1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "llvm/ADT/APSInt.h" 15 #include "llvm/ADT/None.h" 16 #include "llvm/ADT/Optional.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallString.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/ADT/TinyPtrVector.h" 21 #include "llvm/ADT/Triple.h" 22 #include "llvm/ADT/Twine.h" 23 #include "llvm/BinaryFormat/COFF.h" 24 #include "llvm/BinaryFormat/Dwarf.h" 25 #include "llvm/CodeGen/AsmPrinter.h" 26 #include "llvm/CodeGen/LexicalScopes.h" 27 #include "llvm/CodeGen/MachineFrameInfo.h" 28 #include "llvm/CodeGen/MachineFunction.h" 29 #include "llvm/CodeGen/MachineInstr.h" 30 #include "llvm/CodeGen/MachineModuleInfo.h" 31 #include "llvm/CodeGen/TargetFrameLowering.h" 32 #include "llvm/CodeGen/TargetRegisterInfo.h" 33 #include "llvm/CodeGen/TargetSubtargetInfo.h" 34 #include "llvm/Config/llvm-config.h" 35 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 36 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 37 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 38 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 39 #include "llvm/DebugInfo/CodeView/EnumTables.h" 40 #include "llvm/DebugInfo/CodeView/Line.h" 41 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 42 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 43 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 44 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 45 #include "llvm/IR/Constants.h" 46 #include "llvm/IR/DataLayout.h" 47 #include "llvm/IR/DebugInfoMetadata.h" 48 #include "llvm/IR/Function.h" 49 #include "llvm/IR/GlobalValue.h" 50 #include "llvm/IR/GlobalVariable.h" 51 #include "llvm/IR/Metadata.h" 52 #include "llvm/IR/Module.h" 53 #include "llvm/MC/MCAsmInfo.h" 54 #include "llvm/MC/MCContext.h" 55 #include "llvm/MC/MCSectionCOFF.h" 56 #include "llvm/MC/MCStreamer.h" 57 #include "llvm/MC/MCSymbol.h" 58 #include "llvm/Support/BinaryStreamWriter.h" 59 #include "llvm/Support/Casting.h" 60 #include "llvm/Support/Endian.h" 61 #include "llvm/Support/Error.h" 62 #include "llvm/Support/ErrorHandling.h" 63 #include "llvm/Support/FormatVariadic.h" 64 #include "llvm/Support/Path.h" 65 #include "llvm/Support/Program.h" 66 #include "llvm/Support/SMLoc.h" 67 #include "llvm/Support/ScopedPrinter.h" 68 #include "llvm/Target/TargetLoweringObjectFile.h" 69 #include "llvm/Target/TargetMachine.h" 70 #include <algorithm> 71 #include <cassert> 72 #include <cctype> 73 #include <cstddef> 74 #include <iterator> 75 #include <limits> 76 77 using namespace llvm; 78 using namespace llvm::codeview; 79 80 namespace { 81 class CVMCAdapter : public CodeViewRecordStreamer { 82 public: 83 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 84 : OS(&OS), TypeTable(TypeTable) {} 85 86 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 87 88 void emitIntValue(uint64_t Value, unsigned Size) override { 89 OS->emitIntValueInHex(Value, Size); 90 } 91 92 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 93 94 void AddComment(const Twine &T) override { OS->AddComment(T); } 95 96 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 97 98 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 99 100 std::string getTypeName(TypeIndex TI) override { 101 std::string TypeName; 102 if (!TI.isNoneType()) { 103 if (TI.isSimple()) 104 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 105 else 106 TypeName = std::string(TypeTable.getTypeName(TI)); 107 } 108 return TypeName; 109 } 110 111 private: 112 MCStreamer *OS = nullptr; 113 TypeCollection &TypeTable; 114 }; 115 } // namespace 116 117 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 118 switch (Type) { 119 case Triple::ArchType::x86: 120 return CPUType::Pentium3; 121 case Triple::ArchType::x86_64: 122 return CPUType::X64; 123 case Triple::ArchType::thumb: 124 // LLVM currently doesn't support Windows CE and so thumb 125 // here is indiscriminately mapped to ARMNT specifically. 126 return CPUType::ARMNT; 127 case Triple::ArchType::aarch64: 128 return CPUType::ARM64; 129 default: 130 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 131 } 132 } 133 134 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 135 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 136 137 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 138 std::string &Filepath = FileToFilepathMap[File]; 139 if (!Filepath.empty()) 140 return Filepath; 141 142 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 143 144 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 145 // it textually because one of the path components could be a symlink. 146 if (Dir.startswith("/") || Filename.startswith("/")) { 147 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 148 return Filename; 149 Filepath = std::string(Dir); 150 if (Dir.back() != '/') 151 Filepath += '/'; 152 Filepath += Filename; 153 return Filepath; 154 } 155 156 // Clang emits directory and relative filename info into the IR, but CodeView 157 // operates on full paths. We could change Clang to emit full paths too, but 158 // that would increase the IR size and probably not needed for other users. 159 // For now, just concatenate and canonicalize the path here. 160 if (Filename.find(':') == 1) 161 Filepath = std::string(Filename); 162 else 163 Filepath = (Dir + "\\" + Filename).str(); 164 165 // Canonicalize the path. We have to do it textually because we may no longer 166 // have access the file in the filesystem. 167 // First, replace all slashes with backslashes. 168 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 169 170 // Remove all "\.\" with "\". 171 size_t Cursor = 0; 172 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 173 Filepath.erase(Cursor, 2); 174 175 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 176 // path should be well-formatted, e.g. start with a drive letter, etc. 177 Cursor = 0; 178 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 179 // Something's wrong if the path starts with "\..\", abort. 180 if (Cursor == 0) 181 break; 182 183 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 184 if (PrevSlash == std::string::npos) 185 // Something's wrong, abort. 186 break; 187 188 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 189 // The next ".." might be following the one we've just erased. 190 Cursor = PrevSlash; 191 } 192 193 // Remove all duplicate backslashes. 194 Cursor = 0; 195 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 196 Filepath.erase(Cursor, 1); 197 198 return Filepath; 199 } 200 201 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 202 StringRef FullPath = getFullFilepath(F); 203 unsigned NextId = FileIdMap.size() + 1; 204 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 205 if (Insertion.second) { 206 // We have to compute the full filepath and emit a .cv_file directive. 207 ArrayRef<uint8_t> ChecksumAsBytes; 208 FileChecksumKind CSKind = FileChecksumKind::None; 209 if (F->getChecksum()) { 210 std::string Checksum = fromHex(F->getChecksum()->Value); 211 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 212 memcpy(CKMem, Checksum.data(), Checksum.size()); 213 ChecksumAsBytes = ArrayRef<uint8_t>( 214 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 215 switch (F->getChecksum()->Kind) { 216 case DIFile::CSK_MD5: 217 CSKind = FileChecksumKind::MD5; 218 break; 219 case DIFile::CSK_SHA1: 220 CSKind = FileChecksumKind::SHA1; 221 break; 222 case DIFile::CSK_SHA256: 223 CSKind = FileChecksumKind::SHA256; 224 break; 225 } 226 } 227 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 228 static_cast<unsigned>(CSKind)); 229 (void)Success; 230 assert(Success && ".cv_file directive failed"); 231 } 232 return Insertion.first->second; 233 } 234 235 CodeViewDebug::InlineSite & 236 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 237 const DISubprogram *Inlinee) { 238 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 239 InlineSite *Site = &SiteInsertion.first->second; 240 if (SiteInsertion.second) { 241 unsigned ParentFuncId = CurFn->FuncId; 242 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 243 ParentFuncId = 244 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 245 .SiteFuncId; 246 247 Site->SiteFuncId = NextFuncId++; 248 OS.EmitCVInlineSiteIdDirective( 249 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 250 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 251 Site->Inlinee = Inlinee; 252 InlinedSubprograms.insert(Inlinee); 253 getFuncIdForSubprogram(Inlinee); 254 } 255 return *Site; 256 } 257 258 static StringRef getPrettyScopeName(const DIScope *Scope) { 259 StringRef ScopeName = Scope->getName(); 260 if (!ScopeName.empty()) 261 return ScopeName; 262 263 switch (Scope->getTag()) { 264 case dwarf::DW_TAG_enumeration_type: 265 case dwarf::DW_TAG_class_type: 266 case dwarf::DW_TAG_structure_type: 267 case dwarf::DW_TAG_union_type: 268 return "<unnamed-tag>"; 269 case dwarf::DW_TAG_namespace: 270 return "`anonymous namespace'"; 271 default: 272 return StringRef(); 273 } 274 } 275 276 const DISubprogram *CodeViewDebug::collectParentScopeNames( 277 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 278 const DISubprogram *ClosestSubprogram = nullptr; 279 while (Scope != nullptr) { 280 if (ClosestSubprogram == nullptr) 281 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 282 283 // If a type appears in a scope chain, make sure it gets emitted. The 284 // frontend will be responsible for deciding if this should be a forward 285 // declaration or a complete type. 286 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 287 DeferredCompleteTypes.push_back(Ty); 288 289 StringRef ScopeName = getPrettyScopeName(Scope); 290 if (!ScopeName.empty()) 291 QualifiedNameComponents.push_back(ScopeName); 292 Scope = Scope->getScope(); 293 } 294 return ClosestSubprogram; 295 } 296 297 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 298 StringRef TypeName) { 299 std::string FullyQualifiedName; 300 for (StringRef QualifiedNameComponent : 301 llvm::reverse(QualifiedNameComponents)) { 302 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 303 FullyQualifiedName.append("::"); 304 } 305 FullyQualifiedName.append(std::string(TypeName)); 306 return FullyQualifiedName; 307 } 308 309 struct CodeViewDebug::TypeLoweringScope { 310 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 311 ~TypeLoweringScope() { 312 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 313 // inner TypeLoweringScopes don't attempt to emit deferred types. 314 if (CVD.TypeEmissionLevel == 1) 315 CVD.emitDeferredCompleteTypes(); 316 --CVD.TypeEmissionLevel; 317 } 318 CodeViewDebug &CVD; 319 }; 320 321 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 322 StringRef Name) { 323 // Ensure types in the scope chain are emitted as soon as possible. 324 // This can create otherwise a situation where S_UDTs are emitted while 325 // looping in emitDebugInfoForUDTs. 326 TypeLoweringScope S(*this); 327 SmallVector<StringRef, 5> QualifiedNameComponents; 328 collectParentScopeNames(Scope, QualifiedNameComponents); 329 return formatNestedName(QualifiedNameComponents, Name); 330 } 331 332 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 333 const DIScope *Scope = Ty->getScope(); 334 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 335 } 336 337 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 338 // No scope means global scope and that uses the zero index. 339 // 340 // We also use zero index when the scope is a DISubprogram 341 // to suppress the emission of LF_STRING_ID for the function, 342 // which can trigger a link-time error with the linker in 343 // VS2019 version 16.11.2 or newer. 344 // Note, however, skipping the debug info emission for the DISubprogram 345 // is a temporary fix. The root issue here is that we need to figure out 346 // the proper way to encode a function nested in another function 347 // (as introduced by the Fortran 'contains' keyword) in CodeView. 348 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 349 return TypeIndex(); 350 351 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 352 353 // Check if we've already translated this scope. 354 auto I = TypeIndices.find({Scope, nullptr}); 355 if (I != TypeIndices.end()) 356 return I->second; 357 358 // Build the fully qualified name of the scope. 359 std::string ScopeName = getFullyQualifiedName(Scope); 360 StringIdRecord SID(TypeIndex(), ScopeName); 361 auto TI = TypeTable.writeLeafType(SID); 362 return recordTypeIndexForDINode(Scope, TI); 363 } 364 365 static StringRef removeTemplateArgs(StringRef Name) { 366 // Remove template args from the display name. Assume that the template args 367 // are the last thing in the name. 368 if (Name.empty() || Name.back() != '>') 369 return Name; 370 371 int OpenBrackets = 0; 372 for (int i = Name.size() - 1; i >= 0; --i) { 373 if (Name[i] == '>') 374 ++OpenBrackets; 375 else if (Name[i] == '<') { 376 --OpenBrackets; 377 if (OpenBrackets == 0) 378 return Name.substr(0, i); 379 } 380 } 381 return Name; 382 } 383 384 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 385 assert(SP); 386 387 // Check if we've already translated this subprogram. 388 auto I = TypeIndices.find({SP, nullptr}); 389 if (I != TypeIndices.end()) 390 return I->second; 391 392 // The display name includes function template arguments. Drop them to match 393 // MSVC. We need to have the template arguments in the DISubprogram name 394 // because they are used in other symbol records, such as S_GPROC32_IDs. 395 StringRef DisplayName = removeTemplateArgs(SP->getName()); 396 397 const DIScope *Scope = SP->getScope(); 398 TypeIndex TI; 399 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 400 // If the scope is a DICompositeType, then this must be a method. Member 401 // function types take some special handling, and require access to the 402 // subprogram. 403 TypeIndex ClassType = getTypeIndex(Class); 404 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 405 DisplayName); 406 TI = TypeTable.writeLeafType(MFuncId); 407 } else { 408 // Otherwise, this must be a free function. 409 TypeIndex ParentScope = getScopeIndex(Scope); 410 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 411 TI = TypeTable.writeLeafType(FuncId); 412 } 413 414 return recordTypeIndexForDINode(SP, TI); 415 } 416 417 static bool isNonTrivial(const DICompositeType *DCTy) { 418 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 419 } 420 421 static FunctionOptions 422 getFunctionOptions(const DISubroutineType *Ty, 423 const DICompositeType *ClassTy = nullptr, 424 StringRef SPName = StringRef("")) { 425 FunctionOptions FO = FunctionOptions::None; 426 const DIType *ReturnTy = nullptr; 427 if (auto TypeArray = Ty->getTypeArray()) { 428 if (TypeArray.size()) 429 ReturnTy = TypeArray[0]; 430 } 431 432 // Add CxxReturnUdt option to functions that return nontrivial record types 433 // or methods that return record types. 434 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 435 if (isNonTrivial(ReturnDCTy) || ClassTy) 436 FO |= FunctionOptions::CxxReturnUdt; 437 438 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 439 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 440 FO |= FunctionOptions::Constructor; 441 442 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 443 444 } 445 return FO; 446 } 447 448 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 449 const DICompositeType *Class) { 450 // Always use the method declaration as the key for the function type. The 451 // method declaration contains the this adjustment. 