1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/None.h" 17 #include "llvm/ADT/Optional.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallString.h" 20 #include "llvm/ADT/StringRef.h" 21 #include "llvm/ADT/TinyPtrVector.h" 22 #include "llvm/ADT/Triple.h" 23 #include "llvm/ADT/Twine.h" 24 #include "llvm/BinaryFormat/COFF.h" 25 #include "llvm/BinaryFormat/Dwarf.h" 26 #include "llvm/CodeGen/AsmPrinter.h" 27 #include "llvm/CodeGen/LexicalScopes.h" 28 #include "llvm/CodeGen/MachineFrameInfo.h" 29 #include "llvm/CodeGen/MachineFunction.h" 30 #include "llvm/CodeGen/MachineInstr.h" 31 #include "llvm/CodeGen/MachineModuleInfo.h" 32 #include "llvm/CodeGen/MachineOperand.h" 33 #include "llvm/CodeGen/TargetFrameLowering.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/TargetSubtargetInfo.h" 36 #include "llvm/Config/llvm-config.h" 37 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 38 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 39 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 40 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 41 #include "llvm/DebugInfo/CodeView/EnumTables.h" 42 #include "llvm/DebugInfo/CodeView/Line.h" 43 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 44 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 45 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 46 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 47 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 48 #include "llvm/IR/Constants.h" 49 #include "llvm/IR/DataLayout.h" 50 #include "llvm/IR/DebugInfoMetadata.h" 51 #include "llvm/IR/Function.h" 52 #include "llvm/IR/GlobalValue.h" 53 #include "llvm/IR/GlobalVariable.h" 54 #include "llvm/IR/Metadata.h" 55 #include "llvm/IR/Module.h" 56 #include "llvm/MC/MCAsmInfo.h" 57 #include "llvm/MC/MCContext.h" 58 #include "llvm/MC/MCSectionCOFF.h" 59 #include "llvm/MC/MCStreamer.h" 60 #include "llvm/MC/MCSymbol.h" 61 #include "llvm/Support/BinaryByteStream.h" 62 #include "llvm/Support/BinaryStreamReader.h" 63 #include "llvm/Support/BinaryStreamWriter.h" 64 #include "llvm/Support/Casting.h" 65 #include "llvm/Support/CommandLine.h" 66 #include "llvm/Support/Endian.h" 67 #include "llvm/Support/Error.h" 68 #include "llvm/Support/ErrorHandling.h" 69 #include "llvm/Support/FormatVariadic.h" 70 #include "llvm/Support/Path.h" 71 #include "llvm/Support/Program.h" 72 #include "llvm/Support/SMLoc.h" 73 #include "llvm/Support/ScopedPrinter.h" 74 #include "llvm/Target/TargetLoweringObjectFile.h" 75 #include "llvm/Target/TargetMachine.h" 76 #include <algorithm> 77 #include <cassert> 78 #include <cctype> 79 #include <cstddef> 80 #include <iterator> 81 #include <limits> 82 83 using namespace llvm; 84 using namespace llvm::codeview; 85 86 namespace { 87 class CVMCAdapter : public CodeViewRecordStreamer { 88 public: 89 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 90 : OS(&OS), TypeTable(TypeTable) {} 91 92 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 93 94 void emitIntValue(uint64_t Value, unsigned Size) override { 95 OS->emitIntValueInHex(Value, Size); 96 } 97 98 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 99 100 void AddComment(const Twine &T) override { OS->AddComment(T); } 101 102 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 103 104 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 105 106 std::string getTypeName(TypeIndex TI) override { 107 std::string TypeName; 108 if (!TI.isNoneType()) { 109 if (TI.isSimple()) 110 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 111 else 112 TypeName = std::string(TypeTable.getTypeName(TI)); 113 } 114 return TypeName; 115 } 116 117 private: 118 MCStreamer *OS = nullptr; 119 TypeCollection &TypeTable; 120 }; 121 } // namespace 122 123 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 124 switch (Type) { 125 case Triple::ArchType::x86: 126 return CPUType::Pentium3; 127 case Triple::ArchType::x86_64: 128 return CPUType::X64; 129 case Triple::ArchType::thumb: 130 // LLVM currently doesn't support Windows CE and so thumb 131 // here is indiscriminately mapped to ARMNT specifically. 132 return CPUType::ARMNT; 133 case Triple::ArchType::aarch64: 134 return CPUType::ARM64; 135 default: 136 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 137 } 138 } 139 140 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 141 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 142 143 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 144 std::string &Filepath = FileToFilepathMap[File]; 145 if (!Filepath.empty()) 146 return Filepath; 147 148 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 149 150 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 151 // it textually because one of the path components could be a symlink. 152 if (Dir.startswith("/") || Filename.startswith("/")) { 153 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 154 return Filename; 155 Filepath = std::string(Dir); 156 if (Dir.back() != '/') 157 Filepath += '/'; 158 Filepath += Filename; 159 return Filepath; 160 } 161 162 // Clang emits directory and relative filename info into the IR, but CodeView 163 // operates on full paths. We could change Clang to emit full paths too, but 164 // that would increase the IR size and probably not needed for other users. 165 // For now, just concatenate and canonicalize the path here. 166 if (Filename.find(':') == 1) 167 Filepath = std::string(Filename); 168 else 169 Filepath = (Dir + "\\" + Filename).str(); 170 171 // Canonicalize the path. We have to do it textually because we may no longer 172 // have access the file in the filesystem. 173 // First, replace all slashes with backslashes. 174 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 175 176 // Remove all "\.\" with "\". 177 size_t Cursor = 0; 178 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 179 Filepath.erase(Cursor, 2); 180 181 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 182 // path should be well-formatted, e.g. start with a drive letter, etc. 183 Cursor = 0; 184 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 185 // Something's wrong if the path starts with "\..\", abort. 186 if (Cursor == 0) 187 break; 188 189 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 190 if (PrevSlash == std::string::npos) 191 // Something's wrong, abort. 192 break; 193 194 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 195 // The next ".." might be following the one we've just erased. 196 Cursor = PrevSlash; 197 } 198 199 // Remove all duplicate backslashes. 200 Cursor = 0; 201 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 202 Filepath.erase(Cursor, 1); 203 204 return Filepath; 205 } 206 207 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 208 StringRef FullPath = getFullFilepath(F); 209 unsigned NextId = FileIdMap.size() + 1; 210 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 211 if (Insertion.second) { 212 // We have to compute the full filepath and emit a .cv_file directive. 213 ArrayRef<uint8_t> ChecksumAsBytes; 214 FileChecksumKind CSKind = FileChecksumKind::None; 215 if (F->getChecksum()) { 216 std::string Checksum = fromHex(F->getChecksum()->Value); 217 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 218 memcpy(CKMem, Checksum.data(), Checksum.size()); 219 ChecksumAsBytes = ArrayRef<uint8_t>( 220 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 221 switch (F->getChecksum()->Kind) { 222 case DIFile::CSK_MD5: 223 CSKind = FileChecksumKind::MD5; 224 break; 225 case DIFile::CSK_SHA1: 226 CSKind = FileChecksumKind::SHA1; 227 break; 228 case DIFile::CSK_SHA256: 229 CSKind = FileChecksumKind::SHA256; 230 break; 231 } 232 } 233 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 234 static_cast<unsigned>(CSKind)); 235 (void)Success; 236 assert(Success && ".cv_file directive failed"); 237 } 238 return Insertion.first->second; 239 } 240 241 CodeViewDebug::InlineSite & 242 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 243 const DISubprogram *Inlinee) { 244 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 245 InlineSite *Site = &SiteInsertion.first->second; 246 if (SiteInsertion.second) { 247 unsigned ParentFuncId = CurFn->FuncId; 248 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 249 ParentFuncId = 250 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 251 .SiteFuncId; 252 253 Site->SiteFuncId = NextFuncId++; 254 OS.EmitCVInlineSiteIdDirective( 255 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 256 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 257 Site->Inlinee = Inlinee; 258 InlinedSubprograms.insert(Inlinee); 259 getFuncIdForSubprogram(Inlinee); 260 } 261 return *Site; 262 } 263 264 static StringRef getPrettyScopeName(const DIScope *Scope) { 265 StringRef ScopeName = Scope->getName(); 266 if (!ScopeName.empty()) 267 return ScopeName; 268 269 switch (Scope->getTag()) { 270 case dwarf::DW_TAG_enumeration_type: 271 case dwarf::DW_TAG_class_type: 272 case dwarf::DW_TAG_structure_type: 273 case dwarf::DW_TAG_union_type: 274 return "<unnamed-tag>"; 275 case dwarf::DW_TAG_namespace: 276 return "`anonymous namespace'"; 277 default: 278 return StringRef(); 279 } 280 } 281 282 const DISubprogram *CodeViewDebug::collectParentScopeNames( 283 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 284 const DISubprogram *ClosestSubprogram = nullptr; 285 while (Scope != nullptr) { 286 if (ClosestSubprogram == nullptr) 287 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 288 289 // If a type appears in a scope chain, make sure it gets emitted. The 290 // frontend will be responsible for deciding if this should be a forward 291 // declaration or a complete type. 292 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 293 DeferredCompleteTypes.push_back(Ty); 294 295 StringRef ScopeName = getPrettyScopeName(Scope); 296 if (!ScopeName.empty()) 297 QualifiedNameComponents.push_back(ScopeName); 298 Scope = Scope->getScope(); 299 } 300 return ClosestSubprogram; 301 } 302 303 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 304 StringRef TypeName) { 305 std::string FullyQualifiedName; 306 for (StringRef QualifiedNameComponent : 307 llvm::reverse(QualifiedNameComponents)) { 308 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 309 FullyQualifiedName.append("::"); 310 } 311 FullyQualifiedName.append(std::string(TypeName)); 312 return FullyQualifiedName; 313 } 314 315 struct CodeViewDebug::TypeLoweringScope { 316 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 317 ~TypeLoweringScope() { 318 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 319 // inner TypeLoweringScopes don't attempt to emit deferred types. 320 if (CVD.TypeEmissionLevel == 1) 321 CVD.emitDeferredCompleteTypes(); 322 --CVD.TypeEmissionLevel; 323 } 324 CodeViewDebug &CVD; 325 }; 326 327 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 328 StringRef Name) { 329 // Ensure types in the scope chain are emitted as soon as possible. 330 // This can create otherwise a situation where S_UDTs are emitted while 331 // looping in emitDebugInfoForUDTs. 332 TypeLoweringScope S(*this); 333 SmallVector<StringRef, 5> QualifiedNameComponents; 334 collectParentScopeNames(Scope, QualifiedNameComponents); 335 return formatNestedName(QualifiedNameComponents, Name); 336 } 337 338 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 339 const DIScope *Scope = Ty->getScope(); 340 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 341 } 342 343 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 344 // No scope means global scope and that uses the zero index. 345 // 346 // We also use zero index when the scope is a DISubprogram 347 // to suppress the emission of LF_STRING_ID for the function, 348 // which can trigger a link-time error with the linker in 349 // VS2019 version 16.11.2 or newer. 350 // Note, however, skipping the debug info emission for the DISubprogram 351 // is a temporary fix. The root issue here is that we need to figure out 352 // the proper way to encode a function nested in another function 353 // (as introduced by the Fortran 'contains' keyword) in CodeView. 354 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 355 return TypeIndex(); 356 357 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 358 359 // Check if we've already translated this scope. 360 auto I = TypeIndices.find({Scope, nullptr}); 361 if (I != TypeIndices.end()) 362 return I->second; 363 364 // Build the fully qualified name of the scope. 365 std::string ScopeName = getFullyQualifiedName(Scope); 366 StringIdRecord SID(TypeIndex(), ScopeName); 367 auto TI = TypeTable.writeLeafType(SID); 368 return recordTypeIndexForDINode(Scope, TI); 369 } 370 371 static StringRef removeTemplateArgs(StringRef Name) { 372 // Remove template args from the display name. Assume that the template args 373 // are the last thing in the name. 374 if (Name.empty() || Name.back() != '>') 375 return Name; 376 377 int OpenBrackets = 0; 378 for (int i = Name.size() - 1; i >= 0; --i) { 379 if (Name[i] == '>') 380 ++OpenBrackets; 381 else if (Name[i] == '<') { 382 --OpenBrackets; 383 if (OpenBrackets == 0) 384 return Name.substr(0, i); 385 } 386 } 387 return Name; 388 } 389 390 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 391 assert(SP); 392 393 // Check if we've already translated this subprogram. 394 auto I = TypeIndices.find({SP, nullptr}); 395 if (I != TypeIndices.end()) 396 return I->second; 397 398 // The display name includes function template arguments. Drop them to match 399 // MSVC. We need to have the template arguments in the DISubprogram name 400 // because they are used in other symbol records, such as S_GPROC32_IDs. 401 StringRef DisplayName = removeTemplateArgs(SP->getName()); 402 403 const DIScope *Scope = SP->getScope(); 404 TypeIndex TI; 405 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 406 // If the scope is a DICompositeType, then this must be a method. Member 407 // function types take some special handling, and require access to the 408 // subprogram. 409 TypeIndex ClassType = getTypeIndex(Class); 410 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 411 DisplayName); 412 TI = TypeTable.writeLeafType(MFuncId); 413 } else { 414 // Otherwise, this must be a free function. 415 TypeIndex ParentScope = getScopeIndex(Scope); 416 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 417 TI = TypeTable.writeLeafType(FuncId); 418 } 419 420 return recordTypeIndexForDINode(SP, TI); 421 } 422 423 static bool isNonTrivial(const DICompositeType *DCTy) { 424 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 425 } 426 427 static FunctionOptions 428 getFunctionOptions(const DISubroutineType *Ty, 429 const DICompositeType *ClassTy = nullptr, 430 StringRef SPName = StringRef("")) { 431 FunctionOptions FO = FunctionOptions::None; 432 const DIType *ReturnTy = nullptr; 433 if (auto TypeArray = Ty->getTypeArray()) { 434 if (TypeArray.size()) 435 ReturnTy = TypeArray[0]; 436 } 437 438 // Add CxxReturnUdt option to functions that return nontrivial record types 439 // or methods that return record types. 440 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 441 if (isNonTrivial(ReturnDCTy) || ClassTy) 442 FO |= FunctionOptions::CxxReturnUdt; 443 444 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 445 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 446 FO |= FunctionOptions::Constructor; 447 448 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 449 450 } 451 return FO; 452 } 453 454 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 455 const DICompositeType *Class) { 456 // Always use the method declaration as the key for the function type. The 457 // method declaration contains the this adjustment. 