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 V.Part[N] = 830 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 831 } else if (C == '.') { 832 ++N; 833 if (N >= 4) 834 return V; 835 } else if (N > 0) 836 return V; 837 } 838 return V; 839 } 840 841 void CodeViewDebug::emitCompilerInformation() { 842 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 843 uint32_t Flags = 0; 844 845 // The low byte of the flags indicates the source language. 846 Flags = CurrentSourceLanguage; 847 // TODO: Figure out which other flags need to be set. 848 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 849 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 850 } 851 using ArchType = llvm::Triple::ArchType; 852 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 853 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 854 Arch == ArchType::aarch64) { 855 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 856 } 857 858 OS.AddComment("Flags and language"); 859 OS.emitInt32(Flags); 860 861 OS.AddComment("CPUType"); 862 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 863 864 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 865 const MDNode *Node = *CUs->operands().begin(); 866 const auto *CU = cast<DICompileUnit>(Node); 867 868 StringRef CompilerVersion = CU->getProducer(); 869 Version FrontVer = parseVersion(CompilerVersion); 870 OS.AddComment("Frontend version"); 871 for (int N : FrontVer.Part) { 872 OS.emitInt16(N); 873 } 874 875 // Some Microsoft tools, like Binscope, expect a backend version number of at 876 // least 8.something, so we'll coerce the LLVM version into a form that 877 // guarantees it'll be big enough without really lying about the version. 878 int Major = 1000 * LLVM_VERSION_MAJOR + 879 10 * LLVM_VERSION_MINOR + 880 LLVM_VERSION_PATCH; 881 // Clamp it for builds that use unusually large version numbers. 882 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 883 Version BackVer = {{ Major, 0, 0, 0 }}; 884 OS.AddComment("Backend version"); 885 for (int N : BackVer.Part) 886 OS.emitInt16(N); 887 888 OS.AddComment("Null-terminated compiler version string"); 889 emitNullTerminatedSymbolName(OS, CompilerVersion); 890 891 endSymbolRecord(CompilerEnd); 892 } 893 894 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 895 StringRef S) { 896 StringIdRecord SIR(TypeIndex(0x0), S); 897 return TypeTable.writeLeafType(SIR); 898 } 899 900 static std::string flattenCommandLine(ArrayRef<std::string> Args, 901 StringRef MainFilename) { 902 std::string FlatCmdLine; 903 raw_string_ostream OS(FlatCmdLine); 904 bool PrintedOneArg = false; 905 if (!StringRef(Args[0]).contains("-cc1")) { 906 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 907 PrintedOneArg = true; 908 } 909 for (unsigned i = 0; i < Args.size(); i++) { 910 StringRef Arg = Args[i]; 911 if (Arg.empty()) 912 continue; 913 if (Arg == "-main-file-name" || Arg == "-o") { 914 i++; // Skip this argument and next one. 915 continue; 916 } 917 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 918 continue; 919 if (PrintedOneArg) 920 OS << " "; 921 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 922 PrintedOneArg = true; 923 } 924 OS.flush(); 925 return FlatCmdLine; 926 } 927 928 void CodeViewDebug::emitBuildInfo() { 929 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 930 // build info. The known prefix is: 931 // - Absolute path of current directory 932 // - Compiler path 933 // - Main source file path, relative to CWD or absolute 934 // - Type server PDB file 935 // - Canonical compiler command line 936 // If frontend and backend compilation are separated (think llc or LTO), it's 937 // not clear if the compiler path should refer to the executable for the 938 // frontend or the backend. Leave it blank for now. 939 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 940 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 941 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 942 const auto *CU = cast<DICompileUnit>(Node); 943 const DIFile *MainSourceFile = CU->getFile(); 944 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 945 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 946 BuildInfoArgs[BuildInfoRecord::SourceFile] = 947 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 948 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 949 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 950 getStringIdTypeIdx(TypeTable, ""); 951 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 952 BuildInfoArgs[BuildInfoRecord::BuildTool] = 953 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 954 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 955 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 956 MainSourceFile->getFilename())); 957 } 958 BuildInfoRecord BIR(BuildInfoArgs); 959 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 960 961 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 962 // from the module symbols into the type stream. 963 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 964 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 965 OS.AddComment("LF_BUILDINFO index"); 966 OS.emitInt32(BuildInfoIndex.getIndex()); 967 endSymbolRecord(BIEnd); 968 endCVSubsection(BISubsecEnd); 969 } 970 971 void CodeViewDebug::emitInlineeLinesSubsection() { 972 if (InlinedSubprograms.empty()) 973 return; 974 975 OS.AddComment("Inlinee lines subsection"); 976 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 977 978 // We emit the checksum info for files. This is used by debuggers to 979 // determine if a pdb matches the source before loading it. Visual Studio, 980 // for instance, will display a warning that the breakpoints are not valid if 981 // the pdb does not match the source. 982 OS.AddComment("Inlinee lines signature"); 983 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 984 985 for (const DISubprogram *SP : InlinedSubprograms) { 986 assert(TypeIndices.count({SP, nullptr})); 987 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 988 989 OS.AddBlankLine(); 990 unsigned FileId = maybeRecordFile(SP->getFile()); 991 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 992 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 993 OS.AddBlankLine(); 994 OS.AddComment("Type index of inlined function"); 995 OS.emitInt32(InlineeIdx.getIndex()); 996 OS.AddComment("Offset into filechecksum table"); 997 OS.emitCVFileChecksumOffsetDirective(FileId); 998 OS.AddComment("Starting line number"); 999 OS.emitInt32(SP->getLine()); 1000 } 1001 1002 endCVSubsection(InlineEnd); 1003 } 1004 1005 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 1006 const DILocation *InlinedAt, 1007 const InlineSite &Site) { 1008 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1009 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1010 1011 // SymbolRecord 1012 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1013 1014 OS.AddComment("PtrParent"); 1015 OS.emitInt32(0); 1016 OS.AddComment("PtrEnd"); 1017 OS.emitInt32(0); 1018 OS.AddComment("Inlinee type index"); 1019 OS.emitInt32(InlineeIdx.getIndex()); 1020 1021 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1022 unsigned StartLineNum = Site.Inlinee->getLine(); 1023 1024 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1025 FI.Begin, FI.End); 1026 1027 endSymbolRecord(InlineEnd); 1028 1029 emitLocalVariableList(FI, Site.InlinedLocals); 1030 1031 // Recurse on child inlined call sites before closing the scope. 1032 for (const DILocation *ChildSite : Site.ChildSites) { 1033 auto I = FI.InlineSites.find(ChildSite); 1034 assert(I != FI.InlineSites.end() && 1035 "child site not in function inline site map"); 1036 emitInlinedCallSite(FI, ChildSite, I->second); 1037 } 1038 1039 // Close the scope. 1040 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1041 } 1042 1043 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1044 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1045 // comdat key. A section may be comdat because of -ffunction-sections or 1046 // because it is comdat in the IR. 1047 MCSectionCOFF *GVSec = 1048 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1049 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1050 1051 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1052 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1053 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1054 1055 OS.SwitchSection(DebugSec); 1056 1057 // Emit the magic version number if this is the first time we've switched to 1058 // this section. 1059 if (ComdatDebugSections.insert(DebugSec).second) 1060 emitCodeViewMagicVersion(); 1061 } 1062 1063 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1064 // The only supported thunk ordinal is currently the standard type. 1065 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1066 FunctionInfo &FI, 1067 const MCSymbol *Fn) { 1068 std::string FuncName = 1069 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1070 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1071 1072 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1073 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1074 1075 // Emit S_THUNK32 1076 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1077 OS.AddComment("PtrParent"); 1078 OS.emitInt32(0); 1079 OS.AddComment("PtrEnd"); 1080 OS.emitInt32(0); 1081 OS.AddComment("PtrNext"); 1082 OS.emitInt32(0); 1083 OS.AddComment("Thunk section relative address"); 1084 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1085 OS.AddComment("Thunk section index"); 1086 OS.EmitCOFFSectionIndex(Fn); 1087 OS.AddComment("Code size"); 1088 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1089 OS.AddComment("Ordinal"); 1090 OS.emitInt8(unsigned(ordinal)); 1091 OS.AddComment("Function name"); 1092 emitNullTerminatedSymbolName(OS, FuncName); 1093 // Additional fields specific to the thunk ordinal would go here. 1094 endSymbolRecord(ThunkRecordEnd); 1095 1096 // Local variables/inlined routines are purposely omitted here. The point of 1097 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1098 1099 // Emit S_PROC_ID_END 1100 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1101 1102 endCVSubsection(SymbolsEnd); 1103 } 1104 1105 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1106 FunctionInfo &FI) { 1107 // For each function there is a separate subsection which holds the PC to 1108 // file:line table. 1109 const MCSymbol *Fn = Asm->getSymbol(GV); 1110 assert(Fn); 1111 1112 // Switch to the to a comdat section, if appropriate. 1113 switchToDebugSectionForSymbol(Fn); 1114 1115 std::string FuncName; 1116 auto *SP = GV->getSubprogram(); 1117 assert(SP); 1118 setCurrentSubprogram(SP); 1119 1120 if (SP->isThunk()) { 1121 emitDebugInfoForThunk(GV, FI, Fn); 1122 return; 1123 } 1124 1125 // If we have a display name, build the fully qualified name by walking the 1126 // chain of scopes. 1127 if (!SP->getName().empty()) 1128 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1129 1130 // If our DISubprogram name is empty, use the mangled name. 1131 if (FuncName.empty()) 1132 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1133 1134 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1135 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1136 OS.EmitCVFPOData(Fn); 1137 1138 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1139 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1140 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1141 { 1142 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1143 : SymbolKind::S_GPROC32_ID; 1144 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1145 1146 // These fields are filled in by tools like CVPACK which run after the fact. 1147 OS.AddComment("PtrParent"); 1148 OS.emitInt32(0); 1149 OS.AddComment("PtrEnd"); 1150 OS.emitInt32(0); 1151 OS.AddComment("PtrNext"); 1152 OS.emitInt32(0); 1153 // This is the important bit that tells the debugger where the function 1154 // code is located and what's its size: 1155 OS.AddComment("Code size"); 1156 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1157 OS.AddComment("Offset after prologue"); 1158 OS.emitInt32(0); 1159 OS.AddComment("Offset before epilogue"); 1160 OS.emitInt32(0); 1161 OS.AddComment("Function type index"); 1162 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1163 OS.AddComment("Function section relative address"); 1164 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1165 OS.AddComment("Function section index"); 1166 OS.EmitCOFFSectionIndex(Fn); 1167 OS.AddComment("Flags"); 1168 OS.emitInt8(0); 1169 // Emit the function display name as a null-terminated string. 