452 if (SP->getDeclaration()) 453 SP = SP->getDeclaration(); 454 assert(!SP->getDeclaration() && "should use declaration as key"); 455 456 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 457 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 458 auto I = TypeIndices.find({SP, Class}); 459 if (I != TypeIndices.end()) 460 return I->second; 461 462 // Make sure complete type info for the class is emitted *after* the member 463 // function type, as the complete class type is likely to reference this 464 // member function type. 465 TypeLoweringScope S(*this); 466 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 467 468 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 469 TypeIndex TI = lowerTypeMemberFunction( 470 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 471 return recordTypeIndexForDINode(SP, TI, Class); 472 } 473 474 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 475 TypeIndex TI, 476 const DIType *ClassTy) { 477 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 478 (void)InsertResult; 479 assert(InsertResult.second && "DINode was already assigned a type index"); 480 return TI; 481 } 482 483 unsigned CodeViewDebug::getPointerSizeInBytes() { 484 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 485 } 486 487 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 488 const LexicalScope *LS) { 489 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 490 // This variable was inlined. Associate it with the InlineSite. 491 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 492 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 493 Site.InlinedLocals.emplace_back(Var); 494 } else { 495 // This variable goes into the corresponding lexical scope. 496 ScopeVariables[LS].emplace_back(Var); 497 } 498 } 499 500 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 501 const DILocation *Loc) { 502 if (!llvm::is_contained(Locs, Loc)) 503 Locs.push_back(Loc); 504 } 505 506 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 507 const MachineFunction *MF) { 508 // Skip this instruction if it has the same location as the previous one. 509 if (!DL || DL == PrevInstLoc) 510 return; 511 512 const DIScope *Scope = DL->getScope(); 513 if (!Scope) 514 return; 515 516 // Skip this line if it is longer than the maximum we can record. 517 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 518 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 519 LI.isNeverStepInto()) 520 return; 521 522 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 523 if (CI.getStartColumn() != DL.getCol()) 524 return; 525 526 if (!CurFn->HaveLineInfo) 527 CurFn->HaveLineInfo = true; 528 unsigned FileId = 0; 529 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 530 FileId = CurFn->LastFileId; 531 else 532 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 533 PrevInstLoc = DL; 534 535 unsigned FuncId = CurFn->FuncId; 536 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 537 const DILocation *Loc = DL.get(); 538 539 // If this location was actually inlined from somewhere else, give it the ID 540 // of the inline call site. 541 FuncId = 542 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 543 544 // Ensure we have links in the tree of inline call sites. 545 bool FirstLoc = true; 546 while ((SiteLoc = Loc->getInlinedAt())) { 547 InlineSite &Site = 548 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 549 if (!FirstLoc) 550 addLocIfNotPresent(Site.ChildSites, Loc); 551 FirstLoc = false; 552 Loc = SiteLoc; 553 } 554 addLocIfNotPresent(CurFn->ChildSites, Loc); 555 } 556 557 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 558 /*PrologueEnd=*/false, /*IsStmt=*/false, 559 DL->getFilename(), SMLoc()); 560 } 561 562 void CodeViewDebug::emitCodeViewMagicVersion() { 563 OS.emitValueToAlignment(4); 564 OS.AddComment("Debug section magic"); 565 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 566 } 567 568 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 569 switch (DWLang) { 570 case dwarf::DW_LANG_C: 571 case dwarf::DW_LANG_C89: 572 case dwarf::DW_LANG_C99: 573 case dwarf::DW_LANG_C11: 574 case dwarf::DW_LANG_ObjC: 575 return SourceLanguage::C; 576 case dwarf::DW_LANG_C_plus_plus: 577 case dwarf::DW_LANG_C_plus_plus_03: 578 case dwarf::DW_LANG_C_plus_plus_11: 579 case dwarf::DW_LANG_C_plus_plus_14: 580 return SourceLanguage::Cpp; 581 case dwarf::DW_LANG_Fortran77: 582 case dwarf::DW_LANG_Fortran90: 583 case dwarf::DW_LANG_Fortran95: 584 case dwarf::DW_LANG_Fortran03: 585 case dwarf::DW_LANG_Fortran08: 586 return SourceLanguage::Fortran; 587 case dwarf::DW_LANG_Pascal83: 588 return SourceLanguage::Pascal; 589 case dwarf::DW_LANG_Cobol74: 590 case dwarf::DW_LANG_Cobol85: 591 return SourceLanguage::Cobol; 592 case dwarf::DW_LANG_Java: 593 return SourceLanguage::Java; 594 case dwarf::DW_LANG_D: 595 return SourceLanguage::D; 596 case dwarf::DW_LANG_Swift: 597 return SourceLanguage::Swift; 598 case dwarf::DW_LANG_Rust: 599 return SourceLanguage::Rust; 600 default: 601 // There's no CodeView representation for this language, and CV doesn't 602 // have an "unknown" option for the language field, so we'll use MASM, 603 // as it's very low level. 604 return SourceLanguage::Masm; 605 } 606 } 607 608 void CodeViewDebug::beginModule(Module *M) { 609 // If module doesn't have named metadata anchors or COFF debug section 610 // is not available, skip any debug info related stuff. 611 NamedMDNode *CUs = M->getNamedMetadata("llvm.dbg.cu"); 612 if (!CUs || !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 613 Asm = nullptr; 614 return; 615 } 616 // Tell MMI that we have and need debug info. 617 MMI->setDebugInfoAvailability(true); 618 619 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 620 621 // Get the current source language. 622 const MDNode *Node = *CUs->operands().begin(); 623 const auto *CU = cast<DICompileUnit>(Node); 624 625 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 626 627 collectGlobalVariableInfo(); 628 629 // Check if we should emit type record hashes. 630 ConstantInt *GH = 631 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 632 EmitDebugGlobalHashes = GH && !GH->isZero(); 633 } 634 635 void CodeViewDebug::endModule() { 636 if (!Asm || !MMI->hasDebugInfo()) 637 return; 638 639 // The COFF .debug$S section consists of several subsections, each starting 640 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 641 // of the payload followed by the payload itself. The subsections are 4-byte 642 // aligned. 643 644 // Use the generic .debug$S section, and make a subsection for all the inlined 645 // subprograms. 646 switchToDebugSectionForSymbol(nullptr); 647 648 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 649 emitObjName(); 650 emitCompilerInformation(); 651 endCVSubsection(CompilerInfo); 652 653 emitInlineeLinesSubsection(); 654 655 // Emit per-function debug information. 656 for (auto &P : FnDebugInfo) 657 if (!P.first->isDeclarationForLinker()) 658 emitDebugInfoForFunction(P.first, *P.second); 659 660 // Get types used by globals without emitting anything. 661 // This is meant to collect all static const data members so they can be 662 // emitted as globals. 663 collectDebugInfoForGlobals(); 664 665 // Emit retained types. 666 emitDebugInfoForRetainedTypes(); 667 668 // Emit global variable debug information. 669 setCurrentSubprogram(nullptr); 670 emitDebugInfoForGlobals(); 671 672 // Switch back to the generic .debug$S section after potentially processing 673 // comdat symbol sections. 674 switchToDebugSectionForSymbol(nullptr); 675 676 // Emit UDT records for any types used by global variables. 677 if (!GlobalUDTs.empty()) { 678 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 679 emitDebugInfoForUDTs(GlobalUDTs); 680 endCVSubsection(SymbolsEnd); 681 } 682 683 // This subsection holds a file index to offset in string table table. 684 OS.AddComment("File index to string table offset subsection"); 685 OS.emitCVFileChecksumsDirective(); 686 687 // This subsection holds the string table. 688 OS.AddComment("String table"); 689 OS.emitCVStringTableDirective(); 690 691 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 692 // subsection in the generic .debug$S section at the end. There is no 693 // particular reason for this ordering other than to match MSVC. 694 emitBuildInfo(); 695 696 // Emit type information and hashes last, so that any types we translate while 697 // emitting function info are included. 698 emitTypeInformation(); 699 700 if (EmitDebugGlobalHashes) 701 emitTypeGlobalHashes(); 702 703 clear(); 704 } 705 706 static void 707 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 708 unsigned MaxFixedRecordLength = 0xF00) { 709 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 710 // after a fixed length portion of the record. The fixed length portion should 711 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 712 // overall record size is less than the maximum allowed. 713 SmallString<32> NullTerminatedString( 714 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 715 NullTerminatedString.push_back('\0'); 716 OS.emitBytes(NullTerminatedString); 717 } 718 719 void CodeViewDebug::emitTypeInformation() { 720 if (TypeTable.empty()) 721 return; 722 723 // Start the .debug$T or .debug$P section with 0x4. 724 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 725 emitCodeViewMagicVersion(); 726 727 TypeTableCollection Table(TypeTable.records()); 728 TypeVisitorCallbackPipeline Pipeline; 729 730 // To emit type record using Codeview MCStreamer adapter 731 CVMCAdapter CVMCOS(OS, Table); 732 TypeRecordMapping typeMapping(CVMCOS); 733 Pipeline.addCallbackToPipeline(typeMapping); 734 735 Optional<TypeIndex> B = Table.getFirst(); 736 while (B) { 737 // This will fail if the record data is invalid. 738 CVType Record = Table.getType(*B); 739 740 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 741 742 if (E) { 743 logAllUnhandledErrors(std::move(E), errs(), "error: "); 744 llvm_unreachable("produced malformed type record"); 745 } 746 747 B = Table.getNext(*B); 748 } 749 } 750 751 void CodeViewDebug::emitTypeGlobalHashes() { 752 if (TypeTable.empty()) 753 return; 754 755 // Start the .debug$H section with the version and hash algorithm, currently 756 // hardcoded to version 0, SHA1. 757 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 758 759 OS.emitValueToAlignment(4); 760 OS.AddComment("Magic"); 761 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 762 OS.AddComment("Section Version"); 763 OS.emitInt16(0); 764 OS.AddComment("Hash Algorithm"); 765 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 766 767 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 768 for (const auto &GHR : TypeTable.hashes()) { 769 if (OS.isVerboseAsm()) { 770 // Emit an EOL-comment describing which TypeIndex this hash corresponds 771 // to, as well as the stringified SHA1 hash. 772 SmallString<32> Comment; 773 raw_svector_ostream CommentOS(Comment); 774 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 775 OS.AddComment(Comment); 776 ++TI; 777 } 778 assert(GHR.Hash.size() == 8); 779 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 780 GHR.Hash.size()); 781 OS.emitBinaryData(S); 782 } 783 } 784 785 void CodeViewDebug::emitObjName() { 786 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 787 788 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 789 llvm::SmallString<256> PathStore(PathRef); 790 791 if (PathRef.empty() || PathRef == "-") { 792 // Don't emit the filename if we're writing to stdout or to /dev/null. 793 PathRef = {}; 794 } else { 795 llvm::sys::path::remove_dots(PathStore, /*remove_dot_dot=*/true); 796 PathRef = PathStore; 797 } 798 799 OS.AddComment("Signature"); 800 OS.emitIntValue(0, 4); 801 802 OS.AddComment("Object name"); 803 emitNullTerminatedSymbolName(OS, PathRef); 804 805 endSymbolRecord(CompilerEnd); 806 } 807 808 namespace { 809 struct Version { 810 int Part[4]; 811 }; 812 } // end anonymous namespace 813 814 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 815 // the version number. 816 static Version parseVersion(StringRef Name) { 817 Version V = {{0}}; 818 int N = 0; 819 for (const char C : Name) { 820 if (isdigit(C)) { 821 V.Part[N] *= 10; 822 V.Part[N] += C - '0'; 823 V.Part[N] = 824 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 825 } else if (C == '.') { 826 ++N; 827 if (N >= 4) 828 return V; 829 } else if (N > 0) 830 return V; 831 } 832 return V; 833 } 834 835 void CodeViewDebug::emitCompilerInformation() { 836 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 837 uint32_t Flags = 0; 838 839 // The low byte of the flags indicates the source language. 840 Flags = CurrentSourceLanguage; 841 // TODO: Figure out which other flags need to be set. 842 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 843 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 844 } 845 using ArchType = llvm::Triple::ArchType; 846 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 847 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 848 Arch == ArchType::aarch64) { 849 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 850 } 851 852 OS.AddComment("Flags and language"); 853 OS.emitInt32(Flags); 854 855 OS.AddComment("CPUType"); 856 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 857 858 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 859 const MDNode *Node = *CUs->operands().begin(); 860 const auto *CU = cast<DICompileUnit>(Node); 861 862 StringRef CompilerVersion = CU->getProducer(); 863 Version FrontVer = parseVersion(CompilerVersion); 864 OS.AddComment("Frontend version"); 865 for (int N : FrontVer.Part) { 866 OS.emitInt16(N); 867 } 868 869 // Some Microsoft tools, like Binscope, expect a backend version number of at 870 // least 8.something, so we'll coerce the LLVM version into a form that 871 // guarantees it'll be big enough without really lying about the version. 872 int Major = 1000 * LLVM_VERSION_MAJOR + 873 10 * LLVM_VERSION_MINOR + 874 LLVM_VERSION_PATCH; 875 // Clamp it for builds that use unusually large version numbers. 