458 if (SP->getDeclaration()) 459 SP = SP->getDeclaration(); 460 assert(!SP->getDeclaration() && "should use declaration as key"); 461 462 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 463 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 464 auto I = TypeIndices.find({SP, Class}); 465 if (I != TypeIndices.end()) 466 return I->second; 467 468 // Make sure complete type info for the class is emitted *after* the member 469 // function type, as the complete class type is likely to reference this 470 // member function type. 471 TypeLoweringScope S(*this); 472 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 473 474 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 475 TypeIndex TI = lowerTypeMemberFunction( 476 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 477 return recordTypeIndexForDINode(SP, TI, Class); 478 } 479 480 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 481 TypeIndex TI, 482 const DIType *ClassTy) { 483 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 484 (void)InsertResult; 485 assert(InsertResult.second && "DINode was already assigned a type index"); 486 return TI; 487 } 488 489 unsigned CodeViewDebug::getPointerSizeInBytes() { 490 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 491 } 492 493 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 494 const LexicalScope *LS) { 495 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 496 // This variable was inlined. Associate it with the InlineSite. 497 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 498 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 499 Site.InlinedLocals.emplace_back(Var); 500 } else { 501 // This variable goes into the corresponding lexical scope. 502 ScopeVariables[LS].emplace_back(Var); 503 } 504 } 505 506 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 507 const DILocation *Loc) { 508 if (!llvm::is_contained(Locs, Loc)) 509 Locs.push_back(Loc); 510 } 511 512 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 513 const MachineFunction *MF) { 514 // Skip this instruction if it has the same location as the previous one. 515 if (!DL || DL == PrevInstLoc) 516 return; 517 518 const DIScope *Scope = DL.get()->getScope(); 519 if (!Scope) 520 return; 521 522 // Skip this line if it is longer than the maximum we can record. 523 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 524 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 525 LI.isNeverStepInto()) 526 return; 527 528 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 529 if (CI.getStartColumn() != DL.getCol()) 530 return; 531 532 if (!CurFn->HaveLineInfo) 533 CurFn->HaveLineInfo = true; 534 unsigned FileId = 0; 535 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 536 FileId = CurFn->LastFileId; 537 else 538 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 539 PrevInstLoc = DL; 540 541 unsigned FuncId = CurFn->FuncId; 542 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 543 const DILocation *Loc = DL.get(); 544 545 // If this location was actually inlined from somewhere else, give it the ID 546 // of the inline call site. 547 FuncId = 548 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 549 550 // Ensure we have links in the tree of inline call sites. 551 bool FirstLoc = true; 552 while ((SiteLoc = Loc->getInlinedAt())) { 553 InlineSite &Site = 554 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 555 if (!FirstLoc) 556 addLocIfNotPresent(Site.ChildSites, Loc); 557 FirstLoc = false; 558 Loc = SiteLoc; 559 } 560 addLocIfNotPresent(CurFn->ChildSites, Loc); 561 } 562 563 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 564 /*PrologueEnd=*/false, /*IsStmt=*/false, 565 DL->getFilename(), SMLoc()); 566 } 567 568 void CodeViewDebug::emitCodeViewMagicVersion() { 569 OS.emitValueToAlignment(4); 570 OS.AddComment("Debug section magic"); 571 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 572 } 573 574 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 575 switch (DWLang) { 576 case dwarf::DW_LANG_C: 577 case dwarf::DW_LANG_C89: 578 case dwarf::DW_LANG_C99: 579 case dwarf::DW_LANG_C11: 580 case dwarf::DW_LANG_ObjC: 581 return SourceLanguage::C; 582 case dwarf::DW_LANG_C_plus_plus: 583 case dwarf::DW_LANG_C_plus_plus_03: 584 case dwarf::DW_LANG_C_plus_plus_11: 585 case dwarf::DW_LANG_C_plus_plus_14: 586 return SourceLanguage::Cpp; 587 case dwarf::DW_LANG_Fortran77: 588 case dwarf::DW_LANG_Fortran90: 589 case dwarf::DW_LANG_Fortran95: 590 case dwarf::DW_LANG_Fortran03: 591 case dwarf::DW_LANG_Fortran08: 592 return SourceLanguage::Fortran; 593 case dwarf::DW_LANG_Pascal83: 594 return SourceLanguage::Pascal; 595 case dwarf::DW_LANG_Cobol74: 596 case dwarf::DW_LANG_Cobol85: 597 return SourceLanguage::Cobol; 598 case dwarf::DW_LANG_Java: 599 return SourceLanguage::Java; 600 case dwarf::DW_LANG_D: 601 return SourceLanguage::D; 602 case dwarf::DW_LANG_Swift: 603 return SourceLanguage::Swift; 604 case dwarf::DW_LANG_Rust: 605 return SourceLanguage::Rust; 606 default: 607 // There's no CodeView representation for this language, and CV doesn't 608 // have an "unknown" option for the language field, so we'll use MASM, 609 // as it's very low level. 610 return SourceLanguage::Masm; 611 } 612 } 613 614 void CodeViewDebug::beginModule(Module *M) { 615 // If module doesn't have named metadata anchors or COFF debug section 616 // is not available, skip any debug info related stuff. 617 NamedMDNode *CUs = M->getNamedMetadata("llvm.dbg.cu"); 618 if (!CUs || !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 619 Asm = nullptr; 620 return; 621 } 622 // Tell MMI that we have and need debug info. 623 MMI->setDebugInfoAvailability(true); 624 625 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 626 627 // Get the current source language. 628 const MDNode *Node = *CUs->operands().begin(); 629 const auto *CU = cast<DICompileUnit>(Node); 630 631 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 632 633 collectGlobalVariableInfo(); 634 635 // Check if we should emit type record hashes. 636 ConstantInt *GH = 637 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 638 EmitDebugGlobalHashes = GH && !GH->isZero(); 639 } 640 641 void CodeViewDebug::endModule() { 642 if (!Asm || !MMI->hasDebugInfo()) 643 return; 644 645 // The COFF .debug$S section consists of several subsections, each starting 646 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 647 // of the payload followed by the payload itself. The subsections are 4-byte 648 // aligned. 649 650 // Use the generic .debug$S section, and make a subsection for all the inlined 651 // subprograms. 652 switchToDebugSectionForSymbol(nullptr); 653 654 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 655 emitObjName(); 656 emitCompilerInformation(); 657 endCVSubsection(CompilerInfo); 658 659 emitInlineeLinesSubsection(); 660 661 // Emit per-function debug information. 662 for (auto &P : FnDebugInfo) 663 if (!P.first->isDeclarationForLinker()) 664 emitDebugInfoForFunction(P.first, *P.second); 665 666 // Get types used by globals without emitting anything. 667 // This is meant to collect all static const data members so they can be 668 // emitted as globals. 669 collectDebugInfoForGlobals(); 670 671 // Emit retained types. 672 emitDebugInfoForRetainedTypes(); 673 674 // Emit global variable debug information. 675 setCurrentSubprogram(nullptr); 676 emitDebugInfoForGlobals(); 677 678 // Switch back to the generic .debug$S section after potentially processing 679 // comdat symbol sections. 680 switchToDebugSectionForSymbol(nullptr); 681 682 // Emit UDT records for any types used by global variables. 683 if (!GlobalUDTs.empty()) { 684 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 685 emitDebugInfoForUDTs(GlobalUDTs); 686 endCVSubsection(SymbolsEnd); 687 } 688 689 // This subsection holds a file index to offset in string table table. 690 OS.AddComment("File index to string table offset subsection"); 691 OS.emitCVFileChecksumsDirective(); 692 693 // This subsection holds the string table. 694 OS.AddComment("String table"); 695 OS.emitCVStringTableDirective(); 696 697 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 698 // subsection in the generic .debug$S section at the end. There is no 699 // particular reason for this ordering other than to match MSVC. 700 emitBuildInfo(); 701 702 // Emit type information and hashes last, so that any types we translate while 703 // emitting function info are included. 704 emitTypeInformation(); 705 706 if (EmitDebugGlobalHashes) 707 emitTypeGlobalHashes(); 708 709 clear(); 710 } 711 712 static void 713 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 714 unsigned MaxFixedRecordLength = 0xF00) { 715 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 716 // after a fixed length portion of the record. The fixed length portion should 717 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 718 // overall record size is less than the maximum allowed. 719 SmallString<32> NullTerminatedString( 720 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 721 NullTerminatedString.push_back('\0'); 722 OS.emitBytes(NullTerminatedString); 723 } 724 725 void CodeViewDebug::emitTypeInformation() { 726 if (TypeTable.empty()) 727 return; 728 729 // Start the .debug$T or .debug$P section with 0x4. 730 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 731 emitCodeViewMagicVersion(); 732 733 TypeTableCollection Table(TypeTable.records()); 734 TypeVisitorCallbackPipeline Pipeline; 735 736 // To emit type record using Codeview MCStreamer adapter 737 CVMCAdapter CVMCOS(OS, Table); 738 TypeRecordMapping typeMapping(CVMCOS); 739 Pipeline.addCallbackToPipeline(typeMapping); 740 741 Optional<TypeIndex> B = Table.getFirst(); 742 while (B) { 743 // This will fail if the record data is invalid. 744 CVType Record = Table.getType(*B); 745 746 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 747 748 if (E) { 749 logAllUnhandledErrors(std::move(E), errs(), "error: "); 750 llvm_unreachable("produced malformed type record"); 751 } 752 753 B = Table.getNext(*B); 754 } 755 } 756 757 void CodeViewDebug::emitTypeGlobalHashes() { 758 if (TypeTable.empty()) 759 return; 760 761 // Start the .debug$H section with the version and hash algorithm, currently 762 // hardcoded to version 0, SHA1. 763 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 764 765 OS.emitValueToAlignment(4); 766 OS.AddComment("Magic"); 767 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 768 OS.AddComment("Section Version"); 769 OS.emitInt16(0); 770 OS.AddComment("Hash Algorithm"); 771 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 772 773 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 774 for (const auto &GHR : TypeTable.hashes()) { 775 if (OS.isVerboseAsm()) { 776 // Emit an EOL-comment describing which TypeIndex this hash corresponds 777 // to, as well as the stringified SHA1 hash. 778 SmallString<32> Comment; 779 raw_svector_ostream CommentOS(Comment); 780 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 781 OS.AddComment(Comment); 782 ++TI; 783 } 784 assert(GHR.Hash.size() == 8); 785 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 786 GHR.Hash.size()); 787 OS.emitBinaryData(S); 788 } 789 } 790 791 void CodeViewDebug::emitObjName() { 792 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 793 794 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 795 llvm::SmallString<256> PathStore(PathRef); 796 797 if (PathRef.empty() || PathRef == "-") { 798 // Don't emit the filename if we're writing to stdout or to /dev/null. 799 PathRef = {}; 800 } else { 801 llvm::sys::path::remove_dots(PathStore, /*remove_dot_dot=*/true); 802 PathRef = PathStore; 803 } 804 805 OS.AddComment("Signature"); 806 OS.emitIntValue(0, 4); 807 808 OS.AddComment("Object name"); 809 emitNullTerminatedSymbolName(OS, PathRef); 810 811 endSymbolRecord(CompilerEnd); 812 } 813 814 namespace { 815 struct Version { 816 int Part[4]; 817 }; 818 } // end anonymous namespace 819 820 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 821 // the version number. 822 static Version parseVersion(StringRef Name) { 823 Version V = {{0}}; 824 int N = 0; 825 for (const char C : Name) { 826 if (isdigit(C)) { 827 V.Part[N] *= 10; 828 V.Part[N] += C - '0'; 829 } else if (C == '.') { 830 ++N; 831 if (N >= 4) 832 return V; 833 } else if (N > 0) 834 return V; 835 } 836 return V; 837 } 838 839 void CodeViewDebug::emitCompilerInformation() { 840 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 841 uint32_t Flags = 0; 842 843 // The low byte of the flags indicates the source language. 844 Flags = CurrentSourceLanguage; 845 // TODO: Figure out which other flags need to be set. 846 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 847 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 848 } 849 using ArchType = llvm::Triple::ArchType; 850 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 851 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 852 Arch == ArchType::aarch64) { 853 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 854 } 855 856 OS.AddComment("Flags and language"); 857 OS.emitInt32(Flags); 858 859 OS.AddComment("CPUType"); 860 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 861 862 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 863 const MDNode *Node = *CUs->operands().begin(); 864 const auto *CU = cast<DICompileUnit>(Node); 865 866 StringRef CompilerVersion = CU->getProducer(); 867 Version FrontVer = parseVersion(CompilerVersion); 868 OS.AddComment("Frontend version"); 869 for (int N : FrontVer.Part) { 870 N = std::min<int>(N, std::numeric_limits<uint16_t>::max()); 871 OS.emitInt16(N); 872 } 873 874 // Some Microsoft tools, like Binscope, expect a backend version number of at 875 // least 8.something, so we'll coerce the LLVM version into a form that 876 // guarantees it'll be big enough without really lying about the version. 877 int Major = 1000 * LLVM_VERSION_MAJOR + 878 10 * LLVM_VERSION_MINOR + 879 LLVM_VERSION_PATCH; 880 // Clamp it for builds that use unusually large version numbers. 