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/STLExtras.h" 16 #include "llvm/ADT/SmallBitVector.h" 17 #include "llvm/ADT/SmallString.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/ADT/TinyPtrVector.h" 20 #include "llvm/ADT/Twine.h" 21 #include "llvm/BinaryFormat/COFF.h" 22 #include "llvm/BinaryFormat/Dwarf.h" 23 #include "llvm/CodeGen/AsmPrinter.h" 24 #include "llvm/CodeGen/LexicalScopes.h" 25 #include "llvm/CodeGen/MachineFrameInfo.h" 26 #include "llvm/CodeGen/MachineFunction.h" 27 #include "llvm/CodeGen/MachineInstr.h" 28 #include "llvm/CodeGen/MachineModuleInfo.h" 29 #include "llvm/CodeGen/TargetFrameLowering.h" 30 #include "llvm/CodeGen/TargetLowering.h" 31 #include "llvm/CodeGen/TargetRegisterInfo.h" 32 #include "llvm/CodeGen/TargetSubtargetInfo.h" 33 #include "llvm/Config/llvm-config.h" 34 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 35 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 36 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 37 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 38 #include "llvm/DebugInfo/CodeView/EnumTables.h" 39 #include "llvm/DebugInfo/CodeView/Line.h" 40 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 41 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 42 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 43 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 44 #include "llvm/IR/Constants.h" 45 #include "llvm/IR/DataLayout.h" 46 #include "llvm/IR/DebugInfoMetadata.h" 47 #include "llvm/IR/Function.h" 48 #include "llvm/IR/GlobalValue.h" 49 #include "llvm/IR/GlobalVariable.h" 50 #include "llvm/IR/Metadata.h" 51 #include "llvm/IR/Module.h" 52 #include "llvm/MC/MCAsmInfo.h" 53 #include "llvm/MC/MCContext.h" 54 #include "llvm/MC/MCSectionCOFF.h" 55 #include "llvm/MC/MCStreamer.h" 56 #include "llvm/MC/MCSymbol.h" 57 #include "llvm/Support/BinaryStreamWriter.h" 58 #include "llvm/Support/Casting.h" 59 #include "llvm/Support/Endian.h" 60 #include "llvm/Support/Error.h" 61 #include "llvm/Support/ErrorHandling.h" 62 #include "llvm/Support/FormatVariadic.h" 63 #include "llvm/Support/Path.h" 64 #include "llvm/Support/Program.h" 65 #include "llvm/Support/SMLoc.h" 66 #include "llvm/Support/ScopedPrinter.h" 67 #include "llvm/Target/TargetLoweringObjectFile.h" 68 #include "llvm/Target/TargetMachine.h" 69 #include "llvm/TargetParser/Triple.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(std::move(Var)); 494 } else { 495 // This variable goes into the corresponding lexical scope. 496 ScopeVariables[LS].emplace_back(std::move(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(Align(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 return SourceLanguage::C; 575 case dwarf::DW_LANG_C_plus_plus: 576 case dwarf::DW_LANG_C_plus_plus_03: 577 case dwarf::DW_LANG_C_plus_plus_11: 578 case dwarf::DW_LANG_C_plus_plus_14: 579 return SourceLanguage::Cpp; 580 case dwarf::DW_LANG_Fortran77: 581 case dwarf::DW_LANG_Fortran90: 582 case dwarf::DW_LANG_Fortran95: 583 case dwarf::DW_LANG_Fortran03: 584 case dwarf::DW_LANG_Fortran08: 585 return SourceLanguage::Fortran; 586 case dwarf::DW_LANG_Pascal83: 587 return SourceLanguage::Pascal; 588 case dwarf::DW_LANG_Cobol74: 589 case dwarf::DW_LANG_Cobol85: 590 return SourceLanguage::Cobol; 591 case dwarf::DW_LANG_Java: 592 return SourceLanguage::Java; 593 case dwarf::DW_LANG_D: 594 return SourceLanguage::D; 595 case dwarf::DW_LANG_Swift: 596 return SourceLanguage::Swift; 597 case dwarf::DW_LANG_Rust: 598 return SourceLanguage::Rust; 599 case dwarf::DW_LANG_ObjC: 600 return SourceLanguage::ObjC; 601 case dwarf::DW_LANG_ObjC_plus_plus: 602 return SourceLanguage::ObjCpp; 603 default: 604 // There's no CodeView representation for this language, and CV doesn't 605 // have an "unknown" option for the language field, so we'll use MASM, 606 // as it's very low level. 607 return SourceLanguage::Masm; 608 } 609 } 610 611 void CodeViewDebug::beginModule(Module *M) { 612 // If module doesn't have named metadata anchors or COFF debug section 613 // is not available, skip any debug info related stuff. 614 if (!MMI->hasDebugInfo() || 615 !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 616 Asm = nullptr; 617 return; 618 } 619 620 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 621 622 // Get the current source language. 623 const MDNode *Node = *M->debug_compile_units_begin(); 624 const auto *CU = cast<DICompileUnit>(Node); 625 626 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 627 628 collectGlobalVariableInfo(); 629 630 // Check if we should emit type record hashes. 631 ConstantInt *GH = 632 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 633 EmitDebugGlobalHashes = GH && !GH->isZero(); 634 } 635 636 void CodeViewDebug::endModule() { 637 if (!Asm || !MMI->hasDebugInfo()) 638 return; 639 640 // The COFF .debug$S section consists of several subsections, each starting 641 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 642 // of the payload followed by the payload itself. The subsections are 4-byte 643 // aligned. 644 645 // Use the generic .debug$S section, and make a subsection for all the inlined 646 // subprograms. 647 switchToDebugSectionForSymbol(nullptr); 648 649 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 650 emitObjName(); 651 emitCompilerInformation(); 652 endCVSubsection(CompilerInfo); 653 654 emitInlineeLinesSubsection(); 655 656 // Emit per-function debug information. 657 for (auto &P : FnDebugInfo) 658 if (!P.first->isDeclarationForLinker()) 659 emitDebugInfoForFunction(P.first, *P.second); 660 661 // Get types used by globals without emitting anything. 662 // This is meant to collect all static const data members so they can be 663 // emitted as globals. 664 collectDebugInfoForGlobals(); 665 666 // Emit retained types. 667 emitDebugInfoForRetainedTypes(); 668 669 // Emit global variable debug information. 670 setCurrentSubprogram(nullptr); 671 emitDebugInfoForGlobals(); 672 673 // Switch back to the generic .debug$S section after potentially processing 674 // comdat symbol sections. 675 switchToDebugSectionForSymbol(nullptr); 676 677 // Emit UDT records for any types used by global variables. 678 if (!GlobalUDTs.empty()) { 679 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 680 emitDebugInfoForUDTs(GlobalUDTs); 681 endCVSubsection(SymbolsEnd); 682 } 683 684 // This subsection holds a file index to offset in string table table. 685 OS.AddComment("File index to string table offset subsection"); 686 OS.emitCVFileChecksumsDirective(); 687 688 // This subsection holds the string table. 689 OS.AddComment("String table"); 690 OS.emitCVStringTableDirective(); 691 692 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 693 // subsection in the generic .debug$S section at the end. There is no 694 // particular reason for this ordering other than to match MSVC. 695 emitBuildInfo(); 696 697 // Emit type information and hashes last, so that any types we translate while 698 // emitting function info are included. 699 emitTypeInformation(); 700 701 if (EmitDebugGlobalHashes) 702 emitTypeGlobalHashes(); 703 704 clear(); 705 } 706 707 static void 708 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 709 unsigned MaxFixedRecordLength = 0xF00) { 710 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 711 // after a fixed length portion of the record. The fixed length portion should 712 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 713 // overall record size is less than the maximum allowed. 714 SmallString<32> NullTerminatedString( 715 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 716 NullTerminatedString.push_back('\0'); 717 OS.emitBytes(NullTerminatedString); 718 } 719 720 void CodeViewDebug::emitTypeInformation() { 721 if (TypeTable.empty()) 722 return; 723 724 // Start the .debug$T or .debug$P section with 0x4. 725 OS.switchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 726 emitCodeViewMagicVersion(); 727 728 TypeTableCollection Table(TypeTable.records()); 729 TypeVisitorCallbackPipeline Pipeline; 730 731 // To emit type record using Codeview MCStreamer adapter 732 CVMCAdapter CVMCOS(OS, Table); 733 TypeRecordMapping typeMapping(CVMCOS); 734 Pipeline.addCallbackToPipeline(typeMapping); 735 736 std::optional<TypeIndex> B = Table.getFirst(); 737 while (B) { 738 // This will fail if the record data is invalid. 739 CVType Record = Table.getType(*B); 740 741 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 742 743 if (E) { 744 logAllUnhandledErrors(std::move(E), errs(), "error: "); 745 llvm_unreachable("produced malformed type record"); 746 } 747 748 B = Table.getNext(*B); 749 } 750 } 751 752 void CodeViewDebug::emitTypeGlobalHashes() { 753 if (TypeTable.empty()) 754 return; 755 756 // Start the .debug$H section with the version and hash algorithm, currently 757 // hardcoded to version 0, SHA1. 758 OS.switchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 759 760 OS.emitValueToAlignment(Align(4)); 761 OS.AddComment("Magic"); 762 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 763 OS.AddComment("Section Version"); 764 OS.emitInt16(0); 765 OS.AddComment("Hash Algorithm"); 766 OS.emitInt16(uint16_t(GlobalTypeHashAlg::BLAKE3)); 767 768 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 769 for (const auto &GHR : TypeTable.hashes()) { 770 if (OS.isVerboseAsm()) { 771 // Emit an EOL-comment describing which TypeIndex this hash corresponds 772 // to, as well as the stringified SHA1 hash. 773 SmallString<32> Comment; 774 raw_svector_ostream CommentOS(Comment); 775 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 776 OS.AddComment(Comment); 777 ++TI; 778 } 779 assert(GHR.Hash.size() == 8); 780 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 781 GHR.Hash.size()); 782 OS.emitBinaryData(S); 783 } 784 } 785 786 void CodeViewDebug::emitObjName() { 787 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 788 789 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 790 llvm::SmallString<256> PathStore(PathRef); 791 792 if (PathRef.empty() || PathRef == "-") { 793 // Don't emit the filename if we're writing to stdout or to /dev/null. 794 PathRef = {}; 795 } else { 796 PathRef = PathStore; 797 } 798 799 OS.AddComment("Signature"); 800 OS.emitIntValue(0, 4); 801 802 OS.AddComment("Object name"); 803 emitNullTerminatedSymbolName(OS, PathRef); 804 805 endSymbolRecord(CompilerEnd); 806 } 807 808 namespace { 809 struct Version { 810 int Part[4]; 811 }; 812 } // end anonymous namespace 813 814 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 815 // the version number. 816 static Version parseVersion(StringRef Name) { 817 Version V = {{0}}; 818 int N = 0; 819 for (const char C : Name) { 820 if (isdigit(C)) { 821 V.Part[N] *= 10; 822 V.Part[N] += C - '0'; 823 V.Part[N] = 824 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 825 } else if (C == '.') { 826 ++N; 827 if (N >= 4) 828 return V; 829 } else if (N > 0) 830 return V; 831 } 832 return V; 833 } 834 835 void CodeViewDebug::emitCompilerInformation() { 836 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 837 uint32_t Flags = 0; 838 839 // The low byte of the flags indicates the source language. 840 Flags = CurrentSourceLanguage; 841 // TODO: Figure out which other flags need to be set. 842 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 843 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 844 } 845 using ArchType = llvm::Triple::ArchType; 846 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 847 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 848 Arch == ArchType::aarch64) { 849 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 850 } 851 852 OS.AddComment("Flags and language"); 853 OS.emitInt32(Flags); 854 855 OS.AddComment("CPUType"); 856 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 857 858 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 859 const MDNode *Node = *CUs->operands().begin(); 860 const auto *CU = cast<DICompileUnit>(Node); 861 862 StringRef CompilerVersion = CU->getProducer(); 863 Version FrontVer = parseVersion(CompilerVersion); 864 OS.AddComment("Frontend version"); 865 for (int N : FrontVer.Part) { 866 OS.emitInt16(N); 867 } 868 869 // Some Microsoft tools, like Binscope, expect a backend version number of at 870 // least 8.something, so we'll coerce the LLVM version into a form that 871 // guarantees it'll be big enough without really lying about the version. 872 int Major = 1000 * LLVM_VERSION_MAJOR + 873 10 * LLVM_VERSION_MINOR + 874 LLVM_VERSION_PATCH; 875 // Clamp it for builds that use unusually large version numbers. 876 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 877 Version BackVer = {{ Major, 0, 0, 0 }}; 878 OS.AddComment("Backend version"); 879 for (int N : BackVer.Part) 880 OS.emitInt16(N); 881 882 OS.AddComment("Null-terminated compiler version string"); 883 emitNullTerminatedSymbolName(OS, CompilerVersion); 884 885 endSymbolRecord(CompilerEnd); 886 } 887 888 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 889 StringRef S) { 890 StringIdRecord SIR(TypeIndex(0x0), S); 891 return TypeTable.writeLeafType(SIR); 892 } 893 894 static std::string flattenCommandLine(ArrayRef<std::string> Args, 895 StringRef MainFilename) { 896 std::string FlatCmdLine; 897 raw_string_ostream OS(FlatCmdLine); 898 bool PrintedOneArg = false; 899 if (!StringRef(Args[0]).contains("-cc1")) { 900 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 901 PrintedOneArg = true; 902 } 903 for (unsigned i = 0; i < Args.size(); i++) { 904 StringRef Arg = Args[i]; 905 if (Arg.empty()) 906 continue; 907 if (Arg == "-main-file-name" || Arg == "-o") { 908 i++; // Skip this argument and next one. 909 continue; 910 } 911 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 912 continue; 913 // Skip fmessage-length for reproduciability. 