876 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 877 Version BackVer = {{ Major, 0, 0, 0 }}; 878 OS.AddComment("Backend version"); 879 for (int N : BackVer.Part) 880 OS.emitInt16(N); 881 882 OS.AddComment("Null-terminated compiler version string"); 883 emitNullTerminatedSymbolName(OS, CompilerVersion); 884 885 endSymbolRecord(CompilerEnd); 886 } 887 888 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 889 StringRef S) { 890 StringIdRecord SIR(TypeIndex(0x0), S); 891 return TypeTable.writeLeafType(SIR); 892 } 893 894 static std::string flattenCommandLine(ArrayRef<std::string> Args, 895 StringRef MainFilename) { 896 std::string FlatCmdLine; 897 raw_string_ostream OS(FlatCmdLine); 898 bool PrintedOneArg = false; 899 if (!StringRef(Args[0]).contains("-cc1")) { 900 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 901 PrintedOneArg = true; 902 } 903 for (unsigned i = 0; i < Args.size(); i++) { 904 StringRef Arg = Args[i]; 905 if (Arg.empty()) 906 continue; 907 if (Arg == "-main-file-name" || Arg == "-o") { 908 i++; // Skip this argument and next one. 909 continue; 910 } 911 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 912 continue; 913 if (PrintedOneArg) 914 OS << " "; 915 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 916 PrintedOneArg = true; 917 } 918 OS.flush(); 919 return FlatCmdLine; 920 } 921 922 void CodeViewDebug::emitBuildInfo() { 923 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 924 // build info. The known prefix is: 925 // - Absolute path of current directory 926 // - Compiler path 927 // - Main source file path, relative to CWD or absolute 928 // - Type server PDB file 929 // - Canonical compiler command line 930 // If frontend and backend compilation are separated (think llc or LTO), it's 931 // not clear if the compiler path should refer to the executable for the 932 // frontend or the backend. Leave it blank for now. 933 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 934 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 935 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 936 const auto *CU = cast<DICompileUnit>(Node); 937 const DIFile *MainSourceFile = CU->getFile(); 938 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 939 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 940 BuildInfoArgs[BuildInfoRecord::SourceFile] = 941 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 942 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 943 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 944 getStringIdTypeIdx(TypeTable, ""); 945 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 946 BuildInfoArgs[BuildInfoRecord::BuildTool] = 947 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 948 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 949 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 950 MainSourceFile->getFilename())); 951 } 952 BuildInfoRecord BIR(BuildInfoArgs); 953 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 954 955 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 956 // from the module symbols into the type stream. 957 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 958 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 959 OS.AddComment("LF_BUILDINFO index"); 960 OS.emitInt32(BuildInfoIndex.getIndex()); 961 endSymbolRecord(BIEnd); 962 endCVSubsection(BISubsecEnd); 963 } 964 965 void CodeViewDebug::emitInlineeLinesSubsection() { 966 if (InlinedSubprograms.empty()) 967 return; 968 969 OS.AddComment("Inlinee lines subsection"); 970 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 971 972 // We emit the checksum info for files. This is used by debuggers to 973 // determine if a pdb matches the source before loading it. Visual Studio, 974 // for instance, will display a warning that the breakpoints are not valid if 975 // the pdb does not match the source. 976 OS.AddComment("Inlinee lines signature"); 977 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 978 979 for (const DISubprogram *SP : InlinedSubprograms) { 980 assert(TypeIndices.count({SP, nullptr})); 981 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 982 983 OS.AddBlankLine(); 984 unsigned FileId = maybeRecordFile(SP->getFile()); 985 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 986 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 987 OS.AddBlankLine(); 988 OS.AddComment("Type index of inlined function"); 989 OS.emitInt32(InlineeIdx.getIndex()); 990 OS.AddComment("Offset into filechecksum table"); 991 OS.emitCVFileChecksumOffsetDirective(FileId); 992 OS.AddComment("Starting line number"); 993 OS.emitInt32(SP->getLine()); 994 } 995 996 endCVSubsection(InlineEnd); 997 } 998 999 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 1000 const DILocation *InlinedAt, 1001 const InlineSite &Site) { 1002 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1003 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1004 1005 // SymbolRecord 1006 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1007 1008 OS.AddComment("PtrParent"); 1009 OS.emitInt32(0); 1010 OS.AddComment("PtrEnd"); 1011 OS.emitInt32(0); 1012 OS.AddComment("Inlinee type index"); 1013 OS.emitInt32(InlineeIdx.getIndex()); 1014 1015 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1016 unsigned StartLineNum = Site.Inlinee->getLine(); 1017 1018 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1019 FI.Begin, FI.End); 1020 1021 endSymbolRecord(InlineEnd); 1022 1023 emitLocalVariableList(FI, Site.InlinedLocals); 1024 1025 // Recurse on child inlined call sites before closing the scope. 1026 for (const DILocation *ChildSite : Site.ChildSites) { 1027 auto I = FI.InlineSites.find(ChildSite); 1028 assert(I != FI.InlineSites.end() && 1029 "child site not in function inline site map"); 1030 emitInlinedCallSite(FI, ChildSite, I->second); 1031 } 1032 1033 // Close the scope. 1034 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1035 } 1036 1037 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1038 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1039 // comdat key. A section may be comdat because of -ffunction-sections or 1040 // because it is comdat in the IR. 1041 MCSectionCOFF *GVSec = 1042 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1043 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1044 1045 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1046 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1047 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1048 1049 OS.SwitchSection(DebugSec); 1050 1051 // Emit the magic version number if this is the first time we've switched to 1052 // this section. 1053 if (ComdatDebugSections.insert(DebugSec).second) 1054 emitCodeViewMagicVersion(); 1055 } 1056 1057 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1058 // The only supported thunk ordinal is currently the standard type. 1059 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1060 FunctionInfo &FI, 1061 const MCSymbol *Fn) { 1062 std::string FuncName = 1063 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1064 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1065 1066 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1067 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1068 1069 // Emit S_THUNK32 1070 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1071 OS.AddComment("PtrParent"); 1072 OS.emitInt32(0); 1073 OS.AddComment("PtrEnd"); 1074 OS.emitInt32(0); 1075 OS.AddComment("PtrNext"); 1076 OS.emitInt32(0); 1077 OS.AddComment("Thunk section relative address"); 1078 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1079 OS.AddComment("Thunk section index"); 1080 OS.EmitCOFFSectionIndex(Fn); 1081 OS.AddComment("Code size"); 1082 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1083 OS.AddComment("Ordinal"); 1084 OS.emitInt8(unsigned(ordinal)); 1085 OS.AddComment("Function name"); 1086 emitNullTerminatedSymbolName(OS, FuncName); 1087 // Additional fields specific to the thunk ordinal would go here. 1088 endSymbolRecord(ThunkRecordEnd); 1089 1090 // Local variables/inlined routines are purposely omitted here. The point of 1091 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1092 1093 // Emit S_PROC_ID_END 1094 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1095 1096 endCVSubsection(SymbolsEnd); 1097 } 1098 1099 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1100 FunctionInfo &FI) { 1101 // For each function there is a separate subsection which holds the PC to 1102 // file:line table. 1103 const MCSymbol *Fn = Asm->getSymbol(GV); 1104 assert(Fn); 1105 1106 // Switch to the to a comdat section, if appropriate. 1107 switchToDebugSectionForSymbol(Fn); 1108 1109 std::string FuncName; 1110 auto *SP = GV->getSubprogram(); 1111 assert(SP); 1112 setCurrentSubprogram(SP); 1113 1114 if (SP->isThunk()) { 1115 emitDebugInfoForThunk(GV, FI, Fn); 1116 return; 1117 } 1118 1119 // If we have a display name, build the fully qualified name by walking the 1120 // chain of scopes. 1121 if (!SP->getName().empty()) 1122 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1123 1124 // If our DISubprogram name is empty, use the mangled name. 1125 if (FuncName.empty()) 1126 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1127 1128 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1129 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1130 OS.EmitCVFPOData(Fn); 1131 1132 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1133 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1134 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1135 { 1136 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1137 : SymbolKind::S_GPROC32_ID; 1138 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1139 1140 // These fields are filled in by tools like CVPACK which run after the fact. 1141 OS.AddComment("PtrParent"); 1142 OS.emitInt32(0); 1143 OS.AddComment("PtrEnd"); 1144 OS.emitInt32(0); 1145 OS.AddComment("PtrNext"); 1146 OS.emitInt32(0); 1147 // This is the important bit that tells the debugger where the function 1148 // code is located and what's its size: 1149 OS.AddComment("Code size"); 1150 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1151 OS.AddComment("Offset after prologue"); 1152 OS.emitInt32(0); 1153 OS.AddComment("Offset before epilogue"); 1154 OS.emitInt32(0); 1155 OS.AddComment("Function type index"); 1156 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1157 OS.AddComment("Function section relative address"); 1158 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1159 OS.AddComment("Function section index"); 1160 OS.EmitCOFFSectionIndex(Fn); 1161 OS.AddComment("Flags"); 1162 OS.emitInt8(0); 1163 // Emit the function display name as a null-terminated string. 1164 OS.AddComment("Function name"); 1165 // Truncate the name so we won't overflow the record length field. 1166 emitNullTerminatedSymbolName(OS, FuncName); 1167 endSymbolRecord(ProcRecordEnd); 1168 1169 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1170 // Subtract out the CSR size since MSVC excludes that and we include it. 1171 OS.AddComment("FrameSize"); 1172 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1173 OS.AddComment("Padding"); 1174 OS.emitInt32(0); 1175 OS.AddComment("Offset of padding"); 1176 OS.emitInt32(0); 1177 OS.AddComment("Bytes of callee saved registers"); 1178 OS.emitInt32(FI.CSRSize); 1179 OS.AddComment("Exception handler offset"); 1180 OS.emitInt32(0); 1181 OS.AddComment("Exception handler section"); 1182 OS.emitInt16(0); 1183 OS.AddComment("Flags (defines frame register)"); 1184 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1185 endSymbolRecord(FrameProcEnd); 1186 1187 emitLocalVariableList(FI, FI.Locals); 1188 emitGlobalVariableList(FI.Globals); 1189 emitLexicalBlockList(FI.ChildBlocks, FI); 1190 1191 // Emit inlined call site information. Only emit functions inlined directly 1192 // into the parent function. We'll emit the other sites recursively as part 1193 // of their parent inline site. 1194 for (const DILocation *InlinedAt : FI.ChildSites) { 1195 auto I = FI.InlineSites.find(InlinedAt); 1196 assert(I != FI.InlineSites.end() && 1197 "child site not in function inline site map"); 1198 emitInlinedCallSite(FI, InlinedAt, I->second); 1199 } 1200 1201 for (auto Annot : FI.Annotations) { 1202 MCSymbol *Label = Annot.first; 1203 MDTuple *Strs = cast<MDTuple>(Annot.second); 1204 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1205 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1206 // FIXME: Make sure we don't overflow the max record size. 1207 OS.EmitCOFFSectionIndex(Label); 1208 OS.emitInt16(Strs->getNumOperands()); 1209 for (Metadata *MD : Strs->operands()) { 1210 // MDStrings are null terminated, so we can do EmitBytes and get the 1211 // nice .asciz directive. 1212 StringRef Str = cast<MDString>(MD)->getString(); 1213 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1214 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1215 } 1216 endSymbolRecord(AnnotEnd); 1217 } 1218 1219 for (auto HeapAllocSite : FI.HeapAllocSites) { 1220 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1221 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1222 const DIType *DITy = std::get<2>(HeapAllocSite); 1223 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1224 OS.AddComment("Call site offset"); 1225 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1226 OS.AddComment("Call site section index"); 1227 OS.EmitCOFFSectionIndex(BeginLabel); 1228 OS.AddComment("Call instruction length"); 1229 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1230 OS.AddComment("Type index"); 1231 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1232 endSymbolRecord(HeapAllocEnd); 1233 } 1234 1235 if (SP != nullptr) 1236 emitDebugInfoForUDTs(LocalUDTs); 1237 1238 // We're done with this function. 1239 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1240 } 1241 endCVSubsection(SymbolsEnd); 1242 1243 // We have an assembler directive that takes care of the whole line table. 1244 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1245 } 1246 1247 CodeViewDebug::LocalVarDefRange 1248 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1249 LocalVarDefRange DR; 1250 DR.InMemory = -1; 1251 DR.DataOffset = Offset; 1252 assert(DR.DataOffset == Offset && "truncation"); 1253 DR.IsSubfield = 0; 1254 DR.StructOffset = 0; 1255 DR.CVRegister = CVRegister; 1256 return DR; 1257 } 1258 1259 void CodeViewDebug::collectVariableInfoFromMFTable( 1260 DenseSet<InlinedEntity> &Processed) { 1261 const MachineFunction &MF = *Asm->MF; 1262 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1263 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1264 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1265 1266 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1267 if (!VI.Var) 1268 continue; 1269 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1270 "Expected inlined-at fields to agree"); 1271 1272 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1273 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1274 1275 // If variable scope is not found then skip this variable. 1276 if (!Scope) 1277 continue; 1278 1279 // If the variable has an attached offset expression, extract it. 1280 // FIXME: Try to handle DW_OP_deref as well. 1281 int64_t ExprOffset = 0; 1282 bool Deref = false; 1283 if (VI.Expr) { 1284 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1285 if (VI.Expr->getNumElements() == 1 && 1286 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1287 Deref = true; 1288 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1289 continue; 1290 } 1291 1292 // Get the frame register used and the offset. 1293 Register FrameReg; 1294 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1295 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1296 1297 assert(!FrameOffset.getScalable() && 1298 "Frame offsets with a scalable component are not supported"); 1299 1300 // Calculate the label ranges. 1301 LocalVarDefRange DefRange = 1302 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1303 1304 for (const InsnRange &Range : Scope->getRanges()) { 1305 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1306 const MCSymbol *End = getLabelAfterInsn(Range.second); 1307 End = End ? End : Asm->getFunctionEnd(); 1308 DefRange.Ranges.emplace_back(Begin, End); 1309 } 1310 1311 LocalVariable Var; 1312 Var.DIVar = VI.Var; 1313 Var.DefRanges.emplace_back(std::move(DefRange)); 1314 if (Deref) 1315 Var.UseReferenceType = true; 1316 1317 recordLocalVariable(std::move(Var), Scope); 1318 } 1319 } 1320 1321 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1322 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1323 } 1324 1325 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1326 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1327 } 1328 1329 void CodeViewDebug::calculateRanges( 1330 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1331 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1332 1333 // Calculate the definition ranges. 