported thunk ordinal is currently the standard type. 1019 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1020 FunctionInfo &FI, 1021 const MCSymbol *Fn) { 1022 std::string FuncName = 1023 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1024 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1025 1026 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1027 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1028 1029 // Emit S_THUNK32 1030 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1031 OS.AddComment("PtrParent"); 1032 OS.emitInt32(0); 1033 OS.AddComment("PtrEnd"); 1034 OS.emitInt32(0); 1035 OS.AddComment("PtrNext"); 1036 OS.emitInt32(0); 1037 OS.AddComment("Thunk section relative address"); 1038 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1039 OS.AddComment("Thunk section index"); 1040 OS.EmitCOFFSectionIndex(Fn); 1041 OS.AddComment("Code size"); 1042 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1043 OS.AddComment("Ordinal"); 1044 OS.emitInt8(unsigned(ordinal)); 1045 OS.AddComment("Function name"); 1046 emitNullTerminatedSymbolName(OS, FuncName); 1047 // Additional fields specific to the thunk ordinal would go here. 1048 endSymbolRecord(ThunkRecordEnd); 1049 1050 // Local variables/inlined routines are purposely omitted here. The point of 1051 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1052 1053 // Emit S_PROC_ID_END 1054 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1055 1056 endCVSubsection(SymbolsEnd); 1057 } 1058 1059 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1060 FunctionInfo &FI) { 1061 // For each function there is a separate subsection which holds the PC to 1062 // file:line table. 1063 const MCSymbol *Fn = Asm->getSymbol(GV); 1064 assert(Fn); 1065 1066 // Switch to the to a comdat section, if appropriate. 1067 switchToDebugSectionForSymbol(Fn); 1068 1069 std::string FuncName; 1070 auto *SP = GV->getSubprogram(); 1071 assert(SP); 1072 setCurrentSubprogram(SP); 1073 1074 if (SP->isThunk()) { 1075 emitDebugInfoForThunk(GV, FI, Fn); 1076 return; 1077 } 1078 1079 // If we have a display name, build the fully qualified name by walking the 1080 // chain of scopes. 1081 if (!SP->getName().empty()) 1082 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1083 1084 // If our DISubprogram name is empty, use the mangled name. 1085 if (FuncName.empty()) 1086 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1087 1088 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1089 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1090 OS.EmitCVFPOData(Fn); 1091 1092 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1093 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1094 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1095 { 1096 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1097 : SymbolKind::S_GPROC32_ID; 1098 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1099 1100 // These fields are filled in by tools like CVPACK which run after the fact. 1101 OS.AddComment("PtrParent"); 1102 OS.emitInt32(0); 1103 OS.AddComment("PtrEnd"); 1104 OS.emitInt32(0); 1105 OS.AddComment("PtrNext"); 1106 OS.emitInt32(0); 1107 // This is the important bit that tells the debugger where the function 1108 // code is located and what's its size: 1109 OS.AddComment("Code size"); 1110 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1111 OS.AddComment("Offset after prologue"); 1112 OS.emitInt32(0); 1113 OS.AddComment("Offset before epilogue"); 1114 OS.emitInt32(0); 1115 OS.AddComment("Function type index"); 1116 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1117 OS.AddComment("Function section relative address"); 1118 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1119 OS.AddComment("Function section index"); 1120 OS.EmitCOFFSectionIndex(Fn); 1121 OS.AddComment("Flags"); 1122 OS.emitInt8(0); 1123 // Emit the function display name as a null-terminated string. 1124 OS.AddComment("Function name"); 1125 // Truncate the name so we won't overflow the record length field. 1126 emitNullTerminatedSymbolName(OS, FuncName); 1127 endSymbolRecord(ProcRecordEnd); 1128 1129 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1130 // Subtract out the CSR size since MSVC excludes that and we include it. 1131 OS.AddComment("FrameSize"); 1132 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1133 OS.AddComment("Padding"); 1134 OS.emitInt32(0); 1135 OS.AddComment("Offset of padding"); 1136 OS.emitInt32(0); 1137 OS.AddComment("Bytes of callee saved registers"); 1138 OS.emitInt32(FI.CSRSize); 1139 OS.AddComment("Exception handler offset"); 1140 OS.emitInt32(0); 1141 OS.AddComment("Exception handler section"); 1142 OS.emitInt16(0); 1143 OS.AddComment("Flags (defines frame register)"); 1144 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1145 endSymbolRecord(FrameProcEnd); 1146 1147 emitLocalVariableList(FI, FI.Locals); 1148 emitGlobalVariableList(FI.Globals); 1149 emitLexicalBlockList(FI.ChildBlocks, FI); 1150 1151 // Emit inlined call site information. Only emit functions inlined directly 1152 // into the parent function. We'll emit the other sites recursively as part 1153 // of their parent inline site. 1154 for (const DILocation *InlinedAt : FI.ChildSites) { 1155 auto I = FI.InlineSites.find(InlinedAt); 1156 assert(I != FI.InlineSites.end() && 1157 "child site not in function inline site map"); 1158 emitInlinedCallSite(FI, InlinedAt, I->second); 1159 } 1160 1161 for (auto Annot : FI.Annotations) { 1162 MCSymbol *Label = Annot.first; 1163 MDTuple *Strs = cast<MDTuple>(Annot.second); 1164 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1165 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1166 // FIXME: Make sure we don't overflow the max record size. 1167 OS.EmitCOFFSectionIndex(Label); 1168 OS.emitInt16(Strs->getNumOperands()); 1169 for (Metadata *MD : Strs->operands()) { 1170 // MDStrings are null terminated, so we can do EmitBytes and get the 1171 // nice .asciz directive. 1172 StringRef Str = cast<MDString>(MD)->getString(); 1173 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1174 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1175 } 1176 endSymbolRecord(AnnotEnd); 1177 } 1178 1179 for (auto HeapAllocSite : FI.HeapAllocSites) { 1180 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1181 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1182 const DIType *DITy = std::get<2>(HeapAllocSite); 1183 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1184 OS.AddComment("Call site offset"); 1185 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1186 OS.AddComment("Call site section index"); 1187 OS.EmitCOFFSectionIndex(BeginLabel); 1188 OS.AddComment("Call instruction length"); 1189 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1190 OS.AddComment("Type index"); 1191 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1192 endSymbolRecord(HeapAllocEnd); 1193 } 1194 1195 if (SP != nullptr) 1196 emitDebugInfoForUDTs(LocalUDTs); 1197 1198 // We're done with this function. 1199 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1200 } 1201 endCVSubsection(SymbolsEnd); 1202 1203 // We have an assembler directive that takes care of the whole line table. 1204 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1205 } 1206 1207 CodeViewDebug::LocalVarDefRange 1208 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1209 LocalVarDefRange DR; 1210 DR.InMemory = -1; 1211 DR.DataOffset = Offset; 1212 assert(DR.DataOffset == Offset && "truncation"); 1213 DR.IsSubfield = 0; 1214 DR.StructOffset = 0; 1215 DR.CVRegister = CVRegister; 1216 return DR; 1217 } 1218 1219 void CodeViewDebug::collectVariableInfoFromMFTable( 1220 DenseSet<InlinedEntity> &Processed) { 1221 const MachineFunction &MF = *Asm->MF; 1222 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1223 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1224 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1225 1226 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1227 if (!VI.Var) 1228 continue; 1229 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1230 "Expected inlined-at fields to agree"); 1231 1232 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1233 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1234 1235 // If variable scope is not found then skip this variable. 1236 if (!Scope) 1237 continue; 1238 1239 // If the variable has an attached offset expression, extract it. 1240 // FIXME: Try to handle DW_OP_deref as well. 1241 int64_t ExprOffset = 0; 1242 bool Deref = false; 1243 if (VI.Expr) { 1244 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1245 if (VI.Expr->getNumElements() == 1 && 1246 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1247 Deref = true; 1248 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1249 continue; 1250 } 1251 1252 // Get the frame register used and the offset. 1253 Register FrameReg; 1254 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1255 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1256 1257 assert(!FrameOffset.getScalable() && 1258 "Frame offsets with a scalable component are not supported"); 1259 1260 // Calculate the label ranges. 1261 LocalVarDefRange DefRange = 1262 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1263 1264 for (const InsnRange &Range : Scope->getRanges()) { 1265 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1266 const MCSymbol *End = getLabelAfterInsn(Range.second); 1267 End = End ? End : Asm->getFunctionEnd(); 1268 DefRange.Ranges.emplace_back(Begin, End); 1269 } 1270 1271 LocalVariable Var; 1272 Var.DIVar = VI.Var; 1273 Var.DefRanges.emplace_back(std::move(DefRange)); 1274 if (Deref) 1275 Var.UseReferenceType = true; 1276 1277 recordLocalVariable(std::move(Var), Scope); 1278 } 1279 } 1280 1281 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1282 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1283 } 1284 1285 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1286 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1287 } 1288 1289 void CodeViewDebug::calculateRanges( 1290 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1291 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1292 1293 // Calculate the definition ranges. 1294 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1295 const auto &Entry = *I; 1296 if (!Entry.isDbgValue()) 1297 continue; 1298 const MachineInstr *DVInst = Entry.getInstr(); 1299 assert(DVInst->isDebugValue() && "Invalid History entry"); 1300 // FIXME: Find a way to represent constant variables, since they are 1301 // relatively common. 