914 if (Arg.startswith("-fmessage-length")) 915 continue; 916 if (PrintedOneArg) 917 OS << " "; 918 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 919 PrintedOneArg = true; 920 } 921 OS.flush(); 922 return FlatCmdLine; 923 } 924 925 void CodeViewDebug::emitBuildInfo() { 926 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 927 // build info. The known prefix is: 928 // - Absolute path of current directory 929 // - Compiler path 930 // - Main source file path, relative to CWD or absolute 931 // - Type server PDB file 932 // - Canonical compiler command line 933 // If frontend and backend compilation are separated (think llc or LTO), it's 934 // not clear if the compiler path should refer to the executable for the 935 // frontend or the backend. Leave it blank for now. 936 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 937 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 938 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 939 const auto *CU = cast<DICompileUnit>(Node); 940 const DIFile *MainSourceFile = CU->getFile(); 941 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 942 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 943 BuildInfoArgs[BuildInfoRecord::SourceFile] = 944 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 945 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 946 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 947 getStringIdTypeIdx(TypeTable, ""); 948 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 949 BuildInfoArgs[BuildInfoRecord::BuildTool] = 950 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 951 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 952 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 953 MainSourceFile->getFilename())); 954 } 955 BuildInfoRecord BIR(BuildInfoArgs); 956 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 957 958 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 959 // from the module symbols into the type stream. 960 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 961 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 962 OS.AddComment("LF_BUILDINFO index"); 963 OS.emitInt32(BuildInfoIndex.getIndex()); 964 endSymbolRecord(BIEnd); 965 endCVSubsection(BISubsecEnd); 966 } 967 968 void CodeViewDebug::emitInlineeLinesSubsection() { 969 if (InlinedSubprograms.empty()) 970 return; 971 972 OS.AddComment("Inlinee lines subsection"); 973 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 974 975 // We emit the checksum info for files. This is used by debuggers to 976 // determine if a pdb matches the source before loading it. Visual Studio, 977 // for instance, will display a warning that the breakpoints are not valid if 978 // the pdb does not match the source. 979 OS.AddComment("Inlinee lines signature"); 980 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 981 982 for (const DISubprogram *SP : InlinedSubprograms) { 983 assert(TypeIndices.count({SP, nullptr})); 984 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 985 986 OS.addBlankLine(); 987 unsigned FileId = maybeRecordFile(SP->getFile()); 988 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 989 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 990 OS.addBlankLine(); 991 OS.AddComment("Type index of inlined function"); 992 OS.emitInt32(InlineeIdx.getIndex()); 993 OS.AddComment("Offset into filechecksum table"); 994 OS.emitCVFileChecksumOffsetDirective(FileId); 995 OS.AddComment("Starting line number"); 996 OS.emitInt32(SP->getLine()); 997 } 998 999 endCVSubsection(InlineEnd); 1000 } 1001 1002 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 1003 const DILocation *InlinedAt, 1004 const InlineSite &Site) { 1005 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1006 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1007 1008 // SymbolRecord 1009 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1010 1011 OS.AddComment("PtrParent"); 1012 OS.emitInt32(0); 1013 OS.AddComment("PtrEnd"); 1014 OS.emitInt32(0); 1015 OS.AddComment("Inlinee type index"); 1016 OS.emitInt32(InlineeIdx.getIndex()); 1017 1018 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1019 unsigned StartLineNum = Site.Inlinee->getLine(); 1020 1021 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1022 FI.Begin, FI.End); 1023 1024 endSymbolRecord(InlineEnd); 1025 1026 emitLocalVariableList(FI, Site.InlinedLocals); 1027 1028 // Recurse on child inlined call sites before closing the scope. 1029 for (const DILocation *ChildSite : Site.ChildSites) { 1030 auto I = FI.InlineSites.find(ChildSite); 1031 assert(I != FI.InlineSites.end() && 1032 "child site not in function inline site map"); 1033 emitInlinedCallSite(FI, ChildSite, I->second); 1034 } 1035 1036 // Close the scope. 1037 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1038 } 1039 1040 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1041 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1042 // comdat key. A section may be comdat because of -ffunction-sections or 1043 // because it is comdat in the IR. 1044 MCSectionCOFF *GVSec = 1045 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1046 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1047 1048 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1049 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1050 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1051 1052 OS.switchSection(DebugSec); 1053 1054 // Emit the magic version number if this is the first time we've switched to 1055 // this section. 1056 if (ComdatDebugSections.insert(DebugSec).second) 1057 emitCodeViewMagicVersion(); 1058 } 1059 1060 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1061 // The only supported thunk ordinal is currently the standard type. 1062 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1063 FunctionInfo &FI, 1064 const MCSymbol *Fn) { 1065 std::string FuncName = 1066 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1067 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1068 1069 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1070 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1071 1072 // Emit S_THUNK32 1073 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1074 OS.AddComment("PtrParent"); 1075 OS.emitInt32(0); 1076 OS.AddComment("PtrEnd"); 1077 OS.emitInt32(0); 1078 OS.AddComment("PtrNext"); 1079 OS.emitInt32(0); 1080 OS.AddComment("Thunk section relative address"); 1081 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1082 OS.AddComment("Thunk section index"); 1083 OS.emitCOFFSectionIndex(Fn); 1084 OS.AddComment("Code size"); 1085 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1086 OS.AddComment("Ordinal"); 1087 OS.emitInt8(unsigned(ordinal)); 1088 OS.AddComment("Function name"); 1089 emitNullTerminatedSymbolName(OS, FuncName); 1090 // Additional fields specific to the thunk ordinal would go here. 1091 endSymbolRecord(ThunkRecordEnd); 1092 1093 // Local variables/inlined routines are purposely omitted here. The point of 1094 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1095 1096 // Emit S_PROC_ID_END 1097 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1098 1099 endCVSubsection(SymbolsEnd); 1100 } 1101 1102 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1103 FunctionInfo &FI) { 1104 // For each function there is a separate subsection which holds the PC to 1105 // file:line table. 1106 const MCSymbol *Fn = Asm->getSymbol(GV); 1107 assert(Fn); 1108 1109 // Switch to the to a comdat section, if appropriate. 1110 switchToDebugSectionForSymbol(Fn); 1111 1112 std::string FuncName; 1113 auto *SP = GV->getSubprogram(); 1114 assert(SP); 1115 setCurrentSubprogram(SP); 1116 1117 if (SP->isThunk()) { 1118 emitDebugInfoForThunk(GV, FI, Fn); 1119 return; 1120 } 1121 1122 // If we have a display name, build the fully qualified name by walking the 1123 // chain of scopes. 1124 if (!SP->getName().empty()) 1125 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1126 1127 // If our DISubprogram name is empty, use the mangled name. 1128 if (FuncName.empty()) 1129 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1130 1131 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1132 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1133 OS.emitCVFPOData(Fn); 1134 1135 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1136 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1137 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1138 { 1139 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1140 : SymbolKind::S_GPROC32_ID; 1141 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1142 1143 // These fields are filled in by tools like CVPACK which run after the fact. 1144 OS.AddComment("PtrParent"); 1145 OS.emitInt32(0); 1146 OS.AddComment("PtrEnd"); 1147 OS.emitInt32(0); 1148 OS.AddComment("PtrNext"); 1149 OS.emitInt32(0); 1150 // This is the important bit that tells the debugger where the function 1151 // code is located and what's its size: 1152 OS.AddComment("Code size"); 1153 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1154 OS.AddComment("Offset after prologue"); 1155 OS.emitInt32(0); 1156 OS.AddComment("Offset before epilogue"); 1157 OS.emitInt32(0); 1158 OS.AddComment("Function type index"); 1159 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1160 OS.AddComment("Function section relative address"); 1161 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1162 OS.AddComment("Function section index"); 1163 OS.emitCOFFSectionIndex(Fn); 1164 OS.AddComment("Flags"); 1165 ProcSymFlags ProcFlags = ProcSymFlags::HasOptimizedDebugInfo; 1166 if (FI.HasFramePointer) 1167 ProcFlags |= ProcSymFlags::HasFP; 1168 if (GV->hasFnAttribute(Attribute::NoReturn)) 1169 ProcFlags |= ProcSymFlags::IsNoReturn; 1170 if (GV->hasFnAttribute(Attribute::NoInline)) 1171 ProcFlags |= ProcSymFlags::IsNoInline; 1172 OS.emitInt8(static_cast<uint8_t>(ProcFlags)); 1173 // Emit the function display name as a null-terminated string. 1174 OS.AddComment("Function name"); 1175 // Truncate the name so we won't overflow the record length field. 1176 emitNullTerminatedSymbolName(OS, FuncName); 1177 endSymbolRecord(ProcRecordEnd); 1178 1179 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1180 // Subtract out the CSR size since MSVC excludes that and we include it. 1181 OS.AddComment("FrameSize"); 1182 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1183 OS.AddComment("Padding"); 1184 OS.emitInt32(0); 1185 OS.AddComment("Offset of padding"); 1186 OS.emitInt32(0); 1187 OS.AddComment("Bytes of callee saved registers"); 1188 OS.emitInt32(FI.CSRSize); 1189 OS.AddComment("Exception handler offset"); 1190 OS.emitInt32(0); 1191 OS.AddComment("Exception handler section"); 1192 OS.emitInt16(0); 1193 OS.AddComment("Flags (defines frame register)"); 1194 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1195 endSymbolRecord(FrameProcEnd); 1196 1197 emitLocalVariableList(FI, FI.Locals); 1198 emitGlobalVariableList(FI.Globals); 1199 emitLexicalBlockList(FI.ChildBlocks, FI); 1200 1201 // Emit inlined call site information. Only emit functions inlined directly 1202 // into the parent function. We'll emit the other sites recursively as part 1203 // of their parent inline site. 1204 for (const DILocation *InlinedAt : FI.ChildSites) { 1205 auto I = FI.InlineSites.find(InlinedAt); 1206 assert(I != FI.InlineSites.end() && 1207 "child site not in function inline site map"); 1208 emitInlinedCallSite(FI, InlinedAt, I->second); 1209 } 1210 1211 for (auto Annot : FI.Annotations) { 1212 MCSymbol *Label = Annot.first; 1213 MDTuple *Strs = cast<MDTuple>(Annot.second); 1214 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1215 OS.emitCOFFSecRel32(Label, /*Offset=*/0); 1216 // FIXME: Make sure we don't overflow the max record size. 1217 OS.emitCOFFSectionIndex(Label); 1218 OS.emitInt16(Strs->getNumOperands()); 1219 for (Metadata *MD : Strs->operands()) { 1220 // MDStrings are null terminated, so we can do EmitBytes and get the 1221 // nice .asciz directive. 1222 StringRef Str = cast<MDString>(MD)->getString(); 1223 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1224 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1225 } 1226 endSymbolRecord(AnnotEnd); 1227 } 1228 1229 for (auto HeapAllocSite : FI.HeapAllocSites) { 1230 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1231 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1232 const DIType *DITy = std::get<2>(HeapAllocSite); 1233 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1234 OS.AddComment("Call site offset"); 1235 OS.emitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1236 OS.AddComment("Call site section index"); 1237 OS.emitCOFFSectionIndex(BeginLabel); 1238 OS.AddComment("Call instruction length"); 1239 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1240 OS.AddComment("Type index"); 1241 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1242 endSymbolRecord(HeapAllocEnd); 1243 } 1244 1245 if (SP != nullptr) 1246 emitDebugInfoForUDTs(LocalUDTs); 1247 1248 emitDebugInfoForJumpTables(FI); 1249 1250 // We're done with this function. 1251 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1252 } 1253 endCVSubsection(SymbolsEnd); 1254 1255 // We have an assembler directive that takes care of the whole line table. 1256 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1257 } 1258 1259 CodeViewDebug::LocalVarDef 1260 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1261 LocalVarDef DR; 1262 DR.InMemory = -1; 1263 DR.DataOffset = Offset; 1264 assert(DR.DataOffset == Offset && "truncation"); 1265 DR.IsSubfield = 0; 1266 DR.StructOffset = 0; 1267 DR.CVRegister = CVRegister; 1268 return DR; 1269 } 1270 1271 void CodeViewDebug::collectVariableInfoFromMFTable( 1272 DenseSet<InlinedEntity> &Processed) { 1273 const MachineFunction &MF = *Asm->MF; 1274 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1275 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1276 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1277 1278 for (const MachineFunction::VariableDbgInfo &VI : 1279 MF.getInStackSlotVariableDbgInfo()) { 1280 if (!VI.Var) 1281 continue; 1282 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1283 "Expected inlined-at fields to agree"); 1284 1285 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1286 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1287 1288 // If variable scope is not found then skip this variable. 1289 if (!Scope) 1290 continue; 1291 1292 // If the variable has an attached offset expression, extract it. 1293 // FIXME: Try to handle DW_OP_deref as well. 1294 int64_t ExprOffset = 0; 1295 bool Deref = false; 1296 if (VI.Expr) { 1297 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1298 if (VI.Expr->getNumElements() == 1 && 1299 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1300 Deref = true; 1301 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1302 continue; 1303 } 1304 1305 // Get the frame register used and the offset. 