1170 OS.AddComment("Function name"); 1171 // Truncate the name so we won't overflow the record length field. 1172 emitNullTerminatedSymbolName(OS, FuncName); 1173 endSymbolRecord(ProcRecordEnd); 1174 1175 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1176 // Subtract out the CSR size since MSVC excludes that and we include it. 1177 OS.AddComment("FrameSize"); 1178 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1179 OS.AddComment("Padding"); 1180 OS.emitInt32(0); 1181 OS.AddComment("Offset of padding"); 1182 OS.emitInt32(0); 1183 OS.AddComment("Bytes of callee saved registers"); 1184 OS.emitInt32(FI.CSRSize); 1185 OS.AddComment("Exception handler offset"); 1186 OS.emitInt32(0); 1187 OS.AddComment("Exception handler section"); 1188 OS.emitInt16(0); 1189 OS.AddComment("Flags (defines frame register)"); 1190 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1191 endSymbolRecord(FrameProcEnd); 1192 1193 emitLocalVariableList(FI, FI.Locals); 1194 emitGlobalVariableList(FI.Globals); 1195 emitLexicalBlockList(FI.ChildBlocks, FI); 1196 1197 // Emit inlined call site information. Only emit functions inlined directly 1198 // into the parent function. We'll emit the other sites recursively as part 1199 // of their parent inline site. 1200 for (const DILocation *InlinedAt : FI.ChildSites) { 1201 auto I = FI.InlineSites.find(InlinedAt); 1202 assert(I != FI.InlineSites.end() && 1203 "child site not in function inline site map"); 1204 emitInlinedCallSite(FI, InlinedAt, I->second); 1205 } 1206 1207 for (auto Annot : FI.Annotations) { 1208 MCSymbol *Label = Annot.first; 1209 MDTuple *Strs = cast<MDTuple>(Annot.second); 1210 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1211 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1212 // FIXME: Make sure we don't overflow the max record size. 1213 OS.EmitCOFFSectionIndex(Label); 1214 OS.emitInt16(Strs->getNumOperands()); 1215 for (Metadata *MD : Strs->operands()) { 1216 // MDStrings are null terminated, so we can do EmitBytes and get the 1217 // nice .asciz directive. 1218 StringRef Str = cast<MDString>(MD)->getString(); 1219 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1220 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1221 } 1222 endSymbolRecord(AnnotEnd); 1223 } 1224 1225 for (auto HeapAllocSite : FI.HeapAllocSites) { 1226 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1227 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1228 const DIType *DITy = std::get<2>(HeapAllocSite); 1229 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1230 OS.AddComment("Call site offset"); 1231 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1232 OS.AddComment("Call site section index"); 1233 OS.EmitCOFFSectionIndex(BeginLabel); 1234 OS.AddComment("Call instruction length"); 1235 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1236 OS.AddComment("Type index"); 1237 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1238 endSymbolRecord(HeapAllocEnd); 1239 } 1240 1241 if (SP != nullptr) 1242 emitDebugInfoForUDTs(LocalUDTs); 1243 1244 // We're done with this function. 1245 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1246 } 1247 endCVSubsection(SymbolsEnd); 1248 1249 // We have an assembler directive that takes care of the whole line table. 1250 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1251 } 1252 1253 CodeViewDebug::LocalVarDefRange 1254 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1255 LocalVarDefRange DR; 1256 DR.InMemory = -1; 1257 DR.DataOffset = Offset; 1258 assert(DR.DataOffset == Offset && "truncation"); 1259 DR.IsSubfield = 0; 1260 DR.StructOffset = 0; 1261 DR.CVRegister = CVRegister; 1262 return DR; 1263 } 1264 1265 void CodeViewDebug::collectVariableInfoFromMFTable( 1266 DenseSet<InlinedEntity> &Processed) { 1267 const MachineFunction &MF = *Asm->MF; 1268 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1269 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1270 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1271 1272 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1273 if (!VI.Var) 1274 continue; 1275 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1276 "Expected inlined-at fields to agree"); 1277 1278 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1279 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1280 1281 // If variable scope is not found then skip this variable. 1282 if (!Scope) 1283 continue; 1284 1285 // If the variable has an attached offset expression, extract it. 1286 // FIXME: Try to handle DW_OP_deref as well. 1287 int64_t ExprOffset = 0; 1288 bool Deref = false; 1289 if (VI.Expr) { 1290 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1291 if (VI.Expr->getNumElements() == 1 && 1292 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1293 Deref = true; 1294 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1295 continue; 1296 } 1297 1298 // Get the frame register used and the offset. 1299 Register FrameReg; 1300 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1301 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1302 1303 assert(!FrameOffset.getScalable() && 1304 "Frame offsets with a scalable component are not supported"); 1305 1306 // Calculate the label ranges. 1307 LocalVarDefRange DefRange = 1308 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1309 1310 for (const InsnRange &Range : Scope->getRanges()) { 1311 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1312 const MCSymbol *End = getLabelAfterInsn(Range.second); 1313 End = End ? End : Asm->getFunctionEnd(); 1314 DefRange.Ranges.emplace_back(Begin, End); 1315 } 1316 1317 LocalVariable Var; 1318 Var.DIVar = VI.Var; 1319 Var.DefRanges.emplace_back(std::move(DefRange)); 1320 if (Deref) 1321 Var.UseReferenceType = true; 1322 1323 recordLocalVariable(std::move(Var), Scope); 1324 } 1325 } 1326 1327 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1328 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1329 } 1330 1331 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1332 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1333 } 1334 1335 void CodeViewDebug::calculateRanges( 1336 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1337 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1338 1339 // Calculate the definition ranges. 1340 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1341 const auto &Entry = *I; 1342 if (!Entry.isDbgValue()) 1343 continue; 1344 const MachineInstr *DVInst = Entry.getInstr(); 1345 assert(DVInst->isDebugValue() && "Invalid History entry"); 1346 // FIXME: Find a way to represent constant variables, since they are 1347 // relatively common. 1348 Optional<DbgVariableLocation> Location = 1349 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1350 if (!Location) 1351 continue; 1352 1353 // CodeView can only express variables in register and variables in memory 1354 // at a constant offset from a register. However, for variables passed 1355 // indirectly by pointer, it is common for that pointer to be spilled to a 1356 // stack location. For the special case of one offseted load followed by a 1357 // zero offset load (a pointer spilled to the stack), we change the type of 1358 // the local variable from a value type to a reference type. This tricks the 1359 // debugger into doing the load for us. 1360 if (Var.UseReferenceType) { 1361 // We're using a reference type. Drop the last zero offset load. 1362 if (canUseReferenceType(*Location)) 1363 Location->LoadChain.pop_back(); 1364 else 1365 continue; 1366 } else if (needsReferenceType(*Location)) { 1367 // This location can't be expressed without switching to a reference type. 1368 // Start over using that. 1369 Var.UseReferenceType = true; 1370 Var.DefRanges.clear(); 1371 calculateRanges(Var, Entries); 1372 return; 1373 } 1374 1375 // We can only handle a register or an offseted load of a register. 1376 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1377 continue; 1378 { 1379 LocalVarDefRange DR; 1380 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1381 DR.InMemory = !Location->LoadChain.empty(); 1382 DR.DataOffset = 1383 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1384 if (Location->FragmentInfo) { 1385 DR.IsSubfield = true; 1386 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1387 } else { 1388 DR.IsSubfield = false; 1389 DR.StructOffset = 0; 1390 } 1391 1392 if (Var.DefRanges.empty() || 1393 Var.DefRanges.back().isDifferentLocation(DR)) { 1394 Var.DefRanges.emplace_back(std::move(DR)); 1395 } 1396 } 1397 1398 // Compute the label range. 1399 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1400 const MCSymbol *End; 1401 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1402 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1403 End = EndingEntry.isDbgValue() 1404 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1405 : getLabelAfterInsn(EndingEntry.getInstr()); 1406 } else 1407 End = Asm->getFunctionEnd(); 1408 1409 // If the last range end is our begin, just extend the last range. 1410 // Otherwise make a new range. 1411 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1412 Var.DefRanges.back().Ranges; 1413 if (!R.empty() && R.back().second == Begin) 1414 R.back().second = End; 1415 else 1416 R.emplace_back(Begin, End); 1417 1418 // FIXME: Do more range combining. 1419 } 1420 } 1421 1422 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1423 DenseSet<InlinedEntity> Processed; 1424 // Grab the variable info that was squirreled away in the MMI side-table. 1425 collectVariableInfoFromMFTable(Processed); 1426 1427 for (const auto &I : DbgValues) { 1428 InlinedEntity IV = I.first; 1429 if (Processed.count(IV)) 1430 continue; 1431 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1432 const DILocation *InlinedAt = IV.second; 1433 1434 // Instruction ranges, specifying where IV is accessible. 1435 const auto &Entries = I.second; 1436 1437 LexicalScope *Scope = nullptr; 1438 if (InlinedAt) 1439 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1440 else 1441 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1442 // If variable scope is not found then skip this variable. 1443 if (!Scope) 1444 continue; 1445 1446 LocalVariable Var; 1447 Var.DIVar = DIVar; 1448 1449 calculateRanges(Var, Entries); 1450 recordLocalVariable(std::move(Var), Scope); 1451 } 1452 } 1453 1454 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1455 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1456 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1457 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1458 const Function &GV = MF->getFunction(); 1459 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1460 assert(Insertion.second && "function already has info"); 1461 CurFn = Insertion.first->second.get(); 1462 CurFn->FuncId = NextFuncId++; 1463 CurFn->Begin = Asm->getFunctionBegin(); 1464 1465 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1466 // callee-saved registers were used. For targets that don't use a PUSH 1467 // instruction (AArch64), this will be zero. 1468 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1469 CurFn->FrameSize = MFI.getStackSize(); 1470 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1471 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1472 1473 // For this function S_FRAMEPROC record, figure out which codeview register 1474 // will be the frame pointer. 