1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "llvm/ADT/APSInt.h" 15 #include "llvm/ADT/None.h" 16 #include "llvm/ADT/Optional.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallString.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/ADT/TinyPtrVector.h" 21 #include "llvm/ADT/Triple.h" 22 #include "llvm/ADT/Twine.h" 23 #include "llvm/BinaryFormat/COFF.h" 24 #include "llvm/BinaryFormat/Dwarf.h" 25 #include "llvm/CodeGen/AsmPrinter.h" 26 #include "llvm/CodeGen/LexicalScopes.h" 27 #include "llvm/CodeGen/MachineFrameInfo.h" 28 #include "llvm/CodeGen/MachineFunction.h" 29 #include "llvm/CodeGen/MachineInstr.h" 30 #include "llvm/CodeGen/MachineModuleInfo.h" 31 #include "llvm/CodeGen/TargetFrameLowering.h" 32 #include "llvm/CodeGen/TargetRegisterInfo.h" 33 #include "llvm/CodeGen/TargetSubtargetInfo.h" 34 #include "llvm/Config/llvm-config.h" 35 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 36 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 37 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 38 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 39 #include "llvm/DebugInfo/CodeView/EnumTables.h" 40 #include "llvm/DebugInfo/CodeView/Line.h" 41 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 42 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 43 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 44 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 45 #include "llvm/IR/Constants.h" 46 #include "llvm/IR/DataLayout.h" 47 #include "llvm/IR/DebugInfoMetadata.h" 48 #include "llvm/IR/Function.h" 49 #include "llvm/IR/GlobalValue.h" 50 #include "llvm/IR/GlobalVariable.h" 51 #include "llvm/IR/Metadata.h" 52 #include "llvm/IR/Module.h" 53 #include "llvm/MC/MCAsmInfo.h" 54 #include "llvm/MC/MCContext.h" 55 #include "llvm/MC/MCSectionCOFF.h" 56 #include "llvm/MC/MCStreamer.h" 57 #include "llvm/MC/MCSymbol.h" 58 #include "llvm/Support/BinaryStreamWriter.h" 59 #include "llvm/Support/Casting.h" 60 #include "llvm/Support/Endian.h" 61 #include "llvm/Support/Error.h" 62 #include "llvm/Support/ErrorHandling.h" 63 #include "llvm/Support/FormatVariadic.h" 64 #include "llvm/Support/Path.h" 65 #include "llvm/Support/Program.h" 66 #include "llvm/Support/SMLoc.h" 67 #include "llvm/Support/ScopedPrinter.h" 68 #include "llvm/Target/TargetLoweringObjectFile.h" 69 #include "llvm/Target/TargetMachine.h" 70 #include <algorithm> 71 #include <cassert> 72 #include <cctype> 73 #include <cstddef> 74 #include <iterator> 75 #include <limits> 76 77 using namespace llvm; 78 using namespace llvm::codeview; 79 80 namespace { 81 class CVMCAdapter : public CodeViewRecordStreamer { 82 public: 83 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 84 : OS(&OS), TypeTable(TypeTable) {} 85 86 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 87 88 void emitIntValue(uint64_t Value, unsigned Size) override { 89 OS->emitIntValueInHex(Value, Size); 90 } 91 92 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 93 94 void AddComment(const Twine &T) override { OS->AddComment(T); } 95 96 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 97 98 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 99 100 std::string getTypeName(TypeIndex TI) override { 101 std::string TypeName; 102 if (!TI.isNoneType()) { 103 if (TI.isSimple()) 104 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 105 else 106 TypeName = std::string(TypeTable.getTypeName(TI)); 107 } 108 return TypeName; 109 } 110 111 private: 112 MCStreamer *OS = nullptr; 113 TypeCollection &TypeTable; 114 }; 115 } // namespace 116 117 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 118 switch (Type) { 119 case Triple::ArchType::x86: 120 return CPUType::Pentium3; 121 case Triple::ArchType::x86_64: 122 return CPUType::X64; 123 case Triple::ArchType::thumb: 124 // LLVM currently doesn't support Windows CE and so thumb 125 // here is indiscriminately mapped to ARMNT specifically. 126 return CPUType::ARMNT; 127 case Triple::ArchType::aarch64: 128 return CPUType::ARM64; 129 default: 130 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 131 } 132 } 133 134 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 135 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 136 137 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 138 std::string &Filepath = FileToFilepathMap[File]; 139 if (!Filepath.empty()) 140 return Filepath; 141 142 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 143 144 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 145 // it textually because one of the path components could be a symlink. 146 if (Dir.startswith("/") || Filename.startswith("/")) { 147 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 148 return Filename; 149 Filepath = std::string(Dir); 150 if (Dir.back() != '/') 151 Filepath += '/'; 152 Filepath += Filename; 153 return Filepath; 154 } 155 156 // Clang emits directory and relative filename info into the IR, but CodeView 157 // operates on full paths. We could change Clang to emit full paths too, but 158 // that would increase the IR size and probably not needed for other users. 159 // For now, just concatenate and canonicalize the path here. 160 if (Filename.find(':') == 1) 161 Filepath = std::string(Filename); 162 else 163 Filepath = (Dir + "\\" + Filename).str(); 164 165 // Canonicalize the path. We have to do it textually because we may no longer 166 // have access the file in the filesystem. 167 // First, replace all slashes with backslashes. 168 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 169 170 // Remove all "\.\" with "\". 171 size_t Cursor = 0; 172 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 173 Filepath.erase(Cursor, 2); 174 175 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 176 // path should be well-formatted, e.g. start with a drive letter, etc. 177 Cursor = 0; 178 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 179 // Something's wrong if the path starts with "\..\", abort. 180 if (Cursor == 0) 181 break; 182 183 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 184 if (PrevSlash == std::string::npos) 185 // Something's wrong, abort. 186 break; 187 188 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 189 // The next ".." might be following the one we've just erased. 190 Cursor = PrevSlash; 191 } 192 193 // Remove all duplicate backslashes. 194 Cursor = 0; 195 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 196 Filepath.erase(Cursor, 1); 197 198 return Filepath; 199 } 200 201 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 202 StringRef FullPath = getFullFilepath(F); 203 unsigned NextId = FileIdMap.size() + 1; 204 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 205 if (Insertion.second) { 206 // We have to compute the full filepath and emit a .cv_file directive. 207 ArrayRef<uint8_t> ChecksumAsBytes; 208 FileChecksumKind CSKind = FileChecksumKind::None; 209 if (F->getChecksum()) { 210 std::string Checksum = fromHex(F->getChecksum()->Value); 211 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 212 memcpy(CKMem, Checksum.data(), Checksum.size()); 213 ChecksumAsBytes = ArrayRef<uint8_t>( 214 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 215 switch (F->getChecksum()->Kind) { 216 case DIFile::CSK_MD5: 217 CSKind = FileChecksumKind::MD5; 218 break; 219 case DIFile::CSK_SHA1: 220 CSKind = FileChecksumKind::SHA1; 221 break; 222 case DIFile::CSK_SHA256: 223 CSKind = FileChecksumKind::SHA256; 224 break; 225 } 226 } 227 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 228 static_cast<unsigned>(CSKind)); 229 (void)Success; 230 assert(Success && ".cv_file directive failed"); 231 } 232 return Insertion.first->second; 233 } 234 235 CodeViewDebug::InlineSite & 236 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 237 const DISubprogram *Inlinee) { 238 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 239 InlineSite *Site = &SiteInsertion.first->second; 240 if (SiteInsertion.second) { 241 unsigned ParentFuncId = CurFn->FuncId; 242 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 243 ParentFuncId = 244 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 245 .SiteFuncId; 246 247 Site->SiteFuncId = NextFuncId++; 248 OS.EmitCVInlineSiteIdDirective( 249 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 250 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 251 Site->Inlinee = Inlinee; 252 InlinedSubprograms.insert(Inlinee); 253 getFuncIdForSubprogram(Inlinee); 254 } 255 return *Site; 256 } 257 258 static StringRef getPrettyScopeName(const DIScope *Scope) { 259 StringRef ScopeName = Scope->getName(); 260 if (!ScopeName.empty()) 261 return ScopeName; 262 263 switch (Scope->getTag()) { 264 case dwarf::DW_TAG_enumeration_type: 265 case dwarf::DW_TAG_class_type: 266 case dwarf::DW_TAG_structure_type: 267 case dwarf::DW_TAG_union_type: 268 return "<unnamed-tag>"; 269 case dwarf::DW_TAG_namespace: 270 return "`anonymous namespace'"; 271 default: 272 return StringRef(); 273 } 274 } 275 276 const DISubprogram *CodeViewDebug::collectParentScopeNames( 277 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 278 const DISubprogram *ClosestSubprogram = nullptr; 279 while (Scope != nullptr) { 280 if (ClosestSubprogram == nullptr) 281 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 282 283 // If a type appears in a scope chain, make sure it gets emitted. The 284 // frontend will be responsible for deciding if this should be a forward 285 // declaration or a complete type. 286 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 287 DeferredCompleteTypes.push_back(Ty); 288 289 StringRef ScopeName = getPrettyScopeName(Scope); 290 if (!ScopeName.empty()) 291 QualifiedNameComponents.push_back(ScopeName); 292 Scope = Scope->getScope(); 293 } 294 return ClosestSubprogram; 295 } 296 297 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 298 StringRef TypeName) { 299 std::string FullyQualifiedName; 300 for (StringRef QualifiedNameComponent : 301 llvm::reverse(QualifiedNameComponents)) { 302 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 303 FullyQualifiedName.append("::"); 304 } 305 FullyQualifiedName.append(std::string(TypeName)); 306 return FullyQualifiedName; 307 } 308 309 struct CodeViewDebug::TypeLoweringScope { 310 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 311 ~TypeLoweringScope() { 312 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 313 // inner TypeLoweringScopes don't attempt to emit deferred types. 314 if (CVD.TypeEmissionLevel == 1) 315 CVD.emitDeferredCompleteTypes(); 316 --CVD.TypeEmissionLevel; 317 } 318 CodeViewDebug &CVD; 319 }; 320 321 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 322 StringRef Name) { 323 // Ensure types in the scope chain are emitted as soon as possible. 324 // This can create otherwise a situation where S_UDTs are emitted while 325 // looping in emitDebugInfoForUDTs. 326 TypeLoweringScope S(*this); 327 SmallVector<StringRef, 5> QualifiedNameComponents; 328 collectParentScopeNames(Scope, QualifiedNameComponents); 329 return formatNestedName(QualifiedNameComponents, Name); 330 } 331 332 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 333 const DIScope *Scope = Ty->getScope(); 334 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 335 } 336 337 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 338 // No scope means global scope and that uses the zero index. 339 // 340 // We also use zero index when the scope is a DISubprogram 341 // to suppress the emission of LF_STRING_ID for the function, 342 // which can trigger a link-time error with the linker in 343 // VS2019 version 16.11.2 or newer. 344 // Note, however, skipping the debug info emission for the DISubprogram 345 // is a temporary fix. The root issue here is that we need to figure out 346 // the proper way to encode a function nested in another function 347 // (as introduced by the Fortran 'contains' keyword) in CodeView. 348 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 349 return TypeIndex(); 350 351 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 352 353 // Check if we've already translated this scope. 354 auto I = TypeIndices.find({Scope, nullptr}); 355 if (I != TypeIndices.end()) 356 return I->second; 357 358 // Build the fully qualified name of the scope. 359 std::string ScopeName = getFullyQualifiedName(Scope); 360 StringIdRecord SID(TypeIndex(), ScopeName); 361 auto TI = TypeTable.writeLeafType(SID); 362 return recordTypeIndexForDINode(Scope, TI); 363 } 364 365 static StringRef removeTemplateArgs(StringRef Name) { 366 // Remove template args from the display name. Assume that the template args 367 // are the last thing in the name. 368 if (Name.empty() || Name.back() != '>') 369 return Name; 370 371 int OpenBrackets = 0; 372 for (int i = Name.size() - 1; i >= 0; --i) { 373 if (Name[i] == '>') 374 ++OpenBrackets; 375 else if (Name[i] == '<') { 376 --OpenBrackets; 377 if (OpenBrackets == 0) 378 return Name.substr(0, i); 379 } 380 } 381 return Name; 382 } 383 384 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 385 assert(SP); 386 387 // Check if we've already translated this subprogram. 388 auto I = TypeIndices.find({SP, nullptr}); 389 if (I != TypeIndices.end()) 390 return I->second; 391 392 // The display name includes function template arguments. Drop them to match 393 // MSVC. We need to have the template arguments in the DISubprogram name 394 // because they are used in other symbol records, such as S_GPROC32_IDs. 395 StringRef DisplayName = removeTemplateArgs(SP->getName()); 396 397 const DIScope *Scope = SP->getScope(); 398 TypeIndex TI; 399 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 400 // If the scope is a DICompositeType, then this must be a method. Member 401 // function types take some special handling, and require access to the 402 // subprogram. 403 TypeIndex ClassType = getTypeIndex(Class); 404 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 405 DisplayName); 406 TI = TypeTable.writeLeafType(MFuncId); 407 } else { 408 // Otherwise, this must be a free function. 409 TypeIndex ParentScope = getScopeIndex(Scope); 410 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 411 TI = TypeTable.writeLeafType(FuncId); 412 } 413 414 return recordTypeIndexForDINode(SP, TI); 415 } 416 417 static bool isNonTrivial(const DICompositeType *DCTy) { 418 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 419 } 420 421 static FunctionOptions 422 getFunctionOptions(const DISubroutineType *Ty, 423 const DICompositeType *ClassTy = nullptr, 424 StringRef SPName = StringRef("")) { 425 FunctionOptions FO = FunctionOptions::None; 426 const DIType *ReturnTy = nullptr; 427 if (auto TypeArray = Ty->getTypeArray()) { 428 if (TypeArray.size()) 429 ReturnTy = TypeArray[0]; 430 } 431 432 // Add CxxReturnUdt option to functions that return nontrivial record types 433 // or methods that return record types. 434 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 435 if (isNonTrivial(ReturnDCTy) || ClassTy) 436 FO |= FunctionOptions::CxxReturnUdt; 437 438 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 439 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 440 FO |= FunctionOptions::Constructor; 441 442 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 443 444 } 445 return FO; 446 } 447 448 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 449 const DICompositeType *Class) { 450 // Always use the method declaration as the key for the function type. The 451 // method declaration contains the this adjustment. 