1306 Register FrameReg; 1307 StackOffset FrameOffset = 1308 TFI->getFrameIndexReference(*Asm->MF, VI.getStackSlot(), FrameReg); 1309 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1310 1311 assert(!FrameOffset.getScalable() && 1312 "Frame offsets with a scalable component are not supported"); 1313 1314 // Calculate the label ranges. 1315 LocalVarDef DefRange = 1316 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1317 1318 LocalVariable Var; 1319 Var.DIVar = VI.Var; 1320 1321 for (const InsnRange &Range : Scope->getRanges()) { 1322 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1323 const MCSymbol *End = getLabelAfterInsn(Range.second); 1324 End = End ? End : Asm->getFunctionEnd(); 1325 Var.DefRanges[DefRange].emplace_back(Begin, End); 1326 } 1327 1328 if (Deref) 1329 Var.UseReferenceType = true; 1330 1331 recordLocalVariable(std::move(Var), Scope); 1332 } 1333 } 1334 1335 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1336 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1337 } 1338 1339 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1340 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1341 } 1342 1343 void CodeViewDebug::calculateRanges( 1344 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1345 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1346 1347 // Calculate the definition ranges. 1348 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1349 const auto &Entry = *I; 1350 if (!Entry.isDbgValue()) 1351 continue; 1352 const MachineInstr *DVInst = Entry.getInstr(); 1353 assert(DVInst->isDebugValue() && "Invalid History entry"); 1354 // FIXME: Find a way to represent constant variables, since they are 1355 // relatively common. 1356 std::optional<DbgVariableLocation> Location = 1357 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1358 if (!Location) 1359 { 1360 // When we don't have a location this is usually because LLVM has 1361 // transformed it into a constant and we only have an llvm.dbg.value. We 1362 // can't represent these well in CodeView since S_LOCAL only works on 1363 // registers and memory locations. Instead, we will pretend this to be a 1364 // constant value to at least have it show up in the debugger. 1365 auto Op = DVInst->getDebugOperand(0); 1366 if (Op.isImm()) 1367 Var.ConstantValue = APSInt(APInt(64, Op.getImm()), false); 1368 continue; 1369 } 1370 1371 // CodeView can only express variables in register and variables in memory 1372 // at a constant offset from a register. However, for variables passed 1373 // indirectly by pointer, it is common for that pointer to be spilled to a 1374 // stack location. For the special case of one offseted load followed by a 1375 // zero offset load (a pointer spilled to the stack), we change the type of 1376 // the local variable from a value type to a reference type. This tricks the 1377 // debugger into doing the load for us. 1378 if (Var.UseReferenceType) { 1379 // We're using a reference type. Drop the last zero offset load. 1380 if (canUseReferenceType(*Location)) 1381 Location->LoadChain.pop_back(); 1382 else 1383 continue; 1384 } else if (needsReferenceType(*Location)) { 1385 // This location can't be expressed without switching to a reference type. 1386 // Start over using that. 1387 Var.UseReferenceType = true; 1388 Var.DefRanges.clear(); 1389 calculateRanges(Var, Entries); 1390 return; 1391 } 1392 1393 // We can only handle a register or an offseted load of a register. 1394 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1395 continue; 1396 1397 LocalVarDef DR; 1398 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1399 DR.InMemory = !Location->LoadChain.empty(); 1400 DR.DataOffset = 1401 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1402 if (Location->FragmentInfo) { 1403 DR.IsSubfield = true; 1404 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1405 } else { 1406 DR.IsSubfield = false; 1407 DR.StructOffset = 0; 1408 } 1409 1410 // Compute the label range. 1411 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1412 const MCSymbol *End; 1413 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1414 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1415 End = EndingEntry.isDbgValue() 1416 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1417 : getLabelAfterInsn(EndingEntry.getInstr()); 1418 } else 1419 End = Asm->getFunctionEnd(); 1420 1421 // If the last range end is our begin, just extend the last range. 1422 // Otherwise make a new range. 1423 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1424 Var.DefRanges[DR]; 1425 if (!R.empty() && R.back().second == Begin) 1426 R.back().second = End; 1427 else 1428 R.emplace_back(Begin, End); 1429 1430 // FIXME: Do more range combining. 1431 } 1432 } 1433 1434 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1435 DenseSet<InlinedEntity> Processed; 1436 // Grab the variable info that was squirreled away in the MMI side-table. 1437 collectVariableInfoFromMFTable(Processed); 1438 1439 for (const auto &I : DbgValues) { 1440 InlinedEntity IV = I.first; 1441 if (Processed.count(IV)) 1442 continue; 1443 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1444 const DILocation *InlinedAt = IV.second; 1445 1446 // Instruction ranges, specifying where IV is accessible. 1447 const auto &Entries = I.second; 1448 1449 LexicalScope *Scope = nullptr; 1450 if (InlinedAt) 1451 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1452 else 1453 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1454 // If variable scope is not found then skip this variable. 1455 if (!Scope) 1456 continue; 1457 1458 LocalVariable Var; 1459 Var.DIVar = DIVar; 1460 1461 calculateRanges(Var, Entries); 1462 recordLocalVariable(std::move(Var), Scope); 1463 } 1464 } 1465 1466 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1467 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1468 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1469 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1470 const Function &GV = MF->getFunction(); 1471 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1472 assert(Insertion.second && "function already has info"); 1473 CurFn = Insertion.first->second.get(); 1474 CurFn->FuncId = NextFuncId++; 1475 CurFn->Begin = Asm->getFunctionBegin(); 1476 1477 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1478 // callee-saved registers were used. For targets that don't use a PUSH 1479 // instruction (AArch64), this will be zero. 1480 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1481 CurFn->FrameSize = MFI.getStackSize(); 1482 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1483 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1484 1485 // For this function S_FRAMEPROC record, figure out which codeview register 1486 // will be the frame pointer. 1487 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1488 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1489 if (CurFn->FrameSize > 0) { 1490 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1491 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1492 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1493 } else { 1494 CurFn->HasFramePointer = true; 1495 // If there is an FP, parameters are always relative to it. 1496 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1497 if (CurFn->HasStackRealignment) { 1498 // If the stack needs realignment, locals are relative to SP or VFRAME. 1499 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1500 } else { 1501 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1502 // other stack adjustments. 1503 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1504 } 1505 } 1506 } 1507 1508 // Compute other frame procedure options. 1509 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1510 if (MFI.hasVarSizedObjects()) 1511 FPO |= FrameProcedureOptions::HasAlloca; 1512 if (MF->exposesReturnsTwice()) 1513 FPO |= FrameProcedureOptions::HasSetJmp; 1514 // FIXME: Set HasLongJmp if we ever track that info. 1515 if (MF->hasInlineAsm()) 1516 FPO |= FrameProcedureOptions::HasInlineAssembly; 1517 if (GV.hasPersonalityFn()) { 1518 if (isAsynchronousEHPersonality( 1519 classifyEHPersonality(GV.getPersonalityFn()))) 1520 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1521 else 1522 FPO |= FrameProcedureOptions::HasExceptionHandling; 1523 } 1524 if (GV.hasFnAttribute(Attribute::InlineHint)) 1525 FPO |= FrameProcedureOptions::MarkedInline; 1526 if (GV.hasFnAttribute(Attribute::Naked)) 1527 FPO |= FrameProcedureOptions::Naked; 1528 if (MFI.hasStackProtectorIndex()) { 1529 FPO |= FrameProcedureOptions::SecurityChecks; 1530 if (GV.hasFnAttribute(Attribute::StackProtectStrong) || 1531 GV.hasFnAttribute(Attribute::StackProtectReq)) { 1532 FPO |= FrameProcedureOptions::StrictSecurityChecks; 1533 } 1534 } else if (!GV.hasStackProtectorFnAttr()) { 1535 // __declspec(safebuffers) disables stack guards. 1536 FPO |= FrameProcedureOptions::SafeBuffers; 1537 } 1538 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1539 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1540 if (Asm->TM.getOptLevel() != CodeGenOptLevel::None && !GV.hasOptSize() && 1541 !GV.hasOptNone()) 1542 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1543 if (GV.hasProfileData()) { 1544 FPO |= FrameProcedureOptions::ValidProfileCounts; 1545 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1546 } 1547 // FIXME: Set GuardCfg when it is implemented. 1548 CurFn->FrameProcOpts = FPO; 1549 1550 OS.emitCVFuncIdDirective(CurFn->FuncId); 1551 1552 // Find the end of the function prolog. First known non-DBG_VALUE and 1553 // non-frame setup location marks the beginning of the function body. 1554 // FIXME: is there a simpler a way to do this? Can we just search 1555 // for the first instruction of the function, not the last of the prolog? 1556 DebugLoc PrologEndLoc; 1557 bool EmptyPrologue = true; 1558 for (const auto &MBB : *MF) { 1559 for (const auto &MI : MBB) { 1560 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1561 MI.getDebugLoc()) { 1562 PrologEndLoc = MI.getDebugLoc(); 1563 break; 1564 } else if (!MI.isMetaInstruction()) { 1565 EmptyPrologue = false; 1566 } 1567 } 1568 } 1569 1570 // Record beginning of function if we have a non-empty prologue. 1571 if (PrologEndLoc && !EmptyPrologue) { 1572 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1573 maybeRecordLocation(FnStartDL, MF); 1574 } 1575 1576 // Find heap alloc sites and emit labels around them. 1577 for (const auto &MBB : *MF) { 1578 for (const auto &MI : MBB) { 1579 if (MI.getHeapAllocMarker()) { 1580 requestLabelBeforeInsn(&MI); 1581 requestLabelAfterInsn(&MI); 1582 } 1583 } 1584 } 1585 1586 // Mark branches that may potentially be using jump tables with labels. 1587 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 1588 llvm::Triple::ArchType::thumb; 1589 discoverJumpTableBranches(MF, isThumb); 1590 } 1591 1592 static bool shouldEmitUdt(const DIType *T) { 1593 if (!T) 1594 return false; 1595 1596 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1597 if (T->getTag() == dwarf::DW_TAG_typedef) { 1598 if (DIScope *Scope = T->getScope()) { 1599 switch (Scope->getTag()) { 1600 case dwarf::DW_TAG_structure_type: 1601 case dwarf::DW_TAG_class_type: 1602 case dwarf::DW_TAG_union_type: 1603 return false; 1604 default: 1605 // do nothing. 1606 ; 1607 } 1608 } 1609 } 1610 1611 while (true) { 1612 if (!T || T->isForwardDecl()) 1613 return false; 1614 1615 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1616 if (!DT) 1617 return true; 1618 T = DT->getBaseType(); 1619 } 1620 return true; 1621 } 1622 1623 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1624 // Don't record empty UDTs. 1625 if (Ty->getName().empty()) 1626 return; 1627 if (!shouldEmitUdt(Ty)) 1628 return; 1629 1630 SmallVector<StringRef, 5> ParentScopeNames; 1631 const DISubprogram *ClosestSubprogram = 1632 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1633 1634 std::string FullyQualifiedName = 1635 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1636 1637 if (ClosestSubprogram == nullptr) { 1638 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1639 } else if (ClosestSubprogram == CurrentSubprogram) { 1640 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1641 } 1642 1643 // TODO: What if the ClosestSubprogram is neither null or the current 1644 // subprogram? Currently, the UDT just gets dropped on the floor. 1645 // 1646 // The current behavior is not desirable. To get maximal fidelity, we would 1647 // need to perform all type translation before beginning emission of .debug$S 1648 // and then make LocalUDTs a member of FunctionInfo 1649 } 1650 1651 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1652 // Generic dispatch for lowering an unknown type. 1653 switch (Ty->getTag()) { 1654 case dwarf::DW_TAG_array_type: 1655 return lowerTypeArray(cast<DICompositeType>(Ty)); 1656 case dwarf::DW_TAG_typedef: 1657 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1658 case dwarf::DW_TAG_base_type: 1659 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1660 case dwarf::DW_TAG_pointer_type: 1661 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1662 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1663 [[fallthrough]]; 1664 case dwarf::DW_TAG_reference_type: 1665 case dwarf::DW_TAG_rvalue_reference_type: 1666 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1667 case dwarf::DW_TAG_ptr_to_member_type: 1668 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1669 case dwarf::DW_TAG_restrict_type: 1670 case dwarf::DW_TAG_const_type: 1671 case dwarf::DW_TAG_volatile_type: 1672 // TODO: add support for DW_TAG_atomic_type here 1673 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1674 case dwarf::DW_TAG_subroutine_type: 1675 if (ClassTy) { 1676 // The member function type of a member function pointer has no 1677 // ThisAdjustment. 