1334 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1335 const auto &Entry = *I; 1336 if (!Entry.isDbgValue()) 1337 continue; 1338 const MachineInstr *DVInst = Entry.getInstr(); 1339 assert(DVInst->isDebugValue() && "Invalid History entry"); 1340 // FIXME: Find a way to represent constant variables, since they are 1341 // relatively common. 1342 Optional<DbgVariableLocation> Location = 1343 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1344 if (!Location) 1345 continue; 1346 1347 // CodeView can only express variables in register and variables in memory 1348 // at a constant offset from a register. However, for variables passed 1349 // indirectly by pointer, it is common for that pointer to be spilled to a 1350 // stack location. For the special case of one offseted load followed by a 1351 // zero offset load (a pointer spilled to the stack), we change the type of 1352 // the local variable from a value type to a reference type. This tricks the 1353 // debugger into doing the load for us. 1354 if (Var.UseReferenceType) { 1355 // We're using a reference type. Drop the last zero offset load. 1356 if (canUseReferenceType(*Location)) 1357 Location->LoadChain.pop_back(); 1358 else 1359 continue; 1360 } else if (needsReferenceType(*Location)) { 1361 // This location can't be expressed without switching to a reference type. 1362 // Start over using that. 1363 Var.UseReferenceType = true; 1364 Var.DefRanges.clear(); 1365 calculateRanges(Var, Entries); 1366 return; 1367 } 1368 1369 // We can only handle a register or an offseted load of a register. 1370 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1371 continue; 1372 { 1373 LocalVarDefRange DR; 1374 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1375 DR.InMemory = !Location->LoadChain.empty(); 1376 DR.DataOffset = 1377 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1378 if (Location->FragmentInfo) { 1379 DR.IsSubfield = true; 1380 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1381 } else { 1382 DR.IsSubfield = false; 1383 DR.StructOffset = 0; 1384 } 1385 1386 if (Var.DefRanges.empty() || 1387 Var.DefRanges.back().isDifferentLocation(DR)) { 1388 Var.DefRanges.emplace_back(std::move(DR)); 1389 } 1390 } 1391 1392 // Compute the label range. 1393 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1394 const MCSymbol *End; 1395 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1396 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1397 End = EndingEntry.isDbgValue() 1398 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1399 : getLabelAfterInsn(EndingEntry.getInstr()); 1400 } else 1401 End = Asm->getFunctionEnd(); 1402 1403 // If the last range end is our begin, just extend the last range. 1404 // Otherwise make a new range. 1405 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1406 Var.DefRanges.back().Ranges; 1407 if (!R.empty() && R.back().second == Begin) 1408 R.back().second = End; 1409 else 1410 R.emplace_back(Begin, End); 1411 1412 // FIXME: Do more range combining. 1413 } 1414 } 1415 1416 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1417 DenseSet<InlinedEntity> Processed; 1418 // Grab the variable info that was squirreled away in the MMI side-table. 1419 collectVariableInfoFromMFTable(Processed); 1420 1421 for (const auto &I : DbgValues) { 1422 InlinedEntity IV = I.first; 1423 if (Processed.count(IV)) 1424 continue; 1425 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1426 const DILocation *InlinedAt = IV.second; 1427 1428 // Instruction ranges, specifying where IV is accessible. 1429 const auto &Entries = I.second; 1430 1431 LexicalScope *Scope = nullptr; 1432 if (InlinedAt) 1433 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1434 else 1435 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1436 // If variable scope is not found then skip this variable. 1437 if (!Scope) 1438 continue; 1439 1440 LocalVariable Var; 1441 Var.DIVar = DIVar; 1442 1443 calculateRanges(Var, Entries); 1444 recordLocalVariable(std::move(Var), Scope); 1445 } 1446 } 1447 1448 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1449 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1450 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1451 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1452 const Function &GV = MF->getFunction(); 1453 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1454 assert(Insertion.second && "function already has info"); 1455 CurFn = Insertion.first->second.get(); 1456 CurFn->FuncId = NextFuncId++; 1457 CurFn->Begin = Asm->getFunctionBegin(); 1458 1459 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1460 // callee-saved registers were used. For targets that don't use a PUSH 1461 // instruction (AArch64), this will be zero. 1462 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1463 CurFn->FrameSize = MFI.getStackSize(); 1464 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1465 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1466 1467 // For this function S_FRAMEPROC record, figure out which codeview register 1468 // will be the frame pointer. 1469 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1470 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1471 if (CurFn->FrameSize > 0) { 1472 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1473 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1474 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1475 } else { 1476 // If there is an FP, parameters are always relative to it. 1477 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1478 if (CurFn->HasStackRealignment) { 1479 // If the stack needs realignment, locals are relative to SP or VFRAME. 1480 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1481 } else { 1482 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1483 // other stack adjustments. 1484 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1485 } 1486 } 1487 } 1488 1489 // Compute other frame procedure options. 1490 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1491 if (MFI.hasVarSizedObjects()) 1492 FPO |= FrameProcedureOptions::HasAlloca; 1493 if (MF->exposesReturnsTwice()) 1494 FPO |= FrameProcedureOptions::HasSetJmp; 1495 // FIXME: Set HasLongJmp if we ever track that info. 1496 if (MF->hasInlineAsm()) 1497 FPO |= FrameProcedureOptions::HasInlineAssembly; 1498 if (GV.hasPersonalityFn()) { 1499 if (isAsynchronousEHPersonality( 1500 classifyEHPersonality(GV.getPersonalityFn()))) 1501 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1502 else 1503 FPO |= FrameProcedureOptions::HasExceptionHandling; 1504 } 1505 if (GV.hasFnAttribute(Attribute::InlineHint)) 1506 FPO |= FrameProcedureOptions::MarkedInline; 1507 if (GV.hasFnAttribute(Attribute::Naked)) 1508 FPO |= FrameProcedureOptions::Naked; 1509 if (MFI.hasStackProtectorIndex()) 1510 FPO |= FrameProcedureOptions::SecurityChecks; 1511 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1512 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1513 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1514 !GV.hasOptSize() && !GV.hasOptNone()) 1515 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1516 if (GV.hasProfileData()) { 1517 FPO |= FrameProcedureOptions::ValidProfileCounts; 1518 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1519 } 1520 // FIXME: Set GuardCfg when it is implemented. 1521 CurFn->FrameProcOpts = FPO; 1522 1523 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1524 1525 // Find the end of the function prolog. First known non-DBG_VALUE and 1526 // non-frame setup location marks the beginning of the function body. 1527 // FIXME: is there a simpler a way to do this? Can we just search 1528 // for the first instruction of the function, not the last of the prolog? 1529 DebugLoc PrologEndLoc; 1530 bool EmptyPrologue = true; 1531 for (const auto &MBB : *MF) { 1532 for (const auto &MI : MBB) { 1533 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1534 MI.getDebugLoc()) { 1535 PrologEndLoc = MI.getDebugLoc(); 1536 break; 1537 } else if (!MI.isMetaInstruction()) { 1538 EmptyPrologue = false; 1539 } 1540 } 1541 } 1542 1543 // Record beginning of function if we have a non-empty prologue. 1544 if (PrologEndLoc && !EmptyPrologue) { 1545 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1546 maybeRecordLocation(FnStartDL, MF); 1547 } 1548 1549 // Find heap alloc sites and emit labels around them. 1550 for (const auto &MBB : *MF) { 1551 for (const auto &MI : MBB) { 1552 if (MI.getHeapAllocMarker()) { 1553 requestLabelBeforeInsn(&MI); 1554 requestLabelAfterInsn(&MI); 1555 } 1556 } 1557 } 1558 } 1559 1560 static bool shouldEmitUdt(const DIType *T) { 1561 if (!T) 1562 return false; 1563 1564 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1565 if (T->getTag() == dwarf::DW_TAG_typedef) { 1566 if (DIScope *Scope = T->getScope()) { 1567 switch (Scope->getTag()) { 1568 case dwarf::DW_TAG_structure_type: 1569 case dwarf::DW_TAG_class_type: 1570 case dwarf::DW_TAG_union_type: 1571 return false; 1572 default: 1573 // do nothing. 1574 ; 1575 } 1576 } 1577 } 1578 1579 while (true) { 1580 if (!T || T->isForwardDecl()) 1581 return false; 1582 1583 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1584 if (!DT) 1585 return true; 1586 T = DT->getBaseType(); 1587 } 1588 return true; 1589 } 1590 1591 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1592 // Don't record empty UDTs. 1593 if (Ty->getName().empty()) 1594 return; 1595 if (!shouldEmitUdt(Ty)) 1596 return; 1597 1598 SmallVector<StringRef, 5> ParentScopeNames; 1599 const DISubprogram *ClosestSubprogram = 1600 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1601 1602 std::string FullyQualifiedName = 1603 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1604 1605 if (ClosestSubprogram == nullptr) { 1606 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1607 } else if (ClosestSubprogram == CurrentSubprogram) { 1608 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1609 } 1610 1611 // TODO: What if the ClosestSubprogram is neither null or the current 1612 // subprogram? Currently, the UDT just gets dropped on the floor. 1613 // 1614 // The current behavior is not desirable. To get maximal fidelity, we would 1615 // need to perform all type translation before beginning emission of .debug$S 1616 // and then make LocalUDTs a member of FunctionInfo 1617 } 1618 1619 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1620 // Generic dispatch for lowering an unknown type. 1621 switch (Ty->getTag()) { 1622 case dwarf::DW_TAG_array_type: 1623 return lowerTypeArray(cast<DICompositeType>(Ty)); 1624 case dwarf::DW_TAG_typedef: 1625 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1626 case dwarf::DW_TAG_base_type: 1627 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1628 case dwarf::DW_TAG_pointer_type: 1629 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1630 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1631 LLVM_FALLTHROUGH; 1632 case dwarf::DW_TAG_reference_type: 1633 case dwarf::DW_TAG_rvalue_reference_type: 1634 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1635 case dwarf::DW_TAG_ptr_to_member_type: 1636 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1637 case dwarf::DW_TAG_restrict_type: 1638 case dwarf::DW_TAG_const_type: 1639 case dwarf::DW_TAG_volatile_type: 1640 // TODO: add support for DW_TAG_atomic_type here 1641 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1642 case dwarf::DW_TAG_subroutine_type: 1643 if (ClassTy) { 1644 // The member function type of a member function pointer has no 1645 // ThisAdjustment. 1646 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1647 /*ThisAdjustment=*/0, 1648 /*IsStaticMethod=*/false); 1649 } 1650 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1651 case dwarf::DW_TAG_enumeration_type: 1652 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1653 case dwarf::DW_TAG_class_type: 1654 case dwarf::DW_TAG_structure_type: 1655 return lowerTypeClass(cast<DICompositeType>(Ty)); 1656 case dwarf::DW_TAG_union_type: 1657 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1658 case dwarf::DW_TAG_string_type: 1659 return lowerTypeString(cast<DIStringType>(Ty)); 1660 case dwarf::DW_TAG_unspecified_type: 1661 if (Ty->getName() == "decltype(nullptr)") 1662 return TypeIndex::NullptrT(); 1663 return TypeIndex::None(); 1664 default: 1665 // Use the null type index. 1666 return TypeIndex(); 1667 } 1668 } 1669 1670 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1671 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1672 StringRef TypeName = Ty->getName(); 1673 1674 addToUDTs(Ty); 1675 1676 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1677 TypeName == "HRESULT") 1678 return TypeIndex(SimpleTypeKind::HResult); 1679 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1680 TypeName == "wchar_t") 1681 return TypeIndex(SimpleTypeKind::WideCharacter); 1682 1683 return UnderlyingTypeIndex; 1684 } 1685 1686 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1687 const DIType *ElementType = Ty->getBaseType(); 1688 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1689 // IndexType is size_t, which depends on the bitness of the target. 1690 TypeIndex IndexType = getPointerSizeInBytes() == 8 1691 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1692 : TypeIndex(SimpleTypeKind::UInt32Long); 1693 1694 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1695 1696 // Add subranges to array type. 1697 DINodeArray Elements = Ty->getElements(); 1698 for (int i = Elements.size() - 1; i >= 0; --i) { 1699 const DINode *Element = Elements[i]; 1700 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1701 1702 const DISubrange *Subrange = cast<DISubrange>(Element); 1703 int64_t Count = -1; 1704 1705 // If Subrange has a Count field, use it. 1706 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1707 // where lowerbound is from the LowerBound field of the Subrange, 1708 // or the language default lowerbound if that field is unspecified. 1709 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1710 Count = CI->getSExtValue(); 1711 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1712 // Fortran uses 1 as the default lowerbound; other languages use 0. 1713 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1714 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1715 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1716 Count = UI->getSExtValue() - Lowerbound + 1; 1717 } 1718 1719 // Forward declarations of arrays without a size and VLAs use a count of -1. 1720 // Emit a count of zero in these cases to match what MSVC does for arrays 1721 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1722 // should do for them even if we could distinguish them. 1723 if (Count == -1) 1724 Count = 0; 1725 1726 // Update the element size and element type index for subsequent subranges. 1727 ElementSize *= Count; 1728 1729 // If this is the outermost array, use the size from the array. It will be 1730 // more accurate if we had a VLA or an incomplete element type size. 1731 uint64_t ArraySize = 1732 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1733 1734 StringRef Name = (i == 0) ? Ty->getName() : ""; 1735 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1736 ElementTypeIndex = TypeTable.writeLeafType(AR); 1737 } 1738 1739 return ElementTypeIndex; 1740 } 1741 1742 // This function lowers a Fortran character type (DIStringType). 