452 if (SP->getDeclaration()) 453 SP = SP->getDeclaration(); 454 assert(!SP->getDeclaration() && "should use declaration as key"); 455 456 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 457 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 458 auto I = TypeIndices.find({SP, Class}); 459 if (I != TypeIndices.end()) 460 return I->second; 461 462 // Make sure complete type info for the class is emitted *after* the member 463 // function type, as the complete class type is likely to reference this 464 // member function type. 465 TypeLoweringScope S(*this); 466 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 467 468 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 469 TypeIndex TI = lowerTypeMemberFunction( 470 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 471 return recordTypeIndexForDINode(SP, TI, Class); 472 } 473 474 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 475 TypeIndex TI, 476 const DIType *ClassTy) { 477 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 478 (void)InsertResult; 479 assert(InsertResult.second && "DINode was already assigned a type index"); 480 return TI; 481 } 482 483 unsigned CodeViewDebug::getPointerSizeInBytes() { 484 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 485 } 486 487 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 488 const LexicalScope *LS) { 489 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 490 // This variable was inlined. Associate it with the InlineSite. 491 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 492 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 493 Site.InlinedLocals.emplace_back(Var); 494 } else { 495 // This variable goes into the corresponding lexical scope. 496 ScopeVariables[LS].emplace_back(Var); 497 } 498 } 499 500 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 501 const DILocation *Loc) { 502 if (!llvm::is_contained(Locs, Loc)) 503 Locs.push_back(Loc); 504 } 505 506 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 507 const MachineFunction *MF) { 508 // Skip this instruction if it has the same location as the previous one. 509 if (!DL || DL == PrevInstLoc) 510 return; 511 512 const DIScope *Scope = DL->getScope(); 513 if (!Scope) 514 return; 515 516 // Skip this line if it is longer than the maximum we can record. 517 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 518 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 519 LI.isNeverStepInto()) 520 return; 521 522 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 523 if (CI.getStartColumn() != DL.getCol()) 524 return; 525 526 if (!CurFn->HaveLineInfo) 527 CurFn->HaveLineInfo = true; 528 unsigned FileId = 0; 529 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 530 FileId = CurFn->LastFileId; 531 else 532 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 533 PrevInstLoc = DL; 534 535 unsigned FuncId = CurFn->FuncId; 536 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 537 const DILocation *Loc = DL.get(); 538 539 // If this location was actually inlined from somewhere else, give it the ID 540 // of the inline call site. 541 FuncId = 542 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 543 544 // Ensure we have links in the tree of inline call sites. 545 bool FirstLoc = true; 546 while ((SiteLoc = Loc->getInlinedAt())) { 547 InlineSite &Site = 548 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 549 if (!FirstLoc) 550 addLocIfNotPresent(Site.ChildSites, Loc); 551 FirstLoc = false; 552 Loc = SiteLoc; 553 } 554 addLocIfNotPresent(CurFn->ChildSites, Loc); 555 } 556 557 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 558 /*PrologueEnd=*/false, /*IsStmt=*/false, 559 DL->getFilename(), SMLoc()); 560 } 561 562 void CodeViewDebug::emitCodeViewMagicVersion() { 563 OS.emitValueToAlignment(4); 564 OS.AddComment("Debug section magic"); 565 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 566 } 567 568 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 569 switch (DWLang) { 570 case dwarf::DW_LANG_C: 571 case dwarf::DW_LANG_C89: 572 case dwarf::DW_LANG_C99: 573 case dwarf::DW_LANG_C11: 574 case dwarf::DW_LANG_ObjC: 575 return SourceLanguage::C; 576 case dwarf::DW_LANG_C_plus_plus: 577 case dwarf::DW_LANG_C_plus_plus_03: 578 case dwarf::DW_LANG_C_plus_plus_11: 579 case dwarf::DW_LANG_C_plus_plus_14: 580 return SourceLanguage::Cpp; 581 case dwarf::DW_LANG_Fortran77: 582 case dwarf::DW_LANG_Fortran90: 583 case dwarf::DW_LANG_Fortran95: 584 case dwarf::DW_LANG_Fortran03: 585 case dwarf::DW_LANG_Fortran08: 586 return SourceLanguage::Fortran; 587 case dwarf::DW_LANG_Pascal83: 588 return SourceLanguage::Pascal; 589 case dwarf::DW_LANG_Cobol74: 590 case dwarf::DW_LANG_Cobol85: 591 return SourceLanguage::Cobol; 592 case dwarf::DW_LANG_Java: 593 return SourceLanguage::Java; 594 case dwarf::DW_LANG_D: 595 return SourceLanguage::D; 596 case dwarf::DW_LANG_Swift: 597 return SourceLanguage::Swift; 598 case dwarf::DW_LANG_Rust: 599 return SourceLanguage::Rust; 600 default: 601 // There's no CodeView representation for this language, and CV doesn't 602 // have an "unknown" option for the language field, so we'll use MASM, 603 // as it's very low level. 604 return SourceLanguage::Masm; 605 } 606 } 607 608 void CodeViewDebug::beginModule(Module *M) { 609 // If module doesn't have named metadata anchors or COFF debug section 610 // is not available, skip any debug info related stuff. 611 if (!MMI->hasDebugInfo() || 612 !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 613 Asm = nullptr; 614 return; 615 } 616 617 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 618 619 // Get the current source language. 620 const MDNode *Node = *M->debug_compile_units_begin(); 621 const auto *CU = cast<DICompileUnit>(Node); 622 623 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 624 625 collectGlobalVariableInfo(); 626 627 // Check if we should emit type record hashes. 628 ConstantInt *GH = 629 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 630 EmitDebugGlobalHashes = GH && !GH->isZero(); 631 } 632 633 void CodeViewDebug::endModule() { 634 if (!Asm || !MMI->hasDebugInfo()) 635 return; 636 637 // The COFF .debug$S section consists of several subsections, each starting 638 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 639 // of the payload followed by the payload itself. The subsections are 4-byte 640 // aligned. 641 642 // Use the generic .debug$S section, and make a subsection for all the inlined 643 // subprograms. 644 switchToDebugSectionForSymbol(nullptr); 645 646 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 647 emitObjName(); 648 emitCompilerInformation(); 649 endCVSubsection(CompilerInfo); 650 651 emitInlineeLinesSubsection(); 652 653 // Emit per-function debug information. 654 for (auto &P : FnDebugInfo) 655 if (!P.first->isDeclarationForLinker()) 656 emitDebugInfoForFunction(P.first, *P.second); 657 658 // Get types used by globals without emitting anything. 659 // This is meant to collect all static const data members so they can be 660 // emitted as globals. 661 collectDebugInfoForGlobals(); 662 663 // Emit retained types. 664 emitDebugInfoForRetainedTypes(); 665 666 // Emit global variable debug information. 667 setCurrentSubprogram(nullptr); 668 emitDebugInfoForGlobals(); 669 670 // Switch back to the generic .debug$S section after potentially processing 671 // comdat symbol sections. 672 switchToDebugSectionForSymbol(nullptr); 673 674 // Emit UDT records for any types used by global variables. 675 if (!GlobalUDTs.empty()) { 676 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 677 emitDebugInfoForUDTs(GlobalUDTs); 678 endCVSubsection(SymbolsEnd); 679 } 680 681 // This subsection holds a file index to offset in string table table. 682 OS.AddComment("File index to string table offset subsection"); 683 OS.emitCVFileChecksumsDirective(); 684 685 // This subsection holds the string table. 686 OS.AddComment("String table"); 687 OS.emitCVStringTableDirective(); 688 689 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 690 // subsection in the generic .debug$S section at the end. There is no 691 // particular reason for this ordering other than to match MSVC. 692 emitBuildInfo(); 693 694 // Emit type information and hashes last, so that any types we translate while 695 // emitting function info are included. 696 emitTypeInformation(); 697 698 if (EmitDebugGlobalHashes) 699 emitTypeGlobalHashes(); 700 701 clear(); 702 } 703 704 static void 705 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 706 unsigned MaxFixedRecordLength = 0xF00) { 707 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 708 // after a fixed length portion of the record. The fixed length portion should 709 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 710 // overall record size is less than the maximum allowed. 711 SmallString<32> NullTerminatedString( 712 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 713 NullTerminatedString.push_back('\0'); 714 OS.emitBytes(NullTerminatedString); 715 } 716 717 void CodeViewDebug::emitTypeInformation() { 718 if (TypeTable.empty()) 719 return; 720 721 // Start the .debug$T or .debug$P section with 0x4. 722 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 723 emitCodeViewMagicVersion(); 724 725 TypeTableCollection Table(TypeTable.records()); 726 TypeVisitorCallbackPipeline Pipeline; 727 728 // To emit type record using Codeview MCStreamer adapter 729 CVMCAdapter CVMCOS(OS, Table); 730 TypeRecordMapping typeMapping(CVMCOS); 731 Pipeline.addCallbackToPipeline(typeMapping); 732 733 Optional<TypeIndex> B = Table.getFirst(); 734 while (B) { 735 // This will fail if the record data is invalid. 736 CVType Record = Table.getType(*B); 737 738 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 739 740 if (E) { 741 logAllUnhandledErrors(std::move(E), errs(), "error: "); 742 llvm_unreachable("produced malformed type record"); 743 } 744 745 B = Table.getNext(*B); 746 } 747 } 748 749 void CodeViewDebug::emitTypeGlobalHashes() { 750 if (TypeTable.empty()) 751 return; 752 753 // Start the .debug$H section with the version and hash algorithm, currently 754 // hardcoded to version 0, SHA1. 755 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 756 757 OS.emitValueToAlignment(4); 758 OS.AddComment("Magic"); 759 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 760 OS.AddComment("Section Version"); 761 OS.emitInt16(0); 762 OS.AddComment("Hash Algorithm"); 763 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 764 765 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 766 for (const auto &GHR : TypeTable.hashes()) { 767 if (OS.isVerboseAsm()) { 768 // Emit an EOL-comment describing which TypeIndex this hash corresponds 769 // to, as well as the stringified SHA1 hash. 770 SmallString<32> Comment; 771 raw_svector_ostream CommentOS(Comment); 772 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 773 OS.AddComment(Comment); 774 ++TI; 775 } 776 assert(GHR.Hash.size() == 8); 777 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 778 GHR.Hash.size()); 779 OS.emitBinaryData(S); 780 } 781 } 782 783 void CodeViewDebug::emitObjName() { 784 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 785 786 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 787 llvm::SmallString<256> PathStore(PathRef); 788 789 if (PathRef.empty() || PathRef == "-") { 790 // Don't emit the filename if we're writing to stdout or to /dev/null. 791 PathRef = {}; 792 } else { 793 llvm::sys::path::remove_dots(PathStore, /*remove_dot_dot=*/true); 794 PathRef = PathStore; 795 } 796 797 OS.AddComment("Signature"); 798 OS.emitIntValue(0, 4); 799 800 OS.AddComment("Object name"); 801 emitNullTerminatedSymbolName(OS, PathRef); 802 803 endSymbolRecord(CompilerEnd); 804 } 805 806 namespace { 807 struct Version { 808 int Part[4]; 809 }; 810 } // end anonymous namespace 811 812 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 813 // the version number. 814 static Version parseVersion(StringRef Name) { 815 Version V = {{0}}; 816 int N = 0; 817 for (const char C : Name) { 818 if (isdigit(C)) { 819 V.Part[N] *= 10; 820 V.Part[N] += C - '0'; 821 V.Part[N] = 822 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 823 } else if (C == '.') { 824 ++N; 825 if (N >= 4) 826 return V; 827 } else if (N > 0) 828 return V; 829 } 830 return V; 831 } 832 833 void CodeViewDebug::emitCompilerInformation() { 834 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 835 uint32_t Flags = 0; 836 837 // The low byte of the flags indicates the source language. 838 Flags = CurrentSourceLanguage; 839 // TODO: Figure out which other flags need to be set. 840 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 841 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 842 } 843 using ArchType = llvm::Triple::ArchType; 844 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 845 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 846 Arch == ArchType::aarch64) { 847 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 848 } 849 850 OS.AddComment("Flags and language"); 851 OS.emitInt32(Flags); 852 853 OS.AddComment("CPUType"); 854 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 855 856 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 857 const MDNode *Node = *CUs->operands().begin(); 858 const auto *CU = cast<DICompileUnit>(Node); 859 860 StringRef CompilerVersion = CU->getProducer(); 861 Version FrontVer = parseVersion(CompilerVersion); 862 OS.AddComment("Frontend version"); 863 for (int N : FrontVer.Part) { 864 OS.emitInt16(N); 865 } 866 867 // Some Microsoft tools, like Binscope, expect a backend version number of at 868 // least 8.something, so we'll coerce the LLVM version into a form that 869 // guarantees it'll be big enough without really lying about the version. 870 int Major = 1000 * LLVM_VERSION_MAJOR + 871 10 * LLVM_VERSION_MINOR + 872 LLVM_VERSION_PATCH; 873 // Clamp it for builds that use unusually large version numbers. 874 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 875 Version BackVer = {{ Major, 0, 0, 0 }}; 876 OS.AddComment("Backend version"); 877 for (int N : BackVer.Part) 878 OS.emitInt16(N); 879 880 OS.AddComment("Null-terminated compiler version string"); 881 emitNullTerminatedSymbolName(OS, CompilerVersion); 882 883 endSymbolRecord(CompilerEnd); 884 } 885 886 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 887 StringRef S) { 888 StringIdRecord SIR(TypeIndex(0x0), S); 889 return TypeTable.writeLeafType(SIR); 890 } 891 892 static std::string flattenCommandLine(ArrayRef<std::string> Args, 893 StringRef MainFilename) { 894 std::string FlatCmdLine; 895 raw_string_ostream OS(FlatCmdLine); 896 bool PrintedOneArg = false; 897 if (!StringRef(Args[0]).contains("-cc1")) { 898 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 899 PrintedOneArg = true; 900 } 901 for (unsigned i = 0; i < Args.size(); i++) { 902 StringRef Arg = Args[i]; 903 if (Arg.empty()) 904 continue; 905 if (Arg == "-main-file-name" || Arg == "-o") { 906 i++; // Skip this argument and next one. 907 continue; 908 } 909 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 910 continue; 911 if (PrintedOneArg) 912 OS << " "; 913 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 914 PrintedOneArg = true; 915 } 916 OS.flush(); 917 return FlatCmdLine; 918 } 919 920 void CodeViewDebug::emitBuildInfo() { 921 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 922 // build info. The known prefix is: 923 // - Absolute path of current directory 924 // - Compiler path 925 // - Main source file path, relative to CWD or absolute 926 // - Type server PDB file 927 // - Canonical compiler command line 928 // If frontend and backend compilation are separated (think llc or LTO), it's 929 // not clear if the compiler path should refer to the executable for the 930 // frontend or the backend. Leave it blank for now. 931 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 932 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 933 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 934 const auto *CU = cast<DICompileUnit>(Node); 935 const DIFile *MainSourceFile = CU->getFile(); 936 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 937 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 938 BuildInfoArgs[BuildInfoRecord::SourceFile] = 939 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 940 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 941 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 942 getStringIdTypeIdx(TypeTable, ""); 943 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 944 BuildInfoArgs[BuildInfoRecord::BuildTool] = 945 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 946 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 947 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 948 MainSourceFile->getFilename())); 949 } 950 BuildInfoRecord BIR(BuildInfoArgs); 951 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 952 953 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 954 // from the module symbols into the type stream. 