1678 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1679 /*ThisAdjustment=*/0, 1680 /*IsStaticMethod=*/false); 1681 } 1682 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1683 case dwarf::DW_TAG_enumeration_type: 1684 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1685 case dwarf::DW_TAG_class_type: 1686 case dwarf::DW_TAG_structure_type: 1687 return lowerTypeClass(cast<DICompositeType>(Ty)); 1688 case dwarf::DW_TAG_union_type: 1689 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1690 case dwarf::DW_TAG_string_type: 1691 return lowerTypeString(cast<DIStringType>(Ty)); 1692 case dwarf::DW_TAG_unspecified_type: 1693 if (Ty->getName() == "decltype(nullptr)") 1694 return TypeIndex::NullptrT(); 1695 return TypeIndex::None(); 1696 default: 1697 // Use the null type index. 1698 return TypeIndex(); 1699 } 1700 } 1701 1702 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1703 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1704 StringRef TypeName = Ty->getName(); 1705 1706 addToUDTs(Ty); 1707 1708 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1709 TypeName == "HRESULT") 1710 return TypeIndex(SimpleTypeKind::HResult); 1711 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1712 TypeName == "wchar_t") 1713 return TypeIndex(SimpleTypeKind::WideCharacter); 1714 1715 return UnderlyingTypeIndex; 1716 } 1717 1718 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1719 const DIType *ElementType = Ty->getBaseType(); 1720 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1721 // IndexType is size_t, which depends on the bitness of the target. 1722 TypeIndex IndexType = getPointerSizeInBytes() == 8 1723 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1724 : TypeIndex(SimpleTypeKind::UInt32Long); 1725 1726 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1727 1728 // Add subranges to array type. 1729 DINodeArray Elements = Ty->getElements(); 1730 for (int i = Elements.size() - 1; i >= 0; --i) { 1731 const DINode *Element = Elements[i]; 1732 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1733 1734 const DISubrange *Subrange = cast<DISubrange>(Element); 1735 int64_t Count = -1; 1736 1737 // If Subrange has a Count field, use it. 1738 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1739 // where lowerbound is from the LowerBound field of the Subrange, 1740 // or the language default lowerbound if that field is unspecified. 1741 if (auto *CI = dyn_cast_if_present<ConstantInt *>(Subrange->getCount())) 1742 Count = CI->getSExtValue(); 1743 else if (auto *UI = dyn_cast_if_present<ConstantInt *>( 1744 Subrange->getUpperBound())) { 1745 // Fortran uses 1 as the default lowerbound; other languages use 0. 1746 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1747 auto *LI = dyn_cast_if_present<ConstantInt *>(Subrange->getLowerBound()); 1748 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1749 Count = UI->getSExtValue() - Lowerbound + 1; 1750 } 1751 1752 // Forward declarations of arrays without a size and VLAs use a count of -1. 1753 // Emit a count of zero in these cases to match what MSVC does for arrays 1754 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1755 // should do for them even if we could distinguish them. 1756 if (Count == -1) 1757 Count = 0; 1758 1759 // Update the element size and element type index for subsequent subranges. 1760 ElementSize *= Count; 1761 1762 // If this is the outermost array, use the size from the array. It will be 1763 // more accurate if we had a VLA or an incomplete element type size. 1764 uint64_t ArraySize = 1765 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1766 1767 StringRef Name = (i == 0) ? Ty->getName() : ""; 1768 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1769 ElementTypeIndex = TypeTable.writeLeafType(AR); 1770 } 1771 1772 return ElementTypeIndex; 1773 } 1774 1775 // This function lowers a Fortran character type (DIStringType). 1776 // Note that it handles only the character*n variant (using SizeInBits 1777 // field in DIString to describe the type size) at the moment. 1778 // Other variants (leveraging the StringLength and StringLengthExp 1779 // fields in DIStringType) remain TBD. 1780 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1781 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1782 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1783 StringRef Name = Ty->getName(); 1784 // IndexType is size_t, which depends on the bitness of the target. 1785 TypeIndex IndexType = getPointerSizeInBytes() == 8 1786 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1787 : TypeIndex(SimpleTypeKind::UInt32Long); 1788 1789 // Create a type of character array of ArraySize. 1790 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1791 1792 return TypeTable.writeLeafType(AR); 1793 } 1794 1795 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1796 TypeIndex Index; 1797 dwarf::TypeKind Kind; 1798 uint32_t ByteSize; 1799 1800 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1801 ByteSize = Ty->getSizeInBits() / 8; 1802 1803 SimpleTypeKind STK = SimpleTypeKind::None; 1804 switch (Kind) { 1805 case dwarf::DW_ATE_address: 1806 // FIXME: Translate 1807 break; 1808 case dwarf::DW_ATE_boolean: 1809 switch (ByteSize) { 1810 case 1: STK = SimpleTypeKind::Boolean8; break; 1811 case 2: STK = SimpleTypeKind::Boolean16; break; 1812 case 4: STK = SimpleTypeKind::Boolean32; break; 1813 case 8: STK = SimpleTypeKind::Boolean64; break; 1814 case 16: STK = SimpleTypeKind::Boolean128; break; 1815 } 1816 break; 1817 case dwarf::DW_ATE_complex_float: 1818 // The CodeView size for a complex represents the size of 1819 // an individual component. 1820 switch (ByteSize) { 1821 case 4: STK = SimpleTypeKind::Complex16; break; 1822 case 8: STK = SimpleTypeKind::Complex32; break; 1823 case 16: STK = SimpleTypeKind::Complex64; break; 1824 case 20: STK = SimpleTypeKind::Complex80; break; 1825 case 32: STK = SimpleTypeKind::Complex128; break; 1826 } 1827 break; 1828 case dwarf::DW_ATE_float: 1829 switch (ByteSize) { 1830 case 2: STK = SimpleTypeKind::Float16; break; 1831 case 4: STK = SimpleTypeKind::Float32; break; 1832 case 6: STK = SimpleTypeKind::Float48; break; 1833 case 8: STK = SimpleTypeKind::Float64; break; 1834 case 10: STK = SimpleTypeKind::Float80; break; 1835 case 16: STK = SimpleTypeKind::Float128; break; 1836 } 1837 break; 1838 case dwarf::DW_ATE_signed: 1839 switch (ByteSize) { 1840 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1841 case 2: STK = SimpleTypeKind::Int16Short; break; 1842 case 4: STK = SimpleTypeKind::Int32; break; 1843 case 8: STK = SimpleTypeKind::Int64Quad; break; 1844 case 16: STK = SimpleTypeKind::Int128Oct; break; 1845 } 1846 break; 1847 case dwarf::DW_ATE_unsigned: 1848 switch (ByteSize) { 1849 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1850 case 2: STK = SimpleTypeKind::UInt16Short; break; 1851 case 4: STK = SimpleTypeKind::UInt32; break; 1852 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1853 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1854 } 1855 break; 1856 case dwarf::DW_ATE_UTF: 1857 switch (ByteSize) { 1858 case 1: STK = SimpleTypeKind::Character8; break; 1859 case 2: STK = SimpleTypeKind::Character16; break; 1860 case 4: STK = SimpleTypeKind::Character32; break; 1861 } 1862 break; 1863 case dwarf::DW_ATE_signed_char: 1864 if (ByteSize == 1) 1865 STK = SimpleTypeKind::SignedCharacter; 1866 break; 1867 case dwarf::DW_ATE_unsigned_char: 1868 if (ByteSize == 1) 1869 STK = SimpleTypeKind::UnsignedCharacter; 1870 break; 1871 default: 1872 break; 1873 } 1874 1875 // Apply some fixups based on the source-level type name. 1876 // Include some amount of canonicalization from an old naming scheme Clang 1877 // used to use for integer types (in an outdated effort to be compatible with 1878 // GCC's debug info/GDB's behavior, which has since been addressed). 1879 if (STK == SimpleTypeKind::Int32 && 1880 (Ty->getName() == "long int" || Ty->getName() == "long")) 1881 STK = SimpleTypeKind::Int32Long; 1882 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1883 Ty->getName() == "unsigned long")) 1884 STK = SimpleTypeKind::UInt32Long; 1885 if (STK == SimpleTypeKind::UInt16Short && 1886 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1887 STK = SimpleTypeKind::WideCharacter; 1888 if ((STK == SimpleTypeKind::SignedCharacter || 1889 STK == SimpleTypeKind::UnsignedCharacter) && 1890 Ty->getName() == "char") 1891 STK = SimpleTypeKind::NarrowCharacter; 1892 1893 return TypeIndex(STK); 1894 } 1895 1896 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1897 PointerOptions PO) { 1898 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1899 1900 // Pointers to simple types without any options can use SimpleTypeMode, rather 1901 // than having a dedicated pointer type record. 1902 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1903 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1904 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1905 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1906 ? SimpleTypeMode::NearPointer64 1907 : SimpleTypeMode::NearPointer32; 1908 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1909 } 1910 1911 PointerKind PK = 1912 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1913 PointerMode PM = PointerMode::Pointer; 1914 switch (Ty->getTag()) { 1915 default: llvm_unreachable("not a pointer tag type"); 1916 case dwarf::DW_TAG_pointer_type: 1917 PM = PointerMode::Pointer; 1918 break; 1919 case dwarf::DW_TAG_reference_type: 1920 PM = PointerMode::LValueReference; 1921 break; 1922 case dwarf::DW_TAG_rvalue_reference_type: 1923 PM = PointerMode::RValueReference; 1924 break; 1925 } 1926 1927 if (Ty->isObjectPointer()) 1928 PO |= PointerOptions::Const; 1929 1930 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1931 return TypeTable.writeLeafType(PR); 1932 } 1933 1934 static PointerToMemberRepresentation 1935 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1936 // SizeInBytes being zero generally implies that the member pointer type was 1937 // incomplete, which can happen if it is part of a function prototype. In this 1938 // case, use the unknown model instead of the general model. 1939 if (IsPMF) { 1940 switch (Flags & DINode::FlagPtrToMemberRep) { 1941 case 0: 1942 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1943 : PointerToMemberRepresentation::GeneralFunction; 1944 case DINode::FlagSingleInheritance: 1945 return PointerToMemberRepresentation::SingleInheritanceFunction; 1946 case DINode::FlagMultipleInheritance: 1947 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1948 case DINode::FlagVirtualInheritance: 1949 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1950 } 1951 } else { 1952 switch (Flags & DINode::FlagPtrToMemberRep) { 1953 case 0: 1954 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1955 : PointerToMemberRepresentation::GeneralData; 1956 case DINode::FlagSingleInheritance: 1957 return PointerToMemberRepresentation::SingleInheritanceData; 1958 case DINode::FlagMultipleInheritance: 1959 return PointerToMemberRepresentation::MultipleInheritanceData; 1960 case DINode::FlagVirtualInheritance: 1961 return PointerToMemberRepresentation::VirtualInheritanceData; 1962 } 1963 } 1964 llvm_unreachable("invalid ptr to member representation"); 1965 } 1966 1967 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1968 PointerOptions PO) { 1969 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1970 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1971 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1972 TypeIndex PointeeTI = 1973 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1974 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1975 : PointerKind::Near32; 1976 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1977 : PointerMode::PointerToDataMember; 1978 1979 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1980 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1981 MemberPointerInfo MPI( 1982 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1983 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1984 return TypeTable.writeLeafType(PR); 1985 } 1986 1987 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1988 /// have a translation, use the NearC convention. 1989 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1990 switch (DwarfCC) { 1991 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1992 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1993 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1994 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1995 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1996 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1997 } 1998 return CallingConvention::NearC; 1999 } 2000 2001 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 2002 ModifierOptions Mods = ModifierOptions::None; 2003 PointerOptions PO = PointerOptions::None; 2004 bool IsModifier = true; 2005 const DIType *BaseTy = Ty; 2006 while (IsModifier && BaseTy) { 2007 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 2008 switch (BaseTy->getTag()) { 2009 case dwarf::DW_TAG_const_type: 2010 Mods |= ModifierOptions::Const; 2011 PO |= PointerOptions::Const; 2012 break; 2013 case dwarf::DW_TAG_volatile_type: 2014 Mods |= ModifierOptions::Volatile; 2015 PO |= PointerOptions::Volatile; 2016 break; 2017 case dwarf::DW_TAG_restrict_type: 2018 // Only pointer types be marked with __restrict. There is no known flag 2019 // for __restrict in LF_MODIFIER records. 2020 PO |= PointerOptions::Restrict; 2021 break; 2022 default: 2023 IsModifier = false; 2024 break; 2025 } 2026 if (IsModifier) 2027 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 2028 } 2029 2030 // Check if the inner type will use an LF_POINTER record. If so, the 2031 // qualifiers will go in the LF_POINTER record. This comes up for types like 2032 // 'int *const' and 'int *__restrict', not the more common cases like 'const 2033 // char *'. 2034 if (BaseTy) { 2035 switch (BaseTy->getTag()) { 2036 case dwarf::DW_TAG_pointer_type: 2037 case dwarf::DW_TAG_reference_type: 2038 case dwarf::DW_TAG_rvalue_reference_type: 2039 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2040 case dwarf::DW_TAG_ptr_to_member_type: 2041 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2042 default: 2043 break; 2044 } 2045 } 2046 2047 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2048 2049 // Return the base type index if there aren't any modifiers. For example, the 2050 // metadata could contain restrict wrappers around non-pointer types. 