1475 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1476 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1477 if (CurFn->FrameSize > 0) { 1478 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1479 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1480 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1481 } else { 1482 // If there is an FP, parameters are always relative to it. 1483 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1484 if (CurFn->HasStackRealignment) { 1485 // If the stack needs realignment, locals are relative to SP or VFRAME. 1486 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1487 } else { 1488 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1489 // other stack adjustments. 1490 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1491 } 1492 } 1493 } 1494 1495 // Compute other frame procedure options. 1496 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1497 if (MFI.hasVarSizedObjects()) 1498 FPO |= FrameProcedureOptions::HasAlloca; 1499 if (MF->exposesReturnsTwice()) 1500 FPO |= FrameProcedureOptions::HasSetJmp; 1501 // FIXME: Set HasLongJmp if we ever track that info. 1502 if (MF->hasInlineAsm()) 1503 FPO |= FrameProcedureOptions::HasInlineAssembly; 1504 if (GV.hasPersonalityFn()) { 1505 if (isAsynchronousEHPersonality( 1506 classifyEHPersonality(GV.getPersonalityFn()))) 1507 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1508 else 1509 FPO |= FrameProcedureOptions::HasExceptionHandling; 1510 } 1511 if (GV.hasFnAttribute(Attribute::InlineHint)) 1512 FPO |= FrameProcedureOptions::MarkedInline; 1513 if (GV.hasFnAttribute(Attribute::Naked)) 1514 FPO |= FrameProcedureOptions::Naked; 1515 if (MFI.hasStackProtectorIndex()) 1516 FPO |= FrameProcedureOptions::SecurityChecks; 1517 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1518 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1519 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1520 !GV.hasOptSize() && !GV.hasOptNone()) 1521 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1522 if (GV.hasProfileData()) { 1523 FPO |= FrameProcedureOptions::ValidProfileCounts; 1524 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1525 } 1526 // FIXME: Set GuardCfg when it is implemented. 1527 CurFn->FrameProcOpts = FPO; 1528 1529 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1530 1531 // Find the end of the function prolog. First known non-DBG_VALUE and 1532 // non-frame setup location marks the beginning of the function body. 1533 // FIXME: is there a simpler a way to do this? Can we just search 1534 // for the first instruction of the function, not the last of the prolog? 1535 DebugLoc PrologEndLoc; 1536 bool EmptyPrologue = true; 1537 for (const auto &MBB : *MF) { 1538 for (const auto &MI : MBB) { 1539 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1540 MI.getDebugLoc()) { 1541 PrologEndLoc = MI.getDebugLoc(); 1542 break; 1543 } else if (!MI.isMetaInstruction()) { 1544 EmptyPrologue = false; 1545 } 1546 } 1547 } 1548 1549 // Record beginning of function if we have a non-empty prologue. 1550 if (PrologEndLoc && !EmptyPrologue) { 1551 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1552 maybeRecordLocation(FnStartDL, MF); 1553 } 1554 1555 // Find heap alloc sites and emit labels around them. 1556 for (const auto &MBB : *MF) { 1557 for (const auto &MI : MBB) { 1558 if (MI.getHeapAllocMarker()) { 1559 requestLabelBeforeInsn(&MI); 1560 requestLabelAfterInsn(&MI); 1561 } 1562 } 1563 } 1564 } 1565 1566 static bool shouldEmitUdt(const DIType *T) { 1567 if (!T) 1568 return false; 1569 1570 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1571 if (T->getTag() == dwarf::DW_TAG_typedef) { 1572 if (DIScope *Scope = T->getScope()) { 1573 switch (Scope->getTag()) { 1574 case dwarf::DW_TAG_structure_type: 1575 case dwarf::DW_TAG_class_type: 1576 case dwarf::DW_TAG_union_type: 1577 return false; 1578 default: 1579 // do nothing. 1580 ; 1581 } 1582 } 1583 } 1584 1585 while (true) { 1586 if (!T || T->isForwardDecl()) 1587 return false; 1588 1589 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1590 if (!DT) 1591 return true; 1592 T = DT->getBaseType(); 1593 } 1594 return true; 1595 } 1596 1597 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1598 // Don't record empty UDTs. 1599 if (Ty->getName().empty()) 1600 return; 1601 if (!shouldEmitUdt(Ty)) 1602 return; 1603 1604 SmallVector<StringRef, 5> ParentScopeNames; 1605 const DISubprogram *ClosestSubprogram = 1606 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1607 1608 std::string FullyQualifiedName = 1609 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1610 1611 if (ClosestSubprogram == nullptr) { 1612 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1613 } else if (ClosestSubprogram == CurrentSubprogram) { 1614 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1615 } 1616 1617 // TODO: What if the ClosestSubprogram is neither null or the current 1618 // subprogram? Currently, the UDT just gets dropped on the floor. 1619 // 1620 // The current behavior is not desirable. To get maximal fidelity, we would 1621 // need to perform all type translation before beginning emission of .debug$S 1622 // and then make LocalUDTs a member of FunctionInfo 1623 } 1624 1625 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1626 // Generic dispatch for lowering an unknown type. 1627 switch (Ty->getTag()) { 1628 case dwarf::DW_TAG_array_type: 1629 return lowerTypeArray(cast<DICompositeType>(Ty)); 1630 case dwarf::DW_TAG_typedef: 1631 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1632 case dwarf::DW_TAG_base_type: 1633 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1634 case dwarf::DW_TAG_pointer_type: 1635 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1636 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1637 LLVM_FALLTHROUGH; 1638 case dwarf::DW_TAG_reference_type: 1639 case dwarf::DW_TAG_rvalue_reference_type: 1640 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1641 case dwarf::DW_TAG_ptr_to_member_type: 1642 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1643 case dwarf::DW_TAG_restrict_type: 1644 case dwarf::DW_TAG_const_type: 1645 case dwarf::DW_TAG_volatile_type: 1646 // TODO: add support for DW_TAG_atomic_type here 1647 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1648 case dwarf::DW_TAG_subroutine_type: 1649 if (ClassTy) { 1650 // The member function type of a member function pointer has no 1651 // ThisAdjustment. 1652 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1653 /*ThisAdjustment=*/0, 1654 /*IsStaticMethod=*/false); 1655 } 1656 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1657 case dwarf::DW_TAG_enumeration_type: 1658 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1659 case dwarf::DW_TAG_class_type: 1660 case dwarf::DW_TAG_structure_type: 1661 return lowerTypeClass(cast<DICompositeType>(Ty)); 1662 case dwarf::DW_TAG_union_type: 1663 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1664 case dwarf::DW_TAG_string_type: 1665 return lowerTypeString(cast<DIStringType>(Ty)); 1666 case dwarf::DW_TAG_unspecified_type: 1667 if (Ty->getName() == "decltype(nullptr)") 1668 return TypeIndex::NullptrT(); 1669 return TypeIndex::None(); 1670 default: 1671 // Use the null type index. 1672 return TypeIndex(); 1673 } 1674 } 1675 1676 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1677 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1678 StringRef TypeName = Ty->getName(); 1679 1680 addToUDTs(Ty); 1681 1682 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1683 TypeName == "HRESULT") 1684 return TypeIndex(SimpleTypeKind::HResult); 1685 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1686 TypeName == "wchar_t") 1687 return TypeIndex(SimpleTypeKind::WideCharacter); 1688 1689 return UnderlyingTypeIndex; 1690 } 1691 1692 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1693 const DIType *ElementType = Ty->getBaseType(); 1694 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1695 // IndexType is size_t, which depends on the bitness of the target. 1696 TypeIndex IndexType = getPointerSizeInBytes() == 8 1697 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1698 : TypeIndex(SimpleTypeKind::UInt32Long); 1699 1700 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1701 1702 // Add subranges to array type. 1703 DINodeArray Elements = Ty->getElements(); 1704 for (int i = Elements.size() - 1; i >= 0; --i) { 1705 const DINode *Element = Elements[i]; 1706 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1707 1708 const DISubrange *Subrange = cast<DISubrange>(Element); 1709 int64_t Count = -1; 1710 1711 // If Subrange has a Count field, use it. 1712 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1713 // where lowerbound is from the LowerBound field of the Subrange, 1714 // or the language default lowerbound if that field is unspecified. 1715 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1716 Count = CI->getSExtValue(); 1717 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1718 // Fortran uses 1 as the default lowerbound; other languages use 0. 1719 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1720 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1721 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1722 Count = UI->getSExtValue() - Lowerbound + 1; 1723 } 1724 1725 // Forward declarations of arrays without a size and VLAs use a count of -1. 1726 // Emit a count of zero in these cases to match what MSVC does for arrays 1727 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1728 // should do for them even if we could distinguish them. 1729 if (Count == -1) 1730 Count = 0; 1731 1732 // Update the element size and element type index for subsequent subranges. 1733 ElementSize *= Count; 1734 1735 // If this is the outermost array, use the size from the array. It will be 1736 // more accurate if we had a VLA or an incomplete element type size. 1737 uint64_t ArraySize = 1738 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1739 1740 StringRef Name = (i == 0) ? Ty->getName() : ""; 1741 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1742 ElementTypeIndex = TypeTable.writeLeafType(AR); 1743 } 1744 1745 return ElementTypeIndex; 1746 } 1747 1748 // This function lowers a Fortran character type (DIStringType). 1749 // Note that it handles only the character*n variant (using SizeInBits 1750 // field in DIString to describe the type size) at the moment. 1751 // Other variants (leveraging the StringLength and StringLengthExp 1752 // fields in DIStringType) remain TBD. 1753 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1754 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1755 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1756 StringRef Name = Ty->getName(); 1757 // IndexType is size_t, which depends on the bitness of the target. 1758 TypeIndex IndexType = getPointerSizeInBytes() == 8 1759 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1760 : TypeIndex(SimpleTypeKind::UInt32Long); 1761 1762 // Create a type of character array of ArraySize. 