1302 Optional<DbgVariableLocation> Location = 1303 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1304 if (!Location) 1305 continue; 1306 1307 // CodeView can only express variables in register and variables in memory 1308 // at a constant offset from a register. However, for variables passed 1309 // indirectly by pointer, it is common for that pointer to be spilled to a 1310 // stack location. For the special case of one offseted load followed by a 1311 // zero offset load (a pointer spilled to the stack), we change the type of 1312 // the local variable from a value type to a reference type. This tricks the 1313 // debugger into doing the load for us. 1314 if (Var.UseReferenceType) { 1315 // We're using a reference type. Drop the last zero offset load. 1316 if (canUseReferenceType(*Location)) 1317 Location->LoadChain.pop_back(); 1318 else 1319 continue; 1320 } else if (needsReferenceType(*Location)) { 1321 // This location can't be expressed without switching to a reference type. 1322 // Start over using that. 1323 Var.UseReferenceType = true; 1324 Var.DefRanges.clear(); 1325 calculateRanges(Var, Entries); 1326 return; 1327 } 1328 1329 // We can only handle a register or an offseted load of a register. 1330 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1331 continue; 1332 { 1333 LocalVarDefRange DR; 1334 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1335 DR.InMemory = !Location->LoadChain.empty(); 1336 DR.DataOffset = 1337 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1338 if (Location->FragmentInfo) { 1339 DR.IsSubfield = true; 1340 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1341 } else { 1342 DR.IsSubfield = false; 1343 DR.StructOffset = 0; 1344 } 1345 1346 if (Var.DefRanges.empty() || 1347 Var.DefRanges.back().isDifferentLocation(DR)) { 1348 Var.DefRanges.emplace_back(std::move(DR)); 1349 } 1350 } 1351 1352 // Compute the label range. 1353 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1354 const MCSymbol *End; 1355 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1356 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1357 End = EndingEntry.isDbgValue() 1358 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1359 : getLabelAfterInsn(EndingEntry.getInstr()); 1360 } else 1361 End = Asm->getFunctionEnd(); 1362 1363 // If the last range end is our begin, just extend the last range. 1364 // Otherwise make a new range. 1365 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1366 Var.DefRanges.back().Ranges; 1367 if (!R.empty() && R.back().second == Begin) 1368 R.back().second = End; 1369 else 1370 R.emplace_back(Begin, End); 1371 1372 // FIXME: Do more range combining. 1373 } 1374 } 1375 1376 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1377 DenseSet<InlinedEntity> Processed; 1378 // Grab the variable info that was squirreled away in the MMI side-table. 1379 collectVariableInfoFromMFTable(Processed); 1380 1381 for (const auto &I : DbgValues) { 1382 InlinedEntity IV = I.first; 1383 if (Processed.count(IV)) 1384 continue; 1385 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1386 const DILocation *InlinedAt = IV.second; 1387 1388 // Instruction ranges, specifying where IV is accessible. 1389 const auto &Entries = I.second; 1390 1391 LexicalScope *Scope = nullptr; 1392 if (InlinedAt) 1393 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1394 else 1395 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1396 // If variable scope is not found then skip this variable. 1397 if (!Scope) 1398 continue; 1399 1400 LocalVariable Var; 1401 Var.DIVar = DIVar; 1402 1403 calculateRanges(Var, Entries); 1404 recordLocalVariable(std::move(Var), Scope); 1405 } 1406 } 1407 1408 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1409 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1410 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1411 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1412 const Function &GV = MF->getFunction(); 1413 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1414 assert(Insertion.second && "function already has info"); 1415 CurFn = Insertion.first->second.get(); 1416 CurFn->FuncId = NextFuncId++; 1417 CurFn->Begin = Asm->getFunctionBegin(); 1418 1419 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1420 // callee-saved registers were used. For targets that don't use a PUSH 1421 // instruction (AArch64), this will be zero. 1422 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1423 CurFn->FrameSize = MFI.getStackSize(); 1424 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1425 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1426 1427 // For this function S_FRAMEPROC record, figure out which codeview register 1428 // will be the frame pointer. 1429 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1430 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1431 if (CurFn->FrameSize > 0) { 1432 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1433 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1434 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1435 } else { 1436 // If there is an FP, parameters are always relative to it. 1437 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1438 if (CurFn->HasStackRealignment) { 1439 // If the stack needs realignment, locals are relative to SP or VFRAME. 1440 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1441 } else { 1442 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1443 // other stack adjustments. 1444 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1445 } 1446 } 1447 } 1448 1449 // Compute other frame procedure options. 1450 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1451 if (MFI.hasVarSizedObjects()) 1452 FPO |= FrameProcedureOptions::HasAlloca; 1453 if (MF->exposesReturnsTwice()) 1454 FPO |= FrameProcedureOptions::HasSetJmp; 1455 // FIXME: Set HasLongJmp if we ever track that info. 1456 if (MF->hasInlineAsm()) 1457 FPO |= FrameProcedureOptions::HasInlineAssembly; 1458 if (GV.hasPersonalityFn()) { 1459 if (isAsynchronousEHPersonality( 1460 classifyEHPersonality(GV.getPersonalityFn()))) 1461 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1462 else 1463 FPO |= FrameProcedureOptions::HasExceptionHandling; 1464 } 1465 if (GV.hasFnAttribute(Attribute::InlineHint)) 1466 FPO |= FrameProcedureOptions::MarkedInline; 1467 if (GV.hasFnAttribute(Attribute::Naked)) 1468 FPO |= FrameProcedureOptions::Naked; 1469 if (MFI.hasStackProtectorIndex()) 1470 FPO |= FrameProcedureOptions::SecurityChecks; 1471 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1472 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1473 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1474 !GV.hasOptSize() && !GV.hasOptNone()) 1475 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1476 if (GV.hasProfileData()) { 1477 FPO |= FrameProcedureOptions::ValidProfileCounts; 1478 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1479 } 1480 // FIXME: Set GuardCfg when it is implemented. 1481 CurFn->FrameProcOpts = FPO; 1482 1483 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1484 1485 // Find the end of the function prolog. First known non-DBG_VALUE and 1486 // non-frame setup location marks the beginning of the function body. 1487 // FIXME: is there a simpler a way to do this? Can we just search 1488 // for the first instruction of the function, not the last of the prolog? 1489 DebugLoc PrologEndLoc; 1490 bool EmptyPrologue = true; 1491 for (const auto &MBB : *MF) { 1492 for (const auto &MI : MBB) { 1493 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1494 MI.getDebugLoc()) { 1495 PrologEndLoc = MI.getDebugLoc(); 1496 break; 1497 } else if (!MI.isMetaInstruction()) { 1498 EmptyPrologue = false; 1499 } 1500 } 1501 } 1502 1503 // Record beginning of function if we have a non-empty prologue. 1504 if (PrologEndLoc && !EmptyPrologue) { 1505 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1506 maybeRecordLocation(FnStartDL, MF); 1507 } 1508 1509 // Find heap alloc sites and emit labels around them. 1510 for (const auto &MBB : *MF) { 1511 for (const auto &MI : MBB) { 1512 if (MI.getHeapAllocMarker()) { 1513 requestLabelBeforeInsn(&MI); 1514 requestLabelAfterInsn(&MI); 1515 } 1516 } 1517 } 1518 } 1519 1520 static bool shouldEmitUdt(const DIType *T) { 1521 if (!T) 1522 return false; 1523 1524 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1525 if (T->getTag() == dwarf::DW_TAG_typedef) { 1526 if (DIScope *Scope = T->getScope()) { 1527 switch (Scope->getTag()) { 1528 case dwarf::DW_TAG_structure_type: 1529 case dwarf::DW_TAG_class_type: 1530 case dwarf::DW_TAG_union_type: 1531 return false; 1532 default: 1533 // do nothing. 1534 ; 1535 } 1536 } 1537 } 1538 1539 while (true) { 1540 if (!T || T->isForwardDecl()) 1541 return false; 1542 1543 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1544 if (!DT) 1545 return true; 1546 T = DT->getBaseType(); 1547 } 1548 return true; 1549 } 1550 1551 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1552 // Don't record empty UDTs. 1553 if (Ty->getName().empty()) 1554 return; 1555 if (!shouldEmitUdt(Ty)) 1556 return; 1557 1558 SmallVector<StringRef, 5> ParentScopeNames; 1559 const DISubprogram *ClosestSubprogram = 1560 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1561 1562 std::string FullyQualifiedName = 1563 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1564 1565 if (ClosestSubprogram == nullptr) { 1566 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1567 } else if (ClosestSubprogram == CurrentSubprogram) { 1568 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1569 } 1570 1571 // TODO: What if the ClosestSubprogram is neither null or the current 1572 // subprogram? Currently, the UDT just gets dropped on the floor. 1573 // 1574 // The current behavior is not desirable. To get maximal fidelity, we would 1575 // need to perform all type translation before beginning emission of .debug$S 1576 // and then make LocalUDTs a member of FunctionInfo 1577 } 1578 1579 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1580 // Generic dispatch for lowering an unknown type. 1581 switch (Ty->getTag()) { 1582 case dwarf::DW_TAG_array_type: 1583 return lowerTypeArray(cast<DICompositeType>(Ty)); 1584 case dwarf::DW_TAG_typedef: 1585 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1586 case dwarf::DW_TAG_base_type: 1587 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1588 case dwarf::DW_TAG_pointer_type: 1589 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1590 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1591 LLVM_FALLTHROUGH; 1592 case dwarf::DW_TAG_reference_type: 1593 case dwarf::DW_TAG_rvalue_reference_type: 1594 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1595 case dwarf::DW_TAG_ptr_to_member_type: 1596 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1597 case dwarf::DW_TAG_restrict_type: 1598 case dwarf::DW_TAG_const_type: 1599 case dwarf::DW_TAG_volatile_type: 1600 // TODO: add support for DW_TAG_atomic_type here 1601 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1602 case dwarf::DW_TAG_subroutine_type: 1603 if (ClassTy) { 1604 // The member function type of a member function pointer has no 1605 // ThisAdjustment. 