955 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 956 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 957 OS.AddComment("LF_BUILDINFO index"); 958 OS.emitInt32(BuildInfoIndex.getIndex()); 959 endSymbolRecord(BIEnd); 960 endCVSubsection(BISubsecEnd); 961 } 962 963 void CodeViewDebug::emitInlineeLinesSubsection() { 964 if (InlinedSubprograms.empty()) 965 return; 966 967 OS.AddComment("Inlinee lines subsection"); 968 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 969 970 // We emit the checksum info for files. This is used by debuggers to 971 // determine if a pdb matches the source before loading it. Visual Studio, 972 // for instance, will display a warning that the breakpoints are not valid if 973 // the pdb does not match the source. 974 OS.AddComment("Inlinee lines signature"); 975 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 976 977 for (const DISubprogram *SP : InlinedSubprograms) { 978 assert(TypeIndices.count({SP, nullptr})); 979 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 980 981 OS.AddBlankLine(); 982 unsigned FileId = maybeRecordFile(SP->getFile()); 983 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 984 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 985 OS.AddBlankLine(); 986 OS.AddComment("Type index of inlined function"); 987 OS.emitInt32(InlineeIdx.getIndex()); 988 OS.AddComment("Offset into filechecksum table"); 989 OS.emitCVFileChecksumOffsetDirective(FileId); 990 OS.AddComment("Starting line number"); 991 OS.emitInt32(SP->getLine()); 992 } 993 994 endCVSubsection(InlineEnd); 995 } 996 997 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 998 const DILocation *InlinedAt, 999 const InlineSite &Site) { 1000 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1001 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1002 1003 // SymbolRecord 1004 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1005 1006 OS.AddComment("PtrParent"); 1007 OS.emitInt32(0); 1008 OS.AddComment("PtrEnd"); 1009 OS.emitInt32(0); 1010 OS.AddComment("Inlinee type index"); 1011 OS.emitInt32(InlineeIdx.getIndex()); 1012 1013 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1014 unsigned StartLineNum = Site.Inlinee->getLine(); 1015 1016 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1017 FI.Begin, FI.End); 1018 1019 endSymbolRecord(InlineEnd); 1020 1021 emitLocalVariableList(FI, Site.InlinedLocals); 1022 1023 // Recurse on child inlined call sites before closing the scope. 1024 for (const DILocation *ChildSite : Site.ChildSites) { 1025 auto I = FI.InlineSites.find(ChildSite); 1026 assert(I != FI.InlineSites.end() && 1027 "child site not in function inline site map"); 1028 emitInlinedCallSite(FI, ChildSite, I->second); 1029 } 1030 1031 // Close the scope. 1032 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1033 } 1034 1035 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1036 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1037 // comdat key. A section may be comdat because of -ffunction-sections or 1038 // because it is comdat in the IR. 1039 MCSectionCOFF *GVSec = 1040 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1041 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1042 1043 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1044 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1045 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1046 1047 OS.SwitchSection(DebugSec); 1048 1049 // Emit the magic version number if this is the first time we've switched to 1050 // this section. 1051 if (ComdatDebugSections.insert(DebugSec).second) 1052 emitCodeViewMagicVersion(); 1053 } 1054 1055 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1056 // The only supported thunk ordinal is currently the standard type. 1057 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1058 FunctionInfo &FI, 1059 const MCSymbol *Fn) { 1060 std::string FuncName = 1061 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1062 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1063 1064 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1065 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1066 1067 // Emit S_THUNK32 1068 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1069 OS.AddComment("PtrParent"); 1070 OS.emitInt32(0); 1071 OS.AddComment("PtrEnd"); 1072 OS.emitInt32(0); 1073 OS.AddComment("PtrNext"); 1074 OS.emitInt32(0); 1075 OS.AddComment("Thunk section relative address"); 1076 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1077 OS.AddComment("Thunk section index"); 1078 OS.EmitCOFFSectionIndex(Fn); 1079 OS.AddComment("Code size"); 1080 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1081 OS.AddComment("Ordinal"); 1082 OS.emitInt8(unsigned(ordinal)); 1083 OS.AddComment("Function name"); 1084 emitNullTerminatedSymbolName(OS, FuncName); 1085 // Additional fields specific to the thunk ordinal would go here. 1086 endSymbolRecord(ThunkRecordEnd); 1087 1088 // Local variables/inlined routines are purposely omitted here. The point of 1089 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1090 1091 // Emit S_PROC_ID_END 1092 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1093 1094 endCVSubsection(SymbolsEnd); 1095 } 1096 1097 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1098 FunctionInfo &FI) { 1099 // For each function there is a separate subsection which holds the PC to 1100 // file:line table. 1101 const MCSymbol *Fn = Asm->getSymbol(GV); 1102 assert(Fn); 1103 1104 // Switch to the to a comdat section, if appropriate. 1105 switchToDebugSectionForSymbol(Fn); 1106 1107 std::string FuncName; 1108 auto *SP = GV->getSubprogram(); 1109 assert(SP); 1110 setCurrentSubprogram(SP); 1111 1112 if (SP->isThunk()) { 1113 emitDebugInfoForThunk(GV, FI, Fn); 1114 return; 1115 } 1116 1117 // If we have a display name, build the fully qualified name by walking the 1118 // chain of scopes. 1119 if (!SP->getName().empty()) 1120 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1121 1122 // If our DISubprogram name is empty, use the mangled name. 1123 if (FuncName.empty()) 1124 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1125 1126 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1127 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1128 OS.EmitCVFPOData(Fn); 1129 1130 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1131 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1132 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1133 { 1134 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1135 : SymbolKind::S_GPROC32_ID; 1136 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1137 1138 // These fields are filled in by tools like CVPACK which run after the fact. 1139 OS.AddComment("PtrParent"); 1140 OS.emitInt32(0); 1141 OS.AddComment("PtrEnd"); 1142 OS.emitInt32(0); 1143 OS.AddComment("PtrNext"); 1144 OS.emitInt32(0); 1145 // This is the important bit that tells the debugger where the function 1146 // code is located and what's its size: 1147 OS.AddComment("Code size"); 1148 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1149 OS.AddComment("Offset after prologue"); 1150 OS.emitInt32(0); 1151 OS.AddComment("Offset before epilogue"); 1152 OS.emitInt32(0); 1153 OS.AddComment("Function type index"); 1154 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1155 OS.AddComment("Function section relative address"); 1156 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1157 OS.AddComment("Function section index"); 1158 OS.EmitCOFFSectionIndex(Fn); 1159 OS.AddComment("Flags"); 1160 OS.emitInt8(0); 1161 // Emit the function display name as a null-terminated string. 1162 OS.AddComment("Function name"); 1163 // Truncate the name so we won't overflow the record length field. 1164 emitNullTerminatedSymbolName(OS, FuncName); 1165 endSymbolRecord(ProcRecordEnd); 1166 1167 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1168 // Subtract out the CSR size since MSVC excludes that and we include it. 1169 OS.AddComment("FrameSize"); 1170 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1171 OS.AddComment("Padding"); 1172 OS.emitInt32(0); 1173 OS.AddComment("Offset of padding"); 1174 OS.emitInt32(0); 1175 OS.AddComment("Bytes of callee saved registers"); 1176 OS.emitInt32(FI.CSRSize); 1177 OS.AddComment("Exception handler offset"); 1178 OS.emitInt32(0); 1179 OS.AddComment("Exception handler section"); 1180 OS.emitInt16(0); 1181 OS.AddComment("Flags (defines frame register)"); 1182 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1183 endSymbolRecord(FrameProcEnd); 1184 1185 emitLocalVariableList(FI, FI.Locals); 1186 emitGlobalVariableList(FI.Globals); 1187 emitLexicalBlockList(FI.ChildBlocks, FI); 1188 1189 // Emit inlined call site information. Only emit functions inlined directly 1190 // into the parent function. We'll emit the other sites recursively as part 1191 // of their parent inline site. 1192 for (const DILocation *InlinedAt : FI.ChildSites) { 1193 auto I = FI.InlineSites.find(InlinedAt); 1194 assert(I != FI.InlineSites.end() && 1195 "child site not in function inline site map"); 1196 emitInlinedCallSite(FI, InlinedAt, I->second); 1197 } 1198 1199 for (auto Annot : FI.Annotations) { 1200 MCSymbol *Label = Annot.first; 1201 MDTuple *Strs = cast<MDTuple>(Annot.second); 1202 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1203 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1204 // FIXME: Make sure we don't overflow the max record size. 1205 OS.EmitCOFFSectionIndex(Label); 1206 OS.emitInt16(Strs->getNumOperands()); 1207 for (Metadata *MD : Strs->operands()) { 1208 // MDStrings are null terminated, so we can do EmitBytes and get the 1209 // nice .asciz directive. 1210 StringRef Str = cast<MDString>(MD)->getString(); 1211 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1212 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1213 } 1214 endSymbolRecord(AnnotEnd); 1215 } 1216 1217 for (auto HeapAllocSite : FI.HeapAllocSites) { 1218 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1219 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1220 const DIType *DITy = std::get<2>(HeapAllocSite); 1221 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1222 OS.AddComment("Call site offset"); 1223 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1224 OS.AddComment("Call site section index"); 1225 OS.EmitCOFFSectionIndex(BeginLabel); 1226 OS.AddComment("Call instruction length"); 1227 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1228 OS.AddComment("Type index"); 1229 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1230 endSymbolRecord(HeapAllocEnd); 1231 } 1232 1233 if (SP != nullptr) 1234 emitDebugInfoForUDTs(LocalUDTs); 1235 1236 // We're done with this function. 1237 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1238 } 1239 endCVSubsection(SymbolsEnd); 1240 1241 // We have an assembler directive that takes care of the whole line table. 1242 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1243 } 1244 1245 CodeViewDebug::LocalVarDefRange 1246 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1247 LocalVarDefRange DR; 1248 DR.InMemory = -1; 1249 DR.DataOffset = Offset; 1250 assert(DR.DataOffset == Offset && "truncation"); 1251 DR.IsSubfield = 0; 1252 DR.StructOffset = 0; 1253 DR.CVRegister = CVRegister; 1254 return DR; 1255 } 1256 1257 void CodeViewDebug::collectVariableInfoFromMFTable( 1258 DenseSet<InlinedEntity> &Processed) { 1259 const MachineFunction &MF = *Asm->MF; 1260 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1261 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1262 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1263 1264 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1265 if (!VI.Var) 1266 continue; 1267 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1268 "Expected inlined-at fields to agree"); 1269 1270 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1271 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1272 1273 // If variable scope is not found then skip this variable. 1274 if (!Scope) 1275 continue; 1276 1277 // If the variable has an attached offset expression, extract it. 1278 // FIXME: Try to handle DW_OP_deref as well. 1279 int64_t ExprOffset = 0; 1280 bool Deref = false; 1281 if (VI.Expr) { 1282 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1283 if (VI.Expr->getNumElements() == 1 && 1284 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1285 Deref = true; 1286 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1287 continue; 1288 } 1289 1290 // Get the frame register used and the offset. 1291 Register FrameReg; 1292 StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1293 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1294 1295 assert(!FrameOffset.getScalable() && 1296 "Frame offsets with a scalable component are not supported"); 1297 1298 // Calculate the label ranges. 1299 LocalVarDefRange DefRange = 1300 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1301 1302 for (const InsnRange &Range : Scope->getRanges()) { 1303 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1304 const MCSymbol *End = getLabelAfterInsn(Range.second); 1305 End = End ? End : Asm->getFunctionEnd(); 1306 DefRange.Ranges.emplace_back(Begin, End); 1307 } 1308 1309 LocalVariable Var; 1310 Var.DIVar = VI.Var; 1311 Var.DefRanges.emplace_back(std::move(DefRange)); 1312 if (Deref) 1313 Var.UseReferenceType = true; 1314 1315 recordLocalVariable(std::move(Var), Scope); 1316 } 1317 } 1318 1319 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1320 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1321 } 1322 1323 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1324 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1325 } 1326 1327 void CodeViewDebug::calculateRanges( 1328 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1329 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1330 1331 // Calculate the definition ranges. 