1743 // Note that it handles only the character*n variant (using SizeInBits 1744 // field in DIString to describe the type size) at the moment. 1745 // Other variants (leveraging the StringLength and StringLengthExp 1746 // fields in DIStringType) remain TBD. 1747 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1748 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1749 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1750 StringRef Name = Ty->getName(); 1751 // IndexType is size_t, which depends on the bitness of the target. 1752 TypeIndex IndexType = getPointerSizeInBytes() == 8 1753 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1754 : TypeIndex(SimpleTypeKind::UInt32Long); 1755 1756 // Create a type of character array of ArraySize. 1757 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1758 1759 return TypeTable.writeLeafType(AR); 1760 } 1761 1762 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1763 TypeIndex Index; 1764 dwarf::TypeKind Kind; 1765 uint32_t ByteSize; 1766 1767 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1768 ByteSize = Ty->getSizeInBits() / 8; 1769 1770 SimpleTypeKind STK = SimpleTypeKind::None; 1771 switch (Kind) { 1772 case dwarf::DW_ATE_address: 1773 // FIXME: Translate 1774 break; 1775 case dwarf::DW_ATE_boolean: 1776 switch (ByteSize) { 1777 case 1: STK = SimpleTypeKind::Boolean8; break; 1778 case 2: STK = SimpleTypeKind::Boolean16; break; 1779 case 4: STK = SimpleTypeKind::Boolean32; break; 1780 case 8: STK = SimpleTypeKind::Boolean64; break; 1781 case 16: STK = SimpleTypeKind::Boolean128; break; 1782 } 1783 break; 1784 case dwarf::DW_ATE_complex_float: 1785 switch (ByteSize) { 1786 case 2: STK = SimpleTypeKind::Complex16; break; 1787 case 4: STK = SimpleTypeKind::Complex32; break; 1788 case 8: STK = SimpleTypeKind::Complex64; break; 1789 case 10: STK = SimpleTypeKind::Complex80; break; 1790 case 16: STK = SimpleTypeKind::Complex128; break; 1791 } 1792 break; 1793 case dwarf::DW_ATE_float: 1794 switch (ByteSize) { 1795 case 2: STK = SimpleTypeKind::Float16; break; 1796 case 4: STK = SimpleTypeKind::Float32; break; 1797 case 6: STK = SimpleTypeKind::Float48; break; 1798 case 8: STK = SimpleTypeKind::Float64; break; 1799 case 10: STK = SimpleTypeKind::Float80; break; 1800 case 16: STK = SimpleTypeKind::Float128; break; 1801 } 1802 break; 1803 case dwarf::DW_ATE_signed: 1804 switch (ByteSize) { 1805 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1806 case 2: STK = SimpleTypeKind::Int16Short; break; 1807 case 4: STK = SimpleTypeKind::Int32; break; 1808 case 8: STK = SimpleTypeKind::Int64Quad; break; 1809 case 16: STK = SimpleTypeKind::Int128Oct; break; 1810 } 1811 break; 1812 case dwarf::DW_ATE_unsigned: 1813 switch (ByteSize) { 1814 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1815 case 2: STK = SimpleTypeKind::UInt16Short; break; 1816 case 4: STK = SimpleTypeKind::UInt32; break; 1817 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1818 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1819 } 1820 break; 1821 case dwarf::DW_ATE_UTF: 1822 switch (ByteSize) { 1823 case 1: STK = SimpleTypeKind::Character8; break; 1824 case 2: STK = SimpleTypeKind::Character16; break; 1825 case 4: STK = SimpleTypeKind::Character32; break; 1826 } 1827 break; 1828 case dwarf::DW_ATE_signed_char: 1829 if (ByteSize == 1) 1830 STK = SimpleTypeKind::SignedCharacter; 1831 break; 1832 case dwarf::DW_ATE_unsigned_char: 1833 if (ByteSize == 1) 1834 STK = SimpleTypeKind::UnsignedCharacter; 1835 break; 1836 default: 1837 break; 1838 } 1839 1840 // Apply some fixups based on the source-level type name. 1841 // Include some amount of canonicalization from an old naming scheme Clang 1842 // used to use for integer types (in an outdated effort to be compatible with 1843 // GCC's debug info/GDB's behavior, which has since been addressed). 1844 if (STK == SimpleTypeKind::Int32 && 1845 (Ty->getName() == "long int" || Ty->getName() == "long")) 1846 STK = SimpleTypeKind::Int32Long; 1847 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1848 Ty->getName() == "unsigned long")) 1849 STK = SimpleTypeKind::UInt32Long; 1850 if (STK == SimpleTypeKind::UInt16Short && 1851 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1852 STK = SimpleTypeKind::WideCharacter; 1853 if ((STK == SimpleTypeKind::SignedCharacter || 1854 STK == SimpleTypeKind::UnsignedCharacter) && 1855 Ty->getName() == "char") 1856 STK = SimpleTypeKind::NarrowCharacter; 1857 1858 return TypeIndex(STK); 1859 } 1860 1861 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1862 PointerOptions PO) { 1863 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1864 1865 // Pointers to simple types without any options can use SimpleTypeMode, rather 1866 // than having a dedicated pointer type record. 1867 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1868 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1869 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1870 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1871 ? SimpleTypeMode::NearPointer64 1872 : SimpleTypeMode::NearPointer32; 1873 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1874 } 1875 1876 PointerKind PK = 1877 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1878 PointerMode PM = PointerMode::Pointer; 1879 switch (Ty->getTag()) { 1880 default: llvm_unreachable("not a pointer tag type"); 1881 case dwarf::DW_TAG_pointer_type: 1882 PM = PointerMode::Pointer; 1883 break; 1884 case dwarf::DW_TAG_reference_type: 1885 PM = PointerMode::LValueReference; 1886 break; 1887 case dwarf::DW_TAG_rvalue_reference_type: 1888 PM = PointerMode::RValueReference; 1889 break; 1890 } 1891 1892 if (Ty->isObjectPointer()) 1893 PO |= PointerOptions::Const; 1894 1895 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1896 return TypeTable.writeLeafType(PR); 1897 } 1898 1899 static PointerToMemberRepresentation 1900 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1901 // SizeInBytes being zero generally implies that the member pointer type was 1902 // incomplete, which can happen if it is part of a function prototype. In this 1903 // case, use the unknown model instead of the general model. 1904 if (IsPMF) { 1905 switch (Flags & DINode::FlagPtrToMemberRep) { 1906 case 0: 1907 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1908 : PointerToMemberRepresentation::GeneralFunction; 1909 case DINode::FlagSingleInheritance: 1910 return PointerToMemberRepresentation::SingleInheritanceFunction; 1911 case DINode::FlagMultipleInheritance: 1912 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1913 case DINode::FlagVirtualInheritance: 1914 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1915 } 1916 } else { 1917 switch (Flags & DINode::FlagPtrToMemberRep) { 1918 case 0: 1919 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1920 : PointerToMemberRepresentation::GeneralData; 1921 case DINode::FlagSingleInheritance: 1922 return PointerToMemberRepresentation::SingleInheritanceData; 1923 case DINode::FlagMultipleInheritance: 1924 return PointerToMemberRepresentation::MultipleInheritanceData; 1925 case DINode::FlagVirtualInheritance: 1926 return PointerToMemberRepresentation::VirtualInheritanceData; 1927 } 1928 } 1929 llvm_unreachable("invalid ptr to member representation"); 1930 } 1931 1932 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1933 PointerOptions PO) { 1934 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1935 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1936 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1937 TypeIndex PointeeTI = 1938 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1939 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1940 : PointerKind::Near32; 1941 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1942 : PointerMode::PointerToDataMember; 1943 1944 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1945 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1946 MemberPointerInfo MPI( 1947 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1948 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1949 return TypeTable.writeLeafType(PR); 1950 } 1951 1952 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1953 /// have a translation, use the NearC convention. 1954 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1955 switch (DwarfCC) { 1956 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1957 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1958 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1959 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1960 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1961 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1962 } 1963 return CallingConvention::NearC; 1964 } 1965 1966 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1967 ModifierOptions Mods = ModifierOptions::None; 1968 PointerOptions PO = PointerOptions::None; 1969 bool IsModifier = true; 1970 const DIType *BaseTy = Ty; 1971 while (IsModifier && BaseTy) { 1972 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1973 switch (BaseTy->getTag()) { 1974 case dwarf::DW_TAG_const_type: 1975 Mods |= ModifierOptions::Const; 1976 PO |= PointerOptions::Const; 1977 break; 1978 case dwarf::DW_TAG_volatile_type: 1979 Mods |= ModifierOptions::Volatile; 1980 PO |= PointerOptions::Volatile; 1981 break; 1982 case dwarf::DW_TAG_restrict_type: 1983 // Only pointer types be marked with __restrict. There is no known flag 1984 // for __restrict in LF_MODIFIER records. 1985 PO |= PointerOptions::Restrict; 1986 break; 1987 default: 1988 IsModifier = false; 1989 break; 1990 } 1991 if (IsModifier) 1992 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1993 } 1994 1995 // Check if the inner type will use an LF_POINTER record. If so, the 1996 // qualifiers will go in the LF_POINTER record. This comes up for types like 1997 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1998 // char *'. 1999 if (BaseTy) { 2000 switch (BaseTy->getTag()) { 2001 case dwarf::DW_TAG_pointer_type: 2002 case dwarf::DW_TAG_reference_type: 2003 case dwarf::DW_TAG_rvalue_reference_type: 2004 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2005 case dwarf::DW_TAG_ptr_to_member_type: 2006 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2007 default: 2008 break; 2009 } 2010 } 2011 2012 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2013 2014 // Return the base type index if there aren't any modifiers. For example, the 2015 // metadata could contain restrict wrappers around non-pointer types. 2016 if (Mods == ModifierOptions::None) 2017 return ModifiedTI; 2018 2019 ModifierRecord MR(ModifiedTI, Mods); 2020 return TypeTable.writeLeafType(MR); 2021 } 2022 2023 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2024 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2025 for (const DIType *ArgType : Ty->getTypeArray()) 2026 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2027 2028 // MSVC uses type none for variadic argument. 2029 if (ReturnAndArgTypeIndices.size() > 1 && 2030 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2031 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2032 } 2033 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2034 ArrayRef<TypeIndex> ArgTypeIndices = None; 2035 if (!ReturnAndArgTypeIndices.empty()) { 2036 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 2037 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2038 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2039 } 2040 2041 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2042 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2043 2044 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2045 2046 FunctionOptions FO = getFunctionOptions(Ty); 2047 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2048 ArgListIndex); 2049 return TypeTable.writeLeafType(Procedure); 2050 } 2051 2052 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2053 const DIType *ClassTy, 2054 int ThisAdjustment, 2055 bool IsStaticMethod, 2056 FunctionOptions FO) { 2057 // Lower the containing class type. 2058 TypeIndex ClassType = getTypeIndex(ClassTy); 2059 2060 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2061 2062 unsigned Index = 0; 2063 SmallVector<TypeIndex, 8> ArgTypeIndices; 2064 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2065 if (ReturnAndArgs.size() > Index) { 2066 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2067 } 2068 2069 // If the first argument is a pointer type and this isn't a static method, 2070 // treat it as the special 'this' parameter, which is encoded separately from 2071 // the arguments. 2072 TypeIndex ThisTypeIndex; 2073 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2074 if (const DIDerivedType *PtrTy = 2075 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2076 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2077 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2078 Index++; 2079 } 2080 } 2081 } 2082 2083 while (Index < ReturnAndArgs.size()) 2084 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2085 2086 // MSVC uses type none for variadic argument. 2087 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2088 ArgTypeIndices.back() = TypeIndex::None(); 2089 2090 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2091 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2092 2093 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2094 2095 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2096 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2097 return TypeTable.writeLeafType(MFR); 2098 } 2099 2100 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2101 unsigned VSlotCount = 2102 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2103 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2104 2105 VFTableShapeRecord VFTSR(Slots); 2106 return TypeTable.writeLeafType(VFTSR); 2107 } 2108 2109 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2110 switch (Flags & DINode::FlagAccessibility) { 2111 case DINode::FlagPrivate: return MemberAccess::Private; 2112 case DINode::FlagPublic: return MemberAccess::Public; 2113 case DINode::FlagProtected: return MemberAccess::Protected; 2114 case 0: 2115 // If there was no explicit access control, provide the default for the tag. 2116 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2117 : MemberAccess::Public; 2118 } 2119 llvm_unreachable("access flags are exclusive"); 2120 } 2121 2122 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2123 if (SP->isArtificial()) 2124 return MethodOptions::CompilerGenerated; 2125 2126 // FIXME: Handle other MethodOptions. 2127 2128 return MethodOptions::None; 2129 } 2130 2131 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2132 bool Introduced) { 2133 if (SP->getFlags() & DINode::FlagStaticMember) 2134 return MethodKind::Static; 2135 2136 switch (SP->getVirtuality()) { 2137 case dwarf::DW_VIRTUALITY_none: 2138 break; 2139 case dwarf::DW_VIRTUALITY_virtual: 2140 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2141 case dwarf::DW_VIRTUALITY_pure_virtual: 2142 return Introduced ? MethodKind::PureIntroducingVirtual 2143 : MethodKind::PureVirtual; 2144 default: 2145 llvm_unreachable("unhandled virtuality case"); 2146 } 2147 2148 return MethodKind::Vanilla; 2149 } 2150 2151 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2152 switch (Ty->getTag()) { 2153 case dwarf::DW_TAG_class_type: 2154 return TypeRecordKind::Class; 2155 case dwarf::DW_TAG_structure_type: 2156 return TypeRecordKind::Struct; 2157 default: 2158 llvm_unreachable("unexpected tag"); 2159 } 2160 } 2161 2162 /// Return ClassOptions that should be present on both the forward declaration 2163 /// and the defintion of a tag type. 2164 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2165 ClassOptions CO = ClassOptions::None; 2166 2167 // MSVC always sets this flag, even for local types. Clang doesn't always 2168 // appear to give every type a linkage name, which may be problematic for us. 2169 // FIXME: Investigate the consequences of not following them here. 2170 if (!Ty->getIdentifier().empty()) 2171 CO |= ClassOptions::HasUniqueName; 2172 2173 // Put the Nested flag on a type if it appears immediately inside a tag type. 2174 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2175 // here. That flag is only set on definitions, and not forward declarations. 