1606 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1607 /*ThisAdjustment=*/0, 1608 /*IsStaticMethod=*/false); 1609 } 1610 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1611 case dwarf::DW_TAG_enumeration_type: 1612 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1613 case dwarf::DW_TAG_class_type: 1614 case dwarf::DW_TAG_structure_type: 1615 return lowerTypeClass(cast<DICompositeType>(Ty)); 1616 case dwarf::DW_TAG_union_type: 1617 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1618 case dwarf::DW_TAG_string_type: 1619 return lowerTypeString(cast<DIStringType>(Ty)); 1620 case dwarf::DW_TAG_unspecified_type: 1621 if (Ty->getName() == "decltype(nullptr)") 1622 return TypeIndex::NullptrT(); 1623 return TypeIndex::None(); 1624 default: 1625 // Use the null type index. 1626 return TypeIndex(); 1627 } 1628 } 1629 1630 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1631 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1632 StringRef TypeName = Ty->getName(); 1633 1634 addToUDTs(Ty); 1635 1636 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1637 TypeName == "HRESULT") 1638 return TypeIndex(SimpleTypeKind::HResult); 1639 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1640 TypeName == "wchar_t") 1641 return TypeIndex(SimpleTypeKind::WideCharacter); 1642 1643 return UnderlyingTypeIndex; 1644 } 1645 1646 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1647 const DIType *ElementType = Ty->getBaseType(); 1648 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1649 // IndexType is size_t, which depends on the bitness of the target. 1650 TypeIndex IndexType = getPointerSizeInBytes() == 8 1651 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1652 : TypeIndex(SimpleTypeKind::UInt32Long); 1653 1654 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1655 1656 // Add subranges to array type. 1657 DINodeArray Elements = Ty->getElements(); 1658 for (int i = Elements.size() - 1; i >= 0; --i) { 1659 const DINode *Element = Elements[i]; 1660 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1661 1662 const DISubrange *Subrange = cast<DISubrange>(Element); 1663 int64_t Count = -1; 1664 1665 // If Subrange has a Count field, use it. 1666 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1667 // where lowerbound is from the LowerBound field of the Subrange, 1668 // or the language default lowerbound if that field is unspecified. 1669 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1670 Count = CI->getSExtValue(); 1671 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1672 // Fortran uses 1 as the default lowerbound; other languages use 0. 1673 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1674 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1675 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1676 Count = UI->getSExtValue() - Lowerbound + 1; 1677 } 1678 1679 // Forward declarations of arrays without a size and VLAs use a count of -1. 1680 // Emit a count of zero in these cases to match what MSVC does for arrays 1681 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1682 // should do for them even if we could distinguish them. 1683 if (Count == -1) 1684 Count = 0; 1685 1686 // Update the element size and element type index for subsequent subranges. 1687 ElementSize *= Count; 1688 1689 // If this is the outermost array, use the size from the array. It will be 1690 // more accurate if we had a VLA or an incomplete element type size. 1691 uint64_t ArraySize = 1692 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1693 1694 StringRef Name = (i == 0) ? Ty->getName() : ""; 1695 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1696 ElementTypeIndex = TypeTable.writeLeafType(AR); 1697 } 1698 1699 return ElementTypeIndex; 1700 } 1701 1702 // This function lowers a Fortran character type (DIStringType). 1703 // Note that it handles only the character*n variant (using SizeInBits 1704 // field in DIString to describe the type size) at the moment. 1705 // Other variants (leveraging the StringLength and StringLengthExp 1706 // fields in DIStringType) remain TBD. 1707 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1708 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1709 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1710 StringRef Name = Ty->getName(); 1711 // IndexType is size_t, which depends on the bitness of the target. 1712 TypeIndex IndexType = getPointerSizeInBytes() == 8 1713 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1714 : TypeIndex(SimpleTypeKind::UInt32Long); 1715 1716 // Create a type of character array of ArraySize. 1717 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1718 1719 return TypeTable.writeLeafType(AR); 1720 } 1721 1722 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1723 TypeIndex Index; 1724 dwarf::TypeKind Kind; 1725 uint32_t ByteSize; 1726 1727 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1728 ByteSize = Ty->getSizeInBits() / 8; 1729 1730 SimpleTypeKind STK = SimpleTypeKind::None; 1731 switch (Kind) { 1732 case dwarf::DW_ATE_address: 1733 // FIXME: Translate 1734 break; 1735 case dwarf::DW_ATE_boolean: 1736 switch (ByteSize) { 1737 case 1: STK = SimpleTypeKind::Boolean8; break; 1738 case 2: STK = SimpleTypeKind::Boolean16; break; 1739 case 4: STK = SimpleTypeKind::Boolean32; break; 1740 case 8: STK = SimpleTypeKind::Boolean64; break; 1741 case 16: STK = SimpleTypeKind::Boolean128; break; 1742 } 1743 break; 1744 case dwarf::DW_ATE_complex_float: 1745 switch (ByteSize) { 1746 case 2: STK = SimpleTypeKind::Complex16; break; 1747 case 4: STK = SimpleTypeKind::Complex32; break; 1748 case 8: STK = SimpleTypeKind::Complex64; break; 1749 case 10: STK = SimpleTypeKind::Complex80; break; 1750 case 16: STK = SimpleTypeKind::Complex128; break; 1751 } 1752 break; 1753 case dwarf::DW_ATE_float: 1754 switch (ByteSize) { 1755 case 2: STK = SimpleTypeKind::Float16; break; 1756 case 4: STK = SimpleTypeKind::Float32; break; 1757 case 6: STK = SimpleTypeKind::Float48; break; 1758 case 8: STK = SimpleTypeKind::Float64; break; 1759 case 10: STK = SimpleTypeKind::Float80; break; 1760 case 16: STK = SimpleTypeKind::Float128; break; 1761 } 1762 break; 1763 case dwarf::DW_ATE_signed: 1764 switch (ByteSize) { 1765 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1766 case 2: STK = SimpleTypeKind::Int16Short; break; 1767 case 4: STK = SimpleTypeKind::Int32; break; 1768 case 8: STK = SimpleTypeKind::Int64Quad; break; 1769 case 16: STK = SimpleTypeKind::Int128Oct; break; 1770 } 1771 break; 1772 case dwarf::DW_ATE_unsigned: 1773 switch (ByteSize) { 1774 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1775 case 2: STK = SimpleTypeKind::UInt16Short; break; 1776 case 4: STK = SimpleTypeKind::UInt32; break; 1777 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1778 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1779 } 1780 break; 1781 case dwarf::DW_ATE_UTF: 1782 switch (ByteSize) { 1783 case 2: STK = SimpleTypeKind::Character16; break; 1784 case 4: STK = SimpleTypeKind::Character32; break; 1785 } 1786 break; 1787 case dwarf::DW_ATE_signed_char: 1788 if (ByteSize == 1) 1789 STK = SimpleTypeKind::SignedCharacter; 1790 break; 1791 case dwarf::DW_ATE_unsigned_char: 1792 if (ByteSize == 1) 1793 STK = SimpleTypeKind::UnsignedCharacter; 1794 break; 1795 default: 1796 break; 1797 } 1798 1799 // Apply some fixups based on the source-level type name. 1800 // Include some amount of canonicalization from an old naming scheme Clang 1801 // used to use for integer types (in an outdated effort to be compatible with 1802 // GCC's debug info/GDB's behavior, which has since been addressed). 1803 if (STK == SimpleTypeKind::Int32 && 1804 (Ty->getName() == "long int" || Ty->getName() == "long")) 1805 STK = SimpleTypeKind::Int32Long; 1806 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1807 Ty->getName() == "unsigned long")) 1808 STK = SimpleTypeKind::UInt32Long; 1809 if (STK == SimpleTypeKind::UInt16Short && 1810 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1811 STK = SimpleTypeKind::WideCharacter; 1812 if ((STK == SimpleTypeKind::SignedCharacter || 1813 STK == SimpleTypeKind::UnsignedCharacter) && 1814 Ty->getName() == "char") 1815 STK = SimpleTypeKind::NarrowCharacter; 1816 1817 return TypeIndex(STK); 1818 } 1819 1820 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1821 PointerOptions PO) { 1822 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1823 1824 // Pointers to simple types without any options can use SimpleTypeMode, rather 1825 // than having a dedicated pointer type record. 1826 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1827 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1828 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1829 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1830 ? SimpleTypeMode::NearPointer64 1831 : SimpleTypeMode::NearPointer32; 1832 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1833 } 1834 1835 PointerKind PK = 1836 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1837 PointerMode PM = PointerMode::Pointer; 1838 switch (Ty->getTag()) { 1839 default: llvm_unreachable("not a pointer tag type"); 1840 case dwarf::DW_TAG_pointer_type: 1841 PM = PointerMode::Pointer; 1842 break; 1843 case dwarf::DW_TAG_reference_type: 1844 PM = PointerMode::LValueReference; 1845 break; 1846 case dwarf::DW_TAG_rvalue_reference_type: 1847 PM = PointerMode::RValueReference; 1848 break; 1849 } 1850 1851 if (Ty->isObjectPointer()) 1852 PO |= PointerOptions::Const; 1853 1854 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1855 return TypeTable.writeLeafType(PR); 1856 } 1857 1858 static PointerToMemberRepresentation 1859 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1860 // SizeInBytes being zero generally implies that the member pointer type was 1861 // incomplete, which can happen if it is part of a function prototype. In this 1862 // case, use the unknown model instead of the general model. 1863 if (IsPMF) { 1864 switch (Flags & DINode::FlagPtrToMemberRep) { 1865 case 0: 1866 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1867 : PointerToMemberRepresentation::GeneralFunction; 1868 case DINode::FlagSingleInheritance: 1869 return PointerToMemberRepresentation::SingleInheritanceFunction; 1870 case DINode::FlagMultipleInheritance: 1871 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1872 case DINode::FlagVirtualInheritance: 1873 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1874 } 1875 } else { 1876 switch (Flags & DINode::FlagPtrToMemberRep) { 1877 case 0: 1878 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1879 : PointerToMemberRepresentation::GeneralData; 1880 case DINode::FlagSingleInheritance: 1881 return PointerToMemberRepresentation::SingleInheritanceData; 1882 case DINode::FlagMultipleInheritance: 1883 return PointerToMemberRepresentation::MultipleInheritanceData; 1884 case DINode::FlagVirtualInheritance: 1885 return PointerToMemberRepresentation::VirtualInheritanceData; 1886 } 1887 } 1888 llvm_unreachable("invalid ptr to member representation"); 1889 } 1890 1891 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1892 PointerOptions PO) { 1893 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1894 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1895 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1896 TypeIndex PointeeTI = 1897 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1898 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1899 : PointerKind::Near32; 1900 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1901 : PointerMode::PointerToDataMember; 1902 1903 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1904 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1905 MemberPointerInfo MPI( 1906 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1907 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1908 return TypeTable.writeLeafType(PR); 1909 } 1910 1911 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1912 /// have a translation, use the NearC convention. 