1332 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1333 const auto &Entry = *I; 1334 if (!Entry.isDbgValue()) 1335 continue; 1336 const MachineInstr *DVInst = Entry.getInstr(); 1337 assert(DVInst->isDebugValue() && "Invalid History entry"); 1338 // FIXME: Find a way to represent constant variables, since they are 1339 // relatively common. 1340 Optional<DbgVariableLocation> Location = 1341 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1342 if (!Location) 1343 continue; 1344 1345 // CodeView can only express variables in register and variables in memory 1346 // at a constant offset from a register. However, for variables passed 1347 // indirectly by pointer, it is common for that pointer to be spilled to a 1348 // stack location. For the special case of one offseted load followed by a 1349 // zero offset load (a pointer spilled to the stack), we change the type of 1350 // the local variable from a value type to a reference type. This tricks the 1351 // debugger into doing the load for us. 1352 if (Var.UseReferenceType) { 1353 // We're using a reference type. Drop the last zero offset load. 1354 if (canUseReferenceType(*Location)) 1355 Location->LoadChain.pop_back(); 1356 else 1357 continue; 1358 } else if (needsReferenceType(*Location)) { 1359 // This location can't be expressed without switching to a reference type. 1360 // Start over using that. 1361 Var.UseReferenceType = true; 1362 Var.DefRanges.clear(); 1363 calculateRanges(Var, Entries); 1364 return; 1365 } 1366 1367 // We can only handle a register or an offseted load of a register. 1368 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1369 continue; 1370 { 1371 LocalVarDefRange DR; 1372 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1373 DR.InMemory = !Location->LoadChain.empty(); 1374 DR.DataOffset = 1375 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1376 if (Location->FragmentInfo) { 1377 DR.IsSubfield = true; 1378 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1379 } else { 1380 DR.IsSubfield = false; 1381 DR.StructOffset = 0; 1382 } 1383 1384 if (Var.DefRanges.empty() || 1385 Var.DefRanges.back().isDifferentLocation(DR)) { 1386 Var.DefRanges.emplace_back(std::move(DR)); 1387 } 1388 } 1389 1390 // Compute the label range. 1391 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1392 const MCSymbol *End; 1393 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1394 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1395 End = EndingEntry.isDbgValue() 1396 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1397 : getLabelAfterInsn(EndingEntry.getInstr()); 1398 } else 1399 End = Asm->getFunctionEnd(); 1400 1401 // If the last range end is our begin, just extend the last range. 1402 // Otherwise make a new range. 1403 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1404 Var.DefRanges.back().Ranges; 1405 if (!R.empty() && R.back().second == Begin) 1406 R.back().second = End; 1407 else 1408 R.emplace_back(Begin, End); 1409 1410 // FIXME: Do more range combining. 1411 } 1412 } 1413 1414 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1415 DenseSet<InlinedEntity> Processed; 1416 // Grab the variable info that was squirreled away in the MMI side-table. 1417 collectVariableInfoFromMFTable(Processed); 1418 1419 for (const auto &I : DbgValues) { 1420 InlinedEntity IV = I.first; 1421 if (Processed.count(IV)) 1422 continue; 1423 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1424 const DILocation *InlinedAt = IV.second; 1425 1426 // Instruction ranges, specifying where IV is accessible. 1427 const auto &Entries = I.second; 1428 1429 LexicalScope *Scope = nullptr; 1430 if (InlinedAt) 1431 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1432 else 1433 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1434 // If variable scope is not found then skip this variable. 1435 if (!Scope) 1436 continue; 1437 1438 LocalVariable Var; 1439 Var.DIVar = DIVar; 1440 1441 calculateRanges(Var, Entries); 1442 recordLocalVariable(std::move(Var), Scope); 1443 } 1444 } 1445 1446 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1447 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1448 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1449 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1450 const Function &GV = MF->getFunction(); 1451 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1452 assert(Insertion.second && "function already has info"); 1453 CurFn = Insertion.first->second.get(); 1454 CurFn->FuncId = NextFuncId++; 1455 CurFn->Begin = Asm->getFunctionBegin(); 1456 1457 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1458 // callee-saved registers were used. For targets that don't use a PUSH 1459 // instruction (AArch64), this will be zero. 1460 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1461 CurFn->FrameSize = MFI.getStackSize(); 1462 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1463 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1464 1465 // For this function S_FRAMEPROC record, figure out which codeview register 1466 // will be the frame pointer. 1467 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1468 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1469 if (CurFn->FrameSize > 0) { 1470 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1471 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1472 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1473 } else { 1474 // If there is an FP, parameters are always relative to it. 1475 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1476 if (CurFn->HasStackRealignment) { 1477 // If the stack needs realignment, locals are relative to SP or VFRAME. 1478 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1479 } else { 1480 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1481 // other stack adjustments. 1482 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1483 } 1484 } 1485 } 1486 1487 // Compute other frame procedure options. 1488 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1489 if (MFI.hasVarSizedObjects()) 1490 FPO |= FrameProcedureOptions::HasAlloca; 1491 if (MF->exposesReturnsTwice()) 1492 FPO |= FrameProcedureOptions::HasSetJmp; 1493 // FIXME: Set HasLongJmp if we ever track that info. 1494 if (MF->hasInlineAsm()) 1495 FPO |= FrameProcedureOptions::HasInlineAssembly; 1496 if (GV.hasPersonalityFn()) { 1497 if (isAsynchronousEHPersonality( 1498 classifyEHPersonality(GV.getPersonalityFn()))) 1499 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1500 else 1501 FPO |= FrameProcedureOptions::HasExceptionHandling; 1502 } 1503 if (GV.hasFnAttribute(Attribute::InlineHint)) 1504 FPO |= FrameProcedureOptions::MarkedInline; 1505 if (GV.hasFnAttribute(Attribute::Naked)) 1506 FPO |= FrameProcedureOptions::Naked; 1507 if (MFI.hasStackProtectorIndex()) 1508 FPO |= FrameProcedureOptions::SecurityChecks; 1509 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1510 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1511 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1512 !GV.hasOptSize() && !GV.hasOptNone()) 1513 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1514 if (GV.hasProfileData()) { 1515 FPO |= FrameProcedureOptions::ValidProfileCounts; 1516 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1517 } 1518 // FIXME: Set GuardCfg when it is implemented. 1519 CurFn->FrameProcOpts = FPO; 1520 1521 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1522 1523 // Find the end of the function prolog. First known non-DBG_VALUE and 1524 // non-frame setup location marks the beginning of the function body. 1525 // FIXME: is there a simpler a way to do this? Can we just search 1526 // for the first instruction of the function, not the last of the prolog? 1527 DebugLoc PrologEndLoc; 1528 bool EmptyPrologue = true; 1529 for (const auto &MBB : *MF) { 1530 for (const auto &MI : MBB) { 1531 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1532 MI.getDebugLoc()) { 1533 PrologEndLoc = MI.getDebugLoc(); 1534 break; 1535 } else if (!MI.isMetaInstruction()) { 1536 EmptyPrologue = false; 1537 } 1538 } 1539 } 1540 1541 // Record beginning of function if we have a non-empty prologue. 1542 if (PrologEndLoc && !EmptyPrologue) { 1543 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1544 maybeRecordLocation(FnStartDL, MF); 1545 } 1546 1547 // Find heap alloc sites and emit labels around them. 1548 for (const auto &MBB : *MF) { 1549 for (const auto &MI : MBB) { 1550 if (MI.getHeapAllocMarker()) { 1551 requestLabelBeforeInsn(&MI); 1552 requestLabelAfterInsn(&MI); 1553 } 1554 } 1555 } 1556 } 1557 1558 static bool shouldEmitUdt(const DIType *T) { 1559 if (!T) 1560 return false; 1561 1562 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1563 if (T->getTag() == dwarf::DW_TAG_typedef) { 1564 if (DIScope *Scope = T->getScope()) { 1565 switch (Scope->getTag()) { 1566 case dwarf::DW_TAG_structure_type: 1567 case dwarf::DW_TAG_class_type: 1568 case dwarf::DW_TAG_union_type: 1569 return false; 1570 default: 1571 // do nothing. 1572 ; 1573 } 1574 } 1575 } 1576 1577 while (true) { 1578 if (!T || T->isForwardDecl()) 1579 return false; 1580 1581 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1582 if (!DT) 1583 return true; 1584 T = DT->getBaseType(); 1585 } 1586 return true; 1587 } 1588 1589 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1590 // Don't record empty UDTs. 1591 if (Ty->getName().empty()) 1592 return; 1593 if (!shouldEmitUdt(Ty)) 1594 return; 1595 1596 SmallVector<StringRef, 5> ParentScopeNames; 1597 const DISubprogram *ClosestSubprogram = 1598 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1599 1600 std::string FullyQualifiedName = 1601 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1602 1603 if (ClosestSubprogram == nullptr) { 1604 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1605 } else if (ClosestSubprogram == CurrentSubprogram) { 1606 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1607 } 1608 1609 // TODO: What if the ClosestSubprogram is neither null or the current 1610 // subprogram? Currently, the UDT just gets dropped on the floor. 1611 // 1612 // The current behavior is not desirable. To get maximal fidelity, we would 1613 // need to perform all type translation before beginning emission of .debug$S 1614 // and then make LocalUDTs a member of FunctionInfo 1615 } 1616 1617 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1618 // Generic dispatch for lowering an unknown type. 1619 switch (Ty->getTag()) { 1620 case dwarf::DW_TAG_array_type: 1621 return lowerTypeArray(cast<DICompositeType>(Ty)); 1622 case dwarf::DW_TAG_typedef: 1623 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1624 case dwarf::DW_TAG_base_type: 1625 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1626 case dwarf::DW_TAG_pointer_type: 1627 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1628 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1629 LLVM_FALLTHROUGH; 1630 case dwarf::DW_TAG_reference_type: 1631 case dwarf::DW_TAG_rvalue_reference_type: 1632 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1633 case dwarf::DW_TAG_ptr_to_member_type: 1634 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1635 case dwarf::DW_TAG_restrict_type: 1636 case dwarf::DW_TAG_const_type: 1637 case dwarf::DW_TAG_volatile_type: 1638 // TODO: add support for DW_TAG_atomic_type here 1639 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1640 case dwarf::DW_TAG_subroutine_type: 1641 if (ClassTy) { 1642 // The member function type of a member function pointer has no 1643 // ThisAdjustment. 1644 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1645 /*ThisAdjustment=*/0, 1646 /*IsStaticMethod=*/false); 1647 } 1648 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1649 case dwarf::DW_TAG_enumeration_type: 1650 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1651 case dwarf::DW_TAG_class_type: 1652 case dwarf::DW_TAG_structure_type: 1653 return lowerTypeClass(cast<DICompositeType>(Ty)); 1654 case dwarf::DW_TAG_union_type: 1655 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1656 case dwarf::DW_TAG_string_type: 1657 return lowerTypeString(cast<DIStringType>(Ty)); 1658 case dwarf::DW_TAG_unspecified_type: 1659 if (Ty->getName() == "decltype(nullptr)") 1660 return TypeIndex::NullptrT(); 1661 return TypeIndex::None(); 1662 default: 1663 // Use the null type index. 1664 return TypeIndex(); 1665 } 1666 } 1667 1668 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1669 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1670 StringRef TypeName = Ty->getName(); 1671 1672 addToUDTs(Ty); 1673 1674 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1675 TypeName == "HRESULT") 1676 return TypeIndex(SimpleTypeKind::HResult); 1677 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1678 TypeName == "wchar_t") 1679 return TypeIndex(SimpleTypeKind::WideCharacter); 1680 1681 return UnderlyingTypeIndex; 1682 } 1683 1684 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1685 const DIType *ElementType = Ty->getBaseType(); 1686 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1687 // IndexType is size_t, which depends on the bitness of the target. 1688 TypeIndex IndexType = getPointerSizeInBytes() == 8 1689 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1690 : TypeIndex(SimpleTypeKind::UInt32Long); 1691 1692 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1693 1694 // Add subranges to array type. 1695 DINodeArray Elements = Ty->getElements(); 1696 for (int i = Elements.size() - 1; i >= 0; --i) { 1697 const DINode *Element = Elements[i]; 1698 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1699 1700 const DISubrange *Subrange = cast<DISubrange>(Element); 1701 int64_t Count = -1; 1702 1703 // If Subrange has a Count field, use it. 1704 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1705 // where lowerbound is from the LowerBound field of the Subrange, 1706 // or the language default lowerbound if that field is unspecified. 1707 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt *>()) 1708 Count = CI->getSExtValue(); 1709 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt *>()) { 1710 // Fortran uses 1 as the default lowerbound; other languages use 0. 1711 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1712 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1713 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1714 Count = UI->getSExtValue() - Lowerbound + 1; 1715 } 1716 1717 // Forward declarations of arrays without a size and VLAs use a count of -1. 1718 // Emit a count of zero in these cases to match what MSVC does for arrays 1719 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1720 // should do for them even if we could distinguish them. 1721 if (Count == -1) 1722 Count = 0; 1723 1724 // Update the element size and element type index for subsequent subranges. 1725 ElementSize *= Count; 1726 1727 // If this is the outermost array, use the size from the array. It will be 1728 // more accurate if we had a VLA or an incomplete element type size. 1729 uint64_t ArraySize = 1730 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1731 1732 StringRef Name = (i == 0) ? Ty->getName() : ""; 1733 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1734 ElementTypeIndex = TypeTable.writeLeafType(AR); 1735 } 1736 1737 return ElementTypeIndex; 1738 } 1739 1740 // This function lowers a Fortran character type (DIStringType). 1741 // Note that it handles only the character*n variant (using SizeInBits 1742 // field in DIString to describe the type size) at the moment. 