881 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 882 Version BackVer = {{ Major, 0, 0, 0 }}; 883 OS.AddComment("Backend version"); 884 for (int N : BackVer.Part) 885 OS.emitInt16(N); 886 887 OS.AddComment("Null-terminated compiler version string"); 888 emitNullTerminatedSymbolName(OS, CompilerVersion); 889 890 endSymbolRecord(CompilerEnd); 891 } 892 893 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 894 StringRef S) { 895 StringIdRecord SIR(TypeIndex(0x0), S); 896 return TypeTable.writeLeafType(SIR); 897 } 898 899 static std::string flattenCommandLine(ArrayRef<std::string> Args, 900 StringRef MainFilename) { 901 std::string FlatCmdLine; 902 raw_string_ostream OS(FlatCmdLine); 903 bool PrintedOneArg = false; 904 if (!StringRef(Args[0]).contains("-cc1")) { 905 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 906 PrintedOneArg = true; 907 } 908 for (unsigned i = 0; i < Args.size(); i++) { 909 StringRef Arg = Args[i]; 910 if (Arg.empty()) 911 continue; 912 if (Arg == "-main-file-name" || Arg == "-o") { 913 i++; // Skip this argument and next one. 914 continue; 915 } 916 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 917 continue; 918 if (PrintedOneArg) 919 OS << " "; 920 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 921 PrintedOneArg = true; 922 } 923 OS.flush(); 924 return FlatCmdLine; 925 } 926 927 void CodeViewDebug::emitBuildInfo() { 928 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 929 // build info. The known prefix is: 930 // - Absolute path of current directory 931 // - Compiler path 932 // - Main source file path, relative to CWD or absolute 933 // - Type server PDB file 934 // - Canonical compiler command line 935 // If frontend and backend compilation are separated (think llc or LTO), it's 936 // not clear if the compiler path should refer to the executable for the 937 // frontend or the backend. Leave it blank for now. 938 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 939 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 940 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 941 const auto *CU = cast<DICompileUnit>(Node); 942 const DIFile *MainSourceFile = CU->getFile(); 943 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 944 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 945 BuildInfoArgs[BuildInfoRecord::SourceFile] = 946 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 947 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 948 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 949 getStringIdTypeIdx(TypeTable, ""); 950 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 951 BuildInfoArgs[BuildInfoRecord::BuildTool] = 952 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 953 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 954 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 955 MainSourceFile->getFilename())); 956 } 957 BuildInfoRecord BIR(BuildInfoArgs); 958 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 959 960 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 961 // from the module symbols into the type stream. 962 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 963 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 964 OS.AddComment("LF_BUILDINFO index"); 965 OS.emitInt32(BuildInfoIndex.getIndex()); 966 endSymbolRecord(BIEnd); 967 endCVSubsection(BISubsecEnd); 968 } 969 970 void CodeViewDebug::emitInlineeLinesSubsection() { 971 if (InlinedSubprograms.empty()) 972 return; 973 974 OS.AddComment("Inlinee lines subsection"); 975 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 976 977 // We emit the checksum info for files. This is used by debuggers to 978 // determine if a pdb matches the source before loading it. Visual Studio, 979 // for instance, will display a warning that the breakpoints are not valid if 980 // the pdb does not match the source. 981 OS.AddComment("Inlinee lines signature"); 982 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 983 984 for (const DISubprogram *SP : InlinedSubprograms) { 985 assert(TypeIndices.count({SP, nullptr})); 986 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 987 988 OS.AddBlankLine(); 989 unsigned FileId = maybeRecordFile(SP->getFile()); 990 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 991 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 992 OS.AddBlankLine(); 993 OS.AddComment("Type index of inlined function"); 994 OS.emitInt32(InlineeIdx.getIndex()); 995 OS.AddComment("Offset into filechecksum table"); 996 OS.emitCVFileChecksumOffsetDirective(FileId); 997 OS.AddComment("Starting line number"); 998 OS.emitInt32(SP->getLine()); 999 } 1000 1001 endCVSubsection(InlineEnd); 1002 } 1003 1004 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 1005 const DILocation *InlinedAt, 1006 const InlineSite &Site) { 1007 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1008 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1009 1010 // SymbolRecord 1011 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1012 1013 OS.AddComment("PtrParent"); 1014 OS.emitInt32(0); 1015 OS.AddComment("PtrEnd"); 1016 OS.emitInt32(0); 1017 OS.AddComment("Inlinee type index"); 1018 OS.emitInt32(InlineeIdx.getIndex()); 1019 1020 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1021 unsigned StartLineNum = Site.Inlinee->getLine(); 1022 1023 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1024 FI.Begin, FI.End); 1025 1026 endSymbolRecord(InlineEnd); 1027 1028 emitLocalVariableList(FI, Site.InlinedLocals); 1029 1030 // Recurse on child inlined call sites before closing the scope. 1031 for (const DILocation *ChildSite : Site.ChildSites) { 1032 auto I = FI.InlineSites.find(ChildSite); 1033 assert(I != FI.InlineSites.end() && 1034 "child site not in function inline site map"); 1035 emitInlinedCallSite(FI, ChildSite, I->second); 1036 } 1037 1038 // Close the scope. 1039 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1040 } 1041 1042 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1043 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1044 // comdat key. A section may be comdat because of -ffunction-sections or 1045 // because it is comdat in the IR. 1046 MCSectionCOFF *GVSec = 1047 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1048 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1049 1050 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1051 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1052 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1053 1054 OS.SwitchSection(DebugSec); 1055 1056 // Emit the magic version number if this is the first time we've switched to 1057 // this section. 1058 if (ComdatDebugSections.insert(DebugSec).second) 1059 emitCodeViewMagicVersion(); 1060 } 1061 1062 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1063 // The only supported thunk ordinal is currently the standard type. 1064 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1065 FunctionInfo &FI, 1066 const MCSymbol *Fn) { 1067 std::string FuncName = 1068 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1069 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1070 1071 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1072 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1073 1074 // Emit S_THUNK32 1075 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1076 OS.AddComment("PtrParent"); 1077 OS.emitInt32(0); 1078 OS.AddComment("PtrEnd"); 1079 OS.emitInt32(0); 1080 OS.AddComment("PtrNext"); 1081 OS.emitInt32(0); 1082 OS.AddComment("Thunk section relative address"); 1083 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1084 OS.AddComment("Thunk section index"); 1085 OS.EmitCOFFSectionIndex(Fn); 1086 OS.AddComment("Code size"); 1087 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1088 OS.AddComment("Ordinal"); 1089 OS.emitInt8(unsigned(ordinal)); 1090 OS.AddComment("Function name"); 1091 emitNullTerminatedSymbolName(OS, FuncName); 1092 // Additional fields specific to the thunk ordinal would go here. 1093 endSymbolRecord(ThunkRecordEnd); 1094 1095 // Local variables/inlined routines are purposely omitted here. The point of 1096 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1097 1098 // Emit S_PROC_ID_END 1099 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1100 1101 endCVSubsection(SymbolsEnd); 1102 } 1103 1104 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1105 FunctionInfo &FI) { 1106 // For each function there is a separate subsection which holds the PC to 1107 // file:line table. 1108 const MCSymbol *Fn = Asm->getSymbol(GV); 1109 assert(Fn); 1110 1111 // Switch to the to a comdat section, if appropriate. 1112 switchToDebugSectionForSymbol(Fn); 1113 1114 std::string FuncName; 1115 auto *SP = GV->getSubprogram(); 1116 assert(SP); 1117 setCurrentSubprogram(SP); 1118 1119 if (SP->isThunk()) { 1120 emitDebugInfoForThunk(GV, FI, Fn); 1121 return; 1122 } 1123 1124 // If we have a display name, build the fully qualified name by walking the 1125 // chain of scopes. 1126 if (!SP->getName().empty()) 1127 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1128 1129 // If our DISubprogram name is empty, use the mangled name. 1130 if (FuncName.empty()) 1131 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1132 1133 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1134 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1135 OS.EmitCVFPOData(Fn); 1136 1137 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1138 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1139 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1140 { 1141 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1142 : SymbolKind::S_GPROC32_ID; 1143 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1144 1145 // These fields are filled in by tools like CVPACK which run after the fact. 1146 OS.AddComment("PtrParent"); 1147 OS.emitInt32(0); 1148 OS.AddComment("PtrEnd"); 1149 OS.emitInt32(0); 1150 OS.AddComment("PtrNext"); 1151 OS.emitInt32(0); 1152 // This is the important bit that tells the debugger where the function 1153 // code is located and what's its size: 1154 OS.AddComment("Code size"); 1155 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1156 OS.AddComment("Offset after prologue"); 1157 OS.emitInt32(0); 1158 OS.AddComment("Offset before epilogue"); 1159 OS.emitInt32(0); 1160 OS.AddComment("Function type index"); 1161 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1162 OS.AddComment("Function section relative address"); 1163 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1164 OS.AddComment("Function section index"); 1165 OS.EmitCOFFSectionIndex(Fn); 1166 OS.AddComment("Flags"); 1167 OS.emitInt8(0); 1168 // Emit the function display name as a null-terminated string. 1169 OS.AddComment("Function name"); 1170 // Truncate the name so we won't overflow the record length field. 1171 emitNullTerminatedSymbolName(OS, FuncName); 1172 endSymbolRecord(ProcRecordEnd); 1173 1174 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1175 // Subtract out the CSR size since MSVC excludes that and we include it. 1176 OS.AddComment("FrameSize"); 1177 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1178 OS.AddComment("Padding"); 1179 OS.emitInt32(0); 1180 OS.AddComment("Offset of padding"); 1181 OS.emitInt32(0); 1182 OS.AddComment("Bytes of callee saved registers"); 1183 OS.emitInt32(FI.CSRSize); 1184 OS.AddComment("Exception handler offset"); 1185 OS.emitInt32(0); 1186 OS.AddComment("Exception handler section"); 1187 OS.emitInt16(0); 1188 OS.AddComment("Flags (defines frame register)"); 1189 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1190 endSymbolRecord(FrameProcEnd); 1191 1192 emitLocalVariableList(FI, FI.Locals); 1193 emitGlobalVariableList(FI.Globals); 1194 emitLexicalBlockList(FI.ChildBlocks, FI); 1195 1196 // Emit inlined call site information. Only emit functions inlined directly 1197 // into the parent function. We'll emit the other sites recursively as part 1198 // of their parent inline site. 1199 for (const DILocation *InlinedAt : FI.ChildSites) { 1200 auto I = FI.InlineSites.find(InlinedAt); 1201 assert(I != FI.InlineSites.end() && 1202 "child site not in function inline site map"); 1203 emitInlinedCallSite(FI, InlinedAt, I->second); 1204 } 1205 1206 for (auto Annot : FI.Annotations) { 1207 MCSymbol *Label = Annot.first; 1208 MDTuple *Strs = cast<MDTuple>(Annot.second); 1209 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1210 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1211 // FIXME: Make sure we don't overflow the max record size. 1212 OS.EmitCOFFSectionIndex(Label); 1213 OS.emitInt16(Strs->getNumOperands()); 1214 for (Metadata *MD : Strs->operands()) { 1215 // MDStrings are null terminated, so we can do EmitBytes and get the 1216 // nice .asciz directive. 1217 StringRef Str = cast<MDString>(MD)->getString(); 1218 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1219 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1220 } 1221 endSymbolRecord(AnnotEnd); 1222 } 1223 1224 for (auto HeapAllocSite : FI.HeapAllocSites) { 1225 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1226 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1227 const DIType *DITy = std::get<2>(HeapAllocSite); 1228 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1229 OS.AddComment("Call site offset"); 1230 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1231 OS.AddComment("Call site section index"); 1232 OS.EmitCOFFSectionIndex(BeginLabel); 1233 OS.AddComment("Call instruction length"); 1234 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1235 OS.AddComment("Type index"); 1236 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1237 endSymbolRecord(HeapAllocEnd); 1238 } 1239 1240 if (SP != nullptr) 1241 emitDebugInfoForUDTs(LocalUDTs); 1242 1243 // We're done with this function. 1244 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1245 } 1246 endCVSubsection(SymbolsEnd); 1247 1248 // We have an assembler directive that takes care of the whole line table. 1249 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1250 } 1251 1252 CodeViewDebug::LocalVarDefRange 1253 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1254 LocalVarDefRange DR; 1255 DR.InMemory = -1; 1256 DR.DataOffset = Offset; 1257 assert(DR.DataOffset == Offset && "truncation"); 1258 DR.IsSubfield = 0; 1259 DR.StructOffset = 0; 1260 DR.CVRegister = CVRegister; 1261 return DR; 1262 } 1263 1264 void CodeViewDebug::collectVariableInfoFromMFTable( 1265 DenseSet<InlinedEntity> &Processed) { 1266 const MachineFunction &MF = *Asm->MF; 1267 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1268 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1269 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1270 1271 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1272 if (!VI.Var) 1273 continue; 1274 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1275 "Expected inlined-at fields to agree"); 1276 1277 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1278 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1279 1280 // If variable scope is not found then skip this variable. 1281 if (!Scope) 1282 continue; 1283 1284 // If the variable has an attached offset expression, extract it. 1285 // FIXME: Try to handle DW_OP_deref as well. 1286 int64_t ExprOffset = 0; 1287 bool Deref = false; 1288 if (VI.Expr) { 1289 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1290 if (VI.Expr->getNumElements() == 1 && 1291 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1292 Deref = true; 1293 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1294 continue; 1295 } 1296 1297 // Get the frame register used and the offset. 1298 Register FrameReg; 1299 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1300 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1301 1302 assert(!FrameOffset.getScalable() && 1303 "Frame offsets with a scalable component are not supported"); 1304 1305 // Calculate the label ranges. 1306 LocalVarDefRange DefRange = 1307 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1308 1309 for (const InsnRange &Range : Scope->getRanges()) { 1310 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1311 const MCSymbol *End = getLabelAfterInsn(Range.second); 1312 End = End ? End : Asm->getFunctionEnd(); 1313 DefRange.Ranges.emplace_back(Begin, End); 1314 } 1315 1316 LocalVariable Var; 1317 Var.DIVar = VI.Var; 1318 Var.DefRanges.emplace_back(std::move(DefRange)); 1319 if (Deref) 1320 Var.UseReferenceType = true; 1321 1322 recordLocalVariable(std::move(Var), Scope); 1323 } 1324 } 1325 1326 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1327 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1328 } 1329 1330 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1331 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1332 } 1333 1334 void CodeViewDebug::calculateRanges( 1335 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1336 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1337 1338 // Calculate the definition ranges. 