1913 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1914 switch (DwarfCC) { 1915 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1916 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1917 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1918 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1919 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1920 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1921 } 1922 return CallingConvention::NearC; 1923 } 1924 1925 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1926 ModifierOptions Mods = ModifierOptions::None; 1927 PointerOptions PO = PointerOptions::None; 1928 bool IsModifier = true; 1929 const DIType *BaseTy = Ty; 1930 while (IsModifier && BaseTy) { 1931 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1932 switch (BaseTy->getTag()) { 1933 case dwarf::DW_TAG_const_type: 1934 Mods |= ModifierOptions::Const; 1935 PO |= PointerOptions::Const; 1936 break; 1937 case dwarf::DW_TAG_volatile_type: 1938 Mods |= ModifierOptions::Volatile; 1939 PO |= PointerOptions::Volatile; 1940 break; 1941 case dwarf::DW_TAG_restrict_type: 1942 // Only pointer types be marked with __restrict. There is no known flag 1943 // for __restrict in LF_MODIFIER records. 1944 PO |= PointerOptions::Restrict; 1945 break; 1946 default: 1947 IsModifier = false; 1948 break; 1949 } 1950 if (IsModifier) 1951 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1952 } 1953 1954 // Check if the inner type will use an LF_POINTER record. If so, the 1955 // qualifiers will go in the LF_POINTER record. This comes up for types like 1956 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1957 // char *'. 1958 if (BaseTy) { 1959 switch (BaseTy->getTag()) { 1960 case dwarf::DW_TAG_pointer_type: 1961 case dwarf::DW_TAG_reference_type: 1962 case dwarf::DW_TAG_rvalue_reference_type: 1963 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1964 case dwarf::DW_TAG_ptr_to_member_type: 1965 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1966 default: 1967 break; 1968 } 1969 } 1970 1971 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1972 1973 // Return the base type index if there aren't any modifiers. For example, the 1974 // metadata could contain restrict wrappers around non-pointer types. 1975 if (Mods == ModifierOptions::None) 1976 return ModifiedTI; 1977 1978 ModifierRecord MR(ModifiedTI, Mods); 1979 return TypeTable.writeLeafType(MR); 1980 } 1981 1982 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1983 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1984 for (const DIType *ArgType : Ty->getTypeArray()) 1985 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1986 1987 // MSVC uses type none for variadic argument. 1988 if (ReturnAndArgTypeIndices.size() > 1 && 1989 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1990 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1991 } 1992 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1993 ArrayRef<TypeIndex> ArgTypeIndices = None; 1994 if (!ReturnAndArgTypeIndices.empty()) { 1995 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1996 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1997 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1998 } 1999 2000 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2001 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2002 2003 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2004 2005 FunctionOptions FO = getFunctionOptions(Ty); 2006 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2007 ArgListIndex); 2008 return TypeTable.writeLeafType(Procedure); 2009 } 2010 2011 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2012 const DIType *ClassTy, 2013 int ThisAdjustment, 2014 bool IsStaticMethod, 2015 FunctionOptions FO) { 2016 // Lower the containing class type. 2017 TypeIndex ClassType = getTypeIndex(ClassTy); 2018 2019 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2020 2021 unsigned Index = 0; 2022 SmallVector<TypeIndex, 8> ArgTypeIndices; 2023 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2024 if (ReturnAndArgs.size() > Index) { 2025 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2026 } 2027 2028 // If the first argument is a pointer type and this isn't a static method, 2029 // treat it as the special 'this' parameter, which is encoded separately from 2030 // the arguments. 2031 TypeIndex ThisTypeIndex; 2032 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2033 if (const DIDerivedType *PtrTy = 2034 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2035 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2036 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2037 Index++; 2038 } 2039 } 2040 } 2041 2042 while (Index < ReturnAndArgs.size()) 2043 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2044 2045 // MSVC uses type none for variadic argument. 2046 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2047 ArgTypeIndices.back() = TypeIndex::None(); 2048 2049 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2050 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2051 2052 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2053 2054 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2055 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2056 return TypeTable.writeLeafType(MFR); 2057 } 2058 2059 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2060 unsigned VSlotCount = 2061 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2062 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2063 2064 VFTableShapeRecord VFTSR(Slots); 2065 return TypeTable.writeLeafType(VFTSR); 2066 } 2067 2068 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2069 switch (Flags & DINode::FlagAccessibility) { 2070 case DINode::FlagPrivate: return MemberAccess::Private; 2071 case DINode::FlagPublic: return MemberAccess::Public; 2072 case DINode::FlagProtected: return MemberAccess::Protected; 2073 case 0: 2074 // If there was no explicit access control, provide the default for the tag. 2075 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2076 : MemberAccess::Public; 2077 } 2078 llvm_unreachable("access flags are exclusive"); 2079 } 2080 2081 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2082 if (SP->isArtificial()) 2083 return MethodOptions::CompilerGenerated; 2084 2085 // FIXME: Handle other MethodOptions. 2086 2087 return MethodOptions::None; 2088 } 2089 2090 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2091 bool Introduced) { 2092 if (SP->getFlags() & DINode::FlagStaticMember) 2093 return MethodKind::Static; 2094 2095 switch (SP->getVirtuality()) { 2096 case dwarf::DW_VIRTUALITY_none: 2097 break; 2098 case dwarf::DW_VIRTUALITY_virtual: 2099 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2100 case dwarf::DW_VIRTUALITY_pure_virtual: 2101 return Introduced ? MethodKind::PureIntroducingVirtual 2102 : MethodKind::PureVirtual; 2103 default: 2104 llvm_unreachable("unhandled virtuality case"); 2105 } 2106 2107 return MethodKind::Vanilla; 2108 } 2109 2110 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2111 switch (Ty->getTag()) { 2112 case dwarf::DW_TAG_class_type: 2113 return TypeRecordKind::Class; 2114 case dwarf::DW_TAG_structure_type: 2115 return TypeRecordKind::Struct; 2116 default: 2117 llvm_unreachable("unexpected tag"); 2118 } 2119 } 2120 2121 /// Return ClassOptions that should be present on both the forward declaration 2122 /// and the defintion of a tag type. 2123 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2124 ClassOptions CO = ClassOptions::None; 2125 2126 // MSVC always sets this flag, even for local types. Clang doesn't always 2127 // appear to give every type a linkage name, which may be problematic for us. 2128 // FIXME: Investigate the consequences of not following them here. 2129 if (!Ty->getIdentifier().empty()) 2130 CO |= ClassOptions::HasUniqueName; 2131 2132 // Put the Nested flag on a type if it appears immediately inside a tag type. 2133 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2134 // here. That flag is only set on definitions, and not forward declarations. 2135 const DIScope *ImmediateScope = Ty->getScope(); 2136 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2137 CO |= ClassOptions::Nested; 2138 2139 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2140 // type only when it has an immediate function scope. Clang never puts enums 2141 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2142 // always in function, class, or file scopes. 2143 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2144 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2145 CO |= ClassOptions::Scoped; 2146 } else { 2147 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2148 Scope = Scope->getScope()) { 2149 if (isa<DISubprogram>(Scope)) { 2150 CO |= ClassOptions::Scoped; 2151 break; 2152 } 2153 } 2154 } 2155 2156 return CO; 2157 } 2158 2159 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2160 switch (Ty->getTag()) { 2161 case dwarf::DW_TAG_class_type: 2162 case dwarf::DW_TAG_structure_type: 2163 case dwarf::DW_TAG_union_type: 2164 case dwarf::DW_TAG_enumeration_type: 2165 break; 2166 default: 2167 return; 2168 } 2169 2170 if (const auto *File = Ty->getFile()) { 2171 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2172 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2173 2174 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2175 TypeTable.writeLeafType(USLR); 2176 } 2177 } 2178 2179 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2180 ClassOptions CO = getCommonClassOptions(Ty); 2181 TypeIndex FTI; 2182 unsigned EnumeratorCount = 0; 2183 2184 if (Ty->isForwardDecl()) { 2185 CO |= ClassOptions::ForwardReference; 2186 } else { 2187 ContinuationRecordBuilder ContinuationBuilder; 2188 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2189 for (const DINode *Element : Ty->getElements()) { 2190 // We assume that the frontend provides all members in source declaration 2191 // order, which is what MSVC does. 2192 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2193 // FIXME: Is it correct to always emit these as unsigned here? 2194 EnumeratorRecord ER(MemberAccess::Public, 2195 APSInt(Enumerator->getValue(), true), 2196 Enumerator->getName()); 2197 ContinuationBuilder.writeMemberType(ER); 2198 EnumeratorCount++; 2199 } 2200 } 2201 FTI = TypeTable.insertRecord(ContinuationBuilder); 2202 } 2203 2204 std::string FullName = getFullyQualifiedName(Ty); 2205 2206 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2207 getTypeIndex(Ty->getBaseType())); 2208 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2209 2210 addUDTSrcLine(Ty, EnumTI); 2211 2212 return EnumTI; 2213 } 2214 2215 //===----------------------------------------------------------------------===// 2216 // ClassInfo 2217 //===----------------------------------------------------------------------===// 2218 2219 struct llvm::ClassInfo { 2220 struct MemberInfo { 2221 const DIDerivedType *MemberTypeNode; 2222 uint64_t BaseOffset; 2223 }; 2224 // [MemberInfo] 2225 using MemberList = std::vector<MemberInfo>; 2226 2227 using MethodsList = TinyPtrVector<const DISubprogram *>; 2228 // MethodName -> MethodsList 2229 using MethodsMap = MapVector<MDString *, MethodsList>; 2230 2231 /// Base classes. 