2051 if (Mods == ModifierOptions::None) 2052 return ModifiedTI; 2053 2054 ModifierRecord MR(ModifiedTI, Mods); 2055 return TypeTable.writeLeafType(MR); 2056 } 2057 2058 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2059 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2060 for (const DIType *ArgType : Ty->getTypeArray()) 2061 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2062 2063 // MSVC uses type none for variadic argument. 2064 if (ReturnAndArgTypeIndices.size() > 1 && 2065 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2066 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2067 } 2068 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2069 ArrayRef<TypeIndex> ArgTypeIndices = std::nullopt; 2070 if (!ReturnAndArgTypeIndices.empty()) { 2071 auto ReturnAndArgTypesRef = ArrayRef(ReturnAndArgTypeIndices); 2072 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2073 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2074 } 2075 2076 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2077 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2078 2079 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2080 2081 FunctionOptions FO = getFunctionOptions(Ty); 2082 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2083 ArgListIndex); 2084 return TypeTable.writeLeafType(Procedure); 2085 } 2086 2087 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2088 const DIType *ClassTy, 2089 int ThisAdjustment, 2090 bool IsStaticMethod, 2091 FunctionOptions FO) { 2092 // Lower the containing class type. 2093 TypeIndex ClassType = getTypeIndex(ClassTy); 2094 2095 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2096 2097 unsigned Index = 0; 2098 SmallVector<TypeIndex, 8> ArgTypeIndices; 2099 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2100 if (ReturnAndArgs.size() > Index) { 2101 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2102 } 2103 2104 // If the first argument is a pointer type and this isn't a static method, 2105 // treat it as the special 'this' parameter, which is encoded separately from 2106 // the arguments. 2107 TypeIndex ThisTypeIndex; 2108 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2109 if (const DIDerivedType *PtrTy = 2110 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2111 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2112 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2113 Index++; 2114 } 2115 } 2116 } 2117 2118 while (Index < ReturnAndArgs.size()) 2119 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2120 2121 // MSVC uses type none for variadic argument. 2122 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2123 ArgTypeIndices.back() = TypeIndex::None(); 2124 2125 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2126 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2127 2128 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2129 2130 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2131 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2132 return TypeTable.writeLeafType(MFR); 2133 } 2134 2135 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2136 unsigned VSlotCount = 2137 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2138 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2139 2140 VFTableShapeRecord VFTSR(Slots); 2141 return TypeTable.writeLeafType(VFTSR); 2142 } 2143 2144 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2145 switch (Flags & DINode::FlagAccessibility) { 2146 case DINode::FlagPrivate: return MemberAccess::Private; 2147 case DINode::FlagPublic: return MemberAccess::Public; 2148 case DINode::FlagProtected: return MemberAccess::Protected; 2149 case 0: 2150 // If there was no explicit access control, provide the default for the tag. 2151 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2152 : MemberAccess::Public; 2153 } 2154 llvm_unreachable("access flags are exclusive"); 2155 } 2156 2157 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2158 if (SP->isArtificial()) 2159 return MethodOptions::CompilerGenerated; 2160 2161 // FIXME: Handle other MethodOptions. 2162 2163 return MethodOptions::None; 2164 } 2165 2166 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2167 bool Introduced) { 2168 if (SP->getFlags() & DINode::FlagStaticMember) 2169 return MethodKind::Static; 2170 2171 switch (SP->getVirtuality()) { 2172 case dwarf::DW_VIRTUALITY_none: 2173 break; 2174 case dwarf::DW_VIRTUALITY_virtual: 2175 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2176 case dwarf::DW_VIRTUALITY_pure_virtual: 2177 return Introduced ? MethodKind::PureIntroducingVirtual 2178 : MethodKind::PureVirtual; 2179 default: 2180 llvm_unreachable("unhandled virtuality case"); 2181 } 2182 2183 return MethodKind::Vanilla; 2184 } 2185 2186 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2187 switch (Ty->getTag()) { 2188 case dwarf::DW_TAG_class_type: 2189 return TypeRecordKind::Class; 2190 case dwarf::DW_TAG_structure_type: 2191 return TypeRecordKind::Struct; 2192 default: 2193 llvm_unreachable("unexpected tag"); 2194 } 2195 } 2196 2197 /// Return ClassOptions that should be present on both the forward declaration 2198 /// and the defintion of a tag type. 2199 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2200 ClassOptions CO = ClassOptions::None; 2201 2202 // MSVC always sets this flag, even for local types. Clang doesn't always 2203 // appear to give every type a linkage name, which may be problematic for us. 2204 // FIXME: Investigate the consequences of not following them here. 2205 if (!Ty->getIdentifier().empty()) 2206 CO |= ClassOptions::HasUniqueName; 2207 2208 // Put the Nested flag on a type if it appears immediately inside a tag type. 2209 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2210 // here. That flag is only set on definitions, and not forward declarations. 2211 const DIScope *ImmediateScope = Ty->getScope(); 2212 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2213 CO |= ClassOptions::Nested; 2214 2215 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2216 // type only when it has an immediate function scope. Clang never puts enums 2217 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2218 // always in function, class, or file scopes. 2219 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2220 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2221 CO |= ClassOptions::Scoped; 2222 } else { 2223 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2224 Scope = Scope->getScope()) { 2225 if (isa<DISubprogram>(Scope)) { 2226 CO |= ClassOptions::Scoped; 2227 break; 2228 } 2229 } 2230 } 2231 2232 return CO; 2233 } 2234 2235 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2236 switch (Ty->getTag()) { 2237 case dwarf::DW_TAG_class_type: 2238 case dwarf::DW_TAG_structure_type: 2239 case dwarf::DW_TAG_union_type: 2240 case dwarf::DW_TAG_enumeration_type: 2241 break; 2242 default: 2243 return; 2244 } 2245 2246 if (const auto *File = Ty->getFile()) { 2247 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2248 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2249 2250 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2251 TypeTable.writeLeafType(USLR); 2252 } 2253 } 2254 2255 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2256 ClassOptions CO = getCommonClassOptions(Ty); 2257 TypeIndex FTI; 2258 unsigned EnumeratorCount = 0; 2259 2260 if (Ty->isForwardDecl()) { 2261 CO |= ClassOptions::ForwardReference; 2262 } else { 2263 ContinuationRecordBuilder ContinuationBuilder; 2264 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2265 for (const DINode *Element : Ty->getElements()) { 2266 // We assume that the frontend provides all members in source declaration 2267 // order, which is what MSVC does. 2268 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2269 // FIXME: Is it correct to always emit these as unsigned here? 2270 EnumeratorRecord ER(MemberAccess::Public, 2271 APSInt(Enumerator->getValue(), true), 2272 Enumerator->getName()); 2273 ContinuationBuilder.writeMemberType(ER); 2274 EnumeratorCount++; 2275 } 2276 } 2277 FTI = TypeTable.insertRecord(ContinuationBuilder); 2278 } 2279 2280 std::string FullName = getFullyQualifiedName(Ty); 2281 2282 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2283 getTypeIndex(Ty->getBaseType())); 2284 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2285 2286 addUDTSrcLine(Ty, EnumTI); 2287 2288 return EnumTI; 2289 } 2290 2291 //===----------------------------------------------------------------------===// 2292 // ClassInfo 2293 //===----------------------------------------------------------------------===// 2294 2295 struct llvm::ClassInfo { 2296 struct MemberInfo { 2297 const DIDerivedType *MemberTypeNode; 2298 uint64_t BaseOffset; 2299 }; 2300 // [MemberInfo] 2301 using MemberList = std::vector<MemberInfo>; 2302 2303 using MethodsList = TinyPtrVector<const DISubprogram *>; 2304 // MethodName -> MethodsList 2305 using MethodsMap = MapVector<MDString *, MethodsList>; 2306 2307 /// Base classes. 2308 std::vector<const DIDerivedType *> Inheritance; 2309 2310 /// Direct members. 2311 MemberList Members; 2312 // Direct overloaded methods gathered by name. 2313 MethodsMap Methods; 2314 2315 TypeIndex VShapeTI; 2316 2317 std::vector<const DIType *> NestedTypes; 2318 }; 2319 2320 void CodeViewDebug::clear() { 2321 assert(CurFn == nullptr); 2322 FileIdMap.clear(); 2323 FnDebugInfo.clear(); 2324 FileToFilepathMap.clear(); 2325 LocalUDTs.clear(); 2326 GlobalUDTs.clear(); 2327 TypeIndices.clear(); 2328 CompleteTypeIndices.clear(); 2329 ScopeGlobals.clear(); 2330 CVGlobalVariableOffsets.clear(); 2331 } 2332 2333 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2334 const DIDerivedType *DDTy) { 2335 if (!DDTy->getName().empty()) { 2336 Info.Members.push_back({DDTy, 0}); 2337 2338 // Collect static const data members with values. 2339 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2340 DINode::FlagStaticMember) { 2341 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2342 isa<ConstantFP>(DDTy->getConstant()))) 2343 StaticConstMembers.push_back(DDTy); 2344 } 2345 2346 return; 2347 } 2348 2349 // An unnamed member may represent a nested struct or union. Attempt to 2350 // interpret the unnamed member as a DICompositeType possibly wrapped in 2351 // qualifier types. Add all the indirect fields to the current record if that 2352 // succeeds, and drop the member if that fails. 2353 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2354 uint64_t Offset = DDTy->getOffsetInBits(); 2355 const DIType *Ty = DDTy->getBaseType(); 2356 bool FullyResolved = false; 2357 while (!FullyResolved) { 2358 switch (Ty->getTag()) { 2359 case dwarf::DW_TAG_const_type: 2360 case dwarf::DW_TAG_volatile_type: 2361 // FIXME: we should apply the qualifier types to the indirect fields 2362 // rather than dropping them. 2363 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2364 break; 2365 default: 2366 FullyResolved = true; 2367 break; 2368 } 2369 } 2370 2371 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2372 if (!DCTy) 2373 return; 2374 2375 ClassInfo NestedInfo = collectClassInfo(DCTy); 2376 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2377 Info.Members.push_back( 2378 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2379 } 2380 2381 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2382 ClassInfo Info; 2383 // Add elements to structure type. 2384 DINodeArray Elements = Ty->getElements(); 2385 for (auto *Element : Elements) { 2386 // We assume that the frontend provides all members in source declaration 2387 // order, which is what MSVC does. 2388 if (!Element) 2389 continue; 2390 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2391 Info.Methods[SP->getRawName()].push_back(SP); 2392 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2393 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2394 collectMemberInfo(Info, DDTy); 2395 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2396 Info.Inheritance.push_back(DDTy); 2397 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2398 DDTy->getName() == "__vtbl_ptr_type") { 2399 Info.VShapeTI = getTypeIndex(DDTy); 2400 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2401 Info.NestedTypes.push_back(DDTy); 2402 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2403 // Ignore friend members. It appears that MSVC emitted info about 2404 // friends in the past, but modern versions do not. 2405 } 2406 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2407 Info.NestedTypes.push_back(Composite); 2408 } 2409 // Skip other unrecognized kinds of elements. 2410 } 2411 return Info; 2412 } 2413 2414 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2415 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2416 // if a complete type should be emitted instead of a forward reference. 2417 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2418 !Ty->isForwardDecl(); 2419 } 2420 2421 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2422 // Emit the complete type for unnamed structs. C++ classes with methods 2423 // which have a circular reference back to the class type are expected to 2424 // be named by the front-end and should not be "unnamed". C unnamed 2425 // structs should not have circular references. 2426 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2427 // If this unnamed complete type is already in the process of being defined 2428 // then the description of the type is malformed and cannot be emitted 2429 // into CodeView correctly so report a fatal error. 2430 auto I = CompleteTypeIndices.find(Ty); 2431 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2432 report_fatal_error("cannot debug circular reference to unnamed type"); 2433 return getCompleteTypeIndex(Ty); 2434 } 2435 2436 // First, construct the forward decl. Don't look into Ty to compute the 2437 // forward decl options, since it might not be available in all TUs. 2438 TypeRecordKind Kind = getRecordKind(Ty); 2439 ClassOptions CO = 2440 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2441 std::string FullName = getFullyQualifiedName(Ty); 2442 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2443 FullName, Ty->getIdentifier()); 2444 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2445 if (!Ty->isForwardDecl()) 2446 DeferredCompleteTypes.push_back(Ty); 2447 return FwdDeclTI; 2448 } 2449 2450 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2451 // Construct the field list and complete type record. 