1339 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1340 const auto &Entry = *I; 1341 if (!Entry.isDbgValue()) 1342 continue; 1343 const MachineInstr *DVInst = Entry.getInstr(); 1344 assert(DVInst->isDebugValue() && "Invalid History entry"); 1345 // FIXME: Find a way to represent constant variables, since they are 1346 // relatively common. 1347 Optional<DbgVariableLocation> Location = 1348 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1349 if (!Location) 1350 continue; 1351 1352 // CodeView can only express variables in register and variables in memory 1353 // at a constant offset from a register. However, for variables passed 1354 // indirectly by pointer, it is common for that pointer to be spilled to a 1355 // stack location. For the special case of one offseted load followed by a 1356 // zero offset load (a pointer spilled to the stack), we change the type of 1357 // the local variable from a value type to a reference type. This tricks the 1358 // debugger into doing the load for us. 1359 if (Var.UseReferenceType) { 1360 // We're using a reference type. Drop the last zero offset load. 1361 if (canUseReferenceType(*Location)) 1362 Location->LoadChain.pop_back(); 1363 else 1364 continue; 1365 } else if (needsReferenceType(*Location)) { 1366 // This location can't be expressed without switching to a reference type. 1367 // Start over using that. 1368 Var.UseReferenceType = true; 1369 Var.DefRanges.clear(); 1370 calculateRanges(Var, Entries); 1371 return; 1372 } 1373 1374 // We can only handle a register or an offseted load of a register. 1375 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1376 continue; 1377 { 1378 LocalVarDefRange DR; 1379 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1380 DR.InMemory = !Location->LoadChain.empty(); 1381 DR.DataOffset = 1382 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1383 if (Location->FragmentInfo) { 1384 DR.IsSubfield = true; 1385 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1386 } else { 1387 DR.IsSubfield = false; 1388 DR.StructOffset = 0; 1389 } 1390 1391 if (Var.DefRanges.empty() || 1392 Var.DefRanges.back().isDifferentLocation(DR)) { 1393 Var.DefRanges.emplace_back(std::move(DR)); 1394 } 1395 } 1396 1397 // Compute the label range. 1398 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1399 const MCSymbol *End; 1400 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1401 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1402 End = EndingEntry.isDbgValue() 1403 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1404 : getLabelAfterInsn(EndingEntry.getInstr()); 1405 } else 1406 End = Asm->getFunctionEnd(); 1407 1408 // If the last range end is our begin, just extend the last range. 1409 // Otherwise make a new range. 1410 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1411 Var.DefRanges.back().Ranges; 1412 if (!R.empty() && R.back().second == Begin) 1413 R.back().second = End; 1414 else 1415 R.emplace_back(Begin, End); 1416 1417 // FIXME: Do more range combining. 1418 } 1419 } 1420 1421 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1422 DenseSet<InlinedEntity> Processed; 1423 // Grab the variable info that was squirreled away in the MMI side-table. 1424 collectVariableInfoFromMFTable(Processed); 1425 1426 for (const auto &I : DbgValues) { 1427 InlinedEntity IV = I.first; 1428 if (Processed.count(IV)) 1429 continue; 1430 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1431 const DILocation *InlinedAt = IV.second; 1432 1433 // Instruction ranges, specifying where IV is accessible. 1434 const auto &Entries = I.second; 1435 1436 LexicalScope *Scope = nullptr; 1437 if (InlinedAt) 1438 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1439 else 1440 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1441 // If variable scope is not found then skip this variable. 1442 if (!Scope) 1443 continue; 1444 1445 LocalVariable Var; 1446 Var.DIVar = DIVar; 1447 1448 calculateRanges(Var, Entries); 1449 recordLocalVariable(std::move(Var), Scope); 1450 } 1451 } 1452 1453 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1454 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1455 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1456 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1457 const Function &GV = MF->getFunction(); 1458 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1459 assert(Insertion.second && "function already has info"); 1460 CurFn = Insertion.first->second.get(); 1461 CurFn->FuncId = NextFuncId++; 1462 CurFn->Begin = Asm->getFunctionBegin(); 1463 1464 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1465 // callee-saved registers were used. For targets that don't use a PUSH 1466 // instruction (AArch64), this will be zero. 1467 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1468 CurFn->FrameSize = MFI.getStackSize(); 1469 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1470 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1471 1472 // For this function S_FRAMEPROC record, figure out which codeview register 1473 // will be the frame pointer. 1474 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1475 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1476 if (CurFn->FrameSize > 0) { 1477 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1478 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1479 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1480 } else { 1481 // If there is an FP, parameters are always relative to it. 1482 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1483 if (CurFn->HasStackRealignment) { 1484 // If the stack needs realignment, locals are relative to SP or VFRAME. 1485 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1486 } else { 1487 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1488 // other stack adjustments. 1489 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1490 } 1491 } 1492 } 1493 1494 // Compute other frame procedure options. 1495 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1496 if (MFI.hasVarSizedObjects()) 1497 FPO |= FrameProcedureOptions::HasAlloca; 1498 if (MF->exposesReturnsTwice()) 1499 FPO |= FrameProcedureOptions::HasSetJmp; 1500 // FIXME: Set HasLongJmp if we ever track that info. 1501 if (MF->hasInlineAsm()) 1502 FPO |= FrameProcedureOptions::HasInlineAssembly; 1503 if (GV.hasPersonalityFn()) { 1504 if (isAsynchronousEHPersonality( 1505 classifyEHPersonality(GV.getPersonalityFn()))) 1506 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1507 else 1508 FPO |= FrameProcedureOptions::HasExceptionHandling; 1509 } 1510 if (GV.hasFnAttribute(Attribute::InlineHint)) 1511 FPO |= FrameProcedureOptions::MarkedInline; 1512 if (GV.hasFnAttribute(Attribute::Naked)) 1513 FPO |= FrameProcedureOptions::Naked; 1514 if (MFI.hasStackProtectorIndex()) 1515 FPO |= FrameProcedureOptions::SecurityChecks; 1516 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1517 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1518 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1519 !GV.hasOptSize() && !GV.hasOptNone()) 1520 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1521 if (GV.hasProfileData()) { 1522 FPO |= FrameProcedureOptions::ValidProfileCounts; 1523 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1524 } 1525 // FIXME: Set GuardCfg when it is implemented. 1526 CurFn->FrameProcOpts = FPO; 1527 1528 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1529 1530 // Find the end of the function prolog. First known non-DBG_VALUE and 1531 // non-frame setup location marks the beginning of the function body. 1532 // FIXME: is there a simpler a way to do this? Can we just search 1533 // for the first instruction of the function, not the last of the prolog? 1534 DebugLoc PrologEndLoc; 1535 bool EmptyPrologue = true; 1536 for (const auto &MBB : *MF) { 1537 for (const auto &MI : MBB) { 1538 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1539 MI.getDebugLoc()) { 1540 PrologEndLoc = MI.getDebugLoc(); 1541 break; 1542 } else if (!MI.isMetaInstruction()) { 1543 EmptyPrologue = false; 1544 } 1545 } 1546 } 1547 1548 // Record beginning of function if we have a non-empty prologue. 1549 if (PrologEndLoc && !EmptyPrologue) { 1550 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1551 maybeRecordLocation(FnStartDL, MF); 1552 } 1553 1554 // Find heap alloc sites and emit labels around them. 1555 for (const auto &MBB : *MF) { 1556 for (const auto &MI : MBB) { 1557 if (MI.getHeapAllocMarker()) { 1558 requestLabelBeforeInsn(&MI); 1559 requestLabelAfterInsn(&MI); 1560 } 1561 } 1562 } 1563 } 1564 1565 static bool shouldEmitUdt(const DIType *T) { 1566 if (!T) 1567 return false; 1568 1569 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1570 if (T->getTag() == dwarf::DW_TAG_typedef) { 1571 if (DIScope *Scope = T->getScope()) { 1572 switch (Scope->getTag()) { 1573 case dwarf::DW_TAG_structure_type: 1574 case dwarf::DW_TAG_class_type: 1575 case dwarf::DW_TAG_union_type: 1576 return false; 1577 default: 1578 // do nothing. 1579 ; 1580 } 1581 } 1582 } 1583 1584 while (true) { 1585 if (!T || T->isForwardDecl()) 1586 return false; 1587 1588 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1589 if (!DT) 1590 return true; 1591 T = DT->getBaseType(); 1592 } 1593 return true; 1594 } 1595 1596 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1597 // Don't record empty UDTs. 1598 if (Ty->getName().empty()) 1599 return; 1600 if (!shouldEmitUdt(Ty)) 1601 return; 1602 1603 SmallVector<StringRef, 5> ParentScopeNames; 1604 const DISubprogram *ClosestSubprogram = 1605 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1606 1607 std::string FullyQualifiedName = 1608 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1609 1610 if (ClosestSubprogram == nullptr) { 1611 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1612 } else if (ClosestSubprogram == CurrentSubprogram) { 1613 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1614 } 1615 1616 // TODO: What if the ClosestSubprogram is neither null or the current 1617 // subprogram? Currently, the UDT just gets dropped on the floor. 1618 // 1619 // The current behavior is not desirable. To get maximal fidelity, we would 1620 // need to perform all type translation before beginning emission of .debug$S 1621 // and then make LocalUDTs a member of FunctionInfo 1622 } 1623 1624 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1625 // Generic dispatch for lowering an unknown type. 1626 switch (Ty->getTag()) { 1627 case dwarf::DW_TAG_array_type: 1628 return lowerTypeArray(cast<DICompositeType>(Ty)); 1629 case dwarf::DW_TAG_typedef: 1630 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1631 case dwarf::DW_TAG_base_type: 1632 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1633 case dwarf::DW_TAG_pointer_type: 1634 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1635 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1636 LLVM_FALLTHROUGH; 1637 case dwarf::DW_TAG_reference_type: 1638 case dwarf::DW_TAG_rvalue_reference_type: 1639 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1640 case dwarf::DW_TAG_ptr_to_member_type: 1641 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1642 case dwarf::DW_TAG_restrict_type: 1643 case dwarf::DW_TAG_const_type: 1644 case dwarf::DW_TAG_volatile_type: 1645 // TODO: add support for DW_TAG_atomic_type here 1646 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1647 case dwarf::DW_TAG_subroutine_type: 1648 if (ClassTy) { 1649 // The member function type of a member function pointer has no 1650 // ThisAdjustment. 1651 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1652 /*ThisAdjustment=*/0, 1653 /*IsStaticMethod=*/false); 1654 } 1655 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1656 case dwarf::DW_TAG_enumeration_type: 1657 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1658 case dwarf::DW_TAG_class_type: 1659 case dwarf::DW_TAG_structure_type: 1660 return lowerTypeClass(cast<DICompositeType>(Ty)); 1661 case dwarf::DW_TAG_union_type: 1662 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1663 case dwarf::DW_TAG_string_type: 1664 return lowerTypeString(cast<DIStringType>(Ty)); 1665 case dwarf::DW_TAG_unspecified_type: 1666 if (Ty->getName() == "decltype(nullptr)") 1667 return TypeIndex::NullptrT(); 1668 return TypeIndex::None(); 1669 default: 1670 // Use the null type index. 1671 return TypeIndex(); 1672 } 1673 } 1674 1675 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1676 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1677 StringRef TypeName = Ty->getName(); 1678 1679 addToUDTs(Ty); 1680 1681 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1682 TypeName == "HRESULT") 1683 return TypeIndex(SimpleTypeKind::HResult); 1684 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1685 TypeName == "wchar_t") 1686 return TypeIndex(SimpleTypeKind::WideCharacter); 1687 1688 return UnderlyingTypeIndex; 1689 } 1690 1691 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1692 const DIType *ElementType = Ty->getBaseType(); 1693 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1694 // IndexType is size_t, which depends on the bitness of the target. 1695 TypeIndex IndexType = getPointerSizeInBytes() == 8 1696 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1697 : TypeIndex(SimpleTypeKind::UInt32Long); 1698 1699 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1700 1701 // Add subranges to array type. 1702 DINodeArray Elements = Ty->getElements(); 1703 for (int i = Elements.size() - 1; i >= 0; --i) { 1704 const DINode *Element = Elements[i]; 1705 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1706 1707 const DISubrange *Subrange = cast<DISubrange>(Element); 1708 int64_t Count = -1; 1709 1710 // If Subrange has a Count field, use it. 1711 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1712 // where lowerbound is from the LowerBound field of the Subrange, 1713 // or the language default lowerbound if that field is unspecified. 1714 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1715 Count = CI->getSExtValue(); 1716 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1717 // Fortran uses 1 as the default lowerbound; other languages use 0. 1718 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1719 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1720 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1721 Count = UI->getSExtValue() - Lowerbound + 1; 1722 } 1723 1724 // Forward declarations of arrays without a size and VLAs use a count of -1. 1725 // Emit a count of zero in these cases to match what MSVC does for arrays 1726 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1727 // should do for them even if we could distinguish them. 1728 if (Count == -1) 1729 Count = 0; 1730 1731 // Update the element size and element type index for subsequent subranges. 1732 ElementSize *= Count; 1733 1734 // If this is the outermost array, use the size from the array. It will be 1735 // more accurate if we had a VLA or an incomplete element type size. 1736 uint64_t ArraySize = 1737 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1738 1739 StringRef Name = (i == 0) ? Ty->getName() : ""; 1740 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1741 ElementTypeIndex = TypeTable.writeLeafType(AR); 1742 } 1743 1744 return ElementTypeIndex; 1745 } 1746 1747 // This function lowers a Fortran character type (DIStringType). 