2232 std::vector<const DIDerivedType *> Inheritance; 2233 2234 /// Direct members. 2235 MemberList Members; 2236 // Direct overloaded methods gathered by name. 2237 MethodsMap Methods; 2238 2239 TypeIndex VShapeTI; 2240 2241 std::vector<const DIType *> NestedTypes; 2242 }; 2243 2244 void CodeViewDebug::clear() { 2245 assert(CurFn == nullptr); 2246 FileIdMap.clear(); 2247 FnDebugInfo.clear(); 2248 FileToFilepathMap.clear(); 2249 LocalUDTs.clear(); 2250 GlobalUDTs.clear(); 2251 TypeIndices.clear(); 2252 CompleteTypeIndices.clear(); 2253 ScopeGlobals.clear(); 2254 CVGlobalVariableOffsets.clear(); 2255 } 2256 2257 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2258 const DIDerivedType *DDTy) { 2259 if (!DDTy->getName().empty()) { 2260 Info.Members.push_back({DDTy, 0}); 2261 2262 // Collect static const data members with values. 2263 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2264 DINode::FlagStaticMember) { 2265 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2266 isa<ConstantFP>(DDTy->getConstant()))) 2267 StaticConstMembers.push_back(DDTy); 2268 } 2269 2270 return; 2271 } 2272 2273 // An unnamed member may represent a nested struct or union. Attempt to 2274 // interpret the unnamed member as a DICompositeType possibly wrapped in 2275 // qualifier types. Add all the indirect fields to the current record if that 2276 // succeeds, and drop the member if that fails. 2277 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2278 uint64_t Offset = DDTy->getOffsetInBits(); 2279 const DIType *Ty = DDTy->getBaseType(); 2280 bool FullyResolved = false; 2281 while (!FullyResolved) { 2282 switch (Ty->getTag()) { 2283 case dwarf::DW_TAG_const_type: 2284 case dwarf::DW_TAG_volatile_type: 2285 // FIXME: we should apply the qualifier types to the indirect fields 2286 // rather than dropping them. 2287 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2288 break; 2289 default: 2290 FullyResolved = true; 2291 break; 2292 } 2293 } 2294 2295 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2296 if (!DCTy) 2297 return; 2298 2299 ClassInfo NestedInfo = collectClassInfo(DCTy); 2300 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2301 Info.Members.push_back( 2302 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2303 } 2304 2305 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2306 ClassInfo Info; 2307 // Add elements to structure type. 2308 DINodeArray Elements = Ty->getElements(); 2309 for (auto *Element : Elements) { 2310 // We assume that the frontend provides all members in source declaration 2311 // order, which is what MSVC does. 2312 if (!Element) 2313 continue; 2314 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2315 Info.Methods[SP->getRawName()].push_back(SP); 2316 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2317 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2318 collectMemberInfo(Info, DDTy); 2319 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2320 Info.Inheritance.push_back(DDTy); 2321 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2322 DDTy->getName() == "__vtbl_ptr_type") { 2323 Info.VShapeTI = getTypeIndex(DDTy); 2324 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2325 Info.NestedTypes.push_back(DDTy); 2326 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2327 // Ignore friend members. It appears that MSVC emitted info about 2328 // friends in the past, but modern versions do not. 2329 } 2330 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2331 Info.NestedTypes.push_back(Composite); 2332 } 2333 // Skip other unrecognized kinds of elements. 2334 } 2335 return Info; 2336 } 2337 2338 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2339 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2340 // if a complete type should be emitted instead of a forward reference. 2341 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2342 !Ty->isForwardDecl(); 2343 } 2344 2345 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2346 // Emit the complete type for unnamed structs. C++ classes with methods 2347 // which have a circular reference back to the class type are expected to 2348 // be named by the front-end and should not be "unnamed". C unnamed 2349 // structs should not have circular references. 2350 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2351 // If this unnamed complete type is already in the process of being defined 2352 // then the description of the type is malformed and cannot be emitted 2353 // into CodeView correctly so report a fatal error. 2354 auto I = CompleteTypeIndices.find(Ty); 2355 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2356 report_fatal_error("cannot debug circular reference to unnamed type"); 2357 return getCompleteTypeIndex(Ty); 2358 } 2359 2360 // First, construct the forward decl. Don't look into Ty to compute the 2361 // forward decl options, since it might not be available in all TUs. 2362 TypeRecordKind Kind = getRecordKind(Ty); 2363 ClassOptions CO = 2364 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2365 std::string FullName = getFullyQualifiedName(Ty); 2366 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2367 FullName, Ty->getIdentifier()); 2368 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2369 if (!Ty->isForwardDecl()) 2370 DeferredCompleteTypes.push_back(Ty); 2371 return FwdDeclTI; 2372 } 2373 2374 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2375 // Construct the field list and complete type record. 2376 TypeRecordKind Kind = getRecordKind(Ty); 2377 ClassOptions CO = getCommonClassOptions(Ty); 2378 TypeIndex FieldTI; 2379 TypeIndex VShapeTI; 2380 unsigned FieldCount; 2381 bool ContainsNestedClass; 2382 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2383 lowerRecordFieldList(Ty); 2384 2385 if (ContainsNestedClass) 2386 CO |= ClassOptions::ContainsNestedClass; 2387 2388 // MSVC appears to set this flag by searching any destructor or method with 2389 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2390 // the members, however special member functions are not yet emitted into 2391 // debug information. For now checking a class's non-triviality seems enough. 2392 // FIXME: not true for a nested unnamed struct. 2393 if (isNonTrivial(Ty)) 2394 CO |= ClassOptions::HasConstructorOrDestructor; 2395 2396 std::string FullName = getFullyQualifiedName(Ty); 2397 2398 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2399 2400 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2401 SizeInBytes, FullName, Ty->getIdentifier()); 2402 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2403 2404 addUDTSrcLine(Ty, ClassTI); 2405 2406 addToUDTs(Ty); 2407 2408 return ClassTI; 2409 } 2410 2411 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2412 // Emit the complete type for unnamed unions. 2413 if (shouldAlwaysEmitCompleteClassType(Ty)) 2414 return getCompleteTypeIndex(Ty); 2415 2416 ClassOptions CO = 2417 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2418 std::string FullName = getFullyQualifiedName(Ty); 2419 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2420 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2421 if (!Ty->isForwardDecl()) 2422 DeferredCompleteTypes.push_back(Ty); 2423 return FwdDeclTI; 2424 } 2425 2426 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2427 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2428 TypeIndex FieldTI; 2429 unsigned FieldCount; 2430 bool ContainsNestedClass; 2431 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2432 lowerRecordFieldList(Ty); 2433 2434 if (ContainsNestedClass) 2435 CO |= ClassOptions::ContainsNestedClass; 2436 2437 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2438 std::string FullName = getFullyQualifiedName(Ty); 2439 2440 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2441 Ty->getIdentifier()); 2442 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2443 2444 addUDTSrcLine(Ty, UnionTI); 2445 2446 addToUDTs(Ty); 2447 2448 return UnionTI; 2449 } 2450 2451 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2452 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2453 // Manually count members. MSVC appears to count everything that generates a 2454 // field list record. Each individual overload in a method overload group 2455 // contributes to this count, even though the overload group is a single field 2456 // list record. 2457 unsigned MemberCount = 0; 2458 ClassInfo Info = collectClassInfo(Ty); 2459 ContinuationRecordBuilder ContinuationBuilder; 2460 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2461 2462 // Create base classes. 2463 for (const DIDerivedType *I : Info.Inheritance) { 2464 if (I->getFlags() & DINode::FlagVirtual) { 2465 // Virtual base. 2466 unsigned VBPtrOffset = I->getVBPtrOffset(); 2467 // FIXME: Despite the accessor name, the offset is really in bytes. 2468 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2469 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2470 ? TypeRecordKind::IndirectVirtualBaseClass 2471 : TypeRecordKind::VirtualBaseClass; 2472 VirtualBaseClassRecord VBCR( 2473 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2474 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2475 VBTableIndex); 2476 2477 ContinuationBuilder.writeMemberType(VBCR); 2478 MemberCount++; 2479 } else { 2480 assert(I->getOffsetInBits() % 8 == 0 && 2481 "bases must be on byte boundaries"); 2482 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2483 getTypeIndex(I->getBaseType()), 2484 I->getOffsetInBits() / 8); 2485 ContinuationBuilder.writeMemberType(BCR); 2486 MemberCount++; 2487 } 2488 } 2489 2490 // Create members. 