2452 TypeRecordKind Kind = getRecordKind(Ty); 2453 ClassOptions CO = getCommonClassOptions(Ty); 2454 TypeIndex FieldTI; 2455 TypeIndex VShapeTI; 2456 unsigned FieldCount; 2457 bool ContainsNestedClass; 2458 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2459 lowerRecordFieldList(Ty); 2460 2461 if (ContainsNestedClass) 2462 CO |= ClassOptions::ContainsNestedClass; 2463 2464 // MSVC appears to set this flag by searching any destructor or method with 2465 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2466 // the members, however special member functions are not yet emitted into 2467 // debug information. For now checking a class's non-triviality seems enough. 2468 // FIXME: not true for a nested unnamed struct. 2469 if (isNonTrivial(Ty)) 2470 CO |= ClassOptions::HasConstructorOrDestructor; 2471 2472 std::string FullName = getFullyQualifiedName(Ty); 2473 2474 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2475 2476 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2477 SizeInBytes, FullName, Ty->getIdentifier()); 2478 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2479 2480 addUDTSrcLine(Ty, ClassTI); 2481 2482 addToUDTs(Ty); 2483 2484 return ClassTI; 2485 } 2486 2487 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2488 // Emit the complete type for unnamed unions. 2489 if (shouldAlwaysEmitCompleteClassType(Ty)) 2490 return getCompleteTypeIndex(Ty); 2491 2492 ClassOptions CO = 2493 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2494 std::string FullName = getFullyQualifiedName(Ty); 2495 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2496 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2497 if (!Ty->isForwardDecl()) 2498 DeferredCompleteTypes.push_back(Ty); 2499 return FwdDeclTI; 2500 } 2501 2502 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2503 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2504 TypeIndex FieldTI; 2505 unsigned FieldCount; 2506 bool ContainsNestedClass; 2507 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2508 lowerRecordFieldList(Ty); 2509 2510 if (ContainsNestedClass) 2511 CO |= ClassOptions::ContainsNestedClass; 2512 2513 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2514 std::string FullName = getFullyQualifiedName(Ty); 2515 2516 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2517 Ty->getIdentifier()); 2518 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2519 2520 addUDTSrcLine(Ty, UnionTI); 2521 2522 addToUDTs(Ty); 2523 2524 return UnionTI; 2525 } 2526 2527 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2528 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2529 // Manually count members. MSVC appears to count everything that generates a 2530 // field list record. Each individual overload in a method overload group 2531 // contributes to this count, even though the overload group is a single field 2532 // list record. 2533 unsigned MemberCount = 0; 2534 ClassInfo Info = collectClassInfo(Ty); 2535 ContinuationRecordBuilder ContinuationBuilder; 2536 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2537 2538 // Create base classes. 2539 for (const DIDerivedType *I : Info.Inheritance) { 2540 if (I->getFlags() & DINode::FlagVirtual) { 2541 // Virtual base. 2542 unsigned VBPtrOffset = I->getVBPtrOffset(); 2543 // FIXME: Despite the accessor name, the offset is really in bytes. 2544 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2545 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2546 ? TypeRecordKind::IndirectVirtualBaseClass 2547 : TypeRecordKind::VirtualBaseClass; 2548 VirtualBaseClassRecord VBCR( 2549 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2550 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2551 VBTableIndex); 2552 2553 ContinuationBuilder.writeMemberType(VBCR); 2554 MemberCount++; 2555 } else { 2556 assert(I->getOffsetInBits() % 8 == 0 && 2557 "bases must be on byte boundaries"); 2558 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2559 getTypeIndex(I->getBaseType()), 2560 I->getOffsetInBits() / 8); 2561 ContinuationBuilder.writeMemberType(BCR); 2562 MemberCount++; 2563 } 2564 } 2565 2566 // Create members. 2567 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2568 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2569 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2570 StringRef MemberName = Member->getName(); 2571 MemberAccess Access = 2572 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2573 2574 if (Member->isStaticMember()) { 2575 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2576 ContinuationBuilder.writeMemberType(SDMR); 2577 MemberCount++; 2578 continue; 2579 } 2580 2581 // Virtual function pointer member. 2582 if ((Member->getFlags() & DINode::FlagArtificial) && 2583 Member->getName().startswith("_vptr$")) { 2584 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2585 ContinuationBuilder.writeMemberType(VFPR); 2586 MemberCount++; 2587 continue; 2588 } 2589 2590 // Data member. 2591 uint64_t MemberOffsetInBits = 2592 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2593 if (Member->isBitField()) { 2594 uint64_t StartBitOffset = MemberOffsetInBits; 2595 if (const auto *CI = 2596 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2597 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2598 } 2599 StartBitOffset -= MemberOffsetInBits; 2600 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2601 StartBitOffset); 2602 MemberBaseType = TypeTable.writeLeafType(BFR); 2603 } 2604 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2605 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2606 MemberName); 2607 ContinuationBuilder.writeMemberType(DMR); 2608 MemberCount++; 2609 } 2610 2611 // Create methods 2612 for (auto &MethodItr : Info.Methods) { 2613 StringRef Name = MethodItr.first->getString(); 2614 2615 std::vector<OneMethodRecord> Methods; 2616 for (const DISubprogram *SP : MethodItr.second) { 2617 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2618 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2619 2620 unsigned VFTableOffset = -1; 2621 if (Introduced) 2622 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2623 2624 Methods.push_back(OneMethodRecord( 2625 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2626 translateMethodKindFlags(SP, Introduced), 2627 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2628 MemberCount++; 2629 } 2630 assert(!Methods.empty() && "Empty methods map entry"); 2631 if (Methods.size() == 1) 2632 ContinuationBuilder.writeMemberType(Methods[0]); 2633 else { 2634 // FIXME: Make this use its own ContinuationBuilder so that 2635 // MethodOverloadList can be split correctly. 2636 MethodOverloadListRecord MOLR(Methods); 2637 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2638 2639 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2640 ContinuationBuilder.writeMemberType(OMR); 2641 } 2642 } 2643 2644 // Create nested classes. 2645 for (const DIType *Nested : Info.NestedTypes) { 2646 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2647 ContinuationBuilder.writeMemberType(R); 2648 MemberCount++; 2649 } 2650 2651 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2652 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2653 !Info.NestedTypes.empty()); 2654 } 2655 2656 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2657 if (!VBPType.getIndex()) { 2658 // Make a 'const int *' type. 2659 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2660 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2661 2662 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2663 : PointerKind::Near32; 2664 PointerMode PM = PointerMode::Pointer; 2665 PointerOptions PO = PointerOptions::None; 2666 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2667 VBPType = TypeTable.writeLeafType(PR); 2668 } 2669 2670 return VBPType; 2671 } 2672 2673 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2674 // The null DIType is the void type. Don't try to hash it. 2675 if (!Ty) 2676 return TypeIndex::Void(); 2677 2678 // Check if we've already translated this type. Don't try to do a 2679 // get-or-create style insertion that caches the hash lookup across the 2680 // lowerType call. It will update the TypeIndices map. 2681 auto I = TypeIndices.find({Ty, ClassTy}); 2682 if (I != TypeIndices.end()) 2683 return I->second; 2684 2685 TypeLoweringScope S(*this); 2686 TypeIndex TI = lowerType(Ty, ClassTy); 2687 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2688 } 2689 2690 codeview::TypeIndex 2691 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2692 const DISubroutineType *SubroutineTy) { 2693 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2694 "this type must be a pointer type"); 2695 2696 PointerOptions Options = PointerOptions::None; 2697 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2698 Options = PointerOptions::LValueRefThisPointer; 2699 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2700 Options = PointerOptions::RValueRefThisPointer; 2701 2702 // Check if we've already translated this type. If there is no ref qualifier 2703 // on the function then we look up this pointer type with no associated class 2704 // so that the TypeIndex for the this pointer can be shared with the type 2705 // index for other pointers to this class type. If there is a ref qualifier 2706 // then we lookup the pointer using the subroutine as the parent type. 2707 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2708 if (I != TypeIndices.end()) 2709 return I->second; 2710 2711 TypeLoweringScope S(*this); 2712 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2713 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2714 } 2715 2716 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2717 PointerRecord PR(getTypeIndex(Ty), 2718 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2719 : PointerKind::Near32, 2720 PointerMode::LValueReference, PointerOptions::None, 2721 Ty->getSizeInBits() / 8); 2722 return TypeTable.writeLeafType(PR); 2723 } 2724 2725 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2726 // The null DIType is the void type. Don't try to hash it. 2727 if (!Ty) 2728 return TypeIndex::Void(); 2729 2730 // Look through typedefs when getting the complete type index. Call 2731 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2732 // emitted only once. 2733 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2734 (void)getTypeIndex(Ty); 2735 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2736 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2737 2738 // If this is a non-record type, the complete type index is the same as the 2739 // normal type index. Just call getTypeIndex. 2740 switch (Ty->getTag()) { 2741 case dwarf::DW_TAG_class_type: 2742 case dwarf::DW_TAG_structure_type: 2743 case dwarf::DW_TAG_union_type: 2744 break; 2745 default: 2746 return getTypeIndex(Ty); 2747 } 2748 2749 const auto *CTy = cast<DICompositeType>(Ty); 2750 2751 TypeLoweringScope S(*this); 2752 2753 // Make sure the forward declaration is emitted first. It's unclear if this 2754 // is necessary, but MSVC does it, and we should follow suit until we can show 2755 // otherwise. 2756 // We only emit a forward declaration for named types. 2757 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2758 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2759 2760 // Just use the forward decl if we don't have complete type info. This 2761 // might happen if the frontend is using modules and expects the complete 2762 // definition to be emitted elsewhere. 2763 if (CTy->isForwardDecl()) 2764 return FwdDeclTI; 2765 } 2766 2767 // Check if we've already translated the complete record type. 2768 // Insert the type with a null TypeIndex to signify that the type is currently 2769 // being lowered. 2770 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2771 if (!InsertResult.second) 2772 return InsertResult.first->second; 2773 2774 TypeIndex TI; 2775 switch (CTy->getTag()) { 2776 case dwarf::DW_TAG_class_type: 2777 case dwarf::DW_TAG_structure_type: 2778 TI = lowerCompleteTypeClass(CTy); 2779 break; 2780 case dwarf::DW_TAG_union_type: 2781 TI = lowerCompleteTypeUnion(CTy); 2782 break; 2783 default: 2784 llvm_unreachable("not a record"); 2785 } 2786 2787 // Update the type index associated with this CompositeType. This cannot 2788 // use the 'InsertResult' iterator above because it is potentially 2789 // invalidated by map insertions which can occur while lowering the class 2790 // type above. 2791 CompleteTypeIndices[CTy] = TI; 2792 return TI; 2793 } 2794 2795 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2796 /// and do this until fixpoint, as each complete record type typically 2797 /// references 2798 /// many other record types. 2799 void CodeViewDebug::emitDeferredCompleteTypes() { 2800 SmallVector<const DICompositeType *, 4> TypesToEmit; 2801 while (!DeferredCompleteTypes.empty()) { 2802 std::swap(DeferredCompleteTypes, TypesToEmit); 2803 for (const DICompositeType *RecordTy : TypesToEmit) 2804 getCompleteTypeIndex(RecordTy); 2805 TypesToEmit.clear(); 2806 } 2807 } 2808 2809 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2810 ArrayRef<LocalVariable> Locals) { 2811 // Get the sorted list of parameters and emit them first. 2812 SmallVector<const LocalVariable *, 6> Params; 2813 for (const LocalVariable &L : Locals) 2814 if (L.DIVar->isParameter()) 2815 Params.push_back(&L); 2816 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2817 return L->DIVar->getArg() < R->DIVar->getArg(); 2818 }); 2819 for (const LocalVariable *L : Params) 2820 emitLocalVariable(FI, *L); 2821 2822 // Next emit all non-parameters in the order that we found them. 2823 for (const LocalVariable &L : Locals) { 2824 if (!L.DIVar->isParameter()) { 2825 if (L.ConstantValue) { 2826 // If ConstantValue is set we will emit it as a S_CONSTANT instead of a 2827 // S_LOCAL in order to be able to represent it at all. 2828 const DIType *Ty = L.DIVar->getType(); 2829 APSInt Val(*L.ConstantValue); 2830 emitConstantSymbolRecord(Ty, Val, std::string(L.DIVar->getName())); 2831 } else { 2832 emitLocalVariable(FI, L); 2833 } 2834 } 2835 } 2836 } 2837 2838 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2839 const LocalVariable &Var) { 2840 // LocalSym record, see SymbolRecord.h for more info. 2841 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2842 2843 LocalSymFlags Flags = LocalSymFlags::None; 2844 if (Var.DIVar->isParameter()) 2845 Flags |= LocalSymFlags::IsParameter; 2846 if (Var.DefRanges.empty()) 2847 Flags |= LocalSymFlags::IsOptimizedOut; 2848 2849 OS.AddComment("TypeIndex"); 2850 TypeIndex TI = Var.UseReferenceType 2851 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2852 : getCompleteTypeIndex(Var.DIVar->getType()); 2853 OS.emitInt32(TI.getIndex()); 2854 OS.AddComment("Flags"); 2855 OS.emitInt16(static_cast<uint16_t>(Flags)); 2856 // Truncate the name so we won't overflow the record length field. 