1748 // Note that it handles only the character*n variant (using SizeInBits 1749 // field in DIString to describe the type size) at the moment. 1750 // Other variants (leveraging the StringLength and StringLengthExp 1751 // fields in DIStringType) remain TBD. 1752 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1753 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1754 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1755 StringRef Name = Ty->getName(); 1756 // IndexType is size_t, which depends on the bitness of the target. 1757 TypeIndex IndexType = getPointerSizeInBytes() == 8 1758 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1759 : TypeIndex(SimpleTypeKind::UInt32Long); 1760 1761 // Create a type of character array of ArraySize. 1762 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1763 1764 return TypeTable.writeLeafType(AR); 1765 } 1766 1767 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1768 TypeIndex Index; 1769 dwarf::TypeKind Kind; 1770 uint32_t ByteSize; 1771 1772 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1773 ByteSize = Ty->getSizeInBits() / 8; 1774 1775 SimpleTypeKind STK = SimpleTypeKind::None; 1776 switch (Kind) { 1777 case dwarf::DW_ATE_address: 1778 // FIXME: Translate 1779 break; 1780 case dwarf::DW_ATE_boolean: 1781 switch (ByteSize) { 1782 case 1: STK = SimpleTypeKind::Boolean8; break; 1783 case 2: STK = SimpleTypeKind::Boolean16; break; 1784 case 4: STK = SimpleTypeKind::Boolean32; break; 1785 case 8: STK = SimpleTypeKind::Boolean64; break; 1786 case 16: STK = SimpleTypeKind::Boolean128; break; 1787 } 1788 break; 1789 case dwarf::DW_ATE_complex_float: 1790 switch (ByteSize) { 1791 case 2: STK = SimpleTypeKind::Complex16; break; 1792 case 4: STK = SimpleTypeKind::Complex32; break; 1793 case 8: STK = SimpleTypeKind::Complex64; break; 1794 case 10: STK = SimpleTypeKind::Complex80; break; 1795 case 16: STK = SimpleTypeKind::Complex128; break; 1796 } 1797 break; 1798 case dwarf::DW_ATE_float: 1799 switch (ByteSize) { 1800 case 2: STK = SimpleTypeKind::Float16; break; 1801 case 4: STK = SimpleTypeKind::Float32; break; 1802 case 6: STK = SimpleTypeKind::Float48; break; 1803 case 8: STK = SimpleTypeKind::Float64; break; 1804 case 10: STK = SimpleTypeKind::Float80; break; 1805 case 16: STK = SimpleTypeKind::Float128; break; 1806 } 1807 break; 1808 case dwarf::DW_ATE_signed: 1809 switch (ByteSize) { 1810 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1811 case 2: STK = SimpleTypeKind::Int16Short; break; 1812 case 4: STK = SimpleTypeKind::Int32; break; 1813 case 8: STK = SimpleTypeKind::Int64Quad; break; 1814 case 16: STK = SimpleTypeKind::Int128Oct; break; 1815 } 1816 break; 1817 case dwarf::DW_ATE_unsigned: 1818 switch (ByteSize) { 1819 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1820 case 2: STK = SimpleTypeKind::UInt16Short; break; 1821 case 4: STK = SimpleTypeKind::UInt32; break; 1822 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1823 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1824 } 1825 break; 1826 case dwarf::DW_ATE_UTF: 1827 switch (ByteSize) { 1828 case 2: STK = SimpleTypeKind::Character16; break; 1829 case 4: STK = SimpleTypeKind::Character32; break; 1830 } 1831 break; 1832 case dwarf::DW_ATE_signed_char: 1833 if (ByteSize == 1) 1834 STK = SimpleTypeKind::SignedCharacter; 1835 break; 1836 case dwarf::DW_ATE_unsigned_char: 1837 if (ByteSize == 1) 1838 STK = SimpleTypeKind::UnsignedCharacter; 1839 break; 1840 default: 1841 break; 1842 } 1843 1844 // Apply some fixups based on the source-level type name. 1845 // Include some amount of canonicalization from an old naming scheme Clang 1846 // used to use for integer types (in an outdated effort to be compatible with 1847 // GCC's debug info/GDB's behavior, which has since been addressed). 1848 if (STK == SimpleTypeKind::Int32 && 1849 (Ty->getName() == "long int" || Ty->getName() == "long")) 1850 STK = SimpleTypeKind::Int32Long; 1851 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1852 Ty->getName() == "unsigned long")) 1853 STK = SimpleTypeKind::UInt32Long; 1854 if (STK == SimpleTypeKind::UInt16Short && 1855 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1856 STK = SimpleTypeKind::WideCharacter; 1857 if ((STK == SimpleTypeKind::SignedCharacter || 1858 STK == SimpleTypeKind::UnsignedCharacter) && 1859 Ty->getName() == "char") 1860 STK = SimpleTypeKind::NarrowCharacter; 1861 1862 return TypeIndex(STK); 1863 } 1864 1865 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1866 PointerOptions PO) { 1867 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1868 1869 // Pointers to simple types without any options can use SimpleTypeMode, rather 1870 // than having a dedicated pointer type record. 1871 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1872 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1873 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1874 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1875 ? SimpleTypeMode::NearPointer64 1876 : SimpleTypeMode::NearPointer32; 1877 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1878 } 1879 1880 PointerKind PK = 1881 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1882 PointerMode PM = PointerMode::Pointer; 1883 switch (Ty->getTag()) { 1884 default: llvm_unreachable("not a pointer tag type"); 1885 case dwarf::DW_TAG_pointer_type: 1886 PM = PointerMode::Pointer; 1887 break; 1888 case dwarf::DW_TAG_reference_type: 1889 PM = PointerMode::LValueReference; 1890 break; 1891 case dwarf::DW_TAG_rvalue_reference_type: 1892 PM = PointerMode::RValueReference; 1893 break; 1894 } 1895 1896 if (Ty->isObjectPointer()) 1897 PO |= PointerOptions::Const; 1898 1899 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1900 return TypeTable.writeLeafType(PR); 1901 } 1902 1903 static PointerToMemberRepresentation 1904 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1905 // SizeInBytes being zero generally implies that the member pointer type was 1906 // incomplete, which can happen if it is part of a function prototype. In this 1907 // case, use the unknown model instead of the general model. 1908 if (IsPMF) { 1909 switch (Flags & DINode::FlagPtrToMemberRep) { 1910 case 0: 1911 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1912 : PointerToMemberRepresentation::GeneralFunction; 1913 case DINode::FlagSingleInheritance: 1914 return PointerToMemberRepresentation::SingleInheritanceFunction; 1915 case DINode::FlagMultipleInheritance: 1916 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1917 case DINode::FlagVirtualInheritance: 1918 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1919 } 1920 } else { 1921 switch (Flags & DINode::FlagPtrToMemberRep) { 1922 case 0: 1923 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1924 : PointerToMemberRepresentation::GeneralData; 1925 case DINode::FlagSingleInheritance: 1926 return PointerToMemberRepresentation::SingleInheritanceData; 1927 case DINode::FlagMultipleInheritance: 1928 return PointerToMemberRepresentation::MultipleInheritanceData; 1929 case DINode::FlagVirtualInheritance: 1930 return PointerToMemberRepresentation::VirtualInheritanceData; 1931 } 1932 } 1933 llvm_unreachable("invalid ptr to member representation"); 1934 } 1935 1936 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1937 PointerOptions PO) { 1938 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1939 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1940 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1941 TypeIndex PointeeTI = 1942 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1943 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1944 : PointerKind::Near32; 1945 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1946 : PointerMode::PointerToDataMember; 1947 1948 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1949 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1950 MemberPointerInfo MPI( 1951 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1952 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1953 return TypeTable.writeLeafType(PR); 1954 } 1955 1956 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1957 /// have a translation, use the NearC convention. 1958 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1959 switch (DwarfCC) { 1960 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1961 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1962 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1963 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1964 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1965 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1966 } 1967 return CallingConvention::NearC; 1968 } 1969 1970 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1971 ModifierOptions Mods = ModifierOptions::None; 1972 PointerOptions PO = PointerOptions::None; 1973 bool IsModifier = true; 1974 const DIType *BaseTy = Ty; 1975 while (IsModifier && BaseTy) { 1976 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1977 switch (BaseTy->getTag()) { 1978 case dwarf::DW_TAG_const_type: 1979 Mods |= ModifierOptions::Const; 1980 PO |= PointerOptions::Const; 1981 break; 1982 case dwarf::DW_TAG_volatile_type: 1983 Mods |= ModifierOptions::Volatile; 1984 PO |= PointerOptions::Volatile; 1985 break; 1986 case dwarf::DW_TAG_restrict_type: 1987 // Only pointer types be marked with __restrict. There is no known flag 1988 // for __restrict in LF_MODIFIER records. 1989 PO |= PointerOptions::Restrict; 1990 break; 1991 default: 1992 IsModifier = false; 1993 break; 1994 } 1995 if (IsModifier) 1996 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1997 } 1998 1999 // Check if the inner type will use an LF_POINTER record. If so, the 2000 // qualifiers will go in the LF_POINTER record. This comes up for types like 2001 // 'int *const' and 'int *__restrict', not the more common cases like 'const 2002 // char *'. 2003 if (BaseTy) { 2004 switch (BaseTy->getTag()) { 2005 case dwarf::DW_TAG_pointer_type: 2006 case dwarf::DW_TAG_reference_type: 2007 case dwarf::DW_TAG_rvalue_reference_type: 2008 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2009 case dwarf::DW_TAG_ptr_to_member_type: 2010 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2011 default: 2012 break; 2013 } 2014 } 2015 2016 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2017 2018 // Return the base type index if there aren't any modifiers. For example, the 2019 // metadata could contain restrict wrappers around non-pointer types. 2020 if (Mods == ModifierOptions::None) 2021 return ModifiedTI; 2022 2023 ModifierRecord MR(ModifiedTI, Mods); 2024 return TypeTable.writeLeafType(MR); 2025 } 2026 2027 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2028 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2029 for (const DIType *ArgType : Ty->getTypeArray()) 2030 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2031 2032 // MSVC uses type none for variadic argument. 2033 if (ReturnAndArgTypeIndices.size() > 1 && 2034 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2035 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2036 } 2037 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2038 ArrayRef<TypeIndex> ArgTypeIndices = None; 2039 if (!ReturnAndArgTypeIndices.empty()) { 2040 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 2041 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2042 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2043 } 2044 2045 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2046 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2047 2048 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2049 2050 FunctionOptions FO = getFunctionOptions(Ty); 2051 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2052 ArgListIndex); 2053 return TypeTable.writeLeafType(Procedure); 2054 } 2055 2056 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2057 const DIType *ClassTy, 2058 int ThisAdjustment, 2059 bool IsStaticMethod, 2060 FunctionOptions FO) { 2061 // Lower the containing class type. 2062 TypeIndex ClassType = getTypeIndex(ClassTy); 2063 2064 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2065 2066 unsigned Index = 0; 2067 SmallVector<TypeIndex, 8> ArgTypeIndices; 2068 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2069 if (ReturnAndArgs.size() > Index) { 2070 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2071 } 2072 2073 // If the first argument is a pointer type and this isn't a static method, 2074 // treat it as the special 'this' parameter, which is encoded separately from 2075 // the arguments. 2076 TypeIndex ThisTypeIndex; 2077 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2078 if (const DIDerivedType *PtrTy = 2079 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2080 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2081 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2082 Index++; 2083 } 2084 } 2085 } 2086 2087 while (Index < ReturnAndArgs.size()) 2088 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2089 2090 // MSVC uses type none for variadic argument. 2091 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2092 ArgTypeIndices.back() = TypeIndex::None(); 2093 2094 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2095 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2096 2097 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2098 2099 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2100 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2101 return TypeTable.writeLeafType(MFR); 2102 } 2103 2104 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2105 unsigned VSlotCount = 2106 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2107 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2108 2109 VFTableShapeRecord VFTSR(Slots); 2110 return TypeTable.writeLeafType(VFTSR); 2111 } 2112 2113 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2114 switch (Flags & DINode::FlagAccessibility) { 2115 case DINode::FlagPrivate: return MemberAccess::Private; 2116 case DINode::FlagPublic: return MemberAccess::Public; 2117 case DINode::FlagProtected: return MemberAccess::Protected; 2118 case 0: 2119 // If there was no explicit access control, provide the default for the tag. 2120 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2121 : MemberAccess::Public; 2122 } 2123 llvm_unreachable("access flags are exclusive"); 2124 } 2125 2126 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2127 if (SP->isArtificial()) 2128 return MethodOptions::CompilerGenerated; 2129 2130 // FIXME: Handle other MethodOptions. 