2491 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2492 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2493 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2494 StringRef MemberName = Member->getName(); 2495 MemberAccess Access = 2496 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2497 2498 if (Member->isStaticMember()) { 2499 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2500 ContinuationBuilder.writeMemberType(SDMR); 2501 MemberCount++; 2502 continue; 2503 } 2504 2505 // Virtual function pointer member. 2506 if ((Member->getFlags() & DINode::FlagArtificial) && 2507 Member->getName().startswith("_vptr$")) { 2508 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2509 ContinuationBuilder.writeMemberType(VFPR); 2510 MemberCount++; 2511 continue; 2512 } 2513 2514 // Data member. 2515 uint64_t MemberOffsetInBits = 2516 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2517 if (Member->isBitField()) { 2518 uint64_t StartBitOffset = MemberOffsetInBits; 2519 if (const auto *CI = 2520 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2521 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2522 } 2523 StartBitOffset -= MemberOffsetInBits; 2524 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2525 StartBitOffset); 2526 MemberBaseType = TypeTable.writeLeafType(BFR); 2527 } 2528 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2529 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2530 MemberName); 2531 ContinuationBuilder.writeMemberType(DMR); 2532 MemberCount++; 2533 } 2534 2535 // Create methods 2536 for (auto &MethodItr : Info.Methods) { 2537 StringRef Name = MethodItr.first->getString(); 2538 2539 std::vector<OneMethodRecord> Methods; 2540 for (const DISubprogram *SP : MethodItr.second) { 2541 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2542 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2543 2544 unsigned VFTableOffset = -1; 2545 if (Introduced) 2546 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2547 2548 Methods.push_back(OneMethodRecord( 2549 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2550 translateMethodKindFlags(SP, Introduced), 2551 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2552 MemberCount++; 2553 } 2554 assert(!Methods.empty() && "Empty methods map entry"); 2555 if (Methods.size() == 1) 2556 ContinuationBuilder.writeMemberType(Methods[0]); 2557 else { 2558 // FIXME: Make this use its own ContinuationBuilder so that 2559 // MethodOverloadList can be split correctly. 2560 MethodOverloadListRecord MOLR(Methods); 2561 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2562 2563 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2564 ContinuationBuilder.writeMemberType(OMR); 2565 } 2566 } 2567 2568 // Create nested classes. 2569 for (const DIType *Nested : Info.NestedTypes) { 2570 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2571 ContinuationBuilder.writeMemberType(R); 2572 MemberCount++; 2573 } 2574 2575 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2576 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2577 !Info.NestedTypes.empty()); 2578 } 2579 2580 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2581 if (!VBPType.getIndex()) { 2582 // Make a 'const int *' type. 2583 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2584 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2585 2586 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2587 : PointerKind::Near32; 2588 PointerMode PM = PointerMode::Pointer; 2589 PointerOptions PO = PointerOptions::None; 2590 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2591 VBPType = TypeTable.writeLeafType(PR); 2592 } 2593 2594 return VBPType; 2595 } 2596 2597 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2598 // The null DIType is the void type. Don't try to hash it. 2599 if (!Ty) 2600 return TypeIndex::Void(); 2601 2602 // Check if we've already translated this type. Don't try to do a 2603 // get-or-create style insertion that caches the hash lookup across the 2604 // lowerType call. It will update the TypeIndices map. 2605 auto I = TypeIndices.find({Ty, ClassTy}); 2606 if (I != TypeIndices.end()) 2607 return I->second; 2608 2609 TypeLoweringScope S(*this); 2610 TypeIndex TI = lowerType(Ty, ClassTy); 2611 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2612 } 2613 2614 codeview::TypeIndex 2615 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2616 const DISubroutineType *SubroutineTy) { 2617 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2618 "this type must be a pointer type"); 2619 2620 PointerOptions Options = PointerOptions::None; 2621 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2622 Options = PointerOptions::LValueRefThisPointer; 2623 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2624 Options = PointerOptions::RValueRefThisPointer; 2625 2626 // Check if we've already translated this type. If there is no ref qualifier 2627 // on the function then we look up this pointer type with no associated class 2628 // so that the TypeIndex for the this pointer can be shared with the type 2629 // index for other pointers to this class type. If there is a ref qualifier 2630 // then we lookup the pointer using the subroutine as the parent type. 2631 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2632 if (I != TypeIndices.end()) 2633 return I->second; 2634 2635 TypeLoweringScope S(*this); 2636 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2637 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2638 } 2639 2640 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2641 PointerRecord PR(getTypeIndex(Ty), 2642 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2643 : PointerKind::Near32, 2644 PointerMode::LValueReference, PointerOptions::None, 2645 Ty->getSizeInBits() / 8); 2646 return TypeTable.writeLeafType(PR); 2647 } 2648 2649 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2650 // The null DIType is the void type. Don't try to hash it. 2651 if (!Ty) 2652 return TypeIndex::Void(); 2653 2654 // Look through typedefs when getting the complete type index. Call 2655 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2656 // emitted only once. 2657 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2658 (void)getTypeIndex(Ty); 2659 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2660 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2661 2662 // If this is a non-record type, the complete type index is the same as the 2663 // normal type index. Just call getTypeIndex. 2664 switch (Ty->getTag()) { 2665 case dwarf::DW_TAG_class_type: 2666 case dwarf::DW_TAG_structure_type: 2667 case dwarf::DW_TAG_union_type: 2668 break; 2669 default: 2670 return getTypeIndex(Ty); 2671 } 2672 2673 const auto *CTy = cast<DICompositeType>(Ty); 2674 2675 TypeLoweringScope S(*this); 2676 2677 // Make sure the forward declaration is emitted first. It's unclear if this 2678 // is necessary, but MSVC does it, and we should follow suit until we can show 2679 // otherwise. 2680 // We only emit a forward declaration for named types. 2681 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2682 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2683 2684 // Just use the forward decl if we don't have complete type info. This 2685 // might happen if the frontend is using modules and expects the complete 2686 // definition to be emitted elsewhere. 2687 if (CTy->isForwardDecl()) 2688 return FwdDeclTI; 2689 } 2690 2691 // Check if we've already translated the complete record type. 2692 // Insert the type with a null TypeIndex to signify that the type is currently 2693 // being lowered. 2694 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2695 if (!InsertResult.second) 2696 return InsertResult.first->second; 2697 2698 TypeIndex TI; 2699 switch (CTy->getTag()) { 2700 case dwarf::DW_TAG_class_type: 2701 case dwarf::DW_TAG_structure_type: 2702 TI = lowerCompleteTypeClass(CTy); 2703 break; 2704 case dwarf::DW_TAG_union_type: 2705 TI = lowerCompleteTypeUnion(CTy); 2706 break; 2707 default: 2708 llvm_unreachable("not a record"); 2709 } 2710 2711 // Update the type index associated with this CompositeType. This cannot 2712 // use the 'InsertResult' iterator above because it is potentially 2713 // invalidated by map insertions which can occur while lowering the class 2714 // type above. 2715 CompleteTypeIndices[CTy] = TI; 2716 return TI; 2717 } 2718 2719 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2720 /// and do this until fixpoint, as each complete record type typically 2721 /// references 2722 /// many other record types. 2723 void CodeViewDebug::emitDeferredCompleteTypes() { 2724 SmallVector<const DICompositeType *, 4> TypesToEmit; 2725 while (!DeferredCompleteTypes.empty()) { 2726 std::swap(DeferredCompleteTypes, TypesToEmit); 2727 for (const DICompositeType *RecordTy : TypesToEmit) 2728 getCompleteTypeIndex(RecordTy); 2729 TypesToEmit.clear(); 2730 } 2731 } 2732 2733 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2734 ArrayRef<LocalVariable> Locals) { 2735 // Get the sorted list of parameters and emit them first. 2736 SmallVector<const LocalVariable *, 6> Params; 2737 for (const LocalVariable &L : Locals) 2738 if (L.DIVar->isParameter()) 2739 Params.push_back(&L); 2740 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2741 return L->DIVar->getArg() < R->DIVar->getArg(); 2742 }); 2743 for (const LocalVariable *L : Params) 2744 emitLocalVariable(FI, *L); 2745 2746 // Next emit all non-parameters in the order that we found them. 2747 for (const LocalVariable &L : Locals) 2748 if (!L.DIVar->isParameter()) 2749 emitLocalVariable(FI, L); 2750 } 2751 2752 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2753 const LocalVariable &Var) { 2754 // LocalSym record, see SymbolRecord.h for more info. 2755 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2756 2757 LocalSymFlags Flags = LocalSymFlags::None; 2758 if (Var.DIVar->isParameter()) 2759 Flags |= LocalSymFlags::IsParameter; 2760 if (Var.DefRanges.empty()) 2761 Flags |= LocalSymFlags::IsOptimizedOut; 2762 2763 OS.AddComment("TypeIndex"); 2764 TypeIndex TI = Var.UseReferenceType 2765 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2766 : getCompleteTypeIndex(Var.DIVar->getType()); 2767 OS.emitInt32(TI.getIndex()); 2768 OS.AddComment("Flags"); 2769 OS.emitInt16(static_cast<uint16_t>(Flags)); 2770 // Truncate the name so we won't overflow the record length field. 2771 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2772 endSymbolRecord(LocalEnd); 2773 2774 // Calculate the on disk prefix of the appropriate def range record. The 2775 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2776 // should be big enough to hold all forms without memory allocation. 