2857 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2858 endSymbolRecord(LocalEnd); 2859 2860 // Calculate the on disk prefix of the appropriate def range record. The 2861 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2862 // should be big enough to hold all forms without memory allocation. 2863 SmallString<20> BytePrefix; 2864 for (const auto &Pair : Var.DefRanges) { 2865 LocalVarDef DefRange = Pair.first; 2866 const auto &Ranges = Pair.second; 2867 BytePrefix.clear(); 2868 if (DefRange.InMemory) { 2869 int Offset = DefRange.DataOffset; 2870 unsigned Reg = DefRange.CVRegister; 2871 2872 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2873 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2874 // instead. In frames without stack realignment, $T0 will be the CFA. 2875 if (RegisterId(Reg) == RegisterId::ESP) { 2876 Reg = unsigned(RegisterId::VFRAME); 2877 Offset += FI.OffsetAdjustment; 2878 } 2879 2880 // If we can use the chosen frame pointer for the frame and this isn't a 2881 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2882 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2883 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2884 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2885 (bool(Flags & LocalSymFlags::IsParameter) 2886 ? (EncFP == FI.EncodedParamFramePtrReg) 2887 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2888 DefRangeFramePointerRelHeader DRHdr; 2889 DRHdr.Offset = Offset; 2890 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2891 } else { 2892 uint16_t RegRelFlags = 0; 2893 if (DefRange.IsSubfield) { 2894 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2895 (DefRange.StructOffset 2896 << DefRangeRegisterRelSym::OffsetInParentShift); 2897 } 2898 DefRangeRegisterRelHeader DRHdr; 2899 DRHdr.Register = Reg; 2900 DRHdr.Flags = RegRelFlags; 2901 DRHdr.BasePointerOffset = Offset; 2902 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2903 } 2904 } else { 2905 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2906 if (DefRange.IsSubfield) { 2907 DefRangeSubfieldRegisterHeader DRHdr; 2908 DRHdr.Register = DefRange.CVRegister; 2909 DRHdr.MayHaveNoName = 0; 2910 DRHdr.OffsetInParent = DefRange.StructOffset; 2911 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2912 } else { 2913 DefRangeRegisterHeader DRHdr; 2914 DRHdr.Register = DefRange.CVRegister; 2915 DRHdr.MayHaveNoName = 0; 2916 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2917 } 2918 } 2919 } 2920 } 2921 2922 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2923 const FunctionInfo& FI) { 2924 for (LexicalBlock *Block : Blocks) 2925 emitLexicalBlock(*Block, FI); 2926 } 2927 2928 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2929 /// lexical block scope. 2930 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2931 const FunctionInfo& FI) { 2932 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2933 OS.AddComment("PtrParent"); 2934 OS.emitInt32(0); // PtrParent 2935 OS.AddComment("PtrEnd"); 2936 OS.emitInt32(0); // PtrEnd 2937 OS.AddComment("Code size"); 2938 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2939 OS.AddComment("Function section relative address"); 2940 OS.emitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2941 OS.AddComment("Function section index"); 2942 OS.emitCOFFSectionIndex(FI.Begin); // Func Symbol 2943 OS.AddComment("Lexical block name"); 2944 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2945 endSymbolRecord(RecordEnd); 2946 2947 // Emit variables local to this lexical block. 2948 emitLocalVariableList(FI, Block.Locals); 2949 emitGlobalVariableList(Block.Globals); 2950 2951 // Emit lexical blocks contained within this block. 2952 emitLexicalBlockList(Block.Children, FI); 2953 2954 // Close the lexical block scope. 2955 emitEndSymbolRecord(SymbolKind::S_END); 2956 } 2957 2958 /// Convenience routine for collecting lexical block information for a list 2959 /// of lexical scopes. 2960 void CodeViewDebug::collectLexicalBlockInfo( 2961 SmallVectorImpl<LexicalScope *> &Scopes, 2962 SmallVectorImpl<LexicalBlock *> &Blocks, 2963 SmallVectorImpl<LocalVariable> &Locals, 2964 SmallVectorImpl<CVGlobalVariable> &Globals) { 2965 for (LexicalScope *Scope : Scopes) 2966 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2967 } 2968 2969 /// Populate the lexical blocks and local variable lists of the parent with 2970 /// information about the specified lexical scope. 2971 void CodeViewDebug::collectLexicalBlockInfo( 2972 LexicalScope &Scope, 2973 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2974 SmallVectorImpl<LocalVariable> &ParentLocals, 2975 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2976 if (Scope.isAbstractScope()) 2977 return; 2978 2979 // Gather information about the lexical scope including local variables, 2980 // global variables, and address ranges. 2981 bool IgnoreScope = false; 2982 auto LI = ScopeVariables.find(&Scope); 2983 SmallVectorImpl<LocalVariable> *Locals = 2984 LI != ScopeVariables.end() ? &LI->second : nullptr; 2985 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2986 SmallVectorImpl<CVGlobalVariable> *Globals = 2987 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2988 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2989 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2990 2991 // Ignore lexical scopes which do not contain variables. 2992 if (!Locals && !Globals) 2993 IgnoreScope = true; 2994 2995 // Ignore lexical scopes which are not lexical blocks. 2996 if (!DILB) 2997 IgnoreScope = true; 2998 2999 // Ignore scopes which have too many address ranges to represent in the 3000 // current CodeView format or do not have a valid address range. 3001 // 3002 // For lexical scopes with multiple address ranges you may be tempted to 3003 // construct a single range covering every instruction where the block is 3004 // live and everything in between. Unfortunately, Visual Studio only 3005 // displays variables from the first matching lexical block scope. If the 3006 // first lexical block contains exception handling code or cold code which 3007 // is moved to the bottom of the routine creating a single range covering 3008 // nearly the entire routine, then it will hide all other lexical blocks 3009 // and the variables they contain. 3010 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 3011 IgnoreScope = true; 3012 3013 if (IgnoreScope) { 3014 // This scope can be safely ignored and eliminating it will reduce the 3015 // size of the debug information. Be sure to collect any variable and scope 3016 // information from the this scope or any of its children and collapse them 3017 // into the parent scope. 3018 if (Locals) 3019 ParentLocals.append(Locals->begin(), Locals->end()); 3020 if (Globals) 3021 ParentGlobals.append(Globals->begin(), Globals->end()); 3022 collectLexicalBlockInfo(Scope.getChildren(), 3023 ParentBlocks, 3024 ParentLocals, 3025 ParentGlobals); 3026 return; 3027 } 3028 3029 // Create a new CodeView lexical block for this lexical scope. If we've 3030 // seen this DILexicalBlock before then the scope tree is malformed and 3031 // we can handle this gracefully by not processing it a second time. 3032 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 3033 if (!BlockInsertion.second) 3034 return; 3035 3036 // Create a lexical block containing the variables and collect the 3037 // lexical block information for the children. 3038 const InsnRange &Range = Ranges.front(); 3039 assert(Range.first && Range.second); 3040 LexicalBlock &Block = BlockInsertion.first->second; 3041 Block.Begin = getLabelBeforeInsn(Range.first); 3042 Block.End = getLabelAfterInsn(Range.second); 3043 assert(Block.Begin && "missing label for scope begin"); 3044 assert(Block.End && "missing label for scope end"); 3045 Block.Name = DILB->getName(); 3046 if (Locals) 3047 Block.Locals = std::move(*Locals); 3048 if (Globals) 3049 Block.Globals = std::move(*Globals); 3050 ParentBlocks.push_back(&Block); 3051 collectLexicalBlockInfo(Scope.getChildren(), 3052 Block.Children, 3053 Block.Locals, 3054 Block.Globals); 3055 } 3056 3057 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3058 const Function &GV = MF->getFunction(); 3059 assert(FnDebugInfo.count(&GV)); 3060 assert(CurFn == FnDebugInfo[&GV].get()); 3061 3062 collectVariableInfo(GV.getSubprogram()); 3063 3064 // Build the lexical block structure to emit for this routine. 3065 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3066 collectLexicalBlockInfo(*CFS, 3067 CurFn->ChildBlocks, 3068 CurFn->Locals, 3069 CurFn->Globals); 3070 3071 // Clear the scope and variable information from the map which will not be 3072 // valid after we have finished processing this routine. This also prepares 3073 // the map for the subsequent routine. 3074 ScopeVariables.clear(); 3075 3076 // Don't emit anything if we don't have any line tables. 3077 // Thunks are compiler-generated and probably won't have source correlation. 3078 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3079 FnDebugInfo.erase(&GV); 3080 CurFn = nullptr; 3081 return; 3082 } 3083 3084 // Find heap alloc sites and add to list. 3085 for (const auto &MBB : *MF) { 3086 for (const auto &MI : MBB) { 3087 if (MDNode *MD = MI.getHeapAllocMarker()) { 3088 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3089 getLabelAfterInsn(&MI), 3090 dyn_cast<DIType>(MD))); 3091 } 3092 } 3093 } 3094 3095 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 3096 llvm::Triple::ArchType::thumb; 3097 collectDebugInfoForJumpTables(MF, isThumb); 3098 3099 CurFn->Annotations = MF->getCodeViewAnnotations(); 3100 3101 CurFn->End = Asm->getFunctionEnd(); 3102 3103 CurFn = nullptr; 3104 } 3105 3106 // Usable locations are valid with non-zero line numbers. A line number of zero 3107 // corresponds to optimized code that doesn't have a distinct source location. 3108 // In this case, we try to use the previous or next source location depending on 3109 // the context. 3110 static bool isUsableDebugLoc(DebugLoc DL) { 3111 return DL && DL.getLine() != 0; 3112 } 3113 3114 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3115 DebugHandlerBase::beginInstruction(MI); 3116 3117 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3118 if (!Asm || !CurFn || MI->isDebugInstr() || 3119 MI->getFlag(MachineInstr::FrameSetup)) 3120 return; 3121 3122 // If the first instruction of a new MBB has no location, find the first 3123 // instruction with a location and use that. 3124 DebugLoc DL = MI->getDebugLoc(); 3125 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3126 for (const auto &NextMI : *MI->getParent()) { 3127 if (NextMI.isDebugInstr()) 3128 continue; 3129 DL = NextMI.getDebugLoc(); 3130 if (isUsableDebugLoc(DL)) 3131 break; 3132 } 3133 // FIXME: Handle the case where the BB has no valid locations. This would 3134 // probably require doing a real dataflow analysis. 3135 } 3136 PrevInstBB = MI->getParent(); 3137 3138 // If we still don't have a debug location, don't record a location. 3139 if (!isUsableDebugLoc(DL)) 3140 return; 3141 3142 maybeRecordLocation(DL, Asm->MF); 3143 } 3144 3145 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3146 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3147 *EndLabel = MMI->getContext().createTempSymbol(); 3148 OS.emitInt32(unsigned(Kind)); 3149 OS.AddComment("Subsection size"); 3150 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3151 OS.emitLabel(BeginLabel); 3152 return EndLabel; 3153 } 3154 3155 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3156 OS.emitLabel(EndLabel); 3157 // Every subsection must be aligned to a 4-byte boundary. 3158 OS.emitValueToAlignment(Align(4)); 3159 } 3160 3161 static StringRef getSymbolName(SymbolKind SymKind) { 3162 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3163 if (EE.Value == SymKind) 3164 return EE.Name; 3165 return ""; 3166 } 3167 3168 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3169 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3170 *EndLabel = MMI->getContext().createTempSymbol(); 3171 OS.AddComment("Record length"); 3172 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3173 OS.emitLabel(BeginLabel); 3174 if (OS.isVerboseAsm()) 3175 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3176 OS.emitInt16(unsigned(SymKind)); 3177 return EndLabel; 3178 } 3179 3180 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3181 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3182 // an extra copy of every symbol record in LLD. This increases object file 3183 // size by less than 1% in the clang build, and is compatible with the Visual 3184 // C++ linker. 