2131 2132 return MethodOptions::None; 2133 } 2134 2135 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2136 bool Introduced) { 2137 if (SP->getFlags() & DINode::FlagStaticMember) 2138 return MethodKind::Static; 2139 2140 switch (SP->getVirtuality()) { 2141 case dwarf::DW_VIRTUALITY_none: 2142 break; 2143 case dwarf::DW_VIRTUALITY_virtual: 2144 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2145 case dwarf::DW_VIRTUALITY_pure_virtual: 2146 return Introduced ? MethodKind::PureIntroducingVirtual 2147 : MethodKind::PureVirtual; 2148 default: 2149 llvm_unreachable("unhandled virtuality case"); 2150 } 2151 2152 return MethodKind::Vanilla; 2153 } 2154 2155 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2156 switch (Ty->getTag()) { 2157 case dwarf::DW_TAG_class_type: 2158 return TypeRecordKind::Class; 2159 case dwarf::DW_TAG_structure_type: 2160 return TypeRecordKind::Struct; 2161 default: 2162 llvm_unreachable("unexpected tag"); 2163 } 2164 } 2165 2166 /// Return ClassOptions that should be present on both the forward declaration 2167 /// and the defintion of a tag type. 2168 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2169 ClassOptions CO = ClassOptions::None; 2170 2171 // MSVC always sets this flag, even for local types. Clang doesn't always 2172 // appear to give every type a linkage name, which may be problematic for us. 2173 // FIXME: Investigate the consequences of not following them here. 2174 if (!Ty->getIdentifier().empty()) 2175 CO |= ClassOptions::HasUniqueName; 2176 2177 // Put the Nested flag on a type if it appears immediately inside a tag type. 2178 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2179 // here. That flag is only set on definitions, and not forward declarations. 2180 const DIScope *ImmediateScope = Ty->getScope(); 2181 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2182 CO |= ClassOptions::Nested; 2183 2184 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2185 // type only when it has an immediate function scope. Clang never puts enums 2186 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2187 // always in function, class, or file scopes. 2188 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2189 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2190 CO |= ClassOptions::Scoped; 2191 } else { 2192 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2193 Scope = Scope->getScope()) { 2194 if (isa<DISubprogram>(Scope)) { 2195 CO |= ClassOptions::Scoped; 2196 break; 2197 } 2198 } 2199 } 2200 2201 return CO; 2202 } 2203 2204 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2205 switch (Ty->getTag()) { 2206 case dwarf::DW_TAG_class_type: 2207 case dwarf::DW_TAG_structure_type: 2208 case dwarf::DW_TAG_union_type: 2209 case dwarf::DW_TAG_enumeration_type: 2210 break; 2211 default: 2212 return; 2213 } 2214 2215 if (const auto *File = Ty->getFile()) { 2216 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2217 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2218 2219 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2220 TypeTable.writeLeafType(USLR); 2221 } 2222 } 2223 2224 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2225 ClassOptions CO = getCommonClassOptions(Ty); 2226 TypeIndex FTI; 2227 unsigned EnumeratorCount = 0; 2228 2229 if (Ty->isForwardDecl()) { 2230 CO |= ClassOptions::ForwardReference; 2231 } else { 2232 ContinuationRecordBuilder ContinuationBuilder; 2233 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2234 for (const DINode *Element : Ty->getElements()) { 2235 // We assume that the frontend provides all members in source declaration 2236 // order, which is what MSVC does. 2237 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2238 // FIXME: Is it correct to always emit these as unsigned here? 2239 EnumeratorRecord ER(MemberAccess::Public, 2240 APSInt(Enumerator->getValue(), true), 2241 Enumerator->getName()); 2242 ContinuationBuilder.writeMemberType(ER); 2243 EnumeratorCount++; 2244 } 2245 } 2246 FTI = TypeTable.insertRecord(ContinuationBuilder); 2247 } 2248 2249 std::string FullName = getFullyQualifiedName(Ty); 2250 2251 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2252 getTypeIndex(Ty->getBaseType())); 2253 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2254 2255 addUDTSrcLine(Ty, EnumTI); 2256 2257 return EnumTI; 2258 } 2259 2260 //===----------------------------------------------------------------------===// 2261 // ClassInfo 2262 //===----------------------------------------------------------------------===// 2263 2264 struct llvm::ClassInfo { 2265 struct MemberInfo { 2266 const DIDerivedType *MemberTypeNode; 2267 uint64_t BaseOffset; 2268 }; 2269 // [MemberInfo] 2270 using MemberList = std::vector<MemberInfo>; 2271 2272 using MethodsList = TinyPtrVector<const DISubprogram *>; 2273 // MethodName -> MethodsList 2274 using MethodsMap = MapVector<MDString *, MethodsList>; 2275 2276 /// Base classes. 2277 std::vector<const DIDerivedType *> Inheritance; 2278 2279 /// Direct members. 2280 MemberList Members; 2281 // Direct overloaded methods gathered by name. 2282 MethodsMap Methods; 2283 2284 TypeIndex VShapeTI; 2285 2286 std::vector<const DIType *> NestedTypes; 2287 }; 2288 2289 void CodeViewDebug::clear() { 2290 assert(CurFn == nullptr); 2291 FileIdMap.clear(); 2292 FnDebugInfo.clear(); 2293 FileToFilepathMap.clear(); 2294 LocalUDTs.clear(); 2295 GlobalUDTs.clear(); 2296 TypeIndices.clear(); 2297 CompleteTypeIndices.clear(); 2298 ScopeGlobals.clear(); 2299 CVGlobalVariableOffsets.clear(); 2300 } 2301 2302 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2303 const DIDerivedType *DDTy) { 2304 if (!DDTy->getName().empty()) { 2305 Info.Members.push_back({DDTy, 0}); 2306 2307 // Collect static const data members with values. 2308 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2309 DINode::FlagStaticMember) { 2310 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2311 isa<ConstantFP>(DDTy->getConstant()))) 2312 StaticConstMembers.push_back(DDTy); 2313 } 2314 2315 return; 2316 } 2317 2318 // An unnamed member may represent a nested struct or union. Attempt to 2319 // interpret the unnamed member as a DICompositeType possibly wrapped in 2320 // qualifier types. Add all the indirect fields to the current record if that 2321 // succeeds, and drop the member if that fails. 2322 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2323 uint64_t Offset = DDTy->getOffsetInBits(); 2324 const DIType *Ty = DDTy->getBaseType(); 2325 bool FullyResolved = false; 2326 while (!FullyResolved) { 2327 switch (Ty->getTag()) { 2328 case dwarf::DW_TAG_const_type: 2329 case dwarf::DW_TAG_volatile_type: 2330 // FIXME: we should apply the qualifier types to the indirect fields 2331 // rather than dropping them. 2332 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2333 break; 2334 default: 2335 FullyResolved = true; 2336 break; 2337 } 2338 } 2339 2340 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2341 if (!DCTy) 2342 return; 2343 2344 ClassInfo NestedInfo = collectClassInfo(DCTy); 2345 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2346 Info.Members.push_back( 2347 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2348 } 2349 2350 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2351 ClassInfo Info; 2352 // Add elements to structure type. 2353 DINodeArray Elements = Ty->getElements(); 2354 for (auto *Element : Elements) { 2355 // We assume that the frontend provides all members in source declaration 2356 // order, which is what MSVC does. 2357 if (!Element) 2358 continue; 2359 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2360 Info.Methods[SP->getRawName()].push_back(SP); 2361 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2362 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2363 collectMemberInfo(Info, DDTy); 2364 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2365 Info.Inheritance.push_back(DDTy); 2366 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2367 DDTy->getName() == "__vtbl_ptr_type") { 2368 Info.VShapeTI = getTypeIndex(DDTy); 2369 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2370 Info.NestedTypes.push_back(DDTy); 2371 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2372 // Ignore friend members. It appears that MSVC emitted info about 2373 // friends in the past, but modern versions do not. 2374 } 2375 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2376 Info.NestedTypes.push_back(Composite); 2377 } 2378 // Skip other unrecognized kinds of elements. 2379 } 2380 return Info; 2381 } 2382 2383 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2384 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2385 // if a complete type should be emitted instead of a forward reference. 2386 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2387 !Ty->isForwardDecl(); 2388 } 2389 2390 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2391 // Emit the complete type for unnamed structs. C++ classes with methods 2392 // which have a circular reference back to the class type are expected to 2393 // be named by the front-end and should not be "unnamed". C unnamed 2394 // structs should not have circular references. 2395 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2396 // If this unnamed complete type is already in the process of being defined 2397 // then the description of the type is malformed and cannot be emitted 2398 // into CodeView correctly so report a fatal error. 2399 auto I = CompleteTypeIndices.find(Ty); 2400 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2401 report_fatal_error("cannot debug circular reference to unnamed type"); 2402 return getCompleteTypeIndex(Ty); 2403 } 2404 2405 // First, construct the forward decl. Don't look into Ty to compute the 2406 // forward decl options, since it might not be available in all TUs. 2407 TypeRecordKind Kind = getRecordKind(Ty); 2408 ClassOptions CO = 2409 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2410 std::string FullName = getFullyQualifiedName(Ty); 2411 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2412 FullName, Ty->getIdentifier()); 2413 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2414 if (!Ty->isForwardDecl()) 2415 DeferredCompleteTypes.push_back(Ty); 2416 return FwdDeclTI; 2417 } 2418 2419 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2420 // Construct the field list and complete type record. 2421 TypeRecordKind Kind = getRecordKind(Ty); 2422 ClassOptions CO = getCommonClassOptions(Ty); 2423 TypeIndex FieldTI; 2424 TypeIndex VShapeTI; 2425 unsigned FieldCount; 2426 bool ContainsNestedClass; 2427 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2428 lowerRecordFieldList(Ty); 2429 2430 if (ContainsNestedClass) 2431 CO |= ClassOptions::ContainsNestedClass; 2432 2433 // MSVC appears to set this flag by searching any destructor or method with 2434 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2435 // the members, however special member functions are not yet emitted into 2436 // debug information. For now checking a class's non-triviality seems enough. 2437 // FIXME: not true for a nested unnamed struct. 2438 if (isNonTrivial(Ty)) 2439 CO |= ClassOptions::HasConstructorOrDestructor; 2440 2441 std::string FullName = getFullyQualifiedName(Ty); 2442 2443 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2444 2445 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2446 SizeInBytes, FullName, Ty->getIdentifier()); 2447 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2448 2449 addUDTSrcLine(Ty, ClassTI); 2450 2451 addToUDTs(Ty); 2452 2453 return ClassTI; 2454 } 2455 2456 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2457 // Emit the complete type for unnamed unions. 2458 if (shouldAlwaysEmitCompleteClassType(Ty)) 2459 return getCompleteTypeIndex(Ty); 2460 2461 ClassOptions CO = 2462 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2463 std::string FullName = getFullyQualifiedName(Ty); 2464 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2465 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2466 if (!Ty->isForwardDecl()) 2467 DeferredCompleteTypes.push_back(Ty); 2468 return FwdDeclTI; 2469 } 2470 2471 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2472 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2473 TypeIndex FieldTI; 2474 unsigned FieldCount; 2475 bool ContainsNestedClass; 2476 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2477 lowerRecordFieldList(Ty); 2478 2479 if (ContainsNestedClass) 2480 CO |= ClassOptions::ContainsNestedClass; 2481 2482 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2483 std::string FullName = getFullyQualifiedName(Ty); 2484 2485 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2486 Ty->getIdentifier()); 2487 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2488 2489 addUDTSrcLine(Ty, UnionTI); 2490 2491 addToUDTs(Ty); 2492 2493 return UnionTI; 2494 } 2495 2496 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2497 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2498 // Manually count members. MSVC appears to count everything that generates a 2499 // field list record. Each individual overload in a method overload group 2500 // contributes to this count, even though the overload group is a single field 2501 // list record. 2502 unsigned MemberCount = 0; 2503 ClassInfo Info = collectClassInfo(Ty); 2504 ContinuationRecordBuilder ContinuationBuilder; 2505 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2506 2507 // Create base classes. 2508 for (const DIDerivedType *I : Info.Inheritance) { 2509 if (I->getFlags() & DINode::FlagVirtual) { 2510 // Virtual base. 2511 unsigned VBPtrOffset = I->getVBPtrOffset(); 2512 // FIXME: Despite the accessor name, the offset is really in bytes. 2513 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2514 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2515 ? TypeRecordKind::IndirectVirtualBaseClass 2516 : TypeRecordKind::VirtualBaseClass; 2517 VirtualBaseClassRecord VBCR( 2518 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2519 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2520 VBTableIndex); 2521 2522 ContinuationBuilder.writeMemberType(VBCR); 2523 MemberCount++; 2524 } else { 2525 assert(I->getOffsetInBits() % 8 == 0 && 2526 "bases must be on byte boundaries"); 2527 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2528 getTypeIndex(I->getBaseType()), 2529 I->getOffsetInBits() / 8); 2530 ContinuationBuilder.writeMemberType(BCR); 2531 MemberCount++; 2532 } 2533 } 2534 2535 // Create members. 2536 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2537 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2538 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2539 StringRef MemberName = Member->getName(); 2540 MemberAccess Access = 2541 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2542 2543 if (Member->isStaticMember()) { 2544 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2545 ContinuationBuilder.writeMemberType(SDMR); 2546 MemberCount++; 2547 continue; 2548 } 2549 2550 // Virtual function pointer member. 2551 if ((Member->getFlags() & DINode::FlagArtificial) && 2552 Member->getName().startswith("_vptr$")) { 2553 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2554 ContinuationBuilder.writeMemberType(VFPR); 2555 MemberCount++; 2556 continue; 2557 } 2558 2559 // Data member. 2560 uint64_t MemberOffsetInBits = 2561 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2562 if (Member->isBitField()) { 2563 uint64_t StartBitOffset = MemberOffsetInBits; 2564 if (const auto *CI = 2565 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2566 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2567 } 2568 StartBitOffset -= MemberOffsetInBits; 2569 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2570 StartBitOffset); 2571 MemberBaseType = TypeTable.writeLeafType(BFR); 2572 } 2573 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2574 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2575 MemberName); 2576 ContinuationBuilder.writeMemberType(DMR); 2577 MemberCount++; 2578 } 2579 2580 // Create methods 2581 for (auto &MethodItr : Info.Methods) { 2582 StringRef Name = MethodItr.first->getString(); 2583 2584 std::vector<OneMethodRecord> Methods; 2585 for (const DISubprogram *SP : MethodItr.second) { 2586 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2587 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2588 2589 unsigned VFTableOffset = -1; 2590 if (Introduced) 2591 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2592 2593 Methods.push_back(OneMethodRecord( 2594 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2595 translateMethodKindFlags(SP, Introduced), 2596 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2597 MemberCount++; 2598 } 2599 assert(!Methods.empty() && "Empty methods map entry"); 2600 if (Methods.size() == 1) 2601 ContinuationBuilder.writeMemberType(Methods[0]); 2602 else { 2603 // FIXME: Make this use its own ContinuationBuilder so that 2604 // MethodOverloadList can be split correctly. 