1763 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1764 1765 return TypeTable.writeLeafType(AR); 1766 } 1767 1768 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1769 TypeIndex Index; 1770 dwarf::TypeKind Kind; 1771 uint32_t ByteSize; 1772 1773 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1774 ByteSize = Ty->getSizeInBits() / 8; 1775 1776 SimpleTypeKind STK = SimpleTypeKind::None; 1777 switch (Kind) { 1778 case dwarf::DW_ATE_address: 1779 // FIXME: Translate 1780 break; 1781 case dwarf::DW_ATE_boolean: 1782 switch (ByteSize) { 1783 case 1: STK = SimpleTypeKind::Boolean8; break; 1784 case 2: STK = SimpleTypeKind::Boolean16; break; 1785 case 4: STK = SimpleTypeKind::Boolean32; break; 1786 case 8: STK = SimpleTypeKind::Boolean64; break; 1787 case 16: STK = SimpleTypeKind::Boolean128; break; 1788 } 1789 break; 1790 case dwarf::DW_ATE_complex_float: 1791 switch (ByteSize) { 1792 case 2: STK = SimpleTypeKind::Complex16; break; 1793 case 4: STK = SimpleTypeKind::Complex32; break; 1794 case 8: STK = SimpleTypeKind::Complex64; break; 1795 case 10: STK = SimpleTypeKind::Complex80; break; 1796 case 16: STK = SimpleTypeKind::Complex128; break; 1797 } 1798 break; 1799 case dwarf::DW_ATE_float: 1800 switch (ByteSize) { 1801 case 2: STK = SimpleTypeKind::Float16; break; 1802 case 4: STK = SimpleTypeKind::Float32; break; 1803 case 6: STK = SimpleTypeKind::Float48; break; 1804 case 8: STK = SimpleTypeKind::Float64; break; 1805 case 10: STK = SimpleTypeKind::Float80; break; 1806 case 16: STK = SimpleTypeKind::Float128; break; 1807 } 1808 break; 1809 case dwarf::DW_ATE_signed: 1810 switch (ByteSize) { 1811 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1812 case 2: STK = SimpleTypeKind::Int16Short; break; 1813 case 4: STK = SimpleTypeKind::Int32; break; 1814 case 8: STK = SimpleTypeKind::Int64Quad; break; 1815 case 16: STK = SimpleTypeKind::Int128Oct; break; 1816 } 1817 break; 1818 case dwarf::DW_ATE_unsigned: 1819 switch (ByteSize) { 1820 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1821 case 2: STK = SimpleTypeKind::UInt16Short; break; 1822 case 4: STK = SimpleTypeKind::UInt32; break; 1823 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1824 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1825 } 1826 break; 1827 case dwarf::DW_ATE_UTF: 1828 switch (ByteSize) { 1829 case 1: STK = SimpleTypeKind::Character8; break; 1830 case 2: STK = SimpleTypeKind::Character16; break; 1831 case 4: STK = SimpleTypeKind::Character32; break; 1832 } 1833 break; 1834 case dwarf::DW_ATE_signed_char: 1835 if (ByteSize == 1) 1836 STK = SimpleTypeKind::SignedCharacter; 1837 break; 1838 case dwarf::DW_ATE_unsigned_char: 1839 if (ByteSize == 1) 1840 STK = SimpleTypeKind::UnsignedCharacter; 1841 break; 1842 default: 1843 break; 1844 } 1845 1846 // Apply some fixups based on the source-level type name. 1847 // Include some amount of canonicalization from an old naming scheme Clang 1848 // used to use for integer types (in an outdated effort to be compatible with 1849 // GCC's debug info/GDB's behavior, which has since been addressed). 1850 if (STK == SimpleTypeKind::Int32 && 1851 (Ty->getName() == "long int" || Ty->getName() == "long")) 1852 STK = SimpleTypeKind::Int32Long; 1853 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1854 Ty->getName() == "unsigned long")) 1855 STK = SimpleTypeKind::UInt32Long; 1856 if (STK == SimpleTypeKind::UInt16Short && 1857 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1858 STK = SimpleTypeKind::WideCharacter; 1859 if ((STK == SimpleTypeKind::SignedCharacter || 1860 STK == SimpleTypeKind::UnsignedCharacter) && 1861 Ty->getName() == "char") 1862 STK = SimpleTypeKind::NarrowCharacter; 1863 1864 return TypeIndex(STK); 1865 } 1866 1867 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1868 PointerOptions PO) { 1869 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1870 1871 // Pointers to simple types without any options can use SimpleTypeMode, rather 1872 // than having a dedicated pointer type record. 1873 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1874 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1875 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1876 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1877 ? SimpleTypeMode::NearPointer64 1878 : SimpleTypeMode::NearPointer32; 1879 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1880 } 1881 1882 PointerKind PK = 1883 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1884 PointerMode PM = PointerMode::Pointer; 1885 switch (Ty->getTag()) { 1886 default: llvm_unreachable("not a pointer tag type"); 1887 case dwarf::DW_TAG_pointer_type: 1888 PM = PointerMode::Pointer; 1889 break; 1890 case dwarf::DW_TAG_reference_type: 1891 PM = PointerMode::LValueReference; 1892 break; 1893 case dwarf::DW_TAG_rvalue_reference_type: 1894 PM = PointerMode::RValueReference; 1895 break; 1896 } 1897 1898 if (Ty->isObjectPointer()) 1899 PO |= PointerOptions::Const; 1900 1901 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1902 return TypeTable.writeLeafType(PR); 1903 } 1904 1905 static PointerToMemberRepresentation 1906 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1907 // SizeInBytes being zero generally implies that the member pointer type was 1908 // incomplete, which can happen if it is part of a function prototype. In this 1909 // case, use the unknown model instead of the general model. 1910 if (IsPMF) { 1911 switch (Flags & DINode::FlagPtrToMemberRep) { 1912 case 0: 1913 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1914 : PointerToMemberRepresentation::GeneralFunction; 1915 case DINode::FlagSingleInheritance: 1916 return PointerToMemberRepresentation::SingleInheritanceFunction; 1917 case DINode::FlagMultipleInheritance: 1918 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1919 case DINode::FlagVirtualInheritance: 1920 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1921 } 1922 } else { 1923 switch (Flags & DINode::FlagPtrToMemberRep) { 1924 case 0: 1925 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1926 : PointerToMemberRepresentation::GeneralData; 1927 case DINode::FlagSingleInheritance: 1928 return PointerToMemberRepresentation::SingleInheritanceData; 1929 case DINode::FlagMultipleInheritance: 1930 return PointerToMemberRepresentation::MultipleInheritanceData; 1931 case DINode::FlagVirtualInheritance: 1932 return PointerToMemberRepresentation::VirtualInheritanceData; 1933 } 1934 } 1935 llvm_unreachable("invalid ptr to member representation"); 1936 } 1937 1938 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1939 PointerOptions PO) { 1940 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1941 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1942 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1943 TypeIndex PointeeTI = 1944 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1945 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1946 : PointerKind::Near32; 1947 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1948 : PointerMode::PointerToDataMember; 1949 1950 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1951 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1952 MemberPointerInfo MPI( 1953 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1954 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1955 return TypeTable.writeLeafType(PR); 1956 } 1957 1958 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1959 /// have a translation, use the NearC convention. 1960 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1961 switch (DwarfCC) { 1962 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1963 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1964 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1965 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1966 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1967 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1968 } 1969 return CallingConvention::NearC; 1970 } 1971 1972 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1973 ModifierOptions Mods = ModifierOptions::None; 1974 PointerOptions PO = PointerOptions::None; 1975 bool IsModifier = true; 1976 const DIType *BaseTy = Ty; 1977 while (IsModifier && BaseTy) { 1978 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1979 switch (BaseTy->getTag()) { 1980 case dwarf::DW_TAG_const_type: 1981 Mods |= ModifierOptions::Const; 1982 PO |= PointerOptions::Const; 1983 break; 1984 case dwarf::DW_TAG_volatile_type: 1985 Mods |= ModifierOptions::Volatile; 1986 PO |= PointerOptions::Volatile; 1987 break; 1988 case dwarf::DW_TAG_restrict_type: 1989 // Only pointer types be marked with __restrict. There is no known flag 1990 // for __restrict in LF_MODIFIER records. 1991 PO |= PointerOptions::Restrict; 1992 break; 1993 default: 1994 IsModifier = false; 1995 break; 1996 } 1997 if (IsModifier) 1998 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1999 } 2000 2001 // Check if the inner type will use an LF_POINTER record. If so, the 2002 // qualifiers will go in the LF_POINTER record. This comes up for types like 2003 // 'int *const' and 'int *__restrict', not the more common cases like 'const 2004 // char *'. 2005 if (BaseTy) { 2006 switch (BaseTy->getTag()) { 2007 case dwarf::DW_TAG_pointer_type: 2008 case dwarf::DW_TAG_reference_type: 2009 case dwarf::DW_TAG_rvalue_reference_type: 2010 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2011 case dwarf::DW_TAG_ptr_to_member_type: 2012 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2013 default: 2014 break; 2015 } 2016 } 2017 2018 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2019 2020 // Return the base type index if there aren't any modifiers. For example, the 2021 // metadata could contain restrict wrappers around non-pointer types. 2022 if (Mods == ModifierOptions::None) 2023 return ModifiedTI; 2024 2025 ModifierRecord MR(ModifiedTI, Mods); 2026 return TypeTable.writeLeafType(MR); 2027 } 2028 2029 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2030 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2031 for (const DIType *ArgType : Ty->getTypeArray()) 2032 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2033 2034 // MSVC uses type none for variadic argument. 2035 if (ReturnAndArgTypeIndices.size() > 1 && 2036 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2037 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2038 } 2039 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2040 ArrayRef<TypeIndex> ArgTypeIndices = None; 2041 if (!ReturnAndArgTypeIndices.empty()) { 2042 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 2043 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2044 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2045 } 2046 2047 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2048 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2049 2050 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2051 2052 FunctionOptions FO = getFunctionOptions(Ty); 2053 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2054 ArgListIndex); 2055 return TypeTable.writeLeafType(Procedure); 2056 } 2057 2058 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2059 const DIType *ClassTy, 2060 int ThisAdjustment, 2061 bool IsStaticMethod, 2062 FunctionOptions FO) { 2063 // Lower the containing class type. 2064 TypeIndex ClassType = getTypeIndex(ClassTy); 2065 2066 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2067 2068 unsigned Index = 0; 2069 SmallVector<TypeIndex, 8> ArgTypeIndices; 2070 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2071 if (ReturnAndArgs.size() > Index) { 2072 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2073 } 2074 2075 // If the first argument is a pointer type and this isn't a static method, 2076 // treat it as the special 'this' parameter, which is encoded separately from 2077 // the arguments. 2078 TypeIndex ThisTypeIndex; 2079 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2080 if (const DIDerivedType *PtrTy = 2081 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2082 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2083 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2084 Index++; 2085 } 2086 } 2087 } 2088 2089 while (Index < ReturnAndArgs.size()) 2090 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2091 2092 // MSVC uses type none for variadic argument. 