1743 // Other variants (leveraging the StringLength and StringLengthExp 1744 // fields in DIStringType) remain TBD. 1745 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1746 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1747 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1748 StringRef Name = Ty->getName(); 1749 // IndexType is size_t, which depends on the bitness of the target. 1750 TypeIndex IndexType = getPointerSizeInBytes() == 8 1751 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1752 : TypeIndex(SimpleTypeKind::UInt32Long); 1753 1754 // Create a type of character array of ArraySize. 1755 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1756 1757 return TypeTable.writeLeafType(AR); 1758 } 1759 1760 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1761 TypeIndex Index; 1762 dwarf::TypeKind Kind; 1763 uint32_t ByteSize; 1764 1765 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1766 ByteSize = Ty->getSizeInBits() / 8; 1767 1768 SimpleTypeKind STK = SimpleTypeKind::None; 1769 switch (Kind) { 1770 case dwarf::DW_ATE_address: 1771 // FIXME: Translate 1772 break; 1773 case dwarf::DW_ATE_boolean: 1774 switch (ByteSize) { 1775 case 1: STK = SimpleTypeKind::Boolean8; break; 1776 case 2: STK = SimpleTypeKind::Boolean16; break; 1777 case 4: STK = SimpleTypeKind::Boolean32; break; 1778 case 8: STK = SimpleTypeKind::Boolean64; break; 1779 case 16: STK = SimpleTypeKind::Boolean128; break; 1780 } 1781 break; 1782 case dwarf::DW_ATE_complex_float: 1783 switch (ByteSize) { 1784 case 2: STK = SimpleTypeKind::Complex16; break; 1785 case 4: STK = SimpleTypeKind::Complex32; break; 1786 case 8: STK = SimpleTypeKind::Complex64; break; 1787 case 10: STK = SimpleTypeKind::Complex80; break; 1788 case 16: STK = SimpleTypeKind::Complex128; break; 1789 } 1790 break; 1791 case dwarf::DW_ATE_float: 1792 switch (ByteSize) { 1793 case 2: STK = SimpleTypeKind::Float16; break; 1794 case 4: STK = SimpleTypeKind::Float32; break; 1795 case 6: STK = SimpleTypeKind::Float48; break; 1796 case 8: STK = SimpleTypeKind::Float64; break; 1797 case 10: STK = SimpleTypeKind::Float80; break; 1798 case 16: STK = SimpleTypeKind::Float128; break; 1799 } 1800 break; 1801 case dwarf::DW_ATE_signed: 1802 switch (ByteSize) { 1803 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1804 case 2: STK = SimpleTypeKind::Int16Short; break; 1805 case 4: STK = SimpleTypeKind::Int32; break; 1806 case 8: STK = SimpleTypeKind::Int64Quad; break; 1807 case 16: STK = SimpleTypeKind::Int128Oct; break; 1808 } 1809 break; 1810 case dwarf::DW_ATE_unsigned: 1811 switch (ByteSize) { 1812 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1813 case 2: STK = SimpleTypeKind::UInt16Short; break; 1814 case 4: STK = SimpleTypeKind::UInt32; break; 1815 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1816 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1817 } 1818 break; 1819 case dwarf::DW_ATE_UTF: 1820 switch (ByteSize) { 1821 case 1: STK = SimpleTypeKind::Character8; break; 1822 case 2: STK = SimpleTypeKind::Character16; break; 1823 case 4: STK = SimpleTypeKind::Character32; break; 1824 } 1825 break; 1826 case dwarf::DW_ATE_signed_char: 1827 if (ByteSize == 1) 1828 STK = SimpleTypeKind::SignedCharacter; 1829 break; 1830 case dwarf::DW_ATE_unsigned_char: 1831 if (ByteSize == 1) 1832 STK = SimpleTypeKind::UnsignedCharacter; 1833 break; 1834 default: 1835 break; 1836 } 1837 1838 // Apply some fixups based on the source-level type name. 1839 // Include some amount of canonicalization from an old naming scheme Clang 1840 // used to use for integer types (in an outdated effort to be compatible with 1841 // GCC's debug info/GDB's behavior, which has since been addressed). 1842 if (STK == SimpleTypeKind::Int32 && 1843 (Ty->getName() == "long int" || Ty->getName() == "long")) 1844 STK = SimpleTypeKind::Int32Long; 1845 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1846 Ty->getName() == "unsigned long")) 1847 STK = SimpleTypeKind::UInt32Long; 1848 if (STK == SimpleTypeKind::UInt16Short && 1849 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1850 STK = SimpleTypeKind::WideCharacter; 1851 if ((STK == SimpleTypeKind::SignedCharacter || 1852 STK == SimpleTypeKind::UnsignedCharacter) && 1853 Ty->getName() == "char") 1854 STK = SimpleTypeKind::NarrowCharacter; 1855 1856 return TypeIndex(STK); 1857 } 1858 1859 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1860 PointerOptions PO) { 1861 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1862 1863 // Pointers to simple types without any options can use SimpleTypeMode, rather 1864 // than having a dedicated pointer type record. 1865 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1866 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1867 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1868 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1869 ? SimpleTypeMode::NearPointer64 1870 : SimpleTypeMode::NearPointer32; 1871 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1872 } 1873 1874 PointerKind PK = 1875 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1876 PointerMode PM = PointerMode::Pointer; 1877 switch (Ty->getTag()) { 1878 default: llvm_unreachable("not a pointer tag type"); 1879 case dwarf::DW_TAG_pointer_type: 1880 PM = PointerMode::Pointer; 1881 break; 1882 case dwarf::DW_TAG_reference_type: 1883 PM = PointerMode::LValueReference; 1884 break; 1885 case dwarf::DW_TAG_rvalue_reference_type: 1886 PM = PointerMode::RValueReference; 1887 break; 1888 } 1889 1890 if (Ty->isObjectPointer()) 1891 PO |= PointerOptions::Const; 1892 1893 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1894 return TypeTable.writeLeafType(PR); 1895 } 1896 1897 static PointerToMemberRepresentation 1898 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1899 // SizeInBytes being zero generally implies that the member pointer type was 1900 // incomplete, which can happen if it is part of a function prototype. In this 1901 // case, use the unknown model instead of the general model. 1902 if (IsPMF) { 1903 switch (Flags & DINode::FlagPtrToMemberRep) { 1904 case 0: 1905 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1906 : PointerToMemberRepresentation::GeneralFunction; 1907 case DINode::FlagSingleInheritance: 1908 return PointerToMemberRepresentation::SingleInheritanceFunction; 1909 case DINode::FlagMultipleInheritance: 1910 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1911 case DINode::FlagVirtualInheritance: 1912 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1913 } 1914 } else { 1915 switch (Flags & DINode::FlagPtrToMemberRep) { 1916 case 0: 1917 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1918 : PointerToMemberRepresentation::GeneralData; 1919 case DINode::FlagSingleInheritance: 1920 return PointerToMemberRepresentation::SingleInheritanceData; 1921 case DINode::FlagMultipleInheritance: 1922 return PointerToMemberRepresentation::MultipleInheritanceData; 1923 case DINode::FlagVirtualInheritance: 1924 return PointerToMemberRepresentation::VirtualInheritanceData; 1925 } 1926 } 1927 llvm_unreachable("invalid ptr to member representation"); 1928 } 1929 1930 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1931 PointerOptions PO) { 1932 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1933 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1934 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1935 TypeIndex PointeeTI = 1936 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1937 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1938 : PointerKind::Near32; 1939 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1940 : PointerMode::PointerToDataMember; 1941 1942 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1943 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1944 MemberPointerInfo MPI( 1945 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1946 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1947 return TypeTable.writeLeafType(PR); 1948 } 1949 1950 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1951 /// have a translation, use the NearC convention. 1952 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1953 switch (DwarfCC) { 1954 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1955 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1956 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1957 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1958 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1959 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1960 } 1961 return CallingConvention::NearC; 1962 } 1963 1964 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1965 ModifierOptions Mods = ModifierOptions::None; 1966 PointerOptions PO = PointerOptions::None; 1967 bool IsModifier = true; 1968 const DIType *BaseTy = Ty; 1969 while (IsModifier && BaseTy) { 1970 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1971 switch (BaseTy->getTag()) { 1972 case dwarf::DW_TAG_const_type: 1973 Mods |= ModifierOptions::Const; 1974 PO |= PointerOptions::Const; 1975 break; 1976 case dwarf::DW_TAG_volatile_type: 1977 Mods |= ModifierOptions::Volatile; 1978 PO |= PointerOptions::Volatile; 1979 break; 1980 case dwarf::DW_TAG_restrict_type: 1981 // Only pointer types be marked with __restrict. There is no known flag 1982 // for __restrict in LF_MODIFIER records. 1983 PO |= PointerOptions::Restrict; 1984 break; 1985 default: 1986 IsModifier = false; 1987 break; 1988 } 1989 if (IsModifier) 1990 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1991 } 1992 1993 // Check if the inner type will use an LF_POINTER record. If so, the 1994 // qualifiers will go in the LF_POINTER record. This comes up for types like 1995 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1996 // char *'. 1997 if (BaseTy) { 1998 switch (BaseTy->getTag()) { 1999 case dwarf::DW_TAG_pointer_type: 2000 case dwarf::DW_TAG_reference_type: 2001 case dwarf::DW_TAG_rvalue_reference_type: 2002 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2003 case dwarf::DW_TAG_ptr_to_member_type: 2004 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2005 default: 2006 break; 2007 } 2008 } 2009 2010 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2011 2012 // Return the base type index if there aren't any modifiers. For example, the 2013 // metadata could contain restrict wrappers around non-pointer types. 2014 if (Mods == ModifierOptions::None) 2015 return ModifiedTI; 2016 2017 ModifierRecord MR(ModifiedTI, Mods); 2018 return TypeTable.writeLeafType(MR); 2019 } 2020 2021 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2022 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2023 for (const DIType *ArgType : Ty->getTypeArray()) 2024 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2025 2026 // MSVC uses type none for variadic argument. 2027 if (ReturnAndArgTypeIndices.size() > 1 && 2028 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2029 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2030 } 2031 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2032 ArrayRef<TypeIndex> ArgTypeIndices = None; 2033 if (!ReturnAndArgTypeIndices.empty()) { 2034 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 2035 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2036 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2037 } 2038 2039 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2040 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2041 2042 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2043 2044 FunctionOptions FO = getFunctionOptions(Ty); 2045 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2046 ArgListIndex); 2047 return TypeTable.writeLeafType(Procedure); 2048 } 2049 2050 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2051 const DIType *ClassTy, 2052 int ThisAdjustment, 2053 bool IsStaticMethod, 2054 FunctionOptions FO) { 2055 // Lower the containing class type. 2056 TypeIndex ClassType = getTypeIndex(ClassTy); 2057 2058 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2059 2060 unsigned Index = 0; 2061 SmallVector<TypeIndex, 8> ArgTypeIndices; 2062 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2063 if (ReturnAndArgs.size() > Index) { 2064 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2065 } 2066 2067 // If the first argument is a pointer type and this isn't a static method, 2068 // treat it as the special 'this' parameter, which is encoded separately from 2069 // the arguments. 2070 TypeIndex ThisTypeIndex; 2071 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2072 if (const DIDerivedType *PtrTy = 2073 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2074 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2075 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2076 Index++; 2077 } 2078 } 2079 } 2080 2081 while (Index < ReturnAndArgs.size()) 2082 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2083 2084 // MSVC uses type none for variadic argument. 2085 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2086 ArgTypeIndices.back() = TypeIndex::None(); 2087 2088 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2089 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2090 2091 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2092 2093 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2094 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2095 return TypeTable.writeLeafType(MFR); 2096 } 2097 2098 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2099 unsigned VSlotCount = 2100 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2101 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2102 2103 VFTableShapeRecord VFTSR(Slots); 2104 return TypeTable.writeLeafType(VFTSR); 2105 } 2106 2107 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2108 switch (Flags & DINode::FlagAccessibility) { 2109 case DINode::FlagPrivate: return MemberAccess::Private; 2110 case DINode::FlagPublic: return MemberAccess::Public; 2111 case DINode::FlagProtected: return MemberAccess::Protected; 2112 case 0: 2113 // If there was no explicit access control, provide the default for the tag. 2114 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2115 : MemberAccess::Public; 2116 } 2117 llvm_unreachable("access flags are exclusive"); 2118 } 2119 2120 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2121 if (SP->isArtificial()) 2122 return MethodOptions::CompilerGenerated; 2123 2124 // FIXME: Handle other MethodOptions. 2125 2126 return MethodOptions::None; 2127 } 2128 2129 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2130 bool Introduced) { 2131 if (SP->getFlags() & DINode::FlagStaticMember) 2132 return MethodKind::Static; 2133 2134 switch (SP->getVirtuality()) { 2135 case dwarf::DW_VIRTUALITY_none: 2136 break; 2137 case dwarf::DW_VIRTUALITY_virtual: 2138 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2139 case dwarf::DW_VIRTUALITY_pure_virtual: 2140 return Introduced ? MethodKind::PureIntroducingVirtual 2141 : MethodKind::PureVirtual; 2142 default: 2143 llvm_unreachable("unhandled virtuality case"); 2144 } 2145 2146 return MethodKind::Vanilla; 2147 } 2148 2149 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2150 switch (Ty->getTag()) { 2151 case dwarf::DW_TAG_class_type: 2152 return TypeRecordKind::Class; 2153 case dwarf::DW_TAG_structure_type: 2154 return TypeRecordKind::Struct; 2155 default: 2156 llvm_unreachable("unexpected tag"); 2157 } 2158 } 2159 2160 /// Return ClassOptions that should be present on both the forward declaration 2161 /// and the defintion of a tag type. 2162 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2163 ClassOptions CO = ClassOptions::None; 2164 2165 // MSVC always sets this flag, even for local types. Clang doesn't always 2166 // appear to give every type a linkage name, which may be problematic for us. 2167 // FIXME: Investigate the consequences of not following them here. 2168 if (!Ty->getIdentifier().empty()) 2169 CO |= ClassOptions::HasUniqueName; 2170 2171 // Put the Nested flag on a type if it appears immediately inside a tag type. 2172 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2173 // here. That flag is only set on definitions, and not forward declarations. 