2777 SmallString<20> BytePrefix; 2778 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2779 BytePrefix.clear(); 2780 if (DefRange.InMemory) { 2781 int Offset = DefRange.DataOffset; 2782 unsigned Reg = DefRange.CVRegister; 2783 2784 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2785 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2786 // instead. In frames without stack realignment, $T0 will be the CFA. 2787 if (RegisterId(Reg) == RegisterId::ESP) { 2788 Reg = unsigned(RegisterId::VFRAME); 2789 Offset += FI.OffsetAdjustment; 2790 } 2791 2792 // If we can use the chosen frame pointer for the frame and this isn't a 2793 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2794 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2795 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2796 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2797 (bool(Flags & LocalSymFlags::IsParameter) 2798 ? (EncFP == FI.EncodedParamFramePtrReg) 2799 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2800 DefRangeFramePointerRelHeader DRHdr; 2801 DRHdr.Offset = Offset; 2802 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2803 } else { 2804 uint16_t RegRelFlags = 0; 2805 if (DefRange.IsSubfield) { 2806 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2807 (DefRange.StructOffset 2808 << DefRangeRegisterRelSym::OffsetInParentShift); 2809 } 2810 DefRangeRegisterRelHeader DRHdr; 2811 DRHdr.Register = Reg; 2812 DRHdr.Flags = RegRelFlags; 2813 DRHdr.BasePointerOffset = Offset; 2814 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2815 } 2816 } else { 2817 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2818 if (DefRange.IsSubfield) { 2819 DefRangeSubfieldRegisterHeader DRHdr; 2820 DRHdr.Register = DefRange.CVRegister; 2821 DRHdr.MayHaveNoName = 0; 2822 DRHdr.OffsetInParent = DefRange.StructOffset; 2823 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2824 } else { 2825 DefRangeRegisterHeader DRHdr; 2826 DRHdr.Register = DefRange.CVRegister; 2827 DRHdr.MayHaveNoName = 0; 2828 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2829 } 2830 } 2831 } 2832 } 2833 2834 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2835 const FunctionInfo& FI) { 2836 for (LexicalBlock *Block : Blocks) 2837 emitLexicalBlock(*Block, FI); 2838 } 2839 2840 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2841 /// lexical block scope. 2842 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2843 const FunctionInfo& FI) { 2844 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2845 OS.AddComment("PtrParent"); 2846 OS.emitInt32(0); // PtrParent 2847 OS.AddComment("PtrEnd"); 2848 OS.emitInt32(0); // PtrEnd 2849 OS.AddComment("Code size"); 2850 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2851 OS.AddComment("Function section relative address"); 2852 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2853 OS.AddComment("Function section index"); 2854 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2855 OS.AddComment("Lexical block name"); 2856 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2857 endSymbolRecord(RecordEnd); 2858 2859 // Emit variables local to this lexical block. 2860 emitLocalVariableList(FI, Block.Locals); 2861 emitGlobalVariableList(Block.Globals); 2862 2863 // Emit lexical blocks contained within this block. 2864 emitLexicalBlockList(Block.Children, FI); 2865 2866 // Close the lexical block scope. 2867 emitEndSymbolRecord(SymbolKind::S_END); 2868 } 2869 2870 /// Convenience routine for collecting lexical block information for a list 2871 /// of lexical scopes. 2872 void CodeViewDebug::collectLexicalBlockInfo( 2873 SmallVectorImpl<LexicalScope *> &Scopes, 2874 SmallVectorImpl<LexicalBlock *> &Blocks, 2875 SmallVectorImpl<LocalVariable> &Locals, 2876 SmallVectorImpl<CVGlobalVariable> &Globals) { 2877 for (LexicalScope *Scope : Scopes) 2878 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2879 } 2880 2881 /// Populate the lexical blocks and local variable lists of the parent with 2882 /// information about the specified lexical scope. 2883 void CodeViewDebug::collectLexicalBlockInfo( 2884 LexicalScope &Scope, 2885 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2886 SmallVectorImpl<LocalVariable> &ParentLocals, 2887 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2888 if (Scope.isAbstractScope()) 2889 return; 2890 2891 // Gather information about the lexical scope including local variables, 2892 // global variables, and address ranges. 2893 bool IgnoreScope = false; 2894 auto LI = ScopeVariables.find(&Scope); 2895 SmallVectorImpl<LocalVariable> *Locals = 2896 LI != ScopeVariables.end() ? &LI->second : nullptr; 2897 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2898 SmallVectorImpl<CVGlobalVariable> *Globals = 2899 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2900 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2901 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2902 2903 // Ignore lexical scopes which do not contain variables. 2904 if (!Locals && !Globals) 2905 IgnoreScope = true; 2906 2907 // Ignore lexical scopes which are not lexical blocks. 2908 if (!DILB) 2909 IgnoreScope = true; 2910 2911 // Ignore scopes which have too many address ranges to represent in the 2912 // current CodeView format or do not have a valid address range. 2913 // 2914 // For lexical scopes with multiple address ranges you may be tempted to 2915 // construct a single range covering every instruction where the block is 2916 // live and everything in between. Unfortunately, Visual Studio only 2917 // displays variables from the first matching lexical block scope. If the 2918 // first lexical block contains exception handling code or cold code which 2919 // is moved to the bottom of the routine creating a single range covering 2920 // nearly the entire routine, then it will hide all other lexical blocks 2921 // and the variables they contain. 2922 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2923 IgnoreScope = true; 2924 2925 if (IgnoreScope) { 2926 // This scope can be safely ignored and eliminating it will reduce the 2927 // size of the debug information. Be sure to collect any variable and scope 2928 // information from the this scope or any of its children and collapse them 2929 // into the parent scope. 2930 if (Locals) 2931 ParentLocals.append(Locals->begin(), Locals->end()); 2932 if (Globals) 2933 ParentGlobals.append(Globals->begin(), Globals->end()); 2934 collectLexicalBlockInfo(Scope.getChildren(), 2935 ParentBlocks, 2936 ParentLocals, 2937 ParentGlobals); 2938 return; 2939 } 2940 2941 // Create a new CodeView lexical block for this lexical scope. If we've 2942 // seen this DILexicalBlock before then the scope tree is malformed and 2943 // we can handle this gracefully by not processing it a second time. 2944 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2945 if (!BlockInsertion.second) 2946 return; 2947 2948 // Create a lexical block containing the variables and collect the the 2949 // lexical block information for the children. 2950 const InsnRange &Range = Ranges.front(); 2951 assert(Range.first && Range.second); 2952 LexicalBlock &Block = BlockInsertion.first->second; 2953 Block.Begin = getLabelBeforeInsn(Range.first); 2954 Block.End = getLabelAfterInsn(Range.second); 2955 assert(Block.Begin && "missing label for scope begin"); 2956 assert(Block.End && "missing label for scope end"); 2957 Block.Name = DILB->getName(); 2958 if (Locals) 2959 Block.Locals = std::move(*Locals); 2960 if (Globals) 2961 Block.Globals = std::move(*Globals); 2962 ParentBlocks.push_back(&Block); 2963 collectLexicalBlockInfo(Scope.getChildren(), 2964 Block.Children, 2965 Block.Locals, 2966 Block.Globals); 2967 } 2968 2969 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2970 const Function &GV = MF->getFunction(); 2971 assert(FnDebugInfo.count(&GV)); 2972 assert(CurFn == FnDebugInfo[&GV].get()); 2973 2974 collectVariableInfo(GV.getSubprogram()); 2975 2976 // Build the lexical block structure to emit for this routine. 2977 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2978 collectLexicalBlockInfo(*CFS, 2979 CurFn->ChildBlocks, 2980 CurFn->Locals, 2981 CurFn->Globals); 2982 2983 // Clear the scope and variable information from the map which will not be 2984 // valid after we have finished processing this routine. This also prepares 2985 // the map for the subsequent routine. 2986 ScopeVariables.clear(); 2987 2988 // Don't emit anything if we don't have any line tables. 2989 // Thunks are compiler-generated and probably won't have source correlation. 2990 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2991 FnDebugInfo.erase(&GV); 2992 CurFn = nullptr; 2993 return; 2994 } 2995 2996 // Find heap alloc sites and add to list. 2997 for (const auto &MBB : *MF) { 2998 for (const auto &MI : MBB) { 2999 if (MDNode *MD = MI.getHeapAllocMarker()) { 3000 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3001 getLabelAfterInsn(&MI), 3002 dyn_cast<DIType>(MD))); 3003 } 3004 } 3005 } 3006 3007 CurFn->Annotations = MF->getCodeViewAnnotations(); 3008 3009 CurFn->End = Asm->getFunctionEnd(); 3010 3011 CurFn = nullptr; 3012 } 3013 3014 // Usable locations are valid with non-zero line numbers. A line number of zero 3015 // corresponds to optimized code that doesn't have a distinct source location. 3016 // In this case, we try to use the previous or next source location depending on 3017 // the context. 3018 static bool isUsableDebugLoc(DebugLoc DL) { 3019 return DL && DL.getLine() != 0; 3020 } 3021 3022 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3023 DebugHandlerBase::beginInstruction(MI); 3024 3025 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3026 if (!Asm || !CurFn || MI->isDebugInstr() || 3027 MI->getFlag(MachineInstr::FrameSetup)) 3028 return; 3029 3030 // If the first instruction of a new MBB has no location, find the first 3031 // instruction with a location and use that. 3032 DebugLoc DL = MI->getDebugLoc(); 3033 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3034 for (const auto &NextMI : *MI->getParent()) { 3035 if (NextMI.isDebugInstr()) 3036 continue; 3037 DL = NextMI.getDebugLoc(); 3038 if (isUsableDebugLoc(DL)) 3039 break; 3040 } 3041 // FIXME: Handle the case where the BB has no valid locations. This would 3042 // probably require doing a real dataflow analysis. 3043 } 3044 PrevInstBB = MI->getParent(); 3045 3046 // If we still don't have a debug location, don't record a location. 3047 if (!isUsableDebugLoc(DL)) 3048 return; 3049 3050 maybeRecordLocation(DL, Asm->MF); 3051 } 3052 3053 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3054 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3055 *EndLabel = MMI->getContext().createTempSymbol(); 3056 OS.emitInt32(unsigned(Kind)); 3057 OS.AddComment("Subsection size"); 3058 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3059 OS.emitLabel(BeginLabel); 3060 return EndLabel; 3061 } 3062 3063 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3064 OS.emitLabel(EndLabel); 3065 // Every subsection must be aligned to a 4-byte boundary. 