2093 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2094 ArgTypeIndices.back() = TypeIndex::None(); 2095 2096 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2097 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2098 2099 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2100 2101 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2102 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2103 return TypeTable.writeLeafType(MFR); 2104 } 2105 2106 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2107 unsigned VSlotCount = 2108 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2109 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2110 2111 VFTableShapeRecord VFTSR(Slots); 2112 return TypeTable.writeLeafType(VFTSR); 2113 } 2114 2115 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2116 switch (Flags & DINode::FlagAccessibility) { 2117 case DINode::FlagPrivate: return MemberAccess::Private; 2118 case DINode::FlagPublic: return MemberAccess::Public; 2119 case DINode::FlagProtected: return MemberAccess::Protected; 2120 case 0: 2121 // If there was no explicit access control, provide the default for the tag. 2122 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2123 : MemberAccess::Public; 2124 } 2125 llvm_unreachable("access flags are exclusive"); 2126 } 2127 2128 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2129 if (SP->isArtificial()) 2130 return MethodOptions::CompilerGenerated; 2131 2132 // FIXME: Handle other MethodOptions. 2133 2134 return MethodOptions::None; 2135 } 2136 2137 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2138 bool Introduced) { 2139 if (SP->getFlags() & DINode::FlagStaticMember) 2140 return MethodKind::Static; 2141 2142 switch (SP->getVirtuality()) { 2143 case dwarf::DW_VIRTUALITY_none: 2144 break; 2145 case dwarf::DW_VIRTUALITY_virtual: 2146 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2147 case dwarf::DW_VIRTUALITY_pure_virtual: 2148 return Introduced ? MethodKind::PureIntroducingVirtual 2149 : MethodKind::PureVirtual; 2150 default: 2151 llvm_unreachable("unhandled virtuality case"); 2152 } 2153 2154 return MethodKind::Vanilla; 2155 } 2156 2157 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2158 switch (Ty->getTag()) { 2159 case dwarf::DW_TAG_class_type: 2160 return TypeRecordKind::Class; 2161 case dwarf::DW_TAG_structure_type: 2162 return TypeRecordKind::Struct; 2163 default: 2164 llvm_unreachable("unexpected tag"); 2165 } 2166 } 2167 2168 /// Return ClassOptions that should be present on both the forward declaration 2169 /// and the defintion of a tag type. 2170 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2171 ClassOptions CO = ClassOptions::None; 2172 2173 // MSVC always sets this flag, even for local types. Clang doesn't always 2174 // appear to give every type a linkage name, which may be problematic for us. 2175 // FIXME: Investigate the consequences of not following them here. 2176 if (!Ty->getIdentifier().empty()) 2177 CO |= ClassOptions::HasUniqueName; 2178 2179 // Put the Nested flag on a type if it appears immediately inside a tag type. 2180 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2181 // here. That flag is only set on definitions, and not forward declarations. 2182 const DIScope *ImmediateScope = Ty->getScope(); 2183 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2184 CO |= ClassOptions::Nested; 2185 2186 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2187 // type only when it has an immediate function scope. Clang never puts enums 2188 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2189 // always in function, class, or file scopes. 2190 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2191 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2192 CO |= ClassOptions::Scoped; 2193 } else { 2194 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2195 Scope = Scope->getScope()) { 2196 if (isa<DISubprogram>(Scope)) { 2197 CO |= ClassOptions::Scoped; 2198 break; 2199 } 2200 } 2201 } 2202 2203 return CO; 2204 } 2205 2206 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2207 switch (Ty->getTag()) { 2208 case dwarf::DW_TAG_class_type: 2209 case dwarf::DW_TAG_structure_type: 2210 case dwarf::DW_TAG_union_type: 2211 case dwarf::DW_TAG_enumeration_type: 2212 break; 2213 default: 2214 return; 2215 } 2216 2217 if (const auto *File = Ty->getFile()) { 2218 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2219 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2220 2221 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2222 TypeTable.writeLeafType(USLR); 2223 } 2224 } 2225 2226 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2227 ClassOptions CO = getCommonClassOptions(Ty); 2228 TypeIndex FTI; 2229 unsigned EnumeratorCount = 0; 2230 2231 if (Ty->isForwardDecl()) { 2232 CO |= ClassOptions::ForwardReference; 2233 } else { 2234 ContinuationRecordBuilder ContinuationBuilder; 2235 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2236 for (const DINode *Element : Ty->getElements()) { 2237 // We assume that the frontend provides all members in source declaration 2238 // order, which is what MSVC does. 2239 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2240 // FIXME: Is it correct to always emit these as unsigned here? 2241 EnumeratorRecord ER(MemberAccess::Public, 2242 APSInt(Enumerator->getValue(), true), 2243 Enumerator->getName()); 2244 ContinuationBuilder.writeMemberType(ER); 2245 EnumeratorCount++; 2246 } 2247 } 2248 FTI = TypeTable.insertRecord(ContinuationBuilder); 2249 } 2250 2251 std::string FullName = getFullyQualifiedName(Ty); 2252 2253 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2254 getTypeIndex(Ty->getBaseType())); 2255 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2256 2257 addUDTSrcLine(Ty, EnumTI); 2258 2259 return EnumTI; 2260 } 2261 2262 //===----------------------------------------------------------------------===// 2263 // ClassInfo 2264 //===----------------------------------------------------------------------===// 2265 2266 struct llvm::ClassInfo { 2267 struct MemberInfo { 2268 const DIDerivedType *MemberTypeNode; 2269 uint64_t BaseOffset; 2270 }; 2271 // [MemberInfo] 2272 using MemberList = std::vector<MemberInfo>; 2273 2274 using MethodsList = TinyPtrVector<const DISubprogram *>; 2275 // MethodName -> MethodsList 2276 using MethodsMap = MapVector<MDString *, MethodsList>; 2277 2278 /// Base classes. 2279 std::vector<const DIDerivedType *> Inheritance; 2280 2281 /// Direct members. 2282 MemberList Members; 2283 // Direct overloaded methods gathered by name. 2284 MethodsMap Methods; 2285 2286 TypeIndex VShapeTI; 2287 2288 std::vector<const DIType *> NestedTypes; 2289 }; 2290 2291 void CodeViewDebug::clear() { 2292 assert(CurFn == nullptr); 2293 FileIdMap.clear(); 2294 FnDebugInfo.clear(); 2295 FileToFilepathMap.clear(); 2296 LocalUDTs.clear(); 2297 GlobalUDTs.clear(); 2298 TypeIndices.clear(); 2299 CompleteTypeIndices.clear(); 2300 ScopeGlobals.clear(); 2301 CVGlobalVariableOffsets.clear(); 2302 } 2303 2304 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2305 const DIDerivedType *DDTy) { 2306 if (!DDTy->getName().empty()) { 2307 Info.Members.push_back({DDTy, 0}); 2308 2309 // Collect static const data members with values. 2310 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2311 DINode::FlagStaticMember) { 2312 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2313 isa<ConstantFP>(DDTy->getConstant()))) 2314 StaticConstMembers.push_back(DDTy); 2315 } 2316 2317 return; 2318 } 2319 2320 // An unnamed member may represent a nested struct or union. Attempt to 2321 // interpret the unnamed member as a DICompositeType possibly wrapped in 2322 // qualifier types. Add all the indirect fields to the current record if that 2323 // succeeds, and drop the member if that fails. 2324 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2325 uint64_t Offset = DDTy->getOffsetInBits(); 2326 const DIType *Ty = DDTy->getBaseType(); 2327 bool FullyResolved = false; 2328 while (!FullyResolved) { 2329 switch (Ty->getTag()) { 2330 case dwarf::DW_TAG_const_type: 2331 case dwarf::DW_TAG_volatile_type: 2332 // FIXME: we should apply the qualifier types to the indirect fields 2333 // rather than dropping them. 2334 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2335 break; 2336 default: 2337 FullyResolved = true; 2338 break; 2339 } 2340 } 2341 2342 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2343 if (!DCTy) 2344 return; 2345 2346 ClassInfo NestedInfo = collectClassInfo(DCTy); 2347 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2348 Info.Members.push_back( 2349 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2350 } 2351 2352 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2353 ClassInfo Info; 2354 // Add elements to structure type. 2355 DINodeArray Elements = Ty->getElements(); 2356 for (auto *Element : Elements) { 2357 // We assume that the frontend provides all members in source declaration 2358 // order, which is what MSVC does. 2359 if (!Element) 2360 continue; 2361 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2362 Info.Methods[SP->getRawName()].push_back(SP); 2363 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2364 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2365 collectMemberInfo(Info, DDTy); 2366 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2367 Info.Inheritance.push_back(DDTy); 2368 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2369 DDTy->getName() == "__vtbl_ptr_type") { 2370 Info.VShapeTI = getTypeIndex(DDTy); 2371 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2372 Info.NestedTypes.push_back(DDTy); 2373 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2374 // Ignore friend members. It appears that MSVC emitted info about 2375 // friends in the past, but modern versions do not. 2376 } 2377 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2378 Info.NestedTypes.push_back(Composite); 2379 } 2380 // Skip other unrecognized kinds of elements. 2381 } 2382 return Info; 2383 } 2384 2385 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2386 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2387 // if a complete type should be emitted instead of a forward reference. 2388 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2389 !Ty->isForwardDecl(); 2390 } 2391 2392 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2393 // Emit the complete type for unnamed structs. C++ classes with methods 2394 // which have a circular reference back to the class type are expected to 2395 // be named by the front-end and should not be "unnamed". C unnamed 2396 // structs should not have circular references. 2397 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2398 // If this unnamed complete type is already in the process of being defined 2399 // then the description of the type is malformed and cannot be emitted 2400 // into CodeView correctly so report a fatal error. 2401 auto I = CompleteTypeIndices.find(Ty); 2402 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2403 report_fatal_error("cannot debug circular reference to unnamed type"); 2404 return getCompleteTypeIndex(Ty); 2405 } 2406 2407 // First, construct the forward decl. Don't look into Ty to compute the 2408 // forward decl options, since it might not be available in all TUs. 2409 TypeRecordKind Kind = getRecordKind(Ty); 2410 ClassOptions CO = 2411 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2412 std::string FullName = getFullyQualifiedName(Ty); 2413 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2414 FullName, Ty->getIdentifier()); 2415 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2416 if (!Ty->isForwardDecl()) 2417 DeferredCompleteTypes.push_back(Ty); 2418 return FwdDeclTI; 2419 } 2420 2421 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2422 // Construct the field list and complete type record. 2423 TypeRecordKind Kind = getRecordKind(Ty); 2424 ClassOptions CO = getCommonClassOptions(Ty); 2425 TypeIndex FieldTI; 2426 TypeIndex VShapeTI; 2427 unsigned FieldCount; 2428 bool ContainsNestedClass; 2429 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2430 lowerRecordFieldList(Ty); 2431 2432 if (ContainsNestedClass) 2433 CO |= ClassOptions::ContainsNestedClass; 2434 2435 // MSVC appears to set this flag by searching any destructor or method with 2436 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2437 // the members, however special member functions are not yet emitted into 2438 // debug information. For now checking a class's non-triviality seems enough. 2439 // FIXME: not true for a nested unnamed struct. 