2605 MethodOverloadListRecord MOLR(Methods); 2606 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2607 2608 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2609 ContinuationBuilder.writeMemberType(OMR); 2610 } 2611 } 2612 2613 // Create nested classes. 2614 for (const DIType *Nested : Info.NestedTypes) { 2615 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2616 ContinuationBuilder.writeMemberType(R); 2617 MemberCount++; 2618 } 2619 2620 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2621 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2622 !Info.NestedTypes.empty()); 2623 } 2624 2625 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2626 if (!VBPType.getIndex()) { 2627 // Make a 'const int *' type. 2628 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2629 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2630 2631 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2632 : PointerKind::Near32; 2633 PointerMode PM = PointerMode::Pointer; 2634 PointerOptions PO = PointerOptions::None; 2635 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2636 VBPType = TypeTable.writeLeafType(PR); 2637 } 2638 2639 return VBPType; 2640 } 2641 2642 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2643 // The null DIType is the void type. Don't try to hash it. 2644 if (!Ty) 2645 return TypeIndex::Void(); 2646 2647 // Check if we've already translated this type. Don't try to do a 2648 // get-or-create style insertion that caches the hash lookup across the 2649 // lowerType call. It will update the TypeIndices map. 2650 auto I = TypeIndices.find({Ty, ClassTy}); 2651 if (I != TypeIndices.end()) 2652 return I->second; 2653 2654 TypeLoweringScope S(*this); 2655 TypeIndex TI = lowerType(Ty, ClassTy); 2656 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2657 } 2658 2659 codeview::TypeIndex 2660 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2661 const DISubroutineType *SubroutineTy) { 2662 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2663 "this type must be a pointer type"); 2664 2665 PointerOptions Options = PointerOptions::None; 2666 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2667 Options = PointerOptions::LValueRefThisPointer; 2668 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2669 Options = PointerOptions::RValueRefThisPointer; 2670 2671 // Check if we've already translated this type. If there is no ref qualifier 2672 // on the function then we look up this pointer type with no associated class 2673 // so that the TypeIndex for the this pointer can be shared with the type 2674 // index for other pointers to this class type. If there is a ref qualifier 2675 // then we lookup the pointer using the subroutine as the parent type. 2676 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2677 if (I != TypeIndices.end()) 2678 return I->second; 2679 2680 TypeLoweringScope S(*this); 2681 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2682 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2683 } 2684 2685 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2686 PointerRecord PR(getTypeIndex(Ty), 2687 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2688 : PointerKind::Near32, 2689 PointerMode::LValueReference, PointerOptions::None, 2690 Ty->getSizeInBits() / 8); 2691 return TypeTable.writeLeafType(PR); 2692 } 2693 2694 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2695 // The null DIType is the void type. Don't try to hash it. 2696 if (!Ty) 2697 return TypeIndex::Void(); 2698 2699 // Look through typedefs when getting the complete type index. Call 2700 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2701 // emitted only once. 2702 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2703 (void)getTypeIndex(Ty); 2704 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2705 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2706 2707 // If this is a non-record type, the complete type index is the same as the 2708 // normal type index. Just call getTypeIndex. 2709 switch (Ty->getTag()) { 2710 case dwarf::DW_TAG_class_type: 2711 case dwarf::DW_TAG_structure_type: 2712 case dwarf::DW_TAG_union_type: 2713 break; 2714 default: 2715 return getTypeIndex(Ty); 2716 } 2717 2718 const auto *CTy = cast<DICompositeType>(Ty); 2719 2720 TypeLoweringScope S(*this); 2721 2722 // Make sure the forward declaration is emitted first. It's unclear if this 2723 // is necessary, but MSVC does it, and we should follow suit until we can show 2724 // otherwise. 2725 // We only emit a forward declaration for named types. 2726 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2727 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2728 2729 // Just use the forward decl if we don't have complete type info. This 2730 // might happen if the frontend is using modules and expects the complete 2731 // definition to be emitted elsewhere. 2732 if (CTy->isForwardDecl()) 2733 return FwdDeclTI; 2734 } 2735 2736 // Check if we've already translated the complete record type. 2737 // Insert the type with a null TypeIndex to signify that the type is currently 2738 // being lowered. 2739 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2740 if (!InsertResult.second) 2741 return InsertResult.first->second; 2742 2743 TypeIndex TI; 2744 switch (CTy->getTag()) { 2745 case dwarf::DW_TAG_class_type: 2746 case dwarf::DW_TAG_structure_type: 2747 TI = lowerCompleteTypeClass(CTy); 2748 break; 2749 case dwarf::DW_TAG_union_type: 2750 TI = lowerCompleteTypeUnion(CTy); 2751 break; 2752 default: 2753 llvm_unreachable("not a record"); 2754 } 2755 2756 // Update the type index associated with this CompositeType. This cannot 2757 // use the 'InsertResult' iterator above because it is potentially 2758 // invalidated by map insertions which can occur while lowering the class 2759 // type above. 2760 CompleteTypeIndices[CTy] = TI; 2761 return TI; 2762 } 2763 2764 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2765 /// and do this until fixpoint, as each complete record type typically 2766 /// references 2767 /// many other record types. 2768 void CodeViewDebug::emitDeferredCompleteTypes() { 2769 SmallVector<const DICompositeType *, 4> TypesToEmit; 2770 while (!DeferredCompleteTypes.empty()) { 2771 std::swap(DeferredCompleteTypes, TypesToEmit); 2772 for (const DICompositeType *RecordTy : TypesToEmit) 2773 getCompleteTypeIndex(RecordTy); 2774 TypesToEmit.clear(); 2775 } 2776 } 2777 2778 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2779 ArrayRef<LocalVariable> Locals) { 2780 // Get the sorted list of parameters and emit them first. 2781 SmallVector<const LocalVariable *, 6> Params; 2782 for (const LocalVariable &L : Locals) 2783 if (L.DIVar->isParameter()) 2784 Params.push_back(&L); 2785 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2786 return L->DIVar->getArg() < R->DIVar->getArg(); 2787 }); 2788 for (const LocalVariable *L : Params) 2789 emitLocalVariable(FI, *L); 2790 2791 // Next emit all non-parameters in the order that we found them. 2792 for (const LocalVariable &L : Locals) 2793 if (!L.DIVar->isParameter()) 2794 emitLocalVariable(FI, L); 2795 } 2796 2797 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2798 const LocalVariable &Var) { 2799 // LocalSym record, see SymbolRecord.h for more info. 2800 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2801 2802 LocalSymFlags Flags = LocalSymFlags::None; 2803 if (Var.DIVar->isParameter()) 2804 Flags |= LocalSymFlags::IsParameter; 2805 if (Var.DefRanges.empty()) 2806 Flags |= LocalSymFlags::IsOptimizedOut; 2807 2808 OS.AddComment("TypeIndex"); 2809 TypeIndex TI = Var.UseReferenceType 2810 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2811 : getCompleteTypeIndex(Var.DIVar->getType()); 2812 OS.emitInt32(TI.getIndex()); 2813 OS.AddComment("Flags"); 2814 OS.emitInt16(static_cast<uint16_t>(Flags)); 2815 // Truncate the name so we won't overflow the record length field. 2816 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2817 endSymbolRecord(LocalEnd); 2818 2819 // Calculate the on disk prefix of the appropriate def range record. The 2820 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2821 // should be big enough to hold all forms without memory allocation. 2822 SmallString<20> BytePrefix; 2823 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2824 BytePrefix.clear(); 2825 if (DefRange.InMemory) { 2826 int Offset = DefRange.DataOffset; 2827 unsigned Reg = DefRange.CVRegister; 2828 2829 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2830 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2831 // instead. In frames without stack realignment, $T0 will be the CFA. 2832 if (RegisterId(Reg) == RegisterId::ESP) { 2833 Reg = unsigned(RegisterId::VFRAME); 2834 Offset += FI.OffsetAdjustment; 2835 } 2836 2837 // If we can use the chosen frame pointer for the frame and this isn't a 2838 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2839 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2840 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2841 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2842 (bool(Flags & LocalSymFlags::IsParameter) 2843 ? (EncFP == FI.EncodedParamFramePtrReg) 2844 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2845 DefRangeFramePointerRelHeader DRHdr; 2846 DRHdr.Offset = Offset; 2847 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2848 } else { 2849 uint16_t RegRelFlags = 0; 2850 if (DefRange.IsSubfield) { 2851 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2852 (DefRange.StructOffset 2853 << DefRangeRegisterRelSym::OffsetInParentShift); 2854 } 2855 DefRangeRegisterRelHeader DRHdr; 2856 DRHdr.Register = Reg; 2857 DRHdr.Flags = RegRelFlags; 2858 DRHdr.BasePointerOffset = Offset; 2859 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2860 } 2861 } else { 2862 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2863 if (DefRange.IsSubfield) { 2864 DefRangeSubfieldRegisterHeader DRHdr; 2865 DRHdr.Register = DefRange.CVRegister; 2866 DRHdr.MayHaveNoName = 0; 2867 DRHdr.OffsetInParent = DefRange.StructOffset; 2868 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2869 } else { 2870 DefRangeRegisterHeader DRHdr; 2871 DRHdr.Register = DefRange.CVRegister; 2872 DRHdr.MayHaveNoName = 0; 2873 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2874 } 2875 } 2876 } 2877 } 2878 2879 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2880 const FunctionInfo& FI) { 2881 for (LexicalBlock *Block : Blocks) 2882 emitLexicalBlock(*Block, FI); 2883 } 2884 2885 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2886 /// lexical block scope. 2887 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2888 const FunctionInfo& FI) { 2889 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2890 OS.AddComment("PtrParent"); 2891 OS.emitInt32(0); // PtrParent 2892 OS.AddComment("PtrEnd"); 2893 OS.emitInt32(0); // PtrEnd 2894 OS.AddComment("Code size"); 2895 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2896 OS.AddComment("Function section relative address"); 2897 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2898 OS.AddComment("Function section index"); 2899 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2900 OS.AddComment("Lexical block name"); 2901 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2902 endSymbolRecord(RecordEnd); 2903 2904 // Emit variables local to this lexical block. 2905 emitLocalVariableList(FI, Block.Locals); 2906 emitGlobalVariableList(Block.Globals); 2907 2908 // Emit lexical blocks contained within this block. 2909 emitLexicalBlockList(Block.Children, FI); 2910 2911 // Close the lexical block scope. 2912 emitEndSymbolRecord(SymbolKind::S_END); 2913 } 2914 2915 /// Convenience routine for collecting lexical block information for a list 2916 /// of lexical scopes. 2917 void CodeViewDebug::collectLexicalBlockInfo( 2918 SmallVectorImpl<LexicalScope *> &Scopes, 2919 SmallVectorImpl<LexicalBlock *> &Blocks, 2920 SmallVectorImpl<LocalVariable> &Locals, 2921 SmallVectorImpl<CVGlobalVariable> &Globals) { 2922 for (LexicalScope *Scope : Scopes) 2923 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2924 } 2925 2926 /// Populate the lexical blocks and local variable lists of the parent with 2927 /// information about the specified lexical scope. 2928 void CodeViewDebug::collectLexicalBlockInfo( 2929 LexicalScope &Scope, 2930 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2931 SmallVectorImpl<LocalVariable> &ParentLocals, 2932 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2933 if (Scope.isAbstractScope()) 2934 return; 2935 2936 // Gather information about the lexical scope including local variables, 2937 // global variables, and address ranges. 2938 bool IgnoreScope = false; 2939 auto LI = ScopeVariables.find(&Scope); 2940 SmallVectorImpl<LocalVariable> *Locals = 2941 LI != ScopeVariables.end() ? &LI->second : nullptr; 2942 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2943 SmallVectorImpl<CVGlobalVariable> *Globals = 2944 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2945 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2946 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2947 2948 // Ignore lexical scopes which do not contain variables. 2949 if (!Locals && !Globals) 2950 IgnoreScope = true; 2951 2952 // Ignore lexical scopes which are not lexical blocks. 2953 if (!DILB) 2954 IgnoreScope = true; 2955 2956 // Ignore scopes which have too many address ranges to represent in the 2957 // current CodeView format or do not have a valid address range. 2958 // 2959 // For lexical scopes with multiple address ranges you may be tempted to 2960 // construct a single range covering every instruction where the block is 2961 // live and everything in between. Unfortunately, Visual Studio only 2962 // displays variables from the first matching lexical block scope. If the 2963 // first lexical block contains exception handling code or cold code which 2964 // is moved to the bottom of the routine creating a single range covering 2965 // nearly the entire routine, then it will hide all other lexical blocks 2966 // and the variables they contain. 2967 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2968 IgnoreScope = true; 2969 2970 if (IgnoreScope) { 2971 // This scope can be safely ignored and eliminating it will reduce the 2972 // size of the debug information. Be sure to collect any variable and scope 2973 // information from the this scope or any of its children and collapse them 2974 // into the parent scope. 