2176 const DIScope *ImmediateScope = Ty->getScope(); 2177 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2178 CO |= ClassOptions::Nested; 2179 2180 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2181 // type only when it has an immediate function scope. Clang never puts enums 2182 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2183 // always in function, class, or file scopes. 2184 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2185 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2186 CO |= ClassOptions::Scoped; 2187 } else { 2188 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2189 Scope = Scope->getScope()) { 2190 if (isa<DISubprogram>(Scope)) { 2191 CO |= ClassOptions::Scoped; 2192 break; 2193 } 2194 } 2195 } 2196 2197 return CO; 2198 } 2199 2200 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2201 switch (Ty->getTag()) { 2202 case dwarf::DW_TAG_class_type: 2203 case dwarf::DW_TAG_structure_type: 2204 case dwarf::DW_TAG_union_type: 2205 case dwarf::DW_TAG_enumeration_type: 2206 break; 2207 default: 2208 return; 2209 } 2210 2211 if (const auto *File = Ty->getFile()) { 2212 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2213 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2214 2215 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2216 TypeTable.writeLeafType(USLR); 2217 } 2218 } 2219 2220 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2221 ClassOptions CO = getCommonClassOptions(Ty); 2222 TypeIndex FTI; 2223 unsigned EnumeratorCount = 0; 2224 2225 if (Ty->isForwardDecl()) { 2226 CO |= ClassOptions::ForwardReference; 2227 } else { 2228 ContinuationRecordBuilder ContinuationBuilder; 2229 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2230 for (const DINode *Element : Ty->getElements()) { 2231 // We assume that the frontend provides all members in source declaration 2232 // order, which is what MSVC does. 2233 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2234 // FIXME: Is it correct to always emit these as unsigned here? 2235 EnumeratorRecord ER(MemberAccess::Public, 2236 APSInt(Enumerator->getValue(), true), 2237 Enumerator->getName()); 2238 ContinuationBuilder.writeMemberType(ER); 2239 EnumeratorCount++; 2240 } 2241 } 2242 FTI = TypeTable.insertRecord(ContinuationBuilder); 2243 } 2244 2245 std::string FullName = getFullyQualifiedName(Ty); 2246 2247 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2248 getTypeIndex(Ty->getBaseType())); 2249 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2250 2251 addUDTSrcLine(Ty, EnumTI); 2252 2253 return EnumTI; 2254 } 2255 2256 //===----------------------------------------------------------------------===// 2257 // ClassInfo 2258 //===----------------------------------------------------------------------===// 2259 2260 struct llvm::ClassInfo { 2261 struct MemberInfo { 2262 const DIDerivedType *MemberTypeNode; 2263 uint64_t BaseOffset; 2264 }; 2265 // [MemberInfo] 2266 using MemberList = std::vector<MemberInfo>; 2267 2268 using MethodsList = TinyPtrVector<const DISubprogram *>; 2269 // MethodName -> MethodsList 2270 using MethodsMap = MapVector<MDString *, MethodsList>; 2271 2272 /// Base classes. 2273 std::vector<const DIDerivedType *> Inheritance; 2274 2275 /// Direct members. 2276 MemberList Members; 2277 // Direct overloaded methods gathered by name. 2278 MethodsMap Methods; 2279 2280 TypeIndex VShapeTI; 2281 2282 std::vector<const DIType *> NestedTypes; 2283 }; 2284 2285 void CodeViewDebug::clear() { 2286 assert(CurFn == nullptr); 2287 FileIdMap.clear(); 2288 FnDebugInfo.clear(); 2289 FileToFilepathMap.clear(); 2290 LocalUDTs.clear(); 2291 GlobalUDTs.clear(); 2292 TypeIndices.clear(); 2293 CompleteTypeIndices.clear(); 2294 ScopeGlobals.clear(); 2295 CVGlobalVariableOffsets.clear(); 2296 } 2297 2298 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2299 const DIDerivedType *DDTy) { 2300 if (!DDTy->getName().empty()) { 2301 Info.Members.push_back({DDTy, 0}); 2302 2303 // Collect static const data members with values. 2304 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2305 DINode::FlagStaticMember) { 2306 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2307 isa<ConstantFP>(DDTy->getConstant()))) 2308 StaticConstMembers.push_back(DDTy); 2309 } 2310 2311 return; 2312 } 2313 2314 // An unnamed member may represent a nested struct or union. Attempt to 2315 // interpret the unnamed member as a DICompositeType possibly wrapped in 2316 // qualifier types. Add all the indirect fields to the current record if that 2317 // succeeds, and drop the member if that fails. 2318 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2319 uint64_t Offset = DDTy->getOffsetInBits(); 2320 const DIType *Ty = DDTy->getBaseType(); 2321 bool FullyResolved = false; 2322 while (!FullyResolved) { 2323 switch (Ty->getTag()) { 2324 case dwarf::DW_TAG_const_type: 2325 case dwarf::DW_TAG_volatile_type: 2326 // FIXME: we should apply the qualifier types to the indirect fields 2327 // rather than dropping them. 2328 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2329 break; 2330 default: 2331 FullyResolved = true; 2332 break; 2333 } 2334 } 2335 2336 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2337 if (!DCTy) 2338 return; 2339 2340 ClassInfo NestedInfo = collectClassInfo(DCTy); 2341 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2342 Info.Members.push_back( 2343 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2344 } 2345 2346 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2347 ClassInfo Info; 2348 // Add elements to structure type. 2349 DINodeArray Elements = Ty->getElements(); 2350 for (auto *Element : Elements) { 2351 // We assume that the frontend provides all members in source declaration 2352 // order, which is what MSVC does. 2353 if (!Element) 2354 continue; 2355 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2356 Info.Methods[SP->getRawName()].push_back(SP); 2357 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2358 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2359 collectMemberInfo(Info, DDTy); 2360 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2361 Info.Inheritance.push_back(DDTy); 2362 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2363 DDTy->getName() == "__vtbl_ptr_type") { 2364 Info.VShapeTI = getTypeIndex(DDTy); 2365 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2366 Info.NestedTypes.push_back(DDTy); 2367 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2368 // Ignore friend members. It appears that MSVC emitted info about 2369 // friends in the past, but modern versions do not. 2370 } 2371 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2372 Info.NestedTypes.push_back(Composite); 2373 } 2374 // Skip other unrecognized kinds of elements. 2375 } 2376 return Info; 2377 } 2378 2379 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2380 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2381 // if a complete type should be emitted instead of a forward reference. 2382 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2383 !Ty->isForwardDecl(); 2384 } 2385 2386 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2387 // Emit the complete type for unnamed structs. C++ classes with methods 2388 // which have a circular reference back to the class type are expected to 2389 // be named by the front-end and should not be "unnamed". C unnamed 2390 // structs should not have circular references. 2391 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2392 // If this unnamed complete type is already in the process of being defined 2393 // then the description of the type is malformed and cannot be emitted 2394 // into CodeView correctly so report a fatal error. 2395 auto I = CompleteTypeIndices.find(Ty); 2396 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2397 report_fatal_error("cannot debug circular reference to unnamed type"); 2398 return getCompleteTypeIndex(Ty); 2399 } 2400 2401 // First, construct the forward decl. Don't look into Ty to compute the 2402 // forward decl options, since it might not be available in all TUs. 2403 TypeRecordKind Kind = getRecordKind(Ty); 2404 ClassOptions CO = 2405 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2406 std::string FullName = getFullyQualifiedName(Ty); 2407 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2408 FullName, Ty->getIdentifier()); 2409 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2410 if (!Ty->isForwardDecl()) 2411 DeferredCompleteTypes.push_back(Ty); 2412 return FwdDeclTI; 2413 } 2414 2415 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2416 // Construct the field list and complete type record. 2417 TypeRecordKind Kind = getRecordKind(Ty); 2418 ClassOptions CO = getCommonClassOptions(Ty); 2419 TypeIndex FieldTI; 2420 TypeIndex VShapeTI; 2421 unsigned FieldCount; 2422 bool ContainsNestedClass; 2423 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2424 lowerRecordFieldList(Ty); 2425 2426 if (ContainsNestedClass) 2427 CO |= ClassOptions::ContainsNestedClass; 2428 2429 // MSVC appears to set this flag by searching any destructor or method with 2430 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2431 // the members, however special member functions are not yet emitted into 2432 // debug information. For now checking a class's non-triviality seems enough. 2433 // FIXME: not true for a nested unnamed struct. 2434 if (isNonTrivial(Ty)) 2435 CO |= ClassOptions::HasConstructorOrDestructor; 2436 2437 std::string FullName = getFullyQualifiedName(Ty); 2438 2439 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2440 2441 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2442 SizeInBytes, FullName, Ty->getIdentifier()); 2443 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2444 2445 addUDTSrcLine(Ty, ClassTI); 2446 2447 addToUDTs(Ty); 2448 2449 return ClassTI; 2450 } 2451 2452 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2453 // Emit the complete type for unnamed unions. 2454 if (shouldAlwaysEmitCompleteClassType(Ty)) 2455 return getCompleteTypeIndex(Ty); 2456 2457 ClassOptions CO = 2458 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2459 std::string FullName = getFullyQualifiedName(Ty); 2460 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2461 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2462 if (!Ty->isForwardDecl()) 2463 DeferredCompleteTypes.push_back(Ty); 2464 return FwdDeclTI; 2465 } 2466 2467 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2468 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2469 TypeIndex FieldTI; 2470 unsigned FieldCount; 2471 bool ContainsNestedClass; 2472 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2473 lowerRecordFieldList(Ty); 2474 2475 if (ContainsNestedClass) 2476 CO |= ClassOptions::ContainsNestedClass; 2477 2478 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2479 std::string FullName = getFullyQualifiedName(Ty); 2480 2481 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2482 Ty->getIdentifier()); 2483 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2484 2485 addUDTSrcLine(Ty, UnionTI); 2486 2487 addToUDTs(Ty); 2488 2489 return UnionTI; 2490 } 2491 2492 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2493 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2494 // Manually count members. MSVC appears to count everything that generates a 2495 // field list record. Each individual overload in a method overload group 2496 // contributes to this count, even though the overload group is a single field 2497 // list record. 2498 unsigned MemberCount = 0; 2499 ClassInfo Info = collectClassInfo(Ty); 2500 ContinuationRecordBuilder ContinuationBuilder; 2501 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2502 2503 // Create base classes. 2504 for (const DIDerivedType *I : Info.Inheritance) { 2505 if (I->getFlags() & DINode::FlagVirtual) { 2506 // Virtual base. 2507 unsigned VBPtrOffset = I->getVBPtrOffset(); 2508 // FIXME: Despite the accessor name, the offset is really in bytes. 2509 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2510 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2511 ? TypeRecordKind::IndirectVirtualBaseClass 2512 : TypeRecordKind::VirtualBaseClass; 2513 VirtualBaseClassRecord VBCR( 2514 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2515 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2516 VBTableIndex); 2517 2518 ContinuationBuilder.writeMemberType(VBCR); 2519 MemberCount++; 2520 } else { 2521 assert(I->getOffsetInBits() % 8 == 0 && 2522 "bases must be on byte boundaries"); 2523 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2524 getTypeIndex(I->getBaseType()), 2525 I->getOffsetInBits() / 8); 2526 ContinuationBuilder.writeMemberType(BCR); 2527 MemberCount++; 2528 } 2529 } 2530 2531 // Create members. 2532 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2533 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2534 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2535 StringRef MemberName = Member->getName(); 2536 MemberAccess Access = 2537 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2538 2539 if (Member->isStaticMember()) { 2540 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2541 ContinuationBuilder.writeMemberType(SDMR); 2542 MemberCount++; 2543 continue; 2544 } 2545 2546 // Virtual function pointer member. 2547 if ((Member->getFlags() & DINode::FlagArtificial) && 2548 Member->getName().startswith("_vptr$")) { 2549 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2550 ContinuationBuilder.writeMemberType(VFPR); 2551 MemberCount++; 2552 continue; 2553 } 2554 2555 // Data member. 2556 uint64_t MemberOffsetInBits = 2557 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2558 if (Member->isBitField()) { 2559 uint64_t StartBitOffset = MemberOffsetInBits; 2560 if (const auto *CI = 2561 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2562 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2563 } 2564 StartBitOffset -= MemberOffsetInBits; 2565 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2566 StartBitOffset); 2567 MemberBaseType = TypeTable.writeLeafType(BFR); 2568 } 2569 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2570 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2571 MemberName); 2572 ContinuationBuilder.writeMemberType(DMR); 2573 MemberCount++; 2574 } 2575 2576 // Create methods 2577 for (auto &MethodItr : Info.Methods) { 2578 StringRef Name = MethodItr.first->getString(); 2579 2580 std::vector<OneMethodRecord> Methods; 2581 for (const DISubprogram *SP : MethodItr.second) { 2582 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2583 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2584 2585 unsigned VFTableOffset = -1; 2586 if (Introduced) 2587 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2588 2589 Methods.push_back(OneMethodRecord( 2590 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2591 translateMethodKindFlags(SP, Introduced), 2592 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2593 MemberCount++; 2594 } 2595 assert(!Methods.empty() && "Empty methods map entry"); 2596 if (Methods.size() == 1) 2597 ContinuationBuilder.writeMemberType(Methods[0]); 2598 else { 2599 // FIXME: Make this use its own ContinuationBuilder so that 2600 // MethodOverloadList can be split correctly. 