2174 const DIScope *ImmediateScope = Ty->getScope(); 2175 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2176 CO |= ClassOptions::Nested; 2177 2178 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2179 // type only when it has an immediate function scope. Clang never puts enums 2180 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2181 // always in function, class, or file scopes. 2182 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2183 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2184 CO |= ClassOptions::Scoped; 2185 } else { 2186 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2187 Scope = Scope->getScope()) { 2188 if (isa<DISubprogram>(Scope)) { 2189 CO |= ClassOptions::Scoped; 2190 break; 2191 } 2192 } 2193 } 2194 2195 return CO; 2196 } 2197 2198 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2199 switch (Ty->getTag()) { 2200 case dwarf::DW_TAG_class_type: 2201 case dwarf::DW_TAG_structure_type: 2202 case dwarf::DW_TAG_union_type: 2203 case dwarf::DW_TAG_enumeration_type: 2204 break; 2205 default: 2206 return; 2207 } 2208 2209 if (const auto *File = Ty->getFile()) { 2210 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2211 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2212 2213 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2214 TypeTable.writeLeafType(USLR); 2215 } 2216 } 2217 2218 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2219 ClassOptions CO = getCommonClassOptions(Ty); 2220 TypeIndex FTI; 2221 unsigned EnumeratorCount = 0; 2222 2223 if (Ty->isForwardDecl()) { 2224 CO |= ClassOptions::ForwardReference; 2225 } else { 2226 ContinuationRecordBuilder ContinuationBuilder; 2227 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2228 for (const DINode *Element : Ty->getElements()) { 2229 // We assume that the frontend provides all members in source declaration 2230 // order, which is what MSVC does. 2231 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2232 // FIXME: Is it correct to always emit these as unsigned here? 2233 EnumeratorRecord ER(MemberAccess::Public, 2234 APSInt(Enumerator->getValue(), true), 2235 Enumerator->getName()); 2236 ContinuationBuilder.writeMemberType(ER); 2237 EnumeratorCount++; 2238 } 2239 } 2240 FTI = TypeTable.insertRecord(ContinuationBuilder); 2241 } 2242 2243 std::string FullName = getFullyQualifiedName(Ty); 2244 2245 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2246 getTypeIndex(Ty->getBaseType())); 2247 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2248 2249 addUDTSrcLine(Ty, EnumTI); 2250 2251 return EnumTI; 2252 } 2253 2254 //===----------------------------------------------------------------------===// 2255 // ClassInfo 2256 //===----------------------------------------------------------------------===// 2257 2258 struct llvm::ClassInfo { 2259 struct MemberInfo { 2260 const DIDerivedType *MemberTypeNode; 2261 uint64_t BaseOffset; 2262 }; 2263 // [MemberInfo] 2264 using MemberList = std::vector<MemberInfo>; 2265 2266 using MethodsList = TinyPtrVector<const DISubprogram *>; 2267 // MethodName -> MethodsList 2268 using MethodsMap = MapVector<MDString *, MethodsList>; 2269 2270 /// Base classes. 2271 std::vector<const DIDerivedType *> Inheritance; 2272 2273 /// Direct members. 2274 MemberList Members; 2275 // Direct overloaded methods gathered by name. 2276 MethodsMap Methods; 2277 2278 TypeIndex VShapeTI; 2279 2280 std::vector<const DIType *> NestedTypes; 2281 }; 2282 2283 void CodeViewDebug::clear() { 2284 assert(CurFn == nullptr); 2285 FileIdMap.clear(); 2286 FnDebugInfo.clear(); 2287 FileToFilepathMap.clear(); 2288 LocalUDTs.clear(); 2289 GlobalUDTs.clear(); 2290 TypeIndices.clear(); 2291 CompleteTypeIndices.clear(); 2292 ScopeGlobals.clear(); 2293 CVGlobalVariableOffsets.clear(); 2294 } 2295 2296 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2297 const DIDerivedType *DDTy) { 2298 if (!DDTy->getName().empty()) { 2299 Info.Members.push_back({DDTy, 0}); 2300 2301 // Collect static const data members with values. 2302 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2303 DINode::FlagStaticMember) { 2304 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2305 isa<ConstantFP>(DDTy->getConstant()))) 2306 StaticConstMembers.push_back(DDTy); 2307 } 2308 2309 return; 2310 } 2311 2312 // An unnamed member may represent a nested struct or union. Attempt to 2313 // interpret the unnamed member as a DICompositeType possibly wrapped in 2314 // qualifier types. Add all the indirect fields to the current record if that 2315 // succeeds, and drop the member if that fails. 2316 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2317 uint64_t Offset = DDTy->getOffsetInBits(); 2318 const DIType *Ty = DDTy->getBaseType(); 2319 bool FullyResolved = false; 2320 while (!FullyResolved) { 2321 switch (Ty->getTag()) { 2322 case dwarf::DW_TAG_const_type: 2323 case dwarf::DW_TAG_volatile_type: 2324 // FIXME: we should apply the qualifier types to the indirect fields 2325 // rather than dropping them. 2326 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2327 break; 2328 default: 2329 FullyResolved = true; 2330 break; 2331 } 2332 } 2333 2334 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2335 if (!DCTy) 2336 return; 2337 2338 ClassInfo NestedInfo = collectClassInfo(DCTy); 2339 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2340 Info.Members.push_back( 2341 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2342 } 2343 2344 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2345 ClassInfo Info; 2346 // Add elements to structure type. 2347 DINodeArray Elements = Ty->getElements(); 2348 for (auto *Element : Elements) { 2349 // We assume that the frontend provides all members in source declaration 2350 // order, which is what MSVC does. 2351 if (!Element) 2352 continue; 2353 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2354 Info.Methods[SP->getRawName()].push_back(SP); 2355 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2356 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2357 collectMemberInfo(Info, DDTy); 2358 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2359 Info.Inheritance.push_back(DDTy); 2360 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2361 DDTy->getName() == "__vtbl_ptr_type") { 2362 Info.VShapeTI = getTypeIndex(DDTy); 2363 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2364 Info.NestedTypes.push_back(DDTy); 2365 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2366 // Ignore friend members. It appears that MSVC emitted info about 2367 // friends in the past, but modern versions do not. 2368 } 2369 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2370 Info.NestedTypes.push_back(Composite); 2371 } 2372 // Skip other unrecognized kinds of elements. 2373 } 2374 return Info; 2375 } 2376 2377 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2378 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2379 // if a complete type should be emitted instead of a forward reference. 2380 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2381 !Ty->isForwardDecl(); 2382 } 2383 2384 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2385 // Emit the complete type for unnamed structs. C++ classes with methods 2386 // which have a circular reference back to the class type are expected to 2387 // be named by the front-end and should not be "unnamed". C unnamed 2388 // structs should not have circular references. 2389 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2390 // If this unnamed complete type is already in the process of being defined 2391 // then the description of the type is malformed and cannot be emitted 2392 // into CodeView correctly so report a fatal error. 2393 auto I = CompleteTypeIndices.find(Ty); 2394 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2395 report_fatal_error("cannot debug circular reference to unnamed type"); 2396 return getCompleteTypeIndex(Ty); 2397 } 2398 2399 // First, construct the forward decl. Don't look into Ty to compute the 2400 // forward decl options, since it might not be available in all TUs. 2401 TypeRecordKind Kind = getRecordKind(Ty); 2402 ClassOptions CO = 2403 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2404 std::string FullName = getFullyQualifiedName(Ty); 2405 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2406 FullName, Ty->getIdentifier()); 2407 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2408 if (!Ty->isForwardDecl()) 2409 DeferredCompleteTypes.push_back(Ty); 2410 return FwdDeclTI; 2411 } 2412 2413 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2414 // Construct the field list and complete type record. 2415 TypeRecordKind Kind = getRecordKind(Ty); 2416 ClassOptions CO = getCommonClassOptions(Ty); 2417 TypeIndex FieldTI; 2418 TypeIndex VShapeTI; 2419 unsigned FieldCount; 2420 bool ContainsNestedClass; 2421 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2422 lowerRecordFieldList(Ty); 2423 2424 if (ContainsNestedClass) 2425 CO |= ClassOptions::ContainsNestedClass; 2426 2427 // MSVC appears to set this flag by searching any destructor or method with 2428 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2429 // the members, however special member functions are not yet emitted into 2430 // debug information. For now checking a class's non-triviality seems enough. 2431 // FIXME: not true for a nested unnamed struct. 2432 if (isNonTrivial(Ty)) 2433 CO |= ClassOptions::HasConstructorOrDestructor; 2434 2435 std::string FullName = getFullyQualifiedName(Ty); 2436 2437 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2438 2439 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2440 SizeInBytes, FullName, Ty->getIdentifier()); 2441 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2442 2443 addUDTSrcLine(Ty, ClassTI); 2444 2445 addToUDTs(Ty); 2446 2447 return ClassTI; 2448 } 2449 2450 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2451 // Emit the complete type for unnamed unions. 2452 if (shouldAlwaysEmitCompleteClassType(Ty)) 2453 return getCompleteTypeIndex(Ty); 2454 2455 ClassOptions CO = 2456 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2457 std::string FullName = getFullyQualifiedName(Ty); 2458 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2459 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2460 if (!Ty->isForwardDecl()) 2461 DeferredCompleteTypes.push_back(Ty); 2462 return FwdDeclTI; 2463 } 2464 2465 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2466 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2467 TypeIndex FieldTI; 2468 unsigned FieldCount; 2469 bool ContainsNestedClass; 2470 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2471 lowerRecordFieldList(Ty); 2472 2473 if (ContainsNestedClass) 2474 CO |= ClassOptions::ContainsNestedClass; 2475 2476 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2477 std::string FullName = getFullyQualifiedName(Ty); 2478 2479 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2480 Ty->getIdentifier()); 2481 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2482 2483 addUDTSrcLine(Ty, UnionTI); 2484 2485 addToUDTs(Ty); 2486 2487 return UnionTI; 2488 } 2489 2490 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2491 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2492 // Manually count members. MSVC appears to count everything that generates a 2493 // field list record. Each individual overload in a method overload group 2494 // contributes to this count, even though the overload group is a single field 2495 // list record. 2496 unsigned MemberCount = 0; 2497 ClassInfo Info = collectClassInfo(Ty); 2498 ContinuationRecordBuilder ContinuationBuilder; 2499 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2500 2501 // Create base classes. 2502 for (const DIDerivedType *I : Info.Inheritance) { 2503 if (I->getFlags() & DINode::FlagVirtual) { 2504 // Virtual base. 2505 unsigned VBPtrOffset = I->getVBPtrOffset(); 2506 // FIXME: Despite the accessor name, the offset is really in bytes. 2507 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2508 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2509 ? TypeRecordKind::IndirectVirtualBaseClass 2510 : TypeRecordKind::VirtualBaseClass; 2511 VirtualBaseClassRecord VBCR( 2512 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2513 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2514 VBTableIndex); 2515 2516 ContinuationBuilder.writeMemberType(VBCR); 2517 MemberCount++; 2518 } else { 2519 assert(I->getOffsetInBits() % 8 == 0 && 2520 "bases must be on byte boundaries"); 2521 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2522 getTypeIndex(I->getBaseType()), 2523 I->getOffsetInBits() / 8); 2524 ContinuationBuilder.writeMemberType(BCR); 2525 MemberCount++; 2526 } 2527 } 2528 2529 // Create members. 2530 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2531 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2532 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2533 StringRef MemberName = Member->getName(); 2534 MemberAccess Access = 2535 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2536 2537 if (Member->isStaticMember()) { 2538 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2539 ContinuationBuilder.writeMemberType(SDMR); 2540 MemberCount++; 2541 continue; 2542 } 2543 2544 // Virtual function pointer member. 2545 if ((Member->getFlags() & DINode::FlagArtificial) && 2546 Member->getName().startswith("_vptr$")) { 2547 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2548 ContinuationBuilder.writeMemberType(VFPR); 2549 MemberCount++; 2550 continue; 2551 } 2552 2553 // Data member. 2554 uint64_t MemberOffsetInBits = 2555 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2556 if (Member->isBitField()) { 2557 uint64_t StartBitOffset = MemberOffsetInBits; 2558 if (const auto *CI = 2559 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2560 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2561 } 2562 StartBitOffset -= MemberOffsetInBits; 2563 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2564 StartBitOffset); 2565 MemberBaseType = TypeTable.writeLeafType(BFR); 2566 } 2567 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2568 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2569 MemberName); 2570 ContinuationBuilder.writeMemberType(DMR); 2571 MemberCount++; 2572 } 2573 2574 // Create methods 2575 for (auto &MethodItr : Info.Methods) { 2576 StringRef Name = MethodItr.first->getString(); 2577 2578 std::vector<OneMethodRecord> Methods; 2579 for (const DISubprogram *SP : MethodItr.second) { 2580 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2581 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2582 2583 unsigned VFTableOffset = -1; 2584 if (Introduced) 2585 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2586 2587 Methods.push_back(OneMethodRecord( 2588 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2589 translateMethodKindFlags(SP, Introduced), 2590 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2591 MemberCount++; 2592 } 2593 assert(!Methods.empty() && "Empty methods map entry"); 2594 if (Methods.size() == 1) 2595 ContinuationBuilder.writeMemberType(Methods[0]); 2596 else { 2597 // FIXME: Make this use its own ContinuationBuilder so that 2598 // MethodOverloadList can be split correctly. 