2440 if (isNonTrivial(Ty)) 2441 CO |= ClassOptions::HasConstructorOrDestructor; 2442 2443 std::string FullName = getFullyQualifiedName(Ty); 2444 2445 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2446 2447 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2448 SizeInBytes, FullName, Ty->getIdentifier()); 2449 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2450 2451 addUDTSrcLine(Ty, ClassTI); 2452 2453 addToUDTs(Ty); 2454 2455 return ClassTI; 2456 } 2457 2458 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2459 // Emit the complete type for unnamed unions. 2460 if (shouldAlwaysEmitCompleteClassType(Ty)) 2461 return getCompleteTypeIndex(Ty); 2462 2463 ClassOptions CO = 2464 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2465 std::string FullName = getFullyQualifiedName(Ty); 2466 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2467 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2468 if (!Ty->isForwardDecl()) 2469 DeferredCompleteTypes.push_back(Ty); 2470 return FwdDeclTI; 2471 } 2472 2473 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2474 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2475 TypeIndex FieldTI; 2476 unsigned FieldCount; 2477 bool ContainsNestedClass; 2478 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2479 lowerRecordFieldList(Ty); 2480 2481 if (ContainsNestedClass) 2482 CO |= ClassOptions::ContainsNestedClass; 2483 2484 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2485 std::string FullName = getFullyQualifiedName(Ty); 2486 2487 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2488 Ty->getIdentifier()); 2489 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2490 2491 addUDTSrcLine(Ty, UnionTI); 2492 2493 addToUDTs(Ty); 2494 2495 return UnionTI; 2496 } 2497 2498 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2499 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2500 // Manually count members. MSVC appears to count everything that generates a 2501 // field list record. Each individual overload in a method overload group 2502 // contributes to this count, even though the overload group is a single field 2503 // list record. 2504 unsigned MemberCount = 0; 2505 ClassInfo Info = collectClassInfo(Ty); 2506 ContinuationRecordBuilder ContinuationBuilder; 2507 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2508 2509 // Create base classes. 2510 for (const DIDerivedType *I : Info.Inheritance) { 2511 if (I->getFlags() & DINode::FlagVirtual) { 2512 // Virtual base. 2513 unsigned VBPtrOffset = I->getVBPtrOffset(); 2514 // FIXME: Despite the accessor name, the offset is really in bytes. 2515 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2516 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2517 ? TypeRecordKind::IndirectVirtualBaseClass 2518 : TypeRecordKind::VirtualBaseClass; 2519 VirtualBaseClassRecord VBCR( 2520 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2521 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2522 VBTableIndex); 2523 2524 ContinuationBuilder.writeMemberType(VBCR); 2525 MemberCount++; 2526 } else { 2527 assert(I->getOffsetInBits() % 8 == 0 && 2528 "bases must be on byte boundaries"); 2529 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2530 getTypeIndex(I->getBaseType()), 2531 I->getOffsetInBits() / 8); 2532 ContinuationBuilder.writeMemberType(BCR); 2533 MemberCount++; 2534 } 2535 } 2536 2537 // Create members. 2538 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2539 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2540 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2541 StringRef MemberName = Member->getName(); 2542 MemberAccess Access = 2543 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2544 2545 if (Member->isStaticMember()) { 2546 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2547 ContinuationBuilder.writeMemberType(SDMR); 2548 MemberCount++; 2549 continue; 2550 } 2551 2552 // Virtual function pointer member. 2553 if ((Member->getFlags() & DINode::FlagArtificial) && 2554 Member->getName().startswith("_vptr$")) { 2555 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2556 ContinuationBuilder.writeMemberType(VFPR); 2557 MemberCount++; 2558 continue; 2559 } 2560 2561 // Data member. 2562 uint64_t MemberOffsetInBits = 2563 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2564 if (Member->isBitField()) { 2565 uint64_t StartBitOffset = MemberOffsetInBits; 2566 if (const auto *CI = 2567 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2568 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2569 } 2570 StartBitOffset -= MemberOffsetInBits; 2571 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2572 StartBitOffset); 2573 MemberBaseType = TypeTable.writeLeafType(BFR); 2574 } 2575 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2576 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2577 MemberName); 2578 ContinuationBuilder.writeMemberType(DMR); 2579 MemberCount++; 2580 } 2581 2582 // Create methods 2583 for (auto &MethodItr : Info.Methods) { 2584 StringRef Name = MethodItr.first->getString(); 2585 2586 std::vector<OneMethodRecord> Methods; 2587 for (const DISubprogram *SP : MethodItr.second) { 2588 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2589 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2590 2591 unsigned VFTableOffset = -1; 2592 if (Introduced) 2593 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2594 2595 Methods.push_back(OneMethodRecord( 2596 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2597 translateMethodKindFlags(SP, Introduced), 2598 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2599 MemberCount++; 2600 } 2601 assert(!Methods.empty() && "Empty methods map entry"); 2602 if (Methods.size() == 1) 2603 ContinuationBuilder.writeMemberType(Methods[0]); 2604 else { 2605 // FIXME: Make this use its own ContinuationBuilder so that 2606 // MethodOverloadList can be split correctly. 2607 MethodOverloadListRecord MOLR(Methods); 2608 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2609 2610 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2611 ContinuationBuilder.writeMemberType(OMR); 2612 } 2613 } 2614 2615 // Create nested classes. 2616 for (const DIType *Nested : Info.NestedTypes) { 2617 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2618 ContinuationBuilder.writeMemberType(R); 2619 MemberCount++; 2620 } 2621 2622 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2623 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2624 !Info.NestedTypes.empty()); 2625 } 2626 2627 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2628 if (!VBPType.getIndex()) { 2629 // Make a 'const int *' type. 2630 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2631 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2632 2633 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2634 : PointerKind::Near32; 2635 PointerMode PM = PointerMode::Pointer; 2636 PointerOptions PO = PointerOptions::None; 2637 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2638 VBPType = TypeTable.writeLeafType(PR); 2639 } 2640 2641 return VBPType; 2642 } 2643 2644 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2645 // The null DIType is the void type. Don't try to hash it. 2646 if (!Ty) 2647 return TypeIndex::Void(); 2648 2649 // Check if we've already translated this type. Don't try to do a 2650 // get-or-create style insertion that caches the hash lookup across the 2651 // lowerType call. It will update the TypeIndices map. 2652 auto I = TypeIndices.find({Ty, ClassTy}); 2653 if (I != TypeIndices.end()) 2654 return I->second; 2655 2656 TypeLoweringScope S(*this); 2657 TypeIndex TI = lowerType(Ty, ClassTy); 2658 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2659 } 2660 2661 codeview::TypeIndex 2662 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2663 const DISubroutineType *SubroutineTy) { 2664 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2665 "this type must be a pointer type"); 2666 2667 PointerOptions Options = PointerOptions::None; 2668 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2669 Options = PointerOptions::LValueRefThisPointer; 2670 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2671 Options = PointerOptions::RValueRefThisPointer; 2672 2673 // Check if we've already translated this type. If there is no ref qualifier 2674 // on the function then we look up this pointer type with no associated class 2675 // so that the TypeIndex for the this pointer can be shared with the type 2676 // index for other pointers to this class type. If there is a ref qualifier 2677 // then we lookup the pointer using the subroutine as the parent type. 2678 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2679 if (I != TypeIndices.end()) 2680 return I->second; 2681 2682 TypeLoweringScope S(*this); 2683 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2684 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2685 } 2686 2687 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2688 PointerRecord PR(getTypeIndex(Ty), 2689 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2690 : PointerKind::Near32, 2691 PointerMode::LValueReference, PointerOptions::None, 2692 Ty->getSizeInBits() / 8); 2693 return TypeTable.writeLeafType(PR); 2694 } 2695 2696 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2697 // The null DIType is the void type. Don't try to hash it. 2698 if (!Ty) 2699 return TypeIndex::Void(); 2700 2701 // Look through typedefs when getting the complete type index. Call 2702 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2703 // emitted only once. 2704 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2705 (void)getTypeIndex(Ty); 2706 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2707 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2708 2709 // If this is a non-record type, the complete type index is the same as the 2710 // normal type index. Just call getTypeIndex. 2711 switch (Ty->getTag()) { 2712 case dwarf::DW_TAG_class_type: 2713 case dwarf::DW_TAG_structure_type: 2714 case dwarf::DW_TAG_union_type: 2715 break; 2716 default: 2717 return getTypeIndex(Ty); 2718 } 2719 2720 const auto *CTy = cast<DICompositeType>(Ty); 2721 2722 TypeLoweringScope S(*this); 2723 2724 // Make sure the forward declaration is emitted first. It's unclear if this 2725 // is necessary, but MSVC does it, and we should follow suit until we can show 2726 // otherwise. 2727 // We only emit a forward declaration for named types. 2728 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2729 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2730 2731 // Just use the forward decl if we don't have complete type info. This 2732 // might happen if the frontend is using modules and expects the complete 2733 // definition to be emitted elsewhere. 2734 if (CTy->isForwardDecl()) 2735 return FwdDeclTI; 2736 } 2737 2738 // Check if we've already translated the complete record type. 2739 // Insert the type with a null TypeIndex to signify that the type is currently 2740 // being lowered. 2741 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2742 if (!InsertResult.second) 2743 return InsertResult.first->second; 2744 2745 TypeIndex TI; 2746 switch (CTy->getTag()) { 2747 case dwarf::DW_TAG_class_type: 2748 case dwarf::DW_TAG_structure_type: 2749 TI = lowerCompleteTypeClass(CTy); 2750 break; 2751 case dwarf::DW_TAG_union_type: 2752 TI = lowerCompleteTypeUnion(CTy); 2753 break; 2754 default: 2755 llvm_unreachable("not a record"); 2756 } 2757 2758 // Update the type index associated with this CompositeType. This cannot 2759 // use the 'InsertResult' iterator above because it is potentially 2760 // invalidated by map insertions which can occur while lowering the class 2761 // type above. 2762 CompleteTypeIndices[CTy] = TI; 2763 return TI; 2764 } 2765 2766 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2767 /// and do this until fixpoint, as each complete record type typically 2768 /// references 2769 /// many other record types. 