2601 MethodOverloadListRecord MOLR(Methods); 2602 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2603 2604 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2605 ContinuationBuilder.writeMemberType(OMR); 2606 } 2607 } 2608 2609 // Create nested classes. 2610 for (const DIType *Nested : Info.NestedTypes) { 2611 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2612 ContinuationBuilder.writeMemberType(R); 2613 MemberCount++; 2614 } 2615 2616 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2617 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2618 !Info.NestedTypes.empty()); 2619 } 2620 2621 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2622 if (!VBPType.getIndex()) { 2623 // Make a 'const int *' type. 2624 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2625 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2626 2627 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2628 : PointerKind::Near32; 2629 PointerMode PM = PointerMode::Pointer; 2630 PointerOptions PO = PointerOptions::None; 2631 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2632 VBPType = TypeTable.writeLeafType(PR); 2633 } 2634 2635 return VBPType; 2636 } 2637 2638 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2639 // The null DIType is the void type. Don't try to hash it. 2640 if (!Ty) 2641 return TypeIndex::Void(); 2642 2643 // Check if we've already translated this type. Don't try to do a 2644 // get-or-create style insertion that caches the hash lookup across the 2645 // lowerType call. It will update the TypeIndices map. 2646 auto I = TypeIndices.find({Ty, ClassTy}); 2647 if (I != TypeIndices.end()) 2648 return I->second; 2649 2650 TypeLoweringScope S(*this); 2651 TypeIndex TI = lowerType(Ty, ClassTy); 2652 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2653 } 2654 2655 codeview::TypeIndex 2656 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2657 const DISubroutineType *SubroutineTy) { 2658 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2659 "this type must be a pointer type"); 2660 2661 PointerOptions Options = PointerOptions::None; 2662 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2663 Options = PointerOptions::LValueRefThisPointer; 2664 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2665 Options = PointerOptions::RValueRefThisPointer; 2666 2667 // Check if we've already translated this type. If there is no ref qualifier 2668 // on the function then we look up this pointer type with no associated class 2669 // so that the TypeIndex for the this pointer can be shared with the type 2670 // index for other pointers to this class type. If there is a ref qualifier 2671 // then we lookup the pointer using the subroutine as the parent type. 2672 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2673 if (I != TypeIndices.end()) 2674 return I->second; 2675 2676 TypeLoweringScope S(*this); 2677 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2678 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2679 } 2680 2681 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2682 PointerRecord PR(getTypeIndex(Ty), 2683 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2684 : PointerKind::Near32, 2685 PointerMode::LValueReference, PointerOptions::None, 2686 Ty->getSizeInBits() / 8); 2687 return TypeTable.writeLeafType(PR); 2688 } 2689 2690 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2691 // The null DIType is the void type. Don't try to hash it. 2692 if (!Ty) 2693 return TypeIndex::Void(); 2694 2695 // Look through typedefs when getting the complete type index. Call 2696 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2697 // emitted only once. 2698 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2699 (void)getTypeIndex(Ty); 2700 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2701 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2702 2703 // If this is a non-record type, the complete type index is the same as the 2704 // normal type index. Just call getTypeIndex. 2705 switch (Ty->getTag()) { 2706 case dwarf::DW_TAG_class_type: 2707 case dwarf::DW_TAG_structure_type: 2708 case dwarf::DW_TAG_union_type: 2709 break; 2710 default: 2711 return getTypeIndex(Ty); 2712 } 2713 2714 const auto *CTy = cast<DICompositeType>(Ty); 2715 2716 TypeLoweringScope S(*this); 2717 2718 // Make sure the forward declaration is emitted first. It's unclear if this 2719 // is necessary, but MSVC does it, and we should follow suit until we can show 2720 // otherwise. 2721 // We only emit a forward declaration for named types. 2722 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2723 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2724 2725 // Just use the forward decl if we don't have complete type info. This 2726 // might happen if the frontend is using modules and expects the complete 2727 // definition to be emitted elsewhere. 2728 if (CTy->isForwardDecl()) 2729 return FwdDeclTI; 2730 } 2731 2732 // Check if we've already translated the complete record type. 2733 // Insert the type with a null TypeIndex to signify that the type is currently 2734 // being lowered. 2735 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2736 if (!InsertResult.second) 2737 return InsertResult.first->second; 2738 2739 TypeIndex TI; 2740 switch (CTy->getTag()) { 2741 case dwarf::DW_TAG_class_type: 2742 case dwarf::DW_TAG_structure_type: 2743 TI = lowerCompleteTypeClass(CTy); 2744 break; 2745 case dwarf::DW_TAG_union_type: 2746 TI = lowerCompleteTypeUnion(CTy); 2747 break; 2748 default: 2749 llvm_unreachable("not a record"); 2750 } 2751 2752 // Update the type index associated with this CompositeType. This cannot 2753 // use the 'InsertResult' iterator above because it is potentially 2754 // invalidated by map insertions which can occur while lowering the class 2755 // type above. 2756 CompleteTypeIndices[CTy] = TI; 2757 return TI; 2758 } 2759 2760 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2761 /// and do this until fixpoint, as each complete record type typically 2762 /// references 2763 /// many other record types. 2764 void CodeViewDebug::emitDeferredCompleteTypes() { 2765 SmallVector<const DICompositeType *, 4> TypesToEmit; 2766 while (!DeferredCompleteTypes.empty()) { 2767 std::swap(DeferredCompleteTypes, TypesToEmit); 2768 for (const DICompositeType *RecordTy : TypesToEmit) 2769 getCompleteTypeIndex(RecordTy); 2770 TypesToEmit.clear(); 2771 } 2772 } 2773 2774 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2775 ArrayRef<LocalVariable> Locals) { 2776 // Get the sorted list of parameters and emit them first. 2777 SmallVector<const LocalVariable *, 6> Params; 2778 for (const LocalVariable &L : Locals) 2779 if (L.DIVar->isParameter()) 2780 Params.push_back(&L); 2781 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2782 return L->DIVar->getArg() < R->DIVar->getArg(); 2783 }); 2784 for (const LocalVariable *L : Params) 2785 emitLocalVariable(FI, *L); 2786 2787 // Next emit all non-parameters in the order that we found them. 2788 for (const LocalVariable &L : Locals) 2789 if (!L.DIVar->isParameter()) 2790 emitLocalVariable(FI, L); 2791 } 2792 2793 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2794 const LocalVariable &Var) { 2795 // LocalSym record, see SymbolRecord.h for more info. 2796 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2797 2798 LocalSymFlags Flags = LocalSymFlags::None; 2799 if (Var.DIVar->isParameter()) 2800 Flags |= LocalSymFlags::IsParameter; 2801 if (Var.DefRanges.empty()) 2802 Flags |= LocalSymFlags::IsOptimizedOut; 2803 2804 OS.AddComment("TypeIndex"); 2805 TypeIndex TI = Var.UseReferenceType 2806 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2807 : getCompleteTypeIndex(Var.DIVar->getType()); 2808 OS.emitInt32(TI.getIndex()); 2809 OS.AddComment("Flags"); 2810 OS.emitInt16(static_cast<uint16_t>(Flags)); 2811 // Truncate the name so we won't overflow the record length field. 2812 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2813 endSymbolRecord(LocalEnd); 2814 2815 // Calculate the on disk prefix of the appropriate def range record. The 2816 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2817 // should be big enough to hold all forms without memory allocation. 2818 SmallString<20> BytePrefix; 2819 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2820 BytePrefix.clear(); 2821 if (DefRange.InMemory) { 2822 int Offset = DefRange.DataOffset; 2823 unsigned Reg = DefRange.CVRegister; 2824 2825 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2826 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2827 // instead. In frames without stack realignment, $T0 will be the CFA. 2828 if (RegisterId(Reg) == RegisterId::ESP) { 2829 Reg = unsigned(RegisterId::VFRAME); 2830 Offset += FI.OffsetAdjustment; 2831 } 2832 2833 // If we can use the chosen frame pointer for the frame and this isn't a 2834 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2835 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2836 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2837 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2838 (bool(Flags & LocalSymFlags::IsParameter) 2839 ? (EncFP == FI.EncodedParamFramePtrReg) 2840 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2841 DefRangeFramePointerRelHeader DRHdr; 2842 DRHdr.Offset = Offset; 2843 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2844 } else { 2845 uint16_t RegRelFlags = 0; 2846 if (DefRange.IsSubfield) { 2847 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2848 (DefRange.StructOffset 2849 << DefRangeRegisterRelSym::OffsetInParentShift); 2850 } 2851 DefRangeRegisterRelHeader DRHdr; 2852 DRHdr.Register = Reg; 2853 DRHdr.Flags = RegRelFlags; 2854 DRHdr.BasePointerOffset = Offset; 2855 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2856 } 2857 } else { 2858 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2859 if (DefRange.IsSubfield) { 2860 DefRangeSubfieldRegisterHeader DRHdr; 2861 DRHdr.Register = DefRange.CVRegister; 2862 DRHdr.MayHaveNoName = 0; 2863 DRHdr.OffsetInParent = DefRange.StructOffset; 2864 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2865 } else { 2866 DefRangeRegisterHeader DRHdr; 2867 DRHdr.Register = DefRange.CVRegister; 2868 DRHdr.MayHaveNoName = 0; 2869 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2870 } 2871 } 2872 } 2873 } 2874 2875 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2876 const FunctionInfo& FI) { 2877 for (LexicalBlock *Block : Blocks) 2878 emitLexicalBlock(*Block, FI); 2879 } 2880 2881 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2882 /// lexical block scope. 2883 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2884 const FunctionInfo& FI) { 2885 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2886 OS.AddComment("PtrParent"); 2887 OS.emitInt32(0); // PtrParent 2888 OS.AddComment("PtrEnd"); 2889 OS.emitInt32(0); // PtrEnd 2890 OS.AddComment("Code size"); 2891 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2892 OS.AddComment("Function section relative address"); 2893 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2894 OS.AddComment("Function section index"); 2895 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2896 OS.AddComment("Lexical block name"); 2897 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2898 endSymbolRecord(RecordEnd); 2899 2900 // Emit variables local to this lexical block. 2901 emitLocalVariableList(FI, Block.Locals); 2902 emitGlobalVariableList(Block.Globals); 2903 2904 // Emit lexical blocks contained within this block. 2905 emitLexicalBlockList(Block.Children, FI); 2906 2907 // Close the lexical block scope. 2908 emitEndSymbolRecord(SymbolKind::S_END); 2909 } 2910 2911 /// Convenience routine for collecting lexical block information for a list 2912 /// of lexical scopes. 2913 void CodeViewDebug::collectLexicalBlockInfo( 2914 SmallVectorImpl<LexicalScope *> &Scopes, 2915 SmallVectorImpl<LexicalBlock *> &Blocks, 2916 SmallVectorImpl<LocalVariable> &Locals, 2917 SmallVectorImpl<CVGlobalVariable> &Globals) { 2918 for (LexicalScope *Scope : Scopes) 2919 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2920 } 2921 2922 /// Populate the lexical blocks and local variable lists of the parent with 2923 /// information about the specified lexical scope. 2924 void CodeViewDebug::collectLexicalBlockInfo( 2925 LexicalScope &Scope, 2926 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2927 SmallVectorImpl<LocalVariable> &ParentLocals, 2928 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2929 if (Scope.isAbstractScope()) 2930 return; 2931 2932 // Gather information about the lexical scope including local variables, 2933 // global variables, and address ranges. 2934 bool IgnoreScope = false; 2935 auto LI = ScopeVariables.find(&Scope); 2936 SmallVectorImpl<LocalVariable> *Locals = 2937 LI != ScopeVariables.end() ? &LI->second : nullptr; 2938 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2939 SmallVectorImpl<CVGlobalVariable> *Globals = 2940 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2941 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2942 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2943 2944 // Ignore lexical scopes which do not contain variables. 2945 if (!Locals && !Globals) 2946 IgnoreScope = true; 2947 2948 // Ignore lexical scopes which are not lexical blocks. 2949 if (!DILB) 2950 IgnoreScope = true; 2951 2952 // Ignore scopes which have too many address ranges to represent in the 2953 // current CodeView format or do not have a valid address range. 2954 // 2955 // For lexical scopes with multiple address ranges you may be tempted to 2956 // construct a single range covering every instruction where the block is 2957 // live and everything in between. Unfortunately, Visual Studio only 2958 // displays variables from the first matching lexical block scope. If the 2959 // first lexical block contains exception handling code or cold code which 2960 // is moved to the bottom of the routine creating a single range covering 2961 // nearly the entire routine, then it will hide all other lexical blocks 2962 // and the variables they contain. 2963 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2964 IgnoreScope = true; 2965 2966 if (IgnoreScope) { 2967 // This scope can be safely ignored and eliminating it will reduce the 2968 // size of the debug information. Be sure to collect any variable and scope 2969 // information from the this scope or any of its children and collapse them 2970 // into the parent scope. 