2599 MethodOverloadListRecord MOLR(Methods); 2600 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2601 2602 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2603 ContinuationBuilder.writeMemberType(OMR); 2604 } 2605 } 2606 2607 // Create nested classes. 2608 for (const DIType *Nested : Info.NestedTypes) { 2609 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2610 ContinuationBuilder.writeMemberType(R); 2611 MemberCount++; 2612 } 2613 2614 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2615 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2616 !Info.NestedTypes.empty()); 2617 } 2618 2619 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2620 if (!VBPType.getIndex()) { 2621 // Make a 'const int *' type. 2622 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2623 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2624 2625 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2626 : PointerKind::Near32; 2627 PointerMode PM = PointerMode::Pointer; 2628 PointerOptions PO = PointerOptions::None; 2629 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2630 VBPType = TypeTable.writeLeafType(PR); 2631 } 2632 2633 return VBPType; 2634 } 2635 2636 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2637 // The null DIType is the void type. Don't try to hash it. 2638 if (!Ty) 2639 return TypeIndex::Void(); 2640 2641 // Check if we've already translated this type. Don't try to do a 2642 // get-or-create style insertion that caches the hash lookup across the 2643 // lowerType call. It will update the TypeIndices map. 2644 auto I = TypeIndices.find({Ty, ClassTy}); 2645 if (I != TypeIndices.end()) 2646 return I->second; 2647 2648 TypeLoweringScope S(*this); 2649 TypeIndex TI = lowerType(Ty, ClassTy); 2650 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2651 } 2652 2653 codeview::TypeIndex 2654 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2655 const DISubroutineType *SubroutineTy) { 2656 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2657 "this type must be a pointer type"); 2658 2659 PointerOptions Options = PointerOptions::None; 2660 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2661 Options = PointerOptions::LValueRefThisPointer; 2662 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2663 Options = PointerOptions::RValueRefThisPointer; 2664 2665 // Check if we've already translated this type. If there is no ref qualifier 2666 // on the function then we look up this pointer type with no associated class 2667 // so that the TypeIndex for the this pointer can be shared with the type 2668 // index for other pointers to this class type. If there is a ref qualifier 2669 // then we lookup the pointer using the subroutine as the parent type. 2670 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2671 if (I != TypeIndices.end()) 2672 return I->second; 2673 2674 TypeLoweringScope S(*this); 2675 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2676 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2677 } 2678 2679 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2680 PointerRecord PR(getTypeIndex(Ty), 2681 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2682 : PointerKind::Near32, 2683 PointerMode::LValueReference, PointerOptions::None, 2684 Ty->getSizeInBits() / 8); 2685 return TypeTable.writeLeafType(PR); 2686 } 2687 2688 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2689 // The null DIType is the void type. Don't try to hash it. 2690 if (!Ty) 2691 return TypeIndex::Void(); 2692 2693 // Look through typedefs when getting the complete type index. Call 2694 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2695 // emitted only once. 2696 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2697 (void)getTypeIndex(Ty); 2698 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2699 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2700 2701 // If this is a non-record type, the complete type index is the same as the 2702 // normal type index. Just call getTypeIndex. 2703 switch (Ty->getTag()) { 2704 case dwarf::DW_TAG_class_type: 2705 case dwarf::DW_TAG_structure_type: 2706 case dwarf::DW_TAG_union_type: 2707 break; 2708 default: 2709 return getTypeIndex(Ty); 2710 } 2711 2712 const auto *CTy = cast<DICompositeType>(Ty); 2713 2714 TypeLoweringScope S(*this); 2715 2716 // Make sure the forward declaration is emitted first. It's unclear if this 2717 // is necessary, but MSVC does it, and we should follow suit until we can show 2718 // otherwise. 2719 // We only emit a forward declaration for named types. 2720 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2721 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2722 2723 // Just use the forward decl if we don't have complete type info. This 2724 // might happen if the frontend is using modules and expects the complete 2725 // definition to be emitted elsewhere. 2726 if (CTy->isForwardDecl()) 2727 return FwdDeclTI; 2728 } 2729 2730 // Check if we've already translated the complete record type. 2731 // Insert the type with a null TypeIndex to signify that the type is currently 2732 // being lowered. 2733 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2734 if (!InsertResult.second) 2735 return InsertResult.first->second; 2736 2737 TypeIndex TI; 2738 switch (CTy->getTag()) { 2739 case dwarf::DW_TAG_class_type: 2740 case dwarf::DW_TAG_structure_type: 2741 TI = lowerCompleteTypeClass(CTy); 2742 break; 2743 case dwarf::DW_TAG_union_type: 2744 TI = lowerCompleteTypeUnion(CTy); 2745 break; 2746 default: 2747 llvm_unreachable("not a record"); 2748 } 2749 2750 // Update the type index associated with this CompositeType. This cannot 2751 // use the 'InsertResult' iterator above because it is potentially 2752 // invalidated by map insertions which can occur while lowering the class 2753 // type above. 2754 CompleteTypeIndices[CTy] = TI; 2755 return TI; 2756 } 2757 2758 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2759 /// and do this until fixpoint, as each complete record type typically 2760 /// references 2761 /// many other record types. 2762 void CodeViewDebug::emitDeferredCompleteTypes() { 2763 SmallVector<const DICompositeType *, 4> TypesToEmit; 2764 while (!DeferredCompleteTypes.empty()) { 2765 std::swap(DeferredCompleteTypes, TypesToEmit); 2766 for (const DICompositeType *RecordTy : TypesToEmit) 2767 getCompleteTypeIndex(RecordTy); 2768 TypesToEmit.clear(); 2769 } 2770 } 2771 2772 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2773 ArrayRef<LocalVariable> Locals) { 2774 // Get the sorted list of parameters and emit them first. 2775 SmallVector<const LocalVariable *, 6> Params; 2776 for (const LocalVariable &L : Locals) 2777 if (L.DIVar->isParameter()) 2778 Params.push_back(&L); 2779 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2780 return L->DIVar->getArg() < R->DIVar->getArg(); 2781 }); 2782 for (const LocalVariable *L : Params) 2783 emitLocalVariable(FI, *L); 2784 2785 // Next emit all non-parameters in the order that we found them. 2786 for (const LocalVariable &L : Locals) 2787 if (!L.DIVar->isParameter()) 2788 emitLocalVariable(FI, L); 2789 } 2790 2791 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2792 const LocalVariable &Var) { 2793 // LocalSym record, see SymbolRecord.h for more info. 2794 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2795 2796 LocalSymFlags Flags = LocalSymFlags::None; 2797 if (Var.DIVar->isParameter()) 2798 Flags |= LocalSymFlags::IsParameter; 2799 if (Var.DefRanges.empty()) 2800 Flags |= LocalSymFlags::IsOptimizedOut; 2801 2802 OS.AddComment("TypeIndex"); 2803 TypeIndex TI = Var.UseReferenceType 2804 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2805 : getCompleteTypeIndex(Var.DIVar->getType()); 2806 OS.emitInt32(TI.getIndex()); 2807 OS.AddComment("Flags"); 2808 OS.emitInt16(static_cast<uint16_t>(Flags)); 2809 // Truncate the name so we won't overflow the record length field. 2810 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2811 endSymbolRecord(LocalEnd); 2812 2813 // Calculate the on disk prefix of the appropriate def range record. The 2814 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2815 // should be big enough to hold all forms without memory allocation. 2816 SmallString<20> BytePrefix; 2817 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2818 BytePrefix.clear(); 2819 if (DefRange.InMemory) { 2820 int Offset = DefRange.DataOffset; 2821 unsigned Reg = DefRange.CVRegister; 2822 2823 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2824 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2825 // instead. In frames without stack realignment, $T0 will be the CFA. 2826 if (RegisterId(Reg) == RegisterId::ESP) { 2827 Reg = unsigned(RegisterId::VFRAME); 2828 Offset += FI.OffsetAdjustment; 2829 } 2830 2831 // If we can use the chosen frame pointer for the frame and this isn't a 2832 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2833 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2834 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2835 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2836 (bool(Flags & LocalSymFlags::IsParameter) 2837 ? (EncFP == FI.EncodedParamFramePtrReg) 2838 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2839 DefRangeFramePointerRelHeader DRHdr; 2840 DRHdr.Offset = Offset; 2841 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2842 } else { 2843 uint16_t RegRelFlags = 0; 2844 if (DefRange.IsSubfield) { 2845 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2846 (DefRange.StructOffset 2847 << DefRangeRegisterRelSym::OffsetInParentShift); 2848 } 2849 DefRangeRegisterRelHeader DRHdr; 2850 DRHdr.Register = Reg; 2851 DRHdr.Flags = RegRelFlags; 2852 DRHdr.BasePointerOffset = Offset; 2853 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2854 } 2855 } else { 2856 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2857 if (DefRange.IsSubfield) { 2858 DefRangeSubfieldRegisterHeader DRHdr; 2859 DRHdr.Register = DefRange.CVRegister; 2860 DRHdr.MayHaveNoName = 0; 2861 DRHdr.OffsetInParent = DefRange.StructOffset; 2862 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2863 } else { 2864 DefRangeRegisterHeader DRHdr; 2865 DRHdr.Register = DefRange.CVRegister; 2866 DRHdr.MayHaveNoName = 0; 2867 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2868 } 2869 } 2870 } 2871 } 2872 2873 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2874 const FunctionInfo& FI) { 2875 for (LexicalBlock *Block : Blocks) 2876 emitLexicalBlock(*Block, FI); 2877 } 2878 2879 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2880 /// lexical block scope. 2881 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2882 const FunctionInfo& FI) { 2883 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2884 OS.AddComment("PtrParent"); 2885 OS.emitInt32(0); // PtrParent 2886 OS.AddComment("PtrEnd"); 2887 OS.emitInt32(0); // PtrEnd 2888 OS.AddComment("Code size"); 2889 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2890 OS.AddComment("Function section relative address"); 2891 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2892 OS.AddComment("Function section index"); 2893 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2894 OS.AddComment("Lexical block name"); 2895 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2896 endSymbolRecord(RecordEnd); 2897 2898 // Emit variables local to this lexical block. 2899 emitLocalVariableList(FI, Block.Locals); 2900 emitGlobalVariableList(Block.Globals); 2901 2902 // Emit lexical blocks contained within this block. 2903 emitLexicalBlockList(Block.Children, FI); 2904 2905 // Close the lexical block scope. 2906 emitEndSymbolRecord(SymbolKind::S_END); 2907 } 2908 2909 /// Convenience routine for collecting lexical block information for a list 2910 /// of lexical scopes. 2911 void CodeViewDebug::collectLexicalBlockInfo( 2912 SmallVectorImpl<LexicalScope *> &Scopes, 2913 SmallVectorImpl<LexicalBlock *> &Blocks, 2914 SmallVectorImpl<LocalVariable> &Locals, 2915 SmallVectorImpl<CVGlobalVariable> &Globals) { 2916 for (LexicalScope *Scope : Scopes) 2917 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2918 } 2919 2920 /// Populate the lexical blocks and local variable lists of the parent with 2921 /// information about the specified lexical scope. 2922 void CodeViewDebug::collectLexicalBlockInfo( 2923 LexicalScope &Scope, 2924 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2925 SmallVectorImpl<LocalVariable> &ParentLocals, 2926 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2927 if (Scope.isAbstractScope()) 2928 return; 2929 2930 // Gather information about the lexical scope including local variables, 2931 // global variables, and address ranges. 2932 bool IgnoreScope = false; 2933 auto LI = ScopeVariables.find(&Scope); 2934 SmallVectorImpl<LocalVariable> *Locals = 2935 LI != ScopeVariables.end() ? &LI->second : nullptr; 2936 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2937 SmallVectorImpl<CVGlobalVariable> *Globals = 2938 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2939 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2940 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2941 2942 // Ignore lexical scopes which do not contain variables. 2943 if (!Locals && !Globals) 2944 IgnoreScope = true; 2945 2946 // Ignore lexical scopes which are not lexical blocks. 2947 if (!DILB) 2948 IgnoreScope = true; 2949 2950 // Ignore scopes which have too many address ranges to represent in the 2951 // current CodeView format or do not have a valid address range. 2952 // 2953 // For lexical scopes with multiple address ranges you may be tempted to 2954 // construct a single range covering every instruction where the block is 2955 // live and everything in between. Unfortunately, Visual Studio only 2956 // displays variables from the first matching lexical block scope. If the 2957 // first lexical block contains exception handling code or cold code which 2958 // is moved to the bottom of the routine creating a single range covering 2959 // nearly the entire routine, then it will hide all other lexical blocks 2960 // and the variables they contain. 2961 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2962 IgnoreScope = true; 2963 2964 if (IgnoreScope) { 2965 // This scope can be safely ignored and eliminating it will reduce the 2966 // size of the debug information. Be sure to collect any variable and scope 2967 // information from the this scope or any of its children and collapse them 2968 // into the parent scope. 2969 if (Locals) 2970 ParentLocals.append(Locals->begin(), Locals->end()); 2971 if (Globals) 2972 ParentGlobals.append(Globals->begin(), Globals->end()); 2973 collectLexicalBlockInfo(Scope.getChildren(), 2974 ParentBlocks, 2975 ParentLocals, 2976 ParentGlobals); 2977 return; 2978 } 2979 2980 // Create a new CodeView lexical block for this lexical scope. If we've 2981 // seen this DILexicalBlock before then the scope tree is malformed and 2982 // we can handle this gracefully by not processing it a second time. 2983 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2984 if (!BlockInsertion.second) 2985 return; 2986 2987 // Create a lexical block containing the variables and collect the the 2988 // lexical block information for the children. 