2975 if (Locals) 2976 ParentLocals.append(Locals->begin(), Locals->end()); 2977 if (Globals) 2978 ParentGlobals.append(Globals->begin(), Globals->end()); 2979 collectLexicalBlockInfo(Scope.getChildren(), 2980 ParentBlocks, 2981 ParentLocals, 2982 ParentGlobals); 2983 return; 2984 } 2985 2986 // Create a new CodeView lexical block for this lexical scope. If we've 2987 // seen this DILexicalBlock before then the scope tree is malformed and 2988 // we can handle this gracefully by not processing it a second time. 2989 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2990 if (!BlockInsertion.second) 2991 return; 2992 2993 // Create a lexical block containing the variables and collect the the 2994 // lexical block information for the children. 2995 const InsnRange &Range = Ranges.front(); 2996 assert(Range.first && Range.second); 2997 LexicalBlock &Block = BlockInsertion.first->second; 2998 Block.Begin = getLabelBeforeInsn(Range.first); 2999 Block.End = getLabelAfterInsn(Range.second); 3000 assert(Block.Begin && "missing label for scope begin"); 3001 assert(Block.End && "missing label for scope end"); 3002 Block.Name = DILB->getName(); 3003 if (Locals) 3004 Block.Locals = std::move(*Locals); 3005 if (Globals) 3006 Block.Globals = std::move(*Globals); 3007 ParentBlocks.push_back(&Block); 3008 collectLexicalBlockInfo(Scope.getChildren(), 3009 Block.Children, 3010 Block.Locals, 3011 Block.Globals); 3012 } 3013 3014 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3015 const Function &GV = MF->getFunction(); 3016 assert(FnDebugInfo.count(&GV)); 3017 assert(CurFn == FnDebugInfo[&GV].get()); 3018 3019 collectVariableInfo(GV.getSubprogram()); 3020 3021 // Build the lexical block structure to emit for this routine. 3022 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3023 collectLexicalBlockInfo(*CFS, 3024 CurFn->ChildBlocks, 3025 CurFn->Locals, 3026 CurFn->Globals); 3027 3028 // Clear the scope and variable information from the map which will not be 3029 // valid after we have finished processing this routine. This also prepares 3030 // the map for the subsequent routine. 3031 ScopeVariables.clear(); 3032 3033 // Don't emit anything if we don't have any line tables. 3034 // Thunks are compiler-generated and probably won't have source correlation. 3035 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3036 FnDebugInfo.erase(&GV); 3037 CurFn = nullptr; 3038 return; 3039 } 3040 3041 // Find heap alloc sites and add to list. 3042 for (const auto &MBB : *MF) { 3043 for (const auto &MI : MBB) { 3044 if (MDNode *MD = MI.getHeapAllocMarker()) { 3045 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3046 getLabelAfterInsn(&MI), 3047 dyn_cast<DIType>(MD))); 3048 } 3049 } 3050 } 3051 3052 CurFn->Annotations = MF->getCodeViewAnnotations(); 3053 3054 CurFn->End = Asm->getFunctionEnd(); 3055 3056 CurFn = nullptr; 3057 } 3058 3059 // Usable locations are valid with non-zero line numbers. A line number of zero 3060 // corresponds to optimized code that doesn't have a distinct source location. 3061 // In this case, we try to use the previous or next source location depending on 3062 // the context. 3063 static bool isUsableDebugLoc(DebugLoc DL) { 3064 return DL && DL.getLine() != 0; 3065 } 3066 3067 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3068 DebugHandlerBase::beginInstruction(MI); 3069 3070 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3071 if (!Asm || !CurFn || MI->isDebugInstr() || 3072 MI->getFlag(MachineInstr::FrameSetup)) 3073 return; 3074 3075 // If the first instruction of a new MBB has no location, find the first 3076 // instruction with a location and use that. 3077 DebugLoc DL = MI->getDebugLoc(); 3078 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3079 for (const auto &NextMI : *MI->getParent()) { 3080 if (NextMI.isDebugInstr()) 3081 continue; 3082 DL = NextMI.getDebugLoc(); 3083 if (isUsableDebugLoc(DL)) 3084 break; 3085 } 3086 // FIXME: Handle the case where the BB has no valid locations. This would 3087 // probably require doing a real dataflow analysis. 3088 } 3089 PrevInstBB = MI->getParent(); 3090 3091 // If we still don't have a debug location, don't record a location. 3092 if (!isUsableDebugLoc(DL)) 3093 return; 3094 3095 maybeRecordLocation(DL, Asm->MF); 3096 } 3097 3098 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3099 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3100 *EndLabel = MMI->getContext().createTempSymbol(); 3101 OS.emitInt32(unsigned(Kind)); 3102 OS.AddComment("Subsection size"); 3103 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3104 OS.emitLabel(BeginLabel); 3105 return EndLabel; 3106 } 3107 3108 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3109 OS.emitLabel(EndLabel); 3110 // Every subsection must be aligned to a 4-byte boundary. 3111 OS.emitValueToAlignment(4); 3112 } 3113 3114 static StringRef getSymbolName(SymbolKind SymKind) { 3115 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3116 if (EE.Value == SymKind) 3117 return EE.Name; 3118 return ""; 3119 } 3120 3121 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3122 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3123 *EndLabel = MMI->getContext().createTempSymbol(); 3124 OS.AddComment("Record length"); 3125 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3126 OS.emitLabel(BeginLabel); 3127 if (OS.isVerboseAsm()) 3128 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3129 OS.emitInt16(unsigned(SymKind)); 3130 return EndLabel; 3131 } 3132 3133 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3134 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3135 // an extra copy of every symbol record in LLD. This increases object file 3136 // size by less than 1% in the clang build, and is compatible with the Visual 3137 // C++ linker. 3138 OS.emitValueToAlignment(4); 3139 OS.emitLabel(SymEnd); 3140 } 3141 3142 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3143 OS.AddComment("Record length"); 3144 OS.emitInt16(2); 3145 if (OS.isVerboseAsm()) 3146 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3147 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3148 } 3149 3150 void CodeViewDebug::emitDebugInfoForUDTs( 3151 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3152 #ifndef NDEBUG 3153 size_t OriginalSize = UDTs.size(); 3154 #endif 3155 for (const auto &UDT : UDTs) { 3156 const DIType *T = UDT.second; 3157 assert(shouldEmitUdt(T)); 3158 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3159 OS.AddComment("Type"); 3160 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3161 assert(OriginalSize == UDTs.size() && 3162 "getCompleteTypeIndex found new UDTs!"); 3163 emitNullTerminatedSymbolName(OS, UDT.first); 3164 endSymbolRecord(UDTRecordEnd); 3165 } 3166 } 3167 3168 void CodeViewDebug::collectGlobalVariableInfo() { 3169 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3170 GlobalMap; 3171 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3172 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3173 GV.getDebugInfo(GVEs); 3174 for (const auto *GVE : GVEs) 3175 GlobalMap[GVE] = &GV; 3176 } 3177 3178 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3179 for (const MDNode *Node : CUs->operands()) { 3180 const auto *CU = cast<DICompileUnit>(Node); 3181 for (const auto *GVE : CU->getGlobalVariables()) { 3182 const DIGlobalVariable *DIGV = GVE->getVariable(); 3183 const DIExpression *DIE = GVE->getExpression(); 3184 3185 if ((DIE->getNumElements() == 2) && 3186 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3187 // Record the constant offset for the variable. 3188 // 3189 // A Fortran common block uses this idiom to encode the offset 3190 // of a variable from the common block's starting address. 3191 CVGlobalVariableOffsets.insert( 3192 std::make_pair(DIGV, DIE->getElement(1))); 3193 3194 // Emit constant global variables in a global symbol section. 3195 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3196 CVGlobalVariable CVGV = {DIGV, DIE}; 3197 GlobalVariables.emplace_back(std::move(CVGV)); 3198 } 3199 3200 const auto *GV = GlobalMap.lookup(GVE); 3201 if (!GV || GV->isDeclarationForLinker()) 3202 continue; 3203 3204 DIScope *Scope = DIGV->getScope(); 3205 SmallVector<CVGlobalVariable, 1> *VariableList; 3206 if (Scope && isa<DILocalScope>(Scope)) { 3207 // Locate a global variable list for this scope, creating one if 3208 // necessary. 3209 auto Insertion = ScopeGlobals.insert( 3210 {Scope, std::unique_ptr<GlobalVariableList>()}); 3211 if (Insertion.second) 3212 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3213 VariableList = Insertion.first->second.get(); 3214 } else if (GV->hasComdat()) 3215 // Emit this global variable into a COMDAT section. 3216 VariableList = &ComdatVariables; 3217 else 3218 // Emit this global variable in a single global symbol section. 3219 VariableList = &GlobalVariables; 3220 CVGlobalVariable CVGV = {DIGV, GV}; 3221 VariableList->emplace_back(std::move(CVGV)); 3222 } 3223 } 3224 } 3225 3226 void CodeViewDebug::collectDebugInfoForGlobals() { 3227 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3228 const DIGlobalVariable *DIGV = CVGV.DIGV; 3229 const DIScope *Scope = DIGV->getScope(); 3230 getCompleteTypeIndex(DIGV->getType()); 3231 getFullyQualifiedName(Scope, DIGV->getName()); 3232 } 3233 3234 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3235 const DIGlobalVariable *DIGV = CVGV.DIGV; 3236 const DIScope *Scope = DIGV->getScope(); 3237 getCompleteTypeIndex(DIGV->getType()); 3238 getFullyQualifiedName(Scope, DIGV->getName()); 3239 } 3240 } 3241 3242 void CodeViewDebug::emitDebugInfoForGlobals() { 3243 // First, emit all globals that are not in a comdat in a single symbol 3244 // substream. MSVC doesn't like it if the substream is empty, so only open 3245 // it if we have at least one global to emit. 3246 switchToDebugSectionForSymbol(nullptr); 3247 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3248 OS.AddComment("Symbol subsection for globals"); 3249 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3250 emitGlobalVariableList(GlobalVariables); 3251 emitStaticConstMemberList(); 3252 endCVSubsection(EndLabel); 3253 } 3254 3255 // Second, emit each global that is in a comdat into its own .debug$S 3256 // section along with its own symbol substream. 3257 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3258 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3259 MCSymbol *GVSym = Asm->getSymbol(GV); 3260 OS.AddComment("Symbol subsection for " + 3261 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3262 switchToDebugSectionForSymbol(GVSym); 3263 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3264 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3265 emitDebugInfoForGlobal(CVGV); 3266 endCVSubsection(EndLabel); 3267 } 3268 } 3269 3270 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3271 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3272 for (const MDNode *Node : CUs->operands()) { 3273 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3274 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3275 getTypeIndex(RT); 3276 // FIXME: Add to global/local DTU list. 3277 } 3278 } 3279 } 3280 } 3281 3282 // Emit each global variable in the specified array. 3283 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3284 for (const CVGlobalVariable &CVGV : Globals) { 3285 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3286 emitDebugInfoForGlobal(CVGV); 3287 } 3288 } 3289 3290 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3291 const std::string &QualifiedName) { 3292 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3293 OS.AddComment("Type"); 3294 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3295 3296 OS.AddComment("Value"); 3297 3298 // Encoded integers shouldn't need more than 10 bytes. 3299 uint8_t Data[10]; 3300 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3301 CodeViewRecordIO IO(Writer); 3302 cantFail(IO.mapEncodedInteger(Value)); 3303 StringRef SRef((char *)Data, Writer.getOffset()); 3304 OS.emitBinaryData(SRef); 3305 3306 OS.AddComment("Name"); 3307 emitNullTerminatedSymbolName(OS, QualifiedName); 3308 endSymbolRecord(SConstantEnd); 3309 } 3310 3311 void CodeViewDebug::emitStaticConstMemberList() { 3312 for (const DIDerivedType *DTy : StaticConstMembers) { 3313 const DIScope *Scope = DTy->getScope(); 3314 3315 APSInt Value; 3316 if (const ConstantInt *CI = 3317 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3318 Value = APSInt(CI->getValue(), 3319 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3320 else if (const ConstantFP *CFP = 3321 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3322 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3323 else 3324 llvm_unreachable("cannot emit a constant without a value"); 3325 3326 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3327 getFullyQualifiedName(Scope, DTy->getName())); 3328 } 3329 } 3330 3331 static bool isFloatDIType(const DIType *Ty) { 3332 if (isa<DICompositeType>(Ty)) 3333 return false; 3334 3335 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3336 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3337 if (T == dwarf::DW_TAG_pointer_type || 3338 T == dwarf::DW_TAG_ptr_to_member_type || 3339 T == dwarf::DW_TAG_reference_type || 3340 T == dwarf::DW_TAG_rvalue_reference_type) 3341 return false; 3342 assert(DTy->getBaseType() && "Expected valid base type"); 3343 return isFloatDIType(DTy->getBaseType()); 3344 } 3345 3346 auto *BTy = cast<DIBasicType>(Ty); 3347 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3348 } 3349 3350 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3351 const DIGlobalVariable *DIGV = CVGV.DIGV; 3352 3353 const DIScope *Scope = DIGV->getScope(); 3354 // For static data members, get the scope from the declaration. 3355 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3356 DIGV->getRawStaticDataMemberDeclaration())) 3357 Scope = MemberDecl->getScope(); 3358 // For Fortran, the scoping portion is elided in its name so that we can 3359 // reference the variable in the command line of the VS debugger. 3360 std::string QualifiedName = 3361 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3362 : getFullyQualifiedName(Scope, DIGV->getName()); 3363 3364 if (const GlobalVariable *GV = 3365 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3366 // DataSym record, see SymbolRecord.h for more info. Thread local data 3367 // happens to have the same format as global data. 3368 MCSymbol *GVSym = Asm->getSymbol(GV); 3369 SymbolKind DataSym = GV->isThreadLocal() 3370 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3371 : SymbolKind::S_GTHREAD32) 3372 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3373 : SymbolKind::S_GDATA32); 3374 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3375 OS.AddComment("Type"); 3376 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3377 OS.AddComment("DataOffset"); 3378 3379 uint64_t Offset = 0; 3380 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3381 // Use the offset seen while collecting info on globals. 3382 Offset = CVGlobalVariableOffsets[DIGV]; 3383 OS.EmitCOFFSecRel32(GVSym, Offset); 3384 3385 OS.AddComment("Segment"); 3386 OS.EmitCOFFSectionIndex(GVSym); 3387 OS.AddComment("Name"); 3388 const unsigned LengthOfDataRecord = 12; 3389 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3390 endSymbolRecord(DataEnd); 3391 } else { 3392 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3393 assert(DIE->isConstant() && 3394 "Global constant variables must contain a constant expression."); 3395 3396 // Use unsigned for floats. 3397 bool isUnsigned = isFloatDIType(DIGV->getType()) 3398 ? true 3399 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3400 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3401 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3402 } 3403 } 3404