2971 if (Locals) 2972 ParentLocals.append(Locals->begin(), Locals->end()); 2973 if (Globals) 2974 ParentGlobals.append(Globals->begin(), Globals->end()); 2975 collectLexicalBlockInfo(Scope.getChildren(), 2976 ParentBlocks, 2977 ParentLocals, 2978 ParentGlobals); 2979 return; 2980 } 2981 2982 // Create a new CodeView lexical block for this lexical scope. If we've 2983 // seen this DILexicalBlock before then the scope tree is malformed and 2984 // we can handle this gracefully by not processing it a second time. 2985 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2986 if (!BlockInsertion.second) 2987 return; 2988 2989 // Create a lexical block containing the variables and collect the the 2990 // lexical block information for the children. 2991 const InsnRange &Range = Ranges.front(); 2992 assert(Range.first && Range.second); 2993 LexicalBlock &Block = BlockInsertion.first->second; 2994 Block.Begin = getLabelBeforeInsn(Range.first); 2995 Block.End = getLabelAfterInsn(Range.second); 2996 assert(Block.Begin && "missing label for scope begin"); 2997 assert(Block.End && "missing label for scope end"); 2998 Block.Name = DILB->getName(); 2999 if (Locals) 3000 Block.Locals = std::move(*Locals); 3001 if (Globals) 3002 Block.Globals = std::move(*Globals); 3003 ParentBlocks.push_back(&Block); 3004 collectLexicalBlockInfo(Scope.getChildren(), 3005 Block.Children, 3006 Block.Locals, 3007 Block.Globals); 3008 } 3009 3010 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3011 const Function &GV = MF->getFunction(); 3012 assert(FnDebugInfo.count(&GV)); 3013 assert(CurFn == FnDebugInfo[&GV].get()); 3014 3015 collectVariableInfo(GV.getSubprogram()); 3016 3017 // Build the lexical block structure to emit for this routine. 3018 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3019 collectLexicalBlockInfo(*CFS, 3020 CurFn->ChildBlocks, 3021 CurFn->Locals, 3022 CurFn->Globals); 3023 3024 // Clear the scope and variable information from the map which will not be 3025 // valid after we have finished processing this routine. This also prepares 3026 // the map for the subsequent routine. 3027 ScopeVariables.clear(); 3028 3029 // Don't emit anything if we don't have any line tables. 3030 // Thunks are compiler-generated and probably won't have source correlation. 3031 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3032 FnDebugInfo.erase(&GV); 3033 CurFn = nullptr; 3034 return; 3035 } 3036 3037 // Find heap alloc sites and add to list. 3038 for (const auto &MBB : *MF) { 3039 for (const auto &MI : MBB) { 3040 if (MDNode *MD = MI.getHeapAllocMarker()) { 3041 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3042 getLabelAfterInsn(&MI), 3043 dyn_cast<DIType>(MD))); 3044 } 3045 } 3046 } 3047 3048 CurFn->Annotations = MF->getCodeViewAnnotations(); 3049 3050 CurFn->End = Asm->getFunctionEnd(); 3051 3052 CurFn = nullptr; 3053 } 3054 3055 // Usable locations are valid with non-zero line numbers. A line number of zero 3056 // corresponds to optimized code that doesn't have a distinct source location. 3057 // In this case, we try to use the previous or next source location depending on 3058 // the context. 3059 static bool isUsableDebugLoc(DebugLoc DL) { 3060 return DL && DL.getLine() != 0; 3061 } 3062 3063 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3064 DebugHandlerBase::beginInstruction(MI); 3065 3066 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3067 if (!Asm || !CurFn || MI->isDebugInstr() || 3068 MI->getFlag(MachineInstr::FrameSetup)) 3069 return; 3070 3071 // If the first instruction of a new MBB has no location, find the first 3072 // instruction with a location and use that. 3073 DebugLoc DL = MI->getDebugLoc(); 3074 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3075 for (const auto &NextMI : *MI->getParent()) { 3076 if (NextMI.isDebugInstr()) 3077 continue; 3078 DL = NextMI.getDebugLoc(); 3079 if (isUsableDebugLoc(DL)) 3080 break; 3081 } 3082 // FIXME: Handle the case where the BB has no valid locations. This would 3083 // probably require doing a real dataflow analysis. 3084 } 3085 PrevInstBB = MI->getParent(); 3086 3087 // If we still don't have a debug location, don't record a location. 3088 if (!isUsableDebugLoc(DL)) 3089 return; 3090 3091 maybeRecordLocation(DL, Asm->MF); 3092 } 3093 3094 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3095 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3096 *EndLabel = MMI->getContext().createTempSymbol(); 3097 OS.emitInt32(unsigned(Kind)); 3098 OS.AddComment("Subsection size"); 3099 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3100 OS.emitLabel(BeginLabel); 3101 return EndLabel; 3102 } 3103 3104 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3105 OS.emitLabel(EndLabel); 3106 // Every subsection must be aligned to a 4-byte boundary. 3107 OS.emitValueToAlignment(4); 3108 } 3109 3110 static StringRef getSymbolName(SymbolKind SymKind) { 3111 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3112 if (EE.Value == SymKind) 3113 return EE.Name; 3114 return ""; 3115 } 3116 3117 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3118 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3119 *EndLabel = MMI->getContext().createTempSymbol(); 3120 OS.AddComment("Record length"); 3121 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3122 OS.emitLabel(BeginLabel); 3123 if (OS.isVerboseAsm()) 3124 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3125 OS.emitInt16(unsigned(SymKind)); 3126 return EndLabel; 3127 } 3128 3129 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3130 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3131 // an extra copy of every symbol record in LLD. This increases object file 3132 // size by less than 1% in the clang build, and is compatible with the Visual 3133 // C++ linker. 3134 OS.emitValueToAlignment(4); 3135 OS.emitLabel(SymEnd); 3136 } 3137 3138 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3139 OS.AddComment("Record length"); 3140 OS.emitInt16(2); 3141 if (OS.isVerboseAsm()) 3142 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3143 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3144 } 3145 3146 void CodeViewDebug::emitDebugInfoForUDTs( 3147 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3148 #ifndef NDEBUG 3149 size_t OriginalSize = UDTs.size(); 3150 #endif 3151 for (const auto &UDT : UDTs) { 3152 const DIType *T = UDT.second; 3153 assert(shouldEmitUdt(T)); 3154 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3155 OS.AddComment("Type"); 3156 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3157 assert(OriginalSize == UDTs.size() && 3158 "getCompleteTypeIndex found new UDTs!"); 3159 emitNullTerminatedSymbolName(OS, UDT.first); 3160 endSymbolRecord(UDTRecordEnd); 3161 } 3162 } 3163 3164 void CodeViewDebug::collectGlobalVariableInfo() { 3165 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3166 GlobalMap; 3167 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3168 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3169 GV.getDebugInfo(GVEs); 3170 for (const auto *GVE : GVEs) 3171 GlobalMap[GVE] = &GV; 3172 } 3173 3174 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3175 for (const MDNode *Node : CUs->operands()) { 3176 const auto *CU = cast<DICompileUnit>(Node); 3177 for (const auto *GVE : CU->getGlobalVariables()) { 3178 const DIGlobalVariable *DIGV = GVE->getVariable(); 3179 const DIExpression *DIE = GVE->getExpression(); 3180 3181 if ((DIE->getNumElements() == 2) && 3182 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3183 // Record the constant offset for the variable. 3184 // 3185 // A Fortran common block uses this idiom to encode the offset 3186 // of a variable from the common block's starting address. 3187 CVGlobalVariableOffsets.insert( 3188 std::make_pair(DIGV, DIE->getElement(1))); 3189 3190 // Emit constant global variables in a global symbol section. 3191 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3192 CVGlobalVariable CVGV = {DIGV, DIE}; 3193 GlobalVariables.emplace_back(std::move(CVGV)); 3194 } 3195 3196 const auto *GV = GlobalMap.lookup(GVE); 3197 if (!GV || GV->isDeclarationForLinker()) 3198 continue; 3199 3200 DIScope *Scope = DIGV->getScope(); 3201 SmallVector<CVGlobalVariable, 1> *VariableList; 3202 if (Scope && isa<DILocalScope>(Scope)) { 3203 // Locate a global variable list for this scope, creating one if 3204 // necessary. 3205 auto Insertion = ScopeGlobals.insert( 3206 {Scope, std::unique_ptr<GlobalVariableList>()}); 3207 if (Insertion.second) 3208 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3209 VariableList = Insertion.first->second.get(); 3210 } else if (GV->hasComdat()) 3211 // Emit this global variable into a COMDAT section. 3212 VariableList = &ComdatVariables; 3213 else 3214 // Emit this global variable in a single global symbol section. 3215 VariableList = &GlobalVariables; 3216 CVGlobalVariable CVGV = {DIGV, GV}; 3217 VariableList->emplace_back(std::move(CVGV)); 3218 } 3219 } 3220 } 3221 3222 void CodeViewDebug::collectDebugInfoForGlobals() { 3223 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3224 const DIGlobalVariable *DIGV = CVGV.DIGV; 3225 const DIScope *Scope = DIGV->getScope(); 3226 getCompleteTypeIndex(DIGV->getType()); 3227 getFullyQualifiedName(Scope, DIGV->getName()); 3228 } 3229 3230 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3231 const DIGlobalVariable *DIGV = CVGV.DIGV; 3232 const DIScope *Scope = DIGV->getScope(); 3233 getCompleteTypeIndex(DIGV->getType()); 3234 getFullyQualifiedName(Scope, DIGV->getName()); 3235 } 3236 } 3237 3238 void CodeViewDebug::emitDebugInfoForGlobals() { 3239 // First, emit all globals that are not in a comdat in a single symbol 3240 // substream. MSVC doesn't like it if the substream is empty, so only open 3241 // it if we have at least one global to emit. 3242 switchToDebugSectionForSymbol(nullptr); 3243 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3244 OS.AddComment("Symbol subsection for globals"); 3245 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3246 emitGlobalVariableList(GlobalVariables); 3247 emitStaticConstMemberList(); 3248 endCVSubsection(EndLabel); 3249 } 3250 3251 // Second, emit each global that is in a comdat into its own .debug$S 3252 // section along with its own symbol substream. 3253 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3254 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3255 MCSymbol *GVSym = Asm->getSymbol(GV); 3256 OS.AddComment("Symbol subsection for " + 3257 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3258 switchToDebugSectionForSymbol(GVSym); 3259 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3260 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3261 emitDebugInfoForGlobal(CVGV); 3262 endCVSubsection(EndLabel); 3263 } 3264 } 3265 3266 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3267 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3268 for (const MDNode *Node : CUs->operands()) { 3269 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3270 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3271 getTypeIndex(RT); 3272 // FIXME: Add to global/local DTU list. 3273 } 3274 } 3275 } 3276 } 3277 3278 // Emit each global variable in the specified array. 3279 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3280 for (const CVGlobalVariable &CVGV : Globals) { 3281 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3282 emitDebugInfoForGlobal(CVGV); 3283 } 3284 } 3285 3286 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3287 const std::string &QualifiedName) { 3288 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3289 OS.AddComment("Type"); 3290 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3291 3292 OS.AddComment("Value"); 3293 3294 // Encoded integers shouldn't need more than 10 bytes. 3295 uint8_t Data[10]; 3296 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3297 CodeViewRecordIO IO(Writer); 3298 cantFail(IO.mapEncodedInteger(Value)); 3299 StringRef SRef((char *)Data, Writer.getOffset()); 3300 OS.emitBinaryData(SRef); 3301 3302 OS.AddComment("Name"); 3303 emitNullTerminatedSymbolName(OS, QualifiedName); 3304 endSymbolRecord(SConstantEnd); 3305 } 3306 3307 void CodeViewDebug::emitStaticConstMemberList() { 3308 for (const DIDerivedType *DTy : StaticConstMembers) { 3309 const DIScope *Scope = DTy->getScope(); 3310 3311 APSInt Value; 3312 if (const ConstantInt *CI = 3313 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3314 Value = APSInt(CI->getValue(), 3315 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3316 else if (const ConstantFP *CFP = 3317 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3318 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3319 else 3320 llvm_unreachable("cannot emit a constant without a value"); 3321 3322 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3323 getFullyQualifiedName(Scope, DTy->getName())); 3324 } 3325 } 3326 3327 static bool isFloatDIType(const DIType *Ty) { 3328 if (isa<DICompositeType>(Ty)) 3329 return false; 3330 3331 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3332 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3333 if (T == dwarf::DW_TAG_pointer_type || 3334 T == dwarf::DW_TAG_ptr_to_member_type || 3335 T == dwarf::DW_TAG_reference_type || 3336 T == dwarf::DW_TAG_rvalue_reference_type) 3337 return false; 3338 assert(DTy->getBaseType() && "Expected valid base type"); 3339 return isFloatDIType(DTy->getBaseType()); 3340 } 3341 3342 auto *BTy = cast<DIBasicType>(Ty); 3343 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3344 } 3345 3346 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3347 const DIGlobalVariable *DIGV = CVGV.DIGV; 3348 3349 const DIScope *Scope = DIGV->getScope(); 3350 // For static data members, get the scope from the declaration. 3351 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3352 DIGV->getRawStaticDataMemberDeclaration())) 3353 Scope = MemberDecl->getScope(); 3354 // For Fortran, the scoping portion is elided in its name so that we can 3355 // reference the variable in the command line of the VS debugger. 3356 std::string QualifiedName = 3357 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3358 : getFullyQualifiedName(Scope, DIGV->getName()); 3359 3360 if (const GlobalVariable *GV = 3361 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3362 // DataSym record, see SymbolRecord.h for more info. Thread local data 3363 // happens to have the same format as global data. 3364 MCSymbol *GVSym = Asm->getSymbol(GV); 3365 SymbolKind DataSym = GV->isThreadLocal() 3366 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3367 : SymbolKind::S_GTHREAD32) 3368 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3369 : SymbolKind::S_GDATA32); 3370 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3371 OS.AddComment("Type"); 3372 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3373 OS.AddComment("DataOffset"); 3374 3375 uint64_t Offset = 0; 3376 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3377 // Use the offset seen while collecting info on globals. 3378 Offset = CVGlobalVariableOffsets[DIGV]; 3379 OS.EmitCOFFSecRel32(GVSym, Offset); 3380 3381 OS.AddComment("Segment"); 3382 OS.EmitCOFFSectionIndex(GVSym); 3383 OS.AddComment("Name"); 3384 const unsigned LengthOfDataRecord = 12; 3385 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3386 endSymbolRecord(DataEnd); 3387 } else { 3388 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3389 assert(DIE->isConstant() && 3390 "Global constant variables must contain a constant expression."); 3391 3392 // Use unsigned for floats. 3393 bool isUnsigned = isFloatDIType(DIGV->getType()) 3394 ? true 3395 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3396 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3397 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3398 } 3399 } 3400