2989 const InsnRange &Range = Ranges.front(); 2990 assert(Range.first && Range.second); 2991 LexicalBlock &Block = BlockInsertion.first->second; 2992 Block.Begin = getLabelBeforeInsn(Range.first); 2993 Block.End = getLabelAfterInsn(Range.second); 2994 assert(Block.Begin && "missing label for scope begin"); 2995 assert(Block.End && "missing label for scope end"); 2996 Block.Name = DILB->getName(); 2997 if (Locals) 2998 Block.Locals = std::move(*Locals); 2999 if (Globals) 3000 Block.Globals = std::move(*Globals); 3001 ParentBlocks.push_back(&Block); 3002 collectLexicalBlockInfo(Scope.getChildren(), 3003 Block.Children, 3004 Block.Locals, 3005 Block.Globals); 3006 } 3007 3008 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3009 const Function &GV = MF->getFunction(); 3010 assert(FnDebugInfo.count(&GV)); 3011 assert(CurFn == FnDebugInfo[&GV].get()); 3012 3013 collectVariableInfo(GV.getSubprogram()); 3014 3015 // Build the lexical block structure to emit for this routine. 3016 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3017 collectLexicalBlockInfo(*CFS, 3018 CurFn->ChildBlocks, 3019 CurFn->Locals, 3020 CurFn->Globals); 3021 3022 // Clear the scope and variable information from the map which will not be 3023 // valid after we have finished processing this routine. This also prepares 3024 // the map for the subsequent routine. 3025 ScopeVariables.clear(); 3026 3027 // Don't emit anything if we don't have any line tables. 3028 // Thunks are compiler-generated and probably won't have source correlation. 3029 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3030 FnDebugInfo.erase(&GV); 3031 CurFn = nullptr; 3032 return; 3033 } 3034 3035 // Find heap alloc sites and add to list. 3036 for (const auto &MBB : *MF) { 3037 for (const auto &MI : MBB) { 3038 if (MDNode *MD = MI.getHeapAllocMarker()) { 3039 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3040 getLabelAfterInsn(&MI), 3041 dyn_cast<DIType>(MD))); 3042 } 3043 } 3044 } 3045 3046 CurFn->Annotations = MF->getCodeViewAnnotations(); 3047 3048 CurFn->End = Asm->getFunctionEnd(); 3049 3050 CurFn = nullptr; 3051 } 3052 3053 // Usable locations are valid with non-zero line numbers. A line number of zero 3054 // corresponds to optimized code that doesn't have a distinct source location. 3055 // In this case, we try to use the previous or next source location depending on 3056 // the context. 3057 static bool isUsableDebugLoc(DebugLoc DL) { 3058 return DL && DL.getLine() != 0; 3059 } 3060 3061 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3062 DebugHandlerBase::beginInstruction(MI); 3063 3064 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3065 if (!Asm || !CurFn || MI->isDebugInstr() || 3066 MI->getFlag(MachineInstr::FrameSetup)) 3067 return; 3068 3069 // If the first instruction of a new MBB has no location, find the first 3070 // instruction with a location and use that. 3071 DebugLoc DL = MI->getDebugLoc(); 3072 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3073 for (const auto &NextMI : *MI->getParent()) { 3074 if (NextMI.isDebugInstr()) 3075 continue; 3076 DL = NextMI.getDebugLoc(); 3077 if (isUsableDebugLoc(DL)) 3078 break; 3079 } 3080 // FIXME: Handle the case where the BB has no valid locations. This would 3081 // probably require doing a real dataflow analysis. 3082 } 3083 PrevInstBB = MI->getParent(); 3084 3085 // If we still don't have a debug location, don't record a location. 3086 if (!isUsableDebugLoc(DL)) 3087 return; 3088 3089 maybeRecordLocation(DL, Asm->MF); 3090 } 3091 3092 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3093 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3094 *EndLabel = MMI->getContext().createTempSymbol(); 3095 OS.emitInt32(unsigned(Kind)); 3096 OS.AddComment("Subsection size"); 3097 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3098 OS.emitLabel(BeginLabel); 3099 return EndLabel; 3100 } 3101 3102 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3103 OS.emitLabel(EndLabel); 3104 // Every subsection must be aligned to a 4-byte boundary. 3105 OS.emitValueToAlignment(4); 3106 } 3107 3108 static StringRef getSymbolName(SymbolKind SymKind) { 3109 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3110 if (EE.Value == SymKind) 3111 return EE.Name; 3112 return ""; 3113 } 3114 3115 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3116 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3117 *EndLabel = MMI->getContext().createTempSymbol(); 3118 OS.AddComment("Record length"); 3119 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3120 OS.emitLabel(BeginLabel); 3121 if (OS.isVerboseAsm()) 3122 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3123 OS.emitInt16(unsigned(SymKind)); 3124 return EndLabel; 3125 } 3126 3127 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3128 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3129 // an extra copy of every symbol record in LLD. This increases object file 3130 // size by less than 1% in the clang build, and is compatible with the Visual 3131 // C++ linker. 3132 OS.emitValueToAlignment(4); 3133 OS.emitLabel(SymEnd); 3134 } 3135 3136 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3137 OS.AddComment("Record length"); 3138 OS.emitInt16(2); 3139 if (OS.isVerboseAsm()) 3140 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3141 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3142 } 3143 3144 void CodeViewDebug::emitDebugInfoForUDTs( 3145 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3146 #ifndef NDEBUG 3147 size_t OriginalSize = UDTs.size(); 3148 #endif 3149 for (const auto &UDT : UDTs) { 3150 const DIType *T = UDT.second; 3151 assert(shouldEmitUdt(T)); 3152 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3153 OS.AddComment("Type"); 3154 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3155 assert(OriginalSize == UDTs.size() && 3156 "getCompleteTypeIndex found new UDTs!"); 3157 emitNullTerminatedSymbolName(OS, UDT.first); 3158 endSymbolRecord(UDTRecordEnd); 3159 } 3160 } 3161 3162 void CodeViewDebug::collectGlobalVariableInfo() { 3163 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3164 GlobalMap; 3165 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3166 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3167 GV.getDebugInfo(GVEs); 3168 for (const auto *GVE : GVEs) 3169 GlobalMap[GVE] = &GV; 3170 } 3171 3172 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3173 for (const MDNode *Node : CUs->operands()) { 3174 const auto *CU = cast<DICompileUnit>(Node); 3175 for (const auto *GVE : CU->getGlobalVariables()) { 3176 const DIGlobalVariable *DIGV = GVE->getVariable(); 3177 const DIExpression *DIE = GVE->getExpression(); 3178 // Don't emit string literals in CodeView, as the only useful parts are 3179 // generally the filename and line number, which isn't possible to output 3180 // in CodeView. String literals should be the only unnamed GlobalVariable 3181 // with debug info. 3182 if (DIGV->getName().empty()) continue; 3183 3184 if ((DIE->getNumElements() == 2) && 3185 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3186 // Record the constant offset for the variable. 3187 // 3188 // A Fortran common block uses this idiom to encode the offset 3189 // of a variable from the common block's starting address. 3190 CVGlobalVariableOffsets.insert( 3191 std::make_pair(DIGV, DIE->getElement(1))); 3192 3193 // Emit constant global variables in a global symbol section. 3194 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3195 CVGlobalVariable CVGV = {DIGV, DIE}; 3196 GlobalVariables.emplace_back(std::move(CVGV)); 3197 } 3198 3199 const auto *GV = GlobalMap.lookup(GVE); 3200 if (!GV || GV->isDeclarationForLinker()) 3201 continue; 3202 3203 DIScope *Scope = DIGV->getScope(); 3204 SmallVector<CVGlobalVariable, 1> *VariableList; 3205 if (Scope && isa<DILocalScope>(Scope)) { 3206 // Locate a global variable list for this scope, creating one if 3207 // necessary. 3208 auto Insertion = ScopeGlobals.insert( 3209 {Scope, std::unique_ptr<GlobalVariableList>()}); 3210 if (Insertion.second) 3211 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3212 VariableList = Insertion.first->second.get(); 3213 } else if (GV->hasComdat()) 3214 // Emit this global variable into a COMDAT section. 3215 VariableList = &ComdatVariables; 3216 else 3217 // Emit this global variable in a single global symbol section. 3218 VariableList = &GlobalVariables; 3219 CVGlobalVariable CVGV = {DIGV, GV}; 3220 VariableList->emplace_back(std::move(CVGV)); 3221 } 3222 } 3223 } 3224 3225 void CodeViewDebug::collectDebugInfoForGlobals() { 3226 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3227 const DIGlobalVariable *DIGV = CVGV.DIGV; 3228 const DIScope *Scope = DIGV->getScope(); 3229 getCompleteTypeIndex(DIGV->getType()); 3230 getFullyQualifiedName(Scope, DIGV->getName()); 3231 } 3232 3233 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3234 const DIGlobalVariable *DIGV = CVGV.DIGV; 3235 const DIScope *Scope = DIGV->getScope(); 3236 getCompleteTypeIndex(DIGV->getType()); 3237 getFullyQualifiedName(Scope, DIGV->getName()); 3238 } 3239 } 3240 3241 void CodeViewDebug::emitDebugInfoForGlobals() { 3242 // First, emit all globals that are not in a comdat in a single symbol 3243 // substream. MSVC doesn't like it if the substream is empty, so only open 3244 // it if we have at least one global to emit. 3245 switchToDebugSectionForSymbol(nullptr); 3246 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3247 OS.AddComment("Symbol subsection for globals"); 3248 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3249 emitGlobalVariableList(GlobalVariables); 3250 emitStaticConstMemberList(); 3251 endCVSubsection(EndLabel); 3252 } 3253 3254 // Second, emit each global that is in a comdat into its own .debug$S 3255 // section along with its own symbol substream. 3256 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3257 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3258 MCSymbol *GVSym = Asm->getSymbol(GV); 3259 OS.AddComment("Symbol subsection for " + 3260 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3261 switchToDebugSectionForSymbol(GVSym); 3262 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3263 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3264 emitDebugInfoForGlobal(CVGV); 3265 endCVSubsection(EndLabel); 3266 } 3267 } 3268 3269 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3270 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3271 for (const MDNode *Node : CUs->operands()) { 3272 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3273 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3274 getTypeIndex(RT); 3275 // FIXME: Add to global/local DTU list. 3276 } 3277 } 3278 } 3279 } 3280 3281 // Emit each global variable in the specified array. 3282 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3283 for (const CVGlobalVariable &CVGV : Globals) { 3284 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3285 emitDebugInfoForGlobal(CVGV); 3286 } 3287 } 3288 3289 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3290 const std::string &QualifiedName) { 3291 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3292 OS.AddComment("Type"); 3293 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3294 3295 OS.AddComment("Value"); 3296 3297 // Encoded integers shouldn't need more than 10 bytes. 3298 uint8_t Data[10]; 3299 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3300 CodeViewRecordIO IO(Writer); 3301 cantFail(IO.mapEncodedInteger(Value)); 3302 StringRef SRef((char *)Data, Writer.getOffset()); 3303 OS.emitBinaryData(SRef); 3304 3305 OS.AddComment("Name"); 3306 emitNullTerminatedSymbolName(OS, QualifiedName); 3307 endSymbolRecord(SConstantEnd); 3308 } 3309 3310 void CodeViewDebug::emitStaticConstMemberList() { 3311 for (const DIDerivedType *DTy : StaticConstMembers) { 3312 const DIScope *Scope = DTy->getScope(); 3313 3314 APSInt Value; 3315 if (const ConstantInt *CI = 3316 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3317 Value = APSInt(CI->getValue(), 3318 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3319 else if (const ConstantFP *CFP = 3320 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3321 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3322 else 3323 llvm_unreachable("cannot emit a constant without a value"); 3324 3325 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3326 getFullyQualifiedName(Scope, DTy->getName())); 3327 } 3328 } 3329 3330 static bool isFloatDIType(const DIType *Ty) { 3331 if (isa<DICompositeType>(Ty)) 3332 return false; 3333 3334 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3335 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3336 if (T == dwarf::DW_TAG_pointer_type || 3337 T == dwarf::DW_TAG_ptr_to_member_type || 3338 T == dwarf::DW_TAG_reference_type || 3339 T == dwarf::DW_TAG_rvalue_reference_type) 3340 return false; 3341 assert(DTy->getBaseType() && "Expected valid base type"); 3342 return isFloatDIType(DTy->getBaseType()); 3343 } 3344 3345 auto *BTy = cast<DIBasicType>(Ty); 3346 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3347 } 3348 3349 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3350 const DIGlobalVariable *DIGV = CVGV.DIGV; 3351 3352 const DIScope *Scope = DIGV->getScope(); 3353 // For static data members, get the scope from the declaration. 3354 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3355 DIGV->getRawStaticDataMemberDeclaration())) 3356 Scope = MemberDecl->getScope(); 3357 // For Fortran, the scoping portion is elided in its name so that we can 3358 // reference the variable in the command line of the VS debugger. 3359 std::string QualifiedName = 3360 (moduleIsInFortran()) ? std::string(DIGV->getName()) 3361 : getFullyQualifiedName(Scope, DIGV->getName()); 3362 3363 if (const GlobalVariable *GV = 3364 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3365 // DataSym record, see SymbolRecord.h for more info. Thread local data 3366 // happens to have the same format as global data. 3367 MCSymbol *GVSym = Asm->getSymbol(GV); 3368 SymbolKind DataSym = GV->isThreadLocal() 3369 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3370 : SymbolKind::S_GTHREAD32) 3371 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3372 : SymbolKind::S_GDATA32); 3373 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3374 OS.AddComment("Type"); 3375 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3376 OS.AddComment("DataOffset"); 3377 3378 uint64_t Offset = 0; 3379 if (CVGlobalVariableOffsets.find(DIGV) != CVGlobalVariableOffsets.end()) 3380 // Use the offset seen while collecting info on globals. 3381 Offset = CVGlobalVariableOffsets[DIGV]; 3382 OS.EmitCOFFSecRel32(GVSym, Offset); 3383 3384 OS.AddComment("Segment"); 3385 OS.EmitCOFFSectionIndex(GVSym); 3386 OS.AddComment("Name"); 3387 const unsigned LengthOfDataRecord = 12; 3388 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3389 endSymbolRecord(DataEnd); 3390 } else { 3391 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3392 assert(DIE->isConstant() && 3393 "Global constant variables must contain a constant expression."); 3394 3395 // Use unsigned for floats. 3396 bool isUnsigned = isFloatDIType(DIGV->getType()) 3397 ? true 3398 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3399 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3400 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3401 } 3402 } 3403