2770 void CodeViewDebug::emitDeferredCompleteTypes() { 2771 SmallVector<const DICompositeType *, 4> TypesToEmit; 2772 while (!DeferredCompleteTypes.empty()) { 2773 std::swap(DeferredCompleteTypes, TypesToEmit); 2774 for (const DICompositeType *RecordTy : TypesToEmit) 2775 getCompleteTypeIndex(RecordTy); 2776 TypesToEmit.clear(); 2777 } 2778 } 2779 2780 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2781 ArrayRef<LocalVariable> Locals) { 2782 // Get the sorted list of parameters and emit them first. 2783 SmallVector<const LocalVariable *, 6> Params; 2784 for (const LocalVariable &L : Locals) 2785 if (L.DIVar->isParameter()) 2786 Params.push_back(&L); 2787 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2788 return L->DIVar->getArg() < R->DIVar->getArg(); 2789 }); 2790 for (const LocalVariable *L : Params) 2791 emitLocalVariable(FI, *L); 2792 2793 // Next emit all non-parameters in the order that we found them. 2794 for (const LocalVariable &L : Locals) 2795 if (!L.DIVar->isParameter()) 2796 emitLocalVariable(FI, L); 2797 } 2798 2799 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2800 const LocalVariable &Var) { 2801 // LocalSym record, see SymbolRecord.h for more info. 2802 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2803 2804 LocalSymFlags Flags = LocalSymFlags::None; 2805 if (Var.DIVar->isParameter()) 2806 Flags |= LocalSymFlags::IsParameter; 2807 if (Var.DefRanges.empty()) 2808 Flags |= LocalSymFlags::IsOptimizedOut; 2809 2810 OS.AddComment("TypeIndex"); 2811 TypeIndex TI = Var.UseReferenceType 2812 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2813 : getCompleteTypeIndex(Var.DIVar->getType()); 2814 OS.emitInt32(TI.getIndex()); 2815 OS.AddComment("Flags"); 2816 OS.emitInt16(static_cast<uint16_t>(Flags)); 2817 // Truncate the name so we won't overflow the record length field. 2818 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2819 endSymbolRecord(LocalEnd); 2820 2821 // Calculate the on disk prefix of the appropriate def range record. The 2822 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2823 // should be big enough to hold all forms without memory allocation. 2824 SmallString<20> BytePrefix; 2825 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2826 BytePrefix.clear(); 2827 if (DefRange.InMemory) { 2828 int Offset = DefRange.DataOffset; 2829 unsigned Reg = DefRange.CVRegister; 2830 2831 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2832 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2833 // instead. In frames without stack realignment, $T0 will be the CFA. 2834 if (RegisterId(Reg) == RegisterId::ESP) { 2835 Reg = unsigned(RegisterId::VFRAME); 2836 Offset += FI.OffsetAdjustment; 2837 } 2838 2839 // If we can use the chosen frame pointer for the frame and this isn't a 2840 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2841 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2842 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2843 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2844 (bool(Flags & LocalSymFlags::IsParameter) 2845 ? (EncFP == FI.EncodedParamFramePtrReg) 2846 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2847 DefRangeFramePointerRelHeader DRHdr; 2848 DRHdr.Offset = Offset; 2849 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2850 } else { 2851 uint16_t RegRelFlags = 0; 2852 if (DefRange.IsSubfield) { 2853 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2854 (DefRange.StructOffset 2855 << DefRangeRegisterRelSym::OffsetInParentShift); 2856 } 2857 DefRangeRegisterRelHeader DRHdr; 2858 DRHdr.Register = Reg; 2859 DRHdr.Flags = RegRelFlags; 2860 DRHdr.BasePointerOffset = Offset; 2861 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2862 } 2863 } else { 2864 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2865 if (DefRange.IsSubfield) { 2866 DefRangeSubfieldRegisterHeader DRHdr; 2867 DRHdr.Register = DefRange.CVRegister; 2868 DRHdr.MayHaveNoName = 0; 2869 DRHdr.OffsetInParent = DefRange.StructOffset; 2870 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2871 } else { 2872 DefRangeRegisterHeader DRHdr; 2873 DRHdr.Register = DefRange.CVRegister; 2874 DRHdr.MayHaveNoName = 0; 2875 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2876 } 2877 } 2878 } 2879 } 2880 2881 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2882 const FunctionInfo& FI) { 2883 for (LexicalBlock *Block : Blocks) 2884 emitLexicalBlock(*Block, FI); 2885 } 2886 2887 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2888 /// lexical block scope. 2889 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2890 const FunctionInfo& FI) { 2891 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2892 OS.AddComment("PtrParent"); 2893 OS.emitInt32(0); // PtrParent 2894 OS.AddComment("PtrEnd"); 2895 OS.emitInt32(0); // PtrEnd 2896 OS.AddComment("Code size"); 2897 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2898 OS.AddComment("Function section relative address"); 2899 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2900 OS.AddComment("Function section index"); 2901 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2902 OS.AddComment("Lexical block name"); 2903 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2904 endSymbolRecord(RecordEnd); 2905 2906 // Emit variables local to this lexical block. 2907 emitLocalVariableList(FI, Block.Locals); 2908 emitGlobalVariableList(Block.Globals); 2909 2910 // Emit lexical blocks contained within this block. 2911 emitLexicalBlockList(Block.Children, FI); 2912 2913 // Close the lexical block scope. 2914 emitEndSymbolRecord(SymbolKind::S_END); 2915 } 2916 2917 /// Convenience routine for collecting lexical block information for a list 2918 /// of lexical scopes. 2919 void CodeViewDebug::collectLexicalBlockInfo( 2920 SmallVectorImpl<LexicalScope *> &Scopes, 2921 SmallVectorImpl<LexicalBlock *> &Blocks, 2922 SmallVectorImpl<LocalVariable> &Locals, 2923 SmallVectorImpl<CVGlobalVariable> &Globals) { 2924 for (LexicalScope *Scope : Scopes) 2925 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2926 } 2927 2928 /// Populate the lexical blocks and local variable lists of the parent with 2929 /// information about the specified lexical scope. 2930 void CodeViewDebug::collectLexicalBlockInfo( 2931 LexicalScope &Scope, 2932 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2933 SmallVectorImpl<LocalVariable> &ParentLocals, 2934 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2935 if (Scope.isAbstractScope()) 2936 return; 2937 2938 // Gather information about the lexical scope including local variables, 2939 // global variables, and address ranges. 2940 bool IgnoreScope = false; 2941 auto LI = ScopeVariables.find(&Scope); 2942 SmallVectorImpl<LocalVariable> *Locals = 2943 LI != ScopeVariables.end() ? &LI->second : nullptr; 2944 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2945 SmallVectorImpl<CVGlobalVariable> *Globals = 2946 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2947 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2948 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2949 2950 // Ignore lexical scopes which do not contain variables. 2951 if (!Locals && !Globals) 2952 IgnoreScope = true; 2953 2954 // Ignore lexical scopes which are not lexical blocks. 2955 if (!DILB) 2956 IgnoreScope = true; 2957 2958 // Ignore scopes which have too many address ranges to represent in the 2959 // current CodeView format or do not have a valid address range. 2960 // 2961 // For lexical scopes with multiple address ranges you may be tempted to 2962 // construct a single range covering every instruction where the block is 2963 // live and everything in between. Unfortunately, Visual Studio only 2964 // displays variables from the first matching lexical block scope. If the 2965 // first lexical block contains exception handling code or cold code which 2966 // is moved to the bottom of the routine creating a single range covering 2967 // nearly the entire routine, then it will hide all other lexical blocks 2968 // and the variables they contain. 2969 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2970 IgnoreScope = true; 2971 2972 if (IgnoreScope) { 2973 // This scope can be safely ignored and eliminating it will reduce the 2974 // size of the debug information. Be sure to collect any variable and scope 2975 // information from the this scope or any of its children and collapse them 2976 // into the parent scope. 2977 if (Locals) 2978 ParentLocals.append(Locals->begin(), Locals->end()); 2979 if (Globals) 2980 ParentGlobals.append(Globals->begin(), Globals->end()); 2981 collectLexicalBlockInfo(Scope.getChildren(), 2982 ParentBlocks, 2983 ParentLocals, 2984 ParentGlobals); 2985 return; 2986 } 2987 2988 // Create a new CodeView lexical block for this lexical scope. If we've 2989 // seen this DILexicalBlock before then the scope tree is malformed and 2990 // we can handle this gracefully by not processing it a second time. 2991 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2992 if (!BlockInsertion.second) 2993 return; 2994 2995 // Create a lexical block containing the variables and collect the the 2996 // lexical block information for the children. 2997 const InsnRange &Range = Ranges.front(); 2998 assert(Range.first && Range.second); 2999 LexicalBlock &Block = BlockInsertion.first->second; 3000 Block.Begin = getLabelBeforeInsn(Range.first); 3001 Block.End = getLabelAfterInsn(Range.second); 3002 assert(Block.Begin && "missing label for scope begin"); 3003 assert(Block.End && "missing label for scope end"); 3004 Block.Name = DILB->getName(); 3005 if (Locals) 3006 Block.Locals = std::move(*Locals); 3007 if (Globals) 3008 Block.Globals = std::move(*Globals); 3009 ParentBlocks.push_back(&Block); 3010 collectLexicalBlockInfo(Scope.getChildren(), 3011 Block.Children, 3012 Block.Locals, 3013 Block.Globals); 3014 } 3015 3016 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3017 const Function &GV = MF->getFunction(); 3018 assert(FnDebugInfo.count(&GV)); 3019 assert(CurFn == FnDebugInfo[&GV].get()); 3020 3021 collectVariableInfo(GV.getSubprogram()); 3022 3023 // Build the lexical block structure to emit for this routine. 3024 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3025 collectLexicalBlockInfo(*CFS, 3026 CurFn->ChildBlocks, 3027 CurFn->Locals, 3028 CurFn->Globals); 3029 3030 // Clear the scope and variable information from the map which will not be 3031 // valid after we have finished processing this routine. This also prepares 3032 // the map for the subsequent routine. 3033 ScopeVariables.clear(); 3034 3035 // Don't emit anything if we don't have any line tables. 3036 // Thunks are compiler-generated and probably won't have source correlation. 3037 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3038 FnDebugInfo.erase(&GV); 3039 CurFn = nullptr; 3040 return; 3041 } 3042 3043 // Find heap alloc sites and add to list. 3044 for (const auto &MBB : *MF) { 3045 for (const auto &MI : MBB) { 3046 if (MDNode *MD = MI.getHeapAllocMarker()) { 3047 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3048 getLabelAfterInsn(&MI), 3049 dyn_cast<DIType>(MD))); 3050 } 3051 } 3052 } 3053 3054 CurFn->Annotations = MF->getCodeViewAnnotations(); 3055 3056 CurFn->End = Asm->getFunctionEnd(); 3057 3058 CurFn = nullptr; 3059 } 3060 3061 // Usable locations are valid with non-zero line numbers. A line number of zero 3062 // corresponds to optimized code that doesn't have a distinct source location. 3063 // In this case, we try to use the previous or next source location depending on 3064 // the context. 3065 static bool isUsableDebugLoc(DebugLoc DL) { 3066 return DL && DL.getLine() != 0; 3067 } 3068 3069 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3070 DebugHandlerBase::beginInstruction(MI); 3071 3072 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3073 if (!Asm || !CurFn || MI->isDebugInstr() || 3074 MI->getFlag(MachineInstr::FrameSetup)) 3075 return; 3076 3077 // If the first instruction of a new MBB has no location, find the first 3078 // instruction with a location and use that. 3079 DebugLoc DL = MI->getDebugLoc(); 3080 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3081 for (const auto &NextMI : *MI->getParent()) { 3082 if (NextMI.isDebugInstr()) 3083 continue; 3084 DL = NextMI.getDebugLoc(); 3085 if (isUsableDebugLoc(DL)) 3086 break; 3087 } 3088 // FIXME: Handle the case where the BB has no valid locations. This would 3089 // probably require doing a real dataflow analysis. 