3185 OS.emitValueToAlignment(Align(4)); 3186 OS.emitLabel(SymEnd); 3187 } 3188 3189 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3190 OS.AddComment("Record length"); 3191 OS.emitInt16(2); 3192 if (OS.isVerboseAsm()) 3193 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3194 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3195 } 3196 3197 void CodeViewDebug::emitDebugInfoForUDTs( 3198 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3199 #ifndef NDEBUG 3200 size_t OriginalSize = UDTs.size(); 3201 #endif 3202 for (const auto &UDT : UDTs) { 3203 const DIType *T = UDT.second; 3204 assert(shouldEmitUdt(T)); 3205 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3206 OS.AddComment("Type"); 3207 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3208 assert(OriginalSize == UDTs.size() && 3209 "getCompleteTypeIndex found new UDTs!"); 3210 emitNullTerminatedSymbolName(OS, UDT.first); 3211 endSymbolRecord(UDTRecordEnd); 3212 } 3213 } 3214 3215 void CodeViewDebug::collectGlobalVariableInfo() { 3216 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3217 GlobalMap; 3218 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3219 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3220 GV.getDebugInfo(GVEs); 3221 for (const auto *GVE : GVEs) 3222 GlobalMap[GVE] = &GV; 3223 } 3224 3225 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3226 for (const MDNode *Node : CUs->operands()) { 3227 const auto *CU = cast<DICompileUnit>(Node); 3228 for (const auto *GVE : CU->getGlobalVariables()) { 3229 const DIGlobalVariable *DIGV = GVE->getVariable(); 3230 const DIExpression *DIE = GVE->getExpression(); 3231 // Don't emit string literals in CodeView, as the only useful parts are 3232 // generally the filename and line number, which isn't possible to output 3233 // in CodeView. String literals should be the only unnamed GlobalVariable 3234 // with debug info. 3235 if (DIGV->getName().empty()) continue; 3236 3237 if ((DIE->getNumElements() == 2) && 3238 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3239 // Record the constant offset for the variable. 3240 // 3241 // A Fortran common block uses this idiom to encode the offset 3242 // of a variable from the common block's starting address. 3243 CVGlobalVariableOffsets.insert( 3244 std::make_pair(DIGV, DIE->getElement(1))); 3245 3246 // Emit constant global variables in a global symbol section. 3247 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3248 CVGlobalVariable CVGV = {DIGV, DIE}; 3249 GlobalVariables.emplace_back(std::move(CVGV)); 3250 } 3251 3252 const auto *GV = GlobalMap.lookup(GVE); 3253 if (!GV || GV->isDeclarationForLinker()) 3254 continue; 3255 3256 DIScope *Scope = DIGV->getScope(); 3257 SmallVector<CVGlobalVariable, 1> *VariableList; 3258 if (Scope && isa<DILocalScope>(Scope)) { 3259 // Locate a global variable list for this scope, creating one if 3260 // necessary. 3261 auto Insertion = ScopeGlobals.insert( 3262 {Scope, std::unique_ptr<GlobalVariableList>()}); 3263 if (Insertion.second) 3264 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3265 VariableList = Insertion.first->second.get(); 3266 } else if (GV->hasComdat()) 3267 // Emit this global variable into a COMDAT section. 3268 VariableList = &ComdatVariables; 3269 else 3270 // Emit this global variable in a single global symbol section. 3271 VariableList = &GlobalVariables; 3272 CVGlobalVariable CVGV = {DIGV, GV}; 3273 VariableList->emplace_back(std::move(CVGV)); 3274 } 3275 } 3276 } 3277 3278 void CodeViewDebug::collectDebugInfoForGlobals() { 3279 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3280 const DIGlobalVariable *DIGV = CVGV.DIGV; 3281 const DIScope *Scope = DIGV->getScope(); 3282 getCompleteTypeIndex(DIGV->getType()); 3283 getFullyQualifiedName(Scope, DIGV->getName()); 3284 } 3285 3286 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3287 const DIGlobalVariable *DIGV = CVGV.DIGV; 3288 const DIScope *Scope = DIGV->getScope(); 3289 getCompleteTypeIndex(DIGV->getType()); 3290 getFullyQualifiedName(Scope, DIGV->getName()); 3291 } 3292 } 3293 3294 void CodeViewDebug::emitDebugInfoForGlobals() { 3295 // First, emit all globals that are not in a comdat in a single symbol 3296 // substream. MSVC doesn't like it if the substream is empty, so only open 3297 // it if we have at least one global to emit. 3298 switchToDebugSectionForSymbol(nullptr); 3299 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3300 OS.AddComment("Symbol subsection for globals"); 3301 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3302 emitGlobalVariableList(GlobalVariables); 3303 emitStaticConstMemberList(); 3304 endCVSubsection(EndLabel); 3305 } 3306 3307 // Second, emit each global that is in a comdat into its own .debug$S 3308 // section along with its own symbol substream. 3309 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3310 const GlobalVariable *GV = cast<const GlobalVariable *>(CVGV.GVInfo); 3311 MCSymbol *GVSym = Asm->getSymbol(GV); 3312 OS.AddComment("Symbol subsection for " + 3313 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3314 switchToDebugSectionForSymbol(GVSym); 3315 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3316 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3317 emitDebugInfoForGlobal(CVGV); 3318 endCVSubsection(EndLabel); 3319 } 3320 } 3321 3322 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3323 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3324 for (const MDNode *Node : CUs->operands()) { 3325 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3326 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3327 getTypeIndex(RT); 3328 // FIXME: Add to global/local DTU list. 3329 } 3330 } 3331 } 3332 } 3333 3334 // Emit each global variable in the specified array. 3335 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3336 for (const CVGlobalVariable &CVGV : Globals) { 3337 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3338 emitDebugInfoForGlobal(CVGV); 3339 } 3340 } 3341 3342 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3343 const std::string &QualifiedName) { 3344 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3345 OS.AddComment("Type"); 3346 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3347 3348 OS.AddComment("Value"); 3349 3350 // Encoded integers shouldn't need more than 10 bytes. 3351 uint8_t Data[10]; 3352 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3353 CodeViewRecordIO IO(Writer); 3354 cantFail(IO.mapEncodedInteger(Value)); 3355 StringRef SRef((char *)Data, Writer.getOffset()); 3356 OS.emitBinaryData(SRef); 3357 3358 OS.AddComment("Name"); 3359 emitNullTerminatedSymbolName(OS, QualifiedName); 3360 endSymbolRecord(SConstantEnd); 3361 } 3362 3363 void CodeViewDebug::emitStaticConstMemberList() { 3364 for (const DIDerivedType *DTy : StaticConstMembers) { 3365 const DIScope *Scope = DTy->getScope(); 3366 3367 APSInt Value; 3368 if (const ConstantInt *CI = 3369 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3370 Value = APSInt(CI->getValue(), 3371 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3372 else if (const ConstantFP *CFP = 3373 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3374 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3375 else 3376 llvm_unreachable("cannot emit a constant without a value"); 3377 3378 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3379 getFullyQualifiedName(Scope, DTy->getName())); 3380 } 3381 } 3382 3383 static bool isFloatDIType(const DIType *Ty) { 3384 if (isa<DICompositeType>(Ty)) 3385 return false; 3386 3387 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3388 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3389 if (T == dwarf::DW_TAG_pointer_type || 3390 T == dwarf::DW_TAG_ptr_to_member_type || 3391 T == dwarf::DW_TAG_reference_type || 3392 T == dwarf::DW_TAG_rvalue_reference_type) 3393 return false; 3394 assert(DTy->getBaseType() && "Expected valid base type"); 3395 return isFloatDIType(DTy->getBaseType()); 3396 } 3397 3398 auto *BTy = cast<DIBasicType>(Ty); 3399 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3400 } 3401 3402 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3403 const DIGlobalVariable *DIGV = CVGV.DIGV; 3404 3405 const DIScope *Scope = DIGV->getScope(); 3406 // For static data members, get the scope from the declaration. 3407 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3408 DIGV->getRawStaticDataMemberDeclaration())) 3409 Scope = MemberDecl->getScope(); 3410 // For static local variables and Fortran, the scoping portion is elided 3411 // in its name so that we can reference the variable in the command line 3412 // of the VS debugger. 3413 std::string QualifiedName = 3414 (moduleIsInFortran() || (Scope && isa<DILocalScope>(Scope))) 3415 ? std::string(DIGV->getName()) 3416 : getFullyQualifiedName(Scope, DIGV->getName()); 3417 3418 if (const GlobalVariable *GV = 3419 dyn_cast_if_present<const GlobalVariable *>(CVGV.GVInfo)) { 3420 // DataSym record, see SymbolRecord.h for more info. Thread local data 3421 // happens to have the same format as global data. 3422 MCSymbol *GVSym = Asm->getSymbol(GV); 3423 SymbolKind DataSym = GV->isThreadLocal() 3424 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3425 : SymbolKind::S_GTHREAD32) 3426 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3427 : SymbolKind::S_GDATA32); 3428 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3429 OS.AddComment("Type"); 3430 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3431 OS.AddComment("DataOffset"); 3432 3433 uint64_t Offset = 0; 3434 if (CVGlobalVariableOffsets.contains(DIGV)) 3435 // Use the offset seen while collecting info on globals. 3436 Offset = CVGlobalVariableOffsets[DIGV]; 3437 OS.emitCOFFSecRel32(GVSym, Offset); 3438 3439 OS.AddComment("Segment"); 3440 OS.emitCOFFSectionIndex(GVSym); 3441 OS.AddComment("Name"); 3442 const unsigned LengthOfDataRecord = 12; 3443 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3444 endSymbolRecord(DataEnd); 3445 } else { 3446 const DIExpression *DIE = cast<const DIExpression *>(CVGV.GVInfo); 3447 assert(DIE->isConstant() && 3448 "Global constant variables must contain a constant expression."); 3449 3450 // Use unsigned for floats. 3451 bool isUnsigned = isFloatDIType(DIGV->getType()) 3452 ? true 3453 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3454 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3455 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3456 } 3457 } 3458 3459 void forEachJumpTableBranch( 3460 const MachineFunction *MF, bool isThumb, 3461 const std::function<void(const MachineJumpTableInfo &, const MachineInstr &, 3462 int64_t)> &Callback) { 3463 auto JTI = MF->getJumpTableInfo(); 3464 if (JTI && !JTI->isEmpty()) { 3465 #ifndef NDEBUG 3466 auto UsedJTs = llvm::SmallBitVector(JTI->getJumpTables().size()); 3467 #endif 3468 for (const auto &MBB : *MF) { 3469 // Search for indirect branches... 3470 const auto LastMI = MBB.getFirstTerminator(); 3471 if (LastMI != MBB.end() && LastMI->isIndirectBranch()) { 3472 if (isThumb) { 3473 // ... that directly use jump table operands. 3474 // NOTE: ARM uses pattern matching to lower its BR_JT SDNode to 3475 // machine instructions, hence inserting a JUMP_TABLE_DEBUG_INFO node 3476 // interferes with this process *but* the resulting pseudo-instruction 3477 // uses a Jump Table operand, so extract the jump table index directly 3478 // from that. 3479 for (const auto &MO : LastMI->operands()) { 3480 if (MO.isJTI()) { 3481 unsigned Index = MO.getIndex(); 3482 #ifndef NDEBUG 3483 UsedJTs.set(Index); 3484 #endif 3485 Callback(*JTI, *LastMI, Index); 3486 break; 3487 } 3488 } 3489 } else { 3490 // ... that have jump table debug info. 3491 // NOTE: The debug info is inserted as a JUMP_TABLE_DEBUG_INFO node 3492 // when lowering the BR_JT SDNode to an indirect branch. 3493 for (auto I = MBB.instr_rbegin(), E = MBB.instr_rend(); I != E; ++I) { 3494 if (I->isJumpTableDebugInfo()) { 3495 unsigned Index = I->getOperand(0).getImm(); 3496 #ifndef NDEBUG 3497 UsedJTs.set(Index); 3498 #endif 3499 Callback(*JTI, *LastMI, Index); 3500 break; 3501 } 3502 } 3503 } 3504 } 3505 } 3506 #ifndef NDEBUG 3507 assert(UsedJTs.all() && 3508 "Some of jump tables were not used in a debug info instruction"); 3509 #endif 3510 } 3511 } 3512 3513 void CodeViewDebug::discoverJumpTableBranches(const MachineFunction *MF, 3514 bool isThumb) { 3515 forEachJumpTableBranch( 3516 MF, isThumb, 3517 [this](const MachineJumpTableInfo &, const MachineInstr &BranchMI, 3518 int64_t) { requestLabelBeforeInsn(&BranchMI); }); 3519 } 3520 3521 void CodeViewDebug::collectDebugInfoForJumpTables(const MachineFunction *MF, 3522 bool isThumb) { 3523 forEachJumpTableBranch( 3524 MF, isThumb, 3525 [this, MF](const MachineJumpTableInfo &JTI, const MachineInstr &BranchMI, 3526 int64_t JumpTableIndex) { 3527 // For label-difference jump tables, find the base expression. 3528 // Otherwise the jump table uses an absolute address (so no base 3529 // is required). 3530 const MCSymbol *Base; 3531 uint64_t BaseOffset = 0; 3532 const MCSymbol *Branch = getLabelBeforeInsn(&BranchMI); 3533 JumpTableEntrySize EntrySize; 3534 switch (JTI.getEntryKind()) { 3535 case MachineJumpTableInfo::EK_Custom32: 3536 case MachineJumpTableInfo::EK_GPRel32BlockAddress: 3537 case MachineJumpTableInfo::EK_GPRel64BlockAddress: 3538 llvm_unreachable( 3539 "EK_Custom32, EK_GPRel32BlockAddress, and " 3540 "EK_GPRel64BlockAddress should never be emitted for COFF"); 3541 case MachineJumpTableInfo::EK_BlockAddress: 3542 // Each entry is an absolute address. 3543 EntrySize = JumpTableEntrySize::Pointer; 3544 Base = nullptr; 3545 break; 3546 case MachineJumpTableInfo::EK_Inline: 3547 case MachineJumpTableInfo::EK_LabelDifference32: 3548 case MachineJumpTableInfo::EK_LabelDifference64: 3549 // Ask the AsmPrinter. 3550 std::tie(Base, BaseOffset, Branch, EntrySize) = 3551 Asm->getCodeViewJumpTableInfo(JumpTableIndex, &BranchMI, Branch); 3552 break; 3553 } 3554 3555 CurFn->JumpTables.push_back( 3556 {EntrySize, Base, BaseOffset, Branch, 3557 MF->getJTISymbol(JumpTableIndex, MMI->getContext()), 3558 JTI.getJumpTables()[JumpTableIndex].MBBs.size()}); 3559 }); 3560 } 3561 3562 void CodeViewDebug::emitDebugInfoForJumpTables(const FunctionInfo &FI) { 3563 for (auto JumpTable : FI.JumpTables) { 3564 MCSymbol *JumpTableEnd = beginSymbolRecord(SymbolKind::S_ARMSWITCHTABLE); 3565 if (JumpTable.Base) { 3566 OS.AddComment("Base offset"); 3567 OS.emitCOFFSecRel32(JumpTable.Base, JumpTable.BaseOffset); 3568 OS.AddComment("Base section index"); 3569 OS.emitCOFFSectionIndex(JumpTable.Base); 3570 } else { 3571 OS.AddComment("Base offset"); 3572 OS.emitInt32(0); 3573 OS.AddComment("Base section index"); 3574 OS.emitInt16(0); 3575 } 3576 OS.AddComment("Switch type"); 3577 OS.emitInt16(static_cast<uint16_t>(JumpTable.EntrySize)); 3578 OS.AddComment("Branch offset"); 3579 OS.emitCOFFSecRel32(JumpTable.Branch, /*Offset=*/0); 3580 OS.AddComment("Table offset"); 3581 OS.emitCOFFSecRel32(JumpTable.Table, /*Offset=*/0); 3582 OS.AddComment("Branch section index"); 3583 OS.emitCOFFSectionIndex(JumpTable.Branch); 3584 OS.AddComment("Table section index"); 3585 OS.emitCOFFSectionIndex(JumpTable.Table); 3586 OS.AddComment("Entries count"); 3587 OS.emitInt32(JumpTable.TableSize); 3588 endSymbolRecord(JumpTableEnd); 3589 } 3590 } 3591