3066 OS.emitValueToAlignment(4); 3067 } 3068 3069 static StringRef getSymbolName(SymbolKind SymKind) { 3070 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3071 if (EE.Value == SymKind) 3072 return EE.Name; 3073 return ""; 3074 } 3075 3076 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3077 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3078 *EndLabel = MMI->getContext().createTempSymbol(); 3079 OS.AddComment("Record length"); 3080 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3081 OS.emitLabel(BeginLabel); 3082 if (OS.isVerboseAsm()) 3083 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3084 OS.emitInt16(unsigned(SymKind)); 3085 return EndLabel; 3086 } 3087 3088 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3089 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3090 // an extra copy of every symbol record in LLD. This increases object file 3091 // size by less than 1% in the clang build, and is compatible with the Visual 3092 // C++ linker. 3093 OS.emitValueToAlignment(4); 3094 OS.emitLabel(SymEnd); 3095 } 3096 3097 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3098 OS.AddComment("Record length"); 3099 OS.emitInt16(2); 3100 if (OS.isVerboseAsm()) 3101 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3102 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3103 } 3104 3105 void CodeViewDebug::emitDebugInfoForUDTs( 3106 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3107 #ifndef NDEBUG 3108 size_t OriginalSize = UDTs.size(); 3109 #endif 3110 for (const auto &UDT : UDTs) { 3111 const DIType *T = UDT.second; 3112 assert(shouldEmitUdt(T)); 3113 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3114 OS.AddComment("Type"); 3115 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3116 assert(OriginalSize == UDTs.size() && 3117 "getCompleteTypeIndex found new UDTs!"); 3118 emitNullTerminatedSymbolName(OS, UDT.first); 3119 endSymbolRecord(UDTRecordEnd); 3120 } 3121 } 3122 3123 void CodeViewDebug::collectGlobalVariableInfo() { 3124 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3125 GlobalMap; 3126 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3127 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3128 GV.getDebugInfo(GVEs); 3129 for (const auto *GVE : GVEs) 3130 GlobalMap[GVE] = &GV; 3131 } 3132 3133 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3134 for (const MDNode *Node : CUs->operands()) { 3135 const auto *CU = cast<DICompileUnit>(Node); 3136 for (const auto *GVE : CU->getGlobalVariables()) { 3137 const DIGlobalVariable *DIGV = GVE->getVariable(); 3138 const DIExpression *DIE = GVE->getExpression(); 3139 3140 if ((DIE->getNumElements() == 2) && 3141 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3142 // Record the constant offset for the variable. 3143 // 3144 // A Fortran common block uses this idiom to encode the offset 3145 // of a variable from the common block's starting address. 3146 CVGlobalVariableOffsets.insert( 3147 std::make_pair(DIGV, DIE->getElement(1))); 3148 3149 // Emit constant global variables in a global symbol section. 3150 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3151 CVGlobalVariable CVGV = {DIGV, DIE}; 3152 GlobalVariables.emplace_back(std::move(CVGV)); 3153 } 3154 3155 const auto *GV = GlobalMap.lookup(GVE); 3156 if (!GV || GV->isDeclarationForLinker()) 3157 continue; 3158 3159 DIScope *Scope = DIGV->getScope(); 3160 SmallVector<CVGlobalVariable, 1> *VariableList; 3161 if (Scope && isa<DILocalScope>(Scope)) { 3162 // Locate a global variable list for this scope, creating one if 3163 // necessary. 3164 auto Insertion = ScopeGlobals.insert( 3165 {Scope, std::unique_ptr<GlobalVariableList>()}); 3166 if (Insertion.second) 3167 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3168 VariableList = Insertion.first->second.get(); 3169 } else if (GV->hasComdat()) 3170 // Emit this global variable into a COMDAT section. 3171 VariableList = &ComdatVariables; 3172 else 3173 // Emit this global variable in a single global symbol section. 3174 VariableList = &GlobalVariables; 3175 CVGlobalVariable CVGV = {DIGV, GV}; 3176 VariableList->emplace_back(std::move(CVGV)); 3177 } 3178 } 3179 } 3180 3181 void CodeViewDebug::collectDebugInfoForGlobals() { 3182 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3183 const DIGlobalVariable *DIGV = CVGV.DIGV; 3184 const DIScope *Scope = DIGV->getScope(); 3185 getCompleteTypeIndex(DIGV->getType()); 3186 getFullyQualifiedName(Scope, DIGV->getName()); 3187 } 3188 3189 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3190 const DIGlobalVariable *DIGV = CVGV.DIGV; 3191 const DIScope *Scope = DIGV->getScope(); 3192 getCompleteTypeIndex(DIGV->getType()); 3193 getFullyQualifiedName(Scope, DIGV->getName()); 3194 } 3195 } 3196 3197 void CodeViewDebug::emitDebugInfoForGlobals() { 3198 // First, emit all globals that are not in a comdat in a single symbol 3199 // substream. MSVC doesn't like it if the substream is empty, so only open 3200 // it if we have at least one global to emit. 3201 switchToDebugSectionForSymbol(nullptr); 3202 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3203 OS.AddComment("Symbol subsection for globals"); 3204 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3205 emitGlobalVariableList(GlobalVariables); 3206 emitStaticConstMemberList(); 3207 endCVSubsection(EndLabel); 3208 } 3209 3210 // Second, emit each global that is in a comdat into its own .debug$S 3211 // section along with its own symbol substream. 3212 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3213 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3214 MCSymbol *GVSym = Asm->getSymbol(GV); 3215 OS.AddComment("Symbol subsection for " + 3216 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3217 switchToDebugSectionForSymbol(GVSym); 3218 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3219 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3220 emitDebugInfoForGlobal(CVGV); 3221 endCVSubsection(EndLabel); 3222 } 3223 } 3224 3225 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3226 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3227 for (const MDNode *Node : CUs->operands()) { 3228 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3229 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3230 getTypeIndex(RT); 3231 // FIXME: Add to global/local DTU list. 3232 } 3233 } 3234 } 3235 } 3236 3237 // Emit each global variable in the specified array. 3238 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3239 for (const CVGlobalVariable &CVGV : Globals) { 3240 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3241 emitDebugInfoForGlobal(CVGV); 3242 } 3243 } 3244 3245 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3246 const std::string &QualifiedName) { 3247 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3248 OS.AddComment("Type"); 3249 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3250 3251 OS.AddComment("Value"); 3252 3253 // Encoded integers shouldn't need more than 10 bytes. 3254 uint8_t Data[10]; 3255 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3256 CodeViewRecordIO IO(Writer); 3257 cantFail(IO.mapEncodedInteger(Value)); 3258 StringRef SRef((char *)Data, Writer.getOffset()); 3259 OS.emitBinaryData(SRef); 3260 3261 OS.AddComment("Name"); 3262 emitNullTerminatedSymbolName(OS, QualifiedName); 3263 endSymbolRecord(SConstantEnd); 3264 } 3265 3266 void CodeViewDebug::emitStaticConstMemberList() { 3267 for (const DIDerivedType *DTy : StaticConstMembers) { 3268 const DIScope *Scope = DTy->getScope(); 3269 3270 APSInt Value; 3271 if (const ConstantInt *CI = 3272 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3273 Value = APSInt(CI->getValue(), 3274 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3275 else if (const ConstantFP *CFP = 3276 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3277 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3278 else 3279 llvm_unreachable("cannot emit a constant without a value"); 3280 3281 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3282 getFullyQualifiedName(Scope, DTy->getName())); 3283 } 3284 } 3285 3286 static bool isFloatDIType(const DIType *Ty) { 3287 if (isa<DICompositeType>(Ty)) 3288 return false; 3289 3290 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3291 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3292 if (T == dwarf::DW_TAG_pointer_type || 3293 T == dwarf::DW_TAG_ptr_to_member_type || 3294 T == dwarf::DW_TAG_reference_type || 3295 T == dwarf::DW_TAG_rvalue_reference_type) 3296 return false; 3297 assert(DTy->getBaseType() && "Expected valid base type"); 3298 return isFloatDIType(DTy->getBaseType()); 3299 } 3300 3301 auto *BTy = cast<DIBasicType>(Ty); 3302 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3303 } 3304 3305 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3306 const DIGlobalVariable *DIGV = CVGV.DIGV; 3307 3308 const DIScope *Scope = DIGV->getScope(); 3309 // For static data members, get the scope from the declaration. 3310 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3311 DIGV->getRawStaticDataMemberDeclaration())) 3312 Scope = MemberDecl->getScope(); 3313 // For Fortran, the scoping portion is elided in its name so that we can 3314 // reference the variable in the command line of the VS debugger. 3315 std::string QualifiedName = 3316 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3317 : getFullyQualifiedName(Scope, DIGV->getName()); 3318 3319 if (const GlobalVariable *GV = 3320 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3321 // DataSym record, see SymbolRecord.h for more info. Thread local data 3322 // happens to have the same format as global data. 3323 MCSymbol *GVSym = Asm->getSymbol(GV); 3324 SymbolKind DataSym = GV->isThreadLocal() 3325 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3326 : SymbolKind::S_GTHREAD32) 3327 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3328 : SymbolKind::S_GDATA32); 3329 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3330 OS.AddComment("Type"); 3331 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3332 OS.AddComment("DataOffset"); 3333 3334 uint64_t Offset = 0; 3335 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3336 // Use the offset seen while collecting info on globals. 3337 Offset = CVGlobalVariableOffsets[DIGV]; 3338 OS.EmitCOFFSecRel32(GVSym, Offset); 3339 3340 OS.AddComment("Segment"); 3341 OS.EmitCOFFSectionIndex(GVSym); 3342 OS.AddComment("Name"); 3343 const unsigned LengthOfDataRecord = 12; 3344 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3345 endSymbolRecord(DataEnd); 3346 } else { 3347 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3348 assert(DIE->isConstant() && 3349 "Global constant variables must contain a constant expression."); 3350 3351 // Use unsigned for floats. 3352 bool isUnsigned = isFloatDIType(DIGV->getType()) 3353 ? true 3354 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3355 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3356 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3357 } 3358 } 3359