3090 } 3091 PrevInstBB = MI->getParent(); 3092 3093 // If we still don't have a debug location, don't record a location. 3094 if (!isUsableDebugLoc(DL)) 3095 return; 3096 3097 maybeRecordLocation(DL, Asm->MF); 3098 } 3099 3100 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3101 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3102 *EndLabel = MMI->getContext().createTempSymbol(); 3103 OS.emitInt32(unsigned(Kind)); 3104 OS.AddComment("Subsection size"); 3105 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3106 OS.emitLabel(BeginLabel); 3107 return EndLabel; 3108 } 3109 3110 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3111 OS.emitLabel(EndLabel); 3112 // Every subsection must be aligned to a 4-byte boundary. 3113 OS.emitValueToAlignment(4); 3114 } 3115 3116 static StringRef getSymbolName(SymbolKind SymKind) { 3117 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3118 if (EE.Value == SymKind) 3119 return EE.Name; 3120 return ""; 3121 } 3122 3123 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3124 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3125 *EndLabel = MMI->getContext().createTempSymbol(); 3126 OS.AddComment("Record length"); 3127 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3128 OS.emitLabel(BeginLabel); 3129 if (OS.isVerboseAsm()) 3130 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3131 OS.emitInt16(unsigned(SymKind)); 3132 return EndLabel; 3133 } 3134 3135 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3136 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3137 // an extra copy of every symbol record in LLD. This increases object file 3138 // size by less than 1% in the clang build, and is compatible with the Visual 3139 // C++ linker. 3140 OS.emitValueToAlignment(4); 3141 OS.emitLabel(SymEnd); 3142 } 3143 3144 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3145 OS.AddComment("Record length"); 3146 OS.emitInt16(2); 3147 if (OS.isVerboseAsm()) 3148 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3149 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3150 } 3151 3152 void CodeViewDebug::emitDebugInfoForUDTs( 3153 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3154 #ifndef NDEBUG 3155 size_t OriginalSize = UDTs.size(); 3156 #endif 3157 for (const auto &UDT : UDTs) { 3158 const DIType *T = UDT.second; 3159 assert(shouldEmitUdt(T)); 3160 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3161 OS.AddComment("Type"); 3162 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3163 assert(OriginalSize == UDTs.size() && 3164 "getCompleteTypeIndex found new UDTs!"); 3165 emitNullTerminatedSymbolName(OS, UDT.first); 3166 endSymbolRecord(UDTRecordEnd); 3167 } 3168 } 3169 3170 void CodeViewDebug::collectGlobalVariableInfo() { 3171 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3172 GlobalMap; 3173 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3174 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3175 GV.getDebugInfo(GVEs); 3176 for (const auto *GVE : GVEs) 3177 GlobalMap[GVE] = &GV; 3178 } 3179 3180 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3181 for (const MDNode *Node : CUs->operands()) { 3182 const auto *CU = cast<DICompileUnit>(Node); 3183 for (const auto *GVE : CU->getGlobalVariables()) { 3184 const DIGlobalVariable *DIGV = GVE->getVariable(); 3185 const DIExpression *DIE = GVE->getExpression(); 3186 3187 if ((DIE->getNumElements() == 2) && 3188 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3189 // Record the constant offset for the variable. 3190 // 3191 // A Fortran common block uses this idiom to encode the offset 3192 // of a variable from the common block's starting address. 3193 CVGlobalVariableOffsets.insert( 3194 std::make_pair(DIGV, DIE->getElement(1))); 3195 3196 // Emit constant global variables in a global symbol section. 3197 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3198 CVGlobalVariable CVGV = {DIGV, DIE}; 3199 GlobalVariables.emplace_back(std::move(CVGV)); 3200 } 3201 3202 const auto *GV = GlobalMap.lookup(GVE); 3203 if (!GV || GV->isDeclarationForLinker()) 3204 continue; 3205 3206 DIScope *Scope = DIGV->getScope(); 3207 SmallVector<CVGlobalVariable, 1> *VariableList; 3208 if (Scope && isa<DILocalScope>(Scope)) { 3209 // Locate a global variable list for this scope, creating one if 3210 // necessary. 3211 auto Insertion = ScopeGlobals.insert( 3212 {Scope, std::unique_ptr<GlobalVariableList>()}); 3213 if (Insertion.second) 3214 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3215 VariableList = Insertion.first->second.get(); 3216 } else if (GV->hasComdat()) 3217 // Emit this global variable into a COMDAT section. 3218 VariableList = &ComdatVariables; 3219 else 3220 // Emit this global variable in a single global symbol section. 3221 VariableList = &GlobalVariables; 3222 CVGlobalVariable CVGV = {DIGV, GV}; 3223 VariableList->emplace_back(std::move(CVGV)); 3224 } 3225 } 3226 } 3227 3228 void CodeViewDebug::collectDebugInfoForGlobals() { 3229 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3230 const DIGlobalVariable *DIGV = CVGV.DIGV; 3231 const DIScope *Scope = DIGV->getScope(); 3232 getCompleteTypeIndex(DIGV->getType()); 3233 getFullyQualifiedName(Scope, DIGV->getName()); 3234 } 3235 3236 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3237 const DIGlobalVariable *DIGV = CVGV.DIGV; 3238 const DIScope *Scope = DIGV->getScope(); 3239 getCompleteTypeIndex(DIGV->getType()); 3240 getFullyQualifiedName(Scope, DIGV->getName()); 3241 } 3242 } 3243 3244 void CodeViewDebug::emitDebugInfoForGlobals() { 3245 // First, emit all globals that are not in a comdat in a single symbol 3246 // substream. MSVC doesn't like it if the substream is empty, so only open 3247 // it if we have at least one global to emit. 3248 switchToDebugSectionForSymbol(nullptr); 3249 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3250 OS.AddComment("Symbol subsection for globals"); 3251 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3252 emitGlobalVariableList(GlobalVariables); 3253 emitStaticConstMemberList(); 3254 endCVSubsection(EndLabel); 3255 } 3256 3257 // Second, emit each global that is in a comdat into its own .debug$S 3258 // section along with its own symbol substream. 3259 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3260 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3261 MCSymbol *GVSym = Asm->getSymbol(GV); 3262 OS.AddComment("Symbol subsection for " + 3263 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3264 switchToDebugSectionForSymbol(GVSym); 3265 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3266 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3267 emitDebugInfoForGlobal(CVGV); 3268 endCVSubsection(EndLabel); 3269 } 3270 } 3271 3272 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3273 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3274 for (const MDNode *Node : CUs->operands()) { 3275 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3276 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3277 getTypeIndex(RT); 3278 // FIXME: Add to global/local DTU list. 3279 } 3280 } 3281 } 3282 } 3283 3284 // Emit each global variable in the specified array. 3285 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3286 for (const CVGlobalVariable &CVGV : Globals) { 3287 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3288 emitDebugInfoForGlobal(CVGV); 3289 } 3290 } 3291 3292 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3293 const std::string &QualifiedName) { 3294 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3295 OS.AddComment("Type"); 3296 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3297 3298 OS.AddComment("Value"); 3299 3300 // Encoded integers shouldn't need more than 10 bytes. 3301 uint8_t Data[10]; 3302 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3303 CodeViewRecordIO IO(Writer); 3304 cantFail(IO.mapEncodedInteger(Value)); 3305 StringRef SRef((char *)Data, Writer.getOffset()); 3306 OS.emitBinaryData(SRef); 3307 3308 OS.AddComment("Name"); 3309 emitNullTerminatedSymbolName(OS, QualifiedName); 3310 endSymbolRecord(SConstantEnd); 3311 } 3312 3313 void CodeViewDebug::emitStaticConstMemberList() { 3314 for (const DIDerivedType *DTy : StaticConstMembers) { 3315 const DIScope *Scope = DTy->getScope(); 3316 3317 APSInt Value; 3318 if (const ConstantInt *CI = 3319 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3320 Value = APSInt(CI->getValue(), 3321 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3322 else if (const ConstantFP *CFP = 3323 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3324 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3325 else 3326 llvm_unreachable("cannot emit a constant without a value"); 3327 3328 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3329 getFullyQualifiedName(Scope, DTy->getName())); 3330 } 3331 } 3332 3333 static bool isFloatDIType(const DIType *Ty) { 3334 if (isa<DICompositeType>(Ty)) 3335 return false; 3336 3337 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3338 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3339 if (T == dwarf::DW_TAG_pointer_type || 3340 T == dwarf::DW_TAG_ptr_to_member_type || 3341 T == dwarf::DW_TAG_reference_type || 3342 T == dwarf::DW_TAG_rvalue_reference_type) 3343 return false; 3344 assert(DTy->getBaseType() && "Expected valid base type"); 3345 return isFloatDIType(DTy->getBaseType()); 3346 } 3347 3348 auto *BTy = cast<DIBasicType>(Ty); 3349 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3350 } 3351 3352 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3353 const DIGlobalVariable *DIGV = CVGV.DIGV; 3354 3355 const DIScope *Scope = DIGV->getScope(); 3356 // For static data members, get the scope from the declaration. 3357 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3358 DIGV->getRawStaticDataMemberDeclaration())) 3359 Scope = MemberDecl->getScope(); 3360 // For Fortran, the scoping portion is elided in its name so that we can 3361 // reference the variable in the command line of the VS debugger. 3362 std::string QualifiedName = 3363 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3364 : getFullyQualifiedName(Scope, DIGV->getName()); 3365 3366 if (const GlobalVariable *GV = 3367 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3368 // DataSym record, see SymbolRecord.h for more info. Thread local data 3369 // happens to have the same format as global data. 3370 MCSymbol *GVSym = Asm->getSymbol(GV); 3371 SymbolKind DataSym = GV->isThreadLocal() 3372 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3373 : SymbolKind::S_GTHREAD32) 3374 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3375 : SymbolKind::S_GDATA32); 3376 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3377 OS.AddComment("Type"); 3378 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3379 OS.AddComment("DataOffset"); 3380 3381 uint64_t Offset = 0; 3382 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3383 // Use the offset seen while collecting info on globals. 3384 Offset = CVGlobalVariableOffsets[DIGV]; 3385 OS.EmitCOFFSecRel32(GVSym, Offset); 3386 3387 OS.AddComment("Segment"); 3388 OS.EmitCOFFSectionIndex(GVSym); 3389 OS.AddComment("Name"); 3390 const unsigned LengthOfDataRecord = 12; 3391 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3392 endSymbolRecord(DataEnd); 3393 } else { 3394 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3395 assert(DIE->isConstant() && 3396 "Global constant variables must contain a constant expression."); 3397 3398 // Use unsigned for floats. 3399 bool isUnsigned = isFloatDIType(DIGV->getType()) 3400 ? true 3401 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3402 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3403 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3404 } 3405 } 3406